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_GAUSSIAN_DISTRIBUTION_H_
16 #define ABSL_RANDOM_GAUSSIAN_DISTRIBUTION_H_
17
18 // absl::gaussian_distribution implements the Ziggurat algorithm
19 // for generating random gaussian numbers.
20 //
21 // Implementation based on "The Ziggurat Method for Generating Random Variables"
22 // by George Marsaglia and Wai Wan Tsang: http://www.jstatsoft.org/v05/i08/
23 //
24
25 #include <cmath>
26 #include <cstdint>
27 #include <istream>
28 #include <limits>
29 #include <type_traits>
30
31 #include "absl/base/config.h"
32 #include "absl/random/internal/fast_uniform_bits.h"
33 #include "absl/random/internal/generate_real.h"
34 #include "absl/random/internal/iostream_state_saver.h"
35
36 namespace absl {
37 ABSL_NAMESPACE_BEGIN
38 namespace random_internal {
39
40 // absl::gaussian_distribution_base implements the underlying ziggurat algorithm
41 // using the ziggurat tables generated by the gaussian_distribution_gentables
42 // binary.
43 //
44 // The specific algorithm has some of the improvements suggested by the
45 // 2005 paper, "An Improved Ziggurat Method to Generate Normal Random Samples",
46 // Jurgen A Doornik. (https://www.doornik.com/research/ziggurat.pdf)
47 class ABSL_DLL gaussian_distribution_base {
48 public:
49 template <typename URBG>
50 inline double zignor(URBG& g); // NOLINT(runtime/references)
51
52 private:
53 friend class TableGenerator;
54
55 template <typename URBG>
56 inline double zignor_fallback(URBG& g, // NOLINT(runtime/references)
57 bool neg);
58
59 // Constants used for the gaussian distribution.
60 static constexpr double kR = 3.442619855899; // Start of the tail.
61 static constexpr double kRInv = 0.29047645161474317; // ~= (1.0 / kR) .
62 static constexpr double kV = 9.91256303526217e-3;
63 static constexpr uint64_t kMask = 0x07f;
64
65 // The ziggurat tables store the pdf(f) and inverse-pdf(x) for equal-area
66 // points on one-half of the normal distribution, where the pdf function,
67 // pdf = e ^ (-1/2 *x^2), assumes that the mean = 0 & stddev = 1.
68 //
69 // These tables are just over 2kb in size; larger tables might improve the
70 // distributions, but also lead to more cache pollution.
71 //
72 // x = {3.71308, 3.44261, 3.22308, ..., 0}
73 // f = {0.00101, 0.00266, 0.00554, ..., 1}
74 struct Tables {
75 double x[kMask + 2];
76 double f[kMask + 2];
77 };
78 static const Tables zg_;
79 random_internal::FastUniformBits<uint64_t> fast_u64_;
80 };
81
82 } // namespace random_internal
83
84 // absl::gaussian_distribution:
85 // Generates a number conforming to a Gaussian distribution.
86 template <typename RealType = double>
87 class gaussian_distribution : random_internal::gaussian_distribution_base {
88 public:
89 using result_type = RealType;
90
91 class param_type {
92 public:
93 using distribution_type = gaussian_distribution;
94
95 explicit param_type(result_type mean = 0, result_type stddev = 1)
mean_(mean)96 : mean_(mean), stddev_(stddev) {}
97
98 // Returns the mean distribution parameter. The mean specifies the location
99 // of the peak. The default value is 0.0.
mean()100 result_type mean() const { return mean_; }
101
102 // Returns the deviation distribution parameter. The default value is 1.0.
stddev()103 result_type stddev() const { return stddev_; }
104
105 friend bool operator==(const param_type& a, const param_type& b) {
106 return a.mean_ == b.mean_ && a.stddev_ == b.stddev_;
107 }
108
109 friend bool operator!=(const param_type& a, const param_type& b) {
110 return !(a == b);
111 }
112
113 private:
114 result_type mean_;
115 result_type stddev_;
116
117 static_assert(
118 std::is_floating_point<RealType>::value,
119 "Class-template absl::gaussian_distribution<> must be parameterized "
120 "using a floating-point type.");
121 };
122
gaussian_distribution()123 gaussian_distribution() : gaussian_distribution(0) {}
124
125 explicit gaussian_distribution(result_type mean, result_type stddev = 1)
param_(mean,stddev)126 : param_(mean, stddev) {}
127
gaussian_distribution(const param_type & p)128 explicit gaussian_distribution(const param_type& p) : param_(p) {}
129
reset()130 void reset() {}
131
132 // Generating functions
133 template <typename URBG>
operator()134 result_type operator()(URBG& g) { // NOLINT(runtime/references)
135 return (*this)(g, param_);
136 }
137
138 template <typename URBG>
139 result_type operator()(URBG& g, // NOLINT(runtime/references)
140 const param_type& p);
141
param()142 param_type param() const { return param_; }
param(const param_type & p)143 void param(const param_type& p) { param_ = p; }
144
result_type(min)145 result_type(min)() const {
146 return -std::numeric_limits<result_type>::infinity();
147 }
