Neumann Series and Resolvent Kernel

An integral equation of the second kind, $\phi(x)=f(x)+\lambda\int_a^b K(x,y)\,\phi(y)\,\mathrm{d}y$ , hides its unknown $\phi$ on both sides of the equals sign: outside the integral as the answer, inside it as part of the machinery that produces the answer. This self-reference is the honest form of almost every problem in which a source, a boundary, and a propagation law meet — invert a differential operator with a Green function and the boundary data plus forcing collapse into exactly this shape, $K$ carrying the Green function and $\lambda$ the coupling. The Neumann series is the most direct way in. Treat $\lambda K$ as a small feedback and substitute the equation into its own right-hand side, over and over. Powers of $\lambda$ pile up, each multiplying an iterated operator $K^n$ , and the whole thing telescopes into $\phi=\sum_{n\ge0}\lambda^n K^n f$ — the operator version of $1/(1-r)$ , convergent absolutely and uniformly under the contraction estimate $|\lambda|\,\lVert K\rVert<1$ . Resumming the kernels once, before any particular $f$ is chosen, defines the resolvent kernel $R(x,y;\lambda)$ : a single object that inverts $\mathbb{1}-\lambda K$ for all time, so any new source is handled by one quadrature $\phi=f+\lambda\int R\,f$ . Its poles in the complex $\lambda$ -plane sit exactly at the eigenvalues, encoding resonance — and the same iterate-and-resum move reappears as the Born series in scattering and the Dyson series in quantum dynamics.

Key takeaways

The Neumann series solves $\phi=f+\lambda K\phi$ by iterating the substitution, producing a power series in $\lambda$ whose coefficients are the iterated integral operators $K^n$ .
Resumming the kernel factor gives the resolvent kernel $R(x,y;\lambda)$ , which delivers the solution in one stroke as $\phi=f+\lambda\int R\,f\,\mathrm{d}y$ for any inhomogeneity $f$ .
Convergence is guaranteed whenever $|\lambda|\,\lVert K\rVert<1$ ; the series is literally the operator identity $(\mathbb{1}-\lambda K)^{-1}=\sum_{n\ge0}\lambda^n K^n$ .
The sharp radius is $\rho=1/r(K)=\lvert\lambda_1\rvert$ , the smallest-modulus characteristic value; beyond it the resolvent continues as a meromorphic function with poles at the eigenvalues.
For Volterra kernels the shrinking domain forces factorial decay of the iterated kernels, so the resolvent is entire and the series converges for every $\lambda$ .
This is the integral-equation ancestor of perturbation theory, the Born series, and the Dyson series.
Key idea: Iterating $\phi=f+\lambda K\phi$ converts an implicit integral equation into the explicit operator geometric series $(\mathbb{1}-\lambda K)^{-1}f=\sum_n\lambda^nK^nf$ ; resumming the kernel factor gives the resolvent that inverts $\mathbb{1}-\lambda K$ for all inhomogeneities at once.

Where this lives today

The iterate-and-resum idea is quietly everywhere a linear inverse must be built cheaply. In 5G massive-MIMO receivers, base stations invert large, diagonally dominant Gram matrices to separate hundreds of user streams in real time; hardware detectors approximate $(\mathbb{1}-\lambda K)^{-1}$ by the first few Neumann terms, trading an $O(N^3)$ inversion for a handful of matrix–vector products — the contraction condition $|\lambda|\lVert K\rVert<1$ is literally the design constraint. The same truncated series drives Neumann-series preconditioners in numerical linear algebra and Jacobi and Richardson iterations. In imaging, the Born series is the Neumann series of the Lippmann–Schwinger equation; its convergence — engineered via the preconditioned "convergent Born series" methods of the 2010s — underlies diffraction tomography, ultrasound, and seismic full-waveform inversion. In quantum field theory the Dyson series is the Volterra case, entire in the coupling by time ordering, and the Hamiltonian resolvent $(z-H)^{-1}$ — the same object with $\lambda\to1/z$ — is the central tool of scattering theory. The open frontier is the resonant region: continuing the series past its first pole, where Borel summation, resurgence, and Fredholm determinants take over from naive iteration.

