tolerance

class brainstate.environ.tolerance(*, env=None)

Get numerical tolerance based on current precision.

This function returns an appropriate tolerance value for numerical comparisons based on the current precision setting.

Parameters:

env (EnvironmentState | None) – The environment state to query. If None, uses the global environment.

Returns:

Tolerance value as a scalar array.

Return type:

Array

Examples

>>> import brainstate.environ as env
>>> import jax.numpy as jnp
>>>
>>> # Different tolerances for different precisions
>>> env.set(precision=64)
>>> tol64 = env.tolerance()
>>> print(f"64-bit tolerance: {tol64}")  # 1e-12
>>>
>>> env.set(precision=32)
>>> tol32 = env.tolerance()
>>> print(f"32-bit tolerance: {tol32}")  # 1e-5
>>>
>>> # Use in numerical comparisons
>>> def are_close(a, b):
...     return jnp.abs(a - b) < env.tolerance()

Using custom environment:

>>> import brainstate.environ as env
>>>
>>> custom_env = env.EnvironmentState()
>>> env.set(precision=64, env=custom_env)
>>> print(env.tolerance(env=custom_env))  # 1e-12

Notes

Tolerance values: - 64-bit: 1e-12 - 32-bit: 1e-5 - 16-bit and below: 1e-2