basin 1.1.0

Numerical optimization in pure Rust, with pluggable linear-algebra backends and WASM support.
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
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
//! GETACT: LINCOA's active-set step and the incremental QR factorization of the
//! active linear-constraint normals (Powell 2015, §3; PRIMA `getact.f90`).
//!
//! GETACT solves the linearly-constrained projected problem (Powell 2015,
//! eqs. 3.3–3.5)
//!
//! ```text
//! min ‖d + g‖   subject to   a_jᵀ d ≤ 0   for j in the nearly-active set,
//! ```
//!
//! whose solution `psd` is the projected steepest-descent direction that drives
//! the truncated-CG of [`trstep`](super::trstep). It uses the Goldfarb–Idnani
//! (1983) active-set method and maintains a QR factorization of the active
//! constraint normals.
//!
//! The active set and its `Q R` factorization are **warm-started** — GETACT
//! reuses them within a TRSTEP solve and across driver iterations — so they live
//! in [`ActiveSetQr`], owned by the driver ([`LincoaWork`](super::driver)) and
//! threaded by `&mut` into the free `trstep`/`getact` functions rather than
//! hidden behind the immutable [`TrustRegionSubproblem`](crate::solver::powell::TrustRegionSubproblem)
//! seam (which stays unchanged for NEWUOA/BOBYQA). This mirrors PRIMA, where
//! `iact`/`nact`/`qfac`/`rfac` are `lincob` locals threaded as inout.
//!
//! The QR is held as bespoke `Vec<F>` scratch (column-major `n × n` `Q` and
//! `R`), maintained by Givens rotations on add/drop of an active normal
//! (`qradd_Rfull` / `qrexc_Rfull` in PRIMA `powalg.f90`) — there is no QR in
//! basin's `linalg` tier and this active-set factorization is solver-internal,
//! exactly the precedent set by the model's `H`-algebra.

use crate::core::math::Scalar;

/// Incremental QR factorization of the active linear-constraint normals,
/// warm-started across TRSTEP and driver iterations.
///
/// LINCOA normalizes every constraint normal to unit length at init, so the
/// columns of `A` referenced by `iact[..nact]` are unit vectors; `qfac`/`rfac`
/// factor the `n × nact` matrix of those active normals: column-major `n × n`
/// `Q` (orthogonal) and `R` (upper-triangular, positive diagonal on the leading
/// `nact × nact` block), with `Q[:, ..nact] · R[..nact, ..nact] = A[:, iact]`.
pub(crate) struct ActiveSetQr<F = f64> {
    /// Number of variables `n`.
    pub(crate) n: usize,
    /// Indices of the active constraints; the first [`nact`](Self::nact) entries
    /// are the active set, the rest scratch. Length `m`.
    pub(crate) iact: Vec<usize>,
    /// Size of the active set (`≤ n`).
    pub(crate) nact: usize,
    /// Orthogonal factor `Q` (`n × n`, column-major: `Q[r,c] = qfac[r + c*n]`).
    pub(crate) qfac: Vec<F>,
    /// Upper-triangular factor `R` (`n × n`, column-major); only the leading
    /// `nact × nact` block is meaningful.
    pub(crate) rfac: Vec<F>,
}

impl<F: Scalar> ActiveSetQr<F> {
    /// Cold start: empty active set, `Q = I`, `R = 0` (PRIMA `lincob.f90`
    /// initialization of `nact`/`qfac`/`rfac`/`iact`).
    pub(crate) fn new(n: usize, m: usize) -> Self {
        let mut qfac = vec![F::zero(); n * n];
        for i in 0..n {
            qfac[i + i * n] = F::one();
        }
        Self {
            n,
            iact: (0..m).collect(),
            nact: 0,
            qfac,
            rfac: vec![F::zero(); n * n],
        }
    }

    /// Overwrite `qfac` with the identity (PRIMA `qfac = eye(n)`).
    fn set_qfac_identity(&mut self) {
        let n = self.n;
        for x in self.qfac.iter_mut() {
            *x = F::zero();
        }
        for i in 0..n {
            self.qfac[i + i * n] = F::one();
        }
    }
}

