A question about Metropolis-Hastings sampling

[fancyerii]

I have a question about this line of code:

            #MH ratio
            self.alpha[i] = self.proposal_pdf(x_new) / self.proposal_pdf(x_curr) #q(x|x')/q(x'|x)

proposal_pdf is:

    def proposal_pdf(self, x):
        #q(x) = N(x; mu, Sigma)
        return multivariate_normal.pdf(x, self.proposal_mu, self.proposal_sigma)

q(x|x') = proposal_pdf(x_new).

According to the book:

it seems q(x|x') = N(x; x', Sigma) = multivariate_normal.pdf(x, x_new, self.proposal_sigma)

[vsmolyakov]

Thanks! You are right, I updated the code in the following commit: 1500392

It looks like the proposal_pdf was always sampling from the same self.proposal_mu mean, I updated it so that the mean is a parameter, in other words:
q(x|x') = N(x; x', Sigma) = multivariate_normal.pdf(x_curr, x_new, self.proposal_sigma)
This resulted in a slightly higher MH acceptance ratio: 0.1674 (new) compared to 0.1586 (old).
Thanks!

[vsmolyakov]

resolved