devn00b
/
EQ2EMu


			
				
					
						
						
							123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211
							<!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>Tests and Examples</title>
</head>

<body lang="en">
  <h1>Tests and Examples</h1>

  <h2>A first example</h2>

  <p>This example shows how to design a function which takes a polynomial and
  a value and returns the sign of this polynomial at this point. This
  function is a filter: if the answer is not guaranteed, the functions says
  so. The reason of using a filter rather than a simple evaluation function
  is: computations with floating-point numbers will incur approximations and
  it can be enough to change the sign of the polynomial. So, in order to
  validate the result, the function will use interval arithmetic.</p>

  <p>The first step is the inclusion of the appropriate headers. Because the
  function will handle floating-point bounds, the easiest solution is:</p>
  <pre>
#include &lt;boost/numeric/interval.hpp&gt;
</pre>

  <p>Now, let's begin the function. The polynomial is given by the array of
  its coefficients and its size (strictly greater to its degree). In order to
  simplify the code, two namespaces of the library are included.</p>
  <pre>
int sign_polynomial(double x, double P[], int sz) {
  using namespace boost::numeric;
  using namespace interval_lib;
</pre>

  <p>Then we can define the interval type. Since no special behavior is
  required, the default policies are enough:</p>
  <pre>
  typedef interval&lt;double&gt; I;
</pre>

  <p>For the evaluation, let's just use the Horner scheme with interval
  arithmetic. The library overloads all the arithmetic operators and provides
  mixed operations, so the only difference between the code with and without
  interval arithmetic lies in the type of the iterated value
  <code>y</code>:</p>
  <pre>
  I y = P[sz - 1];
  for(int i = sz - 2; i &gt;= 0; i--)
    y = y * x + P[i];
</pre>

  <p>The last step is the computation of the sign of <code>y</code>. It is
  done by choosing an appropriate comparison scheme and then doing the
  comparison with the usual operators:</p>
  <pre>
  using namespace compare::certain;
  if (y &gt; 0.) return 1;
  if (y &lt; 0.) return -1;
  return 0;
}
</pre>

  <p>The answer <code>0</code> does not mean the polynomial is zero at this
  point. It only means the answer is not known since <code>y</code> contains
  zero and thus does not have a precise sign.</p>

  <p>Now we have the expected function. However, due to the poor
  implementations of floating-point rounding in most of the processors, it
  can be useful to say to optimize the code; or rather, to let the library
  optimize it. The main condition for this optimization is that the interval
  code should not be mixed with floating-point code. In this example, it is
  the case, since all the operations done in the functions involve the
  library. So the code can be rewritten:</p>
  <pre>
int sign_polynomial(double x, double P[], int sz) {
  using namespace boost::numeric;
  using namespace interval_lib;
  typedef interval&lt;double&gt; I_aux;

  I_aux::traits_type::rounding rnd;
  typedef unprotect&lt;I_aux&gt;::type I;

  I y = P[sz - 1];
  for(int i = sz - 2; i &gt;= 0; i--) 
    y = y * x + P[i];

  using namespace compare::certain;
  if (y &gt; 0.) return 1;
  if (y &lt; 0.) return -1;
  return 0;
}
</pre>

  <p>The difference between this code and the previous is the use of another
  interval type. This new type <code>I</code> indicates to the library that
  all the computations can be done without caring for the rounding mode. And
  because of that, it is up to the function to care about it: a rounding
  object need to be alive whenever the optimized type is used.</p>

  <h2>Other tests and examples</h2>

  <p>In <code>libs/numeric/interval/test/</code> and
  <code>libs/numeric/interval/examples/</code> are some test and example
  programs.. The examples illustrate a few uses of intervals. For a general
  description and considerations on using this library, and some potential
  domains of application, please read this <a href=
  "guide.htm">mini-guide</a>.</p>

  <h3>Tests</h3>

  <p>The test programs are as follows. Please note that they require the use
  of the Boost.test library and can be automatically tested by using
  <code>bjam</code> (except for interval_test.cpp).</p>

