<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
"http://www.w3.org/TR/html4/loose.dtd">
<html>
<head>
<meta http-equiv="Content-Language" content="en-us">
<meta http-equiv="Content-Type" content="text/html; charset=us-ascii">
<link rel="stylesheet" type="text/css" href="../../../../boost.css">
<title>Boost Interval Arithmetic Library</title>
</head>
<body lang="en">
<h1><img src="../../../../boost.png" alt="boost.png (6897 bytes)" align=
"middle"> Interval Arithmetic Library</h1>
<center>
<table width="80%" summary="">
<tbody>
<tr>
<td><b>Contents of this page:</b><br>
<a href="#intro">Introduction</a><br>
<a href="#synopsis">Synopsis</a><br>
<a href="#interval">Template class <code>interval</code></a><br>
<a href="#opers">Operations and functions</a><br>
<a href="#interval_lib">Interval support library</a><br>
<!--<a href="#compil">Compilation notes</a><br>-->
<a href="#dangers">Common pitfalls and dangers</a><br>
<a href="#rationale">Rationale</a><br>
<a href="#acks">History and Acknowledgments</a></td>
<td><b>Other pages associated with this page:</b><br>
<a href="rounding.htm">Rounding policies</a><br>
<a href="checking.htm">Checking policies</a><br>
<a href="policies.htm">Policies manipulation</a><br>
<a href="comparisons.htm">Comparisons</a><br>
<a href="numbers.htm">Base number type requirements</a><br>
<a href="guide.htm">Choosing your own interval type</a><br>
<a href="examples.htm">Test and example programs</a><br>
<a href="includes.htm">Headers inclusion</a><br>
<a href="todo.htm">Some items on the todo list</a></td>
</tr>
</tbody>
</table>
</center>
<h2 id="intro">Introduction and Overview</h2>
<p>As implied by its name, this library is intended to help manipulating
mathematical intervals. It consists of a single header <<a href=
"../../../../boost/numeric/interval.hpp">boost/numeric/interval.hpp</a>>
and principally a type which can be used as <code>interval<T></code>.
In fact, this interval template is declared as
<code>interval<T,Policies></code> where <code>Policies</code> is a
policy class that controls the various behaviours of the interval class;
<code>interval<T></code> just happens to pick the default policies
for the type <code>T</code>.</p>
<p><span style="color: #FF0000; font-weight: bold">Warning!</span>
Guaranteed interval arithmetic for native floating-point format is not
supported on every combination of processor, operating system, and
compiler. This is a list of systems known to work correctly when using
<code>interval<float></code> and <code>interval<double></code>
with basic arithmetic operators.</p>
<ul>
<li>x86-like hardware is supported by the library with GCC, Visual C++
≥ 7.1, Intel compiler (≥ 8 on Windows), CodeWarrior (≥ 9), as
long as the traditional x87 floating-point unit is used for
floating-point computations (no <code>-mfpmath=sse2</code> support).</li>
<li>Sparc hardware is supported with GCC and Sun compiler.</li>
<li>PowerPC hardware is supported with GCC and CodeWarrior, when
floating-point computations are not done with the Altivec unit.</li>
<li>Alpha hardware is supported with GCC, except maybe for the square
root. The options <code>-mfp-rounding-mode=d -mieee</code> have to be
used.</li>
</ul>
<p>The previous list is not exhaustive. And even if a system does not
provide guaranteed computations for hardware floating-point types, the
interval library is still usable with user-defined types and for doing box
arithmetic.</p>
<h3>Interval Arithmetic</h3>
<p>An interval is a pair of numbers which represents all the numbers
between these two. (Intervals are considered closed so the bounds are
included.) The purpose of this library is to extend the usual arithmetic
functions to intervals. These intervals will be written [<i>a</i>,<i>b</i>]
to represent all the numbers between <i>a</i> and <i>b</i> (included).
<i>a</i> and <i>b</i> can be infinite (but they can not be the same
infinite) and <i>a</i> ≤ <i>b</i>.</p>
<p>The fundamental property of interval arithmetic is the
<em><strong>inclusion property</strong></em>:</p>
<dl>
<dd>``if <i>f</i> is a function on a set of numbers, <i>f</i> can be
extended to a new function defined on intervals. This new function
<i>f</i> takes one interval argument and returns an interval result such
as: ∀ <i>x</i> ∈ [<i>a</i>,<i>b</i>], <i>f</i>(<i>x</i>)
∈ <i>f</i>([<i>a</i>,<i>b</i>]).''</dd>
</dl>
<p>Such a property is not limited to functions with only one argument.
Whenever possible, the interval result should be the smallest one able to
satisfy the property (it is not really useful if the new functions always
answer [-∞,+∞]).</p>
<p>There are at least two reasons a user would like to use this library.
The obvious one is when the user has to compute with intervals. One example
is when input data have some builtin imprecision: instead of a number, an
input variable can be passed as an interval. Another example application is
to solve equations, by bisecting an interval until the interval width is
small enough. A third example application is in computer graphics, where
computations with boxes, segments or rays can be reduced to computations
with points via intervals.</p>
<p>Another common reason to use interval arithmetic is when the computer
doesn't produce exact results: by using intervals, it is possible to
quantify the propagation of rounding errors. This approach is used often in
numerical computation. For example, let's assume the computer stores
numbers with ten decimal significant digits. To the question 1 + 1E-100 -
1, the computer will answer 0 although the correct answer would be 1E-100.