The physical picture

Read $\phi=f+\lambda K\phi$ as a feedback loop. A drive $f$ enters a box that outputs $\phi$ ; a filtered, scaled copy $\lambda K\phi$ is fed back and added to the drive. Unroll the loop one pass at a time and the response assembles itself: the raw drive $f$ , then its first echo $\lambda Kf$ , then the echo of the echo $\lambda^2K^2f$ , and so on. If the loop gain $\lambda K$ is a genuine contraction, the echoes shrink geometrically and their sum is finite — exactly $1+r+r^2+\cdots=1/(1-r)$ with the scalar $r$ promoted to the operator $\lambda K$ . The resolvent is the closed-loop transfer function: it reports, once and for all, how a unit impulse of source at $y$ becomes response at $x$ after every bounce is summed. The first figure makes the collapse concrete on a kernel simple enough to solve by hand, $K(x,y)=xy$ on the unit interval. Each echo carries a factor $\tfrac13$ of the one before, so the partial sums $1,\tfrac43,\tfrac{13}{9},\dots$ climb geometrically to the exact answer $\phi=\tfrac32 x$ .

Figure

Geometric collapse of the Neumann series for the kernel K(x,y)=xy on the unit interval, with source f(x)=x and λ=1. Each iterate λⁿKⁿf shrinks by the factor 1/3 (the kernel eigenvalue), so the partial sums 1, 1.333, 1.444, 1.481, 1.494, … climb to the exact solution φ = 3x/2.

The two kernel families behave differently, for a geometric reason. For a Fredholm kernel the limits $a,b$ are fixed, so the operator norm is a fixed number and the echoes shrink only if $|\lambda|$ is small enough — convergence is conditional. For a Volterra kernel the upper limit is the running point $x$ , so each successive integral covers a shrinking triangle; the iterated kernels pick up a factor $1/n!$ and the echoes die factorially however large $\lambda$ is. The second figure shows the payoff of resumming: the analytic structure of the resolvent in the complex $\lambda$ -plane, where the series lives inside a disc reaching out to the first pole.

Figure

Analytic structure of the resolvent in the complex λ-plane. The Neumann series converges inside the disc of radius ρ = 1/r(K), which reaches out to the nearest pole — the smallest-modulus characteristic value λ₁. For the degenerate kernel K(x,y)=xy the single characteristic value is λ₁ = 3, so ρ = 3; a generic kernel has further poles λ₂, λ₃, … marching outward.

Resonance is the physics behind the poles. As $\lambda$ approaches a characteristic value $\lambda_n$ , the operator $\mathbb{1}-\lambda K$ becomes singular and the response to a generic source blows up — precisely as a driven oscillator does at a natural mode. The resolvent's pole at $\lambda_n$ is the fingerprint of the blow-up.

The core equations

The equation the method solves is a linear integral equation of the second kind, so named because the unknown appears both alone and under the integral. Fixed limits $a,b$ make it Fredholm; a running upper limit makes it Volterra.

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

where: $\phi(x)$ — unknown function to be determined; $f(x)$ — given inhomogeneity, the source or forcing; $K(x,y)$ — given kernel of the integral operator; $\lambda$ — complex coupling parameter; $a,b$ — fixed integration limits (Fredholm), while an upper limit $x$ would make it Volterra. Iterating that self-reference generates the Neumann series. Every term is the drive filtered $n$ times, and collecting them is the operator analogue of expanding $1/(1-r)$ as $\sum r^n$ — hence the same sum is the inverse of $\mathbb{1}-\lambda K$ .

\phi=\sum_{n=0}^{\infty}\lambda^n K^n f=(\mathbb{1}-\lambda K)^{-1}f \qquad (2)

where: $K^n$ — the $n$ -fold composition of the integral operator; $K^0=\mathbb{1}$ is the identity; $(\mathbb{1}-\lambda K)^{-1}$ — the resolvent operator. Each power $K^n$ acts as an integral against an iterated kernel, defined next. The $n$ -fold composition is bookkept by a recursion on the kernels. Composing $K$ with the $n$ -th iterated kernel produces the $(n{+}1)$ -th, with an intermediate variable $z$ swept over the domain.

