Modified Green's Functions and Solvability

A Green's function is the workhorse of linear physics: it is the kernel $G(x,\xi)$ that inverts a differential operator $L$ , so that $u=\int G\,f$ solves $Lu=f$ for any forcing $f$ . Writing $u=L^{-1}f$ this way silently assumes $L$ has an inverse — but a startling number of the operators physicists care most about do not. The Laplacian with pure Neumann (no-flux) walls annihilates constants; the Schrödinger operator $H_0-E_0$ annihilates its own eigenstate at a resonance; the fluctuation operator around a soliton annihilates the translations and rotations that move the soliton for free. Each has a nonzero $\psi_0$ with $L\psi_0=0$ obeying the boundary conditions — a zero mode — and for every one of them the ordinary Green's function simply fails to exist. The obstruction is a sharp dichotomy, the Fredholm alternative. For a self-adjoint $L$ exactly one of two worlds holds. Either $L$ has no zero modes and $Lu=f$ has a unique solution for every $f$ ; or $L$ kills some $\psi_0$ , and then $Lu=f$ is solvable only when the source is orthogonal to every zero mode, $\langle\psi_0,f\rangle=0$ , and the solution is fixed only up to additions of zero modes. A point source $\delta(x-\xi)$ almost never passes this test, which is precisely why $LG=\delta$ has no solution. The modified (or generalized) Green's function repairs the machine. One subtracts from $\delta$ exactly its projection onto the kernel, leaving a source the operator can invert, then fixes the leftover freedom by demanding $G_{\mathrm M}$ be orthogonal to the zero modes. The result is the integral-operator twin of the Moore–Penrose pseudoinverse: it inverts $L$ on the orthogonal complement of the kernel and is the zero map on the kernel itself. This one construction stitches together Sturm–Liouville theory, the spectral theorem, degenerate perturbation theory, and the removal of secular terms in multiple-scale analysis.

Key takeaways

When a self-adjoint operator $L$ has a zero eigenvalue, its Green's function does not exist: $L\,G=\delta$ is unsolvable because the point source $\delta$ overlaps the zero mode $\psi_0$ .
The Fredholm alternative sets the rule: $L\,u=f$ has a solution if and only if the source is orthogonal to every zero mode, $\langle\psi_0,f\rangle=0$ .
The cure is the modified Green's function $G_{\mathrm M}$ , defined by $L\,G_{\mathrm M}=\delta(x-\xi)-\sum_i\psi_0^{(i)}(x)\psi_0^{(i)}(\xi)$ — the point source with its kernel projection removed.
Spectrally $G_{\mathrm M}=\sum_{\lambda_n\neq0}\psi_n(x)\psi_n(\xi)/\lambda_n$ : invert eigenvalue by eigenvalue on the nonzero spectrum and omit the zero mode outright — the exact integral-operator Moore–Penrose pseudoinverse.
Uniqueness needs a second condition, $\langle\psi_0^{(i)},G_{\mathrm M}\rangle=0$ , which removes the residual zero-mode freedom and makes $G_{\mathrm M}$ symmetric.
Key idea: a singular self-adjoint operator cannot be inverted, only pseudo-inverted — project the source off the kernel, invert on the complement, and $G_{\mathrm M}=\sum_{\lambda_n\neq 0}\psi_n(x)\psi_n(\xi)/\lambda_n$ is the unique symmetric kernel that does exactly that.

The core insight

A singular self-adjoint operator cannot be inverted, only pseudo-inverted: project the source off the kernel, invert on the orthogonal complement, and the modified Green's function $G_{\mathrm M}=\sum_{\lambda_n\neq 0}\psi_n(x)\psi_n(\xi)/\lambda_n$ is the unique symmetric kernel doing precisely that.

Where this lives today

The modified Green's function is the everyday tool wherever a symmetry produces a zero mode. In computational field theory it is the propagator one must use when quantizing fluctuations around a soliton, kink, or instanton: the translational and rotational symmetries that move the classical solution for free appear as exact zero modes of the fluctuation operator, and both the one-loop determinant (a pseudo-determinant on the orthogonal complement) and the two-point function are defined only after those directions are removed and handled by collective coordinates. Lattice gauge theory meets the same issue numerically whenever a near-zero Dirac eigenvalue appears, and deflation algorithms — projecting the low modes out before inverting — are literally the modified Green's function in code, keeping the conjugate-gradient solver from stalling. In electrostatics on closed manifolds (a torus, a periodic simulation cell) the constant is always a zero mode, so Poisson solvers enforce charge neutrality and return a potential defined only up to a constant — the reason particle-mesh Ewald codes subtract a uniform background. And in applied mathematics the same condition drives the Lyapunov–Schmidt reduction and secular-term removal in multiple-scale and Poincaré–Lindstedt expansions, where orthogonality to the resonant mode is what keeps a perturbation series from diverging.

