2.3 Sturm–Liouville Boundary Value Problems
Two-point boundary value problems, regular Sturm–Liouville problems in self-adjoint form, real simple eigenvalues and weighted orthogonality, the Sturm comparison and oscillation theorems, the Rayleigh quotient, eigenfunction expansions and the classical singular problems of Legendre and Bessel.
Separate variables in the heat equation $u_t=u_{xx}$ for a rod $0\le x\le L$ whose ends are kept at temperature zero: $u=X(x)T(t)$ gives $T'/T=X''/X=-\lambda$, and the spatial factor must satisfy $$X''+\lambda X=0,\qquad X(0)=X(L)=0.$$ This is not an initial value problem: the data sit at two points. For most $\lambda$ the only solution is $X\equiv0$; only for the special values $\lambda=(n\pi/L)^2$ is there a nonzero solution $\sin(n\pi x/L)$, and the temperature is then built as a superposition $\sum b_ne^{-\lambda_nt}\sin(n\pi x/L)$. The same picture appears for a vibrating string of variable density, the radial part of a drum, the quantum particle in a well and the buckling of a column. In each case one meets a Sturm–Liouville problem $$(p\,y')'+q\,y+\lambda\,r\,y=0,\qquad\text{boundary conditions at }x=a\text{ and }x=b,$$ and the questions are always the same: for which $\lambda$ are there nonzero solutions, what do those eigenfunctions look like, and can every reasonable function be expanded in them, as in a Fourier series?
This section answers these questions. We first see why boundary value problems behave so differently from the initial value problems of §2.1 (Theorem 2.21). Then we put the equation in self-adjoint form, which makes the eigenvalues real and the eigenfunctions orthogonal with respect to the weight $r$; prove the Sturm comparison theorem (larger coefficients mean faster oscillation) and deduce that the $n$-th eigenfunction has exactly $n-1$ interior zeros; estimate eigenvalues with the Rayleigh quotient; and expand functions in eigenfunctions, including the singular problems of Legendre and Bessel whose orthogonality was announced in §2.2. CSIR NET asks for eigenvalues under given boundary conditions, for the weight function of orthogonality, for numbers of zeros, for bounds on the first eigenvalue and, in Part C, for true/false statements about simplicity, sign and orthogonality — each one answered by a theorem below or by one of the counterexamples in the ⚠️ box.
Boundary Value Problems: When Existence Can Fail
Let $p_1,p_0,f$ be continuous on a closed bounded interval $[a,b]$. A (linear) two-point boundary value problem (BVP) is $$y''+p_1(x)y'+p_0(x)y=f(x)\quad(a\lt x\lt b),\qquad B_1[y]=\gamma_1,\quad B_2[y]=\gamma_2,$$ where each boundary form $B_i[y]=\alpha_{i1}y(a)+\alpha_{i2}y'(a)+\alpha_{i3}y(b)+\alpha_{i4}y'(b)$ has real coefficients and the two coefficient vectors are linearly independent. The problem is homogeneous if $f\equiv0$ and $\gamma_1=\gamma_2=0$. The conditions are separated if $B_1$ involves only the end point $a$ and $B_2$ only the end point $b$:
$$\alpha_1y(a)+\alpha_2y'(a)=0,\qquad\beta_1y(b)+\beta_2y'(b)=0,\qquad(\alpha_1,\alpha_2)\ne(0,0),\ (\beta_1,\beta_2)\ne(0,0).$$At an end point such a condition is called Dirichlet ($y=0$), Neumann ($y'=0$) or Robin (both coefficients nonzero). The periodic conditions $y(a)=y(b)$, $y'(a)=y'(b)$ are the standard example of conditions that are not separated.
For an initial value problem Theorem 2.1 guarantees exactly one solution, whatever the data. For a boundary value problem the two conditions are imposed at different points, and the question becomes one of linear algebra in the two-dimensional solution space:
- The following are equivalent: (a) $\Delta\ne0$; (b) the homogeneous BVP has only the trivial solution; (c) for every continuous $f$ and all $\gamma_1,\gamma_2$ the BVP has exactly one solution.
- If $\Delta=0$, the homogeneous BVP has a nontrivial solution, and for given $f,\gamma_1,\gamma_2$ the BVP has either no solution or infinitely many; in the second case the solutions are $y^*+y_h$, where $y^*$ is one of them and $y_h$ runs through the solutions of the homogeneous BVP.
Proof. The coefficients extend continuously to a slightly larger open interval, so Theorem 2.1 provides a particular solution $y_p$ of the equation on $[a,b]$, and by Theorem 2.2(3) every solution is $y=y_p+c_1y_1+c_2y_2$. Since the $B_i$ are linear, $y$ satisfies the boundary conditions if and only if $$\begin{pmatrix}B_1[y_1]&B_1[y_2]\\B_2[y_1]&B_2[y_2]\end{pmatrix}\begin{pmatrix}c_1\\c_2\end{pmatrix}=\begin{pmatrix}\gamma_1-B_1[y_p]\\\gamma_2-B_2[y_p]\end{pmatrix}.\qquad(\ast)$$ (1) If $\Delta\ne0$, $(\ast)$ has exactly one solution $(c_1,c_2)$ for every right-hand side: (a)⇒(c). (c)⇒(b) is the special case $f=0$, $\gamma=0$, whose solution $y\equiv0$ is then the only one. (b)⇒(a): if $\Delta=0$, the homogeneous system ($y_p=0$, $\gamma=0$) has a solution $(c_1,c_2)\ne(0,0)$, and $c_1y_1+c_2y_2$ is a nontrivial solution of the homogeneous BVP because $y_1,y_2$ are independent. (2) If $\Delta=0$, the matrix of $(\ast)$ is singular: the system has no solution or an affine line (or plane) of solutions, and the difference of two solutions solves the homogeneous BVP. Finally, another fundamental set is $(\tilde y_1,\tilde y_2)=(y_1,y_2)C$ with $\det C\ne0$; the matrix in $(\ast)$ is multiplied on the right by $C$, so $\Delta$ is multiplied by $\det C\ne0$. ∎
So the dichotomy for boundary value problems is exactly the one for a square linear system $Mc=d$: uniqueness for the homogeneous problem is equivalent to existence for every right-hand side. When uniqueness fails, existence requires a compatibility condition on $f$ — we find it below with Green's formula, and §2.4 turns it into the Fredholm alternative.