(K^n f)(x)=\int_a^b K_n(x,y)\,f(y)\,\mathrm{d}y,\qquad K_{n+1}(x,y)=\int_a^b K(x,z)\,K_n(z,y)\,\mathrm{d}z,\quad K_1=K \qquad (3)

where: $K_n(x,y)$ — the $n$ -th iterated kernel; $z$ — intermediate integration variable; $K_1(x,y)=K(x,y)$ seeds the recursion. The iterated kernel $K_n$ is the amplitude for propagating from $y$ to $x$ in exactly $n$ steps. Summing those iterated kernels against powers of $\lambda$ — before any source is named — defines the resolvent, or solving, kernel. It is the single object that carries all of the infinite series.

R(x,y;\lambda)=\sum_{n=0}^{\infty}\lambda^n K_{n+1}(x,y) \qquad (4)

where: $R(x,y;\lambda)$ — resolvent kernel; $K_{n+1}$ — the $(n{+}1)$ -th iterated kernel, so the sum starts from $K_1=K$ at $n=0$ . It is analytic in $\lambda$ inside the disc of convergence. With the resolvent in hand the solution is a single quadrature, valid for any inhomogeneity $f$ at once. This is the whole return on summing the series, and the resolvent itself obeys an equation of the same second-kind form.

\phi(x)=f(x)+\lambda\int_a^b R(x,y;\lambda)\,f(y)\,\mathrm{d}y,\qquad R=K+\lambda\int_a^b K(x,z)\,R(z,y;\lambda)\,\mathrm{d}z \qquad (5)

where: $R(x,y;\lambda)$ — resolvent kernel summing the Neumann series; the second relation is the resolvent's own integral equation, structurally identical to (1) with $f\to K$ and $\phi\to R$ . Invert once: every new source then costs one integral. Inside the disc the series defines $R$ . Outside it, the resolvent is continued by the ratio of two entire functions — the Fredholm first minor over the Fredholm determinant — exposing the poles the series could never reach.

R(x,y;\lambda)=\frac{D(x,y;\lambda)}{D(\lambda)} \qquad (6)

where: $D(\lambda)$ — the Fredholm determinant, entire in $\lambda$ , whose zeros are the characteristic values $\lambda_n$ ; $D(x,y;\lambda)$ — the first Fredholm minor, also entire. Because both are entire, their ratio is meromorphic on the whole plane, and the poles of $R$ sit exactly at the zeros of $D(\lambda)$ — the eigenvalues of the problem.

Derivation AIterated substitution and the operator geometric series

The goal is to turn the implicit relation $\phi=f+\lambda K\phi$ , in which the unknown appears on both sides, into an explicit convergent series by the most elementary move available: substitute the equation into its own right-hand side again and again, trading one factor of the unknown for one more known term at each pass, then show the leftover remainder vanishes under a norm condition so the finite pattern becomes an honest infinite sum equal to $(\mathbb{1}-\lambda K)^{-1}f$ .

Step 1. Start from $\phi=f+\lambda K\phi$ and notice it is a fixed-point statement: $\phi$ equals a fixed drive plus a linear operator applied to $\phi$ itself. Why: nothing has been assumed small yet — this is just the equation, rewritten so that the operator $T\phi=f+\lambda K\phi$ has $\phi$ as its fixed point.

Step 2. Substitute the whole right-hand side in for the $\phi$ it contains: $\phi=f+\lambda K(f+\lambda K\phi)=f+\lambda Kf+\lambda^2K^2\phi$ . Why: each substitution converts one factor of the unknown into a fully known term $\lambda^n K^n f$ plus a higher-order remainder still carrying $\phi$ — the series is being extruded one term at a time.

Step 3. Repeat $N$ times to reach the pattern $\phi=\sum_{n=0}^{N}\lambda^n K^n f+\lambda^{N+1}K^{N+1}\phi$ , a finite partial sum plus a single remainder term. Why: the structure is now transparent — an explicit polynomial in $\lambda K$ acting on $f$ , and one lump that still hides the unknown; everything hinges on whether that lump can be discarded.

Step 4. Bound the remainder with the submultiplicative operator norm: $\lVert\lambda^{N+1}K^{N+1}\phi\rVert\le(|\lambda|\,\lVert K\rVert)^{N+1}\lVert\phi\rVert$ , which tends to zero exactly when $|\lambda|\,\lVert K\rVert<1$ . Why: this is the contraction condition; the geometric factor $(|\lambda|\lVert K\rVert)^{N+1}$ crushes the tail, leaving $\phi=\sum_{n\ge0}\lambda^nK^nf$ as a convergent limit.

