{
  "id": "numerical-integration",
  "title": "Numerical Integration",
  "lang": "en",
  "source_lang": "en",
  "summary": "Numerical integration approximates definite integrals through weighted evaluations, adaptive subdivision, transformations, or sampling, with accuracy governed by smoothness, dimension, and numerical error.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Numerical quadrature",
    "Quadrature",
    "Numerical cubature"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/numerical-integration/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/numerical-integration/",
      "json": "https://aiwiki.page/zh/numerical-integration.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Numerical integration is the computation of approximate values of definite integrals using finitely many evaluations of a function or finitely many measured data points. It is a central subject of numerical analysis, used when an antiderivative is unavailable, inconvenient, or insufficient to determine an integral accurately. The term quadrature is especially common for one-dimensional integration; cubature often denotes integration in several dimensions.\n\nThe subject encompasses simple weighted sums, high-order formulas based on polynomials, adaptive algorithms that concentrate evaluations in difficult regions, and probabilistic methods for high-dimensional problems. Its central task is not merely to produce a number, but to relate computational cost to approximation error and numerical reliability. The most effective method depends on the integrand’s smoothness, singularities, oscillation, dimensionality, and the way function values are obtained."
    },
    {
      "heading": "Mathematical formulation and sources of error",
      "text": "For a finite interval, the standard problem is to approximate\n\n\\[I(f)=\\int_a^b f(x)\\,dx\\]\n\nby a quadrature rule\n\n\\[Q_n(f)=\\sum_{i=1}^{n}w_i f(x_i),\\]\n\nwhere \\(x_i\\) are nodes and \\(w_i\\) are weights. The nodes may be fixed in advance or selected adaptively after earlier evaluations. Weighted integration replaces \\(dx\\) by \\(w(x)\\,dx\\), while multiple integration replaces the interval by a region in a higher-dimensional space. Integrals defined through measure theory, including expectations under probability measures, fit the same general framework, although their computational treatment can differ substantially.\n\nA basic property is the degree of exactness: a rule has degree \\(m\\) if it integrates every polynomial of degree at most \\(m\\) exactly. This is an algebraic property, not a universal accuracy ranking. A high-degree rule may perform poorly near a discontinuity, while a lower-degree rule applied on carefully refined subintervals may be highly accurate. Likewise, a rule’s asymptotic order describes its behavior within a specified family of increasingly fine discretizations and under stated smoothness assumptions. An estimate such as \\(O(h^4)\\), written using asymptotic notation, does not specify the size of its hidden constant.\n\nThe total computational error has several sources. Quadrature error is the difference between the exact integral and the ideal weighted sum. Errors in evaluating \\(f\\), rounding errors in floating-point arithmetic, and uncertainty in supplied data are separate contributions. If the evaluated values satisfy \\(\\widetilde f_i=f(x_i)+\\delta_i\\), then\n\n\\[\\left|\\sum_iw_i\\widetilde f_i-\\sum_iw_i f(x_i)\\right| \\leq \\sum_i|w_i|\\,|\\delta_i|.\\]\n\nPositive weights summing to \\(b-a\\) therefore provide a useful stability property: uniformly bounded absolute perturbations are not amplified beyond the interval length. Large positive and negative weights can behave much less favorably.\n\nRelative accuracy also depends on cancellation within the integral. When \\(I(f)\\ne0\\), a useful relative sensitivity measure is\n\n\\[\\kappa(f)=\\frac{\\int_a^b|f(x)|\\,dx}{|I(f)|}.\\]\n\nA large value indicates that substantial positive and negative contributions nearly cancel. This is a property of the problem’s conditioning, distinct from the stability of a particular algorithm. No finite set of samples can guarantee accurate integration over all continuous functions without further assumptions: an unsampled narrow peak can change the integral while leaving every sampled value unchanged. Reliable error bounds therefore require information such as derivative bounds, analyticity, bounded variation, or a probabilistic sampling model."
