...one of the most highly
regarded and expertly designed C++ library projects in the
world.
— Herb Sutter and Andrei
Alexandrescu, C++
Coding Standards
Motivation
Synopsis
Function templates description
Definition
Requirements on types
Preconditions
Postconditions
Complexity
Example
Notes
Rationale
Note about performance
Acknowledgements
The minmax library is composed of two headers, <boost/algorithm/minmax.hpp> and <boost/algorithm/minmax_element.hpp>. (See rationale.) The problem solved by this library is that simultaneous min and max computation requires only one comparison, but using std::min and std::max forces two comparisons. Worse, to compute the minimum and maximum elements of a range of n elements requires only 3n/2+1 comparisons, instead of the 2n (in two passes) forced by std::min_element and std::max_element. I always thought it is a waste to have to call two functions to compute the extent of a range, performing two passes over the input, when one should be enough. The present library solves both problems.
The first file implements the function templates minmax as straightforward extensions of the C++ standard. As it returns a pair of const&, we must use the Boost.tuple library to construct such pairs. (Please note: the intent is not to fix the known defaults of std::min and std::max, but to add one more algorithms that combines both; see the rationale.)
The second file implements the function templates minmax_element. In a second part, it also proposes variants that can usually not be computed by the minmax algorithm, and which are more flexible in case some elements are equal. Those variants could have been also provided with policy-based design, but I ruled against that (see rationale).
If you are interested about performance, you will see that minmax_element is just slightly less efficient than a single min_element or max_element, and thus twice as efficient as two separate calls to min_element and max_element. From a theoretical standpoint, all the minmax_element functions perform at most 3n/2+1 comparisons and exactly n increments of the ForwardIterator.
#include <boost/tuple/tuple.hpp> namespace boost { template <class T> tuple<T const&, T const&> > minmax(const T& a, const T& b); template <class T, class BinaryPredicate> tuple<T const&, T const&> > minmax(const T& a, const T& b, BinaryPredicate comp); }
#include <utility> // for std::pair namespace boost { template <class ForwardIterator> std::pair<ForwardIterator,ForwardIterator> minmax_element(ForwardIterator first, ForwardIterator last); template <class ForwardIterator, class BinaryPredicate> std::pair<ForwardIterator,ForwardIterator> minmax_element(ForwardIterator first, ForwardIterator last, BinaryPredicate comp); }In addition, there are a bunch of extensions which specify which element(s) you want to pick in case of equal elements. They are:
The minmax_element is semantically equivalent to first_min_first_max_element.
First_min_element and first_max_element find the smallest and largest elements in the range [first, last). If there are several instance of these elements, the first one is returned. They are identical to std::min_element and std::max_elementand are only included in this library for symmetry.
Last_min_element and last_max_element find the smallest and largest elements in the range [first, last). They are almost identical to std::min_element and std::max_element, except that they return the last instance of the largest element (and not the first, as first_min_element and last_max_element would).
The family of algorithms comprising first_min_first_max_element, first_min_last_max_element, last_min_first_max_element, and last_min_last_max_element can be described generically as follows (using which and what for first or last): which_min_what_max_element finds the (first or last, according to which) smallest element and the (first or last, according to what) largest element in the range [first, last). The first version is semantically equivalent to:
std::make_pair(boost::which_min_element(first,last), boost::what_max_element(first,last)),and the second version to:
std::make_pair(boost::which_min_element(first,last,comp), boost::what_max_element(first,last,comp)).
Note: the first_min_last_max_element can also be described
as finding the first and last elements in the range if it were stably sorted.
