File size: 36,056 Bytes
b200bda
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
# contributed to mpmath by Kristopher L. Kuhlman, February 2017
# contributed to mpmath by Guillermo Navas-Palencia, February 2022

class InverseLaplaceTransform(object):
    r"""
    Inverse Laplace transform methods are implemented using this
    class, in order to simplify the code and provide a common
    infrastructure.

    Implement a custom inverse Laplace transform algorithm by
    subclassing :class:`InverseLaplaceTransform` and implementing the
    appropriate methods. The subclass can then be used by
    :func:`~mpmath.invertlaplace` by passing it as the *method*
    argument.
    """

    def __init__(self, ctx):
        self.ctx = ctx

    def calc_laplace_parameter(self, t, **kwargs):
        r"""
        Determine the vector of Laplace parameter values needed for an
        algorithm, this will depend on the choice of algorithm (de
        Hoog is default), the algorithm-specific parameters passed (or
        default ones), and desired time.
        """
        raise NotImplementedError

    def calc_time_domain_solution(self, fp):
        r"""
        Compute the time domain solution, after computing the
        Laplace-space function evaluations at the abscissa required
        for the algorithm. Abscissa computed for one algorithm are
        typically not useful for another algorithm.
        """
        raise NotImplementedError


class FixedTalbot(InverseLaplaceTransform):

    def calc_laplace_parameter(self, t, **kwargs):
        r"""The "fixed" Talbot method deforms the Bromwich contour towards
        `-\infty` in the shape of a parabola. Traditionally the Talbot
        algorithm has adjustable parameters, but the "fixed" version
        does not. The `r` parameter could be passed in as a parameter,
        if you want to override the default given by (Abate & Valko,
        2004).

        The Laplace parameter is sampled along a parabola opening
        along the negative imaginary axis, with the base of the
        parabola along the real axis at
        `p=\frac{r}{t_\mathrm{max}}`. As the number of terms used in
        the approximation (degree) grows, the abscissa required for
        function evaluation tend towards `-\infty`, requiring high
        precision to prevent overflow.  If any poles, branch cuts or
        other singularities exist such that the deformed Bromwich
        contour lies to the left of the singularity, the method will
        fail.

        **Optional arguments**

        :class:`~mpmath.calculus.inverselaplace.FixedTalbot.calc_laplace_parameter`
        recognizes the following keywords

        *tmax*
            maximum time associated with vector of times
            (typically just the time requested)
        *degree*
            integer order of approximation (M = number of terms)
        *r*
            abscissa for `p_0` (otherwise computed using rule
            of thumb `2M/5`)

        The working precision will be increased according to a rule of
        thumb. If 'degree' is not specified, the working precision and
        degree are chosen to hopefully achieve the dps of the calling
        context. If 'degree' is specified, the working precision is
        chosen to achieve maximum resulting precision for the
        specified degree.

        .. math ::

            p_0=\frac{r}{t}

        .. math ::

            p_i=\frac{i r \pi}{Mt_\mathrm{max}}\left[\cot\left(
            \frac{i\pi}{M}\right) + j \right] \qquad 1\le i <M

        where `j=\sqrt{-1}`, `r=2M/5`, and `t_\mathrm{max}` is the
        maximum specified time.

        """

        # required
        # ------------------------------
        # time of desired approximation
        self.t = self.ctx.convert(t)

        # optional
        # ------------------------------
        # maximum time desired (used for scaling) default is requested
        # time.
        self.tmax = self.ctx.convert(kwargs.get('tmax', self.t))

        # empirical relationships used here based on a linear fit of
        # requested and delivered dps for exponentially decaying time
        # functions for requested dps up to 512.

        if 'degree' in kwargs:
            self.degree = kwargs['degree']
            self.dps_goal = self.degree
        else:
            self.dps_goal = int(1.72*self.ctx.dps)
            self.degree = max(12, int(1.38*self.dps_goal))

        M = self.degree

        # this is adjusting the dps of the calling context hopefully
        # the caller doesn't monkey around with it between calling
        # this routine and calc_time_domain_solution()
        self.dps_orig = self.ctx.dps
        self.ctx.dps = self.dps_goal

