primal-barracuda 0.4.0

barraCuda — BARrier-free Rust Abstracted Cross-platform Unified Dimensional Algebra. Vendor-agnostic GPU compute via WGSL.
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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Symmetric Eigenvalue Decomposition (f64 CPU)
//!
//! Computes eigenvalues and eigenvectors of a symmetric matrix:
//! A = V·D·Vᵀ where D is diagonal (eigenvalues) and V is orthogonal (eigenvectors).
//!
//! # Algorithm
//!
//! Uses the Jacobi eigenvalue algorithm with Givens rotations.
//! Converges quadratically for distinct eigenvalues.
//!
//! # Applications
//!
//! - Principal Component Analysis (PCA)
//! - Spectral clustering
//! - Vibration analysis
//! - Quantum mechanics (Hamiltonian diagonalization)
//! - hotSpring HFB nuclear physics solver
//!
//! # References
//!
//! - Golub & Van Loan, "Matrix Computations", Section 8.4
//! - scipy.linalg.eigh

use crate::error::{BarracudaError, Result};

/// Result of symmetric eigenvalue decomposition A = V·D·Vᵀ.
#[derive(Debug, Clone)]
pub struct EighDecomposition {
    /// Eigenvalues (n elements, sorted ascending by default)
    pub eigenvalues: Vec<f64>,
    /// Eigenvectors as columns of V (n×n, row-major)
    /// Column j is the eigenvector for eigenvalue j
    pub eigenvectors: Vec<f64>,
    /// Matrix dimension
    pub n: usize,
}

impl EighDecomposition {
    /// Get eigenvector for eigenvalue at index i.
    #[must_use]
    pub fn eigenvector(&self, i: usize) -> Option<Vec<f64>> {
        if i >= self.n {
            return None;
        }
        let mut v = vec![0.0; self.n];
        for (row, v_row) in v.iter_mut().enumerate() {
            *v_row = self.eigenvectors[row * self.n + i];
        }
        Some(v)
    }

    /// Sort eigenvalues and eigenvectors in descending order.
    /// Useful for PCA where largest eigenvalues matter most.
    pub fn sort_descending(&mut self) {
        let n = self.n;
        let mut perm: Vec<usize> = (0..n).collect();
        perm.sort_by(|&a, &b| {
            self.eigenvalues[b]
                .partial_cmp(&self.eigenvalues[a])
                .unwrap_or(std::cmp::Ordering::Equal)
        });

        let new_vals: Vec<f64> = perm.iter().map(|&i| self.eigenvalues[i]).collect();
        let mut new_vecs = vec![0.0; n * n];
        for (new_col, &old_col) in perm.iter().enumerate() {
            for row in 0..n {
                new_vecs[row * n + new_col] = self.eigenvectors[row * n + old_col];
            }
        }

        self.eigenvalues = new_vals;
        self.eigenvectors = new_vecs;
    }

    /// Reconstruct A = V·D·Vᵀ.
    /// Useful for verification.
    #[must_use]
    pub fn reconstruct(&self) -> Vec<f64> {
        let mut a = vec![0.0; self.n * self.n];

        for i in 0..self.n {
            for j in 0..self.n {
                let mut sum = 0.0;
                for k in 0..self.n {
                    // A[i,j] = Σₖ V[i,k] * D[k,k] * V[j,k]
                    let vik = self.eigenvectors[i * self.n + k];
                    let vjk = self.eigenvectors[j * self.n + k];
                    sum = (vik * self.eigenvalues[k]).mul_add(vjk, sum);
                }
                a[i * self.n + j] = sum;
            }
        }

        a
    }

    /// Compute trace (sum of eigenvalues).
    #[must_use]
    pub fn trace(&self) -> f64 {
        self.eigenvalues.iter().sum()
    }

    /// Compute determinant (product of eigenvalues).
    #[must_use]
    pub fn det(&self) -> f64 {
        self.eigenvalues.iter().product()
    }
}

