Fredholm Integral Equations

An integral equation hides its unknown under an integral sign, and the single most consequential fact about it is where the limits sit. When they are fixed constants $a$ and $b$ — the same domain for every value of the free variable — the equation is of Fredholm type, and its integral operator behaves, spectrally and computationally, almost exactly like a finite matrix. That is the miracle of the subject: an object in an infinite-dimensional function space nonetheless obeys determinants, eigenvalues, and a solvability dichotomy lifted verbatim from first-year linear algebra. Contrast a Volterra equation, whose upper limit is the running variable and whose operator is effectively lower-triangular with no eigenvalues to speak of. Linear Fredholm equations split into two kinds by one structural question: does the unknown appear outside the integral? In the first kind it does not, and solving means inverting a smoothing operator — generically ill-posed, amplifying noise without bound. In the second kind it does, so the equation is a perturbation of the identity, $(I-\lambda K)\phi=f$ , well-posed for all but a discrete set of $\lambda$ . Those exceptional values are the characteristic values, reciprocals of the operator's eigenvalues, and the whole theory organizes around them. Three tools do the work — the Neumann series for small $\lambda$ , the reduction to a matrix for separable kernels, and the spectral expansion for symmetric ones — and together they turn scattering, potential theory, and Sturm–Liouville problems into one language.

Key takeaways

A Fredholm equation relates an unknown $\phi(x)$ to its integral against a kernel $K(x,t)$ over a fixed domain $[a,b]$ — unlike a Volterra equation, whose upper limit is the variable $x$ .
First-kind equations $\int_a^b K\phi\,\mathrm{d}t=f$ put the unknown only inside the integral and are typically ill-posed; second-kind equations $\phi=f+\lambda K\phi$ are perturbations of the identity and well-behaved.
For a square-integrable kernel the second-kind equation is solved by the Liouville–Neumann series in powers of $\lambda$ , which converges geometrically whenever $|\lambda|\,\|K\|_{HS}<1$ .
A degenerate (separable) kernel $K=\sum_n g_n(x)h_n(t)$ collapses the problem exactly onto a finite linear system $(I-\lambda A)\mathbf c=\mathbf b$ ; the characteristic values are the roots of $\det(I-\lambda A)=0$ .
The Fredholm alternative rules solvability: either $(I-\lambda K)\phi=f$ is uniquely solvable for every $f$ , or $f$ must be orthogonal to every solution of the adjoint homogeneous equation; a symmetric kernel gives real characteristic values and orthonormal eigenfunctions.
Key idea: on a fixed domain the operator $K\phi(x)=\int_a^b K(x,t)\phi(t)\,\mathrm{d}t$ is compact, so $(I-\lambda K)\phi=f$ is solvable exactly when $1/\lambda$ is not an eigenvalue of $K$ , and the whole finite-dimensional toolkit — determinants, eigenvalues, the alternative — carries over verbatim.

The core insight

On a fixed domain the integral operator $K\phi(x)=\int_a^b K(x,t)\phi(t)\,\mathrm{d}t$ is compact; the second-kind equation $(I-\lambda K)\phi=f$ is therefore solvable exactly when $1/\lambda$ is not an eigenvalue of $K$ , and the entire finite-dimensional toolkit — determinants, eigenvalues, the Fredholm alternative — carries over verbatim.

Where this lives today

Fredholm theory is quietly load-bearing across computational physics. In quantum scattering the Lippmann–Schwinger equation $\psi=\phi_{\text{in}}+G_0^+U\psi$ is a second-kind Fredholm equation whose Neumann series is the Born series and whose resolvent poles mark bound states; modern few-body and nuclear codes solve it as exactly that. In electromagnetics and acoustics the boundary element method recasts the Dirichlet and Neumann problems as second-kind boundary integral equations, and the alternative reproduces the classical solvability conditions — the backbone of antenna and scattering solvers. First-kind equations govern the inverse problems of imaging — CT inverting the Radon transform, geophysical inversion, image deblurring — all ill-posed, all stabilized by the Tikhonov and truncated-SVD regularization the singular-value analysis demands. And in random matrix theory the Fredholm determinant is the star: the Tracy–Widom distribution and the gap probabilities of determinantal point processes are literally Fredholm determinants of the Airy and sine kernels — the 1903 construction, now the language of universality.

