File size: 24,922 Bytes
7885a28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
import numpy as np
import scipy.sparse as sps
from ._numdiff import approx_derivative, group_columns
from ._hessian_update_strategy import HessianUpdateStrategy
from scipy.sparse.linalg import LinearOperator
from scipy._lib._array_api import array_namespace
from scipy._lib import array_api_extra as xpx


FD_METHODS = ('2-point', '3-point', 'cs')


def _wrapper_fun(fun, args=()):
    ncalls = [0]

    def wrapped(x):
        ncalls[0] += 1
        # Send a copy because the user may overwrite it.
        # Overwriting results in undefined behaviour because
        # fun(self.x) will change self.x, with the two no longer linked.
        fx = fun(np.copy(x), *args)
        # Make sure the function returns a true scalar
        if not np.isscalar(fx):
            try:
                fx = np.asarray(fx).item()
            except (TypeError, ValueError) as e:
                raise ValueError(
                    "The user-provided objective function "
                    "must return a scalar value."
                ) from e
        return fx
    return wrapped, ncalls


def _wrapper_grad(grad, fun=None, args=(), finite_diff_options=None):
    ncalls = [0]

    if callable(grad):
        def wrapped(x, **kwds):
            # kwds present to give function same signature as numdiff variant
            ncalls[0] += 1
            return np.atleast_1d(grad(np.copy(x), *args))
        return wrapped, ncalls

    elif grad in FD_METHODS:
        def wrapped1(x, f0=None):
            ncalls[0] += 1
            return approx_derivative(
                fun, x, f0=f0, **finite_diff_options
            )

        return wrapped1, ncalls


def _wrapper_hess(hess, grad=None, x0=None, args=(), finite_diff_options=None):
    if callable(hess):
        H = hess(np.copy(x0), *args)
        ncalls = [1]

        if sps.issparse(H):
            def wrapped(x, **kwds):
                ncalls[0] += 1
                return sps.csr_matrix(hess(np.copy(x), *args))

            H = sps.csr_matrix(H)

        elif isinstance(H, LinearOperator):
            def wrapped(x, **kwds):
                ncalls[0] += 1
                return hess(np.copy(x), *args)

        else:  # dense
            def wrapped(x, **kwds):
                ncalls[0] += 1
                return np.atleast_2d(np.asarray(hess(np.copy(x), *args)))

            H = np.atleast_2d(np.asarray(H))

        return wrapped, ncalls, H
    elif hess in FD_METHODS:
        ncalls = [0]

        def wrapped1(x, f0=None):
            return approx_derivative(
                grad, x, f0=f0, **finite_diff_options
            )

        return wrapped1, ncalls, None


class ScalarFunction:
    """Scalar function and its derivatives.

    This class defines a scalar function F: R^n->R and methods for
    computing or approximating its first and second derivatives.

    Parameters
    ----------
    fun : callable
        evaluates the scalar function. Must be of the form ``fun(x, *args)``,
        where ``x`` is the argument in the form of a 1-D array and ``args`` is
        a tuple of any additional fixed parameters needed to completely specify
        the function. Should return a scalar.
    x0 : array-like
        Provides an initial set of variables for evaluating fun. Array of real
        elements of size (n,), where 'n' is the number of independent
        variables.
    args : tuple, optional
        Any additional fixed parameters needed to completely specify the scalar
        function.
    grad : {callable, '2-point', '3-point', 'cs'}
        Method for computing the gradient vector.
        If it is a callable, it should be a function that returns the gradient
        vector:

            ``grad(x, *args) -> array_like, shape (n,)``

        where ``x`` is an array with shape (n,) and ``args`` is a tuple with
        the fixed parameters.
        Alternatively, the keywords  {'2-point', '3-point', 'cs'} can be used
        to select a finite difference scheme for numerical estimation of the
        gradient with a relative step size. These finite difference schemes
        obey any specified `bounds`.
    hess : {callable, '2-point', '3-point', 'cs', HessianUpdateStrategy}
        Method for computing the Hessian matrix. If it is callable, it should
        return the  Hessian matrix:

