Separable and Degenerate Kernels

An integral equation ties an unknown function to itself through a kernel, and in general that couples every point of the function to every other — an honestly infinite-dimensional problem. But a whole class of kernels quietly cheats. A kernel is separable, or degenerate, when it is a finite sum of products of one-variable functions, $K(x,t)=\sum_{i=1}^n a_i(x)\,b_i(t)$ . The moment a kernel has this form, the operator can only ever produce combinations of the $n$ fixed shapes $a_1(x),\dots,a_n(x)$ : its range is the finite-dimensional $\mathrm{span}\{a_i\}$ , and it is a matrix wearing the costume of an operator. The infinite problem collapses to an $n\times n$ one. The payoff is total. A Fredholm equation of the second kind with a degenerate kernel forces its own solution into the form $\phi=f+\lambda\sum_i c_i a_i$ , leaving only $n$ numbers $c_i$ to find; projecting onto the readers $b_j$ turns the equation into a linear system $(\mathbf I-\lambda\mathbf A)\mathbf c=\mathbf g$ . Every structural feature of the operator — spectrum, resolvent, and the Fredholm alternative that decides when solutions exist — is then read off a single small matrix and its determinant. This is not a toy: it is the finite-rank skeleton of the whole theory of compact operators, since any square-integrable kernel is an $L^2$ -limit of degenerate ones. The machinery here underlies the numerical solution of integral equations, separable potentials in scattering, and the Schmidt decomposition of quantum states. It is where functional analysis becomes linear algebra you can do by hand.

Key takeaways

A kernel is separable (degenerate) when it is a finite sum $K(x,t)=\sum_{i=1}^n a_i(x)\,b_i(t)$ ; the operator then has rank $\le n$ and its range is $\mathrm{span}\{a_i\}$ .
A Fredholm equation of the second kind with such a kernel collapses to an $n\times n$ linear system $(\mathbf I-\lambda\mathbf A)\mathbf c=\mathbf g$ for the projection coefficients $c_i=\int b_i\phi$ .
Solvability is governed entirely by $\mathbf M=\mathbf I-\lambda\mathbf A$ with $A_{ij}=\int b_i a_j$ : unique solution when $\det\mathbf M\ne0$ , and the characteristic values are the roots of the Fredholm determinant $d(\lambda)=\det\mathbf M$ .
The Fredholm alternative becomes the ordinary linear-algebra alternative — either $\mathbf M$ is invertible, or the homogeneous problem has eigenfunctions and the forcing must be orthogonal to the transposed null space.
Truncating a Hilbert–Schmidt (or Mercer) expansion approximates any compact operator by a degenerate one with controlled $L^2$ error, the basis of Nyström and Galerkin schemes.
Key idea: if $K(x,t)=\sum_{i=1}^n a_i(x)b_i(t)$ then the operator has rank $\le n$ ; its range lies in $\mathrm{span}\{a_i\}$ , so the integral equation reduces to an $n\times n$ linear system in the coefficients $c_i=\int b_i\phi$ .

The core insight

If $K(x,t)=\sum_{i=1}^n a_i(x)b_i(t)$ then the integral operator has rank $\le n$ and its range is $\mathrm{span}\{a_i\}$ , so a second-kind Fredholm equation reduces exactly to an $n\times n$ linear system for the coefficients $c_i=\int b_i\phi$ — spectrum, resolvent, and Fredholm alternative all read off one small matrix.

Where this lives today

Degenerate kernels are the quiet workhorse behind several live areas. In nuclear and atomic scattering, separable potentials $V(\mathbf r,\mathbf r')=\sum_i g_i(\mathbf r)\lambda_i g_i(\mathbf r')$ turn the Lippmann–Schwinger equation into an algebraic system for the $T$ -matrix; the Yamaguchi potential, a rank-one deuteron model from 1954, still gives closed-form phase shifts and a bound-state condition that is a single determinant set to zero, and modern few-body (Faddeev) calculations lean on finite-rank expansions for the same tractability. In numerical analysis, degenerate-kernel, Nyström, and Galerkin methods are the standard route to solving integral equations on a computer — from boundary-element electromagnetics to radiative transfer — replacing a continuous operator by a finite-rank one with an $L^2$ error bound certified in advance. In quantum information, a pure bipartite state of finite Schmidt rank is exactly a degenerate kernel $\rho(x,x')=\sum_i\lambda_i\psi_i(x)\psi_i^*(x')$ , whose reduced density operator is a finite matrix and whose eigenvalues give the entanglement entropy directly. And in photonics, projecting a Green's function onto a finite mode set yields a degenerate kernel and the coupled-mode equations. Wherever an operator is well approximated by a handful of dominant modes, the degenerate kernel is its exact language.

