Jerry Avins <
[email protected]> writes:
> Randy Yates wrote:
>> Jerry Avins <
[email protected]> writes:
>>
>>> Randy Yates wrote:
>>>> "bharat pathak" <
[email protected]> writes:
>>>>
>>>>> Randy,
>>>>>
>>>>> The output wordlength is briefly touched upon in fir.pdf
>>>>> equations 86 and 87. It just talks about one part of the story,
>>>>> and that part is overflow, which you have mentioned in your doc.
>>>> Well, I mention maintaining precision and avoiding overflow. In other
>>>> words, there are at least two "parts of the story" there. Maybe you have
>>>> more parts?
>>> How about the number of worthwhile (significant) bits? When I measure
>>> the diameter of a tree as 2 inches, it's naive to cite it's
>>> circumference as 6.283185307179586476925286766559 inches. For this
>>> kind of measurement, fractional Angstrom units don't mean much.
>>>
>>> When later truncation creates so little noise that the noise from the
>>> original quantization masks it, any more bits are simply wasted.
>>
>> If you can make some statistical assumptions about the input data, then
>> perhaps you can get away with fewer bits. But of course in general,
>> using any fewer bits would lose information.
>
> It would lose something, but would what is lost be information?
>
> Would truncating the tree diameter from
> 6.283185307179586476925286766559 to 6.283, thus reporting it to the
> nearest thousandth of an inch, (or 25 microns) lose information? What
> is the meaning of 6.2831853071795864 +/-
> .001? It's like giving baking time to milliseconds when the oven
> temperature is known only within a few degrees.
>
> DSP is about numbers, and numbers can have any presumed
> precision. When those numbers are used to describe something, there is
> a limit to how precise the description can meaningfully be.
I see your point, Jerry. I don't see how you could generalize that
type of property though since it depends very much on the specifics
of the application. But yes, there are certainly times when using
less than L + log_2(alpha) bits is perfectly acceptable. In fact
you MUST do that, e.g., when using the SSE2 architecture since its
"accumulators" are only 32 bits wide (and the input data is 16
bits).
--
% Randy Yates % "And all that I can do
%% Fuquay-Varina, NC % is say I'm sorry,
%%% 919-577-9882 % that's the way it goes..."
%%%% <
[email protected]> % Getting To The Point', *Balance of Power*, ELO
http://www.digitalsignallabs.com