Sturm-Liouville Theory

Sturm-Liouville theory is the point where linear algebra breaks free of finite dimensions. A real symmetric matrix has real eigenvalues and an orthonormal eigenbasis in which it becomes diagonal — the spectral theorem. Sturm-Liouville theory promotes that theorem, word for word, to the infinite-dimensional world of functions. The matrix becomes a second-order differential operator $\mathcal{L}=-(py')'-qy$ ; the eigenvectors become eigenfunctions; the dot product becomes an integral weighted by a density $w(x)$ ; and the diagonalizing basis becomes a generalized Fourier series. The single hypothesis that carries the whole edifice is self-adjointness: the demand that $\mathcal{L}$ be symmetric under the weighted inner product. The first surprise is that any nondegenerate second-order linear ODE can be dragged into this form. Multiply $P y''+Q y'+R y=0$ by the right integrating factor and it becomes the self-adjoint $(py')'+qy+\lambda w y=0$ , with $p$ , $q$ and the weight $w$ read off from $P,Q,R$ . Then one integration by parts (Green's identity) shows the operator is symmetric provided a boundary term vanishes, and the boundary conditions that kill that term — separated, periodic, or a regularity demand at a singular endpoint — are precisely the ones that deliver real eigenvalues and orthogonal eigenfunctions. The payoff is enormous. The regular problem has a discrete, real, increasing spectrum $\lambda_0<\lambda_1<\cdots\to+\infty$ whose $n$ -th eigenfunction has exactly $n$ interior zeros, and the eigenfunctions are complete, so any reasonable function expands in them. Essentially every special function of physics — Legendre, Bessel, Hermite, Laguerre, Chebyshev — is the eigenfunctions of some Sturm-Liouville problem, their orthogonality a one-line corollary. This is the ODE backbone of separation of variables, of quantum bound states, and of the spectral theory of self-adjoint operators.

Key takeaways

A Sturm-Liouville problem is a second-order ODE in the self-adjoint form $\frac{\mathrm{d}}{\mathrm{d}x}\!\big(p\,y'\big)+q\,y+\lambda\,w\,y=0$ on $[a,b]$ , with self-adjoint (separated, periodic, or singular) boundary conditions.
Any nondegenerate $Py''+Qy'+Ry=0$ reaches this form via the integrating factor $p=\exp\!\int(Q/P)\,\mathrm{d}x$ , whereupon the weight is $w=p/P$ .
Self-adjointness under $\langle f,g\rangle_w=\int_a^b f^*g\,w\,\mathrm{d}x$ forces the eigenvalues to be real, distinct-eigenvalue eigenfunctions to be $w$ -orthogonal, and (regular case) each eigenvalue to be simple.
The regular spectrum is discrete and increasing, $\lambda_0<\lambda_1<\cdots\to+\infty$ , and the oscillation theorem says $\phi_n$ has exactly $n$ interior zeros.
The eigenfunctions are complete in $L^2_w[a,b]$ , so any admissible $f$ has a convergent generalized Fourier expansion $f=\sum_n c_n\phi_n$ with $c_n=\langle\phi_n,f\rangle_w/\lVert\phi_n\rVert_w^2$ .
Key idea: cast the ODE in self-adjoint form $-(py')'-qy=\lambda w y$ ; then $\mathcal{L}$ is Hermitian under $\langle f,g\rangle_w=\int f^*g\,w\,\mathrm{d}x$ , and self-adjointness alone guarantees real eigenvalues, $w$ -orthogonal eigenfunctions, and a complete eigenbasis — the spectral theorem for differential operators.

The core insight

Cast the ODE in self-adjoint form $-(py')'-qy=\lambda w y$ ; then the operator is Hermitian under $\langle f,g\rangle_w=\int f^*g\,w\,\mathrm{d}x$ , and self-adjointness alone guarantees real eigenvalues, $w$ -orthogonal eigenfunctions, and a complete eigenbasis — the spectral theorem for differential operators.

Where this lives today

Sturm-Liouville theory is quietly load-bearing across modern computation and physics. High-accuracy eigenvalue solvers such as SLEIGN2 and MATSLISE compute regular and singular spectra to ten or more digits, powering quantum-well design, molecular vibrational spectra, and waveguide-mode analysis in photonics, where the transverse modes of an optical fibre are literally SL eigenfunctions. The oscillation theorem is used numerically as a "shooting count": sweep $\lambda$ , count sign changes, and each new node marks the next state. Inverse Sturm-Liouville theory — reconstructing $q(x)$ from the spectrum via Gelfand-Levitan-Marchenko — underlies quantum inverse scattering, the integrability of the Korteweg-de Vries equation, and geophysical imaging. On the discrete side, the graph Laplacian behind spectral clustering and manifold learning is the finite-difference SL operator, so machine learning invokes eigenfunction completeness every time it embeds data in a Laplacian's leading eigenvectors. And in fundamental physics the framework is exactly the spectral theory of self-adjoint Schrödinger operators, where essential self-adjointness decides whether a quantum Hamiltonian even has well-defined dynamics.

