emltree 0.1.1

Compile elementary-function formulas into pure EML (Exp-Minus-Log) form

emltree

Compile elementary-function formulas into pure EML (Exp-Minus-Log) form, where the only operator is

eml(x, y) = exp(x) − ln(y)

and the only constant is 1.

Paper

This project is an implementation of ideas introduced in:

Andrzej Odrzywolek. All elementary functions from a single operator. arXiv:2603.21852 (2026). https://arxiv.org/abs/2603.21852

The paper proves that the single binary operator eml(x, y) = exp(x) − ln(y) together with the constant 1 generates the entire scientific-calculator basis (arithmetic, roots, logs, all trig and hyperbolic functions, the constants e, π, i). Every elementary expression becomes a binary tree whose only internal node is eml. This package is the inverse direction: it takes an ordinary formula and compiles it into that tree. A sibling Rust crate, OxiEML, covers the Rust ecosystem; a zero-dependency JS/TypeScript port lives in js/ (npm: emltree); this package is the Python side.

Quick start

uv venv --python 3.11
uv pip install -e ".[dev]"

# Compile a formula
uv run emltree "sin(x) + exp(y)" -v x -v y

# ASCII tree
uv run emltree "log(x)" -v x -f tree

# Numerical sanity check
uv run emltree "sin(x)**2 + cos(x)**2" -v x=0.9 --verify

Library use

from emltree import compile_formula, evaluate

tree = compile_formula("sqrt(x**2 + y**2)", variables=["x", "y"])
print(tree.to_nested())
print(evaluate(tree, {"x": 3.0, "y": 4.0}))   # ≈ 5+0j

The returned tree is an immutable ADT (One / Var / Eml) — walk it, hash it, render as RPN (exp(x) is x 1 E), or ship it into an FPGA/analog circuit as the paper suggests.

Evaluation is vectorised: pass numpy arrays as bindings and the tree evaluates elementwise.

import numpy as np
xs = np.linspace(-2, 2, 1000)
evaluate(tree, {"x": xs, "y": xs})   # array of 1000 values

Why the trees are large

This compiler is compositional, not optimal. Every primitive bottoms out in exp / ln / sub, so sin(x) produces a tree with hundreds of nodes. The paper's direct-search results (Table 4) are vastly shorter — integrating that search is one of the open directions below.

Numerical caveats

• Branch cuts: outside their real domains (asin(2), acosh(-2), log of negatives, …) results flow through complex branch cuts and may land on a non-principal branch — or, where float fuzz compounds, off-sheet entirely (paper §4.1). On the usual real domains everything matches sympy to ~1e-7.
• Addition overflow: add_'s expansion applies exp() to its second operand, so adding values past ~709 overflows float64. Integer/decimal constants avoid this internally (binary decomposition with a multiplicative odd step), but x + y with huge y is an inherent ceiling of the encoding.

Tests

uv run pytest

Contributing

Contributions are very welcome — this is an early-stage package and there's plenty of room to improve. A few directions that would make a great first PR:

• Shorter trees — hand-curated or searched EML identities to replace the naive compositional expansions (see paper Table 4).
• More primitives — abs, sign, floor, ceil, erf, etc. (many require tricks; see the paper's supplementary).
• Better output — LaTeX, GraphViz / Mermaid, Jupyter rich repr.
• Torch evaluation — numpy bindings already vectorise; a torch evaluator would make EML trees differentiable end-to-end.
• Optional Rust backend — PyO3 bindings to OxiEML for expensive search and symbolic regression.
• Docs & examples — walk-throughs of the paper's identities, notebooks showing symbolic-regression use cases.

Please open an issue to discuss before sending a large PR. Bug reports, documentation fixes, and additional test cases are also very welcome — no contribution is too small.

Development

uv venv --python 3.11
uv pip install -e ".[dev]"
uv run pytest

License

MIT — see LICENSE.

Citation

If you use emltree in academic work, please cite both the paper and the software:

@article{odrzywolek2026eml,
  title   = {All elementary functions from a single operator},
  author  = {Odrzywolek, Andrzej},
  journal = {arXiv preprint arXiv:2603.21852},
  year    = {2026}
}

@software{hsiao2026emltree,
  author  = {Hsiao, Wei-Chien},
  title   = {emltree: a Python compiler from elementary formulas to EML trees},
  year    = {2026},
  url     = {https://github.com/gba3124/emltree},
  version = {0.1.1}
}