{
  "id": "interval-arithmetic",
  "title": "Interval Arithmetic",
  "lang": "en",
  "source_lang": "en",
  "summary": "Interval arithmetic computes guaranteed enclosures of numerical results, accounting for bounded input uncertainty and rounding errors.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Interval analysis",
    "interval computation"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/interval-arithmetic/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/interval-arithmetic/",
      "json": "https://aiwiki.page/zh/interval-arithmetic.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Interval arithmetic is a system of numerical calculation in which quantities are represented by intervals rather than single numbers. An interval records lower and upper bounds within which a value is known to lie, and arithmetic operations produce bounds containing every possible result consistent with those inputs. When implemented with correctly controlled rounding, it can account for errors introduced by floating-point arithmetic as well as uncertainty already present in the data.\n\nThe central purpose is not merely to estimate an answer, but to establish an enclosure whose validity follows from mathematical rules. This makes interval arithmetic a foundation of validated numerics: numerical methods that can certify the existence of solutions, exclude impossible cases, and provide rigorous error bounds. Its applications include nonlinear equations, linear systems, global optimization, differential equations, and computer-assisted proofs. Its principal difficulty is that guaranteed bounds can become unnecessarily wide, particularly when calculations repeatedly use the same uncertain quantities."
    },
    {
      "heading": "Intervals and the meaning of an enclosure",
      "text": "A bounded, closed real interval is a set\n\n\\[X=[\\underline{x},\\overline{x}] =\\{x\\in\\mathbb{R}:\\underline{x}\\leq x\\leq\\overline{x}\\}.\\]\n\nIts endpoints are real numbers, with \\(\\underline{x}\\leq\\overline{x}\\). A point interval, such as \\([3,3]\\), represents an exact number. An interval with positive width can represent an uncertain measurement, an approximation error, or a whole domain over which a function is being investigated. These interpretations use the same arithmetic, although their scientific meanings differ.\n\nCommon descriptors are the width, midpoint, and radius:\n\n\\[w(X)=\\overline{x}-\\underline{x},\\qquad m(X)=\\frac{\\underline{x}+\\overline{x}}{2},\\qquad r(X)=\\frac{\\overline{x}-\\underline{x}}{2}.\\]\n\nThus \\(X=[m(X)-r(X),m(X)+r(X)]\\). Endpoint representation is convenient for directed rounding, while midpoint–radius representation is often useful for error estimates and matrix bounds. In machine arithmetic, converting between these representations must itself preserve enclosure; a naively rounded midpoint and radius need not do so.\n\nFor a real-valued function \\(f\\), an interval calculation seeks an interval \\(F(X)\\) satisfying\n\n\\[\\{f(x):x\\in X\\}\\subseteq F(X).\\]\n\nThe set on the left is the exact range. Its smallest enclosing interval is its interval hull. An interval evaluation may produce this hull, but it may instead produce a larger enclosure. The difference is crucial: validity means that no possible result is omitted, whereas sharpness concerns how little extra material is included.\n\nAn interval does not ordinarily assign probabilities to its elements. The statement \\(x\\in[2,3]\\) does not imply that \\(x\\) follows a uniform probability distribution, nor that values near the midpoint are more likely. Likewise, an interval enclosure is not automatically a confidence interval. A confidence interval is associated with a statistical procedure and its coverage properties; an interval-arithmetic bound is a deterministic consequence of specified assumptions. Statistical bounds can supply inputs to interval calculations, but their probabilistic qualifications remain attached to those inputs.\n\nIn several dimensions, intervals are combined into a box,\n\n\\[X=X_1\\times\\cdots\\times X_n,\\]\n\na Cartesian product of coordinate intervals. Boxes are inexpensive to store and manipulate, but usually discard relations among coordinates. If the only possible pairs satisfy \\(x_2=x_1\\), replacing that set by a box also admits pairs that violate the relation. Much of interval analysis concerns recovering, preserving, or compensating for information lost through such representations."
