-
-
Notifications
You must be signed in to change notification settings - Fork 5.1k
/
rk.py
599 lines (520 loc) · 22.2 KB
/
rk.py
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
import numpy as np
from .base import OdeSolver, DenseOutput
from .common import (validate_max_step, validate_tol, select_initial_step,
norm, warn_extraneous, validate_first_step)
from . import dop853_coefficients
# Multiply steps computed from asymptotic behaviour of errors by this.
SAFETY = 0.9
MIN_FACTOR = 0.2 # Minimum allowed decrease in a step size.
MAX_FACTOR = 10 # Maximum allowed increase in a step size.
def rk_step(fun, t, y, f, h, A, B, C, K):
"""Perform a single Runge-Kutta step.
This function computes a prediction of an explicit Runge-Kutta method and
also estimates the error of a less accurate method.
Notation for Butcher tableau is as in [1]_.
Parameters
----------
fun : callable
Right-hand side of the system.
t : float
Current time.
y : ndarray, shape (n,)
Current state.
f : ndarray, shape (n,)
Current value of the derivative, i.e., ``fun(x, y)``.
h : float
Step to use.
A : ndarray, shape (n_stages, n_stages)
Coefficients for combining previous RK stages to compute the next
stage. For explicit methods the coefficients at and above the main
diagonal are zeros.
B : ndarray, shape (n_stages,)
Coefficients for combining RK stages for computing the final
prediction.
C : ndarray, shape (n_stages,)
Coefficients for incrementing time for consecutive RK stages.
The value for the first stage is always zero.
K : ndarray, shape (n_stages + 1, n)
Storage array for putting RK stages here. Stages are stored in rows.
The last row is a linear combination of the previous rows with
coefficients
Returns
-------
y_new : ndarray, shape (n,)
Solution at t + h computed with a higher accuracy.
f_new : ndarray, shape (n,)
Derivative ``fun(t + h, y_new)``.
References
----------
.. [1] E. Hairer, S. P. Norsett G. Wanner, "Solving Ordinary Differential
Equations I: Nonstiff Problems", Sec. II.4.
"""
K[0] = f
for s, (a, c) in enumerate(zip(A[1:], C[1:]), start=1):
dy = np.dot(K[:s].T, a[:s]) * h
K[s] = fun(t + c * h, y + dy)
y_new = y + h * np.dot(K[:-1].T, B)
f_new = fun(t + h, y_new)
K[-1] = f_new
return y_new, f_new
class RungeKutta(OdeSolver):
"""Base class for explicit Runge-Kutta methods."""
C: np.ndarray = NotImplemented
A: np.ndarray = NotImplemented
B: np.ndarray = NotImplemented
E: np.ndarray = NotImplemented
P: np.ndarray = NotImplemented
order: int = NotImplemented
error_estimator_order: int = NotImplemented
n_stages: int = NotImplemented
def __init__(self, fun, t0, y0, t_bound, max_step=np.inf,
rtol=1e-3, atol=1e-6, vectorized=False,
first_step=None, **extraneous):
warn_extraneous(extraneous)
super().__init__(fun, t0, y0, t_bound, vectorized,
support_complex=True)
self.y_old = None
self.max_step = validate_max_step(max_step)
self.rtol, self.atol = validate_tol(rtol, atol, self.n)
self.f = self.fun(self.t, self.y)
if first_step is None:
self.h_abs = select_initial_step(
self.fun, self.t, self.y, self.f, self.direction,
self.error_estimator_order, self.rtol, self.atol)
else:
self.h_abs = validate_first_step(first_step, t0, t_bound)
self.K = np.empty((self.n_stages + 1, self.n), dtype=self.y.dtype)
self.error_exponent = -1 / (self.error_estimator_order + 1)
self.h_previous = None
def _estimate_error(self, K, h):
return np.dot(K.T, self.E) * h
def _estimate_error_norm(self, K, h, scale):
return norm(self._estimate_error(K, h) / scale)
def _step_impl(self):
t = self.t
y = self.y
max_step = self.max_step
rtol = self.rtol
atol = self.atol
min_step = 10 * np.abs(np.nextafter(t, self.direction * np.inf) - t)
if self.h_abs > max_step:
h_abs = max_step
elif self.h_abs < min_step:
h_abs = min_step
else:
h_abs = self.h_abs
step_accepted = False
