{
  "id": "maximum-principle",
  "title": "Maximum Principle",
  "lang": "en",
  "source_lang": "en",
  "summary": "A family of theorems controlling interior extrema of solutions to differential equations through boundary values, positivity, and comparison.",
  "category": "mathematics",
  "subcategory": "analysis",
  "type": "concept",
  "aliases": [
    "Maximum principles"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/maximum-principle/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/maximum-principle/",
      "json": "https://aiwiki.page/zh/maximum-principle.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The maximum principle is a family of results in mathematical analysis asserting that solutions of certain equations cannot develop unrestricted interior maxima or minima. In its simplest form, a harmonic function on a bounded region takes its largest and smallest values on the boundary, unless it is constant. More general versions apply to elliptic and parabolic partial differential equations, holomorphic functions, discrete networks, and nonlinear equations.\n\nMaximum principles turn local information about an equation into global information about its solutions. They establish uniqueness, preserve positivity, compare different solutions, and provide bounds without requiring an explicit formula. Their validity depends on the equation’s structure, the sign of certain coefficients, the regularity of the solution, and the geometry of the domain. Weak maximum principles locate extrema on a boundary; strong maximum principles explain when an interior extremum forces constancy; boundary point lemmas describe the derivative at a boundary extremum. A different theorem with the same name, the Pontryagin maximum principle, gives necessary conditions in optimal control."
    },
    {
      "heading": "Classical elliptic formulation",
      "text": "Let \\(\\Omega\\) be a bounded domain in \\(\\mathbb R^n\\), and suppose\n\n\\[u\\in C^2(\\Omega)\\cap C(\\overline{\\Omega}).\\]\n\nHere a domain is an open connected set, and \\(\\overline{\\Omega}\\) is its closure. The simplest weak maximum principle states that\n\n\\[\\Delta u\\geq 0\\quad\\text{in }\\Omega \\quad\\Longrightarrow\\quad \\max_{\\overline{\\Omega}}u=\\max_{\\partial\\Omega}u,\\]\n\nwhere \\(\\Delta\\) is the Laplace operator. Functions satisfying this differential inequality are called subharmonic. Reversing the inequality gives a minimum principle for superharmonic functions. A solution of Laplace’s equation, \\(\\Delta u=0\\), satisfies both principles.\n\nThe local calculation behind the theorem uses the Hessian matrix. At a smooth interior maximum, the gradient vanishes and the Hessian is negative semidefinite, so \\(\\Delta u\\leq0\\). This immediately excludes an interior maximum if \\(\\Delta u>0\\), but not if \\(\\Delta u=0\\). To handle the non-strict inequality, replace \\(u\\) by\n\n\\[u_\\varepsilon(x)=u(x)+\\varepsilon |x|^2.\\]\n\nSince \\(\\Delta u_\\varepsilon\\geq2n\\varepsilon>0\\), its maximum lies on the boundary. Boundedness of \\(\\Omega\\) permits passage to the limit as \\(\\varepsilon\\downarrow0\\), yielding the desired statement.\n\nA broader version concerns a linear second-order operator\n\n\\[Lu=\\sum_{i,j=1}^{n}a_{ij}(x)D_{ij}u +\\sum_{i=1}^{n}b_i(x)D_i u+c(x)u.\\]\n\nThe symmetric coefficient matrix \\(A=(a_{ij})\\) is uniformly elliptic if constants \\(0<\\lambda\\leq\\Lambda\\) satisfy\n\n\\[\\lambda|\\xi|^2\\leq \\xi^{T}A(x)\\xi\\leq\\Lambda|\\xi|^2\\]\n\nfor every \\(x\\) and vector \\(\\xi\\). Thus its eigenvalues remain positive and uniformly separated from zero. Under standard boundedness assumptions on the coefficients, if \\(c\\leq0\\) and \\(Lu\\geq0\\), then\n\n\\[\\sup_\\Omega u\\leq\\max\\!\\left\\{0,\\sup_{\\partial\\Omega}u\\right\\}.\\]\n\nWhen \\(c=0\\), the zero on the right is unnecessary.\n\nThe qualification involving zero is important. A nonpositive zeroth-order coefficient controls positive maxima, but it does not generally force a negative maximum to occur on the boundary. The maximum principle is therefore an order property tied to a particular sign convention, not merely an application of the extreme value theorem. Continuity on a compact closure guarantees that an extremum exists; the differential equation determines where it can occur."
