Removing noise can't be achieved, it is a conjecture. Partial noise reduction is however possible and each application is a specific case. In the attached sheet, we compare the simple finite differences with an 11 points old style filtering, but still very valid. If you play a bit, you will discover that the 11 points has less tendency to distort the underlying data to the detriment of increasing the noise in the top region, which in turn can be isolated and treated otherwise. The naive sigmoidal shape is a good representation of a difficult data set to exploit.
Image filtering, speech recognition ... blabla: all that is BS !
No matter what you do, you "temper". Unintentional tempering is one thing, but where is the border line between "tempering and forgery".
jmG
