Eigenfunction Expansions and Completeness

A vast swathe of physics — the plucked string, heat leaking from a rod, the bound states of an atom, the potential inside a charged sphere — collapses onto a single mathematical move: expand an arbitrary function in the eigenfunctions of a differential operator. Sturm–Liouville theory is what makes that move legitimate and systematic. It picks out a broad class of second-order operators that are self-adjoint, proves their eigenfunctions are mutually orthogonal, proves they are complete — enough of them to reconstruct any reasonable function — and pins down the precise sense in which the resulting series converges. It is, line for line, the infinite-dimensional continuation of a fact you already trust: a Hermitian matrix has an orthonormal eigenbasis, and any vector is the sum of its projections onto that basis. Two ideas do all the work, and they are not the same idea. Orthogonality is cheap: it follows the instant the operator is self-adjoint, and it hands you a formula for the expansion coefficients for free. Completeness is deep: it is the separate guarantee that no function slips through the cracks — that the coefficients you computed actually rebuild the function you started with. The gap between them is where most physical intuition quietly goes wrong. This article assembles the machinery in order. It builds the weighted Hilbert space where the operator lives, derives orthogonality and real eigenvalues from a single integration by parts (Lagrange's identity), shows why the Fourier coefficients are the best mean-square approximation, and ties completeness to the Green's function of the operator through the spectral theorem. Along the way it insists on the one distinction physicists most often blur — convergence in mean square is not convergence everywhere — with the Gibbs overshoot standing as the permanent, quantifiable warning.

Key takeaways

A regular Sturm–Liouville operator $\mathcal{L}y=-\tfrac{1}{w}[(py')'-qy]$ on $[a,b]$ with self-adjoint boundary conditions has real, discrete eigenvalues $\lambda_0<\lambda_1<\cdots\to\infty$ and eigenfunctions $\{y_n\}$ orthogonal under the weight $w$ .
Orthogonality is automatic from self-adjointness (Lagrange's identity); it gives the coefficient formula $c_n=\langle y_n,f\rangle_w$ but does not by itself guarantee reconstruction.
Completeness is the deeper statement that $\{y_n\}$ is an orthonormal basis of $L^2_w[a,b]$ : every square-integrable $f$ equals $\sum_n c_n y_n$ in the mean-square sense, with no nonzero function orthogonal to all of them.
Parseval's identity $\|f\|_w^2=\sum_n|c_n|^2$ is equivalent to completeness; Bessel's inequality $\sum_n|c_n|^2\le\|f\|_w^2$ holds always, complete or not.
Mean-square convergence is weaker than pointwise: partial sums overshoot jumps by a fixed 8.95% (Gibbs), term-by-term differentiation needs uniform convergence, and Fourier, Bessel, Legendre and Hermite series are all one completeness theorem with kernel $\sum_n y_n(x)\overline{y_n(x')}=\delta(x-x')/w(x')$ .
Key idea: A self-adjoint Sturm–Liouville operator is the infinite-dimensional Hermitian matrix — its eigenfunctions are orthogonal under the weighted inner product and complete in $L^2_w$ , so every square-integrable function is the mean-square sum of its generalised Fourier series $f=\sum_n \langle y_n,f\rangle_w\,y_n$ , with Parseval's identity expressing conservation of the norm across the spectral decomposition.

Where this lives today

The completeness theorem is quietly load-bearing across modern physics and computation. Every use of the closure relation $\sum_n|n\rangle\langle n|=\mathbb{1}$ in quantum mechanics — inserting a complete set of states, expanding a wavepacket, deriving a sum rule — is Sturm–Liouville completeness for the Hamiltonian, and Parseval's identity is conservation of total probability across eigenstates. Numerical physics leans on it even harder: spectral methods (Chebyshev, Legendre, and Fourier Galerkin schemes) solve PDEs by expanding the unknown in a complete orthogonal basis and evolving the coefficients, achieving exponential accuracy for smooth solutions precisely because completeness guarantees the truncation error is controlled in $L^2$ — the same reason they suffer Gibbs ringing at shocks. The spectrum itself carries geometric information: Mark Kac's 1966 question "can one hear the shape of a drum?" asks whether the Sturm–Liouville (Helmholtz) eigenvalues of a membrane determine its shape, answered no in 1992 when Gordon, Webb and Wolpert exhibited two differently shaped isospectral drums. And the modes are physical objects one now measures directly — the resonant patterns of a vibrating plate (Chladni figures), the normal modes read out of a gravitational-wave ringdown, and the diffusion eigenmodes underpinning MRI and heat-signature analysis are all eigenfunctions of a self-adjoint operator, expanded and reconstructed by the theorem above.

