Skip to content

Interval Arithmetic

Arithmetic on interval enclosures that preserves inclusion of every admissible exact result, with outward rounding in finite-precision implementations.

Version
v2 · 2026-10-03 · History
Domain-specific #
13342
Domain group
Formal Sciences
Origin domain
Mathematics
Subdomains
Numerical Analysis, Validated Numerics → Mathematics
Aliases
Interval computation, Interval methods

Core Idea

Interval arithmetic replaces a single real operand by a closed interval containing every value under consideration, and replaces each arithmetic operation by an operation on intervals that contains every possible exact result. If \(x\in X\) and \(y\in Y\), then the interval result \(X\mathbin\circ Y\) must contain \(x\mathbin\circ y\) for a supported operation \(\circ\). In exact real arithmetic, the interval operation can be defined by the set of all pairwise results. In a finite-precision computer, its lower endpoint is rounded down and its upper endpoint up, preserving the inclusion despite endpoint rounding.[1]

The important invariant is not that the reported bounds are narrow or that the true value sits at their midpoint. It is that the true result cannot lie outside them, provided the input enclosures and all operations are sound. Replacing each operation in an expression by its interval counterpart gives a natural interval extension \(F\) with \(x_i\in X_i\Rightarrow f(x_1,\ldots,x_n)\in F(X_1,\ldots,X_n)\).[1] That makes interval arithmetic a building block of validated numerics; a separate theorem is still needed to infer, for example, existence or uniqueness of a root.

The distinction between safety and precision is central. Repeated occurrences of the same variable can be treated as if they varied independently, widening a correct bound. For \(X=[3.14,3.15]\), ordinary interval subtraction yields \(X-X=[-0.01,0.01]\), even though \(x-x=0\) when both occurrences are the same \(x\). This dependency loss is not caused only by machine rounding: it is already present in set operations that forget the equality constraint.[1]

Structural Signature

Sig role-phrases: admissible real-value enclosure → lifted interval operation → inclusion invariant → outward finite-precision realization.

  • Admissible real-value enclosure: Each operand \(X=[\underline x,\overline x]\) includes every exact value intended by the problem. It may be a point interval when the operand is exactly represented. If the true input was excluded before arithmetic starts, later inclusion claims about that input have no basis.[1]
  • Lifted interval operation: Addition, subtraction, multiplication, and suitably defined division act on sets of possible operands. For ordinary bounded intervals, the result encloses \(\{x\circ y:x\in X,y\in Y\}\), not merely an operation on midpoints. The simple division rule assumes the denominator interval excludes zero; exceptional divisions require an explicitly different treatment.[1]
  • Inclusion invariant: Every admissible exact operand choice maps to a result within the interval output. When valid operations are composed in a natural interval extension, the same implication holds for the full expression. This guarantee is the defining test; a narrow-looking numerical estimate without it is not a rigorous enclosure.[1]
  • Outward finite-precision realization: In an implementation whose endpoint format cannot represent every real number, the exact lower bound is rounded toward \(-\infty\) and the exact upper bound toward \(+\infty\). This is a condition on machine interval arithmetic, not an extra axiom needed when interval-set operations are exact.[1]

These roles have different logical jobs. The first says what the computation promises to include; the second defines the transformed operands; the third is the invariant obtained; the fourth protects that invariant against representational rounding. Dependency overestimation is an important failure of tightness, not a fifth condition that every valid interval must exhibit.

What It Is Not

It is not ordinary floating-point arithmetic with an error bar added afterward. A central estimate can be accurate in practice and still omit some admissible exact result. Interval arithmetic requires an inclusion-preserving operation at each step, with properly enclosed initial inputs and outward endpoint rounding where finite precision is used.[1]

It is not automatically numerical certification. The arithmetic provides rigorous bounds on evaluated expressions. To prove that \(f\) has exactly one root in an interval, an interval-Newton theorem additionally requires differentiability, a derivative enclosure excluding zero, and an inclusion test for its Newton image. A bare interval around a guessed root is not that certificate.[1]

It is not a promise of the smallest possible bound for an expression. A single exact interval operation on independent operands may give the exact set range, yet repeated occurrences of one variable can lose their correlation across operations. \(X-X\) may remain wide even though the symbolic expression \(x-x\) is identically zero.[1]