The physical picture

Picture the integral operator as a machine: you feed in a function, it grinds through $\int K(x,t)\,(\,\cdot\,)\,\mathrm dt$ , and hands back another. For a generic kernel the output can be anything. A degenerate kernel jams the machine into a tiny showroom: whatever you feed it, the answer is a blend of only $n$ fixed display models $a_1(x),\dots,a_n(x)$ . The functions $b_i(t)$ are the sensors — they read $n$ numbers $c_i=\int b_i\phi$ off the input, and those $n$ numbers are the only thing the operator ever notices about $\phi$ . Solving the equation therefore means finding $n$ numbers and reconstructing the answer from an explicit formula — the first figure shows this collapse.

Figure

An arbitrary input φ(t) lives in an infinite-dimensional space, but a degenerate kernel reads off only n numbers cᵢ = ∫ bᵢ(t)φ(t)dt through its sensors bᵢ, and every output is a blend of the n fixed shapes aᵢ(x) — the operator has rank ≤ n and its whole range is span of the aᵢ

The matrix $A_{ij}=\int b_i a_j$ is the whole story of how the building blocks feed back — how much of shape $a_j$ the sensor $b_i$ detects. Its eigenvalues are the eigenvalues of the operator, and the values of $\lambda$ where $\det(\mathbf I-\lambda\mathbf A)$ vanishes are exactly the characteristic values: the resonances where the homogeneous equation acquires solutions and the resolvent blows up. There is no hidden infinite spectrum; the operator has at most $n$ nonzero eigenvalues, the roots of a degree- $n$ polynomial. The second figure makes this concrete for the simplest kernel of all, $K(x,t)=xt$ on $[0,1]$ , whose one characteristic value sits at $\lambda_0=3$ .

Figure

The resolvent amplification 1/(1 − λ/3) for the rank-one kernel K(x,t) = xt on the interval 0 to 1: it is smooth and finite everywhere except at the single characteristic value λ₀ = 3, where it diverges — that pole is the lone eigenvalue of the operator, with eigenfunction φ₀(x) = x

The core equations

The definition is the whole engine. A kernel is degenerate when it splits into a finite sum of products of one-variable functions, and we always take the two families to be linearly independent so that $n$ is the true rank.

K(x,t)=\sum_{i=1}^{n} a_i(x)\,b_i(t) \qquad (1)

where: $K(x,t)$ — kernel of the integral operator on $a\le x,t\le b$ ; $a_i(x),\,b_i(t)$ — single-variable functions, each family linearly independent; $n$ — the (finite) rank of the operator, which can only build outputs from the $n$ shapes $a_i(x)$ . The problem we solve is the Fredholm equation of the second kind — the unknown appears both outside and inside the integral, with a spectral parameter $\lambda$ weighting the kernel term.

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

where: $\phi(x)$ — the unknown function; $f(x)$ — the given forcing (inhomogeneous term); $\lambda$ — the spectral/coupling parameter. When $f\equiv0$ this is the homogeneous eigenproblem; the values of $\lambda$ admitting nonzero $\phi$ are the characteristic values. Substituting (1) into (2) forces the solution into a rigid shape: it is $f$ plus a fixed combination of the $a_i$ , so the only freedom left is $n$ numbers.

\phi(x)=f(x)+\lambda\sum_{i=1}^{n} c_i\,a_i(x),\qquad c_i\equiv\int_a^b b_i(t)\,\phi(t)\,\mathrm{d}t \qquad (3)

where: $c_i$ — the projection coefficients, the $n$ genuine unknowns. Solving the equation is finding $n$ numbers, then writing $\phi$ down explicitly. Two families of integrals carry all the data from the functions into the algebra: the matrix $\mathbf A$ of how the blocks overlap the readers, and the vector $\mathbf g$ of how the forcing projects onto the readers.

A_{ij}=\int_a^b b_i(t)\,a_j(t)\,\mathrm{d}t,\qquad g_i=\int_a^b b_i(t)\,f(t)\,\mathrm{d}t \qquad (4)

where: $A_{ij}$ — the feedback matrix, how block $a_j$ registers on sensor $b_i$ ; $g_i$ — the projection of the forcing onto sensor $b_i$ . Ordinary definite integrals — pure numbers. Projecting the ansatz (3) onto each $b_j$ closes the loop and yields the finite linear system that replaces the integral equation entirely.