/// `TINYCV` (PRIMA `consts.F90`): `2^-200`, the tiny constraint value a freshly
/// deactivated constraint's residual is floored to.
fn tinycv<F: Scalar>() -> F {
    F::from_f64(2.0).expect("2.0 representable").powi(-200)
}

/// Fortran `sign(a, b)`: `|a|` if `b ≥ 0` (including `+0`), else `-|a|`.
fn fsign<F: Scalar>(a: F, b: F) -> F {
    if b >= F::zero() { a.abs() } else { -a.abs() }
}

/// Givens rotation `(c, s)` such that `G·[x0, x1]ᵀ = [r, 0]ᵀ` with
/// `G = [[c, s], [-s, c]]` (PRIMA `linalg.f90::planerot`). Kept orthogonal for
/// the NaN/Inf/zero edge cases.
fn planerot<F: Scalar>(x0: F, x1: F) -> (F, F) {
    let (zero, one) = (F::zero(), F::one());
    if x0.is_nan() || x1.is_nan() {
        return (one, zero);
    }
    if x0.is_infinite() && x1.is_infinite() {
        let inv_sqrt2 = F::from_f64(2.0).expect("2.0 representable").sqrt().recip();
        return (fsign(inv_sqrt2, x0), fsign(inv_sqrt2, x1));
    }
    if x0.abs() <= zero && x1.abs() <= zero {
        return (one, zero);
    }
    let eps = F::epsilon();
    if x1.abs() <= eps * x0.abs() {
        return (fsign(one, x0), zero);
    }
    if x0.abs() <= eps * x1.abs() {
        return (zero, fsign(one, x1));
    }
    let r = (x0 * x0 + x1 * x1).sqrt();
    (x0 / r, x1 / r)
}

/// Append the unit column `c` (length `n`) as the new last active column,
/// updating the QR and incrementing `nact` (PRIMA `qradd_Rfull`). `Q` is `n × n`
/// column-major, `R` is `n × n` column-major. It is assumed `c` is not in the
/// span of the existing active columns (always true in LINCOA).
fn qradd<F: Scalar>(c: &[F], qfac: &mut [F], rfac: &mut [F], nact: &mut usize, n: usize) {
    let na = *nact;
    // cq = Qᵀ c.
    let mut cq = vec![F::zero(); n];
    for (k, cqk) in cq.iter_mut().enumerate() {
        let mut s = F::zero();
        for r in 0..n {
            s = s + c[r] * qfac[r + k * n];
        }
        *cqk = s;
    }
    // Rotate columns [na .. n-1] of Q so they stay orthogonal to c, zeroing
    // cq[k+1] into cq[k] for k = n-2 down to na.
    let mut kk = n as isize - 2;
    while kk >= na as isize {
        let k = kk as usize;
        if cq[k + 1].abs() > F::zero() {
            let (cc, ss) = planerot(cq[k], cq[k + 1]);
            for r in 0..n {
                let a = qfac[r + k * n];
                let b = qfac[r + (k + 1) * n];
                qfac[r + k * n] = cc * a + ss * b;
                qfac[r + (k + 1) * n] = cc * b - ss * a;
            }
            cq[k] = (cq[k] * cq[k] + cq[k + 1] * cq[k + 1]).sqrt();
        }
        kk -= 1;
    }
    // New R column: R[0..na, na] = Qᵀ c over the (unchanged) first na columns.
    for i in 0..na {
        let mut s = F::zero();
        for r in 0..n {
            s = s + c[r] * qfac[r + i * n];
        }
        rfac[i + na * n] = s;
    }
    // Keep the new diagonal entry positive.
    if cq[na] < F::zero() {
        for r in 0..n {
            qfac[r + na * n] = -qfac[r + na * n];
        }
    }
    rfac[na + na * n] = cq[na].abs();
    *nact = na + 1;
}

