test/
stats.rs

1#![allow(missing_docs)]
2
3use std::mem;
4
5#[cfg(test)]
6mod tests;
7
8fn local_sort(v: &mut [f64]) {
9    v.sort_by(|x: &f64, y: &f64| x.total_cmp(y));
10}
11
12/// Trait that provides simple descriptive statistics on a univariate set of numeric samples.
13pub trait Stats {
14    /// Sum of the samples.
15    ///
16    /// Note: this method sacrifices performance at the altar of accuracy
17    /// Depends on IEEE 754 arithmetic guarantees. See proof of correctness at:
18    /// ["Adaptive Precision Floating-Point Arithmetic and Fast Robust Geometric
19    /// Predicates"][paper]
20    ///
21    /// [paper]: https://www.cs.cmu.edu/~quake-papers/robust-arithmetic.ps
22    fn sum(&self) -> f64;
23
24    /// Minimum value of the samples.
25    fn min(&self) -> f64;
26
27    /// Maximum value of the samples.
28    fn max(&self) -> f64;
29
30    /// Arithmetic mean (average) of the samples: sum divided by sample-count.
31    ///
32    /// See: <https://en.wikipedia.org/wiki/Arithmetic_mean>
33    fn mean(&self) -> f64;
34
35    /// Median of the samples: value separating the lower half of the samples from the higher half.
36    /// Equal to `self.percentile(50.0)`.
37    ///
38    /// See: <https://en.wikipedia.org/wiki/Median>
39    fn median(&self) -> f64;
40
41    /// Variance of the samples: bias-corrected mean of the squares of the differences of each
42    /// sample from the sample mean. Note that this calculates the _sample variance_ rather than the
43    /// population variance, which is assumed to be unknown. It therefore corrects the `(n-1)/n`
44    /// bias that would appear if we calculated a population variance, by dividing by `(n-1)` rather
45    /// than `n`.
46    ///
47    /// See: <https://en.wikipedia.org/wiki/Variance>
48    fn var(&self) -> f64;
49
50    /// Standard deviation: the square root of the sample variance.
51    ///
52    /// Note: this is not a robust statistic for non-normal distributions. Prefer the
53    /// `median_abs_dev` for unknown distributions.
54    ///
55    /// See: <https://en.wikipedia.org/wiki/Standard_deviation>
56    fn std_dev(&self) -> f64;
57
58    /// Standard deviation as a percent of the mean value. See `std_dev` and `mean`.
59    ///
60    /// Note: this is not a robust statistic for non-normal distributions. Prefer the
61    /// `median_abs_dev_pct` for unknown distributions.
62    fn std_dev_pct(&self) -> f64;
63
64    /// Scaled median of the absolute deviations of each sample from the sample median. This is a
65    /// robust (distribution-agnostic) estimator of sample variability. Use this in preference to
66    /// `std_dev` if you cannot assume your sample is normally distributed. Note that this is scaled
67    /// by the constant `1.4826` to allow its use as a consistent estimator for the standard
68    /// deviation.
69    ///
70    /// See: <https://en.wikipedia.org/wiki/Median_absolute_deviation>
71    fn median_abs_dev(&self) -> f64;
72
73    /// Median absolute deviation as a percent of the median. See `median_abs_dev` and `median`.
74    fn median_abs_dev_pct(&self) -> f64;
75
76    /// Percentile: the value below which `pct` percent of the values in `self` fall. For example,
77    /// percentile(95.0) will return the value `v` such that 95% of the samples `s` in `self`
78    /// satisfy `s <= v`.
79    ///
80    /// Calculated by linear interpolation between closest ranks.
81    ///
82    /// See: <https://en.wikipedia.org/wiki/Percentile>
83    fn percentile(&self, pct: f64) -> f64;
84
85    /// Quartiles of the sample: three values that divide the sample into four equal groups, each
86    /// with 1/4 of the data. The middle value is the median. See `median` and `percentile`. This
87    /// function may calculate the 3 quartiles more efficiently than 3 calls to `percentile`, but
88    /// is otherwise equivalent.
89    ///
90    /// See also: <https://en.wikipedia.org/wiki/Quartile>
91    fn quartiles(&self) -> (f64, f64, f64);
92
93    /// Inter-quartile range: the difference between the 25th percentile (1st quartile) and the 75th
94    /// percentile (3rd quartile). See `quartiles`.
95    ///
96    /// See also: <https://en.wikipedia.org/wiki/Interquartile_range>
97    fn iqr(&self) -> f64;
98}
99
100/// Extracted collection of all the summary statistics of a sample set.
101#[derive(Debug, Clone, PartialEq, Copy)]
102#[allow(missing_docs)]
103pub struct Summary {
104    pub sum: f64,
105    pub min: f64,
106    pub max: f64,
107    pub mean: f64,
108    pub median: f64,
109    pub var: f64,
110    pub std_dev: f64,
111    pub std_dev_pct: f64,
112    pub median_abs_dev: f64,
113    pub median_abs_dev_pct: f64,
114    pub quartiles: (f64, f64, f64),
115    pub iqr: f64,
116}
117
118impl Summary {
119    /// Constructs a new summary of a sample set.
120    pub fn new(samples: &[f64]) -> Summary {
121        Summary {
122            sum: samples.sum(),
123            min: samples.min(),
124            max: samples.max(),
125            mean: samples.mean(),
126            median: samples.median(),
127            var: samples.var(),
128            std_dev: samples.std_dev(),
129            std_dev_pct: samples.std_dev_pct(),
130            median_abs_dev: samples.median_abs_dev(),
131            median_abs_dev_pct: samples.median_abs_dev_pct(),
132            quartiles: samples.quartiles(),
133            iqr: samples.iqr(),
134        }
135    }
136}
137
138impl Stats for [f64] {
139    // FIXME #11059 handle NaN, inf and overflow
140    fn sum(&self) -> f64 {
141        let mut partials = vec![];
142
143        for &x in self {
144            let mut x = x;
145            let mut j = 0;
146            // This inner loop applies `hi`/`lo` summation to each
147            // partial so that the list of partial sums remains exact.