The physical picture

Hold a Hermitian matrix $A$ in mind. Its eigenvectors are orthogonal and span $\mathbb{C}^N$ , so any vector is $v=\sum_n (e_n^\dagger v)\,e_n$ — you resolve $v$ into components along a privileged set of axes. Sturm–Liouville theory keeps this picture intact and swaps only the furniture: vectors become functions, the dot product becomes $\int \overline{f}\,g\,w\,\mathrm{d}x$ , and the matrix becomes a self-adjoint differential operator $\mathcal{L}$ . The eigenfunctions are the new axes; completeness is the claim that they really do span the whole space of functions, leaving no direction unaccounted for. The first figure shows those axes concretely — the lowest three Dirichlet modes of the simplest operator $-\mathrm{d}^2/\mathrm{d}x^2$ , each carrying one more interior zero than the last, exactly as the oscillation theorem demands.

Figure

The first three Dirichlet eigenfunctions of −y″ = λy on the interval 0 to L: y₁, y₂, y₃ carry 0, 1, 2 interior zeros (the oscillation theorem), with eigenvalues λₙ = (nπ/L)² equal to 1, 4, 9 in units of (π/L)² — they are orthogonal under ∫₀ᴸ yₘ yₙ dx = 0 and complete in L²

Why does a weight $w(x)$ intrude at all? Because forcing a general second-order operator into self-adjoint form fixes a particular measure $w\,\mathrm{d}x$ under which — and only under which — $\mathcal{L}$ is symmetric. Orthogonality is therefore always orthogonality against that weight, never against plain $\mathrm{d}x$ : Legendre polynomials are orthogonal with $w=1$ , Chebyshev with $w=(1-x^2)^{-1/2}$ , Hermite functions with $w=e^{-x^2}$ , Bessel functions with $w=\rho$ . The weight is not decoration; it is where the geometry of the problem lives. The deeper question — why the eigenfunctions are not merely orthogonal but complete — is answered by the operator's inverse: integrating twice against the Green's function is a compact, self-adjoint operation, and the spectral theorem for compact operators hands back an orthonormal eigenbasis automatically. Completeness is inherited from compactness, which is inherited from the smoothing power of integration. What that completeness cannot promise is pointwise perfection: the second figure shows the permanent scar it leaves at a discontinuity.

Figure

A unit square wave rebuilt from its Fourier sine series: the 8-term partial sum (harmonics k = 1, 3, …, 15) converges to the wave in mean square, yet overshoots each jump by a fixed 8.95% of the jump height — the Gibbs phenomenon, which never disappears as N grows, only narrows toward the discontinuity

The core equations

The stage is a weighted Hilbert space and one operator on it. The eigenvalue problem below is the object of study; every subsequent equation either characterizes its spectrum or expands a function on its eigenfunctions.

\mathcal{L}y=-\frac{1}{w(x)}\left[\frac{\mathrm{d}}{\mathrm{d}x}\!\left(p(x)\frac{\mathrm{d}y}{\mathrm{d}x}\right)-q(x)\,y\right]=\lambda\,y \qquad (1)

where: $p(x)>0$ — continuously differentiable coefficient (a "tension"); $q(x)$ — continuous potential-like coefficient; $w(x)>0$ — positive weight defining the inner product; $\lambda$ — eigenvalue (spectral parameter); $y(x)$ — eigenfunction obeying the boundary conditions. Reading it: $\mathcal{L}$ is the continuous analogue of a Hermitian matrix, and $\lambda,y$ are its eigenvalue and eigenvector. Orthogonality is meaningless without an inner product, and the correct one is weighted. This is the geometry $\mathcal{L}$ imposes on the function space; it is the measure under which $\mathcal{L}$ is symmetric.

\langle f,g\rangle_w=\int_a^b \overline{f(x)}\,g(x)\,w(x)\,\mathrm{d}x,\qquad \|f\|_w^2=\langle f,f\rangle_w \qquad (2)

where: $\overline{f}$ — complex conjugate of $f$ (the inner product is conjugate-linear in the first slot); $w(x)$ — weight that makes $\mathcal{L}$ self-adjoint; $\|f\|_w$ — the $L^2_w$ norm, whose finiteness defines the space. Self-adjointness is not automatic — it costs a boundary condition. Lagrange's identity isolates exactly the surface term that must vanish for $\mathcal{L}$ to be symmetric.

\langle g,\mathcal{L}f\rangle_w-\langle \mathcal{L}g,f\rangle_w=\Big[\,p\big(\overline{g}'\,f-\overline{g}\,f'\big)\Big]_a^b \qquad (3)

where: the bracket $[\cdots]_a^b$ — the bilinear boundary (Wronskian-like) term; it vanishes for separated self-adjoint boundary conditions such as $\alpha y(a)+\beta y'(a)=0$ , making $\langle g,\mathcal{L}f\rangle_w=\langle \mathcal{L}g,f\rangle_w$ . The payoff of self-adjointness is orthogonality of eigenfunctions belonging to distinct eigenvalues — the property that makes the coefficient formula work.

