xref: /aosp_15_r20/external/angle/third_party/abseil-cpp/absl/random/poisson_distribution.h (revision 8975f5c5ed3d1c378011245431ada316dfb6f244)
1 // Copyright 2017 The Abseil Authors.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //      https://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 
15 #ifndef ABSL_RANDOM_POISSON_DISTRIBUTION_H_
16 #define ABSL_RANDOM_POISSON_DISTRIBUTION_H_
17 
18 #include <cassert>
19 #include <cmath>
20 #include <cstdint>
21 #include <istream>
22 #include <limits>
23 #include <ostream>
24 
25 #include "absl/base/config.h"
26 #include "absl/random/internal/fast_uniform_bits.h"
27 #include "absl/random/internal/fastmath.h"
28 #include "absl/random/internal/generate_real.h"
29 #include "absl/random/internal/iostream_state_saver.h"
30 #include "absl/random/internal/traits.h"
31 
32 namespace absl {
33 ABSL_NAMESPACE_BEGIN
34 
35 // absl::poisson_distribution:
36 // Generates discrete variates conforming to a Poisson distribution.
37 //   p(n) = (mean^n / n!) exp(-mean)
38 //
39 // Depending on the parameter, the distribution selects one of the following
40 // algorithms:
41 // * The standard algorithm, attributed to Knuth, extended using a split method
42 // for larger values
43 // * The "Ratio of Uniforms as a convenient method for sampling from classical
44 // discrete distributions", Stadlober, 1989.
45 // http://www.sciencedirect.com/science/article/pii/0377042790903495
46 //
47 // NOTE: param_type.mean() is a double, which permits values larger than
48 // poisson_distribution<IntType>::max(), however this should be avoided and
49 // the distribution results are limited to the max() value.
50 //
51 // The goals of this implementation are to provide good performance while still
52 // being thread-safe: This limits the implementation to not using lgamma
53 // provided by <math.h>.
54 //
55 template <typename IntType = int>
56 class poisson_distribution {
57  public:
58   using result_type = IntType;
59 
60   class param_type {
61    public:
62     using distribution_type = poisson_distribution;
63     explicit param_type(double mean = 1.0);
64 
mean()65     double mean() const { return mean_; }
66 
67     friend bool operator==(const param_type& a, const param_type& b) {
68       return a.mean_ == b.mean_;
69     }
70 
71     friend bool operator!=(const param_type& a, const param_type& b) {
72       return !(a == b);
73     }
74 
75    private:
76     friend class poisson_distribution;
77 
78     double mean_;
79     double emu_;  // e ^ -mean_
80     double lmu_;  // ln(mean_)
81     double s_;
82     double log_k_;
83     int split_;
84 
85     static_assert(random_internal::IsIntegral<IntType>::value,
86                   "Class-template absl::poisson_distribution<> must be "
87                   "parameterized using an integral type.");
88   };
89 
poisson_distribution()90   poisson_distribution() : poisson_distribution(1.0) {}
91 
poisson_distribution(double mean)92   explicit poisson_distribution(double mean) : param_(mean) {}
93 
poisson_distribution(const param_type & p)94   explicit poisson_distribution(const param_type& p) : param_(p) {}
95 
reset()96   void reset() {}
97 
98   // generating functions
99   template <typename URBG>
operator()100   result_type operator()(URBG& g) {  // NOLINT(runtime/references)
101     return (*this)(g, param_);
102   }
103 
104   template <typename URBG>
105   result_type operator()(URBG& g,  // NOLINT(runtime/references)
106                          const param_type& p);
107 
param()108   param_type param() const { return param_; }
param(const param_type & p)109   void param(const param_type& p) { param_ = p; }
110 
result_type(min)111   result_type(min)() const { return 0; }
result_type(max)112   result_type(max)() const { return (std::numeric_limits<result_type>::max)(); }
113 
mean()114   double mean() const { return param_.mean(); }
115 
116   friend bool operator==(const poisson_distribution& a,
117                          const poisson_distribution& b) {
118     return a.param_ == b.param_;
119   }
120   friend bool operator!=(const poisson_distribution& a,
121                          const poisson_distribution& b) {
122     return a.param_ != b.param_;
123   }
124 
125  private:
126   param_type param_;
127   random_internal::FastUniformBits<uint64_t> fast_u64_;
128 };
129 
130 // -----------------------------------------------------------------------------
131 // Implementation details follow
132 // -----------------------------------------------------------------------------
133 
134 template <typename IntType>
param_type(double mean)135 poisson_distribution<IntType>::param_type::param_type(double mean)
136     : mean_(mean), split_(0) {
137   assert(mean >= 0);
138   assert(mean <=
139          static_cast<double>((std::numeric_limits<result_type>::max)()));
140   // As a defensive measure, avoid large values of the mean.  The rejection
141   // algorithm used does not support very large values well.  It my be worth
142   // changing algorithms to better deal with these cases.
143   assert(mean <= 1e10);
144   if (mean_ < 10) {
145     // For small lambda, use the knuth method.