Step 5. Recognise the result as the operator geometric series $(\mathbb{1}-\lambda K)^{-1}=\sum_{n\ge0}\lambda^nK^n$ , and note that the partial sums are exactly the Banach fixed-point iterates of $T$ . Why: the same condition $|\lambda|\lVert K\rVert<1$ makes $T$ a contraction ( $\lVert T\phi_1-T\phi_2\rVert\le|\lambda|\lVert K\rVert\,\lVert\phi_1-\phi_2\rVert$ ), so the fixed-point theorem certifies both existence and uniqueness of the solution the series builds.

Go deeper · Derivation B · 1 minSumming the kernel factor into the resolvent

Derivation BSumming the kernel factor into the resolvent

The aim is to pay the cost of the infinite series once, in advance, by resumming the iterated kernels into a single solving kernel $R(x,y;\lambda)$ so that any inhomogeneity is dispatched by one quadrature $\phi=f+\lambda\int R\,f$ , and then to show this $R$ obeys an integral equation of the very same second-kind form — a self-contained characterisation that survives beyond the radius of convergence.

Step 1. Write every power as an integral against an iterated kernel, $(K^nf)(x)=\int K_n(x,y)f(y)\,\mathrm{d}y$ , with the recursion $K_{n+1}(x,y)=\int K(x,z)K_n(z,y)\,\mathrm{d}z$ and $K_1=K$ . Why: composing integral operators is a convolution-like operation, so the $n$ -th term of the Neumann series is a single integral whose kernel encodes $n$ propagation steps.

Step 2. Peel off the $n=0$ term and shift the index by $n=m+1$ : $\phi=f+\sum_{n\ge1}\lambda^nK^nf=f+\lambda\int\Big(\sum_{m\ge0}\lambda^mK_{m+1}(x,y)\Big)f(y)\,\mathrm{d}y$ . Why: pulling out one factor of $\lambda$ and one integration exposes a common kernel-sum that multiplies $f$ — the same bracketed sum for every source, which is precisely what we want to compute only once.

Step 3. Name the bracket the resolvent kernel, $R(x,y;\lambda)=\sum_{m\ge0}\lambda^mK_{m+1}(x,y)$ , so the solution is $\phi(x)=f(x)+\lambda\int R(x,y;\lambda)f(y)\,\mathrm{d}y$ . Why: the interchange of sum and integral is justified by the uniform convergence established in the previous derivation, and it collapses an infinite stack of integrals into one — invert once, then apply to any $f$ .

Step 4. Feed the kernel recursion back into the definition of $R$ and resum: $R=K+\lambda\int K(x,z)\,R(z,y;\lambda)\,\mathrm{d}z$ . Why: replacing each $K_{n+1}$ by $\int K\,K_n$ and re-collecting the sum reproduces $R$ under the integral, giving an equation identical in form to $\phi=f+\lambda K\phi$ with $f\to K$ and $\phi\to R$ — a closed characterisation that no longer references the series.

Step 5. Verify consistency by substituting $\phi=f+\lambda\int R f$ into the original equation and using the resolvent equation to cancel terms back to $\phi=f+\lambda K\phi$ . Why: this closes the logic — the resolvent formula is not a guess but the exact inverse of $\mathbb{1}-\lambda K$ , and the check confirms $\mathbb{1}+\lambda R=(\mathbb{1}-\lambda K)^{-1}$ .

Go deeper · Derivation C · 1 minThe radius of convergence set by the spectrum

Derivation CThe radius of convergence set by the spectrum

The goal is to pin down exactly how far the Neumann series reaches by treating it as a genuine power series in $\lambda$ with operator-valued coefficients, applying the Cauchy–Hadamard root test, and translating the growth rate of $\lVert K^n\rVert$ into spectral language through Gelfand's formula — so the abstract radius becomes the concrete statement that the disc of convergence stops at the nearest pole, the smallest-modulus characteristic value.