        # Abate & Valko rule of thumb for r parameter
        self.r = kwargs.get('r', self.ctx.fraction(2, 5)*M)

        self.theta = self.ctx.linspace(0.0, self.ctx.pi, M+1)

        self.cot_theta = self.ctx.matrix(M, 1)
        self.cot_theta[0] = 0  # not used

        # all but time-dependent part of p
        self.delta = self.ctx.matrix(M, 1)
        self.delta[0] = self.r

        for i in range(1, M):
            self.cot_theta[i] = self.ctx.cot(self.theta[i])
            self.delta[i] = self.r*self.theta[i]*(self.cot_theta[i] + 1j)

        self.p = self.ctx.matrix(M, 1)
        self.p = self.delta/self.tmax

        # NB: p is complex (mpc)

    def calc_time_domain_solution(self, fp, t, manual_prec=False):
        r"""The fixed Talbot time-domain solution is computed from the
        Laplace-space function evaluations using

        .. math ::

            f(t,M)=\frac{2}{5t}\sum_{k=0}^{M-1}\Re \left[
            \gamma_k \bar{f}(p_k)\right]

        where

        .. math ::

            \gamma_0 = \frac{1}{2}e^{r}\bar{f}(p_0)

        .. math ::

            \gamma_k = e^{tp_k}\left\lbrace 1 + \frac{jk\pi}{M}\left[1 +
            \cot \left( \frac{k \pi}{M} \right)^2 \right] - j\cot\left(
            \frac{k \pi}{M}\right)\right \rbrace \qquad 1\le k<M.

        Again, `j=\sqrt{-1}`.

        Before calling this function, call
        :class:`~mpmath.calculus.inverselaplace.FixedTalbot.calc_laplace_parameter`
        to set the parameters and compute the required coefficients.

        **References**

        1. Abate, J., P. Valko (2004). Multi-precision Laplace
           transform inversion. *International Journal for Numerical
           Methods in Engineering* 60:979-993,
           http://dx.doi.org/10.1002/nme.995
        2. Talbot, A. (1979). The accurate numerical inversion of
           Laplace transforms. *IMA Journal of Applied Mathematics*
           23(1):97, http://dx.doi.org/10.1093/imamat/23.1.97
        """

        # required
        # ------------------------------
        self.t = self.ctx.convert(t)

        # assume fp was computed from p matrix returned from
        # calc_laplace_parameter(), so is already a list or matrix of
        # mpmath 'mpc' types

        # these were computed in previous call to
        # calc_laplace_parameter()
        theta = self.theta
        delta = self.delta
        M = self.degree
        p = self.p
        r = self.r

        ans = self.ctx.matrix(M, 1)
        ans[0] = self.ctx.exp(delta[0])*fp[0]/2

        for i in range(1, M):
            ans[i] = self.ctx.exp(delta[i])*fp[i]*(
                1 + 1j*theta[i]*(1 + self.cot_theta[i]**2) -
                1j*self.cot_theta[i])

        result = self.ctx.fraction(2, 5)*self.ctx.fsum(ans)/self.t

        # setting dps back to value when calc_laplace_parameter was
        # called, unless flag is set.
        if not manual_prec:
            self.ctx.dps = self.dps_orig

        return result.real


# ****************************************

class Stehfest(InverseLaplaceTransform):

    def calc_laplace_parameter(self, t, **kwargs):
        r"""
        The Gaver-Stehfest method is a discrete approximation of the
        Widder-Post inversion algorithm, rather than a direct
        approximation of the Bromwich contour integral.

        The method abscissa along the real axis, and therefore has
        issues inverting oscillatory functions (which have poles in
        pairs away from the real axis).

        The working precision will be increased according to a rule of
        thumb. If 'degree' is not specified, the working precision and
        degree are chosen to hopefully achieve the dps of the calling
        context. If 'degree' is specified, the working precision is
        chosen to achieve maximum resulting precision for the
        specified degree.

        .. math ::

            p_k = \frac{k \log 2}{t} \qquad 1 \le k \le M
        """

        # required
        # ------------------------------
        # time of desired approximation
        self.t = self.ctx.convert(t)

        # optional
        # ------------------------------

        # empirical relationships used here based on a linear fit of
        # requested and delivered dps for exponentially decaying time
        # functions for requested dps up to 512.

        if 'degree' in kwargs:
            self.degree = kwargs['degree']
            self.dps_goal = int(1.38*self.degree)
        else:
            self.dps_goal = int(2.93*self.ctx.dps)
            self.degree = max(16, self.dps_goal)

        # _coeff routine requires even degree
        if self.degree % 2 > 0:
            self.degree += 1

        M = self.degree

        # this is adjusting the dps of the calling context
        # hopefully the caller doesn't monkey around with it
        # between calling this routine and calc_time_domain_solution()
        self.dps_orig = self.ctx.dps
        self.ctx.dps = self.dps_goal

        self.V = self._coeff()
        self.p = self.ctx.matrix(self.ctx.arange(1, M+1))*self.ctx.ln2/self.t

