gprx 0.2.0

Gaussian process regression in Rust: exact GPR, sparse GPR (VFE), and SVGP.
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
//! Trainer for collapsed variational SGPR.

use std::marker::PhantomData;

use crate::error::GprError;
use crate::policy::{JitterPolicy, KernelExp, with_kernel_exp};
use crate::sparse::{SparseCore, SparseSpec};
use crate::transform::{UnfittedTarget, UnfittedTransform};

use crate::kernel::{KernelSpec, ModelKernel, ModelKernelParts, PointKernel};
use crate::likelihood::GaussianLikelihood;
use crate::optimizer::{Fixed, Lbfgs, Optimizer};
use crate::precision::{DoublePrecision, GpScalar};
use crate::sgpr::SgprObjective;

use super::factor::assemble_fitted;
use super::fitted::FittedSgpr;
use super::{FixedInducing, FreeInducing, InducingLayout};
use crate::kernel::NoSupply;

/// Represents the trainer for collapsed variational SGPR at an inducing set the caller chooses.
///
/// The type parameters are the optimizer `O` (default [`Lbfgs`]; [`Fixed`]
/// for [`Sgpr<Fixed, I>::factor`]), the inducing layout `I`, the precision
/// `P` (default [`DoublePrecision`]), and the kernel `K` (default
/// [`KernelSpec`]). The default [`FixedInducing`] searches kernel and
/// likelihood `θ` only. [`Self::with_inducing`]`(`[`FreeInducing`]`)`
/// searches `θ` and `Z` together. [`Sgpr<Fixed, I>::factor`] prepares the
/// VFE system at the current `θ` with no search.
///
/// `K` selects how the inducing points are given. A [`KernelSpec`] model
/// takes their coordinates `Z` (column-major, `m` points by `d` features)
/// next to the training `X`. A [`crate::kernel::DistanceKernel`] model has
/// no coordinates to place them at: its inducing points are training
/// points, named by their indices, and the caller supplies per slot the
/// `n × m` squared distances from the training points to them (`K_mm`
/// reads the inducing points' rows of that block). A prediction then takes
/// the `m × q` squared distances from the inducing points to the queries.
/// Such a model keeps its inducing points fixed; [`FreeInducing`] needs
/// coordinates to move.
///
/// # Examples
///
/// ```rust
/// use gprx::kernel::{KernelSpec, RbfKernel};
/// use gprx::{
///     Fixed, FixedInducing, FreeInducing, GaussianLikelihood, JitterPolicy, KernelExp, Lbfgs,
///     Sgpr, SinglePrecision,
/// };
///
/// # fn main() -> Result<(), gprx::GprError> {
/// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
/// let likelihood = GaussianLikelihood::new(0.1)?;
/// let mut sgpr: Sgpr<Lbfgs, FixedInducing> =
///     Sgpr::new(kernel, likelihood).with_math(KernelExp::FastApprox);
/// assert_eq!(sgpr.math(), KernelExp::FastApprox);
/// assert!(matches!(sgpr.jitter_policy(), JitterPolicy::Adaptive(_)));
/// // θ is the kernel's log-lengthscale, then the likelihood's log σn²; `Z` is not in it.
/// assert_eq!(sgpr.num_params(), sgpr.kernel().num_params() + 1);
/// let mut theta = [0.0; 2];
/// sgpr.get_params(&mut theta)?;
/// theta[0] = 0.5_f64.ln();
/// sgpr.set_params(&theta)?;
/// assert!((sgpr.likelihood().log_noise_variance() - theta[1]).abs() < 1e-12);
///
/// // Column-major `X` (n = 4, d = 1) and the inducing coordinates `Z` (m = 2).
/// let (x, y, z) = ([0.0, 1.0, 2.0, 3.0], [0.0, 1.0, 0.5, 0.25], [0.5, 2.5]);
/// let fitted = sgpr.clone().fit(&x, 4, 1, &y, &z, 2).map_err(|(_, e)| e)?;
/// assert_eq!(fitted.n(), 4);
/// assert_eq!(fitted.m(), 2);
///
/// // `Z` searched with θ: two more parameters.
/// let free = sgpr.clone().with_inducing(FreeInducing).fit(&x, 4, 1, &y, &z, 2).map_err(|(_, e)| e)?;
/// assert_eq!(free.num_params(), 4);
///
/// // `f32` storage, factored at the current θ.
/// let single = sgpr
///     .with_precision::<SinglePrecision>()
///     .with_optimizer(Fixed)
///     .factor(&x, 4, 1, &y, &z, 2)
///     .map_err(|(_, e)| e)?;
/// assert_eq!(single.predict(&[1.5], 1, 1)?.mean.len(), 1);
/// # Ok(())
/// # }
/// ```
#[derive(Clone, Debug)]
pub struct Sgpr<O = Lbfgs, I = FixedInducing, P = DoublePrecision, K: ModelKernel = KernelSpec> {
    pub(super) spec: SparseSpec<K::Supply>,
    pub(super) optimizer: O,
    pub(super) inducing: PhantomData<I>,
    pub(super) _precision: PhantomData<P>,
    pub(super) _kernel: PhantomData<K>,
}

