cinx

cinx is a minimal, non-nonsense library for differentiable and fast cubic spline interpolation in JAX.

It's designed to be fully JAX-native, supporting GPU/TPU/CPU execution, and allowing you to backpropagate through your interpolation steps for seamless integration into your machine learning workflows.

Usage

cinx provides a straightforward interp function that handles both single-variable and multi-variable interpolation. Its pure JAX design means you can trivially use JAX transformations like jax.vmap for batch processing and jax.grad for gradient calculations.

Basic 1D Interpolation

import cinx
import jax.numpy as jnp

xp = jnp.linspace(0, 1, 5)
fp = jnp.sin(xp * jnp.pi)
x = jnp.linspace(0, 1, 100)

f = cinx.interp(x, xp, fp)

Multi-variate Interpolation

import jax.numpy as jnp
from cinx import interp

theta = jnp.linspace(0, 2 * jnp.pi, 5)
X_data = jnp.stack((theta, jnp.sin(theta)), axis=-1)
fp = jnp.dot(X_data, rot_matrix(jnp.pi / 4).T)

xp = jnp.linspace(0, 1, len(fp))
x = jnp.linspace(0, 1, 100)

f = interp(x, xp, fp)

Batch Interpolation

import jax
import jax.numpy as jnp
from cinx import interp

theta = jnp.linspace(0, 2 * jnp.pi, 5)
rot_matrix = lambda a: jnp.array([[jnp.cos(a), -jnp.sin(a)], 
                                  [jnp.sin(a), jnp.cos(a)]])
X_data = jnp.stack((theta, jnp.sin(theta)), axis=-1)

angles_batch = jnp.linspace(0, 2 * jnp.pi, 10)
fp = jnp.stack([jnp.dot(X_data, rot_matrix(a).T) for a in angles_batch], axis=0)

xp = jnp.linspace(0, 1, len(fp[0]))
x = jnp.linspace(0, 1, 100)

interp_vmap = jax.vmap(interp, in_axes=(None, None, 0))
f = interp_vmap(x, xp, fp)

Installation

cinx can be installed from PyPI via pip:

pip install cinx

Citation

You don't have to, but if you use cinx in your research and need to reference it, please cite it as follows:

@software{alonso_zdyb_cinx,
  author = {Alonso, Albert and Zdyb, Frans},
  title = {cinx: Minimal Cubic Spline Interpolation in JAX},
  url = {https://github.com/alonfnt/cinx},
  version = {0.0.1},
  year = {2025}
}