    },
    {
      "heading": "Strong maximum principles and boundary point lemmas",
      "text": "The strong maximum principle strengthens a boundary bound into a rigidity statement. For a uniformly elliptic operator with \\(c=0\\), if \\(Lu\\geq0\\) on a connected domain and \\(u\\) attains its maximum at an interior point, then \\(u\\) is constant. If \\(c\\leq0\\), a standard formulation requires the attained maximum to be nonnegative. The weak principle allows an interior maximum equal to the boundary maximum; the strong principle says that this can happen only in the exceptional constant case.\n\nConnectedness is essential because the conclusion propagates through the domain. A function can take different constant values on different connected components without violating the local differential equation. Uniform ellipticity supplies the mechanism for propagation: the equation controls curvature in every spatial direction, preventing a nonconstant subsolution from possessing an interior plateau at its greatest value. Proofs typically use auxiliary functions, comparison on balls or annuli, and a geometric argument connecting regions where the solution is strictly below its maximum.\n\nThe closely related Hopf boundary point lemma concerns a boundary maximum. For example, suppose \\(\\Delta u\\geq0\\), \\(u(x)<u(x_0)\\) for \\(x\\in\\Omega\\), and \\(x_0\\in\\partial\\Omega\\). If the domain has an interior tangent ball at \\(x_0\\), and the relevant normal derivative exists, then\n\n\\[\\frac{\\partial u}{\\partial\\nu}(x_0)>0,\\]\n\nwhere \\(\\nu\\) is the outward unit normal. Thus a nonconstant subharmonic function approaches its boundary maximum with a strictly positive outward slope. At a boundary minimum of a superharmonic function, the inequality is reversed.\n\nA typical proof constructs an exponential barrier in an annulus, such as\n\n\\[v(x)=e^{-\\alpha|x-y|^2}-e^{-\\alpha R^2},\\]\n\nwhere the sphere of radius \\(R\\) centered at \\(y\\) touches the boundary at \\(x_0\\). For sufficiently large \\(\\alpha\\), the elliptic operator applied to \\(v\\) has a controlled sign near the touching point. Comparison between \\(u\\) and a suitable multiple of \\(v\\) then forces the strict derivative inequality.\n\nThe geometric hypothesis cannot simply be omitted. At cusps or other irregular boundary points, a solution may reach its boundary extremum with zero normal growth, or a normal direction may not be well defined. The strong principle and boundary point lemma therefore answer different questions: the former excludes interior extrema of nonconstant solutions, while the latter quantifies boundary behavior under additional geometric and differentiability assumptions. Together they are central tools in positivity arguments, boundary uniqueness, and symmetry proofs."
