In a course on PDE's, one would typically come across some treatment of Sturm-Liouville problems, which are roughly in the form

$$\frac{d}{dx} \left [p(x) \frac{dy}{dx} \right ]+y \left (q(x)+\lambda w(x)\right )=0$$

I'll omit the details for each function as that is not the focus of the question and can be found in any book/online source.

But to me this seems rather arbitrary and I have not been successful in finding an explanation motivating such a form other than being thrown a number of examples and showing that they are in the above form for certain choices of $p(x),q(x)$ and $w(x)$.

I was wondering, what is the motivation for such a form? Or did Sturm/Liouville really just look at a number of ODE's and 'realised' that they were all special cases of this form?

  • 1
    $\begingroup$ Doesn't this form simply pop out when you try and solve a second order PDE (the most common type) using separation of variables? $\endgroup$
    – Paul
    Sep 5, 2016 at 13:17
  • $\begingroup$ I don't know what was SL original motivation, so let's just mention some likely reasons it's important: 1) it’s (roughly speaking) the most general second order linear ODE one can encounter 2) it contains as subcases so many of the functions we know and love; Bessel functions, Airy functions, Legendre, Hermite, Laguerre, and Chebyshev polynomials (and I’m sure I forgot some). 3) it pops up alot in applied mathematics, espesially physics, where a big chung of the DE one encounter are linear and second order. I'm sure there are examples from other fields also. $\endgroup$
    – Winther
    Sep 5, 2016 at 15:50

2 Answers 2


For me, it's fairly simple, really. Suppose you have the problem $$ L u = a(x) u''(x) + b(x) u'(x) + c(x) u = f(x), $$ with boundary conditions $$ B u = \begin{pmatrix} \alpha_{11} u(x_1) + \alpha_{12} u'(x_1) \\ \alpha_{21} u(x_2) + \alpha_{22} u'(x_2) \end{pmatrix} = 0, $$ where $a$, $b$, $c$ and $f$ are smooth enough, in order for all the following calculations to make sense, and $\alpha_{11}\alpha_{22} - \alpha_{12}\alpha_{21} \neq 0$. In a very broad sense, we are thinking of an operator $L$ that takes a function $u$ in the space $X = \{ u \text{ smooth enough such that } B u = 0 \}$ and send it to the function $f(x)$ in a space $Y$ (with the needed characteristics, I'm not going to get technical here): $$ L: X \longrightarrow Y. $$ In this setting, a fair question is "What are the properties of $L$?"

If we define the inner product as $$ \langle u(x), v(x) \rangle := \int_{x_1}^{x_2} u(x)v(x) dx $$ and the adjoint $L^*$ by the right hand side of the equation $$ \langle u(x), L^* v(x) \rangle := \langle L u(x), v(x) \rangle, $$ then, with the proper definitions, we can study $L$ in the same way we studied matrices. So, let's take a look to the expression $\langle L u(x), v(x) \rangle$. Expanding, we have \begin{multline} \langle L u(x), v(x) \rangle = \big(a u'v - u (a v)' + u(bv)\big)\big|_{x_1}^{x_2} \\ + \int_{x_1}^{x_2}\big(a v'' - (2 a' - b)v' + (a'' - b' + c)v\big)u dx \end{multline} Then, $$ L^* v = a v'' + (2 a' - b)v' + (a'' - b' + c)v $$ and $B = B^*$ (can you prove this?).

As in matrices, a very special kind of operator is the self-adjoint operator, $L = L^*$. Why? Well, among other things, because all its eignevalues are real and all its eigenfunctions are orthogonal (you should prove this also). Now, what does the self-adjoint operator looks like for second order ODEs? The function $a$ and $b$ must satisfy \begin{align} 2 a' - b &= b,\\ a'' - b' + c &= c. \end{align} Thus, $a' = b$, and $L$ can be written as $$ L u = a u'' + a' u' + c u = (a u')' + c u, $$ which is a Sturm-Liouville operator!

So, Sturm-Liouville operators are self-adjoint operators and have all the fantastic properties of self-adjoint operators. Great! But, they seem rather rare, as the condition $a' = b$ seems pretty restrictive. Why on Earth, then, we take so much effort in studying them? Well, what if instead of the regular inner product that I took, I use the inner product $$ \langle u(x), v(x) \rangle_{w(x)} = \int_{x_1}^{x_2} u(x) v(x) w(x) dx, $$ where $w(x) > 0$?

Then, $$ w L u = w a u'' + w b u ' + w c u = (w a u')' + (w b - (w a)')u' + w c, $$ and if $w$ satisfy the ODE $$ a w' + (b - a') w = 0, $$ the operator $wL$ is selfadjoint!

In conclusion,

With the proper inner product ($w(x) = e^{\int \frac{b-a'}{a} dx}$), every second order ODE can be studied as a Sturm-Liouville problem, where the operator is self-adjoint!

With one stroke, we've enclosed all second order ODEs into one, which happens to be self-adjoint. Those guys really knew what they were doing!


I would say that the main motivation for this problem comes from the associated variational framework. For simplicity let's consider the problem to be augmented with homogeneous boundary conditions, namely $y(0) = y(L) =0$. We then define the space $$ H^1_0((0,L)) = \{f : (0,L) \to \mathbb{R} : f, f' \in L^2((0,L)), f(0)=f(L) =0\}, $$ which is well-defined via the theory of traces of Sobolev functions.

We can then define the functionals $E,J : H^1_0((0,L)) \to \mathbb{R}$ via $$ E(y) = \int_0^L p |y'|^2 + q |y|^2 \\ J(y) = \int_0^L w|y|^2. $$ Now the key feature is that part of the S-L equations is the Euler-Lagrange equation for $E$, i.e. we have that critical points of $E$ must satisfy $$ 0 = DE(y) = -(p y')' + qy, $$ which just comes from integrating by parts in the expression $$ 0 = \frac{d}{dt} E(y + t \psi) = 2\int_0^L p y' \psi' + q y \psi $$ and then setting $t=0$.

Where does $J$ come in? Well, the variational derivative of it is $$ DJ(y) = wy, $$ which can be seen by a similar argument to the one above. Now, if instead of attempting to minimize $E$ we look for constrained minimizers, i.e. we look for $$ \min\{ E(y) : J(y) =1 \} $$ then the theory of Lagrange multipliers requires that $$ DE(y) = \lambda DJ(y) \text{ for some }\lambda \in \mathbb{R}, $$ which is equivalent to $$ -(py')' + qy = \lambda w y. $$ This is the form you describe above, up to flipping the sign on $q$, which doesn't really matter since in SL theory we don't assume it has a sign.

The above is really for "global minimizers" of $E$ subject to the constraint that $J=1$, but it can be pushed farther to get all possible "eigenvalues" $\lambda$. For instance, once the global minimizer $y_1$ is known with eigenvalue $\lambda_1$, we can consider $$ \min\{ E(y) : J(y)=1 \text{ and } \int_0^L w y y_1 =0 \}. $$ This will also produce a solution to $$ -(py')' + qy = \lambda w y. $$ but for $y_2$, $\lambda_2$ with $\lambda_2 \ge \lambda_1$.

Proceeding in this way one can ultimately view all solutions to the S-L equations in this form. All of this was done under the assumption that $y(0)=y(L) = 0$, but other boundary conditions work in this framework too.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.