Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: John's asinh split versus convert-to-sortable...



> Date: Fri, 23 Mar 2012 12:24:59 +0100
> From: Arnold Neumaier <Arnold.Neumaier@xxxxxxxxxxxx>
> To: "Kreinovich, Vladik" <vladik@xxxxxxxx>
> CC: 'Dan Zuras Intervals' <intervals08@xxxxxxxxxxxxxx>, 
>  "stds-1788@xxxxxxxxxxxxxxxxx" <stds-1788@xxxxxxxxxxxxxxxxx>
> Subject: Re: John's asinh split versus convert-to-sortable...
> 
> On 03/22/2012 10:06 PM, Kreinovich, Vladik wrote:
> > I like this reformulation, it only used sqrt which is usually very fast.
> >
> > 		split(x,y):
> > 			u = x + sqrt(1 + x^2);
> > 			v = y + sqrt(1 + y^2);
> > 			t = sqrt(u*v);
> > 			return s = (t - 1/t)/2;
> 
> This produces inf when x or y > 1e160, not a good splitting point.
> 


	Arnold,

	You refer to an overflow of u*v in the formula for t.
	From the list of optimizations I presented there was:


	"I'll set aside problems with over/underflow in t =
	sqrt(u*v) (which happen in u*v, of course) other than
	to say one can always go to:

		t = sqrt(u)*sqrt(v)

	which is not ideal but easier to express than the
	usual prescaling crap that also solves the problem."


	All these things are level 2 issues for a function
	which is well defined at level 1.


				Dan