There are some problems about the inverse Gaussian (Wald distribution)

[J.Cui]

There are some problems about the inverse Gaussian process. When I estimate the parameters of an oil leakage process as an inverse Gaussian process, that is, the leakage increment obeys the inverse Gaussian distribution,

with pm.Model() as model:
    w = pm.Uniform('w', lower=0, upper=100)
    k = pm.Uniform('k', lower=0, upper=100)
    q = pm.Uniform('q', lower=0, upper=20)
    lamb = pm.Uniform("lamb", lower=0, upper=1000)
    deltaLam = pow(time_data[2], q) - pow(time_data[1], q)

    mu = pm.TruncatedNormal("mu", mu=w, tau=k ** 2,lower=0)
    mu_delta = mu * deltaLam
    la = mu ** 3 * deltaLam ** 2 / lamb

    Y_obs = pm.Wald(
        "Y_obs", mu=mu_delta, lam=la, observed=data)
    start = pm.find_MAP(fmin=optimize.fmin_powell)
    step = pm.Metropolis()
    trace = pm.sample(2000, step=step, chains=3, tune=1000, cores=1)

pm.traceplot(trace)
plt.show()
for RV in model.basic_RVs:
    print(RV.name, RV.logp(model.test_point))
print(model.logp(model.test_point))

[J.Cui]

[J.Cui]

[J.Cui]

[J.Cui]

[ricardoV94]

Looks like you are hitting numerical precision issues, possibly due to an extreme combination of data and/or parameters.

Can you narrow down which value(s) from your observations leads to an underflow of -inf for the likelihood?

[J.Cui]

Thank you！ I found some problems with my data. After changing the data, there is no -inf