Regular Sturm–Liouville Problems and Self-Adjoint Form
For $p\in C^1[a,b]$ and $q\in C[a,b]$ the Sturm–Liouville operator is $L[y]=(p\,y')'+q\,y$. A regular Sturm–Liouville (SL) problem is
$$(p\,y')'+q\,y+\lambda\,r\,y=0\quad(a\lt x\lt b),\qquad\alpha_1y(a)+\alpha_2y'(a)=0,\quad\beta_1y(b)+\beta_2y'(b)=0,$$where $[a,b]$ is a closed bounded interval, $p\in C^1[a,b]$, $q,r\in C[a,b]$, $p\gt0$ and $r\gt0$ on all of $[a,b]$, and $\alpha_1,\alpha_2,\beta_1,\beta_2$ are real with $(\alpha_1,\alpha_2)\ne(0,0)\ne(\beta_1,\beta_2)$. The function $r$ is the weight. A (real or complex) number $\lambda$ is an eigenvalue if the problem has a solution $y\not\equiv0$, called an eigenfunction belonging to $\lambda$; the eigenfunctions together with $0$ form the eigenspace. If instead $p(a)=p(b)$ and the conditions are periodic, $y(a)=y(b)$, $y'(a)=y'(b)$, we speak of a periodic SL problem.
The heat-equation problem $X''+\lambda X=0$, $X(0)=X(L)=0$ is the regular SL problem with $p=r=1$, $q=0$. The point of the special form $(py')'$ — rather than $a_2y''+a_1y'$ — is the symmetry identity of Theorem 2.23. Every second-order equation can be brought into this form:
Proof. $p$ is positive and $p'=\dfrac{a_1}{a_2}\,p$ is continuous. Hence $(py')'=py''+p'y'=\dfrac{p}{a_2}\left(a_2y''+a_1y'\right)$, and multiplying the whole equation by the positive function $\mu=p/a_2$ gives $(py')'+\dfrac{a_0p}{a_2}y+\lambda\dfrac{bp}{a_2}y=0$. Multiplication by a nonvanishing function does not change the set of solutions. ∎
For example, $y''+2y'+\lambda y=0$ becomes $(e^{2x}y')'+\lambda e^{2x}y=0$ (weight $e^{2x}$), Hermite's equation $y''-2xy'+\lambda y=0$ becomes $(e^{-x^2}y')'+\lambda e^{-x^2}y=0$, and Legendre's equation $(1-x^2)y''-2xy'+\lambda y=0$ is already self-adjoint, with $p=1-x^2$, $r=1$.
- Lagrange's identity: $\ u\,L[v]-v\,L[u]=\dfrac{d}{dx}\Big[p\,(uv'-u'v)\Big]$.
- Green's formula: $\ \displaystyle\int_a^b\big(u\,L[v]-v\,L[u]\big)\,dx=\Big[p\,(uv'-u'v)\Big]_a^b.$
- Symmetry: if $u$ and $v$ both satisfy the separated conditions of Definition 2.11, or both satisfy the periodic conditions and $p(a)=p(b)$, the boundary term vanishes, so $\displaystyle\int_a^bu\,L[v]\,dx=\int_a^bv\,L[u]\,dx$.
Proof. (1) By the product rule (the two terms $u'v'$ cancel), $$\big[p(uv'-u'v)\big]'=p'(uv'-u'v)+p\,(uv''-u''v)$$ $$=u\,(p'v'+pv'')-v\,(p'u'+pu'')=u\,(pv')'-v\,(pu')',$$ and adding $u\cdot qv-v\cdot qu=0$ gives $uL[v]-vL[u]$. (2) Integrate (1). (3) At $x=a$ the vector $(\alpha_1,\alpha_2)\ne(0,0)$ satisfies $\alpha_1u(a)+\alpha_2u'(a)=0$ and $\alpha_1v(a)+\alpha_2v'(a)=0$, i.e. it lies in the kernel of $\begin{pmatrix}u(a)&u'(a)\\v(a)&v'(a)\end{pmatrix}$; so this matrix is singular and $u(a)v'(a)-u'(a)v(a)=0$. The same argument at $b$ kills the other half of the boundary term. In the periodic case $u,v,u',v'$ take equal values at $a$ and $b$, and $p(a)=p(b)$, so $p(uv'-u'v)$ has equal values at both ends. ∎
In the language of linear algebra, $L$ is a symmetric operator on the space of $C^2$ functions satisfying the boundary conditions, with respect to the inner product $\int_a^bu\bar v\,dx$. Symmetric matrices have real eigenvalues and orthogonal eigenvectors; the same proofs work here.
- every eigenvalue is real, and every eigenspace has a basis of real eigenfunctions;
- eigenfunctions $\varphi,\psi$ belonging to distinct eigenvalues $\lambda\ne\mu$ are orthogonal with respect to the weight $r$: $\displaystyle\int_a^br(x)\,\varphi(x)\,\overline{\psi(x)}\,dx=0$;
- for a regular problem (separated conditions) every eigenvalue is simple: any two eigenfunctions belonging to the same eigenvalue are proportional. (For periodic conditions an eigenspace can be two-dimensional.)
Proof. (1) Let $L[\varphi]=-\lambda r\varphi$ with $\varphi\not\equiv0$ satisfying the boundary conditions. Since $p,q,r$ and the boundary coefficients are real, $\bar\varphi$ satisfies the same conditions and $L[\bar\varphi]=-\bar\lambda r\bar\varphi$. Theorem 2.23(3) with $u=\bar\varphi$, $v=\varphi$ gives $$0=\int_a^b\big(\bar\varphi L[\varphi]-\varphi L[\bar\varphi]\big)dx=(\bar\lambda-\lambda)\int_a^br\,|\varphi|^2\,dx.$$ The integral is positive because $r\gt0$ and $|\varphi|^2$ is continuous and not identically zero; hence $\bar\lambda=\lambda$. Now $L$ has real coefficients and $\lambda$ is real, so $\operatorname{Re}\varphi$ and $\operatorname{Im}\varphi$ are again solutions satisfying the conditions; they span a space containing $\varphi$, so real eigenfunctions span each eigenspace. (2) With $L[\varphi]=-\lambda r\varphi$, $L[\bar\psi]=-\mu r\bar\psi$ ($\mu$ real), Theorem 2.23(3) gives $0=\int(\bar\psi L[\varphi]-\varphi L[\bar\psi])=(\mu-\lambda)\int r\varphi\bar\psi$, and $\mu\ne\lambda$. (3) Let $\varphi,\psi$ belong to the same $\lambda$. Both solve the same equation $y''+\frac{p'}{p}y'+\frac{q+\lambda r}{p}y=0$, whose coefficients are continuous because $p\gt0$. As in the proof of Theorem 2.23(3), the condition at $a$ forces $W(\varphi,\psi)(a)=\varphi(a)\psi'(a)-\varphi'(a)\psi(a)=0$, so $\varphi,\psi$ are dependent by Theorem 2.4. ∎
| conditions | eigenvalues | eigenfunctions |
|---|---|---|
| $y(0)=y(L)=0$ | $\left(\frac{n\pi}{L}\right)^2,\ n\ge1$ | $\sin\frac{n\pi x}{L}$ |
| $y'(0)=y'(L)=0$ | $\left(\frac{n\pi}{L}\right)^2,\ n\ge0$ | $\cos\frac{n\pi x}{L}$ (for $n=0$: $1$) |
| $y(0)=0,\ y'(L)=0$ | $\left(\frac{(2n-1)\pi}{2L}\right)^2,\ n\ge1$ | $\sin\frac{(2n-1)\pi x}{2L}$ |
| $y'(0)=0,\ y(L)=0$ | $\left(\frac{(2n-1)\pi}{2L}\right)^2,\ n\ge1$ | $\cos\frac{(2n-1)\pi x}{2L}$ |
| periodic on $[-L,L]$ | $\left(\frac{n\pi}{L}\right)^2,\ n\ge0$ | $1$; $\cos\frac{n\pi x}{L}$ and $\sin\frac{n\pi x}{L}$ for $n\ge1$ |
The Sturm Comparison Theorem and Oscillation of Eigenfunctions
In §2.1 we compared two solutions of one equation (Sturm separation). Now we compare solutions of two different equations. The tool is an identity found by Mauro Picone in 1909.
