flair-rs 0.2.1

Wasm-compilable implement of time series forecasting algorithm FLAIR.
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
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
// SPDX-License-Identifier: Apache-2.0
//
// Derived from nalgebra v0.35.0 (Copyright 2020 Sébastien Crozet, Apache-2.0).
// Ported to Vec<f64>-based standalone implementation. See NOTICE for details.

extern crate alloc;
use alloc::vec;
use alloc::vec::Vec;
use libm::{sqrt, fabs, hypot};

/// Thin SVD — nalgebra-0.35.0 algorithm ported to `Vec<f64>`.
///
/// Origin:
/// - <https://github.com/dimforge/nalgebra/blob/v0.35.0/src/linalg/svd.rs>
/// - <https://github.com/dimforge/nalgebra/blob/v0.35.0/src/linalg/bidiagonal.rs>
/// - <https://github.com/dimforge/nalgebra/blob/v0.35.0/src/linalg/givens.rs>
///
/// nalgebra's type system (`ComplexField`, `DefaultAllocator`, `OMatrix`/`OVector` generics)
/// pulls in a large dependency tree, so all types are replaced with plain `Vec<f64>`.
/// The algorithm itself (Householder bidiagonalization → Givens QR iteration) is a
/// direct port of the nalgebra implementation.
///
/// Public API:
/// - [`svd`]     : thin SVD → `(U[m×n], s[n], Vt[n×n])`
/// - [`svdvals`] : singular values only (equivalent to `scipy.linalg.svdvals`)

// ─────────────────────────────────────────────────────────────────────────────
// Givens rotation
//
// Origin: nalgebra/src/linalg/givens.rs (Apache-2.0)
// https://github.com/dimforge/nalgebra/blob/v0.35.0/src/linalg/givens.rs
// ─────────────────────────────────────────────────────────────────────────────

/// Givens rotation `G` such that `G * [a, b]^T = [r, 0]^T`.
#[derive(Clone, Copy, Debug)]
struct Givens {
    c: f64,
    s: f64,
}

impl Givens {
    fn identity() -> Self {
        Self { c: 1.0, s: 0.0 }
    }

    /// Constructs a Givens rotation from unnormalized (c, s) and returns the norm.
    ///
    /// Port of nalgebra `GivensRotation::new` / `try_new` (eps=0) for f64:
    /// `sign0 = signum(c)`, `denom = hypot(c, s)`,
    /// `norm = sign0·denom`, `c_out = |c|/denom`, `s_out = s/norm`.
    /// Returns identity rotation and norm=0 when `denom == 0`.
    fn new(c: f64, s: f64) -> (Self, f64) {
        let mod0 = fabs(c);
        let sign0 = if c >= 0.0 { 1.0 } else { -1.0 };
        let denom = sqrt(mod0 * mod0 + s * s);
        if denom > 0.0 {
            let norm = sign0 * denom;
            (Self { c: mod0 / denom, s: s / norm }, norm)
        } else {
            (Self::identity(), 0.0)
        }
    }

    /// Returns `(G, r)` such that `G * [a, b]^T = [r, 0]^T`, or `None` if `b == 0`.
    ///
    /// Port of nalgebra `GivensRotation::cancel_y` for f64:
    /// `c = |a|/denom`, `s = -b/(signum(a)·denom)`, `r = signum(a)·denom`.
    fn cancel_y(a: f64, b: f64) -> Option<(Self, f64)> {
        if b == 0.0 {
            return None;
        }
        let mod0 = fabs(a);
        let sign0 = if a >= 0.0 { 1.0 } else { -1.0 };
        let denom = sqrt(mod0 * mod0 + b * b);
        let c = mod0 / denom;
        let s = -b / (sign0 * denom);
        let r = sign0 * denom;
        Some((Self { c, s }, r))
    }

    /// Returns `(G, r)` such that `G * [a, b]^T = [0, r]^T`, or `None` if `a == 0`.
    ///
    /// Port of nalgebra `GivensRotation::cancel_x` for f64:
    /// `c = |b|/denom`, `s = a·signum(b)/denom`, `r = signum(b)·denom`.
    fn cancel_x(a: f64, b: f64) -> Option<(Self, f64)> {
        if a == 0.0 {
            return None;
        }
        let mod1 = fabs(b);
        let sign1 = if b >= 0.0 { 1.0 } else { -1.0 };
        let denom = sqrt(mod1 * mod1 + a * a);
        let c = mod1 / denom;
        let s = (a * sign1) / denom;
        let r = sign1 * denom;
        Some((Self { c, s }, r))
    }

    /// Inverse (transpose) of this rotation.
    fn inverse(self) -> Self {
        Self { c: self.c, s: -self.s }
    }

    /// Right-applies this rotation to column pair `(ci, cj)`: `lhs <- lhs * G`.
    ///
    /// Port of nalgebra `GivensRotation::rotate_rows` for f64:
    /// `new_ci = c*ci + s*cj`, `new_cj = -s*ci + c*cj`.
    fn rotate_rows(self, mat: &mut [Vec<f64>], ci: usize, cj: usize) {
        for row in mat.iter_mut() {
            let a = row[ci];
            let b = row[cj];
            row[ci] = self.c * a + self.s * b;
            row[cj] = -self.s * a + self.c * b;
        }
    }

