File size: 34,727 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
/* Implementation of Gauss's hypergeometric function for complex values.
 *
 * This implementation is based on the Fortran implementation by Shanjie Zhang and
 * Jianming Jin included in specfun.f [1]_.  Computation of Gauss's hypergeometric
 * function involves handling a patchwork of special cases. By default the Zhang and
 * Jin implementation has been followed as closely as possible except for situations where
 * an improvement was obvious. We've attempted to document the reasons behind decisions
 * made by Zhang and Jin and to document the reasons for deviating from their implementation
 * when this has been done. References to the NIST Digital Library of Mathematical
 * Functions [2]_ have been added where they are appropriate. The review paper by
 * Pearson et al [3]_ is an excellent resource for best practices for numerical
 * computation of hypergeometric functions. We have followed this review paper
 * when making improvements to and correcting defects in Zhang and Jin's
 * implementation. When Pearson et al propose several competing alternatives for a
 * given case, we've used our best judgment to decide on the method to use.
 *
 * Author: Albert Steppi
 *
 * Distributed under the same license as Scipy.
 *
 * References
 * ----------
 * .. [1] S. Zhang and J.M. Jin, "Computation of Special Functions", Wiley 1996
 * .. [2] NIST Digital Library of Mathematical Functions. http://dlmf.nist.gov/,
 *        Release 1.1.1 of 2021-03-15. F. W. J. Olver, A. B. Olde Daalhuis,
 *        D. W. Lozier, B. I. Schneider, R. F. Boisvert, C. W. Clark, B. R. Miller,
 *        B. V. Saunders, H. S. Cohl, and M. A. McClain, eds.
 * .. [3] Pearson, J.W., Olver, S. & Porter, M.A.
 *        "Numerical methods for the computation of the confluent and Gauss
 *        hypergeometric functions."
 *        Numer Algor 74, 821-866 (2017). https://doi.org/10.1007/s11075-016-0173-0
 * .. [4] Raimundas Vidunas, "Degenerate Gauss Hypergeometric Functions",
 *        Kyushu Journal of Mathematics, 2007, Volume 61, Issue 1, Pages 109-135,
 * .. [5] López, J.L., Temme, N.M. New series expansions of the Gauss hypergeometric
 *        function. Adv Comput Math 39, 349-365 (2013).
 *        https://doi.org/10.1007/s10444-012-9283-y
 * """
 */

#pragma once

#include "config.h"
#include "error.h"
#include "tools.h"

#include "binom.h"
#include "cephes/gamma.h"
#include "cephes/lanczos.h"
#include "cephes/poch.h"
#include "cephes/hyp2f1.h"
#include "digamma.h"

namespace xsf {
namespace detail {
    constexpr double hyp2f1_EPS = 1e-15;
    /* The original implementation in SciPy from Zhang and Jin used 1500 for the
     * maximum number of series iterations in some cases and 500 in others.
     * Through the empirical results on the test cases in
     * scipy/special/_precompute/hyp2f1_data.py, it was determined that these values
     * can lead to early termination of series which would have eventually converged
     * at a reasonable level of accuracy. We've bumped the iteration limit to 3000,
     * and may adjust it again based on further analysis. */
    constexpr std::uint64_t hyp2f1_MAXITER = 3000;

    XSF_HOST_DEVICE inline double four_gammas_lanczos(double u, double v, double w, double x) {
        /* Compute ratio of gamma functions using lanczos approximation.
         *
         * Computes gamma(u)*gamma(v)/(gamma(w)*gamma(x))
         *
         * It is assumed that x = u + v - w, but it is left to the user to
         * ensure this.
         *
         * The lanczos approximation takes the form
         *
         * gamma(x) = factor(x) * lanczos_sum_expg_scaled(x)
         *
         * where factor(x) = ((x + lanczos_g - 0.5)/e)**(x - 0.5).
         *
         * The formula above is only valid for x >= 0.5, but can be extended to
         * x < 0.5 with the reflection principle.
         *
         * Using the lanczos approximation when computing this ratio of gamma functions
         * allows factors to be combined analytically to avoid underflow and overflow
         * and produce a more accurate result. The condition x = u + v - w makes it
         * possible to cancel the factors in the expression
         *
         * factor(u) * factor(v) / (factor(w) * factor(x))
         *
         * by taking one factor and absorbing it into the others. Currently, this
         * implementation takes the factor corresponding to the argument with largest
         * absolute value and absorbs it into the others.
         *
         * Since this is only called internally by four_gammas. It is assumed that
         * |u| >= |v| and |w| >= |x|.
         */

        /* The below implementation may incorrectly return finite results
         * at poles of the gamma function. Handle these cases explicitly. */
        if ((u == std::trunc(u) && u <= 0) || (v == std::trunc(v) && v <= 0)) {
            /* Return nan if numerator has pole. Diverges to +- infinity
             * depending on direction so value is undefined. */
            return std::numeric_limits<double>::quiet_NaN();
        }
        if ((w == std::trunc(w) && w <= 0) || (x == std::trunc(x) && x <= 0)) {
            // Return 0 if denominator has pole but not numerator.
            return 0.0;
        }

        double result = 1.0;
        double ugh, vgh, wgh, xgh, u_prime, v_prime, w_prime, x_prime;

        if (u >= 0.5) {
            result *= cephes::lanczos_sum_expg_scaled(u);
            ugh = u + cephes::lanczos_g - 0.5;
            u_prime = u;
        } else {
            result /= cephes::lanczos_sum_expg_scaled(1 - u) * std::sin(M_PI * u) * M_1_PI;
            ugh = 0.5 - u + cephes::lanczos_g;
            u_prime = 1 - u;
        }