\langle y_m,y_n\rangle_w=0\quad (\lambda_m\neq\lambda_n),\qquad \langle y_m,y_n\rangle_w=\delta_{mn}\ \text{(after normalising)} \qquad (4)

where: $\delta_{mn}$ — Kronecker delta; $\lambda_m,\lambda_n$ — distinct real eigenvalues. Reading it: the eigenfunctions form an orthonormal set once each is scaled to unit $\|\cdot\|_w$ . With an orthonormal set in hand, the generalised Fourier expansion is the projection of $f$ onto each axis — identical in form to $v=\sum_n(e_n^\dagger v)e_n$ .

f(x)=\sum_{n=0}^{\infty} c_n\,y_n(x),\qquad c_n=\langle y_n,f\rangle_w=\int_a^b \overline{y_n(x)}\,f(x)\,w(x)\,\mathrm{d}x \qquad (5)

where: $c_n$ — generalised Fourier coefficient; the equality holds in the mean-square ( $L^2_w$ ) sense, guaranteed only when the set is complete. Before completeness is known, only an inequality is available. Bessel's inequality says the energy captured by any finite set of projections cannot exceed the total.

\sum_{n=0}^{\infty}|c_n|^2 \;\le\; \|f\|_w^2 \qquad (6)

where: $|c_n|^2$ — energy in the $n$ -th mode. Reading it: the sum of squared coefficients is bounded by the squared norm — always true, whether or not the basis is complete. Completeness is precisely the upgrade of Bessel's $\le$ to Parseval's $=$ : no energy is lost, so nothing was missed.

\sum_{n=0}^{\infty}|c_n|^2 \;=\; \|f\|_w^2 \quad\Longleftrightarrow\quad \{y_n\}\ \text{is complete} \qquad (7)

where: the equality — Parseval's identity — is logically equivalent to completeness, which is equivalent in turn to the mean-square convergence of the partial sums to $f$ . Spelling out that convergence: it is convergence of the norm of the error to zero, not convergence at every point.

\lim_{N\to\infty}\Big\|f-\sum_{n=0}^{N}c_n y_n\Big\|_w=0 \qquad (8)

where: $\|\cdot\|_w$ — the $L^2_w$ norm. Reading it: the averaged squared error vanishes; pointwise the partial sums may still overshoot (Gibbs) or miss isolated points. Completeness has a pointwise avatar as a distribution — the resolution of the identity, the kernel that reproduces any function.

\sum_{n=0}^{\infty} y_n(x)\,\overline{y_n(x')}=\frac{\delta(x-x')}{w(x')} \qquad (9)

where: $\delta(x-x')$ — Dirac delta; the factor $1/w(x')$ — undoes the weight in the coefficient integral so that summing reproduces $f(x)$ exactly. Finally, the operator's resolvent packages the whole spectrum: its poles sit at the eigenvalues and its residues are built from the eigenfunctions.

G(x,x';\lambda)=\sum_{n=0}^{\infty}\frac{y_n(x)\,\overline{y_n(x')}}{\lambda_n-\lambda} \qquad (10)

where: $G$ — resolvent (Green's function) solving $(\mathcal{L}-\lambda)G=\delta(x-x')/w(x)$ with the problem's boundary conditions; the poles at $\lambda=\lambda_n$ encode the discrete spectrum and their residues rebuild the completeness relation (9).

Derivation AReal eigenvalues and weighted orthogonality from self-adjointness

The goal is to extract, from the single fact that $\mathcal{L}$ is self-adjoint on the weighted space, the two structural properties on which the entire expansion rests: that every eigenvalue is real, and that eigenfunctions belonging to different eigenvalues are orthogonal under the weight $w$ . Nothing more than Lagrange's identity and the conjugate-linearity of the inner product is needed; the boundary conditions enter only through the vanishing of one surface term, and once that term is gone the algebra is forced.

Step 1. Take two eigenpairs, $\mathcal{L}y_n=\lambda_n y_n$ and $\mathcal{L}y_m=\lambda_m y_m$ , with boundary conditions that kill the Lagrange bracket, so $\langle y_m,\mathcal{L}y_n\rangle_w=\langle \mathcal{L}y_m,y_n\rangle_w$ . Why: self-adjointness is exactly the licence to move $\mathcal{L}$ from one side of the inner product to the other without penalty — the whole argument is the consequence of that single move.

Step 2. Insert the eigenvalue equations. The right slot gives $\lambda_n\langle y_m,y_n\rangle_w$ ; the left slot, because the inner product is conjugate-linear in its first argument, gives $\overline{\lambda_m}\langle y_m,y_n\rangle_w$ . Hence $(\lambda_n-\overline{\lambda_m})\langle y_m,y_n\rangle_w=0$ . Why: the conjugate on $\lambda_m$ is the crux — it is what forces reality when $m=n$ and orthogonality when $m\neq n$ , and dropping it is the most common error in the whole subject.

Step 3. Set $m=n$ . Since $\|y_n\|_w^2>0$ , the factor $(\lambda_n-\overline{\lambda_n})$ must vanish, so $\lambda_n=\overline{\lambda_n}$ : every eigenvalue is real. Why: a self-adjoint operator cannot have a complex eigenvalue any more than a Hermitian matrix can — the weighted norm being strictly positive is what enforces it.

