Python epsilon is not the smallest number

问题:

What does sys.float_info.epsilon return?

On my system I get:

>>> sys.float_info.epsilon
2.220446049250313e-16
>>> sys.float_info.epsilon / 2
1.1102230246251565e-16
>>> 0 < sys.float_info.epsilon / 2 < sys.float_info.epsilon
True

How is this possible?

EDIT:

You are all right, I thought epsilon does what min does. So I actually meant sys.float_info.min.

EDIT2

Everybody and especially John Kugelman, thanks for your answers!

Some playing around I did to clarify things to myself:

>>> float.hex(sys.float_info.epsilon)
'0x1.0000000000000p-52'
>>> float.hex(sys.float_info.min)
'0x1.0000000000000p-1022'
>>> float.hex(1 + a)
'0x1.0000000000001p+0'
>>> float.fromhex('0x0.0000000000001p+0') == sys.float_info.epsilon
True
>>> float.hex(sys.float_info.epsilon * sys.float_info.min)
'0x0.0000000000001p-1022'

So epsilon * min gives the number with the smallest positive significand (or mantissa) and the smallest exponent.

回答1:

epsilon is the difference between 1 and the next representable float. That's not the same as the smallest float, which would be the closest number to 0, not 1.

There are two smallest floats, depending on your criteria. min is the smallest normalized float. The smallest subnormal float is min * epsilon.

>>> sys.float_info.min
2.2250738585072014e-308
>>> sys.float_info.min * sys.float_info.epsilon
5e-324

Note the distinction between normalized and subnormal floats: min is not actually the smallest float, it's just the smallest one with full precision. Subnormal numbers cover the range between 0 and min, but they lose a lot of precision. Notice that 5e-324 has only one significant digit. Subnormals are also much slower to work with, up to 100x slower than normalized floats.

>>> (sys.float_info.min * sys.float_info.epsilon) / 2
0.0
>>> 4e-324
5e-324
>>> 5e-325
0.0

These tests confirm that 5e-324 truly is the smallest float. Dividing by two underflows to 0.

See also: What is the range of values a float can have in Python?

回答2:

You actually want sys.float_info.min ("minimum positive normalized float"), which on machine gives me .2250738585072014e-308.

epsilon is:

difference between 1 and the least value greater than 1 that is representable as a float

See the docs for more info on the fields of sys.float_info.

回答3:

Your last expression is possible, because for any real, positive number, 0 < num/2 < num.

From the docs:

difference between 1 and the least value greater than 1 that is representable as a float

回答4:

sys.float_info is defined as

difference between 1 and the least value greater than 1 that is representable as a float

on this page.

回答5:

The documentation defines sys.float_info.epsilon as the

difference between 1 and the least value greater than 1 that is representable as a float

However, the gap between successive floats is bigger for bigger floats, so the gap between epsilon and the next smaller float is a lot smaller than epsilon. In particular, the next smaller float is not 0.

回答6:

Like every answer says, it's the difference between 1 and the next greatest value that can be represented, if you tried to add half of it to 1, you'll get 1 back

>>> (1 + (sys.float_info.epsilon/2)) == 1
True

Additionally if you try to add two thirds of it to 1, you'll get the same value:

>>> (1 + sys.float_info.epsilon) == (1 + (sys.float_info.epsilon * (2./3)))
True