gam 0.3.115

Generalized penalized likelihood engine
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
//! #1011 — deterministic two-sided enclosures for a block-SPD log-determinant.
//!
//! For the bordered-arrow evidence at frontier atom counts, the dense border
//! Schur factor is the scaling wall. This module computes CERTIFIED bounds
//! `lower ≤ log|S| − log|D| ≤ upper` from exact moments of the
//! block-preconditioned residual, with no randomness and no estimator
//! variance — an enclosure, refinable until it is tighter than the consuming
//! decision's margin (topology-race Δ, EFS step tolerance, …).
//!
//! Math (derivation on issue #1011): with `D = blockdiag(S_11..S_KK)`,
//! `S_ii = L_i L_iᵀ`, and `E = D^{-1/2}(S − D)D^{-1/2}`:
//! * `I + E = D^{-1/2} S D^{-1/2} ≻ 0` ⇒ every eigenvalue `λ_a > −1`;
//! * `E` has ZERO diagonal blocks ⇒ `tr E = Σ λ_a = 0`;
//! * `p₂ = tr E² = Σ_{i≠j} ‖Ẽ_ij‖_F²` and
//!   `p₃ = tr E³ = Σ_{i≠j≠k≠i} tr(Ẽ_ij Ẽ_jk Ẽ_ki)` are EXACT block
//!   contractions of `Ẽ_ij = L_i⁻¹ S_ij L_j⁻ᵀ` — never forming `E` densely;
//! * a spectral-radius certificate `ρ = min(√p₂, max_i Σ_{j≠i} ‖Ẽ_ij‖_F)`
//!   (block Gershgorin via `‖·‖₂ ≤ ‖·‖_F`), required `< 1`.
//!
//! Per-eigenvalue inequalities, valid for ALL `λ > −1` (alternating-series
//! remainder for `λ ≥ 0`, monotone tail for `λ < 0`):
//! `log(1+λ) ≤ λ − λ²/2 + λ³/3`, and on `[−ρ, ρ]` the cubic remainder obeys
//! `R(λ) ≥ −ρ²λ²/(4(1−ρ))`. Summing with `Σλ = 0`:
//!
//! ```text
//! order 3:  upper = −p₂/2 + p₃/3
//!           lower = upper − ρ²·p₂ / (4(1−ρ))
//! order 2:  upper = −p₂/2 + ρ·p₂/3          (λ³ ≤ ρλ² for λ≥0; λ³<0≤ρλ² else)
//!           lower = −p₂/2 − ρ·p₂ / (3(1−ρ))
//! ```
//!
//! The gap scales as `ρ·p₂` (order 2) / `ρ²·p₂` (order 3): preconditioner
//! quality drives certainty, and absorbing the worst off-diagonal pair into
//! `D` is the refinement step when the gap is too wide. `ρ ≥ 1` is an
//! explicit refusal (`Err`), never a silent fallback.

use faer::Side;
use ndarray::Array2;

use crate::linalg::faer_ndarray::FaerCholesky;
use crate::linalg::triangular::forward_substitution_lower_matrix;

/// A certified enclosure of `log|S|` for a block-partitioned SPD matrix.
#[derive(Debug, Clone)]
pub struct LogdetEnclosure {
    /// Exact `log|D| = Σ_i log|S_ii|` from the per-block Cholesky factors.
    pub block_diag_logdet: f64,
    /// Certified lower bound on `log|S|` (i.e. `block_diag_logdet + correction_lower`).
    pub lower: f64,
    /// Certified upper bound on `log|S|`.
    pub upper: f64,
    /// The spectral-radius certificate used (`< 1` or this struct would not exist).
    pub rho: f64,
    /// Exact second moment `tr(E²)`.
    pub p2: f64,
    /// Exact third moment `tr(E³)` when the order-3 enclosure was requested.
    pub p3: Option<f64>,
}

impl LogdetEnclosure {
    /// Width of the enclosure — compare against the consuming decision's margin.
    pub fn gap(&self) -> f64 {
        self.upper - self.lower
    }