Step 4. Now take $m\neq n$ with reality established, so $\overline{\lambda_m}=\lambda_m$ and $(\lambda_n-\lambda_m)\langle y_m,y_n\rangle_w=0$ . Distinct eigenvalues make the prefactor nonzero, forcing $\langle y_m,y_n\rangle_w=0$ ; normalising each $y_n$ yields an orthonormal set. Why: this is the entire source of orthogonality — it is a theorem about the operator, not a lucky property of any particular special function. *Verified: reproduces Arfken–Weber–Harris Ch. 8 and Stone & Goldbart §8.2; the boxed result is $(\lambda_m-\overline{\lambda_n})\langle y_m,y_n\rangle_w=0$ , real eigenvalues and $\langle y_m,y_n\rangle_w=0$ for $\lambda_m\neq\lambda_n$ .*

Go deeper · Derivation B · 1 minWhy the Fourier coefficients are the best mean-square approximation

Derivation BWhy the Fourier coefficients are the best mean-square approximation

The goal is to prove that among all possible coefficient choices, the generalised Fourier coefficients $c_n=\langle y_n,f\rangle_w$ are the unique ones that minimise the mean-square error of the truncated expansion, and to read off Bessel's inequality, Parseval's identity, and the exact meaning of completeness from a single completed square. This is the projection theorem in disguise: the best approximation from a subspace is the orthogonal projection onto it, and the error is whatever is orthogonal to the subspace.

Step 1. Fix $N$ and expand the squared error $E_N=\|f-\sum_{n\le N}a_n y_n\|_w^2$ for arbitrary coefficients $a_n$ . Using orthonormality $\langle y_m,y_n\rangle_w=\delta_{mn}$ to collapse the cross terms gives $E_N=\|f\|_w^2-\sum_n\overline{a_n}c_n-\sum_n a_n\overline{c_n}+\sum_n|a_n|^2$ , with $c_n=\langle y_n,f\rangle_w$ . Why: orthonormality is what turns a messy double sum into a clean sum over single modes — without it there is no separation into independent contributions.

Step 2. Complete the square mode by mode: $|a_n|^2-\overline{a_n}c_n-a_n\overline{c_n}=|a_n-c_n|^2-|c_n|^2$ . Substituting, $E_N=\|f\|_w^2-\sum_{n\le N}|c_n|^2+\sum_{n\le N}|a_n-c_n|^2$ . Why: the error splits into a part fixed by $f$ and a part you control, and the controllable part is a sum of squares that can only be reduced, never made negative.

Step 3. The only term you can influence is $\sum_n|a_n-c_n|^2\ge 0$ , minimised uniquely at $a_n=c_n$ . The minimum error is $E_N^{\min}=\|f\|_w^2-\sum_{n\le N}|c_n|^2\ge 0$ , which rearranges to Bessel's inequality $\sum_{n\le N}|c_n|^2\le\|f\|_w^2$ . Why: this proves in one line both that the Fourier coefficients are optimal and that the captured energy is bounded by the total — Bessel is just the non-negativity of a leftover.

Step 4. Let $N\to\infty$ . By definition $\{y_n\}$ is complete iff $E_N^{\min}\to 0$ , i.e. iff $\sum_n|c_n|^2=\|f\|_w^2$ (Parseval) for every $f$ . Why: completeness is thereby recast as a purely quantitative statement — the optimal error is driven to zero — which is what makes it checkable, and equivalent to Parseval rather than a separate mystery. *Verified: standard projection-theorem result, Riley–Hobson–Bence §17; the completed-square identity $E_N=\|f\|_w^2-\sum|c_n|^2+\sum|a_n-c_n|^2$ yields Bessel ( $\le$ ) and, in the limit, Parseval ( $=$ ).*

Go deeper · Derivation C · 1 minThe completeness relation from the resolvent Green function

Derivation CThe completeness relation from the resolvent Green function

The goal is to show that completeness is not an article of faith but a consequence of the analytic structure of the operator's resolvent, by building the Green's function $G(x,x';\lambda)$ spectrally, reading off its poles and residues, and summing them with a contour integral to recover the delta-function kernel $\delta(x-x')/w(x')$ . The same argument reveals why completeness ultimately rests on compactness: the resolvent is a compact self-adjoint integral operator, and the spectral theorem for such operators supplies an orthonormal eigenbasis of the whole space.

Step 1. Expand the resolvent in the eigenbasis: applying $(\mathcal{L}-\lambda)$ to the ansatz $G=\sum_n y_n(x)\overline{y_n(x')}/(\lambda_n-\lambda)$ and using $\mathcal{L}y_n=\lambda_n y_n$ reproduces $\delta(x-x')/w(x)$ , fixing the coefficients uniquely. Why: this is the continuous analogue of writing a matrix inverse $(A-\lambda I)^{-1}=\sum_n e_n e_n^\dagger/(\lambda_n-\lambda)$ — the resolvent is diagonal in the eigenbasis.

