Fundamental Solutions and Green Operators

A Green operator is an inverse only after its input space, output domain, and boundary, initial, support, or asymptotic conditions have been fixed. A fundamental solution is the free-space distributional seed $P E=\delta_0$ for a constant-coefficient operator. Convolution with that seed can produce a particular solution, while a Green kernel incorporates the extra conditions that select one solution from all particular-plus-homogeneous possibilities.

Required background. Test-Function Spaces, Distributions, Support, and Convergence supplies distributional differentiation, support, and convergence.

Helpful background. Symbols, Characteristics, and PDE Type supplies principal-symbol reasoning.

Fundamental solutions, kernels, and Green operators

Let

$$P_B: \mathcal D_B(P)\subset X \longrightarrow Y_0\subseteq Y$$

be a linear differential operator with a domain $\mathcal D_B(P)$ encoding the homogeneous condition denoted by $B$, after any inhomogeneous data have been lifted. The codomain $Y_0$ is the selected space of compatible data; it may be a proper subspace of a larger ambient space $Y$.

Object	Defining relation	What selects it
Fundamental solution $E$	$P E=\delta_0$ on free space	A distribution class and, when needed, support or asymptotic behavior
Green operator $G_B$	$P_B G_B f=f$ for $f\in Y_0$	Spaces, operator domain, and boundary, initial, support, or asymptotic conditions
Distributional Green kernel $K_B(x,y)$	$P_xK_B(x,y)=\delta^{(d)}(x-y)$ in a Lebesgue coordinate chart	The same data as the Green operator, expressed in kernel form

The subscript $B$ is a reminder that the differential expression alone does not determine the inverse. Terminology varies: many sources call $E$ a “free-space Green function.” This page reserves fundamental solution for the free-space point-source distribution whose associated kernel $E(x-y)$ is translation invariant, and Green kernel for the kernel of a selected operator inverse.

Fundamental solutions are distributional

For a nonzero constant-coefficient operator

$$P(\partial) = \sum_{|\alpha|\leq m} c_\alpha\partial^\alpha \quad\text{on }\mathbb R^d,$$

a fundamental solution is a distribution $E\in\mathcal D'(\mathbb R^d)$ satisfying

$$P E=\delta_0.$$

Equivalently, with $P^{\mathsf t}$ the formal transpose,

$$\langle E,P^{\mathsf t}\varphi\rangle = \varphi(0) \qquad \text{for every }\varphi\in C_c^\infty(\mathbb R^d).$$

This test-function identity includes the normalization at the singular point. Checking only $P E=0$ away from the origin is insufficient.

The term is unrelated to the fundamental matrix of a homogeneous first-order ODE system, developed in Linear ODEs, Evolution Operators, and Wronskians. Dyatlov, Dyatlov 2022, Definition 9.2 and Remark 9.3, printed p. 98, PDF, gives the distributional definition and its transpose formulation.

Green operators are right inverses on declared data

A right Green operator is a linear map

$$G_B:Y_0\longrightarrow\mathcal D_B(P)$$

such that

$$P_B G_B = I_{Y_0}.$$

This relation states existence for every datum in $Y_0$. It does not yet say that

$$G_B P_B=I_{\mathcal D_B(P)}.$$

The second identity follows when the selected realization $P_B$ is also injective. If $0\neq h\in\ker P_B$, then $G_BP_Bh=0\neq h$, so no right inverse can also be a left inverse on the full domain. Likewise, if $P_B$ is not onto $Y$, no right inverse can be defined on all of $Y$.

When $G_B:C_c^\infty(V)\to\mathcal D'(U)$ is sequentially continuous, the Schwartz kernel theorem gives a unique distribution $K_B(x,y)\in\mathcal D'(U\times V)$ such that

$$\langle G_B f,\psi\rangle = \langle K_B(x,y),\, \psi(x)f(y) \rangle.$$

When the kernel and data are ordinary functions in Euclidean coordinates, this becomes

$$(G_Bf)(x) = \int_V K_B(x,y)f(y)\,\mathrm dy.$$

The distributional statement, including the Lebesgue identity kernel $\delta(x-y)$, is Theorem 7.6 and Proposition 7.7 of Dyatlov 2022, §§7.2.1–7.2.2, printed pp. 80–82, PDF. The manifold-level formulation continues in Distributional Kernels and Distributions on Manifolds.

