Haversine Formula¶

A half-angle spherical-trigonometry relation that converts two latitude–longitude positions into their central angle and great-circle arc distance, with explicit radius, angle-unit, and numerical-boundary controls.

Version: v2 · 2026-09-06 · History
Domain-specific #: 1988
Origin domain: spherical trigonometry
Subdomain: great-circle distance computation
Aliases: Great-circle haversine formula

Core Idea¶

The haversine formula converts the latitude and longitude of two points on a sphere into their central angle and hence the length of the shorter great-circle arc between them. It does so through the half-angle function

\[ \operatorname{hav}(x)=\sin^2\!\left(\frac{x}{2}\right)=\frac{1-\cos x}{2}. \]

For latitudes \(\varphi_1,\varphi_2\), longitude difference \(\Delta\lambda\), and sphere radius \(R\), form

\[ a=\sin^2\!\left(\frac{\varphi_2-\varphi_1}{2}\right) +\cos\varphi_1\cos\varphi_2 \sin^2\!\left(\frac{\Delta\lambda}{2}\right), \]

recover the central angle

\[ \delta=2\operatorname{atan2}(\sqrt a,\sqrt{1-a}), \]

and return \(d=R\delta\). The equivalent principal-branch expression \(2\arcsin\sqrt a\) is mathematically valid for \(0\le a\le1\). The two-argument arctangent spelling makes the numerator, denominator, and selected range explicit.

The abstraction is not merely the displayed equation. A valid instance binds coordinates to one spherical reference surface, converts angular units consistently, normalizes longitude difference, controls floating-point drift outside \([0,1]\), recovers the principal central angle in \([0,\pi]\), and declares whether a spherical approximation is adequate. Its invariant is exact on the assumed sphere: the output is the sphere's metric distance along a shortest great-circle arc. When the sphere is only a model of Earth, model error is separate from arithmetic error.

The formula remains useful because its half-angle accumulation avoids the severe loss of resolution that can occur when a tiny separation is inferred by applying \(\arccos\) to a rounded cosine near one. Roger Sinnott's classic numerical note emphasized that virtue^[1]; modern software still uses spherical distance when speed and an explicitly approximate Earth model are acceptable. The stable role system and its failure boundaries make the Haversine Formula an autonomous domain-specific abstraction rather than a bare glyph sequence.

Structural Signature¶

The defining roles are:

the spherical reference surface — a sphere with declared radius \(R>0\), not an unspoken ellipsoid or terrain surface;
the first angular position — latitude \(\varphi_1\) and longitude \(\lambda_1\) on that sphere;
the second angular position — latitude \(\varphi_2\) and longitude \(\lambda_2\) in the same datum and angular convention;
the wrapped longitude difference — \(\Delta\lambda\), reduced to an equivalent interval such as \([-\pi,\pi]\) when needed;
the half-angle accumulator — the dimensionless value \(a\) built from squared half-angle sines and the latitude cosine product;
the range guard — numerical enforcement of \(0\le a\le1\) against roundoff without hiding invalid inputs;
the principal central angle — \(\delta\in[0,\pi]\), recovered from \(a\);
the radius scaling — \(d=R\delta\), which turns an angle in radians into an arc length in the units of \(R\);
the shortest-arc invariant — the returned scalar equals the great-circle metric between the positions on the declared sphere;
the approximation declaration — an explicit account of whether the sphere is the object itself or a surrogate for an ellipsoid, geoid, or physical route surface.

A computation qualifies only when these roles are recoverable. Merely calling a function haversine is insufficient if its inputs, radius, angle units, output units, branch, or surface model are unknown.

What It Is Not¶

The Haversine Formula is not a general ellipsoidal geodesic solver. Earth is more accurately modeled for geodetic work by an ellipsoid of revolution, and authoritative geodesy guidance distinguishes the computationally convenient sphere from the ellipsoid required for more exact long-distance work^[2]. Karney's algorithms solve the direct and inverse geodesic problems on an ellipsoid robustly and accurately; substituting a mean Earth radius into the spherical formula does not reproduce that solution^[3].

It is not the ordinary planar distance formula applied to latitude and longitude. Degrees of longitude correspond to different surface lengths at different latitudes, longitude wraps at the antimeridian, and straight lines in a latitude–longitude chart are generally not great-circle geodesics.

It is not chord distance. The straight three-dimensional chord joining two radius vectors is \(2R\sin(\delta/2)\), while the spherical surface distance is \(R\delta\). They agree only to first order for small \(\delta\).

It is not a bearing, route, or travel distance. A scalar distance does not select an initial azimuth, enumerate intermediate points, account for barriers, or price a path through a transportation network. For antipodal points the distance is well-defined as \(\pi R\), but infinitely many great circles realize it, so no unique route or bearing follows.

