Log probability returns complex value

Questions
#1

with pm.Model() as gw_model:
neta = pm.Normal(‘neta’ , mu = 0.1 , sd = 0.2)
mass = pm.Normal(‘mass’ , mu = 25 , sd = 100)
amplitude = 10**-3
Mass = 30
distance = 1
Neta = 0.25
tc=0
phic =0

with gw_model :
phase = 2np.pifrequenciestc + phic + 3/(128 * neta * (np.pimass)(5/3))frequencies(-4/3)
exponent = 1jphase + 1j(np.pi/4)
strain = (amplitude * mass*(5/6))/(distancenp.pi**(2/3)) ((5*neta)/(24))**(-7/6)*np.exp(exponent)
diff = np.linalg.norm(waveform - strain)
likelihood = pm.Normal(‘obs’ , mu = diff
, sd =1 , observed = np.random.normal( 0 , 1, len(waveform)))

The model above is a very simplified model of a gravitational wave signal, I’m looking at using probabilistic programming for parameter estimation. The likelihood function at the most basic revolves around taking the norm of the difference between singals. The example above shows what I’m trying to do however when i evaluate the log probability at any point it returns a complex value.

i.e.
gw_model.logp({‘mass’: 100 , ‘neta’:0.1})
returns
array((-10766.848421086901-16136.450316121165j))

This (I think is leading to me being unable to evaluate the gradients so can’t use VI or sampling) ). I’m new to PyMC3 so may have made a mistake somewhere

the entire code can be found at
https://github.com/rgreen1995/Probabilistic-Programming/blob/master/PyMC3/PyMC_make_model.ipynb
Thanks

#2

I think some of the functions you define returns complex value. Did you check each of the output in numpy?

#3

yes, some of the functions do return complex values but I take the norm of those and feed them into the likelihood funciton. So the likelihood evaluation should just be a normal distribution evaluated at the difference between the signals ?