The physical picture

Take a matrix intuition and stretch it until it becomes a differential operator. A real symmetric matrix $A=A^{\mathsf T}$ has real eigenvalues and mutually orthogonal eigenvectors along which $A$ merely stretches. The Sturm-Liouville operator $-(py')'-qy$ is that symmetric matrix in the continuum limit: its "eigenvectors" are functions, its "components" are the values $y(x)$ , and the dot product $\sum_i u_i v_i$ becomes the integral $\int u\,v\,w\,\mathrm{d}x$ . The weight $w$ is the continuum analogue of the right measure on the index set — the density against which orthogonality is judged, which is why the naive integral $\int u\,v\,\mathrm{d}x$ gives the wrong answer for Hermite or Laguerre functions. The eigenfunctions are standing waves, and the boundary conditions pin their nodes exactly as a guitar string's length fixes its harmonics. Larger eigenvalue means faster oscillation means more nodes — the oscillation theorem, the most visual fact in the subject. The first figure draws it for the prototype, the vibrating string, where the eigenvalues are exactly $\lambda_n=(n\pi/L)^2$ .

Figure

Eigenfunctions are standing waves and higher eigenvalues pack in more nodes — for the vibrating string −y″ = λy on the interval of length L = 1 with y = 0 at both ends, the spectrum is λn = n²π²: λ₁ = π² ≈ 9.87 (0 interior nodes), λ₂ = 4π² ≈ 39.5 (1 node), λ₃ = 9π² ≈ 88.8 (2 nodes)

The weight $w$ also encodes the geometry: $w=x$ for a circular drumhead (Bessel), $w=e^{-x^2}$ for the quantum oscillator (Hermite). Completeness is the deep, non-elementary part — the claim that these standing-wave shapes can reconstruct any reasonable function, exactly as Fourier sines do. Indeed ordinary Fourier analysis is just the case $p=w=1$ , $q=0$ ; Sturm-Liouville theory is "Fourier analysis for an arbitrary second-order operator." The second figure makes the matrix-to-operator promotion explicit and shows the discrete eigenvalue ladder of the Legendre operator.

Figure

The spectral theorem promoted to functions — a real symmetric matrix A = Aᵀ has real eigenvalues and orthonormal eigenvectors, and the Sturm-Liouville operator inherits exactly this structure; for the Legendre operator −d/dx((1−x²)y′) = λy the ladder is λ = ℓ(ℓ+1) = 0, 2, 6, 12, 20 for ℓ = 0…4, with the ℓ-th eigenfunction Pℓ carrying ℓ interior zeros

The core equations

The whole theory is generated by one canonical form and the inner product that makes it symmetric. Everything downstream — reality, orthogonality, completeness, the variational principle — is a consequence squeezed out of these few lines. The self-adjoint (Sturm-Liouville) form is the normal form every second-order eigenvalue problem is driven toward. It packages the operator so that a single integration by parts exposes its symmetry.

\frac{\mathrm{d}}{\mathrm{d}x}\!\left(p(x)\frac{\mathrm{d}y}{\mathrm{d}x}\right)+q(x)\,y+\lambda\,w(x)\,y=0 \qquad (1)

where: $p(x)$ — leading coefficient; the regular case requires $p>0$ and $p\in C^1[a,b]$ ; $q(x)$ — zeroth-order (potential) coefficient, continuous on $[a,b]$ ; $w(x)$ — weight (density), $w>0$ in the regular case, which sets the inner product; $\lambda$ — the spectral (eigenvalue) parameter; $y(x)$ — the eigenfunction. Reading: the operator $\mathcal{L}y=-(py')'-qy$ returns $\lambda$ times $y$ , weighted by $w$ — an eigenvalue equation with a measure attached. Any nondegenerate second-order linear ODE reaches form (1) after multiplication by an integrating factor. This is why the theory is universal rather than a special trick for a few named equations.

P y''+Q y'+R y=0\ \Longrightarrow\ p=\exp\!\int\frac{Q}{P}\,\mathrm{d}x,\quad w=\frac{p}{P},\quad q=\frac{R}{P}\,p \qquad (2)

where: $P,Q,R$ — coefficients of the general equation ( $P\neq0$ on the interior); $p$ — the integrating factor, itself the SL leading coefficient; $w=p/P$ — the weight inherited by the $\lambda$ -term. Reading: one integral yields the triple $(p,q,w)$ ; the weight is never chosen but dictated by the equation. The weight defines the Hilbert space $L^2_w[a,b]$ and its inner product — the arena in which $\mathcal{L}$ is symmetric. Change the weight and you change what "orthogonal" means.