step_rejected = False
while not step_accepted:
if h_abs < min_step:
return False, self.TOO_SMALL_STEP
h = h_abs * self.direction
t_new = t + h
if self.direction * (t_new - self.t_bound) > 0:
t_new = self.t_bound
h = t_new - t
h_abs = np.abs(h)
y_new, f_new = rk_step(self.fun, t, y, self.f, h, self.A,
self.B, self.C, self.K)
scale = atol + np.maximum(np.abs(y), np.abs(y_new)) * rtol
error_norm = self._estimate_error_norm(self.K, h, scale)
if error_norm < 1:
if error_norm == 0:
factor = MAX_FACTOR
else:
factor = min(MAX_FACTOR,
SAFETY * error_norm ** self.error_exponent)
if step_rejected:
factor = min(1, factor)
h_abs *= factor
step_accepted = True
else:
h_abs *= max(MIN_FACTOR,
SAFETY * error_norm ** self.error_exponent)
step_rejected = True
self.h_previous = h
self.y_old = y
self.t = t_new
self.y = y_new
self.h_abs = h_abs
self.f = f_new
return True, None
def _dense_output_impl(self):
Q = self.K.T.dot(self.P)
return RkDenseOutput(self.t_old, self.t, self.y_old, Q)
class RK23(RungeKutta):
"""Explicit Runge-Kutta method of order 3(2).
This uses the Bogacki-Shampine pair of formulas [1]_. The error is controlled
assuming accuracy of the second-order method, but steps are taken using the
third-order accurate formula (local extrapolation is done). A cubic Hermite
polynomial is used for the dense output.
Can be applied in the complex domain.
Parameters
----------
fun : callable
Right-hand side of the system: the time derivative of the state ``y``
at time ``t``. The calling signature is ``fun(t, y)``, where ``t`` is a
scalar and ``y`` is an ndarray with ``len(y) = len(y0)``. ``fun`` must
return an array of the same shape as ``y``. See `vectorized` for more
information.
t0 : float
Initial time.
y0 : array_like, shape (n,)
Initial state.
t_bound : float
Boundary time - the integration won't continue beyond it. It also
determines the direction of the integration.
first_step : float or None, optional
Initial step size. Default is ``None`` which means that the algorithm
should choose.
max_step : float, optional
Maximum allowed step size. Default is np.inf, i.e., the step size is not
bounded and determined solely by the solver.
rtol, atol : float and array_like, optional
Relative and absolute tolerances. The solver keeps the local error
estimates less than ``atol + rtol * abs(y)``. Here `rtol` controls a
relative accuracy (number of correct digits), while `atol` controls
absolute accuracy (number of correct decimal places). To achieve the
desired `rtol`, set `atol` to be smaller than the smallest value that
can be expected from ``rtol * abs(y)`` so that `rtol` dominates the
allowable error. If `atol` is larger than ``rtol * abs(y)`` the
number of correct digits is not guaranteed. Conversely, to achieve the
desired `atol` set `rtol` such that ``rtol * abs(y)`` is always smaller
than `atol`. If components of y have different scales, it might be
beneficial to set different `atol` values for different components by
passing array_like with shape (n,) for `atol`. Default values are
1e-3 for `rtol` and 1e-6 for `atol`.
vectorized : bool, optional
Whether `fun` may be called in a vectorized fashion. False (default)
is recommended for this solver.
If ``vectorized`` is False, `fun` will always be called with ``y`` of
shape ``(n,)``, where ``n = len(y0)``.
If ``vectorized`` is True, `fun` may be called with ``y`` of shape
``(n, k)``, where ``k`` is an integer. In this case, `fun` must behave
such that ``fun(t, y)[:, i] == fun(t, y[:, i])`` (i.e. each column of
the returned array is the time derivative of the state corresponding
with a column of ``y``).
Setting ``vectorized=True`` allows for faster finite difference
approximation of the Jacobian by methods 'Radau' and 'BDF', but
will result in slower execution for this solver.