The physical picture

The fastest way to feel what a Fredholm equation is is to discretize it and watch it become a matrix. Sample $[a,b]$ at nodes $t_j$ with weights $w_j$ ; then $\int_a^b K(x,t)\phi(t)\,\mathrm{d}t$ becomes $\sum_j K(x_i,t_j)\,w_j\,\phi(t_j)$ — a matrix $K_{ij}=K(x_i,t_j)w_j$ hitting a vector $\phi_j$ . A first-kind equation is then "solve $K\boldsymbol\phi=\boldsymbol f$ " for a near-singular $K$ ; a second-kind equation is the friendlier "solve $(I-\lambda K)\boldsymbol\phi=\boldsymbol f$ ", a small deformation of the identity. Every theorem below is the continuum limit of a matrix fact, and the first figure fixes the dictionary.

Figure

Discretizing the fixed domain turns the integral into a matrix acting on a vector — ∫ₐᵇ K(x,t) φ(t) dt ≈ Σⱼ K(xᵢ,tⱼ) wⱼ φ(tⱼ) — so the second-kind equation becomes the linear system (I − λK)φ = f and every Fredholm theorem is the continuum limit of a matrix fact

Why is the second kind tame and the first kind treacherous? For a smooth or $L^2$ kernel $K$ is compact, with eigenvalues $\mu_n$ that accumulate only at $0$ . Inverting $K$ alone — the first kind — divides spectral coefficients by those $\mu_n\to0$ , an unbounded operation that magnifies the tiniest noise. But $I-\lambda K$ has eigenvalues $1-\lambda\mu_n$ piling up at $1$ , away from zero except at the discrete characteristic values $\lambda_n=1/\mu_n$ , which march off to infinity. The second figure makes this spectrum concrete for a kernel we solve exactly below.

Figure

Spectrum of the compact operator for the min(x,t) Green kernel on the unit interval — eigenvalues μₙ = 1/(n²π²) accumulate at 0 (μ₁ ≈ 0.1013) with trace Σ μₙ = 1/6 ≈ 0.1667, while the characteristic values λₙ = n²π² run off to infinity (λ₁ ≈ 9.87)

The Fredholm alternative is the survival of that finite-dimensional dichotomy: either $(I-\lambda K)$ is invertible with a unique solution for every $f$ , or it has a nontrivial null space and the problem is solvable only when $f$ is orthogonal to the adjoint null space. There is no continuous-spectrum middle ground — a compact operator does not permit one — which is why the theory feels like linear algebra rather than analysis.

The core equations

The whole subject hangs on one equation and its kin. The second-kind equation places the unknown both outside and inside the integral, making it a perturbation of the identity; the parameter $\lambda$ is the dial whose special values expose the spectrum.

\phi(x)=f(x)+\lambda\int_a^b K(x,t)\,\phi(t)\,\mathrm{d}t \qquad (1)

where: $\phi(x)$ — unknown; $f(x)$ — given source; $K(x,t)$ — kernel on $[a,b]\times[a,b]$ ; $\lambda$ — complex parameter whose inverse acts as an eigenvalue; $[a,b]$ — fixed domain, the defining feature. Reading: $\phi$ is the source plus a $\lambda$ -weighted smearing of itself; setting $f\equiv0$ gives the eigenvalue problem $\phi=\lambda K\phi$ . Drop the outside copy of $\phi$ and you have the first kind, where the unknown survives only under the integral — harmless-looking but the ill-posed member of the family.

\int_a^b K(x,t)\,\phi(t)\,\mathrm{d}t = f(x) \qquad (2)

where: $\phi(t)$ — unknown, appearing only inside the integral. Reading: recovering $\phi$ means inverting the compact smoother $K$ , whose singular values $\sigma_n\to0$ make the inverse unbounded, so errors in $f$ blow up unless one regularizes. For small $|\lambda|$ the second-kind equation is solved by feeding it back into itself, generating the Liouville–Neumann series whose coefficients are the iterated kernels of the operator powers $K^n$ .

\phi(x)=\sum_{n=0}^{\infty}\lambda^n \int_a^b K_n(x,t)\,f(t)\,\mathrm{d}t,\qquad K_1=K,\;\; K_{n}(x,t)=\int_a^b K(x,s)\,K_{n-1}(s,t)\,\mathrm{d}s \qquad (3)

where: $K_n(x,t)$ — the $n$ -th iterated kernel, kernel of $K^n$ ; the $n=0$ term is $f$ . Reading: an expansion of $(I-\lambda K)^{-1}$ that converges below the reciprocal spectral radius. Summing the iterated kernels packages the series into the resolvent kernel, turning the solution into one integral against the data.

