Suppose I have two floating point numbers, x
and y
, with their values being very close.
There's a discrete number of floating point numbers representable on a computer, so we can enumerate them in increasing order: f_1, f_2, f_3, ...
. I wish to find the distance of x
and y
in this list (i.e. are they 1, 2, 3, ... or n
discrete steps apart?)
Is it possible to do this by only using arithmetic operations (+-*/
), and not looking at the binary representation? I'm primarily interested in how this works on x86.
Is the following approximation correct, assuming that y > x
and that x
and y
are only a few steps (say, < 100) apart? (Probably not ...)
(y-x) / x / eps
Here eps
denotes the machine epsilon. (The machine epsilon is the difference between 1.0 and the next smallest floating point number.)
Floats are lexicographically ordered, therefore:
int steps(float a, float b){
int ai = *(int*)&a; // reinterpret as integer
int bi = *(int*)&b; // reinterpret as integer
return bi - ai;
}
steps(5.0e-1, 5.0000054e-1); // returns 9
Such a technique is used when comparing floating point numbers.
You do not have to examine the binary representation directly, but you do have to rely on it to get an exact answer, I think.
Start by using frexp() to break x into exponent exp
and mantissa. I believe the next float bigger than x is x + eps * 2^(exp-1)
. (The "-1" is because frexp returns a mantissa in the range [1/2, 1) and not [1, 2).)
If x and y have the same exponent, you are basically done. Otherwise you need to count how many steps there are per power of 2, which is just 1.0/eps
. In other words, the number of steps between 2^n and 2^(n+1) is 1.0/eps
.
So, for y > x, count how many steps there are from x to the next power of two; then count how many more steps it takes to get to the largest power of 2 less than y; then count how many more steps it takes to get from there up to y. All of these are pretty easily expressible in terms of eps
, I think.
来源:https://stackoverflow.com/questions/6188822/finding-the-discrete-difference-between-close-floating-point-numbers