The 64-ulp spread threshold still classified distinct representable
predictors such as [1e15, 1e15 + 1, 1e15 + 2] as constant and dropped
them. A column is now constant only when every raw value is the same
double; the 1e15 case is added to the offset regression test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The small-k product also needs (1 - p)^(n - k) to stay a normal number;
otherwise it returned 0 for BINOM.DIST(15,1100,0.5,FALSE) (exact
2.1367646606291324e-298) and the cumulative form built on it. Those cases
now use the logarithmic formula. Reference grid gets k = 10..25 at n = 1100.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A centred column was dropped when its length fell under 1e-12 of the raw
column, which also removed real variation such as [1e13, 1e13+1, 1e13+2].
Constant columns are now recognised on the raw values, within a few ulps.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* functions: MODE counts through an index, FREQUENCY bins by binary search
MODE/MODE.SNGL/MODE.MULT searched the list of distinct numbers seen so far
for every input, and FREQUENCY scanned the sorted bins for every data value,
so both were quadratic in computed arrays. MODE now keys counts by value
(-0 folded into 0, NaN never merged) and keeps first-occurrence order;
FREQUENCY finds the first bin at or above each value with partition_point
over the order it already sorts. Results are unchanged.
* address review: FREQUENCY keeps NaN data in the overflow bin and NaN bins empty
CodeRabbit pointed out that an error or NaN result passed the accuracy test and
that a zero reference allowed any finite value (the bound was infinite). Errors
now count as infinite error and a zero reference allows 1e-300.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
CodeRabbit pointed out that the drop test used the centred column length, so a
constant column such as [0.1; 3] (centred to about -1.4e-17) was scaled to a unit
vector and kept, giving a meaningless coefficient and wrong df. A centred column
shorter than 1e-12 of its uncentred length is now treated as collinear with the
intercept. The suggested 1e-6 threshold was not used: it would also drop columns
that sit far from zero, such as 1e9 + i.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The mass came from ln_gamma differences (relative error 1.1e-13 at n = 100,
1.5e-11 at n = 10,000) and the cumulative forms summed up to 10,000 of them.
The mass now comes from Loader's saddle-point formula, the cumulative sum
walks away from the mode by the ratio of neighbouring terms and stops when
terms stop mattering, and the incomplete beta / gamma branches take the same
mass as their prefactor.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The normal equations square the condition number, so a degree-5 polynomial on
x = 1..12 came back with errors up to 2.6e-6 and x, x^2 near 1e6 gave slope
2.99976 and intercept 251 for y = 3x + 7. LINEST, LOGEST, TREND and GROWTH now
factor the centred, column-scaled data with Householder QR. Collinear columns
are still dropped under the same relative tolerance.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>