Version: SMASH-1.5
random.cc
Go to the documentation of this file.
1 /*
2  *
3  * Copyright (c) 2014-2018
4  * SMASH Team
5  *
6  * GNU General Public License (GPLv3 or later)
7  *
8  */
9 
10 #include "smash/random.h"
11 #include <random>
12 #include "smash/logging.h"
13 
14 namespace smash {
15 /*thread_local (see #3075)*/ random::Engine random::engine;
16 
17 random::BesselSampler::BesselSampler(const double poisson_mean1,
18  const double poisson_mean2,
19  const int fixed_difference)
20  : a_(2.0 * std::sqrt(poisson_mean1 * poisson_mean2)),
21  N_(std::abs(fixed_difference)),
22  N_is_positive_(fixed_difference >= 0) {
23  const auto &log = logger<LogArea::GrandcanThermalizer>();
24  assert(poisson_mean1 >= 0.0);
25  assert(poisson_mean2 >= 0.0);
26  log.debug("Bessel sampler", ": Poisson mean N1 = ", poisson_mean1,
27  ", Poisson mean N2 = ", poisson_mean2, ", N1 - N2 fixed to ",
28  fixed_difference);
29  m_ = 0.5 * (std::sqrt(a_ * a_ + N_ * N_) - N_);
30  if (m_ >= m_switch_method_) {
31  mu_ = 0.5 * a_ * r_(N_, a_);
32  const double mean_sqr = mu_ * (1.0 + 0.5 * a_ * r_(N_ + 1, a_));
33  sigma_ = std::sqrt(mean_sqr - mu_ * mu_);
34  log.debug("m = ", m_, " -> using gaussian sampling with mean = ", mu_,
35  ", sigma = ", sigma_);
36  } else {
37  log.debug("m = ", m_, " -> using direct sampling method");
38  std::vector<double> probabilities;
39  double wi = 1.0, sum = 0.0;
40  int i = 0;
41  do {
42  sum += wi;
43  probabilities.push_back(wi);
44  wi *= 0.25 * a_ * a_ / (i + 1) / (N_ + i + 1);
45  i++;
46  } while (wi > negligible_probability_);
47  i = 0;
48  for (double p : probabilities) {
49  p /= sum;
50  log.debug("Probability (", i, ") = ", p);
51  i++;
52  }
53  dist_.reset_weights(probabilities);
54  }
55 }
56 
57 std::pair<int, int> random::BesselSampler::sample() {
58  const int N_smaller = (m_ >= m_switch_method_)
59  ? std::round(random::normal(mu_, sigma_))
60  : dist_();
61  return N_is_positive_ ? std::make_pair(N_smaller + N_, N_smaller)
62  : std::make_pair(N_smaller, N_smaller + N_);
63 }
64 
65 double random::BesselSampler::r_(int n, double a) {
66  const double a_inv = 1.0 / a;
67  double res = 0.0;
68  // |x - continued fraction of order n| < 2^(-n+1), see the book
69  // "Continued fractions" by Khinchin. For 10^-16 = ~2^-50 precision
70  // 50 iterations should be sufficient. However, I found that for some
71  // numerical reason at least 100 terms are needed.
72  int i = 200;
73  for (; i > 0; i--) {
74  res = 1.0 / (a_inv * 2 * (n + i) + res);
75  }
76  // Check the known property of r(n,a) function, see iref{Yuan2000}.
77  assert(a / (std::sqrt(a * a + (n + 1) * (n + 1)) + n + 1) <= res);
78  assert(res <= a / (std::sqrt(a * a + n * n) + n));
79  return res;
80 }
81 
82 } // namespace smash
random::discrete_dist< double > dist_
Vector to store tabulated values of probabilities for small m case (m <6).
Definition: random.h:407
static constexpr double negligible_probability_
Probabilities smaller than negligibly_probability are neglected.
Definition: random.h:429
STL namespace.
double mu_
Mean of the Bessel distribution.
Definition: random.h:432
std::pair< int, int > sample()
Sample two numbers from given Poissonians with a fixed difference.
Definition: random.cc:57
BesselSampler(const double poisson_mean1, const double poisson_mean2, const int fixed_difference)
Construct a BesselSampler.
Definition: random.cc:17
static constexpr double m_switch_method_
Switching mode to normal approximation.
Definition: random.h:426
const int N_
First parameter of Bessel distribution (= in Yuan2000).
Definition: random.h:416
double sigma_
Standard deviation of the Bessel distribution.
Definition: random.h:435
Engine engine
The engine that is used commonly by all distributions.
Definition: random.cc:15
constexpr int p
Proton.
void reset_weights(const std::vector< T > &plist)
Reset the discrete distribution from a new probability list.
Definition: random.h:277
std::mt19937_64 Engine
The random number engine used is the Mersenne Twister.
Definition: random.h:27
double normal(const T &mean, const T &sigma)
Returns a random number drawn from a normal distribution.
Definition: random.h:247
static double r_(int n, double a)
Compute the ratio of two Bessel functions r(n,a) = bessel_I(n+1,a)/bessel_I(n,a) using the continued ...
Definition: random.cc:65
double m_
Mode of the Bessel function, see Yuan2000 for details.
Definition: random.h:410
constexpr int n
Neutron.
const double a_
Second parameter of Bessel distribution, see Yuan2000 for details.
Definition: random.h:413
Definition: action.h:24