rmath-py 0.1.5

Drop-in numerical accelerator for Python, powered by Rust.

Drop-in numerical accelerator for the Python computing ecosystem.

---

RMath is a high-speed accelerator that offloads heavy mathematical workloads to Rust, seamlessly integrating back into Python via PyO3. Array operations, linear algebra, statistics, calculus, autodiff, and signal processing all execute outside the GIL on a Rayon thread pool.

Drop-in accelerator, not a replacement. Interop with NumPy, PyTorch, JAX, pandas, and scikit-learn via zero-copy NumPy bridges.

Install

pip install rmath-py

Pre-built wheels are available for Windows, Linux, and macOS. No Rust toolchain required.

Architecture

rmath
│
├── Core Types
│   ├── Scalar          — atomic f64 math unit
│   ├── Vector          — 1D optimized operations
│   ├── Array           — N-dimensional compute engine
│   ├── Tensor          — autodiff-enabled (forward + reverse mode)
│   ├── Dual            — forward-mode automatic differentiation
│   └── LazyArray       — memory-efficient large-scale data
│
├── Math Domains
│   ├── linalg          — LU, QR, Cholesky, SVD (via faer)
│   ├── stats           — descriptive + inferential statistics
│   ├── geometry        — 3D transforms, quaternions, convex hull
│   ├── signal          — FFT, convolution, spectral analysis
│   └── special         — gamma, beta, error functions
│
├── ML & Autodiff
│   ├── nn              — activations, loss, normalization
│   ├── Tensor          — reverse-mode gradient tracking
│   └── Dual            — forward-mode differentiation
│
├── Utilities
│   ├── constants       — mathematical and physical constants
│   └── loop_range      — lazy pipeline engine
│
└── Interop
    ├── NumPy           — from_numpy / to_numpy
    ├── PyTorch         — from_torch / to_torch
    ├── JAX             — from_jax / to_jax
    └── pandas          — from_dataframe / to_dataframe

Modules

Module	Description
rmath.array	N-dimensional array with automatic storage tiering (stack / heap / mmap)
rmath.vector	1-D parallel engine — trig, reductions, sorting, filtering, complex numbers
rmath.scalar	Precision f64 math — 80+ functions mirroring Python's math module
rmath.linalg	Matrix solvers (LU, QR, Cholesky, SVD) via faer
rmath.stats	Descriptive and inferential statistics — Welford's algorithm, distributions, regression
rmath.calculus	Automatic differentiation (dual numbers), numerical integration, root-finding
rmath.geometry	3D transforms, quaternions, convex hull
rmath.signal	FFT, convolution, spectral analysis
rmath.nn	Activation functions (GELU, Softmax), loss, normalization layers
rmath.special	Gamma, beta, and error functions
rmath.constants	Mathematical and physical constants

---

Domain Examples

1. Scientific Computing / Numerical Analysis

Solving a linear system + residual validation

import rmath as rm

A = rm.Array([[4.0, 2.0], [1.0, 3.0]])
b = rm.Array([[1.0], [2.0]])

x = rm.linalg.solve(A, b)

# Validate: residual should be ~0
residual = A.matmul(x).sub(b).norm_frobenius()
print("Solution:", x)
print("Residual:", residual)

Eigendecomposition + reconstruction check

A = rm.Array([[2.0, 1.0], [1.0, 2.0]])

eigvals, eigvecs = rm.linalg.eigh(A)

# Reconstruct: A = V * Lambda * V^T
Lambda = rm.Array.zeros(2, 2)
Lambda[0, 0] = eigvals[0]
Lambda[1, 1] = eigvals[1]

reconstructed = eigvecs.matmul(Lambda).matmul(eigvecs.t())
error = A.sub(reconstructed).norm_frobenius()
print("Reconstruction error:", error)  # ~1e-16

---

2. Data Science / Data Analysis

Load, extract columns, correlate

import rmath as rm

data = rm.Array([[25, 50000], [30, 60000], [22, 45000], [35, 80000]])

ages = rm.Vector(data.get_col(0))
income = rm.Vector(data.get_col(1))

print("Mean age:", ages.mean())
print("Income std:", income.std_dev())

corr = rm.stats.correlation(ages, income)
print("Correlation:", corr)  # 0.98

---

3. Statistics / Research

Descriptive statistics

import rmath as rm

v = rm.Vector([2, 4, 4, 4, 5, 5, 7, 9])

print("Mean:", v.mean())          # 5.0
print("Variance:", v.variance())  # 4.57
print("Std Dev:", v.std_dev())    # 2.14

Hypothesis testing

group1 = rm.Vector([20, 22, 19, 24, 25])
group2 = rm.Vector([30, 29, 35, 32, 31])

t_stat, p_value = rm.stats.t_test_independent(group1, group2)
print("t-stat:", t_stat)   # -6.12
print("p-value:", p_value) # 0.001