- Either $v$ has a zero in the open interval $(a,b)$, or $v$ is a constant multiple of $u$ on $[a,b]$. If $q_2\not\equiv q_1$ on $[a,b]$, the second alternative is impossible: $v$ vanishes somewhere in $(a,b)$.
- (The case $p_1=p_2=1$.) If $q_2\ge q_1$, then between two zeros of a nontrivial solution of $y''+q_1y=0$ there is a zero of every nontrivial solution of $y''+q_2y=0$, unless that solution is a multiple of the first one.
- (Zero spacing.) If $0\lt m^2\le Q(x)\le M^2$ on an interval $J$, consecutive zeros $x_1\lt x_2$ in $J$ of a nontrivial solution of $u''+Q(x)u=0$ satisfy $$\frac{\pi}{M}\le x_2-x_1\le\frac{\pi}{m},$$ and every closed subinterval of $J$ of length $\pi/m$ contains at least one zero.
Proof. Picone's identity. On any interval where $v\ne0$, $$\frac{d}{dx}\left[\frac{u}{v}\big(p_1u'v-p_2uv'\big)\right]=(q_2-q_1)u^2+(p_1-p_2)u'^2+p_2\left(u'-\frac{uv'}{v}\right)^2.$$ To check it, expand the left side as $\frac{d}{dx}\big[p_1uu'\big]-\frac{d}{dx}\Big[p_2\frac{u^2v'}{v}\Big]$. The first term is $p_1u'^2+u(p_1u')'=p_1u'^2-q_1u^2$. For the second, $$\frac{d}{dx}\Big[\frac{u^2}{v}\,p_2v'\Big]=\frac{2uu'v-u^2v'}{v^2}\,p_2v'+\frac{u^2}{v}(p_2v')'$$ $$=2p_2\frac{uu'v'}{v}-p_2\frac{u^2v'^2}{v^2}-q_2u^2.$$ Subtracting, the left side equals $$p_1u'^2+(q_2-q_1)u^2-2p_2\frac{uu'v'}{v}+p_2\frac{u^2v'^2}{v^2},$$ and since $p_2\big(u'-\frac{uv'}{v}\big)^2=p_2u'^2-2p_2\frac{uu'v'}{v}+p_2\frac{u^2v'^2}{v^2}$, this is the right side.
(1) Suppose $v$ has no zero in $(a,b)$. Fix small $\varepsilon\gt0$ and integrate Picone's identity over $[a+\varepsilon,b-\varepsilon]$. The bracket is $w=p_1uu'-p_2\,u\cdot\frac{u}{v}\,v'$. As $x\to a^+$: if $v(a)\ne0$, $w\to0$ because $u(a)=0$; if $v(a)=0$, then $v'(a)\ne0$ (zeros are simple, Theorem 2.10) and $\frac{u}{v}\to\frac{u'(a)}{v'(a)}$ by l'Hôpital's rule, so again $w\to0$ because of the factor $u$. The same holds at $b$. Letting $\varepsilon\to0$, $$\int_a^b\Big[(q_2-q_1)u^2+(p_1-p_2)u'^2+p_2\left(u'-\tfrac{uv'}{v}\right)^2\Big]dx=0.$$ All three terms are nonnegative and continuous on $(a,b)$, so each vanishes identically. From the last one, $\left(\frac uv\right)'=\frac{u'v-uv'}{v^2}=0$, so $u=cv$ on $(a,b)$ and, by continuity, on $[a,b]$; $c\ne0$ since $u\not\equiv0$. Thus $v=u/c$. Moreover $(q_2-q_1)u^2\equiv0$; since $u$ has only finitely many zeros in $[a,b]$ (Theorem 2.10) and $q_2-q_1$ is continuous, $q_2\equiv q_1$. So if $q_2\not\equiv q_1$, $v$ must vanish in $(a,b)$.
(2) is (1) with $p_1=p_2=1$. (3) Upper bound. Suppose $x_2-x_1\gt\pi/m$, and choose $c$ with $x_1\lt c\lt c+\pi/m\lt x_2$. Apply (2) to $\tilde u=\sin m(x-c)$, which solves $\tilde u''+m^2\tilde u=0$ and vanishes at $c$ and $c+\pi/m$, and to our $u$ (coefficient $Q\ge m^2$): $u$ has a zero in $(c,c+\pi/m)$ or is a multiple of $\tilde u$ and vanishes at $c$ — either way a zero strictly between $x_1$ and $x_2$, contradicting consecutiveness. The same comparison on any $[c,c+\pi/m]\subseteq J$ gives the last statement. Lower bound. Apply (2) to $u$ (zeros $x_1,x_2$, coefficient $Q$) and $\tilde v=\sin M(x-x_1)$ (coefficient $M^2\ge Q$): $\tilde v$ has a zero in $(x_1,x_2)$ or is a multiple of $u$. Its zeros are $x_1+k\pi/M$; if $x_2-x_1\lt\pi/M$ it has none in $(x_1,x_2)$, so $\tilde v=cu$ and $\tilde v(x_2)=0$, i.e. $x_2-x_1$ is a positive multiple of $\pi/M$ — a contradiction. ∎
The comparison theorem turns qualitative information about the coefficients into quantitative information about zeros. Applied to an eigenvalue problem, it explains the "one more zero for each eigenvalue" pattern you saw in the chapter teaser. The cleanest proof of the full oscillation theorem uses the Prüfer substitution $y=\rho\sin\theta$, $py'=\rho\cos\theta$, which replaces the linear equation by a first-order equation for the angle $\theta$ alone (Exercise III.8).
- the eigenvalues form an infinite, strictly increasing sequence $\lambda_1\lt\lambda_2\lt\lambda_3\lt\cdots$ of real numbers with $\lambda_n\to+\infty$ (in particular there are only finitely many negative eigenvalues);
- an eigenfunction $\varphi_n$ belonging to $\lambda_n$ has exactly $n-1$ zeros in the open interval $(a,b)$;
- between two consecutive zeros of $\varphi_n$ in $[a,b]$ there is a zero of $\varphi_{n+1}$;
- (Asymptotics, statement only.) $\dfrac{\lambda_n}{n^2}\to\dfrac{\pi^2}{\ell^2}$, where $\ell=\displaystyle\int_a^b\sqrt{\frac{r}{p}}\,dx$; for Dirichlet conditions at both ends, more precisely $\lambda_n=\dfrac{n^2\pi^2}{\ell^2}+O(1)$.