        # NB: p is real (mpf)

    def _coeff(self):
        r"""Salzer summation weights (aka, "Stehfest coefficients")
        only depend on the approximation order (M) and the precision"""

        M = self.degree
        M2 = int(M/2)  # checked earlier that M is even

        V = self.ctx.matrix(M, 1)

        # Salzer summation weights
        # get very large in magnitude and oscillate in sign,
        # if the precision is not high enough, there will be
        # catastrophic cancellation
        for k in range(1, M+1):
            z = self.ctx.matrix(min(k, M2)+1, 1)
            for j in range(int((k+1)/2), min(k, M2)+1):
                z[j] = (self.ctx.power(j, M2)*self.ctx.fac(2*j)/
                        (self.ctx.fac(M2-j)*self.ctx.fac(j)*
                         self.ctx.fac(j-1)*self.ctx.fac(k-j)*
                         self.ctx.fac(2*j-k)))
            V[k-1] = self.ctx.power(-1, k+M2)*self.ctx.fsum(z)

        return V

    def calc_time_domain_solution(self, fp, t, manual_prec=False):
        r"""Compute time-domain Stehfest algorithm solution.

        .. math ::

            f(t,M) = \frac{\log 2}{t} \sum_{k=1}^{M} V_k \bar{f}\left(
            p_k \right)

        where

        .. math ::

            V_k = (-1)^{k + N/2} \sum^{\min(k,N/2)}_{i=\lfloor(k+1)/2 \rfloor}
            \frac{i^{\frac{N}{2}}(2i)!}{\left(\frac{N}{2}-i \right)! \, i! \,
            \left(i-1 \right)! \, \left(k-i\right)! \, \left(2i-k \right)!}

        As the degree increases, the abscissa (`p_k`) only increase
        linearly towards `\infty`, but the Stehfest coefficients
        (`V_k`) alternate in sign and increase rapidly in sign,
        requiring high precision to prevent overflow or loss of
        significance when evaluating the sum.

        **References**

        1. Widder, D. (1941). *The Laplace Transform*. Princeton.
        2. Stehfest, H. (1970). Algorithm 368: numerical inversion of
           Laplace transforms. *Communications of the ACM* 13(1):47-49,
           http://dx.doi.org/10.1145/361953.361969

        """

        # required
        self.t = self.ctx.convert(t)

        # assume fp was computed from p matrix returned from
        # calc_laplace_parameter(), so is already
        # a list or matrix of mpmath 'mpf' types

        result = self.ctx.fdot(self.V, fp)*self.ctx.ln2/self.t

        # setting dps back to value when calc_laplace_parameter was called
        if not manual_prec:
            self.ctx.dps = self.dps_orig

        # ignore any small imaginary part
        return result.real


# ****************************************

class deHoog(InverseLaplaceTransform):

    def calc_laplace_parameter(self, t, **kwargs):
        r"""the de Hoog, Knight & Stokes algorithm is an
        accelerated form of the Fourier series numerical
        inverse Laplace transform algorithms.

        .. math ::

            p_k = \gamma + \frac{jk}{T} \qquad 0 \le k < 2M+1

        where

        .. math ::

            \gamma = \alpha - \frac{\log \mathrm{tol}}{2T},

        `j=\sqrt{-1}`, `T = 2t_\mathrm{max}` is a scaled time,
        `\alpha=10^{-\mathrm{dps\_goal}}` is the real part of the
        rightmost pole or singularity, which is chosen based on the
        desired accuracy (assuming the rightmost singularity is 0),
        and `\mathrm{tol}=10\alpha` is the desired tolerance, which is
        chosen in relation to `\alpha`.`

        When increasing the degree, the abscissa increase towards
        `j\infty`, but more slowly than the fixed Talbot
        algorithm. The de Hoog et al. algorithm typically does better
        with oscillatory functions of time, and less well-behaved
        functions. The method tends to be slower than the Talbot and
        Stehfest algorithsm, especially so at very high precision
        (e.g., `>500` digits precision).