Step 2. Read off the analytic structure: $G$ is meromorphic in $\lambda$ with simple poles at each eigenvalue $\lambda_n$ , and $\operatorname{Res}_{\lambda=\lambda_n}G=-\,y_n(x)\overline{y_n(x')}$ . Why: the physics of the spectrum is now encoded as complex-analytic data — where the poles sit is the spectrum, and what their residues are is the eigenprojectors.

Step 3. Integrate $G$ around a large contour $C$ enclosing every pole: $-\tfrac{1}{2\pi i}\oint_C G\,\mathrm{d}\lambda=\sum_n y_n(x)\overline{y_n(x')}$ , the sum of all residues. Why: a contour integral is the natural machine for summing residues, and here it assembles the complete set of eigenprojectors into a single object.

Step 4. Evaluate the same contour integral from the large- $|\lambda|$ behaviour of $G$ : for a regular problem the resolvent tends to the inverse of the identity, whose kernel is $\delta(x-x')/w(x')$ . Equating the two evaluations gives $\sum_n y_n(x)\overline{y_n(x')}=\delta(x-x')/w(x')$ . Why: the two ways of computing one integral must agree, and their agreement is exactly the completeness relation — the resolution of the identity in the eigenbasis. *Verified: Stone & Goldbart §9 and Courant–Hilbert Vol. I; the residue $-y_n(x)\overline{y_n(x')}$ and the closure kernel $\delta(x-x')/w(x')$ are the standard results, with compactness of $G$ supplying completeness via the Hilbert–Schmidt spectral theorem.*

Limiting cases

Every classical orthogonal-function expansion is one instance of the same completeness theorem, distinguished only by interval, weight, and boundary behaviour. The last row is the cautionary instance — where completeness holds but pointwise reconstruction fails.

Case	Result	Meaning
Fourier series	$\mathcal{L}=-\mathrm{d}^2/\mathrm{d}x^2$ , $[0,2\pi]$ , $w=1$ , periodic BC	Eigenfunctions $e^{inx}$ , eigenvalues $n^2$ ; classical completeness in $L^2[0,2\pi]$ .
Fourier–Bessel	$\mathcal{L}y=-\tfrac{1}{x}(xy')'+\tfrac{m^2}{x^2}y$ , $w=x$ , $[0,a]$	Singular at $x=0$ ; eigenfunctions $J_m(\alpha_{mk}\rho/a)$ with $\alpha_{mk}$ the zeros of $J_m$ ; complete on $[0,a]$ .
Legendre	$\mathcal{L}y=-[(1-x^2)y']'$ , $w=1$ , $[-1,1]$	Doubly singular endpoints ( $p=1-x^2$ vanishes); eigenvalues $n(n+1)$ , eigenfunctions $P_n$ , complete in $L^2[-1,1]$ .
Hermite (singular, whole line)	$\mathcal{L}y=-y''+x^2y$ , $w=1$ , $(-\infty,\infty)$	Quantum oscillator; discrete spectrum $2n+1$ , Hermite-function eigenstates — complete despite the infinite interval because the potential confines.
Failure of pointwise convergence	$f=\operatorname{sgn}(x)$ : $S_N$ overshoots by $8.95\%$	Mean-square convergence still holds; the partial-sum overshoot near a jump is fixed at $\approx 8.95\%$ of the jump (Gibbs), narrowing but never shrinking.

Common trapOrthogonality is not completeness, and convergence is not pointwise

Two traps hide in the word "expand." First, orthogonality alone lets you always compute $c_n=\langle y_n,f\rangle_w$ and write $\sum_n c_n y_n$ — but if the set is not complete, that series converges to the orthogonal projection of $f$ onto the closed span of $\{y_n\}$ , which can be a different function from $f$ . Dropping the constant mode from a cosine basis, or an even function's odd partners, silently loses whatever lived in the missing directions. Completeness is the separate guarantee that no direction is missing. Second, even with a complete basis, " $f=\sum_n c_n y_n$ " is a statement about the $L^2_w$ norm, not about every point: at a jump the partial sums overshoot by 8.95% no matter how many terms you keep, and term-by-term differentiation is illegitimate unless the differentiated series converges uniformly. Mean-square equality tolerates disagreement on sets of measure zero and permanent overshoot on shrinking neighbourhoods.

Worked examples

1 · Fourier sine expansion of a constant, and Parseval as the Basel sum (fully worked)

Expand $f(x)=1$ on $[0,\pi]$ in the Dirichlet eigenfunctions $\{\sin nx\}$ of $-\mathrm{d}^2/\mathrm{d}x^2$ , and use Parseval's identity to evaluate $\sum_{n\ \mathrm{odd}}1/n^2$ .

Step 1. The orthonormal basis. On $[0,\pi]$ with $w=1$ , $\int_0^\pi \sin^2(nx)\,\mathrm{d}x=\pi/2$ , so the normalised eigenfunctions are $\varphi_n(x)=\sqrt{2/\pi}\,\sin nx$ . This is the SL problem $-y''=\lambda y$ , $y(0)=y(\pi)=0$ , with $\lambda_n=n^2$ .

