Change boost::math defaults to avoid double to long double propagation

[wds15]

Description

During the PR, #1255, the default policy boost_policy_t has been changed to avoid the propagation of input arguments from double to long double which is the default policy for boost::math. This boost_policy_t affects the boost::math functions in stan::math of

• rising_factorial
• falling_factorial
• digamma
• lgamma

The reason to avoid the propagation to long double is performance. Without the over-propagation the lgamma function runs significantly faster (~3x); the digamma function runs faster as well (2x). Negative binomial models, which use the digamma function, run now 2x faster while giving the same results. Thus, the slightly reduced precision due to avoiding the propagation does not seem to have noticeable speed precision impacts.

Moreover, the boost::math folks are considering themselves to change the default behaviour of boost::math to avoid the over propagation in future releases (see boostorg/math#241).

Therefore, we should consider to avoid the over-propagation for all boost math functions. This can be achieved by adding -DBOOST_MATH_PROMOTE_DOUBLE_POLICY=false to our makefiles.

Example

NA

Expected Output

Faster model inferences which involve boost::math functions

Current Version:

v2.20.0