array-api-signal

Signal processing primitives (windows, filters, integration) for the Python Array API Standard.

array-api-signal is a lightweight extension library that brings scientific computing capabilities—similar to scipy.signal and scipy.integrate—to the Python Array API ecosystem.

It allows you to write signal processing code that runs unmodified on NumPy, PyTorch, CuPy, JAX, and * Dask*.

Why?

The Python Array API Standard unifies basic array operations (sin, matmul, reshape), and array-api-extra provides common utilities (pad, kron).

However, scientific domains often need higher-level primitives like Simpson's rule integration, **window generation **, or smoothing filters. array-api-signal fills this gap, providing a single namespace that:

1. Implements missing primitives (like xp.simpson) using backend-agnostic operations.
2. Re-exports array-api-extra functions (like xp.pad).
3. Dispatches standard API calls (like xp.sin) to the underlying backend.

This library is designed to be the "SciPy" companion to array-api-compat's "NumPy".

Installation

pip install array-api-signal

Quick Start: The Unified Namespace

The core philosophy is the Unified Namespace. You rarely need to import numpy or torch directly for your processing logic.

import array_api_signal as aps
import numpy as np
import torch

# 1. Start with any array (NumPy or Torch)
x_np = np.linspace(0, 10, 100)
x_torch = torch.linspace(0, 10, 100, device='cuda')

# 2. Get the unified 'xp' namespace
# This object has EVERYTHING: standard API + signal extensions + extra utils
xp = aps.array_namespace(x_torch)

# 3. Write backend-agnostic code
y = xp.sin(x_torch)                # Calls torch.sin (Standard API)
y = xp.pad(y, (2, 2))              # Calls array_api_extra.pad (Extra API)
area = xp.simpson(y, dx=0.1)       # Calls array_api_signal.simpson (Signal API)

print(f"Area on GPU: {area}")      # Returns a Torch tensor

API Reference

1. Integration (xp.integrate)

Robust numerical integration with support for native backend optimizations (e.g., torch.trapezoid).

• xp.trapezoid(y, x=None, dx=1.0, axis=-1): Composite trapezoidal rule.
• xp.simpson(y, x=None, dx=1.0, axis=-1): Composite Simpson's rule. Matches scipy.integrate.simpson behavior (using the trapezoidal rule for the last interval if samples are even).

2. Signal Processing (xp.signal)

Filtering and convolution tools. Dispatches to scipy.ndimage (NumPy) or torch.nn.functional (Torch) for maximum performance.

• xp.gaussian_filter1d(input, sigma, ...): 1D Gaussian smoothing.
• xp.convolve(in1, in2, mode='full'): General 1D convolution (via FFT).

3. Window Functions (xp.windows)

Backend-agnostic window generation. If you use the unified xp object, these functions automatically return arrays on the correct backend.

• xp.hann(M, sym=True): Hann window.
• xp.hamming(M, sym=True): Hamming window.
• xp.tukey(M, alpha=0.5, sym=True): Tukey (tapered cosine) window.

4. Special Functions (xp.special)

Backend-native special functions the Array API standard omits.

• xp.i0e(x): Exponentially scaled modified Bessel function I_0(x) * exp(-|x|).
• xp.ndtr(x): Standard normal CDF Phi(x).
• xp.logsumexp(x, axis=None, keepdims=False): Overflow-safe log(sum(exp(x))).
• xp.interp(x, x_ref, f_ref): Edge-clamped 1-D linear interpolation (matches numpy.interp, including infinite f_ref values).

Backend support: logsumexp and interp have portable fallbacks and work on every backend. i0e and ndtr dispatch to scipy.special (NumPy) or torch (PyTorch); they have no portable Array API formulation, so on other backends (JAX, CuPy, Dask) they raise NotImplementedError.

5. Extensions from array-api-extra

We re-export widely used extensions so you only need one import.

• xp.pad
• xp.sinc
• xp.cov
• xp.kron

Contributing

We welcome contributions! If you have implemented a scipy function (like a specific filter, window, or transform) using only standard Array API primitives, please submit a PR.

Running Tests

We use pytest to verify correctness against scipy (Gold Standard) across all installed backends.

# Run tests
pytest tests/

# Test specific backend
pytest tests/ --backend=torch