8
$\begingroup$

I am trying to implement HMC with a non-diagonal mass matrix, but am getting tripped up by some of the terminology.

According to BDA3 and Neal's review, the kinetic energy term (that I guess is always used due to convenience) is

$$ K(p) = \frac{p^T M^{-1} p}{2} \,.$$

This is also recognizable called a multivariate normal with zero mean and covariance matrix $M$. BDA3 (pg 301) says

To keep it simple, we commonly use a diagonal mass matrix, M. If so, the components of φ are independent, with φj ∼ N(0,Mjj) for each dimension j = 1, . . . , d. It can be useful for M to roughly scale with the inverse covariance matrix of the posterior distribution, (var(θ|y))^−1.

(I'm reading N(0, M)) as a multivariate normal with mean zero and covariance M.)

The part tripping me up is where says that that "it can be useful for $M$ to roughly scale with the inverse covariance matrix of the posterior distribution...".

And then also just before that the momentum sample that starts the leapfrog steps ($\phi$) is drawn from a multivariate normal with covariance matrix $M$.

So which is it? To construct a good M for HMC, do I estimate the covariance or precision matrix of the posterior? Even though $M$ is the covariance matrix of the kinetic energy, using an $M$ that is an estimate of the precision matrix of the posterior will yield a more efficient algorithm?

Secondary question: what is the intuition that could guide me here?

  • Do you want to use a precision matrix so that the momentum pushes orthogonally to the potential/posterior to improve mixing?

  • OR do you want the momentum to push towards the high probability mass part of the posterior (because thats where you want to draw most samples from).

p.s. The reason I'm not using the identity matrix for $M$ is because for my problem I happen to able to obtain decent estimate of the covariance matrix of my pretty high dimensional (~1000) posterior beforehand.

$\endgroup$
6
$\begingroup$

A linear transformation of the position variables is equivalent to the inverse linear transformation of the momentum variables. Ideally, you want to sample from a (transformed) distribution whose covariance matrix is the identity matrix, and this is obtained by the transformation indicated above.

For details, there is a nice explanation in Neal's "MCMC using Hamiltonian dynamics", Chapter 5 of the Handbook of Markov Chain Monte Carlo, Section 4.1 ("Effect of linear transformations"). The chapter is available here.

Neal explains:

Suppose we have an estimate, $\Sigma$, of the covariance matrix for $q$, and suppose also that $q$ has at least a roughly Gaussian distribution. How can we use this information to improve the performance of HMC? One way is to transform the variables so that their covariance matrix is close to the identity, by finding the Cholesky decomposition, $\Sigma = LL^T$, with $L$ being lower-triangular, and letting $q^\prime = L^{−1}q$. [$\ldots$]

An equivalent way to make use of the estimated covariance $\Sigma$ is to keep the original $q$ variables, but use the kinetic energy function $K(p) = p^T \Sigma p/2$ — ie, we let the momentum variables have covariance $\Sigma^{−1}$. The equivalence can be seen by transforming this kinetic energy to correspond to a transformation to $q^\prime = L^{−1} q$ (see equation (4.1)), which gives $K(p^\prime) = (p^\prime)^T{M^\prime}^{−1}p^\prime$ with $M^\prime = (L^{−1}(LL^T)(L^{−1})^T)^{−1} = I$.

To give some intuition, suppose that the target pdf is cigar-shaped pointing in one direction which is not axis-aligned. You can either rotate and rescale the space, so that the cigar becomes a ball, and then draw momenta from a unit multivariate normal, or equivalently you can keep the original space and draw your momenta so that they are aligned with the cigar (e.g., with most of the velocity along the major axis of the cigar, so that you can explore it quickly).

$\endgroup$
  • $\begingroup$ I must confess that I read that (a few times) and then decided to ask this question... If you can connect Neal's explanation there with my questions you will get all the points $\endgroup$ – bill_e Aug 10 '16 at 4:28
  • $\begingroup$ @bill_e: Your first question is fully answered above: "we let the momentum variables have covariance $\Sigma^{-1}$". So, yes, $M = \Sigma^{-1}$. The answer to the second question is that you want to send $p \rightarrow L^T p$ so that the distribution of the momenta is aligned to the axes of the covariance matrix (which is equivalent to reparametrize the original space so that the covariance matrix is the identity). $\endgroup$ – lacerbi Aug 10 '16 at 4:35
  • $\begingroup$ Posted steps of an implementation of this, just to be crystal crystal clear. Is it right? Very specifically the inverse (or lack of) on $\hat{\Sigma}$. Yes, the cigar thing is making sense. I'll think about it more. $\endgroup$ – bill_e Aug 10 '16 at 4:58
4
$\begingroup$

A simple way to see why $M$ should be the inverse covariance of the distribution you want to sample from is to consider sampling from a multivariate Normal with aribitrary mean $\mu$ and covariance $\Sigma$. In this case the Hamiltonian equations of motion can be solved exactly (i.e., no need for leapfrog integration). Now, for $M=\Sigma^{-1}$ two magic things happen: (i) the equations of motion for each coordinate decouple from the rest, and (ii) the matrices $\Sigma$ and $M$ cancel each other and disappear from the equations of motion. The solution is a set of oscilators with equal frequency, which can be argued to yield the fastest possible mixing. See some details in eqs (2.31)-(2.35) here.

In a generic distribution this approach would be just an approximation.

$\endgroup$
0
$\begingroup$

Salient part of linearly transforming momentum using estimated covariance.

Given an estimate $\hat{\Sigma}$ of the covariance matrix of the posterior HMC is to sample from:

  1. Draw $\phi \sim N(0, \hat{\Sigma}^{-1})$

  2. Simulate Hamiltonian dynamics. (Repeating L times)

    A. Half step: $\phi \leftarrow \phi + \frac{1}{2}\epsilon \frac{d}{d\theta}\mathrm{log}p(\theta \mid y)$.

    B. Full step: $\theta \leftarrow \theta + \epsilon \hat{\Sigma}\phi$.

    C. Half step: $\phi \leftarrow \phi + \frac{1}{2}\epsilon \frac{d}{d\theta}\mathrm{log} p(\theta \mid y)$.

  3. Accept/Reject.

(If this is right, don't upvote this, upvote @lacerbi)

$\endgroup$
  • 1
    $\begingroup$ subbed $M^{-1}$ for $\hat{\Sigma}$ $\endgroup$ – bill_e Aug 10 '16 at 5:14
  • $\begingroup$ Yeah, I think this is right. See for example Eq. (4) of this paper. $\endgroup$ – lacerbi Aug 10 '16 at 15:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.