Python: sort function breaks in the presence of nan
Question:
sorted([2, float('nan'), 1])
returns [2, nan, 1]
(At least on Activestate Python 3.1 implementation.)
I understand nan
is a weird object, so I wouldn’t be surprised if it shows up in random places in the sort result. But it also messes up the sort for the non-nan numbers in the container, which is really unexpected.
I asked a related question about max
, and based on that I understand why sort
works like this. But should this be considered a bug?
Documentation just says “Return a new sorted list […]” without specifying any details.
EDIT:
I now agree that this isn’t in violation of the IEEE standard. However, it’s a bug from any common sense viewpoint, I think. Even Microsoft, which isn’t known to admit their mistakes often, has recognized this one as a bug, and fixed it in the latest version: http://connect.microsoft.com/VisualStudio/feedback/details/363379/bug-in-list-double-sort-in-list-which-contains-double-nan.
Anyway, I ended up following @khachik’s answer:
sorted(list_, key = lambda x : float('-inf') if math.isnan(x) else x)
I suspect it results in a performance hit compared to the language doing that by default, but at least it works (barring any bugs that I introduced).
Answers:
IEEE754 is the standard that defines floating point operations in this instance. This standard defines the compare operation of operands, at least one of which is a NaN, to be an error. Hence, this is not a bug. You need to deal with the NaNs before operating on your array.
I’m not sure about the bug, but the workaround may be the following:
sorted(
(2, 1, float('nan')),
lambda x,y: x is float('nan') and -1
or (y is float('nan') and 1
or cmp(x,y)))
which results in:
('nan', 1, 2)
Or remove nan
s before sorting or anything else.
The problem is that there’s no correct order if the list
contains a NAN
, since a sequence a1, a2, a3, ..., an
is sorted if a1 <= a2 <= a3 <= ... <= an
. If any of these a values is a NAN
then the sorted property breaks, since for all a, a <= NAN and NAN <= a
are both false
.
The previous answers are useful, but perhaps not clear regarding the root of the problem.
In any language, sort applies a given ordering, defined by a comparison function or in some other way, over the domain of the input values. For example, less-than, a.k.a. operator <,
could be used throughout if and only if less than defines a suitable ordering over the input values.
But this is specifically NOT true for floating point values and less-than:
“NaN is unordered: it is not equal to, greater than, or less than anything, including itself.” (Clear prose from GNU C manual, but applies to all modern IEEE754
based floating point)
So the possible solutions are:
- remove the NaNs first, making the input domain well defined via <
(or the other sorting function being used)
- define a custom comparison function (a.k.a. predicate) that does
define an ordering for NaN, such as less than any number, or greater
than any number.
Either approach can be used, in any language.
Practically, considering python, I would prefer to remove the NaNs if you either don’t care much about fastest performance or if removing NaNs is a desired behavior in context.
Otherwise you could use a suitable predicate function via “cmp” in older python versions, or via this and functools.cmp_to_key()
. The latter is a bit more awkward, naturally, than removing the NaNs first. And care will be required to avoid worse performance, when defining this predicate function.
Assuming you want to keep the NaNs and order them as the lowest “values”, here is a workaround working both with non-unique nan, unique numpy nan, numerical and non numerical objects:
def is_nan(x):
return (x is np.nan or x != x)
list_ = [2, float('nan'), 'z', 1, 'a', np.nan, 4, float('nan')]
sorted(list_, key = lambda x : float('-inf') if is_nan(x) else x)
# [nan, nan, nan, 1, 2, 4, 'a', 'z']
Regardless of standards, there are many cases where a user-defined ordering of float and NA
values is useful. For instance, I was sorting stock returns and wanted highest to lowest with NA
last (since those were irrelevant). There are 4 possible combinations
- Ascending float values,
NA
values last
- Ascending float values,
NA
values first
- Descending float values,
NA
values last
- Descending float values,
NA
values first
Here is a function that covers all scenarios by conditionally replacing NA
values with +/- inf
import math
def sort_with_na(x, reverse=False, na_last=True):
"""Intelligently sort iterable with NA values
For reliable behavior with NA values, we should change the NAs to +/- inf
to guarantee their order rather than relying on the built-in
``sorted(reverse=True)`` which will have no effect. To use the ``reverse``
parameter or other kwargs, use functools.partial in your lambda i.e.