    },
    {
      "heading": "Comparison, uniqueness, and elliptic boundary problems",
      "text": "The maximum principle becomes especially useful when applied to the difference of two functions. Suppose \\(u,v\\in C^2(\\Omega)\\cap C(\\overline{\\Omega})\\), and \\(L\\) is a linear uniformly elliptic operator satisfying a weak maximum principle. If\n\n\\[Lu\\geq Lv\\quad\\text{in }\\Omega, \\qquad u\\leq v\\quad\\text{on }\\partial\\Omega,\\]\n\nthen \\(L(u-v)\\geq0\\), and consequently\n\n\\[u\\leq v\\quad\\text{throughout }\\Omega.\\]\n\nThis comparison principle orders entire solutions using their differential inequalities and boundary values.\n\nFor a Dirichlet boundary-value problem,\n\n\\[Lu=f\\quad\\text{in }\\Omega,\\qquad u=g\\quad\\text{on }\\partial\\Omega,\\]\n\ntwo solutions have a difference satisfying \\(L(u-v)=0\\) and zero boundary data. Comparison first gives \\(u\\leq v\\), and then \\(v\\leq u\\), establishing uniqueness. When \\(c=0\\), the same argument gives stability with respect to boundary data:\n\n\\[\\|u-v\\|_{L^\\infty(\\Omega)} \\leq \\|g-h\\|_{L^\\infty(\\partial\\Omega)}\\]\n\nfor solutions with the same forcing and boundary values \\(g,h\\). This estimate explains why errors in boundary measurements cannot be amplified in the supremum norm by harmonic extension.\n\nComparison also establishes signs for solutions of Poisson’s equation. Under the convention\n\n\\[-\\Delta u=f\\geq0,\\qquad u|_{\\partial\\Omega}=0,\\]\n\nthe minimum principle gives \\(u\\geq0\\). On a connected domain, if \\(f\\) is not identically zero and the solution is classical, the strong principle gives \\(u>0\\) in the interior. In a ball \\(B_R(0)\\), the explicit solution of \\(-\\Delta u=1\\) with zero boundary values is\n\n\\[u(x)=\\frac{R^2-|x|^2}{2n}.\\]\n\nThis quadratic function serves as a barrier for more complicated forcing terms and domains.\n\nBarriers are functions chosen to satisfy a convenient differential inequality while dominating the prescribed boundary values. They can bound solutions, prove continuity at boundary points, and support existence constructions. In Perron’s method, an appropriate envelope of subharmonic functions is used to construct a harmonic solution. Boundary regularity is a separate issue: uniqueness may follow from comparison even when the existence of a solution continuous up to every boundary point requires additional hypotheses.\n\nThe same positivity structure appears in Green’s functions. For the Dirichlet problem associated with \\(-\\Delta\\), the Green function is nonnegative, reflecting that nonnegative forcing produces a nonnegative solution. Consequently, increasing the forcing or boundary data increases the solution. These are order statements, not merely estimates of absolute magnitude, and they remain useful even when no explicit representation of the solution is available."
    },
    {
      "heading": "Parabolic equations and the direction of time",
      "text": "For the heat equation, the maximum principle incorporates the initial time as well as the spatial boundary. Let\n\n\\[Q_T=\\Omega\\times(0,T],\\]\n\nwith \\(\\Omega\\) bounded. Its parabolic boundary is\n\n\\[\\partial_pQ_T= (\\overline{\\Omega}\\times\\{0\\}) \\cup(\\partial\\Omega\\times[0,T]).\\]\n\nUnlike the ordinary topological boundary of a space–time cylinder, this set excludes the interior of the final-time face. Initial values and lateral boundary values determine the evolution; final values do not play the same role.\n\nIf \\(u\\) is continuous on the closed cylinder, sufficiently differentiable in its interior, and\n\n\\[u_t-\\kappa\\Delta u\\leq0,\\qquad \\kappa>0,\\]\n\nthen\n\n\\[\\sup_{\\overline{Q_T}}u\\leq\\sup_{\\partial_pQ_T}u.\\]\n\nOne proof perturbs the function to \\(u_\\varepsilon=u-\\varepsilon t\\). At a positive-time spatial maximum that exceeds all earlier boundary values, the spatial Hessian is negative semidefinite and the time derivative has the opposing sign needed to contradict\n\n\\[(u_\\varepsilon)_t-\\kappa\\Delta u_\\varepsilon<0.\\]\n\nThe argument also accommodates a maximum at the final time by using a one-sided time derivative.\n\nPhysically, pure diffusion cannot create a new temperature above both the initial maximum and the largest imposed boundary temperature. Likewise, nonnegative initial and boundary data remain nonnegative. Applying the principle to differences proves uniqueness and supremum-norm stability for the initial-boundary-value problem. A strong parabolic maximum principle says, in particular, that an interior maximum at positive time forces constancy throughout the relevant connected portion of the preceding space–time region, not necessarily at future times.\n\nFor equations with reaction terms,\n\n\\[u_t-\\kappa\\Delta u=f(u),\\]\n\na literal assertion that maxima never increase is generally false. Spatially constant solutions satisfy the ordinary differential equation \\(y'=f(y)\\), and may grow. Comparison instead uses a suitable scalar solution as a time-dependent upper bound. If \\(f\\) is locally Lipschitz continuous, ordered bounded solutions can often be compared by linearizing their difference and absorbing a bounded zeroth-order term through an exponential change of variables.\n\nThis distinction separates diffusion from reaction. The heat equation preserves a fixed range under appropriate boundary data, whereas reaction–diffusion equations may preserve a moving interval, exhibit decay, or develop large amplitudes. Maximum-principle arguments still apply, but the comparison function must reflect the actual equation. On unbounded spatial domains, additional boundedness or growth assumptions are also needed; continuity and a formal differential inequality alone do not control behavior at spatial infinity."