    },
    {
      "heading": "Interpolatory rules and composite formulas",
      "text": "A broad class of quadrature formulas is obtained by interpolating the integrand at chosen nodes. If\n\n\\[p(x)=\\sum_{i=1}^{n}f(x_i)\\ell_i(x),\\]\n\nwhere \\(\\ell_i\\) are the Lagrange interpolation polynomials, integrating \\(p\\) gives\n\n\\[Q_n(f)=\\sum_{i=1}^{n}f(x_i)\\int_a^b\\ell_i(x)\\,dx.\\]\n\nThus each weight is the integral of a basis polynomial. The resulting rule integrates every polynomial of degree at most \\(n-1\\) exactly; symmetry can increase the degree further. This construction connects quadrature directly to approximation theory, but pointwise interpolation error and integrated error need not behave identically.\n\nThe Newton–Cotes formulas use equally spaced nodes. Closed formulas include both endpoints; open formulas omit them. The trapezoidal rule is\n\n\\[T=\\frac{b-a}{2}[f(a)+f(b)],\\]\n\nand Simpson’s rule is\n\n\\[S=\\frac{b-a}{6} \\left[f(a)+4f\\!\\left(\\frac{a+b}{2}\\right)+f(b)\\right].\\]\n\nTheir degrees of exactness are one and three, respectively. For sufficiently smooth functions, their single-interval errors are\n\n\\[I-T=-\\frac{(b-a)^3}{12}f''(\\xi), \\qquad I-S=-\\frac{(b-a)^5}{2880}f^{(4)}(\\eta),\\]\n\nfor suitable interior points \\(\\xi\\) and \\(\\eta\\). The formulas specify both the required derivatives and the scaling with interval width.\n\nComposite rules apply a simple formula separately on many panels. With \\(x_j=a+jh\\) and \\(h=(b-a)/N\\), the composite trapezoidal rule is\n\n\\[T_N=h\\left[\\frac{f(a)+f(b)}2+\\sum_{j=1}^{N-1}f(x_j)\\right].\\]\n\nFor bounded second derivative, its error is \\(O(h^2)\\). Composite Simpson integration, using an even number of equal subintervals, has error \\(O(h^4)\\) when the fourth derivative is bounded. For example, the trapezoidal approximation to \\(\\int_0^1x^2\\,dx\\) is exactly\n\n\\[T_N=\\frac13+\\frac{1}{6N^2},\\]\n\nwhereas Simpson’s rule integrates the same function exactly. Such examples distinguish refinement-based convergence from polynomial exactness.\n\nIncreasing the degree of a single equally spaced interpolant is not generally as dependable as subdividing the interval. High-degree interpolation can oscillate strongly near endpoints, and sufficiently high-order closed Newton–Cotes rules contain negative weights. Composite low-order rules avoid some of these difficulties and are especially useful when data are already available on a uniform grid. For irregularly spaced measurements, integrating a piecewise-linear interpolant gives\n\n\\[\\sum_j\\frac{x_{j+1}-x_j}{2}(y_j+y_{j+1}).\\]\n\nThis computes the area under that interpolant; it does not recover unresolved structure between observations. Measurement uncertainty and assumptions about interpolation remain part of the integration problem."
    },
    {
      "heading": "Extrapolation, periodicity, and spectral convergence",
      "text": "Quadrature accuracy can sometimes be improved without changing the underlying sampling rule. For sufficiently smooth functions, the composite trapezoidal rule has an expansion described by the Euler–Maclaurin formula:\n\n\\[T(h)=I+\\frac{h^2}{12}[f'(b)-f'(a)] -\\frac{h^4}{720}[f'''(b)-f'''(a)]+\\cdots.\\]\n\nThe leading coefficients depend on differences between derivatives at the endpoints. This expansion explains why a rule based on straight-line segments can support much higher-order approximations: results on successively finer grids contain predictable error terms that can be canceled algebraically.\n\nRichardson extrapolation combines approximations with different step sizes. If \\(A(h)=I+ch^p+O(h^{p+r})\\), then\n\n\\[\\frac{2^pA(h/2)-A(h)}{2^p-1}\\]\n\neliminates the leading error. Applied to trapezoidal estimates with \\(p=2\\), the first extrapolation gives\n\n\\[\\frac{4T(h/2)-T(h)}3,\\]\n\nwhich is Simpson’s formula on the associated grid. Romberg integration repeats this process in a triangular table. Setting \\(R_{k,0}=T(h_0/2^k)\\), its recurrence is\n\n\\[R_{k,j}=R_{k,j-1} +\\frac{R_{k,j-1}-R_{k-1,j-1}}{4^j-1}.\\]\n\nPreviously evaluated function values are reused as the grid is refined.\n\nExtrapolation depends on the assumed error expansion. Endpoint singularities, discontinuities, and inadequate resolution can replace the expected even powers of \\(h\\) with different behavior. A table may then converge slowly or show misleading agreement. Eventually rounding error limits useful extrapolation even for smooth functions, because subtracting nearby approximations magnifies uncertainty in their differences. The existence of increasingly high formal orders does not imply that every additional extrapolation improves an actual floating-point result.