Extended precision math, accurate and performant
66 Stars
Updated Last
1 Year Ago
Started In
May 2018


Copyright © 2015-2023 by Jeffrey Sarnoff.

This work is released under The MIT License.

For multiprecision numerical computing using values with 25..2,500 digits. With arithmetic and higher level mathematics, this package offers you the best balance of performance and accuracy.

This package uses the Arb C Library, and adapts some C library interface work from Nemo (see below). Here is a presentation by the designer and architect of the Arb C library.

Travis Build Status   Docs   Docs


ArbNumerics exports three types: ArbFloat, ArbReal, ArbComplex. ArbFloat is an extended precision floating point type. Math using ArbFloat is expected to be very near the veridical value, and often is the closest value for the precision in use. ArbReal is an interval-valued quantity formed of an ArbFloat (the midpoint) and a radius. Math functions with ArbReal are assured to enclose the veridical value. This assurance extends to multiple function applications. ArbComplex is an ArbReal pair (real, imaginary). The same enclosure assurance applies.

While the bounds of an ArbReal or ArbComplex are available, the default is to show these values as digit sequences which almost assuredly are accurate, in a round to nearest sense, to the precision displayed. Math with ArbFloat does not provide the assurance one gets using ArbReal, as an ArbFloat is a point value. While some effort has been taken to provide you with more reliable results from math with ArbFloat values than would be the case using the underlying library itself, ArbReal or ArbComplex are suggested for work that is important to you. ArbFloat is appropriate when exactness is not required during development, or with applications that are approximating something at increasing precisions.


pkg> add Readables
pkg> add ArbNumerics


When updating ArbNumerics, do pkg> gc to prevent accruing a great deal of unused diskspace.


using ArbNumerics or, if you installed Readables, using ArbNumerics, Readables


If you want to work with bit-level precision, first do setextrabits(0).

Otherwise, some extra bits are used to assist with printing values rounded to the last digit displayed. You can find out how many extra bits are used with extrabits(). If you want to change the number of extra bits used, call setextrabits with the desired number of extra bits.

You can set the internal working precision (which is the same as the displayed precision with setextrabits(0)) to a given number of bits or a given number of decimal digits:

setworkingprecision(ArbFloat, bits=250), setworkingprecision(ArbReal, digits=100)

The type can be any of ArbFloat, ArbReal, ArbComplex. All types share the same precision so interconversion makes sense.

You can set the external displayed precision (which is the the same as the internal precision with setextrabits(0)) to a given number of bits or a given number of decimal digits:

setprecision(ArbFloat, bits=250), setworkingprecision(ArbReal, digits=100)

The type can be any of ArbFloat, ArbReal, ArbComplex. All types share the same precision so interconversion makes sense.

Using ArbFloat

Reading the sections that follow gives you a good platform from which to develop.

  • consider using ArbNumerics, Readables
julia> ArbFloat(pi, digits=30, base=10)

julia> readable(ans)

information about using Readables


Initially, the default precision is set to 106 bits. All ArbNumeric types use the same default precision. You can change this to e.g. 750 bits: setprecision(ArbFloat, 750) or setprecision(ArbReal, 750) or setprecision(ArbComplex, 750). Change one default the others follow automatically. This is done to preserve internal consistency. Sometimes more than one type is used within a function. The minimal precision allowed is 24 bits. There is no maximum. The underlying C library calculates more rapidly than BigFloat at any precision.

The precision in use may be set globally, as with BigFloats, or it may be given with the constructor. For most purposes, you should work with a type at one, two, or three precisions. It is helps clarity to convert precisions explicitly, however, it is not necessary.

Constructors using the default precision

julia> a = ArbFloat(3)

julia> b = ArbReal(pi)

julia> c = one(ArbComplex)
1.0000000000000000000000000000000 + 0im

Constructors using a specified precision

julia> BITS = 53;

julia> a = sqrt(ArbFloat(2, BITS))

julia> b = ArbReal(pi, BITS)

julia> c = ArbComplex(a, b, BITS)
1.414213562373095 + 3.141592653589793im
julia> DIGITS = 78;

julia> ArbFloat(pi, bits4digits(DIGITS))

julia> DIGITS == length(string(ans)) - 1 # (-1 for the decimal point)

changing precision

julia> a = ArbFloat(2, 25)
julia> a = ArbFloat(a, 50)

julia> precision = 25
julia> a = ArbFloat(2, precision)
julia> precision = 50
julia> a = ArbFloat(a, precision)


julia> a = sqrt(ArbFloat(2))

julia> b = ArbReal(a)

julia> c = ArbComplex(a, b)
1.414213562373095048801688724210 + 1.414213562373095048801688724210im

julia> Float64(a)

julia> Float32(b)

julia> Float16(c)