impl<K: ModelKernel> Sgpr<Lbfgs, FixedInducing, DoublePrecision, K> {
    /// Builds a trainer with identity transforms, the current kernel `θ`, and [`Lbfgs`].
    ///
    /// Inducing coordinates are an argument of [`Sgpr::fit`] /
    /// [`Sgpr<Fixed>::factor`], not of this constructor. Call
    /// [`Self::with_optimizer`] to switch to [`Fixed`] or another
    /// [`Optimizer`].
    ///
    /// See the example on [`Sgpr`].
    pub fn new(kernel: K, likelihood: GaussianLikelihood) -> Self {
        Self {
            spec: SparseSpec::new(<K as ModelKernelParts>::into_spec(kernel), likelihood),
            optimizer: Lbfgs::new(),
            inducing: PhantomData,
            _precision: PhantomData,
            _kernel: PhantomData,
        }
    }
}

impl<O, I, P, K: ModelKernel> Sgpr<O, I, P, K> {
    /// The same settings under new type parameters, with `map` applied to
    /// the optimizer.
    fn retype<O2, I2, P2>(self, map: impl FnOnce(O) -> O2) -> Sgpr<O2, I2, P2, K> {
        Sgpr {
            spec: self.spec,
            optimizer: map(self.optimizer),
            inducing: PhantomData,
            _precision: PhantomData,
            _kernel: PhantomData,
        }
    }

    /// Replaces the optimizer type parameter.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{Fixed, GaussianLikelihood, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
    /// let likelihood = GaussianLikelihood::new(0.1)?;
    /// let fitted = Sgpr::new(kernel, likelihood)
    ///     .with_optimizer(Fixed)
    ///     .factor(&[0.0, 1.0], 2, 1, &[0.0, 1.0], &[0.0, 1.0], 2)
    ///     .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.m(), fitted.n());
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_optimizer<O2>(self, optimizer: O2) -> Sgpr<O2, I, P, K> {
        self.retype(|_| optimizer)
    }

    /// Selects the storage precision.
    ///
    /// Omitting it leaves [`DoublePrecision`].
    ///
    /// See the example on [`Sgpr`].
    pub fn with_precision<P2: GpScalar>(self) -> Sgpr<O, I, P2, K> {
        self.retype(|optimizer| optimizer)
    }

    /// Selects the kernel `exp`.
    ///
    /// Omitting it leaves [`KernelExp::Accurate`].
    ///
    /// `fit`, `factor`, and predict use the same `exp`. Hyperparameter
    /// `exp(θ)` is unchanged.
    ///
    /// See the example on [`Sgpr`].
    pub fn with_math(mut self, math: KernelExp) -> Self {
        self.spec.math = math;
        self
    }

    /// Returns the kernel `exp` mode.
    ///
    /// See the example on [`Sgpr`].
    pub fn math(&self) -> KernelExp {
        self.spec.math
    }

    /// Replaces the jitter retries for factoring `K_mm = k(Z, Z)`.
    ///
    /// The default is `JitterPolicy::adaptive(1e-8, 10.0, 5, 1e-3)`, not
    /// the [`JitterPolicy::default`] (no retry) of [`crate::Gpr`]: the
    /// Exact system `K + σn² I` carries the observation noise, `K_mm`
    /// does not, so close inducing points leave it singular in floating
    /// point. The policy applies wherever `K_mm` is factored: fit, factor,
    /// `set_params`, predict, and the online updates.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{Fixed, GaussianLikelihood, JitterPolicy, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let policy = JitterPolicy::fixed(1e-6)?;
    /// let fitted = Sgpr::new(
    ///     KernelSpec::from(RbfKernel::new(1.0)?),
    ///     GaussianLikelihood::new(0.1)?,
    /// )
    /// .with_jitter_policy(policy)
    /// .with_optimizer(Fixed)
    ///     .factor(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[0.5, 2.5], 2)
    /// .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.jitter_policy(), policy);
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_jitter_policy(mut self, policy: JitterPolicy) -> Self {
        self.spec.jitter = policy;
        self
    }

