6.3 Series Solutions, Legendre and Bessel Functions, Sturm Theory
Power series solutions at ordinary points, the Frobenius method at regular singular points, Legendre polynomials and Bessel functions with their recurrences and orthogonality, and Sturm's separation and comparison theorems with regular Sturm–Liouville problems — the variable-coefficient toolkit of GATE MA.
The constant-coefficient and Cauchy–Euler equations of §6.1 can be solved with exponentials and powers. Most second-order equations that arise in physics cannot: Legendre's equation comes from separating Laplace's equation in spherical coordinates, Bessel's equation from separating the wave or heat equation in a disc or a cylinder, Airy's equation from optics and quantum tunnelling. For such equations the next best thing to a closed formula is a power series whose coefficients obey a recurrence. Where the coefficients of the equation are analytic, ordinary power series work; at a regular singular point Frobenius' idea of multiplying by $x^r$ rescues the method; and two families of solutions — Legendre polynomials and Bessel functions — are important enough to have their own catalogue of identities. When no formula is available at all, Sturm's theorems still say a great deal about the zeros of solutions, and the Sturm–Liouville theory explains why the eigenfunctions of these problems are orthogonal.
GATE MA asks this material in a very recognisable way: classify a singular point and find the indicial roots; give a lower bound for the radius of convergence of a series solution; compute a coefficient of a series solution (a favourite NAT); evaluate $\int_{-1}^1 x^kP_mP_n\,dx$ or a value $P_n(0)$; simplify an expression with Bessel recurrences or $J_{\pm1/2}$; decide, in an MSQ, how many zeros a solution of $y''+q(x)y=0$ can have on an interval; and find the eigenvalues of a Sturm–Liouville problem. Throughout, the equation is written $y''+P(x)y'+Q(x)y=0$ or $P_0(x)y''+P_1(x)y'+P_2(x)y=0$, and "analytic at $x_0$" means "equal to a convergent power series in $x-x_0$ near $x_0$".
Power Series Solutions at Ordinary Points
Informally: at a regular singular point $P$ has at worst a simple pole and $Q$ at worst a double pole. Thus $x=0$ is a regular singular point of Bessel's equation $x^2y''+xy'+(x^2-\nu^2)y=0$ (here $xP=1$ and $x^2Q=x^2-\nu^2$), while it is an irregular singular point of $x^3y''+y=0$ ($x^2Q=1/x$ is not analytic).
Proof (idea; the full majorant proof is in the 📚 box). Formal part. Take $x_0=0$, write $P=\sum p_kx^k$, $Q=\sum q_kx^k$ and substitute $y=\sum a_kx^k$. The coefficient of $x^k$ in $y''+Py'+Qy$ is
$$(k+2)(k+1)a_{k+2}+\sum_{j=0}^{k}\Big[(j+1)a_{j+1}p_{k-j}+a_jq_{k-j}\Big],$$and setting it to zero expresses $a_{k+2}$ through $a_0,\dots,a_{k+1}$. So $a_0,a_1$ determine the whole series. Convergence. One compares $|a_k|$ with the coefficients of an explicitly solvable "majorant" equation (Cauchy's method) and shows that the series converges wherever the series of $P$ and $Q$ do. A slicker route uses complex analysis: the same existence theorem holds for $y''+P(z)y'+Q(z)y=0$ on the disc $|z|\lt R$ of $\mathbb{C}$, the solution is holomorphic there, and the Taylor series of a holomorphic function converges on every disc contained in its domain. Uniqueness and independence follow from the uniqueness theorem of §6.1: a solution is fixed by $y(x_0)$ and $y'(x_0)$, and the Wronskian of the two basic solutions at $x_0$ is $\det I=1\ne0$. ∎
- Locate and classify the singular points (zeros of $P_0$, also complex ones). The radius guaranteed about $x_0$ is the distance to the nearest one.
- Substitute $y=\sum a_kx^k$; shift indices so that every sum runs over the same power $x^k$; read off the recurrence (often a two-term recurrence $a_{k+2}=c(k)\,a_k$, which splits even and odd coefficients).
- Write the two basic solutions and, if possible, recognise them (an exponential, a polynomial, $1/(1+x^2)$, …). A polynomial solution appears exactly when the recurrence factor vanishes for some $k$.
- If the recurrence gives $\lim|a_{k+2}/a_k|=L$, the even and odd series converge for $|x|\lt1/\sqrt L$ — a useful check against step 1.
Regular Singular Points: the Frobenius Method
At a singular point a power series may not exist. For the Cauchy–Euler equation $x^2y''+axy'+by=0$ the solutions are powers $x^r$ with $r(r-1)+ar+b=0$, often non-integer or with a $\log x$. Frobenius' method combines the two ideas: near a regular singular point, look for $x^r$ times a power series.
- always: $y_1=x^{r_1}\sum_{k\ge0}a_kx^k$;
- if $r_1-r_2\notin\mathbb{Z}$: $y_2=x^{r_2}\sum_{k\ge0}b_kx^k$;
- if $r_1=r_2$: $y_2=y_1\log x+x^{r_1}\sum_{k\ge1}b_kx^k$ — a logarithm is unavoidable;
- if $r_1-r_2=N\in\mathbb{N}$: $y_2=c\,y_1\log x+x^{r_2}\sum_{k\ge0}b_kx^k$ for some constant $c$, which may be zero.
Proof (the algebra in full, convergence sketched). Substitute $y=\sum_{k\ge0}a_kx^{k+r}$ ($a_0=1$) into $x^2y''+x(xP)y'+(x^2Q)y=0$. The coefficient of $x^{k+r}$ is
$$F(r+k)\,a_k+\sum_{j=0}^{k-1}\Big[(r+j)\,p_{k-j}+q_{k-j}\Big]a_j=0 .\qquad(\ast)$$For $k=0$ this is $F(r)=0$: $r$ must be an exponent. Since $F(s)=(s-r_1)(s-r_2)$, we have $F(r_1+k)=k(k+r_1-r_2)\ne0$ for every $k\ge1$, so $(\ast)$ determines all $a_k$ for $r=r_1$: this is (1). For $r=r_2$ the same works unless $F(r_2+k)=0$ for some $k\ge1$, i.e. unless $r_1-r_2=k$ is a positive integer: this gives (2). In case (4), at $k=N$ the left side of $(\ast)$ reads $0\cdot b_N+(\text{known})$; if the known part happens to vanish, $b_N$ is free and $c=0$, otherwise no series of this form exists and a logarithm is needed.
