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
// Copyright 2019-2024 argmin developers
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://apache.org/licenses/LICENSE-2.0> or the MIT license <LICENSE-MIT or
// http://opensource.org/licenses/MIT>, at your option. This file may not be
// copied, modified, or distributed except according to those terms.

use crate::core::{
    ArgminFloat, CostFunction, Error, Executor, Gradient, Hessian, IterState, OptimizationResult,
    Problem, Solver, TerminationReason, TerminationStatus, TrustRegionRadius, KV,
};
use argmin_math::{
    ArgminAdd, ArgminDot, ArgminL2Norm, ArgminMul, ArgminSub, ArgminWeightedDot, ArgminZeroLike,
};
#[cfg(feature = "serde1")]
use serde::{Deserialize, Serialize};

/// # SR1 Trust region method
///
/// A Quasi-Newton method which uses symmetric rank 1 (SR1) updating of the Hessian in a trust
/// region framework. An initial parameter vector must be provided, initial cost, gradient and
/// Hessian are optional and will be computed if not provided.
/// Requires a [trust region sub problem](`crate::solver::trustregion`).
///
/// ## Requirements on the optimization problem
///
/// The optimization problem is required to implement [`CostFunction`], [`Gradient`] and
/// [`Hessian`].
///
/// ## Reference
///
/// Jorge Nocedal and Stephen J. Wright (2006). Numerical Optimization.
/// Springer. ISBN 0-387-30303-0.
#[derive(Clone)]
#[cfg_attr(feature = "serde1", derive(Serialize, Deserialize))]
pub struct SR1TrustRegion<R, F> {
    /// parameter for skipping rule
    denominator_factor: F,
    /// subproblem
    subproblem: R,
    /// Radius
    radius: F,
    /// eta \in (0, 10^-3)
    eta: F,
    /// Tolerance for the stopping criterion based on the change of the norm on the gradient
    tol_grad: F,
}

impl<R, F> SR1TrustRegion<R, F>
where
    F: ArgminFloat,
{
    /// Construct a new instance of [`SR1TrustRegion`]
    ///
    /// # Example
    ///
    /// ```
    /// # use argmin::solver::quasinewton::SR1TrustRegion;
    /// # let subproblem = ();
    /// let subproblem = argmin::solver::trustregion::Steihaug::new().with_max_iters(20);
    /// # // The next line defines the type of `subproblem`. This is done here hidden in order to
    /// # // not litter the docs. When all of this is fed into an Executor, the compiler will
    /// # // figure out the types.
    /// # let subproblem: argmin::solver::trustregion::Steihaug<Vec<f64>, f64> = subproblem;
    /// let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem);
    /// ```
    pub fn new(subproblem: R) -> Self {
        SR1TrustRegion {
            denominator_factor: float!(1e-8),
            subproblem,
            radius: float!(1.0),
            eta: float!(0.5 * 1e-3),
            tol_grad: float!(1e-3),
        }
    }

    /// Set denominator factor
    ///
    /// If the denominator of the update is below the `denominator_factor` (scaled with other
    /// factors derived from the parameter vectors and the gradients), then the update of the
    /// inverse Hessian will be skipped.
    ///
    /// Must be in `(0, 1)` and defaults to `1e-8`.
    ///
    /// # Example
    ///
    /// ```
    /// # use argmin::solver::quasinewton::SR1TrustRegion;
    /// # use argmin::core::Error;
    /// # fn main() -> Result<(), Error> {
    /// # let subproblem = ();
    /// let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem).with_denominator_factor(1e-7)?;
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_denominator_factor(mut self, denominator_factor: F) -> Result<Self, Error> {
        if denominator_factor <= float!(0.0) || denominator_factor >= float!(1.0) {
            return Err(argmin_error!(
                InvalidParameter,
                "`SR1TrustRegion`: denominator_factor must be in (0, 1)."
            ));
        }
        self.denominator_factor = denominator_factor;
        Ok(self)
    }

    /// Set initial radius
    ///
    /// Defaults to 1.0.
    ///
    /// # Example
    ///
    /// ```
    /// # use argmin::solver::quasinewton::SR1TrustRegion;
    /// # use argmin::core::Error;
    /// # fn main() -> Result<(), Error> {
    /// # let subproblem = ();
    /// let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem).with_radius(2.0);
    /// # Ok(())
    /// # }
    /// ```
    #[must_use]
    pub fn with_radius(mut self, radius: F) -> Self {
        self.radius = radius.abs();
        self
    }