    /// Left-applies this rotation to row pair `(ri, rj)`: `rhs <- G * rhs`.
    ///
    /// Port of nalgebra `GivensRotation::rotate` for f64:
    /// `new_ri = c*ri - s*rj`, `new_rj = s*ri + c*rj`
    /// (note opposite sign convention from `rotate_rows`).
    fn rotate(self, mat: &mut [Vec<f64>], ri: usize, rj: usize) {
        let ncols = mat[0].len();
        for k in 0..ncols {
            let a = mat[ri][k];
            let b = mat[rj][k];
            mat[ri][k] = self.c * a - self.s * b;
            mat[rj][k] = self.s * a + self.c * b;
        }
    }
}

/// Right-applies `G` to column pair `(c0, c1)` of a local 2×3 submatrix.
///
/// Equivalent to nalgebra `rot.rotate_rows(subm.fixed_columns_mut::<2>(c0))`.
fn subm_rotate_rows(subm: &mut [[f64; 3]; 2], g: Givens, c0: usize, c1: usize) {
    for r in 0..2 {
        let a = subm[r][c0];
        let b = subm[r][c1];
        subm[r][c0] = g.c * a + g.s * b;
        subm[r][c1] = -g.s * a + g.c * b;
    }
}

/// Left-applies `G` to rows 0..2 of a local 2×3 submatrix over columns `cstart..cend`.
///
/// Equivalent to nalgebra `rot.rotate(subm.fixed_columns_mut::<2>(cstart))`.
fn subm_rotate(subm: &mut [[f64; 3]; 2], g: Givens, cstart: usize, cend: usize) {
    for col in cstart..cend {
        let a = subm[0][col];
        let b = subm[1][col];
        subm[0][col] = g.c * a - g.s * b;
        subm[1][col] = g.s * a + g.c * b;
    }
}

// ─────────────────────────────────────────────────────────────────────────────
// Householder reflection axis construction
//
// Origin: nalgebra/src/linalg/householder.rs `reflection_axis_mut` (Apache-2.0)
// https://github.com/dimforge/nalgebra/blob/v0.35.0/src/linalg/householder.rs
//
// `ComplexField::to_exp` / `unscale_mut` / `normalize_mut` are expanded for f64.
// Complex conjugate is identity for reals and omitted.
// ─────────────────────────────────────────────────────────────────────────────

/// Replaces `column` in-place with the unit Householder axis that maps
/// `column` to `(±‖column‖, 0, …, 0)`.
///
/// Returns `(reflection_norm, not_zero)`.
/// `reflection_norm` is the signed norm placed on the diagonal/off-diagonal.
/// When `not_zero == false` the reflection is unnecessary.
///
/// Port of nalgebra `reflection_axis_mut` for f64.
fn reflection_axis_mut(column: &mut [f64]) -> (f64, bool) {
    let reflection_sq_norm: f64 = column.iter().map(|x| x * x).sum();
    let reflection_norm = sqrt(reflection_sq_norm);

    // to_exp() for reals: modulus = |x|, sign = signum(x) (sign = 1 when x = 0)
    let x0 = column[0];
    let sign = if x0 >= 0.0 { 1.0 } else { -1.0 };
    let modulus = fabs(x0);
    let signed_norm = sign * reflection_norm;
    let factor = (reflection_sq_norm + modulus * reflection_norm) * 2.0;
    column[0] += signed_norm;

    if factor != 0.0 {
        let inv = 1.0 / sqrt(factor);
        for c in column.iter_mut() {
            *c *= inv;
        }
        // Second normalization pass to enforce exact unit length (see nalgebra comment).
        let nrm: f64 = sqrt(column.iter().map(|x| x * x).sum::<f64>());
        if nrm > 0.0 {
            let inv2 = 1.0 / nrm;
            for c in column.iter_mut() {
                *c *= inv2;
            }
        }
        (-signed_norm, true)
    } else {
        (signed_norm, false)
    }
}

/// Applies a Householder reflection defined by a unit axis to the columns of `mat`.
///
/// `col_new = sign · (I − 2·axis·axisᵀ) · col_old`
///
/// Port of nalgebra `Reflection::reflect_with_sign` for f64 (bias = 0).
/// Acts on `mat[from_row..][from_col..ncols]`.
fn reflect_cols_with_sign(
    mat: &mut [Vec<f64>],
    axis: &[f64],
    from_row: usize,
    from_col: usize,
    ncols: usize,
    sign: f64,
) {
    for j in from_col..ncols {
        let dot: f64 = axis
            .iter()
            .enumerate()
            .map(|(i, ai)| ai * mat[from_row + i][j])
            .sum();
        let factor = sign * (-2.0) * dot;
        for (i, ai) in axis.iter().enumerate() {
            mat[from_row + i][j] = factor * ai + sign * mat[from_row + i][j];
        }
    }
}

/// Applies a Householder reflection defined by a unit axis to the rows of `mat` (right-side / Vt).
///
/// `row_new = sign · row_old · (I − 2·axis·axisᵀ)`
///
/// Port of nalgebra `Reflection::reflect_rows_with_sign` for f64 (bias = 0).
/// Acts on `mat[from_row..nrows][from_col..]`.
fn reflect_rows_with_sign(
    mat: &mut [Vec<f64>],
    axis: &[f64],
    from_row: usize,
    nrows: usize,
    from_col: usize,
    sign: f64,
) {
    for i in from_row..nrows {
        let dot: f64 = axis
            .iter()
            .enumerate()
            .map(|(j, aj)| aj * mat[i][from_col + j])
            .sum();
        let factor = sign * (-2.0) * dot;
        for (j, aj) in axis.iter().enumerate() {
            mat[i][from_col + j] = factor * aj + sign * mat[i][from_col + j];
        }
    }
}