For (3), let $r$ vary near $r_1$, define $a_k(r)$ by $(\ast)$ for $k\ge1$ (possible since $F(r+k)\ne0$ for $r$ near $r_1$) and set $Y(x,r)=\sum a_k(r)x^{k+r}$. Then all coefficients except the first cancel, and
$$x^2Y''+x(xP)Y'+(x^2Q)Y=F(r)\,x^r=(r-r_1)^2x^r .$$Differentiating with respect to $r$ and putting $r=r_1$, the right side gives $2(r-r_1)x^r+(r-r_1)^2x^r\log x=0$, so $\partial Y/\partial r\,(x,r_1)=y_1\log x+\sum_ka_k'(r_1)x^{k+r_1}$ is a solution; it is independent of $y_1$ because of the logarithm. A variant of this differentiation (applied to $(r-r_2)Y$) produces case (4). Convergence of all these series for $|x|\lt R$ is proved by the majorant method, as for Theorem 6.28 (📚 box). ∎
Two remarks that GATE uses constantly. At an ordinary point $p_0=q_0=0$, so the exponents are $0$ and $1$ — and no logarithm occurs (Exercise I.8). And the Cauchy–Euler equation $x^2y''+axy'+by=0$ is the case where $xP$ and $x^2Q$ are constants: all $p_k,q_k$ with $k\ge1$ vanish, the series stop after one term, and the solutions are $x^{r_1},x^{r_2}$ (or $x^{r_1},x^{r_1}\log x$ when $r_1=r_2$).
Legendre Polynomials
The first few are $P_0=1$, $P_1=x$, $P_2=\tfrac12(3x^2-1)$, $P_3=\tfrac12(5x^3-3x)$, $P_4=\tfrac18(35x^4-30x^2+3)$, $P_5=\tfrac18(63x^5-70x^3+15x)$. The recurrence comes from substituting $y=\sum a_kx^k$: the coefficient of $x^k$ is $(k+2)(k+1)a_{k+2}-\left[k(k+1)-\nu(\nu+1)\right]a_k$, and $k(k+1)-\nu(\nu+1)=(k-\nu)(k+\nu+1)$.
Proof. Let $u=(x^2-1)^n$, so $(x^2-1)u'=2nxu$. Differentiate this identity $n+1$ times with Leibniz' rule; only three terms survive on each side:
$$(x^2-1)u^{(n+2)}+2(n+1)x\,u^{(n+1)}+n(n+1)u^{(n)}=2nx\,u^{(n+1)}+2n(n+1)u^{(n)} .$$Rearranged, $(1-x^2)w''-2xw'+n(n+1)w=0$ for $w=u^{(n)}$: the right-hand side of Rodrigues' formula solves Legendre's equation, and it is a polynomial of degree $n$ with leading coefficient $\frac{2n(2n-1)\cdots(n+1)}{2^nn!}=\frac{(2n)!}{2^n(n!)^2}$. Polynomial solutions form a one-dimensional space, because only one of the two series of Definition 6.17 terminates (the other has $a_{k+2}/a_k\to1$ and nonzero coefficients for all large $k$). So Rodrigues' polynomial is a multiple of $P_n$, and it remains to check its value at $1$: writing $u=(x-1)^n(x+1)^n$, every term of $u^{(n)}$ in Leibniz' rule contains a factor $x-1$ except $n!\,(x+1)^n$, which equals $n!\,2^n$ at $x=1$. Parity: $u$ is even, so $u^{(n)}$ has the parity of $n$. Finally, putting $x=1$ in Legendre's equation gives $-2P_n'(1)+n(n+1)P_n(1)=0$. ∎
Proof. Write $G(x,t)=(1-2xt+t^2)^{-1/2}=(1-u)^{-1/2}$ with $u=2xt-t^2$ and expand by the binomial series: the coefficient $g_n(x)$ of $t^n$ is a polynomial in $x$ of degree $n$. Differentiating, $(1-2xt+t^2)\,G_t=(x-t)\,G$; inserting $G=\sum g_nt^n$ and comparing the coefficients of $t^n$,
$$(n+1)g_{n+1}-2xn\,g_n+(n-1)g_{n-1}=x\,g_n-g_{n-1},$$which is Bonnet's recurrence for $g_n$. Also $g_0=1$ and $g_1=x$ (from $G=1+xt+O(t^2)$). A similar computation with $G_x$, namely $(1-2xt+t^2)G_x=tG$, gives the derivative identities. To see that $g_n=P_n$, one checks by differentiation that $G$ satisfies $(1-x^2)G_{xx}-2xG_x+t\,\partial_t^2(tG)=0$; comparing coefficients shows that $g_n$ solves Legendre's equation with $\nu=n$. It is a polynomial, and $g_n(1)=1$ because $G(1,t)=(1-t)^{-1}=\sum t^n$; by the uniqueness in Theorem 6.30, $g_n=P_n$. ∎
- $\displaystyle\int_{-1}^{1}P_m(x)P_n(x)\,dx=\frac{2}{2n+1}\,\delta_{mn}$.
- $\int_{-1}^1q(x)P_n(x)\,dx=0$ for every polynomial $q$ of degree $\lt n$. Every polynomial $f$ of degree $N$ is $f=\sum_{n=0}^Nc_nP_n$ with $c_n=\frac{2n+1}{2}\int_{-1}^1fP_n\,dx$; more generally, for $f$ square-integrable on $[-1,1]$, $\sum c_nP_n$ converges to $f$ in the mean, and the partial sum of degree $N$ is the best polynomial approximation of degree $\le N$ in the $L^2$ sense.
