hyperreal provides exact rational arithmetic, symbolic real values, and lazy
computable real approximation.
It is useful when code needs more information than an f64 can provide:
structural sign facts, exact zero/nonzero knowledge, exact rational access,
bounded sign refinement, and recognizable forms such as pi, e, square
roots, logarithms, and rational trig constants.
Numeric Model
hyperreal is built around three layers that deliberately keep exact and
symbolic information available before approximation:
Rational: is the exact arithmetic base. It stores arbitrary-precision numerator/denominator values and performs exact reduction, dyadic detection, square extraction, shared-denominator dot products, and exact IEEE-754 import.Computable: is the lazy approximation layer. It represents exact-real expression graphs such as sums, products, inverses, roots, logs, trig kernels, and shared constants. It approximates only when a caller asks for a binary precision, then caches the result and conservative sign/magnitude facts.Real: is the public symbolic scalar. It stores an exact rational scale plus a compact symbolic class and, when needed, aComputablecertificate. Common classes include exact one, powers/products ofpiande, selected square roots, logarithms, trig forms, and factored constant products.RealStructuralFacts: conservative public facts about sign, zero status, magnitude, and exact-rational state.Simple: a small Lisp-like expression parser, enabled by the optionalsimplefeature.
Relationship to Other Crates
realistic_blasuseshyperreal::Realas its default exact/symbolic scalar backend. It forwardshyperrealstructural facts through itsScalartype and adds vector, matrix, transform, and retained-geometry facts around them.liminalcan consumehyperreal::Realdirectly, using structural facts, finitef64approximations, and bounded sign refinement before robust fallback.hypersolveis the experimental solver layer. Its current direction is to evaluate constraints through symbolic references to variables, reuse reductions across iterations, and route repeated residual and geometry kernels throughhyperrealandrealistic_blasinstead of rebuilding scalar expressions from scratch.
hyperreal owns scalar representation and approximation. It does not own vector
or matrix algebra, and it does not decide geometry topology. The stack is
layered intentionally: scalar facts live here, object-level facts live in
realistic_blas/geometry layers, and decision procedures live above them.
Problems This Stack Targets
hyperreal and the surrounding stack are aimed at problems where ordinary
floating-point arithmetic is fast but loses too much information too early.
The target workloads tend to have many cheap structural decisions, a smaller
number of expensive exact or high-precision decisions, and repeated kernels
where cached structure can be reused.
Representative problem families include:
- robust geometric predicates such as orientation, sidedness, intersection, and clearance tests
- CAD-style geometric constraint solving, where many residuals share variables, transforms, and symbolic subexpressions
- PCB routing and autorouting constraints, where exact topology and clearance decisions matter before approximate coordinates are useful
- toolpath planning and manufacturing geometry, where small sign or ordering errors can change topology
- matrix/vector transform stacks with many affine, diagonal, triangular, homogeneous, point, or direction-specialized cases
- exact-rational or symbolic scalar workloads that should remain exact until a caller explicitly asks for digits
The implementation strategy is intentionally thin:
- Preserve exact rational, dyadic, symbolic, sign, zero, magnitude, and approximation-cache facts at the scalar layer.
- Carry inexpensive object facts at higher layers, such as known point/direction
w, affine form, diagonal or triangular matrix structure, retained transform facts, shared determinant/cofactor work, and known coordinate zeros. - Reduce symbolically before approximating. Common forms such as exact
rationals,
pi,e, roots, logarithms, and selected trig constants should simplify or classify without entering a generic approximation kernel. - Defer approximation until a decision or output precision requires it, then cache the result and any conservative sign or magnitude information.
- Prefer deterministic fast paths guarded by cheap facts over speculative probing inside dense loops.
- Keep similar functions flat: a fast path should improve the family it targets without making neighboring functions erratic.
