numeris 0.2.0

Pure-Rust numerical algorithms library — no-std compatible
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
# numeris

Pure-Rust numerical algorithms library, no-std compatible. Similar in scope to SciPy, suitable for embedded targets (no heap allocation, no FPU assumptions) while being highly performant on desktop/server hardware via SIMD intrinsics.

**Alpha software** — APIs are unstable and may change without notice.

## Features

- **Fixed-size matrices** — stack-allocated, const-generic `Matrix<T, M, N>`
- **Dynamic matrices** — heap-allocated `DynMatrix<T>` with runtime dimensions (optional `alloc` feature)
- **Linear algebra** — LU, Cholesky, QR, SVD decompositions; symmetric eigendecomposition; real Schur decomposition
- **ODE integration** — fixed-step RK4, 7 adaptive Runge-Kutta solvers with dense output, RODAS4 stiff solver
- **Optimization** — root finding (Brent, Newton), BFGS minimization, Gauss-Newton and Levenberg-Marquardt least squares
- **Complex number support** — all decompositions work with `Complex<f32>` / `Complex<f64>` (optional feature)
- **Quaternions** — unit quaternion rotations, SLERP, Euler angles, rotation matrices
- **Norms** — L1, L2, Frobenius, infinity, one norms
- **SIMD acceleration** — NEON (aarch64) and SSE2 (x86_64) intrinsics for matmul, dot products, and element-wise ops; zero-cost scalar fallback for integers and unsupported architectures
- **No-std / embedded** — runs without `std` or heap; float math falls back to software `libm`

## Quick start

```toml
[dependencies]
numeris = "0.1"
```

```rust
use numeris::{Matrix, Vector};

// Solve a linear system Ax = b
let a = Matrix::new([
    [2.0_f64, 1.0, -1.0],
    [-3.0, -1.0, 2.0],
    [-2.0, 1.0, 2.0],
]);
let b = Vector::from_array([8.0, -11.0, -3.0]);
let x = a.solve(&b).unwrap(); // x = [2, 3, -1]

// Cholesky decomposition of a symmetric positive-definite matrix
let spd = Matrix::new([[4.0, 2.0], [2.0, 3.0]]);
let chol = spd.cholesky().unwrap();
let inv = chol.inverse();

// QR decomposition and least-squares
let a = Matrix::new([
    [1.0_f64, 0.0],
    [1.0, 1.0],
    [1.0, 2.0],
]);
let b = Vector::from_array([1.0, 2.0, 4.0]);
let x = a.qr().unwrap().solve(&b); // least-squares fit

// Symmetric eigendecomposition
let sym = Matrix::new([[4.0_f64, 1.0], [1.0, 3.0]]);
let eig = sym.eig_symmetric().unwrap();
let eigenvalues = eig.eigenvalues();   // sorted ascending
let eigenvectors = eig.eigenvectors(); // columns = eigenvectors

// General eigenvalues via Schur decomposition
let a = Matrix::new([[0.0_f64, -1.0], [1.0, 0.0]]); // 90° rotation
let (re, im) = a.eigenvalues().unwrap(); // eigenvalues ±i

// Quaternion rotation
use numeris::Quaternion;
let q = Quaternion::from_axis_angle(
    &Vector::from_array([0.0, 0.0, 1.0]),
    std::f64::consts::FRAC_PI_2,
);
let v = Vector::from_array([1.0, 0.0, 0.0]);
let rotated = q * v; // [0, 1, 0]
```

## Dynamic matrices

When dimensions aren't known at compile time, use `DynMatrix` (requires `alloc`, included with default `std` feature):

```rust
use numeris::{DynMatrix, DynVector};

// Runtime-sized matrix
let a = DynMatrix::from_rows(3, 3, &[
    2.0_f64, 1.0, -1.0,
    -3.0, -1.0, 2.0,
    -2.0, 1.0, 2.0,
]);
let b = DynVector::from_slice(&[8.0, -11.0, -3.0]);
let x = a.solve(&b).unwrap();

// Convert between fixed and dynamic
use numeris::Matrix;
let fixed = Matrix::new([[1.0, 2.0], [3.0, 4.0]]);
let dynamic: DynMatrix<f64> = fixed.into();
let back: Matrix<f64, 2, 2> = (&dynamic).try_into().unwrap();

// Mixed arithmetic
let result = fixed * &dynamic; // Matrix * DynMatrix → DynMatrix
```

Type aliases: `DynMatrixf64`, `DynMatrixf32`, `DynVectorf64`, `DynVectorf32`, `DynMatrixz64` (complex, requires `complex` feature), etc.