- $P_n$ has exactly $n$ zeros in $(-1,1)$, all simple.
Proof. (1), $m\ne n$: Legendre's equation is $\left((1-x^2)y'\right)'+n(n+1)y=0$. Multiply the equation for $P_n$ by $P_m$, the one for $P_m$ by $P_n$, subtract and integrate:
$$\big[n(n+1)-m(m+1)\big]\int_{-1}^1P_mP_n\,dx=\Big[(1-x^2)\big(P_mP_n'-P_nP_m'\big)\Big]_{-1}^{1}=0,$$because $1-x^2$ vanishes at $\pm1$ — no boundary condition is needed. Since $n(n+1)\ne m(m+1)$, the integral is $0$. For $m=n$ use Rodrigues: with $u=(x^2-1)^n$, every $u^{(k)}$, $k\lt n$, vanishes at $\pm1$, so $n$ integrations by parts give $\int(u^{(n)})^2=(-1)^n\int u\,u^{(2n)}=(2n)!\int_{-1}^1(1-x^2)^n\,dx$. The last integral equals $\frac{2^{2n+1}(n!)^2}{(2n+1)!}$ (substitute $x=\cos\theta$, or use the Beta function), so $\int P_n^2=\frac{(2n)!\,2^{2n+1}(n!)^2}{(2n+1)!\,4^n(n!)^2}=\frac{2}{2n+1}$.
(2) $P_0,\dots,P_{n-1}$ have degrees $0,\dots,n-1$, so they span the polynomials of degree $\lt n$; apply (1). The coefficient formula follows by multiplying $f=\sum c_kP_k$ by $P_n$ and integrating; the mean convergence and best-approximation statements are the Hilbert-space facts of §8.3 for the orthogonal basis $\{P_n\}$ of $L^2[-1,1]$. (3) Let $x_1,\dots,x_k$ be the points of $(-1,1)$ where $P_n$ changes sign. If $k\lt n$, then $P_n(x)\prod_i(x-x_i)$ has constant sign on $(-1,1)$ and is not identically zero, so its integral is nonzero — contradicting (2). Hence $k\ge n$; as $\deg P_n=n$, the $n$ sign changes are all the zeros, and they are simple. ∎
Bessel Functions
The Frobenius recurrence for exponent $r=\nu$ is $k(k+2\nu)a_k=-a_{k-2}$ (with $a_1=0$), so only even $k$ occur and $a_{2m}=\frac{(-1)^m}{4^mm!\,(\nu+1)(\nu+2)\cdots(\nu+m)}a_0$; the choice $a_0=\frac{1}{2^\nu\Gamma(\nu+1)}$ produces $J_\nu$. In particular $J_0(x)=1-\frac{x^2}{4}+\frac{x^4}{64}-\cdots$ and $J_1(x)=\frac x2-\frac{x^3}{16}+\cdots$. The series converges for all $x$.
- $J_\nu$ solves Bessel's equation of order $\nu$ on $(0,\infty)$; so does $J_{-\nu}$.
- If $\nu\notin\mathbb{Z}$, then $J_\nu$ and $J_{-\nu}$ are linearly independent: the general solution is $c_1J_\nu+c_2J_{-\nu}$. This includes the half-integers, where $r_1-r_2=2\nu$ is an odd integer.
- If $\nu=n\in\{0,1,2,\dots\}$, then $J_{-n}=(-1)^nJ_n$, and a second solution $Y_n$ (the Bessel function of the second kind) contains the term $\frac2\pi J_n(x)\log x$ and is unbounded as $x\to0^+$. A solution bounded near $0$ is a multiple of $J_n$.
- $J_{1/2}(x)=\sqrt{\dfrac{2}{\pi x}}\,\sin x$ and $J_{-1/2}(x)=\sqrt{\dfrac{2}{\pi x}}\,\cos x$.
Proof. (1) The equation only involves $\nu^2$, and the series with $\nu$ replaced by $-\nu$ satisfies the same recurrence. (2) Near $0$, $J_\nu\sim\frac{(x/2)^\nu}{\Gamma(\nu+1)}\to0$ while $J_{-\nu}\sim\frac{(x/2)^{-\nu}}{\Gamma(1-\nu)}$ is unbounded ($\Gamma(1-\nu)$ is finite and nonzero for $\nu\notin\mathbb{Z}$), so neither is a multiple of the other. (3) For $\nu=-n$ the terms with $m\lt n$ vanish because $1/\Gamma(m-n+1)=0$; putting $m=j+n$,
$$J_{-n}(x)=\sum_{j\ge0}\frac{(-1)^{j+n}}{(j+n)!\,j!}\left(\frac x2\right)^{2j+n}=(-1)^nJ_n(x).$$So $J_{-n}$ gives nothing new, and the exponents $\pm n$ differ by the integer $2n$ (or coincide, for $n=0$); the second solution of Theorem 6.29 contains a logarithm, which makes it unbounded at $0$. Since $J_n$ is bounded, a solution bounded near $0$ cannot contain $Y_n$. (4) Using $\Gamma(m+\tfrac32)=\frac{(2m+1)!\sqrt\pi}{2^{2m+1}m!}$ (induction from $\Gamma(\tfrac32)=\tfrac{\sqrt\pi}{2}$),
$$J_{1/2}(x)=\sum_{m\ge0}\frac{(-1)^m\,2^{2m+1}}{(2m+1)!\,\sqrt\pi}\left(\frac x2\right)^{2m+1/2}=\sqrt{\frac{2}{\pi x}}\sum_{m\ge0}\frac{(-1)^mx^{2m+1}}{(2m+1)!}=\sqrt{\frac{2}{\pi x}}\sin x,$$and $J_{-1/2}$ in the same way (or from Example 4, where $x^{-1/2}\cos x$ and $x^{-1/2}\sin x$ were found to be solutions). ∎
Proof. Differentiate termwise (the series converge uniformly on compact subsets of $(0,\infty)$). For the first identity, $x^\nu J_\nu=\sum_m\frac{(-1)^mx^{2m+2\nu}}{2^{2m+\nu}m!\,\Gamma(m+\nu+1)}$, and since $(2m+2\nu)/\Gamma(m+\nu+1)=2/\Gamma(m+\nu)$,
