gam_solve/psis.rs
1//! Pareto-smoothed importance-sampling utilities.
2//!
3//! The implementation is intentionally self-contained: it estimates the
4//! generalized-Pareto tail shape `k` from the largest positive weights and
5//! replaces only that empirical tail by monotone GPD expected quantiles. The
6//! returned `k_hat` has the usual GPD tail interpretation: values near zero
7//! indicate light tails, `k > 0.5` indicates that the fitted tail has infinite
8//! variance, and larger values mark increasingly unstable upper tails. Consumers
9//! decide whether that tail is a draw-wise PSIS reliability diagnostic or another
10//! influence diagnostic based on what the supplied weights represent.
11//!
12//! The shape is recovered with the Zhang–Stephens (2009) empirical-Bayes
13//! profile estimator — the same GPD tail fit used by `loo`/ArviZ for draw-wise
14//! PSIS diagnostics.
15//! Crucially it is consistent across the entire `k ∈ (−∞, ∞)` range, including
16//! the dangerous `k ≥ 0.5` regime where the GPD variance is infinite and the
17//! older method-of-moments form `k = ½(1 − μ²/Var)` is structurally capped
18//! below `0.5` and so cannot fire a heavy-tail gate.
19
20#[derive(Debug, Clone)]
21pub struct PsisResult {
22 pub smoothed: Vec<f64>,
23 pub k_hat: f64,
24 pub tail_start: usize,
25 pub tail_count: usize,
26}
27
28pub const MIN_TAIL_COUNT: usize = 5;
29const MAX_TAIL_FRACTION: f64 = 0.2;
30
31/// Pareto-smooth a non-negative weight vector and report the fitted GPD tail
32/// shape. Non-tail observations are left bit-identical; only the largest tail
33/// observations are replaced by sorted GPD expected quantiles and then clipped
34/// to be non-decreasing in the original sorted order.
35pub fn pareto_smooth_weights(weights: &[f64]) -> Option<PsisResult> {
36 if weights.len() < MIN_TAIL_COUNT * 2 || weights.iter().any(|w| !w.is_finite() || *w < 0.0) {
37 return None;
38 }
39 let mut indexed: Vec<(usize, f64)> = weights.iter().copied().enumerate().collect();
40 indexed.sort_by(|a, b| a.1.total_cmp(&b.1));
41 let n = indexed.len();
42 let tail_count = ((n as f64).sqrt().ceil() as usize)
43 .max(MIN_TAIL_COUNT)
44 .min(((MAX_TAIL_FRACTION * n as f64).ceil() as usize).max(MIN_TAIL_COUNT))
45 .min(n - 1);
46 let tail_start = n - tail_count;
47 let threshold = indexed[tail_start - 1].1;
48 let excesses: Vec<f64> = indexed[tail_start..]
49 .iter()
50 .map(|(_, w)| (w - threshold).max(0.0))
51 .collect();
52 let (k_hat, sigma_hat) = fit_gpd_moments(&excesses)?;
53 let mut smoothed = weights.to_vec();
54 let mut previous = threshold;
55 for (rank, (original_idx, _)) in indexed[tail_start..].iter().enumerate() {
56 let p = (rank as f64 + 0.5) / tail_count as f64;
57 let q = threshold + gpd_quantile(p, k_hat, sigma_hat).max(0.0);
58 let monotone = q.max(previous);
59 smoothed[*original_idx] = monotone;
60 previous = monotone;
61 }
62 Some(PsisResult {
63 smoothed,
64 k_hat,
65 tail_start,
66 tail_count,
67 })
68}
69
70/// Fit a generalized-Pareto distribution `1 − (1 + k·x/σ)^(−1/k)` to positive
71/// excesses with the Zhang–Stephens (2009) empirical-Bayes profile estimator.