Proof (sketch). Since a nontrivial solution $y$ never has $y=y'=0$ at the same point, we can write $y=\rho\sin\theta$, $py'=\rho\cos\theta$ with $\rho\gt0$ and a continuous angle $\theta(x)$. Differentiating $\tan\theta=y/(py')$ and using $(py')'=-(q+\lambda r)y$ gives (Exercise III.8) the Prüfer equation $$\theta'=\frac{1}{p}\cos^2\theta+(q+\lambda r)\sin^2\theta.$$ Step 1 (zeros). $y(x)=0$ iff $\theta(x)\in\pi\mathbb{Z}$, and there $\theta'=1/p\gt0$: the angle crosses each multiple of $\pi$ only once and only upwards. So the number of zeros of $y$ in $(a,x)$ is the number of multiples of $\pi$ in the interval $(\theta(a),\theta(x))$. Step 2 (boundary conditions as angles). The condition at $a$ fixes $\theta(a)=\alpha\in[0,\pi)$ (with $\alpha_1\sin\alpha+\alpha_2\cos\alpha/p(a)=0$); the condition at $b$ says $\theta(b)\equiv\beta\pmod\pi$ for a fixed $\beta\in(0,\pi]$. Let $\theta(x;\lambda)$ be the solution with $\theta(a)=\alpha$. Step 3 (monotonicity). The right-hand side of the Prüfer equation increases with $\lambda$, so by a comparison argument for first-order equations $\lambda\mapsto\theta(b;\lambda)$ is continuous and strictly increasing; comparing with constant-coefficient equations one shows $\theta(b;\lambda)\to0^+$ as $\lambda\to-\infty$ and $\theta(b;\lambda)\to+\infty$ as $\lambda\to+\infty$. Step 4 (counting). Hence for each $n\ge1$ the equation $\theta(b;\lambda)=\beta+(n-1)\pi$ has exactly one root $\lambda_n$; these are all the eigenvalues, and they increase to $+\infty$: (1). For $\lambda=\lambda_n$ the angle runs from $\alpha\lt\pi$ to $\beta+(n-1)\pi\le n\pi$, crossing exactly $\pi,2\pi,\dots,(n-1)\pi$ in the interior: (2). (3) follows from Theorem 2.25(1) with $p_1=p_2=p$, $q_1=q+\lambda_nr$ and $q_2=q+\lambda_{n+1}r$, since $q_2-q_1=(\lambda_{n+1}-\lambda_n)r\gt0$. (4) comes from the Liouville transformation $\xi=\int_a^x\sqrt{r/p}$, which turns the equation into $z''+(\lambda-\tilde Q(\xi))z=0$ on an interval of length $\ell$. ∎
The Rayleigh Quotient
Multiplying the eigenvalue equation by the eigenfunction and integrating by parts expresses each eigenvalue as a ratio of "energies". This gives signs of eigenvalues at a glance and, through trial functions, sharp numerical bounds for $\lambda_1$.
- If $\varphi$ is a real eigenfunction belonging to $\lambda$, then $\lambda=R[\varphi]$.
- If $q\le0$ on $[a,b]$, $\alpha_1\alpha_2\le0$ and $\beta_1\beta_2\ge0$ (in particular for Dirichlet and Neumann conditions), then every eigenvalue is $\ge0$; and $\lambda=0$ is an eigenvalue only if $q\equiv0$ and $\alpha_1=\beta_1=0$ (Neumann at both ends), with constant eigenfunction.
- Minimum principle (proof sketched). $\lambda_1=\min R[u]$, the minimum being taken over all piecewise $C^1$ functions $u\not\equiv0$ that satisfy the Dirichlet conditions of the problem (if any) — where a condition is of Robin type, write the boundary term as $\frac{\beta_1}{\beta_2}p(b)u(b)^2-\frac{\alpha_1}{\alpha_2}p(a)u(a)^2$; Neumann ends impose nothing. The minimum is attained exactly at the multiples of $\varphi_1$. In particular $R[u]\ge\lambda_1$ for every admissible trial function.
Proof. (1) Multiply $(p\varphi')'+q\varphi=-\lambda r\varphi$ by $\varphi$ and integrate, using $\int_a^b\varphi\,(p\varphi')'\,dx=\big[p\varphi\varphi'\big]_a^b-\int_a^bp\varphi'^2\,dx$: $$\big[p\varphi\varphi'\big]_a^b-\int_a^bp\varphi'^2\,dx+\int_a^bq\varphi^2\,dx=-\lambda\int_a^br\varphi^2\,dx,$$ which rearranges to $\lambda=R[\varphi]$; the denominator is positive. (2) Look at the boundary term $-p(b)\varphi(b)\varphi'(b)+p(a)\varphi(a)\varphi'(a)$. If $\beta_2\ne0$, then $\varphi'(b)=-\frac{\beta_1}{\beta_2}\varphi(b)$ and the first part equals $p(b)\frac{\beta_1\beta_2}{\beta_2^2}\varphi(b)^2\ge0$; if $\beta_2=0$ then $\varphi(b)=0$ and it vanishes. Similarly $p(a)\varphi(a)\varphi'(a)=-p(a)\frac{\alpha_1\alpha_2}{\alpha_2^2}\varphi(a)^2\ge0$ or $0$. With $-q\ge0$ the numerator is $\ge\int p\varphi'^2\ge0$, so $\lambda\ge0$. If $\lambda=0$, then $\int p\varphi'^2=0$, so $\varphi$ is a nonzero constant; then $\int q\varphi^2=0$ forces $q\equiv0$, and a nonzero constant satisfies $\alpha_1\varphi(a)+\alpha_2\varphi'(a)=\alpha_1\varphi(a)=0$ only if $\alpha_1=0$; likewise $\beta_1=0$. (3) Sketch. Let $\{\varphi_n\}$ be the eigenfunctions, normalised by $\int r\varphi_n^2=1$, and expand an admissible $u=\sum c_n\varphi_n$ (Theorem 2.28). Integrating by parts as in (1), the numerator of $R[u]$ equals $\sum\lambda_nc_n^2$, while the denominator is $\sum c_n^2$ by Parseval's identity; hence $R[u]=\frac{\sum\lambda_nc_n^2}{\sum c_n^2}\ge\lambda_1$, with equality iff $c_n=0$ whenever $\lambda_n\gt\lambda_1$. Making this rigorous for piecewise $C^1$ trial functions is the subject of the variational methods of Chapter 6. ∎
Three consequences are used constantly. Monotonicity: increasing $p$, decreasing $q$, or (for positive eigenvalues) decreasing $r$ increases every Rayleigh quotient, hence increases $\lambda_1$. Domain monotonicity: shrinking the interval of a Dirichlet problem raises $\lambda_1$ (extend a trial function by $0$). Upper bounds: any trial function gives $\lambda_1\le R[u]$ — for $y''+\lambda y=0$, $y(0)=y(1)=0$, the trial $u=x(1-x)$ gives $R=\frac{1/3}{1/30}=10$, very close to $\lambda_1=\pi^2\approx9.8696$.