\n\nPeriodic functions exhibit a particularly important exception to the usual low-order assessment of the trapezoidal rule. If the function and its relevant derivatives agree at both endpoints, the corresponding Euler–Maclaurin boundary terms vanish. For an infinitely differentiable periodic function, convergence can be faster than every fixed algebraic power; for a function analytic in a complex strip around the real axis, geometric convergence follows under appropriate bounds. A complementary explanation uses the Fourier series: equally spaced samples integrate low Fourier modes exactly, while unresolved modes whose frequencies are multiples of the grid size contribute aliasing error.\n\nClenshaw–Curtis quadrature obtains another rapidly convergent family by sampling at cosine-spaced nodes on \\([-1,1]\\). The interpolant is expressed in Chebyshev polynomials and integrated analytically. Coefficients can be computed efficiently using algorithms related to the fast Fourier transform. Doubling the number of subintervals preserves existing nodes, making refinement economical. For analytic integrands, convergence is typically geometric; for finitely smooth integrands, coefficient decay and quadrature convergence become algebraic. These methods show how node placement and the structure of a function can matter more than a rule’s nominal degree alone."
    },
    {
      "heading": "Gaussian quadrature and orthogonal polynomials",
      "text": "Gaussian quadrature chooses both nodes and weights to maximize polynomial exactness. For\n\n\\[I_w(f)=\\int_a^b f(x)w(x)\\,dx,\\]\n\nwith a positive weight and suitable finite moments, an \\(n\\)-node Gaussian rule is exact through degree \\(2n-1\\). This is the highest possible degree for an \\(n\\)-node rule in this setting. To see the obstruction at degree \\(2n\\), consider\n\n\\[q(x)=\\prod_{i=1}^{n}(x-x_i)^2.\\]\n\nThe quadrature evaluates \\(q\\) as zero, but its weighted integral is positive when the measure has sufficiently rich support.\n\nThe Gaussian nodes are the zeros of the degree-\\(n\\) orthogonal polynomial associated with \\(w\\). Orthogonality uses the inner product\n\n\\[\\langle p,q\\rangle=\\int_a^b p(x)q(x)w(x)\\,dx.\\]\n\nFor any polynomial \\(p\\) of degree at most \\(2n-1\\), division by the orthogonal polynomial \\(p_n\\) gives \\(p=sp_n+r\\), where \\(s\\) and \\(r\\) have degree at most \\(n-1\\). Orthogonality makes the integral of \\(sp_n\\) zero, and the Gaussian nodes make its quadrature value zero. Interpolatory exactness handles \\(r\\). This argument explains the exceptional degree without requiring the interpolant itself to have degree \\(2n-1\\).\n\nGauss–Legendre quadrature uses weight \\(1\\) on \\([-1,1]\\). Its two-node rule is\n\n\\[\\int_{-1}^{1}f(x)\\,dx \\approx f\\!\\left(-\\frac1{\\sqrt3}\\right) +f\\!\\left(\\frac1{\\sqrt3}\\right),\\]\n\nwhich integrates cubics exactly. Gauss–Jacobi rules use \\((1-x)^\\alpha(1+x)^\\beta\\), with \\(\\alpha,\\beta>-1\\); Gauss–Laguerre uses \\(x^\\alpha e^{-x}\\) on \\([0,\\infty)\\); and one common Gauss–Hermite convention uses \\(e^{-x^2}\\) on the real line. Matching a known endpoint or tail factor to the weight can leave a much smoother function for approximation.\n\nFor sufficiently differentiable \\(f\\), the Gauss–Legendre error is\n\n\\[I(f)-Q_n(f)= \\frac{2^{2n+1}(n!)^4} {(2n+1)[(2n)!]^3}f^{(2n)}(\\xi).\\]\n\nThis formula establishes exactness and supplies bounds when the derivative is controlled, but high derivatives may be difficult to estimate. For analytic functions, approximation in a complex neighborhood often gives more useful convergence bounds. Gaussian weights are positive, which also contributes to stable summation.\n\nNodes and weights can be computed using the three-term recurrence relation of orthogonal polynomials. The Golub–Welsch method, published in 1969, forms a symmetric tridiagonal Jacobi matrix: its eigenvalues give the nodes, and components of normalized eigenvectors determine the weights. Related Gauss–Radau rules prescribe one endpoint, while Gauss–Lobatto rules prescribe both, sacrificing some degree of exactness in exchange for endpoint inclusion."