(\mathbf{I}-\lambda\mathbf{A})\,\mathbf{c}=\mathbf{g} \qquad (5)

where: $\mathbf c=(c_1,\dots,c_n)^\top$ — the coefficient vector; $\mathbf I$ — the $n\times n$ identity; $\mathbf g=(g_1,\dots,g_n)^\top$ . The entire integral equation is now one small matrix inversion. The determinant of that matrix is the finite-rank Fredholm determinant, a polynomial of degree at most $n$ whose roots are precisely the characteristic values of the operator.

d(\lambda)\equiv\det(\mathbf{I}-\lambda\mathbf{A}),\qquad d(\lambda_k)=0\ \Leftrightarrow\ \mu_k=1/\lambda_k\in\mathrm{spec}(\mathbf A) \qquad (6)

where: $d(\lambda)$ — Fredholm determinant, degree $\le n$ ; $\lambda_k$ — characteristic values; $\mu_k=1/\lambda_k$ — the eigenvalues of $\mathbf A$ and of the operator. The spectrum of the operator is the spectrum of a small matrix. Away from those roots the system inverts, and back-substitution reveals that the resolvent kernel — which solves the equation for any forcing — is itself degenerate of rank $\le n$ .

\phi(x)=f(x)+\lambda\int_a^b R(x,t;\lambda)\,f(t)\,\mathrm{d}t,\qquad R(x,t;\lambda)=\sum_{i,j}a_i(x)\,[\mathbf M^{-1}]_{ij}\,b_j(t) \qquad (7)

where: $R(x,t;\lambda)$ — resolvent kernel; $\mathbf M=\mathbf I-\lambda\mathbf A$ . Solving once inverts the operator for all $f$ at once; the inverse is still finite-rank, with poles at the characteristic values.

Derivation AFrom integral equation to a finite linear system

The goal is to convert the Fredholm equation of the second kind — posed in the infinite-dimensional space of square-integrable functions — into an ordinary $n\times n$ matrix equation whose solution reconstructs the unknown exactly, and to watch the celebrated Fredholm alternative emerge as nothing grander than whether one small matrix is invertible. We insert the degenerate kernel, note that the solution is trapped in a finite-dimensional affine space, project onto the readers, and split into the two cases decided by the determinant.

Step 1. Insert (1) into (2). Because the sum is finite and $x$ and $t$ separate, the $x$ -dependence factors straight out of the integral: $\lambda\int_a^b K(x,t)\phi(t)\,\mathrm dt=\lambda\sum_i a_i(x)\int_a^b b_i(t)\phi(t)\,\mathrm dt=\lambda\sum_i a_i(x)\,c_i$ , with $c_i\equiv\int_a^b b_i\phi$ . Why: the $t$ -integrals are ordinary numbers independent of $x$ , so the entire integral term is a fixed combination of the $a_i(x)$ — this is the whole reason the problem is finite.

Step 2. Read off the shape of the answer. The equation now says $\phi=f+\lambda\sum_i c_i a_i$ , so every solution lies in the affine space $f+\mathrm{span}\{a_1,\dots,a_n\}$ ; only the $n$ coefficients $c_i$ remain unknown. Why: no matter how complicated $f$ is, the correction the operator adds is confined to an $n$ -dimensional subspace, so we have already reduced infinitely many unknowns to $n$ .

Step 3. Project onto the readers. Apply $\int_a^b b_j(\cdot)\,\mathrm dt$ to the ansatz and use the definitions in (4): $c_j=g_j+\lambda\sum_i A_{ji}c_i$ , that is $\sum_i(\delta_{ji}-\lambda A_{ji})c_i=g_j$ , or in matrix form $(\mathbf I-\lambda\mathbf A)\mathbf c=\mathbf g$ . Why: projecting turns the functional statement into $n$ scalar equations for the $n$ scalars $c_i$ — the integral equation has become linear algebra with no approximation.

Step 4. Case one, $d(\lambda)=\det(\mathbf I-\lambda\mathbf A)\ne0$ . The matrix inverts, $\mathbf c=(\mathbf I-\lambda\mathbf A)^{-1}\mathbf g$ is unique, and (3) gives the unique $\phi$ . Why: a nonsingular reduced matrix is exactly the statement that $\mathbf I-\lambda K$ is invertible as an operator, so the resolvent exists.