Eigenfunction Expansions and Singular Problems: Legendre and Bessel
- (Best approximation; Bessel's inequality.) For each $N$, among all sums $\sum_{n=1}^Na_n\varphi_n$ the distance $\big\|f-\sum_{n\le N}a_n\varphi_n\big\|_r$ is smallest exactly when $a_n=c_n$, and $$\sum_{n=1}^{\infty}c_n^2\,\|\varphi_n\|_r^2\le\|f\|_r^2.$$
- (Completeness; stated.) Equality holds — Parseval's identity — so $\big\|f-\sum_{n\le N}c_n\varphi_n\big\|_r\to0$. In particular, a continuous $f$ orthogonal to every $\varphi_n$ is identically zero.
- (Pointwise and uniform convergence; stated.) If $f$ is piecewise $C^1$ on $[a,b]$, the series converges to $\tfrac12\big(f(x^+)+f(x^-)\big)$ at every $x\in(a,b)$. If $f$ is continuous, piecewise $C^1$, and satisfies the conditions of Dirichlet type ($f(a)=0$ if $\alpha_2=0$, $f(b)=0$ if $\beta_2=0$), the series converges to $f$ absolutely and uniformly on $[a,b]$.
Proof of (1); sketch of (2)–(3). Write $S=\sum_{n\le N}a_n\varphi_n$. By orthogonality, $\|S\|_r^2=\sum a_n^2\|\varphi_n\|^2$ and $\langle f,S\rangle_r=\sum a_nc_n\|\varphi_n\|^2$, so $$\|f-S\|_r^2=\|f\|_r^2-2\sum_{n\le N}a_nc_n\|\varphi_n\|^2+\sum_{n\le N}a_n^2\|\varphi_n\|^2=\|f\|_r^2-\sum_{n\le N}c_n^2\|\varphi_n\|^2+\sum_{n\le N}(a_n-c_n)^2\|\varphi_n\|^2.$$ The last sum is $\ge0$ and vanishes only for $a_n=c_n$. Taking $a_n=c_n$ shows $\sum_{n\le N}c_n^2\|\varphi_n\|^2\le\|f\|^2$ for every $N$; let $N\to\infty$. For (2)–(3) one shows (§2.4) that the solution operator $f\mapsto u=\int G(\cdot,t)f(t)\,dt$ of the boundary value problem is a compact symmetric integral operator whose eigenfunctions are the $\varphi_n$; the Hilbert–Schmidt theorem for symmetric kernels (§7.2) then gives completeness. Pointwise convergence is proved by comparing the expansion with an ordinary Fourier series, as for $\sin nx$ and $\cos nx$. ∎
Bessel's inequality alone already forces $c_n\|\varphi_n\|\to0$. The faster the coefficients decay, the better the convergence: for the sine expansion on $[0,\pi]$, $x(\pi-x)$ (which satisfies the boundary conditions) has coefficients $\frac{8}{\pi n^3}$ for odd $n$, while $f\equiv1$ (which does not) has $\frac{4}{\pi n}$ and its partial sums overshoot by about $18\%$ near the ends — the Gibbs phenomenon. Interactive 2 shows both.
- Self-adjoint form first (Theorem 2.22): read off $p$, $q$ and the weight $r$. The weight is what orthogonality and the coefficients $c_n$ use.
- Sign of $\lambda$ from the Rayleigh quotient (Theorem 2.27(2)) — it often removes the cases $\lambda\lt0$ or $\lambda=0$ at once; otherwise treat $\lambda\lt0$, $\lambda=0$, $\lambda\gt0$ separately (all eigenvalues are real).
- Impose the condition at one end to reduce to one solution up to a factor; the condition at the other end gives the characteristic equation ($\sin\mu L=0$, $\tan\mu=-\mu/h$, …).
- Check with theory: eigenvalues increase to $\infty$, $\varphi_n$ has $n-1$ interior zeros, separated conditions give simple eigenvalues, $\lambda_n\approx(n\pi/\ell)^2$.
- Expansions: $c_n=\int f\varphi_nr\,/\int\varphi_n^2r$; uniform convergence when $f$ is continuous, piecewise $C^1$ and meets the Dirichlet-type conditions.
- Estimates: trial functions in the Rayleigh quotient give upper bounds for $\lambda_1$; comparison of coefficients (or Theorem 2.25) gives two-sided bounds.
Many important equations are Sturm–Liouville equations for which the hypotheses of Definition 2.11 fail at an end point: $p$ vanishes there, or the interval is infinite. The boundary condition is then replaced by a condition of boundedness or growth, and the whole theory survives as long as the boundary term of Green's formula still vanishes.
| name | self-adjoint equation | conditions | eigenpairs | weight |
|---|---|---|---|---|
| Legendre | $\big((1-x^2)y'\big)'+\lambda y=0$ | bounded on $(-1,1)$ | $n(n+1)$, $P_n$ | $1$ |
| Chebyshev | $\big(\sqrt{1-x^2}\,y'\big)'+\frac{\lambda}{\sqrt{1-x^2}}y=0$ | bounded on $(-1,1)$ | $n^2$, $T_n$ | $\frac{1}{\sqrt{1-x^2}}$ |
| Bessel | $(xy')'+\left(\lambda x-\frac{\nu^2}{x}\right)y=0$ | bounded at $0$, $y(1)=0$ | $j_{\nu,k}^2$, $J_\nu(j_{\nu,k}x)$ | $x$ |
| Hermite | $\big(e^{-x^2}y'\big)'+\lambda e^{-x^2}y=0$ | polynomial growth on $\mathbb{R}$ | $2n$, $H_n$ | $e^{-x^2}$ |
| Laguerre | $\big(xe^{-x}y'\big)'+\lambda e^{-x}y=0$ | polynomial growth on $(0,\infty)$ | $n$, $L_n$ | $e^{-x}$ |
Proof. Apply Green's formula on $[a+\varepsilon,b-\varepsilon]$ to $L[\varphi]=-\lambda r\varphi$, $L[\psi]=-\mu r\psi$: $$(\lambda-\mu)\int_{a+\varepsilon}^{b-\varepsilon}r\varphi\psi\,dx=\Big[p(\varphi\psi'-\varphi'\psi)\Big]_{a+\varepsilon}^{b-\varepsilon}.$$ The integral converges as $\varepsilon\to0$ (by Cauchy–Schwarz, since $\int r\varphi^2$, $\int r\psi^2$ are finite) and the right side tends to $0$; divide by $\lambda-\mu\ne0$. Legendre: $p=1-x^2$ vanishes at $\pm1$ while $P_m,P_n$ and their derivatives are bounded, so the limits are $0$; the eigenvalues $m(m+1)\ne n(n+1)$ for $m\ne n\ge0$, and by Theorem 2.20(c) of §2.2 these are the only values for which a bounded solution exists. For the norm, square the generating function $(1-2xt+t^2)^{-1/2}=\sum P_nt^n$ and integrate over $[-1,1]$; the cross terms vanish by orthogonality: $$\sum_{n\ge0}t^{2n}\int_{-1}^1P_n^2\,dx=\int_{-1}^1\frac{dx}{1-2xt+t^2}=\frac1t\log\frac{1+t}{1-t}=\sum_{n\ge0}\frac{2\,t^{2n}}{2n+1}\qquad(|t|\lt1),$$ and comparing coefficients gives $\int P_n^2=\frac2{2n+1}$. Bessel: $u_k(x)=J_\nu(j_kx)$ satisfies $(xu_k')'+\left(j_k^2x-\frac{\nu^2}{x}\right)u_k=0$ (substitute $s=j_kx$ in Bessel's equation and divide by $x$); near $0$, $u_k=O(x^\nu)$ and $u_k'=O(x^{\nu-1})$ for $\nu\gt0$ (both bounded for $\nu=0$), so $x\,(u_ku_m'-u_k'u_m)=O(x^{2\nu})\to0$ for $\nu\gt0$, and for $\nu=0$ it tends to $0$ because $u_k,u_k'$ are bounded and $x\to0$; at $x=1$ both vanish. Hermite and Laguerre: $p$ contains $e^{-x^2}$ or $xe^{-x}$, which kills every polynomial at the ends. The Bessel norm is derived in Exercise IV.5. ∎
- A BVP can have no solution with perfectly smooth data. $y''+y=x$, $y(0)=y(\pi)=0$ has none; $y''+y=\cos x$ with the same conditions has infinitely many. Existence for every $f$ $\iff$ uniqueness for the homogeneous problem (Theorem 2.21).