    },
    {
      "heading": "Adaptive quadrature and error estimation",
      "text": "Adaptive quadrature varies the distribution of nodes according to observed local difficulty. A typical algorithm approximates the integral on an interval, estimates the local error, and subdivides where that estimate is largest. Smooth regions may require few evaluations, while layers, cusps, or nearby singularities receive greater resolution. This differs from uniform refinement, which spends the same effort everywhere whether or not the function needs it. Adaptivity is especially valuable when each function evaluation is expensive and the locations of difficult features are not known beforehand.\n\nAdaptive Simpson integration compares a Simpson estimate \\(S\\) on one interval with the sum \\(S_2\\) of Simpson estimates on its two halves. In the regime where the leading smooth-error model applies, halving the panel width reduces the combined error by approximately \\(16\\). Consequently,\n\n\\[I-S_2\\approx\\frac{S_2-S}{15},\\]\n\nand the corrected approximation is \\(S_2+(S_2-S)/15\\). The difference supplies an economical error indicator because several function values are shared. It is an asymptotic estimate, however, not an unconditional bound.\n\nMany general-purpose algorithms instead use Gauss–Kronrod quadrature. A Kronrod extension retains the nodes of an \\(n\\)-point Gaussian rule and adds \\(n+1\\) nodes. The resulting \\(2n+1\\)-point rule generally has substantially higher polynomial exactness, while the difference between the two estimates indicates local error. Common pairs contain 7 and 15 nodes or 10 and 21 nodes. Reusing the Gaussian nodes makes the pair much less expensive than two unrelated rules of comparable sizes.\n\nGlobal adaptive routines maintain a collection of active subintervals and sum their integral estimates. They also combine local error indicators, often by summation, and repeatedly split the interval with the largest estimated contribution. A common target is\n\n\\[E_{\\mathrm{est}}\\leq \\max\\!\\left(\\varepsilon_{\\mathrm{abs}}, \\varepsilon_{\\mathrm{rel}}|Q|\\right),\\]\n\nwhere \\(Q\\) is the current integral approximation. Absolute tolerance is essential when the integral is zero or small because of cancellation. Practical termination also depends on evaluation budgets, subdivision limits, and whether floating-point arithmetic can still represent distinct interior nodes.\n\nError estimation is vulnerable to unresolved features. A narrow peak can lie between every sampled node, and oscillations can cause two rules to agree for the wrong reason. Known discontinuities or singular points are therefore commonly used as explicit subdivision boundaries. More elaborate estimators examine several differences, variation measures, or extrapolation sequences, but none removes the need for assumptions about the function class. A successful termination status means that an algorithm’s internal criteria have been met; it is not, by itself, a mathematical proof of the reported error."