/// Cyclically move active column `icon` to the last (`ncols-1`) position,
/// re-triangularizing the QR by Givens rotations (PRIMA `qrexc_Rfull`).
/// `ncols` is the current active count; `n` the matrix dimension.
fn qrexc<F: Scalar>(qfac: &mut [F], rfac: &mut [F], n: usize, ncols: usize, icon: usize) {
    if icon + 1 >= ncols {
        return;
    }
    for kk in icon..(ncols - 1) {
        let r_kp1_kp1 = rfac[(kk + 1) + (kk + 1) * n];
        let r_k_kp1 = rfac[kk + (kk + 1) * n];
        let (cc, ss) = planerot(r_kp1_kp1, r_k_kp1);
        let hypt = r_kp1_kp1.hypot(r_k_kp1);
        // Q(:, [kk, kk+1]) = Q(:, [kk+1, kk]) · Gᵀ.
        for r in 0..n {
            let qk = qfac[r + kk * n];
            let qk1 = qfac[r + (kk + 1) * n];
            qfac[r + kk * n] = cc * qk1 + ss * qk;
            qfac[r + (kk + 1) * n] = cc * qk - ss * qk1;
        }
        // R([kk, kk+1], kk..ncols) = G · R([kk+1, kk], kk..ncols).
        for j in kk..ncols {
            let rkj = rfac[kk + j * n];
            let rk1j = rfac[(kk + 1) + j * n];
            rfac[kk + j * n] = cc * rk1j + ss * rkj;
            rfac[(kk + 1) + j * n] = cc * rkj - ss * rk1j;
        }
        // Swap columns kk, kk+1 of R for rows 0..=kk+1.
        for r in 0..=(kk + 1) {
            rfac.swap(r + kk * n, r + (kk + 1) * n);
        }
        rfac[kk + kk * n] = hypt;
        rfac[(kk + 1) + kk * n] = F::zero();
    }
}

/// Solve the upper-triangular system `R[..k, ..k] · x = b` (PRIMA `solve` on the
/// `istriu` branch). `R` is `n × n` column-major; `k = b.len()`.
fn rsolve_upper<F: Scalar>(rfac: &[F], n: usize, b: &[F]) -> Vec<F> {
    let k = b.len();
    let mut x = vec![F::zero(); k];
    for i in (0..k).rev() {
        let mut s = b[i];
        for j in (i + 1)..k {
            s = s - rfac[i + j * n] * x[j];
        }
        x[i] = s / rfac[i + i * n];
    }
    x
}

/// Least-squares multipliers `vlam = R[..nact,..nact]⁻¹ (Qᵀ g)` over the active
/// columns (PRIMA `lsqr_Rfull`).
fn lsqr_mult<F: Scalar>(g: &[F], qfac: &[F], rfac: &[F], n: usize, nact: usize) -> Vec<F> {
    let mut x = vec![F::zero(); nact];
    for (i, xi) in x.iter_mut().enumerate() {
        let mut s = F::zero();
        for r in 0..n {
            s = s + g[r] * qfac[r + i * n];
        }
        *xi = s;
    }
    for i in (0..nact).rev() {
        for j in (i + 1)..nact {
            x[i] = x[i] - rfac[i + j * n] * x[j];
        }
        x[i] = x[i] / rfac[i + i * n];
    }
    x
}

/// Add constraint `l` (gradient `c`) to the active set (PRIMA `addact`).
fn addact<F: Scalar>(
    l: usize,
    c: &[F],
    warm: &mut ActiveSetQr<F>,
    resact: &mut [F],
    resnew: &mut [F],
    vlam: &mut [F],
) {
    let n = warm.n;
    qradd(c, &mut warm.qfac, &mut warm.rfac, &mut warm.nact, n);
    let na = warm.nact; // already incremented
    warm.iact[na - 1] = l;
    resact[na - 1] = resnew[l];
    resnew[l] = F::zero();
    vlam[na - 1] = F::zero();
}