With the help of interval arithmetic, the computer will answer [0,1E-9].
This is quite a huge interval for such a little result, but the precision
is now known, without having to compute error propagation.</p>
<h3>Numbers, rounding, and exceptional behavior</h3>
<p>The <em><strong>base number type</strong></em> is the type that holds
the bounds of the interval. In order to successfully use interval
arithmetic, the base number type must present some <a href=
"rounding.htm">characteristics</a>. Firstly, due to the definition of an
interval, the base numbers have to be totally ordered so, for instance,
<code>complex<T></code> is not usable as base number type for
intervals. The mathematical functions for the base number type should also
be compatible with the total order (for instance if x>y and z>t, then
it should also hold that x+z > y+t), so modulo types are not usable
either.</p>
<p>Secondly, the computations must be exact or provide some rounding
methods (for instance, toward minus or plus infinity) if we want to
guarantee the inclusion property. Note that we also may explicitely specify
no rounding, for instance if the base number type is exact, i.e. the result
of a mathematic operation is always computed and represented without loss
of precision. If the number type is not exact, we may still explicitely
specify no rounding, with the obvious consequence that the inclusion
property is no longuer guaranteed.</p>
<p>Finally, because heavy loss of precision is always possible, some
numbers have to represent infinities or an exceptional behavior must be
defined. The same situation also occurs for NaN (<i>Not a Number</i>).</p>
<p>Given all this, one may want to limit the template argument T of the
class template <code>interval</code> to the floating point types
<code>float</code>, <code>double</code>, and <code>long double</code>, as
defined by the IEEE-754 Standard. Indeed, if the interval arithmetic is
intended to replace the arithmetic provided by the floating point unit of a
processor, these types are the best choice. Unlike
<code>std::complex</code>, however, we don't want to limit <code>T</code>
to these types. This is why we allow the rounding and exceptional behaviors
to be given by the two policies (rounding and checking). We do nevertheless
provide highly optimized rounding and checking class specializations for
the above-mentioned floating point types.</p>
<h3>Operations and functions</h3>
<p>It is straightforward to define the elementary arithmetic operations on
intervals, being guided by the inclusion property. For instance, if [a,b]
and [c,d] are intervals, [a,b]+[c,d] can be computed by taking the smallest
interval that contains all the numbers x+y for x in [a,b] and y in [c,d];
in this case, rounding a+c down and b+d up will suffice. Other operators
and functions are similarly defined (see their definitions below).</p>
<h3>Comparisons</h3>
<p>It is also possible to define some comparison operators. Given two
intervals, the result is a tri-state boolean type
{<i>false</i>,<i>true,indeterminate</i>}. The answers <i>false</i> and
<i>true</i> are easy to manipulate since they can directly be mapped on the
boolean <i>true</i> and <i>false</i>. But it is not the case for the answer
<em>indeterminate</em> since comparison operators are supposed to be
boolean functions. So, what to do in order to obtain boolean answers?</p>
<p>One solution consists of deciding to adopt an exceptional behavior, such
as a failed assertion or raising an exception. In this case, the
exceptional behavior will be triggered when the result is
indeterminate.</p>
<p>Another solution is to map <em>indeterminate</em> always to
<i>false,</i> or always to <i>true</i>. If <i>false</i> is chosen, the
comparison will be called "<i>certain</i>;" indeed, the result of
[<i>a</i>,<i>b</i>] < [<i>c</i>,<i>d</i>] will be <i>true</i> if and
only if: ∀ <i>x</i> ∈ [<i>a</i>,<i>b</i>] ∀ <i>y</i>
∈ [<i>c</i>,<i>d</i>], <i>x</i> < <i>y</i>. If <i>true</i> is
chosen, the comparison will be called "<i>possible</i>;" indeed, the result
of [<i>a</i>,<i>b</i>] < [<i>c</i>,<i>d</i>] will be <i>true</i> if and
only if: ∃ <i>x</i> ∈ [<i>a</i>,<i>b</i>] ∃ <i>y</i>
∈ [<i>c</i>,<i>d</i>], <i>x</i> < <i>y</i>.</p>
<p>Since any of these solution has a clearly defined semantics, it is not
clear that we should enforce either of them. For this reason, the default
behavior consists to mimic the real comparisons by throwing an exception in
the indeterminate case. Other behaviors can be selected bu using specific
comparison namespace. There is also a bunch of explicitely named comparison
functions. See <a href="comparisons.htm">comparisons</a> pages for further
details.</p>
<h3>Overview of the library, and usage</h3>
<p>This library provides two quite distinct levels of usage. One is to use
the basic class template <code>interval<T></code> without specifying
the policy. This only requires to know and understand the concepts
developed above and the content of the namespace boost. In addition to the
class <code>interval<T></code>, this level of usage provides
arithmetic operators (<code>+</code>, <code>-</code>, <code>*</code>,
<code>/</code>), algebraic and piecewise-algebraic functions
(<code>abs</code>, <code>square</code>, <code>sqrt</code>,
<code>pow</code>), transcendental and trigonometric functions
(<code>exp</code>, <code>log</code>, <code>sin</code>, <code>cos</code>,
<code>tan</code>, <code>asin</code>, <code>acos</code>, <code>atan</code>,
<code>sinh</code>, <code>cosh</code>, <code>tanh</code>,
<code>asinh</code>, <code>acosh</code>, <code>atanh</code>), and the
standard comparison operators (<code><</code>, <code><=</code>,
<code>></code>, <code>>=</code>, <code>==</code>, <code>!=</code>),
as well as several interval-specific functions (<code>min</code>,
<code>max</code>, which have a different meaning than <code>std::min</code>
and <code>std::max</code>; <code>lower</code>, <code>upper</code>,
<code>width</code>, <code>median</code>, <code>empty</code>,
<code>singleton</code>, <code>equal</code>, <code>in</code>,
<code>zero_in</code>, <code>subset</code>, <code>proper_subset</code>,
<code>overlap</code>, <code>intersection</code>, <code>hull</code>,
<code>bisect</code>).</p>
<p>For some functions which take several parameters of type
<code>interval<T></code>, all combinations of argument types
<code>T</code> and <code>interval<T></code> which contain at least
one <code>interval<T></code>, are considered in order to avoid a
conversion from the arguments of type <code>T</code> to a singleton of type
<code>interval<T></code>. This is done for efficiency reasons (the
fact that an argument is a singleton sometimes renders some tests
unnecessary).</p>
<p>A somewhat more advanced usage of this library is to hand-pick the
policies <code>Rounding</code> and <code>Checking</code> and pass them to
<code>interval<T, Policies></code> through the use of <code>Policies
:= boost::numeric::interval_lib::policies<Rounding,Checking></code>.
