{
  "id": "greens-function",
  "title": "Green's Function",
  "lang": "en",
  "source_lang": "en",
  "summary": "A Green's function represents the response of a linear differential operator to a point source, enabling solutions to be constructed by integration.",
  "category": "mathematics",
  "subcategory": "analysis",
  "type": "concept",
  "aliases": [
    "Green function"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/greens-function/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/greens-function/",
      "json": "https://aiwiki.page/zh/greens-function.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A Green's function is a kernel that describes the response of a linear differential equation to a source concentrated at a single point in space, time, or both. Once this response is known, the solution for a distributed source can often be obtained by integrating the Green's function against that source. It is therefore a continuous analogue of the entries of an inverse matrix, with boundary conditions, initial conditions, and other restrictions incorporated into its definition.\n\nGreen's functions connect differential equations with integral representations, spectral methods, and physical response theory. They are used to calculate electrostatic potentials, heat diffusion, wave propagation, quantum amplitudes, and the expected occupation of states in stochastic systems. A Green's function need not be an ordinary function: singular sources and propagating wavefronts frequently require the language of distributions. Its existence and uniqueness depend on the operator and the conditions imposed, not merely on the differential expression appearing in the equation."
    },
    {
      "heading": "Origins and the inverse-operator viewpoint",
      "text": "The method takes its name from George Green, whose privately published An Essay on the Application of Mathematical Analysis to the Theories of Electricity and Magnetism appeared in 1828. Green developed integral identities and potential-theoretic techniques for determining fields from sources and boundary data. These ideas subsequently became central to mathematical physics. Their modern formulation extends beyond the electrostatic problems that motivated Green to general linear operators on function spaces.\n\nConsider a linear problem\n\n\\[L u(x)=f(x),\\qquad x\\in\\Omega,\\]\n\ntogether with homogeneous boundary conditions \\(Bu=0\\). Here \\(L\\) may be an ordinary or partial differential operator, and \\(B\\) specifies the boundary restrictions. A Green's function \\(G(x,y)\\) satisfies, formally,\n\n\\[L_xG(x,y)=\\delta(x-y),\\qquad B_xG(x,y)=0,\\]\n\nwhere the subscript indicates the variable on which an operator acts. The symbol \\(\\delta\\) denotes the Dirac delta distribution. The point \\(y\\) is the source location, while \\(x\\) is the observation location.\n\nThe corresponding solution is\n\n\\[u(x)=\\int_\\Omega G(x,y)f(y)\\,dy.\\]\n\nIndeed, applying \\(L_x\\) under the integral gives \\(Lu(x)=f(x)\\), provided the operations are justified in the relevant classical or generalized sense. The formula expresses superposition: each infinitesimal piece of the source contributes its own point-source response. For vector-valued equations, \\(G\\) becomes a matrix-valued kernel that also records coupling between components.\n\nThe matrix analogy makes the roles of the two variables explicit. If a nonsingular matrix \\(A\\) defines \\(Au=f\\), then\n\n\\[u_i=\\sum_j(A^{-1})_{ij}f_j.\\]\n\nIntegration replaces summation, and \\(G(x,y)\\) replaces \\((A^{-1})_{ij}\\). In an infinite-dimensional setting, however, an inverse may be unbounded, may exist only between specified function spaces, or may not possess an ordinary function as its kernel.\n\nA Green's function consequently belongs to a complete operator problem rather than to a differential formula alone. The expression \\(-d^2/dx^2\\), for example, produces different kernels on the real line, on a finite interval with fixed endpoint values, and on a periodic interval. Some choices admit an inverse; others possess a nontrivial null space and require an additional normalization or a compatibility condition. This dependence on the operator domain is fundamental in functional analysis."