Step 2. The coefficients. $c_n=\langle\varphi_n,f\rangle=\sqrt{2/\pi}\int_0^\pi \sin(nx)\,\mathrm{d}x=\sqrt{2/\pi}\,\dfrac{1-\cos n\pi}{n}=\sqrt{2/\pi}\,\dfrac{1-(-1)^n}{n}$ . For even $n$ this vanishes; for odd $n$ it is $c_n=\sqrt{2/\pi}\,(2/n)$ . In classical (un-normalised) form, $1=\dfrac{4}{\pi}\sum_{n\ \mathrm{odd}}\dfrac{\sin nx}{n}$ — the square-wave series drawn in the second figure.

Step 3. Parseval's ledger. The norm is $\|f\|^2=\int_0^\pi 1\,\mathrm{d}x=\pi$ . The coefficient energy is $\sum_n|c_n|^2=\sum_{n\ \mathrm{odd}}\dfrac{2}{\pi}\cdot\dfrac{4}{n^2}=\dfrac{8}{\pi}\sum_{n\ \mathrm{odd}}\dfrac{1}{n^2}$ . Completeness demands these be equal:

\frac{8}{\pi}\sum_{n\ \mathrm{odd}}\frac{1}{n^2}=\pi \;\Longrightarrow\; \sum_{n\ \mathrm{odd}}\frac{1}{n^2}=\frac{\pi^2}{8}.

Step 4. Sanity and consequence. Numerically $\sum_{n\ \mathrm{odd}}1/n^2=1.2337\ldots$ and $\pi^2/8=1.23370\ldots$ — agreement to five digits. Splitting all integers into odd and even, $\sum_{n\ge1}1/n^2=\sum_{\mathrm{odd}}+\tfrac14\sum_{n\ge1}$ gives $\sum 1/n^2=\tfrac43\cdot\tfrac{\pi^2}{8}=\pi^2/6$ , the Basel value. Completeness of the sine basis literally computes $\sum_{\mathrm{odd}}1/n^2=\pi^2/8$ ; yet the same series overshoots the constant near the endpoints by 8.95% — mean-square equality, not pointwise.

2 · Heat conduction in a rod: modes as a low-pass filter (fully worked)

A 10 cm aluminium rod ( $\kappa=9.7\times10^{-5}\ \mathrm{m^2/s}$ ) is held at $0\,^\circ\mathrm{C}$ at both ends. Its initial temperature profile $f(x)$ is expanded in $\sin(n\pi x/L)$ . Find the decay time of each mode and the ratio of the first two.

Step 1. Separation and completeness. $u_t=\kappa u_{xx}$ with $u(0,t)=u(L,t)=0$ separates into the SL problem $-X''=\lambda X$ , eigenfunctions $\sin(n\pi x/L)$ , eigenvalues $\lambda_n=(n\pi/L)^2$ , weight $w=1$ . Completeness in $L^2[0,L]$ guarantees any initial $f$ expands as $f(x)=\sum_n b_n\sin(n\pi x/L)$ , and the evolved solution is $u(x,t)=\sum_n b_n e^{-\kappa\lambda_n t}\sin(n\pi x/L)$ .

Step 2. Eigenvalues. With $L=0.10\ \mathrm{m}$ , $\lambda_n=(n\pi/0.10)^2=n^2\pi^2\times100\ \mathrm{m^{-2}}$ : $\lambda_1=986.96$ , $\lambda_2=3947.8$ , $\lambda_3=8882.6\ \mathrm{m^{-2}}$ .

Step 3. Decay times $\tau_n=1/(\kappa\lambda_n)$ . $\tau_1=1/(9.7\times10^{-5}\times986.96)=\mathbf{10.4\ s}$ ; $\tau_2=\mathbf{2.61\ s}$ ; $\tau_3=\mathbf{1.16\ s}$ . Because $\lambda_n\propto n^2$ , the decay accelerates quadratically with mode number.

Step 4. The physics read-off. $\tau_1/\tau_2=\lambda_2/\lambda_1=4$ exactly, and $\tau_1/\tau_3=9$ . Rough features of the initial profile (high $n$ ) wash out in about a second, while the smooth fundamental lingers for ten — diffusion is a low-pass filter, and completeness is what lets us decompose an arbitrary profile into modes each decaying at its own rate $e^{-\kappa\lambda_n t}$ .

3 · A Legendre expansion terminates exactly

Expand $f(x)=x^2$ on $[-1,1]$ in Legendre polynomials, whose SL normalisation is $\int_{-1}^{1}P_n^2\,\mathrm{d}x=2/(2n+1)$ , weight $w=1$ .

Step 1. Coefficient formula. $c_n=\dfrac{2n+1}{2}\displaystyle\int_{-1}^{1}x^2 P_n(x)\,\mathrm{d}x$ . Parity kills every odd $n$ ; and since $x^2$ is a degree-2 polynomial, it is orthogonal to all $P_n$ with $n>2$ , so only $c_0$ and $c_2$ survive.