    },
    {
      "heading": "Arithmetic operations and algebraic structure",
      "text": "For two intervals \\(X=[a,b]\\) and \\(Y=[c,d]\\), an elementary operation is defined by taking all pairwise results and enclosing them. In exact real arithmetic,\n\n\\[X+Y=[a+c,b+d],\\]\n\n\\[X-Y=[a-d,b-c],\\]\n\nand\n\n\\[XY= [\\min(ac,ad,bc,bd),\\max(ac,ad,bc,bd)].\\]\n\nThese formulas give the exact interval ranges of the corresponding binary operations when the two operands vary independently. For example,\n\n\\[[1,2]+[3,5]=[4,7],\\qquad [-2,3][4,5]=[-10,15].\\]\n\nAlthough implementations often use sign-dependent shortcuts for multiplication, those shortcuts must remain equivalent to the enclosing definition.\n\nIf \\(0\\notin Y\\), division is defined through the reciprocal:\n\n\\[\\frac{X}{Y}=X\\left[\\frac1d,\\frac1c\\right].\\]\n\nThe formula applies whether \\(Y\\) is strictly positive or strictly negative. Division by an interval containing zero requires different treatment. For example, the real quotients of \\(1\\) by nonzero elements of \\([-1,1]\\) form\n\n\\[(-\\infty,-1]\\cup[1,\\infty).\\]\n\nThis is disconnected and cannot be represented exactly by a single interval. A system may return its hull, the entire real line, retain two separate components, or use an exception mechanism. These choices are part of the interval system’s semantics, not interchangeable computational details.\n\nInterval addition and multiplication are commutative and associative in exact endpoint arithmetic, but intervals do not form a field analogous to the real numbers. For a nondegenerate interval \\(X\\), subtraction does not cancel uncertainty:\n\n\\[X-X=[-w(X),w(X)],\\]\n\nrather than \\([0,0]\\). Full distributivity also fails. For \\(X=[1,2]\\), \\(Y=[1,1]\\), and \\(Z=[-1,-1]\\),\n\n\\[X(Y+Z)=[0,0], \\qquad XY+XZ=[-1,1].\\]\n\nThe replacement is subdistributivity,\n\n\\[X(Y+Z)\\subseteq XY+XZ.\\]\n\nThe right-hand expression permits the two occurrences of \\(X\\) to vary independently.\n\nA central property is inclusion isotonicity: enlarging the inputs cannot shrink the exact interval result. For example,\n\n\\[X\\subseteq X',\\quad Y\\subseteq Y' \\quad\\Longrightarrow\\quad X+Y\\subseteq X'+Y'.\\]\n\nComparable statements hold for multiplication and for division on appropriate domains. This property supports algorithms that progressively narrow boxes or subdivide domains.\n\nSet intersection is another important operation. If two independent arguments establish that a quantity lies in \\(X\\) and in \\(Y\\), it lies in \\(X\\cap Y\\). An empty intersection can certify inconsistency. Interval algorithms therefore often alternate arithmetic evaluation with intersections, using several valid but differently obtained bounds to improve one another."