72///
73/// The GPD is reparameterised by the single scalar `b = −k/σ`. For a fixed
74/// `b` the shape has the closed-form profile MLE
75/// `k(b) = mean_i log(1 − b·xᵢ)`, and the profile log-likelihood (up to an
76/// additive constant) is `ℓ(b) = n·(log(−b/k(b)) − k(b) − 1)`. Zhang &
77/// Stephens place a diffuse grid of candidate `b` values around `1/x_max`
78/// (the boundary of the admissible region) plus a quartile-scaled spread, then
79/// average `b` under the softmax of `ℓ` — an empirical-Bayes posterior mean
80/// rather than a single maximiser, which is markedly more stable in small
81/// samples. The shape is read back off the posterior-mean `b` and shrunk
82/// toward `0.5` by a weak `N(0.5, …)`-style prior worth `PRIOR_K`
83/// pseudo-observations (negligible once `n ≫ PRIOR_K`).
84///
85/// Unlike the method-of-moments identity `k = ½(1 − μ²/Var)` — which is bounded
86/// above by `0.5` for every real sample and so can never report the heavy tails
87/// (`k > 0.7`) the diagnostic exists to flag — this estimator is consistent for
88/// `k` up to and beyond `1`, recovering the true shape to a couple percent from
89/// a large exact sample.
90///
91/// Returns `(k_hat, sigma_hat)`, or `None` when there are too few positive
92/// excesses or the data are degenerate (all equal / non-finite).
93pub fn fit_gpd_moments(excesses: &[f64]) -> Option<(f64, f64)> {
94 let mut x: Vec<f64> = excesses
95 .iter()
96 .copied()
97 .filter(|v| v.is_finite() && *v > 0.0)
98 .collect();
99 if x.len() < MIN_TAIL_COUNT {
100 return None;
101 }
102 x.sort_by(f64::total_cmp);
103 let n = x.len();
104 let nf = n as f64;
105
106 // Admissible region is `b < 1/x_max`; the smallest order statistic sets the
107 // quartile scale (Zhang & Stephens use the lower-quartile observation).
108 let x_max = x[n - 1];
109 let q_idx = ((nf / 4.0 + 0.5).floor() as usize)
110 .saturating_sub(1)
111 .min(n - 1);
112 let x_star = x[q_idx];
113 if !(x_max.is_finite() && x_max > 0.0 && x_star.is_finite() && x_star > 0.0) {
114 return None;
115 }
116
117 // Weakly-informative prior: `PRIOR_BS` controls grid spread, `PRIOR_K`
118 // pseudo-observations shrink the final shape toward 0.5 (Zhang–Stephens / loo).
119 const PRIOR_BS: f64 = 3.0;
120 const PRIOR_K: f64 = 10.0;
121 let m_est = 30 + (nf.sqrt() as usize);
122
123 // Profile log-likelihood `ℓ(b)` on the candidate grid (the `n·` factor and
124 // additive constants are retained so the softmax weights match loo/ArviZ).
125 let mut b_grid = Vec::with_capacity(m_est);
126 let mut len_scale = Vec::with_capacity(m_est);
127 let mut max_ls = f64::NEG_INFINITY;
128 for j in 1..=m_est {
129 let b =
130 (1.0 - (m_est as f64 / (j as f64 - 0.5)).sqrt()) / (PRIOR_BS * x_star) + 1.0 / x_max;
131 let k = profile_shape(b, &x);
132 let arg = k.map(|k| -(b / k));
133 let ls = if let (Some(k), Some(arg)) = (k, arg) {
134 if arg.is_finite() && arg > 0.0 {
135 nf * (arg.ln() - k - 1.0)
136 } else {
137 f64::NEG_INFINITY
138 }
139 } else {
140 f64::NEG_INFINITY
141 };
142 if ls > max_ls {
143 max_ls = ls;
144 }
145 b_grid.push(b);
146 len_scale.push(ls);
147 }
148 if !max_ls.is_finite() {
149 return None;
150 }
151
152 // Posterior mean of `b` under the (numerically stable) softmax of `ℓ`.