result_type(max)148 result_type(max)() const {
149 return std::numeric_limits<result_type>::infinity();
150 }
151
mean()152 result_type mean() const { return param_.mean(); }
stddev()153 result_type stddev() const { return param_.stddev(); }
154
155 friend bool operator==(const gaussian_distribution& a,
156 const gaussian_distribution& b) {
157 return a.param_ == b.param_;
158 }
159 friend bool operator!=(const gaussian_distribution& a,
160 const gaussian_distribution& b) {
161 return a.param_ != b.param_;
162 }
163
164 private:
165 param_type param_;
166 };
167
168 // --------------------------------------------------------------------------
169 // Implementation details only below
170 // --------------------------------------------------------------------------
171
172 template <typename RealType>
173 template <typename URBG>
174 typename gaussian_distribution<RealType>::result_type
operator()175 gaussian_distribution<RealType>::operator()(
176 URBG& g, // NOLINT(runtime/references)
177 const param_type& p) {
178 return p.mean() + p.stddev() * static_cast<result_type>(zignor(g));
179 }
180
181 template <typename CharT, typename Traits, typename RealType>
182 std::basic_ostream<CharT, Traits>& operator<<(
183 std::basic_ostream<CharT, Traits>& os, // NOLINT(runtime/references)
184 const gaussian_distribution<RealType>& x) {
185 auto saver = random_internal::make_ostream_state_saver(os);
186 os.precision(random_internal::stream_precision_helper<RealType>::kPrecision);
187 os << x.mean() << os.fill() << x.stddev();
188 return os;
189 }
190
191 template <typename CharT, typename Traits, typename RealType>
192 std::basic_istream<CharT, Traits>& operator>>(
193 std::basic_istream<CharT, Traits>& is, // NOLINT(runtime/references)
194 gaussian_distribution<RealType>& x) { // NOLINT(runtime/references)
195 using result_type = typename gaussian_distribution<RealType>::result_type;
196 using param_type = typename gaussian_distribution<RealType>::param_type;
197
198 auto saver = random_internal::make_istream_state_saver(is);
199 auto mean = random_internal::read_floating_point<result_type>(is);
200 if (is.fail()) return is;
201 auto stddev = random_internal::read_floating_point<result_type>(is);
202 if (!is.fail()) {
203 x.param(param_type(mean, stddev));
204 }
205 return is;
206 }
207
208 namespace random_internal {
209
210 template <typename URBG>
zignor_fallback(URBG & g,bool neg)211 inline double gaussian_distribution_base::zignor_fallback(URBG& g, bool neg) {
212 using random_internal::GeneratePositiveTag;
213 using random_internal::GenerateRealFromBits;
214
215 // This fallback path happens approximately 0.05% of the time.
216 double x, y;
217 do {
218 // kRInv = 1/r, U(0, 1)
219 x = kRInv *
220 std::log(GenerateRealFromBits<double, GeneratePositiveTag, false>(
221 fast_u64_(g)));
222 y = -std::log(
223 GenerateRealFromBits<double, GeneratePositiveTag, false>(fast_u64_(g)));
224 } while ((y + y) < (x * x));
225 return neg ? (x - kR) : (kR - x);
226 }
227
228 template <typename URBG>
zignor(URBG & g)229 inline double gaussian_distribution_base::zignor(
230 URBG& g) { // NOLINT(runtime/references)
231 using random_internal::GeneratePositiveTag;
232 using random_internal::GenerateRealFromBits;
233 using random_internal::GenerateSignedTag;
234
235 while (true) {
236 // We use a single uint64_t to generate both a double and a strip.
237 // These bits are unused when the generated double is > 1/2^5.
238 // This may introduce some bias from the duplicated low bits of small
239 // values (those smaller than 1/2^5, which all end up on the left tail).
240 uint64_t bits = fast_u64_(g);
241 int i = static_cast<int>(bits & kMask); // pick a random strip
242 double j = GenerateRealFromBits<double, GenerateSignedTag, false>(
243 bits); // U(-1, 1)
244 const double x = j * zg_.x[i];
245
246 // Retangular box. Handles >97% of all cases.
247 // For any given box, this handles between 75% and 99% of values.
248 // Equivalent to U(01) < (x[i+1] / x[i]), and when i == 0, ~93.5%
249 if (std::abs(x) < zg_.x[i + 1]) {
250 return x;
251 }
252
253 // i == 0: Base box. Sample using a ratio of uniforms.
254 if (i == 0) {
255 // This path happens about 0.05% of the time.
256 return zignor_fallback(g, j < 0);
257 }
258
259 // i > 0: Wedge samples using precomputed values.
260 double v = GenerateRealFromBits<double, GeneratePositiveTag, false>(
261 fast_u64_(g)); // U(0, 1)
262 if ((zg_.f[i + 1] + v * (zg_.f[i] - zg_.f[i + 1])) <
263 std::exp(-0.5 * x * x)) {
264 return x;
265 }
266
267 // The wedge was missed; reject the value and try again.
268 }
269 }
270
271 } // namespace random_internal
272 ABSL_NAMESPACE_END
273 } // namespace absl
274
275 #endif // ABSL_RANDOM_GAUSSIAN_DISTRIBUTION_H_
276