\langle f,g\rangle_w=\int_a^b f^*(x)\,g(x)\,w(x)\,\mathrm{d}x,\qquad \lVert f\rVert_w^2=\langle f,f\rangle_w \qquad (3)

where: $f^*$ — complex conjugate of $f$ ; $w(x)$ — the weight from (1); $\lVert f\rVert_w$ — the weighted norm. Reading: orthogonality, normalization, and expansion coefficients all carry the factor $w$ , not $\mathrm{d}x$ alone. Green's (Lagrange) identity is the engine of the subject: two integrations by parts turn $f^*\mathcal{L}g-g\,\mathcal{L}f^*$ into a pure boundary term, the bilinear concomitant.

\int_a^b\!\left(f^*\mathcal{L}g-g\,\mathcal{L}f^*\right)\mathrm{d}x=-\Big[\,p\,\big(f^*g'-g\,f^{*\prime}\big)\Big]_a^b \qquad (4)

where: $\mathcal{L}$ — the SL operator $-(py')'-qy$ ; the bracket — the concomitant at the endpoints. Reading: $\mathcal{L}$ is self-adjoint precisely when this boundary term vanishes — a statement about the domain (the boundary conditions), not just the formula for $\mathcal{L}$ . The conditions that annihilate the concomitant in the regular case are separated — one relation at each endpoint. Periodic conditions and, for singular problems, boundedness or decay do the same job.

\alpha_1\,y(a)+\alpha_2\,y'(a)=0,\qquad \beta_1\,y(b)+\beta_2\,y'(b)=0 \qquad (5)

where: $\alpha_i,\beta_i$ — real constants with $(\alpha_1,\alpha_2)\neq(0,0)$ and $(\beta_1,\beta_2)\neq(0,0)$ . Reading: each condition mixes value and slope at one endpoint; because both eigenfunctions obey the same relations, the concomitant vanishes at $a$ and $b$ separately. The headline result: eigenfunctions of distinct eigenvalues are orthogonal in the weighted product — not by assumption but forced by self-adjointness.

(\lambda_m-\lambda_n)\int_a^b \phi_m^*\,\phi_n\,w\,\mathrm{d}x=0\ \Longrightarrow\ \langle\phi_m,\phi_n\rangle_w=0\quad(m\neq n) \qquad (6)

where: $\phi_n$ — the eigenfunction with eigenvalue $\lambda_n$ ; $\lambda_m\neq\lambda_n$ — distinct eigenvalues. Reading: since the prefactor is nonzero, the weighted overlap integral itself must vanish. Completeness upgrades orthogonality to a basis: any admissible $f$ is the weighted sum of its projections onto the eigenfunctions — a generalized Fourier series.

f(x)=\sum_{n=0}^{\infty} c_n\,\phi_n(x),\qquad c_n=\frac{\langle\phi_n,f\rangle_w}{\langle\phi_n,\phi_n\rangle_w} \qquad (7)

where: $c_n$ — the expansion (Fourier) coefficient, the projection onto $\phi_n$ ; $\{\phi_n\}$ — the complete orthogonal eigenbasis of $L^2_w[a,b]$ . Reading: this is exactly the "component along a basis vector," with the weighted inner product doing the projecting. The Rayleigh quotient both proves reality of $\lambda$ and gives a variational estimate of the lowest eigenvalue — the workhorse of numerical spectral estimation.

\lambda_0\le R[y]=\frac{\displaystyle\int_a^b\!\big(p\,|y'|^2-q\,|y|^2\big)\,\mathrm{d}x-\big[p\,y^*y'\big]_a^b}{\displaystyle\int_a^b w\,|y|^2\,\mathrm{d}x} \qquad (8)

where: $y$ — any admissible trial function satisfying the boundary conditions; equality holds only at $y=\phi_0$ . Reading: the numerator is real for self-adjoint BCs, so $\lambda\in\mathbb{R}$ ; minimizing $R[y]$ bounds the ground eigenvalue from above. Finally, completeness as a resolution of the identity gives the closure relation and, dividing by $\lambda_n-\lambda$ , the spectral Green's function that solves inhomogeneous problems.

\sum_{n}\frac{\phi_n(x)\,\phi_n^*(\xi)}{\lVert\phi_n\rVert_w^2}=\frac{\delta(x-\xi)}{w(\xi)},\qquad G(x,\xi)=\sum_{n}\frac{\phi_n(x)\,\phi_n^*(\xi)}{(\lambda_n-\lambda)\,\lVert\phi_n\rVert_w^2} \qquad (9)

where: $\delta$ — the Dirac delta, the $1/w$ reflecting the weighted measure; $G(x,\xi)$ — the kernel solving $(\mathcal{L}-\lambda w)y=f$ via $y(x)=\int G(x,\xi)f(\xi)\,\mathrm{d}\xi$ . Reading: the poles of $G$ at $\lambda=\lambda_n$ are resonances — the operator is non-invertible exactly on its spectrum.