For all the other function templates, versions with two template parameters:
The complexity of all the other algorithms is linear. They all perform exactly n increment operations, and zero comparisons if [first,last) is empty, otherwise :
int main() { using namespace std; boost::tuple<int const&, int const&> result1 = boost::minmax(1, 0); assert( result1.get<0>() == 0 ); assert( result1.get<1>() == 1 ); list<int> L; generate_n(front_inserter(L), 1000, rand); typedef list<int>::const_iterator iterator; pair< iterator, iterator > result2 = boost::minmax_element(L.begin(), L.end()); cout << "The smallest element is " << *(result2.first) << endl; cout << "The largest element is " << *(result2.second) << endl; assert( result2.first == std::min_element(L.begin(), L.end()); assert( result2.second == std::max_element(L.begin(), L.end()); }
[2] These algorithms always perform at least 3n/2-2 comparisons, which is a lower bound on the number of comparisons in any case (Cormen, Leiserson, Rivest: "Introduction to Algorithms", section 9.1, Exercise 9.1-). The algorithms essentially compare the elements in pairs, performing 1 comparison for the first two elements, then 3 comparisons for each remaining pair of elements (one to order the elements and one for updating each the minimum and and the maximum). When the number of elements is odd, the last one needs to be compared to the current minimum and also to the current maximum. In addition, for minmax, in cases where equality of the two members in the pair could occur, and the update stores the second, we save the first to check at the end if the update should have stored the first (in case of equality). It's hard to predict if the last comparison is performed or not, hence the at most in both cases.
[3] These algorithms always perform at least 3n/2-2 comparisons, which is a lower bound on the number of comparisons in any case. The method is the same as in note [2] above, and like above, when the number of elements is odd, the last one needs to be compared to the current minimum and also to the current maximum. We can avoid the latter comparison if the former is successful, hence the at most instead of exactly in the odd case.
This was the design originally proposed and approved in the formal review. As the need for Boost.tuple became clear (due to the limitations of std::pair), it became also annoying to require another library for minmax_element which does not need it. Hence the separation into two header files.
I am aware of the problems with std::min and std::max, and all the debate that has been going on (please consult Alexandrescu's paper and the links therein). But I don't see the purpose of this library as fixing something that is part of the C++ standard. I humbly think it's beyond the scope of this library. Rather, I am following the way of the standard in simply providing one more function of the same family. If someone else wants to fix std::min, their fix would probably apply to boost::minmax as well.
In a first version of the library, I proposed _if versions of all the algorithms (well, not all, because that would be too much). However, there is simply no reason to do so, and all the versions I had were just as fast implemented using the excellent <boost/iterator_adaptors.hpp> library. Namely, a call to min_element_if(first, last, pred) would be just as well implemented by:
// equivalent to min_element_if(first, last, pred) min_element(boost::make_filter_iterator(first, last, pred), boost::make_filter_iterator(last, last, pred));Arguably, the min_element_if version is somewhat shorter, but the overhead of iterator adaptors is not large, and they get rid of a lot of code (think of all the combinations between first/last and doubling them with _if variants!).
This rationale is somewhat historical, but explains why there are all these first/last_min/max_element functions.
The C++ standard mandates that std::min_element and std::max_element return the first instance of the smallest and largest elements (as opposed to, say, the last). This arbitrary choice has some consistency: In the case of v of type vector<int>, for instance, it is true that std::min_element(v.begin(),v.end(),std::less<int>()) == std::max_element(v.begin(),v.end(),std::greater<int>()).
There is of course nothing wrong with this: it's simply a matter of choice. Yet another way to specify min_element and max_element is to define them as the first and the last elements if the range was stably sorted. (The stable sort is necessary to disambiguate between iterators that have the same value.) In that case, min should return the first instance and max should return the last. Then, both functions are related by reverse_iterator(std::first_min_element(v.begin(),v.end(),std::less<int>())) == std::last_max_element(v.rbegin(),v.rend(),std::greater<int>()). This definition is subtly different from the previous one.