    /// Set eta
    ///
    /// A step is taken if the actual reduction over the predicted reduction exceeds eta.
    /// Must be in (0, 10^-3) and defaults to 0.5 * 10^-3.
    ///
    /// # Example
    ///
    /// ```
    /// # use argmin::solver::quasinewton::SR1TrustRegion;
    /// # use argmin::core::Error;
    /// # fn main() -> Result<(), Error> {
    /// # let subproblem = ();
    /// let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem).with_eta(10e-4)?;
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_eta(mut self, eta: F) -> Result<Self, Error> {
        if eta >= float!(10e-3) || eta <= float!(0.0) {
            return Err(argmin_error!(
                InvalidParameter,
                "SR1TrustRegion: eta must be in (0, 10^-3)."
            ));
        }
        self.eta = eta;
        Ok(self)
    }

    /// The algorithm stops if the norm of the gradient is below `tol_grad`.
    ///
    /// The provided value must be non-negative. Defaults to `10^-3`.
    ///
    /// # Example
    ///
    /// ```
    /// # use argmin::solver::quasinewton::SR1TrustRegion;
    /// # use argmin::core::Error;
    /// # fn main() -> Result<(), Error> {
    /// # let subproblem = ();
    /// let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem).with_tolerance_grad(1e-6)?;
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_tolerance_grad(mut self, tol_grad: F) -> Result<Self, Error> {
        if tol_grad < float!(0.0) {
            return Err(argmin_error!(
                InvalidParameter,
                "`SR1TrustRegion`: gradient tolerance must be >= 0."
            ));
        }
        self.tol_grad = tol_grad;
        Ok(self)
    }
}

impl<O, R, P, G, B, F> Solver<O, IterState<P, G, (), B, (), F>> for SR1TrustRegion<R, F>
where
    O: CostFunction<Param = P, Output = F>
        + Gradient<Param = P, Gradient = G>
        + Hessian<Param = P, Hessian = B>,
    P: Clone
        + ArgminSub<P, P>
        + ArgminAdd<P, P>
        + ArgminDot<P, F>
        + ArgminDot<P, B>
        + ArgminL2Norm<F>
        + ArgminZeroLike,
    G: Clone + ArgminL2Norm<F> + ArgminDot<P, F> + ArgminSub<G, P>,
    B: Clone + ArgminDot<P, P> + ArgminAdd<B, B> + ArgminMul<F, B>,
    R: Clone + TrustRegionRadius<F> + Solver<O, IterState<P, G, (), B, (), F>>,
    F: ArgminFloat + ArgminL2Norm<F>,
{
    fn name(&self) -> &str {
        "SR1 trust region"
    }

    fn init(
        &mut self,
        problem: &mut Problem<O>,
        mut state: IterState<P, G, (), B, (), F>,
    ) -> Result<(IterState<P, G, (), B, (), F>, Option<KV>), Error> {
        let param = state.take_param().ok_or_else(argmin_error_closure!(
            NotInitialized,
            concat!(
                "`SR1TrustRegion` requires an initial parameter vector. ",
                "Please provide an initial guess via `Executor`s `configure` method."
            )
        ))?;

        let cost = state.get_cost();
        let cost = if cost.is_infinite() {
            problem.cost(&param)?
        } else {
            cost
        };

        let grad = state
            .take_gradient()
            .map(Result::Ok)
            .unwrap_or_else(|| problem.gradient(&param))?;

        let hessian = state
            .take_hessian()
            .map(Result::Ok)
            .unwrap_or_else(|| problem.hessian(&param))?;

        Ok((
            state
                .param(param)
                .cost(cost)
                .gradient(grad)
                .hessian(hessian),
            None,
        ))
    }

    fn next_iter(
        &mut self,
        problem: &mut Problem<O>,
        mut state: IterState<P, G, (), B, (), F>,
    ) -> Result<(IterState<P, G, (), B, (), F>, Option<KV>), Error> {
        let xk = state.take_param().ok_or_else(argmin_error_closure!(
            PotentialBug,
            "`SR1TrustRegion`: Parameter vector in state not set."
        ))?;

        let cost = state.get_cost();

        let prev_grad = state.take_gradient().ok_or_else(argmin_error_closure!(
            PotentialBug,
            "`SR1TrustRegion`: Gradient in state not set."
        ))?;

        let hessian = state.take_hessian().ok_or_else(argmin_error_closure!(
            PotentialBug,
            "`SR1TrustRegion`: Hessian in state not set."
        ))?;

        self.subproblem.set_radius(self.radius);

        let OptimizationResult {
            problem: sub_problem,
            state: mut sub_state,
            ..
        } = Executor::new(problem.take_problem().unwrap(), self.subproblem.clone())
            .configure(|config| {
                config
                    .param(xk.zero_like())
                    .hessian(hessian.clone())
                    .gradient(prev_grad.clone())
                    .cost(cost)
            })
            .ctrlc(false)
            .run()?;

