rdrobust-core 0.1.1

Pure-Rust numerical core for robust local-polynomial regression-discontinuity estimation (rdrobust).
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
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
//! Shared variance machinery: residuals (nn / hc0 / hc1) and the sandwich meat,
//! used by both the bandwidth engine (`bw.rs`) and the main estimator
//! (`rdrobust.rs`). Ports of `rdrobust_res` (functions.R:146-210) and the
//! non-cluster branch of `rdrobust_vce` (functions.R:380-392).

use nalgebra::{DMatrix, DVector};

use crate::Vce;

/// Exact integer factorial as f64 (R `factorial(nu)`; factorial(0)=1).
pub(crate) fn factorial(n: usize) -> f64 {
    (1..=n).map(|k| k as f64).product()
}

/// Run-length structure of a SORTED running variable (R `dups`/`dupsid`,
/// rdrobust.R:207-214). `dups[i]` = tie-block size of `x[i]`; `dupsid[i]` = its
/// 1-based rank within that block.
pub(crate) fn compute_dups_dupsid(x: &[f64]) -> (Vec<usize>, Vec<usize>) {
    let n = x.len();
    let mut dups = vec![0usize; n];
    let mut dupsid = vec![0usize; n];
    let mut i = 0;
    while i < n {
        let mut j = i + 1;
        while j < n && x[j] == x[i] {
            j += 1;
        }
        let len = j - i;
        for k in i..j {
            dups[k] = len;
            dupsid[k] = k - i + 1;
        }
        i = j;
    }
    (dups, dupsid)
}

/// Nearest-neighbor leave-one-out residuals for every column of `d` (R
/// `rdrobust_res` nn branch, functions.R:153-187). `x` is the windowed SORTED
/// running variable; `dups`/`dupsid` are the window-subset of the full-side
/// run-length structure (R subsets `dups[ind]`, it does not recompute).
pub(crate) fn nn_residuals(
    x: &[f64],
    d: &DMatrix<f64>,
    dups: &[usize],
    dupsid: &[usize],
    matches: usize,
) -> DMatrix<f64> {
    let n = x.len();
    let ncols = d.ncols();
    let nn_tol_eps = f64::EPSILON.sqrt();
    // An empty window (every point outside the kernel support, e.g. a pilot
    // bandwidth of 0 when one side collapses onto the cutoff) has no residuals.
    // `n - 1` would underflow; the loop below never runs for n == 0 anyway.
    let target = matches.min(n.saturating_sub(1)) as isize;
    let mut res = DMatrix::zeros(n, ncols);

    for pos in 1..=n {
        let p = pos as isize;
        let mut rpos = dups[pos - 1] as isize - dupsid[pos - 1] as isize;
        let mut lpos = dupsid[pos - 1] as isize - 1;
        while lpos + rpos < target {
            if p - lpos - 1 <= 0 {
                rpos += dups[(p + rpos) as usize] as isize;
            } else if p + rpos + 1 > n as isize {
                lpos += dups[(p - lpos - 2) as usize] as isize;
            } else {
                let dleft = x[pos - 1] - x[(p - lpos - 2) as usize];
                let dright = x[(p + rpos) as usize] - x[pos - 1];
                let nn_tol = dleft.max(dright) * nn_tol_eps;
                if dleft - dright > nn_tol {
                    rpos += dups[(p + rpos) as usize] as isize;
                } else if dright - dleft > nn_tol {
                    lpos += dups[(p - lpos - 2) as usize] as isize;
                } else {
                    rpos += dups[(p + rpos) as usize] as isize;
                    lpos += dups[(p - lpos - 2) as usize] as isize;
                }
            }
        }
        let lo0 = (p - lpos - 1) as usize;
        let hi0 = (p + rpos - 1) as usize;
        let ji = (lpos + rpos) as f64;
        for col in 0..ncols {
            let mut sum = 0.0;
            for k in lo0..=hi0 {
                sum += d[(k, col)];
            }
            let d_j = sum - d[(pos - 1, col)];
            res[(pos - 1, col)] = (ji / (ji + 1.0)).sqrt() * (d[(pos - 1, col)] - d_j / ji);
        }
    }
    res
}