Appropriate policies can be fabricated by using the various classes
provided in the namespace <code>boost::numeric::interval_lib</code> as
detailed in section <a href="#interval_lib">Interval Support Library</a>.
It is also possible to choose the comparison scheme by overloading
operators through namespaces.</p>
<h2><a name="synopsis" id="synopsis"></a>Synopsis</h2>
<pre>
namespace boost {
namespace numeric {
namespace interval_lib {
/* this declaration is necessary for the declaration of interval */
template <class T> struct default_policies;
/* ... ; the full synopsis of namespace interval_lib can be found <a href=
"#interval_lib">here</a> */
} // namespace interval_lib
/* template interval_policies; class definition can be found <a href=
"policies.htm">here</a> */
template<class Rounding, class Checking>
struct interval_policies;
/* template class interval; class definition can be found <a href=
"#interval">here</a> */
template<class T, class Policies = typename interval_lib::default_policies<T>::type > class interval;
/* arithmetic operators involving intervals */
template <class T, class Policies> interval<T, Policies> operator+(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> operator-(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> operator+(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> operator+(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> interval<T, Policies> operator+(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> operator-(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> operator-(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> interval<T, Policies> operator-(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> operator*(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> operator*(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> interval<T, Policies> operator*(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> operator/(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> operator/(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> interval<T, Policies> operator/(const T& r, const interval<T, Policies>& x);
/* algebraic functions: sqrt, abs, square, pow, nth_root */
template <class T, class Policies> interval<T, Policies> abs(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> sqrt(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> square(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> pow(const interval<T, Policies>& x, int y);
template <class T, class Policies> interval<T, Policies> nth_root(const interval<T, Policies>& x, int y);
/* transcendental functions: exp, log */
template <class T, class Policies> interval<T, Policies> exp(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> log(const interval<T, Policies>& x);
/* fmod, for trigonometric function argument reduction (see below) */
template <class T, class Policies> interval<T, Policies> fmod(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> fmod(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> interval<T, Policies> fmod(const T& x, const interval<T, Policies>& y);
/* trigonometric functions */
template <class T, class Policies> interval<T, Policies> sin(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> cos(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> tan(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> asin(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> acos(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> atan(const interval<T, Policies>& x);
/* hyperbolic trigonometric functions */
template <class T, class Policies> interval<T, Policies> sinh(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> cosh(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> tanh(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> asinh(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> acosh(const interval<T, Policies>& x);
template <class T, class Policies> interval<T, Policies> atanh(const interval<T, Policies>& x);
/* min, max external functions (NOT std::min/max, see below) */
template <class T, class Policies> interval<T, Policies> max(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> max(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> interval<T, Policies> max(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> min(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> min(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> interval<T, Policies> min(const T& x, const interval<T, Policies>& y);
/* bounds-related interval functions */
template <class T, class Policies> T lower(const interval<T, Policies>& x);
template <class T, class Policies> T upper(const interval<T, Policies>& x);
template <class T, class Policies> T width(const interval<T, Policies>& x);
template <class T, class Policies> T median(const interval<T, Policies>& x);
template <class T, class Policies> T norm(const interval<T, Policies>& x);
/* bounds-related interval functions */
template <class T, class Policies> bool empty(const interval<T, Policies>& b);
template <class T, class Policies> bool singleton(const interval<T, Policies>& x);
template <class T, class Policies> bool equal(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool in(const T& r, const interval<T, Policies>& b);
template <class T, class Policies> bool zero_in(const interval<T, Policies>& b);
template <class T, class Policies> bool subset(const interval<T, Policies>& a, const interval<T, Policies>& b);
template <class T, class Policies> bool proper_subset(const interval<T, Policies>& a, const interval<T, Policies>& b);
template <class T, class Policies> bool overlap(const interval<T, Policies>& x, const interval<T, Policies>& y);
/* set manipulation interval functions */
template <class T, class Policies> interval<T, Policies> intersection(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> hull(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> hull(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> interval<T, Policies> hull(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> interval<T, Policies> hull(const T& x, const T& y);
template <class T, class Policies> std::pair<interval<T, Policies>, interval<T, Policies> > bisect(const interval<T, Policies>& x);
/* interval comparison operators */
template<class T, class Policies> bool operator<(const interval<T, Policies>& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator<(const interval<T, Policies>& x, const T& y);
template<class T, class Policies> bool operator<(const T& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator<=(const interval<T, Policies>& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator<=(const interval<T, Policies>& x, const T& y);
template<class T, class Policies> bool operator<=(const T& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator>(const interval<T, Policies>& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator>(const interval<T, Policies>& x, const T& y);