\phi(x)=f(x)+\lambda\int_a^b R(x,t;\lambda)\,f(t)\,\mathrm{d}t,\qquad R(x,t;\lambda)=\sum_{n=0}^{\infty}\lambda^{n}K_{n+1}(x,t) \qquad (4)

where: $R(x,t;\lambda)$ — resolvent kernel, meromorphic in $\lambda$ with poles at the characteristic values. Reading: knowing $R$ solves the equation for every source $f$ at once. Fredholm's own construction writes the resolvent as a ratio of two entire functions, extending it past the Neumann disc to the whole $\lambda$ -plane.

R(x,t;\lambda)=\frac{D(x,t;\lambda)}{D(\lambda)},\qquad D(\lambda)=\sum_{n=0}^{\infty}\frac{(-\lambda)^n}{n!}\int_a^b\!\!\cdots\!\!\int_a^b \det\big[K(s_i,s_j)\big]\,\mathrm{d}s_1\cdots \mathrm{d}s_n \qquad (5)

where: $D(\lambda)$ — Fredholm determinant, entire in $\lambda$ ; $D(x,t;\lambda)$ — Fredholm minor; $\det[K(s_i,s_j)]$ — $n\times n$ determinant of kernel values. Reading: the characteristic values are the zeros of $D(\lambda)$ — the continuum analogue of eigenvalues as roots of the characteristic polynomial. A degenerate kernel — a finite sum of products of a function of $x$ and a function of $t$ — makes all of this exact: the integral equation becomes a genuine matrix equation of finite rank.

K(x,t)=\sum_{n=1}^{N} g_n(x)\,h_n(t)\;\Longrightarrow\; (I-\lambda A)\,\mathbf{c}=\mathbf{b},\quad A_{mn}=\int_a^b h_m(t)\,g_n(t)\,\mathrm{d}t,\; b_m=\int_a^b h_m(t)\,f(t)\,\mathrm{d}t \qquad (6)

where: $c_n$ — coefficients in $\phi=f+\lambda\sum_n c_n g_n$ ; $A$ — $N\times N$ matrix of kernel moments; $\mathbf b$ — source moments. Reading: characteristic values are the roots of $\det(I-\lambda A)=0$ , a polynomial of degree $\le N$ . Solvability of the second-kind equation is ruled by the Fredholm alternative, the lift of the rank–nullity dichotomy to compact operators.

\text{Either } (I-\lambda K)\phi=f \text{ is uniquely solvable } \forall f,\; \text{ or } \;\langle f,\psi\rangle=0 \;\;\forall\,\psi:\ (I-\bar\lambda K^\dagger)\psi=0 \qquad (7)

where: $K^\dagger$ — adjoint, kernel $\overline{K(t,x)}$ ; $\psi$ — adjoint homogeneous solution. Reading: when the homogeneous equation has nontrivial solutions, the source must be orthogonal to every adjoint null $\psi$ . Finally, a symmetric (Hermitian) kernel gives a compact self-adjoint operator, and the spectral theorem hands over an orthonormal eigenbasis in which the kernel expands.

K(x,t)=\sum_{n}\frac{\phi_n(x)\,\phi_n(t)}{\lambda_n},\qquad K\phi_n=\frac{1}{\lambda_n}\phi_n,\quad \langle\phi_m,\phi_n\rangle=\delta_{mn} \qquad (8)

where: $\lambda_n$ — characteristic values, real with $|\lambda_n|\to\infty$ ; $\phi_n$ — orthonormal eigenfunctions. Reading: the Hilbert–Schmidt expansion is the continuum version of diagonalizing a symmetric matrix.

Derivation AReduction of a degenerate kernel to linear algebra

The goal here is to show that when the kernel separates into a finite sum $K(x,t)=\sum_{n=1}^N g_n(x)h_n(t)$ , the second-kind Fredholm equation is not merely approximated by a matrix problem but is exactly equivalent to one of size $N\times N$ , the characteristic values appearing as the roots of a determinant. This is the cleanest demonstration of the article's thesis — a Fredholm operator is a matrix — and the workhorse method for hand-solvable examples.