    },
    {
      "heading": "The maximum modulus principle in complex analysis",
      "text": "In complex analysis, the corresponding central theorem concerns the modulus of a holomorphic function. If \\(f\\) is holomorphic on a connected open set and \\(|f|\\) has a local maximum at an interior point, then \\(f\\) is constant. Consequently, if \\(f\\) is holomorphic on a bounded domain and continuous on its closure,\n\n\\[\\max_{\\overline{\\Omega}}|f| = \\max_{\\partial\\Omega}|f|.\\]\n\nThe modulus is essential: complex numbers themselves have no order compatible with their field operations, so a complex-valued function has no ordinary maximum in the real-valued sense.\n\nThe theorem is closely connected to harmonicity. Writing \\(f=u+iv\\), the Cauchy–Riemann equations imply that \\(u\\) and \\(v\\) are harmonic. Moreover,\n\n\\[\\Delta |f|^2=4|f'|^2\\geq0.\\]\n\nThus \\(|f|^2\\) is subharmonic. If it attains an interior maximum, the strong maximum principle makes it constant, and the displayed identity then gives \\(f'=0\\). Alternatively, Cauchy’s integral formula expresses \\(f\\) at the center of a disk as an average of its boundary values; equality in the resulting modulus bound forces rigidity.\n\nWhere \\(f\\neq0\\), the function \\(\\log|f|\\) is harmonic. At zeros it is naturally interpreted as an extended-valued subharmonic function with value \\(-\\infty\\). This logarithmic viewpoint is especially useful for growth estimates, the distribution of zeros, and comparison with harmonic functions. It also explains the minimum modulus principle: if \\(f\\) has no zeros, the maximum modulus principle applied to \\(1/f\\) prevents a nonconstant \\(f\\) from having a positive interior minimum of its modulus.\n\nSeveral fundamental results follow. A bounded entire function is constant by Liouville’s theorem, which can be obtained from Cauchy estimates. The maximum modulus principle also supplies a standard proof of the fundamental theorem of algebra: if a nonconstant polynomial had no zero, its reciprocal would be entire and bounded, hence constant. In another direction, Schwarz’s lemma applies the principle to a holomorphic self-map of the unit disk fixing the origin and yields \\(|f(z)|\\leq|z|\\).\n\nBoundary and growth hypotheses remain indispensable. A holomorphic function on an unbounded region may grow without bound while having small values on its finite boundary. Extensions such as the Phragmén–Lindelöf principle compensate by imposing restrictions on growth at infinity. The complex maximum modulus principle therefore belongs to the same broad family as elliptic maximum principles, while exploiting the stronger rigidity supplied by complex differentiability."