/// Delete the `icon`-th active constraint (PRIMA `delact`), reducing `nact`.
fn delact<F: Scalar>(
    icon: usize,
    warm: &mut ActiveSetQr<F>,
    resact: &mut [F],
    resnew: &mut [F],
    vlam: &mut [F],
) {
    let n = warm.n;
    let nact = warm.nact;
    qrexc(&mut warm.qfac, &mut warm.rfac, n, nact, icon);
    // Rotate iact/resact/vlam: position icon goes to the (now last) slot.
    let li = warm.iact[icon];
    let ri = resact[icon];
    let vi = vlam[icon];
    for t in icon..(nact - 1) {
        warm.iact[t] = warm.iact[t + 1];
        resact[t] = resact[t + 1];
        vlam[t] = vlam[t + 1];
    }
    warm.iact[nact - 1] = li;
    resact[nact - 1] = ri;
    vlam[nact - 1] = vi;
    resnew[warm.iact[nact - 1]] = resact[nact - 1].max(tinycv());
    warm.nact = nact - 1;
}

/// `psd ← −Q[:, nact..n] (Q[:, nact..n]ᵀ g)` — minus the projection of `g` onto
/// the range of the inactive `Q` columns (PRIMA's projection of `-G`).
fn project_neg_g<F: Scalar>(g: &[F], qfac: &[F], n: usize, nact: usize, psd: &mut [F]) {
    let zero = F::zero();
    for p in psd.iter_mut() {
        *p = zero;
    }
    for col in nact..n {
        let mut coeff = zero;
        for r in 0..n {
            coeff = coeff + g[r] * qfac[r + col * n];
        }
        for r in 0..n {
            psd[r] = psd[r] - coeff * qfac[r + col * n];
        }
    }
}

/// GETACT: compute the projected steepest-descent direction `psd` for the
/// linearly-constrained trust-region subproblem and (warm-)update the active set
/// and its QR factorization (Powell 2015, §3; PRIMA `getact`).
///
/// - `amat`: `n × m` column-major unit constraint normals (`a_j = amat[.. , j]`).
/// - `g`: `∇Q` at the current CG iterate (length `n`).
/// - `delta`: trust-region radius (defines the nearly-active band `0.2·Δ`).
/// - `resact`/`resnew`: TRSTEP-level residual scratch (length `m`); kept up to
///   date here (only permuted, values preserved across add/drop).
/// - `psd`: output direction (length `n`).
#[allow(clippy::too_many_arguments)]
pub(crate) fn getact<F: Scalar>(
    amat: &[F],
    n: usize,
    m: usize,
    delta: F,
    g: &[F],
    warm: &mut ActiveSetQr<F>,
    resact: &mut [F],
    resnew: &mut [F],
    psd: &mut [F],
) {
    let (zero, one) = (F::zero(), F::one());
    let two = F::from_f64(2.0).expect("2.0 representable");
    let ten = F::from_f64(10.0).expect("10.0 representable");
    let eps = F::epsilon();
    let realmax = F::max_value();

    // Quick return when there are no constraints.
    if m == 0 {
        warm.nact = 0;
        warm.set_qfac_identity();
        for i in 0..n {
            psd[i] = -g[i];
        }
        return;
    }

    let gg: F = (0..n).fold(zero, |a, i| a + g[i] * g[i]);
    let tdel = F::from_f64(0.2).expect("0.2 representable") * delta;
    if warm.nact == 0 {
        warm.set_qfac_identity();
    }

    let mut vlam = vec![zero; n];

    // Remove active constraints whose residuals exceed TDEL.
    let mut icon = warm.nact;
    while icon >= 1 {
        if resact[icon - 1] > tdel {
            delact(icon - 1, warm, resact, resnew, &mut vlam);
        }
        icon -= 1;
    }

    // Remove active constraints with nonnegative Lagrange multipliers; set the
    // surviving multipliers.
    while warm.nact > 0 {
        let v = lsqr_mult(g, &warm.qfac, &warm.rfac, n, warm.nact);
        vlam[..warm.nact].copy_from_slice(&v);
        if !(0..warm.nact).any(|i| vlam[i] >= zero) {
            break;
        }
        // icon = last index with vlam >= 0.
        let icon = (0..warm.nact)
            .filter(|&i| vlam[i] >= zero)
            .next_back()
            .unwrap();
        delact(icon, warm, resact, resnew, &mut vlam);
    }