    },
    {
      "heading": "Point sources, singularities, and fundamental solutions",
      "text": "The Dirac delta is characterized by\n\n\\[\\int \\delta(x-y)\\varphi(x)\\,dx=\\varphi(y)\\]\n\nfor suitable smooth test functions \\(\\varphi\\). It is not an ordinary function with an infinite value at one point. Rather, it is a linear functional describing evaluation. The equation \\(L_xG=\\delta_y\\) therefore means that derivatives of \\(G\\) are interpreted distributionally when its singularity prevents an ordinary pointwise interpretation.\n\nA fundamental solution is commonly a distribution solving \\(LE=\\delta\\) in the whole space, usually for an operator with constant coefficients. The term “Green's function” often emphasizes the additional boundary, initial, or radiation conditions of a particular problem. Usage varies, and whole-space fundamental solutions are also frequently called Green's functions. The distinction matters because a fundamental solution alone generally does not satisfy the boundary conditions of a bounded domain.\n\nFor the negative Laplacian in three-dimensional space,\n\n\\[-\\Delta\\Phi(x)=\\delta(x), \\qquad \\Phi(x)=\\frac{1}{4\\pi|x|}.\\]\n\nAway from the origin, \\(\\Delta\\Phi=0\\). Its delta source is detected by the flux through a small sphere:\n\n\\[-\\int_{\\partial B_\\varepsilon} \\frac{\\partial\\Phi}{\\partial n}\\,dS=1.\\]\n\nThe divergence theorem thus identifies the normalization of the singularity. For the convention \\(+\\Delta G=\\delta\\), the corresponding kernel has the opposite sign.\n\nIn dimension \\(n\\geq3\\), the whole-space kernel of \\(-\\Delta\\) is\n\n\\[\\Phi_n(x)= \\frac{|x|^{2-n}}{(n-2)|S^{n-1}|},\\]\n\nwhere \\(|S^{n-1}|\\) is the surface measure of the unit sphere. In two dimensions it is\n\n\\[\\Phi_2(x)=-\\frac{1}{2\\pi}\\log|x|,\\]\n\nand in one dimension a fundamental solution is \\(-|x|/2\\). These different forms explain why planar potential problems have logarithmic behavior rather than the inverse-distance behavior familiar in three dimensions.\n\nSingularity does not necessarily prevent integration. The three-dimensional kernel \\(1/|x-y|\\), for example, is locally integrable because volume near the source supplies a factor proportional to \\(r^2\\,dr\\). Its derivatives are more singular and may require principal-value interpretations or distributional boundary terms.\n\nFor a smooth domain, an interior Green's function for \\(-\\Delta\\) can be written\n\n\\[G(x,y)=\\Phi_n(x-y)-H(x,y),\\]\n\nwhere \\(H\\), as a function of \\(x\\), is a harmonic function chosen to enforce the boundary condition. The fundamental singularity describes the local source response; the harmonic correction encodes the global geometry. For uniformly elliptic operators with variable coefficients, local singular behavior similarly reflects the principal differential part, although explicit formulas are much less common."
    },
    {
      "heading": "Boundary conditions, reciprocity, and solvability",
      "text": "A boundary-value problem specifies not only an equation inside a domain but also how its solution behaves at the boundary. Dirichlet conditions prescribe the value of the solution, while Neumann conditions prescribe a normal derivative or flux. Robin conditions combine these quantities. Changing the boundary condition changes the Green's function even when the interior differential operator remains identical.\n\nFor a smooth bounded domain and the Dirichlet problem\n\n\\[-\\Delta u=f,\\qquad u|_{\\partial\\Omega}=g,\\]\n\nlet \\(G\\) satisfy \\(-\\Delta_yG(x,y)=\\delta(x-y)\\) and vanish when \\(y\\in\\partial\\Omega\\). Green's second identity, an extension of integration by parts, yields\n\n\\[u(x)= \\int_\\Omega G(x,y)f(y)\\,dy -\\int_{\\partial\\Omega} g(y)\\frac{\\partial G(x,y)}{\\partial n_y}\\,dS_y.\\]\n\nThus the volume source and the boundary values contribute through different kernels. The quantity \\(-\\partial_{n_y}G(x,y)\\) is the Poisson kernel when the usual smoothness assumptions apply.