Step 2. Compute the two survivors with $P_0=1$ , $P_2=(3x^2-1)/2$ . $c_0=\tfrac12\int_{-1}^1 x^2\,\mathrm{d}x=\tfrac12\cdot\tfrac23=\tfrac13$ . $c_2=\tfrac52\int_{-1}^1 x^2\cdot\tfrac{3x^2-1}{2}\,\mathrm{d}x=\tfrac52\cdot\tfrac12\big(3\cdot\tfrac25-\tfrac23\big)=\tfrac52\cdot\tfrac12\cdot\tfrac{8}{15}=\tfrac23$ .

Step 3. Assemble and verify. $\tfrac13 P_0+\tfrac23 P_2=\tfrac13+\tfrac23\cdot\tfrac{3x^2-1}{2}=\tfrac13+\tfrac{3x^2-1}{3}=x^2$ . The series is finite and exact — $x^2=\tfrac13 P_0+\tfrac23 P_2$ — because a polynomial has only finitely many nonzero projections. Completeness is what promises that a non-polynomial $f$ would need the full infinite set, converging in mean square.

Physics of the portrait

Sturm–Liouville theory is where the empirical trigonometric series of the eighteenth century became a theorem: the recognition that Fourier's, Bessel's and Legendre's expansions were not separate tricks but one spectral fact about self-adjoint operators.

Origins · 1 minsturm–liouville theory and how this was discovered

sturm–liouville theory

the spectral theory of self-adjoint second-order operators, 1830s onward · orthogonal eigenfunction expansions and completeness

The theory carries two names because it was one of the great collaborations of nineteenth-century analysis. Joseph Fourier had, from 1807, boldly claimed that any function could be written as a trigonometric series to solve heat conduction — a claim so cavalier that Lagrange and Laplace resisted publishing it, for no one could say in what sense such a series converged or why the coefficients were unique. The rigour came a generation later. Charles-François Sturm, a Genevan working in Paris, published in 1836 a memoir on second-order differential equations that proved the eigenvalues form an increasing sequence and the $n$ -th eigenfunction has exactly $n$ interior zeros — the oscillation theorem visible in the first figure — studying the equations qualitatively rather than solving them. His friend Joseph Liouville supplied the analytic half almost simultaneously (1836–1837), establishing orthogonality, the reality of the spectrum, and the convergence of the eigenfunction expansion. Together they turned Fourier's audacity into structure: a whole class of boundary-value problems whose solutions form a complete orthogonal system. The completeness itself waited for the twentieth century — David Hilbert and Erhard Schmidt (1904–1907) built the spectral theory of symmetric integral operators that made "complete" a proven property via compactness, and Hermann Weyl (1910) extended the theory to singular endpoints and continuous spectra, the setting quantum mechanics would soon require. What began as a physicist's shortcut for a hot rod is now the mathematical backbone of every eigenfunction method in physics.

History

1807

— Fourier submits his memoir expanding heat profiles in trigonometric series

1836

— Sturm and Liouville publish the oscillation, reality, and orthogonality theorems

1899

— Gibbs pins down the roughly 9% overshoot of partial sums at a jump

1907

— Hilbert and Schmidt build the spectral theory of symmetric integral operators

1910

— Weyl classifies singular endpoints, extending completeness to continuous spectra

Going further

Texts. Stone & Goldbart, Mathematics for Physics, Ch. 8 (Linear Differential Operators) and Ch. 9 (Integral Equations) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 8 (Sturm–Liouville Theory) · Hassani, Mathematical Physics, Ch. 19–20 · Byron & Fuller, Mathematics of Classical and Quantum Physics, Ch. 5 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 17 · Courant & Hilbert, Methods of Mathematical Physics, Vol. I, Ch. V–VI · Coddington & Levinson, Theory of Ordinary Differential Equations, Ch. 7–8.

Provenance. Lagrange's identity, the reality/orthogonality argument, the completed-square least-squares identity yielding Bessel and Parseval, and the residue/contour construction of the completeness relation are standard in Stone & Goldbart (Ch. 8–9), Arfken–Weber–Harris (Ch. 8), Hassani (Ch. 19–20), and Courant–Hilbert (Vol. I). A verifier should recompute: (1) Lagrange's identity by two integrations by parts; (2) the conjugation in the first inner-product slot giving real eigenvalues; (3) the Fourier-sine result $\sum_{n\ \mathrm{odd}}1/n^2=\pi^2/8$ from Parseval and hence $\pi^2/6$ ; (4) the heat-mode decay times $\tau_1=10.4$ s, $\tau_2=2.61$ s, $\tau_3=1.16$ s for aluminium with $L=0.10$ m and their exact ratios $4$ and $9$ ; (5) the Legendre coefficients $x^2=\tfrac13P_0+\tfrac23P_2$ ; (6) the Gibbs overshoot constant $\tfrac{2}{\pi}\int_0^\pi\tfrac{\sin t}{t}\,\mathrm{d}t-1\approx 0.0895$ . The Bessel-series normalisation $2/[a^2 J_{m+1}(\alpha_{mk})^2]$ , Legendre eigenvalues $n(n+1)$ , and oscillator eigenvalues $2n+1$ are standard tabulated results.

