Fredholm Theory and Fredholm Alternative

Fredholm's determinant is not a historical curiosity but the working engine of some of the most active mathematical physics of the last thirty years. Its most spectacular modern home is random matrix theory: the probability that a large Gaussian random matrix has no eigenvalue above a threshold is a Fredholm determinant of the Airy kernel, and its value is the Tracy–Widom distribution — the universal law for the largest eigenvalue of the GUE, the longest increasing subsequence of a random permutation, and the fluctuating height of a growing interface in the Kardar–Parisi–Zhang class. That last link is experimental: Takeuchi and Sano's 2010 measurements of turbulent liquid-crystal fronts reproduced the Tracy–Widom curve precisely, so a Fredholm determinant is now readable off a laboratory histogram. The same object controls gap probabilities in integrable systems. In QFT, one-loop effective actions are functional determinants $\det(-\partial^2+m^2+V)$ that, written as $I+K$ and evaluated by Gel'fand–Yaglom or $\zeta$-regularization, are Fredholm determinants in disguise — the descendant of $D(\lambda)=\prod(1-\lambda/\lambda_n)$. And in numerical analysis the boundary-element and Nyström methods for scattering, acoustics, and electrostatics are direct discretizations of a second-kind Fredholm equation, their convergence guaranteed by the 1903 theorems and their conditioning set by proximity to a characteristic value.

Think of $\int_a^b K(x,y)\phi(y)\,\mathrm{d}y$ as an infinite matrix acting on the vector $\phi$. Chop $[a,b]$ into $n$ nodes spaced by $h=(b-a)/n$, and the equation becomes an honest linear system $(I-\lambda h K_{ij})\phi_j=f_i$. Solve it by Cramer's rule and let $n\to\infty$: the determinant and its cofactors converge, term by term, to the Fredholm determinant and minor. What could have been a hopeless divergence of ever-larger matrices is tamed because the many small factors of $h$ conspire into a rapidly convergent series — Fredholm's original and still most concrete route into the theory. The first figure makes the passage from operator to matrix literal.

The alternative says that for a compact $K$ the operator $I-\lambda K$ behaves exactly like a finite matrix: injective if and only if surjective. So either you can always solve, uniquely, or you cannot always solve and, when you can, the answer is non-unique — there is no third case of "injective but not surjective," the pathology that plagues genuinely infinite-dimensional operators. Geometrically the range is closed and equals $\ker(I-\lambda K^\dagger)^{\perp}$: the source $f$ must avoid the "forbidden directions" picked out by the adjoint zero modes, exactly as an undamped oscillator refuses to be driven on resonance. The second figure shows the analytic skeleton — the Fredholm determinant of an actual kernel, its zeros lined up as characteristic values running to infinity.

Because the eigenvalues $\mu_n$ accumulate only at zero, the characteristic values $\lambda_n=1/\mu_n$ run to infinity, so $D$ has only finitely many isolated zeros in any bounded region — which is why a compact operator never grows a continuous spectrum away from the origin. Every qualitative feature is already visible in $\cos\sqrt\lambda$: entire, real zeros spaced ever farther apart, and $D(0)=1$.

The object being solved is the Fredholm equation of the second kind: the unknown appears both outside the integral and inside it. That "unknown outside as well as inside" is what makes the operator $I-\lambda\hat K$ rather than the ill-posed $\hat K$, and it is the entire reason the clean theory below exists.

where: $\phi$ — unknown to be found; $f$ — given source term; $K(x,y)$ — kernel of the operator $\hat K$; $\lambda$ — complex coupling (spectral) parameter; $[a,b]$ — interval, possibly a region of $\mathbb{R}^d$. Read it as $(I-\lambda\hat K)\phi=f$, the infinite-dimensional cousin of a matrix equation. The Fredholm determinant is the continuum limit of $\det(I-\lambda K_n)$: a power series whose $n$-th coefficient is an $n$-fold integral of an $n\times n$ determinant built from the kernel — and it converges for every $\lambda$.

