13.2: Sturm-Liouville Problems
( \newcommand{\kernel}{\mathrm{null}\,}\)
In this section we consider eigenvalue problems of the form
P0(x)y″+P1(x)y′+P2(x)y+λR(x)y=0,B1(y)=0,B2(y)=0,
where
B1(y)=αy(a)+βy′(a)andB2(y)=ρy(b)+δy′(b).
As in Section 13.1, α, β, ρ, and δ are real numbers, with
α2+β2>0andρ2+δ2>0,
P0, P1, P2, and R are continuous, and P0 and R are positive on [a,b].
We say that λ is an eigenvalue of Equation ??? if Equation ??? has a nontrivial solution y. In this case, y is an eigenfunction associated with λ, or a λ-eigenfunction. Solving the eigenvalue problem means finding all eigenvalues and associated eigenfunctions of Equation ???.
Solve the eigenvalue problem
y″+3y′+2y+λy=0,y(0)=0,y(1)=0.
Solution
The characteristic equation of Equation ??? is
r2+3r+2+λ=0,
with zeros
r1=−3+√1−4λ2andr2=−3−√1−4λ2.
If λ<1/4 then r1 and r2 are real and distinct, so the general solution of the differential equation in Equation ??? is
y=c1er1t+c2er2t.
The boundary conditions require that
c1er1+c2er2=0c1er1+c2er2=0.
Since the determinant of this system is er2−er1≠0, the system has only the trivial solution. Therefore λ isn’t an eigenvalue of Equation ???.
If λ=1/4 then r1=r2=−3/2, so the general solution of Equation ??? is
y=e−3x/2(c1+c2x).
The boundary condition y(0)=0 requires that c1=0, so y=c2xe−3x/2 and the boundary condition y(0) requires that c2=0. Therefore λ=1/4 isn’t an eigenvalue of Equation ???.
If λ>1/4 then
r1=−32+iωandr2=−32−iω,
with
ω=√4λ−12or equivalentlyλ=1+4ω24.
In this case the general solution of the differential equation in Equation ??? is
y=e−3x/2(c1cosωx+c2sinωx).
The boundary condition y(0)=0 requires that c1=0, so y=c2e−3x/2sinωx, which holds with c2≠0 if and only if ω=nπ, where n is an integer. We may assume that n is a positive integer. (Why?). From Equation ???, the eigenvalues are λn=(1+4n2π2)/4, with associated eigenfunctions
yn=e−3x/2sinnπx,n=1,2,3,….
Solve the eigenvalue problem
x2y″+xy′+λy=0,y(1)=0,y(2)=0.
Solution
If λ=0, the differential equation in Equation ??? reduces to x(xy′)′=0, so xy′=c1,
y′=c1x,andy=c1lnx+c2.
The boundary condition y(1)=0 requires that c2=0, so y=c1lnx. The boundary condition y(2)=0 requires that c1ln2=0, so c1=0. Therefore zero isn’t an eigenvalue of Equation ???.
If λ<0, we write λ=−k2 with k>0, so Equation ??? becomes
x2y″+xy′−k2y=0,
an Euler equation (Section 7.4) with indicial equation
r2−k2=(r−k)(r+k)=0.
Therefore
y=c1xk+c2x−k.
The boundary conditions require that
2kc1+2−kc2=02kc1+2−kc2=0.
Since the determinant of this system is 2−k−2k≠0, c1=c2=0. Therefore Equation ??? has no negative eigenvalues.
If λ>0 we write λ=k2 with k>0. Then Equation ??? becomes
x2y″+xy′+k2y=0,
an Euler equation with indicial equation
r2+k2=(r−ik)(r+ik)=0,
so
y=c1cos(klnx)+c2sin(klnx).
The boundary condition y(1)=0 requires that c1=0. Therefore y=c2sin(klnx). This holds with c2≠0 if and only if k=nπ/ln2, where n is a positive integer. Hence, the eigenvalues of Equation ??? are λn=(nπ/ln2)2, with associated eigenfunctions
yn=sin(nπln2lnx),n=1,2,3,….
