criterion/stats/univariate/
sample.rs

1use std::{mem, ops};
2
3use cast;
4use rayon::prelude::*;
5use stats::float::Float;
6
7use stats::tuple::{Tuple, TupledDistributionsBuilder};
8use stats::univariate::Percentiles;
9use stats::univariate::Resamples;
10
11/// A collection of data points drawn from a population
12///
13/// Invariants:
14///
15/// - The sample contains at least 2 data points
16/// - The sample contains no `NaN`s
17pub struct Sample<A>([A]);
18
19// TODO(rust-lang/rfcs#735) move this `impl` into a private percentiles module
20impl<A> Sample<A>
21where
22    A: Float,
23{
24    /// Creates a new sample from an existing slice
25    ///
26    /// # Panics
27    ///
28    /// Panics if `slice` contains any `NaN` or if `slice` has less than two elements
29    #[cfg_attr(feature = "cargo-clippy", allow(clippy::new_ret_no_self))]
30    pub fn new(slice: &[A]) -> &Sample<A> {
31        assert!(slice.len() > 1 && slice.iter().all(|x| !x.is_nan()));
32
33        unsafe { mem::transmute(slice) }
34    }
35
36    /// Returns the biggest element in the sample
37    ///
38    /// - Time: `O(length)`
39    pub fn max(&self) -> A {
40        let mut elems = self.iter();
41
42        match elems.next() {
43            Some(&head) => elems.fold(head, |a, &b| a.max(b)),
44            // NB `unreachable!` because `Sample` is guaranteed to have at least one data point
45            None => unreachable!(),
46        }
47    }
48
49    /// Returns the arithmetic average of the sample
50    ///
51    /// - Time: `O(length)`
52    pub fn mean(&self) -> A {
53        let n = self.len();
54
55        self.sum() / A::cast(n)
56    }
57
58    /// Returns the median absolute deviation
59    ///
60    /// The `median` can be optionally passed along to speed up (2X) the computation
61    ///
62    /// - Time: `O(length)`
63    /// - Memory: `O(length)`
64    pub fn median_abs_dev(&self, median: Option<A>) -> A
65    where
66        usize: cast::From<A, Output = Result<usize, cast::Error>>,
67    {
68        let median = median.unwrap_or_else(|| self.percentiles().median());
69
70        // NB Although this operation can be SIMD accelerated, the gain is negligible because the
71        // bottle neck is the sorting operation which is part of the computation of the median
72        let abs_devs = self.iter().map(|&x| (x - median).abs()).collect::<Vec<_>>();
73
74        let abs_devs: &Self = Self::new(&abs_devs);
75
76        abs_devs.percentiles().median() * A::cast(1.4826)
77    }
78
79    /// Returns the median absolute deviation as a percentage of the median
80    ///
81    /// - Time: `O(length)`
82    /// - Memory: `O(length)`
83    pub fn median_abs_dev_pct(&self) -> A
84    where
85        usize: cast::From<A, Output = Result<usize, cast::Error>>,
86    {
87        let _100 = A::cast(100);
88        let median = self.percentiles().median();
89        let mad = self.median_abs_dev(Some(median));
90
91        (mad / median) * _100
92    }
93
94    /// Returns the smallest element in the sample
95    ///
96    /// - Time: `O(length)`
97    pub fn min(&self) -> A {
98        let mut elems = self.iter();
99
100        match elems.next() {
101            Some(&elem) => elems.fold(elem, |a, &b| a.min(b)),
102            // NB `unreachable!` because `Sample` is guaranteed to have at least one data point
103            None => unreachable!(),
104        }
105    }
106
107    /// Returns a "view" into the percentiles of the sample
108    ///
109    /// This "view" makes consecutive computations of percentiles much faster (`O(1)`)
110    ///
111    /// - Time: `O(N log N) where N = length`
112    /// - Memory: `O(length)`
113    pub fn percentiles(&self) -> Percentiles<A>
114    where
115        usize: cast::From<A, Output = Result<usize, cast::Error>>,
116    {
117        use std::cmp::Ordering;
118
119        // NB This function assumes that there are no `NaN`s in the sample
120        fn cmp<T>(a: &T, b: &T) -> Ordering
121        where
122            T: PartialOrd,
123        {
124            if a < b {
125                Ordering::Less