            ``hess(x, *args) -> {LinearOperator, spmatrix, array}, (n, n)``

        where x is a (n,) ndarray and `args` is a tuple with the fixed
        parameters. Alternatively, the keywords {'2-point', '3-point', 'cs'}
        select a finite difference scheme for numerical estimation. Or, objects
        implementing `HessianUpdateStrategy` interface can be used to
        approximate the Hessian.
        Whenever the gradient is estimated via finite-differences, the Hessian
        cannot be estimated with options {'2-point', '3-point', 'cs'} and needs
        to be estimated using one of the quasi-Newton strategies.
    finite_diff_rel_step : None or array_like
        Relative step size to use. The absolute step size is computed as
        ``h = finite_diff_rel_step * sign(x0) * max(1, abs(x0))``, possibly
        adjusted to fit into the bounds. For ``method='3-point'`` the sign
        of `h` is ignored. If None then finite_diff_rel_step is selected
        automatically,
    finite_diff_bounds : tuple of array_like
        Lower and upper bounds on independent variables. Defaults to no bounds,
        (-np.inf, np.inf). Each bound must match the size of `x0` or be a
        scalar, in the latter case the bound will be the same for all
        variables. Use it to limit the range of function evaluation.
    epsilon : None or array_like, optional
        Absolute step size to use, possibly adjusted to fit into the bounds.
        For ``method='3-point'`` the sign of `epsilon` is ignored. By default
        relative steps are used, only if ``epsilon is not None`` are absolute
        steps used.

    Notes
    -----
    This class implements a memoization logic. There are methods `fun`,
    `grad`, hess` and corresponding attributes `f`, `g` and `H`. The following
    things should be considered:

        1. Use only public methods `fun`, `grad` and `hess`.
        2. After one of the methods is called, the corresponding attribute
           will be set. However, a subsequent call with a different argument
           of *any* of the methods may overwrite the attribute.
    """
    def __init__(self, fun, x0, args, grad, hess, finite_diff_rel_step,
                 finite_diff_bounds, epsilon=None):
        if not callable(grad) and grad not in FD_METHODS:
            raise ValueError(
                f"`grad` must be either callable or one of {FD_METHODS}."
            )

        if not (callable(hess) or hess in FD_METHODS
                or isinstance(hess, HessianUpdateStrategy)):
            raise ValueError(
                f"`hess` must be either callable, HessianUpdateStrategy"
                f" or one of {FD_METHODS}."
            )

        if grad in FD_METHODS and hess in FD_METHODS:
            raise ValueError("Whenever the gradient is estimated via "
                             "finite-differences, we require the Hessian "
                             "to be estimated using one of the "
                             "quasi-Newton strategies.")

        self.xp = xp = array_namespace(x0)
        _x = xpx.atleast_nd(xp.asarray(x0), ndim=1, xp=xp)
        _dtype = xp.float64
        if xp.isdtype(_x.dtype, "real floating"):
            _dtype = _x.dtype

        # original arguments
        self._wrapped_fun, self._nfev = _wrapper_fun(fun, args=args)
        self._orig_fun = fun
        self._orig_grad = grad
        self._orig_hess = hess
        self._args = args

        # promotes to floating
        self.x = xp.astype(_x, _dtype)
        self.x_dtype = _dtype
        self.n = self.x.size
        self.f_updated = False
        self.g_updated = False
        self.H_updated = False

        self._lowest_x = None
        self._lowest_f = np.inf

        finite_diff_options = {}
        if grad in FD_METHODS:
            finite_diff_options["method"] = grad
            finite_diff_options["rel_step"] = finite_diff_rel_step
            finite_diff_options["abs_step"] = epsilon
            finite_diff_options["bounds"] = finite_diff_bounds
        if hess in FD_METHODS:
            finite_diff_options["method"] = hess
            finite_diff_options["rel_step"] = finite_diff_rel_step
            finite_diff_options["abs_step"] = epsilon
            finite_diff_options["as_linear_operator"] = True

        # Initial function evaluation
        self._update_fun()

        # Initial gradient evaluation
        self._wrapped_grad, self._ngev = _wrapper_grad(
            grad,
            fun=self._wrapped_fun,
            args=args,
            finite_diff_options=finite_diff_options
        )
        self._update_grad()

        # Hessian evaluation
        if callable(hess):
            self._wrapped_hess, self._nhev, self.H = _wrapper_hess(
                hess, x0=x0, args=args
            )
            self.H_updated = True
        elif hess in FD_METHODS:
            self._wrapped_hess, self._nhev, self.H = _wrapper_hess(
                hess,
                grad=self._wrapped_grad,
                x0=x0,
                finite_diff_options=finite_diff_options
            )
            self._update_grad()
            self.H = self._wrapped_hess(self.x, f0=self.g)
            self.H_updated = True
        elif isinstance(hess, HessianUpdateStrategy):
            self.H = hess
            self.H.initialize(self.n, 'hess')
            self.H_updated = True
            self.x_prev = None
            self.g_prev = None
            self._nhev = [0]

    @property
    def nfev(self):
        return self._nfev[0]

    @property
    def ngev(self):
        return self._ngev[0]

    @property
    def nhev(self):
        return self._nhev[0]

    def _update_x(self, x):
        if isinstance(self._orig_hess, HessianUpdateStrategy):
            self._update_grad()
            self.x_prev = self.x
            self.g_prev = self.g
            # ensure that self.x is a copy of x. Don't store a reference
            # otherwise the memoization doesn't work properly.