sorted(iterable, key=partial(sort_with_na, reverse=True, na_last=False))
:param x: Element to be sorted
:param bool na_last: Whether NA values should come last or first
:param bool reverse: Return ascending if ``False`` else descending
:return bool:
"""
if not math.isnan(x):
return -x if reverse else x
else:
return float('inf') if na_last else float('-inf')
Testing out each of the 4 combinations
from functools import partial
a = [2, float('nan'), 1]
sorted(a, key=sort_with_na) # Default
sorted(a, key=partial(sort_with_na, reverse=False, na_last=True)) # Ascend, NA last
sorted(a, key=partial(sort_with_na, reverse=False, na_last=False)) # Ascend, NA first
sorted(a, key=partial(sort_with_na, reverse=True, na_last=True)) # Descend, NA last
sorted(a, key=partial(sort_with_na, reverse=True, na_last=False)) # Descend, NA first
A resilient sort involves a compare of 2 items and returning: less, equal, greater.
If cmp(a,b)
is "greater", then cmp(b,a)
must be "less".
If cmp(a,b)
is "zero", then cmp(b,a)
must be "zero".
What is missing in answers to date is the case of comparing 2 float
s which are both NANs and preserving the above properties. 2 NANs should compare as equal or perhaps based on some consistent interpretation of their payloads.
Alternate compare algorithm to put all NAN > +inf
if isnan(a)
if isnan(b)
return 0 (or maybe compare payloads/bit patterns)
return 1
if isnan(b) return 1
if a > b return 1
if a < b return -1
return 0
To recap the problems:
NaN
This always returns False
for every comparison, so it stays where it is in the list:
>>> sorted([float('nan'), 0])
[nan, 0]
>>> sorted([0, float('nan')])
[0, nan]
-0.0
This is == to 0.0
, but has a different repr, a different json representation, and slightly different numerical properties. It has the same problem that the positive and negative zeros will stay in the same order they were in the original list:
>>> sorted([0.0, -0.0])
[0.0, -0.0]
>>> sorted([-0.0, 0.0])
[-0.0, 0.0]
Other solution?
@khachik’s solution has inconsistent sorting behavior for NaN
and -inf
>>> key=lambda x: float('-inf') if math.isnan(x) else x
>>> sorted([float('nan'), float('-inf')], key=key)
[nan, -inf]
>>> sorted([float('-inf'), float('nan')], key=key)
[-inf, nan]
Solution: more complex key function.
So there are problems with signs and nans. We can just include those in a key function:
def stable_float_sort_key(x: float):
return math.copysign(1, x), math.isnan(x), x
That works on all the examples above:
>>> sorted([float('nan'), 0.0], key=stable_float_sort_key)
[0.0, nan]
>>> sorted([0.0, float('nan')], key=stable_float_sort_key)
[0.0, nan]
>>> sorted([float('nan'), float('-inf')], key=stable_float_sort_key)
[-inf, nan]
>>> sorted([float('-inf'), float('nan')], key=stable_float_sort_key)
[-inf, nan]
>>> sorted([0.0, -0.0], key=stable_float_sort_key)
[-0.0, 0.0]
>>> sorted([-0.0, 0.0], key=stable_float_sort_key)
[-0.0, 0.0]
Indeed, you can write a hypothesis test showing it’s consistent across all floating point numbers:
import json
from hypothesis import given, settings
from hypothesis import strategies as st
@given(nums=st.lists(st.floats()), random=st.randoms())
@settings(max_examples=10000)
def test_stable_json_sorting(nums, random):
shuffled = list(nums)
random.shuffle(shuffled)
l1 = sorted(nums, key=stable_float_sort_key)
l2 = sorted(shuffled, key=stable_float_sort_key)
assert json.dumps(l1) == json.dumps(l2)
It does have some oddities however, since some NaN’s are negative! For example:
>>> sorted([float('nan'), -0.0, 0.0, float('-nan')], key=stable_float_sort_key)
[-0.0, nan, 0.0, nan]
If that bothers you, you can fix this by switching the ordering:
def stable_float_sort_key(x: float):
return math.isnan(x), math.copysign(1, x), x
That sorts negative numbers first, followed by positive numbers, followed by NaNs.
Does any of this make sense?
Of course other answerers are correct that in some sense none of this makes sense. Comparison of NaNs is a conceptual error of some kind. However, even in cases where the question doesn’t "make sense", you may want to have invariants like serializing sets of floating point numbers generated by the same code to exactly the same JSON representation, despite hash randomization (my use case). That’s more of a formal property of python code, rather than something where there’s a "right answer" per an IEEE standard.
sorted([2, float('nan'), 1])
returns [2, nan, 1]
(At least on Activestate Python 3.1 implementation.)
