Truncated logabsdetjac forward / inverse inconsistency

[penelopeysm]

import Bijectors as B
using Distributions: Normal

dist = truncated(Normal(); lower=6.515552440303498)
fwd = B.bijector(dist)
ivs = B.inverse(fwd)
y = -24.79099078521386

Inverse transform (changed in #325):

julia> B.logabsdetjac(ivs, y)
-24.79099078521386

implemented here

Bijectors.jl/src/bijectors/truncated.jl

Lines 84 to 87 in 8a525f1

	function logabsdetjac(ib::Inverse{<:TruncatedBijector}, y)
	a, b = ib.orig.lb, ib.orig.ub
	return sum(truncated_inv_logabsdetjac.(y, a, b))
	end

Forward transform (not changed, now out of sync):

julia> B.logabsdetjac(fwd, ivs(y))
24.791002373799344

implemented here

Bijectors.jl/src/bijectors/truncated.jl

Lines 51 to 54 in 8a525f1

	function logabsdetjac(b::TruncatedBijector, x)
	a, b = b.lb, b.ub
	return sum(truncated_logabsdetjac.(_clamp.(x, a, b), a, b))
	end

This is causing CI failures on TuringLang/DynamicPPL.jl#853 (and minimised from the failing test there, hence the very specific values above)

[penelopeysm]

Bijectors.jl/src/bijectors/truncated.jl

Lines 56 to 67 in 8a525f1

	function truncated_logabsdetjac(x, a, b)
	lowerbounded, upperbounded = isfinite(a), isfinite(b)
	if lowerbounded && upperbounded
	return -log((x - a) * (b - x) / (b - a))
	elseif lowerbounded
	return -log(x - a)
	elseif upperbounded
	return -log(b - x)
	else
	return zero(x)
	end
	end

with the values above it turns out that

x = 6.515552440320614
a = 6.515552440303498

so x-a is ~ 1e-11, not sure if that is meaningful

[penelopeysm]

Yes it's probably meaningful:

julia> y = -24.79099078521386; a = 6.515552440303498;

julia> log(exp(y))  # reverse transform
-24.79099078521386

julia> log(exp(y) + a - a)  # forward transform
-24.791002373799344