Step 1. Insert the separated kernel into $\phi=f+\lambda K\phi$ : $\phi(x)=f(x)+\lambda\sum_{n=1}^N g_n(x)\int_a^b h_n(t)\phi(t)\,\mathrm{d}t$ . The $x$ -dependence pulls out, leaving only the $N$ numbers $c_n\equiv\int_a^b h_n(t)\phi(t)\,\mathrm{d}t$ unknown. Why: separability freezes the $x$ -dependence into the known $g_n$ , converting a search over a function space into a search over $\mathbb{C}^N$ .

Step 2. Read off the solution's shape: $\phi(x)=f(x)+\lambda\sum_{n=1}^N c_n\,g_n(x)$ — the unknown is $f$ plus a known finite combination of the $g_n$ . Why: the correction $\phi-f$ lives in the finite-dimensional span of the $g_n$ — exactly "finite rank".

Step 3. Close the system: multiply by $h_m(x)$ and integrate, giving $c_m=b_m+\lambda\sum_n A_{mn}c_n$ with $b_m=\int_a^b h_m f$ , $A_{mn}=\int_a^b h_m g_n$ . Why: taking moments against the $h_m$ is the projection that turns the operator equation into scalar equations.

Step 4. Assemble the matrix form $(I-\lambda A)\mathbf c=\mathbf b$ ; by Cramer's rule it has a unique $\mathbf c$ — hence a unique $\phi$ — precisely when $\det(I-\lambda A)\ne0$ . Why: solvability of the infinite-dimensional problem reduces to invertibility of a finite matrix.

Step 5. Locate the characteristic values as the roots of $\det(I-\lambda A)=0$ : there the homogeneous system acquires a nontrivial $\mathbf c$ , so the homogeneous integral equation has a nontrivial solution — the alternative made explicit. Why: the determinant has degree $\le N$ , so a rank- $N$ kernel has at most $N$ characteristic values.

Go deeper · Derivation B · 1 minSpectral expansion of a symmetric kernel

Derivation BSpectral expansion of a symmetric kernel

The aim is to solve the second-kind equation for a real symmetric kernel $K(x,t)=K(t,x)$ by diagonalizing the compact self-adjoint operator it generates and expanding in its eigenbasis. The payoff is a closed-form solution as a sum over eigenmodes, making the poles at the characteristic values and the exact solvability condition when the source hits one visible at a glance — the integral-equation face of the spectral theorem and the bridge to Sturm–Liouville theory.

Step 1. Record the spectral data. Compactness plus self-adjointness (the Hilbert–Schmidt theorem) give a complete orthonormal eigenbasis with real eigenvalues, $K\phi_n=\tfrac{1}{\lambda_n}\phi_n$ , $\langle\phi_m,\phi_n\rangle=\delta_{mn}$ , the $\lambda_n$ real with $|\lambda_n|\to\infty$ . Why: symmetry forbids complex eigenvalues and Jordan blocks, so the operator is genuinely diagonal in an orthonormal basis.

Step 2. Expand the unknown and source: $\phi=\sum_n\alpha_n\phi_n$ , $f=\sum_n f_n\phi_n$ with $f_n=\langle f,\phi_n\rangle$ ; completeness makes the expansions exact. Why: in an eigenbasis the operator acts componentwise, uncoupling the equation into scalars.

Step 3. Project $\phi=f+\lambda K\phi$ onto $\phi_n$ . Since $K\phi=\sum_n(\alpha_n/\lambda_n)\phi_n$ , matching coefficients gives $\alpha_n=f_n+\lambda\,\alpha_n/\lambda_n$ . Why: the operator is now just multiplication by $1/\lambda_n$ on the $n$ -th component.

Step 4. Solve each scalar: $\alpha_n=\dfrac{\lambda_n f_n}{\lambda_n-\lambda}$ , valid whenever $\lambda\ne\lambda_n$ ; the vanishing denominator at $\lambda=\lambda_n$ is a characteristic value. Why: the solution's pole structure in the complex $\lambda$ -plane sits precisely on the spectrum.