    },
    {
      "heading": "Directed rounding and rigorous machine computation",
      "text": "Real-number formulas alone do not make a computer calculation rigorous. If both endpoints are rounded to the nearest machine number, a lower endpoint may move upward or an upper endpoint downward, excluding a valid result. An enclosing implementation instead rounds lower bounds toward negative infinity and upper bounds toward positive infinity. For addition, the machine operation has the form\n\n\\[X+Y= [\\operatorname{rd}(a+c),\\operatorname{ru}(b+d)],\\]\n\nwhere \\(\\operatorname{rd}\\) and \\(\\operatorname{ru}\\) denote downward and upward rounding.\n\nThis distinction matters even for ordinary decimal constants. The decimal fraction \\(0.1\\) has no finite binary floating-point representation. If a program first converts it to one nearby binary number and then creates a point interval, the resulting interval generally represents that binary number, not the exact decimal fraction. To enclose the decimal value, the conversion must produce outward-rounded endpoints directly from its textual or exact rational representation. Similarly, placing an interval around the result of a long, already completed floating-point computation does not retrospectively bound all errors in that computation.\n\nThe IEEE 754 floating-point standard provides rounding directions needed for basic enclosing operations. IEEE 1788-2015 specifies interval arithmetic, including operations and mechanisms for recording qualifications about definedness and continuity. Correct implementation nevertheless requires attention to the entire software environment. Compiler transformations that reassociate expressions, replace operations, or assume a fixed rounding mode can invalidate a proof based on the original sequence of operations. Underflow, overflow, signed zeros, infinities, and exceptional values also require explicit handling.\n\nElementary transcendental functions present additional challenges. For a monotone exponential function, enclosing the endpoint values suffices:\n\n\\[\\exp([a,b])= [\\operatorname{rd}(\\exp a),\\operatorname{ru}(\\exp b)].\\]\n\nFor trigonometric functions, interior extrema must also be considered. If an interval contains a point where sine equals \\(1\\), its upper bound must include \\(1\\), regardless of the endpoint values. For large arguments, even determining whether a critical point lies inside the interval can require rigorous argument reduction and additional precision.\n\nSome implementations change the hardware rounding mode; others use error-free transformations, higher-precision intermediates, or explicit error estimates. No single implementation technique is mandatory. The required property is that every returned bound satisfies the stated enclosure contract. Higher precision can reduce rounding inflation, but it cannot by itself eliminate overestimation caused by lost dependence between variables. Machine rounding error and structural interval overestimation are distinct sources of excess width."
    },
    {
      "heading": "Dependency, expression choice, and the wrapping effect",
      "text": "The dependency problem arises when repeated occurrences of the same variable are treated as independent interval operands. If \\(X=[1,2]\\), direct evaluation gives\n\n\\[X-X=[-1,1],\\]\n\nalthough the function \\(f(x)=x-x\\) is identically zero. The arithmetic operation is correct for two independently varying inputs; the overestimation occurs because the expression requires those inputs to be equal.\n\nEquivalent formulas can therefore have very different interval evaluations. For\n\n\\[f(x)=x^2-x,\\qquad X=[0,1],\\]\n\na dedicated square operation gives \\(X^2=[0,1]\\), so direct subtraction yields\n\n\\[X^2-X=[-1,1].\\]\n\nCompleting the square gives\n\n\\[f(x)=\\left(x-\\frac12\\right)^2-\\frac14,\\]\n\nwhose interval evaluation is exactly\n\n\\[\\left[-\\frac14,0\\right].\\]\n\nThe range is the same mathematical object in both cases; the computational representation determines how much dependency information survives.\n\nSubdivision is a general remedy. A domain is split into smaller intervals, each is evaluated separately, and the results are combined. On narrower domains, repeated occurrences usually introduce less excess width. Under suitable continuity and convergence assumptions, systematic refinement can approach the exact range. The cost can nevertheless grow rapidly with dimension, reflecting the curse of dimensionality. Splitting every coordinate into ten pieces creates \\(10^n\\) boxes in \\(n\\) dimensions.\n\nA related geometric problem is the wrapping effect. Consider a box in the plane transformed by a rotation. Its exact image is a rotated rectangle, but the smallest coordinate-aligned enclosing box generally contains additional points. Repeatedly transforming and reboxing the set can magnify this excess, even though the underlying rotation preserves distances. In dynamical calculations, wrapping can dominate genuine growth of uncertainty and make otherwise valid bounds practically useless.\n\nSeveral representations address these losses. Affine arithmetic expresses quantities using shared noise symbols, preserving linear correlations among occurrences. Taylor models retain a multivariate polynomial together with a rigorous remainder interval, capturing nonlinear dependence over a domain. Other approaches propagate ellipsoids, parallelepipeds, or sets described by linear constraints before converting to intervals when necessary.\n\nThese methods do not make interval bounds universally sharp. Retaining more structure increases storage and computational cost, and nonlinear operations can still introduce approximation errors. The practical choice is therefore not simply between “interval” and “non-interval” computation, but among representations that preserve different kinds of information while maintaining a checkable enclosure."