\n\nFor the upper half-space \\(x_n>0\\), the Dirichlet Green's function for \\(-\\Delta\\) in dimensions \\(n\\geq3\\) follows from the method of images:\n\n\\[G(x,y)=\\Phi_n(x-y)-\\Phi_n(x-y^*),\\]\n\nwhere \\(y^*\\) is the reflection of \\(y\\) across \\(x_n=0\\). Both distances coincide on the boundary, so the terms cancel there. The reflected source lies outside the physical domain and supplies a harmonic correction inside it.\n\nSymmetry follows from the operator rather than from the point-source interpretation alone. For a real, suitably self-adjoint problem,\n\n\\[G(x,y)=G(y,x).\\]\n\nFor a complex self-adjoint operator, the corresponding identity is Hermitian:\n\n\\[G(x,y)=\\overline{G(y,x)}.\\]\n\nNon-self-adjoint equations instead relate a Green's kernel to the kernel of the adjoint problem. Advection and absorbing boundary conditions can therefore alter the reciprocity properties familiar from static potential theory.\n\nPure Neumann problems illustrate the obstruction caused by zero modes. On a connected bounded domain, constants satisfy \\(-\\Delta u=0\\) and \\(\\partial_nu=0\\). For homogeneous Neumann data, a solution to \\(-\\Delta u=f\\) requires\n\n\\[\\int_\\Omega f\\,dx=0.\\]\n\nA normalized Neumann kernel can be defined by\n\n\\[-\\Delta_xG_N(x,y) =\\delta(x-y)-\\frac{1}{|\\Omega|}, \\qquad \\partial_nG_N=0, \\qquad \\int_\\Omega G_N(x,y)\\,dx=0.\\]\n\nIt inverts the operator on zero-mean functions rather than on the entire function space. For nonzero prescribed normal derivative \\(h\\), the compatibility condition becomes \\(\\int_\\Omega f\\,dx=-\\int_{\\partial\\Omega}h\\,dS\\)."
    },
    {
      "heading": "Construction for ordinary differential equations",
      "text": "For a second-order ordinary differential equation, a Green's function can often be constructed directly from homogeneous solutions. Consider a regular Sturm–Liouville operator on an interval \\([a,b]\\),\n\n\\[Lu=-\\frac{d}{dx}\\left(p(x)\\frac{du}{dx}\\right)+q(x)u, \\qquad p(x)>0,\\]\n\nwith separated homogeneous conditions at the endpoints. Away from \\(x=y\\), the kernel satisfies the homogeneous equation \\(L_xG=0\\).\n\nLet \\(u_L\\) satisfy the left endpoint condition and \\(u_R\\) satisfy the right endpoint condition. Their Wronskian is\n\n\\[W(x)=u_L(x)u_R'(x)-u_L'(x)u_R(x).\\]\n\nFor this operator, \\(p(x)W(x)\\) is constant. Provided it is nonzero, the Green's function is\n\n\\[G(x,y)= -\\frac{u_L(x_<)u_R(x_>)}{p(y)W(y)},\\]\n\nwhere \\(x_<=\\min(x,y)\\) and \\(x_>=\\max(x,y)\\). Each branch satisfies the appropriate endpoint condition automatically.\n\nThe matching conditions at the source determine the normalization. The kernel is continuous at \\(x=y\\), while its flux derivative has the jump\n\n\\[p(y)\\left[ \\partial_xG(y^+,y)-\\partial_xG(y^-,y) \\right]=-1.\\]\n\nIntegrating \\(L_xG=\\delta(x-y)\\) across a small interval around \\(y\\) gives this relation. A jump in \\(G\\) itself would generally introduce a derivative of the delta distribution, which is absent from the specified source.\n\nFor the elementary problem\n\n\\[-u''=f,\\qquad u(0)=u(\\ell)=0,\\]\n\nthe homogeneous solutions \\(u_L=x\\) and \\(u_R=\\ell-x\\) give\n\n\\[G(x,y)= \\frac{\\min(x,y)\\,[\\ell-\\max(x,y)]}{\\ell}.\\]\n\nThis kernel is nonnegative and piecewise linear in either variable. Its first derivative has the required downward jump of one. Integrating it against a positive load produces the positive deflection profile associated with this sign convention for the operator.\n\nIf \\(pW=0\\), the left and right solutions are linearly dependent. A nonzero homogeneous solution then satisfies both endpoint conditions, so the inverse fails to exist without restricting the source and fixing the homogeneous ambiguity. For the shifted problem \\(L-\\lambda\\), such failures occur at eigenvalues.\n\nInitial-value equations have a different construction because their kernels are causal rather than symmetric. For\n\n\\[u'(t)+a(t)u(t)=f(t),\\qquad u(t_0)=0,\\]\n\nthe kernel is\n\n\\[G(t,s)=\\Theta(t-s) \\exp\\left[-\\int_s^t a(r)\\,dr\\right].\\]\n\nHere \\(\\Theta\\) is the step function. The solution integrates only over \\(t_0\\leq s\\leq t\\), showing directly how initial conditions select a different inverse from endpoint conditions."