148            for i in 0..partials.len() {
149                let mut y: f64 = partials[i];
150                if x.abs() < y.abs() {
151                    mem::swap(&mut x, &mut y);
152                }
153                // Rounded `x+y` is stored in `hi` with round-off stored in
154                // `lo`. Together `hi+lo` are exactly equal to `x+y`.
155                let hi = x + y;
156                let lo = y - (hi - x);
157                if lo != 0.0 {
158                    partials[j] = lo;
159                    j += 1;
160                }
161                x = hi;
162            }
163            if j >= partials.len() {
164                partials.push(x);
165            } else {
166                partials[j] = x;
167                partials.truncate(j + 1);
168            }
169        }
170        let zero: f64 = 0.0;
171        partials.iter().fold(zero, |p, q| p + *q)
172    }
173
174    fn min(&self) -> f64 {
175        assert!(!self.is_empty());
176        self.iter().fold(self[0], |p, q| p.min(*q))
177    }
178
179    fn max(&self) -> f64 {
180        assert!(!self.is_empty());
181        self.iter().fold(self[0], |p, q| p.max(*q))
182    }
183
184    fn mean(&self) -> f64 {
185        assert!(!self.is_empty());
186        self.sum() / (self.len() as f64)
187    }
188
189    fn median(&self) -> f64 {
190        self.percentile(50_f64)
191    }
192
193    fn var(&self) -> f64 {
194        if self.len() < 2 {
195            0.0
196        } else {
197            let mean = self.mean();
198            let mut v: f64 = 0.0;
199            for s in self {
200                let x = *s - mean;
201                v += x * x;
202            }
203            // N.B., this is _supposed to be_ len-1, not len. If you
204            // change it back to len, you will be calculating a
205            // population variance, not a sample variance.
206            let denom = (self.len() - 1) as f64;
207            v / denom
208        }
209    }
210
211    fn std_dev(&self) -> f64 {
212        self.var().sqrt()
213    }
214
215    fn std_dev_pct(&self) -> f64 {
216        let hundred = 100_f64;
217        (self.std_dev() / self.mean()) * hundred
218    }
219
220    fn median_abs_dev(&self) -> f64 {
221        let med = self.median();
222        let abs_devs: Vec<f64> = self.iter().map(|&v| (med - v).abs()).collect();
223        // This constant is derived by smarter statistics brains than me, but it is
224        // consistent with how R and other packages treat the MAD.
225        let number = 1.4826;
226        abs_devs.median() * number
227    }
228
229    fn median_abs_dev_pct(&self) -> f64 {
230        let hundred = 100_f64;
231        (self.median_abs_dev() / self.median()) * hundred
232    }
233
234    fn percentile(&self, pct: f64) -> f64 {
235        let mut tmp = self.to_vec();
236        local_sort(&mut tmp);
237        percentile_of_sorted(&tmp, pct)
238    }
239
240    fn quartiles(&self) -> (f64, f64, f64) {
241        let mut tmp = self.to_vec();
242        local_sort(&mut tmp);
243        let first = 25_f64;
244        let a = percentile_of_sorted(&tmp, first);
245        let second = 50_f64;
246        let b = percentile_of_sorted(&tmp, second);
247        let third = 75_f64;
248        let c = percentile_of_sorted(&tmp, third);
249        (a, b, c)
250    }
251
252    fn iqr(&self) -> f64 {
253        let (a, _, c) = self.quartiles();
254        c - a
255    }
256}
257
258// Helper function: extract a value representing the `pct` percentile of a sorted sample-set, using
259// linear interpolation. If samples are not sorted, return nonsensical value.
260fn percentile_of_sorted(sorted_samples: &[f64], pct: f64) -> f64 {
261    assert!(!sorted_samples.is_empty());
262    if sorted_samples.len() == 1 {
263        return sorted_samples[0];
264    }
265    let zero: f64 = 0.0;
266    assert!(zero <= pct);
267    let hundred = 100_f64;
268    assert!(pct <= hundred);
269    if pct == hundred {
270        return sorted_samples[sorted_samples.len() - 1];
271    }
272    let length = (sorted_samples.len() - 1) as f64;
273    let rank = (pct / hundred) * length;
274    let lrank = rank.floor();
275    let d = rank - lrank;
276    let n = lrank as usize;
277    let lo = sorted_samples[n];
278    let hi = sorted_samples[n + 1];
279    lo + (hi - lo) * d
280}
281
282/// Winsorize a set of samples, replacing values above the `100-pct` percentile
283/// and below the `pct` percentile with those percentiles themselves. This is a
284/// way of minimizing the effect of outliers, at the cost of biasing the sample.
285/// It differs from trimming in that it does not change the number of samples,
286/// just changes the values of those that are outliers.
287///
288/// See: <https://en.wikipedia.org/wiki/Winsorising>
289pub fn winsorize(samples: &mut [f64], pct: f64) {
290    let mut tmp = samples.to_vec();
291    local_sort(&mut tmp);
292    let lo = percentile_of_sorted(&tmp, pct);
293    let hundred = 100_f64;
294    let hi = percentile_of_sorted(&tmp, hundred - pct);
295    for samp in samples {
296        if *samp > hi {
297            *samp = hi
298        } else if *samp < lo {
299            *samp = lo
300        }
301    }
302}