The physical picture

Replace the operator $L$ by a Hermitian matrix $A$ with one zero eigenvalue, and the whole story becomes elementary linear algebra. The equation $A\mathbf u=\mathbf f$ is solvable only when $\mathbf f$ has no component along the null eigenvector $\mathbf v_0$ ; the pseudoinverse $A^{+}$ then inverts $A$ on the orthogonal complement and sends $\mathbf v_0$ to zero. The modified Green's function is nothing but $A^{+}$ promoted to a continuous kernel, with the null vector $\mathbf v_0$ replaced by the zero-mode function $\psi_0(x)$ . Every eigenvalue $\lambda_n\neq0$ is inverted to $1/\lambda_n$ ; the offending $\lambda_0=0$ , which would demand division by zero, is simply struck from the sum. The first figure is that pseudoinverse spectrum, drawn for a concrete operator.

Figure

The pseudoinverse spectrum for the Neumann Laplacian on a unit interval: L has eigenvalues λ₀ = 0, λ₁ = 9.87, λ₂ = 39.5, λ₃ = 88.8, and G_M inverts each nonzero eigenvalue to 1/λₙ while dropping the λ = 0 zero mode entirely

Why must the ordinary construction fail? A point source $\delta(x-\xi)$ , resolved in the eigenbasis, has some overlap with every eigenfunction, the zero mode included. Inverting term by term produces the fatal contribution $\psi_0(x)\psi_0(\xi)/\lambda_0$ with $\lambda_0=0$ — an infinity. The repair is brutally simple: delete that single term from the source before inverting, leaving a projected delta the operator can actually digest. Physically a zero mode is a free, undamped collective motion — a rigid translation, a global phase, a Goldstone direction. You cannot hold such a mode in static balance: a force along it does not push against a restoring spring, it accelerates the mode forever. Solvability is the demand that the net force along the zero mode vanish; the modified Green's function quietly resolves everything transverse to it. The second figure shows the actual kernel that does this for the no-flux interval.

Figure

The Neumann modified Green's function G_M for source at ξ = 0.5 on the unit interval: a cusp of height 1/12 ≈ 0.083 at the source, symmetric about the midpoint, with equal shaded area above and below zero so that ∫₀¹ G_M dx = 0, its orthogonality to the constant zero mode

The core equations

The starting point is the definition of a Green's function as the kernel that inverts $L$ . It solves the operator equation with a unit point source and inherits the same homogeneous boundary conditions as the physical problem, so that convolving it against any forcing returns a solution.

L\,G(x,\xi)=\delta(x-\xi),\qquad G\ \text{satisfies homogeneous BCs} \qquad (1)

where: $L$ — self-adjoint operator, e.g. Sturm–Liouville $L=-\tfrac{\mathrm d}{\mathrm dx}\!\big(p\tfrac{\mathrm d}{\mathrm dx}\big)+q$ ; $G(x,\xi)$ — kernel of $L^{-1}$ ; $\delta(x-\xi)$ — Dirac point source; $\xi$ — source location. Equation (1) is meaningful only when $L$ is invertible; the rest of this article is what to do when it is not. Solvability of $Lu=f$ is governed by a single orthogonality test against the kernel — the Fredholm alternative in the form that matters for differential operators: a source can be balanced only if it has no component along a direction the operator cannot resist.

L\,u=f\ \text{solvable}\iff \langle\psi_0^{(i)},f\rangle=\int \psi_0^{(i)}(x)\,f(x)\,\mathrm dx=0\ \ \forall i \qquad (2)

where: $\psi_0^{(i)}$ — orthonormal basis of $\ker L$ (the zero modes), $L\psi_0^{(i)}=0$ ; $f$ — inhomogeneous source; $\langle\cdot,\cdot\rangle$ — $L^2$ inner product with the Sturm–Liouville weight. When $\ker L=\{0\}$ the condition is vacuous and (1) is fine; otherwise it is a genuine constraint on $f$ . To restore solvability one subtracts from the source exactly its projection onto the kernel. The right-hand side below is manifestly orthogonal to every zero mode, so the defining equation of $G_{\mathrm M}$ can be solved.

L\,G_{\mathrm M}(x,\xi)=\delta(x-\xi)-\sum_{i}\psi_0^{(i)}(x)\,\psi_0^{(i)}(\xi) \qquad (3)

where: $G_{\mathrm M}$ — the modified (generalized) Green's function; $\sum_i\psi_0^{(i)}(x)\psi_0^{(i)}(\xi)$ — the kernel of the orthogonal projector $P_0$ onto $\ker L$ . The subtracted term is the "background charge" that neutralizes the source. Equation (3) fixes $G_{\mathrm M}$ only up to an arbitrary zero mode, exactly as $A^{+}$ is fixed only after one demands it kill the null space. The uniqueness (gauge-fixing) condition removes that freedom.

