Jerbie.jl

Jerbie.jl is a Julia/Rust tool for automatically improving the floating-point accuracy of numerical expressions, using equality saturation (via egg) to search for mathematically equivalent rewrites and real point-sampling (via rival) to score each candidate's accuracy. It follows the same search+score methodology as Herbie.

Installation

Jerbie.jl is not yet a registered Julia package. Until then, install it directly from source:

using Pkg
Pkg.develop(url = "https://github.com/JuliaSymbolics/Jerbie.jl.git", subdir = "Jerbie.jl")

Jerbie's e-graph search runs through a Rust backend (egg-jerbie/) that must be built locally before use — a prebuilt binary distribution (via BinaryBuilder/Jerbie_jll) is planned but not yet available:

cd egg-jerbie
cargo build --release

Documentation

See the dev documentation for the full API reference and internals guide. It is rebuilt automatically on every push to main.

Relationship to Other Packages

• egg-jerbie — this repo's own Rust crate, a thin FFI layer around the egg equality-saturation library. Jerbie.jl calls into it via ccall to build, saturate, and extract from e-graphs; the rewrite rule set (egg-jerbie/src/herbie_rules.rs) is adapted from Herbie's own rules.
• rival3 — a Rust dependency of egg-jerbie, used to evaluate each rewrite candidate against real-valued sample points to score its floating-point accuracy.
• Herbie — the Racket tool this project's search+score approach is directly modeled on. Herbie solves the same problem (find a more accurate equivalent of a floating-point expression); Jerbie.jl re-implements that methodology for the Julia/Symbolics.jl ecosystem, on top of egg rather than Herbie's own e-graph implementation.

Example

Expressions are written as ordinary Julia code over Symbolics.jl variables:

using Symbolics, Jerbie

@variables x
result = optimize_expr(sqrt(x + 1) - sqrt(x))

This runs the full search (guided by equality saturation) and scores every candidate against 8000 held-out sample points, printing a summary:

*********************************************************************
input: sqrt(1 + x) - sqrt(x)
alternatives:
  [1] 1 / (sqrt(1 + x) + sqrt(x))
  [2] 1 / (1 + sqrt(x))
  [3] sqrt(1 + x) - sqrt(x)
start_score (bits-of-error, original): 29.658448391155197
end_score   (bits-of-error, winner):    0.1634530078147536

The scores are bits-of-error (lower is better) — the winning alternative here recovers about 29.5 bits of accuracy lost to catastrophic cancellation in the original expression, near x = 0.

result itself is a NamedTuple with everything needed to inspect or reuse the run:

julia> propertynames(result)
(:alternatives, :start_score, :end_score, :test_context)

See the documentation for the full walkthrough, including run_improve_with_report and the lower-level e-graph/scoring primitives Jerbie.jl is built from.

Citation

Jerbie.jl builds directly on the following prior work:

@article{2021-egg,
  author     = {Willsey, Max and Nandi, Chandrakana and Wang, Yisu Remy and Flatt, Oliver and Tatlock, Zachary and Panchekha, Pavel},
  title      = {egg: Fast and Extensible Equality Saturation},
  year       = {2021},
  journal    = {Proc. ACM Program. Lang.},
  volume     = {5},
  number     = {POPL},
  articleno  = {23},
  doi        = {10.1145/3434304},
  url        = {https://doi.org/10.1145/3434304}
}

@misc{flatt2021interval,
  author        = {Flatt, Oliver and Panchekha, Pavel},
  title         = {An Interval Arithmetic for Robust Error Estimation},
  year          = {2021},
  eprint        = {2107.05784},
  archivePrefix = {arXiv},
  primaryClass  = {math.NA},
  url           = {https://arxiv.org/abs/2107.05784}
}

@inproceedings{panchekha2015herbie,
  author    = {Panchekha, Pavel and Sanchez-Stern, Alex and Wilcox, James R. and Tatlock, Zachary},
  title     = {Automatically Improving Accuracy for Floating Point Expressions},
  booktitle = {Proceedings of the 36th ACM SIGPLAN Conference on Programming Language Design and Implementation},
  series    = {PLDI '15},
  year      = {2015},
  pages     = {1--11},
  doi       = {10.1145/2737924.2737959},
  url       = {https://dl.acm.org/doi/10.1145/2737924.2737959}
}

Author

@ParthsarthiSingh-glang

AI Assistance

Developed with assistance from Claude (Anthropic). Especially for test generation, parsing, and documentation.