psStats validation tests

2013-10-17

Statistic	czw working trunk r36222M
robust
prior version

2013-10-16b

Trunk version robust median had two issues. First, the cumulative bins were set as (effectively) identical to those of the histogram bins. This is a mistake. Cumulative bounds denote the value that is greater than NUM input values; histogram bounds denote the value that is smaller than all values in that bin. Secondly, the fitLinearSearchForYThenReturnBin was fitting against the midpoint of the cumulative bins. This is not correct for the cumulative distribution. We know only the values of the CDF at those bound values, and so that is what we should be fitting.

There is still an error in the sigma values for this algorithm which is not resolved by this modification.

Statistic	czw working trunk r36222M
robust

2013-10-16

For these tests, I generated 1e5 sample distributions. The mean of the sample distributions was set to 0.0, and the sigma value randomly drawn from the range [0:100]. For each of these distributions, I drew 1e5 samples, and placed them in a psVector type, and passed this to psVectorStats to calculate the sample mean/stdev, clipped mean/stdev, robust median/stdev, and fitted mean/stdev. In addition, the first row shows the "direct" measurements, the mean and standard deviation values calculated directly from the sample values.

Statistic	ipp-20130712 tag		czw working trunk r36222
direct vs input
sample
clipped
robust
fitted