It is also not an unconditional guarantee for every machine code path. If the exact decimal $0.1$ is first converted to a nearby binary float and only then made a point interval, that interval need not contain the intended exact decimal. Rump demonstrates that passing the decimal as a string to the interval constructor repairs this particular input-conversion trap.[1]

Scope of Application

The literal setting is numerical analysis over real-valued quantities whose exact values are represented by point or nonpoint intervals. The basic method applies to arithmetic expressions built from supported operations. Natural interval extensions then bound function values over a domain; interval algorithms can use those enclosures for further tests. The guarantee concerns what is included, not the number of significant digits that remain useful after repeated operations.[1]

Rump's finite-precision product is a direct arithmetic instance: two short positive operand intervals produce an exact product range, and a three-digit implementation widens the endpoints outward. His interval-Newton example is an applied instance: interval arithmetic encloses the derivative and Newton image for \(f(x)=x^2-2\), while the separate theorem turns checked inclusions into a root conclusion.[1]

The method can be extended to vectors, matrices, nonlinear systems, optimization, and computer-assisted proofs, but each extension must specify its interval operations, handling of exceptions, and theorem connecting bounds to the desired claim. The scalar inclusion principle does not by itself establish every advertised application. For division, a denominator interval crossing zero cannot be fed into the simple finite-interval quotient rule used here; one must use a defined extended convention, split the domain, or report that the requested finite bound is unavailable.[1]

Clarity

The signature separates three statements often conflated: the input is enclosed, the arithmetic result is enclosed, and a proposition has been proved. The first is an assumption to establish; the second follows from correctly lifted operations and directed rounding; the third needs a problem-specific implication. In the root example, the implication is supplied by interval Newton under explicit derivative and containment conditions, not by the interval datatype alone.[1]

It also distinguishes a valid but loose interval from an invalid one. If \(X-X\) is broad because two occurrences were treated independently, inclusion still holds; simplification or domain subdivision may make it more useful. If an endpoint was rounded inward or the true decimal input was lost at conversion, inclusion itself fails. Those two outcomes demand different repairs: improve expression form for tightness, but fix input or implementation correctness for soundness.

Manages Complexity

For one operation, interval arithmetic compresses an uncountable set of possible operand pairs into two endpoints while preserving a specified property: containment of all results. Composition lets a program propagate that compact enclosure through a long expression without enumerating every real input. The inclusion invariant gives a short inductive proof obligation for each supported operator rather than a fresh pointwise argument for every input.[1]

This compression intentionally discards correlation. Once two appearances of \(x\) are each represented only by \(X\), the arithmetic does not remember that they must take the same value. The representation thus trades a compact, safe enclosure for possible overestimation. One can recover useful precision by rewriting a repeated-variable expression, retaining additional structure, or subdividing the input range—but that is extra analytic or computational work, not a relaxation of the soundness requirement.[1]

Abstract Reasoning

To check an interval claim, first state the intended exact inputs and prove each lies in its initial interval. Next inspect every operation and function extension, including division domains and any transcendental implementation. Verify that finite-precision lower and upper endpoints are rounded outward. Only then invoke the inclusion principle to conclude that each exact output lies in the computed interval.[1]

From there, ask what the bound actually decides. If an interval evaluation \(F(X)\) excludes zero, then \(f(x)\) cannot be zero for any \(x\in X\)—provided \(F\) is a valid enclosure of \(f\) on all of \(X\). If it contains zero, no root is thereby proved. Interval Newton can make a stronger statement: for differentiable \(f\), a derivative enclosure excluding zero and a Newton image lying inside \(X\) yield a unique root in \(X\); if the image is disjoint from \(X\), there is no root there.[1] These are downstream theorem applications, not the arithmetic's defining axiom.