    },
    {
      "heading": "Fourier methods, spectral expansions, and resolvents",
      "text": "Translation-invariant operators admit kernels depending only on the difference of their arguments:\n\n\\[G(x,y)=g(x-y).\\]\n\nThe solution then becomes a convolution, \\(u=g*f\\). A Fourier transform converts constant-coefficient differentiation into multiplication. With the convention\n\n\\[\\widehat f(k)=\\int_{\\mathbb R^n}e^{-ik\\cdot x}f(x)\\,dx,\\]\n\nan operator \\(L=P(\\partial)\\) has symbol \\(P(ik)\\), and its fundamental solution formally satisfies\n\n\\[\\widehat g(k)=\\frac{1}{P(ik)}.\\]\n\nThis formula is straightforward when the denominator stays away from zero, but zeros require additional interpretation. For \\(-\\Delta\\), the multiplier is \\(1/|k|^2\\), whose treatment at \\(k=0\\) depends on dimension and on the distributional formulation. For wave operators, the symbol vanishes on a characteristic cone. Different prescriptions for approaching those zeros produce different kernels, including retarded, advanced, and time-ordered versions. Algebraic division by the symbol does not by itself specify the physical solution.\n\nOn a bounded domain, spectral theory replaces Fourier waves with eigenfunctions adapted to the boundary. If \\(L\\) is positive and self-adjoint with discrete spectrum,\n\n\\[L\\phi_j=\\lambda_j\\phi_j,\\qquad \\lambda_j>0,\\]\n\nand the \\(\\phi_j\\) form an orthonormal basis in \\(L^2(\\Omega)\\), then formally\n\n\\[G(x,y)=\\sum_j \\frac{\\phi_j(x)\\overline{\\phi_j(y)}}{\\lambda_j}.\\]\n\nThe series may converge distributionally or in a suitable function space rather than uniformly near \\(x=y\\). Its diagonal singularity reflects the collective contribution of high-frequency eigenfunctions.\n\nThe interval kernel for \\(-d^2/dx^2\\) with zero endpoint values, for example, has the expansion\n\n\\[G(x,y)=\\frac{2\\ell}{\\pi^2} \\sum_{m=1}^{\\infty} \\frac{\\sin(m\\pi x/\\ell)\\sin(m\\pi y/\\ell)}{m^2}.\\]\n\nThis represents the same piecewise-linear kernel obtained by derivative matching. It also displays how inversion suppresses modes with large eigenvalues.\n\nA more general object is the resolvent\n\n\\[R(z)=(L-zI)^{-1},\\]\n\ndefined for \\(z\\) outside the spectrum. Its kernel is a parameter-dependent Green's function, with modal coefficients \\(1/(\\lambda_j-z)\\). Isolated eigenvalues appear as poles, whose residues encode spectral projections up to the sign set by the resolvent convention. In unbounded domains, continuous spectra replace pure eigenfunction sums with spectral integrals. Boundary values of the resolvent from different sides of the spectrum distinguish outgoing and incoming responses, making it central to scattering theory and resonance analysis."