Step 5. Reassemble with $\lambda_n/(\lambda_n-\lambda)=1+\lambda/(\lambda_n-\lambda)$ and completeness: $\phi(x)=f(x)+\lambda\sum_n\dfrac{\langle f,\phi_n\rangle}{\lambda_n-\lambda}\,\phi_n(x)$ . At $\lambda=\lambda_k$ this blows up unless $\langle f,\phi_k\rangle=0$ — the Fredholm orthogonality condition. Why: one formula carries both the generic solution and, in its breakdown, the exact obstruction to solvability.

Go deeper · Derivation C · 1 minConvergence of the Neumann series

Derivation CConvergence of the Neumann series

The objective is to prove that the Liouville–Neumann series really converges and to pin its radius of convergence to a computable property of the kernel — the Hilbert–Schmidt norm. The argument is a chain of Cauchy–Schwarz estimates bounding the iterated kernels geometrically, so that for small $|\lambda|$ the resolvent series sums to a genuine solution, with analytic continuation taking over beyond the disc.

Step 1. Bound one application of $K$ in $L^2$ . Cauchy–Schwarz in $t$ gives $\big|\int_a^b Kf\,\mathrm{d}t\big|^2\le\big(\int_a^b|K|^2\mathrm{d}t\big)\|f\|^2$ ; integrating over $x$ , $\|Kf\|\le B\|f\|$ with $B^2=\|K\|_{HS}^2=\int_a^b\!\int_a^b|K|^2$ . Why: this turns "compact operator" into a concrete finite constant $B$ we can iterate.

Step 2. Iterate to $\|K^n f\|\le B^n\|f\|$ , since each factor of $K$ multiplies the norm by at most $B\ge\|K\|_{op}$ . Why: geometric growth of the norms is what a geometric series needs, guaranteed once the kernel is square-integrable.

Step 3. Sum the majorant: $\sum_{n=0}^\infty\big\||\lambda|^n K^n f\big\|\le\|f\|\sum_{n=0}^\infty(|\lambda|B)^n=\dfrac{\|f\|}{1-|\lambda|B}$ , convergent for $|\lambda|B<1$ . Why: a convergent majorant certifies the series as a genuine solution operator inside the disc $|\lambda|<1/B$ .

Step 4. Continue past the disc. The series may diverge for larger $|\lambda|$ , but $R=D(x,t;\lambda)/D(\lambda)$ extends the resolvent meromorphically over the whole plane, its only singularities being poles where $D(\lambda)=0$ . Why: convergence is local to a disc, but the true structure — poles at the spectrum, holomorphy elsewhere — is global, seen only by Fredholm's entire-function construction.

Limiting cases

The character of $(I-\lambda K)\phi=f$ is decided by where $\lambda$ sits relative to the spectrum of $K$ and by any special structure of the kernel:

Case	Result	Meaning
Small $\lambda$ (subcritical)	$\vert\lambda\vert\,\Vert K\Vert<1 \Rightarrow \phi=\sum_{n=0}^{\infty}\lambda^n K^n f$	Neumann series converges geometrically; the resolvent is analytic in the disc of radius $1/\Vert K\Vert$ .
Characteristic value	$\lambda=\lambda_k,\ \ \phi_k=\lambda_k K\phi_k$	$1/\lambda_k$ is an eigenvalue of $K$ ; the homogeneous equation has nontrivial solutions and the inhomogeneous one needs the orthogonality condition.
Degenerate kernel	$K=\sum_{n=1}^N g_n(x)h_n(t)$	Reduces exactly to an $N\times N$ system; characteristic values are the roots of $\det(I-\lambda A)=0$ , at most $N$ of them.
Symmetric kernel	$K(x,t)=K(t,x)\Rightarrow \lambda_n\in\mathbb{R},\ \langle\phi_m,\phi_n\rangle=\delta_{mn}$	Compact self-adjoint: real characteristic values, orthonormal eigenfunctions, Hilbert–Schmidt expansion of $K$ .
Volterra limit	upper limit $\to x$ , $K(x,t)=0$ for $t>x$	Operator is quasi-nilpotent, no nonzero eigenvalues; the second-kind equation is always uniquely solvable, series converges for all $\lambda$ .

Common trapConfusing the kind of an equation with the presence of a source

The two-way classification into "kinds" has nothing to do with whether $f$ vanishes. Second kind means the unknown $\phi$ appears outside the integral (regularizing the equation into a perturbation of the identity); it says nothing about the source, and a homogeneous second-kind equation $f\equiv0$ is still second kind — the eigenvalue problem $\phi=\lambda K\phi$ . Likewise, never assume a first-kind equation is solvable just because you can write it down: its solution operator $K^{-1}$ is unbounded because the singular values pile up at zero, so naive inversion amplifies noise catastrophically and a stable answer needs regularization. Keeping "kind" (structure) separate from "homogeneous" (source) is the most common bookkeeping error in the subject.