A useful diagnostic order is therefore: sound inputs → sound operations → useful width → theorem conditions. Rump's naive interval iteration for \(x^2-2\) remains sound but broadens dramatically because the repeated \(X\) values lose dependency. The centered interval-Newton form uses the same interval arithmetic within a better-structured proof procedure.[1]

Knowledge Transfer

Within validated numerics, the inclusion invariant travels from a scalar product to natural interval extensions of functions and to verified algorithms that consume those extensions. The mapped roles remain the same: admissible values, lifted operations, inclusion, and finite-precision endpoint protection. A new algorithm must still prove its own link from those bounds to its conclusion; interval arithmetic supplies reliable premises, not an automatic universal theorem.[1]

Across application domains, an engineering tolerance, a measurement range, or an exact mathematical parameter can all be encoded as an interval if the intended values are genuinely enclosed and the operations are defined for them. The name transfers literally only where this real-interval arithmetic and inclusion semantics are used. A qualitative “range of possibilities” in another field may echo Representation or conservative bounding, but without the real operations and inclusion proof it is an analogy rather than Interval Arithmetic.

Examples

Outward-rounded product

Rump gives \(X=[2.31,2.33]\) and \(Y=[3.74,3.76]\). Because both are positive, the exact set of pairwise products is \([8.6394,8.7608]\). In a three-digit decimal endpoint format, a correct interval result is \([8.63,8.77]\): the lower endpoint moves below the exact minimum and the upper above the exact maximum. The widened interval is not a guess about likely products; it contains every product from the stated input intervals.[1]

Mapped back: Admissible real-value enclosure = \(X\) and \(Y\); Lifted interval operation = all pairwise products with their extreme endpoint products; Inclusion invariant = every \(xy\) lies in the exact range and therefore in \([8.63,8.77]\); Outward finite-precision realization = down-rounding $8.6394$ to $8.63$ and up-rounding $8.7608$ to $8.77$. This example exhibits the arithmetic directly, before any root or optimization theorem is invoked.

Interval bounds inside a root test

For \(f(x)=x^2-2\), Rump starts with \(X=[1,2]\), checks that the derivative interval \(f'(X)=2X\) excludes zero, and evaluates a centered interval-Newton image using interval operations. Its successive displayed enclosures narrow around \(\sqrt2\). The arithmetic certifies the image bounds; Theorem 6.2 supplies the separate conclusion about a root when its image-inclusion condition is checked. This distinction matters because Rump also shows that mechanically intervalizing the ordinary Newton recurrence can widen a correct interval until it is useless.[1]

Mapped back: Admissible real-value enclosure = the current \(X\) and interval for the selected center; Lifted interval operation = interval evaluation of \(f\), \(f'\) and \(\widetilde x-f(\widetilde x)/f'(X)\); Inclusion invariant = computed intervals contain the corresponding exact derivative and Newton-image values; Outward finite-precision realization = the implementation's directed-rounding endpoint bounds. The proof of root uniqueness remains an additional conditional step, not a fifth arithmetic role.

Structural Tensions

  • T1: Guaranteed inclusion vs. actionable tightness. Rounding outward protects the true result, but dependency loss can widen an enclosure until it cannot decide whether a condition holds. Tightening through reformulation or subdivision costs analysis and computation; merely rounding inward to make the output prettier destroys the guarantee. Diagnostic: Is the current bound too wide because of representation/independence loss, and what valid transformation would narrow it while preserving inclusion?
  • T2: Local enclosure vs. global theorem. The arithmetic can guarantee that an evaluated function image is enclosed. A uniqueness or existence conclusion demands hypotheses about the function and an interval-Newton (or other) decision condition. Treating local inclusion as a proof overclaims; demanding a full theorem before calling a single interval product valid underclaims. Diagnostic: Which claim follows from the inclusion invariant alone, and which depends on a separately verified theorem?
  • T3: General-purpose operations vs. exceptional inputs. Simple arithmetic rules make compositional evaluation easy, but division by an interval containing zero and incorrectly converted decimal inputs require explicit handling. Pretending the ordinary rule covers them makes a clean-looking API unsound; rejecting all interval use because exceptions exist discards valid cases. Diagnostic: Were the input and operation domains checked before the enclosure was trusted?