    },
    {
      "heading": "Heat, wave propagation, and causal kernels",
      "text": "Time-dependent Green's functions describe how an impulse at one event influences later events. A retarded kernel vanishes before the source acts; an advanced kernel vanishes after it. This support condition is additional information, not a consequence of the differential equation alone. For evolution problems, it determines the direction in which initial data and forcing propagate.\n\nFor the heat equation\n\n\\[(\\partial_t-\\kappa\\Delta)u=f,\\qquad \\kappa>0,\\]\n\nthe retarded whole-space kernel is\n\n\\[G(x,t;y,s)= \\Theta(t-s) \\frac{\\exp[-|x-y|^2/(4\\kappa(t-s))]} {[4\\pi\\kappa(t-s)]^{n/2}}.\\]\n\nFor each \\(t>s\\), its spatial integral is one. As \\(t\\) decreases to \\(s\\), it approaches a spatial delta distribution. Its width grows proportionally to \\(\\sqrt{\\kappa(t-s)}\\), expressing the characteristic diffusion length.\n\nWith initial temperature \\(u_0\\) at \\(t=0\\), the solution consists of an initial-data term and a forcing term:\n\n\\[u(x,t)= \\int K_t(x,y)u_0(y)\\,dy+ \\int_0^t\\int K_{t-s}(x,y)f(y,s)\\,dy\\,ds,\\]\n\nwhere \\(K_t\\) denotes the spatial heat kernel. This is Duhamel's principle. The Gaussian is positive everywhere at every positive time, so the classical heat equation has instantaneous spatial propagation, despite rapid decay far from the source.\n\nThe wave equation has different support. For\n\n\\[(\\partial_t^2-c^2\\Delta)G =\\delta(t)\\delta(x)\\]\n\nin three spatial dimensions, the retarded kernel is\n\n\\[G_{\\mathrm{ret}}(x,t) =\\frac{\\delta(t-|x|/c)}{4\\pi c^2|x|}.\\]\n\nA point impulse reaches distance \\(r\\) at time \\(r/c\\). The kernel is concentrated on the forward light or sound cone, rather than throughout its interior. This sharp propagation is an instance of Huygens' principle.\n\nThe behavior depends on dimension and on lower-order terms. In two spatial dimensions, the wave kernel has a tail inside the forward cone, so a point impulse continues contributing after the first arrival. Boundaries can introduce reflected fronts and repeated arrivals, while variable coefficients bend propagation paths.\n\nHeat kernels also connect evolution to elliptic inversion. If \\(L\\) is positive, self-adjoint, and bounded below by a strictly positive constant, then\n\n\\[L^{-1}=\\int_0^\\infty e^{-tL}\\,dt.\\]\n\nConsequently, integrating the heat kernel over all positive times gives the Green's kernel of \\(L\\). If zero modes are present or the integral diverges at long times, projection or other modifications are necessary."