## ODE integration

Fixed-step RK4 and 7 adaptive Runge-Kutta solvers with embedded error estimation and dense output:

```rust
use numeris::ode::{RKAdaptive, RKTS54, AdaptiveSettings};
use numeris::Vector;

// Harmonic oscillator: y'' = -y
let y0 = Vector::from_array([1.0_f64, 0.0]);
let tau = 2.0 * std::f64::consts::PI;
let sol = RKTS54::integrate(
    0.0, tau, &y0,
    |_t, y| Vector::from_array([y[1], -y[0]]),
    &AdaptiveSettings::default(),
).unwrap();
assert!((sol.y[0] - 1.0).abs() < 1e-6); // cos(2π) ≈ 1
```

### Explicit solvers

| Solver | Stages | Order | FSAL | Interpolant |
|---|---|---|---|---|
| `RKF45` | 6 | 5(4) | no | — |
| `RKTS54` | 7 | 5(4) | yes | 4th degree |
| `RKV65` | 10 | 6(5) | no | 6th degree |
| `RKV87` | 17 | 8(7) | no | 7th degree |
| `RKV98` | 21 | 9(8) | no | 8th degree |
| `RKV98NoInterp` | 16 | 9(8) | no | — |
| `RKV98Efficient` | 26 | 9(8) | no | 9th degree |

### Stiff solvers (Rosenbrock)

For stiff ODEs (chemical kinetics, circuit simulation, orbital mechanics with drag), use `RODAS4` — a linearly-implicit Rosenbrock method that solves linear systems involving the Jacobian instead of nonlinear Newton iterations:

```rust
use numeris::ode::{Rosenbrock, RODAS4, AdaptiveSettings};
use numeris::{Vector, Matrix};

// Stiff decay: y' = -1000*y, y(0) = 1
let y0 = Vector::from_array([1.0_f64]);
let sol = RODAS4::integrate(
    0.0, 0.01, &y0,
    |_t, y| Vector::from_array([-1000.0 * y[0]]),
    |_t, _y| Matrix::new([[-1000.0]]),  // Jacobian ∂f/∂y
    &AdaptiveSettings::default(),
).unwrap();
// Or use integrate_auto for automatic finite-difference Jacobian
```

| Solver | Stages | Order | L-stable |
|---|---|---|---|
| `RODAS4` | 6 | 4(3) | yes |

## Optimization

Root finding, unconstrained minimization, and nonlinear least squares (requires `optim` feature):

```toml
[dependencies]
numeris = { version = "0.1", features = ["optim"] }
```

```rust
use numeris::optim::{brent, minimize_bfgs, least_squares_lm, RootSettings, BfgsSettings, LmSettings};
use numeris::{Matrix, Vector};

// Root finding: solve x² - 2 = 0
let root = brent(|x| x * x - 2.0, 0.0, 2.0, &RootSettings::default()).unwrap();
assert!((root.x - std::f64::consts::SQRT_2).abs() < 1e-12);

// BFGS minimization: minimize (x-1)² + (y-2)²
let min = minimize_bfgs(
    |x: &Vector<f64, 2>| (x[0] - 1.0).powi(2) + (x[1] - 2.0).powi(2),
    |x: &Vector<f64, 2>| Vector::from_array([2.0 * (x[0] - 1.0), 2.0 * (x[1] - 2.0)]),
    &Vector::from_array([0.0, 0.0]),
    &BfgsSettings::default(),
).unwrap();
assert!((min.x[0] - 1.0).abs() < 1e-6);

// Levenberg-Marquardt: fit y = a * exp(b * x)
let t = [0.0_f64, 1.0, 2.0, 3.0, 4.0];
let y = [2.0, 2.7, 3.65, 4.95, 6.7];
let fit = least_squares_lm(
    |x: &Vector<f64, 2>| {
        let mut r = Vector::<f64, 5>::zeros();
        for i in 0..5 { r[i] = x[0] * (x[1] * t[i]).exp() - y[i]; }
        r
    },
    |x: &Vector<f64, 2>| {
        let mut j = Matrix::<f64, 5, 2>::zeros();
        for i in 0..5 {
            let e = (x[1] * t[i]).exp();
            j[(i, 0)] = e;
            j[(i, 1)] = x[0] * t[i] * e;
        }
        j
    },
    &Vector::from_array([1.0, 0.1]),
    &LmSettings::default(),
).unwrap();
assert!(fit.cost < 0.1);
```