Derivation AReducing any second-order ODE to self-adjoint form

The goal is to show that self-adjoint form is not a lucky feature of a few named equations but a normal form reachable by every nondegenerate second-order linear ODE, and to derive the recipe $p=\exp\!\int(Q/P)\,\mathrm{d}x$ , $w=p/P$ , $q=(R/P)p$ . We demand that the leading two terms collapse into the single package $(py')'$ and find that the required multiplier is exactly a first-order integrating factor.

Step 1. Start from $Py''+Qy'+Ry=0$ and seek a positive $p(x)$ so that multiplying by $\mu=p/P$ turns the first two terms into $(py')'=py''+p'y'$ . Matching $y''$ is automatic; matching $y'$ demands $p'=(p/P)Q$ , i.e. $p'/p=Q/P$ . Why: the whole point of self-adjoint form is that the second-order part is a perfect derivative — the operator is the "gradient of a flux" $py'$ .

Step 2. Integrate $p'/p=Q/P$ : $p(x)=\exp\!\int^x (Q(t)/P(t))\,\mathrm{d}t$ , positive wherever the integral is finite. Why: this is literally the integrating factor of first-order linear theory, so those prerequisites feed straight into Sturm-Liouville.

Step 3. Multiply through by $\mu=p/P$ to get $(py')'+(R/P)p\,y=0$ , so $q=(R/P)p$ ; the eigenvalue term $\lambda y$ is multiplied by the same $\mu$ , giving it the weight $w=p/P$ . Why: the weight is never a modeling choice — it is dictated by $p/P$ , which is exactly why forgetting it (using $\mathrm{d}x$ for Hermite or Laguerre) breaks orthogonality.

Step 4. Verify on Legendre $(1-x^2)y''-2xy'+\ell(\ell+1)y=0$ : with $P=1-x^2$ , $Q=-2x$ , $p=\exp\!\int\!\frac{-2x}{1-x^2}\mathrm{d}x=1-x^2$ , giving $w=1$ and $\lambda=\ell(\ell+1)$ . The endpoints $x=\pm1$ send $p\to0$ , so this is a singular problem with boundedness as its condition. Why: the recipe reproduces the textbook Legendre form and flags where regular theory hands off to singular. *As a check: for Legendre the recipe yields $p=1-x^2$ and $w=1$ , and for Hermite $y''-2xy'+2ny=0$ it gives $p=w=e^{-x^2}$ , matching the standard weight.*

Go deeper · Derivation B · 1 minReal eigenvalues and weighted orthogonality from self-adjointness

Derivation BReal eigenvalues and weighted orthogonality from self-adjointness

The goal is to extract, from Green's identity alone, the two facts that make Sturm-Liouville theory the spectral theorem for differential operators: every eigenvalue is real, and eigenfunctions of distinct eigenvalues are $w$ -orthogonal. We feed two eigenfunctions into the identity, use the boundary conditions to kill the boundary term, and read off reality from the diagonal case and orthogonality from the off-diagonal case.

Step 1. Let $\mathcal{L}\phi_n=\lambda_n w\,\phi_n$ and $\mathcal{L}\phi_m=\lambda_m w\,\phi_m$ , both obeying the same self-adjoint boundary conditions. Because $p,q,w$ are real, conjugating the second gives $\mathcal{L}\phi_m^*=\lambda_m^* w\,\phi_m^*$ . Why: reality of the coefficients lets the complex conjugate of an eigenfunction stay an eigenfunction, with the conjugated eigenvalue.

Step 2. Apply Green's identity to $\phi_m^*$ and $\phi_n$ : the left side is $\int_a^b(\phi_m^*\mathcal{L}\phi_n-\phi_n\mathcal{L}\phi_m^*)\,\mathrm{d}x$ and the right side is the concomitant $[p(\phi_m^*\phi_n'-\phi_n\phi_m^{*\prime})]_a^b$ , which vanishes since both satisfy the same conditions at $a$ and $b$ . Why: this is the single place the boundary conditions enter — the hypothesis, without which the conclusion collapses.

Step 3. Substitute the eigenvalue equations into the (now zero) left side: $(\lambda_n-\lambda_m^*)\langle\phi_m,\phi_n\rangle_w=0$ . Why: the operator is eliminated in favor of eigenvalues and one weighted overlap — the rest is algebra.

Step 4. Set $m=n$ : $(\lambda_n-\lambda_n^*)\lVert\phi_n\rVert_w^2=0$ , and since $w>0$ makes $\lVert\phi_n\rVert_w^2>0$ , we get $\lambda_n=\lambda_n^*$ — every eigenvalue is real. Why: this is the continuum echo of "a symmetric matrix has real eigenvalues," and it makes the eigenvalues meaningful energies or frequencies.

