Talk:Averages/Mean angle

Accuracy

Huh, the mean angle of 90, 180, 270 and 360 is "-90.0 degrees"? Doesn't that sound a little fruity? --Ledrug 23:31, 9 July 2012 (UTC)

Well, if done exactly, the mean of those complex numbers should be 0, and then the angle would be, well, undefined, or perhaps whatever default the library picks (like 0). But what is actually happening in this case is that the mean of those complex numbers is not calculated to be exactly 0, due to floating-point error when converting between degrees and radians, and then converting between rectangular and polar coordinates. It is a very small number that is very close to 0

>>> print sum(rect(1, radians(d)) for d in [90, 180, 270, 360])
-2.28847549044e-17j

The angle is then calculated from whatever nonzero components that this small non-zero number happens to have. Getting an angle from a mean this small is unstable, because any tiny change (perhaps by a different order of calculations, resulting in different behavior of the error) would result in a radically different angle. --Spoon! 05:35, 10 July 2012 (UTC)

Hmm - I did rush this as I was very tired. This draft task might start a little shakily but I did try to gather the essence of something that had not occurred to me before last night and I intend to tidy-up so please criticise away and make/suggest improvements. I am feeling quite sheepish about having computed average times (another circular mean), for so long and not been actively aware of what I had been doing. --Paddy3118 06:55, 10 July 2012 (UTC)

Which reminds me that I should work on a similar task for mean times. --Paddy3118 06:55, 10 July 2012 (UTC)

It's not just a question of numerical stability, the real problem is that "average of directions" is a bad concept. "Mean" is inherently tied to summation, but summing directions is just ridiculous. What happens when you add east to north, do you get west, south, or north by northeast maybe?

Besides--it's moot--to get an angle out of a vector, atan(y/x) is no good: you need atan2(y, x). --Ledrug 07:42, 10 July 2012 (UTC)

Not so ridiculous. If you have flashes at a distance that you think are localized then you could measure their bearing and want some numerical number that indicates where they are likely to be coming from. An average of the individual, circular, bearings would help. I have to read a bit more about opposite directions but there was something about the size of the radius of the resultant for me to grasp. --Paddy3118 08:28, 10 July 2012 (UTC)