  <p><b>add.cpp</b> tests if the additive and subtractive operators and the
  respective _std and _opp rounding functions are correctly implemented. It
  is done by using symbolic expressions as a base type.</p>

  <p><b>cmp.cpp</b>, <b>cmp_lex.cpp</b>, <b>cmp_set.cpp</b>, and
  <b>cmp_tribool.cpp</b> test if the operators <code>&lt;</code>
  <code>&gt;</code> <code>&lt;=</code> <code>&gt;=</code> <code>==</code>
  <code>!=</code> behave correctly for the default, lexicographic, set, and
  tristate comparisons. <b>cmp_exp.cpp</b> tests the explicit comparison
  functions <code>cer..</code> and <code>pos..</code> behave correctly.
  <b>cmp_exn.cpp</b> tests if the various policies correctly detect
  exceptional cases. All these tests use some simple intervals ([1,2] and
  [3,4], [1,3] and [2,4], [1,2] and [2,3], etc).</p>

  <p><b>det.cpp</b> tests if the <code>_std</code> and <code>_opp</code>
  versions in protected and unprotected mode produce the same result when
  Gauss scheme is used on an unstable matrix (in order to exercise rounding).
  The tests are done for <code>interval&lt;float&gt;</code> and
  <code>interval&lt;double&gt;</code>.</p>

  <p><b>fmod.cpp</b> defines a minimalistic version of
  <code>interval&lt;int&gt;</code> and uses it in order to test
  <code>fmod</code> on some specific interval values.</p>

  <p><b>mul.cpp</b> exercises the multiplication, the finite division, the
  square and the square root with some integer intervals leading to exact
  results.</p>

  <p><b>pi.cpp</b> tests if the interval value of &pi; (for <code>int</code>,
  <code>float</code> and <code>double</code> base types) contains the number
  &pi; (defined with 21 decimal digits) and if it is a subset of
  [&pi;&plusmn;1ulp] (in order to ensure some precision).</p>

  <p><b>pow.cpp</b> tests if the <code>pow</code> function behaves correctly
  on some simple test cases.</p>

  <p><b>test_float.cpp</b> exercises the arithmetic operations of the library
  for floating point base types.</p>

  <p><b>interval_test.cpp</b> tests if the interval library respects the
  inclusion property of interval arithmetic by computing some functions and
  operations for both <code>double</code> and
  <code>interval&lt;double&gt;</code>.</p>

  <h2>Examples</h2>

  <p><b>filter.cpp</b> contains filters for computational geometry able to
  find the sign of a determinant. This example is inspired by the article
  <em>Interval arithmetic yields efficient dynamic filters for computational
  geometry</em> by Br&ouml;nnimann, Burnikel and Pion, 2001.</p>

  <p><b>findroot_demo.cpp</b> finds zeros of some functions by using
  dichotomy and even produces gnuplot data for one of them. The processor has
  to correctly handle elementary functions for this example to properly
  work.</p>

  <p><b>horner.cpp</b> is a really basic example of unprotecting the interval
  operations for a whole function (which computes the value of a polynomial
  by using Horner scheme).</p>

  <p><b>io.cpp</b> shows some stream input and output operators for intervals
  .The wide variety of possibilities explains why the library do not
  implement i/o operators and they are left to the user.</p>

  <p><b>newton-raphson.cpp</b> is an implementation of a specialized version
  of Newton-Raphson algorithm for finding the zeros of a function knowing its
  derivative. It exercises unprotecting, full division, some set operations
  and empty intervals.</p>

  <p><b>transc.cpp</b> implements the transcendental part of the rounding
  policy for <code>double</code> by using an external library (the MPFR
  subset of GMP in this case).</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 -->2006-12-24<!--webbot bot="Timestamp" endspan i-checksum="12172" --></p>

  <p><i>Copyright &copy; 2002 Guillaume Melquiond, Sylvain Pion, Herv&eacute;
  Br&ouml;nnimann, Polytechnic University<br>
  Copyright &copy; 2003 Guillaume Melquiond</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>