\langle\psi_0^{(i)},\,G_{\mathrm M}(\cdot,\xi)\rangle=0\quad\forall i \qquad (4)

where: $G_{\mathrm M}(\cdot,\xi)$ — $G_{\mathrm M}$ viewed as a function of its first argument. With (3) and (4) together the modified Green's function is unique and, as a bonus, symmetric. Diagonalizing $L$ gives the cleanest formula of all: build $G_{\mathrm M}$ eigenvalue by eigenvalue, inverting the nonzero spectrum and omitting the zero modes. This is the spectral or eigenfunction representation.

G_{\mathrm M}(x,\xi)=\sum_{\lambda_n\neq 0}\frac{\psi_n(x)\,\psi_n(\xi)}{\lambda_n} \qquad (5)

where: $\psi_n,\lambda_n$ — orthonormal eigenfunctions and eigenvalues, $L\psi_n=\lambda_n\psi_n$ ; $\sum_{\lambda_n\neq0}$ — the sum runs over all nonzero eigenvalues, the zero modes simply left out. The omitted $\lambda=0$ term is exactly what was subtracted from $\delta$ in (3). Once $G_{\mathrm M}$ is in hand and the source passes the Fredholm test, the general solution is the particular integral plus the full homogeneous freedom of the kernel.

u(x)=\int G_{\mathrm M}(x,\xi)\,f(\xi)\,\mathrm d\xi+\sum_i a_i\,\psi_0^{(i)}(x) \qquad (6)

where: $a_i$ — arbitrary constants, the zero-mode freedom; $u(x)$ — the full solution, whose particular part $\int G_{\mathrm M}f$ is orthogonal to the kernel. Different $a_i$ correspond to different physical zero points (an undetermined overall potential, an arbitrary rigid shift). Finally the two identities that certify $G_{\mathrm M}$ as a genuine pseudoinverse. They say $G_{\mathrm M}$ inverts $L$ everywhere except on the kernel, where both compositions reduce to the identity minus the kernel projector.

L\,G_{\mathrm M}=\mathbb{1}-P_0,\qquad G_{\mathrm M}\,L=\mathbb{1}-P_0,\qquad P_0=\sum_i |\psi_0^{(i)}\rangle\langle\psi_0^{(i)}| \qquad (7)

where: $P_0$ — the orthogonal projector onto $\ker L$ ; $\mathbb 1$ — the identity operator. These are the integral-operator versions of the Moore–Penrose relations $AA^{+}=\mathbb 1-P_{\ker}$ , together with $G_{\mathrm M}\psi_0=0$ and $G_{\mathrm M}^{\dagger}=G_{\mathrm M}$ .

Derivation AWhy solvability requires orthogonality to the kernel

The goal of this derivation is to prove the Fredholm alternative in the concrete form (2): that $Lu=f$ can be solved precisely when the source $f$ is orthogonal to every zero mode of $L$ . The whole argument rides on one property — self-adjointness — which lets us slide the operator from one slot of the inner product to the other and land it on a function it annihilates. We establish necessity first, then sufficiency.

Step 1. Take any zero mode $\psi_0\in\ker L$ and pair it with both sides of $Lu=f$ : $\langle\psi_0,Lu\rangle=\langle\psi_0,f\rangle$ . Why: pairing against the kernel is the one probe that "sees" the obstruction — it projects the equation onto the direction the operator cannot balance.

Step 2. Move $L$ across by self-adjointness, $\langle\psi_0,Lu\rangle=\langle L\psi_0,u\rangle$ . Since $L\psi_0=0$ the left side collapses to $\langle 0,u\rangle=0$ . Why: the boundary terms from the integration by parts vanish because u and ψ₀ obey the same homogeneous conditions, so self-adjointness turns "L on u" into "L on ψ₀", which is nothing.

Step 3. Hence if a solution exists then $\langle\psi_0,f\rangle=0$ for every zero mode — necessity. Why: applied to $f=\delta(\cdot-\xi)$ it gives $\psi_0(\xi)\neq0$ generically, the proof that a singular operator has no ordinary Green's function.

Step 4. For sufficiency, restrict $L$ to $(\ker L)^{\perp}$ , where every eigenvalue is nonzero and $L$ is invertible; the spectral sum $u=\sum_{\lambda_n\neq0}\psi_n\langle\psi_n,f\rangle/\lambda_n$ then converges to a solution whenever $\langle\psi_0,f\rangle=0$ . Why: the kernel is the only place the inverse blows up; delete it and orthogonality is not just necessary but enough.