For theoretical purposes, it is useful to rewrite the differential equation in Equation ??? in a different form, provided by the next theorem.
If P0, P1, P2, and R are continuous and P0 and R are positive on a closed interval [a,b], then the equation
P0(x)y″+P1(x)y′+P2(x)y+λR(x)y=0
can be rewritten as
(p(x)y′)′+q(x)y+λr(x)y=0,
where p, p′, q and r are continuous and p and r are positive on [a,b].
- Proof
-
We begin by rewriting Equation ??? as
y″+u(x)y′+v(x)y+λR1(x)y=0,
with u=P1/P0, v=P2/P0, and R1=R/P0. (Note that R1 is positive on [a,b].) Now let p(x)=eU(x), where U is any antiderivative of u. Then p is positive on [a,b] and, since U′=u,
p′(x)=p(x)u(x)
is continuous on [a,b]. Multiplying Equation 13.2.26 by p(x) yields
p(x)y″+p(x)u(x)y′+p(x)v(x)y+λp(x)R1(x)y=0.
Since p is positive on [a,b], this equation has the same solutions as Equation ???. From Equation 13.2.27,
(p(x)y′)′=p(x)y″+p′(x)y′=p(x)y″+p(x)u(x)y′,
so Equation 13.2.28 can be rewritten as in Equation ???, with q(x)=p(x)v(x) and r(x)=p(x)R1(x). This completes the proof.
It is to be understood throughout the rest of this section that p, q, and r have the properties stated in Theorem 13.2.1 . Moreover, whenever we write Ly in a general statement, we mean
Ly=(p(x)y′)′+q(x)y.
The differential equation Equation ??? is called a Sturm-Liouville equation, and the eigenvalue problem
(p(x)y′)′+q(x)y+λr(x)y=0,B1(y)=0,B2(y)=0,
which is equivalent to Equation ???, is called a Sturm-Liouville problem.
Rewrite the eigenvalue problem
y″+3y′+(2+λ)y=0,y(0)=0,y(1)=0
of Theorem 13.2.1 as a Sturm-Liouville problem.
Solution
Comparing Equation ??? to Equation 13.2.26 shows that u(x)=3, so we take U(x)=3x and p(x)=e3x. Multiplying the differential equation in Equation ??? by e3x yields
e3x(y″+3y′)+2e3xy+λe3xy=0.
Since
e3x(y″+3y′)=(e3xy′)′,
Equation ??? is equivalent to the Sturm–Liouville problem
(e3xy′)′+2e3xy+λe3xy=0,y(0)=0,y(1)=0.
Rewrite the eigenvalue problem
x2y″+xy′+λy=0,y(1)=0,y(2)=0
of Theorem 13.2.2 as a Sturm-Liouville problem.
Solution
Dividing the differential equation in Equation ??? by x2 yields
y″+1xy′+λx2y=0.
Comparing this to Equation 13.2.26 shows that u(x)=1/x, so we take U(x)=lnx and p(x)=elnx=x. Multiplying the differntial equation by x yields
xy″+y′+λxy=0.
Since
xy″+y′=(xy′)′,
Equation ??? is equivalent to the Sturm–Liouville problem
(xy′)′+λxy=0,y(1)=0,y(2)=0.
Problems 1–4 of Section 11.1 are Sturm–Liouville problems. (Problem 5 isn’t, although some authors use a definition of Sturm-Liouville problem that does include it.) We were able to find the eigenvalues of Problems 1-4 explicitly because in each problem the coefficients in the boundary conditions satisfy αβ=0 and ρδ=0; that is, each boundary condition involves either y or y′, but not both. If this isn’t true then the eigenvalues can’t in general be expressed exactly by simple formulas; rather, approximate values must be obtained by numerical solution of equations derived by requiring the determinants of certain 2×2 systems of homogeneous equations to be zero. To apply the numerical methods effectively, graphical methods must be used to determine approximate locations of the zeros of these determinants. Then the zeros can be computed accurately by numerical methods.