    // Main Goldfarb–Idnani iteration.
    let mut psdsav = vec![zero; n];
    let mut ddsav = two * gg;
    let maxiter = (2 * (m + n)).max(m);

    for _iter in 0..maxiter {
        if warm.nact >= n {
            for p in psd.iter_mut() {
                *p = zero;
            }
            break;
        }

        project_neg_g(g, &warm.qfac, n, warm.nact, psd);
        let dd: F = (0..n).fold(zero, |a, i| a + psd[i] * psd[i]);
        let dnorm = dd.sqrt();

        if dnorm <= eps || dnorm.is_nan() {
            break;
        }
        if dd >= ddsav {
            for p in psd.iter_mut() {
                *p = zero;
            }
            break;
        }
        let psd_dot_g: F = (0..n).fold(zero, |a, i| a + psd[i] * g[i]);
        let abs_sum: F = (0..n).fold(zero, |a, i| a + psd[i].abs());
        if psd_dot_g > zero || !abs_sum.is_finite() {
            psd.copy_from_slice(&psdsav);
            break;
        }

        psdsav.copy_from_slice(psd);
        ddsav = dd;

        // Pick the most violated nearly-active constraint, if any.
        // apsd[j] = psdᵀ a_j.
        let mut l: Option<usize> = None;
        let mut violmx = -realmax;
        let scale = dnorm / delta;
        for j in 0..m {
            let apsd_j = (0..n).fold(zero, |a, r| a + psd[r] * amat[r + j * n]);
            if resnew[j] > zero
                && resnew[j] <= tdel
                && apsd_j > scale * resnew[j]
                && apsd_j > violmx
            {
                violmx = apsd_j;
                l = Some(j);
            }
        }

        // Terminate if no violation, or a positive VIOLMX is at the rounding
        // level relative to the active residuals.
        let act_inf = (0..warm.nact).fold(zero, |a, ic| {
            let j = warm.iact[ic];
            let apsd_j = (0..n).fold(zero, |b, r| b + psd[r] * amat[r + j * n]);
            a.max(apsd_j.abs())
        });
        let l = match l {
            Some(l) if violmx > (eps * dnorm).max(ten * act_inf) => l,
            _ => break,
        };

        // Add constraint L; column = a_L.
        let c: Vec<F> = (0..n).map(|r| amat[r + l * n]).collect();
        addact(l, &c, warm, resact, resnew, &mut vlam);

        // Drag VIOLMX down to zero, deactivating constraints as multipliers turn
        // nonnegative (PRIMA's VMU loop).
        while violmx > zero && warm.nact > 0 {
            let na = warm.nact;
            let mut rhs = vec![zero; na];
            rhs[na - 1] = one / warm.rfac[(na - 1) + (na - 1) * n];
            let vmu = rsolve_upper(&warm.rfac, n, &rhs);

            let mut vmult = violmx;
            for ic in 0..na {
                if vmu[ic] < zero && vlam[ic] < zero {
                    let f = vlam[ic] / vmu[ic];
                    if f < vmult {
                        vmult = f;
                    }
                }
            }
            // icon = last 1-based index with frac <= vmult (0 = none).
            let mut icon1 = 0usize;
            for ic in 0..na {
                let f = if vmu[ic] < zero && vlam[ic] < zero {
                    vlam[ic] / vmu[ic]
                } else {
                    realmax
                };
                if f <= vmult {
                    icon1 = ic + 1;
                }
            }

            violmx = (violmx - vmult).max(zero);
            for ic in 0..na {
                vlam[ic] = vlam[ic] - vmult * vmu[ic];
            }
            if icon1 >= 1 && icon1 <= na {
                vlam[icon1 - 1] = zero;
            }

