Clarifying Draw Terminology

[gil2rok]

Per my discussion with @WardBrian , in HMC a draw can refer to two distinct concepts:

1. The position variable theta, which denotes a draw from the target distribution
2. The tuple (theta, rho) that contains position and momentum variables accordingly, denoting a draw from the joint distribution.

In Bayes-Kit, the data type DrawAndLogP suggests definition 1 when it is defined here. However, comments from @bob-carpenter in my pull request here suggest definition 2. Additionally, Stan documentation uses the phrase "draw" in HMC to refer to definition 1 here.

My code for the drghmc sampler uses the DrawAndLogP data type, suggesting definition 1, but in its documentation uses a draw to refer to definition 2. I worry this may be confusing.

Lastly, if Bayes-Kit is intended as a pedagogical resource, it may be especially worthwhile to clarify this confusion. Would love to hear other's thoughts on this!

[bob-carpenter]

That's a good point and we should clarify.

• The target density has parameter theta with density p(theta | y), where y is data.
• In HMC, we couple a momentum variable rho of same dimensionality as theta and give rho an independent normal(0, Sigma) distribution where Sigma is either diagonal or dense.
• The joint density is p(theta, rho | y, Sigma) = p(theta | y) * normal(rho | 0, Sigma).
• Thus the parameters of the HMC sampler are (rho, theta)
• A single parameter value and momentum drawn from p(theta, rho) is called a "draw". A sequence (or multi-set) of draws is called a "sample." The literature often overloads "sample" to also refer to a draw, but let's try to be precise and consistent in our doc.
• For user applications, we usually only care about draws from the marginal distribution p(theta). In general with (MC)MC methods, if we have a draw (theta(m), rho(m)) ~ p(theta, rho), where p(theta, rho) propto exp(-H(theta, rho)), then theta(m) is a draw from p(theta).
• So to be precise, we would say (theta(m), rho(m)) is a draw from the joint position/momentum distribution and theta(m) is a draw from the position distribution and hence a draw from the target distribution.

In general, the internal data structures are not relevant for doc and should not be referenced in the doc. That's a developer/programmer detail. What we want to document is the client-facing API. That allows us to fix an API, document it and put tests in place, then later refactor it without changing doc or tests.

[roualdes]

draw from the joint position/momentum distribution

Depending on our desired level of pedantry, some call this joint distribution the canonical distribution.

[bob-carpenter]

The term "canonical distribution" is how the physicists refer to it. The other physics term that's relevant is "phase space," which refers to the coupled (theta, rho) variables. I don't think using either would be helpful for our doc, but I'm also not opposed if you want to mention that's what physicists call these things.