Structural–Framed Character

The core membership test is structural: given specified real intervals and operations, set inclusion and directed endpoint rounding are mathematical conditions. Evaluative weight enters when choosing whether a bound is tight enough for a user's purpose, not when deciding whether inclusion holds. Human practice supplies implementation conventions, rounding-mode controls, and problem formulations, but the pointwise-set definition can be stated independently of a particular institution or programming library.[1]

The method has an institutional and historical home in numerical analysis and validated computing; those origins do not make correctness a community vote. Its vocabulary travels to engineering and computer-assisted proof when the same interval-set operations are actually used, not whenever a writer speaks loosely of a “range.” For import versus recognition, a domain that already computes rigorous real bounds can be recognized as instantiating the method once its operations satisfy inclusion; a domain that merely records uncertainty would need to import the arithmetic and its proof obligations before the label applies.

The candidate portable skeleton is inclusion-preserving enclosure of possibilities: a tractable object safely contains all relevant outcomes even if it overestimates them. Whether that skeleton merits a future prime requires examples and boundaries outside real-number intervals; no such prime is asserted here. The live Representation is a useful comparison for finite-precision endpoint encoding, in which a distinct machine medium represents exact real bounds, but exact interval-set arithmetic operates on intervals as sets and need not instantiate that target/medium mapping. Approximation is also not a safe strict parent: this method can operate on exact point inputs and can return loose bounds with no purpose-relative error tolerance. Its character: a strongly structural but domain-specific arithmetic method whose validity is formal, while its value depends on numerical precision and downstream proof needs.

Structural Core vs. Domain Accent

Portable skeletal relation. The possible higher-order pattern is inclusion-preserving enclosure: replace a set of possibilities by a safe containing set that supports reasoning despite excess width. That is an explicitly future-prime question, not a verified existing prime or a proposed DAG parent. In the machine implementation, endpoints encode exact interval bounds, and live Representation can describe that additional target/medium relation; in exact interval-set arithmetic the operands already are sets, so the representation relation is not required.

Indispensable domain-bound mechanism. Interval Arithmetic's identity requires ordered real intervals, lifted real operations, and a compositional inclusion theorem. In finite precision it additionally requires directed rounding of computed endpoints and sound initialization of exact inputs. Removing these details leaves only a broad enclosure idea, not this named method. Its famous dependency problem arises from interval operations forgetting correlations, rather than from a universal property of all representations.[1]

Prime boundary. The method is reusable across numerical applications, but those applications all instantiate the same real-arithmetic machinery. A qualitative uncertainty interval in social reasoning, a date window, or a categorical range lacks the operative +, −, ×, ÷ semantics and inclusion proof. The named numerical residual has not been demonstrated independently of its mathematical substrate. A future prime about conservative enclosure would need separate cross-domain evidence and boundaries rather than promoting this title by metaphor; none is added here.

No strict typed parent relation is asserted in the current DAG.

Neighborhood in Abstraction Space

Interval Arithmetic sits in a sparse region of the domain-specific corpus (77th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Numerical Root-Finding & Quadrature Methods (7 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With

  • Numerical certification: A theorem-backed proof about a numerical candidate or domain. Interval arithmetic often supplies bounds to such a proof but does not itself conclude existence or uniqueness.
  • Floating-point arithmetic: Uses finite representations and rounding; ordinary nearest rounding of a single estimate does not guarantee a two-endpoint set enclosure.
  • Exact symbolic interval expressions: Exact set operations can satisfy the inclusion invariant without any rounding. Outward rounding is a machine-realization obligation, not the mathematical definition.
  • Dependency-aware exact range calculation: An interval extension can be correct yet overestimate the range of an expression with repeated variables, since ordinary interval operations forget their equality constraint.
  • An interval constructed after unsound input conversion: A pair of bounds is only a valid enclosure for the target values it actually includes. A rounded binary point interval need not contain an intended exact decimal such as $0.1$.[1]

References

[1] Siegfried M. Rump, “Verification methods: Rigorous results using floating-point arithmetic”, Acta Numerica 19 (2010), 287–449. Author-hosted original paper, especially §§5–7 and §9.1: set and interval operations (eqs. 5.1, 5.4–5.5), directed rounding (5.14–5.15), inclusion (5.16–5.17), dependency and interval Newton (Algorithm 6.1 and Theorem 6.2), input conversion, and worked product (9.15). registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z