/// Residual matrix (n_w x ncols) for the given variance estimator.
/// hc0: `D - R beta`; hc1: scaled by `sqrt(n/(n-dof))`, dof = `r_mat.ncols()`
/// (the poly count R passes as `d`); nn: leave-one-out neighbor residuals.
/// `has_cluster` suppresses the hc1 dof scaling (R functions.R:202 — clustering
/// applies its own df factor in the meat, so doubling is avoided); the cluster
/// variants (crv2/crv3) return the RAW residual (functions.R:189-199), the cluster-level
/// leverage adjustment being deferred to the meat.
#[allow(clippy::too_many_arguments)]
pub(crate) fn residuals(
    ex: &[f64],
    r_mat: &DMatrix<f64>,
    inv_g: &DMatrix<f64>,
    rx: &DMatrix<f64>,
    beta: &DMatrix<f64>,
    d: &DMatrix<f64>,
    n_w: usize,
    vce: Vce,
    dups_w: &[usize],
    dupsid_w: &[usize],
    nnmatch: usize,
    has_cluster: bool,
) -> DMatrix<f64> {
    // Clustering remaps every vce to a raw-residual path (rdrobust.R:318-344):
    // nn/hc0/hc1/cr1 -> hc1 (w=1 -> raw), hc2/hc3 -> crv2/crv3 (raw). The cluster
    // leverage adjustment lives in the meat, not the residual.
    if has_cluster {
        return d - r_mat * beta;
    }
    // crv2/crv3 use raw residuals; the hat-matrix adjustment is cluster-wise.
    if vce.cluster_order().is_some() {
        return d - r_mat * beta;
    }
    match vce.hc_order() {
        Some(0) => d - r_mat * beta,
        Some(1) => {
            let dof = r_mat.ncols() as f64;
            let w = (n_w as f64 / (n_w as f64 - dof)).sqrt();
            (d - r_mat * beta) * w
        }
        // hc2/hc3 (R functions.R:203-204): w = sqrt(1/max(1-hii,1e-8)) / (1/...).
        // hii = rowSums((R invG) * (R*W)) -- the hat-matrix diagonal.
        Some(2) | Some(3) => {
            let hc3 = vce.hc_order() == Some(3);
            let rig = r_mat * inv_g; // n_w x (o+1)
            let mut res = d - r_mat * beta;
            let ncols = res.ncols();
            for k in 0..n_w {
                let hii: f64 = (0..r_mat.ncols()).map(|j| rig[(k, j)] * rx[(k, j)]).sum();
                let denom = (1.0 - hii).max(1e-8);
                let w = if hc3 {
                    1.0 / denom
                } else {
                    (1.0 / denom).sqrt()
                };
                for col in 0..ncols {
                    res[(k, col)] *= w;
                }
            }
            res
        }
        // Only NearestNeighbor reaches here: Hc0-3 have hc_order Some(_), and the
        // Cr1-3 variants are handled by the earlier cluster_order check.
        _ => match vce {
            Vce::NearestNeighbor => nn_residuals(ex, d, dups_w, dupsid_w, nnmatch),
            _ => unreachable!("non-NearestNeighbor vce {vce:?} handled above"),
        },
    }
}

/// Covariate FWL partialling inputs for one fit (R functions.R:251-256 /
/// rdrobust.R:659-667). `rx` = R*W (row-scaled design), `ew` = its row weights,
/// `d` = `[Y, T?, Z..]`. Returns `(ZWZ, ZWY)` = (dz x dz, dz x (1+dt)); the main
/// estimator SUMS these across sides before inverting, the bw engine does not.
pub(crate) fn partialling(
    rx: &DMatrix<f64>,
    inv_g: &DMatrix<f64>,
    d: &DMatrix<f64>,
    ew: &[f64],
    dt: usize,
    dz: usize,
) -> (DMatrix<f64>, DMatrix<f64>) {
    let cz = 1 + dt;
    let mut ezw = d.columns(cz, dz).into_owned();
    for i in 0..d.nrows() {
        for j in 0..dz {
            ezw[(i, j)] *= ew[i];
        }
    }
    let u = rx.transpose() * d; // crossprod(R*W, D)
    let zwd = ezw.transpose() * d; // crossprod(eZ*W, D)
    let u_cz = u.columns(cz, dz).into_owned();
    let uigu = u_cz.transpose() * (inv_g * &u);
    let zwz = zwd.columns(cz, dz) - uigu.columns(cz, dz);
    let zwy = zwd.columns(0, 1 + dt) - uigu.columns(0, 1 + dt);
    (zwz, zwy)
}