Linear regression

x = rm.Vector([1, 2, 3, 4, 5])
y = rm.Vector([2, 4, 5, 4, 5])

result = rm.stats.linear_regression(x, y)
print(f"y = {result['slope']}*x + {result['intercept']}")
print(f"R² = {result['r_squared']}")

---

4. Financial / Economic Analysis

Return analysis

import rmath as rm

prices = rm.Vector([100, 102, 101, 105, 110])
diffs = prices.diff()  # [2, -1, 4, 5]
print("Price changes:", list(diffs))

Covariance matrix

data = rm.Array([
    [0.01, 0.02, 0.015],
    [0.03, 0.01, 0.02],
    [0.02, 0.025, 0.03],
])

cov = data.covariance()
print("Covariance matrix:", cov)  # 3x3

---

5. Machine Learning (Autograd)

Gradient computation

import rmath as rm

x = rm.Tensor([1.0, 2.0, 3.0], requires_grad=True)

y = (x * x).sum()
y.backward()

print("Gradients:", x.grad)  # [2.0, 4.0, 6.0]

Simple neural step

w = rm.Tensor.randn(3, requires_grad=True)
x = rm.Tensor([1.0, 2.0, 3.0])

y_pred = (w * x).sum()
target = rm.Tensor([10.0])
loss = ((y_pred - target) * (y_pred - target)).sum()
loss.backward()

print("Loss:", loss.data.to_flat_list()[0])
print("Gradients:", w.grad)

Built-in activations

x = rm.Array([-1.0, 0.0, 2.0])
print(x.relu())  # [0.0, 0.0, 2.0]

---

6. Calculus / Differentiation

Forward-mode autodiff (dual numbers)

import rmath.calculus as rc

x = rc.Dual(2.0, 1.0)  # value=2, seed=1

y = x.sin() * x.exp()

print("f(2)  =", y.value)       # 6.72
print("f'(2) =", y.derivative)  # 3.64

Numerical integration

import rmath as rm

result = rm.calculus.integrate_simpson(lambda x: x * x, 0, 1, 100)
print("Integral of x² from 0 to 1:", result)  # 0.3333...

---

7. Signal Processing

1D Convolution (FFT-accelerated)

import rmath as rm

signal = rm.Vector([1, 2, 3, 4])
kernel = rm.Vector([1, 0, -1])

filtered = rm.signal.convolve(signal, kernel, "full")
print(list(filtered))  # [1, 2, 2, 2, -3, -4]

FFT

signal = rm.Vector([1, 0, 0, 0])

fft_result = rm.signal.fft(signal)
print("Magnitudes:", list(fft_result.to_mags()))  # [1, 1, 1, 1]

---

8. Geometry

Distance & similarity

import rmath as rm

a = rm.Vector([1, 2, 3])
b = rm.Vector([4, 5, 6])

dist = rm.geometry.euclidean_distance(a, b)
cos_sim = rm.geometry.cosine_similarity(a, b)

print("Distance:", dist)            # 5.196
print("Cosine similarity:", cos_sim) # 0.975

---

9. Interoperability

Scikit-Learn Drop-in (Zero-Copy via __array__)

import rmath as rm
from sklearn.linear_model import LinearRegression

# 1. Generate data entirely in RMath (Rust)
X = rm.Array.randn(100, 1)  # 100 samples, 1 feature
Y = rm.Array.randn(100, 1)  # 100 targets

# 2. Pass RMath arrays natively into Scikit-Learn (Python)
model = LinearRegression()

# This "Just Works" because RMath natively exposes the __array__ protocol!
model.fit(X, Y)

print("Scikit-Learn R² Score:", model.score(X, Y))

NumPy roundtrip

import numpy as np
import rmath as rm

np_arr = np.array([[1.0, 2.0], [3.0, 4.0]])
rm_arr = rm.Array.from_numpy(np_arr)
back = rm_arr.to_numpy()

PyTorch bridge

import torch
import rmath as rm

t = torch.tensor([[1.0, 2.0]])
rm_arr = rm.Array.from_torch(t)
back = rm_arr.to_torch()

pandas integration

import pandas as pd
import rmath as rm

data = rm.Array([[1, 2], [3, 4], [5, 6]])
df = data.to_dataframe(columns=["x", "y"])
back = rm.Array.from_dataframe(df)

---

Performance

Benchmarked on Windows (CPython 3.13, AMD64). Medians of 20 runs, 3 warmup.

Vector (1-D) — 167/167 tests passed

Operation	Size	Speedup vs Python
sum_range	100K	6,076x
norm_l1	100K	142x
std_dev	100K	119x
dot	100K	43x
sin (elementwise)	100K	5.8x
sort	100K	4.6x

Average speedup: 50x over pure Python.