    },
    {
      "heading": "Weak solutions and nonlinear comparison",
      "text": "Classical maximum principles assume pointwise second derivatives, but many important equations have only weak solutions. For a uniformly elliptic divergence-form equation, consider the weak inequality\n\n\\[-\\operatorname{div}(A(x)\\nabla u)\\leq0.\\]\n\nIf \\(u\\) belongs to the Sobolev space \\(H^1(\\Omega)\\), this means\n\n\\[\\int_\\Omega A\\nabla u\\cdot\\nabla\\varphi\\,dx\\leq0\\]\n\nfor every nonnegative test function \\(\\varphi\\) vanishing at the boundary. The coefficients may be merely bounded and measurable, provided their ellipticity is controlled.\n\nSuppose the boundary trace of \\(u\\) is at most \\(M\\). The positive part\n\n\\[w=(u-M)^+=\\max\\{u-M,0\\}\\]\n\nthen belongs to \\(H_0^1(\\Omega)\\) and can be used as a test function. Since \\(\\nabla w=\\nabla u\\) where \\(u>M\\), ellipticity gives\n\n\\[\\lambda\\int_\\Omega|\\nabla w|^2\\,dx \\leq \\int_\\Omega A\\nabla u\\cdot\\nabla w\\,dx \\leq0.\\]\n\nThus \\(w=0\\), and \\(u\\leq M\\) almost everywhere. This proof replaces the geometry of a classical maximum point with an energy calculation and the boundary trace.\n\nA major nonlinear example is the p-Laplacian,\n\n\\[\\Delta_pu=\\operatorname{div}\\!\\left(|\\nabla u|^{p-2}\\nabla u\\right), \\qquad 1<p<\\infty.\\]\n\nIts comparison principle relies on the monotonicity inequality\n\n\\[\\bigl(|\\xi|^{p-2}\\xi-|\\eta|^{p-2}\\eta\\bigr) \\cdot(\\xi-\\eta)\\geq0.\\]\n\nTesting the difference of two weak equations with \\((u-v)^+\\) yields an order comparison under appropriate boundary assumptions. The operator is degenerate for \\(p>2\\) and singular for \\(1<p<2\\) at zero gradient, so classical uniformly elliptic proofs do not transfer unchanged. Strong comparison results require additional care.\n\nFor fully nonlinear equations, solutions may lack even weak second derivatives. The theory of viscosity solutions instead tests the equation against smooth functions touching the candidate solution from above or below. Comparison is proved by analyzing possible positive maxima of \\(u-v\\), often using a penalized expression involving two independent spatial variables. A standard device is\n\n\\[u(x)-v(y)-\\frac{|x-y|^2}{2\\varepsilon}.\\]\n\nIt forces the relevant points together as \\(\\varepsilon\\) decreases.\n\nThese methods are central to nonlinear elliptic equations and Hamilton–Jacobi equations, including forms of the Bellman equation. Their common feature is compatibility between the equation and an ordering of functions. Such compatibility is not automatic for arbitrary nonlinear terms or systems of equations. Even when each component of a system resembles a scalar elliptic equation, coupling may destroy comparison unless it satisfies suitable sign conditions."