Step 5. Case two, $d(\lambda_0)=0$ . Now $\lambda_0$ is a characteristic value: the homogeneous system $(\mathbf I-\lambda_0\mathbf A)\mathbf c=0$ has nontrivial solutions, giving eigenfunctions $\phi_h=\lambda_0\sum_i c_i a_i$ , and the inhomogeneous problem is solvable only if $\mathbf g$ is orthogonal to the left null space, $\mathbf g\perp\ker(\mathbf I-\lambda_0\mathbf A^\top)$ . Why: this is the Fredholm alternative in the flesh — either a unique solution, or eigenfunctions plus a solvability condition on the forcing — and it is now just the fundamental theorem of linear algebra applied to a transpose.

Go deeper · Derivation B · 1 minThe resolvent kernel is itself degenerate

Derivation BThe resolvent kernel is itself degenerate

The goal is to show that solving the reduced system once produces a closed-form inverse operator valid for every forcing $f$ , represented by a resolvent kernel $R(x,t;\lambda)$ which is again a finite sum of products — degenerate of rank at most $n$ — with poles precisely at the characteristic values. This makes the analytic structure of the operator, in particular where its inverse ceases to exist, transparent as the vanishing of a small determinant.

Step 1. Solve the reduced system symbolically. Away from a characteristic value, $\mathbf c=\mathbf M^{-1}\mathbf g$ with $\mathbf M=\mathbf I-\lambda\mathbf A$ , so componentwise $c_i=\sum_j[\mathbf M^{-1}]_{ij}\,g_j=\sum_j[\mathbf M^{-1}]_{ij}\int_a^b b_j(t)f(t)\,\mathrm dt$ . Why: we keep $f$ general instead of a specific forcing, so the answer becomes an operator acting on $f$ rather than a single solution.

Step 2. Substitute back into the ansatz (3): $\phi(x)=f(x)+\lambda\sum_{i,j}a_i(x)\,[\mathbf M^{-1}]_{ij}\int_a^b b_j(t)f(t)\,\mathrm dt$ . Reading off the kernel acting on $f$ gives $R(x,t;\lambda)=\sum_{i,j}a_i(x)\,[\mathbf M^{-1}]_{ij}\,b_j(t)$ . Why: the double sum is a finite sum of products, so $R$ is degenerate of rank $\le n$ — the inverse mirrors the forward structure.

Step 3. Locate the singularities. By Cramer's rule $\mathbf M^{-1}=\mathrm{adj}(\mathbf M)/d(\lambda)$ , so $R(x,t;\lambda)$ is a rational function of $\lambda$ whose only poles are the zeros of $d(\lambda)$ . Why: the resolvent is analytic in $\lambda$ except at the finitely many characteristic values, where an eigenfunction appears and the inverse diverges — exactly the pole at $\lambda_0=3$ in the second figure. *For example, $K=xt$ on $[0,1]$ gives $R=xt/(1-\lambda/3)$ , a single pole at $\lambda_0=3$ , confirming that the resolvent formula $R=\sum_{ij}a_i[\mathbf M^{-1}]_{ij}b_j$ captures the correct analytic structure.*

Go deeper · Derivation C · 1 minApproximating any Hilbert-Schmidt kernel by a degenerate one

Derivation CApproximating any Hilbert-Schmidt kernel by a degenerate one

The goal is to explain why degenerate kernels are the computational backbone of the whole theory of integral equations: every square-integrable kernel can be approximated in $L^2$ by a degenerate one to any accuracy, and because the Hilbert–Schmidt norm dominates the operator norm, the solution of the degenerate problem converges to the true solution — the warrant behind Galerkin and Nyström methods and behind the density of finite-rank operators in the compact operators.

Step 1. Expand in a product basis. If $K\in L^2([a,b]^2)$ and $\{e_i\}$ is an orthonormal basis of $L^2[a,b]$ , then $K(x,t)=\sum_{i,j}K_{ij}e_i(x)e_j(t)$ with $K_{ij}=\iint K(x,t)e_i(x)e_j(t)\,\mathrm dx\,\mathrm dt$ , and Parseval gives $\sum_{ij}|K_{ij}|^2=\|K\|_{L^2}^2<\infty$ . Why: square-integrability of the kernel is exactly convergence of the coefficient sum, so the tail can be made arbitrarily small.