    },
    {
      "heading": "Function evaluation, differentiation, and range bounds",
      "text": "A natural interval extension is obtained by replacing real variables and operations in a formula by their interval counterparts. If \\(f\\) is built from supported elementary functions, its natural extension \\(F\\) satisfies\n\n\\[f(X)\\subseteq F(X)\\]\n\nwherever the expression is defined. More generally, a useful interval extension should agree with ordinary evaluation on point inputs, up to unavoidable machine enclosure, and should remain inclusion-preserving. Its sharpness depends strongly on the expression tree.\n\nInformation from derivatives often yields stronger bounds. For a differentiable scalar function on \\(X\\), choose a center \\(m\\in X\\) and let \\(F'(X)\\) enclose all derivative values. The mean value theorem gives the centered enclosure\n\n\\[f(X)\\subseteq f(m)+F'(X)(X-m).\\]\n\nThe point value and all arithmetic must also be enclosed rigorously. For several variables, a corresponding form uses an interval gradient or Jacobian matrix evaluated over the box.\n\nDerivative signs can remove the need for subdivision. If \\(F'(X)\\) is strictly positive, \\(f\\) is increasing throughout \\(X\\), and its range is bounded by its endpoint values. If the derivative interval contains zero, that does not establish the presence of a stationary point: it may merely reflect overestimation. Additional subdivision or derivative analysis can distinguish these cases.\n\nHigher-order expansions provide still more structure. For \\(x,m\\in X\\), a second-order Taylor formula gives\n\n\\[f(x)=f(m)+f'(m)(x-m) +\\frac12 f''(\\xi)(x-m)^2\\]\n\nfor an intermediate point \\(\\xi\\). Replacing \\(f''(\\xi)\\) by an interval enclosure over \\(X\\) yields a rigorous remainder bound. Higher-order Taylor expansions can be treated similarly, provided that the remainder is actually bounded; a truncated series without such a bound is not a certificate.\n\nAutomatic differentiation can propagate interval derivatives through a computational graph. Unlike numerical differentiation based on nearby sample values, it applies differentiation rules to the program’s operations. The resulting bounds can support monotonicity tests, root verification, sensitivity calculations, and optimization. Nondifferentiable operations and branches require separate mathematical treatment; differentiating only the branch taken by a midpoint evaluation need not describe the full interval domain.\n\nDomain restrictions are equally important. For example, the real logarithm requires positive arguments, and the square root requires nonnegative ones. An interval crossing a domain boundary may need splitting, restriction, or a result explicitly qualified as only partially defined. A numerical enclosure is meaningful only together with the domain conditions under which the represented expression exists."
    },
    {
      "heading": "Verified solutions of nonlinear equations",
      "text": "Interval methods distinguish several claims that approximate root-finding often leaves unresolved: whether a root exists, whether it is unique in a region, and whether all roots in a specified domain have been found. A small residual alone proves none of these. If a function is nearly flat, a tiny function value can occur far from a root; if it has a nonzero local minimum, no root may exist at all.\n\nThe simplest exclusion test evaluates an interval extension \\(F(X)\\). If\n\n\\[0\\notin F(X),\\]\n\nthen \\(f\\) has no zero in \\(X\\). The converse is false: inclusion of zero may result from interval overestimation. A basic complete-search strategy therefore evaluates boxes, discards those that cannot contain roots, and subdivides the rest. Derivative information can greatly accelerate this process.\n\nThe scalar interval Newton method combines an approximate solution with a bound on all derivatives in an interval. For \\(m\\in X\\),\n\n\\[N(X)=m-\\frac{f(m)}{F'(X)},\\]\n\nassuming \\(0\\notin F'(X)\\). Every root in \\(X\\) belongs to \\(N(X)\\), so replacing \\(X\\) by \\(X\\cap N(X)\\) preserves all roots in the original interval. If this intersection is empty, the original interval contains no root. Under the standard continuity and differentiability assumptions, the strict inclusion\n\n\\[N(X)\\subset\\operatorname{int}(X)\\]\n\ncertifies existence; the exclusion of zero from the derivative enclosure also ensures uniqueness in \\(X\\).\n\nFor example, let \\(f(x)=x^2-2\\), \\(X=[1,2]\\), and \\(m=3/2\\). Then \\(f(m)=1/4\\), while \\(F'(X)=[2,4]\\), giving\n\n\\[N(X)=\\frac32-\\frac{1/4}{[2,4]} =\\left[\\frac{11}{8},\\frac{23}{16}\\right].\\]\n\nThis interval lies strictly inside \\([1,2]\\), certifying the unique positive root there. Further iterations produce narrower certified bounds. Ordinary Newton’s method can supply a good initial approximation, but the interval step provides the verification.\n\nFor systems, direct interval division is replaced by validated linear algebra or an operator such as the Krawczyk operator:\n\n\\[K(X)=m-Cf(m)+(I-CJ(X))(X-m),\\]\n\nwhere \\(C\\) is a point matrix approximating the inverse Jacobian. A self-inclusion condition establishes existence under the usual hypotheses. A bound making \\(I-CJ(X)\\) a contraction establishes uniqueness. These conditions are sufficient, not necessary: verification may fail because the box is too wide, the Jacobian is nearly singular, or the root is multiple, rather than because the approximate solution is wrong."
