fix(jumpavg): penalize stdev also for size=2

This fix is mainly needed for bisection using PDR values.
The impact on trending is smaller but still beneficial,
as this fix should reduce the amount of false anomalies
for two-band and other unstable tests.

+ Update metadata for 0.4.1 release into PyPI.

Change-Id: Iabab4df50f4c4ad034362820904a237c507fa710
Signed-off-by: Vratko Polak <vrpolak@cisco.com>

1 files changed, 7 insertions, 4 deletions

diff --git a/resources/libraries/python/jumpavg/bit_counting_stats.py b/resources/libraries/python/jumpavg/bit_counting_stats.py
index caece2c8ca..3d1cb8aef0 100644
--- a/resources/libraries/python/jumpavg/bit_counting_stats.py
+++ b/resources/libraries/python/jumpavg/bit_counting_stats.py
@@ -97,10 +97,13 @@ class BitCountingStats(AvgStdevStats):
         if self.size < 2:
             return
         stdev = self.stdev / self.unit
-        # Stdev is considered to be uniformly distributed
-        # from zero to max_value. That is quite a bad expectation,
-        # but resilient to negative samples etc.
-        self.bits += math.log(max_value + 1, 2)
+        # Stdev can be anything between zero and max value.
+        # For size==2, sphere surface is 2 points regardless of radius,
+        # we need to penalize large stdev already when encoding the stdev.
+        # The simplest way is to use the same distribution as with size...
+        self.bits += math.log((stdev + 1) * (stdev + 2), 2)
+        # .. just with added normalization from the max value cut-off.
+        self.bits += math.log(1 - 1 / (max_value + 2), 2)
         # Now we know the samples lie on sphere in size-1 dimensions.
         # So it is (size-2)-sphere, with radius^2 == stdev^2 * size.
         # https://en.wikipedia.org/wiki/N-sphere