Step 2. Truncate to a degenerate kernel. Keep a dominant finite index set $\Lambda_n$ and set $K_n(x,t)=\sum_{(i,j)\in\Lambda_n}K_{ij}e_i(x)e_j(t)$ ; this is degenerate of finite rank, and the error is the discarded tail, $\|K-K_n\|_{L^2}^2=\sum_{(i,j)\notin\Lambda_n}|K_{ij}|^2\to0$ as $n\to\infty$ . Why: a smooth kernel has rapidly decaying $K_{ij}$ , so a modest $n$ already captures it and the remainder is a convergent series' tail.

Step 3. Transfer convergence to the operator and the solution. Since $\|K-K_n\|_{\mathrm{op}}\le\|K-K_n\|_{L^2}\to0$ , the operators converge in norm; wherever $(\mathbf I-\lambda K)$ is invertible, so is $(\mathbf I-\lambda K_n)$ for large $n$ , and the degenerate-kernel solution converges to the true one. Why: this is why one solves integral equations with a finite-rank surrogate — the error is controlled a priori by the truncated Hilbert–Schmidt norm. *The key identities underpinning this step are Bessel's inequality and the Hilbert–Schmidt norm identity $\|K\|_{L^2}^2=\sum_{ij}|K_{ij}|^2$ ; operator-norm domination by the Hilbert–Schmidt norm then guarantees that convergence of the truncated kernel implies convergence of the corresponding operators.*

Limiting cases

Special cases expose the machinery — the rank-one formula you can memorize, the weak-coupling Neumann series, and the resonance where the whole thing breaks down productively.

Case	Result	Meaning
Rank one $K=a(x)b(t)$	$\phi(x)=f(x)+\dfrac{\lambda\,a(x)\int_a^b b(t)f(t)\,\mathrm dt}{1-\lambda\int_a^b b(t)a(t)\,\mathrm dt}$	one coefficient; the sole characteristic value is $\lambda_0=1/\!\int b\,a$ , where the resolvent diverges
$\lambda\to0$ (weak coupling)	$\phi=f+\lambda\!\int K f\,\mathrm dt+O(\lambda^2)$	Neumann series: $(\mathbf I-\lambda\mathbf A)^{-1}=\sum_k\lambda^k\mathbf A^k$ converges for $\|\lambda\|<1/\rho(\mathbf A)$
$d(\lambda_0)=0$ (resonance)	$\phi_h=\lambda_0\sum_i c_i^{(0)}a_i,\ (\mathbf I-\lambda_0\mathbf A)\mathbf c^{(0)}=0$	homogeneous equation gains eigenfunctions; inhomogeneous one needs $f\perp$ transposed eigenfunctions (Fredholm alternative)
Symmetric $b_i=a_i$ , real	$\mathbf A$ symmetric $\Rightarrow$ real $\mu_k$ , orthogonal eigenfunctions	self-adjoint case: real characteristic values, orthogonal eigenfunctions (finite Mercer)

Common trapthe misconception

The separation $K=\sum_i a_i b_i$ is not unique, and this trips people constantly. You must use a minimal representation with the $\{a_i\}$ linearly independent and the $\{b_i\}$ linearly independent; otherwise the rank you assign is wrong, $\mathbf A$ is singular for trivial book-keeping reasons rather than physical ones, and the "extra" zero characteristic values are artefacts. If $a_2=2a_1$ , for instance, the two-term kernel is secretly rank one, and forcing it into a $2\times2$ matrix produces a spurious singularity. A second, deeper trap: this exact reduction is a feature of second-kind equations only. For a first-kind equation $\int_a^b K(x,t)\phi(t)\,\mathrm dt=f(x)$ with a degenerate kernel, the operator has finite rank, so the equation is solvable only if $f\in\mathrm{span}\{a_i\}$ and even then leaves an infinite-dimensional ambiguity in $\phi$ — it is wildly ill-posed, not merely a matrix inversion.

Worked examples

1 · A rank-one equation solved to a number (confidence builder)

Solve $\phi(x)=1+\displaystyle\int_0^1 x t\,\phi(t)\,\mathrm dt$ on $[0,1]$ — that is, $K(x,t)=xt$ , forcing $f(x)=1$ , and $\lambda=1$ .

Step 1. Identify the blocks: $a_1(x)=x$ , $b_1(t)=t$ , so the kernel is rank one with the single unknown $c_1=\int_0^1 t\,\phi(t)\,\mathrm dt$ . The solution must have the form $\phi(x)=1+\lambda c_1 x=1+c_1 x$ .

Step 2. Compute the matrix and forcing entries from (4): $A_{11}=\int_0^1 t\cdot t\,\mathrm dt=\tfrac13$ and $g_1=\int_0^1 t\cdot1\,\mathrm dt=\tfrac12$ .