Worked examples

1 · A degenerate kernel solved three ways (fully worked)

Solve $\phi(x)=x+\lambda\displaystyle\int_0^1 x\,t\,\phi(t)\,\mathrm{d}t$ on $[0,1]$ . Find the general solution, evaluate at $\lambda=1$ , identify the characteristic value, and confirm against the Neumann series.

Step 1. Recognize the rank-one degenerate kernel $K(x,t)=x\,t$ , so $g(x)=x$ , $h(t)=t$ , $N=1$ . The single moment is $A=\int_0^1 t^2\,\mathrm{d}t=\tfrac13$ , and (since $f=x$ ) the source moment is $b=\int_0^1 t^2\,\mathrm{d}t=\tfrac13$ . A = 1/3, b = 1/3.

Step 2. The scalar system $(1-\lambda A)c=b$ gives $c=\dfrac{b}{1-\lambda/3}=\dfrac{1}{3-\lambda}$ , and the solution is $\phi(x)=x+\lambda c\,x=x\Big(1+\dfrac{\lambda}{3-\lambda}\Big)=\dfrac{3x}{3-\lambda}$ . The exact solution is φ(x) = 3x/(3 − λ).

Step 3. Evaluate at $\lambda=1$ : $\phi(x)=\dfrac{3x}{2}=1.5\,x$ . Check by substitution: $x+\int_0^1 x\,t\,(1.5\,t)\,\mathrm{d}t=x+1.5\,x\cdot\tfrac13=\tfrac{3x}{2}$ . Verified: φ = 1.5x at λ = 1.

Step 4. Identify the characteristic value from $\det(I-\lambda A)=1-\lambda/3=0$ , giving $\lambda_1=3$ ; the solution $3x/(3-\lambda)$ blows up there. The eigenfunction is $\phi_1\propto x$ (normalized $\sqrt3\,x$ ), so the operator's one nonzero eigenvalue is $\mu_1=1/\lambda_1=\tfrac13=A$ . Characteristic value λ₁ = 3, eigenfunction ∝ x.

Step 5. Cross-check with the Neumann series. Since $Kf=x\int_0^1 t^2\,\mathrm{d}t=x/3$ and each further $K$ adds a factor $\tfrac13$ , we get $K^n f=x/3^n$ , so $\phi=\sum_{n=0}^\infty\lambda^n\dfrac{x}{3^n}=\dfrac{x}{1-\lambda/3}=\dfrac{3x}{3-\lambda}$ — identical to Step 2, and convergent exactly for $|\lambda|<3=\lambda_1$ . The series reproduces 3x/(3 − λ) and converges for |λ| < 3.

2 · A symmetric kernel and a trace identity

The Green's function of $-\mathrm{d}^2/\mathrm{d}x^2$ on $[0,1]$ with $u(0)=u(1)=0$ is $G(x,t)=x(1-t)$ for $x\le t$ and $t(1-x)$ for $x\ge t$ . The homogeneous equation $u=\lambda\int_0^1 G\,u$ is equivalent to $-u''=\lambda u$ . Find the characteristic values and eigenfunctions, and verify the spectral sum with a trace identity.

Step 1. Convert back to differential form: applying $-\mathrm{d}^2/\mathrm{d}x^2$ to $u=\lambda\int_0^1 G\,u$ and using $-G''=\delta(x-t)$ gives $-u''=\lambda u$ , $u(0)=u(1)=0$ . The kernel is the inverse of the Sturm–Liouville operator.

Step 2. Solve the eigenproblem: $u_n(x)=\sqrt2\,\sin(n\pi x)$ with $\lambda_n=n^2\pi^2$ , orthonormal on $[0,1]$ ; numerically $\lambda_1=\pi^2\approx\mathbf{9.87}$ , $\lambda_2\approx\mathbf{39.5}$ , $\lambda_3\approx\mathbf{88.8}$ . Characteristic values λₙ = n²π².