Go deeper · Derivation B · 1 minBuilding the modified Green's function from the spectrum

Derivation BBuilding the modified Green's function from the spectrum

The goal is to construct $G_{\mathrm M}$ from the eigenbasis and prove it is the integral-operator Moore–Penrose pseudoinverse of $L$ — that it satisfies the defining equation (3), annihilates the kernel, and is symmetric. We take a complete orthonormal eigenbasis $L\psi_n=\lambda_n\psi_n$ with a single zero mode $\psi_0$ ( $\lambda_0=0$ ) for clarity, and let completeness and orthonormality do the work termwise.

Step 1. Write the resolution of the identity $\delta(x-\xi)=\sum_n\psi_n(x)\psi_n(\xi)$ and read off the kernel projector as its $n=0$ term, $P_0(x,\xi)=\psi_0(x)\psi_0(\xi)$ . Why: this exposes the delta as a sum over modes, making visible the single term that must be excised.

Step 2. Define $G_{\mathrm M}(x,\xi)\equiv\sum_{n:\,\lambda_n\neq0}\psi_n(x)\psi_n(\xi)/\lambda_n$ , the spectral sum with the zero mode omitted. Why: dividing by $\lambda_n$ inverts each mode; skipping $\lambda_0=0$ is the only change from the ordinary Green's function and is exactly what keeps the sum finite.

Step 3. Apply $L$ in $x$ : each eigenfunction brings down its eigenvalue, $L_x G_{\mathrm M}=\sum_{\lambda_n\neq0}\psi_n(x)\psi_n(\xi)=\delta(x-\xi)-\psi_0(x)\psi_0(\xi)=\mathbb 1-P_0$ . Why: this recovers the defining equation (3) from the spectral formula, closing the loop between the two definitions.

Step 4. Check it kills the kernel: $\int G_{\mathrm M}(x,\xi)\psi_0(\xi)\,\mathrm d\xi=\sum_{\lambda_n\neq0}\psi_n(x)\langle\psi_n,\psi_0\rangle/\lambda_n=0$ since $\langle\psi_n,\psi_0\rangle=0$ for $n\neq0$ . Why: this is the Penrose condition $A^{+}\mathbf v_0=0$ , so gauge condition (4) holds automatically.

Step 5. Feed a source through it: $u=\int G_{\mathrm M}f$ gives $Lu=(\mathbb 1-P_0)f=f-\psi_0\langle\psi_0,f\rangle$ , which equals $f$ exactly when $\langle\psi_0,f\rangle=0$ . Why: the spectral construction reproduces the Fredholm alternative from an independent starting point — a strong consistency check. Symmetry is manifest in the numerator $\psi_n(x)\psi_n(\xi)$ .

Go deeper · Derivation C · 2 minThe Neumann Laplacian modified Green's function in closed form

Derivation CThe Neumann Laplacian modified Green's function in closed form

The final derivation carries the abstract machine down to a fully explicit answer for the operator $L=-\mathrm d^2/\mathrm dx^2$ on the interval from $0$ to $\ell$ with no-flux boundary conditions $u'(0)=u'(\ell)=0$ . This operator annihilates constants, so its normalized zero mode is $\psi_0=1/\sqrt\ell$ and it is genuinely singular. We build $G_{\mathrm M}$ by direct integration, imposing neutrality, the Neumann conditions, continuity, the correct derivative jump, and orthogonality — and we cross-check the closed form against the eigenfunction sum.

Step 1. Write the modified defining equation with the projector kernel $\psi_0(x)\psi_0(\xi)=1/\ell$ subtracted: $-\partial_x^2 G_{\mathrm M}=\delta(x-\xi)-1/\ell$ . Why: the uniform $-1/\ell$ is a smeared background that makes the total source neutral, $\int_0^\ell(\delta-1/\ell)\,\mathrm dx=1-1=0$ — exactly the Fredholm condition for the Neumann problem, which insists that net flux through insulating walls be zero.

Step 2. Integrate once. The background gives $x/\ell$ and the delta gives a unit downward step, so $\partial_x G_{\mathrm M}=x/\ell-\theta(x-\xi)+A$ . The left wall condition $\partial_x G_{\mathrm M}(0)=0$ forces $A=0$ , and the right wall then checks out automatically: $\partial_x G_{\mathrm M}(\ell)=\ell/\ell-1=0$ . Why: the right-hand boundary condition is satisfied for free precisely because we subtracted $1/\ell$ — a consistency that would fail for the naive unmodified source, which is another face of non-existence.