    /// The enclosure's certified point value when (and only when) the gap is
    /// already below the consuming decision's margin. Below margin a single
    /// `f64` is a lie — the caller must escalate — so this returns the explicit
    /// [`MarginVerdict`] rather than ever fabricating one.
    ///
    /// `decision_margin` is the smallest spread in the consumer's verdict that
    /// matters: the topology-race candidate gap Δ, the EFS step tolerance, an
    /// Armijo slack. A `Decided` value is the enclosure midpoint, which is
    /// within `gap/2 ≤ margin/2` of the truth — tighter than the decision can
    /// resolve, so the verdict is identical to the one the exact logdet would
    /// have produced.
    /// Whether a bare enclosure `gap` is resolved more tightly than a consumer's
    /// `decision_margin` — the predicate behind [`Self::decide_within_margin`],
    /// exposed for consumers that hold only the gap (e.g. the EFS engine, which
    /// receives the cost's enclosure width through `EfsEval`).
    pub fn gap_resolves_margin(gap: f64, decision_margin: f64) -> bool {
        decision_margin.is_finite()
            && decision_margin > 0.0
            && gap.is_finite()
            && gap < decision_margin
    }

    pub fn decide_within_margin(&self, decision_margin: f64) -> MarginVerdict {
        let gap = self.gap();
        if decision_margin.is_finite() && decision_margin > 0.0 && gap < decision_margin {
            MarginVerdict::Decided {
                value: 0.5 * (self.lower + self.upper),
                gap,
                decision_margin,
            }
        } else {
            MarginVerdict::InsufficientMargin {
                gap,
                decision_margin,
            }
        }
    }
}

/// The shared decision-margin contract between an enclosure-valued quantity and
/// its consumer (the topology race, the EFS outer step, the coreset race
/// transfer — all declare a margin and inherit this verdict).
///
/// `Decided` means the enclosure is strictly tighter than the consumer's
/// decision margin, so its midpoint is interchangeable with the exact value for
/// that decision. `InsufficientMargin` is the honesty escalation: the consumer
/// must refine (more moments, pair absorption, a larger coreset) or fall back
/// to the exact dense path — never decide on a point value that the enclosure
/// does not actually pin down.
#[derive(Debug, Clone, Copy, PartialEq)]
pub enum MarginVerdict {
    Decided {
        value: f64,
        gap: f64,
        decision_margin: f64,
    },
    InsufficientMargin {
        gap: f64,
        decision_margin: f64,
    },
}

impl MarginVerdict {
    /// The certified point value when the margin closed, else `None` — the
    /// caller that pattern-matches `None` must escalate to the exact path.
    pub fn decided_value(&self) -> Option<f64> {
        match self {
            MarginVerdict::Decided { value, .. } => Some(*value),
            MarginVerdict::InsufficientMargin { .. } => None,
        }
    }

    pub fn is_decided(&self) -> bool {
        matches!(self, MarginVerdict::Decided { .. })
    }
}

/// Refine a block-preconditioned log-determinant enclosure until its gap falls
/// below `decision_margin`, climbing the deterministic ladder:
///
/// 1. order-2 moments (`tr E²` only) — gap `∝ ρ·p₂`;
/// 2. order-3 moments (`+ tr E³`)   — gap `∝ ρ²·p₂`;
/// 3. **pair absorption**: merge the strongest off-diagonal pair `(i, j)` (the
///    largest `‖Ẽ_ij‖_F`, the term that dominates `ρ`) into a single joint
///    diagonal block, shrinking the residual `E` and `ρ`, then recurse.
///
/// Each rung is an EXACT enclosure (it always contains the truth); the ladder
/// only tightens the gap. When absorption can no longer split a pair (the whole
/// matrix has collapsed to one block, i.e. the exact dense logdet) or the
/// `max_absorptions` budget is spent without closing the margin, the result is
/// [`MarginVerdict::InsufficientMargin`] — the consumer then takes the exact
/// dense path. `decision_margin` is the consumer's declared margin.
pub fn refine_logdet_enclosure_to_margin(
    diag: &[Array2<f64>],
    off: &[(usize, usize, Array2<f64>)],
    decision_margin: f64,
    max_absorptions: usize,
) -> Result<(LogdetEnclosure, MarginVerdict), String> {
    // Rung 1 & 2: moments at the current partition.
    let order2 = block_preconditioned_logdet_enclosure(diag, off, false)?;
    if order2.decide_within_margin(decision_margin).is_decided() {
        let verdict = order2.decide_within_margin(decision_margin);
        return Ok((order2, verdict));
    }
    let order3 = block_preconditioned_logdet_enclosure(diag, off, true)?;
    let verdict3 = order3.decide_within_margin(decision_margin);
    if verdict3.is_decided() {
        return Ok((order3, verdict3));
    }