    },
    {
      "heading": "Quantitative estimates and limits of validity",
      "text": "A basic maximum principle controls a solution by its boundary values when the differential inequality has a favorable sign. Quantitative variants also measure the effect of a forcing term. For a uniformly elliptic operator with \\(c=0\\), an Aleksandrov–Bakelman–Pucci estimate has the schematic form\n\n\\[\\sup_\\Omega u \\leq \\sup_{\\partial\\Omega}u^+ + C\\,\\operatorname{diam}(\\Omega)\\, \\|(Lu)^-\\|_{L^n(\\Omega)},\\]\n\nunder appropriate coefficient hypotheses. Here \\(r^-=\\max\\{-r,0\\}\\), and \\(C\\) depends on the dimension, ellipticity, and relevant bounds on lower-order coefficients. The estimate controls the positive height of a function through the negative part of its differential operator.\n\nUnlike proofs based only on a maximum point, the Aleksandrov–Bakelman–Pucci argument studies a contact set where supporting planes touch a suitable envelope of the function. It relates the range of slopes of those planes to a determinant involving second derivatives. The exponent \\(n\\) reflects this geometric use of \\(n\\)-dimensional volume. Such estimates are important in equations whose coefficients have limited regularity and in fully nonlinear elliptic theory.\n\nAnother quantitative relative is Harnack’s inequality. For a nonnegative harmonic function on a ball \\(B_{2r}\\),\n\n\\[\\sup_{B_r}u\\leq C_n\\inf_{B_r}u.\\]\n\nThe inequality bounds oscillation multiplicatively on a smaller ball and implies that a nonnegative harmonic function vanishing at an interior point is identically zero. Extensions to elliptic equations depend on ellipticity and lower-order bounds. Harnack inequalities provide more information than the qualitative strong maximum principle, although their scope and hypotheses are not identical.\n\nSeveral elementary counterexamples show why the assumptions matter. On \\((0,\\pi)\\), the positive function \\(u(x)=\\sin x\\) satisfies\n\n\\[u''+u=0,\\qquad u(0)=u(\\pi)=0.\\]\n\nIt has a positive interior maximum, so the unrestricted maximum principle fails when the zeroth-order coefficient has the wrong sign. More refined criteria involve the principal Dirichlet eigenvalue: for \\(\\Delta+c\\) with constant \\(c\\), the threshold is governed by the first eigenvalue of \\(-\\Delta\\).\n\nBoundedness of the domain also matters. The harmonic function \\(u(x)=x\\) on \\((0,\\infty)\\) vanishes at its only finite boundary point but is unbounded above. Principles on unbounded domains therefore require additional growth restrictions or barriers at infinity. Other failures arise from insufficient ellipticity, irregular boundaries, or unsuitable coupling in systems. The wave equation, being hyperbolic rather than diffusive, has no general boundary maximum principle of the elliptic type. Maximum principles express specific structural properties, not a universal rule for differential equations."
    },
    {
      "heading": "Probabilistic interpretation",
      "text": "The maximum principle has a natural interpretation through random motion. Let \\(B_t\\) be a standard Brownian motion starting at \\(x\\in\\Omega\\), and let\n\n\\[\\tau=\\inf\\{t\\geq0:B_t\\notin\\Omega\\}\\]\n\nbe its first exit time. For a bounded harmonic function continuous on the closure of a bounded domain, the stopped process \\(u(B_{t\\wedge\\tau})\\) is a martingale, and\n\n\\[u(x)=\\mathbb E_x[u(B_\\tau)].\\]\n\nThe interior value is an expected value of boundary values and therefore cannot exceed their maximum or fall below their minimum.\n\nThe connection follows from Itô’s formula. Standard Brownian motion has infinitesimal generator \\(\\tfrac12\\Delta\\), so the drift term in \\(u(B_t)\\) vanishes precisely when \\(\\Delta u=0\\). Stopping the process before it leaves the domain and applying an appropriate optional-stopping argument yields the representation. If the exit distribution from \\(x\\) is denoted by \\(\\omega^x\\), then\n\n\\[u(x)=\\int_{\\partial\\Omega}g(y)\\,\\omega^x(dy)\\]\n\nfor boundary data \\(g\\). The probability measure \\(\\omega^x\\) is called harmonic measure.\n\nThis representation explains positivity particularly directly. If \\(g\\geq0\\), its expectation is nonnegative. If \\(g\\leq h\\), then the corresponding expectations are ordered. The strong maximum principle reflects a further fact: Brownian motion can explore every open part of a connected domain with positive probability before suitable stopping times. A harmonic function that achieves its greatest value inside cannot average genuinely smaller nearby values and retain that maximum.\n\nMore general diffusion processes lead to operators containing variable second-order coefficients and first-order drift terms. A generator may take the form\n\n\\[\\mathcal L u = \\frac12\\sum_{i,j}a_{ij}(x)D_{ij}u +\\sum_i b_i(x)D_i u.\\]\n\nThe connection between \\(\\mathcal L\\) and a stochastic differential equation allows comparison properties of the equation to be interpreted through transition probabilities. A nonpositive zeroth-order coefficient can correspond to killing or discounting, consistent with the sign restrictions in analytic maximum principles.\n\nThe Feynman–Kac formula extends these ideas to forcing and potential terms by introducing time integrals and exponential weights. For instance, under suitable assumptions, the solution of \\(-\\tfrac12\\Delta u=f\\) with zero boundary data is the expected accumulated value of \\(f(B_t)\\) before exit. Nonnegative forcing then produces a nonnegative solution. Such representations support probabilistic proofs and Monte Carlo methods, but require integrability, stopping-time, and boundary assumptions rather than following from formal averaging alone."