Step 3. Integrate a second time and write the result symmetrically with $x_>=\max(x,\xi)$ : $G_{\mathrm M}=\dfrac{x^2+\xi^2}{2\ell}-x_>+C$ , where continuity at $x=\xi$ has already been used and $C$ is a single constant. Why: casting the answer in terms of $x_>$ makes the symmetry $x\leftrightarrow\xi$ manifest and packages the two-piece $\theta$ -function form into one expression.

Step 4. Fix $C$ by orthogonality to the zero mode, $\int_0^\ell G_{\mathrm M}\,\mathrm dx=0$ . Evaluating, $\int_0^\ell\frac{x^2+\xi^2}{2\ell}\mathrm dx=\frac{\ell^2}{6}+\frac{\xi^2}{2}$ and $\int_0^\ell x_>\,\mathrm dx=\frac{\ell^2}{2}+\frac{\xi^2}{2}$ , whose difference is $-\ell^2/3$ ; setting the total to zero gives $C=\ell/3$ . Hence $G_{\mathrm M}(x,\xi)=\dfrac{x^2+\xi^2}{2\ell}-x_>+\dfrac{\ell}{3}$ . Why: the $\xi^2$ pieces cancel in the integral, leaving a constant equation for $C$ — the gauge condition (4) doing its one job of pinning the undetermined additive level.

Step 5. Cross-check against the spectrum. The Neumann eigenpairs are $\psi_n=\sqrt{2/\ell}\cos(n\pi x/\ell)$ , $\lambda_n=(n\pi/\ell)^2$ for $n\geq1$ , so $G_{\mathrm M}=\frac{2\ell}{\pi^2}\sum_{n\geq1}\frac{\cos(n\pi x/\ell)\cos(n\pi\xi/\ell)}{n^2}$ ; summing this Fourier series with the identity $\sum_{n\geq1}\cos(n\theta)/n^2=\pi^2/6-\pi\theta/2+\theta^2/4$ reproduces the boxed closed form term for term. Why: two independent constructions — direct integration and the spectral sum (5) — landing on the same kernel is the strongest possible verification that $G_{\mathrm M}$ is right. The boxed result:

G_{\mathrm M}(x,\xi)=\frac{x^2+\xi^2}{2\ell}-x_>+\frac{\ell}{3},\qquad x_>=\max(x,\xi)

Limiting cases

Case	Result	Meaning
No zero modes	$\ker L=\{0\}$	The Fredholm test is vacuous; $Lu=f$ has a unique solution for every $f$ and $G_{\mathrm M}$ collapses to the ordinary Green's function $G=\sum_n\psi_n(x)\psi_n(\xi)/\lambda_n$ .
Source orthogonal to kernel	$\langle\psi_0,f\rangle=0$	The inhomogeneous problem is solvable; $\int G_{\mathrm M}f$ is a valid particular solution, unique up to $u\to u+c\,\psi_0$ (the zero-mode freedom).
Source not orthogonal	$\langle\psi_0,f\rangle\neq 0$	No static solution exists at all; the component of the forcing along the zero mode is unbalanced and the response grows secularly. This is the failure that solvability conditions are designed to forbid.
Degenerate kernel	$\dim\ker L=m>1$	Subtract the full projector $P_0=\sum_{i=1}^{m}\psi_0^{(i)}(x)\psi_0^{(i)}(\xi)$ ; solvability now requires $\langle\psi_0^{(i)},f\rangle=0$ for every one of the $m$ zero modes, and $u$ carries $m$ free constants.

Common trapForcing L G = δ when L has a zero mode

Do not try to solve $L\,G=\delta(x-\xi)$ when $L$ has a zero mode — the equation is genuinely insoluble, and any "solution" you write will secretly violate a boundary condition or hide a divergent $\psi_0(x)\psi_0(\xi)/\lambda_0$ term. Two separate repairs are needed and neither can be skipped. First, subtract the zero-mode projection of the source, so the right-hand side becomes $\delta-P_0$ and passes the Fredholm test. Second, impose orthogonality of $G_{\mathrm M}$ to the kernel, condition (4), which removes the residual zero-mode freedom left over in $G_{\mathrm M}$ itself. Omitting the first makes the problem unsolvable; omitting the second leaves it ambiguous by an arbitrary multiple of $\psi_0$ .

Worked examples

1 · The Neumann modified Green's function, evaluated and checked (fully worked)

Take $\ell=1$ and a source at the midpoint $\xi=0.5$ . Evaluate $G_{\mathrm M}(x,0.5)$ , confirm its peak and its zero total area, and check how fast the eigenfunction sum converges to it.