    // Rung 3: absorb the strongest off-diagonal pair and recurse. Each call to
    // `block_preconditioned_logdet_enclosure` is independent and exact, so the
    // recursion only ever sharpens; the worst it can do is reach the single
    // dense block (the exact logdet, gap 0) when the budget allows.
    if max_absorptions == 0 || off.is_empty() {
        return Ok((order3, verdict3));
    }
    let (merged_diag, merged_off) = absorb_strongest_pair(diag, off)?;
    // Absorption strictly reduces the block count, so the recursion terminates.
    refine_logdet_enclosure_to_margin(
        &merged_diag,
        &merged_off,
        decision_margin,
        max_absorptions - 1,
    )
}

/// Merge the two atoms joined by the strongest off-diagonal block (largest
/// `‖S_ij‖_F` in the *whitened* metric, the term that dominates the spectral
/// radius) into a single joint diagonal block, returning the smaller partition.
///
/// The new joint block stacks the two diagonals and the coupling on its
/// off-diagonal; every other block is re-indexed into the contracted layout.
/// The dense matrix `S` is unchanged — only the *partition* coarsens — so the
/// resulting enclosure is over the same `log|S|`, just with a smaller residual
/// `E` (the absorbed pair contributes 0 to the new `E`).
fn absorb_strongest_pair(
    diag: &[Array2<f64>],
    off: &[(usize, usize, Array2<f64>)],
) -> Result<(Vec<Array2<f64>>, Vec<(usize, usize, Array2<f64>)>), String> {
    let k = diag.len();
    // Pick the pair with the largest whitened Frobenius coupling. We whiten by
    // the per-block Cholesky factors so the ranking matches the contribution to
    // ρ, not the raw (un-preconditioned) block norm.
    let mut lowers: Vec<Array2<f64>> = Vec::with_capacity(k);
    for (i, s_ii) in diag.iter().enumerate() {
        let factor = s_ii
            .cholesky(Side::Lower)
            .map_err(|e| format!("absorb_strongest_pair: block {i} is not SPD: {e:?}"))?;
        lowers.push(factor.lower_triangular());
    }
    let mut best: Option<(usize, f64)> = None; // (index into off, whitened ‖·‖_F)
    for (idx, (i, j, s_ij)) in off.iter().enumerate() {
        let e_ij = whitened_off_block(&lowers[*i], &lowers[*j], s_ij);
        let f = frobenius_sq(&e_ij).sqrt();
        match best {
            None => best = Some((idx, f)),
            Some((_, bf)) if f > bf => best = Some((idx, f)),
            _ => {}
        }
    }
    let (best_idx, _) =
        best.ok_or_else(|| "absorb_strongest_pair: no off-diagonal pair to absorb".to_string())?;
    let (a, b, _) = &off[best_idx];
    let (a, b) = (*a.min(b), *a.max(b));
    let (ma, mb) = (diag[a].nrows(), diag[b].nrows());
    let merged_dim = ma + mb;

    // Assemble the joint diagonal block [[S_aa, S_ab], [S_abᵀ, S_bb]].
    let mut joint = Array2::<f64>::zeros((merged_dim, merged_dim));
    joint.slice_mut(ndarray::s![..ma, ..ma]).assign(&diag[a]);
    joint.slice_mut(ndarray::s![ma.., ma..]).assign(&diag[b]);
    if let Some((_, _, s_ab)) = off.iter().find(|(i, j, _)| (*i, *j) == (a, b)) {
        joint.slice_mut(ndarray::s![..ma, ma..]).assign(s_ab);
        let s_ba = s_ab.t().to_owned();
        joint.slice_mut(ndarray::s![ma.., ..ma]).assign(&s_ba);
    }

    // New block ordering: the joint block takes index 0, every other old block
    // `g ∉ {a, b}` keeps its relative order at a shifted index.
    let mut old_to_new = vec![0usize; k];
    let mut new_diag: Vec<Array2<f64>> = Vec::with_capacity(k - 1);
    new_diag.push(joint);
    let mut next = 1usize;
    for g in 0..k {
        if g == a || g == b {
            old_to_new[g] = 0;
        } else {
            old_to_new[g] = next;
            new_diag.push(diag[g].clone());
            next += 1;
        }
    }