Step 5. Take $m\neq n$ with $\lambda_m\neq\lambda_n$ . Reality gives $\lambda_m^*=\lambda_m$ , so $(\lambda_n-\lambda_m)\langle\phi_m,\phi_n\rangle_w=0$ with a nonzero prefactor, forcing $\langle\phi_m,\phi_n\rangle_w=0$ . Why: distinct eigenvalues guarantee orthogonality automatically; a degenerate eigenvalue is handled by Gram-Schmidt within its eigenspace, as for matrices. *As a check: setting $w=1$ recovers the standard Fourier orthogonality $\int_0^L\sin(m\pi x/L)\sin(n\pi x/L)\,\mathrm{d}x=0$ .*

Go deeper · Derivation C · 1 minThe Sturm oscillation and comparison theorem

Derivation CThe Sturm oscillation and comparison theorem

The goal is to prove the node-counting law that gives the spectrum its shape: ordering $\lambda_0<\lambda_1<\cdots$ , the $n$ -th eigenfunction has exactly $n$ interior zeros, with a larger eigenvalue oscillating faster. We compare two solutions at different eigenvalues, build a Wronskian-like quantity with a sign-definite derivative, and integrate between consecutive zeros to trap a new zero of the higher solution.

Step 1. Take $-(py_1')'-qy_1=\lambda_1 w y_1$ and $-(py_2')'-qy_2=\lambda_2 w y_2$ with $\lambda_2>\lambda_1$ . Multiply the first by $y_2$ , the second by $y_1$ , and subtract the first from the second; the $q$ -terms cancel and the result telescopes into $\frac{\mathrm{d}}{\mathrm{d}x}\big[p(y_2y_1'-y_1y_2')\big]=(\lambda_2-\lambda_1)\,w\,y_1y_2$ . Why: the left side is a total derivative, so integrating it leaves only boundary contributions — the same mechanism as Green's identity.

Step 2. Let $x_k<x_{k+1}$ be consecutive zeros of $y_1$ . Integrating over $[x_k,x_{k+1}]$ and using $y_1=0$ there, the concomitant collapses to $\big[p\,y_2\,y_1'\big]_{x_k}^{x_{k+1}}=(\lambda_2-\lambda_1)\int_{x_k}^{x_{k+1}} w\,y_1y_2\,\mathrm{d}x$ . Why: putting the limits at zeros of $y_1$ strips the boundary term to a single controllable sign.

Step 3. Suppose $y_2$ has no zero in $(x_k,x_{k+1})$ , say $y_2>0$ with $y_1>0$ there too. Then the right side is strictly positive, but $y_1'(x_k)>0$ and $y_1'(x_{k+1})<0$ make the left side $\le0$ . Why: the slope of $y_1$ flips sign between adjacent zeros, putting the two sides in irreconcilable conflict.

Step 4. The contradiction forces $y_2$ to vanish in $(x_k,x_{k+1})$ : between any two zeros of the lower solution lies a zero of the higher one. Why: this is the comparison theorem — larger $\lambda$ interlaces extra nodes, so raising the eigenvalue can only increase the zero count.

Step 5. With the boundary conditions fixing the count at the ends, induction gives a discrete increasing spectrum $\lambda_0<\lambda_1<\cdots\to+\infty$ in which $\phi_n$ has exactly $n$ interior zeros. Why: this is the oscillation theorem — why a quantum ground state is nodeless while the $n$ -th excited state has $n$ nodes. *As a check: the string eigenfunctions $\sin(n\pi x/L)$ ( $n=1,2,\dots$ ) carry $n-1$ interior zeros, matching the oscillation theorem under the $\phi_0=\sin(\pi x/L)$ indexing convention.*

Limiting cases

Choosing $p,q,w$ and the interval specializes the one framework to every classical equation of physics; each row is a genuine Sturm-Liouville problem in disguise.

Case	Result	Meaning
Fourier / vibrating string	$-y''=\lambda y,\ \ p=w=1,\ q=0,\ [0,L]$	eigenfunctions $\sin(n\pi x/L)$ , eigenvalues $(n\pi/L)^2$ ; the regular prototype giving ordinary Fourier series
Legendre	$-\frac{\mathrm{d}}{\mathrm{d}x}\!\big[(1-x^2)y'\big]=\lambda y,\ w=1,\ [-1,1]$	singular at $x=\pm1$ where $p=0$ ; $\lambda=\ell(\ell+1)$ , eigenfunctions $P_\ell$ ; boundary condition is boundedness
Hermite	$-(e^{-x^2}y')'=2n\,e^{-x^2}y,\ w=e^{-x^2},\ (-\infty,\infty)$	singular (infinite interval); eigenfunctions $H_n$ , eigenvalues $2n$ ; the quantum harmonic oscillator
Bessel	$-(x y')'+\frac{\nu^2}{x}y=\lambda x\,y,\ w=x,\ [0,R]$	singular at $x=0$ ; eigenfunctions $J_\nu(\sqrt{\lambda}\,x)$ , spectrum set by the zeros of $J_\nu$ at $x=R$
Laguerre	$-(x e^{-x}y')'=n\,e^{-x}y,\ w=e^{-x},\ [0,\infty)$	singular; eigenfunctions $L_n$ , eigenvalues $n$ ; the radial hydrogen problem