It is not the generic haversine function alone. The function \(\operatorname{hav}\) can occur in other spherical-trigonometry relations; this node concerns the point-pair distance relation and its operational obligations.

Scope of Application¶

Literal uses arise whenever positions are represented on a sphere and a scalar shortest surface separation is required: celestial angular separation, idealized planetary or stellar surfaces, introductory great-circle navigation, rapid geographic proximity screening, database measurements under a declared spherical Earth option, simulation on spherical meshes, and validation fixtures for more elaborate geospatial systems.

The formula is often an appropriate approximation when application tolerances dominate Earth-figure error, when inputs are coarse, or when a fast preliminary filter will later hand close cases to an ellipsoidal solver. PostGIS exposes this decision explicitly: its spherical-distance operation is faster but less accurate than its spheroidal counterpart^[4]. That pairing illustrates good use because the model choice is visible rather than buried in a function name.

The scope contracts when centimeter- or meter-scale geodetic accuracy matters across substantial baselines, when datum transformations are unresolved, when altitude or terrain length matters, or when a route must obey a network. It also contracts on nonspherical bodies unless a spherical surrogate and its error budget are justified.

The mathematical relation is scale-independent. Changing \(R\) changes only the final length scale; the dimensionless central angle remains the same. Consequently, one implementation can serve a unit sphere, a celestial sphere, an idealized planet, or a chosen mean-Earth sphere, provided the radius and interpretation are explicit.

Clarity¶

A reader should be able to audit a Haversine calculation by answering eight questions:

Are the two positions latitude–longitude pairs on the same reference surface and datum?
Are all trigonometric inputs in radians, or is conversion performed before evaluation?
How is longitude difference wrapped across the antimeridian?
Which radius is used, in what units, and why is that radius appropriate?
Is \(a\) clamped only for small floating-point excursions after input validation?
Is the recovered angle constrained to the shorter-arc range \([0,\pi]\)?
Is the requested output a central angle, spherical arc length, ellipsoidal distance, chord, bearing, or routed travel distance?
Does the stated error tolerance include both numerical error and surface-model error?

These questions expose common silent failures. Supplying degrees to sine and cosine produces a dimensionally plausible but wrong number. Mixing kilometers and meters in \(R\) rescales every result by a thousand. Subtracting longitudes without considering their periodic equivalence can complicate other formulas, even though the squared half-angle term itself is periodic. Reporting many decimal places cannot repair a poor sphere model.

Manages Complexity¶

The formula compresses a three-dimensional shortest-path problem on a sphere into a fixed scalar pipeline: coordinate differences, two squared half-angle terms, inverse-haversine recovery, and radius scaling. The practitioner need not construct planes through the sphere's center, calculate Cartesian radius vectors, identify their great circle, or integrate an arc for every pair.

It also separates three kinds of error that are often conflated. Input error concerns coordinates, datum, and units. Numerical error concerns finite-precision evaluation, especially values of \(a\) near zero or one. Model error concerns replacing the relevant surface by a sphere. A result can be numerically excellent and geodetically inadequate, or model-appropriate and numerically corrupted. Keeping the roles separate makes method selection and debugging tractable.

For batch work, the same signature enables early decisions: precompute radian latitudes and cosines; reject invalid coordinates; choose a single radius policy; vectorize the accumulator; clamp only within a documented tolerance; and escalate requests requiring spheroidal accuracy. The abstraction therefore manages both mathematical and operational complexity.

Abstract Reasoning¶

The point-pair relation is a specialization of spherical trigonometry. If the north pole and the two positions form a spherical triangle, the included angle at the pole is \(\Delta\lambda\), the relevant side lengths are related to the co-latitudes, and rearranging the spherical law of cosines with \(1-\cos x=2\sin^2(x/2)\) yields the haversine form.

The range of \(a\) provides immediate diagnostics. Exact arithmetic gives \(a=0\) precisely for coincident spherical positions and \(a=1\) for antipodes. Values strictly between them correspond monotonically to \(0<\delta<\pi\). Therefore a materially negative value or a value materially above one signals invalid inputs or an implementation error; a tiny excursion such as \(1+2\epsilon\) may be clamped as roundoff.

The formula's short-distance conditioning follows from scale. For small \(\delta\), \(a=\sin^2(\delta/2)\approx\delta^2/4\). The small quantity is accumulated directly from small squared sines. By contrast, the cosine form first computes \(\cos\delta\approx1-\delta^2/2\) and then subtracts the result from one implicitly through inverse cosine; rounding near one can erase the separation. Near antipodes, however, \(1-a\) is small, so no finite-precision spelling eliminates all sensitivity. A robust implementation treats both endpoints deliberately.