Step 3. Solve the (scalar) reduced system $(1-\lambda A_{11})c_1=g_1$ : $(1-\tfrac13)c_1=\tfrac12$ , so $\tfrac23 c_1=\tfrac12$ and $c_1=\tfrac34$ .

Step 4. Reconstruct and check: $\phi(x)=1+\tfrac34 x$ . Substituting back, $1+x\int_0^1 t(1+\tfrac34 t)\,\mathrm dt=1+x(\tfrac12+\tfrac34\cdot\tfrac13)=1+x(\tfrac12+\tfrac14)=1+\tfrac34 x$ . The solution is $\phi(x)=1+\tfrac34 x$ , exact.

2 · The full spectrum of a rank-two kernel

Find the characteristic values of $K(x,t)=xt+x^2t^2$ on $[0,1]$ — the values of $\lambda$ where the homogeneous equation has nontrivial solutions.

Step 1. Read off $a_1=x,\ b_1=t,\ a_2=x^2,\ b_2=t^2$ , a rank-two kernel. Build $\mathbf A$ from $A_{ij}=\int_0^1 b_i a_j$ : $A_{11}=\int t^2=\tfrac13$ , $A_{12}=\int t^3=\tfrac14$ , $A_{21}=\int t^3=\tfrac14$ , $A_{22}=\int t^4=\tfrac15$ , so $\mathbf A=\begin{pmatrix}1/3 & 1/4\\ 1/4 & 1/5\end{pmatrix}$ .

Step 2. Form the Fredholm determinant $d(\lambda)=\det(\mathbf I-\lambda\mathbf A)=1-(\operatorname{tr}\mathbf A)\lambda+(\det\mathbf A)\lambda^2$ . Here $\operatorname{tr}\mathbf A=\tfrac13+\tfrac15=\tfrac{8}{15}$ and $\det\mathbf A=\tfrac13\cdot\tfrac15-\tfrac1{16}=\tfrac1{15}-\tfrac1{16}=\tfrac1{240}$ , so $d(\lambda)=1-\tfrac{8}{15}\lambda+\tfrac{1}{240}\lambda^2$ .

Step 3. Clear denominators — multiply by $240$ — to get $\lambda^2-128\lambda+240=0$ , whence $\lambda=\tfrac{128\pm\sqrt{128^2-960}}{2}=\tfrac{128\pm\sqrt{15424}}{2}$ .

Step 4. Evaluate: $\sqrt{15424}\approx124.19$ , giving the two characteristic values $\lambda_1\approx1.903$ and $\lambda_2\approx126.1$ . Cross-check via the operator eigenvalues $\mu_k=1/\lambda_k$ : $\mu_1+\mu_2=0.5254+0.00793=0.5333=\operatorname{tr}\mathbf A$ and $\mu_1\mu_2=0.004167=\det\mathbf A$ , both exact. At $\lambda_1,\lambda_2$ the resolvent (7) diverges.

3 · A degenerate kernel with a degenerate eigenvalue

Solve the homogeneous problem for $K(x,t)=\cos(x-t)$ on $[0,2\pi]$ — a rank-two kernel whose spectrum has a repeated (degenerate) eigenvalue.

Step 1. Use the addition formula: $\cos(x-t)=\cos x\cos t+\sin x\sin t$ , so $a_1=\cos x,\ b_1=\cos t,\ a_2=\sin x,\ b_2=\sin t$ — rank two.

Step 2. Compute $\mathbf A$ over $[0,2\pi]$ : $A_{11}=\int_0^{2\pi}\cos^2 t\,\mathrm dt=\pi$ , $A_{22}=\int_0^{2\pi}\sin^2 t\,\mathrm dt=\pi$ , and $A_{12}=A_{21}=\int_0^{2\pi}\cos t\sin t\,\mathrm dt=0$ . Hence $\mathbf A=\pi\,\mathbf I$ .

Step 3. The Fredholm determinant is $d(\lambda)=\det(\mathbf I-\lambda\pi\mathbf I)=(1-\pi\lambda)^2$ , a double root. The single characteristic value is $\lambda_0=1/\pi\approx0.3183$ , with algebraic and geometric multiplicity two.

Step 4. Since $\mathbf A=\pi\mathbf I$ , every vector $\mathbf c$ solves the homogeneous system at $\lambda_0$ , so the eigenspace is the full two-dimensional $\mathrm{span}\{\cos x,\sin x\}$ . Check $\phi=\cos x$ : $\tfrac1\pi\int_0^{2\pi}\cos(x-t)\cos t\,\mathrm dt=\tfrac1\pi(\pi\cos x+0)=\cos x$ . A degenerate kernel producing a doubly degenerate eigenvalue — the multiplicity is the dimension of the reader overlap.

