Fix #2258

[junpenglao]

Related/same issue including: #2076, #2109

[junpenglao]

@aseyboldt I got something funny when testing, using the following code:

with pm.Model() as model:
    a = pm.Uniform('a', lower=0., upper=1.)
    pm.Uniform('b', lower=0., upper=a)
start = {'a': .7, 'b': .5}
pm.sampling._update_start_vals(start, model.test_point, model)
start

The value in start['b_interval__'] is incorrect (it changes every time the above is excuted), however if I do

    pm.Uniform('b', lower=0., upper=a-0.)

Then the value propagates correctly.
That's what I am currently doing in https://github.com/pymc-devs/pymc3/pull/2328/files#diff-2d8e65f96de72b3dad37dd6025425149R140

a, b = draw_values([self.a-0., self.b-0.],
                                 point=point)

but I feel like I am missing something obvious here... Do you know if there is a better way?

[twiecki]

This LGTM. @junpenglao is the above issue resolved after your recent commit?

[junpenglao]

Yes, there is no issue anymore, it just the implementation is not very pretty and potentially confusing. Maybe I should put a comment to explain why, and a warning not to delete the minus 0.

[twiecki]

Good idea.