比较两个双精度浮点数或两个浮点数最有效的方法是什么?

简单地这样做是不正确的:

bool CompareDoubles1 (double A, double B)
{
   return A == B;
}

比如:

bool CompareDoubles2 (double A, double B) 
{
   diff = A - B;
   return (diff < EPSILON) && (-diff < EPSILON);
}

似乎是浪费加工。

有人知道更聪明的浮点比较器吗?


当前回答

正如其他人所指出的那样,使用固定指数(例如0.0000001)对于远离该值的值是无用的。例如,如果你的两个值是10000.000977和10000,那么这两个数字之间没有32位浮点值——10000和10000.000977是你可能得到的最接近的值,而不是位对位相同。这里,小于0.0009是没有意义的;你也可以使用直接等式运算符。

同样地,当两个值的大小接近ε时,相对误差增长到100%。

Thus, trying to mix a fixed point number such as 0.00001 with floating-point values (where the exponent is arbitrary) is a pointless exercise. This will only ever work if you can be assured that the operand values lie within a narrow domain (that is, close to some specific exponent), and if you properly select an epsilon value for that specific test. If you pull a number out of the air ("Hey! 0.00001 is small, so that must be good!"), you're doomed to numerical errors. I've spent plenty of time debugging bad numerical code where some poor schmuck tosses in random epsilon values to make yet another test case work.

如果你从事任何类型的数值编程,并认为你需要达到定点的epsilon,请阅读BRUCE关于比较浮点数的文章。

浮点数比较

其他回答

意识到这是一个老话题,但这篇文章是我发现的关于比较浮点数的最直接的文章之一,如果你想探索更多,它也有更详细的参考资料,它的主要站点涵盖了处理浮点数的完整范围的问题《浮点指南:比较》。

我们可以在浮点公差中找到一篇更实用的文章,并指出有绝对公差测试,在c++中归结为:

bool absoluteToleranceCompare(double x, double y)
{
    return std::fabs(x - y) <= std::numeric_limits<double>::epsilon() ;
}

及相对耐量试验:

bool relativeToleranceCompare(double x, double y)
{
    double maxXY = std::max( std::fabs(x) , std::fabs(y) ) ;
    return std::fabs(x - y) <= std::numeric_limits<double>::epsilon()*maxXY ;
}

文章指出,当x和y较大时,绝对检验失败;当x和y较小时,相对检验失败。假设绝对耐受性和相对耐受性是相同的,综合测试将是这样的:

bool combinedToleranceCompare(double x, double y)
{
    double maxXYOne = std::max( { 1.0, std::fabs(x) , std::fabs(y) } ) ;

    return std::fabs(x - y) <= std::numeric_limits<double>::epsilon()*maxXYOne ;
}

你必须为浮点数比较做这个处理,因为浮点数不能像整数类型那样完美地比较。下面是各种比较运算符的函数。

浮点数等于(==)

我也更喜欢减法技术,而不是依赖于fabs()或abs(),但我必须在从64位PC到ATMega328微控制器(Arduino)的各种架构上快速配置它,才能真正看到它是否会产生很大的性能差异。

所以,让我们忘记这些绝对值的东西,只做一些减法和比较!

从微软的例子修改如下:

/// @brief      See if two floating point numbers are approximately equal.
/// @param[in]  a        number 1
/// @param[in]  b        number 2
/// @param[in]  epsilon  A small value such that if the difference between the two numbers is
///                      smaller than this they can safely be considered to be equal.
/// @return     true if the two numbers are approximately equal, and false otherwise
bool is_float_eq(float a, float b, float epsilon) {
    return ((a - b) < epsilon) && ((b - a) < epsilon);
}
bool is_double_eq(double a, double b, double epsilon) {
    return ((a - b) < epsilon) && ((b - a) < epsilon);
}

使用示例:

constexpr float EPSILON = 0.0001; // 1e-4
is_float_eq(1.0001, 0.99998, EPSILON);

我不完全确定,但在我看来,对基于epsilon的方法的一些批评,正如这个高度好评的答案下面的评论所描述的那样,可以通过使用变量epsilon来解决,根据比较的浮点值缩放,像这样:

float a = 1.0001;
float b = 0.99998;
float epsilon = std::max(std::fabs(a), std::fabs(b)) * 1e-4;

is_float_eq(a, b, epsilon);