Attributes
----------
n : int
Number of equations.
status : string
Current status of the solver: 'running', 'finished' or 'failed'.
t_bound : float
Boundary time.
direction : float
Integration direction: +1 or -1.
t : float
Current time.
y : ndarray
Current state.
t_old : float
Previous time. None if no steps were made yet.
step_size : float
Size of the last successful step. None if no steps were made yet.
nfev : int
Number evaluations of the system's right-hand side.
njev : int
Number of evaluations of the Jacobian. Is always 0 for this solver as it does not use the Jacobian.
nlu : int
Number of LU decompositions. Is always 0 for this solver.
References
----------
.. [1] P. Bogacki, L.F. Shampine, "A 3(2) Pair of Runge-Kutta Formulas",
Appl. Math. Lett. Vol. 2, No. 4. pp. 321-325, 1989.
"""
order = 3
error_estimator_order = 2
n_stages = 3
C = np.array([0, 1/2, 3/4])
A = np.array([
[0, 0, 0],
[1/2, 0, 0],
[0, 3/4, 0]
])
B = np.array([2/9, 1/3, 4/9])
E = np.array([5/72, -1/12, -1/9, 1/8])
P = np.array([[1, -4 / 3, 5 / 9],
[0, 1, -2/3],
[0, 4/3, -8/9],
[0, -1, 1]])
class RK45(RungeKutta):
"""Explicit Runge-Kutta method of order 5(4).
This uses the Dormand-Prince pair of formulas [1]_. The error is controlled
assuming accuracy of the fourth-order method accuracy, but steps are taken
using the fifth-order accurate formula (local extrapolation is done).
A quartic interpolation polynomial is used for the dense output [2]_.
Can be applied in the complex domain.
Parameters
----------
fun : callable
Right-hand side of the system. The calling signature is ``fun(t, y)``.
Here ``t`` is a scalar, and there are two options for the ndarray ``y``:
It can either have shape (n,); then ``fun`` must return array_like with
shape (n,). Alternatively it can have shape (n, k); then ``fun``
must return an array_like with shape (n, k), i.e., each column
corresponds to a single column in ``y``. The choice between the two
options is determined by `vectorized` argument (see below).
t0 : float
Initial time.
y0 : array_like, shape (n,)
Initial state.
t_bound : float
Boundary time - the integration won't continue beyond it. It also
determines the direction of the integration.
first_step : float or None, optional
Initial step size. Default is ``None`` which means that the algorithm
should choose.
max_step : float, optional
Maximum allowed step size. Default is np.inf, i.e., the step size is not
bounded and determined solely by the solver.
rtol, atol : float and array_like, optional
Relative and absolute tolerances. The solver keeps the local error
estimates less than ``atol + rtol * abs(y)``. Here `rtol` controls a
relative accuracy (number of correct digits), while `atol` controls
absolute accuracy (number of correct decimal places). To achieve the
desired `rtol`, set `atol` to be smaller than the smallest value that
can be expected from ``rtol * abs(y)`` so that `rtol` dominates the
allowable error. If `atol` is larger than ``rtol * abs(y)`` the
number of correct digits is not guaranteed. Conversely, to achieve the
desired `atol` set `rtol` such that ``rtol * abs(y)`` is always smaller
than `atol`. If components of y have different scales, it might be
beneficial to set different `atol` values for different components by
passing array_like with shape (n,) for `atol`. Default values are
1e-3 for `rtol` and 1e-6 for `atol`.
vectorized : bool, optional
Whether `fun` is implemented in a vectorized fashion. Default is False.
Attributes
----------
n : int
Number of equations.
status : string
Current status of the solver: 'running', 'finished' or 'failed'.
t_bound : float
Boundary time.
direction : float
Integration direction: +1 or -1.
t : float
Current time.
y : ndarray
Current state.
t_old : float
Previous time. None if no steps were made yet.
step_size : float
Size of the last successful step. None if no steps were made yet.
nfev : int
Number evaluations of the system's right-hand side.
njev : int
Number of evaluations of the Jacobian. Is always 0 for this solver as it does not use the Jacobian.
nlu : int
Number of LU decompositions. Is always 0 for this solver.
References
----------
.. [1] J. R. Dormand, P. J. Prince, "A family of embedded Runge-Kutta
formulae", Journal of Computational and Applied Mathematics, Vol. 6,
No. 1, pp. 19-26, 1980.