- Eigenvalues of $y''+\lambda y=0$ need not be positive. Neumann and periodic conditions have $\lambda=0$; the Robin condition $y'(1)=2y(1)$ (with $y(0)=0$) gives $\lambda_1\approx-3.667$. Positivity needs $q\le0$, $\alpha_1\alpha_2\le0$, $\beta_1\beta_2\ge0$ (Theorem 2.27).
- Periodic problems are not simple. $y''+\lambda y=0$ with periodic conditions on $[-\pi,\pi]$: $\lambda=n^2$ has the two independent eigenfunctions $\cos nx$, $\sin nx$. Simplicity needs separated conditions.
- Orthogonality needs the right weight. The eigenfunctions $x^{-1}\sin(n\pi\log x)$ of $x^2y''+3xy'+\lambda y=0$ on $[1,e]$ are orthogonal for $r=x$, not for $r=1$ ($\int_1^e\varphi_1\varphi_2\,dx\approx0.0553$).
- Without symmetry, eigenvalues can be complex. $y''+\lambda y=0$, $y(0)=0$, $y'(1)=2y'(0)$ is not self-adjoint; its eigenvalues are $\lambda=(2k\pi\pm i\log(2+\sqrt3))^2$, e.g. $\lambda\approx37.74\pm16.55i$ for $k=1$ (and $-(\log(2+\sqrt3))^2$ for $k=0$).
- $q\gt0$ allows negative eigenvalues. $y''+10y+\lambda y=0$, $y(0)=y(\pi)=0$ has $\lambda_n=n^2-10$: three negative eigenvalues. Finitely many, though — never infinitely many (Theorem 2.26(1)).
- "Exactly $n-1$ zeros" is for separated conditions. For the periodic problem both $\cos nx$ and $\sin nx$ have $2n$ zeros in $[-\pi,\pi)$; and zeros at the end points do not count as interior zeros.
- At a singular end, boundedness is the boundary condition. Legendre's equation has a nonzero solution for every $\lambda$; only for $\lambda=n(n+1)$ is one of them bounded on $(-1,1)$.
- The comparison theorem goes one way. $q_2\ge q_1$ everywhere implies faster oscillation of the second equation; more zeros of $v$ does not imply $q_2\ge q_1$ pointwise, and the conclusion fails if the inequality holds only on part of the interval.
- Mean-square is not uniform convergence. The sine series of $f\equiv1$ on $[0,\pi]$ converges in $L^2$ and pointwise on $(0,\pi)$, but equals $0$ at the ends and overshoots to about $1.179$ near them (Gibbs). Uniform convergence needs $f$ continuous, piecewise $C^1$ and meeting the Dirichlet-type conditions.
- The Rayleigh quotient bounds $\lambda_1$ from above only, and only for trial functions satisfying the Dirichlet conditions; a trial function violating $u(a)=0$ can give a value below $\lambda_1$ (e.g. $u\equiv1$ gives $R=0$ for the Dirichlet problem).
Exercises
- 1 For $y''+4y=0$ decide the number of solutions and find them: (a) $y(0)=0$, $y(\tfrac{\pi}{4})=1$; (b) $y(0)=0$, $y(\pi)=1$; (c) $y(0)=0$, $y(\pi)=0$.
- 2 For which real $\lambda$ does $y''+\lambda y=0$, $y(0)=0$, $y(1)=1$ have exactly one solution? What happens for the other values?
- 3 Decide whether $y''+y=f$, $y(0)=y(\pi)=0$ is solvable, and find all solutions when it is, for $f=1$, $f=\cos x$, $f=\sin2x$, $f=x-\tfrac{\pi}{2}$.
- 4 Solve $y''-y=1$, $y(0)=y(1)=0$, and explain without solving why the solution is unique.
- 5 For which $L\gt0$ does $y''+y=0$, $y'(0)=0$, $y(L)=0$ have a nontrivial solution?
- 6 Write in self-adjoint form and identify $p$, $q$, $r$: (a) Chebyshev's equation $(1-x^2)y''-xy'+\lambda y=0$ on $(-1,1)$; (b) Hermite's equation $y''-2xy'+\lambda y=0$; (c) $x^2y''+xy'+(\lambda x^2-\nu^2)y=0$ on $(0,1)$.
- 7 Prove Lagrange's identity $uL[v]-vL[u]=\big(p(uv'-u'v)\big)'$ for $L[y]=(py')'+qy$ by expanding both sides.
- 8 On $[0,1]$, decide whether $\int_0^1(uv''-u''v)\,dx=0$ for all $C^2$ functions $u,v$ satisfying (a) the antiperiodic conditions $y(0)=-y(1)$, $y'(0)=-y'(1)$; (b) $y(0)=0$, $y'(1)=2y'(0)$. For (b) test $u=x+\tfrac{x^2}{2}$, $v=3x+x^3$.
- 1 $y''+\lambda y=0$, $y(0)=0$, $y'(L)=0$: find all eigenpairs, treating $\lambda\le0$ explicitly.
- 2 $y''+\lambda y=0$, $y'(0)=y'(\pi)=0$.
