Floating point arithmetic.

[Amit Bhatia]

Hi there.
I am cross posting this on comp.lang.c as well: sorry for same.
The problem I am facing is as follows:
For example:
double a= 0.15;
double b=2.4;
const double VERYTINY =1.e-10;
I know b/a = 16 and hence the remainder is zero; but I am not
able to find any suitable thing to encode it into in c.
for example (fmod(b,a)>VERYTINY) returns true!
Now for this particular instance, (fmodf(b,a)>VERYTINY)
does return false.
But now if
a=0.15;
b=4.5;
then fmodf and fmod both don't help...

any suggestions on this?
I was pointed to a reference on floating point arithmetic, where they talk
of ulps etc, but is there a small function or fix to deal with this problem
available somewhere?

thanks,
amit.

 

[Robbie Hatley]

The problem I am facing is as follows:
For example:

Example of what? It makes more sense to state the
thing you want to give an example of, before giving
examples of it.

double a= 0.15;
double b=2.4;
const double VERYTINY =1.e-10;
I know b/a = 16 and hence the remainder is zero;
but I am not able to find any suitable thing to
encode it into in c.

Define "encode".

Define "it".

If you want to code b/a in C, the code is:

b/a

for example

Example of what? Please give some clue as to the thing
you're trying to give an example OF, before giving
examples!

(fmod(b,a)>VERYTINY) returns true!

It returns false on my system. However, it seems to me
you're pushing the granularity of type double to the max.
Even if you WERE off by one part per 1e10, so what?
That's one part per ten billion. Who cares?

But if you really DO care, try acquiring a compiler that
supports type "long double". Or, perhaps, use a computer
with a 64-bit processor and compiler that defines "double"
to be 64 bits. Even if 4 or so of those bits were used
for sign and exponent, you'd still have a granularity of
about 1 part per quintillion (1e18). More precision than
you'll EVER need.

Now for this particular instance, (fmodf(b,a)>VERYTINY)
does return false.

What is fmodf? That's not a part of the C or C++ std.
libraries. I don't have it on my compiler (DJGPP) either.

But now if
a=0.15;
b=4.5;
then fmodf and fmod both don't help...

I get b/a = 30 and (fmod(b,a)>VERYTINY) = false.

What do YOU get?

any suggestions on this?

Suggestions on what, precisely?

I was pointed to a reference on floating point
arithmetic, where they talk of ulps etc,

What's "ulps"?

is there a small function or fix to deal with this problem
available somewhere?

What problem are you trying to fix?

--
Cheers,
Robbie Hatley
Tustin, CA, USA
(e-mail address removed)
http://home.pacbell.net/earnur/

 

[John Harrison]

What's "ulps"?

ULP stands for unit in the last place. Its a measure of rounding error and
a common enough term in floating point arithmetic.

john

 

[John Harrison]

Hi there.
I am cross posting this on comp.lang.c as well: sorry for same.
The problem I am facing is as follows:
For example:
double a= 0.15;
double b=2.4;
const double VERYTINY =1.e-10;
I know b/a = 16 and hence the remainder is zero; but I am not
able to find any suitable thing to encode it into in c.
for example (fmod(b,a)>VERYTINY) returns true!
Now for this particular instance, (fmodf(b,a)>VERYTINY)
does return false.
But now if
a=0.15;
b=4.5;
then fmodf and fmod both don't help...

any suggestions on this?

Maybe use a bigger value for VERYTINY.
Maybe use integral or rational aritmetic.
Maybe use fixed point arithmetic.

I was pointed to a reference on floating point arithmetic, where they
talk
of ulps etc, but is there a small function or fix to deal with this
problem
available somewhere?

There is not easy answer to the problem of floating point rounding errors.
The simplest thing is not to write code that depends on any particular
exactness of floating point arithmetic. Obviously this is not always
possible but it is possible more often than people think.

So the asnwer to your question really depends on what kind of problem you
are actually trying to solve.

john

 

[David Harmon]

On Sat, 10 Jul 2004 23:08:01 -0500 in comp.lang.c++, Amit Bhatia

Hi there.
I am cross posting this on comp.lang.c as well: sorry for same.

In fact you did so a few days ago and got some good tips.

The problem I am facing is as follows:
For example:
double a= 0.15;
double b=2.4;
const double VERYTINY =1.e-10;

b = two and two fifths. Fifths can not be represented exactly in base
two floating point format, so this number cannot be stored exactly.
a is 16 times that, so of course has a similar approximation.

And so on. Never expect any form of exactitude from any floating point
operation, and your life will be less frustrating.

I know b/a = 16 and hence the remainder is zero; but I am not
able to find any suitable thing to encode it into in c.

It may be close to zero, or it may be just less than 0.15.

If you persist in this folly then I guess you need to test for both
possibilities.