template<class T, class Policies> bool operator>(const T& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator>=(const interval<T, Policies>& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator>=(const interval<T, Policies>& x, const T& y);
template<class T, class Policies> bool operator>=(const T& x, const interval<T, Policies>& y);
</pre>
<pre>
template<class T, class Policies> bool operator==(const interval<T, Policies>& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator==(const interval<T, Policies>& x, const T& y);
template<class T, class Policies> bool operator==(const T& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator!=(const interval<T, Policies>& x, const interval<T, Policies>& y);
template<class T, class Policies> bool operator!=(const interval<T, Policies>& x, const T& y);
template<class T, class Policies> bool operator!=(const T& x, const interval<T, Policies>& y);
namespace interval_lib {
template<class T, class Policies> interval<T, Policies> division_part1(const interval<T, Policies>& x, const interval<T, Policies& y, bool& b);
template<class T, class Policies> interval<T, Policies> division_part2(const interval<T, Policies>& x, const interval<T, Policies& y, bool b = true);
template<class T, class Policies> interval<T, Policies> multiplicative_inverse(const interval<T, Policies>& x);
template<class I> I add(const typename I::base_type& x, const typename I::base_type& y);
template<class I> I sub(const typename I::base_type& x, const typename I::base_type& y);
template<class I> I mul(const typename I::base_type& x, const typename I::base_type& y);
template<class I> I div(const typename I::base_type& x, const typename I::base_type& y);
} // namespace interval_lib
} // namespace numeric
} // namespace boost
</pre>
<h2><a name="interval" id="interval"></a>Template class
<code>interval</code></h2>The public interface of the template class
interval itself is kept at a simplest minimum:
<pre>
template <class T, class Policies = typename interval_lib::default_policies<T>::type>
class interval
{
public:
typedef T base_type;
typedef Policies traits_type;
interval();
interval(T const &v);
template<class T1> interval(T1 const &v);
interval(T const &l, T const &u);
template<class T1, class T2> interval(T1 const &l, T2 const &u);
interval(interval<T, Policies> const &r);
template<class Policies1> interval(interval<T, Policies1> const &r);
template<class T1, class Policies1> interval(interval<T1, Policies1> const &r);
interval &operator=(T const &v);
template<class T1> interval &operator=(T1 const &v);
interval &operator=(interval<T, Policies> const &r);
template<class Policies1> interval &operator=(interval<T, Policies1> const &r);
template<class T1, class Policies1> interval &operator=(interval<T1, Policies1> const &r);
void assign(T const &l, T const &u);
T const &lower() const;
T const &upper() const;
static interval empty();
static interval whole();
static interval hull(T const &x, T const &y);
interval& operator+= (T const &r);
interval& operator-= (T const &r);
interval& operator*= (T const &r);
interval& operator/= (T const &r);
interval& operator+= (interval const &r);
interval& operator-= (interval const &r);
interval& operator*= (interval const &r);
interval& operator/= (interval const &r);
};
</pre>
<p>The constructors create an interval enclosing their arguments. If there
are two arguments, the first one is assumed to be the left bound and the
second one is the right bound. Consequently, the arguments need to be
ordered. If the property !(l <= u) is not respected, the checking policy
will be used to create an empty interval. If no argument is given, the
created interval is the singleton zero.</p>
<p>If the type of the arguments is the same as the base number type, the
values are directly used for the bounds. If it is not the same type, the
library will use the rounding policy in order to convert the arguments
(<code>conv_down</code> and <code>conv_up</code>) and create an enclosing
interval. When the argument is an interval with a different policy, the
input interval is checked in order to correctly propagate its emptiness (if
empty).</p>
<p>The assignment operators behave similarly, except they obviously take
one argument only. There is also an <code>assign</code> function in order
to directly change the bounds of an interval. It behaves like the
two-arguments constructors if the bounds are not ordered. There is no
assign function that directly takes an interval or only one number as a
parameter; just use the assignment operators in such a case.</p>
<p>The type of the bounds and the policies of the interval type define the
set of values the intervals contain. E.g. with the default policies,
intervals are subsets of the set of real numbers. The static functions
<code>empty</code> and <code>whole</code> produce the intervals/subsets
that are repectively the empty subset and the whole set. They are static
member functions rather than global functions because they cannot guess
their return types. Likewise for <code>hull</code>. <code>empty</code> and
<code>whole</code> involve the checking policy in order to get the bounds
of the resulting intervals.</p>
<h2><a name="opers" id="opers"></a>Operations and Functions</h2>
<p>Some of the following functions expect <code>min</code> and
<code>max</code> to be defined for the base type. Those are the only
requirements for the <code>interval</code> class (but the policies can have
other requirements).</p>
<h4>Operators <code>+</code> <code>-</code> <code>*</code> <code>/</code>
<code>+=</code> <code>-=</code> <code>*=</code> <code>/=</code></h4>
<p>The basic operations are the unary minus and the binary <code>+</code>
<code>-</code> <code>*</code> <code>/</code>. The unary minus takes an
interval and returns an interval. The binary operations take two intervals,
or one interval and a number, and return an interval. If an argument is a
number instead of an interval, you can expect the result to be the same as
if the number was first converted to an interval. This property will be
true for all the following functions and operators.</p>
<p>There are also some assignment operators <code>+=</code> <code>-=</code>
<code>*=</code> <code>/=</code>. There is not much to say: <code>x op=
y</code> is equivalent to <code>x = x op y</code>. If an exception is
thrown during the computations, the l-value is not modified (but it may be
corrupt if an exception is thrown by the base type during an
assignment).</p>
<p>The operators <code>/</code> and <code>/=</code> will try to produce an
empty interval if the denominator is exactly zero. If the denominator
contains zero (but not only zero), the result will be the smallest interval
containing the set of division results; so one of its bound will be
infinite, but it may not be the whole interval.</p>
<h4><code>lower</code> <code>upper</code> <code>median</code>
<code>width</code> <code>norm</code></h4>
<p><code>lower</code>, <code>upper</code>, <code>median</code> respectively
compute the lower bound, the upper bound, and the median number of an
interval (<code>(lower+upper)/2</code> rounded to nearest).
