Upsilon/poincare/src/integral.cpp

#include <poincare/integral.h>
#include <poincare/symbol.h>
#include <poincare/context.h>
#include <poincare/undefined.h>
#include <cmath>
#include <assert.h>
#include <float.h>
#include <stdlib.h>
#include <poincare/integral_layout_node.h>

namespace Poincare {

Expression::Type Integral::type() const {
  return Type::Integral;
}

Expression * Integral::clone() const {
  Integral * a = new Integral(m_operands, true);
  return a;
}

int Integral::polynomialDegree(char symbolName) const {
  if (symbolName == 'x') {
    int da = operand(1)->polynomialDegree(symbolName);
    int db = operand(2)->polynomialDegree(symbolName);
    if (da != 0 || db != 0) {
      return -1;
    }
    return 0;
  }
  return Expression::polynomialDegree(symbolName);
}

Expression Integral::shallowReduce(Context& context, Preferences::AngleUnit angleUnit) {
  Expression * e = Expression::shallowReduce(context, angleUnit);
  if (e != this) {
    return e;
  }
#if MATRIX_EXACT_REDUCING
  if (operand(0)->type() == Type::Matrix || operand(1)->type() == Type::Matrix || operand(2)->type() == Type::Matrix) {
    return replaceWith(new Undefined(), true);
  }
#endif
  return this;
}

template<typename T>
Complex<T> * Integral::templatedApproximate(Context & context, Preferences::AngleUnit angleUnit) const {
  Evaluation<T> * aInput = operand(1)->privateApproximate(T(), context, angleUnit);
  T a = aInput->toScalar();
  delete aInput;
  Evaluation<T> * bInput = operand(2)->privateApproximate(T(), context, angleUnit);
  T b = bInput->toScalar();
  delete bInput;
  if (std::isnan(a) || std::isnan(b)) {
    return new Complex<T>(Complex<T>::Undefined());
  }
#ifdef LAGRANGE_METHOD
  T result = lagrangeGaussQuadrature<T>(a, b, context, angleUnit);
#else
  T result = adaptiveQuadrature<T>(a, b, 0.1, k_maxNumberOfIterations, context, angleUnit);
#endif
  return new Complex<T>(result);
}

LayoutRef Integral::createLayout(Preferences::PrintFloatMode floatDisplayMode, int numberOfSignificantDigits) const {
  return IntegralLayoutRef(
      operand(0)->createLayout(floatDisplayMode, numberOfSignificantDigits),
      operand(1)->createLayout(floatDisplayMode, numberOfSignificantDigits),
      operand(2)->createLayout(floatDisplayMode, numberOfSignificantDigits));
}

template<typename T>
T Integral::functionValueAtAbscissa(T x, Context & context, Preferences::AngleUnit angleUnit) const {
  return operand(0)->approximateWithValueForSymbol('x', x, context, angleUnit);
}

#ifdef LAGRANGE_METHOD

template<typename T>
T Integral::lagrangeGaussQuadrature(T a, T b, Context & context, Preferences::AngleUnit angleUnit) const {
  /* We here use Gauss-Legendre quadrature with n = 5
   * Gauss-Legendre abscissae and weights can be found in
   * C/C++ library source code. */
  const static T x[10]={0.0765265211334973337546404, 0.2277858511416450780804962, 0.3737060887154195606725482, 0.5108670019508270980043641,
   0.6360536807265150254528367, 0.7463319064601507926143051, 0.8391169718222188233945291, 0.9122344282513259058677524,
   0.9639719272779137912676661, 0.9931285991850949247861224};
  const static T w[10]={0.1527533871307258506980843, 0.1491729864726037467878287, 0.1420961093183820513292983, 0.1316886384491766268984945, 0.1181945319615184173123774,
    0.1019301198172404350367501, 0.0832767415767047487247581, 0.0626720483341090635695065, 0.0406014298003869413310400, 0.0176140071391521183118620};
  T xm = 0.5*(a+b);
  T xr = 0.5*(b-a);
  T result = 0;
  for (int j = 0; j < 10; j++) {
    T dx = xr * x[j];
    T evaluationAfterX = functionValueAtAbscissa(xm+dx, context, angleUnit);
    if (std::isnan(evaluationAfterX)) {
      return NAN;
    }
    T evaluationBeforeX = functionValueAtAbscissa(xm-dx, context, angleUnit);
    if (std::isnan(evaluationBeforeX)) {
      return NAN;
    }
    result += w[j]*(evaluationAfterX + evaluationBeforeX);
  }
  result *= xr;
  return result;
}