通过这种方式,epsilon值随浮点值伸缩,因此它的值不会小到不重要。

为了完整起见,让我们添加剩下的:

大于(>)小于(<):

/// @brief      See if floating point number `a` is > `b`
/// @param[in]  a        number 1
/// @param[in]  b        number 2
/// @param[in]  epsilon  a small value such that if `a` is > `b` by this amount, `a` is considered
///             to be definitively > `b`
/// @return     true if `a` is definitively > `b`, and false otherwise
bool is_float_gt(float a, float b, float epsilon) {
    return a > b + epsilon;
}
bool is_double_gt(double a, double b, double epsilon) {
    return a > b + epsilon;
}

/// @brief      See if floating point number `a` is < `b`
/// @param[in]  a        number 1
/// @param[in]  b        number 2
/// @param[in]  epsilon  a small value such that if `a` is < `b` by this amount, `a` is considered
///             to be definitively < `b`
/// @return     true if `a` is definitively < `b`, and false otherwise
bool is_float_lt(float a, float b, float epsilon) {
    return a < b - epsilon;
}
bool is_double_lt(double a, double b, double epsilon) {
    return a < b - epsilon;
}

大于或等于(>=),小于或等于(<=)

/// @brief      Returns true if `a` is definitively >= `b`, and false otherwise
bool is_float_ge(float a, float b, float epsilon) {
    return a > b - epsilon;
}
bool is_double_ge(double a, double b, double epsilon) {
    return a > b - epsilon;
}

/// @brief      Returns true if `a` is definitively <= `b`, and false otherwise
bool is_float_le(float a, float b, float epsilon) {
    return a < b + epsilon;
}
bool is_double_le(double a, double b, double epsilon) {
    return a < b + epsilon;
}

额外的改进:

A good default value for epsilon in C++ is std::numeric_limits<T>::epsilon(), which evaluates to either 0 or FLT_EPSILON, DBL_EPSILON, or LDBL_EPSILON. See here: https://en.cppreference.com/w/cpp/types/numeric_limits/epsilon. You can also see the float.h header for FLT_EPSILON, DBL_EPSILON, and LDBL_EPSILON. See https://en.cppreference.com/w/cpp/header/cfloat and https://www.cplusplus.com/reference/cfloat/ You could template the functions instead, to handle all floating point types: float, double, and long double, with type checks for these types via a static_assert() inside the template. Scaling the epsilon value is a good idea to ensure it works for really large and really small a and b values. This article recommends and explains it: http://realtimecollisiondetection.net/blog/?p=89. So, you should scale epsilon by a scaling value equal to max(1.0, abs(a), abs(b)), as that article explains. Otherwise, as a and/or b increase in magnitude, the epsilon would eventually become so small relative to those values that it becomes lost in the floating point error. So, we scale it to become larger in magnitude like they are. However, using 1.0 as the smallest allowed scaling factor for epsilon also ensures that for really small-magnitude a and b values, epsilon itself doesn't get scaled so small that it also becomes lost in the floating point error. So, we limit the minimum scaling factor to 1.0. If you want to "encapsulate" the above functions into a class, don't. Instead, wrap them up in a namespace if you like in order to namespace them. Ex: if you put all of the stand-alone functions into a namespace called float_comparison, then you could access the is_eq() function like this, for instance: float_comparison::is_eq(1.0, 1.5);. It might also be nice to add comparisons against zero, not just comparisons between two values. So, here is a better type of solution with the above improvements in place: namespace float_comparison { /// Scale the epsilon value to become large for large-magnitude a or b, /// but no smaller than 1.0, per the explanation above, to ensure that /// epsilon doesn't ever fall out in floating point error as a and/or b /// increase in magnitude. template<typename T> static constexpr T scale_epsilon(T a, T b, T epsilon = std::numeric_limits<T>::epsilon()) noexcept { static_assert(std::is_floating_point_v<T>, "Floating point comparisons " "require type float, double, or long double."); T scaling_factor; // Special case for when a or b is infinity if (std::isinf(a) || std::isinf(b)) { scaling_factor = 0; } else { scaling_factor = std::max({(T)1.0, std::abs(a), std::abs(b)}); } T epsilon_scaled = scaling_factor * std::abs(epsilon); return epsilon_scaled; } // Compare two values /// Equal: returns true if a is approximately == b, and false otherwise template<typename T> static constexpr bool is_eq(T a, T b, T epsilon = std::numeric_limits<T>::epsilon()) noexcept { static_assert(std::is_floating_point_v<T>, "Floating point comparisons " "require type float, double, or long double."); // test `a == b` first to see if both a and b are either infinity // or -infinity return a == b || std::abs(a - b) <= scale_epsilon(a, b, epsilon); } /* etc. etc.: is_eq() is_ne() is_lt() is_le() is_gt() is_ge() */ // Compare against zero /// Equal: returns true if a is approximately == 0, and false otherwise template<typename T> static constexpr bool is_eq_zero(T a, T epsilon = std::numeric_limits<T>::epsilon()) noexcept { static_assert(std::is_floating_point_v<T>, "Floating point comparisons " "require type float, double, or long double."); return is_eq(a, (T)0.0, epsilon); } /* etc. etc.: is_eq_zero() is_ne_zero() is_lt_zero() is_le_zero() is_gt_zero() is_ge_zero() */ } // namespace float_comparison

