[Numpy-discussion] Log Arrays

Robert Kern robert.kern@gmail....
Thu May 8 12:57:19 CDT 2008


On Thu, May 8, 2008 at 12:53 PM, Charles R Harris
<charlesr.harris@gmail.com> wrote:
>
> On Thu, May 8, 2008 at 11:18 AM, David Cournapeau <cournape@gmail.com>
> wrote:
>>
>> On Fri, May 9, 2008 at 2:06 AM, Nadav Horesh <nadavh@visionsense.com>
>> wrote:
>> > Is the 80 bits float (float96 on IA32, float128 on AMD64) isn't enough?
>> > It has a 64 bits mantissa and can represent numbers up to nearly 1E(+-)5000.
>>
>> It only make the problem happen later, I think. If you have a GMM with
>> million of samples of high dimension with many clusters, any "linear"
>> representation will fail I think. In a sense, the IEEE format is not
>> adequate for that kind of computation.
>
> David, what you are using is a log(log(x)) representation internally. IEEE
> is *not* linear, it is logarithmic.

*YES*. That is precisely the point. I want 53 bits devoted to the "x"
part of "exp(-x)". The straight IEEE representation is not logarithmic
*enough*.

-- 
Robert Kern

"I have come to believe that the whole world is an enigma, a harmless
enigma that is made terrible by our own mad attempt to interpret it as
though it had an underlying truth."
 -- Umberto Eco


More information about the Numpy-discussion mailing list