| Algorithm | Function | Use case |
|---|---|---|
| Brent's method | `brent` | Bracketed scalar root finding |
| Newton's method | `newton_1d` | Scalar root finding with derivative |
| BFGS | `minimize_bfgs` | Unconstrained minimization |
| Gauss-Newton | `least_squares_gn` | Nonlinear least squares (QR-based) |
| Levenberg-Marquardt | `least_squares_lm` | Nonlinear least squares (damped) |

Finite-difference utilities: `finite_difference_gradient` and `finite_difference_jacobian` for when analytical derivatives aren't available.

## Digital IIR filters

Biquad cascade IIR filters with Butterworth and Chebyshev Type I design (requires `control` feature):

```toml
[dependencies]
numeris = { version = "0.1", features = ["control"] }
```

```rust
use numeris::control::{butterworth_lowpass, chebyshev1_lowpass, BiquadCascade};

// 4th-order Butterworth lowpass at 1 kHz, 8 kHz sample rate (N=2 biquad sections)
let mut lpf: BiquadCascade<f64, 2> = butterworth_lowpass(4, 1000.0, 8000.0).unwrap();
let y = lpf.tick(1.0); // filter one sample

// Bulk processing
let input = [1.0, 0.5, -0.3, 0.8];
let mut output = [0.0; 4];
lpf.reset();
lpf.process(&input, &mut output);

// Chebyshev Type I: steeper rolloff with 1 dB passband ripple
let mut cheb: BiquadCascade<f64, 2> = chebyshev1_lowpass(4, 1.0, 1000.0, 8000.0).unwrap();
```

| Design | Lowpass | Highpass |
|---|---|---|
| Butterworth | `butterworth_lowpass` | `butterworth_highpass` |
| Chebyshev Type I | `chebyshev1_lowpass` | `chebyshev1_highpass` |

All filters are no-std compatible, use no complex number arithmetic, and work with both `f32` and `f64`.

### PID controller

Discrete-time PID controller with trapezoidal integration, derivative-on-measurement (no derivative kick), optional derivative low-pass filter, and anti-windup via back-calculation:

```rust
use numeris::control::Pid;

// PID controller at 1 kHz with output clamping
let mut pid = Pid::new(2.0_f64, 0.5, 0.1, 0.001)
    .with_output_limits(-10.0, 10.0)
    .with_derivative_filter(0.01); // first-order LPF on derivative

// Control loop
let setpoint = 1.0;
let mut measurement = 0.0;
for _ in 0..1000 {
    let u = pid.tick(setpoint, measurement);
    measurement += 0.001 * (-measurement + u); // simple plant
}
assert!((measurement - setpoint).abs() < 0.01);
```

## Complex matrices

Enable the `complex` feature to use decompositions with complex elements:

```toml
[dependencies]
numeris = { version = "0.1", features = ["complex"] }
```

```rust
use numeris::{Complex, Matrix, Vector};

type C = Complex<f64>;

let a = Matrix::new([
    [C::new(2.0, 1.0), C::new(1.0, -1.0)],
    [C::new(1.0, 0.0), C::new(3.0, 2.0)],
]);
let b = Vector::from_array([C::new(5.0, 3.0), C::new(7.0, 4.0)]);
let x = a.solve(&b).unwrap();

// Hermitian positive-definite Cholesky (A = L * L^H)
let hpd = Matrix::new([
    [C::new(4.0, 0.0), C::new(2.0, 1.0)],
    [C::new(2.0, -1.0), C::new(5.0, 0.0)],
]);
let chol = hpd.cholesky().unwrap();
```

Complex support adds zero overhead to real-valued code paths. The `LinalgScalar` trait methods (`modulus`, `conj`, etc.) are `#[inline]` identity functions for `f32`/`f64`, fully erased by the compiler.