I understand nan
is a weird object, so I wouldn’t be surprised if it shows up in random places in the sort result. But it also messes up the sort for the non-nan numbers in the container, which is really unexpected.
I asked a related question about max
, and based on that I understand why sort
works like this. But should this be considered a bug?
Documentation just says “Return a new sorted list […]” without specifying any details.
EDIT:
I now agree that this isn’t in violation of the IEEE standard. However, it’s a bug from any common sense viewpoint, I think. Even Microsoft, which isn’t known to admit their mistakes often, has recognized this one as a bug, and fixed it in the latest version: http://connect.microsoft.com/VisualStudio/feedback/details/363379/bug-in-list-double-sort-in-list-which-contains-double-nan.
Anyway, I ended up following @khachik’s answer:
sorted(list_, key = lambda x : float('-inf') if math.isnan(x) else x)
I suspect it results in a performance hit compared to the language doing that by default, but at least it works (barring any bugs that I introduced).
IEEE754 is the standard that defines floating point operations in this instance. This standard defines the compare operation of operands, at least one of which is a NaN, to be an error. Hence, this is not a bug. You need to deal with the NaNs before operating on your array.
I’m not sure about the bug, but the workaround may be the following:
sorted(
(2, 1, float('nan')),
lambda x,y: x is float('nan') and -1
or (y is float('nan') and 1
or cmp(x,y)))
which results in:
('nan', 1, 2)
Or remove nan
s before sorting or anything else.
The problem is that there’s no correct order if the list
contains a NAN
, since a sequence a1, a2, a3, ..., an
is sorted if a1 <= a2 <= a3 <= ... <= an
. If any of these a values is a NAN
then the sorted property breaks, since for all a, a <= NAN and NAN <= a
are both false
.
The previous answers are useful, but perhaps not clear regarding the root of the problem.
In any language, sort applies a given ordering, defined by a comparison function or in some other way, over the domain of the input values. For example, less-than, a.k.a. operator <,
could be used throughout if and only if less than defines a suitable ordering over the input values.
But this is specifically NOT true for floating point values and less-than:
“NaN is unordered: it is not equal to, greater than, or less than anything, including itself.” (Clear prose from GNU C manual, but applies to all modern IEEE754
based floating point)
So the possible solutions are:
- remove the NaNs first, making the input domain well defined via <
(or the other sorting function being used)- define a custom comparison function (a.k.a. predicate) that does
define an ordering for NaN, such as less than any number, or greater
than any number.
Either approach can be used, in any language.
Practically, considering python, I would prefer to remove the NaNs if you either don’t care much about fastest performance or if removing NaNs is a desired behavior in context.
Otherwise you could use a suitable predicate function via “cmp” in older python versions, or via this and functools.cmp_to_key()
. The latter is a bit more awkward, naturally, than removing the NaNs first. And care will be required to avoid worse performance, when defining this predicate function.
Assuming you want to keep the NaNs and order them as the lowest “values”, here is a workaround working both with non-unique nan, unique numpy nan, numerical and non numerical objects:
def is_nan(x):
return (x is np.nan or x != x)
list_ = [2, float('nan'), 'z', 1, 'a', np.nan, 4, float('nan')]
sorted(list_, key = lambda x : float('-inf') if is_nan(x) else x)
# [nan, nan, nan, 1, 2, 4, 'a', 'z']
Regardless of standards, there are many cases where a user-defined ordering of float and NA
values is useful. For instance, I was sorting stock returns and wanted highest to lowest with NA
last (since those were irrelevant). There are 4 possible combinations
- Ascending float values,
NA
values last - Ascending float values,
NA
values first - Descending float values,
NA
values last - Descending float values,
NA
values first
Here is a function that covers all scenarios by conditionally replacing NA
values with +/- inf
import math
def sort_with_na(x, reverse=False, na_last=True):
"""Intelligently sort iterable with NA values
For reliable behavior with NA values, we should change the NAs to +/- inf
to guarantee their order rather than relying on the built-in
``sorted(reverse=True)`` which will have no effect. To use the ``reverse``
parameter or other kwargs, use functools.partial in your lambda i.e.