    },
    {
      "heading": "Interval linear algebra and perturbation bounds",
      "text": "An interval matrix represents a family of matrices whose entries lie within specified bounds. For an interval matrix \\(\\mathbf A\\) and interval vector \\(\\mathbf b\\), a common task is to enclose the united solution set\n\n\\[\\Sigma(\\mathbf A,\\mathbf b) = \\{x:Ax=b\\text{ for some }A\\in\\mathbf A,\\ b\\in\\mathbf b\\}.\\]\n\nThe quantifiers matter. This set differs from the set of vectors satisfying every system in the family, and from models in which several entries depend on shared parameters. Independent entry intervals can admit matrices that the original physical model never allows.\n\nApplying ordinary Gaussian elimination with interval operations gives an enclosing algorithm only when its operations remain defined, but often produces excessive width. A pivot interval may contain zero even when each matrix in the original family is nonsingular. Dependency introduced during elimination can prevent useful division, and representing an evolving family by entrywise intervals discards relations among its entries.\n\nPreconditioning addresses part of the difficulty. For a fixed system \\(Ax=b\\), take an approximate solution \\(\\widehat{x}\\) and a numerical approximate inverse \\(R\\). Writing \\(x=\\widehat{x}+e\\) yields\n\n\\[e=R(b-A\\widehat{x})+(I-RA)e.\\]\n\nIf an interval box \\(E\\) satisfies\n\n\\[R(b-A\\widehat{x})+(I-RA)E\\subseteq E,\\]\n\nit provides a fixed-point enclosure under the appropriate existence conditions. In particular, if a compatible matrix norm verifies\n\n\\[\\|I-RA\\|<1,\\]\n\nthe mapping is a contraction and the solution is unique. With interval \\(A\\) and \\(b\\), enclosing the same expression uniformly can certify bounds for every admitted system.\n\nThe residual \\(b-A\\widehat{x}\\) is especially useful when \\(\\widehat{x}\\) is already accurate. Computing it rigorously, sometimes at higher precision, can establish a small correction even if the original solution was obtained by a conventional numerical solver. This separation between approximate computation and rigorous verification is a recurring design principle: expensive exactness is concentrated on the certificate rather than every exploratory step.\n\nThe condition number controls how data uncertainty can affect the solution, but it is distinct from interval overestimation. An ill-conditioned system may genuinely admit a wide range of solutions. A well-conditioned system can still acquire poor interval bounds through a weak algorithm. Verification must distinguish these effects.\n\nRelated techniques certify matrix nonsingularity, bound inverse entries, and enclose eigenvalues and eigenvectors. Symmetry, positive definiteness, sparsity, and shared-parameter structure can substantially improve the bounds. Treating these properties as mere independent entry intervals can discard precisely the information that makes the problem tractable."