参见:

The macro forms of some of the functions above in my repo here: utilities.h. UPDATE 29 NOV 2020: it's a work-in-progress, and I'm going to make it a separate answer when ready, but I've produced a better, scaled-epsilon version of all of the functions in C in this file here: utilities.c. Take a look. ADDITIONAL READING I need to do now have done: Floating-point tolerances revisited, by Christer Ericson. VERY USEFUL ARTICLE! It talks about scaling epsilon in order to ensure it never falls out in floating point error, even for really large-magnitude a and/or b values!

在数值软件中,确实有这样的情况,你需要检查两个浮点数是否完全相等。我就一个类似的问题发表了这篇文章

https://stackoverflow.com/a/10973098/1447411

所以你不能说“CompareDoubles1”是错误的。

就数量的规模而言:

如果在某种物理意义上,ε是量的大小(即相对值)的一小部分,而A和B类型在同一意义上具有可比性,那么我认为,下面的观点是相当正确的:

#include <limits>
#include <iomanip>
#include <iostream>

#include <cmath>
#include <cstdlib>
#include <cassert>

template< typename A, typename B >
inline
bool close_enough(A const & a, B const & b,
                  typename std::common_type< A, B >::type const & epsilon)
{
    using std::isless;
    assert(isless(0, epsilon)); // epsilon is a part of the whole quantity
    assert(isless(epsilon, 1));
    using std::abs;
    auto const delta = abs(a - b);
    auto const x = abs(a);
    auto const y = abs(b);
    // comparable generally and |a - b| < eps * (|a| + |b|) / 2
    return isless(epsilon * y, x) && isless(epsilon * x, y) && isless((delta + delta) / (x + y), epsilon);
}

int main()
{
    std::cout << std::boolalpha << close_enough(0.9, 1.0, 0.1) << std::endl;
    std::cout << std::boolalpha << close_enough(1.0, 1.1, 0.1) << std::endl;
    std::cout << std::boolalpha << close_enough(1.1,    1.2,    0.01) << std::endl;
    std::cout << std::boolalpha << close_enough(1.0001, 1.0002, 0.01) << std::endl;
    std::cout << std::boolalpha << close_enough(1.0, 0.01, 0.1) << std::endl;
    return EXIT_SUCCESS;
}

使用任何其他建议都要非常小心。这完全取决于上下文。

我花了很长时间在一个系统中追踪错误,该系统假设|a-b|<epsilon,则a==b。潜在的问题是:

The implicit presumption in an algorithm that if a==b and b==c then a==c. Using the same epsilon for lines measured in inches and lines measured in mils (.001 inch). That is a==b but 1000a!=1000b. (This is why AlmostEqual2sComplement asks for the epsilon or max ULPS). The use of the same epsilon for both the cosine of angles and the length of lines! Using such a compare function to sort items in a collection. (In this case using the builtin C++ operator == for doubles produced correct results.)

就像我说的,这完全取决于上下文和a和b的预期大小。

顺便说一下,std::numeric_limits<double>::epsilon()是“机器epsilon”。它是1.0和下一个用double表示的值之间的差值。我猜它可以用在比较函数中,但只有当期望值小于1时。(这是对@cdv的回答的回应…)

同样,如果你的int算术是双精度的(这里我们在某些情况下使用双精度来保存int值),你的算术是正确的。例如,4.0/2.0将等同于1.0+1.0。只要你不做导致分数(4.0/3.0)的事情,或者不超出int的大小。