        """

        # required
        # ------------------------------
        self.t = self.ctx.convert(t)

        # optional
        # ------------------------------
        self.tmax = kwargs.get('tmax', self.t)

        # empirical relationships used here based on a linear fit of
        # requested and delivered dps for exponentially decaying time
        # functions for requested dps up to 512.

        if 'degree' in kwargs:
            self.degree = kwargs['degree']
            self.dps_goal = int(1.38*self.degree)
        else:
            self.dps_goal = int(self.ctx.dps*1.36)
            self.degree = max(10, self.dps_goal)

        # 2*M+1 terms in approximation
        M = self.degree

        # adjust alpha component of abscissa of convergence for higher
        # precision
        tmp = self.ctx.power(10.0, -self.dps_goal)
        self.alpha = self.ctx.convert(kwargs.get('alpha', tmp))

        # desired tolerance (here simply related to alpha)
        self.tol = self.ctx.convert(kwargs.get('tol', self.alpha*10.0))
        self.np = 2*self.degree+1  # number of terms in approximation

        # this is adjusting the dps of the calling context
        # hopefully the caller doesn't monkey around with it
        # between calling this routine and calc_time_domain_solution()
        self.dps_orig = self.ctx.dps
        self.ctx.dps = self.dps_goal

        # scaling factor (likely tun-able, but 2 is typical)
        self.scale = kwargs.get('scale', 2)
        self.T = self.ctx.convert(kwargs.get('T', self.scale*self.tmax))

        self.p = self.ctx.matrix(2*M+1, 1)
        self.gamma = self.alpha - self.ctx.log(self.tol)/(self.scale*self.T)
        self.p = (self.gamma + self.ctx.pi*
                  self.ctx.matrix(self.ctx.arange(self.np))/self.T*1j)

        # NB: p is complex (mpc)

    def calc_time_domain_solution(self, fp, t, manual_prec=False):
        r"""Calculate time-domain solution for
        de Hoog, Knight & Stokes algorithm.

        The un-accelerated Fourier series approach is:

        .. math ::

            f(t,2M+1) = \frac{e^{\gamma t}}{T} \sum_{k=0}^{2M}{}^{'}
            \Re\left[\bar{f}\left( p_k \right)
            e^{i\pi t/T} \right],

        where the prime on the summation indicates the first term is halved.

        This simplistic approach requires so many function evaluations
        that it is not practical. Non-linear acceleration is
        accomplished via Pade-approximation and an analytic expression
        for the remainder of the continued fraction. See the original
        paper (reference 2 below) a detailed description of the
        numerical approach.

        **References**

        1. Davies, B. (2005). *Integral Transforms and their
           Applications*, Third Edition. Springer.
        2. de Hoog, F., J. Knight, A. Stokes (1982). An improved
           method for numerical inversion of Laplace transforms. *SIAM
           Journal of Scientific and Statistical Computing* 3:357-366,
           http://dx.doi.org/10.1137/0903022

        """

        M = self.degree
        np = self.np
        T = self.T

        self.t = self.ctx.convert(t)

        # would it be useful to try re-using
        # space between e&q and A&B?
        e = self.ctx.zeros(np, M+1)
        q = self.ctx.matrix(2*M, M)
        d = self.ctx.matrix(np, 1)
        A = self.ctx.zeros(np+1, 1)
        B = self.ctx.ones(np+1, 1)

        # initialize Q-D table
        e[:, 0] = 0.0 + 0j
        q[0, 0] = fp[1]/(fp[0]/2)
        for i in range(1, 2*M):
            q[i, 0] = fp[i+1]/fp[i]

        # rhombus rule for filling triangular Q-D table (e & q)
        for r in range(1, M+1):
            # start with e, column 1, 0:2*M-2
            mr = 2*(M-r) + 1
            e[0:mr, r] = q[1:mr+1, r-1] - q[0:mr, r-1] + e[1:mr+1, r-1]
            if not r == M:
                rq = r+1
                mr = 2*(M-rq)+1 + 2
                for i in range(mr):
                    q[i, rq-1] = q[i+1, rq-2]*e[i+1, rq-1]/e[i, rq-1]

        # build up continued fraction coefficients (d)
        d[0] = fp[0]/2
        for r in range(1, M+1):
            d[2*r-1] = -q[0, r-1]  # even terms
            d[2*r]   = -e[0, r]    # odd terms

        # seed A and B for recurrence
        A[0] = 0.0 + 0.0j
        A[1] = d[0]
        B[0:2] = 1.0 + 0.0j

        # base of the power series
        z = self.ctx.expjpi(self.t/T)  # i*pi is already in fcn

        # coefficients of Pade approximation (A & B)
        # using recurrence for all but last term
        for i in range(1, 2*M):
            A[i+1] = A[i] + d[i]*A[i-1]*z
            B[i+1] = B[i] + d[i]*B[i-1]*z

