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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
// Copyright 2013 The Rust Project Developers. See the COPYRIGHT
// file at the top-level directory of this distribution and at
// http://rust-lang.org/COPYRIGHT.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.

//! Sampling from random distributions.
//!
//! This is a generalization of `Rand` to allow parameters to control the
//! exact properties of the generated values, e.g. the mean and standard
//! deviation of a normal distribution. The `Sample` trait is the most
//! general, and allows for generating values that change some state
//! internally. The `IndependentSample` trait is for generating values
//! that do not need to record state.

use std::marker;

use {Rng, Rand};

pub use self::range::Range;
pub use self::gamma::{Gamma, ChiSquared, FisherF, StudentT};
pub use self::normal::{Normal, LogNormal};
pub use self::exponential::Exp;

pub mod range;
pub mod gamma;
pub mod normal;
pub mod exponential;

/// Types that can be used to create a random instance of `Support`.
pub trait Sample<Support> {
    /// Generate a random value of `Support`, using `rng` as the
    /// source of randomness.
    fn sample<R: Rng>(&mut self, rng: &mut R) -> Support;
}

/// `Sample`s that do not require keeping track of state.
///
/// Since no state is recorded, each sample is (statistically)
/// independent of all others, assuming the `Rng` used has this
/// property.
// FIXME maybe having this separate is overkill (the only reason is to
// take &self rather than &mut self)? or maybe this should be the
// trait called `Sample` and the other should be `DependentSample`.
pub trait IndependentSample<Support>: Sample<Support> {
    /// Generate a random value.
    fn ind_sample<R: Rng>(&self, &mut R) -> Support;
}

/// A wrapper for generating types that implement `Rand` via the
/// `Sample` & `IndependentSample` traits.
pub struct RandSample<Sup> {
    _marker: marker::PhantomData<fn() -> Sup>,
}

impl<Sup> Copy for RandSample<Sup> {}
impl<Sup> Clone for RandSample<Sup> {
    fn clone(&self) -> Self { *self }
}

impl<Sup: Rand> Sample<Sup> for RandSample<Sup> {
    fn sample<R: Rng>(&mut self, rng: &mut R) -> Sup { self.ind_sample(rng) }
}

impl<Sup: Rand> IndependentSample<Sup> for RandSample<Sup> {
    fn ind_sample<R: Rng>(&self, rng: &mut R) -> Sup {
        rng.gen()
    }
}

impl<Sup> RandSample<Sup> {
    pub fn new() -> RandSample<Sup> {
        RandSample { _marker: marker::PhantomData }
    }
}

/// A value with a particular weight for use with `WeightedChoice`.
#[derive(Copy)]
#[derive(Clone)]
pub struct Weighted<T> {
    /// The numerical weight of this item
    pub weight: u32,
    /// The actual item which is being weighted
    pub item: T,
}

/// A distribution that selects from a finite collection of weighted items.
///
/// Each item has an associated weight that influences how likely it
/// is to be chosen: higher weight is more likely.
///
/// The `Clone` restriction is a limitation of the `Sample` and
/// `IndependentSample` traits. Note that `&T` is (cheaply) `Clone` for
/// all `T`, as is `u32`, so one can store references or indices into
/// another vector.
///
/// # Example
///
/// ```rust
/// use rand::distributions::{Weighted, WeightedChoice, IndependentSample};
///
/// let mut items = vec!(Weighted { weight: 2, item: 'a' },
///                      Weighted { weight: 4, item: 'b' },
///                      Weighted { weight: 1, item: 'c' });
/// let wc = WeightedChoice::new(&mut items);
/// let mut rng = rand::thread_rng();
/// for _ in 0..16 {
///      // on average prints 'a' 4 times, 'b' 8 and 'c' twice.
///      println!("{}", wc.ind_sample(&mut rng));
/// }
/// ```
pub struct WeightedChoice<'a, T:'a> {
    items: &'a mut [Weighted<T>],
    weight_range: Range<u32>
}

impl<'a, T: Clone> WeightedChoice<'a, T> {
    /// Create a new `WeightedChoice`.
    ///
    /// Panics if:
    /// - `v` is empty
    /// - the total weight is 0
    /// - the total weight is larger than a `u32` can contain.
    pub fn new(items: &'a mut [Weighted<T>]) -> WeightedChoice<'a, T> {
        // strictly speaking, this is subsumed by the total weight == 0 case
        assert!(!items.is_empty(), "WeightedChoice::new called with no items");

        let mut running_total: u32 = 0;

