solow-nonparametric 0.2.0

Nonparametric smoothers (lowess, KDE) for the Solow statistical library
Documentation
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
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
//! Multivariate kernel regression and multivariate kernel density estimation
//! with a product Gaussian kernel and a fixed, user-supplied bandwidth.
//!
//! Two estimators are provided, both mirroring the canonical Python reference
//! (`nonparametric.kernel_regression.KernelReg` and
//! `nonparametric.kernel_density.KDEMultivariate`) for the all-continuous
//! (`var_type = "c…c"`) case with an explicit bandwidth:
//!
//! * [`KernelReg`] — Nadaraya–Watson kernel regression of a scalar response on
//!   `d` continuous predictors, with a local-constant ([`RegType::LocalConstant`])
//!   or local-linear ([`RegType::LocalLinear`]) fit.
//! * [`KdeMultivariate`] — a product-Gaussian multivariate kernel density
//!   estimator for `d` continuous variables.
//!
//! # Kernel
//!
//! The continuous kernel is the standard Gaussian
//! `k(z) = exp(−z²/2) / √(2π)`. The generalized product kernel for an
//! observation `Xᵢ ∈ ℝ^d` evaluated at a point `x` with bandwidth vector
//! `h ∈ ℝ^d` is
//!
//! ```text
//! K_h(Xᵢ, x) = (1 / ∏_s h_s) · ∏_{s=1}^{d} k((X_{is} − x_s) / h_s),
//! ```
//!
//! exactly the (non-normalized) generalized product kernel estimator used by
//! the reference's `gpke`. The leading `1/∏ h_s` factor cancels in the
//! Nadaraya–Watson ratio and in the local-linear normal equations, so the
//! regression mean is insensitive to it; it matters for the density.

use ndarray::{Array1, Array2};
use solow_core::error::{Error, Result};
use solow_linalg::pinv;
use std::f64::consts::PI;

/// Local polynomial order for [`KernelReg`].
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum RegType {
    /// Local-constant (Nadaraya–Watson) fit, the reference's `reg_type='lc'`.
    LocalConstant,
    /// Local-linear fit, the reference's `reg_type='ll'`.
    LocalLinear,
}

/// One-dimensional Gaussian kernel `k(z) = exp(−z²/2)/√(2π)`.
#[inline]
fn gaussian(z: f64) -> f64 {
    (-0.5 * z * z).exp() / (2.0 * PI).sqrt()
}

/// Per-observation generalized product kernel values `K_h(Xᵢ, x)` for every
/// training row `Xᵢ` against a single evaluation point `x`.
///
/// Returns a length-`n` vector whose `i`-th entry is
/// `(1/∏ h_s) · ∏_s k((X_{is} − x_s)/h_s)`, matching `gpke(..., tosum=False)`
/// for an all-continuous Gaussian kernel.
fn gpke_rows(bw: &[f64], data: &Array2<f64>, point: &[f64]) -> Array1<f64> {
    let (n, d) = data.dim();
    let mut prod_h = 1.0;
    for &h in bw.iter() {
        prod_h *= h;
    }
    let mut out = Array1::<f64>::zeros(n);
    for i in 0..n {
        let mut prod = 1.0;
        for s in 0..d {
            prod *= gaussian((data[[i, s]] - point[s]) / bw[s]);
        }
        out[i] = prod / prod_h;
    }
    out
}

/// Validate a feature matrix / bandwidth pair shared by both estimators.
fn validate(data: &Array2<f64>, bw: &[f64]) -> Result<()> {
    let (n, d) = data.dim();
    if n == 0 {
        return Err(Error::Value(
            "kernel estimator requires at least one observation".into(),
        ));
    }
    if d == 0 {
        return Err(Error::Value(
            "kernel estimator requires at least one variable".into(),
        ));
    }
    if bw.len() != d {
        return Err(Error::Shape(format!(
            "bandwidth length {} does not match number of variables {d}",
            bw.len()
        )));
    }
    for &h in bw.iter() {
        if !h.is_finite() || h <= 0.0 {
            return Err(Error::Value(
                "every bandwidth must be positive and finite".into(),
            ));
        }
    }
    if data.iter().any(|v| !v.is_finite()) {
        return Err(Error::Value("kernel estimator data must be finite".into()));
    }
    Ok(())
}