    },
    {
      "heading": "Singular, improper, and oscillatory integrals",
      "text": "An integrable singularity can make ordinary smooth-function formulas converge slowly. For example, \\(\\int_0^1x^{-1/2}\\,dx\\) exists, but a rule requiring the value at zero cannot be applied directly. Under the substitution \\(x=t^2\\),\n\n\\[\\int_0^1 f(x)\\,dx=\\int_0^1 2t\\,f(t^2)\\,dt,\\]\n\nthe singular integrand \\(f(x)=x^{-1/2}\\) becomes the constant \\(2\\). More generally, substitutions can soften algebraic endpoint behavior, while weighted Gaussian rules can incorporate known factors directly. Splitting at an interior singularity converts it into endpoint problems, but evaluating the singular point may still need to be avoided.\n\nInfinite intervals require separate attention to the tail. The transformation \\(x=t/(1-t)\\) gives\n\n\\[\\int_0^\\infty f(x)\\,dx =\\int_0^1 f\\!\\left(\\frac{t}{1-t}\\right)\\frac{dt}{(1-t)^2}.\\]\n\nAlthough the domain becomes finite, the transformed function may develop an endpoint singularity. Alternatively, truncating at a finite \\(R\\) separates quadrature error on \\([0,R]\\) from tail error beyond \\(R\\); the latter needs its own estimate. Gauss–Laguerre and Gauss–Hermite formulas are effective when the decay structure matches their weights. No change of variables automatically makes every improper integral easy.\n\nTanh–sinh quadrature, introduced in 1974 by Hidetosi Takahasi and Masatake Mori, uses\n\n\\[x=\\tanh\\!\\left(\\frac{\\pi}{2}\\sinh t\\right)\\]\n\nfor integration on \\((-1,1)\\), followed by a trapezoidal sum in \\(t\\). The derivative of the transformation decays double-exponentially as \\(|t|\\) grows. For broad classes of endpoint singularities, this produces a rapidly decaying transformed integrand and effective numerical convergence. Its implementation must nevertheless handle nodes extremely close to the endpoints and avoid losing information through rounded endpoint values.\n\nHighly oscillatory integrals introduce a different obstacle. For\n\n\\[I(\\omega)=\\int_a^b g(x)e^{i\\omega\\phi(x)}\\,dx,\\]\n\na generic method may require many samples simply to resolve cancellation as \\(|\\omega|\\) increases. Filon-type methods approximate the slowly varying amplitude and integrate it against the oscillatory factor. Levin-type methods seek an auxiliary function \\(p\\) satisfying approximately\n\n\\[p'(x)+i\\omega\\phi'(x)p(x)=g(x),\\]\n\nso the integral can be recovered from boundary values of \\(p(x)e^{i\\omega\\phi(x)}\\). Stationary points where \\(\\phi'(x)=0\\) need special treatment; local asymptotic behavior there differs from regions of nonstationary phase.\n\nCauchy principal values are distinct from ordinary improper integrals. If \\(a<c<b\\), cancellation around a simple pole can be exposed through\n\n\\[\\operatorname{PV}\\int_a^b\\frac{f(x)}{x-c}\\,dx = \\int_a^b\\frac{f(x)-f(c)}{x-c}\\,dx +f(c)\\log\\!\\left(\\frac{b-c}{c-a}\\right).\\]\n\nFor differentiable \\(f\\), the first integrand has a removable singularity with limiting value \\(f'(c)\\). This identity shows why the precise definition of an integral matters computationally: a principal-value routine computes a different limiting operation from separate one-sided improper integration."
    },
    {
      "heading": "Deterministic integration in several dimensions",
      "text": "For a region \\(D\\subset\\mathbb R^d\\), numerical integration seeks\n\n\\[I=\\int_D f(\\mathbf x)\\,d\\mathbf x.\\]\n\nThe simplest construction applies a one-dimensional rule in each coordinate. A tensor-product rule with \\(m\\) nodes per coordinate uses \\(m^d\\) evaluations. At \\(m=10\\), this means \\(100\\) evaluations in two dimensions but \\(10^{10}\\) in ten dimensions. This exponential growth is a standard manifestation of the curse of dimensionality. It limits direct product rules even when their one-dimensional accuracy is excellent.\n\nRepeated one-dimensional integration is another possibility when the region has convenient bounds. An adaptive outer routine calls an inner integration routine at each selected outer node. This approach accommodates variable limits but creates nested error budgets: the outer algorithm receives approximate function values from the inner computation. If the inner tolerance is too loose, its errors can disrupt outer convergence. If it is unnecessarily tight, substantial work is wasted. Direct multidimensional adaptive cubature instead subdivides regions and uses local multidimensional rules, often with directional indicators to decide which coordinate to split.\n\nGeometry is frequently handled by a change of variables. For a suitable map \\(\\mathbf x=\\Phi(\\mathbf u)\\),\n\n\\[\\int_D f(\\mathbf x)\\,d\\mathbf x = \\int_U f(\\Phi(\\mathbf u)) \\left|\\det D\\Phi(\\mathbf u)\\right|\\,d\\mathbf u.