where: $D(\lambda)$ — Fredholm determinant, entire in $\lambda$; $K\binom{t_1\cdots t_n}{t_1\cdots t_n}$ — the $n\times n$ determinant $\det[K(t_i,t_j)]$; the $1/n!$ removes overcounting of unordered variables. Read it as $\det(I-\lambda\hat K)$ written as a sum over $n$-point minors. Through $\det=\exp\operatorname{tr}\log$ the same determinant is the exponential of a sum of traces of iterated kernels — the Plemelj–Smithies form, the workhorse for computation.

where: $\operatorname{tr}K^n$ — trace of the $n$-th iterated kernel, the continuum version of $\sum_i\mu_i^n$. Read it as $\log D=\operatorname{tr}\log(I-\lambda K)$, so $D=\prod_i(1-\lambda\mu_i)$ once eigenvalues are exposed. To recover the solution one needs the first Fredholm minor — the bordered determinant with an extra row and column at $(x,y)$ — whose ratio to $D$ is the resolvent kernel.

where: $D(x,y;\lambda)$ — first minor, entire in $\lambda$; $R$ — resolvent kernel, a ratio of two entire functions and hence meromorphic, with poles only at the zeros of $D$. Read it as the analytic form of $(I-\lambda K)^{-1}=I+\lambda\hat R$. With the resolvent in hand the solution is explicit whenever $D\neq0$ — a closed-form inverse, not just an existence claim.

where: $R$ — the resolvent kernel of (4). Read it as: apply $I+\lambda\hat R$ to $f$; away from characteristic values the solution exists, is unique, and depends continuously on $f$. Since $D$ is entire of order $\le\tfrac12$ for Hilbert–Schmidt kernels, Hadamard factorization writes it as a product over its zeros, the characteristic values.

where: $\lambda_n$ — characteristic values (zeros of $D$); $\mu_n$ — eigenvalues of $\hat K$, accumulating only at $0$. Read it as: $D$ is nothing but its zeros, which run to infinity because the eigenvalues shrink to nothing. At a zero of $D$ the theory hands over to the Fredholm alternative — a sharp dichotomy plus an orthogonality test.

where: $K^\dagger$ — adjoint kernel $\overline{K(y,x)}$; the range is closed and equals the orthogonal complement of the adjoint null space. Read it as: solvable exactly when $\langle\psi_k,f\rangle=0$ for every solution $\psi_k$ of the adjoint problem $\psi=\bar\lambda K^\dagger\psi$. The whole dichotomy collapses to one number — the Fredholm index — which is zero for every compact perturbation of the identity.

where: $\operatorname{ind}$ — Fredholm index; $\operatorname{coker}=H/\operatorname{ran}$. Read it as: kernel and cokernel have equal finite dimension, so injectivity and surjectivity stand or fall together — the analytic rank–nullity theorem.

The theory's behaviour turns sharply on the class of kernel and the kind of equation:

1 · A rank-one kernel, solved end to end (confidence builder)

Solve $\phi(x)=x+\lambda\int_0^1 xy\,\phi(y)\,\mathrm{d}y$ for all $\lambda$, find the characteristic value, and show what happens there.

Step 1. The kernel $K(x,y)=xy$ is rank one, so only the first trace survives: $\operatorname{tr}K=\int_0^1 t^2\,\mathrm{d}t=\tfrac13$, and every higher minor vanishes. Hence the Fredholm determinant truncates to a single term, $\boldsymbol{D(\lambda)=1-\tfrac{\lambda}{3}}$, with its one characteristic value at $\lambda_1=3$ (equivalently $\mu_1=\tfrac13$, the sole nonzero eigenvalue of $\hat K$).

Step 2. For a rank-one kernel the resolvent is $R(x,y;\lambda)=\dfrac{xy}{1-\lambda/3}$. The solution follows from equation (5): $\phi(x)=x+\lambda\displaystyle\int_0^1\frac{xy}{1-\lambda/3}\,y\,\mathrm{d}y=x+\frac{\lambda x}{1-\lambda/3}\cdot\frac13$, and collecting terms gives $\boldsymbol{\phi(x)=\dfrac{x}{1-\lambda/3}}$.