Consider using ArbReals instead of ArbFloats if you want your results to be rock solid. That way you can examine the enclosures for your results with radius(value) or bounds(value). This is strongly suggested when working with precisions that you are increasing dynamically.


arithmetic functions

  • +,-, *, /
  • square, cube, sqrt, cbrt, hypot
  • pow(x,i), root(x,i) where i is an integer > 0
  • factorial, doublefactorial, risingfactorial
  • binomial

elementary functions

  • exp, expm1, log, log1p
  • sin, cos, tan, csc, sec, cot
  • asin, acos, atan, atan(y,x)
  • sinh, cosh, tanh, csch, sech, coth
  • asinh, acosh, atanh

arithmetic-geometric mean

  • agm, agm1

gamma functions

  • gamma, lgamma
  • rgamma, digamma

error functions

  • erf, erfc, erfi

Bessel functions

  • besselj, besselj0, besselj1
  • bessely, bessely0, bessely1
  • besseli, besselk

Airy functions

  • airyai, airyaiprime
  • airybi, airybiprime

elliptic integrals

  • elliptic_e, elliptic_k
  • elliptic_p, elliptic_pi
  • elliptic_zeta, elliptic_sigma
elliptic integrals of squared modulus
  • elliptic_e2, elliptic_k2
  • elliptic_p2, elliptic_pi2
  • elliptic_zeta2, elliptic_sigma2

elliptic functions

  • weierstrass_p, weierstrass_invp
  • weierstrass_zeta, weierstrass_sigma

hypergeometric functions

  • hypgeom0f1, hypgeom1f1, hypgeom1f2
  • hypgeom0f1reg, hypgeom1f1reg, hypgeom1f2reg (regularized)

other special functions

  • ei, si, ci
  • shi, chi
  • zeta, eta, xi # Reimann
  • lambertw, polylog

linear algebra

  • dot (vectors)
  • det, tr, inv (matrix)
  • * (matrix multiply)
  • see docs for more functions


  • dft, inverse_dft
  • see docs for use



  • midpoint, radius
  • upperbound, lowerbound, bounds
  • upperbound_abs, lowerbound_abs, bounds_abs


  • setball(midpoint, radius)
  • setinterval(lobound, hibound)


  • midpoint, radius = ball(x::ArbReal)
  • lobound, hibound = interval(x::ArbReal)

working with intervals

The radii are kept using an Arb C library internal structure that has a 30 bit unsigned significand and a power-of-2 exponent that is, essentially, a BigInt. All radii are nonnegative. From the Arb documentation:

The mag_t type holds an unsigned floating-point number with a fixed-precision mantissa (30 bits) and an arbitrary-precision exponent ..., suited for representing magnitude bounds. The special values zero and positive infinity are supported, but not NaN. Operations that involve rounding will always produce a valid bound, For performance reasons, no attempt is made to compute the best possible bounds: in general, a bound may be several ulps larger/smaller than the optimal bound.

When constructing intervals , you should scale the radius to be as small as possible while preserving enclosure.

a caution for BigFloat

julia> p=64;setprecision(BigFloat,p);

julia> ArbFloat(pi,p+8)

julia> ArbFloat(pi,p),BigFloat(pi)
(3.141592653589793238, 3.14159265358979323851)

julia> [ArbFloat(pi,p), BigFloat(pi)]
2-element Array{ArbFloat{88},1}:

The Arb C library

  • Arb is a C library for rigorous real and complex arithmetic with arbitrary precision. Fredrik Johansson is Arb's designer and primary author., with contributions from others.

  • Arb tracks numerical errors automatically using the midpoint-radius representation of an interval.

  • Arb is designed to provide evaluands that contain the veridical numerical result.

  • Arb uses algorithms with provable error bounds for multiprecision mathematical functions.

  • The code is thread-safe, portable, and extensively tested. The library outperforms others.


This work develops parts the Arb C library within Julia. It is entirely dependent on Arb by Fredrik Johansson and would not exist without the good work of William Hart, Tommy Hofmann and the Nemo.jl team. The libraries for Arb and Flint, and build file are theirs, used with permission.


For a numeric types like Float64 and ComplexF64 with about twice their precision, Quadmath.jl exports Float128 and ComplexF128. For almost as much precision with better performance, DoubleFloats.jl exports Double64 and ComplexDF64. ValidatedNumerics.jl and other packages available at JuliaIntervals provide an alternative approach to developing correctly contained results. Those packages are very good and worthwhile when you do not require multiprecision numerics.


  • To propose internal changes, please use pull requests.
  • To discuss improvements, please raise a GitHub issue.