 

[Subhash]

Hi there.
I am cross posting this on comp.lang.c as well: sorry for same.
The problem I am facing is as follows:
For example:
double a= 0.15;
double b=2.4;
const double VERYTINY =1.e-10;
I know b/a = 16 and hence the remainder is zero; but I am not
able to find any suitable thing to encode it into in c.
for example (fmod(b,a)>VERYTINY) returns true!
Now for this particular instance, (fmodf(b,a)>VERYTINY)
does return false.
But now if
a=0.15;
b=4.5;
then fmodf and fmod both don't help...

any suggestions on this?
I was pointed to a reference on floating point arithmetic, where they talk
of ulps etc, but is there a small function or fix to deal with this problem
available somewhere?

thanks,
amit.

int main() {
double a = 0.15;
double b = 2.4;
const double VERYTINY =1.e-10;
printf("%d", (fmod(b,a) > VERYTINY));
return 0;
}

This gives me false!

 

[Arijit]

Hi there.
I am cross posting this on comp.lang.c as well: sorry for same.
The problem I am facing is as follows:
For example:
double a= 0.15;
double b=2.4;
const double VERYTINY =1.e-10;
I know b/a = 16 and hence the remainder is zero; but I am not
able to find any suitable thing to encode it into in c.
for example (fmod(b,a)>VERYTINY) returns true!
Now for this particular instance, (fmodf(b,a)>VERYTINY)
does return false.
But now if
a=0.15;
b=4.5;
then fmodf and fmod both don't help...

any suggestions on this?
I was pointed to a reference on floating point arithmetic, where they talk
of ulps etc, but is there a small function or fix to deal with this problem
available somewhere?

thanks,
amit.

Try the following loop

cout.precision(20);
cout << b << endl << a << endl;
while(b>a)
b -= a;
cout << b << endl << a << endl;

and see the value of a and b

Your problem is that the method of not comparing floating
point values but comparing their difference to a certain limit
does not work for modulo arithmetic. The reason is simple.

Assume x(mod y) = 0
That is, x = yp

In your case, x and y are floating point values,
while p is an integer. Since p is an integer, you
can say x is divisible by y.

But a computer cannot store floating point numbers exactly.
so a floating point number x is actually stored as (x+dx)
or as (x-dx). In your case, 4.5 can be exactly stored, but
neither can 2.4 or 0.15 . Now consider what happens when you
divide (x+dx) or (x-dx) by y. (x,y,p > 0)

(x+dx)(mod y) = dx
This is ok, your comparison will work in this case, as your
VERYTINY is > dx

However, (x-dx)(mod y) = (yp-dx)(mod y)
= (y(p-1) + y - dx)(mod y) = y-dx
This y-dx is definitely mmuch greater than your VERYTINY
and therefore you do not get the desired result.

I'll give you an example with integers.

Consider dividing 57 by 19. 57(mod 19) = 0
58(mod 19) = 1 ( x+dx case)
56(mod 19) = 18 (x-dx case)

So you should write
( fmod(b,a)<VERYTINY || (a-fmod(b,a))<VERYTINY )

-Arijit

 

[Miguel]

When using floating-point, usually we don't test for equality, instead do
the kind of test you do (fmod(b, a) > VERYTINY)...
But if we take a look at the case in particular, the test doesn't fail on my
machine, and I suspect the > VERYTINY doesn't fail in yours, except for
negative numbers. That is,

2.4 can only be finitely represented in binary as 2.3999... that is
10.0110011(0011)
0.15 as 0.14(9) (or 0.00100110011(0011))

in my machine, perhaps yours too, fmod(2.4, 0.15) is exactly 0. This is not
surprising if we examine the mantissa bits of both numbers, they are equal.

but 4.5 does have a finite binary representation as 100.1

but fmod(4.5, 0.15) is only 1.66533e-16, and so is smaller than 1e-10... and
I don't know why the test should fail....
Since the case of the denominator being represented exactly implies that a
multiple is represented exactly too

So, given best possible representations of the input values the test
(abs(fmod(b, a)) < VERYTINY)) should work.

One exception arises when input values are themselves approximations, and,
in particular, the dividend is slightly less than it would take to be an
exact multiple. In that case, the remainder will be less than, but very
close to the denominator. So, we must test the result also against it being
very close to the denominator.

So, the full test, that should work for all cases would be (assuming we want
to know about approximate multiples):

(abs(fmod(b,a)) < VERYTINY) || (abs(fmod(b,a) - a) < VERYTINY)

Now, the value you choose for VERYTINY, depends on the application and how
many operations you have done on the input values. Of course, because the
result of fmod is in absolute value always less than a, if you know your
inputs are positive numbers, you can drop both abs function calls.


Miguel Ramos

 

[Miguel]

What is fmodf? That's not a part of the C or C++ std.
libraries. I don't have it on my compiler (DJGPP) either.

actually fmodf does exist, it is part of the C99 standard, and most modern
compilers have it.
but I think we shouldn't be very harsh, it's obvious he meant modf.

 

[Robbie Hatley]

actually fmodf does exist, it is part of the C99 standard,

"fmodf" is a not a keyword in C or C++, nor is it
a part of their standard libraries, according to my
reading. I found reference to "fmodf" in the
documentation for my compiler (DJGPP), where it is
listed as a "non-ANSI extention to fmod". So unless
fmodf has been added to the standard libraries very
recently, it's not standard.

and most modern compilers have it.

