From orthogonal to complete: two different promises
By now this rung has handed you a beautiful object. A regular Sturm-Liouville problem delivers a discrete ladder of real eigenvalues λ_1 < λ_2 < λ_3 < ... marching off to infinity, and to each one a real eigenfunction φ_n(x). The previous guide proved these eigenfunctions are orthogonal with respect to a weight w(x): integrate φ_m φ_n w over the interval and you get zero whenever m is not n. That orthogonality is already powerful — it is exactly what lets you read off a coefficient by integration, the same trick that pins down ordinary Fourier coefficients. But orthogonality alone makes a quieter promise than you might think.
Here is the gap. Orthogonality says: if a function can be built from your eigenfunctions, here is exactly how to find the recipe. It says nothing about whether every reasonable function can be built at all. Think of vectors: the single vector (1,0,0) in three-dimensional space is perfectly orthogonal to everything it should be, yet you cannot reach (0,1,5) using only multiples of it. You are missing directions. Completeness is the separate, deeper promise that nothing is missing — that the eigenfunctions φ_n point in enough directions to reach every function in the space, leaving no leftover that is orthogonal to all of them.
Eigenfunctions as a coordinate system for functions
The whole idea of Sturm-Liouville theory is to treat functions the way you treat vectors. In ordinary space, an orthogonal basis like the unit vectors lets you write any vector as a weighted sum of basis directions, and you find each weight by projecting — a dot product. Here the eigenfunctions φ_n play the role of basis directions, and the inner product is an integral with the weight built in: the overlap of f and g is the integral of f(x) g(x) w(x) over the interval. Two functions are orthogonal when that integral is zero; the squared length of f is the integral of f(x)^2 w(x). This is the weight function w(x) doing real work — it is the metric that decides what counts as perpendicular.
With that inner product, expanding a function is mechanical. To write f(x) = sum of c_n φ_n(x), you find each coefficient by projecting f onto φ_n and dividing by the squared length of φ_n: c_n is the integral of f φ_n w, over the integral of φ_n^2 w. Every cross-term vanishes by orthogonality, so the formula for c_n never sees any φ_m but its own. This is an eigenfunction expansion, and because it generalizes the sine-and-cosine recipe to an arbitrary self-adjoint operator, it is also called a generalized Fourier series. The ordinary Fourier series you already know is just the special case where the operator is the plain second derivative and the eigenfunctions are sines and cosines with weight w = 1.
inner product: <f, g> = integral of f(x) g(x) w(x) dx over [a, b]
expansion: f(x) = sum_n c_n phi_n(x)
coefficient: <f, phi_n> integral of f phi_n w
c_n = -------------- = ----------------------
<phi_n, phi_n> integral of phi_n^2 w
(every cross-term <phi_m, phi_n> = 0 for m != n --> only the n-th survives)What "the series equals the function" really means
Now the honest part. When we write f(x) = sum of c_n φ_n(x), in what sense is that equality true? The safest and most natural answer for Sturm-Liouville theory is mean-square convergence, also called convergence in the L^2 sense. It says the energy of the error goes to zero: take the partial sum S_N of the first N terms, form the difference f - S_N, and the integral of (f - S_N)^2 w shrinks to 0 as N grows. The series fits f better and better on average, in the weighted-area sense — even if at a few stubborn points, like a jump, the values never quite agree.
This is a genuinely weaker promise than pointwise agreement, and the distinction is not a technicality — it is the source of real phenomena. The classic warning is the Gibbs overshoot: expand a square wave and the partial sums always overshoot near the jump by a fixed percentage that never goes away, no matter how many terms you add. Yet mean-square convergence holds perfectly, because that overshoot is squeezed into an ever-thinner sliver whose area shrinks to zero. So "the series equals f" is true in the energy sense even where it visibly disagrees point by point. Be precise about which convergence you mean before claiming the equality.