\\]\n\nThe determinant of the Jacobian matrix accounts for volume distortion. Rectangles, triangles, tetrahedra, spheres, and curved elements therefore call for different mappings or specialized rules. A geometrically valid mapping may still produce a sharply varying Jacobian or transformed integrand, so geometric simplification and numerical regularity are separate considerations.\n\nSparse-grid methods combine tensor-product rules at different levels while omitting many combinations in which every coordinate is simultaneously highly resolved. Smolyak’s construction, introduced in 1963, is a central example. For functions with bounded mixed derivatives, a common type of fixed-dimension estimate has the form\n\n\\[O\\!\\left(N^{-r}(\\log N)^{c(d,r)}\\right),\\]\n\nwhere \\(r\\) describes smoothness and the logarithmic exponent depends on dimension and the construction. These methods can greatly reduce costs in moderate dimensions, but mixed-regularity assumptions and dimension-dependent constants remain important. Anisotropic variants assign greater resolution to influential coordinates.\n\nProblem structure can matter more than nominal dimension. A function may depend mainly on a few variables or admit an accurate low-rank representation such as\n\n\\[f(x_1,\\ldots,x_d)\\approx \\sum_{k=1}^{R}\\prod_{j=1}^{d}g_{jk}(x_j).\\]\n\nOn a product domain, integrating this representation reduces to collections of one-dimensional integrals. Coordinate transformations, separable expansions, and adaptive dimension selection can therefore make apparently high-dimensional problems tractable. Such gains rely on identifying structure; an arbitrary function of many variables has no corresponding guarantee."
    },
    {
      "heading": "Monte Carlo and quasi-Monte Carlo integration",
      "text": "Numerical integration is closely connected to probability because an integral can often be expressed as an expectation. If \\(X\\) is uniformly distributed on a region \\(D\\) of finite volume \\(V\\), then\n\n\\[I=V\\,\\mathbb E[f(X)].\\]\n\nThe Monte Carlo method estimates this expectation using independent samples:\n\n\\[\\widehat I_N=\\frac{V}{N}\\sum_{k=1}^{N}f(X_k).\\]\n\nWhen the variance is finite,\n\n\\[\\operatorname{Var}(\\widehat I_N) =\\frac{V^2\\operatorname{Var}(f(X))}{N}.\\]\n\nThus the root-mean-square error decreases as \\(N^{-1/2}\\). This exponent does not explicitly depend on dimension, although the variance and cost of obtaining samples may depend on it strongly.\n\nUnder the usual independent, identically distributed finite-variance assumptions, the central limit theorem supports approximate uncertainty intervals based on the estimated standard error. These intervals have a probabilistic interpretation, unlike deterministic derivative-based error bounds. They can be unreliable with inadequate sampling of rare events, heavy tails, or infinite variance. The law of large numbers establishes convergence under appropriate integrability assumptions, but does not by itself supply a useful finite-sample error estimate.\n\nImportance sampling draws from a density \\(q\\) rather than uniformly and uses\n\n\\[I=\\mathbb E_q\\!\\left[\\frac{f(X)}{q(X)}\\right].\\]\n\nThe density must cover the relevant support, and finite variance requires suitable control of \\(f^2/q\\). A proposal concentrated where the integrand is important can reduce error dramatically; an inadequately matched proposal can increase it or produce infinite variance. Other variance-reduction methods include stratification, antithetic sampling, and control variates. A control variate subtracts a correlated function whose integral is known, so that only the smaller residual requires stochastic estimation.\n\nQuasi-Monte Carlo integration replaces independent random points with deliberately distributed low-discrepancy points, including Sobol’ sequences and lattice constructions. On the unit cube, the Koksma–Hlawka inequality states\n\n\\[\\left| \\frac1N\\sum_{k=1}^{N}f(\\mathbf x_k) -\\int_{[0,1]^d}f(\\mathbf x)\\,d\\mathbf x \\right| \\leq V_{\\mathrm{HK}}(f)D_N^*,\\]\n\nwhere \\(V_{\\mathrm{HK}}\\) is Hardy–Krause variation and \\(D_N^*\\) is star discrepancy. The bound requires finite variation and is often difficult to evaluate numerically. It explains a route to deterministic convergence rather than supplying a universal practical error bar.\n\nRandomized quasi-Monte Carlo methods combine structured point sets with random shifts or scramblings. Suitable randomizations retain uniform marginal distributions and permit uncertainty estimation from independent randomized replicates. Performance can be substantially better than ordinary Monte Carlo for sufficiently regular functions, especially when effective dimension is low. If samples instead come from Markov chain Monte Carlo, dependence changes the variance of an average: autocorrelation and effective sample size become relevant. Independent-sample error formulas cannot simply be transferred to correlated output."