            _x = xpx.atleast_nd(self.xp.asarray(x), ndim=1, xp=self.xp)
            self.x = self.xp.astype(_x, self.x_dtype)
            self.f_updated = False
            self.g_updated = False
            self.H_updated = False
            self._update_hess()
        else:
            # ensure that self.x is a copy of x. Don't store a reference
            # otherwise the memoization doesn't work properly.
            _x = xpx.atleast_nd(self.xp.asarray(x), ndim=1, xp=self.xp)
            self.x = self.xp.astype(_x, self.x_dtype)
            self.f_updated = False
            self.g_updated = False
            self.H_updated = False

    def _update_fun(self):
        if not self.f_updated:
            fx = self._wrapped_fun(self.x)
            if fx < self._lowest_f:
                self._lowest_x = self.x
                self._lowest_f = fx

            self.f = fx
            self.f_updated = True

    def _update_grad(self):
        if not self.g_updated:
            if self._orig_grad in FD_METHODS:
                self._update_fun()
            self.g = self._wrapped_grad(self.x, f0=self.f)
            self.g_updated = True

    def _update_hess(self):
        if not self.H_updated:
            if self._orig_hess in FD_METHODS:
                self._update_grad()
                self.H = self._wrapped_hess(self.x, f0=self.g)
            elif isinstance(self._orig_hess, HessianUpdateStrategy):
                self._update_grad()
                self.H.update(self.x - self.x_prev, self.g - self.g_prev)
            else:       # should be callable(hess)
                self.H = self._wrapped_hess(self.x)

            self.H_updated = True

    def fun(self, x):
        if not np.array_equal(x, self.x):
            self._update_x(x)
        self._update_fun()
        return self.f

    def grad(self, x):
        if not np.array_equal(x, self.x):
            self._update_x(x)
        self._update_grad()
        return self.g

    def hess(self, x):
        if not np.array_equal(x, self.x):
            self._update_x(x)
        self._update_hess()
        return self.H

    def fun_and_grad(self, x):
        if not np.array_equal(x, self.x):
            self._update_x(x)
        self._update_fun()
        self._update_grad()
        return self.f, self.g


class VectorFunction:
    """Vector function and its derivatives.

    This class defines a vector function F: R^n->R^m and methods for
    computing or approximating its first and second derivatives.

    Notes
    -----
    This class implements a memoization logic. There are methods `fun`,
    `jac`, hess` and corresponding attributes `f`, `J` and `H`. The following
    things should be considered:

        1. Use only public methods `fun`, `jac` and `hess`.
        2. After one of the methods is called, the corresponding attribute
           will be set. However, a subsequent call with a different argument
           of *any* of the methods may overwrite the attribute.
    """
    def __init__(self, fun, x0, jac, hess,
                 finite_diff_rel_step, finite_diff_jac_sparsity,
                 finite_diff_bounds, sparse_jacobian):
        if not callable(jac) and jac not in FD_METHODS:
            raise ValueError(f"`jac` must be either callable or one of {FD_METHODS}.")

        if not (callable(hess) or hess in FD_METHODS
                or isinstance(hess, HessianUpdateStrategy)):
            raise ValueError("`hess` must be either callable,"
                             f"HessianUpdateStrategy or one of {FD_METHODS}.")

        if jac in FD_METHODS and hess in FD_METHODS:
            raise ValueError("Whenever the Jacobian is estimated via "
                             "finite-differences, we require the Hessian to "
                             "be estimated using one of the quasi-Newton "
                             "strategies.")

        self.xp = xp = array_namespace(x0)
        _x = xpx.atleast_nd(xp.asarray(x0), ndim=1, xp=xp)
        _dtype = xp.float64
        if xp.isdtype(_x.dtype, "real floating"):
            _dtype = _x.dtype