Step 3. Sanity check at $\lambda=1$: $\phi(x)=x/(1-\tfrac13)=\tfrac32 x$. Substitute back: $x+\int_0^1 xy\,(\tfrac32 y)\,\mathrm{d}y=x+\tfrac32 x\cdot\tfrac13=x+\tfrac12 x=\tfrac32 x$. The resolvent solution reproduces itself exactly.

Step 4. At $\lambda=3$ the determinant vanishes and the resolvent blows up — a resonance. The homogeneous equation is solved by $\phi_0(x)=x$; the kernel is symmetric so the adjoint zero mode is also $x$, and the alternative demands $\langle x,f\rangle=0$. But $f(x)=x$ gives $\int_0^1 x^2\,\mathrm{d}x=\tfrac13\neq0$, so at $\lambda=3$ the equation has no solution at all — the source is aligned with the forbidden direction, like driving an oscillator dead on resonance.

2 · A degenerate rank-two kernel (fully numeric)

For $K(x,y)=x+y$ on $[0,1]$, build the Fredholm determinant, locate the characteristic values, and solve $\phi(x)=1+\int_0^1(x+y)\phi(y)\,\mathrm{d}y$ (that is, at $\lambda=1$).

Step 1. Write $K=x\cdot1+1\cdot y$, so $a_1=x,b_1=1,a_2=1,b_2=y$ and the projection matrix $A_{jk}=\int_0^1 b_j a_k$ is $A=\begin{pmatrix}\tfrac12&1\\[2pt]\tfrac13&\tfrac12\end{pmatrix}$, with $\operatorname{tr}A=1$ and $\det A=\tfrac14-\tfrac13=-\tfrac1{12}$.

Step 2. The Fredholm determinant of a rank-two kernel is the characteristic polynomial of $A$: $D(\lambda)=1-\lambda\operatorname{tr}A+\lambda^2\det A=\boldsymbol{1-\lambda-\tfrac{\lambda^2}{12}}$, whose roots $\lambda^2+12\lambda-12=0$ give the characteristic values $\boldsymbol{\lambda_\pm=-6\pm4\sqrt3}$ ($\approx0.93$ and $-12.93$). Cross-check: $\operatorname{tr}K^2=\iint(x+y)^2=\tfrac76$, and the trace formula's $\lambda^2$ term $-\tfrac12\operatorname{tr}K^2-\tfrac12(\operatorname{tr}K)^2=-\tfrac7{12}$ matches $\log D$.

Step 3. At $\lambda=1$ the system $(I-\lambda A)c=g$ has $I-A=\begin{pmatrix}\tfrac12&-1\\[2pt]-\tfrac13&\tfrac12\end{pmatrix}$, $\det(I-A)=-\tfrac1{12}=D(1)\neq0$, so we are on the "unique solution" branch. With $f=1$ the data are $g_1=\int_0^1 1=1$, $g_2=\int_0^1 y=\tfrac12$.

Step 4. Invert: $(I-A)^{-1}=-12\begin{pmatrix}\tfrac12&1\\[2pt]\tfrac13&\tfrac12\end{pmatrix}=\begin{pmatrix}-6&-12\\-4&-6\end{pmatrix}$, giving $c_1=-6-6=-12$ and $c_2=-4-3=-7$. Reconstruct $\phi(x)=f+\lambda(a_1 c_1+a_2 c_2)=1+(-12x-7)$, so $\boldsymbol{\phi(x)=-12x-6}$.

Step 5. Verify: $1+\int_0^1(x+y)(-12y-6)\,\mathrm{d}y=1+\big[x(-6-6)+(-4-3)\big]=-12x-6$. The solution checks exactly — and had we sat at $\lambda=\lambda_+=-6+4\sqrt3$ the determinant would vanish and existence would hinge on $f\perp$ the adjoint zero mode.

3 · The Green kernel and a determinant you can name

Show that the Fredholm determinant of $K(x,y)=\min(x,y)$ on $[0,1]$ is $D(\lambda)=\cos\sqrt\lambda$, and read off the characteristic values plotted in the second figure.