        // we convert the list from individual weights to cumulative
        // weights so we can binary search. This *could* drop elements
        // with weight == 0 as an optimisation.
        for item in items.iter_mut() {
            running_total = match running_total.checked_add(item.weight) {
                Some(n) => n,
                None => panic!("WeightedChoice::new called with a total weight \
                               larger than a u32 can contain")
            };

            item.weight = running_total;
        }
        assert!(running_total != 0, "WeightedChoice::new called with a total weight of 0");

        WeightedChoice {
            items: items,
            // we're likely to be generating numbers in this range
            // relatively often, so might as well cache it
            weight_range: Range::new(0, running_total)
        }
    }
}

impl<'a, T: Clone> Sample<T> for WeightedChoice<'a, T> {
    fn sample<R: Rng>(&mut self, rng: &mut R) -> T { self.ind_sample(rng) }
}

impl<'a, T: Clone> IndependentSample<T> for WeightedChoice<'a, T> {
    fn ind_sample<R: Rng>(&self, rng: &mut R) -> T {
        // we want to find the first element that has cumulative
        // weight > sample_weight, which we do by binary since the
        // cumulative weights of self.items are sorted.

        // choose a weight in [0, total_weight)
        let sample_weight = self.weight_range.ind_sample(rng);

        // short circuit when it's the first item
        if sample_weight < self.items[0].weight {
            return self.items[0].item.clone();
        }

        let mut idx = 0;
        let mut modifier = self.items.len();

        // now we know that every possibility has an element to the
        // left, so we can just search for the last element that has
        // cumulative weight <= sample_weight, then the next one will
        // be "it". (Note that this greatest element will never be the
        // last element of the vector, since sample_weight is chosen
        // in [0, total_weight) and the cumulative weight of the last
        // one is exactly the total weight.)
        while modifier > 1 {
            let i = idx + modifier / 2;
            if self.items[i].weight <= sample_weight {
                // we're small, so look to the right, but allow this
                // exact element still.
                idx = i;
                // we need the `/ 2` to round up otherwise we'll drop
                // the trailing elements when `modifier` is odd.
                modifier += 1;
            } else {
                // otherwise we're too big, so go left. (i.e. do
                // nothing)
            }
            modifier /= 2;
        }
        return self.items[idx + 1].item.clone();
    }
}

mod ziggurat_tables;

/// Sample a random number using the Ziggurat method (specifically the
/// ZIGNOR variant from Doornik 2005). Most of the arguments are
/// directly from the paper:
///
/// * `rng`: source of randomness
/// * `symmetric`: whether this is a symmetric distribution, or one-sided with P(x < 0) = 0.
/// * `X`: the $x_i$ abscissae.
/// * `F`: precomputed values of the PDF at the $x_i$, (i.e. $f(x_i)$)
/// * `F_DIFF`: precomputed values of $f(x_i) - f(x_{i+1})$
/// * `pdf`: the probability density function
/// * `zero_case`: manual sampling from the tail when we chose the
///    bottom box (i.e. i == 0)

// the perf improvement (25-50%) is definitely worth the extra code
// size from force-inlining.
#[inline(always)]
fn ziggurat<R: Rng, P, Z>(
            rng: &mut R,
            symmetric: bool,
            x_tab: ziggurat_tables::ZigTable,
            f_tab: ziggurat_tables::ZigTable,
            mut pdf: P,
            mut zero_case: Z)
            -> f64 where P: FnMut(f64) -> f64, Z: FnMut(&mut R, f64) -> f64 {
    const SCALE: f64 = (1u64 << 53) as f64;
    loop {
        // reimplement the f64 generation as an optimisation suggested
        // by the Doornik paper: we have a lot of precision-space
        // (i.e. there are 11 bits of the 64 of a u64 to use after
        // creating a f64), so we might as well reuse some to save
        // generating a whole extra random number. (Seems to be 15%
        // faster.)
        //
        // This unfortunately misses out on the benefits of direct
        // floating point generation if an RNG like dSMFT is
        // used. (That is, such RNGs create floats directly, highly
        // efficiently and overload next_f32/f64, so by not calling it
        // this may be slower than it would be otherwise.)
        // FIXME: investigate/optimise for the above.
        let bits: u64 = rng.gen();
        let i = (bits & 0xff) as usize;
        let f = (bits >> 11) as f64 / SCALE;

        // u is either U(-1, 1) or U(0, 1) depending on if this is a
        // symmetric distribution or not.
        let u = if symmetric {2.0 * f - 1.0} else {f};
        let x = u * x_tab[i];

        let test_x = if symmetric { x.abs() } else {x};

        // algebraically equivalent to |u| < x_tab[i+1]/x_tab[i] (or u < x_tab[i+1]/x_tab[i])
        if test_x < x_tab[i + 1] {
            return x;
        }
        if i == 0 {
            return zero_case(rng, u);
        }
        // algebraically equivalent to f1 + DRanU()*(f0 - f1) < 1
        if f_tab[i + 1] + (f_tab[i] - f_tab[i + 1]) * rng.gen::<f64>() < pdf(x) {
            return x;
        }
    }
}