Weights must not be hidden. If $\mathrm d\mu(y)=\rho(y)\,\mathrm dy$ with smooth positive $\rho$, and $G_B^{(\mu)}$ denotes the function kernel used with $\mathrm d\mu$, then

$$K_B(x,y) = \rho(y)G_B^{(\mu)}(x,y), \qquad \delta_\mu(x,y) = \frac{\delta^{(d)}(x-y)}{\rho(y)}.$$

Consequently $P_xG_B^{(\mu)}=\delta_\mu$, whereas $P_xK_B=\delta^{(d)}$. The free-space and interval examples below use Lebesgue measure; the spectral formula states its measure explicitly.

Free space: a fundamental solution produces a particular solution

Let $E$ be a fundamental solution of a constant-coefficient $P$, and first take $f\in C_c^\infty(\mathbb R^d)$. The convolution

$$u(x) = (E*f)(x) = \left\langle E(y),f(x-y) \right\rangle$$

is well defined and smooth. Constant coefficients let $P$ commute through convolution:

$$\begin{aligned} P(E*f) &= (P E)*f\\ &= \delta_0*f\\ &= f. \end{aligned}$$

Thus $E*f$ is one particular solution of $Pu=f$. More generally, convolution of two distributions is available when their supports sum properly; compact support of one factor is a sufficient condition. It is not defined for two arbitrary distributions. Dyatlov, Dyatlov 2022, Theorem 9.4 and Remark 9.5, printed pp. 98–99, PDF, states the precise support condition.

Fundamental solutions need not be unique. If $P E_1=P E_2=\delta_0$, then

$$P(E_1-E_2)=0.$$

Conversely, adding any homogeneous distribution $h$ with $Ph=0$ gives another fundamental solution. The Malgrange–Ehrenpreis theorem guarantees at least one distributional fundamental solution for every nonzero constant-coefficient operator, but it does not select boundary, support, decay, or radiation data; see Dyatlov 2022, Theorem 9.13, printed pp. 101–102, PDF.

Fourier construction and its stop condition

For a Schwartz function $f$, the site’s Fourier pair is

$$\begin{aligned} \widetilde f(p) &= \int\mathrm d^d x\, e^{+ip\cdot x}f(x), \\ f(x) &= \int \frac{\mathrm d^d p}{(2\pi)^d}\, e^{-ip\cdot x}\widetilde f(p). \end{aligned}$$

The transform extends to tempered distributions by duality, rather than by a pointwise oscillatory integral. For a tempered fundamental solution this gives

$$\mathcal F(\partial^\alpha E) = (-ip)^\alpha\widetilde E.$$

Since $\mathcal F\delta_0=1$, the equation $P E=\delta_0$ becomes

$$P(-ip)\widetilde E(p)=1.$$

The tempting expression

$$\widetilde E(p) = \frac{1}{P(-ip)}$$

is valid as an ordinary multiplier only when the reciprocal is actually defined and has the needed mapping properties. If $P(-ip)$ vanishes, division must be solved in a distribution space and extra support, boundary-value, or asymptotic conditions can select inequivalent answers. The formal quotient is then a question, not a definition.

Modified Helmholtz kernel in three dimensions

For

$$P_m=-\Delta+m^2, \qquad m>0,$$

the Fourier multiplier has no real zero:

$$\mathcal F(P_mE_m) = \left( |\mathbf p|^2+m^2 \right) \widetilde E_m(\mathbf p).$$

The unique tempered inverse is therefore

$$\widetilde E_m(\mathbf p) = \frac{1}{|\mathbf p|^2+m^2},$$

and in $\mathbb R^3$

$$\boxed{ E_m(\mathbf x) = \frac{e^{-m|\mathbf x|}}{4\pi|\mathbf x|}. }$$

Direct differentiation gives $P_mE_m=0$ away from the origin. The delta normalization is fixed by the small-sphere flux. Writing $r=|\mathbf x|$,

$$-\lim_{\varepsilon\downarrow0} \int_{r=\varepsilon} \partial_r E_m\,\mathrm dS = \lim_{\varepsilon\downarrow0} e^{-m\varepsilon} \left( 1+m\varepsilon \right) = 1.$$

The $m^2E_m$ integral over the shrinking ball tends to zero, so

$$\left( -\Delta+m^2 \right)E_m = \delta_0$$

as a distribution. The limit $m\downarrow0$ gives the Coulomb kernel $1/(4\pi r)$. Hunter, Hunter 2014, §§2.6–2.7, printed pp. 33–35, PDF, derives the unit-flux normalization and convolution solution for $-\Delta$; Appendix 5.D.2 identifies the inverse Fourier multiplier of $I-\Delta$ as the exponentially decaying Bessel potential.

