[Bug?] for step on float

[newdep]

Hi Lutz,



I dont understand the result below... do you have a clue?



Norman.







> (for (t 1 0.0 0.1) (println t))

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

-5.551115123e-17

-5.551115123e-17

>

[cormullion]

On my Mac it's

Code Select Expand

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0


so I expect your platform (Linux?) is giving you an optimistic view of zero!



And you know better than I do why you get it twice...! ;-)

[newdep]

Heee odd... seems to be a bug indeed... (linux newlisp 9.2.0 , 32 bits)





double screen vision, I have that also since yesterday evening ;-)

[pjot]

It may be the same bug with (sequence) mentioned before:



http://www.alh.net/newlisp/phpbb/viewtopic.php?t=1538&postdays=0&postorder=asc&start=25">http://www.alh.net/newlisp/phpbb/viewto ... c&start=25">http://www.alh.net/newlisp/phpbb/viewtopic.php?t=1538&postdays=0&postorder=asc&start=25





Option [8].



Peter

[newdep]

..Thanks..



But unacceptable..

We are unable to ever reach 0.0 under linux using floats...  Handy ! ;-)

[Lutz]

yes, on Windows too, but fine on Mac OS X, other BSD's and Solaris. Use format to output the numbers.



Lutz



Ps: There is no precise way to convert from decimal to binary and back. On any OS you can find numbers showing remainders on division.

[Lutz]

when comparing floats you always should compare the difference of the numbers against some other small quantity. Google for: comparing floats



Lutz

[newdep]

Im realy lost...



Seems newlisp is simply "Ignoring" ZERO ;-)



> (setq t 1)

1

> (for (x 10 0) (println (dec 't 0.1)))

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

1.387778781e-16

-0.1

-0.1

[Lutz]

This is the same effect you where showing before. Again,  there is no precise way to convert between decimal to binary and back, use formatting when displaying.



Lutz

[Lutz]

for your last example there is a nice piece here:



http://www.petebecker.com/js/js200006.html">http://www.petebecker.com/js/js200006.html



Lutz

[newdep]

..yes..yes i understand the issue.. ;-)



(for (x 1.0 0.01 0.01) (println x))



That assumes almost zero for my purpose ;-)



Thanks..



Norman.

[Lutz]

The difference between:

Code Select Expand

(for (i 1 10) (println (dec 'x 0.1)))

and

Code Select Expand

(for (i 1 0.0 0.1) (println i))

is, that newLISP when doing for loops and sequences tries to avoid accumulating the rounding error by calculating the total number of iterations first, then calculates the loop variable i as the product of 'step * count' and not as a repeated increment as in the first example. This way newLISP will never over/under run a loop, the way JavaScript would do, as shown in Peter Becker's article.



Your last example with decrementing x multiple times will produce the rounding error in pretty much any programming language after enough iterations depending on the floating point format and precision. Here the example in Python:

Code Select Expand

x = 1
for i in range(1, 11):
    x -= 0.1
    print x

0.9
...
0.1
1.38777878078e-16


Lutz



newLISP shows 1.387778781e-16 because it uses only 9 digits when displaying unformatted.

[newdep]

Ooo its very good that newlisp does NOT round it (visualy)

by default to 0.0 because that is even worse...



I understand the idea, no problem ..thanks for explaining..



Norman.