- 3 $y''+\lambda y=0$, $y(-\pi)=y(\pi)$, $y'(-\pi)=y'(\pi)$, including the dimension of each eigenspace.
- 4 $y''+\lambda y=0$, $y(0)=0$, $y(1)+y'(1)=0$: show $\mu_n-(n-\tfrac12)\pi\to0$ for $\lambda_n=\mu_n^2$.
- 5 $y''+2y'+\lambda y=0$, $y(0)=y(\pi)=0$: eigenpairs and the orthogonality weight.
- 6 $x^2y''+xy'+\lambda y=0$, $y'(1)=0$, $y(e)=0$.
- 7 $y''+\lambda y=0$, $y(0)=0$, $y'(1)=y(1)$: show that $\lambda=0$ is an eigenvalue, that there are no negative eigenvalues, and find the positive ones approximately. Which hypothesis of Theorem 2.27(2) fails?
- 8 $\big((1+x)^2y'\big)'+\lambda y=0$, $y(0)=y(1)=0$.
- 1 Prove that every eigenvalue of a regular SL problem with $q\le0$ and Dirichlet conditions at both ends is strictly positive.
- 2 For $y''+\lambda y=0$, $y(0)=0$, $y'(1)=0$, compute $R[u]$ for $u=x(2-x)$ and for $u=x$, and compare with $\lambda_1$. Why is $u=x$ an admissible trial function although $u'(1)\ne0$?
- 3 Show that every nontrivial solution of $y''+(2+\sin x)y=0$ has infinitely many zeros in $[0,\infty)$, with consecutive zeros between $\pi/\sqrt3$ and $\pi$ apart.
- 4 Let $\varphi_n,\varphi_{n+1}$ be eigenfunctions of a regular SL problem. Prove that between two consecutive zeros of $\varphi_n$ in $[a,b]$ there is a zero of $\varphi_{n+1}$.
- 5 Prove directly that if $\varphi=u+iv$ is a complex eigenfunction of a regular SL problem, then $u$ and $v$ are proportional (so $\varphi$ is a complex multiple of a real eigenfunction).
- 6 With $\varphi_n$ as in Example 4, compute $\int_1^ex\,\varphi_1^2\,dx$ and deduce the normalised eigenfunctions.
- 7 Let $\lambda_1(L)$ be the first eigenvalue of $(py')'+qy+\lambda ry=0$ on $[0,L]$ with Dirichlet conditions. Show that $\lambda_1(L_2)\le\lambda_1(L_1)$ when $L_1\lt L_2$, and illustrate with $y''+\lambda y=0$.
- 8 Derive the Prüfer equation: if $y=\rho\sin\theta$, $py'=\rho\cos\theta$ for a nontrivial solution of $(py')'+(q+\lambda r)y=0$, show $\theta'=\frac1p\cos^2\theta+(q+\lambda r)\sin^2\theta$ and $\rho'=\left(\frac1p-q-\lambda r\right)\rho\sin\theta\cos\theta$.
- 1 Expand $f\equiv1$ on $[0,\pi]$ in the eigenfunctions $\sin nx$ of the Dirichlet problem. What does the series give at $x=0$? Does it converge uniformly on $[0,\pi]$?
- 2 Expand $f(x)=x$ on $[0,1]$ in the eigenfunctions of $y''+\lambda y=0$, $y(0)=0$, $y'(1)=0$, and decide whether the convergence is uniform.
- 3 Show $\int_{-1}^1xP_n\,dx=0$ for $n\ne1$ and compute it for $n=1$. Expand $x^3$ in Legendre polynomials.
- 4 Prove that $P_n$ has $n$ distinct zeros in $(-1,1)$. (Hint: if $P_n$ changes sign only at $x_1,\dots,x_k$ with $k\lt n$, consider $\int_{-1}^1P_n\prod_i(x-x_i)\,dx$.)
- 5 Prove $\int_0^1xJ_\nu(j_kx)J_\nu(j_mx)\,dx=0$ for $k\ne m$ (zeros of $J_\nu$, $\nu\ge0$), and show $\int_0^1xJ_\nu(j_kx)^2\,dx=\tfrac12J_\nu'(j_k)^2=\tfrac12J_{\nu+1}(j_k)^2$. (Hint for the norm: multiply Bessel's equation for $u=J_\nu(\mu x)$ by $2u'$ and integrate over $[0,1]$.)
- 6 Let $T_n(x)=\cos(n\arccos x)$. Show $T_n$ solves $(1-x^2)y''-xy'+n^2y=0$, and prove $\int_{-1}^1T_mT_n\frac{dx}{\sqrt{1-x^2}}=0$ ($m\ne n$) both by substitution and by Theorem 2.29.
- 7 Using Green's formula on $[-R,R]$, prove the orthogonality $\int_{\mathbb{R}}H_mH_ne^{-x^2}dx=0$ ($m\ne n$) of polynomial eigenfunctions of $(e^{-x^2}y')'+\lambda e^{-x^2}y=0$ with distinct eigenvalues.
- 8 The sine expansion of $x(\pi-x)$ on $[0,\pi]$ is $\frac8\pi\sum_{n\text{ odd}}\frac{\sin nx}{n^3}$. Use Parseval's identity to evaluate $\sum_{n\text{ odd}}\frac1{n^6}$.
E1. CSIR Part B GATE MA The eigenvalues of $y''+\lambda y=0$, $y'(0)=0$, $y(\pi)=0$ are (A) $n^2$, $n=1,2,\dots$; (B) $n^2$, $n=0,1,2,\dots$; (C) $\left(n-\tfrac12\right)^2$, $n=1,2,\dots$; (D) $(2n-1)^2$, $n=1,2,\dots$.
E2. CSIR Part B The polynomial eigenfunctions of Laguerre's equation $xy''+(1-x)y'+\lambda y=0$ on $(0,\infty)$ are orthogonal with respect to the weight (A) $1$; (B) $e^{-x}$; (C) $xe^{-x}$; (D) $e^{x}$.
E3. CSIR Part B Let $r$ be continuous on $[0,\pi]$ with $1\le r(x)\le4$, and let $\lambda_1$ be the smallest eigenvalue of $y''+\lambda r(x)y=0$, $y(0)=y(\pi)=0$. Then necessarily (A) $\tfrac14\le\lambda_1\le1$; (B) $1\le\lambda_1\le4$; (C) $\tfrac1{16}\le\lambda_1\le\tfrac14$; (D) $\lambda_1\le0$.