sorted(iterable, key=partial(sort_with_na, reverse=True, na_last=False))
:param x: Element to be sorted
:param bool na_last: Whether NA values should come last or first
:param bool reverse: Return ascending if ``False`` else descending
:return bool:
"""
if not math.isnan(x):
return -x if reverse else x
else:
return float('inf') if na_last else float('-inf')
Testing out each of the 4 combinations
from functools import partial
a = [2, float('nan'), 1]
sorted(a, key=sort_with_na) # Default
sorted(a, key=partial(sort_with_na, reverse=False, na_last=True)) # Ascend, NA last
sorted(a, key=partial(sort_with_na, reverse=False, na_last=False)) # Ascend, NA first
sorted(a, key=partial(sort_with_na, reverse=True, na_last=True)) # Descend, NA last
sorted(a, key=partial(sort_with_na, reverse=True, na_last=False)) # Descend, NA first
A resilient sort involves a compare of 2 items and returning: less, equal, greater.
If cmp(a,b)
is "greater", then cmp(b,a)
must be "less".
If cmp(a,b)
is "zero", then cmp(b,a)
must be "zero".
What is missing in answers to date is the case of comparing 2 float
s which are both NANs and preserving the above properties. 2 NANs should compare as equal or perhaps based on some consistent interpretation of their payloads.
Alternate compare algorithm to put all NAN > +inf
if isnan(a)
if isnan(b)
return 0 (or maybe compare payloads/bit patterns)
return 1
if isnan(b) return 1
if a > b return 1
if a < b return -1
return 0
To recap the problems:
NaN
This always returns False
for every comparison, so it stays where it is in the list:
>>> sorted([float('nan'), 0])
[nan, 0]
>>> sorted([0, float('nan')])
[0, nan]
-0.0
This is == to 0.0
, but has a different repr, a different json representation, and slightly different numerical properties. It has the same problem that the positive and negative zeros will stay in the same order they were in the original list:
>>> sorted([0.0, -0.0])
[0.0, -0.0]
>>> sorted([-0.0, 0.0])
[-0.0, 0.0]
Other solution?
@khachik’s solution has inconsistent sorting behavior for NaN
and -inf
>>> key=lambda x: float('-inf') if math.isnan(x) else x
>>> sorted([float('nan'), float('-inf')], key=key)
[nan, -inf]
>>> sorted([float('-inf'), float('nan')], key=key)
[-inf, nan]
Solution: more complex key function.
So there are problems with signs and nans. We can just include those in a key function:
def stable_float_sort_key(x: float):
return math.copysign(1, x), math.isnan(x), x
That works on all the examples above:
>>> sorted([float('nan'), 0.0], key=stable_float_sort_key)
[0.0, nan]
>>> sorted([0.0, float('nan')], key=stable_float_sort_key)
[0.0, nan]
>>> sorted([float('nan'), float('-inf')], key=stable_float_sort_key)
[-inf, nan]
>>> sorted([float('-inf'), float('nan')], key=stable_float_sort_key)
[-inf, nan]
>>> sorted([0.0, -0.0], key=stable_float_sort_key)
[-0.0, 0.0]
>>> sorted([-0.0, 0.0], key=stable_float_sort_key)
[-0.0, 0.0]
Indeed, you can write a hypothesis test showing it’s consistent across all floating point numbers:
import json
from hypothesis import given, settings
from hypothesis import strategies as st
@given(nums=st.lists(st.floats()), random=st.randoms())
@settings(max_examples=10000)
def test_stable_json_sorting(nums, random):
shuffled = list(nums)
random.shuffle(shuffled)
l1 = sorted(nums, key=stable_float_sort_key)
l2 = sorted(shuffled, key=stable_float_sort_key)
assert json.dumps(l1) == json.dumps(l2)
It does have some oddities however, since some NaN’s are negative! For example:
>>> sorted([float('nan'), -0.0, 0.0, float('-nan')], key=stable_float_sort_key)
[-0.0, nan, 0.0, nan]
If that bothers you, you can fix this by switching the ordering:
def stable_float_sort_key(x: float):
return math.isnan(x), math.copysign(1, x), x
That sorts negative numbers first, followed by positive numbers, followed by NaNs.
Does any of this make sense?
Of course other answerers are correct that in some sense none of this makes sense. Comparison of NaNs is a conceptual error of some kind. However, even in cases where the question doesn’t "make sense", you may want to have invariants like serializing sets of floating point numbers generated by the same code to exactly the same JSON representation, despite hash randomization (my use case). That’s more of a formal property of python code, rather than something where there’s a "right answer" per an IEEE standard.