    /// Returns the jitter retries for factoring `K_mm`.
    ///
    /// See the example on [`Sgpr`].
    pub fn jitter_policy(&self) -> JitterPolicy {
        self.spec.jitter
    }

    /// Replaces the target (`y`) transform.
    ///
    /// Omitting it leaves identity.
    ///
    /// The map is fitted on training `y`. Predictions are mapped back to the
    /// original scale. The negative marginal likelihood and its gradient are
    /// those of the transformed `y`. A single map or a
    /// [`crate::transform::TargetPipeline`].
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::transform::StandardizeTarget;
    /// use gprx::{Fixed, GaussianLikelihood, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let fitted = Sgpr::new(
    ///     KernelSpec::from(RbfKernel::new(1.0)?),
    ///     GaussianLikelihood::new(0.1)?,
    /// )
    /// .with_target_transform(StandardizeTarget::new())
    /// .with_optimizer(Fixed)
    ///     .factor(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[100.0, 101.0, 100.5, 100.25], &[0.5, 2.5], 2)
    /// .map_err(|(_, e)| e)?;
    /// let pred = fitted.predict(&[1.0], 1, 1)?;
    /// assert!(pred.mean[0] > 99.0);
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_target_transform(mut self, transform: impl UnfittedTarget + 'static) -> Self {
        self.spec.y_transform = Box::new(transform);
        self
    }

    /// Returns the observation-noise model.
    ///
    /// See the example on [`Sgpr`].
    pub fn likelihood(&self) -> &GaussianLikelihood {
        &self.spec.likelihood
    }

    /// Returns the concatenated kernel and likelihood parameter count.
    ///
    /// Inducing coordinates are not counted.
    ///
    /// See the example on [`Sgpr`].
    pub fn num_params(&self) -> usize {
        self.spec.theta_len()
    }

    /// Writes kernel `θ` then likelihood `θ` into `out`.
    ///
    /// # Errors
    ///
    /// Returns [`GprError::LengthMismatch`] if `out` is the wrong length
    /// or a custom leaf rejects the write.
    ///
    /// See the example on [`Sgpr`].
    pub fn get_params(&self, out: &mut [f64]) -> Result<(), GprError> {
        self.spec.read_theta(out)
    }

    /// Sets kernel then likelihood `θ` without forming the VFE system.
    ///
    /// `params` is kernel parameters followed by the likelihood parameter,
    /// matching [`Self::get_params`]. Inducing coordinates are not in this
    /// slice.
    ///
    /// # Errors
    ///
    /// Returns [`GprError::LengthMismatch`] if `params` is the wrong
    /// length, [`GprError::InvalidHyperparameter`] or
    /// [`GprError::InvalidNoiseVariance`] if a kernel or likelihood `θ` is
    /// not finite or its `exp` is not a positive finite value, or
    /// [`GprError::InvalidInterval`] if a value is outside the interval of
    /// its parameter (see [`crate::BoundedParam`] and
    /// [`GaussianLikelihood::bounds`]). A custom kernel leaf returns its own
    /// error. Kernel and likelihood `θ` are committed together only after
    /// both writes succeed.
    ///
    /// See the example on [`Sgpr`].
    pub fn set_params(&mut self, params: &[f64]) -> Result<(), GprError> {
        self.spec.write_theta(params)
    }
}

impl<O, I, P, K: PointKernel> Sgpr<O, I, P, K> {
    /// Replaces the input (`X`) transform.
    ///
    /// Omitting it leaves identity.
    ///
    /// The map is fitted on training `X`. `X`, the inducing points `Z`, and
    /// every later query or inserted point go through it, so `Z` is passed
    /// in the same coordinates as `X`. [`crate::FreeInducing`] searches `Z`
    /// in the transformed coordinates; the fitted model reports `Z` in the
    /// original ones. A single map, a [`crate::transform::Pipeline`], or
    /// [`crate::transform::ColumnwiseInput`].
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::transform::StandardizeInput;
    /// use gprx::{Fixed, GaussianLikelihood, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let fitted = Sgpr::new(
    ///     KernelSpec::from(RbfKernel::new(1.0)?),
    ///     GaussianLikelihood::new(0.1)?,
    /// )
    /// .with_input_transform(StandardizeInput::new())
    /// .with_optimizer(Fixed)
    ///     .factor(&[0.0, 10.0, 20.0, 30.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[5.0, 25.0], 2)
    /// .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.z(), &[5.0, 25.0]);
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_input_transform(mut self, transform: impl UnfittedTransform + 'static) -> Self {
        self.spec.x_transform = Box::new(transform);
        self
    }
}