$$\frac{d}{dx}\big(x^\nu J_\nu\big)=\sum_m\frac{(-1)^mx^{2m+2\nu-1}}{2^{2m+\nu-1}m!\,\Gamma(m+\nu)}=x^\nu\sum_m\frac{(-1)^m}{m!\,\Gamma(m+\nu)}\left(\frac x2\right)^{2m+\nu-1}=x^\nu J_{\nu-1}.$$For the second, $x^{-\nu}J_\nu=\sum_m\frac{(-1)^mx^{2m}}{2^{2m+\nu}m!\,\Gamma(m+\nu+1)}$; the $m=0$ term is constant, and with $m=j+1$ the derivative is $-\sum_j\frac{(-1)^jx^{2j+1}}{2^{2j+\nu+1}j!\,\Gamma(j+\nu+2)}=-x^{-\nu}J_{\nu+1}$. Expanding the products: $J_\nu'+\frac\nu xJ_\nu=J_{\nu-1}$ and $J_\nu'-\frac\nu xJ_\nu=-J_{\nu+1}$; subtracting and adding gives the two three-term relations, and $\nu=0$ in the second identity gives $J_0'=-J_1$. The integrals restate the two identities. The generating function: multiply the series $e^{xt/2}=\sum_j\frac{(xt/2)^j}{j!}$ and $e^{-x/(2t)}=\sum_m\frac{(-x/(2t))^m}{m!}$; the coefficient of $t^n$ ($n\ge0$) is $\sum_m\frac{(-1)^m(x/2)^{2m+n}}{m!\,(m+n)!}=J_n(x)$, and $t\mapsto-1/t$ handles $n\lt0$. ∎
Proof. $u(x)=J_\nu(\alpha x)$ satisfies $x^2u''+xu'+(\alpha^2x^2-\nu^2)u=0$ (chain rule), i.e. $(xu')'+\left(\alpha^2x-\frac{\nu^2}{x}\right)u=0$; similarly for $v(x)=J_\nu(\beta x)$ with $\beta$. Multiply the first by $v$, the second by $u$, subtract and integrate over $[\varepsilon,1]$:
$$(\beta^2-\alpha^2)\int_\varepsilon^1xuv\,dx=\Big[x\,(u'v-uv')\Big]_\varepsilon^1 .$$At $x=1$ both $u$ and $v$ vanish. As $\varepsilon\to0^+$: $u=ax^\nu(1+O(x^2))$, $v=bx^\nu(1+O(x^2))$, and the leading terms of $x(u'v-uv')$ cancel, leaving $O(x^{2\nu+2})\to0$ because $\nu\gt-1$. Hence the integral vanishes. For the norm (with $\nu\ge0$), a direct computation using the equation gives $\frac{d}{dx}\Big[(xu')^2+(\alpha^2x^2-\nu^2)u^2\Big]=2\alpha^2xu^2$. Integrate over $[0,1]$: at $x=1$, $u=0$ and $xu'=\alpha J_\nu'(\alpha)$; at $x=0$ the bracket is $0$. So $2\alpha^2\int_0^1xu^2=\alpha^2J_\nu'(\alpha)^2$, and $J_\nu'(\alpha)=\frac\nu\alpha J_\nu(\alpha)-J_{\nu+1}(\alpha)=-J_{\nu+1}(\alpha)$ by Theorem 6.34. Zeros are simple because a solution with $u(\alpha)=u'(\alpha)=0$ at $\alpha\gt0$ vanishes identically. The statements about infinitely many zeros and the gaps follow from Sturm's comparison theorem applied to the normal form (Theorem 6.39(4) below). ∎
Sturm Theory and Sturm–Liouville Problems
Bessel functions oscillate like damped sines, Legendre polynomials have all their zeros in $(-1,1)$ — and yet we proved these facts with special tricks. Sturm's theorems give the general principle: the zeros of solutions of $y''+q(x)y=0$ are controlled by the size of $q$.
Proof. The Wronskian $W=y_1y_2'-y_1'y_2$ never vanishes on $I$ (Abel's formula, §6.1), so it has a constant sign. A common zero would make $W=0$ there, and a zero $a$ of $y_1$ with $y_1'(a)=0$ would make $y_1\equiv0$ by uniqueness. Let $a\lt b$ be consecutive zeros of $y_1$, say $y_1\gt0$ on $(a,b)$. Then $y_1'(a)\gt0$ and $y_1'(b)\lt0$ (simple zeros, and $y_1$ is positive in between). At the zeros, $W(a)=-y_1'(a)y_2(a)$ and $W(b)=-y_1'(b)y_2(b)$; since $W(a)$ and $W(b)$ have the same sign while $y_1'(a)$ and $y_1'(b)$ have opposite signs, $y_2(a)$ and $y_2(b)$ have opposite signs, and $y_2$ vanishes in $(a,b)$ by the intermediate value theorem. If $y_2$ had two zeros in $(a,b)$, the argument with the roles exchanged would give a zero of $y_1$ between them — contradicting consecutiveness. ∎
Proof. With $e=\exp(-\frac12\int P)$ we have $e'=-\frac12Pe$ and $e''=\left(\frac14P^2-\frac12P'\right)e$. Then $y'=e\left(u'-\frac12Pu\right)$ and $y''=e\left(u''-Pu'+\left(\frac14P^2-\frac12P'\right)u\right)$, and $y''+Py'+Qy=e\left(u''+\left(Q-\frac14P^2-\frac12P'\right)u\right)$. The factor $e$ is positive. For Bessel: $\frac14P^2+\frac12P'=\frac1{4x^2}-\frac1{2x^2}=-\frac1{4x^2}$. ∎
Proof. We may take $a,b$ consecutive zeros of $y$ with $y\gt0$ on $(a,b)$ (shrinking $[a,b]$ only strengthens the conclusion), so $y'(a)\gt0\gt y'(b)$. Suppose $z$ has no zero in $(a,b)$; replacing $z$ by $-z$, $z\gt0$ on $(a,b)$ and $z(a),z(b)\ge0$. Then
$$\big(y'z-yz'\big)'=y''z-yz''=(q_2-q_1)\,yz\ \ge0\quad\text{on }(a,b).$$Integrating over $[a,b]$ and using $y(a)=y(b)=0$:
$$y'(b)z(b)-y'(a)z(a)=\int_a^b(q_2-q_1)\,y\,z\,dx .$$The left side is $\le0$ (both terms are), the right side is $\ge0$. So both are $0$: $z(a)=z(b)=0$, and $(q_2-q_1)yz\equiv0$, i.e. $q_1=q_2$ on $(a,b)$ (where $yz\gt0$), hence on $[a,b]$. Then $y,z$ solve the same equation with $y(a)=z(a)=0$, so $W(y,z)(a)=0$ and $z$ is a multiple of $y$. ∎
- If $q\le0$ on $I$, then $y$ has at most one zero in $I$ (it is non-oscillatory).