/// Multivariate Nadaraya–Watson kernel regression with a product Gaussian
/// kernel and a fixed bandwidth.
///
/// The model is `y = g(x) + ε` with `x ∈ ℝ^d`. [`KernelReg::fit`] evaluates the
/// estimated conditional mean `ĝ` on a grid of query points.
///
/// # Example
/// ```
/// use ndarray::{array, Array2};
/// use solow_nonparametric::{KernelReg, RegType};
///
/// let exog = Array2::from_shape_vec((5, 1), vec![0.0, 1.0, 2.0, 3.0, 4.0]).unwrap();
/// let endog = array![0.0, 1.0, 4.0, 9.0, 16.0];
/// let kr = KernelReg::new(endog, exog, vec![0.8], RegType::LocalLinear).unwrap();
/// let grid = Array2::from_shape_vec((2, 1), vec![1.5, 2.5]).unwrap();
/// let mean = kr.fit(&grid).unwrap();
/// assert_eq!(mean.len(), 2);
/// ```
#[derive(Debug, Clone)]
pub struct KernelReg {
    endog: Array1<f64>,
    exog: Array2<f64>,
    bw: Vec<f64>,
    reg_type: RegType,
}

impl KernelReg {
    /// Construct a kernel regression for response `endog` on predictors `exog`
    /// (`n × d`) with bandwidth vector `bw` (length `d`) and the given fit order.
    ///
    /// # Errors
    /// Returns an error on empty data, an `endog`/`exog` row-count mismatch, a
    /// bandwidth-length mismatch, a non-positive/non-finite bandwidth, or
    /// non-finite inputs.
    pub fn new(
        endog: Array1<f64>,
        exog: Array2<f64>,
        bw: Vec<f64>,
        reg_type: RegType,
    ) -> Result<Self> {
        validate(&exog, &bw)?;
        if endog.len() != exog.dim().0 {
            return Err(Error::Shape(format!(
                "endog length {} does not match exog rows {}",
                endog.len(),
                exog.dim().0
            )));
        }
        if endog.iter().any(|v| !v.is_finite()) {
            return Err(Error::Value("endog must be finite".into()));
        }
        Ok(KernelReg {
            endog,
            exog,
            bw,
            reg_type,
        })
    }

    /// Number of predictor variables `d`.
    pub fn k_vars(&self) -> usize {
        self.exog.dim().1
    }

    /// The bandwidth vector in use.
    pub fn bw(&self) -> &[f64] {
        &self.bw
    }

    /// Evaluate the estimated conditional mean at each row of `data_predict`
    /// (`m × d`).
    ///
    /// Returns the length-`m` vector of fitted means, matching the first return
    /// value of the reference's `KernelReg.fit`.
    ///
    /// # Errors
    /// Returns an error if `data_predict` does not have `d` columns.
    pub fn fit(&self, data_predict: &Array2<f64>) -> Result<Array1<f64>> {
        let (m, d) = data_predict.dim();
        if d != self.k_vars() {
            return Err(Error::Shape(format!(
                "data_predict has {d} columns, expected {}",
                self.k_vars()
            )));
        }
        let mut mean = Array1::<f64>::zeros(m);
        let mut row = vec![0.0; d];
        for i in 0..m {
            for s in 0..d {
                row[s] = data_predict[[i, s]];
            }
            mean[i] = match self.reg_type {
                RegType::LocalConstant => self.est_loc_constant(&row),
                RegType::LocalLinear => self.est_loc_linear(&row)?,
            };
        }
        Ok(mean)
    }

    /// Local-constant (Nadaraya–Watson) estimate at one point.
    ///
    /// `ĝ(x) = Σᵢ K_h(Xᵢ,x) yᵢ / Σᵢ K_h(Xᵢ,x)`. The `1/∏ h_s` factor cancels.
    fn est_loc_constant(&self, point: &[f64]) -> f64 {
        let ker = gpke_rows(&self.bw, &self.exog, point);
        let mut numer = 0.0;
        let mut denom = 0.0;
        for (k, &w) in ker.iter().enumerate() {
            numer += w * self.endog[k];
            denom += w;
        }
        numer / denom
    }

