[Gretl-devel] Re: Default col argument for matrix creation functions

On Wed, 10 Jul 2019, Sven Schreiber wrote: This is kind of a long story but I'll try to be brief. Executive summary: it's actually not just an internal thing. At one time we attempted to distinguish between NA (= missing value, a stand-in for an unknown value that in principle could have been observed, but was not observed, coded as the maximum double-precision floating point value) and NaN (not-a-number, for instance the result of an invalid arithmetical operation such as taking the log of a negative number). If one could maintain this distinction consistently, then it would be reasonable to maintain that NA * 0 = 0, even though NaN * 0 ought to be NaN (as it is, per IEEE rules). But the distinction is very difficult to maintain and we never actually succeeded. (For example, in the series context, log(-1) gave a result of NA.) Moreover (a) two-way traffic between series and matrices has become a integral part of the modern hansl idiom and (b) we have farmed out much of our matrix computation to openblas, which has no notion of our "NA" definition and would happily treat it as a valid numerical value. We were therefore obliged to police the series/matrix interface at quite heavy cost: NAs in series had to be converted into NaNs in matrices (on pain of producing numerical nonsense), and (debatably) vice versa. So in 2018 we cut through this mess: we redefined gretl's "NA" as NaN. No more policing required; NaNs propagate through calculation as per IEEE rules (any arithmetical operation with NaN as an operand produces NaN as result). Except... we kept the NA*0 = 0 rules for series, and (semi-consistently, as you noted) for multiplication of matrices by scalars. So my question is: should we now accept that we have no good reason for blocking the application of the IEEE rules for NaNs? Allin