treeweave 0.0.0

Python bindings for the treeweave piecewise-polynomial approximator

treeweave Python bindings

Python (nanobind) wrapper for the treeweave piecewise-polynomial function approximator.

Installation

pip install scikit-build-core nanobind numpy
pip install .          # regular install
# or for development:
pip install -e . --no-build-isolation

Quick start

import math
import numpy as np
import treeweave

# Fit exp(x) on [0, 1] to 1e-10 relative tolerance.
def func(x):
    return math.exp(x[0])   # x is a (dim,) ndarray

approx = treeweave.fit(func, 0.0, 1.0, tol=1e-10)   # dim & out_dim inferred

# Scalar eval
print(approx(0.5))          # ≈ exp(0.5)

# Batch eval
xs = np.linspace(0, 1, 1000)
ys = approx(xs)                  # shape (1000,)
ys = approx(xs, sorted=True)     # 1-D ascending fast path

API

treeweave.fit(f, a, b, tol, *, dim=None, out_dim=None, dtype="f64", ...)

Fits callable f and returns a callable TreeweaveFunction.

• a, b: domain corners (scalars for 1D, length-dim lists otherwise).
• tol: approximation tolerance (drives adaptive tree refinement).
• dim: input dimension; inferred from len(a) (scalar corners ⇒ 1) when omitted.
• out_dim: number of output components; inferred by probing f once at the box midpoint (np.asarray(f(mid)).size, scalar ⇒ 1) when omitted.
• dtype: "f64" (default) or "f32".

The C ABI auto-selects a register-optimal leaf polynomial degree per detected CPU; accuracy is controlled entirely by tol.

TreeweaveFunction

The fitted object is called — there are no named eval methods. A point gives a point result; a batch gives a batch result; two optional keyword flags select alternate batch modes.

Call / property	Description
fn(x)	Evaluate. Point: scalar (dim==1) or (dim,) → scalar / (out_dim,). Batch: (N,) (dim==1) or (N, dim) → (N,) / (N, out_dim)
fn(x, sorted=True)	1-D ascending-batch fast path (requires dim == 1)
fn(x, transposed=True)	Batch returning (out_dim, N) (requires out_dim > 1)
.dim	Input dimensionality
.out_dim	Output dimensionality
.dtype	"f64" or "f32"
.memory_usage	Bytes used by the approximation
.print_stats()	Print internal tree statistics

A point whose length != dim, a batch whose column count != dim, sorted=True with dim != 1, or transposed=True with out_dim == 1 each raise ValueError rather than silently mis-shaping the result.

Running tests

pip install pytest
pytest tests/ -q