        # promotes to floating
        self.x = xp.astype(_x, _dtype)
        self.x_dtype = _dtype

        self.n = self.x.size
        self.nfev = 0
        self.njev = 0
        self.nhev = 0
        self.f_updated = False
        self.J_updated = False
        self.H_updated = False

        finite_diff_options = {}
        if jac in FD_METHODS:
            finite_diff_options["method"] = jac
            finite_diff_options["rel_step"] = finite_diff_rel_step
            if finite_diff_jac_sparsity is not None:
                sparsity_groups = group_columns(finite_diff_jac_sparsity)
                finite_diff_options["sparsity"] = (finite_diff_jac_sparsity,
                                                   sparsity_groups)
            finite_diff_options["bounds"] = finite_diff_bounds
            self.x_diff = np.copy(self.x)
        if hess in FD_METHODS:
            finite_diff_options["method"] = hess
            finite_diff_options["rel_step"] = finite_diff_rel_step
            finite_diff_options["as_linear_operator"] = True
            self.x_diff = np.copy(self.x)
        if jac in FD_METHODS and hess in FD_METHODS:
            raise ValueError("Whenever the Jacobian is estimated via "
                             "finite-differences, we require the Hessian to "
                             "be estimated using one of the quasi-Newton "
                             "strategies.")

        # Function evaluation
        def fun_wrapped(x):
            self.nfev += 1
            return np.atleast_1d(fun(x))

        def update_fun():
            self.f = fun_wrapped(self.x)

        self._update_fun_impl = update_fun
        update_fun()

        self.v = np.zeros_like(self.f)
        self.m = self.v.size

        # Jacobian Evaluation
        if callable(jac):
            self.J = jac(self.x)
            self.J_updated = True
            self.njev += 1

            if (sparse_jacobian or
                    sparse_jacobian is None and sps.issparse(self.J)):
                def jac_wrapped(x):
                    self.njev += 1
                    return sps.csr_matrix(jac(x))
                self.J = sps.csr_matrix(self.J)
                self.sparse_jacobian = True

            elif sps.issparse(self.J):
                def jac_wrapped(x):
                    self.njev += 1
                    return jac(x).toarray()
                self.J = self.J.toarray()
                self.sparse_jacobian = False

            else:
                def jac_wrapped(x):
                    self.njev += 1
                    return np.atleast_2d(jac(x))
                self.J = np.atleast_2d(self.J)
                self.sparse_jacobian = False

            def update_jac():
                self.J = jac_wrapped(self.x)

        elif jac in FD_METHODS:
            self.J = approx_derivative(fun_wrapped, self.x, f0=self.f,
                                       **finite_diff_options)
            self.J_updated = True

            if (sparse_jacobian or
                    sparse_jacobian is None and sps.issparse(self.J)):
                def update_jac():
                    self._update_fun()
                    self.J = sps.csr_matrix(
                        approx_derivative(fun_wrapped, self.x, f0=self.f,
                                          **finite_diff_options))
                self.J = sps.csr_matrix(self.J)
                self.sparse_jacobian = True

            elif sps.issparse(self.J):
                def update_jac():
                    self._update_fun()
                    self.J = approx_derivative(fun_wrapped, self.x, f0=self.f,
                                               **finite_diff_options).toarray()
                self.J = self.J.toarray()
                self.sparse_jacobian = False

            else:
                def update_jac():
                    self._update_fun()
                    self.J = np.atleast_2d(
                        approx_derivative(fun_wrapped, self.x, f0=self.f,
                                          **finite_diff_options))
                self.J = np.atleast_2d(self.J)
                self.sparse_jacobian = False

        self._update_jac_impl = update_jac

        # Define Hessian
        if callable(hess):
            self.H = hess(self.x, self.v)
            self.H_updated = True
            self.nhev += 1

            if sps.issparse(self.H):
                def hess_wrapped(x, v):
                    self.nhev += 1
                    return sps.csr_matrix(hess(x, v))
                self.H = sps.csr_matrix(self.H)

            elif isinstance(self.H, LinearOperator):
                def hess_wrapped(x, v):
                    self.nhev += 1
                    return hess(x, v)

            else:
                def hess_wrapped(x, v):
                    self.nhev += 1
                    return np.atleast_2d(np.asarray(hess(x, v)))
                self.H = np.atleast_2d(np.asarray(self.H))

            def update_hess():
                self.H = hess_wrapped(self.x, self.v)
        elif hess in FD_METHODS:
            def jac_dot_v(x, v):
                return jac_wrapped(x).T.dot(v)

            def update_hess():
                self._update_jac()
                self.H = approx_derivative(jac_dot_v, self.x,
                                           f0=self.J.T.dot(self.v),
                                           args=(self.v,),
                                           **finite_diff_options)
            update_hess()
            self.H_updated = True
        elif isinstance(hess, HessianUpdateStrategy):
            self.H = hess
            self.H.initialize(self.n, 'hess')
            self.H_updated = True
            self.x_prev = None
            self.J_prev = None