#[cfg(test)]
mod tests {

    use {Rng, Rand};
    use super::{RandSample, WeightedChoice, Weighted, Sample, IndependentSample};

    #[derive(PartialEq, Debug)]
    struct ConstRand(usize);
    impl Rand for ConstRand {
        fn rand<R: Rng>(_: &mut R) -> ConstRand {
            ConstRand(0)
        }
    }

    // 0, 1, 2, 3, ...
    struct CountingRng { i: u32 }
    impl Rng for CountingRng {
        fn next_u32(&mut self) -> u32 {
            self.i += 1;
            self.i - 1
        }
        fn next_u64(&mut self) -> u64 {
            self.next_u32() as u64
        }
    }

    #[test]
    fn test_rand_sample() {
        let mut rand_sample = RandSample::<ConstRand>::new();

        assert_eq!(rand_sample.sample(&mut ::test::rng()), ConstRand(0));
        assert_eq!(rand_sample.ind_sample(&mut ::test::rng()), ConstRand(0));
    }
    #[test]
    fn test_weighted_choice() {
        // this makes assumptions about the internal implementation of
        // WeightedChoice, specifically: it doesn't reorder the items,
        // it doesn't do weird things to the RNG (so 0 maps to 0, 1 to
        // 1, internally; modulo a modulo operation).

        macro_rules! t {
            ($items:expr, $expected:expr) => {{
                let mut items = $items;
                let wc = WeightedChoice::new(&mut items);
                let expected = $expected;

                let mut rng = CountingRng { i: 0 };

                for &val in expected.iter() {
                    assert_eq!(wc.ind_sample(&mut rng), val)
                }
            }}
        }

        t!(vec!(Weighted { weight: 1, item: 10}), [10]);

        // skip some
        t!(vec!(Weighted { weight: 0, item: 20},
                Weighted { weight: 2, item: 21},
                Weighted { weight: 0, item: 22},
                Weighted { weight: 1, item: 23}),
           [21,21, 23]);

        // different weights
        t!(vec!(Weighted { weight: 4, item: 30},
                Weighted { weight: 3, item: 31}),
           [30,30,30,30, 31,31,31]);

        // check that we're binary searching
        // correctly with some vectors of odd
        // length.
        t!(vec!(Weighted { weight: 1, item: 40},
                Weighted { weight: 1, item: 41},
                Weighted { weight: 1, item: 42},
                Weighted { weight: 1, item: 43},
                Weighted { weight: 1, item: 44}),
           [40, 41, 42, 43, 44]);
        t!(vec!(Weighted { weight: 1, item: 50},
                Weighted { weight: 1, item: 51},
                Weighted { weight: 1, item: 52},
                Weighted { weight: 1, item: 53},
                Weighted { weight: 1, item: 54},
                Weighted { weight: 1, item: 55},
                Weighted { weight: 1, item: 56}),
           [50, 51, 52, 53, 54, 55, 56]);
    }

    #[test]
    fn test_weighted_clone_initialization() {
        let initial : Weighted<u32> = Weighted {weight: 1, item: 1};
        let clone = initial.clone();
        assert_eq!(initial.weight, clone.weight);
        assert_eq!(initial.item, clone.item);
    }

    #[test] #[should_panic]
    fn test_weighted_clone_change_weight() {
        let initial : Weighted<u32> = Weighted {weight: 1, item: 1};
        let mut clone = initial.clone();
        clone.weight = 5;
        assert_eq!(initial.weight, clone.weight);
    }

    #[test] #[should_panic]
    fn test_weighted_clone_change_item() {
        let initial : Weighted<u32> = Weighted {weight: 1, item: 1};
        let mut clone = initial.clone();
        clone.item = 5;
        assert_eq!(initial.item, clone.item);

    }

    #[test] #[should_panic]
    fn test_weighted_choice_no_items() {
        WeightedChoice::<isize>::new(&mut []);
    }
    #[test] #[should_panic]
    fn test_weighted_choice_zero_weight() {
        WeightedChoice::new(&mut [Weighted { weight: 0, item: 0},
                                  Weighted { weight: 0, item: 1}]);
    }
    #[test] #[should_panic]
    fn test_weighted_choice_weight_overflows() {
        let x = ::std::u32::MAX / 2; // x + x + 2 is the overflow
        WeightedChoice::new(&mut [Weighted { weight: x, item: 0 },
                                  Weighted { weight: 1, item: 1 },
                                  Weighted { weight: x, item: 2 },
                                  Weighted { weight: 1, item: 3 }]);
    }
}