May 21st, 2015, 5:22 am
Quote1. The article is implicitly suggesting that not all what-if cases have been taken care of; is there a nasty parameter value lurking in the wings?No nasty parameters that I am aware of.Quote2. The algorithm is iterative (== recursive?) so convergence is not deterministic nor always assured? see the remarks on page 11.Yes, "iterative" might be the better word. I used "recursive" because Marsaglia did so in his paper on Phi. Convergence is discussed in the JSS version (the referee rightly demanded it)Quote3. Assumptions are implicit/unflagged, e.g. in the C code if (rho > 0.99).Yes, there are some cutoff/branching points, and in the paper I mentioned that they might be optimized.Quote4. How robust is the code? (IEEE 754).For the paper, I tried it in quad-double precision (using the QD library). Mike Staunton also sent me a VBA/Excel implementation. In general, if you re-compute the constants and adapt the check against the upper bound, it should work in other settings as well.Of course, your algorithm might be more flexible, and faster. At the time I wrote the paper, my main motivation was to gain deeper understanding of the bivariate normal distribution (it occurs a lot in CreditMetrics-type models and their parameterization, often in disguise). The algorithm popped up as a bonus.