- If $q\le M^2$ on $I$ ($M\gt0$), consecutive zeros of $y$ are at distance $\ge\pi/M$.
- If $q\ge m^2\gt0$ on $I$, every closed interval $[c,c+\pi/m]\subseteq I$ contains a zero of $y$; consecutive zeros are at distance $\le\pi/m$; and if $I=[c,\infty)$, $y$ has infinitely many zeros.
- Every solution of Bessel's equation of any order $\nu$ has infinitely many zeros in $(0,\infty)$, and the gaps between consecutive zeros tend to $\pi$. For $0\le\nu\lt\frac12$ the gaps are $\lt\pi$, for $\nu=\frac12$ they equal $\pi$, for $\nu\gt\frac12$ they are $\gt\pi$.
Proof. (1) If $y(a)=y(b)=0$ with $a\lt b$, then $(yy')'=y'^2+yy''=y'^2-qy^2\ge y'^2$, and integrating, $\int_a^by'^2\le[yy']_a^b=0$. So $y'\equiv0$ on $[a,b]$, $y$ is constant, hence $y\equiv0$ — a contradiction. (2) Let $a\lt b$ be consecutive zeros with $b-a\lt\pi/M$. Then $z=\sin M(x-a)$ solves $z''+M^2z=0$ with $M^2\ge q$ and has no zero in $(a,b)$; by Theorem 6.38, $q\equiv M^2$ and $z$ is a multiple of $y$, so $b=a+\pi/M$ — a contradiction. (3) Apply Theorem 6.38 with $q_1=m^2$, the solution $\sin m(x-c)$ on $[c,c+\pi/m]$, and $q_2=q$: either $y$ vanishes inside, or $y$ is a multiple of $\sin m(x-c)$ and vanishes at the end points. The other claims follow by placing such intervals end to end. (4) By Theorem 6.37 the zeros on $(0,\infty)$ are those of $u''+\left(1+\frac{1-4\nu^2}{4x^2}\right)u=0$. Given $\varepsilon\gt0$, for $x$ large $q$ lies between $(1-\varepsilon)^2$ and $(1+\varepsilon)^2$, and (2), (3) trap the gaps between $\frac{\pi}{1+\varepsilon}$ and $\frac{\pi}{1-\varepsilon}$. The sign of $1-4\nu^2$ decides whether $q\gt1$, $q=1$ or $q\lt1$ for all $x$, which compares the gaps with $\pi$ exactly. ∎
Proof. $p'=\frac{a_1}{a_0}p$, so $(py')'=py''+\frac{a_1}{a_0}py'=\mu(a_0y''+a_1y')$. ∎ For instance $x^2y''+xy'+\lambda y=0$ becomes $(xy')'+\frac\lambda xy=0$: $p=x$, $q=0$, weight $r=\frac1x$.
- every eigenvalue is real, and the eigenfunctions may be taken real;
- eigenfunctions $\varphi,\psi$ belonging to different eigenvalues are orthogonal with weight $r$: $\int_a^b\varphi\,\psi\,r\,dx=0$;
- every eigenvalue is simple: its eigenfunctions are multiples of one of them;
- the eigenvalues form an increasing sequence $\lambda_1\lt\lambda_2\lt\cdots\to+\infty$, and the eigenfunction $\varphi_n$ of $\lambda_n$ has exactly $n-1$ zeros in $(a,b)$;
- if $q\le0$, $\alpha_1\alpha_2\le0$ and $\beta_1\beta_2\ge0$, all eigenvalues are $\ge0$; the eigenfunctions form an orthogonal basis of the space of square-integrable functions with weight $r$.