    },
    {
      "heading": "Physical response, scattering, and quantum theory",
      "text": "In electrostatics, Poisson's equation takes the form\n\n\\[-\\varepsilon_0\\Delta\\phi=\\rho.\\]\n\nIn unbounded three-dimensional space, the potential is\n\n\\[\\phi(x)=\\frac{1}{4\\pi\\varepsilon_0} \\int\\frac{\\rho(y)}{|x-y|}\\,dy.\\]\n\nConducting surfaces replace the free-space kernel with one satisfying the appropriate boundary values. The method of images therefore represents induced boundary effects without explicitly solving for every surface charge. Only geometries permitting the required reflected or transformed sources yield simple image formulas.\n\nIn time-harmonic acoustics and scalar wave scattering, a common convention is\n\n\\[(\\Delta+k^2)g_k(x)=-\\delta(x), \\qquad g_k(x)=\\frac{e^{ik|x|}}{4\\pi|x|}.\\]\n\nFor time dependence \\(e^{-i\\omega t}\\), this is the outgoing three-dimensional Helmholtz kernel. Its selection requires a radiation condition at infinity. Replacing \\(e^{ikr}\\) by \\(e^{-ikr}\\) gives the incoming response, while the differential equation away from the source remains unchanged.\n\nInhomogeneous media can be treated by separating a reference operator from a perturbation. If\n\n\\[(L_0+V)u=f\\]\n\nand \\(G_0=L_0^{-1}\\), then\n\n\\[u=G_0f-G_0Vu.\\]\n\nThis integral equation is the basis of multiple-scattering expansions and, in quantum scattering, the Lippmann–Schwinger formulation. Successive substitutions produce repeated interactions with \\(V\\). The first Born approximation retains only the first scattering contribution; its accuracy depends on the interaction strength and the relevant propagation scales.\n\nIn quantum mechanics, the energy-domain kernel\n\n\\[G(E)=(E-H+i0)^{-1}\\]\n\nis associated with the Hamiltonian \\(H\\). The limiting prescription indicates approach from the upper half of the complex energy plane and selects the retarded boundary value under standard conventions. Bound states appear as poles of the resolvent; continuous-spectrum boundary values describe propagating scattering states. With this convention, the spectral density is related to\n\n\\[-\\frac{1}{\\pi}\\operatorname{Im}G(E).\\]\n\nDiagonal kernels and traces require appropriate interpretation, particularly in continuum models where singularities may occur.\n\nMany-body theory also uses “Green's function” for time-ordered or retarded correlation functions of quantum field operators. These are not always literal inverses of a simple differential operator. For interacting systems, their equations involve higher-order correlations or an energy-dependent self-energy. The formal Dyson relation\n\n\\[G^{-1}=G_0^{-1}-\\Sigma\\]\n\nencodes the effect of interactions in the self-energy \\(\\Sigma\\). Distinguishing these correlation functions from classical point-source kernels is essential, even though both organize propagation and response."
    },
    {
      "heading": "Probability, occupation times, and covariance",
      "text": "Green's functions have a probabilistic interpretation as accumulated transition densities. Let a Markov diffusion have generator \\(A\\), and let it be killed when it exits a domain \\(\\Omega\\). If \\(p_t^\\Omega(x,y)\\) is its killed transition density, define\n\n\\[G(x,y)=\\int_0^\\infty p_t^\\Omega(x,y)\\,dt\\]\n\nwhenever the integral is meaningful. This kernel is the inverse of \\(-A\\), subject to the absorbing boundary condition. It measures how much time a process starting at \\(x\\) is expected to spend near \\(y\\) before termination.\n\nMore precisely, for suitable nonnegative \\(f\\),\n\n\\[\\mathbb E_x\\left[\\int_0^{\\tau_\\Omega}f(X_t)\\,dt\\right] =\\int_\\Omega G(x,y)f(y)\\,dy,\\]\n\nwhere \\(\\tau_\\Omega\\) is the exit time. The left side accumulates a running cost along a stochastic process, and the right side expresses it through the Green's kernel. Taking \\(f=1\\) gives the expected exit time, which solves a Poisson problem when the required regularity and integrability hold.\n\nFor standard Brownian motion, the generator is \\(A=\\tfrac12\\Delta\\). Its occupation kernel therefore solves\n\n\\[-\\tfrac12\\Delta G=\\delta,\\]\n\nrather than \\(-\\Delta G=\\delta\\). The factor of two is important when comparing probabilistic and potential-theoretic conventions. In the whole space, Brownian motion is recurrent in dimensions one and two, and the unmodified time integral diverges. In dimensions three and higher it is transient, and the occupation kernel is finite away from the source and proportional to \\(|x-y|^{2-n}\\).\n\nA discrete analogue appears for an absorbing Markov chain. If \\(Q\\) is the transition matrix restricted to transient states, then\n\n\\[N=(I-Q)^{-1}=\\sum_{m=0}^{\\infty}Q^m.\\]\n\nThe entry \\(N_{ij}\\) is the expected number of visits to state \\(j\\), starting from state \\(i\\), before absorption, including the initial visit when \\(i=j\\). This is a discrete Green's function. Graph Laplacians similarly yield kernels related to random walks and electrical resistance, although zero modes often require a pseudoinverse.\n\nPositive Green's operators also define covariances of Gaussian random fields. The Gaussian free field has covariance given by an inverse Laplacian with specified boundary conditions. In dimensions two and higher, the continuum field is generally distribution-valued rather than an ordinary random function. Its covariance is meaningful after integration against test functions, even when a pointwise diagonal variance diverges. This provides a direct connection between elliptic inversion, spatial dependence, and Gaussian process models."
