1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
//! Deterministic seedable PRNG (xorshift64), shared crate-wide.
//!
//! One PRNG implementation for the whole crate — `model_selection` splitters,
//! `cluster::KMeans`, `decomposition::randomized_svd`, and `tree` estimators —
//! keeps runs fully reproducible with zero external dependencies: no `rand`
//! crate, identical sequences for identical seeds on every platform.
//!
//! Moved here (from `model_selection::rng`) in v0.7; the old path re-exports
//! this type so existing code keeps compiling.
/// xorshift64 uniform PRNG. Identical seeds produce identical sequences.
///
/// Seed `0` is mapped to a non-zero splinter constant to avoid the degenerate
/// all-zero state.
#[derive(Debug, Clone)]
pub struct Rng {
state: u64,
}
impl Rng {
/// Create a new generator. `seed == 0` is remapped to a fixed non-zero
/// constant (the golden-ratio splinter) so the stream is never stuck at 0.
pub fn new(seed: u64) -> Self {
Self {
state: if seed == 0 {
0x9E37_79B9_7F4A_7C15
} else {
seed
},
}
}
/// Advance the state and return the next raw `u64`.
pub fn next_u64(&mut self) -> u64 {
self.state ^= self.state << 13;
self.state ^= self.state >> 7;
self.state ^= self.state << 17;
self.state
}
/// Uniform `f64` in `[0, 1)`.
pub fn next_unit(&mut self) -> f64 {
(self.next_u64() >> 11) as f64 / (1u64 << 53) as f64
}
/// Uniform `f64` in `[lo, hi)`. Panics if `lo >= hi` or either bound is
/// non-finite.
pub fn next_range(&mut self, lo: f64, hi: f64) -> f64 {
assert!(
lo.is_finite() && hi.is_finite() && lo < hi,
"next_range requires finite lo < hi"
);
// Avoid forming `hi - lo`, which can overflow for finite endpoints
// such as `(-f64::MAX, f64::MAX)`. This convex combination remains
// finite for every finite `lo < hi`.
let unit = self.next_unit();
lo * (1.0 - unit) + hi * unit
}
/// Uniform `usize` in `[0, hi)`. `hi == 0` returns `0` (no-op).
///
/// The float-based mapping is retained for compatibility with the
/// pre-v0.7 model-selection RNG stream; the final clamp protects the
/// documented bound in the extreme `usize` case. Prefer `next_u64` /
/// `next_unit` with integer rejection sampling when unbiased
/// subsampling matters.
pub fn next_usize(&mut self, hi: usize) -> usize {
self.next_index(hi)
}
/// Uniform `usize` in `[0, hi)`. `hi == 0` returns `0`.
pub fn next_index(&mut self, hi: usize) -> usize {
if hi == 0 {
return 0;
}
let index = (self.next_unit() * hi as f64) as usize;
index.min(hi - 1)
}
/// Standard normal sample via the Box–Muller transform.
pub fn next_normal(&mut self) -> f64 {
let u1 = self.next_unit().max(1e-300);
let u2 = self.next_unit();
let r = (-2.0 * u1.ln()).sqrt();
let theta = 2.0 * std::f64::consts::PI * u2;
r * theta.cos()
}
/// Shuffle a slice in place using the Fisher–Yates algorithm.
pub fn shuffle<T>(&mut self, slice: &mut [T]) {
let n = slice.len();
if n < 2 {
return;
}
for i in (1..n).rev() {
let j = self.next_index(i + 1);
slice.swap(i, j);
}
}
}
impl Default for Rng {
fn default() -> Self {
Rng::new(0)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn determinism_same_seed() {
let mut a = Rng::new(42);
let mut b = Rng::new(42);
for _ in 0..100 {
assert_eq!(a.next_u64(), b.next_u64());
}
}
#[test]
fn different_seeds_diverge() {
let mut a = Rng::new(1);
let mut b = Rng::new(2);
let mut diffs = 0;
for _ in 0..10 {
if a.next_u64() != b.next_u64() {
diffs += 1;
}
}
assert!(diffs > 5);
}
#[test]
fn next_unit_in_unit_interval() {
let mut rng = Rng::new(7);
for _ in 0..1000 {
let u = rng.next_unit();
assert!((0.0..1.0).contains(&u), "next_unit out of range: {u}");
}
}
#[test]
fn next_usize_bounded() {
let mut rng = Rng::new(99);
for _ in 0..1000 {
let k = rng.next_usize(10);
assert!(k < 10, "next_usize out of range: {k}");
}
assert_eq!(rng.next_usize(0), 0);
}
#[test]
fn next_index_bounded() {
let mut rng = Rng::new(99);
for _ in 0..1000 {
let k = rng.next_index(10);
assert!(k < 10, "next_index out of range: {k}");
}
assert_eq!(rng.next_index(0), 0);
}
#[test]
fn next_range_bounded() {
let mut rng = Rng::new(3);
for _ in 0..1000 {
let v = rng.next_range(-2.0, 5.0);
assert!((-2.0..5.0).contains(&v), "next_range out of range: {v}");
}
}
#[test]
fn next_range_handles_extreme_finite_bounds() {
let mut rng = Rng::new(123);
for _ in 0..1000 {
let value = rng.next_range(-f64::MAX, f64::MAX);
assert!(value.is_finite());
assert!((-f64::MAX..f64::MAX).contains(&value));
}
}
#[test]
fn next_usize_preserves_legacy_sequence() {
let mut rng = Rng::new(42);
let mut expected = Rng::new(42);
for _ in 0..100 {
let old_style = (expected.next_unit() * 37.0) as usize;
assert_eq!(rng.next_usize(37), old_style.min(36));
}
}
#[test]
fn next_normal_finite() {
let mut rng = Rng::new(11);
for _ in 0..1000 {
assert!(rng.next_normal().is_finite());
}
}
#[test]
fn shuffle_is_permutation() {
let mut rng = Rng::new(5);
let mut data: Vec<usize> = (0..64).collect();
let sorted = data.clone();
rng.shuffle(&mut data);
data.sort_unstable();
assert_eq!(data, sorted);
}
#[test]
fn zero_seed_is_remapped_not_stuck() {
let mut rng = Rng::new(0);
let a = rng.next_u64();
let b = rng.next_u64();
assert_ne!(a, b, "all-zero state would be stuck");
assert_ne!(a, 0);
}
}