/// Compute eigenvalue decomposition of a symmetric matrix.
///
/// Uses the Jacobi eigenvalue algorithm with Givens rotations.
/// Returns eigenvalues in ascending order by default.
///
/// # Arguments
///
/// * `a` - n×n symmetric matrix (row-major)
/// * `n` - Matrix dimension
///
/// # Returns
///
/// `EighDecomposition` containing eigenvalues and eigenvectors
///
/// # Errors
///
/// Returns error if matrix dimensions are invalid.
///
/// # Example
///
/// ```
/// use barracuda::linalg::eigh_f64;
///
/// // Symmetric 2×2 matrix: [[3, 1], [1, 3]]
/// // Eigenvalues: 2, 4
/// let a = vec![3.0, 1.0, 1.0, 3.0];
/// let eig = eigh_f64(&a, 2)?;
///
/// assert!((eig.eigenvalues[0] - 2.0).abs() < 1e-10);
/// assert!((eig.eigenvalues[1] - 4.0).abs() < 1e-10);
/// # Ok::<(), barracuda::error::BarracudaError>(())
/// ```
///
/// # Notes
///
/// - For best precision, ensure the input matrix is exactly symmetric
/// - The algorithm handles repeated eigenvalues but eigenvectors may be
///   arbitrary within the eigenspace
/// - Complexity: O(n³) per sweep, typically 5-10 sweeps for convergence
pub fn eigh_f64(a: &[f64], n: usize) -> Result<EighDecomposition> {
    if a.len() != n * n {
        return Err(BarracudaError::InvalidInput {
            message: format!("Matrix has {} elements, expected {}×{}", a.len(), n, n),
        });
    }

    if n == 0 {
        return Err(BarracudaError::InvalidInput {
            message: "Matrix dimension must be positive".to_string(),
        });
    }

    // Symmetry guard: non-symmetric input produces silently wrong results.
    // Check upper vs lower triangle within machine epsilon tolerance.
    let sym_tol = 1e-10 * a.iter().fold(0.0_f64, |mx, &v| mx.max(v.abs())).max(1e-15);
    for i in 0..n {
        for j in (i + 1)..n {
            let diff = (a[i * n + j] - a[j * n + i]).abs();
            if diff > sym_tol {
                return Err(BarracudaError::InvalidInput {
                    message: format!(
                        "Matrix is not symmetric: A[{i},{j}]={} vs A[{j},{i}]={} (diff={diff:.2e})",
                        a[i * n + j],
                        a[j * n + i],
                    ),
                });
            }
        }
    }

    // Special case: 1×1
    if n == 1 {
        return Ok(EighDecomposition {
            eigenvalues: vec![a[0]],
            eigenvectors: vec![1.0],
            n: 1,
        });
    }

    // Copy matrix (will be transformed to diagonal)
    let mut d = a.to_vec();

    // Initialize eigenvectors as identity
    let mut v = vec![0.0; n * n];
    for i in 0..n {
        v[i * n + i] = 1.0;
    }

    // Jacobi iteration parameters
    const MAX_SWEEPS: usize = 50;
    const TOL: f64 = 1e-14;

    for _ in 0..MAX_SWEEPS {
        // Find largest off-diagonal element
        let mut max_off = 0.0;
        let mut p = 0;
        let mut q = 0;

        for i in 0..n {
            for j in (i + 1)..n {
                let val = d[i * n + j].abs();
                if val > max_off {
                    max_off = val;
                    p = i;
                    q = j;
                }
            }
        }

        // Convergence check
        if max_off < TOL {
            break;
        }

        // Compute rotation angle
        let app = d[p * n + p];
        let aqq = d[q * n + q];
        let apq = d[p * n + q];

        let (c, s) = if (aqq - app).abs() < TOL {
            // θ = π/4
            let c = std::f64::consts::FRAC_1_SQRT_2;
            let s = c * apq.signum();
            (c, s)
        } else {
            let tau = (aqq - app) / (2.0 * apq);
            let t = if tau >= 0.0 {
                1.0 / (tau + (1.0 + tau * tau).sqrt())
            } else {
                1.0 / (tau - (1.0 + tau * tau).sqrt())
            };
            let c = 1.0 / (1.0 + t * t).sqrt();
            let s = t * c;
            (c, s)
        };