Current State
The crate is active, benchmark-driven, and no longer just a direct port of computable real ideas. Current implementation work includes:
- exact rational and dyadic fast paths
- dedicated identity constructors for common exact ones and zeros
- cached internal constants for
pi,tau,e, common square roots, and common logarithms - symbolic classes for selected
pi,e,sqrt,ln,sin(pi*q), andtan(pi*q)forms - exact trig, inverse-trig, logarithm, exponential, and inverse-hyperbolic shortcuts where the input structure is recognizable
- argument reduction and prescaled kernels for transcendental approximation
- structural sign, zero, nonzero, magnitude, and exact-rational queries
- bounded sign refinement through
sign_untilandrefine_sign_until - cached approximation and structural-fact propagation through computable nodes
- borrowed arithmetic paths for
RationalandReal - shared-denominator and signed-product-sum hooks used by matrix/vector callers to delay rational canonicalization
serdesupport for expression structure, excluding transient caches and abort signals- dispatch tracing and targeted benchmark suites for scalar, approximation, symbolic, adversarial, and stack-facing regressions
- source-level READMEs for
Rational,Real, andComputable, plusstructural_facts.txtfor planned and implemented fact propagation
Installation
[]
= "0.10.6"
With the Simple parser and calculator binary:
[]
= { = "0.10.6", = ["simple"] }
Feature flags:
| Feature | Default | Purpose |
|---|---|---|
simple |
no | Enables Simple and the package calculator binary. |
Examples
Exact Rationals
Rational is the exact base layer. It is useful for imported measurements,
test fixtures, small coefficients, and any value that should not become binary
floating-point noise before the rest of the stack sees it.
use Rational;
use TryFrom;
let a = fraction.unwrap;
let b = fraction.unwrap;
assert_eq!;
// Finite floats import by exact IEEE-754 decoding, not by decimal rounding.
let half = try_from.unwrap;
assert_eq!;
// Decimal and fraction strings parse into exact rationals.
let decimal: Rational = "12.125".parse.unwrap;
let fraction: Rational = "97/8".parse.unwrap;
assert_eq!;
Symbolic Reals
Real keeps a rational scale plus a symbolic/computable class. That lets common
forms simplify or expose facts before approximation is needed.
use ;
let x = new.sqrt.unwrap;
let y = new.sqrt.unwrap;
let z = x * y;
let approx: f64 = z.into;
assert!;
let half = new;
let angle = half * pi;
let cosine = angle.cos.unwrap;
// Recognizable symbolic/trig forms can answer facts without a full equality
// proof or high-precision decimal expansion.
let facts = cosine.structural_facts;
assert_eq!;
assert_eq!;
Computable Approximation
Computable is the lazy approximation layer. It stores an expression graph and
only computes a scaled integer approximation when a precision is requested.
use ;
let x = rational.sin;
let scaled = x.approx;
assert_ne!;
// Sign refinement asks for only enough precision to decide the sign down to a
// requested floor. The result may be `None` for unresolved or truly difficult
// cases, so callers can decide whether to refine further or use a fallback.
let near_pi = pi.add;
assert_eq!;
Structural Facts
Structural facts are conservative certificates. They are designed for filters, predicates, and higher-level kernels that want to avoid approximation unless a decision actually requires it.
use ;
let value = new.sqrt.unwrap;
let facts = value.structural_facts;
assert_eq!;
assert_eq!;
assert!;
assert_eq!;
let exact = new;
let exact_facts = exact.structural_facts;
assert_eq!;
assert_eq!;
assert_eq!;
assert!;
Facts are conservative. Missing sign or magnitude information means the fact was not proven cheaply.
Stack-Facing Filters
The surrounding geometry stack uses hyperreal values as scalar certificates:
try cheap structural facts first, use finite approximation when that is enough,
and only then request bounded refinement.
use ;
let offset = pi - new;
assert_eq!;
This pattern is the intended handoff to realistic_blas and liminal: cheap
facts route the common case, approximation is delayed until useful, and bounded
refinement remains available for hard predicate boundaries.
Simple Expressions
Requires the simple feature.
use Simple;
let expr: Simple = "(* (+ pi pi) (sin (/ 1 5)))".parse.unwrap;
let value = expr.evaluate.unwrap;
let _: f64 = value.into;
Simple supports arithmetic, roots, powers, logs, exponentials, trig, inverse
trig, inverse hyperbolic functions, integers, decimals, fractions, pi, and
e.
It can also be useful for small configuration- or test-facing formulas:
use ;
let expr: Simple = "(sqrt (/ 49 64))".parse.unwrap;
let value = expr.evaluate.unwrap;
assert_eq!;
Conversions
Supported conversions include:
- integer types to
RationalandReal - finite
f32/f64toRationalandRealby exact IEEE-754 decoding Realtof32/f64by approximationReal::to_f64_approx()for borrowed finite approximation used by filters
Float import rejects NaN and infinities. to_f64_approx() returns None
when no finite f64 approximation can be produced.