Step 1. Write the two branches from the boxed formula $G_{\mathrm M}=\frac{x^2+\xi^2}{2}-x_>+\frac13$ . For $x\le0.5$ the maximum is $\xi$ , giving $G_{\mathrm M}=\frac{x^2}{2}+\frac{0.25}{2}-0.5+\frac13=\frac{x^2}{2}-0.04167$ ; for $x\ge0.5$ the maximum is $x$ , giving $G_{\mathrm M}=\frac{x^2}{2}-x+0.45833$ . The two branches agree at $x=0.5$ , where $G_{\mathrm M}=0.08333=\mathbf{1/12}$ — the cusp height. The kernel peaks at the source with value 1/12.

Step 2. Read off the wall values: $G_{\mathrm M}(0,0.5)=-0.04167=\mathbf{-1/24}$ and by symmetry $G_{\mathrm M}(1,0.5)=-1/24$ . So the kernel is a symmetric hump, positive near the source and dipping to $-1/24$ at both insulating walls — negative there because material must be depleted relative to the mean to keep the total balanced. The kernel runs from +1/12 at the source down to −1/24 at the walls.

Step 3. Verify orthogonality directly by integrating the two branches. The left half gives $\int_0^{0.5}(\tfrac{x^2}{2}-\tfrac{1}{24})\,\mathrm dx=\tfrac{1}{48}-\tfrac{1}{48}=0$ , and the right half $\int_{0.5}^{1}(\tfrac{x^2}{2}-x+\tfrac{11}{24})\,\mathrm dx=\tfrac18-\tfrac18=0$ as well, so $\int_0^1 G_{\mathrm M}\,\mathrm dx=0$ exactly. The positive area near the source cancels the negative area at the walls: $\langle\psi_0,G_{\mathrm M}\rangle=0$ .

Step 4. Test the spectral sum $G_{\mathrm M}=\frac{2}{\pi^2}\sum_{n\geq1}\frac{\cos(n\pi x)\cos(n\pi/2)}{n^2}$ at the source $x=0.5$ . Only even $n$ survive (since $\cos(n\pi/2)=0$ for odd $n$ , and $=\pm1$ for even $n$ ), so the value is $\frac{2}{\pi^2}\sum_{m\geq1}\frac{1}{(2m)^2}=\frac{2}{\pi^2}\cdot\frac14\cdot\frac{\pi^2}{6}=\frac{1}{12}$ . The Fourier series sums exactly to the closed-form peak 1/12 — the two constructions agree to all orders. Truncating at $n=10$ already reproduces $1/12$ to better than $1\%$ , because the $1/n^2$ tail converges quickly.

2 · Solvability fixes the first-order energy shift

A particle in a one-dimensional box of width $L$ (units $\hbar^2/2m=1$ ) has ground state $\psi_0=\sqrt{2/L}\sin(\pi x/L)$ and energy $E_0=(\pi/L)^2$ . Switch on a weak linear "Stark" ramp $H'=\lambda x$ . Show that the Fredholm alternative alone determines the first-order energy shift, and evaluate it for $L=1$ , $\lambda=0.1$ .

Step 1. The Rayleigh–Schrödinger equation at first order is $(H_0-E_0)\psi_1=(E_1-H')\psi_0$ . The operator on the left is $H_0-E_0$ , which annihilates $\psi_0$ — it is exactly the singular, zero-mode situation of this article, with $\psi_0$ the zero mode. So the right-hand side must pass the Fredholm test. The unknown $\psi_1$ can exist only if $(E_1-H')\psi_0\perp\psi_0$ .

Step 2. Impose $\langle\psi_0,(E_1-H')\psi_0\rangle=0$ . Since $\psi_0$ is normalized this reads $E_1-\langle\psi_0,H'\psi_0\rangle=0$ , so $E_1=\langle\psi_0,H'\psi_0\rangle=\lambda\langle x\rangle$ . Solvability, not any extra assumption, forces $E_1=\lambda\langle x\rangle$ — the standard first-order formula falls straight out of the Fredholm condition.

Step 3. Evaluate the expectation. By symmetry of $\sin^2(\pi x/L)$ about the midpoint, $\langle x\rangle=\frac{2}{L}\int_0^L x\sin^2(\pi x/L)\,\mathrm dx=\frac{2}{L}\cdot\frac{L^2}{4}=\frac{L}{2}$ . Hence $E_1=\lambda L/2$ . For $L=1$ , $\lambda=0.1$ this is $E_1=\mathbf{0.05}$ . The linear ramp shifts the ground-state energy by $\lambda L/2 = 0.05$ .

Step 4. Only now can $\psi_1$ be built, and it is built from the modified Green's function of $H_0-E_0$ — the reduced resolvent $G_{\mathrm M}=\sum_{n\neq0}|\psi_n\rangle\langle\psi_n|/(E_n-E_0)$ — acting on the (now solvable) source: $\psi_1=G_{\mathrm M}(E_1-H')\psi_0$ . The energy shift comes from solvability; the wavefunction correction comes from the modified Green's function — the two halves of degenerate perturbation theory.