        // Apply Givens rotation to D: D' = Jᵀ D J
        // Update rows p and q
        for k in 0..n {
            if k != p && k != q {
                let dkp = d[k * n + p];
                let dkq = d[k * n + q];
                d[k * n + p] = c.mul_add(dkp, -(s * dkq));
                d[p * n + k] = d[k * n + p];
                d[k * n + q] = s.mul_add(dkp, c * dkq);
                d[q * n + k] = d[k * n + q];
            }
        }

        // Update diagonal elements
        let new_pp = (s * s).mul_add(aqq, (c * c).mul_add(app, -(2.0 * c * s * apq)));
        let new_qq = (c * c).mul_add(aqq, (s * s).mul_add(app, 2.0 * c * s * apq));
        d[p * n + p] = new_pp;
        d[q * n + q] = new_qq;
        d[p * n + q] = 0.0;
        d[q * n + p] = 0.0;

        // Apply rotation to eigenvectors: V' = V J
        for k in 0..n {
            let vkp = v[k * n + p];
            let vkq = v[k * n + q];
            v[k * n + p] = c.mul_add(vkp, -(s * vkq));
            v[k * n + q] = s.mul_add(vkp, c * vkq);
        }
    }

    // Extract eigenvalues from diagonal
    let mut eigenvalues: Vec<f64> = (0..n).map(|i| d[i * n + i]).collect();

    // Sort eigenvalues ascending and reorder eigenvectors
    let mut indexed: Vec<(usize, f64)> = eigenvalues.iter().copied().enumerate().collect();
    indexed.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap_or(std::cmp::Ordering::Equal));

    let sorted_vals: Vec<f64> = indexed.iter().map(|(_, val)| *val).collect();
    let mut sorted_vecs = vec![0.0; n * n];
    for (new_idx, (old_idx, _)) in indexed.iter().enumerate() {
        for row in 0..n {
            sorted_vecs[row * n + new_idx] = v[row * n + old_idx];
        }
    }

    eigenvalues = sorted_vals;

    Ok(EighDecomposition {
        eigenvalues,
        eigenvectors: sorted_vecs,
        n,
    })
}

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

    fn approx_eq(a: f64, b: f64, tol: f64) -> bool {
        (a - b).abs() < tol
    }

    #[test]
    fn test_eigh_2x2_simple() {
        // [[3, 1], [1, 3]] has eigenvalues 2, 4
        let a = vec![3.0, 1.0, 1.0, 3.0];
        let eig = eigh_f64(&a, 2).unwrap();

        assert!(approx_eq(eig.eigenvalues[0], 2.0, 1e-10));
        assert!(approx_eq(eig.eigenvalues[1], 4.0, 1e-10));

        // Verify trace
        assert!(approx_eq(eig.trace(), 6.0, 1e-10));

        // Verify determinant
        assert!(approx_eq(eig.det(), 8.0, 1e-10));
    }

    #[test]
    fn test_eigh_identity() {
        let a = vec![1.0, 0.0, 0.0, 1.0];
        let eig = eigh_f64(&a, 2).unwrap();

        assert!(approx_eq(eig.eigenvalues[0], 1.0, 1e-10));
        assert!(approx_eq(eig.eigenvalues[1], 1.0, 1e-10));
    }

    #[test]
    fn test_eigh_diagonal() {
        // Already diagonal: eigenvalues are diagonal elements
        let a = vec![5.0, 0.0, 0.0, 3.0];
        let eig = eigh_f64(&a, 2).unwrap();

        // Sorted ascending: 3, 5
        assert!(approx_eq(eig.eigenvalues[0], 3.0, 1e-10));
        assert!(approx_eq(eig.eigenvalues[1], 5.0, 1e-10));
    }

    #[test]
    fn test_eigh_3x3() {
        // Tridiagonal matrix with known eigenvalues
        // [[2, -1, 0], [-1, 2, -1], [0, -1, 2]]
        // Eigenvalues: 2 - √2, 2, 2 + √2
        let a = vec![2.0, -1.0, 0.0, -1.0, 2.0, -1.0, 0.0, -1.0, 2.0];
        let eig = eigh_f64(&a, 3).unwrap();

        let expected = [
            2.0 - std::f64::consts::SQRT_2,
            2.0,
            2.0 + std::f64::consts::SQRT_2,
        ];

        for (i, &exp) in expected.iter().enumerate() {
            assert!(
                approx_eq(eig.eigenvalues[i], exp, 1e-10),
                "eigenvalue[{}] = {}, expected {}",
                i,
                eig.eigenvalues[i],
                exp
            );
        }