Finite decimal and fraction strings parse losslessly through Rational; the
parser also accepts leading + signs and digit separators where the rational
parser supports them. Scientific notation is not a supported exact text format.
-0.0 imports as exact rational zero, so IEEE signed-zero information is not
preserved.
PartialEq on Real is structural, not a full computer-algebra equality
proof. Two expressions may print/debug similarly and approximate identically
while still comparing unequal if they were built through different computable
expression histories. Use structural facts, exact-rational extraction, or
explicit approximation/refinement when semantic equivalence rather than
representation identity is the question.
Performance Notes
Performance shortcuts are intentionally documented next to the code that uses them. The main techniques are:
- keep exact rational and dyadic values outside generic computable graphs
- build identity values through dedicated constructors and clone cached named constants instead of rebuilding them
- preserve lightweight symbolic classes only where benchmarks show value
- reduce trig and exponential arguments before entering series kernels
- use endpoint and tiny-argument transforms for inverse trig and inverse hyperbolic functions
- answer structural queries from certificates before refining approximations
- use borrowed arithmetic to reduce expression-graph cloning in callers such as
realistic_blasandliminal
Benchmark suites:
The generated benchmark summary is in benchmarks.md.
Hand-maintained profiling anchors and regression goals for the Rational,
Real, and Computable paths are in PERFORMANCE.md.
Run dispatch tracing separately:
The generated trace summary is in dispatch_trace.md.
Development
Common checks:
When adding a shortcut, add a focused correctness test and a benchmark row for
the smallest affected surface. Keep the shortcut only if it improves the target
without regressing broader realistic_blas or liminal benchmarks.
Provenance and Acknowledgements
hyperreal descends from the
realistic project and continues
that project's interest in practical computable real arithmetic.
Special thanks to siefkenj, whose contributions improved realistic and hyperreal.
References
These are the papers and books which contribute ideas or methods to this crate. They are listed here in MLA style for easy citation.
- Bareiss, Erwin H. "Sylvester's Identity and Multistep Integer-Preserving Gaussian Elimination." Mathematics of Computation, vol. 22, no. 103, 1968, pp. 565-578. American Mathematical Society, https://doi.org/10.1090/S0025-5718-1968-0226829-0.
- Boehm, Hans-Juergen, Robert Cartwright, Mark Riggle, and Michael J. O'Donnell. "Exact Real Arithmetic: A Case Study in Higher Order Programming." Proceedings of the 1986 ACM Conference on LISP and Functional Programming, ACM, 1986, pp. 162-173.
- Boehm, Hans-J. "Towards an API for the Real Numbers." Proceedings of the 41st ACM SIGPLAN International Conference on Programming Language Design and Implementation, ACM, 2020, pp. 562-576.
- Brent, Richard P. "Fast Multiple-Precision Evaluation of Elementary Functions." Journal of the ACM, vol. 23, no. 2, 1976, pp. 242-251.
- Brent, Richard P., and Paul Zimmermann. "Modern Computer Arithmetic." Cambridge University Press, 2010.
- Middeke, Johannes, David J. Jeffrey, and Christoph Koutschan. "Common Factors in Fraction-Free Matrix Decompositions." Mathematics in Computer Science, vol. 15, 2021, pp. 589-608.
- Payne, Mary H., and Robert N. Hanek. "Radian Reduction for Trigonometric Functions." ACM SIGNUM Newsletter, vol. 18, no. 1, 1983, pp. 19-24.
- Shewchuk, Jonathan Richard. "Adaptive Precision Floating-Point Arithmetic and Fast Robust Geometric Predicates." Discrete & Computational Geometry, vol. 18, no. 3, 1997, pp. 305-363.
- Smith, Luke, and Joan Powell. "An Alternative Method to Gauss-Jordan Elimination: Minimizing Fraction Arithmetic." The Mathematics Educator, vol. 20, no. 2, 2011, pp. 44-50.
- Yap, Chee-Keng. "Towards Exact Geometric Computation." Computational Geometry, vol. 7, nos. 1-2, 1997, pp. 3-23.
License
(C) https://github.com/timschmidt Apache-2.0 / MIT
(C) https://github.com/tialaramex/realistic/ Apache-2.0
(C) https://github.com/siefkenj Apache-2.0