Step 3. The operator eigenvalues are the reciprocals $\mu_n=1/(n^2\pi^2)$ : $\mu_1\approx\mathbf{0.1013}$ , $\mu_2\approx\mathbf{0.0253}$ , $\mu_3\approx\mathbf{0.0113}$ — decreasing and accumulating at $0$ , exactly the spectrum in the second figure. The eigenvalues of K crowd toward zero.

Step 4. Verify with the trace, which for a continuous symmetric kernel equals both the diagonal integral and the eigenvalue sum: $\sum_n\mu_n=\int_0^1 G(x,x)\,\mathrm{d}x=\int_0^1 x(1-x)\,\mathrm{d}x=\tfrac16$ , while independently $\sum_n 1/(n^2\pi^2)=\tfrac{1}{\pi^2}\cdot\tfrac{\pi^2}{6}=\tfrac16$ by the Basel sum. Both routes give Σ μₙ = 1/6 ≈ 0.1667 — a clean check on the spectral picture.

Physics of the portrait

The theory carries one name because a single 1903 paper reorganized a scattered subject into a matrix-like structure — turning integral equations from a bag of tricks into a branch of spectral theory.

Origins · 1 minIvar Fredholm and how this was discovered

Ivar Fredholm

1866–1927 · founder of the modern theory of integral equations

Erik Ivar Fredholm was a Swedish mathematician, a student of Gösta Mittag-Leffler at Stockholm, who spent much of his career as an actuary for the Swedish state while producing one of the most influential papers in analysis. Integral equations had been studied piecemeal for a century — Abel inverted one in 1823, Carl Neumann summed an iterative series in the 1870s, Volterra treated the variable-limit case in the 1890s — but no one had a general theory of the fixed-domain equation. Fredholm supplied it. In a note of 1900 and the celebrated 1903 memoir "Sur une classe d'équations fonctionnelles" in Acta Mathematica, he introduced the determinant $D(\lambda)$ as a limit of finite determinants, showed it is entire with zeros at the characteristic values, and proved the alternative that bears his name. The stroke of insight was to treat the integral operator as a genuine infinite-dimensional matrix, so that determinants, eigenvalues, and the rank–nullity dichotomy all survive the passage to infinite dimensions. Hilbert seized on it to build the spectral theory of symmetric kernels; Schmidt gave the eigenfunction expansion in 1907; Riesz recast the edifice through compact operators in 1918. Fredholm published little else, but that single paper made him the founder of a field.

History

1823

— Abel inverts an integral equation to solve the tautochrone problem

1877

— Carl Neumann introduces the iterated-series solution in potential theory

1903

— Fredholm publishes the determinant and the alternative in Acta Mathematica

1904

— Hilbert develops the spectral theory of symmetric kernels

1918

— Riesz recasts the theory through compact operators

Going further

Texts. Stone & Goldbart, Mathematics for Physics · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed. · Hassani, Mathematical Physics · Byron & Fuller, Mathematics of Classical and Quantum Physics, ch. 9 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, §23 · Courant & Hilbert, Methods of Mathematical Physics, Vol. I · Tricomi, Integral Equations.

Provenance. Classifications and equations are cross-checked against Stone & Goldbart, Arfken–Weber–Harris, Hassani, Byron–Fuller, Courant–Hilbert Vol. I, and Tricomi. A verifier should confirm: (i) for $K=xt$ on $[0,1]$ with $f=x$ , the exact solution $\phi=3x/(3-\lambda)$ , value $1.5x$ at $\lambda=1$ , characteristic value $\lambda_1=3$ , and Neumann sum $x\sum(\lambda/3)^n$ ; (ii) the symmetric min-kernel spectrum $\lambda_n=n^2\pi^2$ , eigenfunctions $\sqrt2\sin(n\pi x)$ , and trace $\sum 1/\lambda_n=\int_0^1 x(1-x)\,\mathrm{d}x=1/6$ via the Basel sum; (iii) the convergence bound $\|K^n f\|\le B^n\|f\|$ and spectral coefficients $\alpha_n=\lambda_n f_n/(\lambda_n-\lambda)$ . All numerical constants were recomputed for this article.

Prerequisites. Hilbert spaces and $L^2$ · Linear algebra: eigenvalues and determinants · Sturm–Liouville theory and Green's functions · Compact and self-adjoint operators. Leads to. WKB Approximation · Laplace's Method for Integrals · Method of Steepest Descent and Saddle Points.

