m replied to --CELKO--
03-Feb-10 09:34 PM
Why would you expect that? The average of a set of values can frequently be
finite and more precise than the input data. Consider that the average of 2
& 3 is 2 1/2. Since the size of the data set, in this case two, is an
exact value, and following propagation of error formulae, you can see that
there is a gain of significance in this case - this is common and expected.
The behaviour of SQL server is implementation specific, and may be hardware
dependent too (although nearly all platforms have now converged their FP &
BCD etc. behaviour). If you are really interested in more detail, then
doing some research into numerical error theory will help tremendously. If
not, then just cast the result to the desired precision and use it.