<code>width</code> computes the width of an interval
(<code>upper-lower</code> rounded toward plus infinity). <code>norm</code>
computes an upper bound of the interval in absolute value; it is a
mathematical norm (hence the name) similar to the absolute value for real
numbers.</p>
<h4><code>min</code> <code>max</code> <code>abs</code> <code>square</code>
<code>pow</code> <code>nth_root</code> <code>division_part?</code>
<code>multiplicative_inverse</code></h4>
<p>The functions <code>min</code>, <code>max</code> and <code>abs</code>
are also defined. Please do not mistake them for the functions defined in
the standard library (aka <code>a<b?a:b</code>, <code>a>b?a:b</code>,
<code>a<0?-a:a</code>). These functions are compatible with the
elementary property of interval arithmetic. For example,
max([<i>a</i>,<i>b</i>], [<i>c</i>,<i>d</i>]) = {max(<i>x</i>,<i>y</i>)
such that <i>x</i> in [<i>a</i>,<i>b</i>] and <i>y</i> in
[<i>c</i>,<i>d</i>]}. They are not defined in the <code>std</code>
namespace but in the boost namespace in order to avoid conflict with the
other definitions.</p>
<p>The <code>square</code> function is quite particular. As you can expect
from its name, it computes the square of its argument. The reason this
function is provided is: <code>square(x)</code> is not <code>x*x</code> but
only a subset when <code>x</code> contains zero. For example, [-2,2]*[-2,2]
= [-4,4] but [-2,2]² = [0,4]; the result is a smaller interval.
Consequently, <code>square(x)</code> should be used instead of
<code>x*x</code> because of its better accuracy and a small performance
improvement.</p>
<p>As for <code>square</code>, <code>pow</code> provides an efficient and
more accurate way to compute the integer power of an interval. Please note:
when the power is 0 and the interval is not empty, the result is 1, even if
the input interval contains 0. <code>nth_root</code> computes the integer root
of an interval (<code>nth_root(pow(x,k),k)</code> encloses <code>x</code> or
<code>abs(x)</code> depending on whether <code>k</code> is odd or even).
The behavior of <code>nth_root</code> is not defined if the integer argument is
not positive. <code>multiplicative_inverse</code> computes
<code>1/x</code>.</p>
<p>The functions <code>division_part1</code> and
<code>division_part2</code> are useful when the user expects the division
to return disjoint intervals if necessary. For example, the narrowest
closed set containg [2,3] / [-2,1] is not ]-∞,∞[ but the union
of ]-∞,-1] and [2,∞[. When the result of the division is
representable by only one interval, <code>division_part1</code> returns
this interval and sets the boolean reference to <code>false</code>.
However, if the result needs two intervals, <code>division_part1</code>
returns the negative part and sets the boolean reference to
<code>true</code>; a call to <code>division_part2</code> is now needed to
get the positive part. This second function can take the boolean returned
by the first function as last argument. If this bool is not given, its
value is assumed to be true and the behavior of the function is then
undetermined if the division does not produce a second interval.</p>
<h4><code>intersect</code> <code>hull</code> <code>overlap</code>
<code>in</code> <code>zero_in</code> <code>subset</code>
<code>proper_subset</code> <code>empty</code> <code>singleton</code>
<code>equal</code></h4>
<p><code>intersect</code> computes the set intersection of two closed sets,
<code>hull</code> computes the smallest interval which contains the two
parameters; those parameters can be numbers or intervals. If one of the
arguments is an invalid number or an empty interval, the function will only
use the other argument to compute the resulting interval (if allowed by the
checking policy).</p>
<p>There is no union function since the union of two intervals is not an
interval if they do not overlap. If they overlap, the <code>hull</code>
function computes the union.</p>
<p>The function <code>overlap</code> tests if two intervals have some
common subset. <code>in</code> tests if a number is in an interval;
<code>zero_in</code> is a variant which tests if zero is in the interval.
<code>subset</code> tests if the first interval is a subset of the second;
and <code>proper_subset</code> tests if it is a proper subset.