    },
    {
      "heading": "Historical development and computational applications",
      "text": "Numerical integration developed from the geometric problem of quadrature: finding areas by comparison with simpler figures. Archimedes used the method of exhaustion in the third century BCE, notably in determining the area of a parabolic segment. These arguments supplied exact geometric results through limiting constructions rather than modern floating-point algorithms. The emergence of calculus in the seventeenth century clarified the relation between area and antidifferentiation, but did not eliminate the need for numerical approximation. Many definite integrals still could not be evaluated conveniently in closed form.\n\nInterpolatory formulas associated with Isaac Newton and Roger Cotes became part of eighteenth-century computation. Carl Friedrich Gauss introduced his optimal-degree quadrature construction in 1814. Mechanical calculation and, later, electronic computers shifted emphasis toward efficient node generation, automatic refinement, and computable error indicators. Werner Romberg published his extrapolation method in 1955; Aleksandr Kronrod developed embedded extensions of Gaussian formulas in the 1960s. These developments made it possible to organize classical formulas into automatic integration procedures rather than use them only as isolated hand calculations.\n\nThe QUADPACK library, documented in 1983, became an influential implementation of adaptive one-dimensional integration. It organized different procedures for general finite intervals, infinite intervals, known difficult points, oscillatory factors, and related special cases. This organization reflects a lasting computational principle: no single algorithm is uniformly best for all integrands. General-purpose routines are valuable, but they operate within assumptions about sampling, regularity, and error behavior. Specialized structure can justify a different method even when a general routine is available.\n\nIn the finite element method, quadrature evaluates element stiffness matrices, mass matrices, and load vectors. On an affine element, polynomial basis functions and polynomial coefficients can make exactness requirements predictable. Curved mappings and nonlinear material laws complicate the integrands. Underintegration may reduce cost or alleviate particular locking effects, but it can also introduce spurious zero-energy modes. In Bayesian inference, integration appears in posterior expectations and marginal likelihoods, with deterministic quadrature useful in low dimensions and sampling methods prominent in larger problems. Other applications include response integrals in mechanics, transport calculations, and evaluation of convolution and transform formulas.\n\nActual performance depends on more than the number of nodes. Evaluations may involve a simple elementary expression, an expensive simulation, or noisy experimental data. Batch evaluation and parallel computing favor methods that can generate many nodes in advance, while sequential adaptivity introduces scheduling dependencies. Pairwise or compensated summation can reduce accumulation error, especially when there is cancellation. A computational record normally includes the approximation, its estimated error, the number of evaluations, and any termination warning. These quantities distinguish a well-resolved result from a plausible number produced after an exhausted evaluation budget."
    }
  ],
  "links": [
    "integral",
    "mathematical-function",
    "numerical-analysis",
    "antiderivative",
    "measure-theory",
    "polynomial",
    "big-o-notation",
    "floating-point-arithmetic",
    "condition-number",
    "numerical-stability",
    "interpolation",
    "fourier-series",
    "fast-fourier-transform",
    "orthogonal-polynomials",
    "inner-product",
    "recurrence-relation",
    "eigenvalues-and-eigenvectors",
    "tensor-product",
    "curse-of-dimensionality",
    "jacobian-matrix",
    "probability",
    "expected-value",
    "monte-carlo-method",
    "central-limit-theorem",
    "standard-error",
    "law-of-large-numbers",
    "markov-chain-monte-carlo",
    "archimedes",
    "method-of-exhaustion",
    "calculus",
    "carl-friedrich-gauss",
    "finite-element-method",
    "bayesian-inference",
    "marginal-likelihood",
    "parallel-computing"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}