Common traporthogonality without the weight, and self-adjointness without the domain

Two mistakes sink most Sturm-Liouville calculations. The first is dropping the weight: orthogonality is with respect to $\int_a^b f^*g\,w\,\mathrm{d}x$ , not the naive $\int f^*g\,\mathrm{d}x$ . Use $w=1$ for Hermite (true weight $e^{-x^2}$ ) or Laguerre ( $e^{-x}$ ) and you get wrong overlaps and expansions that fail to converge — indeed $\int H_mH_n\,\mathrm{d}x$ diverges outright. The second is treating self-adjointness as a property of the formula $\mathcal{L}=-(py')'-qy$ alone. It is not: $\mathcal{L}$ is only formally symmetric until boundary conditions make the concomitant $[p(f^*g'-gf^{*\prime})]_a^b$ vanish. The same expression with the wrong domain can have complex eigenvalues, an incomplete eigenset, or no discrete spectrum at all. Self-adjointness is operator plus domain, together.

Worked examples

1 · The vibrating string, fully solved (confidence builder)

Solve the regular problem $-y''=\lambda y$ on $[0,1]$ with $y(0)=y(1)=0$ . Find the spectrum, the normalized eigenfunctions, verify orthogonality, and count nodes.

Step 1. Here $p=w=1$ , $q=0$ — the cleanest regular problem. For $\lambda>0$ write $\lambda=k^2$ ; the general solution is $y=A\sin kx+B\cos kx$ , and $y(0)=0$ kills the cosine, $B=0$ . (The Rayleigh quotient rules out $\lambda\le0$ : $\int_0^1|y'|^2>0$ for any nonconstant $y$ vanishing at both ends.)

Step 2. The condition $y(1)=0$ gives $A\sin k=0$ , so $k=n\pi$ for $n=1,2,3,\dots$ The spectrum is therefore $\boldsymbol{\lambda_n=n^2\pi^2}$ : numerically $\lambda_1=\pi^2\approx 9.8696$ , $\lambda_2=4\pi^2\approx 39.478$ , $\lambda_3=9\pi^2\approx 88.826$ — discrete, real, increasing, exactly as the general theory promises.

Step 3. Normalize in the (unweighted, $w=1$ ) inner product: $\int_0^1 A^2\sin^2(n\pi x)\,\mathrm{d}x=A^2/2=1$ gives $A=\sqrt2$ , so the orthonormal eigenfunctions are $\boldsymbol{\phi_n(x)=\sqrt2\,\sin(n\pi x)}$ .

Step 4. Verify orthogonality for $m=1,n=2$ : $\int_0^1 2\sin(\pi x)\sin(2\pi x)\,\mathrm{d}x=0$ (checked numerically to $10^{-17}$ ). Distinct eigenvalues, orthogonal eigenfunctions — the spectral theorem in action.

Step 5. Count nodes: $\phi_n=\sqrt2\sin(n\pi x)$ has interior zeros at $x=1/n,2/n,\dots,(n-1)/n$ , i.e. $n-1$ of them. Indexing the ground state as $\phi_0=\sqrt2\sin(\pi x)$ with zero interior nodes recovers the oscillation-theorem count exactly. Spectrum $\lambda_n=n^2\pi^2$ , eigenfunctions $\sqrt2\sin(n\pi x)$ , node count $= n-1$ .

2 · Rayleigh-Ritz estimate of a ground-state eigenvalue

Estimate the lowest eigenvalue $\lambda_0$ of $-y''=\lambda y$ on $[0,1]$ with $y(0)=y(1)=0$ using the simplest admissible trial function, and compare with the exact value from Example 1.

Step 1. Choose the trial function $y(x)=x(1-x)$ , which satisfies both boundary conditions and is nodeless — the right shape for a ground state. Then $y'=1-2x$ . No sines are assumed; we only exploit that the true $\lambda_0$ minimizes the Rayleigh quotient.

Step 2. Numerator (with $q=0$ , $p=1$ , and vanishing boundary term): $\int_0^1 (y')^2\,\mathrm{d}x=\int_0^1(1-2x)^2\,\mathrm{d}x=\int_0^1(1-4x+4x^2)\,\mathrm{d}x=1-2+\tfrac43=\tfrac13$ .

Step 3. Denominator (with $w=1$ ): $\int_0^1 y^2\,\mathrm{d}x=\int_0^1(x-x^2)^2\,\mathrm{d}x=\int_0^1(x^2-2x^3+x^4)\,\mathrm{d}x=\tfrac13-\tfrac12+\tfrac15=\tfrac{1}{30}$ .

