What self-adjointness was secretly buying us
The previous guide did one careful thing: it rewrote a second-order operator in the self-adjoint form (p u')' + q u, with a positive weight w(x) hanging off to the side. That rewrite looked like bookkeeping, but it was really the price of admission to a theorem. Once a Sturm-Liouville problem is in this form, with sensible boundary conditions, three sweeping promises come true at once — and they are exactly the promises that made sines and cosines so well-behaved in the heat and wave rungs. This guide cashes in all three.
The Sturm-Liouville theorem states them plainly. First, every eigenvalue lambda is a real number. Second, the eigenfunctions belonging to different eigenvalues are orthogonal — not in the plain sense, but with respect to the weight w. Third, the eigenvalues form an increasing sequence lambda_1 < lambda_2 < lambda_3 < ... marching off to infinity, and the eigenfunctions are complete: any reasonable function on the interval can be expanded in them. Each clause earns its own section below.
Why every eigenvalue is real
Reality is not obvious. The operator involves derivatives, and there is no law forbidding a differential equation from having complex eigenvalues — the equation u'' = lambda u with the wrong boundary conditions happily admits them. So why does a self-adjoint problem refuse? The argument is short and worth carrying in your head. Define the weighted inner product of two functions as the integral over [a, b] of f times g-conjugate times the weight function w(x). Self-adjointness means the operator L can be moved from one slot of this inner product to the other for free: (L f, g) equals (f, L g).
Now run the trick. Let L u = -lambda u for an eigenfunction u, and feed u into both slots. On one side, (L u, u) = (-lambda u, u) = -lambda-bar times (u, u). On the other, (u, L u) = (u, -lambda u) = -lambda times (u, u). Self-adjointness forces these two to be equal, so lambda-bar (u, u) = lambda (u, u). The quantity (u, u) is the integral of |u|^2 times w, which is strictly positive because w > 0 and u is not the zero function. Cancel it and you are left with lambda-bar = lambda — a number equal to its own complex conjugate is real. That is the whole proof.
Why different eigenfunctions are orthogonal
The second promise is even more useful, and it comes from the same well. Take two eigenfunctions, u with eigenvalue lambda and v with a different eigenvalue mu. Put each into a slot of the weighted inner product. One side gives (L u, v) = -lambda (u, v); the other gives (u, L v) = -mu (u, v), since the eigenvalues are now known to be real and need no conjugation. Self-adjointness sets them equal: lambda (u, v) = mu (u, v), which rearranges to (lambda - mu)(u, v) = 0. Because lambda and mu were assumed different, the first factor is nonzero, forcing (u, v) = 0.
Read that conclusion carefully: (u, v) = 0 means the integral over [a, b] of u(x) v(x) w(x) dx vanishes. This is orthogonality with respect to the weight, and the weight w(x) is the part beginners forget. On a plain interval with w = 1 it reduces to the familiar orthogonality of sines you met two guides back; but on a disk the radial modes come weighted by w = r, and on the sphere the Legendre modes by another factor entirely. Two eigenfunctions can look thoroughly tangled when plotted, yet still be orthogonal once you remember to weigh the overlap by w.
self-adjoint problem: -(p u')' + q u = lambda w u on [a, b] real eigenvalues: lambda-bar = lambda (feed u into both slots) orthogonality: (lambda - mu)(u, v) = 0 (feed u, v into both slots) weighted inner prod: (u, v) = integral_a^b u(x) v(x) w(x) dx lambda != mu => (u, v) = 0
What orthogonality is for: reading off coefficients
Orthogonality is not a curiosity; it is the engine that makes expansions computable. Suppose you believe a function f can be written as a sum c_1 u_1 + c_2 u_2 + c_3 u_3 + ... in the eigenfunctions — an eigenfunction expansion, the direct generalization of a Fourier series. The coefficients look hopeless: infinitely many unknowns tangled in one equation. Orthogonality untangles them in a single stroke. Take the weighted inner product of both sides with one chosen mode u_n. Every cross term (u_m, u_n) with m not equal to n collapses to zero, and only the c_n (u_n, u_n) term survives.
So the n-th coefficient is just c_n = (f, u_n) / (u_n, u_n) — a ratio of two weighted integrals, computable one mode at a time without ever solving a giant system. This is precisely how the generalized Fourier series of f is built. The denominator (u_n, u_n) is the squared length of the mode in the weighted sense; dividing by it is the normalization step. There is also a tidy way to see whether c_n is the best you can do: the Rayleigh quotient measures how much energy each mode carries, and minimizing it picks out the lowest eigenvalue and its eigenfunction.
- Get the operator self-adjoint. Write the equation as -(p u')' + q u = lambda w u and read off the weight w(x) — this fixes which inner product to use.
- Solve the eigenvalue problem. Find the eigenvalues lambda_n and their eigenfunctions u_n that satisfy the boundary conditions; the theorem promises they are real and orthogonal.
- Project f onto each mode. Compute c_n = (f, u_n) / (u_n, u_n), where both inner products carry the weight w — every other mode drops out automatically.
- Reassemble and attach the dynamics. Sum c_n u_n to rebuild f, then carry each mode forward with its own time factor to solve the original PDE.
Beyond sine and cosine: the special functions
Here is the quiet revolution. Sines and cosines were never special — they are merely the eigenfunctions of the simplest Sturm-Liouville problem, the one with p = 1, q = 0, w = 1 on a straight interval. Change the geometry and the same machinery hands you a different orthogonal family. Separate the Laplacian on a disk and the radial part becomes Bessel's equation, whose eigenfunctions are the Bessel functions, orthogonal with weight w = r. Separate it on a sphere and the polar part becomes Legendre's equation, whose eigenfunctions are the Legendre polynomials, orthogonal on [-1, 1]. Stack the polar and azimuthal pieces together and you get the spherical harmonics.
These families are the eigenfunctions of singular Sturm-Liouville problems — singular because the coefficient p vanishes at an endpoint (p = r dies at the center of the disk; p = 1 - x^2 dies at the poles x = plus or minus 1). That is a genuine subtlety, not a footnote: at a singular endpoint there is no ordinary boundary condition to impose. The right replacement is to demand that the solution simply stay bounded there. Remarkably, boundedness is enough to keep the operator self-adjoint, so the same two-line argument still delivers real eigenvalues and weighted orthogonality. The Bessel and Legendre families are orthogonal for the very same reason the sines were.
One honest caveat closes the rung's central theorem. Real and orthogonal we have proven cleanly; completeness — that the eigenfunctions are enough to build every reasonable function — is the deepest clause and the one our short inner-product argument does not reach. It needs heavier analysis (compactness of the inverse operator), and it is what the next guide takes up in earnest. Until completeness is established you only know the modes don't overlap; you do not yet know they leave no gaps. Orthogonality without completeness is a basis that might be missing some directions — a caution worth keeping as you move on.