// ─────────────────────────────────────────────────────────────────────────────
// Householder bidiagonalization
//
// Origin: nalgebra/src/linalg/bidiagonal.rs
//         `Bidiagonal::new` / `clear_column_unchecked` / `clear_row_unchecked` (Apache-2.0)
// https://github.com/dimforge/nalgebra/blob/v0.35.0/src/linalg/bidiagonal.rs
//
// `OMatrix` replaced with `Vec<Vec<f64>>`. Only the upper-diagonal (m >= n) path is
// implemented, which is the only case used in flair.
// ─────────────────────────────────────────────────────────────────────────────

/// Reduces `a` to upper bidiagonal form using Householder reflections.
///
/// Returns:
/// - `diag`    : diagonal elements (signed, length n)
/// - `off`     : superdiagonal elements (signed, length n-1)
/// - `u_axes`  : U-side unit axes `(col, from_row, axis)`
/// - `vt_axes` : Vt-side unit axes `(row, from_col, axis)`
fn bidiagonalize(
    a: &[Vec<f64>],
) -> (Vec<f64>, Vec<f64>, Vec<(usize, usize, Vec<f64>)>, Vec<(usize, usize, Vec<f64>)>) {
    let m = a.len();
    let n = a[0].len();
    let mut work: Vec<Vec<f64>> = a.to_vec();
    let mut diag = vec![0.0f64; n];
    let mut off = vec![0.0f64; n.saturating_sub(1)];
    let mut u_axes: Vec<(usize, usize, Vec<f64>)> = Vec::new();
    let mut vt_axes: Vec<(usize, usize, Vec<f64>)> = Vec::new();

    assert!(m >= n, "bidiagonalize: requires m >= n");

    for k in 0..n {
        // Left Householder: zero out column k below row k (clear_column_unchecked)
        let mut axis: Vec<f64> = (k..m).map(|i| work[i][k]).collect();
        let (reflection_norm, not_zero) = reflection_axis_mut(&mut axis);
        diag[k] = reflection_norm;
        if not_zero {
            let sign = if reflection_norm >= 0.0 { 1.0 } else { -1.0 };
            reflect_cols_with_sign(&mut work, &axis, k, k + 1, n, sign);
            u_axes.push((k, k, axis));
        }

        // Right Householder: zero out row k right of column k+1 (clear_row_unchecked)
        if k + 1 < n {
            let mut axis: Vec<f64> = (k + 1..n).map(|j| work[k][j]).collect();
            let (reflection_norm, not_zero) = reflection_axis_mut(&mut axis);
            // nalgebra calls axis.conjugate_mut() here; identity for reals.
            off[k] = reflection_norm;
            if not_zero {
                let sign = if reflection_norm >= 0.0 { 1.0 } else { -1.0 };
                reflect_rows_with_sign(&mut work, &axis, k + 1, m, k + 1, sign);
                vt_axes.push((k, k + 1, axis));
            }
        }
    }

    (diag, off, u_axes, vt_axes)
}

/// Assembles U (m×n) from the stored unit axes.
///
/// Port of nalgebra `Bidiagonal::u()`: applies each reflection in reverse order
/// with `sign = signum(diag[i])`.
fn assemble_u(
    m: usize,
    n: usize,
    diag: &[f64],
    u_axes: &[(usize, usize, Vec<f64>)],
) -> Vec<Vec<f64>> {
    let mut u = vec![vec![0.0f64; n]; m];
    for i in 0..n {
        u[i][i] = 1.0;
    }
    for &(col, from_row, ref axis) in u_axes.iter().rev() {
        let sign = if diag[col] >= 0.0 { 1.0 } else { -1.0 };
        reflect_cols_with_sign(&mut u, axis, from_row, col, n, sign);
    }
    u
}

/// Assembles Vt (n×n) from the stored unit axes.
///
/// Port of nalgebra `Bidiagonal::v_t()`: applies each reflection in reverse order
/// with `sign = signum(off[i])`.
fn assemble_vt(
    n: usize,
    off: &[f64],
    vt_axes: &[(usize, usize, Vec<f64>)],
) -> Vec<Vec<f64>> {
    let mut vt = vec![vec![0.0f64; n]; n];
    for i in 0..n {
        vt[i][i] = 1.0;
    }
    for &(row, from_col, ref axis) in vt_axes.iter().rev() {
        let sign = if off[row] >= 0.0 { 1.0 } else { -1.0 };
        reflect_rows_with_sign(&mut vt, axis, row, n, from_col, sign);
    }
    vt
}

// ─────────────────────────────────────────────────────────────────────────────
// Wilkinson shift
//
// Origin: nalgebra/src/linalg/symmetric_eigen.rs `wilkinson_shift` (Apache-2.0)
// https://github.com/dimforge/nalgebra/blob/v0.35.0/src/linalg/symmetric_eigen.rs
// ─────────────────────────────────────────────────────────────────────────────

/// Wilkinson shift: eigenvalue of `[[tmm, tmn], [tmn, tnn]]` closest to `tnn`.
#[inline]
fn wilkinson_shift(tmm: f64, tnn: f64, tmn: f64) -> f64 {
    let d = (tmm - tnn) * 0.5;
    let sign_d = if d >= 0.0 { 1.0 } else { -1.0 };
    tnn - sign_d * tmn * tmn / (fabs(d) + sqrt(d * d + tmn * tmn))
}

// ─────────────────────────────────────────────────────────────────────────────
// 2×2 upper-triangular SVD
//
// Origin: nalgebra/src/linalg/svd.rs `compute_2x2_uptrig_svd` (Apache-2.0)
// Reference: Qiao & Wang, "Computing the Singular Values of 2-by-2 Complex Matrices"
// http://www.cas.mcmaster.ca/sqrl/papers/sqrl5.pdf
// ─────────────────────────────────────────────────────────────────────────────