Ask the tutor: Hand me any kernel and source and I will tell you the kind, find the characteristic values, and solve it by the fastest of the Neumann, degenerate, or spectral methods.

What is the difference between a first-kind and a second-kind equation?

In a second-kind equation the unknown appears both outside and inside the integral, $\phi=f+\lambda K\phi$ , so it is a perturbation of the identity and well-posed for all but a discrete set of $\lambda$ . In a first-kind equation the unknown appears only inside, $K\phi=f$ , so solving means inverting the compact smoother $K$ directly — and since its singular values tend to zero, that inverse is unbounded and the problem is generically ill-posed. The presence or absence of the "outside" copy of $\phi$ is the whole distinction, and it separates a stable problem from an unstable one.

Why does a compact operator make the theory behave like linear algebra?

Compactness lets infinite dimensions imitate finite ones. The Riesz–Schauder theory shows that a compact operator's nonzero spectrum consists only of eigenvalues of finite multiplicity, accumulating only at zero — exactly the spectral picture of a finite matrix, minus a harmless accumulation point at the origin. Consequently the Fredholm alternative holds verbatim, determinants and traces make sense, and no continuous spectrum spoils the dichotomy. An $L^2$ (Hilbert–Schmidt) kernel automatically generates a compact operator, which is why "nice kernel" and "linear-algebra behaviour" are the same statement.

What exactly is a characteristic value, and how does it relate to an eigenvalue?

For the operator $K$ an eigenvalue $\mu$ satisfies $K\phi=\mu\phi$ ; the characteristic value is its reciprocal $\lambda=1/\mu$ , the natural parameter because the equation reads $(I-\lambda K)\phi=f$ . Unique solvability fails exactly when $\lambda$ equals a characteristic value, since then $1/\lambda=\mu$ is a genuine eigenvalue and $I-\lambda K$ has a nontrivial null space. Because the eigenvalues $\mu_n$ of a compact operator accumulate at zero, the characteristic values $\lambda_n=1/\mu_n$ run off to infinity — which is why the Neumann series, controlled by the smallest $|\lambda_n|$ , always has a finite but nonzero radius of convergence.

When should I use the Neumann series versus the resolvent or spectral method?

Use the Neumann series when $|\lambda|$ is small — below $1/\|K\|_{HS}$ — since it converges geometrically and each term is a simple iterated integral; this is perturbation theory and the Born series. Use the degenerate-kernel reduction when the kernel separates into a finite sum of products, because the answer is then exact and finite-dimensional for every $\lambda$ . Use the spectral expansion when the kernel is symmetric with known eigenfunctions: it gives a closed form for all $\lambda$ off the spectrum and makes the solvability condition transparent. The Fredholm determinant is the global fallback when the series has diverged.

Why is inverting a first-kind equation so numerically dangerous?

Because you divide by numbers that shrink to zero. In the singular system of $K$ the formal solution is $\phi=\sum_n\langle f,u_n\rangle\,v_n/\sigma_n$ with $\sigma_n\to0$ . Any component of the data along a large- $n$ singular vector — including broadband measurement noise — is divided by a tiny $\sigma_n$ and blown up without bound, so even data accurate to many digits yields a wildly oscillating, useless $\phi$ . The cure is regularization: Tikhonov replaces $1/\sigma_n$ by $\sigma_n/(\sigma_n^2+\alpha)$ , and truncated SVD discards terms below a threshold — both trading a little bias for a large drop in variance, standard in tomography and deblurring.

How do Fredholm equations connect to Sturm–Liouville and Green's-function problems?

They are two views of one object. A regular Sturm–Liouville problem $\mathcal{L}u=\lambda\rho u$ inverts, via its Green's function $G$ , into $u(x)=\lambda\int_a^b G(x,t)\rho(t)u(t)\,\mathrm{d}t$ — a homogeneous second-kind Fredholm equation with a symmetric kernel. The Sturm–Liouville eigenvalues become the characteristic values, the eigenfunctions become the kernel's orthonormal eigenfunctions, and the Hilbert–Schmidt expansion of $G$ reproduces its bilinear eigenfunction expansion. Worked Example 2 is exactly this for $-u''=\lambda u$ on $[0,1]$ : the min-type Green's function has characteristic values $n^2\pi^2$ and eigenfunctions $\sqrt2\sin(n\pi x)$ . The differential and integral formulations carry identical spectral content.