/// `gamma = inv(ZWZ) ZWY`: chol2inv when SPD, ginv when `covs_drop_coll`.
pub(crate) fn gamma_from(
    zwz: &DMatrix<f64>,
    zwy: &DMatrix<f64>,
    covs_drop_coll: bool,
) -> DMatrix<f64> {
    if covs_drop_coll {
        crate::linalg::ginv(zwz) * zwy
    } else {
        match zwz.clone().cholesky() {
            Some(ch) => ch.inverse() * zwy,
            None => crate::linalg::ginv(zwz) * zwy,
        }
    }
}

/// Sandwich `invG * M * invG` (full matrix), meat `M = crossprod(c(res %*% s) * rx)`
/// (R rdrobust_vce non-cluster branch; the d>0 form reduces to d==0 when s=[1]).
pub(crate) fn v_matrix(
    inv_g: &DMatrix<f64>,
    rx: &DMatrix<f64>,
    res: &DMatrix<f64>,
    s: &DVector<f64>,
) -> DMatrix<f64> {
    let r_comb = res * s; // n-vector
    let mut rsc = rx.clone();
    let nrow = rx.nrows();
    let ncol = rx.ncols();
    for i in 0..nrow {
        let ri = r_comb[i];
        for j in 0..ncol {
            rsc[(i, j)] *= ri;
        }
    }
    let m = rsc.transpose() * &rsc;
    inv_g * (m * inv_g)
}

/// Which cluster-robust meat to build (mirrors R's `vce` -> branch selection).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum ClusterKind {
    /// Liang-Zeger CR1 with the rdrobust small-sample df factor.
    Cr1,
    /// Bell-McCaffrey CRV2 (eigendecomposition half-inverse).
    Crv2,
    /// Pustejovsky-Tipton CRV3 (Sherman-Morrison-Woodbury).
    Crv3,
}

/// Map a `Vce` to its cluster meat (only meaningful when a cluster vector is
/// present). R remaps the user's vce when a cluster is supplied (rdrobust.R:
/// 318-344): `nn`/`hc0`/`hc1`/`cr1` -> CR1, `hc2`/`cr2` -> CRV2, `hc3`/`cr3` ->
/// CRV3. The residual is RAW in every clustered branch (hc1 is suppressed,
/// crv2/crv3 use raw) -- see `residuals`.
pub(crate) fn cluster_kind(vce: Vce) -> ClusterKind {
    match (vce.hc_order(), vce.cluster_order()) {
        (Some(2), _) | (_, Some(2)) => ClusterKind::Crv2,
        (Some(3), _) | (_, Some(3)) => ClusterKind::Crv3,
        _ => ClusterKind::Cr1,
    }
}

/// Group row indices by cluster id (R `.rdrobust_cluster_idx`, functions.R:96-108,
/// numeric path only). Returns groups in ascending-cluster-value order; each is
/// the original 0-based row indices in that cluster, in stable (sorted) order.
/// Group order and intra-group order do not affect the meat (sum of rank-1 terms).
pub(crate) fn cluster_idx(c: &[f64]) -> Vec<Vec<usize>> {
    let n = c.len();
    if n == 0 {
        return Vec::new();
    }
    let mut ord: Vec<usize> = (0..n).collect();
    ord.sort_by(|&a, &b| c[a].total_cmp(&c[b]));
    let mut groups: Vec<Vec<usize>> = Vec::new();
    let mut start = 0;
    while start < n {
        let v = c[ord[start]];
        let mut end = start + 1;
        while end < n && c[ord[end]] == v {
            end += 1;
        }
        groups.push(ord[start..end].to_vec());
        start = end;
    }
    groups
}

/// `solve(m)` with R's fallback: LU inverse, else Moore-Penrose `ginv`.
fn solve_or_ginv(m: &DMatrix<f64>) -> DMatrix<f64> {
    match m.clone().try_inverse() {
        Some(inv) => inv,
        None => crate::linalg::ginv(m),
    }
}

/// Combine the `s`-weighted residual columns of cluster `idx` into a single
/// score vector via `f(l)` = column-`l` k-score, summed as `Σ_l s[l] f(l)`.
fn cluster_score(
    s: &DVector<f64>,
    ncols: usize,
    mut f: impl FnMut(usize) -> DVector<f64>,
) -> DVector<f64> {
    let mut sv = f(0) * s[0];
    for l in 1..ncols {
        sv += f(l) * s[l];
    }
    sv
}

