fxp-jax 0.2.0

Simple fixed-point solver implemented in JAX

Fixed-point solver

FixedPointJAX is a simple implementation of a fixed-point iteration algorithm for root finding in JAX. The implementation allow the user to solve the system of fixed point equations by standard fixed point iterations and the SQUAREM accelerator, see Du and Varadhan (2020).

Installation

pip install fxp-jax

Usage

import jax
import jax.numpy as jnp
from jax import random

from fxp_jax import fxp_root

jax.config.update("jax_enable_x64", True)

accelerator = "squarem"

N = 100

a = random.uniform(random.PRNGKey(111), (N,1))
b = random.uniform(random.PRNGKey(112), (1,1))

def fun(x: jnp.ndarray) -> tuple[jnp.ndarray, jnp.ndarray]:
    y = a + x @ b
    return y, y - x

fxp = fxp_root(
    fun,
)
result = fxp.solve(guess=jnp.zeros_like(a), accelerator=accelerator)

y, z = fxp.fun(result.x)

print('--------------------------------------------------------')
print(f'System of fixed-point equations is solved: {jnp.allclose(result.x, y)}.')
print(f'Roots are zero: {jnp.allclose(z, 0.0)}.')
print('--------------------------------------------------------')