        # "improved remainder" to continued fraction
        brem = (1 + (d[2*M-1] - d[2*M])*z)/2
        # powm1(x,y) computes x^y - 1 more accurately near zero
        rem = brem*self.ctx.powm1(1 + d[2*M]*z/brem,
                                  self.ctx.fraction(1, 2))

        # last term of recurrence using new remainder
        A[np] = A[2*M] + rem*A[2*M-1]
        B[np] = B[2*M] + rem*B[2*M-1]

        # diagonal Pade approximation
        # F=A/B represents accelerated trapezoid rule
        result = self.ctx.exp(self.gamma*self.t)/T*(A[np]/B[np]).real

        # setting dps back to value when calc_laplace_parameter was called
        if not manual_prec:
            self.ctx.dps = self.dps_orig

        return result


# ****************************************

class Cohen(InverseLaplaceTransform):

    def calc_laplace_parameter(self, t, **kwargs):
        r"""The Cohen algorithm accelerates the convergence of the nearly
        alternating series resulting from the application of the trapezoidal
        rule to the Bromwich contour inversion integral.

        .. math ::

            p_k = \frac{\gamma}{2 t} + \frac{\pi i k}{t} \qquad 0 \le k < M

        where

        .. math ::

            \gamma = \frac{2}{3} (d + \log(10) + \log(2 t)),

        `d = \mathrm{dps\_goal}`, which is chosen based on the desired
        accuracy using the method developed in [1] to improve numerical
        stability. The Cohen algorithm shows robustness similar to the de Hoog
        et al. algorithm, but it is faster than the fixed Talbot algorithm.

        **Optional arguments**

        *degree*
            integer order of the approximation (M = number of terms)
        *alpha*
            abscissa for `p_0` (controls the discretization error)

        The working precision will be increased according to a rule of
        thumb. If 'degree' is not specified, the working precision and
        degree are chosen to hopefully achieve the dps of the calling
        context. If 'degree' is specified, the working precision is
        chosen to achieve maximum resulting precision for the
        specified degree.

        **References**

        1. P. Glasserman, J. Ruiz-Mata (2006). Computing the credit loss
        distribution in the Gaussian copula model: a comparison of methods.
        *Journal of Credit Risk* 2(4):33-66, 10.21314/JCR.2006.057

        """
        self.t = self.ctx.convert(t)

        if 'degree' in kwargs:
            self.degree = kwargs['degree']
            self.dps_goal = int(1.5 * self.degree)
        else:
            self.dps_goal = int(self.ctx.dps * 1.74)
            self.degree = max(22, int(1.31 * self.dps_goal))

        M = self.degree + 1

        # this is adjusting the dps of the calling context hopefully
        # the caller doesn't monkey around with it between calling
        # this routine and calc_time_domain_solution()
        self.dps_orig = self.ctx.dps
        self.ctx.dps = self.dps_goal

        ttwo = 2 * self.t
        tmp = self.ctx.dps * self.ctx.log(10) + self.ctx.log(ttwo)
        tmp = self.ctx.fraction(2, 3) * tmp
        self.alpha = self.ctx.convert(kwargs.get('alpha', tmp))

        # all but time-dependent part of p
        a_t = self.alpha / ttwo
        p_t = self.ctx.pi * 1j / self.t

        self.p = self.ctx.matrix(M, 1)
        self.p[0] = a_t

        for i in range(1, M):
            self.p[i] = a_t + i * p_t

    def calc_time_domain_solution(self, fp, t, manual_prec=False):
        r"""Calculate time-domain solution for Cohen algorithm.

        The accelerated nearly alternating series is:

        .. math ::

            f(t, M) = \frac{e^{\gamma / 2}}{t} \left[\frac{1}{2}
            \Re\left(\bar{f}\left(\frac{\gamma}{2t}\right) \right) -
            \sum_{k=0}^{M-1}\frac{c_{M,k}}{d_M}\Re\left(\bar{f}
            \left(\frac{\gamma + 2(k+1) \pi i}{2t}\right)\right)\right],

        where coefficients `\frac{c_{M, k}}{d_M}` are described in [1].