153 let mut weight_sum = 0.0;
154 let mut b_post = 0.0;
155 for (&b, &ls) in b_grid.iter().zip(len_scale.iter()) {
156 let w = if ls.is_finite() {
157 (ls - max_ls).exp()
158 } else {
159 0.0
160 };
161 weight_sum += w;
162 b_post += w * b;
163 }
164 if !(weight_sum.is_finite() && weight_sum > 0.0) {
165 return None;
166 }
167 b_post /= weight_sum;
168 if !b_post.is_finite() || b_post == 0.0 {
169 return None;
170 }
171
172 let k_raw = profile_shape(b_post, &x)?;
173 // The GPD scale is fixed by the fitted `b_post` and the *profile* shape that
174 // generated it: `σ = −k(b_post)/b_post`. This is positive by construction for
175 // a genuine tail, because the admissible profile shape `k_raw = mean log(1 −
176 // b·xᵢ)` always carries the opposite sign of `b_post` (every `1 − b·xᵢ ∈ (0,
177 // 1)` for `b_post > 0`, so `k_raw < 0`; symmetrically `k_raw > 0` for
178 // `b_post < 0`). Computing σ here, before the shape is shrunk, mirrors
179 // loo/ArviZ and keeps the sign relationship intact: a light tail (`k_raw < 0`)
180 // yields `σ > 0` rather than being rejected.
181 let sigma = -k_raw / b_post;
182 // Shrink only the *reported* shape toward 0.5 by `PRIOR_K` pseudo-observations.
183 // This affects the diagnostic value alone, never the scale's validity, so the
184 // weak toward-0.5 prior can no longer flip `k`'s sign relative to `b_post` and
185 // spuriously reject a perfectly valid light-tail fit.
186 let k = (nf * k_raw + PRIOR_K * 0.5) / (nf + PRIOR_K);
187 if !(k.is_finite() && sigma.is_finite() && sigma > 0.0) {
188 return None;
189 }
190 Some((k, sigma))
191}
192
193/// Closed-form profile shape `k(b) = mean_i log(1 − b·xᵢ)` for fixed `b`.
194///
195/// A candidate is admissible only when every `1 - b*x_i` is finite and strictly
196/// positive. Boundary or out-of-domain candidates are rejected before any log is
197/// evaluated, so they cannot leak NaN/Inf through the profile likelihood.
198#[inline]
199fn profile_shape(b: f64, x: &[f64]) -> Option<f64> {
200 if !b.is_finite() || x.is_empty() {
201 return None;
202 }
203 let mut acc = 0.0_f64;
204 for &xi in x {
205 let arg = 1.0 - b * xi;
206 if !(arg.is_finite() && arg > 0.0) {
207 return None;
208 }
209 acc += arg.ln();
210 }
211 Some(acc / x.len() as f64)
212}
213
214#[inline]
215fn gpd_quantile(p: f64, k: f64, sigma: f64) -> f64 {
216 let survival = (1.0 - p).clamp(1e-12, 1.0);
217 if k.abs() < 1e-8 {
218 -sigma * survival.ln()
219 } else {
220 sigma * (survival.powf(-k) - 1.0) / k
221 }
222}
223
224#[cfg(test)]
225mod tests {
226 use super::*;
227
228 fn gpd_sample(u: f64, k: f64, sigma: f64) -> f64 {
229 if k.abs() < 1e-12 {
230 -sigma * (1.0 - u).ln()
231 } else {
232 sigma * ((1.0 - u).powf(-k) - 1.0) / k
233 }
234 }
235
236 #[test]
237 fn psis_k_hat_recovers_known_generalized_pareto_tail() {
238 let k_true = 0.35_f64;
239 let sigma = 1.7_f64;
240 let mut xs = Vec::new();
241 for i in 1..=10_000 {
242 let u = (i as f64 - 0.5) / 10_000.0;
243 xs.push(gpd_sample(u, k_true, sigma));
244 }
245 let (k_hat, sigma_hat) = fit_gpd_moments(&xs).expect("GPD fit should succeed");
246 assert!(
247 (k_hat - k_true).abs() < 0.03,
248 "k_hat={k_hat}, k_true={k_true}"
249 );
250 assert!(
251 (sigma_hat - sigma).abs() < 0.08,
252 "sigma_hat={sigma_hat}, sigma={sigma}"
253 );
254 }
255
256 #[test]
257 fn pareto_smoothing_preserves_nontail_and_reports_heavy_tail() {
258 // A flat block of baseline weights plus a genuine GPD(k=0.7) tail. The
259 // tail is heavy enough to sit in the infinite-variance regime, so a
260 // consistent estimator must report `k_hat > 0.5` for it.