    },
    {
      "heading": "Global optimization and constraint solving",
      "text": "In mathematical optimization, a local method can identify a promising candidate without proving that no better point exists elsewhere. Interval methods supply lower bounds over entire regions and can therefore support global certificates. For a minimization problem over a box \\(X\\), an enclosure\n\n\\[f(X)\\subseteq[\\ell,u]\\]\n\nestablishes \\(f(x)\\geq\\ell\\) everywhere in that box. A rigorously evaluated feasible point supplies an upper bound on the global minimum.\n\nAn interval branch-and-bound algorithm maintains a collection of boxes covering the remaining search domain. Each box receives a lower bound, while the best verified feasible value supplies the current upper bound \\(U\\). A box whose lower bound exceeds \\(U\\) cannot contain a better minimizer and may be discarded. Remaining boxes are contracted or split. If the least lower bound across all remaining boxes is \\(L\\), then the global minimum is enclosed by\n\n\\[L\\leq f_{\\min}\\leq U.\\]\n\nThe gap \\(U-L\\) is an objective-value certificate, distinct from a bound on the minimizer’s location.\n\nBounds need not come only from direct function evaluation. Centered forms, derivative signs, convex relaxations, and Hessian matrix estimates can yield stronger lower bounds. If an unconstrained minimizer is known to lie in the interior, its gradient must vanish, allowing interval root methods to eliminate regions. Boundary faces require separate attention; a stationary-point search confined to the interior cannot certify the full constrained problem.\n\nInterval constraint propagation narrows domains using individual equations or inequalities. From \\(x+y=3\\), for example, one can apply\n\n\\[X\\leftarrow X\\cap(3-Y),\\qquad Y\\leftarrow Y\\cap(3-X).\\]\n\nRepeated propagation across a network of constraints can remove large infeasible regions before branching. Stronger contractors combine several constraints through interval Newton steps, linear relaxations, or monotonicity analysis. Every contraction must preserve all feasible points, although it need not remove every infeasible point.\n\nThe same framework supports parameter estimation and geometric feasibility problems. If measurements are specified by bounded errors, a parameter box can be rejected whenever its model predictions cannot intersect an observed range. Surviving boxes then enclose all parameter values consistent with the stated bounds, rather than representing a posterior distribution.\n\nGlobal guarantees remain conditional on the domain, model, arithmetic, and termination criterion. A finite search domain must be justified when the original problem is unbounded. Strict inequalities and singular boundaries require careful handling. A valid algorithm may also leave a large unresolved region: enclosure supplies soundness, not a promise that every difficult optimization problem will be solved efficiently."
    },
    {
      "heading": "Differential equations and validated trajectories",
      "text": "For an initial-value problem\n\n\\[\\dot{x}=f(t,x),\\qquad x(t_0)\\in X_0,\\]\n\na validated solver seeks a set enclosing every trajectory admitted by the initial set and model assumptions. This differs from attaching heuristic error bars to a single numerical trajectory. The enclosure must account for uncertain initial conditions, rounding errors, discretization errors, and, when included, bounded uncertain parameters.\n\nA fundamental step is to establish an a priori enclosure over a time interval. Writing the equation in integral form gives\n\n\\[x(t)=x_0+\\int_{t_0}^{t}f(s,x(s))\\,ds.\\]\n\nIf a trial box \\(B\\) contains \\(X_0\\) and an interval evaluation verifies an appropriate self-inclusion of the associated integral operator over the step, it can certify that the solution remains in \\(B\\). Continuity and Lipschitz conditions supply the existence and uniqueness framework. A bound on the derivative of \\(f\\) can establish a contraction when the time step is sufficiently small.\n\nOnce such a tube is available, a high-order expansion advances the endpoint enclosure. For one component,\n\n\\[x(t_0+h) = \\sum_{k=0}^{p}\\frac{x^{(k)}(t_0)}{k!}h^k+R_{p+1}.\\]\n\nThe remainder is bounded using derivative information over the entire certified tube, not merely at the initial point. Automatic differentiation can generate high-order coefficients efficiently. Both the polynomial evaluation and remainder calculation require enclosing arithmetic.\n\nWrapping is particularly severe in repeated time stepping. Even a stable system can develop rapidly expanding coordinate boxes if its flow rotates or shears the represented set. Validated integrators therefore use changes of coordinates, matrix factorizations, multiple set representations, or Taylor models to preserve geometric relationships. In genuinely unstable systems, however, some growth reflects real sensitivity and cannot be removed by a better representation.\n\nValidated integration also supports event detection, periodic-orbit verification, and boundary-value problems. An event such as crossing a surface \\(g(x)=0\\) requires bounding the event time and establishing the relevant crossing conditions. For a periodic orbit, integration may construct a return map, after which an interval fixed-point argument verifies a solution. These are existence statements about the mathematical model, not empirical validation of the model’s physical assumptions.\n\nFor partial differential equations, finite-dimensional interval calculations alone do not automatically bound the original infinite-dimensional problem. A rigorous computation must also control truncation, spatial discretization, and unresolved modes, often through analytic inequalities and function-space estimates. The same distinction applies whenever a computer solves a finite approximation to a continuous problem: certifying the finite calculation is only one part of certifying the full model."