    },
    {
      "heading": "Numerical evaluation and limits of the method",
      "text": "Explicit Green's functions exist for relatively few combinations of geometry, coefficients, and boundary conditions. Numerical methods therefore either approximate the kernel itself or solve the corresponding operator equation without forming it. These are distinct computational strategies: constructing a complete response kernel can be useful for many different sources, but it may be wasteful when only one solution is required.\n\nDiscretization by the finite difference method produces a matrix equation \\(Au=f\\). Applying a unit load at one grid point computes a column of \\(A^{-1}\\), the discrete analogue of a point-source response. Source normalization must account for cell volume when comparing that column with a continuum kernel. In the finite element method, source and observation quantities also depend on basis functions and mass matrices, so inverse-matrix entries are not automatically pointwise samples of the continuum Green's function.\n\nEven when \\(A\\) is a sparse matrix, its inverse is usually dense. Explicitly storing all point-to-point responses can therefore require quadratic memory. Sparse factorization, iterative solvers, or matrix-free methods often provide solutions more efficiently. Translation-invariant problems permit convolution to be accelerated using the fast Fourier transform, while hierarchical compression and fast multipole methods exploit the smoothness or low-rank structure of interactions between well-separated regions.\n\nBoundary element methods use a known fundamental solution to transform a homogeneous interior equation into an integral equation on the boundary. This reduces the geometric dimension by one and handles unbounded exterior domains naturally. The resulting matrices are generally dense, and the kernels require specialized quadrature near or on singularities. Corners, edges, closely spaced surfaces, and oscillatory high-frequency kernels introduce additional numerical difficulty.\n\nA Green's representation is not automatically stable. When a spectral parameter approaches an eigenvalue, factors such as \\(1/(\\lambda_j-z)\\) become large, amplifying source errors and numerical roundoff. Recovering a source from measured responses is an inverse problem and may be much less stable than computing the forward field. Smoothing by the Green's operator can suppress precisely the high-frequency information needed for inversion, motivating regularization.\n\nThe method also depends fundamentally on linearity. For a nonlinear equation \\(F(u)=f\\), no source-independent kernel generally reconstructs the solution by superposition. A Green's function for the linearized operator \\(DF[u_0]\\) instead describes perturbations about a chosen background \\(u_0\\). Even in linear problems, rough coefficients, noncompact domains, zero modes, and singular boundaries can require weak formulations in Sobolev spaces. In such settings, the appropriate inverse may exist as an operator between function spaces while lacking the simple pointwise kernel suggested by elementary examples."
    }
  ],
  "links": [
    "inverse-matrix",
    "differential-equation",
    "dirac-delta-distribution",
    "integral",
    "matrix",
    "null-space",
    "functional-analysis",
    "laplace-operator",
    "divergence-theorem",
    "harmonic-function",
    "boundary-value-problem",
    "dirichlet-boundary-condition",
    "neumann-boundary-condition",
    "integration-by-parts",
    "self-adjoint-operator",
    "ordinary-differential-equation",
    "eigenvalues-and-eigenvectors",
    "convolution",
    "fourier-transform",
    "spectral-theorem",
    "orthonormal-basis",
    "resolvent-operator",
    "heat-equation",
    "wave-equation",
    "poisson-equation",
    "quantum-mechanics",
    "stochastic-process",
    "wiener-process",
    "markov-chain",
    "covariance",
    "gaussian-process",
    "finite-difference-method",
    "finite-element-method",
    "sparse-matrix",
    "fast-fourier-transform",
    "inverse-problem",
    "regularization",
    "sobolev-space"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}