Array (N-D) — 161/161 tests passed

Operation	Size	Speedup vs NumPy
transpose	500x200	65x
from_numpy	500x200	38x
gelu	500x200	18x
tanh	500x200	5x
matmul	200x200	competitive

Average speedup: 3.2x over NumPy.

Tensor (Autograd) — 30/30 tests passed

Operation	Size	Speedup vs PyTorch
add (forward)	200x200	8.7x
reshape	200x200	6.6x
mul (forward)	200x200	6.2x
sigmoid (forward)	200x200	6.1x
add (backward)	200x200	5.1x
training_step	100x100	3.2x

Average speedup: 3.98x over PyTorch.

Phase 3: Intelligence & Fusion (v0.1.5) 🚀

The latest release introduces Single-Pass Parallel Kernels for optimizers and elementwise math.

Component	Operation	Gain vs Eager/PyTorch
Adam	.step()	3.3x faster vs PyTorch
SGD	.step() (with momentum)	2.0x faster vs PyTorch
Linear Fusion	(x * 2 + 1) * 3	2.3x faster
Fused Reduction	sum(sin(x))	1.2x faster

Unified Lazy Engine (Loop Fusion)

RMath now supports deferred execution for both memory-based and disk-based arrays. Chain your operations with .lazy() to execute them in a single parallel pass through memory.

import rmath.array as ra

# 1. In-Memory Loop Fusion (3 passes -> 1 pass)
a = ra.randn(2000, 2000)
result = a.lazy().mul(2.0).sin().exp().execute()

# 2. Disk-Streaming Fusion (Math applied during load)
result = ra.LazyArray.open("data.csv").sigmoid().add(1.0).load()

Real-World Data Pipeline — rmath vs NumPy (v0.1.5)

Benchmarked on Windows (CPython 3.13, AMD64). 5 million row financial dataset.

Pipeline Step	rmath Time	rmath Mem	NumPy Time	NumPy Mem	Speedup
Data Generation	0.30s	153 MB	1.31s	137 MB	4.3× faster
Data Cleaning	0.15s	0.5 MB	0.17s	4.8 MB	1.1× faster
Feature Engineering	0.07s	76 MB	0.12s	76 MB	1.8× faster
Descriptive Stats	0.26s	0.07 MB	0.25s	0.03 MB	Comparable
Correlation Analysis	0.038s	0.04 MB	0.43s	0.13 MB	11.2× faster
Segmentation	0.47s	120 MB	0.93s	38 MB	2.0× faster
Linear Signal	0.16s	0.03 MB	0.13s	0.00 MB	NumPy slight edge

rmath wins on 5 of 7 pipeline stages. Data cleaning uses 9× less memory than NumPy (0.5 MB vs 4.8 MB) thanks to zero-allocation filter_where. Full benchmark scripts in benchmarks/pipeline/.

Numerical Accuracy

Algorithm	Module	Guarantee
Kahan compensated summation	Vector + Array	O(eps) error regardless of N
Welford's online variance	Vector + Array	Single-pass, no catastrophic cancellation
Parallel Kahan	Array (N >= 8K)	Chunked Kahan + merge, same accuracy as serial

---

How it works

Python ─── PyO3 FFI ──> Rust core (rayon + faer)
                            |
                 +----------+----------+
                 v          v          v
              Stack       Heap       Mmap
            (<=32 f64)  (Arc-shared) (lazy I/O)

• GIL-free: All Vector reductions, operators, norms, sorting, and statistics release the GIL via py.allow_threads(). Tensor division backward pass runs in pure Rust with no Python::with_gil re-entry.
• Storage tiering: Vectors with 32 or fewer elements live on the stack (zero allocation). Larger vectors use Arc<Vec<f64>> for cheap cloning.
• Parallelism: Rayon parallel iterators activate at 8,192 elements (unified threshold across Vector and Array). Below that threshold, serial iterators avoid thread-pool overhead.
• Autograd: Tensor reads data through Arc<RwLock> with no deep clones on .shape, .dtype, forward ops, or backward passes.
• Interop: to_torch() and to_jax() route through NumPy arrays (single memcpy) instead of N individual Python float allocations.
• Type stubs: Full .pyi stubs ship with the package for IDE autocompletion and type-checking.

Documentation

Full API reference: ay-developerweb.github.io/rmath/portal/

Author

Ayomide Adediran (@Ay-developerweb)

• GitHub: Ay-developerweb
• Email: [e-mail]

Contributing

RMath is built in Rust (src/) and exposed to Python via PyO3.

• Rust source: src/ — core numerical engines
• Python stubs: rmath/*.pyi — type annotations
• Benchmarks: benchmarks/ — automated performance suite

License

MIT