Boundary data choose a different kernel

Free-space translation invariance gives kernels of the form $E(x-y)$. Boundaries generally destroy that form. Consider

$$P_D = -\frac{\mathrm d^2}{\mathrm dx^2} \quad\text{on }(0,L), \qquad u(0)=u(L)=0.$$

Define

$$G_D(x,y) = \frac{x_{<}\left(L-x_{>}\right)}{L}, \qquad x_{<}=\min(x,y), \quad x_{>}=\max(x,y).$$

For fixed $y$, this function is linear on each side of $x=y$, vanishes at $x=0,L$, and is continuous at $x=y$. Its first derivative has jump

$$\partial_xG_D(y^+,y) - \partial_xG_D(y^-,y) = -1.$$

Therefore

$$-\partial_x^2G_D(x,y) = \delta(x-y),$$

and

$$u(x) = \int_0^L G_D(x,y)f(y)\,\mathrm dy$$

solves $-u''=f$ with the two Dirichlet conditions.

The free-space fundamental solution for $-\partial_x^2$ is $E_0(x-y)=-|x-y|/2$. The boundary kernel can be written

$$G_D(x,y) = E_0(x-y) + \frac{x+y}{2} - \frac{xy}{L}.$$

For each fixed $y$, the last two terms are homogeneous solutions in $x$. They are exactly the correction needed to impose the endpoint data. This is the general pattern: the inverse equation fixes the singular part, while boundary or support conditions fix a homogeneous ambiguity.

Zero modes and spectral Green operators

For a self-adjoint realization on $L^2(\mathrm d\mu)$ with a discrete orthonormal eigenbasis $\{\phi_n\}$ and $0\notin\sigma(P_B)$, the inverse has the following formal kernel relative to $\mathrm d\mu$:

$$G_B(x,y) = \sum_n \frac{ \phi_n(x)\overline{\phi_n(y)} }{ \lambda_n },$$

with convergence understood in the appropriate operator or distribution topology. The boundary conditions enter through both $\phi_n$ and $\lambda_n$. This construction continues the mode-expansion discussion in Sturm–Liouville Problems and Eigenfunction Expansions.

If zero modes are present, the displayed inverse does not exist on all data. Let $\Pi_0$ project onto $\ker P_B$. A reduced Green operator instead obeys

$$P_B G_\perp = I-\Pi_0$$

on compatible data, together with a normalization such as $G_\perp f\perp\ker P_B$. For the massless Neumann Laplacian on a connected bounded domain, constants are zero modes, the source must annihilate the constant function, and the solution is unique only modulo a constant. The function-space version is discussed in Weak Solutions, Sobolev Spaces, and Well-Posedness.

First QFT application: free scalar response to a source

Continue the static quadratic scalar problem from the weak-solution page, now on $\mathbb R^3$ with $m>0$ and $J\in\mathcal S(\mathbb R^3)$. The equation

$$\left( -\Delta+m^2 \right)\phi = J$$

has the decaying solution

$$\boxed{ \phi(\mathbf x) = \int_{\mathbb R^3} \frac{ e^{-m|\mathbf x-\mathbf y|} }{ 4\pi|\mathbf x-\mathbf y| } J(\mathbf y)\, \mathrm d^3y. }$$

Equivalently,

$$\widetilde\phi(\mathbf p) = \frac{ \widetilde J(\mathbf p) }{ |\mathbf p|^2+m^2 }.$$

The same construction extends distributionally to the point source $J=q\delta_0$, whose response is the Yukawa profile $q e^{-mr}/(4\pi r)$. This is a classical static, or Euclidean, Green problem. It is not by itself a quantum two-point function or a construction of a path-integral measure.