Step 1. View $R(x,y;\lambda)=\sum_{n\ge0}\lambda^nK_{n+1}(x,y)$ as a power series whose coefficient of $\lambda^n$ is the iterated kernel $K_{n+1}$ , equivalently the operator $K^{n+1}$ . Why: convergence of any power series is governed by the root test applied to the size of its coefficients, here the operator norms $\lVert K^{n}\rVert$ .

Step 2. Apply Cauchy–Hadamard: the radius is $\rho=\big(\limsup_n\lVert K^{n}\rVert^{1/n}\big)^{-1}$ . Why: this is the standard radius for a power series with normed coefficients; the $\limsup$ measures the exponential growth rate of the iterated operators.

Step 3. Invoke Gelfand's spectral-radius formula, $\lim_n\lVert K^{n}\rVert^{1/n}=r(K)=\max_n|\mu_n|$ , the largest modulus among the eigenvalues $\mu_n$ of $K$ . Why: the norm-growth rate equals the spectral radius even for non-self-adjoint $K$ , so the analytic radius is controlled purely by the spectrum, not by the possibly larger norm $\lVert K\rVert$ .

Step 4. The resolvent poles occur where $\mathbb{1}-\lambda K$ is singular, i.e. at the characteristic values $\lambda_n=1/\mu_n$ ; the nearest is $\lambda_1=1/\mu_{\max}$ with $|\lambda_1|=1/r(K)$ , and this equals the radius: $\rho=1/r(K)=|\lambda_1|$ . Why: the disc of convergence of a power series always extends exactly to its nearest singularity, so the series stops precisely at the first pole — the smallest-modulus characteristic value — and generally reaches farther than the crude norm bound $1/\lVert K\rVert$ because $r(K)\le\lVert K\rVert$ .

\rho=\frac{1}{\limsup_n\lVert K^{n}\rVert^{1/n}}=\frac{1}{r(K)}=|\lambda_1|

This result is consistent with the pole structure $R=D(x,y;\lambda)/D(\lambda)$ , where $D(\lambda_1)=0$ . For the worked kernel $K(x,y)=xy$ on the unit interval, $r(K)=\tfrac13$ and $\lambda_1=3$ , giving $\rho=3$ — confirmed numerically against a 64-point Gauss–Legendre discretisation.

Limiting cases

Whether the series converges, and how far, depends on the kernel class and on $\lambda$ :

Case	Result	Meaning
Contraction regime	$\vert\lambda\vert\,\lVert K\rVert<1$	Series converges absolutely and uniformly; the map $T\phi=f+\lambda K\phi$ is a contraction, so a unique solution exists by the Banach fixed-point theorem.
Sharp radius	$\vert\lambda\vert<\vert\lambda_1\vert=1/r(K)$	$\lambda_1$ is the smallest-modulus characteristic value ( $=1/\mu_{\max}$ ); this exact radius is generally larger than the norm bound $1/\lVert K\rVert$ since $r(K)\le\lVert K\rVert$ .
Volterra kernel	$\vert K_n(x,y)\vert\le\dfrac{M^n(x-y)^{n-1}}{(n-1)!}$	Iterated kernels decay factorially; the resolvent is entire and the series converges for every $\lambda$ .
Resonance / eigenvalue	$\lambda\to\lambda_n$	Resolvent $R$ develops a pole and the series diverges; the Fredholm determinant $D(\lambda)$ and minor $D(x,y;\lambda)$ supply the analytic continuation.

Common trapDivergence of the series is not the absence of a solution

Two traps sit close together. First, the Neumann series converges only for $|\lambda|<|\lambda_1|$ ; the moment $\lambda$ passes the nearest characteristic value the series diverges — but a solution can still exist. If $\lambda$ is not itself a characteristic value, the Fredholm alternative guarantees a unique solution, and $R=D(x,y;\lambda)/D(\lambda)$ delivers it by analytic continuation, poles and all — so never read a diverging series as a missing solution. Second, mind the convention on the word eigenvalue. The kernel eigenvalues $\mu$ satisfy $K\phi=\mu\phi$ ; the characteristic values $\lambda$ of the integral equation satisfy $\phi=\lambda K\phi$ , so $\lambda=1/\mu$ . The radius is $|\lambda_1|=1/r(K)$ , the smallest-modulus characteristic value — writing it as $1/|\lambda_1|$ silently swaps the two conventions and inverts the answer. When in doubt, anchor on the norm bound $|\lambda|\,\lVert K\rVert<1$ .