146     split_ = 1;
147     emu_ = std::exp(-mean_);
148   } else if (mean_ <= 50) {
149     // Use split-knuth method.
150     split_ = 1 + static_cast<int>(mean_ / 10.0);
151     emu_ = std::exp(-mean_ / static_cast<double>(split_));
152   } else {
153     // Use ratio of uniforms method.
154     constexpr double k2E = 0.7357588823428846;
155     constexpr double kSA = 0.4494580810294493;
156 
157     lmu_ = std::log(mean_);
158     double a = mean_ + 0.5;
159     s_ = kSA + std::sqrt(k2E * a);
160     const double mode = std::ceil(mean_) - 1;
161     log_k_ = lmu_ * mode - absl::random_internal::StirlingLogFactorial(mode);
162   }
163 }
164 
165 template <typename IntType>
166 template <typename URBG>
167 typename poisson_distribution<IntType>::result_type
operator()168 poisson_distribution<IntType>::operator()(
169     URBG& g,  // NOLINT(runtime/references)
170     const param_type& p) {
171   using random_internal::GeneratePositiveTag;
172   using random_internal::GenerateRealFromBits;
173   using random_internal::GenerateSignedTag;
174 
175   if (p.split_ != 0) {
176     // Use Knuth's algorithm with range splitting to avoid floating-point
177     // errors. Knuth's algorithm is: Ui is a sequence of uniform variates on
178     // (0,1); return the number of variates required for product(Ui) <
179     // exp(-lambda).
180     //
181     // The expected number of variates required for Knuth's method can be
182     // computed as follows:
183     // The expected value of U is 0.5, so solving for 0.5^n < exp(-lambda) gives
184     // the expected number of uniform variates
185     // required for a given lambda, which is:
186     //  lambda = [2, 5,  9, 10, 11, 12, 13, 14, 15, 16, 17]
187     //  n      = [3, 8, 13, 15, 16, 18, 19, 21, 22, 24, 25]
188     //
189     result_type n = 0;
190     for (int split = p.split_; split > 0; --split) {
191       double r = 1.0;
192       do {
193         r *= GenerateRealFromBits<double, GeneratePositiveTag, true>(
194             fast_u64_(g));  // U(-1, 0)
195         ++n;
196       } while (r > p.emu_);
197       --n;
198     }
199     return n;
200   }
201 
202   // Use ratio of uniforms method.
203   //
204   // Let u ~ Uniform(0, 1), v ~ Uniform(-1, 1),
205   //     a = lambda + 1/2,
206   //     s = 1.5 - sqrt(3/e) + sqrt(2(lambda + 1/2)/e),
207   //     x = s * v/u + a.
208   // P(floor(x) = k | u^2 < f(floor(x))/k), where
209   // f(m) = lambda^m exp(-lambda)/ m!, for 0 <= m, and f(m) = 0 otherwise,
210   // and k = max(f).
211   const double a = p.mean_ + 0.5;
212   for (;;) {
213     const double u = GenerateRealFromBits<double, GeneratePositiveTag, false>(
214         fast_u64_(g));  // U(0, 1)
215     const double v = GenerateRealFromBits<double, GenerateSignedTag, false>(
216         fast_u64_(g));  // U(-1, 1)
217 
218     const double x = std::floor(p.s_ * v / u + a);
219     if (x < 0) continue;  // f(negative) = 0
220     const double rhs = x * p.lmu_;
221     // clang-format off
222     double s = (x <= 1.0) ? 0.0
223              : (x == 2.0) ? 0.693147180559945
224              : absl::random_internal::StirlingLogFactorial(x);
225     // clang-format on
226     const double lhs = 2.0 * std::log(u) + p.log_k_ + s;
227     if (lhs < rhs) {
228       return x > static_cast<double>((max)())
229                  ? (max)()
230                  : static_cast<result_type>(x);  // f(x)/k >= u^2
231     }
232   }
233 }
234 
235 template <typename CharT, typename Traits, typename IntType>
236 std::basic_ostream<CharT, Traits>& operator<<(
237     std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
238     const poisson_distribution<IntType>& x) {
239   auto saver = random_internal::make_ostream_state_saver(os);
240   os.precision(random_internal::stream_precision_helper<double>::kPrecision);
241   os << x.mean();
242   return os;
243 }
244 
245 template <typename CharT, typename Traits, typename IntType>
246 std::basic_istream<CharT, Traits>& operator>>(
247     std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
248     poisson_distribution<IntType>& x) {     // NOLINT(runtime/references)
249   using param_type = typename poisson_distribution<IntType>::param_type;
250 
251   auto saver = random_internal::make_istream_state_saver(is);
252   double mean = random_internal::read_floating_point<double>(is);
253   if (!is.fail()) {
254     x.param(param_type(mean));
255   }
256   return is;
257 }
258 
259 ABSL_NAMESPACE_END
260 }  // namespace absl
261 
262 #endif  // ABSL_RANDOM_POISSON_DISTRIBUTION_H_
263