Proof of (1)–(3) and the sign in (5); (4) and the basis property are sketched. Write $Ly=(py')'+qy$. Lagrange's identity: $uLv-vLu=\big(p(uv'-u'v)\big)'$, by expanding both sides. If $u,v$ satisfy the same separated condition at $a$, the vectors $(u(a),u'(a))$ and $(v(a),v'(a))$ are both orthogonal to $(\alpha_1,\alpha_2)\ne0$, hence parallel, so $u(a)v'(a)-u'(a)v(a)=0$; likewise at $b$. Thus $\int_a^b(uLv-vLu)\,dx=0$ for all $u,v$ satisfying the conditions (also complex ones, since the conditions have real coefficients). (1) If $L\varphi=-\lambda r\varphi$, apply this to $u=\bar\varphi$, $v=\varphi$: $(\bar\lambda-\lambda)\int r|\varphi|^2=0$, and $\int r|\varphi|^2\gt0$, so $\lambda$ is real; then the real and imaginary parts of $\varphi$ are real eigenfunctions. (2) With $L\varphi=-\lambda r\varphi$, $L\psi=-\mu r\psi$: $(\lambda-\mu)\int r\varphi\psi=0$. (3) Two eigenfunctions for $\lambda$ satisfy the same condition at $a$, so their Wronskian vanishes at $a$, and solutions of the same linear equation with vanishing Wronskian are dependent. (5, sign) Multiply $L\varphi=-\lambda r\varphi$ by $\varphi$ and integrate by parts:
$$\lambda\int_a^br\varphi^2=\int_a^b\left(p\varphi'^2-q\varphi^2\right)dx+p(a)\varphi(a)\varphi'(a)-p(b)\varphi(b)\varphi'(b) .$$If $\alpha_2\ne0$, then $\varphi'(a)=-\frac{\alpha_1}{\alpha_2}\varphi(a)$ and the term from $a$ is $-p(a)\frac{\alpha_1}{\alpha_2}\varphi(a)^2\ge0$ when $\alpha_1\alpha_2\le0$; if $\alpha_2=0$ it is $0$. Likewise the term from $b$ is $p(b)\frac{\beta_1}{\beta_2}\varphi(b)^2\ge0$ when $\beta_1\beta_2\ge0$. With $q\le0$ every term on the right is $\ge0$, so $\lambda\ge0$. (4) One introduces the Prüfer angle $\theta$ with $p y'=\rho\cos\theta$, $y=\rho\sin\theta$; it increases with $\lambda$ and passes the values $k\pi$ exactly at zeros of $y$, and Sturm comparison makes $\theta(b,\lambda)$ increase continuously from near $0$ to $+\infty$; each time it reaches the value required by the condition at $b$ an eigenvalue appears, with one more interior zero than the previous one. The full argument and the expansion theorem are in the 📚 box. ∎
- The radius is decided in $\mathbb{C}$. $(1+x^2)y''+4xy'+2y=0$ has the solution $\frac1{1+x^2}$, smooth on $\mathbb{R}$, whose Maclaurin series diverges for $|x|\gt1$: the singular points $\pm i$ are invisible on the real line (Example 2(d), Interactive 1).
- Theorem 6.28 gives a lower bound only. For Legendre's equation with $\nu=n$ the guaranteed radius about $0$ is $1$, but the solution $P_n$ is a polynomial whose series converges everywhere; only the other solution really has radius $1$.
- Frobenius needs a regular singular point. At the irregular point $0$ of $x^2y''+(3x-1)y'+y=0$ the formal series $\sum_kk!\,x^k$ satisfies the equation term by term but converges only at $x=0$.
- Integer difference does not force a logarithm (Bessel of order $\frac12$, and every Cauchy–Euler equation, e.g. $x^2y''+xy'-y=0$ with solutions $x$, $x^{-1}$), but equal exponents always do, and $xy''-y=0$ shows that integer difference sometimes does (Example 4).
- $J_{-n}$ is not new. For integer $n$, $J_{-n}=(-1)^nJ_n$; the second solution is $Y_n$, unbounded at $0$. For non-integer $\nu$ (including $\nu=\frac12,\frac32$), $J_\nu,J_{-\nu}$ are a fundamental set.
- Not every solution of Legendre's equation with $\nu=n$ is a polynomial. The second solution $Q_n$ behaves like $\log\frac{1+x}{1-x}$ near $\pm1$; for instance $Q_0=\frac12\log\frac{1+x}{1-x}$ solves the case $n=0$. And for non-integer $\nu$ no nonzero solution is bounded on $(-1,1)$.
- Orthogonality needs the right weight and interval. $\int_0^1J_0(\alpha_1x)J_0(\alpha_2x)\,dx\approx0.196\ne0$; the weight is $x$. Legendre polynomials are orthogonal on $[-1,1]$, not on $[0,1]$: $\int_0^1P_0P_1\,dx=\frac12$.
- Sturm separation needs linear independence and the same equation. $\sin x$ and $2\sin x$ share all their zeros; $\sin x$ and $\sin 2x$ solve different equations and do not interlace ($0,\pi$ are common zeros).
- $q\gt0$ does not imply oscillation: $y''+\frac{1}{4x^2}y=0$ has the zero-free solution $\sqrt x$ on $[1,\infty)$. A positive lower bound $q\ge m^2$ is what forces infinitely many zeros.
- Periodic problems can have double eigenvalues. $y''+\lambda y=0$, $y(0)=y(2\pi)$, $y'(0)=y'(2\pi)$ has $\lambda=n^2$ with the two independent eigenfunctions $\cos nx$, $\sin nx$ ($n\ge1$). Simplicity in Theorem 6.41(3) is for separated conditions.
- Negative eigenvalues happen when the sign conditions fail. With the convention $(py')'+qy+\lambda ry=0$, the problem $y''+\lambda y=0$, $y(0)=0$, $y'(1)=2y(1)$ has $\beta_1\beta_2=2\cdot(-1)\lt0$, and indeed $y=\sinh\mu x$ is an eigenfunction for $\lambda=-\mu^2$ whenever $\tanh\mu=\mu/2$: $\mu\approx1.915$, $\lambda\approx-3.667$. Compare Exercise IV.7, where $\beta_1\beta_2=1\gt0$ and all eigenvalues are positive.
- CSIR Unit 3 · §2.2 Non-Homogeneous Equations and Variation of Parameters — its final part, on series solutions, proves the convergence of power series solutions by Cauchy's method of majorants, treats Frobenius' theorem in all three cases with the logarithmic solutions written out, and collects Legendre's and Bessel's functions with their second solutions $Q_n$ and $Y_n$.
- CSIR Unit 3 · §2.3 Sturm–Liouville Boundary Value Problems — solvability of two-point problems, Lagrange's identity and Green's formula, the Sturm–Picone comparison theorem, the oscillation theorem (infinitely many eigenvalues, $n-1$ zeros) via the Prüfer angle, the Rayleigh quotient and minimum principle, the eigenfunction expansion theorem, and singular problems with Legendre and Bessel expansions.