    /// Local-linear estimate at one point.
    ///
    /// Builds the weighted least-squares normal-equation system on p. 81 of
    /// Li & Racine and solves it with the Moore–Penrose pseudoinverse, exactly
    /// as the reference's `_est_loc_linear` (which divides the kernel weights by
    /// `nobs`; that scale cancels and is omitted here). Returns the intercept
    /// component, i.e. the fitted mean at `point`.
    fn est_loc_linear(&self, point: &[f64]) -> Result<f64> {
        let (n, d) = self.exog.dim();
        let ker = gpke_rows(&self.bw, &self.exog, point);

        // M is (d+1)×(d+1); V is (d+1).
        let dim = d + 1;
        let mut m = Array2::<f64>::zeros((dim, dim));
        let mut v = Array1::<f64>::zeros(dim);

        // Centered predictors Δᵢ = Xᵢ − x.
        // M[0,0] = Σ kᵢ
        // M[0,1+s] = M[1+s,0] = Σ kᵢ Δ_{is}
        // M[1+s,1+t] = Σ kᵢ Δ_{is} Δ_{it}
        // V[0] = Σ kᵢ yᵢ
        // V[1+s] = Σ kᵢ Δ_{is} yᵢ
        let mut sum_k = 0.0;
        let mut m12 = vec![0.0; d];
        let mut m22 = vec![0.0; d * d];
        let mut v0 = 0.0;
        let mut v_rest = vec![0.0; d];
        for i in 0..n {
            let ki = ker[i];
            sum_k += ki;
            let yi = self.endog[i];
            v0 += ki * yi;
            for s in 0..d {
                let ds = self.exog[[i, s]] - point[s];
                let kds = ki * ds;
                m12[s] += kds;
                v_rest[s] += kds * yi;
                for t in 0..d {
                    let dt = self.exog[[i, t]] - point[t];
                    m22[s * d + t] += kds * dt;
                }
            }
        }
        m[[0, 0]] = sum_k;
        v[0] = v0;
        for s in 0..d {
            m[[0, s + 1]] = m12[s];
            m[[s + 1, 0]] = m12[s];
            v[s + 1] = v_rest[s];
            for t in 0..d {
                m[[s + 1, t + 1]] = m22[s * d + t];
            }
        }

        let (mp, _s) = pinv(&m)?;
        let coef = mp.dot(&v);
        Ok(coef[0])
    }
}

/// Product-Gaussian multivariate kernel density estimator with a fixed
/// bandwidth, mirroring the reference's `KDEMultivariate` for all-continuous
/// variables.
///
/// The density at a point `x ∈ ℝ^d` is
///
/// ```text
/// f̂(x) = (1/n) · Σᵢ K_h(Xᵢ, x)
///      = 1 / (n · ∏_s h_s) · Σᵢ ∏_s k((X_{is} − x_s)/h_s).
/// ```
///
/// # Example
/// ```
/// use ndarray::Array2;
/// use solow_nonparametric::KdeMultivariate;
///
/// let data = Array2::from_shape_vec((4, 2), vec![
///     0.0, 0.0, 1.0, 0.5, -0.5, 1.0, 0.3, -0.2,
/// ]).unwrap();
/// let kde = KdeMultivariate::new(data, vec![0.5, 0.5]).unwrap();
/// let pts = Array2::from_shape_vec((1, 2), vec![0.0, 0.0]).unwrap();
/// let dens = kde.pdf(&pts).unwrap();
/// assert!(dens[0] > 0.0);
/// ```
#[derive(Debug, Clone)]
pub struct KdeMultivariate {
    data: Array2<f64>,
    bw: Vec<f64>,
}

impl KdeMultivariate {
    /// Construct an estimator over `data` (`n × d`) with bandwidth vector `bw`
    /// of length `d`.
    ///
    /// # Errors
    /// Returns an error on empty data, a bandwidth-length mismatch, a
    /// non-positive/non-finite bandwidth, or non-finite data.
    pub fn new(data: Array2<f64>, bw: Vec<f64>) -> Result<Self> {
        validate(&data, &bw)?;
        Ok(KdeMultivariate { data, bw })
    }

    /// Number of variables `d`.
    pub fn k_vars(&self) -> usize {
        self.data.dim().1
    }

    /// The bandwidth vector in use.
    pub fn bw(&self) -> &[f64] {
        &self.bw
    }

