Why does jax.numpy have different trigonometric functions from numpy and math?

[hyunwoooh5]

While I was writing a code in jax, I found that my c++ code and python give different results. Also, I checked other libararies or and mathematica and only jax.numpy give the different values:

Is there any reason for this?

[jakevdp]

JAX defaults to 32-bit computation (see http://go/jax-sharp-bits#double-64bit-precision), while math, numpy, and others default to 64-bit computation. If you enable 64-bit computation in JAX, then you should see that the results match:

import jax
jax.config.update('jax_enable_x64', True)

import math
import numpy as np
import jax.numpy as jnp

print(f"{jnp.cos(3.0):.15f}")
print(f"{math.cos(3.0):.15f}")
print(f"{np.cos(3.0):.15f}")

-0.989992496600445
-0.989992496600445
-0.989992496600445