.. [2] L. W. Shampine, "Some Practical Runge-Kutta Formulas", Mathematics
of Computation,, Vol. 46, No. 173, pp. 135-150, 1986.
"""
order = 5
error_estimator_order = 4
n_stages = 6
C = np.array([0, 1/5, 3/10, 4/5, 8/9, 1])
A = np.array([
[0, 0, 0, 0, 0],
[1/5, 0, 0, 0, 0],
[3/40, 9/40, 0, 0, 0],
[44/45, -56/15, 32/9, 0, 0],
[19372/6561, -25360/2187, 64448/6561, -212/729, 0],
[9017/3168, -355/33, 46732/5247, 49/176, -5103/18656]
])
B = np.array([35/384, 0, 500/1113, 125/192, -2187/6784, 11/84])
E = np.array([-71/57600, 0, 71/16695, -71/1920, 17253/339200, -22/525,
1/40])
# Corresponds to the optimum value of c_6 from [2]_.
P = np.array([
[1, -8048581381/2820520608, 8663915743/2820520608,
-12715105075/11282082432],
[0, 0, 0, 0],
[0, 131558114200/32700410799, -68118460800/10900136933,
87487479700/32700410799],
[0, -1754552775/470086768, 14199869525/1410260304,
-10690763975/1880347072],
[0, 127303824393/49829197408, -318862633887/49829197408,
701980252875 / 199316789632],
[0, -282668133/205662961, 2019193451/616988883, -1453857185/822651844],
[0, 40617522/29380423, -110615467/29380423, 69997945/29380423]])
class DOP853(RungeKutta):
"""Explicit Runge-Kutta method of order 8.
This is a Python implementation of "DOP853" algorithm originally written
in Fortran [1]_, [2]_. Note that this is not a literate translation, but
the algorithmic core and coefficients are the same.
Can be applied in the complex domain.
Parameters
----------
fun : callable
Right-hand side of the system. The calling signature is ``fun(t, y)``.
Here, ``t`` is a scalar, and there are two options for the ndarray ``y``:
It can either have shape (n,); then ``fun`` must return array_like with
shape (n,). Alternatively it can have shape (n, k); then ``fun``
must return an array_like with shape (n, k), i.e. each column
corresponds to a single column in ``y``. The choice between the two
options is determined by `vectorized` argument (see below).
t0 : float
Initial time.
y0 : array_like, shape (n,)
Initial state.
t_bound : float
Boundary time - the integration won't continue beyond it. It also
determines the direction of the integration.
first_step : float or None, optional
Initial step size. Default is ``None`` which means that the algorithm
should choose.
max_step : float, optional
Maximum allowed step size. Default is np.inf, i.e. the step size is not
bounded and determined solely by the solver.
rtol, atol : float and array_like, optional
Relative and absolute tolerances. The solver keeps the local error
estimates less than ``atol + rtol * abs(y)``. Here `rtol` controls a
relative accuracy (number of correct digits), while `atol` controls
absolute accuracy (number of correct decimal places). To achieve the
desired `rtol`, set `atol` to be smaller than the smallest value that
can be expected from ``rtol * abs(y)`` so that `rtol` dominates the
allowable error. If `atol` is larger than ``rtol * abs(y)`` the
number of correct digits is not guaranteed. Conversely, to achieve the
desired `atol` set `rtol` such that ``rtol * abs(y)`` is always smaller
than `atol`. If components of y have different scales, it might be
beneficial to set different `atol` values for different components by
passing array_like with shape (n,) for `atol`. Default values are
1e-3 for `rtol` and 1e-6 for `atol`.
vectorized : bool, optional
Whether `fun` is implemented in a vectorized fashion. Default is False.
Attributes
----------
n : int
Number of equations.
status : string
Current status of the solver: 'running', 'finished' or 'failed'.
t_bound : float
Boundary time.
direction : float
Integration direction: +1 or -1.
t : float
Current time.
y : ndarray
Current state.
t_old : float
Previous time. None if no steps were made yet.
step_size : float
Size of the last successful step. None if no steps were made yet.
nfev : int
Number evaluations of the system's right-hand side.
njev : int
Number of evaluations of the Jacobian. Is always 0 for this solver
as it does not use the Jacobian.
nlu : int
Number of LU decompositions. Is always 0 for this solver.