        let sk = sub_state.take_param().ok_or_else(argmin_error_closure!(
            PotentialBug,
            "`SR1TrustRegion`: No parameters returned by line search."
        ))?;

        problem.consume_problem(sub_problem);

        let xksk = xk.add(&sk);
        let dfk1 = problem.gradient(&xksk)?;
        let yk = dfk1.sub(&prev_grad);
        let fk1 = problem.cost(&xksk)?;

        let ared = cost - fk1;
        let tmp1: F = prev_grad.dot(&sk);
        let tmp2: F = sk.weighted_dot(&hessian, &sk);
        let tmp2: F = tmp2.mul(float!(0.5));
        let pred = -tmp1 - tmp2;
        let ap = ared / pred;

        let (xk1, fk1, dfk1) = if ap > self.eta {
            (xksk, fk1, dfk1)
        } else {
            (xk, cost, prev_grad)
        };

        self.radius = if ap > float!(0.75) {
            if sk.l2_norm() <= float!(0.8) * self.radius {
                self.radius
            } else {
                float!(2.0) * self.radius
            }
        } else if ap <= float!(0.75) && ap >= float!(0.1) {
            self.radius
        } else {
            float!(0.5) * self.radius
        };

        let bksk = hessian.dot(&sk);
        let ykbksk = yk.sub(&bksk);
        let skykbksk: F = sk.dot(&ykbksk);

        let hessian_update =
            skykbksk.abs() >= self.denominator_factor * sk.l2_norm() * skykbksk.l2_norm();
        let hessian = if hessian_update {
            let a: B = ykbksk.dot(&ykbksk);
            let b: F = sk.dot(&ykbksk);
            hessian.add(&a.mul(&(float!(1.0) / b)))
        } else {
            hessian
        };

        Ok((
            state.param(xk1).cost(fk1).gradient(dfk1).hessian(hessian),
            Some(kv!["ared" => ared;
                         "pred" => pred;
                         "ap" => ap;
                         "radius" => self.radius;
                         "hessian_update" => hessian_update;]),
        ))
    }

    fn terminate(&mut self, state: &IterState<P, G, (), B, (), F>) -> TerminationStatus {
        if state.get_gradient().unwrap().l2_norm() < self.tol_grad {
            return TerminationStatus::Terminated(TerminationReason::SolverConverged);
        }
        TerminationStatus::NotTerminated
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::core::{test_utils::TestProblem, ArgminError, State};
    use crate::solver::trustregion::CauchyPoint;

    test_trait_impl!(sr1, SR1TrustRegion<CauchyPoint<f64>, f64>);

    #[test]
    fn test_new() {
        #[derive(Eq, PartialEq, Debug)]
        struct MyFakeSubProblem {}

        let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(MyFakeSubProblem {});
        let SR1TrustRegion {
            denominator_factor,
            subproblem,
            radius,
            eta,
            tol_grad,
        } = sr1;

        assert_eq!(denominator_factor.to_ne_bytes(), 1e-8f64.to_ne_bytes());
        assert_eq!(subproblem, MyFakeSubProblem {});
        assert_eq!(radius.to_ne_bytes(), 1.0f64.to_ne_bytes());
        assert_eq!(eta.to_ne_bytes(), (0.5f64 * 1e-3f64).to_ne_bytes());
        assert_eq!(tol_grad.to_ne_bytes(), 1e-3f64.to_ne_bytes());
    }

    #[test]
    fn test_with_denominator_factor() {
        #[derive(Eq, PartialEq, Debug, Clone, Copy)]
        struct MyFakeSubProblem {}

        // correct parameters
        for tol in [f64::EPSILON, 1e-8, 1e-6, 1e-2, 1.0 - f64::EPSILON] {
            let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(MyFakeSubProblem {});
            let res = sr1.with_denominator_factor(tol);
            assert!(res.is_ok());

            let nm = res.unwrap();
            assert_eq!(nm.denominator_factor.to_ne_bytes(), tol.to_ne_bytes());
        }

        // incorrect parameters
        for tol in [-f64::EPSILON, 0.0, -1.0, 1.0] {
            let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(MyFakeSubProblem {});
            let res = sr1.with_denominator_factor(tol);
            assert_error!(
                res,
                ArgminError,
                "Invalid parameter: \"`SR1TrustRegion`: denominator_factor must be in (0, 1).\""
            );
        }
    }

    #[test]
    fn test_with_tolerance_grad() {
        #[derive(Eq, PartialEq, Debug, Clone, Copy)]
        struct MyFakeSubProblem {}

        // correct parameters
        for tol in [1e-6, 0.0, 1e-2, 1.0, 2.0] {
            let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(MyFakeSubProblem {});
            let res = sr1.with_tolerance_grad(tol);
            assert!(res.is_ok());

            let nm = res.unwrap();
            assert_eq!(nm.tol_grad.to_ne_bytes(), tol.to_ne_bytes());
        }

