Sure ! I had understand before when you told me about that, but i am understand more and more about this problem of precision for multiple recursion with some power on them. Solution if i am right : center the data on the time axis and normalise their value the nearer from [-1;1].
Just one question, i have seen you use often iniatialize with "X=Cum (0)" instead of "X=0". It seems the same (X=0, put 0 on all the bar of X). It is because it react differently if QuickAFL is on/off (SetBarsRequired) and with Cum(0) you are sure to fill all the bars ? --- In [email protected], "Fred" <[EMAIL PROTECTED]> wrote: > > Possibly you understand now why in Gaussian Elimination I chose to > arbitrarily set the midpoint along the x-axis to zero and keep the > numbers along the scale in fairly tight range, and perform the calcs > in double precission ... yes ? > > --- In [email protected], "Tom Tom" <michel_b_g@> wrote: > > > > So after further investigation. > > > > > > I code loop in VBS double. Result is not the same than SUM, nor > FOR. > > So SUM or FOR are not exact, but they act not the same in > rounding. I thing > > the arythmetic order to calculate is not the same in FOR and in > SUM so > > rounding are different, so result are different. > > But one or other, result is not accurate. > > If I want to stay in AFL and I had to choose between SUM and FOR, > wich one I > > will go ? As they are both innacurate, SUM seems better ? > > > > > > I test max/min number for AFL after reading the link from Paul : > > http://en.wikipedia.org/wiki/IEEE_754 > > > > Code : > > > > ieee = -1e10; > > _TRACE("IEEE:"+ ieee); // to show exponent > > _TRACE("IEEE:"+ NumToStr(ieee,1.9) ); // to show innacurate digit > > > > And try after with : > > ieee = 1e11; > > > > exponant should go from -126 to +127, but in our case : > > ieee = 1e-10; -> result are infinite > > ieee = 1e11; -> ok for direct trace, but with numtostr there is > rounding and > > do not use exponent notation > > ieee = 1e39; -> result are infinite > > > > So my question, wy exponent go from -10 to 39 !? > > And NumToStr(x,1.9), seems to send back rounding result. Cannot it > send back > > with exponent notation and no rounding like direct _trace ? > > > > Second, accuracy seems 7 digit, but it should be 6 digit ? (max > number from > > the mantisse is 2^23 - 1 = 8388607, that give 7 number so six > digit sure). > > Try: ieee = 99999981 > > 7th digit is ok. 8th digit is wrong. > > > > So AB use wich encoding (not IEEE_754 ?) and difference between > direct > > _trace and _trace with NumToStr come from NumToStr wich make > rounding for > > big exponent ? > > > > Cheers, > > Mich. > > > > _________________________________________________________________ > > Windows Live Messenger sur i-modeĀ : dialoguez avec vos amis > depuis votre > > mobile comme sur PC ! http://mobile.live.fr/messenger/bouygues/ > > >