261 let mut w = vec![1.0; 200];
262 for i in 1..=120 {
263 let u = (i as f64 - 0.5) / 120.0;
264 w.push(1.0 + gpd_sample(u, 0.7, 0.5));
265 }
266 let out = pareto_smooth_weights(&w).expect("PSIS should fit a positive tail");
267 assert_eq!(out.smoothed[0], 1.0);
268 assert!(
269 out.k_hat > 0.5,
270 "genuine GPD(k=0.7) tail should be flagged as heavy (infinite variance); got k_hat={}",
271 out.k_hat
272 );
273 }
274
275 /// Regression for #585 from the *shape-recovery* angle: the moment estimator
276 /// `k = ½(1 − μ²/Var)` is structurally `≤ 0.5`, so heavy tails collapsed
277 /// onto an indistinguishable ~0.5. The Zhang–Stephens estimator must instead
278 /// recover the true shape across the dangerous range — and the fitted shapes
279 /// must stay *strictly ordered*, since a diagnostic that cannot separate
280 /// `k=0.7` from `k=1.0` is useless.
281 #[test]
282 fn psis_k_hat_tracks_and_orders_heavy_tails() {
283 let recover = |k_true: f64| -> f64 {
284 let xs: Vec<f64> = (1..=100_000)
285 .map(|i| gpd_sample((i as f64 - 0.5) / 100_000.0, k_true, 1.0))
286 .collect();
287 fit_gpd_moments(&xs).expect("GPD fit should succeed").0
288 };
289 let mut last = f64::NEG_INFINITY;
290 for &k_true in &[0.5_f64, 0.7, 0.85, 1.0, 1.2] {
291 let k_hat = recover(k_true);
292 assert!(
293 (k_hat - k_true).abs() < 0.05,
294 "k_true={k_true}: fitted k_hat={k_hat} not within 0.05"
295 );
296 assert!(
297 k_hat > last,
298 "k_hat must increase with the true shape: {k_hat} !> {last}"
299 );
300 last = k_hat;
301 }
302 }
303
304 /// The estimator must degrade gracefully: too few positive excesses and
305 /// fully degenerate (all-equal) inputs return `None` rather than NaN/`0.5`.
306 #[test]
307 fn psis_gpd_fit_handles_degenerate_inputs() {
308 assert!(
309 fit_gpd_moments(&[1.0, 2.0, 3.0]).is_none(),
310 "fewer than MIN_TAIL_COUNT"
311 );
312 assert!(
313 fit_gpd_moments(&[0.0, -1.0, f64::NAN, 0.0]).is_none(),
314 "no positive finite excesses"
315 );
316 // All-equal positive excesses are degenerate (x_max == x_star). The fit
317 // must not panic or emit NaN, and — most importantly — must never report
318 // a spuriously *heavy* tail; a near-constant block is the lightest tail
319 // there is, so any returned shape must sit well below the 0.5 gate.
320 if let Some((k_hat, sigma_hat)) = fit_gpd_moments(&[2.0; 50]) {
321 assert!(k_hat.is_finite() && sigma_hat.is_finite() && sigma_hat > 0.0);
322 assert!(
323 k_hat < 0.5,
324 "degenerate equal excesses must not be flagged heavy; got k_hat={k_hat}"
325 );
326 }
327 }
328
329 #[test]
330 fn psis_profile_shape_rejects_inadmissible_candidates() {
331 let x = [0.25, 1.0, 2.0];
332 assert!(
333 profile_shape(0.49, &x).is_some(),
334 "b below 1/x_max is admissible for all excesses"
335 );
336 assert!(
337 profile_shape(0.5, &x).is_none(),
338 "b at 1/x_max puts the largest excess on the log boundary"
339 );
340 assert!(
341 profile_shape(0.75, &x).is_none(),
342 "b above 1/x_max makes at least one log argument negative"
343 );
344 }
345}