Solve the Sturm–Liouville problem
y″+λy=0,y(0)+y′(0)=0,y(1)+3y′(1)=0.
Solution
If λ=0, the differential equation in Equation ??? reduces to y″=0, with general solution y=c1+c2x. The boundary conditions require that
c1+4c2=0c1+4c2=0,
so c1=c2=0. Therefore zero isn’t an eigenvalue of Equation ???.
If λ<0, we write λ=−k2 where k>0, and the differential equation in Equation ??? becomes y″−k2y=0, with general solution
y=c1coshkx+c2sinhkx,
so
y′=k(c1sinhkx+c2coshkx).
The boundary conditions require that
c1+kc2=0(coshk+3ksinhk)c1+(sinhk+3kcoshk)c2=0
The determinant of this system is
DN(k)=|1kcoshk+3ksinhksinhk+3kcoshk|=(1−3k2)sinhk+2kcoshk.
Therefore the system Equation ??? has a nontrivial solution if and only if DN(k)=0 or, equivalently,
tanhk=−2k1−3k2.
The graph of the right side (Figure 13.2.1 ) has a vertical asymptote at k=1/√3. Since the two sides have different signs if k<1/√3, this equation has no solution in (0,1/√3). Figure 13.2.1 shows the graphs of the two sides of Equation ??? on an interval to the right of the vertical asymptote, which is indicated by the dashed line. You can see that the two curves intersect near k0=1.2, Given this estmate, you can use Newton’s to compute k0 more accurately. We computed k0≈1.1219395. Therefore −k20≈−1.2587483 is an eigenvalue of Equation ???. From Equation ??? and the first equation in Equation ???,
y0=k0coshk0x−sinhk0x.
If λ>0 we write λ=k2 where k>0, and differential equation in Equation ??? becomes y″+k2y=0, with general solution
y=coskx+c2sinkx,
so
y′=k(−c1sinkx+c2coskx).
The boundary conditions require that
c1+kc2=0(cosk−3ksink)c1+(sink+3kcosk)c2=0.
The determinant of this system is
DP(k)=|1kcosk−3ksinksink+3kcosk|=(1+3k2)sink+2kcosk.
The system Equation ??? has a nontrivial solution if and only if DP(k)=0 or, equivalently,
tank=−2k1+3k2.
Figure 13.2.2 shows the graphs of the two sides of this equation. You can see from the figure that the graphs intersect at infinitely many points kn≈nπ (n=1, 2, 3,…), where the error in this approximation approaches zero as n→∞. Given this estimate, you can use Newton’s method to compute kn more accurately. We computed
k1≈12.9256856,k2≈16.1765914,k3≈19.3538959,k4≈12.5132570.
The estimates of the corresponding eigenvalues λn=k2n are
λ1≈158.5596361,λ2≈538.1502809,λ3≈587.4953676,λ4≈156.5815998.
From Equation ??? and the first equation in Equation ???,
yn=kncosknx−sinknx
is an eigenfunction associated with λn
Since the differential equations in Equation ??? and Equation \ref{eq:13.2.14} are more complicated than those in Equation \ref{eq:13.2.11} and Equation \ref{eq:13.2.13} respectively, what is the point of Theorem 13.2.1 ? The point is this: to solve a specific problem, it may be better to deal with it directly, as we did in Examples 13.2.1 and 13.2.2 ; however, we’ll see that transforming the general eigenvalue problem Equation \ref{eq:13.2.1} to the Sturm–Liouville problem Equation \ref{eq:13.2.10} leads to results applicable to all eigenvalue problems of the form Equation \ref{eq:13.2.1}.
If
Ly=(p(x)y')'+q(x)y \nonumber
and u and v are twice continuously functions on [a,b] that satisfy the boundary conditions B_{1}(y)=0 and B_{2}(y)=0, then
\label{eq:13.2.21} \int_{a}^{b}[u(x)Lv(x)-v(x)Lu(x)]\,dx=0.