126            } else if a == b {
127                Ordering::Equal
128            } else {
129                Ordering::Greater
130            }
131        }
132
133        let mut v = self.to_vec().into_boxed_slice();
134        v.par_sort_unstable_by(cmp);
135
136        // NB :-1: to intra-crate privacy rules
137        unsafe { mem::transmute(v) }
138    }
139
140    /// Returns the standard deviation of the sample
141    ///
142    /// The `mean` can be optionally passed along to speed up (2X) the computation
143    ///
144    /// - Time: `O(length)`
145    pub fn std_dev(&self, mean: Option<A>) -> A {
146        self.var(mean).sqrt()
147    }
148
149    /// Returns the standard deviation as a percentage of the mean
150    ///
151    /// - Time: `O(length)`
152    pub fn std_dev_pct(&self) -> A {
153        let _100 = A::cast(100);
154        let mean = self.mean();
155        let std_dev = self.std_dev(Some(mean));
156
157        (std_dev / mean) * _100
158    }
159
160    /// Returns the sum of all the elements of the sample
161    ///
162    /// - Time: `O(length)`
163    pub fn sum(&self) -> A {
164        ::stats::sum(self)
165    }
166
167    /// Returns the t score between these two samples
168    ///
169    /// - Time: `O(length)`
170    pub fn t(&self, other: &Sample<A>) -> A {
171        let (x_bar, y_bar) = (self.mean(), other.mean());
172        let (s2_x, s2_y) = (self.var(Some(x_bar)), other.var(Some(y_bar)));
173        let n_x = A::cast(self.len());
174        let n_y = A::cast(other.len());
175        let num = x_bar - y_bar;
176        let den = (s2_x / n_x + s2_y / n_y).sqrt();
177
178        num / den
179    }
180
181    /// Returns the variance of the sample
182    ///
183    /// The `mean` can be optionally passed along to speed up (2X) the computation
184    ///
185    /// - Time: `O(length)`
186    pub fn var(&self, mean: Option<A>) -> A {
187        use std::ops::Add;
188
189        let mean = mean.unwrap_or_else(|| self.mean());
190        let slice = self;
191
192        let sum = slice
193            .iter()
194            .map(|&x| (x - mean).powi(2))
195            .fold(A::cast(0), Add::add);
196
197        sum / A::cast(slice.len() - 1)
198    }
199
200    // TODO Remove the `T` parameter in favor of `S::Output`
201    /// Returns the bootstrap distributions of the parameters estimated by the 1-sample statistic
202    ///
203    /// - Multi-threaded
204    /// - Time: `O(nresamples)`
205    /// - Memory: `O(nresamples)`
206    pub fn bootstrap<T, S>(&self, nresamples: usize, statistic: S) -> T::Distributions
207    where
208        S: Fn(&Sample<A>) -> T,
209        S: Sync,
210        T: Tuple,
211        T: Send,
212        T::Distributions: Send,
213        T::Builder: Send,
214    {
215        (0..nresamples)
216            .into_par_iter()
217            .map_init(
218                || Resamples::new(self),
219                |resamples, _| statistic(resamples.next()),
220            )
221            .fold(
222                || T::Builder::new(0),
223                |mut sub_distributions, sample| {
224                    sub_distributions.push(sample);
225                    sub_distributions
226                },
227            )
228            .reduce(
229                || T::Builder::new(0),
230                |mut a, mut b| {
231                    a.extend(&mut b);
232                    a
233                },
234            )
235            .complete()
236    }
237
238    #[cfg(test)]
239    pub fn iqr(&self) -> A
240    where
241        usize: cast::From<A, Output = Result<usize, cast::Error>>,
242    {
243        self.percentiles().iqr()
244    }
245
246    #[cfg(test)]
247    pub fn median(&self) -> A
248    where
249        usize: cast::From<A, Output = Result<usize, cast::Error>>,
250    {
251        self.percentiles().median()
252    }
253}
254
255impl<A> ops::Deref for Sample<A> {
256    type Target = [A];
257
258    fn deref(&self) -> &[A] {
259        &self.0
260    }
261}
criterion/stats/univariate/sample.rs

criterion/stats/univariate/
sample.rs