    // Re-index the surviving off-diagonal blocks. A block touching `a` or `b`
    // becomes a (joint, g) coupling; we must place its rows into the correct
    // half of the joint block's row span. Blocks between two survivors carry
    // over unchanged (up to re-index). The absorbed (a, b) pair is dropped — it
    // now lives inside the joint diagonal block.
    let mut new_off: Vec<(usize, usize, Array2<f64>)> = Vec::new();
    // Accumulate joint↔survivor couplings keyed by the survivor's NEW index, so
    // the `a`- and `b`-halves of a survivor that couples to BOTH are stacked
    // into one (merged_dim × m_g) block.
    use std::collections::BTreeMap;
    let mut joint_couplings: BTreeMap<usize, Array2<f64>> = BTreeMap::new();
    for (i, j, s_ij) in off {
        let (i, j) = (*i, *j);
        if (i, j) == (a, b) {
            continue; // absorbed into the joint diagonal
        }
        let touches_a = i == a || j == a;
        let touches_b = i == b || j == b;
        if touches_a || touches_b {
            // Identify the survivor `g` and the oriented block S_{joint-half, g}.
            let g = if i == a || i == b { j } else { i };
            let mg = diag[g].nrows();
            let entry = joint_couplings
                .entry(old_to_new[g])
                .or_insert_with(|| Array2::<f64>::zeros((merged_dim, mg)));
            // Orient the stored block so its rows index the absorbed atom and
            // columns index `g` (S_{absorbed, g}); transpose when the stored
            // upper-triangle entry has `g` first.
            let s_half = if i == g {
                s_ij.t().to_owned()
            } else {
                s_ij.clone()
            };
            let row_off = if i == a || j == a { 0 } else { ma };
            entry
                .slice_mut(ndarray::s![row_off..row_off + s_half.nrows(), ..])
                .assign(&s_half);
        } else {
            // Survivor–survivor block: re-index, preserving i<j orientation.
            let (ni, nj) = (old_to_new[i], old_to_new[j]);
            let (ni, nj, block) = if ni < nj {
                (ni, nj, s_ij.clone())
            } else {
                (nj, ni, s_ij.t().to_owned())
            };
            new_off.push((ni, nj, block));
        }
    }
    for (g_new, block) in joint_couplings {
        // joint index is 0 < g_new always.
        new_off.push((0, g_new, block));
    }
    Ok((new_diag, new_off))
}

/// `Ẽ_ij = L_i⁻¹ · S_ij · L_j⁻ᵀ`: forward-solve on the left, then on the
/// right via the transpose identity `(X L_j⁻ᵀ)ᵀ = L_j⁻¹ Xᵀ`.
fn whitened_off_block(l_i: &Array2<f64>, l_j: &Array2<f64>, s_ij: &Array2<f64>) -> Array2<f64> {
    let x = forward_substitution_lower_matrix(l_i, s_ij);
    let xt = x.t().to_owned();
    forward_substitution_lower_matrix(l_j, &xt).t().to_owned()
}