<code>empty</code> and <code>singleton</code> test if an interval is empty
or is a singleton. Finally, <code>equal</code> tests if two intervals are
equal.</p>
<h4><code>sqrt</code> <code>log</code> <code>exp</code> <code>sin</code>
<code>cos</code> <code>tan</code> <code>asin</code> <code>acos</code>
<code>atan</code> <code>sinh</code> <code>cosh</code> <code>tanh</code>
<code>asinh</code> <code>acosh</code> <code>atanh</code>
<code>fmod</code></h4>
<p>The functions <code>sqrt</code>, <code>log</code>, <code>exp</code>,
<code>sin</code>, <code>cos</code>, <code>tan</code>, <code>asin</code>,
<code>acos</code>, <code>atan</code>, <code>sinh</code>, <code>cosh</code>,
<code>tanh</code>, <code>asinh</code>, <code>acosh</code>,
<code>atanh</code> are also defined. There is not much to say; these
functions extend the traditional functions to the intervals and respect the
basic property of interval arithmetic. They use the <a href=
"checking.htm">checking</a> policy to produce empty intervals when the
input interval is strictly outside of the domain of the function.</p>
<p>The function <code>fmod(interval x, interval y)</code> expects the lower
bound of <code>y</code> to be strictly positive and returns an interval
<code>z</code> such as <code>0 <= z.lower() < y.upper()</code> and
such as <code>z</code> is a superset of <code>x-n*y</code> (with
<code>n</code> being an integer). So, if the two arguments are positive
singletons, this function <code>fmod(interval, interval)</code> will behave
like the traditional function <code>fmod(double, double)</code>.</p>
<p>Please note that <code>fmod</code> does not respect the inclusion
property of arithmetic interval. For example, the result of
<code>fmod</code>([13,17],[7,8]) should be [0,8] (since it must contain
[0,3] and [5,8]). But this answer is not really useful when the purpose is
to restrict an interval in order to compute a periodic function. It is the
reason why <code>fmod</code> will answer [5,10].</p>
<h4><code>add</code> <code>sub</code> <code>mul</code>
<code>div</code></h4>
<p>These four functions take two numbers and return the enclosing interval
for the operations. It avoids converting a number to an interval before an
operation, it can result in a better code with poor optimizers.</p>
<h3>Constants</h3>
<p>Some constants are hidden in the
<code>boost::numeric::interval_lib</code> namespace. They need to be
explicitely templated by the interval type. The functions are
<code>pi<I>()</code>, <code>pi_half<I>()</code> and
<code>pi_twice<I>()</code>, and they return an object of interval
type <code>I</code>. Their respective values are π, π/2 and
2π.</p>
<h3>Exception throwing</h3>
<p>The interval class and all the functions defined around this class never
throw any exceptions by themselves. However, it does not mean that an
operation will never throw an exception. For example, let's consider the
copy constructor. As explained before, it is the default copy constructor
generated by the compiler. So it will not throw an exception if the copy
constructor of the base type does not throw an exception.</p>
<p>The same situation applies to all the functions: exceptions will only be
thrown if the base type or one of the two policies throws an exception.</p>
<h2 id="interval_lib">Interval Support Library</h2>
<p>The interval support library consists of a collection of classes that
can be used and combined to fabricate almost various commonly-needed
interval policies. In contrast to the basic classes and functions which are
used in conjunction with <code>interval<T></code> (and the default
policies as the implicit second template parameter in this type), which
belong simply to the namespace <code>boost</code>, these components belong
to the namespace <code>boost::numeric::interval_lib</code>.</p>
<p>We merely give the synopsis here and defer each section to a separate
web page since it is only intended for the advanced user. This allows to
expand on each topic with examples, without unduly stretching the limits of
this document.</p>
<h4>Synopsis</h4>
<pre>
namespace boost {
namespace numeric {
namespace interval_lib {
<span style=
"color: #FF0000">/* built-in rounding policy and its specializations */</span>
template <class T> struct rounded_math;
template <> struct rounded_math<float>;
template <> struct rounded_math<double>;
template <> struct rounded_math<long double>;
<span style=
"color: #FF0000">/* built-in rounding construction blocks */</span>
template <class T> struct rounding_control;
template <class T, class Rounding = rounding_control<T> > struct rounded_arith_exact;
template <class T, class Rounding = rounding_control<T> > struct rounded_arith_std;
template <class T, class Rounding = rounding_control<T> > struct rounded_arith_opp;
template <class T, class Rounding> struct rounded_transc_dummy;
template <class T, class Rounding = rounded_arith_exact<T> > struct rounded_transc_exact;
template <class T, class Rounding = rounded_arith_std <T> > struct rounded_transc_std;
template <class T, class Rounding = rounded_arith_opp <T> > struct rounded_transc_opp;
template <class Rounding> struct save_state;
template <class Rounding> struct save_state_nothing;
<span style="color: #FF0000">/* built-in checking policies */</span>
template <class T> struct checking_base;
template <class T, class Checking = checking_base<T>, class Exception = exception_create_empty> struct checking_no_empty;
template <class T, class Checking = checking_base<T> > struct checking_no_nan;
template <class T, class Checking = checking_base<T>, class Exception = exception_invalid_number> struct checking_catch_nan;
template <class T> struct checking_strict;
<span style=
"color: #FF0000">/* some metaprogramming to manipulate interval policies */</span>
template <class Rounding, class Checking> struct policies;
template <class OldInterval, class NewRounding> struct change_rounding;
template <class OldInterval, class NewChecking> struct change_checking;
template <class OldInterval> struct unprotect;
<span style=
"color: #FF0000">/* constants, need to be explicitly templated */</span>
template<class I> I pi();
template<class I> I pi_half();
template<class I> I pi_twice();
<span style="color: #FF0000">/* interval explicit comparison functions:
* the mode can be cer=certainly or pos=possibly,
* the function lt=less_than, gt=greater_than, le=less_than_or_equal_to, ge=greater_than_or_equal_to
* eq=equal_to, ne= not_equal_to */</span>