#else

template<typename T>
Integral::DetailedResult<T> Integral::kronrodGaussQuadrature(T a, T b, Context & context, Preferences::AngleUnit angleUnit) const {
  static T epsilon = sizeof(T) == sizeof(double) ? DBL_EPSILON : FLT_EPSILON;
  static T max = sizeof(T) == sizeof(double) ? DBL_MAX : FLT_MAX;
  /* We here use Kronrod-Legendre quadrature with n = 21
   * The abscissa and weights are taken from QUADPACK library. */
  const static T wg[5]= {0.066671344308688137593568809893332, 0.149451349150580593145776339657697,
    0.219086362515982043995534934228163, 0.269266719309996355091226921569469, 0.295524224714752870173892994651338};
  const static T xgk[11]= {0.995657163025808080735527280689003, 0.973906528517171720077964012084452,
    0.930157491355708226001207180059508, 0.865063366688984510732096688423493, 0.780817726586416897063717578345042,
    0.679409568299024406234327365114874, 0.562757134668604683339000099272694, 0.433395394129247190799265943165784,
    0.294392862701460198131126603103866, 0.148874338981631210884826001129720, 0.000000000000000000000000000000000};
  const static T wgk[11]= {0.011694638867371874278064396062192, 0.032558162307964727478818972459390,
    0.054755896574351996031381300244580, 0.075039674810919952767043140916190, 0.093125454583697605535065465083366,
    0.109387158802297641899210590325805, 0.123491976262065851077958109831074, 0.134709217311473325928054001771707,
    0.142775938577060080797094273138717, 0.147739104901338491374841515972068, 0.149445554002916905664936468389821};
  T fv1[10];
  T fv2[10];

  T centr = 0.5*(a+b);
  T hlgth = 0.5*(b-a);
  T dhlgth = std::fabs(hlgth);

  DetailedResult<T> errorResult;
  errorResult.integral = NAN;
  errorResult.absoluteError = 0;

  T resg = 0;
  T fc = functionValueAtAbscissa(centr, context, angleUnit);
  if (std::isnan(fc)) {
    return errorResult;
  }
  T resk = wgk[10]*fc;
  T resabs = std::fabs(resk);
  for (int j = 0; j < 10; j++) {
    T absc = hlgth*xgk[j];
    T fval1 = functionValueAtAbscissa(centr-absc, context, angleUnit);
    if (std::isnan(fval1)) {
      return errorResult;
    }
    T fval2 = functionValueAtAbscissa(centr+absc, context, angleUnit);
    if (std::isnan(fval2)) {
      return errorResult;
    }
    fv1[j] = fval1;
    fv2[j] = fval2;
    T fsum = fval1+fval2;
    if (j%2 == 1) {
      resg += wg[j/2]*fsum;
    }
    resk += wgk[j]*fsum;
    resabs += wgk[j]*(std::fabs(fval1)+std::fabs(fval2));
  }

  T reskh = resk*0.5;
  T resasc = wgk[10]*std::fabs(fc-reskh);
  for (int j = 0; j < 10; j++) {
    resasc += wgk[j]*(std::fabs(fv1[j]-reskh)+std::fabs(fv2[j]-reskh));
  }
  T integral = resk*hlgth;
  resabs = resabs*dhlgth;
  resasc = resasc*dhlgth;
  T abserr = std::fabs((resk-resg)*hlgth);
  if (resasc != 0 && abserr != 0) {
    abserr = 1 > std::pow((T)(200*abserr/resasc), (T)1.5)? resasc*std::pow((T)(200*abserr/resasc), (T)1.5) : resasc;
  }
  if (resabs > max/(50.0*epsilon)) {
    abserr = abserr > epsilon*50*resabs ? abserr : epsilon*50*resabs;
  }
  DetailedResult<T> result;
  result.integral = integral;
  result.absoluteError = abserr;
  return result;
}

template<typename T>
T Integral::adaptiveQuadrature(T a, T b, T eps, int numberOfIterations, Context & context, Preferences::AngleUnit angleUnit) const {
  if (shouldStopProcessing()) {
    return NAN;
  }
  DetailedResult<T> quadKG = kronrodGaussQuadrature(a, b, context, angleUnit);
  T result = quadKG.integral;
  if (quadKG.absoluteError <= eps) {
    return result;
  } else if (--numberOfIterations > 0) {
    T m = (a+b)/2;
    return adaptiveQuadrature<T>(a, m, eps/2, numberOfIterations, context, angleUnit) + adaptiveQuadrature<T>(m, b, eps/2, numberOfIterations, context, angleUnit);
  } else {
    return NAN;
  }
}
#endif

}