Physics of the portrait

Behind the tidy reduction stands the man who first tamed an integral equation with a determinant — the degenerate kernel being the case where his infinite determinant collapses to one you can compute by hand.

Origins · 1 minErik Ivar Fredholm and how this was discovered

Erik Ivar Fredholm

1866–1927 · Swedish mathematician, founder of the theory of integral equations

Erik Ivar Fredholm, born in Stockholm in 1866, gave integral equations the structure they still carry. Trained under Gösta Mittag-Leffler, Fredholm attacked the equation $\phi(x)=f(x)+\lambda\int K(x,t)\phi(t)\,\mathrm dt$ by treating the integral as a limit of finite sums — a discretization turning the operator into an $N\times N$ matrix. He showed that the determinant of $\mathbf I-\lambda\mathbf K$ tends, as $N\to\infty$ , to a convergent entire function $d(\lambda)$ , the Fredholm determinant, whose zeros are the characteristic values, and that a unique solution exists precisely when $d(\lambda)\ne0$ . His 1903 Acta Mathematica paper, "Sur une classe d'équations fonctionnelles," established the alternative that bears his name: either the inhomogeneous equation has a unique solution, or the homogeneous one has nontrivial solutions and the forcing must be orthogonal to the transposed eigenfunctions. The degenerate kernel is where this whole apparatus becomes finite and exact — Fredholm's infinite determinant is literally the ordinary $n\times n$ determinant $\det(\mathbf I-\lambda\mathbf A)$ . Hilbert and Erhard Schmidt seized on the work at once, recasting it with eigenvalues and orthonormal expansions and inventing much of what became Hilbert space theory. Fredholm remained in Stockholm as professor of rational mechanics and mathematical physics until his death in 1927.

History

1823

— Abel reduces the tautochrone problem to an integral equation and inverts it

1900

— Fredholm circulates his determinant method for the Dirichlet problem

1903

— Fredholm publishes the theory of d(λ) and the alternative in Acta Mathematica

1904

— Hilbert recasts the theory with eigenvalues and quadratic forms

1907

— Schmidt introduces orthonormal expansions, seeding Hilbert space theory

Going further

Texts. Stone & Goldbart, Mathematics for Physics, ch. on integral equations (degenerate kernels and compact operators) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., ch. 21 · Hassani, Mathematical Physics, chapters on integral equations and operators on Hilbert space · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., ch. 23 · Tricomi, Integral Equations (Dover), sections on kernels of finite rank · Byron & Fuller, Mathematics of Classical and Quantum Physics, on Fredholm theory. The key results are: the reduced system $(\mathbf I-\lambda\mathbf A)\mathbf c=\mathbf g$ from the ansatz; example 1 ( $K=xt$ , $f=1$ , $\lambda=1$ ) gives $\phi=1+\tfrac34 x$ by direct substitution; example 2 yields $\mathbf A=\begin{pmatrix}1/3&1/4\\1/4&1/5\end{pmatrix}$ , $d(\lambda)=1-\tfrac8{15}\lambda+\tfrac1{240}\lambda^2$ , characteristic values $\lambda\approx1.903,\,126.1$ , with trace and determinant cross-checks; example 3 gives $\cos(x-t)$ on $[0,2\pi]\Rightarrow\mathbf A=\pi\mathbf I$ , $\lambda_0=1/\pi$ doubly degenerate; and the resolvent $R=\sum_{ij}a_i[\mathbf M^{-1}]_{ij}b_j$ .

Prerequisites. Linear algebra: determinants and eigenvalues · Inner products on $L^2[a,b]$ · Fredholm integral equations of the second kind. Leads to. Hilbert–Schmidt operators and Mercer's theorem · Fredholm determinant theory · Separable potentials in scattering theory.

Ask the tutor: Hand me any kernel and I will tell you its rank, reduce the equation to a matrix, and walk you through its spectrum and resolvent step by step.

What is the difference between "separable" and "degenerate" — are they the same thing?