Tong, Tong 2006, §§2.7.1–2.7.2, Eqs. (2.174)–(2.177), relates the Klein–Gordon inverse equation to a scalar field driven by a source. In Lorentzian signature, the characteristic mass shell makes the inverse equation alone nonunique. In the site’s normalization,

$$\widetilde\Delta_F(p) = \frac{i}{p^2-m^2+i0}, \qquad (\Box+m^2)\Delta_F = -i\delta^{(d)}.$$

Thus the mathematically delta-normalized Feynman fundamental solution would be $G_F=i\Delta_F$, not $\Delta_F$ itself. This page does not use that choice to define causal response. Support and pole conditions belong to the hyperbolic methods treatment, Hyperbolic Equations and Causal Propagators; the physical taxonomy of ordered correlators and source response belongs to Scalar Propagators, Ordered Correlators, and Sources. Heat evolution and short-time kernels are assigned to Elliptic Boundary Problems and Heat Kernels.

Common pitfalls

Calling the differential expression invertible. An inverse belongs to a map between spaces with a declared domain. Boundary conditions and zero modes can change injectivity and surjectivity without changing the local formula for $P$.

Checking only away from the source. The equation $P E=0$ on $\mathbb R^d\setminus\{0\}$ misses the delta normalization. Verify the test-function identity, derivative jump, or small-sphere flux.

Dividing by a symbol at its zeros. The notation $1/P(-ip)$ is not an ordinary function on the characteristic set. Specify the distribution class and the condition that selects a boundary value.

Convolving arbitrary distributions. A formal convolution can be undefined even when both factors are tempered. Check compactness, proper support, or another convolution theorem.

Confusing a Green kernel with a quantum correlator. The inverse equation is only one ingredient. State choice, operator ordering, support, and pole prescriptions belong to the physical and causal treatments linked above.

Exercises

1. Prove that a right inverse of $P_B$ cannot also be a left inverse on the full domain when $\ker P_B\neq\{0\}$.

   Check

   Choose $0\neq h\in\ker P_B$. For every right inverse $G_B$,

   $$G_BP_Bh = G_B0 = 0,$$

   whereas the left-inverse identity would require $G_BP_Bh=h\neq0$. Thus injectivity is necessary for a two-sided inverse.

2. Verify the Dirichlet interval kernel, including its delta normalization.

   Check

   For $x<y$, $G_D=x(L-y)/L$; for $x>y$, $G_D=y(L-x)/L$. Both pieces vanish at the relevant endpoint and agree at $x=y$. Their derivatives are $(L-y)/L$ and $-y/L$, so the jump is $-1$. The distributional second derivative is therefore $-\delta(x-y)$, giving $-\partial_x^2G_D=\delta(x-y)$.

3. Check the Fourier sign in the modified Helmholtz example.

   Check

   With the forward phase $e^{+i\mathbf p\cdot\mathbf x}$, $\partial_j\mapsto-i p_j$, so

   $$-\Delta+m^2 \longmapsto |\mathbf p|^2+m^2.$$

   Multiplication by $(|\mathbf p|^2+m^2)^{-1}$ gives $\widetilde\phi=\widetilde J/(|\mathbf p|^2+m^2)$ with no additional sign or $(2\pi)^3$ factor; that factor appears only in the inverse-transform measure.

4. If $E$ is a fundamental solution and $h$ satisfies $Ph=0$, show that $(E+h)*f$ solves the same inhomogeneous equation whenever the convolutions exist.

   Check

   Linearity and commutation with convolution give

   $$P\bigl((E+h)*f\bigr) = (PE)*f+(Ph)*f = \delta_0*f+0 = f.$$

   The added term $h*f$ is a homogeneous solution. Extra conditions decide whether it is allowed.

References

• Semyon Dyatlov (2022), Lecture Notes for 18.155: Distributions, Elliptic Regularity, and Applications to PDEs, PDF, §7.2 and §9.1, especially Theorems 7.6, 9.4, and 9.13. This is the structural source for distributional kernels, fundamental solutions, convolution under proper support, and the constant-coefficient existence theorem.
• John K. Hunter (2014), Notes on Partial Differential Equations, PDF, §§2.5–2.7 and Appendix 5.D. This is the teaching source for Green’s identities, the Laplace fundamental solution, flux normalization, convolution solutions, and Bessel potentials.
• David Tong (2006), Quantum Field Theory, §2.7.2. This supplies the QFT-facing source equation and the handoff from inverse kernels to physically selected scalar propagators.