        // Verify trace = 6
        assert!(approx_eq(eig.trace(), 6.0, 1e-10));
    }

    #[test]
    fn test_eigh_reconstruction() {
        let a = vec![4.0, 2.0, 2.0, 3.0];
        let eig = eigh_f64(&a, 2).unwrap();

        let recon = eig.reconstruct();

        for i in 0..4 {
            assert!(
                approx_eq(recon[i], a[i], 1e-10),
                "recon[{}] = {}, a = {}",
                i,
                recon[i],
                a[i]
            );
        }
    }

    #[test]
    fn test_eigh_orthogonality() {
        let a = vec![4.0, 2.0, 2.0, 3.0];
        let eig = eigh_f64(&a, 2).unwrap();

        // V^T V should be identity
        let mut vtv = [0.0; 4];
        for i in 0..2 {
            for j in 0..2 {
                for k in 0..2 {
                    vtv[i * 2 + j] = eig.eigenvectors[k * 2 + i]
                        .mul_add(eig.eigenvectors[k * 2 + j], vtv[i * 2 + j]);
                }
            }
        }

        // Check identity
        assert!(approx_eq(vtv[0], 1.0, 1e-10));
        assert!(approx_eq(vtv[1], 0.0, 1e-10));
        assert!(approx_eq(vtv[2], 0.0, 1e-10));
        assert!(approx_eq(vtv[3], 1.0, 1e-10));
    }

    #[test]
    fn test_eigh_sort_descending() {
        let a = vec![1.0, 0.0, 0.0, 5.0];
        let mut eig = eigh_f64(&a, 2).unwrap();

        // Initially ascending: 1, 5
        assert!(approx_eq(eig.eigenvalues[0], 1.0, 1e-10));
        assert!(approx_eq(eig.eigenvalues[1], 5.0, 1e-10));

        eig.sort_descending();

        // Now descending: 5, 1
        assert!(approx_eq(eig.eigenvalues[0], 5.0, 1e-10));
        assert!(approx_eq(eig.eigenvalues[1], 1.0, 1e-10));
    }

    #[test]
    fn test_eigh_eigenvector() {
        let a = vec![3.0, 1.0, 1.0, 3.0];
        let eig = eigh_f64(&a, 2).unwrap();

        let v0 = eig.eigenvector(0).unwrap();
        let v1 = eig.eigenvector(1).unwrap();

        // v0 and v1 should be orthogonal
        let dot: f64 = v0.iter().zip(v1.iter()).map(|(a, b)| a * b).sum();
        assert!(approx_eq(dot, 0.0, 1e-10));

        // v0 and v1 should be unit vectors
        let norm0: f64 = v0.iter().map(|x| x * x).sum::<f64>().sqrt();
        let norm1: f64 = v1.iter().map(|x| x * x).sum::<f64>().sqrt();
        assert!(approx_eq(norm0, 1.0, 1e-10));
        assert!(approx_eq(norm1, 1.0, 1e-10));
    }

    #[test]
    fn test_eigh_1x1() {
        let a = vec![7.0];
        let eig = eigh_f64(&a, 1).unwrap();

        assert_eq!(eig.eigenvalues.len(), 1);
        assert!(approx_eq(eig.eigenvalues[0], 7.0, 1e-10));
    }

    #[test]
    fn test_eigh_negative_eigenvalues() {
        // [[0, 1], [1, 0]] has eigenvalues -1, 1
        let a = vec![0.0, 1.0, 1.0, 0.0];
        let eig = eigh_f64(&a, 2).unwrap();

        assert!(approx_eq(eig.eigenvalues[0], -1.0, 1e-10));
        assert!(approx_eq(eig.eigenvalues[1], 1.0, 1e-10));
    }

    #[test]
    fn test_eigh_errors() {
        assert!(eigh_f64(&[], 0).is_err());
        assert!(eigh_f64(&[1.0, 2.0, 3.0], 2).is_err());
    }
}