Worked examples

1 · A degenerate kernel solved end to end (confidence builder)

Solve the Fredholm equation $\phi(x)=x+\lambda\int_0^1 xy\,\phi(y)\,\mathrm{d}y$ for the separable kernel $K(x,y)=xy$ . Find the eigenvalue, the resolvent, the radius of convergence, and the explicit solution at $\lambda=1$ , and watch the Neumann series converge.

Step 1. The kernel is rank one, so the operator acts as $(K\phi)(x)=x\int_0^1 y\,\phi(y)\,\mathrm{d}y$ . Try $\phi=x$ : then $\int_0^1 y\cdot y\,\mathrm{d}y=\tfrac13$ , so $K x=\tfrac13 x$ . The single non-zero kernel eigenvalue is $\mu=\mathbf{1/3}$ , the spectral radius is $r(K)=\tfrac13$ , and the characteristic value is $\lambda_1=1/\mu=\mathbf{3}$ .

Step 2. Build the iterated kernels. $K_1=xy$ ; $K_2(x,y)=\int_0^1 xz\cdot zy\,\mathrm{d}z=xy\int_0^1 z^2\,\mathrm{d}z=\tfrac13xy$ ; and in general $K_{n+1}(x,y)=\tfrac{1}{3^n}xy$ . The resolvent sums the geometric series $R(x,y;\lambda)=\sum_{n\ge0}\lambda^n\tfrac{1}{3^n}xy=\dfrac{xy}{1-\lambda/3}=\mathbf{\dfrac{3xy}{3-\lambda}}$ , with its single pole at $\lambda=3=\lambda_1$ . The radius of convergence is $\rho=1/r(K)=\mathbf{3}$ .

Step 3. Solve at $\lambda=1$ with $f=x$ using the resolvent formula: $\phi(x)=x+\int_0^1\dfrac{3xy}{3-1}\,y\,\mathrm{d}y=x+\dfrac{3x}{2}\int_0^1 y^2\,\mathrm{d}y=x+\dfrac{3x}{2}\cdot\dfrac13=\mathbf{\dfrac32 x}$ . Check by direct substitution: with $\phi=cx$ , the equation gives $cx=x+x\int_0^1 y\cdot cy\,\mathrm{d}y=x+\tfrac{c}{3}x$ , so $c(1-\tfrac13)=1$ and $c=\tfrac32$ . The two routes agree.

Step 4. Watch the raw Neumann series: $K^nf=(\tfrac13)^n x$ , so the partial sums are $1,\tfrac43,\tfrac{13}{9},\tfrac{40}{27},\dots=1,1.333,1.444,1.481,\dots\to\tfrac32$ , each term exactly $\tfrac13$ of the last — the geometric collapse drawn in the first figure. Solution $\phi=\tfrac32x$ ; resolvent $R=3xy/(3-\lambda)$ ; radius $\rho=3$ .

2 · A Volterra kernel with an entire resolvent

Solve the Volterra equation $\phi(x)=1+\lambda\int_0^x\phi(y)\,\mathrm{d}y$ (constant kernel $K=1$ , running upper limit). Show the resolvent is entire and the series converges for every $\lambda$ .

Step 1. Compute the iterated kernels on the shrinking triangle $0\le y\le x$ . $K_1=1$ ; $K_2(x,y)=\int_y^x 1\,\mathrm{d}z=(x-y)$ ; $K_3(x,y)=\int_y^x(z-y)\,\mathrm{d}z=\tfrac{(x-y)^2}{2}$ ; and by induction $K_n(x,y)=\dfrac{(x-y)^{n-1}}{(n-1)!}$ . The factorial in the denominator is the hallmark of the Volterra structure — the domain of each new integral is smaller than the last.

Step 2. Sum the resolvent: $R(x,y;\lambda)=\sum_{n\ge0}\lambda^n\dfrac{(x-y)^n}{n!}=\mathbf{e^{\lambda(x-y)}}$ . Because the majorant is the exponential series, it converges for all $\lambda$ — the resolvent is an entire function of $\lambda$ , with no poles anywhere. There is no finite radius to speak of; $r(K)=0$ effectively, since the iterated kernels decay factorially.