/// SVD of a 2×2 upper-triangular matrix `[[m11, m12], [0, m22]]`.
///
/// Returns `(u_rot, [s1, s2], vt_rot)`.
fn svd_2x2_uptrig(
    m11: f64,
    m12: f64,
    m22: f64,
    compute_u: bool,
    compute_v: bool,
) -> (Option<Givens>, [f64; 2], Option<Givens>) {
    let denom = hypot(m11 + m22, m12) + hypot(m11 - m22, m12);

    // v1 is the singular value closest to m22 (avoids cancellation; see paper NOTE).
    let mut v1 = m11 * m22 * 2.0 / denom;
    let mut v2 = 0.5 * denom;

    let mut u_rot = None;
    let mut v_rot = None;

    if compute_u || compute_v {
        // Origin uses GivensRotation::new (not cancel_y); sgn_v is multiplied directly.
        let (csv, sgn_v) = Givens::new(m11 * m12, v1 * v1 - m11 * m11);
        v1 *= sgn_v;
        v2 *= sgn_v;

        if compute_v {
            v_rot = Some(csv);
        }

        let cu = (m11 * csv.c + m12 * csv.s) / v1;
        let su = (m22 * csv.s) / v1;
        let (csu, sgn_u) = Givens::new(cu, su);
        v1 *= sgn_u;
        v2 *= sgn_u;

        if compute_u {
            u_rot = Some(csu);
        }
    }

    (u_rot, [v1, v2], v_rot)
}

// ─────────────────────────────────────────────────────────────────────────────
// Subproblem splitting + Givens zeroing
//
// Origin: nalgebra/src/linalg/svd.rs `delimit_subproblem` /
//         `cancel_horizontal_off_diagonal_elt` /
//         `cancel_vertical_off_diagonal_elt` (Apache-2.0)
// ─────────────────────────────────────────────────────────────────────────────

fn cancel_horizontal(
    diag: &mut [f64],
    off: &mut [f64],
    u: &mut Option<Vec<Vec<f64>>>,
    _vt: &mut Option<Vec<Vec<f64>>>,
    i: usize,
    end: usize,
) {
    let mut vx = off[i];
    let mut vy = diag[i + 1];
    off[i] = 0.0;
    for k in i..end {
        if let Some((rot, norm)) = Givens::cancel_x(vx, vy) {
            diag[k + 1] = norm;
            // upper_diagonal = true → update U side
            if let Some(u) = u {
                rot.inverse().rotate_rows(u, i, k + 1);
            }
            if k + 1 != end {
                vx = -rot.s * off[k + 1];
                vy = diag[k + 2];
                off[k + 1] *= rot.c;
            }
        } else {
            break;
        }
    }
}

fn cancel_vertical(
    diag: &mut [f64],
    off: &mut [f64],
    _u: &mut Option<Vec<Vec<f64>>>,
    vt: &mut Option<Vec<Vec<f64>>>,
    i: usize,
) {
    let mut vx = diag[i];
    let mut vy = off[i];
    off[i] = 0.0;
    for k in (0..=i).rev() {
        if let Some((rot, norm)) = Givens::cancel_y(vx, vy) {
            diag[k] = norm;
            // upper_diagonal = true → update Vt side
            if let Some(vt) = vt {
                rot.rotate(vt, k, i + 1);
            }
            if k > 0 {
                vx = diag[k - 1];
                vy = rot.s * off[k - 1];
                off[k - 1] *= rot.c;
            }
        } else {
            break;
        }
    }
}

fn delimit_subproblem(
    diag: &mut Vec<f64>,
    off: &mut Vec<f64>,
    u: &mut Option<Vec<Vec<f64>>>,
    vt: &mut Option<Vec<Vec<f64>>>,
    end: usize,
    eps: f64,
) -> (usize, usize) {
    let mut n = end;
    while n > 0 {
        let m = n - 1;
        if fabs(off[m]) <= eps * (fabs(diag[n]) + fabs(diag[m])) {
            off[m] = 0.0;
        } else if fabs(diag[m]) <= eps {
            diag[m] = 0.0;
            cancel_horizontal(diag, off, u, vt, m, m + 1);
            if m != 0 {
                cancel_vertical(diag, off, u, vt, m - 1);
            }
        } else if fabs(diag[n]) <= eps {
            diag[n] = 0.0;
            cancel_vertical(diag, off, u, vt, m);
        } else {
            break;
        }
        n -= 1;
    }

    if n == 0 {
        return (0, 0);
    }

    let mut new_start = n - 1;
    while new_start > 0 {
        let m = new_start - 1;
        if fabs(off[m]) <= eps * (fabs(diag[new_start]) + fabs(diag[m])) {
            off[m] = 0.0;
            break;
        } else if fabs(diag[m]) <= eps {
            diag[m] = 0.0;
            cancel_horizontal(diag, off, u, vt, m, n);
            if m != 0 {
                cancel_vertical(diag, off, u, vt, m - 1);
            }
            break;
        }
        new_start -= 1;
    }

    (new_start, n)
}

// ─────────────────────────────────────────────────────────────────────────────
// Public API
// ─────────────────────────────────────────────────────────────────────────────