Metric reasoning follows on the declared sphere: non-negativity, symmetry, identity of indiscernibles, and triangle inequality belong to great-circle distance. The displayed coordinate formula is an evaluator of that metric, not the reason those axioms hold. This distinction supports the proposed placement under Metric without confusing one computational chart with the abstract concept of metric.

Knowledge Transfer¶

The formula transfers literally across practices that share spherical angular coordinates and great-circle separation. A navigator comparing waypoints, an astronomer measuring angular separation after converting coordinates to a common spherical frame, and a simulation measuring nodes on a spherical shell instantiate the same roles. Only the radius, coordinate frame, and acceptable approximation change.

It also transfers as a software contract. A geospatial database, scientific library, embedded navigation routine, or spreadsheet can expose the same inputs, central-angle intermediate, radius policy, and error boundaries. Test cases at coincidence, a quarter circle, the antimeridian, a near-zero separation, and antipodes remain meaningful across implementations.

What does not transfer literally is the idea that “haversine” means any distance between coordinate-like records. Sequence dissimilarity, graph distance, and planar Euclidean distance may all instantiate Metric, but they have no latitude cosines, half-angle spherical identity, or great-circle invariant. Their relation is through the parent abstraction, not through this formula.

Examples¶

Quarter-circle check. Take two points on the equator at longitudes \(0\) and \(\pi/2\). Both latitudes are zero, so

\[ a=0+1\cdot1\cdot\sin^2(\pi/4)=\tfrac12, \qquad \delta=2\operatorname{atan2}(1/\sqrt2,1/\sqrt2)=\frac\pi2. \]

The distance is \(\pi R/2\), one quarter of the sphere's circumference. This fixture catches wrong angle units, missing half angles, and incorrect radius scaling.

One degree at the equator. With \(\Delta\lambda=1^\circ=\pi/180\) and equal zero latitudes, \(a=\sin^2(\pi/360)\approx0.0000761524\), \(\delta=\pi/180\), and for \(R=6371.0088\) km, \(d\approx111.195\) km^[5]. The numerical result is a spherical-model output; it should not be relabeled an exact WGS 84 ellipsoidal distance.

Antimeridian equivalence. Points at longitudes \(179.9^\circ\) and \(-179.9^\circ\) differ physically by \(0.2^\circ\), not \(359.8^\circ\). Wrapping the difference improves clarity and supports downstream bearing calculations. The haversine term is periodic and gives the same scalar distance for equivalent differences, but explicit wrapping keeps the coordinate contract auditable.

Antipodal boundary. For \((\varphi,\lambda)=(0,0)\) and \((0,\pi)\), \(a=1\), \(\delta=\pi\), and \(d=\pi R\). The distance is unique; a shortest great-circle path is not. If roundoff produces \(a=1+\epsilon\), clamping to one is appropriate after validating the inputs.

Non-example—road mileage. Feeding two city coordinates into the formula returns idealized spherical surface separation. It does not include roads, borders, traffic, elevation, or permitted routes and therefore cannot answer a travel-distance query.

Structural Tensions¶

Speed versus Earth-model fidelity. A sphere makes evaluation simple and fast; an ellipsoid better represents geodetic distance. Diagnostic: state the tolerance and compare a representative worst-case pair against a trusted ellipsoidal solver before adopting the spherical shortcut.

Short-distance stability versus antipodal sensitivity. The half-angle form preserves small separations well, while \(1-a\) loses relative resolution near antipodes. Diagnostic: include both regimes in tests and use a method with documented all-range behavior when the application depends on nearly antipodal accuracy.

Convenient default radius versus semantic ambiguity. A mean radius makes an API easy to call but hides which sphere is being measured. Diagnostic: record the radius, units, and policy in the result's provenance.

Scalar certainty versus route nonuniqueness. The distance metric can be definite even when direction or path is not, especially at antipodes. Diagnostic: do not infer azimuth or route from the distance-only result.

Defensive clamping versus error concealment. Clamping tiny floating-point drift protects inverse functions; clamping a grossly invalid accumulator masks bad data or code. Diagnostic: clamp only within a declared tolerance after range and finiteness checks.

Structural–Framed Character¶

The Haversine Formula is strongly structural–framed. Its structure is the typed chain from two spherical positions through a dimensionless accumulator and principal central angle to an arc length. Its invariants, limiting cases, and numerical failure regions are mathematically testable. A different symbol set or programming language leaves the identity intact.

The frame is nonetheless indispensable: sphere, angular coordinates, spherical trigonometry, radians, radius, great circles, and finite-precision inverse functions. Removing those commitments leaves only the broad idea of measuring separation, already represented by Metric. The candidate is therefore domain-specific, not prime.