Most compilers have lots of added non-ANSI functions.
Useful, but not standard.

it's obvious he meant modf.

Probably.

--
Cheers,
Robbie Hatley
Tustin, CA, USA
(e-mail address removed)
http://home.pacbell.net/earnur/

 

[John Harrison]

"fmodf" is a not a keyword in C or C++, nor is it
a part of their standard libraries, according to my
reading. I found reference to "fmodf" in the
documentation for my compiler (DJGPP), where it is
listed as a "non-ANSI extention to fmod". So unless
fmodf has been added to the standard libraries very
recently, it's not standard.


Most compilers have lots of added non-ANSI functions.
Useful, but not standard.


Probably.

Did you not read Miguel's post? fmodf is defined in the C99 standard,
section 7.12.10.1. Therefore it is a standard C function. You are talking
about the older ANSI standard I guess.

john

 

[P.J. Plauger]

Did you not read Miguel's post? fmodf is defined in the C99 standard,
section 7.12.10.1. Therefore it is a standard C function. You are talking
about the older ANSI standard I guess.

Even C89 provided for *f and *l versions of the standard math functions.
They were permitted but not required. So fmodf has been around for a
long time. None of which alters the fact that it *still* probably
should have been modf in the posting.

P.J. Plauger
Dinkumware, Ltd.
http://www.dinkumware.com

 

[Amit Bhatia]

Guys,
I was indeed using fmodf and fmod: I use g++ version 3.2 on rh 9.0 (with
-ansi compilation flag) to find the remainder when two double precision
values are divided one of which is exact multiple of the other.
thanks, for pointing me out that it is not exactly computable and I need to
be a bit careful while doing this stuff. Anyway, I am using a slightly
different check now which was suggested by Arijit and which works well.

thanks,
amit.

 

[Robbie Hatley]

John Harrison chastised me thusly:

Did you not read Miguel's post?

Oh, I read it; I just didn't believe it, because the
books available to me said otherwise.

fmodf is defined in the C99 standard, section
7.12.10.1. Therefore it is a standard C function.

Quoting section numbers on my ass, eh?

You are talking about the older ANSI standard I guess.

Yes, I suppose my books are all based on C90, not C99.

I see you're going to throw the book at me, so I'd
better get this standard you guys keep talking about.

::: gets standard :::

Now, let me see... Yes, here we go...

Pursuant to ISO/IEC 9899:1999, §7.12.10.1, paragraph 1,
the header math.h shall contain:

double fmod (double x, double y)
float fmodf (float x, float y)
long double fmodl (long double x, long double y)

OK, it seems that you guys were right.

But wait, that's the C standard, not C++.

Pursuant to ISO/IEC 14882, §26.5, paragraph 6, I see
that in standard C++, the <cmath> header includes:

double fmod (double x, double y)
float fmod (float x, float y)
long double fmod (long double x, long double y)

Uses overloaded versions of fmod, instead of fmodf and
fmodl.

§26.5, paragraph 2 does say "the contents of these
headers (<cmath> and <cstdlib>) are the same as the C
Standard Library headers math.h and stdlib.h, with the
following additions...", which would seem to imply that
fmodf, fmodl, and many other such library functions
are a part of standard C++. But then, 14882 was
written in 1998, so "Standard C" actually meant C90,
so one could argue that fmodf, fmodl, etc. are therefore
NOT part of standard C++. So perhaps I was right, after
all.

--
Feeling argumentative,
Robbie Hatley
Tustin, CA, USA
email: lonewolfintj at pacbell dot net
web: home dot pacbell dot net slant earnur slant

 

[P.J. Plauger]

Pursuant to ISO/IEC 9899:1999, §7.12.10.1, paragraph 1,
the header math.h shall contain:

double fmod (double x, double y)
float fmodf (float x, float y)
long double fmodl (long double x, long double y)

OK, it seems that you guys were right.

But wait, that's the C standard, not C++.

Pursuant to ISO/IEC 14882, §26.5, paragraph 6, I see
that in standard C++, the <cmath> header includes:

double fmod (double x, double y)
float fmod (float x, float y)
long double fmod (long double x, long double y)

Uses overloaded versions of fmod, instead of fmodf and
fmodl.

§26.5, paragraph 2 does say "the contents of these
headers (<cmath> and <cstdlib>) are the same as the C
Standard Library headers math.h and stdlib.h, with the
following additions...", which would seem to imply that
fmodf, fmodl, and many other such library functions
are a part of standard C++. But then, 14882 was
written in 1998, so "Standard C" actually meant C90,
so one could argue that fmodf, fmodl, etc. are therefore
NOT part of standard C++. So perhaps I was right, after
all.

Except that you aren't. The float and long double math
functions that are optional in C90 are explicitly required
in C++98.

P.J. Plauger
Dinkumware, Ltd.
http://www.dinkumware.com