/// Thin SVD: `A = U * diag(s) * Vt`, singular values in descending order.
///
/// - `a`: m×n matrix (slice of row vectors); accepts m >= n and m < n
/// - Returns: `(U[m×k], s[k], Vt[k×n])` where k = min(m, n)
///
/// Equivalent to `numpy.linalg.svd(a, full_matrices=False)`.
pub fn svd(a: &[Vec<f64>]) -> (Vec<Vec<f64>>, Vec<f64>, Vec<Vec<f64>>) {
    let m = a.len();
    let n = if m > 0 { a[0].len() } else { 0 };
    if m >= n {
        svd_impl(a, true, true)
    } else {
        // m < n: compute on A^T (n×m) and swap U, Vt
        // svd(A^T) = (U', s, Vt')  →  svd(A) = (Vt'^T, s, U'^T)
        let at = transpose(a, m, n);
        let (u_t, s, vt_t) = svd_impl(&at, true, true);
        let u = transpose(&vt_t, n, m);   // Vt'^T: m×k
        let vt = transpose(&u_t, m, n);   // U'^T:  k×n
        (u, s, vt)
    }
}

/// Returns singular values only (U and Vt are not computed).
///
/// Equivalent to `scipy.linalg.svdvals(a)`.
pub fn svdvals(a: &[Vec<f64>]) -> Vec<f64> {
    let m = a.len();
    let n = if m > 0 { a[0].len() } else { 0 };
    if m >= n {
        let (_, s, _) = svd_impl(a, false, false);
        s
    } else {
        let at = transpose(a, m, n);
        let (_, s, _) = svd_impl(&at, false, false);
        s
    }
}

fn transpose(a: &[Vec<f64>], rows: usize, cols: usize) -> Vec<Vec<f64>> {
    let mut out = vec![vec![0.0f64; rows]; cols];
    for i in 0..rows {
        for j in 0..cols {
            out[j][i] = a[i][j];
        }
    }
    out
}

// ─────────────────────────────────────────────────────────────────────────────
// Internal implementation
//
// Origin: nalgebra/src/linalg/svd.rs `SVD::try_new_unordered` (Apache-2.0)
// ─────────────────────────────────────────────────────────────────────────────

fn svd_impl(
    a: &[Vec<f64>],
    compute_u: bool,
    compute_v: bool,
) -> (Vec<Vec<f64>>, Vec<f64>, Vec<Vec<f64>>) {
    let m = a.len();
    assert!(m > 0, "cannot compute SVD of empty matrix");
    let n = a[0].len();
    assert!(m >= n, "svd_impl: requires m >= n (thin SVD)");
    let dim = n;

    // Step 1: scale by amax (nalgebra `try_new_unordered` prologue)
    let amax = a
        .iter()
        .flat_map(|r: &Vec<f64>| r.iter())
        .cloned()
        .fold(0.0_f64, |acc: f64, x: f64| if fabs(x) > acc { fabs(x) } else { acc });
    let scale = if amax == 0.0 { 1.0 } else { amax };
    let a_scaled: Vec<Vec<f64>> = a
        .iter()
        .map(|r: &Vec<f64>| r.iter().map(|x| x / scale).collect())
        .collect();

    // Step 2: Householder bidiagonalization
    let (mut diag, mut off, u_refl, vt_refl) = bidiagonalize(&a_scaled);

    // Step 3: assemble U and Vt
    let mut u_mat: Option<Vec<Vec<f64>>> = if compute_u {
        Some(assemble_u(m, dim, &diag, &u_refl))
    } else {
        None
    };
    let mut vt_mat: Option<Vec<Vec<f64>>> = if compute_v {
        Some(assemble_vt(dim, &off, &vt_refl))
    } else {
        None
    };

    // Signs are absorbed into U/Vt; take absolute values on the bidiagonal.
    for d in diag.iter_mut() {
        *d = fabs(*d);
    }
    for e in off.iter_mut() {
        *e = fabs(*e);
    }

    // Step 4: QR iteration (Golub-Reinsch implicit-shift)
    // Origin: nalgebra/src/linalg/svd.rs `SVD::try_new_unordered` main loop
    let eps = f64::EPSILON * 5.0;

    let (mut start, mut end) = delimit_subproblem(
        &mut diag, &mut off, &mut u_mat, &mut vt_mat, dim - 1, eps,
    );

    // nalgebra uses max_niter=0 (unlimited); we cap with a large bound as a safety net.
    let max_niter = 100 * dim + 1000;
    let mut niter = 0usize;

    while end != start {
        let subdim = end - start + 1;

        if subdim > 2 {
            // Implicit QR step with Wilkinson shift
            let m_idx = end - 1;
            let n_idx = end;

            let dm = diag[m_idx];
            let dn = diag[n_idx];
            let fm = off[m_idx];
            let tmm = dm * dm + if m_idx > 0 { off[m_idx - 1] * off[m_idx - 1] } else { 0.0 };
            let tmn = dm * fm;
            let tnn = dn * dn + fm * fm;
            let shift = wilkinson_shift(tmm, tnn, tmn);

            let mut vx = diag[start] * diag[start] - shift;
            let mut vy = diag[start] * off[start];

            for k in start..n_idx {
                let m12 = if k == n_idx - 1 { 0.0 } else { off[k + 1] };

                // Local 2×3 submatrix (nalgebra `Matrix2x3`):
                //   [[d[k], off[k], 0], [0, d[k+1], m12]]
                // nalgebra applies rotate_rows / rotate to fixed-column views of this submatrix.
                let mut subm: [[f64; 3]; 2] = [
                    [diag[k], off[k], 0.0],
                    [0.0, diag[k + 1], m12],
                ];

                if let Some((rot1, norm1)) = Givens::cancel_y(vx, vy) {
                    // rot1.inverse().rotate_rows(subm.cols(0,1)): right-apply to column pair (0,1)
                    let inv1 = rot1.inverse();
                    subm_rotate_rows(&mut subm, inv1, 0, 1);

                    if k > start {
                        off[k - 1] = norm1;
                    }

                    let (rot2, norm2) = Givens::cancel_y(subm[0][0], subm[1][0])
                        .unwrap_or((Givens::identity(), subm[0][0]));