- Proof
-
Integration by parts yields
\begin{aligned} \int_{a}^{b}[u(x)Lv(x)-v(x)Lu(x)]\,dx&= \int_{a}^{b}[u(x)(p(x)v'(x))'-v(x)(p(x)u'(x))']\,dx\\[4pt] &= p(x)[u(x)v'(x)-u'(x)v(x)]\bigg|_{a}^{b}\\[4pt] &-\int_{a}^{b}p(x)[u'(x)v'(x)-u'(x)v'(x)]\,dx.\end{aligned} \nonumber
Since the last integral equals zero,
\label{eq:13.2.22} \int_{a}^{b}[u(x)Lv(x)-v(x)Lu(x)]\,dx = p(x)[u(x)v'(x)-u'(x)v'(x)]\bigg|_{a}^{b}.
By assumption, B_{1}(u)=B_{1}(v)=0 and B_{2}(u)=B_{2}(v)=0. Therefore
\begin{aligned} \alpha u(a)+\beta u'(a)&=0\\[4pt] \alpha v(a)+\beta v'(a)&=0\\[4pt] \end{aligned} \quad \quad \text{and} \quad \quad \begin{gathered} \rho u(b)+\delta u'(b)=0\phantom{.}\\[4pt] \rho v(b)+\delta v'(b)=0. \end{gathered} \nonumber
Since \alpha^{2}+\beta^{2}>0 and \rho^{2}+\delta^{2}>0, the determinants of these two systems must both be zero; that is,
u(a)v'(a)-u'(a)v(a)=u(b)v'(b)-u'(b)v(b)=0. \nonumber
This and Equation \ref{eq:13.2.22} imply Equation \ref{eq:13.2.21}, which completes the proof.
The next theorem shows that a Sturm–Liouville problem has no complex eigenvalues.
If \lambda=p+qi with q\ne0 then the boundary value problem
Ly+\lambda r(x)y=0,\quad B_{1}(y)=0,\quad B_{2}(y)=0 \nonumber
has only the trivial solution.
- Proof
-
For this theorem to make sense, we must consider complex-valued solutions of
\label{eq:13.2.23} Ly+(p+iq)r(x,y)y=0.
If y=u+iv where u and v are real-valued and twice differentiable, we define y'=u'+iv' and y''=u''+iv''. We say that y is a solution of Equation \ref{eq:13.2.23} if the real and imaginary parts of the left side of Equation \ref{eq:13.2.23} are both zero. Since Ly=(p(x)'y)'+q(x)y and p, q, and r are real-valued,
\begin{aligned} Ly+\lambda r(x)y&=L(u+iv)+(p+iq)r(x)(u+iv)\\[4pt] &=Lu+r(x)(pu-qv)+i[Lv+r(x)(pu+qv)],\end{aligned} \nonumber
so Ly+\lambda r(x)y=0 if and only if
\begin{aligned} Lu+r(x)(pu-qv)&=0\\[4pt] Lv+r(x)(qu+pv)&=0.\end{aligned} \nonumber
Multiplying the first equation by v and the second by u yields
\begin{aligned} vLu+r(x)(puv-qv^{2})&=0\\[4pt] uLv+r(x)(qu^{2}+puv)&=0.\end{aligned} \nonumber
Subtracting the first equation from the second yields
uLv-vLu+qr(x)(u^{2}+v^{2})=0, \nonumber
so
\label{eq:13.2.24} \int_{a}^{b}[u(x)Lv(x)-v(x)Lu(x)]\,dx+ \int_{a}^{b}r(x)[u^{2}(x)+v^{2}(x)]\,dx=0.