template <class T, class Policies> bool cerlt(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cerlt(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool cerlt(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cerle(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cerle(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool cerle(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cergt(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cergt(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool cergt(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cerge(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cerge(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool cerge(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cereq(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cereq(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool cereq(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cerne(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool cerne(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool cerne(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool poslt(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool poslt(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool poslt(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool posle(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool posle(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool posle(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool posgt(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool posgt(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool posgt(const T& x, const interval<T, Policies> & y);
template <class T, class Policies> bool posge(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool posge(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool posge(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool poseq(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool poseq(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool poseq(const T& x, const interval<T, Policies>& y);
template <class T, class Policies> bool posne(const interval<T, Policies>& x, const interval<T, Policies>& y);
template <class T, class Policies> bool posne(const interval<T, Policies>& x, const T& y);
template <class T, class Policies> bool posne(const T& x, const interval<T, Policies>& y);
<span style="color: #FF0000">/* comparison namespaces */</span>
namespace compare {
namespace certain;
namespace possible;
namespace lexicographic;
namespace set;
namespace tribool;
} // namespace compare
} // namespace interval_lib
} // namespace numeric
} // namespace boost
</pre>
<p>Each component of the interval support library is detailed in its own
page.</p>
<ul>
<li><a href="comparisons.htm">Comparisons</a></li>
<li><a href="rounding.htm">Rounding</a></li>
<li><a href="checking.htm">Checking</a></li>
</ul>
<h2 id="dangers">Common Pitfalls and Dangers</h2>
<h4>Comparisons</h4>
<p>One of the biggest problems is problably the correct use of the
comparison functions and operators. First, functions and operators do not
try to know if two intervals are the same mathematical object. So, if the
comparison used is "certain", then <code>x != x</code> is always true
unless <code>x</code> is a singleton interval; and the same problem arises
with <code>cereq</code> and <code>cerne</code>.</p>
<p>Another misleading interpretation of the comparison is: you cannot
always expect [a,b] < [c,d] to be !([a,b] >= [c,d]) since the
comparison is not necessarily total. Equality and less comparison should be
seen as two distincts relational operators. However the default comparison
operators do respect this property since they throw an exception whenever
[a,b] and [c,d] overlap.</p>
<h4>Interval values and references</h4>
<p>This problem is a corollary of the previous problem with <code>x !=
x</code>. All the functions of the library only consider the value of an
interval and not the reference of an interval. In particular, you should
not expect (unless <code>x</code> is a singleton) the following values to
be equal: <code>x/x</code> and 1, <code>x*x</code> and
<code>square(x)</code>, <code>x-x</code> and 0, etc. So the main cause of
wide intervals is that interval arithmetic does not identify different
occurences of the same variable. So, whenever possible, the user has to
rewrite the formulas to eliminate multiple occurences of the same variable.
For example, <code>square(x)-2*x</code> is far less precise than
<code>square(x-1)-1</code>.</p>
<h4>Unprotected rounding</h4>
<p>As explained in <a href="rounding.htm#perf">this section</a>, a good way
to speed up computations when the base type is a basic floating-point type
is to unprotect the intervals at the hot spots of the algorithm. This
method is safe and really an improvement for interval computations. But
please remember that any basic floating-point operation executed inside the
unprotection blocks will probably have an undefined behavior (but only for
the current thread). And do not forget to create a rounding object as
explained in the <a href="rounding.htm#perfexp">example</a>.</p>
<h2 id="rationale">Rationale</h2>
<p>The purpose of this library is to provide an efficient and generalized
way to deal with interval arithmetic through the use of a templatized class
<code>boost::numeric::interval</code>. The big contention for which we provide a
rationale is the format of this class template.</p>
<p>It would have been easier to provide a class interval whose base type is
double. Or to follow <code>std::complex</code> and allow only
specializations for <code>float</code>, <code>double</code>, and <code>long
double</code>. We decided not to do this to allow intervals on custom
types, e.g. fixed-precision bigfloat library types (MPFR, etc), rational
numbers, and so on.</p>
<p><strong>Policy design.</strong> Although it was tempting to make it a
class template with only one template argument, the diversity of uses for
an interval arithmetic practically forced us to use policies. The behavior
of this class can be fixed by two policies. These policies are packaged
into a single policy class, rather than making <code>interval</code> with
three template parameters. This is both for ease of use (the policy class
can be picked by default) and for readability.</p>
<p>The first policy provides all the mathematical functions on the base
type needed to define the functions on the interval type. The second one
sets the way exceptional cases encountered during computations are
handled.</p>
<p>We could foresee situations where any combination of these policies
would be appropriate. Moreover, we wanted to enable the user of the library
to reuse the <code>interval</code> class template while at the same time
choosing his own behavior. See this <a href="guide.htm">page</a> for some
examples.</p>
<p><strong>Rounding policy.</strong> The library provides specialized
implementations of the rounding policy for the primitive types float and
double. In order for these implementations to be correct and fast, the
library needs to work a lot with rounding modes. Some processors are
directly dealt with and some mecanisms are provided in order to speed up
the computations. It seems to be heavy and hazardous optimizations for a
gain of only a few computer cycles; but in reality, the speed-up factor can
easily go past 2 or 3 depending on the computer. Moreover, these
optimizations do not impact the interface in any major way (with the design
we have chosen, everything can be added by specialization or by passing
different template parameters).</p>
<p><strong>Pred/succ.</strong> In a previous version, two functions
<code>pred</code> and <code>succ</code>, with various corollaries like
<code>widen</code>, were supplied. The intent was to enlarge the interval
by one ulp (as little as possible), e.g. to ensure the inclusion property.
