# Re: [R] Matrix calculations in R--erroneous?

From: Peter Muhlberger <pmuhl830_at_gmail.com>
Date: Sat 08 Oct 2005 - 06:28:42 EST

Hi Thomas: Thanks! Yes, the function
(yMat2[5,9]-yMat1[5,9]*logitShape()[5,9]) appears to be producing a value of -1.102216e-16 rather than 0. I would have thought it would approach 0 from above given that all input values are at or above zero, but evidently not.

The max function won't do the trick because I need the entire matrix. I could do one cell at a time, but this is part of a ML routine that needs to be evaluated hundreds of thousands of times, so I can't afford to slow it down that much.

I guess I can add 1e-15 rather than e-323, but wonder what that might end up doing to my estimates. Guess I'll find out.

Cheers, Peter

On 10/7/05 1:12 PM, "Thomas Lumley" <tlumley@u.washington.edu> wrote:

> On Fri, 7 Oct 2005, Peter Muhlberger wrote:
>
>> Does anyone know how -log(x) can equal 743 but -log(x+0)=Inf? That's what
>> the following stream of calculations suggest:
>>
>> Browse> -log ( 1e-323+yMat2 - yMat1 * logitShape(matrix(parsList\$Xs,
>> nrow = numXs, ncol=numOfCurves), matrix(means, nrow = numXs,
>> ncol=numOfCurves, byrow=TRUE), matrix(sigmas, nrow = numXs,
>> ncol=numOfCurves, byrow=TRUE)) )[5,9]
>>  Inf
>>
>> Yet:
>>
>> Browse> logitShape(matrix(parsList\$Xs, nrow = numXs, ncol=numOfCurves),
>> matrix(means, nrow = numXs, ncol=numOfCurves, byrow=TRUE), matrix(sigmas,
>> nrow = numXs, ncol=numOfCurves, byrow=TRUE))[5,9]
>>  1
>>
>> So, the logitShape component equals 1.
>
> to within 2e-16
>
>> Browse> yMat1[5,9]
>>  1
>>
>> So yMat1[5,9]*logitShape()[5,9]=1
>
> to within 2e-16
>
>> Browse> yMat2[5,9]
>>  1
>
> to within 2e-16
>
>> So, yMat2[5,9]-yMat1[5,9]*logitShape()[5,9]=0
>
> to within a few parts in 10^16
>
> You haven't actually shown us yMat2[5,9]-yMat1[5,9]*logitShape()[5,9],
> though
>
>> Browse> -log ( 1e-323)
>>  743.7469
>>
>> So, -log( 1e-323)=743 while -log( 1e-323+0)=Inf ?
>>
>
> If "0" is really of the order of 1e-16 then this isn't surprising. If the
> only point of 1e-323 is as a guard value for 0 then use max(1e-323,
> yMat2[5,9]-yMat1[5,9]*logitShape()[5,9])
>
>
> -thomas
>
>

R-help@stat.math.ethz.ch mailing list