impl<O, I, P> Sgpr<O, I, P> {
    /// Replaces the inducing-point type parameter.
    ///
    /// [`FixedInducing`] (the default) keeps `Z` fixed. [`FreeInducing`]
    /// appends column-major `Z` to the parameter vector and searches it with
    /// kernel and likelihood `θ`.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{FreeInducing, GaussianLikelihood, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
    /// let likelihood = GaussianLikelihood::new(0.1)?;
    /// let fitted = Sgpr::new(kernel, likelihood)
    ///     .with_inducing(FreeInducing)
    ///     .fit(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[0.5, 2.5], 2)
    ///     .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.num_params(), 4);
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_inducing<I2>(self, _inducing: I2) -> Sgpr<O, I2, P> {
        self.retype(|optimizer| optimizer)
    }

    /// Returns the kernel whose hyperparameters this trainer owns.
    ///
    /// See the example on [`Sgpr`].
    pub fn kernel(&self) -> &KernelSpec {
        &self.spec.kernel
    }
}

impl<O, I, P, C: crate::kernel::PointUse> Sgpr<O, I, P, crate::kernel::DistanceKernel<C>> {
    /// Returns a copy of the kernel whose hyperparameters this trainer owns.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{RbfKernel, ScalarDistance};
    /// use gprx::{GaussianLikelihood, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let image = ScalarDistance::new();
    /// let sgpr = Sgpr::new(image.kernel(RbfKernel::new(1.0)?), GaussianLikelihood::new(0.1)?);
    /// assert_eq!(sgpr.to_kernel().num_params(), 1);
    /// # Ok(())
    /// # }
    /// ```
    pub fn to_kernel(&self) -> crate::kernel::DistanceKernel<C> {
        <crate::kernel::DistanceKernel<C> as ModelKernelParts>::from_spec(self.spec.kernel.clone())
    }
}

impl<O, P> Sgpr<O, FixedInducing, P>
where
    P: GpScalar,
    O: Clone + for<'a> Optimizer<SgprObjective<'a, O, FixedInducing, P>>,
{
    /// Factors the VFE system and searches kernel and likelihood `θ`.
    ///
    /// `x` and `z` are column-major (`n` / `m` points by `d` features). `Z`
    /// is supplied by the caller and is not moved. Likelihood noise is not
    /// added to `K_mm`.
    ///
    /// # Errors
    ///
    /// Returns [`GprError::EmptyInput`] when `n`, `d`, or `m` is zero.
    /// Returns [`GprError::DimensionMismatch`] when `z` is packed with a
    /// different feature count than `x` (the same `n_cols` is required).
    /// Returns [`GprError::LengthMismatch`] when `x`, `y`, or `z` has the
    /// wrong length, and [`GprError::NonFiniteInput`] when a value is `NaN`
    /// or `Inf`. [`GprError::CholeskyFailed`] when `K_mm` or the VFE `B`
    /// matrix cannot be factored. [`GprError::OptimizationNotConverged`] when the solver
    /// stops without a finite best vector.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{GaussianLikelihood, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
    /// let likelihood = GaussianLikelihood::new(0.1)?;
    /// let fitted = Sgpr::new(kernel, likelihood)
    ///     .fit(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[0.5, 2.5], 2)
    ///     .map_err(|(_, e)| e)?;
    /// let nlml = fitted.neg_log_marginal_likelihood()?;
    /// assert!(nlml.is_finite());
    /// # Ok(())
    /// # }
    /// ```
    #[allow(clippy::result_large_err)] // failure returns the trainer so the caller can retry
    pub fn fit(
        self,
        x: &[f64],
        n_rows: usize,
        n_cols: usize,
        y: &[f64],
        z: &[f64],
        n_inducing: usize,
    ) -> Result<FittedSgpr<O, FixedInducing, P>, (Self, GprError)> {
        let core = match SparseCore::prepare(&self.spec, x, n_rows, n_cols, y, z, n_inducing) {
            Ok(core) => core,
            Err(err) => return Err((self, err)),
        };
        match with_kernel_exp!(self.spec.math, M => assemble_fitted::<_, _, M, _, KernelSpec>(
            core,
            self.optimizer.clone(),
        )) {
            Ok(mut fitted) => match fitted.optimize_hyperparameters() {
                Ok(()) => Ok(fitted),
                Err(err) => Err((fitted.into_trainer(), err)),
            },
            Err(err) => Err((self, err)),
        }
    }
}