Step 3. Solve with $f=1$ : $\phi(x)=1+\lambda\int_0^x e^{\lambda(x-y)}\,\mathrm{d}y=1+e^{\lambda x}-1=\mathbf{e^{\lambda x}}$ . Cross-check by differentiating the original equation: $\phi'=\lambda\phi$ with $\phi(0)=1$ gives $\phi=e^{\lambda x}$ immediately, and a direct quadrature at $\lambda=0.7$ reproduces it to machine precision. Resolvent $R=e^{\lambda(x-y)}$ (entire); solution $\phi=e^{\lambda x}$ , convergent for every $\lambda$ .

Physics of the portrait

The Neumann series is where a nineteenth-century trick of successive approximation grew, in barely forty years, into the operator theory of eigenvalues and spectra.

Origins · 1 minCarl Neumann and how this was discovered

Carl Neumann

1832–1925 · potential theory and integral equations

Carl Gottfried Neumann, professor at Leipzig and a founder of Mathematische Annalen, gave the series its name in his 1877 study of the logarithmic and Newtonian potentials, where he solved boundary-value problems by successive approximation — building the solution as an infinite sum of repeatedly-integrated corrections. The idea was not born fully formed: Joseph Liouville had used iterated integration for differential equations as early as 1837, which is why the object is often called the Liouville–Neumann series. His method worked when the iterations converged and fell silent when they did not; the general theory came in a rush at the turn of the century. Vito Volterra showed in the 1890s that for his triangular kernels the series always converges, the factorial-decay case. Then in 1903 the Swedish actuary-turned-mathematician Erik Ivar Fredholm published, in Acta Mathematica, the determinant $D(\lambda)$ and minor $D(x,y;\lambda)$ that continue the resolvent past the radius of convergence and locate its poles at the eigenvalues — the first complete theory of second-kind equations. David Hilbert and Erhard Schmidt, between 1904 and 1907, recast the subject as the spectral theory of symmetric kernels, coining the word eigenvalue in this setting and building the Hilbert space that quantum mechanics would later inhabit. Stefan Banach's fixed-point theorem of 1922 gave Neumann's convergence its cleanest proof. A working trick had become a cathedral of analysis.

History

1837

— Liouville solves differential equations by iterated integration (successive approximation)

1877

— Carl Neumann sums the series for the logarithmic and Newtonian potentials

1896

— Volterra proves unconditional convergence for triangular (Volterra) kernels

1903

— Fredholm publishes the determinant and resolvent theory of second-kind equations

1904

— Hilbert develops the spectral theory of symmetric kernels and eigenvalues

1922

— Banach's fixed-point theorem certifies the contraction-mapping convergence

Going further

Texts. Stone & Goldbart, Mathematics for Physics, ch. on integral equations · Arfken & Weber, Mathematical Methods for Physicists, Integral Equations chapter · Byron & Fuller, Mathematics of Classical and Quantum Physics, Ch. 9 · Hassani, Mathematical Physics, integral-equations chapters · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering · Courant & Hilbert, Methods of Mathematical Physics, Vol. I · Tricomi, Integral Equations. The key numerical results are as follows. For the degenerate kernel $K=xy$ on $[0,1]$ : $\mu=\tfrac13$ , $\lambda_1=3$ , $\rho=3$ , $R=3xy/(3-\lambda)$ , and $\phi=\tfrac32x$ at $\lambda=1$ (confirmed against a 64-point Gauss–Legendre discretisation and by direct substitution). For the Volterra kernel $K=1$ : $K_n=(x-y)^{n-1}/(n-1)!$ , $R=e^{\lambda(x-y)}$ , and $\phi=e^{\lambda x}$ (confirmed by quadrature at $\lambda=0.7$ ).

Builds on. Linear Integral Operators and Kernels · Green Functions · Operator Norms and Banach Spaces · Eigenvalue Problems for Compact Operators. Leads to. WKB Approximation · Borel Summation and Resurgence · Asymptotic Series and Poincaré Expansions.

Ask the tutor: Ask me to solve a specific integral equation by both the raw Neumann series and its resolvent, and I will show you term by term where they agree and where the pole in $\lambda$ first bites.

How is the Neumann series related to perturbation theory and the Born series?

