latent etc (#72)

* latent etc

* include latent-joint

* update jointof

* docstring

* bump version

Author: cscherrer

Project.toml

@@ -1,7 +1,7 @@
 name = "MeasureBase"
 uuid = "fa1605e6-acd5-459c-a1e6-7e635759db14"
 authors = ["Chad Scherrer <[email protected]> and contributors"]
-version = "0.12.1"
+version = "0.12.2"
 
 [deps]
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"

src/MeasureBase.jl

@@ -127,6 +127,7 @@ include("standard/stdmeasure.jl")
 include("standard/stduniform.jl")
 include("standard/stdexponential.jl")
 include("standard/stdlogistic.jl")
+include("latent-joint.jl")
 
 include("rand.jl")
 

src/latent-joint.jl

@@ -0,0 +1,62 @@
+_LATENT_DOCSTRING = """
+Some Probabilistic Programming Languages (PPLs) like Tilde.jl make a distinction
+between a _latent space_, often a namespace represented as a named tuple, and
+the space containing the return value, which we refer to as the _maifest space_.
+The distinction is that computations are done in terms of the latent space,
+while the resulting value is in the manifest space.
+
+To simplify many manipulations involving these concepts, we introduce the
+concept of a _joint space_. For example, suppose `m()` is a measure with latent
+space `NamedTuple{(:a, :b)}` that returns `a - b`, so the latent value `(a = 3,
+b = 4)` is mapped to the manifest value `0.75`. Then the corresponding value in
+the joint space is the pair `(a = 3, b = 4) => 0.75`.
+
+One of the many goals of probabilistic programming is to blur the line between
+"built in" measures like `Normal()` and those defined in terms of a model from a
+PPL. To accommodate this, we extend these concepts to general measures. 
+
+For many measures, it's convenient to work directly in the manifest space, and
+there's no need for such separation. However, it's important to be able to
+manipulate measures programmatically, with minimal special cases. Because of
+this, we introduce fall-back methods
+
+    latentof(m) = m
+    manifestof(m) = m
+
+The default implementation of `jointof` is then a push-forward through the
+function `x -> (x => x)`. For example,
+
+    julia> rand(MeasureBase.jointof(StdUniform()))
+    0.346439=>0.346439    
+"""
+
+"""
+    latentof(m)
+
+$_LATENT_DOCSTRING
+"""
+latentof(m) = m
+
+"""
+    manifestof(m)
+
+$_LATENT_DOCSTRING
+"""
+manifestof(m) = m
+
+"""
+    jointof(m)
+
+$_LATENT_DOCSTRING
+"""
+function jointof(m)
+    fwd(x) = x => x
+    
+    function back(p::Pair)
+        x,y = p
+        @assert x === y
+        return x
+    end
+
+    PushforwardMeasure(fwd, back, m, NoVolCorr())
+end