Completeness, Bessel's inequality, and Parseval's equality
Completeness has a sharp, measurable signature. Project f onto the first N eigenfunctions and the leftover f - S_N is orthogonal to all of them; a short Pythagorean calculation then shows the energy of the partial sum can never exceed the energy of f. Written out, the sum of c_n^2 times the squared length of φ_n is at most the integral of f^2 w. This is Bessel's inequality, and it holds for any orthogonal family, complete or not. The family is complete precisely when that inequality tightens into an equality for every f — when the coefficients capture all of the energy and the leftover has none.
That equality is Parseval's identity, and it is the cleanest face of completeness: the total energy of f equals the total energy of its coefficients, the integral of f^2 w equals the sum of c_n^2 times φ_n's squared length. Nothing leaks away into directions the eigenfunctions cannot see. The deep guarantee that this actually happens is the Sturm-Liouville theorem itself: for a regular Sturm-Liouville problem on a finite interval, the eigenfunctions are automatically complete in the weighted L^2 space. You do not have to check it case by case — self-adjointness plus the right boundary conditions buys completeness wholesale.
Using the expansion to solve a PDE
Completeness is what makes separation of variables finish. Recall the pattern from the previous rung: separating a PDE produced a spatial eigenvalue problem, and matching the initial data demanded that you expand the starting profile in the eigenfunctions. Completeness is the guarantee that this expansion can always be done — that whatever the initial temperature or displacement, it lives inside the span of the φ_n and can be reached. Without completeness you might solve the equation for a whole shelf of nice profiles and yet be unable to match the one profile your problem actually hands you.
- Build the basis. Solve the Sturm-Liouville eigenvalue problem from separation to get eigenvalues λ_n and orthogonal eigenfunctions φ_n(x), with the weight w(x) inherited from the operator.
- Expand the data. Write the initial profile f(x) as f = sum of c_n φ_n, finding each c_n by the projection formula — the integral of f φ_n w over the integral of φ_n^2 w. Completeness promises this represents f in the mean-square sense.
- Evolve each mode. Each eigenfunction rides its own time factor (a decaying exponential for the heat equation, an oscillation for the wave equation), set by its λ_n. Attach that factor to every term.
- Superpose. Add the evolved modes back up. The result is the full solution u(x,t) — and the very same series, frozen at t = 0, reproduces the data, closing the loop.
One honest caveat about that elegant loop. The interval here is finite and the endpoints regular, which is why the eigenvalues form a discrete ladder and the expansion is a sum. On an infinite domain, or with a singular endpoint, the ladder can dissolve into a continuum — and then the sum becomes an integral, the eigenfunction expansion becomes a transform, which is the very next rung's story. The bookkeeping changes, but the spirit is identical: project onto a complete family, evolve each piece, superpose.
Beyond sines: the special-function families ahead
The payoff of building the theory abstractly is that the same machine runs on far more than sines. Change the geometry and the separated spatial operator changes with it, but it is still a Sturm-Liouville operator — so it still has real eigenvalues, still has orthogonal eigenfunctions with respect to some weight, and on a finite domain is still complete. Only the cast of eigenfunctions changes. Separation on a disk or cylinder hands you Bessel functions; separation on a sphere hands you Legendre polynomials and, in the full angular problem, spherical harmonics. These are not exotic newcomers — they are simply the eigenfunctions that the new coordinate system selects.
And the expansion logic travels untouched. A radial profile gets written as a Bessel-function expansion; a profile on the sphere gets written as a spherical-harmonic expansion; each carries its own weight (an r in the radial integral, a sin θ on the sphere) that came straight from the geometry, and each is governed by its own completeness theorem. Even the Hermite and Laguerre functions, which live on infinite or half-infinite intervals, fall out of singular Sturm-Liouville problems. The next guide meets these families face to face — but you already know the one thing that matters most about all of them: each is a complete, weighted-orthogonal basis, so each lets you expand, project, and superpose exactly as you just did with sines.