impl<O, P> Sgpr<O, FreeInducing, P>
where
    P: GpScalar,
    O: Clone + for<'a> Optimizer<SgprObjective<'a, O, FreeInducing, P>>,
{
    /// Factors the VFE system and searches kernel `θ`, likelihood `θ`, and `Z`.
    ///
    /// `x` and `z` are column-major. The initial `Z` is the start of the
    /// joint search. Params are kernel `θ`, likelihood `θ`, then column-major
    /// `Z`. Coordinate intervals are the training box opened by a slack of
    /// 10% of each feature range, at least `0.1`.
    ///
    /// # Errors
    ///
    /// Same input errors as `fit` with [`FixedInducing`], plus
    /// [`GprError::CoordGradientUnsupported`] for Matérn with `ν = 1/2` or a
    /// `Custom` leaf that does not implement the squared-distance derivatives
    /// of [`KernelTerm`](crate::kernel::KernelTerm).
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{FreeInducing, GaussianLikelihood, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
    /// let likelihood = GaussianLikelihood::new(0.1)?;
    /// let fitted = Sgpr::new(kernel, likelihood)
    ///     .with_inducing(FreeInducing)
    ///     .fit(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[0.2, 0.4], 2)
    ///     .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.num_params(), 4);
    /// # Ok(())
    /// # }
    /// ```
    #[allow(clippy::result_large_err)]
    pub fn fit(
        self,
        x: &[f64],
        n_rows: usize,
        n_cols: usize,
        y: &[f64],
        z: &[f64],
        n_inducing: usize,
    ) -> Result<FittedSgpr<O, FreeInducing, P>, (Self, GprError)> {
        let core = match SparseCore::prepare(&self.spec, x, n_rows, n_cols, y, z, n_inducing) {
            Ok(core) => core,
            Err(err) => return Err((self, err)),
        };
        match with_kernel_exp!(self.spec.math, M => assemble_fitted::<_, _, M, _, KernelSpec>(
            core,
            self.optimizer.clone(),
        )) {
            Ok(mut fitted) => match fitted.optimize_hyperparameters() {
                Ok(()) => Ok(fitted),
                Err(err) => Err((fitted.into_trainer(), err)),
            },
            Err(err) => Err((self, err)),
        }
    }
}

impl<I: InducingLayout<NoSupply>, P> Sgpr<Fixed, I, P>
where
    P: GpScalar,
{
    /// Factors `K_mm = k(Z, Z)` at the current `θ` without a search.
    ///
    /// `x` and `z` are column-major (`n` / `m` points by `d` features). `Z`
    /// is supplied by the caller and is not moved. Likelihood noise is not
    /// added to `K_mm`. Also forms the VFE factors used by
    /// [`FittedSgpr::predict`] and
    /// [`FittedSgpr::neg_log_marginal_likelihood`].
    ///
    /// # Errors
    ///
    /// Returns [`GprError::EmptyInput`] when `n`, `d`, or `m` is zero.
    /// Returns [`GprError::DimensionMismatch`] when `z` is packed with a
    /// different feature count than `x` (the same `n_cols` is required).
    /// Returns [`GprError::LengthMismatch`] when `x`, `y`, or `z` has the
    /// wrong length, and [`GprError::NonFiniteInput`] when a value is `NaN`
    /// or `Inf`. [`GprError::CholeskyFailed`] when `K_mm` or the VFE `B`
    /// matrix cannot be factored.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{Fixed, GaussianLikelihood, Sgpr};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
    /// let likelihood = GaussianLikelihood::new(0.1)?;
    /// let fitted = Sgpr::new(kernel, likelihood)
    ///     .with_optimizer(Fixed)
    ///     .factor(&[0.0, 1.0], 2, 1, &[0.0, 1.0], &[0.0, 1.0], 2)
    ///     .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.m(), fitted.n());
    /// # Ok(())
    /// # }
    /// ```
    #[allow(clippy::result_large_err)] // failure returns the trainer so the caller can retry
    pub fn factor(
        self,
        x: &[f64],
        n_rows: usize,
        n_cols: usize,
        y: &[f64],
        z: &[f64],
        n_inducing: usize,
    ) -> Result<FittedSgpr<Fixed, I, P>, (Self, GprError)> {
        let core = match SparseCore::prepare(&self.spec, x, n_rows, n_cols, y, z, n_inducing) {
            Ok(core) => core,
            Err(err) => return Err((self, err)),
        };
        match with_kernel_exp!(self.spec.math, M => assemble_fitted::<_, _, M, _, KernelSpec>(
            core,
            Fixed,
        )) {
            Ok(fitted) => Ok(fitted),
            Err(err) => Err((self, err)),
        }
    }
}