PyMC3 variables are theano tensors, which means you can not apply numpy/scipy functions on them directly.
The most direct (and many cases slow) way to go around it is to wrap the numpy/scipy function as a theano function: http://deeplearning.net/software/theano/extending/extending_theano.html#as-op
a pymc3 example:
https://github.com/pymc-devs/pymc3/blob/master/pymc3/examples/disaster_model_theano_op.py
The ideal way is to rewrite the numpy/scipy function as a theano function. It could be a bit involve at times, but you can have a look at the doc for an example (which also finding the root of some function): http://docs.pymc.io/advanced_theano.html