They are the same construction wearing different clothes. Perturbation theory expands an unknown in powers of a small coupling; the Neumann series does exactly that with $\lambda K$ as the small quantity, each order adding one more action of the operator. In scattering, the Lippmann–Schwinger equation $\psi=\psi_0+G_0V\psi$ is a second-kind equation with kernel $G_0V$ , and its Neumann series is precisely the Born series, the $n$ -th term describing $n$ successive scatterings; the first Born approximation is just the truncation $\psi\approx\psi_0+G_0V\psi_0$ . So "resum the Neumann series" and "sum the perturbation series" are one instruction.

What is the difference between a kernel eigenvalue and a characteristic value?

It is a convention that trips up almost everyone. A kernel eigenvalue $\mu$ solves $K\phi=\mu\phi$ — the ordinary eigenvalue of the operator. A characteristic value $\lambda$ of the integral equation solves the homogeneous problem $\phi=\lambda K\phi$ , i.e. $K\phi=(1/\lambda)\phi$ , so $\lambda=1/\mu$ . Classical Fredholm theory (Tricomi, Courant–Hilbert) uses the characteristic values, and it is at those $\lambda$ that the resolvent has poles. The radius of convergence is therefore $|\lambda_1|=1/r(K)$ , the smallest-modulus characteristic value, which equals the reciprocal of the largest kernel eigenvalue. If a source writes the radius as $1/|\lambda_1|$ , it is quietly using $\lambda$ to mean the kernel eigenvalue $\mu$ instead — check which convention is in force before trusting the formula.

Why does a Volterra equation always converge but a Fredholm one only sometimes?

The difference is the upper limit of integration. In a Fredholm equation it is fixed at $b$ , so every iterated integral covers the same full domain and $\lVert K^n\rVert$ decays at best geometrically — the series converges only for $|\lambda|<|\lambda_1|$ . In a Volterra equation the upper limit is the running point $x$ , so each integration is confined to a smaller triangle. That shrinking domain injects a factor $1/n!$ into the iterated kernels, $K_n=(x-y)^{n-1}/(n-1)!$ for the constant kernel, and the factorial beats any power of $\lambda$ . The series is dominated by the exponential and converges everywhere, making the resolvent entire with no poles.

If the series diverges, is there still a solution?

Usually yes. Divergence only means $|\lambda|$ has exceeded the nearest pole of the resolvent; it says nothing about solvability. The Fredholm alternative settles it: if $\lambda$ is not a characteristic value, $\mathbb{1}-\lambda K$ is invertible and there is a unique solution for every $f$ — you just cannot reach it by summing the divergent series. Instead use the Fredholm representation $R=D(x,y;\lambda)/D(\lambda)$ , a ratio of entire functions that continues the resolvent to the whole plane, poles included. Only when $\lambda$ lands exactly on a characteristic value does solvability turn conditional, requiring $f$ orthogonal to the adjoint null space.

What is the relation between the resolvent kernel and the resolvent operator?

The resolvent operator is $(\mathbb{1}-\lambda K)^{-1}$ , the abstract inverse mapping any source $f$ to the solution $\phi$ . The resolvent kernel $R(x,y;\lambda)$ is its integral representation minus the identity: since $\phi=f+\lambda\int R\,f$ , one has $(\mathbb{1}-\lambda K)^{-1}=\mathbb{1}+\lambda R$ . Spectral theory more often uses the operator resolvent $(z-K)^{-1}$ of a complex $z$ ; setting $z=1/\lambda$ connects them, with poles in $z$ at the kernel eigenvalues $\mu_n$ and poles in $\lambda$ at the characteristic values $1/\mu_n$ — the same information in two coordinates.

How many terms of the series do I actually need to sum?

The error after $N$ terms is bounded by the geometric tail, roughly $(|\lambda|/|\lambda_1|)^{N+1}$ times the leading term, so the rate is set by how close $\lambda$ sits to the first pole. Deep inside the disc — as in the worked $K=xy$ example at $\lambda=1$ with $|\lambda_1|=3$ — each term shrinks by a factor of $\tfrac13$ and three or four terms give three-figure accuracy. As $\lambda$ approaches $\lambda_1$ the ratio nears one and convergence crawls, so practitioners switch to the resolvent's closed form or to continuation methods. The truncated series is cheap only when the problem is comfortably contractive.