Subject: Generalized Lambert W function - premature commit
From: David Billinghurst
Date: Thu, 24 May 2012 23:49:31 +1000
On 24/05/2012 11:30 PM, Richard Fateman wrote:
> On 5/24/2012 3:20 AM, David Billinghurst wrote:
>>
>>
>> I have just checked in a generic version using the bigfloat package.
>> Much neater. I removed the guard digits from the bfloat calcs - it
>> was a nasty hack that didn't always work - so the results may change
>> slightly. I will just make a note in the manual that the final
>> digits may be wrong, and to suggest users check using higher
>> precision if it matters.
>
> I hope you are talking about just the final digits of the bfloat of
> Lambert W, not the final digits of bfloats.
>
> The cost of getting the rounding correct on bfloat add/subtract is
> substantial, but (when I wrote it) deemed essential
> because IEEE standard was doing the right thing. I hope that has not
> been removed! (Actually, I wrote a slimmer
> faster non-round-to-nearest version also. That may be hanging around
> in the code somewhere...)
>
> RJF
>
Just the final digits on some bfloat values for the Lambert W function.
I had increased the precision locally, but it wasn't sufficient for all
cases and made the generic routine untidy.
I wouldn't presume to fiddle with the core bfloat routines.