## Cargo features

| Feature | Default | Description |
|---|---|---|
| `std` | yes | Implies `alloc`. Uses hardware FPU via system libm. |
| `alloc` | via `std` | Enables `DynMatrix` / `DynVector` (heap-allocated, runtime-sized). |
| `ode` | yes | ODE integration (RK4, adaptive solvers). |
| `optim` | no | Optimization (root finding, BFGS, Gauss-Newton, LM). |
| `control` | no | Digital IIR filters (Butterworth, Chebyshev Type I). |
| `libm` | baseline | Pure-Rust software float math. Always available as fallback. |
| `complex` | no | Adds `Complex<f32>` / `Complex<f64>` support via `num-complex`. |
| `all` | no | All features: `std` + `ode` + `optim` + `control` + `complex`. |

```bash
# Default (std + ode)
cargo build

# All features
cargo build --features all

# No-std for embedded
cargo build --no-default-features --features libm

# No-std with dynamic matrices
cargo build --no-default-features --features "libm,alloc"

# With optimization and complex support
cargo build --features "optim,complex"
```

## Module overview

### `matrix` — Fixed-size matrix

`Matrix<T, M, N>` with `[[T; N]; M]` row-major storage.

- Arithmetic: `+`, `-`, `*` (matrix and scalar), negation, element-wise multiply/divide
- Indexing: `m[(i, j)]`, row/column access, block extraction and insertion
- Square: `trace`, `det`, `diag`, `from_diag`, `pow`, `is_symmetric`, `eye`
- Norms: `frobenius_norm`, `norm_inf`, `norm_one`
- Utilities: `transpose`, `from_fn`, `map`, `swap_rows`, `swap_cols`, `sum`, `abs`
- Iteration: `iter()`, `iter_mut()`, `as_slice()`, `IntoIterator`

Vectors are type aliases: `Vector<T, N>` = `Matrix<T, 1, N>` (row vector), `ColumnVector<T, N>` = `Matrix<T, N, 1>`.

Vector-specific: `dot`, `cross`, `outer`, `norm`, `norm_l1`, `normalize`.

#### Size aliases

Convenience aliases are provided for common sizes (all re-exported from the crate root):

| Square | Rectangular (examples) | Vectors |
|---|---|---|
| `Matrix1` .. `Matrix6` | `Matrix2x3`, `Matrix3x4`, `Matrix4x6`, ... | `Vector1` .. `Vector6` |
| | All M×N combinations for M,N ∈ 1..6, M≠N | `ColumnVector1` .. `ColumnVector6` |

```rust
use numeris::{Matrix3, Matrix4x3, Vector3};

let rotation: Matrix3<f64> = Matrix3::eye();
let points: Matrix4x3<f64> = Matrix4x3::zeros(); // 4 rows, 3 cols
let v: Vector3<f64> = Vector3::from_array([1.0, 2.0, 3.0]);
```

### `dynmatrix` — Dynamic matrix (requires `alloc`)

`DynMatrix<T>` with `Vec<T>` row-major storage and runtime dimensions.

- Same arithmetic, norms, block ops, and utilities as fixed `Matrix`
- Mixed ops: `Matrix * DynMatrix`, `DynMatrix + Matrix`, etc. → `DynMatrix`
- `DynVector<T>` newtype with single-index access and `dot`
- Conversions: `From<Matrix>` → `DynMatrix`, `TryFrom<&DynMatrix>` → `Matrix`
- Full linalg: `DynLu`, `DynCholesky`, `DynQr`, `DynSvd`, `DynSymmetricEigen`, `DynSchur` wrappers + convenience methods

Type aliases: `DynMatrixf64`, `DynMatrixf32`, `DynVectorf64`, `DynVectorf32`, `DynMatrixi32`, `DynMatrixi64`, `DynMatrixu32`, `DynMatrixu64`, `DynMatrixz64`, `DynMatrixz32` (complex).

