Understanding the initial parameter in jax.nn.softmax

[ayaka14732]

I have difficulty understanding the initial parameter in jax.nn.softmax:

    initial: The minimum value used to shift the input array. Must be present
      when :code:`where` is not None.

Is there a formula for that?

[jakevdp]

The where argument lets you pass a mask specifying which values should contribute to the aggregation. When you do this, there is the possibility that where contains all False values, in which case the return value is ambiguous. For this reason, you must set initial when you specify where, so that there is a non-ambiguous return value for this case.

This follows the API of numpy aggregations like max and min, which have the same requirements.