Physics of the portrait

The modified Green's function is the meeting point of three lineages: the integral-equation revolution of the 1900s, the spectral theory that followed it, and the linear-algebra pseudoinverse that turned out to be the same idea in finite dimensions. The name on the central theorem belongs to a Stockholm actuary-mathematician whose short 1903 paper reorganized the whole subject.

Origins · 1 minIvar Fredholm and how this was discovered

Ivar Fredholm

1866–1927, Swedish mathematician and actuary · integral equations and the solvability alternative

Erik Ivar Fredholm studied in Stockholm under Mittag-Leffler and worked much of his career as an insurance actuary while holding a professorship at Stockholm University. His fame rests almost entirely on one paper — the 1903 Acta Mathematica memoir on integral equations of the form $\phi(x)+\int K(x,y)\phi(y)\,\mathrm dy=f(x)$ — in which he showed that such an equation behaves exactly like a finite linear system: either a unique solution for every $f$ , or a nontrivial homogeneous solution and solvability only for $f$ orthogonal to the transposed homogeneous solutions. That dichotomy, the Fredholm alternative, is what governs when a Green's function exists and when it must be modified. He reached it by treating the integral operator as the continuous limit of a matrix and carrying the determinant along — his "Fredholm determinant" still has the eigenvalues as its zeros. Hilbert seized on the work, building from it (1904–1910) the spectral theory of symmetric kernels that underlies every modern treatment, and Erhard Schmidt gave it its Hilbert-space polish. Fredholm published little else of comparable weight, but the alternative bearing his name is one of the load-bearing theorems of mathematical physics, quietly deciding the solvability of boundary-value problems, perturbation expansions, and quantum scattering alike.

History

1877

— Rayleigh's Theory of Sound frames solvability conditions for perturbed normal modes

1903

— Fredholm publishes his theory of integral equations and the alternative that bears his name

1920

— E. H. Moore introduces the general reciprocal of a matrix, the first pseudoinverse

1924

— Courant and Hilbert present the generalized (modified) Green's function for singular problems

1955

— Penrose gives the four defining conditions of the generalized inverse, fixing the modern form

Going further

Texts. Stone and Goldbart, Mathematics for Physics, chapter on Green functions (modified Green function and the Fredholm alternative, the closest match to this article's viewpoint) · Arfken, Weber and Harris, Mathematical Methods for Physicists, 7th ed., Green's-function chapter · Hassani, Mathematical Physics, on Sturm–Liouville systems and generalized Green's functions · Byron and Fuller, Mathematics of Classical and Quantum Physics, on the Fredholm theory of integral equations · Courant and Hilbert, Methods of Mathematical Physics, Vol. I, on the generalized Green's function · Riley, Hobson and Bence, Mathematical Methods for Physics and Engineering, 3rd ed., ODE Green's-function sections. The key results in this article are: the Fredholm solvability condition follows from self-adjointness and $L\psi_0=0$ , with boundary terms vanishing under the shared homogeneous conditions; the Neumann modified Green's function on $[0,\ell]$ is obtained by direct integration — neutrality $\int(\delta-1/\ell)=0$ , both Neumann conditions, the unit derivative jump, and orthogonality $\int G_{\mathrm M}=0$ — giving $G_{\mathrm M}=(x^2+\xi^2)/2\ell-x_>+\ell/3$ ; the eigenfunction sum $\frac{2\ell}{\pi^2}\sum_{n\geq1}\cos(n\pi x/\ell)\cos(n\pi\xi/\ell)/n^2$ reproduces that closed form via $\sum\cos(n\theta)/n^2=\pi^2/6-\pi\theta/2+\theta^2/4$ ; and the perturbation result $E_1=\lambda L/2=0.05$ for $L=1$ , $\lambda=0.1$ follows from $\langle x\rangle=L/2$ in the box. The only numbers claimed are exact geometric factors ( $1/12$ , $1/24$ , $\ell/3$ ) and the elementary expectation $L/2$ .

Prerequisites. Two-Point Boundary-Value Problems · Linear Second-Order ODEs · The Wronskian and Linear Independence · Variation of Parameters. Leads to. Existence and Uniqueness Theorems for ODEs · The Method of Frobenius.

Ask the tutor: Want me to build the modified Green's function for your own singular operator — Neumann, periodic, or a resonant Sturm–Liouville case — and check its solvability step by step?

When exactly does an ordinary Green's function fail to exist?