For kernels the two words are synonyms: a kernel is separable, or degenerate, when it can be written as a finite sum of products $\sum_{i=1}^n a_i(x)b_i(t)$ . "Separable" emphasizes the structure — the variables $x$ and $t$ come apart into independent factors — while "degenerate" emphasizes the consequence — the operator has finite rank $n$ , so its infinite-dimensional Fredholm determinant degenerates into an ordinary $n\times n$ one. Beware the collision with other uses of "separable" (separable differential equations, separable Hilbert spaces, separable quantum states); in the theory of integral equations it means exactly the finite-sum-of-products property, nothing else.

Why does the solution have to be f plus a combination of just the aᵢ?

Because the operator can produce nothing else. When you insert the degenerate kernel into the integral term, the integral over $t$ collapses each product $a_i(x)b_i(t)$ into $a_i(x)$ times a pure number $c_i=\int b_i\phi$ . So the whole integral term is $\lambda\sum_i c_i a_i(x)$ — a fixed combination of the $a_i$ , independent of $x$ except through the $a_i$ themselves. The equation $\phi=f+\lambda\sum_i c_i a_i$ then literally displays the answer: it is the forcing $f$ plus something in $\mathrm{span}\{a_i\}$ . No solution can escape that affine space, which is why only the $n$ coefficients are left to determine.

How is the Fredholm alternative "just linear algebra" here?

The Fredholm alternative for the integral equation says: either the inhomogeneous equation has a unique solution for every forcing, or the homogeneous equation has nontrivial solutions and the inhomogeneous one is solvable only for forcings orthogonal to the homogeneous solutions of the transposed equation. Once the degenerate kernel has reduced everything to $(\mathbf I-\lambda\mathbf A)\mathbf c=\mathbf g$ , this is exactly the fundamental theorem of linear algebra about a square matrix $\mathbf M=\mathbf I-\lambda\mathbf A$ : either $\det\mathbf M\ne0$ (unique solution), or $\det\mathbf M=0$ , in which case $\mathbf M\mathbf c=\mathbf g$ is solvable iff $\mathbf g\perp\ker\mathbf M^\top$ . The deep operator theorem becomes an undergraduate fact about matrices.

Do the characteristic values λ and the operator eigenvalues μ mean the same thing?

They are reciprocals, and the distinction matters. The operator eigenvalue equation is $K\phi=\mu\phi$ , i.e. $\int K(x,t)\phi(t)\,\mathrm dt=\mu\phi(x)$ , so $\mu$ is an eigenvalue of the operator (and of the matrix $\mathbf A$ ). The characteristic value $\lambda$ is the coupling parameter at which the homogeneous second-kind equation $\phi=\lambda\int K\phi$ has solutions, so $\lambda=1/\mu$ . Textbooks differ on which they call "eigenvalue," which causes endless confusion; the safe habit is to state the equation. A rank- $n$ kernel has at most $n$ nonzero $\mu$ and correspondingly at most $n$ finite characteristic values $\lambda_k=1/\mu_k$ , the roots of $\det(\mathbf I-\lambda\mathbf A)=0$ .

Why does the reduction fail for first-kind equations?

A first-kind equation $\int_a^b K(x,t)\phi(t)\,\mathrm dt=f(x)$ has the unknown only inside the integral. With a degenerate kernel the left-hand side is $\sum_i a_i(x)c_i$ , which lives entirely in the $n$ -dimensional $\mathrm{span}\{a_i\}$ . So the equation is satisfiable only if $f$ lies in that span — otherwise no solution exists. And even when it does, only the $n$ numbers $c_i$ are pinned down; the part of $\phi$ orthogonal to every $b_i$ is completely free, an infinite-dimensional ambiguity. A finite-rank operator is neither injective nor surjective, so the problem is ill-posed. The clean matrix reduction is special to the second kind, where the standalone $\phi(x)$ term supplies the missing invertibility.

When is the matrix A symmetric, and why is that the nice case?

The matrix $A_{ij}=\int b_i a_j$ is symmetric when the readers equal the shapes, $b_i=a_i$ and $K(x,t)=K(t,x)=\sum_i a_i(x)a_i(t)$ — the self-adjoint case. Then $\mathbf A$ is a real symmetric (indeed Gram) matrix, so its eigenvalues $\mu_k$ are real, its eigenvectors are orthogonal, and the operator's eigenfunctions form an orthogonal set. This is the finite-rank version of Mercer's theorem and of the spectral theorem for compact self-adjoint operators: the kernel is a finite sum $\sum_k \mu_k\,u_k(x)u_k(t)$ over its own orthonormal eigenfunctions. Non-symmetric kernels can have complex characteristic values and non-orthogonal eigenfunctions, and one must then track the transposed problem for the Fredholm alternative.