            def update_hess():
                self._update_jac()
                # When v is updated before x was updated, then x_prev and
                # J_prev are None and we need this check.
                if self.x_prev is not None and self.J_prev is not None:
                    delta_x = self.x - self.x_prev
                    delta_g = self.J.T.dot(self.v) - self.J_prev.T.dot(self.v)
                    self.H.update(delta_x, delta_g)

        self._update_hess_impl = update_hess

        if isinstance(hess, HessianUpdateStrategy):
            def update_x(x):
                self._update_jac()
                self.x_prev = self.x
                self.J_prev = self.J
                _x = xpx.atleast_nd(self.xp.asarray(x), ndim=1, xp=self.xp)
                self.x = self.xp.astype(_x, self.x_dtype)
                self.f_updated = False
                self.J_updated = False
                self.H_updated = False
                self._update_hess()
        else:
            def update_x(x):
                _x = xpx.atleast_nd(self.xp.asarray(x), ndim=1, xp=self.xp)
                self.x = self.xp.astype(_x, self.x_dtype)
                self.f_updated = False
                self.J_updated = False
                self.H_updated = False

        self._update_x_impl = update_x

    def _update_v(self, v):
        if not np.array_equal(v, self.v):
            self.v = v
            self.H_updated = False

    def _update_x(self, x):
        if not np.array_equal(x, self.x):
            self._update_x_impl(x)

    def _update_fun(self):
        if not self.f_updated:
            self._update_fun_impl()
            self.f_updated = True

    def _update_jac(self):
        if not self.J_updated:
            self._update_jac_impl()
            self.J_updated = True

    def _update_hess(self):
        if not self.H_updated:
            self._update_hess_impl()
            self.H_updated = True

    def fun(self, x):
        self._update_x(x)
        self._update_fun()
        return self.f

    def jac(self, x):
        self._update_x(x)
        self._update_jac()
        return self.J

    def hess(self, x, v):
        # v should be updated before x.
        self._update_v(v)
        self._update_x(x)
        self._update_hess()
        return self.H


class LinearVectorFunction:
    """Linear vector function and its derivatives.

    Defines a linear function F = A x, where x is N-D vector and
    A is m-by-n matrix. The Jacobian is constant and equals to A. The Hessian
    is identically zero and it is returned as a csr matrix.
    """
    def __init__(self, A, x0, sparse_jacobian):
        if sparse_jacobian or sparse_jacobian is None and sps.issparse(A):
            self.J = sps.csr_matrix(A)
            self.sparse_jacobian = True
        elif sps.issparse(A):
            self.J = A.toarray()
            self.sparse_jacobian = False
        else:
            # np.asarray makes sure A is ndarray and not matrix
            self.J = np.atleast_2d(np.asarray(A))
            self.sparse_jacobian = False

        self.m, self.n = self.J.shape

        self.xp = xp = array_namespace(x0)
        _x = xpx.atleast_nd(xp.asarray(x0), ndim=1, xp=xp)
        _dtype = xp.float64
        if xp.isdtype(_x.dtype, "real floating"):
            _dtype = _x.dtype

        # promotes to floating
        self.x = xp.astype(_x, _dtype)
        self.x_dtype = _dtype

        self.f = self.J.dot(self.x)
        self.f_updated = True

        self.v = np.zeros(self.m, dtype=float)
        self.H = sps.csr_matrix((self.n, self.n))

    def _update_x(self, x):
        if not np.array_equal(x, self.x):
            _x = xpx.atleast_nd(self.xp.asarray(x), ndim=1, xp=self.xp)
            self.x = self.xp.astype(_x, self.x_dtype)
            self.f_updated = False

    def fun(self, x):
        self._update_x(x)
        if not self.f_updated:
            self.f = self.J.dot(x)
            self.f_updated = True
        return self.f

    def jac(self, x):
        self._update_x(x)
        return self.J

    def hess(self, x, v):
        self._update_x(x)
        self.v = v
        return self.H


class IdentityVectorFunction(LinearVectorFunction):
    """Identity vector function and its derivatives.

    The Jacobian is the identity matrix, returned as a dense array when
    `sparse_jacobian=False` and as a csr matrix otherwise. The Hessian is
    identically zero and it is returned as a csr matrix.
    """
    def __init__(self, x0, sparse_jacobian):
        n = len(x0)
        if sparse_jacobian or sparse_jacobian is None:
            A = sps.eye(n, format='csr')
            sparse_jacobian = True
        else:
            A = np.eye(n)
            sparse_jacobian = False
        super().__init__(A, x0, sparse_jacobian)