References
----------
.. [1] E. Hairer, S. P. Norsett G. Wanner, "Solving Ordinary Differential
Equations I: Nonstiff Problems", Sec. II.
.. [2] `Page with original Fortran code of DOP853
<http://www.unige.ch/~hairer/software.html>`_.
"""
n_stages = dop853_coefficients.N_STAGES
order = 8
error_estimator_order = 7
A = dop853_coefficients.A[:n_stages, :n_stages]
B = dop853_coefficients.B
C = dop853_coefficients.C[:n_stages]
E3 = dop853_coefficients.E3
E5 = dop853_coefficients.E5
D = dop853_coefficients.D
A_EXTRA = dop853_coefficients.A[n_stages + 1:]
C_EXTRA = dop853_coefficients.C[n_stages + 1:]
def __init__(self, fun, t0, y0, t_bound, max_step=np.inf,
rtol=1e-3, atol=1e-6, vectorized=False,
first_step=None, **extraneous):
super().__init__(fun, t0, y0, t_bound, max_step, rtol, atol,
vectorized, first_step, **extraneous)
self.K_extended = np.empty((dop853_coefficients.N_STAGES_EXTENDED,
self.n), dtype=self.y.dtype)
self.K = self.K_extended[:self.n_stages + 1]
def _estimate_error(self, K, h): # Left for testing purposes.
err5 = np.dot(K.T, self.E5)
err3 = np.dot(K.T, self.E3)
denom = np.hypot(np.abs(err5), 0.1 * np.abs(err3))
correction_factor = np.ones_like(err5)
mask = denom > 0
correction_factor[mask] = np.abs(err5[mask]) / denom[mask]
return h * err5 * correction_factor
def _estimate_error_norm(self, K, h, scale):
err5 = np.dot(K.T, self.E5) / scale
err3 = np.dot(K.T, self.E3) / scale
err5_norm_2 = np.linalg.norm(err5)**2
err3_norm_2 = np.linalg.norm(err3)**2
if err5_norm_2 == 0 and err3_norm_2 == 0:
return 0.0
denom = err5_norm_2 + 0.01 * err3_norm_2
return np.abs(h) * err5_norm_2 / np.sqrt(denom * len(scale))
def _dense_output_impl(self):
K = self.K_extended
h = self.h_previous
for s, (a, c) in enumerate(zip(self.A_EXTRA, self.C_EXTRA),
start=self.n_stages + 1):
dy = np.dot(K[:s].T, a[:s]) * h
K[s] = self.fun(self.t_old + c * h, self.y_old + dy)
F = np.empty((dop853_coefficients.INTERPOLATOR_POWER, self.n),
dtype=self.y_old.dtype)
f_old = K[0]
delta_y = self.y - self.y_old
F[0] = delta_y
F[1] = h * f_old - delta_y
F[2] = 2 * delta_y - h * (self.f + f_old)
F[3:] = h * np.dot(self.D, K)
return Dop853DenseOutput(self.t_old, self.t, self.y_old, F)
class RkDenseOutput(DenseOutput):
def __init__(self, t_old, t, y_old, Q):
super().__init__(t_old, t)
self.h = t - t_old
self.Q = Q
self.order = Q.shape[1] - 1
self.y_old = y_old
def _call_impl(self, t):
x = (t - self.t_old) / self.h
if t.ndim == 0:
p = np.tile(x, self.order + 1)
p = np.cumprod(p)
else:
p = np.tile(x, (self.order + 1, 1))
p = np.cumprod(p, axis=0)
y = self.h * np.dot(self.Q, p)
if y.ndim == 2:
y += self.y_old[:, None]
else:
y += self.y_old
return y
class Dop853DenseOutput(DenseOutput):
def __init__(self, t_old, t, y_old, F):
super().__init__(t_old, t)
self.h = t - t_old
self.F = F
self.y_old = y_old
def _call_impl(self, t):
x = (t - self.t_old) / self.h
if t.ndim == 0:
y = np.zeros_like(self.y_old)
else:
x = x[:, None]
y = np.zeros((len(x), len(self.y_old)), dtype=self.y_old.dtype)
for i, f in enumerate(reversed(self.F)):
y += f
if i % 2 == 0:
y *= x
else:
y *= 1 - x
y += self.y_old
return y.T