/// Cluster-robust MEAT `w * M` (k x k). Mirrors the cluster branches of
/// `rdrobust_vce` (functions.R:363-592). The caller applies the bread
/// (`inv_g * (w*M) * inv_g`). `rx` (n x k) is the weighted score design (R*W);
/// `res` is n x (1+d); `s` is length (1+d). For CRV2/CRV3, `sqrt_rx` (R*sqrt(W),
/// n x k) and `inv_g` (k x k bread of that design) are required. `k_override`
/// sets the CR1 df dimension (else `rx.ncols()`).
#[allow(clippy::too_many_arguments)]
pub(crate) fn vce_meat_cluster(
    s: &DVector<f64>,
    rx: &DMatrix<f64>,
    res: &DMatrix<f64>,
    groups: &[Vec<usize>],
    kind: ClusterKind,
    inv_g: Option<&DMatrix<f64>>,
    sqrt_rx: Option<&DMatrix<f64>>,
    k_override: Option<usize>,
) -> DMatrix<f64> {
    let k = rx.ncols();
    let ncols = res.ncols();
    let mut m = DMatrix::zeros(k, k);

    match kind {
        ClusterKind::Cr1 => {
            // M = Σ_g sv_g sv_g'; sv_g = Σ_l s[l] (rx_g' res_g[,l]).
            for idx in groups {
                let xi = rows(rx, idx); // n_g x k
                let ri = rows(res, idx); // n_g x ncols
                let sv = cluster_score(s, ncols, |l| {
                    (xi.transpose() * ri.column(l)).column(0).into_owned()
                });
                m += &sv * sv.transpose();
            }
            let n = res.nrows() as f64;
            let k_df = k_override.unwrap_or(k) as f64;
            let g = groups.len() as f64;
            let w = ((n - 1.0) / (n - k_df)) * (g / (g - 1.0));
            m * w
        }
        ClusterKind::Crv3 => {
            // SMW: score_g = u_g + L_g (G - L_g)^{-1} u_g.  w = 1.
            let sqrt_rx = sqrt_rx.expect("crv3 requires sqrt_rx");
            let inv_g = inv_g.expect("crv3 requires inv_g");
            let g_gram = solve_or_ginv(inv_g);
            for idx in groups {
                let srx_g = rows(sqrt_rx, idx); // n_g x k
                let sw_g = srx_g.column(0).into_owned(); // sqrt(eW)
                let ri = rows(res, idx);
                let l_g = srx_g.transpose() * &srx_g; // k x k
                let m_g = solve_or_ginv(&(&g_gram - &l_g));
                let sv = cluster_score(s, ncols, |l| {
                    let e = ri.column(l);
                    let swe =
                        DVector::from_iterator(idx.len(), (0..idx.len()).map(|i| sw_g[i] * e[i]));
                    let u_g = srx_g.transpose() * swe; // k x 1
                    &u_g + &l_g * (&m_g * &u_g)
                });
                m += &sv * sv.transpose();
            }
            m
        }
        ClusterKind::Crv2 => {
            // Bell-McCaffrey: score_g = u_g + L_g R' V diag(c) V' R u_g.  w = 1.
            let sqrt_rx = sqrt_rx.expect("crv2 requires sqrt_rx");
            let inv_g = inv_g.expect("crv2 requires inv_g");
            // C_half = chol(invG) (R upper); nalgebra gives lower L (L L' = invG),
            // so C_half = L', tC_half = L. crv2_ok = chol succeeded.
            let c_half_lower = nalgebra::Cholesky::new(inv_g.clone());
            for idx in groups {
                let srx_g = rows(sqrt_rx, idx);
                let sw_g = srx_g.column(0).into_owned();
                let ri = rows(res, idx);
                let l_g = srx_g.transpose() * &srx_g; // k x k

                let u_of = |l: usize| -> DVector<f64> {
                    let e = ri.column(l);
                    let swe =
                        DVector::from_iterator(idx.len(), (0..idx.len()).map(|i| sw_g[i] * e[i]));
                    srx_g.transpose() * swe
                };