    },
    {
      "heading": "Discrete principles and numerical approximation",
      "text": "A discrete maximum principle already appears on a finite weighted graph. Suppose every interior vertex satisfies\n\n\\[u_i=\\sum_j p_{ij}u_j, \\qquad p_{ij}\\geq0,\\qquad \\sum_jp_{ij}=1.\\]\n\nEach value is a convex combination of neighboring values. If an interior vertex attains the global maximum, every neighbor with positive weight must have the same value. Repeating this argument propagates the maximum along positive-weight paths. If every interior vertex is connected to the prescribed boundary, the maximum and minimum are controlled by boundary values, and the discrete Dirichlet problem is unique.\n\nThis is the graph counterpart of harmonic averaging. The same coefficients can be interpreted as transition probabilities of a Markov chain. The solution at an interior vertex is the expected boundary value where the chain first reaches the boundary. Thus graph-theoretic, probabilistic, and analytic maximum principles share an identical order-preserving mechanism.\n\nFor the finite difference method, the standard five-point discretization of Laplace’s equation on a uniform square grid gives\n\n\\[u_{i,j}=\\frac14\\left( u_{i+1,j}+u_{i-1,j}+u_{i,j+1}+u_{i,j-1} \\right).\\]\n\nIts nonnegative averaging weights imply a discrete maximum principle. In matrix form, discretizations of \\(-\\Delta\\) commonly produce an M-matrix after boundary conditions are imposed. A nonsingular M-matrix has a nonnegative inverse, so nonnegative discrete forcing gives a nonnegative discrete solution. This provides an algebraic route to comparison and uniqueness.\n\nTime discretization introduces additional restrictions. For the heat equation in \\(d\\) dimensions, forward Euler with centered spatial differences has the update\n\n\\[u_i^{m+1} = (1-2dr)u_i^m +r\\sum_{j\\sim i}u_j^m, \\qquad r=\\frac{\\kappa\\Delta t}{h^2}.\\]\n\nThe update is a convex combination when\n\n\\[0\\leq r\\leq\\frac{1}{2d}.\\]\n\nUnder this condition, no new maximum is generated. If the time step is too large, the central weight becomes negative and overshoots may occur. Backward Euler with the standard diffusion stencil retains the corresponding monotonicity without this explicit time-step restriction.\n\nThe finite element method presents geometric complications. For piecewise linear elements and isotropic diffusion, nonobtuse simplicial meshes provide familiar sufficient sign conditions for a discrete maximum principle; weaker conditions are possible in particular settings. Anisotropic diffusion changes the relevant mesh geometry. Higher-order approximation or a small residual alone does not guarantee preservation of bounds.\n\nDiscrete maximum principles are therefore distinct from general numerical stability. A method may be stable in an energy norm yet produce negative concentrations or temperatures above all prescribed data. Monotone discretizations, flux corrections, and suitable mesh design address this stronger requirement of preserving the ordering and range properties of the continuous problem."