        1. H. Cohen, F. Rodriguez Villegas, D. Zagier (2000). Convergence
        acceleration of alternating series. *Experiment. Math* 9(1):3-12

        """
        self.t = self.ctx.convert(t)

        n = self.degree
        M = n + 1

        A = self.ctx.matrix(M, 1)
        for i in range(M):
            A[i] = fp[i].real

        d = (3 + self.ctx.sqrt(8)) ** n
        d = (d + 1 / d) / 2
        b = -self.ctx.one
        c = -d
        s = 0

        for k in range(n):
            c = b - c
            s = s + c * A[k + 1]
            b = 2 * (k + n) * (k - n) * b / ((2 * k + 1) * (k + self.ctx.one))

        result = self.ctx.exp(self.alpha / 2) / self.t * (A[0] / 2 - s / d)

        # setting dps back to value when calc_laplace_parameter was
        # called, unless flag is set.
        if not manual_prec:
            self.ctx.dps = self.dps_orig

        return result


# ****************************************

class LaplaceTransformInversionMethods(object):
    def __init__(ctx, *args, **kwargs):
        ctx._fixed_talbot = FixedTalbot(ctx)
        ctx._stehfest = Stehfest(ctx)
        ctx._de_hoog = deHoog(ctx)
        ctx._cohen = Cohen(ctx)

    def invertlaplace(ctx, f, t, **kwargs):
        r"""Computes the numerical inverse Laplace transform for a
        Laplace-space function at a given time.  The function being
        evaluated is assumed to be a real-valued function of time.

        The user must supply a Laplace-space function `\bar{f}(p)`,
        and a desired time at which to estimate the time-domain
        solution `f(t)`.

        A few basic examples of Laplace-space functions with known
        inverses (see references [1,2]) :

        .. math ::

            \mathcal{L}\left\lbrace f(t) \right\rbrace=\bar{f}(p)

        .. math ::

            \mathcal{L}^{-1}\left\lbrace \bar{f}(p) \right\rbrace = f(t)

        .. math ::

            \bar{f}(p) = \frac{1}{(p+1)^2}

        .. math ::

            f(t) = t e^{-t}

        >>> from mpmath import *
        >>> mp.dps = 15; mp.pretty = True
        >>> tt = [0.001, 0.01, 0.1, 1, 10]
        >>> fp = lambda p: 1/(p+1)**2
        >>> ft = lambda t: t*exp(-t)
        >>> ft(tt[0]),ft(tt[0])-invertlaplace(fp,tt[0],method='talbot')
        (0.000999000499833375, 8.57923043561212e-20)
        >>> ft(tt[1]),ft(tt[1])-invertlaplace(fp,tt[1],method='talbot')
        (0.00990049833749168, 3.27007646698047e-19)
        >>> ft(tt[2]),ft(tt[2])-invertlaplace(fp,tt[2],method='talbot')
        (0.090483741803596, -1.75215800052168e-18)
        >>> ft(tt[3]),ft(tt[3])-invertlaplace(fp,tt[3],method='talbot')
        (0.367879441171442, 1.2428864009344e-17)
        >>> ft(tt[4]),ft(tt[4])-invertlaplace(fp,tt[4],method='talbot')
        (0.000453999297624849, 4.04513489306658e-20)

        The methods also work for higher precision:

        >>> mp.dps = 100; mp.pretty = True
        >>> nstr(ft(tt[0]),15),nstr(ft(tt[0])-invertlaplace(fp,tt[0],method='talbot'),15)
        ('0.000999000499833375', '-4.96868310693356e-105')
        >>> nstr(ft(tt[1]),15),nstr(ft(tt[1])-invertlaplace(fp,tt[1],method='talbot'),15)
        ('0.00990049833749168', '1.23032291513122e-104')

        .. math ::

            \bar{f}(p) = \frac{1}{p^2+1}

        .. math ::

            f(t) = \mathrm{J}_0(t)

        >>> mp.dps = 15; mp.pretty = True
        >>> fp = lambda p: 1/sqrt(p*p + 1)
        >>> ft = lambda t: besselj(0,t)
        >>> ft(tt[0]),ft(tt[0])-invertlaplace(fp,tt[0],method='dehoog')
        (0.999999750000016, -6.09717765032273e-18)
        >>> ft(tt[1]),ft(tt[1])-invertlaplace(fp,tt[1],method='dehoog')
        (0.99997500015625, -5.61756281076169e-17)

        .. math ::

            \bar{f}(p) = \frac{\log p}{p}

        .. math ::

            f(t) = -\gamma -\log t

        >>> mp.dps = 15; mp.pretty = True
        >>> fp = lambda p: log(p)/p
        >>> ft = lambda t: -euler-log(t)
        >>> ft(tt[0]),ft(tt[0])-invertlaplace(fp,tt[0],method='stehfest')
        (6.3305396140806, -1.92126634837863e-16)
        >>> ft(tt[1]),ft(tt[1])-invertlaplace(fp,tt[1],method='stehfest')
        (4.02795452108656, -4.81486093200704e-16)