        // incorrect parameters
        for tol in [-f64::EPSILON, -1.0, -100.0, -42.0] {
            let sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(MyFakeSubProblem {});
            let res = sr1.with_tolerance_grad(tol);
            assert_error!(
                res,
                ArgminError,
                "Invalid parameter: \"`SR1TrustRegion`: gradient tolerance must be >= 0.\""
            );
        }
    }

    #[test]
    fn test_init() {
        let subproblem = CauchyPoint::new();

        let param: Vec<f64> = vec![-1.0, 1.0];

        let mut sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem);

        // Forgot to initialize the parameter vector
        let state: IterState<Vec<f64>, Vec<f64>, (), Vec<Vec<f64>>, (), f64> = IterState::new();
        let problem = TestProblem::new();
        let res = sr1.init(&mut Problem::new(problem), state);
        assert_error!(
            res,
            ArgminError,
            concat!(
                "Not initialized: \"`SR1TrustRegion` requires an initial parameter vector. Please ",
                "provide an initial guess via `Executor`s `configure` method.\""
            )
        );

        // All good.
        let state: IterState<Vec<f64>, Vec<f64>, (), Vec<Vec<f64>>, (), f64> =
            IterState::new().param(param.clone());
        let problem = TestProblem::new();
        let (mut state_out, kv) = sr1.init(&mut Problem::new(problem), state).unwrap();

        assert!(kv.is_none());

        let s_param = state_out.take_param().unwrap();

        for (s, p) in s_param.iter().zip(param.iter()) {
            assert_eq!(s.to_ne_bytes(), p.to_ne_bytes());
        }

        let s_grad = state_out.take_gradient().unwrap();

        for (s, p) in s_grad.iter().zip(param.iter()) {
            assert_eq!(s.to_ne_bytes(), p.to_ne_bytes());
        }

        assert_eq!(state_out.get_cost().to_ne_bytes(), 1.0f64.to_ne_bytes())
    }

    #[test]
    fn test_init_provided_cost() {
        let subproblem = CauchyPoint::new();

        let param: Vec<f64> = vec![-1.0, 1.0];

        let mut sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem);

        let state: IterState<Vec<f64>, Vec<f64>, (), Vec<Vec<f64>>, (), f64> =
            IterState::new().param(param).cost(1234.0);

        let problem = TestProblem::new();
        let (state_out, kv) = sr1.init(&mut Problem::new(problem), state).unwrap();

        assert!(kv.is_none());

        assert_eq!(state_out.get_cost().to_ne_bytes(), 1234.0f64.to_ne_bytes())
    }

    #[test]
    fn test_init_provided_grad() {
        let subproblem = CauchyPoint::new();

        let param: Vec<f64> = vec![-1.0, 1.0];
        let gradient: Vec<f64> = vec![4.0, 9.0];

        let mut sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem);

        let state: IterState<Vec<f64>, Vec<f64>, (), Vec<Vec<f64>>, (), f64> =
            IterState::new().param(param).gradient(gradient.clone());

        let problem = TestProblem::new();
        let (mut state_out, kv) = sr1.init(&mut Problem::new(problem), state).unwrap();

        assert!(kv.is_none());

        let s_grad = state_out.take_gradient().unwrap();

        for (s, g) in s_grad.iter().zip(gradient.iter()) {
            assert_eq!(s.to_ne_bytes(), g.to_ne_bytes());
        }
    }

    #[test]
    fn test_init_provided_hessian() {
        let subproblem = CauchyPoint::new();

        let param: Vec<f64> = vec![-1.0, 1.0];
        let gradient: Vec<f64> = vec![4.0, 9.0];
        let hessian: Vec<Vec<f64>> = vec![vec![1.0, 2.0], vec![3.0, 4.0]];

        let mut sr1: SR1TrustRegion<_, f64> = SR1TrustRegion::new(subproblem);

        let state: IterState<Vec<f64>, Vec<f64>, (), Vec<Vec<f64>>, (), f64> = IterState::new()
            .param(param)
            .gradient(gradient)
            .hessian(hessian.clone());

        let problem = TestProblem::new();
        let (mut state_out, kv) = sr1.init(&mut Problem::new(problem), state).unwrap();

        assert!(kv.is_none());

        let s_hessian = state_out.take_hessian().unwrap();

        for (s1, g1) in s_hessian.iter().zip(hessian.iter()) {
            for (s, g) in s1.iter().zip(g1.iter()) {
                assert_eq!(s.to_ne_bytes(), g.to_ne_bytes());
            }
        }
    }
}