Prerequisites. Linear Second-Order ODEs · Two-Point Boundary-Value Problems · The Wronskian and Linear Independence · Existence and Uniqueness Theorems for ODEs. Leads to. The Method of Frobenius · Variation of Parameters · Ordinary and Singular Points.

Ask the tutor: Ask me to expand any function you like in a chosen eigenbasis and watch Parseval's ledger balance term by term — or to show exactly where and why the Gibbs overshoot refuses to die.

What is the difference between orthogonality and completeness, in one sentence each?

Orthogonality says the eigenfunctions point in mutually perpendicular directions, so you can compute the component of any function along each one by a single inner-product integral, $c_n=\langle y_n,f\rangle_w$ . Completeness says those directions exhaust the space, so nothing is left over — the components you computed actually rebuild the function, and no nonzero function is perpendicular to all of them. Orthogonality is a property of the eigenfunctions among themselves; completeness is a property of the eigenfunctions relative to the whole function space. You can have the first without the second (drop a mode), and that is exactly when the expansion silently converges to the wrong function.

Why must the inner product carry a weight w(x)?

Because "self-adjoint" is defined relative to an inner product, and a generic second-order operator is symmetric only under one specific measure. When you push a second-order ODE into Sturm–Liouville form $-(py')'+qy=\lambda w y$ , the function $w(x)$ that appears multiplying $\lambda$ is precisely the weight under which $\mathcal{L}$ becomes self-adjoint. Orthogonality then means $\int_a^b y_m y_n\,w\,\mathrm{d}x=0$ , not $\int y_m y_n\,\mathrm{d}x=0$ . This is why Legendre polynomials are orthogonal with $w=1$ , Hermite functions with $w=e^{-x^2}$ , and Bessel functions with $w=\rho$ : each weight is dictated by the operator, not chosen for convenience.

In what sense does the series equal the function, and when can I trust it pointwise?

By default the equality $f=\sum_n c_n y_n$ is in the mean-square ( $L^2_w$ ) sense: the norm of the error tends to zero, but individual points may disagree. Where $f$ is smooth and compatible with the boundary conditions, the convergence upgrades to uniform and you may treat the equality pointwise. Where $f$ has a jump, the partial sums overshoot by a fixed 8.95% of the jump forever (Gibbs), and at the jump itself the series converges to the midpoint of the two one-sided limits. Practical rule: trust pointwise values away from discontinuities, distrust them near jumps, and never differentiate the series term by term unless you have checked the differentiated series converges uniformly.

How does the Green's function "contain" completeness?

The resolvent $G(x,x';\lambda)=\sum_n y_n(x)\overline{y_n(x')}/(\lambda_n-\lambda)$ is the operator's inverse at spectral parameter $\lambda$ , and it is a meromorphic function of $\lambda$ with a simple pole at each eigenvalue whose residue is the corresponding eigenprojector $-y_n(x)\overline{y_n(x')}$ . Summing all the residues with a contour integral assembles the full set of projectors, and for a regular problem that sum equals the identity kernel $\delta(x-x')/w(x')$ — which is the completeness relation. Structurally, completeness holds because $G$ is a compact self-adjoint integral operator, and the spectral theorem for such operators guarantees its eigenfunctions form an orthonormal basis of the whole space.

Does completeness still work on infinite intervals or at singular endpoints?

Yes, though the analysis is more delicate and the spectrum can change character. When $p(x)$ vanishes at an endpoint (Legendre at $x=\pm1$ , Bessel at $x=0$ ) or the interval is infinite (Hermite on the whole line), the explicit boundary condition is often replaced by a requirement of square-integrability or regularity, and Weyl's limit-point/limit-circle classification decides what is needed. If the potential confines — as $x^2$ does for the quantum oscillator — the spectrum stays discrete and the eigenfunctions remain complete despite the infinite domain. If it does not, a continuous spectrum appears and the sum $\sum_n$ becomes an integral $\int\mathrm{d}\lambda$ , but the completeness paradigm survives as the spectral decomposition of the operator.

Why does the Gibbs overshoot never go away, no matter how many terms I keep?

Because the overshoot is not an error that decays; it is a fixed feature of how a smooth partial sum approximates a discontinuous target. Each partial sum $S_N$ is smooth, and near a jump it must swing above and below the target to "turn the corner"; as $N$ grows the swing moves closer to the jump and gets narrower, but its height converges to a constant — the maximum of $S_N$ approaches $1.0895$ times the jump value, an overshoot of $\tfrac{2}{\pi}\int_0^\pi\tfrac{\sin t}{t}\,\mathrm{d}t-1\approx8.95\%$ . This does not contradict mean-square convergence, because the overshoot occupies an ever-shrinking interval, so its contribution to the $L^2$ error still tends to zero. It is the sharpest reminder that completeness controls the averaged error, not the worst-case pointwise error.