### `linalg` — Decompositions

All decompositions provide both free functions (operating on `&mut impl MatrixMut<T>`) and wrapper structs with `solve()`, `inverse()`, `det()`.

| Decomposition | Free function | Fixed struct | Dynamic struct | Notes |
|---|---|---|---|---|
| LU | `lu_in_place` | `LuDecomposition` | `DynLu` | Partial pivoting |
| Cholesky | `cholesky_in_place` | `CholeskyDecomposition` | `DynCholesky` | A = LL^H (Hermitian) |
| QR | `qr_in_place` | `QrDecomposition` | `DynQr` | Householder reflections, least-squares |
| SVD | `bidiagonalize` + `bidiagonal_qr` | `SvdDecomposition` | `DynSvd` | Rectangular M×N, singular values, rank |
| Symmetric Eigen | `tridiagonalize` + `tridiagonal_qr_*` | `SymmetricEigen` | `DynSymmetricEigen` | Real eigenvalues, orthogonal eigenvectors |
| Schur | `hessenberg` + `francis_qr` | `SchurDecomposition` | `DynSchur` | Quasi-upper-triangular, general eigenvalues |

Convenience methods on `Matrix`: `a.lu()`, `a.cholesky()`, `a.qr()`, `a.svd()`, `a.solve(&b)`, `a.inverse()`, `a.det()`, `a.eig_symmetric()`, `a.eigenvalues_symmetric()`, `a.schur()`, `a.eigenvalues()`.

Same convenience methods on `DynMatrix`: `a.lu()`, `a.cholesky()`, `a.qr()`, `a.svd()`, `a.solve(&b)`, `a.inverse()`, `a.det()`, `a.eig_symmetric()`, `a.eigenvalues_symmetric()`, `a.schur()`, `a.eigenvalues()`.

### `ode` — ODE integration

Fixed-step `rk4` / `rk4_step` and 7 adaptive Runge-Kutta solvers via the `RKAdaptive` trait. PI step-size controller (Söderlind & Wang 2006). Dense output / interpolation available for most solvers (gated behind `std`). For stiff systems, `RODAS4` provides an L-stable Rosenbrock method via the `Rosenbrock` trait — accepts user-supplied or automatic finite-difference Jacobians.

### `optim` — Optimization (requires `optim` feature)

- **Root finding**: `brent` (bracketed, superlinear convergence), `newton_1d` (with derivative)
- **Minimization**: `minimize_bfgs` (BFGS quasi-Newton with Armijo line search)
- **Least squares**: `least_squares_gn` (Gauss-Newton via QR), `least_squares_lm` (Levenberg-Marquardt via damped normal equations)
- **Finite differences**: `finite_difference_gradient`, `finite_difference_jacobian` for numerical derivatives
- All algorithms use `FloatScalar` bound (real-valued), configurable via settings structs with `Default` impls for `f32` and `f64`

### `control` — Digital filters and controllers (requires `control` feature)

Biquad cascade filters designed via the bilinear transform, and a discrete-time PID controller. No `complex` feature dependency.

- `Biquad<T>` — single second-order section, Direct Form II Transposed
- `BiquadCascade<T, N>` — cascade of `N` biquad sections (filter order ≤ 2N)
- Design functions: `butterworth_lowpass`, `butterworth_highpass`, `chebyshev1_lowpass`, `chebyshev1_highpass`
- Supports arbitrary even/odd filter orders; odd-order uses degenerate first-order last section
- `tick`, `process`, `process_inplace` for sample-by-sample or bulk filtering
- `Pid<T>` — PID controller with derivative filtering, output clamping, anti-windup back-calculation

### `quaternion` — Unit quaternion rotations

`Quaternion<T>` with scalar-first convention `[w, x, y, z]`.

- Construction: `new`, `identity`, `from_axis_angle`, `from_euler`, `from_rotation_matrix`, `rotx`, `roty`, `rotz`
- Operations: `*` (Hamilton product), `* Vector3` (rotation), `conjugate`, `inverse`, `normalize`, `slerp`
- Conversion: `to_rotation_matrix`, `to_axis_angle`, `to_euler`

### `traits` — Element traits

| Trait | Bounds | Used by |
|---|---|---|
| `Scalar` | `Copy + PartialEq + Debug + Zero + One + Num` | All matrix ops |
| `FloatScalar` | `Scalar + Float + LinalgScalar<Real=Self>` | Quaternions, ordered comparisons |
| `LinalgScalar` | `Scalar` + `modulus`, `conj`, `re`, `lsqrt`, `lln`, `from_real` | Decompositions, norms |
| `MatrixRef<T>` | Read-only `get(row, col)` | Generic algorithms |
| `MatrixMut<T>` | Adds `get_mut(row, col)` | In-place decompositions |

## Module plan

Checked items are implemented; unchecked are potential future work.

