How to set precision for double data type variables in C?

Question

Possible Duplicate:
Most effective way for float and double comparison

In my program some double variables take values of the form 1.00000001. Equality check of these variables with 1 obviously fails.

I wanted to know how can I reduce the precision of double type variables in c, so that equality with integers works.

http://stackoverflow.com/questions/17333/most-effective-way-for-float-and-double-comparison — Lucas, Jul 05 '12 at 14:44
@J.N. For `1.5` equality should return false. Let us say, anything after fifth decimal place should be ignored. — stressed_geek, Jul 05 '12 at 14:46
Why the fifth decimal place? There are two problems with that. First, a number very slightly below 1.00001 and a number very slightly above 1.00001 will be reported as not equivalent even though you might want otherwise (because they differ only because of rounding errors, not because of actual difference if computed with mathematical exactness). Second, you should have a basis for setting the threshold for considering numbers to be equivalent. You should know how much error can be produced in the values you are comparing and how much error your decision can tolerate. — Eric Postpischil, Jul 05 '12 at 14:54

Felice Pollano · Answer 1 · 2012-07-05T14:55:10.580

3

You should almost never check floating point values for exact equality, expecially when they comes from comutation. You should check the absolute value of the difference from the compare value being less than a certain epsilon. The "precision" of the double is given by the internal number representation and you can't change it. How to exactly choose epsilon can be difficult, there is some comment to that answer discussing that, read it, but you eventually come with the pratical epsilon based equality.

edited Jul 05 '12 at 14:55

answered Jul 05 '12 at 14:43

Felice Pollano

32,832
9
75
115

The double variables are returned by a linear program, therefore I can't do without the equality check. – stressed_geek Jul 05 '12 at 14:47
The question linked by @Lucas suggest being extremely careful when doing so. – J.N. Jul 05 '12 at 14:48
@stressed_geek you just have to change your if from == to abs(a-x) < epsilon ... – Felice Pollano Jul 05 '12 at 14:48
this suggestion is wrong in many ways, `abs(a-x) < epsilon` is a good idea only for numbers close to zero (both `a` and `x`) otherwise relative epsilon or ulps. And you sometimes have to compare doubles using `==` – unkulunkulu Jul 05 '12 at 14:49
@unkulunkulu I agree, but you'd better say why ;) – J.N. Jul 05 '12 at 14:50
@unkulunkulu everybody would appreciate to know why. – Felice Pollano Jul 05 '12 at 14:50
@FelicePollano, complete answer is here http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm – unkulunkulu Jul 05 '12 at 14:51
@unkulunkulu and J.N comparing doubles for exact equality is always wrong. – Felice Pollano Jul 05 '12 at 14:51
@FelicePollano, say I have an input from the user and I want to count the numbers that are non-zero, in this case comparing directly is the only way to go. Also, recently I had a queue simulation system at my hands and I had to estimate the waiting times distribution, sometimes the client gets served immediately after arriving , I used direct comparison and I was right. – unkulunkulu Jul 05 '12 at 14:52
@unkulunkulu I remove the never with "almost never" – Felice Pollano Jul 05 '12 at 14:56

score 2 · Accepted Answer · answered Jul 05 '12 at 14:46

There's no portable way, no.

With the GNU C library, you can use this API to change the rounding mode.

But in general, it's better to express it with code, so that your expectations become clear and portable:

#define EQUALITY_EPSILON 1e-3  /* Or whatever. */

if(fabs(x - y) <= EQUALITY_EPSILON)
{
}

japreiss · Answer 3 · 2012-07-05T14:54:00.427

1

You should avoid checking for equality in floating-point comparisons. Instead, use a precision value epsilon like so:

if (fabs(a - b) < epsilon)
{
    // treat a and b as equal
}

The choice of epsilon is probably complicated, but my knowledge doesn't go that far.

Do a Google search for "What Every Programmer Should Know About Floating-Point Arithmetic". It's a well-known article that covers all this stuff.

edited Jul 05 '12 at 14:54

answered Jul 05 '12 at 14:46

japreiss

11,111
2
40
77

1

This is absolutely not correct code. abs is the integer absolute value function. The floating-point value a-b will be converted to integer (by truncation), and the absolute value of that will be compared to epsilon. Any value of a-b less than 1 will be reported as less than epsilon, regardless of how small epsilon is (as long as it is positive). The correct function to use is fabs. – Eric Postpischil Jul 05 '12 at 14:51
1

Oops, I'm used to C++ where it's overloaded. Fixed in edit. – japreiss Jul 05 '12 at 14:53

score 0 · Answer 4 · answered Jul 05 '12 at 14:48

It doesn't make a lot of sense to 'reduce the precision of double type variables'. You should probably think instead of using functions such as ceil, floor, and round to make an integer out of your double in a way that you have good control over and then use that integer in your comparisons.

score 0 · Answer 5 · answered Jul 05 '12 at 14:54

0

in C99

#include <stdio.h>
#include <math.h>

int main(){
    double d = 1.00000001;
    int i = (int)round(d);

    printf("%d\n", i);

    return 0;
}

answered Jul 05 '12 at 14:54

BLUEPIXY

39,699
7
33
70

How to set precision for double data type variables in C?

5 Answers5