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
#![allow(missing_docs)]
#![allow(deprecated)] // Float

use std::cmp::Ordering::{self, Equal, Greater, Less};
use std::mem;

#[cfg(test)]
mod tests;

fn local_cmp(x: f64, y: f64) -> Ordering {
    // arbitrarily decide that NaNs are larger than everything.
    if y.is_nan() {
        Less
    } else if x.is_nan() {
        Greater
    } else if x < y {
        Less
    } else if x == y {
        Equal
    } else {
        Greater
    }
}

fn local_sort(v: &mut [f64]) {
    v.sort_by(|x: &f64, y: &f64| local_cmp(*x, *y));
}

/// Trait that provides simple descriptive statistics on a univariate set of numeric samples.
pub trait Stats {
    /// Sum of the samples.
    ///
    /// Note: this method sacrifices performance at the altar of accuracy
    /// Depends on IEEE-754 arithmetic guarantees. See proof of correctness at:
    /// ["Adaptive Precision Floating-Point Arithmetic and Fast Robust Geometric
    /// Predicates"][paper]
    ///
    /// [paper]: http://www.cs.cmu.edu/~quake-papers/robust-arithmetic.ps
    fn sum(&self) -> f64;

    /// Minimum value of the samples.
    fn min(&self) -> f64;

    /// Maximum value of the samples.
    fn max(&self) -> f64;

    /// Arithmetic mean (average) of the samples: sum divided by sample-count.
    ///
    /// See: <https://en.wikipedia.org/wiki/Arithmetic_mean>
    fn mean(&self) -> f64;

    /// Median of the samples: value separating the lower half of the samples from the higher half.
    /// Equal to `self.percentile(50.0)`.
    ///
    /// See: <https://en.wikipedia.org/wiki/Median>
    fn median(&self) -> f64;

    /// Variance of the samples: bias-corrected mean of the squares of the differences of each
    /// sample from the sample mean. Note that this calculates the _sample variance_ rather than the
    /// population variance, which is assumed to be unknown. It therefore corrects the `(n-1)/n`
    /// bias that would appear if we calculated a population variance, by dividing by `(n-1)` rather
    /// than `n`.
    ///
    /// See: <https://en.wikipedia.org/wiki/Variance>
    fn var(&self) -> f64;

    /// Standard deviation: the square root of the sample variance.
    ///
    /// Note: this is not a robust statistic for non-normal distributions. Prefer the
    /// `median_abs_dev` for unknown distributions.
    ///
    /// See: <https://en.wikipedia.org/wiki/Standard_deviation>
    fn std_dev(&self) -> f64;

    /// Standard deviation as a percent of the mean value. See `std_dev` and `mean`.
    ///
    /// Note: this is not a robust statistic for non-normal distributions. Prefer the
    /// `median_abs_dev_pct` for unknown distributions.
    fn std_dev_pct(&self) -> f64;

    /// Scaled median of the absolute deviations of each sample from the sample median. This is a
    /// robust (distribution-agnostic) estimator of sample variability. Use this in preference to
    /// `std_dev` if you cannot assume your sample is normally distributed. Note that this is scaled
    /// by the constant `1.4826` to allow its use as a consistent estimator for the standard
    /// deviation.
    ///
    /// See: <http://en.wikipedia.org/wiki/Median_absolute_deviation>
    fn median_abs_dev(&self) -> f64;

    /// Median absolute deviation as a percent of the median. See `median_abs_dev` and `median`.
    fn median_abs_dev_pct(&self) -> f64;

    /// Percentile: the value below which `pct` percent of the values in `self` fall. For example,
    /// percentile(95.0) will return the value `v` such that 95% of the samples `s` in `self`
    /// satisfy `s <= v`.
    ///
    /// Calculated by linear interpolation between closest ranks.
    ///
    /// See: <http://en.wikipedia.org/wiki/Percentile>
    fn percentile(&self, pct: f64) -> f64;

    /// Quartiles of the sample: three values that divide the sample into four equal groups, each
    /// with 1/4 of the data. The middle value is the median. See `median` and `percentile`. This
    /// function may calculate the 3 quartiles more efficiently than 3 calls to `percentile`, but
    /// is otherwise equivalent.
    ///
    /// See also: <https://en.wikipedia.org/wiki/Quartile>
    fn quartiles(&self) -> (f64, f64, f64);

    /// Inter-quartile range: the difference between the 25th percentile (1st quartile) and the 75th
    /// percentile (3rd quartile). See `quartiles`.
    ///
    /// See also: <https://en.wikipedia.org/wiki/Interquartile_range>
    fn iqr(&self) -> f64;
}

/// Extracted collection of all the summary statistics of a sample set.
#[derive(Clone, PartialEq, Copy)]
#[allow(missing_docs)]
pub struct Summary {
    pub sum: f64,
    pub min: f64,
    pub max: f64,
    pub mean: f64,
    pub median: f64,
    pub var: f64,
    pub std_dev: f64,
    pub std_dev_pct: f64,
    pub median_abs_dev: f64,
    pub median_abs_dev_pct: f64,
    pub quartiles: (f64, f64, f64),
    pub iqr: f64,
}

impl Summary {
    /// Construct a new summary of a sample set.
    pub fn new(samples: &[f64]) -> Summary {
        Summary {
            sum: samples.sum(),
            min: samples.min(),
            max: samples.max(),
            mean: samples.mean(),
            median: samples.median(),
            var: samples.var(),
            std_dev: samples.std_dev(),
            std_dev_pct: samples.std_dev_pct(),
            median_abs_dev: samples.median_abs_dev(),
            median_abs_dev_pct: samples.median_abs_dev_pct(),
            quartiles: samples.quartiles(),
            iqr: samples.iqr(),
        }
    }
}