                    // rot2.rotate(subm.cols(1,2)): left-apply to rows 0..2, columns 1..3
                    subm_rotate(&mut subm, rot2, 1, 3);
                    subm[0][0] = norm2;

                    // Update Vt (upper_diagonal=true → rot1.rotate(v_t.rows(k, k+1)))
                    if let Some(ref mut vt) = vt_mat {
                        rot1.rotate(vt, k, k + 1);
                    }
                    // Update U (upper_diagonal=true → rot2.inverse().rotate_rows(u.cols(k, k+1)))
                    if let Some(ref mut u) = u_mat {
                        rot2.inverse().rotate_rows(u, k, k + 1);
                    }

                    diag[k] = subm[0][0];
                    diag[k + 1] = subm[1][1];
                    off[k] = subm[0][1];
                    if k != n_idx - 1 {
                        off[k + 1] = subm[1][2];
                    }
                    vx = subm[0][1];
                    vy = subm[0][2];
                } else {
                    break;
                }
            }
        } else if subdim == 2 {
            // Closed-form SVD for the remaining 2×2 block
            let (u2, s2, v2) = svd_2x2_uptrig(
                diag[start],
                off[start],
                diag[start + 1],
                compute_u,
                compute_v,
            );
            diag[start] = s2[0];
            diag[start + 1] = s2[1];
            off[start] = 0.0;

            if let (Some(u), Some(rot)) = (&mut u_mat, u2) {
                rot.rotate_rows(u as &mut Vec<Vec<f64>>, start, start + 1);
            }
            if let (Some(vt), Some(rot)) = (&mut vt_mat, v2) {
                rot.inverse().rotate(vt as &mut Vec<Vec<f64>>, start, start + 1);
            }
            end -= 1;
        }

        let sub = delimit_subproblem(
            &mut diag, &mut off, &mut u_mat, &mut vt_mat, end, eps,
        );
        start = sub.0;
        end = sub.1;

        niter += 1;
        assert!(niter < max_niter, "svd: QR iteration did not converge");
    }

    // Step 5: restore scale
    for d in diag.iter_mut() {
        *d *= scale;
    }

    // Flip sign of any negative singular values, reflecting into the corresponding U column.
    for i in 0..dim {
        if diag[i] < 0.0 {
            diag[i] = -diag[i];
            if let Some(ref mut u) = u_mat {
                for row in (u as &mut Vec<Vec<f64>>).iter_mut() {
                    row[i] = -row[i];
                }
            }
        }
    }

    // Step 6: sort descending (nalgebra `sort_by_singular_values`)
    let mut idx: Vec<usize> = (0..dim).collect();
    idx.sort_unstable_by(|&a, &b| {
        diag[b].partial_cmp(&diag[a]).unwrap()
    });

    let sorted_s: Vec<f64> = idx.iter().map(|&i| diag[i]).collect();

    let sorted_u = u_mat.map(|u| {
        let mut su = vec![vec![0.0f64; dim]; m];
        for (new_j, &old_j) in idx.iter().enumerate() {
            for i in 0..m {
                su[i][new_j] = u[i][old_j];
            }
        }
        su
    });

    let sorted_vt = vt_mat.map(|vt| {
        let mut svt = vec![vec![0.0f64; n]; dim];
        for (new_i, &old_i) in idx.iter().enumerate() {
            svt[new_i] = vt[old_i].clone();
        }
        svt
    });

    (
        sorted_u.unwrap_or_default(),
        sorted_s,
        sorted_vt.unwrap_or_default(),
    )
}

// ─────────────────────────────────────────────────────────────────────────────
// Tests — compared against nalgebra reference
// ─────────────────────────────────────────────────────────────────────────────

#[cfg(test)]
mod tests {
    use super::*;
    extern crate std;
    use std::println;

    fn nalgebra_singvals(a: &[Vec<f64>]) -> Vec<f64> {
        use nalgebra::DMatrix;
        let m = a.len();
        let n = a[0].len();
        let dm = DMatrix::from_fn(m, n, |r, c| a[r][c]);
        let sv = dm.svd(false, false);
        sv.singular_values.iter().copied().collect()
    }

    fn reconstruction_error(a: &[Vec<f64>], u: &[Vec<f64>], s: &[f64], vt: &[Vec<f64>]) -> f64 {
        let m = a.len();
        let n = a[0].len();
        let k = s.len();
        let mut err = 0.0f64;
        for i in 0..m {
            for j in 0..n {
                let rec: f64 = (0..k).map(|r| u[i][r] * s[r] * vt[r][j]).sum();
                err += (a[i][j] - rec).powi(2);
            }
        }
        err.sqrt()
    }

    fn orthogonality_error_u(u: &[Vec<f64>]) -> f64 {
        let m = u.len();
        let k = u[0].len();
        let mut err = 0.0f64;
        for i in 0..k {
            for j in 0..k {
                let dot: f64 = (0..m).map(|r| u[r][i] * u[r][j]).sum();
                let expected = if i == j { 1.0 } else { 0.0 };
                err += (dot - expected).powi(2);
            }
        }
        err.sqrt()
    }

    fn orthogonality_error_vt(vt: &[Vec<f64>]) -> f64 {
        let k = vt.len();
        let n = vt[0].len();
        let mut err = 0.0f64;
        for i in 0..k {
            for j in 0..k {
                let dot: f64 = (0..n).map(|c| vt[i][c] * vt[j][c]).sum();
                let expected = if i == j { 1.0 } else { 0.0 };
                err += (dot - expected).powi(2);
            }
        }
        err.sqrt()
    }