fn frobenius_sq(a: &Array2<f64>) -> f64 {
    a.iter().map(|v| v * v).sum()
}

/// Certified two-sided enclosure of `log|S|` for a block-SPD matrix given as
/// per-atom diagonal blocks plus upper-triangle off-diagonal blocks
/// (`off[(i, j)]` with `i < j`; `S_ji = S_ijᵀ` by symmetry). Pass
/// `use_third_moment = true` for the order-3 enclosure (extra `O(triples)`
/// work, gap `∝ ρ²p₂` instead of `∝ ρp₂`).
///
/// Errors when a diagonal block is not SPD or when the spectral-radius
/// certificate fails (`ρ ≥ 1`): the caller must refine the partition
/// (absorb the offending pair into one joint diagonal block) — the bound
/// machinery never silently degrades.
pub fn block_preconditioned_logdet_enclosure(
    diag: &[Array2<f64>],
    off: &[(usize, usize, Array2<f64>)],
    use_third_moment: bool,
) -> Result<LogdetEnclosure, String> {
    let k = diag.len();
    if k == 0 {
        return Err("block_preconditioned_logdet_enclosure: no diagonal blocks".to_string());
    }
    // Exact per-block factors and log|D|.
    let mut lowers: Vec<Array2<f64>> = Vec::with_capacity(k);
    let mut block_diag_logdet = 0.0_f64;
    for (i, s_ii) in diag.iter().enumerate() {
        if s_ii.nrows() != s_ii.ncols() {
            return Err(format!(
                "block_preconditioned_logdet_enclosure: diagonal block {i} is not square"
            ));
        }
        let factor = s_ii.cholesky(Side::Lower).map_err(|e| {
            format!("block_preconditioned_logdet_enclosure: block {i} is not SPD: {e:?}")
        })?;
        let l = factor.lower_triangular();
        for d in 0..l.nrows() {
            block_diag_logdet += 2.0 * l[[d, d]].ln();
        }
        lowers.push(l);
    }
    // Whitened off-diagonal blocks (upper triangle), p₂, and block-Gershgorin
    // row sums.
    let mut whitened: Vec<(usize, usize, Array2<f64>)> = Vec::with_capacity(off.len());
    let mut p2 = 0.0_f64;
    let mut row_sums = vec![0.0_f64; k];
    for (i, j, s_ij) in off {
        let (i, j) = (*i, *j);
        if i >= j || j >= k {
            return Err(format!(
                "block_preconditioned_logdet_enclosure: off-block ({i},{j}) must satisfy i<j<K={k}"
            ));
        }
        if s_ij.nrows() != lowers[i].nrows() || s_ij.ncols() != lowers[j].nrows() {
            return Err(format!(
                "block_preconditioned_logdet_enclosure: off-block ({i},{j}) shape mismatch"
            ));
        }
        let e_ij = whitened_off_block(&lowers[i], &lowers[j], s_ij);
        let f2 = frobenius_sq(&e_ij);
        // E_ij and its transpose E_ji contribute equally to tr(E²).
        p2 += 2.0 * f2;
        let f = f2.sqrt();
        row_sums[i] += f;
        row_sums[j] += f;
        whitened.push((i, j, e_ij));
    }
    let gershgorin = row_sums.iter().fold(0.0_f64, |a, &b| a.max(b));
    let rho = p2.sqrt().min(gershgorin);
    if !(rho < 1.0) {
        return Err(format!(
            "block_preconditioned_logdet_enclosure: spectral-radius certificate failed \
             (ρ = {rho:.6} ≥ 1); refine the block partition (absorb the strongest \
             off-diagonal pair into the preconditioner) and retry"
        ));
    }

    // Optional exact third moment over ordered distinct triples. Lookup
    // returns the (a,b) whitened block, transposing the stored upper-triangle
    // entry when needed.
    let p3 = if use_third_moment {
        let get = |a: usize, b: usize| -> Option<Array2<f64>> {
            for (i, j, e) in &whitened {
                if *i == a && *j == b {
                    return Some(e.clone());
                }
                if *i == b && *j == a {
                    return Some(e.t().to_owned());
                }
            }
            None
        };
        let mut acc = 0.0_f64;
        for a in 0..k {
            for b in 0..k {
                if b == a {
                    continue;
                }
                let Some(e_ab) = get(a, b) else { continue };
                for c in 0..k {
                    if c == a || c == b {
                        continue;
                    }
                    let (Some(e_bc), Some(e_ca)) = (get(b, c), get(c, a)) else {
                        continue;
                    };
                    // tr(E_ab · E_bc · E_ca)
                    let prod = e_ab.dot(&e_bc);
                    for r in 0..prod.nrows() {
                        for s in 0..prod.ncols() {
                            acc += prod[[r, s]] * e_ca[[s, r]];
                        }
                    }
                }
            }
        }
        Some(acc)
    } else {
        None
    };

    let (corr_lower, corr_upper) = match p3 {
        Some(p3) => {
            let upper = -p2 / 2.0 + p3 / 3.0;
            let lower = upper - rho * rho * p2 / (4.0 * (1.0 - rho));
            (lower, upper)
        }
        None => {
            let upper = -p2 / 2.0 + rho * p2 / 3.0;
            let lower = -p2 / 2.0 - rho * p2 / (3.0 * (1.0 - rho));
            (lower, upper)
        }
    };
    Ok(LogdetEnclosure {
        block_diag_logdet,
        lower: block_diag_logdet + corr_lower,
        upper: block_diag_logdet + corr_upper,
        rho,
        p2,
        p3,
    })
}

