On Thu, Feb 07, 2002 at 02:31:26PM -0500, Michael Wiktowy wrote: > If I recall my sadistics ... I mean, statistics classes correctly, Hops > per request and Connection rate are not standard distributions so > standard deviation is a misleading parameter to calculate. For example, > ~98% of all values should be within +- 3 stddev of the mean in a > standard/normal distribution. Does that mean that we have >1% negative > hops per request, etc. if mean<3*stddev? Nope.
This is of course true, but sums of equidistributed random variables can be approximated by a normal distribution (see central convergence theorem). So it should work OK to use a normal confidence interval for larger HTL values, and for smaller ones I just decided to do so as well at the lack of another model. Standard deviation is simply an estimate of the square root of the variance, it is not specific to the normal distribution. > These are Poisson processes I believe. Info about them can be found at: > http://www.math.uah.edu/stat/poisson/index.html > Maybe some sort of simple histogram would be a more meaningful > diagnostic tool for these types of parameters (rather than going > overboard and calculating the shape and rate parameter that are > intuitive only to few). Poisson processes are events that occur randomly and independently at any time. Diagnostics values like incomingRequests are probably more or less poisson processes (more or less because they are actually not independent or equally probable to occur at all times), and one could therefore expect to see something of a poisson distribution of the number requests per hour. I do not see, however, how the amount of time taken per hop could be a poisson distribution. -- Oskar Sandberg oskar at freenetproject.org _______________________________________________ Devl mailing list Devl at freenetproject.org http://lists.freenetproject.org/mailman/listinfo/devl
