Error in computing probability of exp and geometric distribution #1141
Description
Exp and Geometric distributions are defined on the positive axis of numbers. When the value is less than zero, the probability (including pmf, pdf, and cdf) of these distributions should be zero. However, currently tfp didn't consider the case that the value is less than zero and give some unreasonable results. For example,
>>> geo = tfp.distributions.Geometric(probs=0.5)
>>> geo.prob(-2.5) # a prob which is larger than one
<tf.Tensor: shape=(), dtype=float32, numpy=4.0>
>>> exp_dist = tfp.distributions.Exponential(0.5)
>>> exp_dist .cdf(-2.5) # a prob which is less than zero
<tf.Tensor: shape=(), dtype=float32, numpy=-2.490343> There should be some special treatments for this case.