#[cfg(test)]
mod tests {
    use super::*;
    use ndarray::Array2;

    /// Deterministic block-SPD fixture: strong SPD diagonal blocks, weak
    /// off-diagonal coupling (so the certificate ρ < 1 holds), assembled
    /// densely for the oracle.
    fn fixture(
        k: usize,
        m: usize,
        coupling: f64,
    ) -> (
        Vec<Array2<f64>>,
        Vec<(usize, usize, Array2<f64>)>,
        Array2<f64>,
    ) {
        let dim = k * m;
        let mut dense = Array2::<f64>::zeros((dim, dim));
        let mut diag = Vec::new();
        let mut off = Vec::new();
        for i in 0..k {
            let mut d = Array2::<f64>::zeros((m, m));
            for r in 0..m {
                for c in 0..m {
                    let v = if r == c {
                        3.0 + 0.4 * (i as f64) + 0.2 * (r as f64)
                    } else {
                        0.3 * ((r + 2 * c + i) as f64 * 0.7).sin()
                    };
                    d[[r, c]] = v;
                }
            }
            // Symmetrize and make diagonally dominant ⇒ SPD.
            let mut sym = Array2::<f64>::zeros((m, m));
            for r in 0..m {
                for c in 0..m {
                    sym[[r, c]] = 0.5 * (d[[r, c]] + d[[c, r]]);
                }
                sym[[r, r]] += 1.0;
            }
            for r in 0..m {
                for c in 0..m {
                    dense[[i * m + r, i * m + c]] = sym[[r, c]];
                }
            }
            diag.push(sym);
        }
        for i in 0..k {
            for j in (i + 1)..k {
                let mut b = Array2::<f64>::zeros((m, m));
                for r in 0..m {
                    for c in 0..m {
                        b[[r, c]] =
                            coupling * ((r as f64) - (c as f64) + (i + j) as f64 * 0.31).cos();
                    }
                }
                for r in 0..m {
                    for c in 0..m {
                        dense[[i * m + r, j * m + c]] = b[[r, c]];
                        dense[[j * m + c, i * m + r]] = b[[r, c]];
                    }
                }
                off.push((i, j, b));
            }
        }
        (diag, off, dense)
    }

    fn dense_logdet(s: &Array2<f64>) -> f64 {
        let l = s
            .cholesky(Side::Lower)
            .expect("oracle fixture must be SPD")
            .lower_triangular();
        (0..l.nrows()).map(|d| 2.0 * l[[d, d]].ln()).sum()
    }

    /// Containment: the enclosure must contain the dense truth at both
    /// orders, and the order-3 gap must not exceed the order-2 gap.
    #[test]
    fn enclosure_contains_dense_truth_and_order3_tightens() {
        let (diag, off, dense) = fixture(4, 3, 0.08);
        let truth = dense_logdet(&dense);
        let e2 =
            block_preconditioned_logdet_enclosure(&diag, &off, false).expect("order-2 enclosure");
        let e3 =
            block_preconditioned_logdet_enclosure(&diag, &off, true).expect("order-3 enclosure");
        assert!(
            e2.lower <= truth && truth <= e2.upper,
            "order-2 enclosure [{}, {}] must contain dense log|S| = {}",
            e2.lower,
            e2.upper,
            truth
        );
        assert!(
            e3.lower <= truth && truth <= e3.upper,
            "order-3 enclosure [{}, {}] must contain dense log|S| = {}",
            e3.lower,
            e3.upper,
            truth
        );
        assert!(
            e3.gap() <= e2.gap() + 1e-12,
            "order-3 gap {} must not exceed order-2 gap {}",
            e3.gap(),
            e2.gap()
        );
        // The enclosure is non-vacuous: the correction is genuinely bounded
        // away from the trivial ±∞ and the gap shrinks with ρ²p₂.
        assert!(e3.gap() < 0.5 * e2.gap() + 1e-9 || e2.gap() < 1e-9);
    }