                let sv = match &c_half_lower {
                    None => cluster_score(s, ncols, u_of), // chol failed -> raw score
                    Some(ch) => {
                        let lower = ch.l(); // tC_half
                        let upper = lower.transpose(); // C_half
                        let f_sq = &upper * &l_g * &lower; // C_half L_g tC_half (symmetric)
                        let eig = nalgebra::SymmetricEigen::new(0.5 * (&f_sq + f_sq.transpose()));
                        let kk = f_sq.nrows();
                        let mut c_coef = DVector::zeros(kk);
                        for j in 0..kk {
                            let sigma2 = eig.eigenvalues[j].max(0.0);
                            c_coef[j] = if sigma2 < 1e-14 {
                                0.0
                            } else {
                                (1.0 / (1.0 - sigma2).max(1e-8).sqrt() - 1.0) / sigma2
                            };
                        }
                        let v_g = &eig.eigenvectors;
                        let lt_chalf = &l_g * &lower; // L_g tC_half
                        cluster_score(s, ncols, |l| {
                            let u_g = u_of(l);
                            let ru_g = &upper * &u_g; // C_half u_g
                                                      // adj = V diag(c) V' (R u_g)
                            let vt_ru = v_g.transpose() * &ru_g;
                            let scaled =
                                DVector::from_iterator(kk, (0..kk).map(|j| c_coef[j] * vt_ru[j]));
                            let adj = v_g * scaled;
                            &u_g + &lt_chalf * adj
                        })
                    }
                };
                m += &sv * sv.transpose();
            }
            m
        }
    }
}

/// Decoupled cluster-robust meat for the bias-corrected (h != b) estimator
/// (R `.rdrobust_vce_qq_cluster`, functions.R:607-683). Returns `M` (k x k,
/// k = ncol(q)); the caller wraps `inv_g_p * M * inv_g_p`. Uses the mixed-bandwidth
/// score `Q` with leverage from the q-regression at bandwidth b.
#[allow(clippy::too_many_arguments)]
pub(crate) fn vce_meat_qq_cluster(
    q: &DMatrix<f64>,
    r_q: &DMatrix<f64>,
    w_b: &[f64],
    inv_g_q: &DMatrix<f64>,
    res: &DMatrix<f64>,
    groups: &[Vec<usize>],
    crv2: bool,
    s: &DVector<f64>,
) -> DMatrix<f64> {
    let k = q.ncols();
    let ncols = res.ncols();
    let mut m = DMatrix::zeros(k, k);
    let g_q = solve_or_ginv(inv_g_q);

    for idx in groups {
        let q_g = rows(q, idx); // n_g x k
        let r_g = rows(r_q, idx); // n_g x k_R
        let wb_g: Vec<f64> = idx.iter().map(|&i| w_b[i]).collect();
        let sqrt_wb: Vec<f64> = wb_g.iter().map(|v| v.sqrt()).collect();
        let mut sr_g = r_g.clone();
        for i in 0..idx.len() {
            for j in 0..sr_g.ncols() {
                sr_g[(i, j)] *= sqrt_wb[i];
            }
        }
        let l_g = sr_g.transpose() * &sr_g; // k_R x k_R
        let p_g = q_g.transpose() * &r_g; // k x k_R
        let ri = rows(res, idx);