    fn assert_svd_quality(label: &str, a: &[Vec<f64>], tol_recon: f64, tol_orth: f64) {
        let (u, s, vt) = svd(a);
        let recon   = reconstruction_error(a, &u, &s, &vt);
        let orth_u  = orthogonality_error_u(&u);
        let orth_vt = orthogonality_error_vt(&vt);
        assert!(recon   < tol_recon, "{label}: reconstruction error {recon:.2e} >= {tol_recon:.2e}");
        assert!(orth_u  < tol_orth,  "{label}: U orthogonality {orth_u:.2e} >= {tol_orth:.2e}");
        assert!(orth_vt < tol_orth,  "{label}: Vt orthogonality {orth_vt:.2e} >= {tol_orth:.2e}");
    }

    #[test]
    fn test_svd_basic() {
        let a = vec![vec![1.0_f64, 2.0], vec![3.0, 4.0], vec![5.0, 6.0]];
        let (u, s, vt) = svd(&a);
        assert_eq!(u.len(), 3);
        assert_eq!(u[0].len(), 2);
        assert_eq!(s.len(), 2);
        assert_eq!(vt.len(), 2);
        assert_eq!(vt[0].len(), 2);
        assert!(s[0] >= s[1], "singular values not in descending order");
        assert!(s[0] > 0.0);
        println!("test_svd_basic: s = {:?}", s);
    }

    #[test]
    fn test_svdvals_basic() {
        let a = vec![vec![1.0_f64, 2.0], vec![3.0, 4.0], vec![5.0, 6.0]];
        let s = svdvals(&a);
        assert_eq!(s.len(), 2);
        assert!(s[0] >= s[1]);
        assert!(s[0] > 0.0);
        println!("test_svdvals_basic: s = {:?}", s);
    }

    #[test]
    fn test_against_nalgebra() {
        // nalgebra reference: [9.52551809, 0.51430058]
        let a = vec![vec![1.0_f64, 2.0], vec![3.0, 4.0], vec![5.0, 6.0]];
        let s = svdvals(&a);
        assert!((s[0] - 9.5255).abs() < 1e-4, "s[0] expected ~9.5255, got {}", s[0]);
        assert!((s[1] - 0.5143).abs() < 1e-4, "s[1] expected ~0.5143, got {}", s[1]);

        // Verify β = Vt^T * diag(1/s) * U^T * y recovers OLS coefficients.
        let x_rows: Vec<Vec<f64>> = (0..30).map(|i| vec![1.0, i as f64 / 30.0]).collect();
        let y_lin: Vec<f64> = x_rows.iter().map(|r| 2.0 + 3.0 * r[1]).collect();
        let (u, s2, vt) = svd(&x_rows);
        let m = x_rows.len();
        let k = s2.len();
        let nf = vt[0].len();
        let uty: Vec<f64> = (0..k).map(|j| (0..m).map(|i| u[i][j] * y_lin[i]).sum()).collect();
        let mut beta = vec![0.0f64; nf];
        for col in 0..nf {
            beta[col] = (0..k).map(|j| vt[j][col] * uty[j] / s2[j]).sum::<f64>();
        }
        assert!((beta[0] - 2.0).abs() < 1e-10, "intercept expected 2.0, got {}", beta[0]);
        assert!((beta[1] - 3.0).abs() < 1e-10, "slope expected 3.0, got {}", beta[1]);
    }

    /// Accuracy of the trailing 2×2 block handler.
    #[test]
    fn svd_2x2_trailing_block() {
        let cases: &[(&str, Vec<Vec<f64>>)] = &[
            ("3x3-trailing-2x2",
             vec![vec![1.0f64, 0.0, 0.0],
                  vec![0.0,    1.0, 1.0],
                  vec![0.0,    1.0, -1.0]]),
            ("3x3-high-cond",
             vec![vec![1e3f64, 1.0,  0.0],
                  vec![0.0,    1.0,  1.0],
                  vec![0.0,    1e-3, 1.0]]),
            ("3x3-dense",
             vec![vec![4.0f64, 3.0, 2.0],
                  vec![3.0,    2.0, 1.0],
                  vec![1.0,    1.0, 5.0]]),
        ];
        for (label, a) in cases {
            let s_ref = nalgebra_singvals(a);
            let s = svdvals(a);
            for (i, (&sr, &sf)) in s_ref.iter().zip(s.iter()).enumerate() {
                let err = if sr > 1e-8 { (sr - sf).abs() / sr } else { (sr - sf).abs() };
                assert!(err < 1e-6,
                    "{label} s[{i}]: nalgebra={sr:.6e} flair={sf:.6e} err={err:.2e}");
            }
            assert_svd_quality(label, a, 1e-11, 1e-11);
        }
    }

    /// High-condition-number diagonal matrix.
    #[test]
    fn svd_high_condition_number() {
        let scales = [1e6f64, 1e3, 1.0, 1e-3];
        let a: Vec<Vec<f64>> = (0..4).map(|i|
            (0..4).map(|j| if i == j { scales[i] } else { 0.0 }).collect()
        ).collect();
        let s_ref = nalgebra_singvals(&a);
        let s = svdvals(&a);
        for (i, (&sr, &sf)) in s_ref.iter().zip(s.iter()).enumerate() {
            let rel = (sr - sf).abs() / sr.max(1e-15);
            assert!(rel < 1e-6,
                "high-cond s[{i}]: nalgebra={sr:.6e} flair={sf:.6e} rel={rel:.2e}");
        }
        assert_svd_quality("high-cond", &a, 1e-11, 1e-11);
    }

