utaal / hdrhist Goto Github PK

The current implementation of the summary() function returns pairs (quantile, value) that are not correct, namely the value returned is the value corresponding to the next bucket. Here more precise example:

If we create a simple HDRHist with as input the following values, repeated each n_times times (tot number of values = 10001):

n_times  value
2400       1       // 100% >= 1  -- 0% < 1
200        2       //  76% >= 2  -- 24% < 2
7400       3       //  74% >= 3  -- 26% < 3   ---> 3 is the 25p
1          5       //0.01% >= 5

CCDF output for this hist from HDRHist is:

value      freq     n_samples
2	    1	        2400     // 100% of values is >= 1, 2400 elements in bucket [1,2[
3	    0.76         200     // 76% of values is >= 2, 200 elements in bucket [2,3[
4	    0.74        7400     // 74% of values is >= 3, 7400 elements in bucket [3,4[
5	    0.0001	   0
6	    0.0001	   1     // 0.0001% of values is >= 5, 1 elements in bucket [5,6[
7	    0.00009        0     // this value is IMO completely redundant

It's clear now that the 25p (0.25q) is the value 3, since we know that 26% of results is < 3 (meaning that also 25% is less than 3). However, in the current implementation of summary(), we will go over our ccdf iterator and pick up the first elements such that element.frequency <= 0.75, and this will make us pick the value 4 instead of the value 3.
We should instead pick up the value from the previous element in the ccdf iterator.