        **Options**

        :func:`~mpmath.invertlaplace` recognizes the following optional
        keywords valid for all methods:

        *method*
            Chooses numerical inverse Laplace transform algorithm
            (described below).
        *degree*
            Number of terms used in the approximation

        **Algorithms**

        Mpmath implements four numerical inverse Laplace transform
        algorithms, attributed to: Talbot, Stehfest, and de Hoog,
        Knight and Stokes. These can be selected by using
        *method='talbot'*, *method='stehfest'*, *method='dehoog'* or
        *method='cohen'* or by passing the classes *method=FixedTalbot*,
        *method=Stehfest*, *method=deHoog*, or *method=Cohen*. The functions
        :func:`~mpmath.invlaptalbot`, :func:`~mpmath.invlapstehfest`,
        :func:`~mpmath.invlapdehoog`, and :func:`~mpmath.invlapcohen`
        are also available as shortcuts.

        All four algorithms implement a heuristic balance between the
        requested precision and the precision used internally for the
        calculations. This has been tuned for a typical exponentially
        decaying function and precision up to few hundred decimal
        digits.

        The Laplace transform converts the variable time (i.e., along
        a line) into a parameter given by the right half of the
        complex `p`-plane.  Singularities, poles, and branch cuts in
        the complex `p`-plane contain all the information regarding
        the time behavior of the corresponding function. Any numerical
        method must therefore sample `p`-plane "close enough" to the
        singularities to accurately characterize them, while not
        getting too close to have catastrophic cancellation, overflow,
        or underflow issues. Most significantly, if one or more of the
        singularities in the `p`-plane is not on the left side of the
        Bromwich contour, its effects will be left out of the computed
        solution, and the answer will be completely wrong.

        *Talbot*

        The fixed Talbot method is high accuracy and fast, but the
        method can catastrophically fail for certain classes of time-domain
        behavior, including a Heaviside step function for positive
        time (e.g., `H(t-2)`), or some oscillatory behaviors. The
        Talbot method usually has adjustable parameters, but the
        "fixed" variety implemented here does not. This method
        deforms the Bromwich integral contour in the shape of a
        parabola towards `-\infty`, which leads to problems
        when the solution has a decaying exponential in it (e.g., a
        Heaviside step function is equivalent to multiplying by a
        decaying exponential in Laplace space).

        *Stehfest*

        The Stehfest algorithm only uses abscissa along the real axis
        of the complex `p`-plane to estimate the time-domain
        function. Oscillatory time-domain functions have poles away
        from the real axis, so this method does not work well with
        oscillatory functions, especially high-frequency ones. This
        method also depends on summation of terms in a series that
        grows very large, and will have catastrophic cancellation
        during summation if the working precision is too low.

        *de Hoog et al.*

        The de Hoog, Knight, and Stokes method is essentially a
        Fourier-series quadrature-type approximation to the Bromwich
        contour integral, with non-linear series acceleration and an
        analytical expression for the remainder term. This method is
        typically one of the most robust. This method also involves the
        greatest amount of overhead, so it is typically the slowest of the
        four methods at high precision.

        *Cohen*

        The Cohen method is a trapezoidal rule approximation to the Bromwich
        contour integral, with linear acceleration for alternating
        series. This method is as robust as the de Hoog et al method and the
        fastest of the four methods at high precision, and is therefore the
        default method.

        **Singularities**

        All numerical inverse Laplace transform methods have problems
        at large time when the Laplace-space function has poles,
        singularities, or branch cuts to the right of the origin in
        the complex plane. For simple poles in `\bar{f}(p)` at the
        `p`-plane origin, the time function is constant in time (e.g.,
        `\mathcal{L}\left\lbrace 1 \right\rbrace=1/p` has a pole at
        `p=0`). A pole in `\bar{f}(p)` to the left of the origin is a
        decreasing function of time (e.g., `\mathcal{L}\left\lbrace
        e^{-t/2} \right\rbrace=1/(p+1/2)` has a pole at `p=-1/2`), and
        a pole to the right of the origin leads to an increasing
        function in time (e.g., `\mathcal{L}\left\lbrace t e^{t/4}
        \right\rbrace = 1/(p-1/4)^2` has a pole at `p=1/4`).  When
        singularities occur off the real `p` axis, the time-domain
        function is oscillatory. For example `\mathcal{L}\left\lbrace
        \mathrm{J}_0(t) \right\rbrace=1/\sqrt{p^2+1}` has a branch cut
        starting at `p=j=\sqrt{-1}` and is a decaying oscillatory
        function, This range of behaviors is illustrated in Duffy [3]
        Figure 4.10.4, p. 228.