Structural Core vs. Domain Accent¶

The structural core is pairwise measurement under an invariant distance rule: choose two objects, evaluate a symmetric nonnegative separation, return zero only at coincidence, and preserve the geometry's triangle inequality. That core belongs to Metric.

The domain accent supplies every distinctive obligation of this node: latitude and longitude, periodic longitude difference, spherical reference radius, haversed half angles, latitude cosine weighting, principal central-angle recovery, great-circle arc scaling, short-distance conditioning, antipodal sensitivity, and sphere-versus-ellipsoid error. These are not decorative examples of Metric; together they determine whether the computation is valid.

The subtraction test therefore leaves an autonomous residual. Subtracting Metric does not tell a practitioner how coordinates enter, why half angles improve small-separation computation, which branch to select, how radius controls units, or when an ellipsoidal algorithm is required. Reconstructing those rules would restate the Haversine Formula.

The formula specializes Metric: its output is the geodesic distance on a sphere expressed through one coordinate relation. Metric supplies the general axiomatic meaning of distance, while the Haversine Formula supplies the spherical evaluator and its model contract.

It is also related to Representation because latitude and longitude encode points in a periodic angular chart, and to Numerical Stability because algebraically equivalent forms can behave differently under floating-point rounding. Those explanatory relations do not require extra taxonomic parents. Inverse Trigonometric Functions is a domain-specific neighbor because the recovery step selects a principal angle, but it neither subsumes nor is subsumed by the whole formula.

Relationships to Other Abstractions¶

Current abstraction Haversine Formula Domain-specific

Parents (1) — more general patterns this builds on

Haversine Formula is a kind of Metric Prime

The formula specializes Metric: its output is the geodesic distance on a sphere expressed through one coordinate relation.

Hierarchy path (1) — routes to 1 parentless root

Haversine Formula → Metric → Function (Mapping)

Neighborhood in Abstraction Space¶

Haversine Formula sits in a sparse region of the domain-specific corpus (82^nd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Wiechel Projection — 0.88
Hammer Projection — 0.88
Trilateration — 0.81
World Geographic Reference System — 0.80
Karlsruhe Metric — 0.80

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Law of haversines: a more general relation among sides and angles of spherical triangles; the point-pair formula is a coordinate special case.
Spherical law of cosines: an algebraically equivalent point-pair relation whose direct inverse-cosine evaluation can lose short-distance resolution in finite precision.
Vincenty's formulae: iterative ellipsoidal direct and inverse geodesic methods, not the spherical haversine relation.
Karney's geodesic algorithms: robust high-accuracy algorithms for an ellipsoid of revolution, used when the spherical model is inadequate.
Chord length: straight-line separation through ambient three-dimensional space rather than surface arc length.
Rhumb-line distance: distance along a constant-bearing curve, generally longer than the great-circle route.
Euclidean distance on a map projection: a planar measurement whose distortion depends on the projection and location.
Cosine distance: a vector-similarity dissimilarity in many data-science contexts; the shared word “cosine” does not make it spherical surface distance.
Travel or network distance: a constrained route length rather than a free spherical geodesic.

References¶

[1] Sinnott, R. W. “Virtues of the Haversine”. Sky & Telescope, 1984. The short numerical note that popularized the haversine form for its retention of precision at small angular separations, demonstrated by computing the Mizar-Alcor separation on a low-precision machine. Bibliographic caution: the page is given as 158 by NASA ADS and as 159 by other citation sources; the article's References list currently says 159. registry ↩

[2] Burkard, Richard K. Geodesy for the Layman. Defense Mapping Agency Technical Report TR 80-003, Fourth Revision, 16 March 1984, 1984. The official geodesy primer's statement of the model choice: the sphere is mathematically simple and satisfactory for many purposes, but measurement of long distances spanning continents and oceans requires the ellipsoid of revolution. registry ↩

[3] Karney. “Algorithms for geodesics”. Journal of Geodesy, 2012. Supplies the accurate and always-converging direct and inverse geodesic solutions on an ellipsoid of revolution — round-off error under 15 nanometres — against which any spherical shortcut is judged; the paper itself does not evaluate the spherical approximation. registry ↩

[4] PostGIS Project Steering Committee and contributors. “ST_DistanceSphere”. PostGIS 3.6 Manual, 2026. The reference documentation states the trade-off at the API surface: ST_DistanceSphere 'uses a spherical earth and radius derived from the spheroid defined by the SRID' and is 'faster than ST_DistanceSpheroid, but less accurate'. registry ↩

[5] Moritz, Helmut. “Geodetic Reference System 1980,”. Journal of Geodesy, 2000. The reference system whose defining parameters (a = 6378137 m, 1/f = 298.257222101) yield the arithmetic mean radius R1 = 6371.0088 km used in the worked example; the value is the IUGG mean radius, not a measured constant. registry ↩