Precisely when the operator $L$ , together with its boundary conditions, has a nonzero function $\psi_0$ that it annihilates — a zero mode. The defining equation $LG=\delta$ then contradicts itself: pairing both sides with $\psi_0$ gives zero on the left (by self-adjointness and $L\psi_0=0$ ) but $\psi_0(\xi)\neq0$ on the right. No kernel can satisfy an impossible equation. Common culprits are the Laplacian with pure Neumann conditions (constants are killed), any Sturm–Liouville operator sitting exactly at one of its eigenvalues, and $H_0-E_0$ in quantum perturbation theory. Whenever a symmetry or a resonance gives you a free, undamped mode, expect the ordinary Green's function to break and the modified one to be required.

What is the difference between the modified Green's function and the ordinary one?

The ordinary Green's function inverts $L$ completely: $LG=\delta$ . The modified one inverts $L$ everywhere except on the kernel: $LG_{\mathrm M}=\delta-P_0$ , where $P_0$ projects onto the zero modes. In the spectral picture the difference is a single omission — the ordinary function is $\sum_n\psi_n\psi_n/\lambda_n$ over all modes, which diverges if any $\lambda_n=0$ , while the modified function is the same sum with the zero modes left out. Equivalently, $G_{\mathrm M}$ is the Moore–Penrose pseudoinverse of $L$ rather than a true inverse. When there are no zero modes the two coincide exactly, so the modified Green's function is the strictly more general object.

Why do we need a second condition, orthogonality of G_M to the kernel?

Because the defining equation $LG_{\mathrm M}=\delta-P_0$ does not pin $G_{\mathrm M}$ down completely. If $G_{\mathrm M}$ is one solution, so is $G_{\mathrm M}+c(\xi)\psi_0(x)$ for any function $c$ , since $L$ kills the extra piece. That leftover freedom is exactly the zero-mode ambiguity, and it is harmless physically but fatal for uniqueness. Demanding $\langle\psi_0,G_{\mathrm M}\rangle=0$ — that $G_{\mathrm M}$ have no component along the kernel in its first argument — removes the ambiguity, singles out one symmetric kernel, and mirrors the linear-algebra requirement that a pseudoinverse annihilate the null space. Without it, "the" modified Green's function is not well defined.

How is this the same as the Moore–Penrose pseudoinverse of a matrix?

It is literally the continuous version. For a Hermitian matrix $A$ with a zero eigenvalue, the pseudoinverse $A^{+}$ inverts $A$ on the span of the nonzero eigenvectors and sends the null eigenvector to zero, so that $AA^{+}=\mathbb 1-P_{\ker}$ . Replace the eigenvectors $\mathbf v_n$ by eigenfunctions $\psi_n(x)$ , the eigenvalues by $\lambda_n$ , and the matrix sum by an integral kernel, and $A^{+}=\sum_{\lambda_n\neq0}\mathbf v_n\mathbf v_n^{\dagger}/\lambda_n$ becomes $G_{\mathrm M}=\sum_{\lambda_n\neq0}\psi_n(x)\psi_n(\xi)/\lambda_n$ . The four Penrose conditions translate one to one: $LG_{\mathrm M}=G_{\mathrm M}L=\mathbb 1-P_0$ , $G_{\mathrm M}\psi_0=0$ , and $G_{\mathrm M}^{\dagger}=G_{\mathrm M}$ . Every intuition about singular matrices carries straight over.

What happens physically if the source violates the solvability condition?

There is simply no steady or static solution. The component of the forcing that lies along the zero mode has nothing to push against — the zero mode is, by definition, a direction with no restoring tendency — so instead of settling into equilibrium, the corresponding collective coordinate accelerates or accumulates without bound. In a diffusion problem with insulating walls, a net source means material piles up forever and no steady concentration exists. In perturbation theory, an uncanceled resonant term produces secular growth $\sim t$ that wrecks the expansion. The modified Green's function returns only the transverse response, the part orthogonal to the kernel; the longitudinal part genuinely has no equilibrium, and the honest answer is that the steady problem is unsolvable until the offending source component is removed.

Does the modified Green's function keep the symmetry of the ordinary one?

Yes. For a self-adjoint operator, $G_{\mathrm M}(x,\xi)=G_{\mathrm M}(\xi,x)$ , just like the ordinary Green's function. In the spectral form this is obvious — the numerator $\psi_n(x)\psi_n(\xi)$ is symmetric under swapping the two arguments, and dividing by the real eigenvalue $\lambda_n$ does not spoil it. In the defining-equation picture, symmetry follows because both the delta and the subtracted projector $P_0$ are symmetric kernels and $L$ is self-adjoint. Symmetry is not automatic if you botch the gauge condition, though: choosing an asymmetric additive term would break it. The orthogonality condition (4) is precisely what selects the symmetric representative out of the family of solutions.