Step 1. The kernel $\min(x,y)$ is the Green's function of $-\mathrm{d}^2/\mathrm{d}x^2$ with $u(0)=0,\ u'(1)=0$, so applying $\hat K$ inverts the second derivative under those conditions and the eigenvalue problem $\hat K\phi=\mu\phi$ becomes $-\phi''=\mu^{-1}\phi$ with the same boundary conditions.

Step 2. Those conditions select $\phi_n(x)=\sin\!\big((n-\tfrac12)\pi x\big)$ with $-\phi_n''=(n-\tfrac12)^2\pi^2\phi_n$, so $\mu_n=1/\big[(n-\tfrac12)^2\pi^2\big]$ and the characteristic values are $\lambda_n=(n-\tfrac12)^2\pi^2$ — $\pi^2/4\approx2.47$, $9\pi^2/4\approx22.2$, $25\pi^2/4\approx61.7$, marching to infinity exactly as compactness demands.

Step 3. Assemble the Hadamard product (6): $D(\lambda)=\prod_{n\ge1}\big(1-\tfrac{\lambda}{(n-1/2)^2\pi^2}\big)$. This is precisely the cosine factorization $\cos w=\prod_{n\ge1}\big(1-\tfrac{w^2}{(n-1/2)^2\pi^2}\big)$, so with $w=\sqrt\lambda$ the closed form is $\boldsymbol{D(\lambda)=\cos\sqrt\lambda}$ — entire (even in $\sqrt\lambda$, no branch cut), equal to $1$ at $\lambda=0$, vanishing exactly at the $\lambda_n$.

Step 4. Every feature of the second figure follows. A resonantly forced $-u''$ at $\mu^{-1}=\lambda_n$ is solvable only if the source is orthogonal to $\sin((n-\tfrac12)\pi x)$ — the Fredholm alternative for this classic boundary-value problem.

Fredholm theory runs in a lineage from Abel's tautochrone through Neumann's iterations to Hilbert's spectral theory, but its decisive step — the determinant that made an integral operator behave like a finite matrix — is the work of one Swedish mathematician who published remarkably little and changed the field permanently.

Texts. Stone & Goldbart, Mathematics for Physics, ch. on integral equations and Green's functions · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., ch. 21 · Hassani, Mathematical Physics, ch. on integral equations · Byron & Fuller, Mathematics of Classical and Quantum Physics, ch. 9 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., ch. 23 · F. Smithies, Integral Equations (Cambridge Tracts, 1958) · Reed & Simon, Methods of Modern Mathematical Physics, vols. I and IV.

Provenance. The determinant and minor series, the Plemelj–Smithies formula, the resolvent identity, the Hadamard product, and the index-zero alternative are standard. A verifier should recompute: (1) the discretized-determinant series and its Hadamard-bound convergence (Smithies ch. 5); (2) $\log D=-\sum\lambda^n\operatorname{tr}K^n/n$ from $\det=\exp\operatorname{tr}\log$; (3) $D=1-\lambda/3$ for $K=xy$, with $\phi=x/(1-\lambda/3)$ and resonance at $\lambda=3$; (4) $K=x+y$: $D=1-\lambda-\lambda^2/12$, characteristic values $-6\pm4\sqrt3$, solution $\phi=-12x-6$ at $\lambda=1,\ f=1$; (5) $D=\cos\sqrt\lambda$ for $K=\min(x,y)$ with $\lambda_n=(n-\tfrac12)^2\pi^2$. All numbers are exact textbook values; none fabricated.

Prerequisites. Compact and Hilbert–Schmidt operators · Hilbert spaces and adjoint operators · Determinants and Cramer's rule · Order Symbols and Gauge Functions. Leads to. WKB Approximation · Borel Summation and Resurgence · Bohr–Sommerfeld Quantization.

Ask the tutor: Hand me any kernel — separable, symmetric, Volterra, or a Green's function — and I will build its Fredholm determinant, find the characteristic values, and tell you exactly when your equation can be solved.