Since
B_{1}(y)=B_{1}(u+iv)=B_{1}(u)+iB_{1}(v) \nonumber
and
B_{2}(y)=B_{2}(u+iv)=B_{2}(u)+iB_{2}(v), \nonumber
B_{1}(y)=0 and B_{2}(y)=0 implies that
B_{1}(u)=B_{2}(u)=B_{1}(v)=B_{2}(v)=0. \nonumber
Therefore Theorem 13.2.2 implies that first integral in Equation \ref{eq:13.2.24} equals zero, so Equation \ref{eq:13.2.24} reduces to
q\int_{a}^{b}r(x)[u^{2}(x)+v^{2}(x)]\,dx =0.\nonumber
Since r is positive on [a,b] and q\ne0 by assumption, this implies that u\equiv0 and v\equiv0 on [a,b]. Therefore y\equiv0 on [a,b], which completes the proof.
If \lambda_{1} and \lambda_{2} are distinct eigenvalues of the Sturm–Liouville problem
\label{eq:13.2.25} Ly+\lambda r(x)y=0,\quad B_{1}(y)=0,\quad B_{2}(y)=0
with associated eigenfunctions u and v respectively, then
\label{eq:13.2.26} \int_{a}^{b}r(x)u(x)v(x)\,dx=0.
- Proof
-
Since u and v satisfy the boundary conditions in Equation \ref{eq:13.2.25}, Theorem 13.2.2 implies that
\int_{a}^{b}[u(x)Lv(x)-v(x)Lu(x)]\,dx=0. \nonumber
Since Lu=-\lambda_{1}ru and Lv=-\lambda_{2}rv, this implies that
(\lambda_{1}-\lambda_{2})\int_{a}^{b}r(x)u(x)v(x)\,dx=0. \nonumber
Since \lambda_{1}\ne\lambda_{2}, this implies Equation \ref{eq:13.2.26}, which completes the proof.
If u and v are any integrable functions on [a,b] and
\int_{a}^{b} r(x)u(x)v(x)\,dx=0, \nonumber
we say that u and v orthogonal on [a,b] with respect to r=r(x).
Theorem 13.1.1 implies the next theorem.
If u\not\equiv0 and v both satisfy
Ly+\lambda r(x)y=0,\quad B_{1}(y)=0,\quad B_{2}(y)=0, \nonumber
then v=cu for some constant c.
We’ve now proved parts of the next theorem. A complete proof is beyond the scope of this book.
The set of all eigenvalues of the Sturm–Liouville problem
Ly+\lambda r(x)y=0,\quad B_{1}(y)=0,\quad B_{2}(y)=0 \nonumber
can be ordered as
\lambda_{1}<\lambda_{2}<\cdots<\lambda_{n}<\cdots, \nonumber
and
\lim_{n\to\infty} \lambda_{n}=\infty. \nonumber
For each n, if y_{n} is an arbitrary \lambda_{n}-eigenfunction, then every \lambda_{n}-eigenfunction is a constant multiple of y_{n}. If m\ne n, y_{m} and y_{n} are orthogonal [a,b] with respect to r=r(x); that is,
\label{eq:13.2.27} \int_{a}^{b} r(x)y_{m}(x)y_{n}(x)\,dx=0.
You may want to verify Equation \ref{eq:13.2.27} for the eigenfunctions obtained in Examples 13.2.1 and 13.2.2 .
In conclusion, we mention the next theorem. The proof is beyond the scope of this book.
Let \lambda_{1}<\lambda_{2}<\cdots<\lambda_{n}<\cdots be the eigenvalues of the Sturm–Liouville problem
Ly+\lambda r(x)y=0,\quad B_{1}(y)=0,\quad B_{2}(y)=0,\nonumber
with associated eigenvectors y_{1}, y_{2}, …, y_{n}, …. Suppose f is piecewise smooth (Definition 11.2.3) on [a,b]. For each n, let
c_{n}=\frac{ \int_{a}^{b} r(x)f(x)y_{n}(x) \, dx}{ \int_{a}^{b} r(x)y_{n}^{2}(x)\,dx}.\nonumber
Then
\frac{f(x-)+f(x+)}{2}=\sum_{n=1}^{\infty}c_{n}y_{n}(x) \nonumber
for all x in the open interval (a,b).