    /// Rank-1 matrix.
    #[test]
    fn svd_rank1_matrix() {
        let u_vec = [1.0f64, 2.0, 3.0];
        let v_vec = [4.0f64, 5.0, 6.0];
        let a: Vec<Vec<f64>> = (0..3).map(|i|
            (0..3).map(|j| u_vec[i] * v_vec[j]).collect()
        ).collect();
        let s = svdvals(&a);
        let expected_s0 = (u_vec.iter().map(|&v| v * v).sum::<f64>()).sqrt()
                        * (v_vec.iter().map(|&v| v * v).sum::<f64>()).sqrt();
        assert!((s[0] - expected_s0).abs() / expected_s0 < 1e-8,
            "rank-1 s[0]: expected {expected_s0:.6} got {:.6}", s[0]);
        assert!(s[1] < 1e-8, "rank-1 s[1] should be ~0, got {}", s[1]);
        assert!(s[2] < 1e-8, "rank-1 s[2] should be ~0, got {}", s[2]);
        assert_svd_quality("rank-1", &a, 1e-10, 1e-10);
    }

    /// Near-singular matrix.
    #[test]
    fn svd_near_singular() {
        let eps = 1e-10f64;
        let a = vec![vec![1.0f64, 1.0], vec![1.0, 1.0 + eps]];
        let s_ref = nalgebra_singvals(&a);
        let s = svdvals(&a);
        let rel0 = (s_ref[0] - s[0]).abs() / s_ref[0];
        assert!(rel0 < 1e-8, "near-singular s[0] rel={rel0:.2e}");
        let abs1 = (s_ref[1] - s[1]).abs();
        assert!(abs1 < 1e-10,
            "near-singular s[1] abs={abs1:.2e} (ref={:.4e} got={:.4e})", s_ref[1], s[1]);
    }

    /// Rank-1 dominant 12×10 matrix (optshrink scenario).
    #[test]
    fn svd_rank1_dominant_period_matrix() {
        let p = 12usize;
        let nc = 10usize;
        let mut a: Vec<Vec<f64>> = vec![vec![0.0; nc]; p];
        for ph in 0..p {
            for ci in 0..nc {
                let signal = (1.0 + 0.1 * ph as f64) * (10.0 + ci as f64);
                let noise  = 0.3 * ((ph * nc + ci) as f64 * 1.7321 % 1.0 - 0.5);
                a[ph][ci] = signal + noise;
            }
        }
        let s_ref = nalgebra_singvals(&a);
        let s = svdvals(&a);
        let rel0 = (s_ref[0] - s[0]).abs() / s_ref[0];
        assert!(rel0 < 1e-4, "period-matrix s[0] rel={rel0:.2e}");
        for (i, (&sr, &sf)) in s_ref.iter().zip(s.iter()).enumerate() {
            let rel = (sr - sf).abs() / sr.max(1e-6);
            assert!(rel < 0.10,
                "period-matrix s[{i}]: nalgebra={sr:.4} flair={sf:.4} rel={rel:.4}");
        }
        assert_svd_quality("period-matrix", &a, 1e-8, 1e-8);
    }

    /// Tall matrix 50×3 (ridge_sa scenario).
    #[test]
    fn svd_tall_ridge_design() {
        let a: Vec<Vec<f64>> = (0..50).map(|i| {
            let t = i as f64 / 50.0;
            vec![1.0, t, (t * 6.28318).sin()]
        }).collect();
        let s_ref = nalgebra_singvals(&a);
        let s = svdvals(&a);
        for (i, (&sr, &sf)) in s_ref.iter().zip(s.iter()).enumerate() {
            let rel = (sr - sf).abs() / sr.max(1e-10);
            assert!(rel < 1e-6,
                "tall-ridge s[{i}]: nalgebra={sr:.6} flair={sf:.6} rel={rel:.2e}");
        }
        assert_svd_quality("tall-ridge", &a, 1e-10, 1e-10);
    }

    /// Full reconstruction vs nalgebra: verifies (U, s, Vt) reconstruction and orthogonality.
    #[test]
    fn svd_full_reconstruction_vs_nalgebra() {
        let cases: &[(&str, Vec<Vec<f64>>)] = &[
            ("3x2", vec![vec![1.0f64,2.0],vec![3.0,4.0],vec![5.0,6.0]]),
            ("4x3", vec![vec![1.0f64,2.0,3.0],vec![4.0,5.0,6.0],
                         vec![7.0,8.0,9.0],vec![10.0,11.0,12.0]]),
            ("5x5-rand", vec![vec![2.0f64,1.0,0.5,0.25,0.1],
                              vec![1.0,3.0,1.0,0.5, 0.2],
                              vec![0.5,1.0,4.0,1.0, 0.3],
                              vec![0.25,0.5,1.0,5.0,0.4],
                              vec![0.1,0.2,0.3,0.4,6.0]]),
        ];
        for (label, a) in cases {
            let s_ref = nalgebra_singvals(a);
            let s = svdvals(a);
            for (i, (&sr, &sf)) in s_ref.iter().zip(s.iter()).enumerate() {
                let rel = (sr - sf).abs() / sr.max(1e-12);
                assert!(rel < 1e-8,
                    "{label} s[{i}]: nalgebra={sr:.8e} flair={sf:.8e} rel={rel:.2e}");
            }
            assert_svd_quality(label, a, 1e-10, 1e-10);
        }
    }
}