        // T2: per-cluster operator depends only on L_g/G_q -- compute once.
        let mut use_simple = false;
        let mut lambda: Option<DMatrix<f64>> = None;
        let mut m_g: Option<DMatrix<f64>> = None;
        if crv2 {
            let eig_l = nalgebra::SymmetricEigen::new(0.5 * (&l_g + l_g.transpose()));
            let kr = l_g.nrows();
            let sigma: Vec<f64> = (0..kr)
                .map(|j| eig_l.eigenvalues[j].max(0.0).sqrt())
                .collect();
            let max_sig = sigma.iter().cloned().fold(0.0_f64, f64::max);
            let keep: Vec<usize> = (0..kr).filter(|&j| sigma[j] > max_sig * 1e-10).collect();
            let r = keep.len();
            if r == 0 {
                use_simple = true;
            } else {
                // V_r (kr x r), Dsig (r x r)
                let mut v_r = DMatrix::zeros(kr, r);
                for (col, &j) in keep.iter().enumerate() {
                    v_r.set_column(col, &eig_l.eigenvectors.column(j));
                }
                let sig: Vec<f64> = keep.iter().map(|&j| sigma[j]).collect();
                let dsig = DMatrix::from_diagonal(&DVector::from_vec(sig.clone()));
                let mut t_mat = &dsig * (v_r.transpose() * (inv_g_q * &v_r)) * &dsig;
                t_mat = 0.5 * (&t_mat + t_mat.transpose());
                let eig_t = nalgebra::SymmetricEigen::new(t_mat);
                let mut gamma = DVector::zeros(r);
                for j in 0..r {
                    let tau = eig_t.eigenvalues[j].clamp(0.0, 1.0 - 1e-10);
                    gamma[j] = 1.0 / (1.0 - tau).sqrt() - 1.0;
                }
                let w_g = &eig_t.eigenvectors;
                let dinv =
                    DMatrix::from_diagonal(&DVector::from_iterator(r, sig.iter().map(|v| 1.0 / v)));
                let a = &v_r * &dinv * w_g; // kr x r
                                            // Lambda = A diag(gamma) A'
                let mut a_scaled = a.clone();
                for col in 0..r {
                    for row in 0..a_scaled.nrows() {
                        a_scaled[(row, col)] *= gamma[col];
                    }
                }
                lambda = Some(&a_scaled * a.transpose());
            }
        } else {
            m_g = Some(solve_or_ginv(&(&g_q - &l_g)));
        }

        let sv = cluster_score(s, ncols, |l| {
            let e_gl = ri.column(l).into_owned();
            let u_qq = q_g.transpose() * &e_gl; // k x 1
            if use_simple {
                u_qq.column(0).into_owned()
            } else {
                let wbe =
                    DVector::from_iterator(idx.len(), (0..idx.len()).map(|i| wb_g[i] * e_gl[i]));
                let u_q = r_g.transpose() * wbe; // k_R x 1
                let adj = if crv2 {
                    lambda.as_ref().unwrap() * &u_q
                } else {
                    m_g.as_ref().unwrap() * &u_q
                };
                (&u_qq + &p_g * adj).column(0).into_owned()
            }
        });
        m += &sv * sv.transpose();
    }
    m
}

/// Subset matrix rows by index (R `M[idx,,drop=FALSE]`).
fn rows(m: &DMatrix<f64>, idx: &[usize]) -> DMatrix<f64> {
    let mut out = DMatrix::zeros(idx.len(), m.ncols());
    for (r, &i) in idx.iter().enumerate() {
        out.set_row(r, &m.row(i));
    }
    out
}

/// Full sandwiched cluster V `inv_g * (w*M) * inv_g` for the direct (h==b style)
/// path when a cluster vector is present.
#[allow(clippy::too_many_arguments)]
pub(crate) fn v_cluster(
    inv_g: &DMatrix<f64>,
    rx: &DMatrix<f64>,
    res: &DMatrix<f64>,
    s: &DVector<f64>,
    groups: &[Vec<usize>],
    kind: ClusterKind,
    sqrt_rx: Option<&DMatrix<f64>>,
    k_override: Option<usize>,
) -> DMatrix<f64> {
    let crv_inv_g = match kind {
        ClusterKind::Cr1 => None,
        _ => Some(inv_g),
    };
    let meat = vce_meat_cluster(s, rx, res, groups, kind, crv_inv_g, sqrt_rx, k_override);
    inv_g * (meat * inv_g)
}

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

    #[test]
    fn nn_residuals_empty_window_is_empty() {
        let d = DMatrix::<f64>::zeros(0, 1);
        let res = nn_residuals(&[], &d, &[], &[], 3);
        assert_eq!((res.nrows(), res.ncols()), (0, 1));
    }

    #[test]
    fn nn_residuals_single_point_is_nan_like_r() {
        // Ji = 0 -> sqrt(0/1) * (y - 0/0): R yields NaN too.
        let d = DMatrix::from_row_slice(1, 1, &[2.0]);
        let res = nn_residuals(&[0.5], &d, &[1], &[1], 3);
        assert!(res[(0, 0)].is_nan());
    }

    #[test]
    fn dups_dupsid_runs() {
        let (d, id) = compute_dups_dupsid(&[1.0, 1.0, 2.0, 3.0, 3.0, 3.0]);
        assert_eq!(d, vec![2, 2, 1, 3, 3, 3]);
        assert_eq!(id, vec![1, 2, 1, 1, 2, 3]);
    }
}