impl Stats for [f64] {
    // FIXME #11059 handle NaN, inf and overflow
    fn sum(&self) -> f64 {
        let mut partials = vec![];

        for &x in self {
            let mut x = x;
            let mut j = 0;
            // This inner loop applies `hi`/`lo` summation to each
            // partial so that the list of partial sums remains exact.
            for i in 0..partials.len() {
                let mut y: f64 = partials[i];
                if x.abs() < y.abs() {
                    mem::swap(&mut x, &mut y);
                }
                // Rounded `x+y` is stored in `hi` with round-off stored in
                // `lo`. Together `hi+lo` are exactly equal to `x+y`.
                let hi = x + y;
                let lo = y - (hi - x);
                if lo != 0.0 {
                    partials[j] = lo;
                    j += 1;
                }
                x = hi;
            }
            if j >= partials.len() {
                partials.push(x);
            } else {
                partials[j] = x;
                partials.truncate(j + 1);
            }
        }
        let zero: f64 = 0.0;
        partials.iter().fold(zero, |p, q| p + *q)
    }

    fn min(&self) -> f64 {
        assert!(!self.is_empty());
        self.iter().fold(self[0], |p, q| p.min(*q))
    }

    fn max(&self) -> f64 {
        assert!(!self.is_empty());
        self.iter().fold(self[0], |p, q| p.max(*q))
    }

    fn mean(&self) -> f64 {
        assert!(!self.is_empty());
        self.sum() / (self.len() as f64)
    }

    fn median(&self) -> f64 {
        self.percentile(50 as f64)
    }

    fn var(&self) -> f64 {
        if self.len() < 2 {
            0.0
        } else {
            let mean = self.mean();
            let mut v: f64 = 0.0;
            for s in self {
                let x = *s - mean;
                v = v + x * x;
            }
            // N.B., this is _supposed to be_ len-1, not len. If you
            // change it back to len, you will be calculating a
            // population variance, not a sample variance.
            let denom = (self.len() - 1) as f64;
            v / denom
        }
    }

    fn std_dev(&self) -> f64 {
        self.var().sqrt()
    }

    fn std_dev_pct(&self) -> f64 {
        let hundred = 100 as f64;
        (self.std_dev() / self.mean()) * hundred
    }

    fn median_abs_dev(&self) -> f64 {
        let med = self.median();
        let abs_devs: Vec<f64> = self.iter().map(|&v| (med - v).abs()).collect();
        // This constant is derived by smarter statistics brains than me, but it is
        // consistent with how R and other packages treat the MAD.
        let number = 1.4826;
        abs_devs.median() * number
    }

    fn median_abs_dev_pct(&self) -> f64 {
        let hundred = 100 as f64;
        (self.median_abs_dev() / self.median()) * hundred
    }

    fn percentile(&self, pct: f64) -> f64 {
        let mut tmp = self.to_vec();
        local_sort(&mut tmp);
        percentile_of_sorted(&tmp, pct)
    }

    fn quartiles(&self) -> (f64, f64, f64) {
        let mut tmp = self.to_vec();
        local_sort(&mut tmp);
        let first = 25f64;
        let a = percentile_of_sorted(&tmp, first);
        let second = 50f64;
        let b = percentile_of_sorted(&tmp, second);
        let third = 75f64;
        let c = percentile_of_sorted(&tmp, third);
        (a, b, c)
    }

    fn iqr(&self) -> f64 {
        let (a, _, c) = self.quartiles();
        c - a
    }
}

// Helper function: extract a value representing the `pct` percentile of a sorted sample-set, using
// linear interpolation. If samples are not sorted, return nonsensical value.
fn percentile_of_sorted(sorted_samples: &[f64], pct: f64) -> f64 {
    assert!(!sorted_samples.is_empty());
    if sorted_samples.len() == 1 {
        return sorted_samples[0];
    }
    let zero: f64 = 0.0;
    assert!(zero <= pct);
    let hundred = 100f64;
    assert!(pct <= hundred);
    if pct == hundred {
        return sorted_samples[sorted_samples.len() - 1];
    }
    let length = (sorted_samples.len() - 1) as f64;
    let rank = (pct / hundred) * length;
    let lrank = rank.floor();
    let d = rank - lrank;
    let n = lrank as usize;
    let lo = sorted_samples[n];
    let hi = sorted_samples[n + 1];
    lo + (hi - lo) * d
}

/// Winsorize a set of samples, replacing values above the `100-pct` percentile
/// and below the `pct` percentile with those percentiles themselves. This is a
/// way of minimizing the effect of outliers, at the cost of biasing the sample.
/// It differs from trimming in that it does not change the number of samples,
/// just changes the values of those that are outliers.
///
/// See: <http://en.wikipedia.org/wiki/Winsorising>
pub fn winsorize(samples: &mut [f64], pct: f64) {
    let mut tmp = samples.to_vec();
    local_sort(&mut tmp);
    let lo = percentile_of_sorted(&tmp, pct);
    let hundred = 100 as f64;
    let hi = percentile_of_sorted(&tmp, hundred - pct);
    for samp in samples {
        if *samp > hi {
            *samp = hi
        } else if *samp < lo {
            *samp = lo
        }
    }
}