One or both of us has got hold of the wrong end of some sticks here;-(

Brian Bessley wrote:

> But in any case it's a moot point because the (input) "phase space" values 
> aren't integers,

The checksum refers to input FFT values (not the "phase space" inputs to the 
IFFT).

> and adding (FFT run length) strictly positive values 
> together is likely to cause a loss of precision even if they were. (Run 
> length = 2^20 is likely to lose about 20 bits of precision in the sum, 
> compared with the precision of the inputs, if you're thinking in fixed point 
> terms.

Now have mentioned a simple example of error combination which I wanted
to use, namely 2^20 irrational numbers which should sum to an integer.

With 20 binary places, the maximum error in the representation of each number
is ~2^-20.
If all the errors were the same sign then our "integer" result may be in 
jeapardy.
But this is not the case, and random walk/Central Limit Theorem arguments tell
us that the total arror is normally distributed about zero with standard 
deviation
2^-10. (The variance of the sum is the sum of the variances).

BTW if we are worried about the integer overflowing, we can always "balance"
the representaion by subtracting 2^power as described before.

> When working in floating point, you won't actually get a wraparound 
> type error like integer overflow, but you do lose the least significant bit 
> of the mantissa every time you add two values with the same sign and the same 
> exponent as the exponent of the result has to increase.)

This is why fixed point is the natural method for addition, while floating point
is appropriate for multiplication.

> 
> In fact, as the FFT run length increases, the checksum criterion becomes 
> increasingly poor as a predictor of accuracy compared with the roundoff error 
> check criterion, as the imprecision of the checksum is proportional to the 
> logarithm of the number of elements in the computation whereas the roundoff 
> error increases much more slowly.

To use the checksum, it seems obvious to use as much precision as necessary to
obtain an exact integer.

> 
> Also, the "special features" of the transform used for the L-L test give us 
> the "subtract 2" part of the L-L algorithm for free; this may interfere in 
> some way with the exact equality mentioned in the basic theory.

Both the subtract 2 and modulo operations may be trivially performed in
"real space".

David Eddy


_________________________________________________________________
Be one of the first to try Windows Live Mail.
http://ideas.live.com/programpage.aspx?versionId=5d21c51a-b161-4314-9b0e-4911fb2b2e6d
_______________________________________________
Prime mailing list
[email protected]
http://hogranch.com/mailman/listinfo/prime

Reply via email to