When the start points are not provided the init function generates them by drawing from the prior. However the theta_mu values are drawn from a normal with the prior mean (theta_mu_mean) and $\sigma$ the diagonal of the prior theta_mu_var.
The start points should instead be sampled from a normal with $\mu$ = theta_mu_mean and $\sigma$ = sqrt(diag(theta_mu_var))
When the start points are not provided the init function generates them by drawing from the prior. However the theta_mu values are drawn from a normal with the prior mean (theta_mu_mean) and $\sigma$ the diagonal of the prior theta_mu_var.
The start points should instead be sampled from a normal with $\mu$ =
theta_mu_mean
and $\sigma$ =sqrt(diag(theta_mu_var))