On 4/23/07, Tracy Harms <[EMAIL PROTECTED]> wrote:
The valuable sort of examination can be found in the following paper, for example:
...
http://www.cs.berkeley.edu/~wkahan/JAVAhurt.pdf
Personally, I am not particularly impressed. I tried to reproduce one of his graphs on page 13. I believe I was able to reproduce his equations on page 12: gz=: *: + ] * %:@>:@*: Fz=: 1 + gz + [EMAIL PROTECTED] ray=:1 :'* ^@((m*0j0.1p1)&*)' z=:[EMAIL PROTECTED],~j.,|:@:((i:4) ray"0) With r=: 0.01*1+i.100 I see what I think is the "bad" result he was talking about. plot Fz z r [ require 'plot' That said, with r=:0.01*1+i.300 I see other quirks, and without a deeper analyzis of what's going on, I'm not sure why I should or should not consider these quirks "errors". Frankly, I find his axis labels on his graphs on page 13 meaningless, and I'm really sure whether or not the quirk I saw with r=: 0.01*1+i.100 is the same one he's trying to describe. For example, how dependent is his displayed result on his value for delta R? Anyways... if he has a good point, I don't think he's described it very well That said, if I understand his point, he expects that Fz 0j1 0j_1 be 0j1p1 0j_1p1 instead of 0j1p1 0j1p1, but when I try to pin down why that should be the case with "negative zero" support, I can only presume that he's implicitly assuming some particular implementation of some trancendental function -- where zero or negative zero would be an intermediate result in that implementation. However, when I look at this closer gz 0j1 0j_1 _1 _1 I don't see any reason why I should expect negative zero to matter at all. Anyways... I am not convinced that I understand his point -- nor am I convinced he has a good point. -- Raul ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
