What is the best way to sample uniformly from an purely open or closed interval.

[abhinavprao]

jax.random.uniform allows two keywords - minval (inclusive) and maxval (exclusive) to specify the scope of the form [minval, maxval) for the unifrom distribution.

What is the best way to achieve purely open scope (minval, maxval) or purely closed scope [minval, maxval]? My network completely breaks at the boundaries, so I would like to avoid it.

Do we have to rely on limiting values/precision of the dtype?

xs = jax.random.uniform(key, 1000, minval = minx+TINIEST_FLOAT, maxval=maxx)

[jakevdp]

You can see how jax.random.uniform is implemented here: https://github.com/google/jax/blob/16b3f00e42122d6be07a5bb1f8025fe4e3d1dd47/jax/_src/random.py#L425-L436

floats here are floating point numbers with an exponent of zero, lying between 0 (inclusive) and 1 (exclusive): they are then shifted to be between minval (inclusive) and maxval (exclusive).

Generating uniform random numbers with an exclusive minimum value would not be easy to do directly: it would essentially involve generating all mantissas except zero, which is difficult to do in a non-iterative fashion given source randomness coming from random bits.

The easiest way to do this will probably be using a variation of what you suggested at the end of your question; for example with minval = jnp.nextafter(minx, jnp.inf).

[abhinavprao]

Thanks! that makes sense