    /// Evaluate the density at each row of `data_predict` (`m × d`).
    ///
    /// # Errors
    /// Returns an error if `data_predict` does not have `d` columns.
    pub fn pdf(&self, data_predict: &Array2<f64>) -> Result<Array1<f64>> {
        let (m, d) = data_predict.dim();
        if d != self.k_vars() {
            return Err(Error::Shape(format!(
                "data_predict has {d} columns, expected {}",
                self.k_vars()
            )));
        }
        let nobs = self.data.dim().0 as f64;
        let mut out = Array1::<f64>::zeros(m);
        let mut row = vec![0.0; d];
        for i in 0..m {
            for s in 0..d {
                row[s] = data_predict[[i, s]];
            }
            let ker = gpke_rows(&self.bw, &self.data, &row);
            out[i] = ker.sum() / nobs;
        }
        Ok(out)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use ndarray::array;

    fn col(v: Vec<f64>) -> Array2<f64> {
        let n = v.len();
        Array2::from_shape_vec((n, 1), v).unwrap()
    }

    #[test]
    fn loc_constant_matches_manual_nw() {
        let exog = col(vec![0.0, 1.0, 2.0, 3.0]);
        let endog = array![1.0, 2.0, 1.0, 3.0];
        let h = 0.7;
        let kr =
            KernelReg::new(endog.clone(), exog.clone(), vec![h], RegType::LocalConstant).unwrap();
        let grid = col(vec![1.5]);
        let got = kr.fit(&grid).unwrap()[0];
        // Manual Nadaraya-Watson.
        let mut num = 0.0;
        let mut den = 0.0;
        for i in 0..4 {
            let w = gaussian((exog[[i, 0]] - 1.5) / h);
            num += w * endog[i];
            den += w;
        }
        assert!((got - num / den).abs() < 1e-12);
    }

    #[test]
    fn loc_linear_reproduces_exact_line() {
        // On exactly-linear data the local-linear estimator is exact.
        let exog = col(vec![0.0, 1.0, 2.0, 3.0, 4.0, 5.0]);
        let endog: Array1<f64> = exog.column(0).mapv(|x| 2.0 * x - 1.0);
        let kr = KernelReg::new(endog, exog, vec![0.6], RegType::LocalLinear).unwrap();
        let grid = col(vec![1.3, 2.7, 4.1]);
        let got = kr.fit(&grid).unwrap();
        for (k, &t) in [1.3, 2.7, 4.1].iter().enumerate() {
            assert!(
                (got[k] - (2.0 * t - 1.0)).abs() < 1e-8,
                "k={k} got={}",
                got[k]
            );
        }
    }

    #[test]
    fn pdf_integrates_to_one_2d() {
        // Coarse trapezoid over a grid should integrate near 1.
        let data = Array2::from_shape_vec(
            (5, 2),
            vec![0.0, 0.0, 1.0, 0.5, -0.5, 1.0, 0.3, -0.2, -0.8, -0.6],
        )
        .unwrap();
        let kde = KdeMultivariate::new(data, vec![0.6, 0.6]).unwrap();
        let lo = -4.0;
        let hi = 4.0;
        let ng = 80;
        let step = (hi - lo) / (ng as f64 - 1.0);
        let axis: Vec<f64> = (0..ng).map(|i| lo + step * i as f64).collect();
        let mut grid_pts = Vec::with_capacity(ng * ng * 2);
        for &a in &axis {
            for &b in &axis {
                grid_pts.push(a);
                grid_pts.push(b);
            }
        }
        let pts = Array2::from_shape_vec((ng * ng, 2), grid_pts).unwrap();
        let dens = kde.pdf(&pts).unwrap();
        let area: f64 = dens.sum() * step * step;
        assert!((area - 1.0).abs() < 1e-2, "area = {area}");
    }

    #[test]
    fn bad_bandwidth_errors() {
        let exog = col(vec![0.0, 1.0, 2.0]);
        let endog = array![0.0, 1.0, 2.0];
        assert!(
            KernelReg::new(endog.clone(), exog.clone(), vec![0.0], RegType::LocalLinear).is_err()
        );
        assert!(
            KernelReg::new(endog, exog, vec![0.5, 0.5], RegType::LocalLinear).is_err(),
            "bandwidth length mismatch should error"
        );
    }
}