- [x] **matrix** — Fixed-size matrix (stack-allocated, const-generic dimensions), size aliases up to 6×6
- [x] **linalg** — LU, Cholesky, QR, SVD decompositions; symmetric eigendecomposition; real Schur decomposition; solvers, inverse, determinant; complex support
- [x] **quaternion** — Unit quaternion for rotations (SLERP, Euler, axis-angle, rotation matrices)
- [x] **ode** — ODE integration (RK4, 7 adaptive solvers, dense output, RODAS4 stiff solver)
- [x] **dynmatrix** — Heap-allocated runtime-sized matrix/vector (`alloc` feature)
- [ ] **interp** — Interpolation (linear, cubic spline, Hermite)
- [x] **optim** — Optimization (Brent, Newton, BFGS, Gauss-Newton, Levenberg-Marquardt)
- [ ] **quad** — Numerical quadrature / integration
- [ ] **fft** — Fast Fourier Transform
- [ ] **special** — Special functions (Bessel, gamma, erf, etc.)
- [ ] **stats** — Statistics and distributions
- [ ] **poly** — Polynomial operations and root-finding
- [x] **control** — Digital IIR filters (Butterworth, Chebyshev), PID controllers, state-space systems, discrete-time control (ZOH, Tustin bilinear transform)

## Performance

numeris is designed for two use cases: no-std embedded systems and high-performance desktop/server computing.

**SIMD acceleration** is always-on for f32/f64 — no feature flag needed. On aarch64 (NEON) and x86_64 (SSE2/AVX/AVX-512), matrix multiply, dot products, and element-wise operations use hardware SIMD intrinsics via `core::arch`. Matrix multiply uses register-blocked micro-kernels (inspired by [nano-gemm](https://github.com/sarah-quinones/nano-gemm) by Sarah Quinones) that accumulate MR×NR tiles in SIMD registers across the full k-loop, writing C only once — reducing memory traffic by up to O(n) vs. naive implementations. Integer and complex types fall back to scalar loops at zero cost (dead-code eliminated at monomorphization via `TypeId` dispatch).

SSE2 and NEON are always-on baselines. AVX and AVX-512 are compile-time opt-in via `-C target-cpu=native`; dispatch selects the widest available ISA.

| Architecture | ISA | f64 tile (MR×NR) | f32 tile (MR×NR) |
|---|---|---|---|
| aarch64 | NEON (128-bit) | 4×4 | 8×4 |
| x86_64 | SSE2 (128-bit) | 4×4 | 8×4 |
| x86_64 | AVX (256-bit) | 8×4 | 16×4 |
| x86_64 | AVX-512 (512-bit) | 16×4 | 32×4 |
| other | scalar fallback | 4×4 | 4×4 |

## Design decisions

- **Stack-allocated**: `[[T; N]; M]` storage, no heap. Dimensions are const generics.
- **Heap-allocated**: `DynMatrix` uses `Vec<T>` for runtime dimensions, behind `alloc` feature.
- **`num-traits`**: Generic numeric bounds with `default-features = false`.
- **In-place algorithms**: Decompositions operate on `&mut impl MatrixMut<T>`, avoiding allocator/storage trait complexity. Both `Matrix` and `DynMatrix` implement `MatrixMut`, so the same free functions work for both.
- **Integer matrices**: Work with `Scalar` (all basic ops). Float-only operations (`det`, norms, decompositions) require `LinalgScalar` or `FloatScalar`.
- **Complex support**: Additive, behind a feature flag. Zero cost for real-only usage.

## Acknowledgments

The register-blocked SIMD matrix multiply micro-kernels are inspired by [nano-gemm](https://github.com/sarah-quinones/nano-gemm) and [faer](https://github.com/sarah-quinones/faer-rs) by Sarah Quinones.

## License

MIT