    },
    {
      "heading": "Historical development, software, and proof practice",
      "text": "The use of upper and lower bounds predates electronic computing, but modern interval arithmetic developed as a systematic response to numerical error and uncertain data. Mieczysław Warmus published work on interval calculations in 1956; Teruo Sunaga’s 1958 work presented interval algebra and applications to numerical analysis. Ramon E. Moore developed the subject extensively in his 1962 doctoral dissertation and the 1966 book Interval Analysis. These contributions established intervals as computational objects rather than occasional estimates appended to ordinary calculations.\n\nLater work expanded the field into verified linear algebra, nonlinear systems, optimization, and differential equations. Researchers including Götz Alefeld, Jürgen Herzberger, Ulrich Kulisch, Arnold Neumaier, and Siegfried Rump developed mathematical methods, implementation techniques, and software for reliable numerical computation. Rudolf Krawczyk’s operator became an important tool for verifying nonlinear solutions without directly constructing an interval inverse Jacobian.\n\nSoftware takes several forms. INTLAB provides interval and verified numerical methods in a MATLAB environment. C-XSC supports extended scientific computing in C++, while libraries such as MPFI combine intervals with multiple-precision endpoints. Other systems emphasize validated differential equations, constraint solving, or global optimization. Their capabilities are not identical: supported functions, rounding guarantees, treatment of exceptional inputs, and available verification algorithms vary.\n\nComputer-assisted mathematical proofs illustrate the distinction between finding a result and certifying it. Interval computations can establish inequalities on large collections of subdomains or verify the hypotheses of an analytic theorem. Warwick Tucker’s computer-assisted proof concerning the Lorenz attractor, published in 2002, is a prominent example of rigorous numerical methods in dynamical systems. Such work combines interval calculations with substantial mathematical arguments; interval arithmetic is not a substitute for identifying the theorem that the computation must establish.\n\nThe trusted components of a certificate include more than the interval formulas. Input interpretation, the implementation of elementary functions, compiler behavior, branch coverage, and the treatment of excluded or singular domains can all affect correctness. Formal verification can reduce reliance on unexamined software by checking arithmetic procedures or the generated certificate in a proof assistant. It does not remove the need to state the mathematical model and assumptions accurately.\n\nA characteristic successful workflow uses fast conventional computation to locate candidates, then a smaller rigorous calculation to verify them. Conversely, failure to verify is not generally a disproof: insufficient precision, a poor coordinate system, excessive dependency, or a nearly singular problem may prevent the sufficient conditions from succeeding. The output must distinguish a certified result, a certified exclusion, and an unresolved case."
    }
  ],
  "links": [
    "floating-point-arithmetic",
    "real-number",
    "probability-distribution",
    "confidence-interval",
    "cartesian-product",
    "field-mathematics",
    "exponential-function",
    "trigonometric-functions",
    "curse-of-dimensionality",
    "derivative",
    "mean-value-theorem",
    "gradient",
    "jacobian-matrix",
    "taylor-series",
    "automatic-differentiation",
    "newtons-method",
    "matrix",
    "gaussian-elimination",
    "norm-mathematics",
    "condition-number",
    "eigenvalues-and-eigenvectors",
    "mathematical-optimization",
    "branch-and-bound",
    "hessian-matrix",
    "lipschitz-continuity",
    "boundary-value-problem",
    "partial-differential-equation",
    "mathematical-proof",
    "formal-verification"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}