Step 4. The Rayleigh quotient is $R[y]=\dfrac{1/3}{1/30}=\boldsymbol{10}$ . Since $R[y]\ge\lambda_0$ for every admissible trial function, this is a rigorous upper bound. The exact value is $\lambda_0=\pi^2\approx 9.8696$ , so a one-line polynomial guess lands within 1.3 % ( $(10-\pi^2)/\pi^2=1.32\%$ ), always from above — the variational principle behind real numerical eigensolvers. $R[y]=10\ge\lambda_0=\pi^2\approx 9.87$ , error $1.3\%$ .

3 · Reading off the weight: Hermite to self-adjoint form

Put the Hermite equation $y''-2xy'+2ny=0$ into Sturm-Liouville form and identify $p,q,w$ and the orthogonality.

Step 1. Identify $P=1$ , $Q=-2x$ , $R=2n$ . The recipe gives $p=\exp\!\int(-2x)\,\mathrm{d}x=e^{-x^2}$ , so $w=p/P=e^{-x^2}$ and the self-adjoint form is $-\big(e^{-x^2}y'\big)'=2n\,e^{-x^2}y$ . The weight is the Gaussian $e^{-x^2}$ , not $\mathrm{d}x$ .

Step 2. Sturm-Liouville theory then guarantees $\int_{-\infty}^{\infty} H_m(x)H_n(x)\,e^{-x^2}\,\mathrm{d}x=0$ for $m\neq n$ , with normalization $\sqrt{\pi}\,2^n n!$ for $m=n$ (e.g. $n=2$ gives $8\sqrt\pi\approx 14.18$ ). The unweighted integral $\int H_mH_n\,\mathrm{d}x$ diverges — the weight is doing essential work.

Physics of the portrait

Sturm-Liouville theory is one of the rare pieces of mathematical physics that carries two names because it was genuinely built by two people working side by side in 1830s Paris — a collaboration so tight that the founding results appear across a shared run of papers rather than in any single author's memoir.

Origins · 1 minSturm and Liouville and how this was discovered

Sturm and Liouville

Jacques Charles François Sturm 1803-1855 · Joseph Liouville 1809-1882 · the 1836-1837 Paris memoirs founding spectral theory

Jacques Charles François Sturm, born in Geneva, and Joseph Liouville, born in Saint-Omer, were young colleagues in Paris when they created, essentially together, the first spectral theory of differential operators. Sturm came to it from mathematical physics — the flow of heat in a bar — and in his 1836 memoir proved the oscillation and comparison theorems that bear his name, showing that eigenvalues form a discrete increasing sequence without ever solving the equations. Liouville, who had just founded the Journal de Mathématiques Pures et Appliquées in 1836 ("Liouville's Journal"), supplied the analytic backbone: he proved that the eigenfunction expansions converge, establishing the completeness that turns orthogonality into a usable basis. Their joint papers of 1836-1837 fused geometry and analysis into one theory — Sturm the geometer of oscillations, Liouville the analyst of convergence. Its full reach appeared only later, when Weyl extended it to singular endpoints and quantum mechanics revealed Schrödinger's energy levels to be Sturm-Liouville eigenvalues.

History

1836

— Sturm's memoir proves the oscillation and comparison theorems for second-order linear equations

1836

— Liouville founds the Journal de Mathématiques, the venue for the joint work

1837

— Sturm and Liouville establish real eigenvalues, orthogonality, and convergence of the eigenfunction expansion

1910

— Weyl classifies singular endpoints by the limit-point and limit-circle alternative

1926

— Schrödinger's wave equation recasts quantum energy levels as a Sturm-Liouville eigenvalue problem

Going further

Texts. Stone & Goldbart, Mathematics for Physics, Ch. 8 (the cleanest modern treatment) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 8 (self-adjoint ODEs) and Ch. 9 · Hassani, Mathematical Physics, Ch. 19 (Sturm-Liouville systems) · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 17 · Byron & Fuller, Mathematics of Classical and Quantum Physics, Ch. 5 · Courant & Hilbert, Methods of Mathematical Physics, Vol. I, Ch. V-VI · Coddington & Levinson, Theory of Ordinary Differential Equations, Ch. 7-8 (rigorous oscillation theory). Key anchor values: the Legendre normalization is $2/(2\ell+1)$ ; the Hermite, Laguerre, and Bessel weights are $e^{-x^2}$ , $e^{-x}$ , and $x$ respectively. The reduction recipe gives $p=1-x^2$ , $w=1$ for Legendre. Green's identity yields $(\lambda_n-\lambda_m^*)\langle\phi_m,\phi_n\rangle_w=0$ . The oscillation derivation turns on the identity $\frac{\mathrm{d}}{\mathrm{d}x}[p(y_2y_1'-y_1y_2')]=(\lambda_2-\lambda_1)wy_1y_2$ . The string spectrum is $\lambda_n=n^2\pi^2$ and the Rayleigh-Ritz bound gives $R[x(1-x)]=10\ge\pi^2\approx9.8696$ .

