Sorting in JavaScript: Shouldn't returning a boolean be enough for a comparison function?
TL;DR
I have always successfully sorted my arrays like this
No, you have not. And didn't notice it. A quick counterexample:
> [1,1,0,2].sort(function(a, b){ return a>b })
Array [0, 1, 2, 1]
// in Opera 12. Results may vary between sorting algorithm implementations
why?
Because your comparison function does return false
(or 0
, equivalently) even when b
is larger than a
. But 0
implies that the two elements are considered equal - and the sorting algorithm believes that.
In-Depth Explanation
Comparison functions in JavaScript
How do comparison functions work?
The Array::sort
method can take an optional, custom comparison function as its argument. That function takes two arguments (commonly referred to as a
and b
) which it should compare, and is supposed to return a number
-
> 0
whena
is considered larger thanb
and should be sorted after it -
== 0
whena
is considered equal tob
and it doesn't matter which comes first -
< 0
whena
is considered smaller thanb
and should be sorted before it
If it does not return a number, the result will be cast to a number (which is handy for booleans). The returned number does not need to be exactly -1
or 0
or 1
(though it typically is).
Consistent ordering
To be consistent, the compare function would need to fulfill the equation
comp(a, b) == -1 * comp(b, a)
// or, if values other than -1, 0 and 1 are considered:
comp(a, b) * comp(b, a) <= 0
If that requirement is broken, the sort will behave undefined.
Citing the ES5.1 spec on sort
(same thing in the ES6 spec):
If
comparefn
is […] not a consistent comparison function for the elements of this array, the behaviour of sort is implementation-defined.A function
comparefn
is a consistent comparison function for a set of valuesS
if all of the requirements below are met for all valuesa
,b
, andc
(possibly the same value) in the setS
: The notationa <CF b
meanscomparefn(a,b) < 0
;a =CF b
meanscomparefn(a,b) = 0
(of either sign); anda >CF b
meanscomparefn(a,b) > 0
.Calling
comparefn(a,b)
always returns the same valuev
when given a specific pair of valuesa
andb
as its two arguments. Furthermore,Type(v)
is Number, andv
is notNaN
. Note that this implies that exactly one ofa <CF b
,a =CF b
, anda >CF b
will be true for a given pair ofa
andb
.
- Calling
comparefn(a,b)
does not modify the this object.a =CF a
(reflexivity)- If
a =CF b
, thenb =CF a
(symmetry)- If
a =CF b
andb =CF c
, thena =CF c
(transitivity of=CF
)- If
a <CF b
andb <CF c
, thena <CF c
(transitivity of<CF
)- If
a >CF b
andb >CF c
, thena >CF c
(transitivity of>CF
)NOTE: The above conditions are necessary and sufficient to ensure that
comparefn
divides the setS
into equivalence classes and that these equivalence classes are totally ordered.
Uh, what does this mean? Why should I care?
A sorting algorithm needs to compare items of the array with each other. To do a good and efficient job, it must not need to compare each item to every other, but needs to be able to reason about their ordering. For that to work well, there are a few rules that a custom comparison function needs to abide. A trivial one is that an item a
is equal to itself (compare(a, a) == 0
) - that's the first item in the list above (reflexivity). Yes, this is a bit mathematical, but pays of well.
The most important one is transitivity. It says that when the algorithm has compared two values a
and b
, and also b
with c
, and has found out by applying the comparison function that e.g. a = b
and b < c
, then it can expect that a < c
holds as well. This seems only logical, and is required for a well-defined, consistent ordering.
But your comparison function does fail this. Lets look at this example:
function compare(a, b) { return Number(a > b); }
compare(0, 2) == 0 // ah, 2 and 0 are equal
compare(1, 0) == 1 // ah, 1 is larger than 0
// let's conclude: 1 is also larger than 2
Ooops. And that is why a sorting algorithm can fail (in the spec, this is be "implementation-dependent behaviour" - i.e. unpredictable results) when it is invoked with a comparison function that is not consistent.
Why is the wrong solution so common?
Because in many other languages, there are sorting algorithms that don't expect a three-way comparison but merely a boolean smaller-than operator. C++ std::sort
is a good example of that. It will simply be applied twice with swapped arguments if an equality needs to be determined. Admittedly, this can be more efficient and is less error-prone, but needs more calls to the comparison function if the operator cannot be inlined.
CounterExamples
I have tested my comparison function, and it works!
Only by sheer luck, if you tried some random example. Or because your test suite is flawed - incorrect and/or incomplete.
Here is the small script I used to find the above minimal counterexample:
function perms(n, i, arr, cb) {
// calls callback with all possible arrays of length n
if (i >= n) return cb(arr);
for (var j=0; j<n; j++) {
arr[i] = j;
perms(n, i+1, arr, cb);
}
}
for (var i=2; ; i++) // infinite loop
perms(i, 0, [], function(a) {
if ( a.slice().sort(function(a,b){ return a>b }).toString()
!= a.slice().sort(function(a,b){ return a-b }).toString() )
// you can also console.log() all of them, but remove the loop!
throw a.toString();
});
What comparison function is correct?
Use no comparison function at all, when you want a lexicographic sort. Items in the array will be stringified if necessary.
A generic comparison function that works like the relational operators can be implemented as
function(a, b) {
if (a > b) return 1;
if (a < b) return -1;
/* else */ return 0;
}
With a few tricks, this can be minified to the equivalent function(a,b){return +(a>b)||-(a<b)}
.
For numbers, you can simply return their difference, which abides all the laws above:
function(a, b) {
return a - b; // but make sure only numbers are passed (to avoid NaN)
}
If you want to sort in reverse, just take the appropriate one and swap a
with b
.
If you want to sort composite types (objects etc), replace each a
and each b
with an access of the properties in question, or a method call or whatever you want to sort by.
The sort
function expects a function that expects two arguments a
and b
, and returns:
- A negative number if a comes before b
- A positive number if a comes after b
- Zero if relative order of a and b does not matter
In order to sort numbers in ascending order return a - b
will produce the correct return values; for example:
a b ret
1 2 -1
3 2 1
2 2 0
On the other hand return a > b
produces the following return values:
a b ret implied
1 2 false 0
3 2 true 1
2 2 false 0
In the above example, the sort function is told that 1 and 2 are same (and placing 1 before 2 or 2 before 1 does not matter). This will produce incorrect result, for example (in Chrome 49):
console.log([5, 8, 7, 1, 2, 3, 4, 6, 9, 10, 11, 12, 13].sort(function(a, b) {
return a > b;
}));
// [4, 5, 3, 1, 2, 6, 7, 8, 9, 10, 11, 12, 13]