The definition problem surfaces when one tries to design a minmax_element, using the procedure proposed in (Cormen, Leiserson, Rivest: "Introduction to Algorithms", section 9.1). It should be possible to derive an algorithm using only 3n/2 comparisons if [first,last) has n elements, but if one tries to write a function called first_min_first_max_element() which returns both std::min_element and std::max_element in a pair, the trivial implementation does not work. The problem, rather subtly, is about equal elements: I had to think for a while to find a way to perform only three comparisons per pair and return the first min and first max elements. For a long time, it seemed any attempts at doing so would consume four comparisons per pair in the worst case. This implementation achieves three.
It is not possible (or even desirable) to change the meaning of max_element, but it is still beneficial to provide a function called minmax_element, which returns a pair of min_element and max_element. Although it is easy enough to call min_element and max_element, this performs 2(n-1) comparisons, and necessitates two passes over the input. In contrast, minmax_element will perform the fewer comparisons and perform a single pass over the input. The savings can be significant when the iterator type is not a raw pointer, or even is just a model of the InputIterator concept (although in that case the interface would have to be changed, as the return type could not be copied, so one could e.g. return a value).
In order to benefit from all the variants of the algorithm, I propose to introduce both first_min_element and last_min_element, and their counterparts first_max_element and last_max_element. Then I also propose all the variants algorithms: first_min_last_max_element and last_min_first_max_element, which perform only at most 3n/2 comparisons, and only a single pass on the input. In fact, it can be proven that computing minmax requires at least 3(n/2)-2 comparisons in any instance of the problem (Cormen, Leiserson, Rivest, 2nd edition, section 9.1). The implementation I give does not perform unnecessary comparisons (whose result could have been computed by transitivity from previous comparisons).
It appears that first_min_last_max_element may be just a tad slower than first_min_element alone, still much less than first_min_element and last_max_element called separately. [2]
The minmax algorithms are useful in computing the extent of a range. In computer graphics, we need a bounding box of a set of objects. In that case the need for a single pass is even more stringent as all three directions must be done at once. Food for thoughts: there is matter for a nice generic programming library with stackable update_min and update_max function objects which store a reference to the min_resultand max_result variables, in conjunction with the for_each algorithm).
I believe many standard sequential algorithms could be reformulated with accumulators (and many others, such as in statistics, expectation / variance / etc.). It seems that there is room for another library, but I do not see it competing with minmax, rather extending several algorithms (including minmax) to the accumulator framework. However, I felt it is beyond the scope of this library to provide such accumulators.
True, and I could have gone that way, with the default policy for min_element and max_element to pick the first occurence of the result. This would have thinned the number of combinations of the minmax_element variants. But it would also have meant to change the interface of boost::minmax_element. One of the goals of the minmax_element algorithm is its eventual addition to the C++ standard, in connection with std::min_element and std::max_element (and I feel that it would be quite natural given the shortness of the implementation, and the not quite trivial detail which is needed to get it right). So changing the interface by adding policies would have meant unfortunately to depart from the standard and created an obstacle towards that goal. Besides, the code remains rather readable and simple without policies. So I am quite happy to keep it like this.
One minmax_element implementation, which performs 3(n/2)+O(log n) comparisons on the average when the elements are random_shuffled, was suggested by my student Marc Glisse. The current one, which performs 3(n/2)+1 comparisons in the worst case, was suggested by John Iacono.
Finally, Matthew Wilson and Jeremy Siek contributed pre-review comments, while Gennadiy Rozental, John Maddock, Craig Henderson, Gary Powell participated in the review of the library, managed by Thomas Witt. In particular, Gennadiy suggested a factorization of the code; while I haven't followed it all the way, his suggestions do make the code more readable and still work with older compilers. Late after the review, as I finally scrounged to add the library for a release, Eric Niebler noted the bad behavior of std::pair for minmax and suggested to use Boost.tuple instead. All my thanks for the excellent advice and reviews from all.
© Copyright Hervé Brönnimann, Polytechnic University, 2002--2004. Use, modification, and distribution is subject to the Boost Software License, Version 1.0. (See accompanying file License_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)