Prerequisites. Linear Second-Order ODEs · Two-Point Boundary-Value Problems · The Wronskian and Linear Independence · Integrating Factors and Exact Equations · Variation of Parameters. Leads to. Separation of Variables (PDEs) · Special Functions (Legendre, Bessel, Hermite, Laguerre) · The Method of Frobenius · Spectral Theory of Self-Adjoint Operators.

Ask the tutor: Bring any second-order ODE — a quantum well, a vibrating membrane, a special-function equation — and I will walk you through casting it in self-adjoint form, reading off its weight, and proving its eigenfunctions orthogonal and complete.

What actually makes a Sturm-Liouville operator "self-adjoint" rather than just symmetric?

Two things together: the differential expression must be in the form $-(py')'-qy$ (which any second-order equation can reach), and the functions in its domain must satisfy boundary conditions that annihilate the concomitant $[p(f^*g'-gf^{*\prime})]_a^b$ in Green's identity. The formula alone gives only formal symmetry. Separated conditions, periodic conditions, and (for singular endpoints) boundedness or decay all do the job. Without the right domain the eigenvalues can be complex and the eigenfunctions non-orthogonal or incomplete, so self-adjointness is genuinely a property of the operator and its boundary conditions taken as one object.

Why does the weight function w(x) appear, and how do I know what it is?

The weight is not a modeling choice — it is forced by the equation. Multiplying $Py''+Qy'+Ry=0$ by the integrating factor $p=\exp\!\int(Q/P)\,\mathrm{d}x$ to reach self-adjoint form, the eigenvalue term picks up $w=p/P$ , and that same $w$ defines the inner product $\int f^*g\,w\,\mathrm{d}x$ in which $\mathcal{L}$ is symmetric. So compute $p$ and divide by $P$ : Legendre gives $w=1$ , Hermite $w=e^{-x^2}$ , Bessel $w=x$ . Using the wrong weight breaks orthogonality and can make the natural integrals diverge.

How is this the "spectral theorem" if the space is infinite-dimensional?

The finite-dimensional spectral theorem says a real symmetric matrix has real eigenvalues and an orthonormal eigenbasis. Sturm-Liouville theory reproduces every clause: self-adjointness gives real eigenvalues (Derivation 2, diagonal case), distinct eigenvalues give orthogonal eigenfunctions (off-diagonal case), and completeness — the non-elementary, infinite-dimensional part — supplies the "basis" clause. What is new is that the spectrum can be infinite (still discrete for regular problems) and, for singular problems, can include a continuous part. Otherwise it is the same theorem, promoted from vectors to functions.

Regular versus singular Sturm-Liouville problems — what changes?

A regular problem lives on a finite interval with $p>0$ and $w>0$ throughout, continuous $p,p',q,w$ , and separated boundary conditions; it has the cleanest theorems — a discrete, simple, real spectrum $\lambda_0<\lambda_1<\cdots\to+\infty$ with $\phi_n$ having $n$ interior zeros. A singular problem has $p$ or $w$ vanishing at an endpoint (Legendre at $x=\pm1$ , Bessel at $x=0$ ), an infinite interval (Hermite, Laguerre), or coefficients that blow up. There the explicit boundary condition is replaced by a regularity demand — boundedness or square-integrability — and the spectrum may acquire a continuous component. Weyl's limit-point/limit-circle classification (1910) tells you which case you are in.

Why does the n-th eigenfunction have exactly n nodes?

Because a larger eigenvalue makes solutions oscillate faster, and the boundary conditions ratchet the count up by one at each step. The Sturm comparison theorem (Derivation 3) shows that between any two zeros of a lower-eigenvalue solution lies a zero of a higher-eigenvalue one. Climbing from the nodeless ground state $\phi_0$ , each new eigenvalue interlaces exactly one additional interior zero, giving $\phi_n$ precisely $n$ . In quantum mechanics this is the node theorem: the ground state is nodeless, the first excited state has one node, and so on — you can identify a state just by counting zero-crossings.

How is separation of variables secretly a Sturm-Liouville problem?

When you separate variables in the heat, wave, Laplace, or Schrödinger equation, the spatial factor satisfies a second-order ODE whose separation constant is the eigenvalue — and that ODE, with the boundary conditions inherited from the domain, is a Sturm-Liouville problem. A 1D string gives $-X''=\lambda X$ ; spherical coordinates give Legendre; cylindrical coordinates give Bessel. Completeness is exactly what guarantees that summing the separated solutions can match arbitrary initial and boundary data — the eigenfunctions form a basis, so the coefficients are weighted projections. Without it, separation of variables would be a guess rather than a general method.