        In general as `p \rightarrow \infty` `t \rightarrow 0` and
        vice-versa. All numerical inverse Laplace transform methods
        require their abscissa to shift closer to the origin for
        larger times. If the abscissa shift left of the rightmost
        singularity in the Laplace domain, the answer will be
        completely wrong (the effect of singularities to the right of
        the Bromwich contour are not included in the results).

        For example, the following exponentially growing function has
        a pole at `p=3`:

        .. math ::

            \bar{f}(p)=\frac{1}{p^2-9}

        .. math ::

            f(t)=\frac{1}{3}\sinh 3t

        >>> mp.dps = 15; mp.pretty = True
        >>> fp = lambda p: 1/(p*p-9)
        >>> ft = lambda t: sinh(3*t)/3
        >>> tt = [0.01,0.1,1.0,10.0]
        >>> ft(tt[0]),invertlaplace(fp,tt[0],method='talbot')
        (0.0100015000675014, 0.0100015000675014)
        >>> ft(tt[1]),invertlaplace(fp,tt[1],method='talbot')
        (0.101506764482381, 0.101506764482381)
        >>> ft(tt[2]),invertlaplace(fp,tt[2],method='talbot')
        (3.33929164246997, 3.33929164246997)
        >>> ft(tt[3]),invertlaplace(fp,tt[3],method='talbot')
        (1781079096920.74, -1.61331069624091e-14)

        **References**

        1. [DLMF]_ section 1.14 (http://dlmf.nist.gov/1.14T4)
        2. Cohen, A.M. (2007). Numerical Methods for Laplace Transform
           Inversion, Springer.
        3. Duffy, D.G. (1998). Advanced Engineering Mathematics, CRC Press.

        **Numerical Inverse Laplace Transform Reviews**

        1. Bellman, R., R.E. Kalaba, J.A. Lockett (1966). *Numerical
           inversion of the Laplace transform: Applications to Biology,
           Economics, Engineering, and Physics*. Elsevier.
        2. Davies, B., B. Martin (1979). Numerical inversion of the
           Laplace transform: a survey and comparison of methods. *Journal
           of Computational Physics* 33:1-32,
           http://dx.doi.org/10.1016/0021-9991(79)90025-1
        3. Duffy, D.G. (1993). On the numerical inversion of Laplace
           transforms: Comparison of three new methods on characteristic
           problems from applications. *ACM Transactions on Mathematical
           Software* 19(3):333-359, http://dx.doi.org/10.1145/155743.155788
        4. Kuhlman, K.L., (2013). Review of Inverse Laplace Transform
           Algorithms for Laplace-Space Numerical Approaches, *Numerical
           Algorithms*, 63(2):339-355.
           http://dx.doi.org/10.1007/s11075-012-9625-3

        """

        rule = kwargs.get('method', 'cohen')
        if type(rule) is str:
            lrule = rule.lower()
            if lrule == 'talbot':
                rule = ctx._fixed_talbot
            elif lrule == 'stehfest':
                rule = ctx._stehfest
            elif lrule == 'dehoog':
                rule = ctx._de_hoog
            elif rule == 'cohen':
                rule = ctx._cohen
            else:
                raise ValueError("unknown invlap algorithm: %s" % rule)
        else:
            rule = rule(ctx)

        # determine the vector of Laplace-space parameter
        # needed for the requested method and desired time
        rule.calc_laplace_parameter(t, **kwargs)

        # compute the Laplace-space function evalutations
        # at the required abscissa.
        fp = [f(p) for p in rule.p]

        # compute the time-domain solution from the
        # Laplace-space function evaluations
        return rule.calc_time_domain_solution(fp, t)

    # shortcuts for the above function for specific methods
    def invlaptalbot(ctx, *args, **kwargs):
        kwargs['method'] = 'talbot'
        return ctx.invertlaplace(*args, **kwargs)

    def invlapstehfest(ctx, *args, **kwargs):
        kwargs['method'] = 'stehfest'
        return ctx.invertlaplace(*args, **kwargs)

    def invlapdehoog(ctx, *args, **kwargs):
        kwargs['method'] = 'dehoog'
        return ctx.invertlaplace(*args, **kwargs)

    def invlapcohen(ctx, *args, **kwargs):
        kwargs['method'] = 'cohen'
        return ctx.invertlaplace(*args, **kwargs)


# ****************************************

if __name__ == '__main__':
    import doctest
    doctest.testmod()