Two Labs
Exercises
- 1 Find the recurrence and the first four nonzero terms of the two basic power series solutions of Airy's equation $y''-xy=0$ about $0$. What is their radius of convergence?
- 2 Show that $y''-2xy'+4y=0$ has a polynomial solution and find it. Is the other basic solution a polynomial?
- 3 Without solving, give the radius of convergence guaranteed for series solutions of $(x^2+2x+2)y''+xy'+y=0$ about $x_0=0$ and about $x_0=1$.
- 4 Classify all singular points of $x^2(x+1)^2y''+(x^2-1)y'+2y=0$.
- 5 Find the exponents at $0$ of $2x^2y''-xy'+(1+x)y=0$ and the first three terms of two independent Frobenius solutions.
- 6 Show that $xy''+y'-y=0$ has exponents $0,0$ at the origin, find the Frobenius solution $y_1$ in closed summation form, and state the form of a second solution.
- 7 Find the Frobenius solution of $xy''-y=0$ with exponent $1$, and prove that every solution independent of it contains $\log x$.
- 8 Prove that at an ordinary point (viewed as a "regular singular point"), the indicial equation is $r(r-1)=0$, and explain why no logarithm appears although the exponents differ by $1$.
- 9 Give an example of an equation with a regular singular point at $0$ whose exponents differ by $2$ and which has two independent solutions of the form $x^r\sum a_kx^k$, or show that none exists.
- 1 Compute $P_3$ from Rodrigues' formula and check $P_3(1)=1$ and $P_3'(1)=6$.
- 2 Express $x^4$ as a combination of $P_0,\dots,P_4$.
- 3 Evaluate $\int_{-1}^1P_3(x)^2\,dx$, $\int_{-1}^1x^2P_3(x)\,dx$ and $\int_{-1}^1x^3P_3(x)\,dx$.
- 4 Find $P_6(0)$ and $P_7(0)$.
- 5 Prove $P_n(1)=1$ and $P_n(-x)=(-1)^nP_n(x)$ directly from the generating function.
- 6 Show that $\int_{-1}^1x\,P_n(x)P_n'(x)\,dx=\frac{2n}{2n+1}$.
- 7 Find the solution of $(1-x^2)y''-2xy'+12y=0$ that is bounded on $(-1,1)$ with $y(1)=2$.
- 8 (Approximation.) Find the polynomial $p$ of degree $\le2$ minimising $\int_{-1}^1\big(|x|-p(x)\big)^2dx$.
- 1 Prove $J_0'=-J_1$ directly from the series.
- 2 Express $J_2$ and $J_3$ in terms of $J_0$ and $J_1$.
- 3 Show that $J_{-1/2}(x)=\sqrt{\frac{2}{\pi x}}\cos x$ and compute the Wronskian of $J_{1/2}$ and $J_{-1/2}$.
- 4 Evaluate $\int xJ_0(x)\,dx$, $\int x^2J_1(x)\,dx$ and $\int J_1(x)\,dx$.
- 5 Show that between two consecutive positive zeros of $J_0$ there is exactly one zero of $J_1$.
- 6 Show that $y=J_0(2\sqrt x)$ solves $xy''+y'+y=0$.
- 7 Let $\alpha_n$ be the positive zeros of $J_0$. Find the coefficients $c_n$ in the Fourier–Bessel expansion $1=\sum_nc_nJ_0(\alpha_nx)$ on $(0,1)$.
- 1 Verify the separation theorem for $\sin x$ and $\cos x$, and compute their Wronskian.
- 2 Show that every nontrivial solution of $y''+\left(4+\frac{x^2}{1+x^2}\right)y=0$ has infinitely many zeros in $[0,\infty)$, with consecutive zeros at distance between $\pi/\sqrt5$ and $\pi/2$.
- 3 Show that a nontrivial solution of $y''-(1+x^2)y=0$ has at most one real zero.
- 4 Use the normal form to show that consecutive positive zeros of $J_0$ are less than $\pi$ apart, and those of $J_2$ more than $\pi$ apart.
- 5 Find all eigenvalues and eigenfunctions of $y''+\lambda y=0$, $y'(0)=0$, $y'(L)=0$.
- 6 Put Chebyshev's equation $(1-x^2)y''-xy'+\lambda y=0$ into self-adjoint form on $(-1,1)$ and state the weight for which its polynomial solutions are orthogonal.
- 7 For $y''+\lambda y=0$, $y(0)=0$, $y(1)+y'(1)=0$, show that the eigenvalues satisfy $\tan\sqrt\lambda=-\sqrt\lambda$ and compute $\lambda_1$ to three decimals.
- 8 Without solving, prove that every eigenvalue of $(xy')'+\frac\lambda xy=0$, $y(1)=y(e)=0$, is positive.
- 9 Prove or disprove: the eigenvalues of $y''+\lambda y=0$, $y(0)=y(2\pi)$, $y'(0)=y'(2\pi)$ are all simple.
E1. GATE MA · MCQ A fundamental set of solutions of $9x^2y''+9xy'+(9x^2-4)y=0$ on $(0,\infty)$ is (A) $J_{2/3}(x),\ J_{-2/3}(x)$; (B) $J_{2/3}(x),\ J_{3/2}(x)$; (C) $J_{2/3}(3x),\ J_{-2/3}(3x)$; (D) $J_{2/3}(x),\ xJ_{2/3}(x)$.
E2. GATE MA · MCQ If $x^3=aP_3(x)+bP_2(x)+cP_1(x)+dP_0(x)$, then $a+b+c+d$ equals (A) $\frac25$; (B) $1$; (C) $\frac35$; (D) $\frac65$.
E3. GATE MA · MSQ Consider $x^2(1-x)y''-x(1+x)y'+y=0$. Which of the following is/are true? (One or more options may be correct.) (A) $x=0$ is a regular singular point. (B) The exponents at $x=0$ are $1$ and $-1$. (C) Every solution on $(0,1)$ is of the form $x^r\sum_{k\ge0}a_kx^k$. (D) The equation has a Frobenius solution about $0$ whose series converges for $|x|\lt1$.