E4. CSIR Part C Consider $y''+\lambda y=0$ on $[0,2\pi]$ with $y(0)=y(2\pi)$, $y'(0)=y'(2\pi)$. Which of the following are true? (One or more options may be correct.) (A) Every eigenvalue is simple. (B) $0$ is an eigenvalue. (C) Eigenfunctions belonging to distinct eigenvalues are orthogonal on $[0,2\pi]$ with weight $1$. (D) The eigenvalues are exactly $n^2$, $n=0,1,2,\dots$
E5. CSIR Part C Let $p\in C^1[0,1]$, $q,r\in C[0,1]$, $p\gt0$, $r\gt0$, and let $\lambda_1\lt\lambda_2\lt\cdots$ be the eigenvalues of $(py')'+qy+\lambda ry=0$, $y(0)=y(1)=0$, with eigenfunctions $\varphi_n$. Which of the following are true? (One or more options may be correct.) (A) $\lambda_n\to\infty$. (B) $\varphi_3$ has exactly three zeros in $(0,1)$. (C) If $q\lt0$ on $[0,1]$, every $\lambda_n$ is positive. (D) Between two consecutive zeros of $\varphi_2$ in $[0,1]$ there is a zero of $\varphi_3$.
E6. CSIR Part C Let $q$ be continuous on $[0,\infty)$ with $4\le q(x)\le9$, and let $y\not\equiv0$ solve $y''+q(x)y=0$. Which of the following are true? (One or more options may be correct.) (A) $y$ has infinitely many zeros in $[0,\infty)$. (B) Any two consecutive zeros of $y$ are at least $\pi/3$ apart. (C) $y$ has at least two zeros in $[0,\pi]$. (D) $y$ has at most three zeros in $[0,\pi]$.
Suggested plan — 3 lectures × 60 min.
- Boundary value problems and self-adjointness. Start from separation of variables in the heat equation (two minutes, no PDE theory needed). Theorem 2.21 as "a $2\times2$ linear system"; Example 1 on the board — ask the class to guess the answer to (b) before computing. Definition 2.11; Theorem 2.22 with Hermite and Laguerre as quick drills; Lagrange's identity and Green's formula (Theorem 2.23) — stress the determinant argument for separated conditions. Theorem 2.24 with the matrix analogy.
- Eigenvalues and zeros. Example 2 (do Dirichlet fully, fill in the table together), Example 3 (draw $\tan\mu$ and $-\mu$). Theorem 2.25: prove Picone's identity carefully once — it is the hardest computation of the chapter — then part (3) and Example 5. Theorem 2.26 via the Prüfer angle; project Interactive 1, first Dirichlet, then Robin with $h$ moving below $-1/\pi$.
- Rayleigh, expansions, singular problems. Theorem 2.27 and Example 6 (have students compute the trial value for $x(1-x)$ themselves). Theorem 2.28 with the best-approximation proof; Interactive 2 (sine basis: hump, then ramp — ask why the error near $\pi$ refuses to shrink). Theorem 2.29: derive Legendre orthogonality and the norm from the generating function; Example 7. Finish with E4 and E5 live.
Discussion questions. (1) A BVP and an IVP for the same equation differ only in where the two conditions are imposed. Why does that change existence so drastically? (2) Which property of a symmetric matrix corresponds to "eigenfunctions are orthogonal with weight $r$", and where does $r$ come from in the matrix picture (generalised eigenproblem $Ax=\lambda Bx$)? (3) Why does a heavier string have lower eigenvalues, and what does the Rayleigh quotient say about a string that is heavier only near one end?
Misconceptions to watch for. Forgetting the cases $\lambda\le0$ (and missing $\lambda=0$ for Neumann problems); assuming eigenvalues of $y''+\lambda y=0$ are always positive; using the coefficient $p$ instead of the weight $r$; claiming simplicity for periodic problems; counting end-point zeros in "$n-1$ zeros"; treating a Rayleigh quotient as a lower bound; using a trial function that violates a Dirichlet condition.
Suggested homework. I(1, 3, 6, 8), II(1, 4, 5, 7), III(2, 3, 4, 8), IV(1, 3, 4, 8); Exam corner E2, E6.
Research question. How much does the spectrum of a non-uniform string reveal about its density? For $y''+\lambda\rho(x)^2y=0$, $y(0)=y(1)=0$, the first-order prediction of Theorem 2.26(4) is $\lambda_n\approx(n\pi/\ell)^2$ with $\ell=\int_0^1\rho\,dx$. How large is the error, and does it settle down to a constant that carries more information about $\rho$?
Why it matters. Instrument makers shape strings, bars and air columns so that their overtones fall on musical ratios; seismologists infer the layering of the Earth from its free oscillations; and in quantum mechanics the spectrum of $-y''+q(x)y$ is what experiments measure. Whether a spectrum determines the coefficient is the inverse Sturm–Liouville problem: Ambartsumian (1929) showed that the Neumann spectrum $\{n^2\pi^2\}$ forces $q\equiv0$, and Borg (1946) that in general one needs two spectra.
What to do.
- For $\rho=1+x$ ($\ell=\tfrac32$), compute $\lambda_1,\dots,\lambda_8$ by shooting (solve $y''=-\lambda(1+x)^2y$, $y(0)=0$, $y'(0)=1$ and find the roots of $\lambda\mapsto y(1;\lambda)$ with a bracketing root-finder), and tabulate $\lambda_n-(n\pi/\ell)^2$.
- The Liouville transformation $\xi=\int_0^x\rho$, $z=\rho^{1/2}y$ turns the equation into $z_{\xi\xi}+(\lambda-\tilde Q(\xi))z=0$ with $\tilde Q=\dfrac{(\rho^{1/2})_{\xi\xi}}{\rho^{1/2}}$. Compute $\tilde Q$ for $\rho=1+x$ and the mean value $\frac1\ell\int_0^\ell\tilde Q\,d\xi$. Compare it with the limit of your differences.
- Repeat for $\rho=1/(1+x)$, for which the eigenvalues are known exactly (it is Exercise II.8 in disguise). What is $\tilde Q$ there?
- Construct two different densities with the same $\ell$ and compare their spectra: which eigenvalues tell them apart?
Tools: Python with SciPy (solve_ivp, brentq) or any ODE solver; Interactive 1 for intuition about the Prüfer angle.
Hand in: A table of $\lambda_n$, $(n\pi/\ell)^2$ and their difference, your computation of the mean of $\tilde Q$, and a paragraph answering the research question.
Going further. Read about Kac's question "Can one hear the shape of a drum?" (1966) and its negative answer for planar domains (Gordon, Webb and Wolpert, 1992), and about the Gel'fand–Levitan method that reconstructs $q$ from spectral data.
🧪 Self-check quiz
Q1. The eigenvalues of $y''+\lambda y=0$, $y(0)=y(1)=0$ are
Q2. In self-adjoint form, Hermite's equation $y''-2xy'+\lambda y=0$ has weight function
Q3. For a regular Sturm–Liouville problem on $[a,b]$, an eigenfunction belonging to the fourth eigenvalue $\lambda_4$ has, in the open interval $(a,b)$,
Q4. For $y''+\lambda y=0$ on $[-\pi,\pi]$ with periodic conditions, the eigenspace of $\lambda=4$ has dimension
Q5. For $y''+\lambda y=0$, $y(0)=y(1)=0$, the Rayleigh quotient of the trial function $u=x(1-x)$ equals $10$. This shows that