            // Reduce the active set so all surviving multipliers are negative.
            let mut ic = warm.nact;
            while ic >= 1 {
                if vlam[ic - 1] >= zero {
                    delact(ic - 1, warm, resact, resnew, &mut vlam);
                }
                ic -= 1;
            }
        }

        if warm.nact == 0 {
            break;
        }
    }

    // NACT may be 0 here; reset to the unconstrained projection.
    if warm.nact == 0 {
        warm.set_qfac_identity();
        for i in 0..n {
            psd[i] = -g[i];
        }
    }
}

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

    fn at<F: Copy>(a: &[F], r: usize, c: usize, n: usize) -> F {
        a[r + c * n]
    }

    /// `QᵀQ == I` to tolerance.
    fn assert_orthonormal(qfac: &[f64], n: usize) {
        for i in 0..n {
            for j in 0..n {
                let mut s = 0.0;
                for r in 0..n {
                    s += at(qfac, r, i, n) * at(qfac, r, j, n);
                }
                let want = if i == j { 1.0 } else { 0.0 };
                assert!((s - want).abs() < 1e-12, "QᵀQ[{i},{j}] = {s}");
            }
        }
    }

    /// `R` upper-triangular with positive diagonal on the leading block.
    fn assert_upper_tri_pos(rfac: &[f64], n: usize, nact: usize) {
        for c in 0..nact {
            for r in 0..n {
                if r > c {
                    assert!(at(rfac, r, c, n).abs() < 1e-12, "R[{r},{c}] not zero");
                }
            }
            assert!(at(rfac, c, c, n) > 0.0, "R[{c},{c}] not positive");
        }
    }

    /// `Q[:, ..nact] · R[..nact, ..nact]` reproduces the active normals.
    fn assert_qr_reconstructs(warm: &ActiveSetQr<f64>, amat: &[f64], n: usize) {
        for (ic, &j) in warm.iact[..warm.nact].iter().enumerate() {
            for r in 0..n {
                let mut s = 0.0;
                for k in 0..warm.nact {
                    s += at(&warm.qfac, r, k, n) * at(&warm.rfac, k, ic, n);
                }
                assert!(
                    (s - at(amat, r, j, n)).abs() < 1e-10,
                    "QR col {ic} (constraint {j}) row {r}: {s} vs {}",
                    at(amat, r, j, n)
                );
            }
        }
    }

    /// `qradd` builds a correct QR of a sequence of independent unit columns.
    #[test]
    fn qradd_builds_valid_qr() {
        let n = 3;
        // Three independent (non-orthogonal) unit normals.
        let cols = [[1.0, 0.0, 0.0], [0.6, 0.8, 0.0], [0.0, 0.6, 0.8]];
        let mut warm = ActiveSetQr::<f64>::new(n, 3);
        let mut amat = vec![0.0; n * 3];
        for (j, c) in cols.iter().enumerate() {
            for r in 0..n {
                amat[r + j * n] = c[r];
            }
            qradd(c, &mut warm.qfac, &mut warm.rfac, &mut warm.nact, n);
            warm.iact[warm.nact - 1] = j;
        }
        assert_eq!(warm.nact, 3);
        assert_orthonormal(&warm.qfac, n);
        assert_upper_tri_pos(&warm.rfac, n, 3);
        assert_qr_reconstructs(&warm, &amat, n);
    }

    /// `qrexc` preserves a valid QR of the remaining columns after a deletion.
    #[test]
    fn qrexc_preserves_qr_after_deletion() {
        let n = 3;
        let cols = [[1.0, 0.0, 0.0], [0.6, 0.8, 0.0], [0.0, 0.6, 0.8]];
        let mut warm = ActiveSetQr::<f64>::new(n, 3);
        let mut amat = vec![0.0; n * 3];
        for (j, c) in cols.iter().enumerate() {
            for r in 0..n {
                amat[r + j * n] = c[r];
            }
            qradd(c, &mut warm.qfac, &mut warm.rfac, &mut warm.nact, n);
            warm.iact[warm.nact - 1] = j;
        }
        // Delete the middle active column (icon = 1): it cycles to the end and nact drops.
        let mut resact = vec![0.0; 3];
        let mut resnew = vec![1.0; 3];
        let mut vlam = vec![0.0; n];
        delact(1, &mut warm, &mut resact, &mut resnew, &mut vlam);
        assert_eq!(warm.nact, 2);
        assert_eq!(&warm.iact[..2], &[0, 2]); // constraint 1 was dropped
        assert_orthonormal(&warm.qfac, n);
        assert_upper_tri_pos(&warm.rfac, n, 2);
        assert_qr_reconstructs(&warm, &amat, n);
    }