E4. GATE MA · MSQ Let $y\not\equiv0$ solve $y''+(2+\cos x)y=0$ on $\mathbb{R}$. Which of the following is/are true? (One or more options may be correct.) (A) $y$ has infinitely many zeros in $[0,\infty)$. (B) Consecutive zeros of $y$ are less than $\pi$ apart. (C) Consecutive zeros of $y$ are less than $\pi/\sqrt3$ apart. (D) $y$ has at most $7$ zeros in $[0,4\pi]$.
E5. GATE MA · NAT Let $y=\sum_{k\ge0}a_kx^k$ be the solution of $y''+xy'+3y=0$ with $y(0)=1$, $y'(0)=0$. The value of $a_6$ is ______ (correct to two decimal places).
E6. GATE MA · NAT The smallest eigenvalue $\lambda$ of $x^2y''+xy'+\lambda y=0$, $y(1)=0$, $y(e^2)=0$, is ______ (correct to two decimal places).
Suggested plan — 3 lectures × 60 min.
- Series and Frobenius. Definition 6.15 with three quick classifications; Theorem 6.28 and the four-step strategy; Example 1 on the board (let the class discover $e^{-x^2/2}$). Project Interactive 1 on $\frac1{1+x^2}$ and ask where the series "knows" to stop converging. Definition 6.16, Theorem 6.29 with the recurrence $(\ast)$ and the $\partial/\partial r$ trick for equal exponents; Examples 3 and 4.
- Legendre and Bessel. Definition 6.17 and the terminating recurrence; Rodrigues (Theorem 6.30) with the Leibniz computation; generating function and Bonnet (Theorem 6.31); orthogonality via the self-adjoint form, the norm via Rodrigues, zeros (Theorem 6.32); Example 5 as a quick-fire drill. Definition 6.18, Theorems 6.33–6.35, Example 6.
- Sturm theory. Theorem 6.36 via the Wronskian sign; the normal form; Theorem 6.38 with its two-line integral proof and Theorem 6.39. Project Interactive 2: first $q=1$, then $2+\cos x$ with the readouts, then Bessel $\nu=3$. Definition 6.19, Theorems 6.40–6.41 and Example 7; finish with GATE corner E4 and E6.
Discussion questions. (1) Why does the radius of convergence of a series solution "see" complex singular points of the equation, and what does this say about Taylor series of real functions in general? (2) The Legendre and Bessel orthogonality proofs needed no boundary conditions at $\pm1$ or at $0$. What replaced them? (3) In Interactive 2, why can a red zero never overtake a blue zero, however $\theta$ is changed?
Misconceptions to watch for. Taking the radius as the distance to the nearest real singular point; applying Frobenius at irregular points; "integer difference ⇒ logarithm" (and its converse); using $J_{-n}$ as a second solution; forgetting the weight $x$ in Bessel orthogonality or using $[0,1]$ for Legendre; concluding oscillation from $q\gt0$; claiming simple eigenvalues for periodic problems; mixing up the direction of the comparison theorem.
Suggested homework. I(1, 3, 5, 7, 9), II(2, 3, 4, 7, 8), III(2, 4, 5, 7), IV(2, 4, 5, 7, 9); GATE corner E1, E3, E5.
Research question. Why does a drum not sound like a guitar string? The overtones of a string are whole-number multiples of its fundamental; what are the overtones of a circular membrane, and what role do the zeros of Bessel functions play?
Why it matters. Orchestral timpani, tabla and mridangam makers, loudspeaker designers and engineers who test vibrating plates and MEMS membranes all work with the modes of a disc. Separating the wave equation $u_{tt}=c^2\Delta u$ on the disc $r\le a$ (with $u=0$ on the rim) in polar coordinates gives modes $J_m(kr)\cos m\theta\,\cos ckt$, and the rim condition $J_m(ka)=0$ quantises $k$: the frequencies are $\frac{c\,j_{m,n}}{2\pi a}$, where $j_{m,n}$ is the $n$-th positive zero of $J_m$.
What to do.
- Separate variables: show that $u=R(r)\Theta(\theta)T(t)$ leads to Bessel's equation $r^2R''+rR'+(k^2r^2-m^2)R=0$, and explain why only $J_m(kr)$ (not $Y_m$) is allowed.
- Compute the zeros $j_{m,n}$ for $m=0,1,2,3$ and $n=1,2,3$, either with a library or by bisection on the series. Use Theorem 6.39(4) to predict the spacing and compare.
- List the eight lowest frequencies in your table as ratios $j_{m,n}/j_{0,1}$. Which, if any, are close to integers or simple fractions? Compare with a string, whose ratios are $1,2,3,\dots$
- Timpani are struck off-centre and their air cavity lowers some modes; players report the modes $(1,1),(2,1),(3,1),(4,1)$ in a ratio close to $2:3:4:5$. Compute the ideal-membrane ratios of these four modes and discuss how close they are.
Tools: Python with scipy.special.jn_zeros (or your own series and bisection); Interactive 2 (Bessel presets) for the spacing of zeros.
Hand in: the separation of variables; a table of $j_{m,n}$; the ratio list; a paragraph comparing a drum, a string and a tuned timpano.
Going further. Read about Mark Kac's question "Can one hear the shape of a drum?" (1966) and the 1992 answer by Gordon, Webb and Wolpert: two differently shaped drums with exactly the same frequencies.
🧪 Self-check quiz
Q1. For $x^2y''+(\sin x)\,y'+y=0$, the point $x=0$ is
Q2. The radius of convergence guaranteed for power series solutions of $(x^2+9)y''+y=0$ about $x_0=4$ is
Q3. $\int_{-1}^{1}P_5(x)^2\,dx$ equals
Q4. $J_{1/2}(x)$ equals
Q5. For a regular Sturm–Liouville problem with eigenvalues $\lambda_1\lt\lambda_2\lt\cdots$, the eigenfunction of $\lambda_4$ has, in the open interval $(a,b)$,