On Sat, 17 Dec 2016, Allin Cottrell wrote: > I've added in git a function named ecdf which takes a series or vector > argument and returns the empirical CDF in the form of a two column matrix > with unique sorted values of the input in column 1 and cumulative relative > frequency in column 2. > > This is not very difficult to do in hansl, but it's substantially faster in > C (the differential depending on the dimensions of the problem) and I think > it may be worth having. Test script below. > > <hansl> > function matrix hansl_ecdf (const matrix M) > matrix ret = values(M) ~ 0 > scalar n = rows(M) > loop i=1..rows(ret) -q > ret[i,2] = sumc(M .<= ret[i,1])/n > endloop > return ret > end function > > scalar N = 3000 > matrix M = zeros(N, 1) > loop i=1..N -q > M[i] = randint(1, 200) > endloop > > scalar t1=0 > scalar t2=0 > > loop 500 -q > set stopwatch > matrix ec1 = ecdf(M) > t1 += $stopwatch > matrix ec2 = hansl_ecdf(M) > t2 += $stopwatch > endloop > > printf "built-in: %.3fs\n", t1 > printf "hansl: %.3fs\n", t2 > </hansl> > > hansl_ecdf() could probably be improved upon, but it seems like a "natural" > solution for hansl users. Output on the machine I'm at: > > built-in: 0.204s > hansl: 1.610s > > (Besides, the built-in version handles both series and vectors, and > automatically drops NAs or NaNs from the calculation.) > > If I've missed a reason why this is redundant (which sometimes happens!) > please let me know. Otherwise I'll go ahead and document it. The following is, in fact, a bit faster than yours, but still slower than C (unsurprisingly); but, as you say, it doesn't support matrices, although it'd be noce to extend aggregate to do so <hansl> function matrix hansl_ecdf(series x) ret = aggregate(const,x) a = cum(ret[,2]) scalar n = a[rows(a)] ret[,2] = a ./ n return ret end function </hansl>