    /// With no constraints, GETACT returns the steepest-descent direction `-g`.
    #[test]
    fn getact_unconstrained_is_neg_g() {
        let n = 2;
        let g = [1.0, -2.0];
        let mut warm = ActiveSetQr::<f64>::new(n, 0);
        let mut psd = vec![0.0; n];
        getact(&[], n, 0, 1.0, &g, &mut warm, &mut [], &mut [], &mut psd);
        assert_eq!(psd, vec![-1.0, 2.0]);
        assert_eq!(warm.nact, 0);
    }

    /// A gradient pushing into a single nearly-active constraint activates it,
    /// and `psd` becomes the projection onto that constraint's null space.
    #[test]
    fn getact_activates_single_constraint() {
        let n = 2;
        let m = 1;
        // Constraint x ≤ b, normal a = (1, 0). Gradient g = (-1, 0): -g = (1, 0)
        // points straight at the constraint, so it must become active and psd → 0.
        let amat = vec![1.0, 0.0];
        let g = [-1.0, 0.0];
        let mut warm = ActiveSetQr::<f64>::new(n, m);
        let mut resact = vec![0.0; m];
        let mut resnew = vec![1e-6; m]; // nearly active (small positive residual)
        let mut psd = vec![0.0; n];
        getact(
            &amat,
            n,
            m,
            1.0,
            &g,
            &mut warm,
            &mut resact,
            &mut resnew,
            &mut psd,
        );
        assert_eq!(warm.nact, 1);
        assert_eq!(warm.iact[0], 0);
        assert!(
            psd[0].abs() < 1e-12 && psd[1].abs() < 1e-12,
            "psd = {psd:?}"
        );
    }

    /// GETACT postconditions on a 3-D problem with two nearly-active constraints:
    /// `psd` is orthogonal to the active normals, a descent direction, and
    /// `‖psd‖² ≤ 2‖g‖²`; the QR stays consistent.
    #[test]
    fn getact_postconditions_hold() {
        let n = 3;
        let m = 2;
        // a1 = (1,0,0), a2 = (0,1,0), both nearly active; g pushes into both.
        let amat = vec![1.0, 0.0, 0.0, 0.0, 1.0, 0.0];
        let g = [-1.0, -1.0, -0.5];
        let mut warm = ActiveSetQr::<f64>::new(n, m);
        let mut resact = vec![0.0; m];
        let mut resnew = vec![1e-4; m];
        let mut psd = vec![0.0; n];
        getact(
            &amat,
            n,
            m,
            1.0,
            &g,
            &mut warm,
            &mut resact,
            &mut resnew,
            &mut psd,
        );

        let gg: f64 = g.iter().map(|x| x * x).sum();
        let dd: f64 = psd.iter().map(|x| x * x).sum();
        let psd_dot_g: f64 = psd.iter().zip(&g).map(|(a, b)| a * b).sum();
        assert!(dd <= 2.0 * gg + 1e-12, "‖psd‖² = {dd}, 2gg = {}", 2.0 * gg);
        assert!(psd_dot_g <= 1e-10, "psdᵀg = {psd_dot_g}");
        // psd ⟂ each active normal.
        for ic in 0..warm.nact {
            let j = warm.iact[ic];
            let ap: f64 = (0..n).map(|r| psd[r] * amat[r + j * n]).sum();
            assert!(ap.abs() < 1e-10, "psdᵀa_{j} = {ap}");
        }
        assert_orthonormal(&warm.qfac, n);
        // Expected: both constraints active, psd ≈ (0, 0, 0.5).
        assert_eq!(warm.nact, 2);
        assert!((psd[2] - 0.5).abs() < 1e-10, "psd = {psd:?}");
    }
}