    /// The block-diagonal case is exact: zero coupling ⇒ enclosure collapses
    /// to the exact log|D| at width 0.
    #[test]
    fn zero_coupling_is_exact() {
        let (diag, _off, dense) = fixture(3, 2, 0.0);
        let truth = dense_logdet(&dense);
        let e = block_preconditioned_logdet_enclosure(&diag, &[], true).expect("enclosure");
        assert!((e.lower - truth).abs() < 1e-10 && (e.upper - truth).abs() < 1e-10);
        assert!(e.gap() < 1e-12);
    }

    /// Strong coupling must REFUSE (ρ ≥ 1), never emit a wrong enclosure.
    #[test]
    fn failed_radius_certificate_refuses() {
        let (diag, off, _dense) = fixture(3, 2, 5.0);
        let err = block_preconditioned_logdet_enclosure(&diag, &off, false)
            .expect_err("ρ ≥ 1 must refuse");
        assert!(err.contains("spectral-radius certificate failed"));
    }

    /// The margin verdict refuses to fabricate a point value when the gap
    /// exceeds the declared decision margin, and decides (with a midpoint that
    /// the dense truth brackets) when it is below.
    #[test]
    fn margin_verdict_is_honest_about_the_gap() {
        let (diag, off, dense) = fixture(4, 3, 0.08);
        let truth = dense_logdet(&dense);
        let e = block_preconditioned_logdet_enclosure(&diag, &off, true).expect("enclosure");
        // A margin tighter than the gap must escalate, never decide.
        let tight = e.gap() * 0.5;
        assert!(!e.decide_within_margin(tight).is_decided());
        assert!(e.decide_within_margin(tight).decided_value().is_none());
        // A margin wider than the gap decides, and the midpoint is within
        // gap/2 of the truth — i.e. interchangeable for that decision.
        let loose = e.gap() * 2.0 + 1e-9;
        let verdict = e.decide_within_margin(loose);
        assert!(verdict.is_decided());
        let value = verdict.decided_value().expect("decided");
        assert!((value - truth).abs() <= 0.5 * e.gap() + 1e-12);
    }

    /// Pair absorption preserves the dense `log|S|` it encloses while shrinking
    /// the residual: the absorbed pair drops out of `E`, so the gap can only
    /// tighten, and the enclosure still brackets the truth.
    #[test]
    fn pair_absorption_preserves_truth_and_tightens() {
        let (diag, off, dense) = fixture(4, 3, 0.14);
        let truth = dense_logdet(&dense);
        let before =
            block_preconditioned_logdet_enclosure(&diag, &off, true).expect("pre-absorption");
        let (mdiag, moff) = absorb_strongest_pair(&diag, &off).expect("absorb");
        assert_eq!(
            mdiag.len(),
            diag.len() - 1,
            "one fewer block after absorption"
        );
        let after =
            block_preconditioned_logdet_enclosure(&mdiag, &moff, true).expect("post-absorption");
        assert!(
            after.lower <= truth && truth <= after.upper,
            "absorbed enclosure [{}, {}] must still contain log|S| = {truth}",
            after.lower,
            after.upper
        );
        assert!(
            after.gap() <= before.gap() + 1e-9,
            "absorption must not widen the gap ({} vs {})",
            after.gap(),
            before.gap()
        );
    }

    /// The refinement ladder closes a margin that order-3 alone cannot, by
    /// absorbing pairs — and the decided value brackets the dense truth.
    #[test]
    fn refinement_ladder_closes_margin_via_absorption() {
        let (diag, off, dense) = fixture(5, 2, 0.16);
        let truth = dense_logdet(&dense);
        let order3 =
            block_preconditioned_logdet_enclosure(&diag, &off, true).expect("order-3 enclosure");
        // Choose a margin between the dense-exact (0) and the order-3 gap so the
        // ladder must climb past moments into absorption.
        let margin = order3.gap() * 0.5;
        let (enc, verdict) =
            refine_logdet_enclosure_to_margin(&diag, &off, margin, 8).expect("ladder");
        assert!(
            verdict.is_decided(),
            "ladder must close the margin via absorption"
        );
        assert!(
            enc.lower <= truth && truth <= enc.upper,
            "refined enclosure [{}, {}] must contain log|S| = {truth}",
            enc.lower,
            enc.upper
        );
        assert!(enc.gap() < margin);
    }
}