    },
    {
      "heading": "Historical development and the optimal-control meaning",
      "text": "Maximum principles developed through the study of potential theory, harmonic functions, and complex function theory in the nineteenth century. Their importance became clearer as differential equations were studied through qualitative properties rather than explicit formulas alone. Harmonic averaging connected local differential equations with global boundary values, while complex analysis supplied a particularly rigid maximum modulus theorem.\n\nIn 1924, Eberhard Hopf published a strong maximum principle for linear second-order elliptic equations. The boundary point result bearing his name subsequently became a standard instrument in elliptic theory. During the twentieth century, maximum principles were extended to parabolic equations, operators with rough coefficients, nonlinear problems, and weak solution settings. The Aleksandrov–Bakelman–Pucci estimate introduced a powerful geometric approach to quantitative bounds, while viscosity solution theory made comparison central to equations for which classical differentiability is unavailable.\n\nA distinct development occurred in optimal control theory. The Pontryagin maximum principle, developed in the 1950s by Lev Pontryagin and collaborators, concerns the choice of a control that optimizes a dynamical system. Despite the shared name, it does not generally say that a solution’s maximum lies on a boundary. It gives necessary conditions for an optimal trajectory by introducing an auxiliary costate and a Hamiltonian.\n\nFor a normal optimal-control problem maximizing\n\n\\[J=\\Phi(x(T))+\\int_0^T\\ell(x(t),a(t),t)\\,dt, \\qquad \\dot x=f(x,a,t),\\]\n\ndefine\n\n\\[H(x,p,a,t)=p\\cdot f(x,a,t)+\\ell(x,a,t).\\]\n\nUnder suitable hypotheses, an optimal control \\(a^*(t)\\) and associated costate \\(p(t)\\) satisfy\n\n\\[\\dot x=\\frac{\\partial H}{\\partial p}, \\qquad \\dot p=-\\frac{\\partial H}{\\partial x},\\]\n\ntogether with\n\n\\[H(x^*,p,a^*,t)=\\max_{a\\in U}H(x^*,p,a,t)\\]\n\nfor almost every time. With a free terminal state and no additional terminal constraints, the transversality condition is \\(p(T)=\\nabla\\Phi(x(T))\\). More general statements include a multiplier that accommodates abnormal extremals.\n\nThe maximization here is over instantaneous admissible controls, not over spatial points. It supplies necessary conditions and is not, without further assumptions such as appropriate concavity, a sufficient certificate of global optimality. Its relationship to the calculus of variations and Lagrange multipliers differs from the boundary comparison arguments of elliptic theory.\n\nThere is nevertheless an indirect connection. Dynamic programming leads to Hamilton–Jacobi–Bellman equations for the value function, and uniqueness for those equations is often established by comparison principles in the viscosity sense. Optimal control thus contains both meanings: Pontryagin’s pointwise Hamiltonian maximization along an extremal trajectory, and differential-equation maximum or comparison principles controlling the associated value function."
    }
  ],
  "links": [
    "mathematical-analysis",
    "harmonic-function",
    "partial-differential-equation",
    "laplace-operator",
    "laplace-equation",
    "hessian-matrix",
    "gradient",
    "eigenvalues-and-eigenvectors",
    "extreme-value-theorem",
    "boundary-value-problem",
    "poisson-equation",
    "greens-function",
    "heat-equation",
    "diffusion",
    "ordinary-differential-equation",
    "lipschitz-continuity",
    "complex-analysis",
    "cauchy-riemann-equations",
    "liouvilles-theorem-complex-analysis",
    "fundamental-theorem-of-algebra",
    "weak-solution",
    "sobolev-space",
    "almost-everywhere",
    "bellman-equation",
    "wave-equation",
    "wiener-process",
    "martingale",
    "expected-value",
    "itos-formula",
    "stochastic-differential-equation",
    "feynman-kac-formula",
    "monte-carlo-method",
    "convex-combination",
    "markov-chain",
    "finite-difference-method",
    "finite-element-method",
    "numerical-stability",
    "optimal-control",
    "calculus-of-variations",
    "lagrange-multipliers",
    "value-function"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}