Since making interval a template of T, we could not define <i>ulp</i> for a
random parameter. In turn, rounding policies let us eliminate entirely the
use of ulp while making the intervals tighter (if a result is a
representable singleton, there is no use to widen the interval). We decided
to drop those functions.</p>
<p><strong>Specialization of <code>std::less</code>.</strong> Since the
operator <code><</code> depends on the comparison namespace locally
chosen by the user, it is not possible to correctly specialize
<code>std::less</code>. So you have to explicitely provide such a class to
all the algorithms and templates that could require it (for example,
<code>std::map</code>).</p>
<p><strong>Input/output.</strong> The interval library does not include I/O
operators. Printing an interval value allows a lot of customization: some
people may want to output the bounds, others may want to display the median
and the width of intervals, and so on. The example file io.cpp shows some
possibilities and may serve as a foundation in order for the user to define
her own operators.</p>
<p><strong>Mixed operations with integers.</strong> When using and reusing
template codes, it is common there are operations like <code>2*x</code>.
However, the library does not provide them by default because the
conversion from <code>int</code> to the base number type is not always
correct (think about the conversion from a 32bit integer to a single
precision floating-point number). So the functions have been put in a
separate header and the user needs to include them explicitely if she wants
to benefit from these mixed operators. Another point, there is no mixed
comparison operators due to the technical way they are defined.</p>
<p><strong>Interval-aware functions.</strong> All the functions defined by
the library are obviously aware they manipulate intervals and they do it
accordingly to general interval arithmetic principles. Consequently they
may have a different behavior than the one commonly encountered with
functions not interval-aware. For example, <code>max</code> is defined by
canonical set extension and the result is not always one of the two
arguments (if the intervals do not overlap, then the result is one of the
two intervals).</p>
<p>This behavior is different from <code>std::max</code> which returns a
reference on one of its arguments. So if the user expects a reference to be
returned, she should use <code>std::max</code> since it is exactly what
this function does. Please note that <code>std::max</code> will throw an
exception when the intervals overlap. This behavior does not predate the
one described by the C++ standard since the arguments are not "equivalent"
and it allows to have an equivalence between <code>a <= b</code> and
<code>&b == &std::max(a,b)</code>(some particular cases may be
implementation-defined). However it is different from the one described by
SGI since it does not return the first argument even if "neither is greater
than the other".</p>
<h2 id="acks">History and Acknowledgments</h2>
<p>This library was mostly inspired by previous work from Jens Maurer. Some
discussions about his work are reproduced <a href=
"http://www.mscs.mu.edu/%7Egeorgec/IFAQ/maurer1.html">here</a>. Jeremy Siek
and Maarten Keijzer provided some rounding control for MSVC and Sparc
platforms.</p>
<p>Guillaume Melquiond, Hervé Brönnimann and Sylvain Pion
started from the library left by Jens and added the policy design.
Guillaume and Sylvain worked hard on the code, especially the porting and
mostly tuning of the rounding modes to the different architectures.
Guillaume did most of the coding, while Sylvain and Hervé have
provided some useful comments in order for this library to be written.
Hervé reorganized and wrote chapters of the documentation based on
Guillaume's great starting point.</p>
<p>This material is partly based upon work supported by the National
Science Foundation under NSF CAREER Grant CCR-0133599. Any opinions,
findings and conclusions or recommendations expressed in this material are
those of the author(s) and do not necessarily reflect the views of the
National Science Foundation (NSF).</p>
<hr>
<p><a href="http://validator.w3.org/check?uri=referer"><img border="0" src=
"../../../../doc/images/valid-html401.png" alt="Valid HTML 4.01 Transitional"
height="31" width="88"></a></p>
<p>Revised
<!--webbot bot="Timestamp" s-type="EDITED" s-format="%Y-%m-%d" startspan -->2006-12-25<!--webbot bot="Timestamp" endspan i-checksum="12174" --></p>
<p><i>Copyright © 2002 Guillaume Melquiond, Sylvain Pion, Hervé
Brönnimann, Polytechnic University<br>
Copyright © 2003-2006 Guillaume Melquiond, ENS Lyon</i></p>
<p><i>Distributed under the Boost Software License, Version 1.0. (See
accompanying file <a href="../../../../LICENSE_1_0.txt">LICENSE_1_0.txt</a>
or copy at <a href=
"http://www.boost.org/LICENSE_1_0.txt">http://www.boost.org/LICENSE_1_0.txt</a>)</i></p>
</body>
</html>