        if (v >= 0.5) {
            result *= cephes::lanczos_sum_expg_scaled(v);
            vgh = v + cephes::lanczos_g - 0.5;
            v_prime = v;
        } else {
            result /= cephes::lanczos_sum_expg_scaled(1 - v) * std::sin(M_PI * v) * M_1_PI;
            vgh = 0.5 - v + cephes::lanczos_g;
            v_prime = 1 - v;
        }

        if (w >= 0.5) {
            result /= cephes::lanczos_sum_expg_scaled(w);
            wgh = w + cephes::lanczos_g - 0.5;
            w_prime = w;
        } else {
            result *= cephes::lanczos_sum_expg_scaled(1 - w) * std::sin(M_PI * w) * M_1_PI;
            wgh = 0.5 - w + cephes::lanczos_g;
            w_prime = 1 - w;
        }

        if (x >= 0.5) {
            result /= cephes::lanczos_sum_expg_scaled(x);
            xgh = x + cephes::lanczos_g - 0.5;
            x_prime = x;
        } else {
            result *= cephes::lanczos_sum_expg_scaled(1 - x) * std::sin(M_PI * x) * M_1_PI;
            xgh = 0.5 - x + cephes::lanczos_g;
            x_prime = 1 - x;
        }

        if (std::abs(u) >= std::abs(w)) {
            // u has greatest absolute value. Absorb ugh into the others.
            if (std::abs((v_prime - u_prime) * (v - 0.5)) < 100 * ugh and v > 100) {
                /* Special case where base is close to 1. Condition taken from
                 * Boost's beta function implementation. */
                result *= std::exp((v - 0.5) * std::log1p((v_prime - u_prime) / ugh));
            } else {
                result *= std::pow(vgh / ugh, v - 0.5);
            }

            if (std::abs((u_prime - w_prime) * (w - 0.5)) < 100 * wgh and u > 100) {
                result *= std::exp((w - 0.5) * std::log1p((u_prime - w_prime) / wgh));
            } else {
                result *= std::pow(ugh / wgh, w - 0.5);
            }

            if (std::abs((u_prime - x_prime) * (x - 0.5)) < 100 * xgh and u > 100) {
                result *= std::exp((x - 0.5) * std::log1p((u_prime - x_prime) / xgh));
            } else {
                result *= std::pow(ugh / xgh, x - 0.5);
            }
        } else {
            // w has greatest absolute value. Absorb wgh into the others.
            if (std::abs((u_prime - w_prime) * (u - 0.5)) < 100 * wgh and u > 100) {
                result *= std::exp((u - 0.5) * std::log1p((u_prime - w_prime) / wgh));
            } else {
                result *= pow(ugh / wgh, u - 0.5);
            }
            if (std::abs((v_prime - w_prime) * (v - 0.5)) < 100 * wgh and v > 100) {
                result *= std::exp((v - 0.5) * std::log1p((v_prime - w_prime) / wgh));
            } else {
                result *= std::pow(vgh / wgh, v - 0.5);
            }
            if (std::abs((w_prime - x_prime) * (x - 0.5)) < 100 * xgh and x > 100) {
                result *= std::exp((x - 0.5) * std::log1p((w_prime - x_prime) / xgh));
            } else {
                result *= std::pow(wgh / xgh, x - 0.5);
            }
        }
        // This exhausts all cases because we assume |u| >= |v| and |w| >= |x|.

        return result;
    }

    XSF_HOST_DEVICE inline double four_gammas(double u, double v, double w, double x) {
        double result;

        // Without loss of generality, ensure |u| >= |v| and |w| >= |x|.
        if (std::abs(v) > std::abs(u)) {
            std::swap(u, v);
        }
        if (std::abs(x) > std::abs(w)) {
            std::swap(x, w);
        }
        /* Direct ratio tends to be more accurate for arguments in this range. Range
         * chosen empirically based on the relevant benchmarks in
         * scipy/special/_precompute/hyp2f1_data.py */
        if (std::abs(u) <= 100 && std::abs(v) <= 100 && std::abs(w) <= 100 && std::abs(x) <= 100) {
            result = cephes::Gamma(u) * cephes::Gamma(v) * (cephes::rgamma(w) * cephes::rgamma(x));
            if (std::isfinite(result) && result != 0.0) {
                return result;
            }
        }
        result = four_gammas_lanczos(u, v, w, x);
        if (std::isfinite(result) && result != 0.0) {
            return result;
        }
        // If overflow or underflow, try again with logs.
        result = std::exp(cephes::lgam(v) - cephes::lgam(x) + cephes::lgam(u) - cephes::lgam(w));
        result *= cephes::gammasgn(u) * cephes::gammasgn(w) * cephes::gammasgn(v) * cephes::gammasgn(x);
        return result;
    }

    class HypergeometricSeriesGenerator {
        /* Maclaurin series for hyp2f1.
         *
         * Series is convergent for |z| < 1 but is only practical for numerical
         * computation when |z| < 0.9.
         */
      public:
        XSF_HOST_DEVICE HypergeometricSeriesGenerator(double a, double b, double c, std::complex<double> z)
            : a_(a), b_(b), c_(c), z_(z), term_(1.0), k_(0) {}

        XSF_HOST_DEVICE std::complex<double> operator()() {
            std::complex<double> output = term_;
            term_ = term_ * (a_ + k_) * (b_ + k_) / ((k_ + 1) * (c_ + k_)) * z_;
            ++k_;
            return output;
        }

      private:
        double a_, b_, c_;
        std::complex<double> z_, term_;
        std::uint64_t k_;
    };

    class Hyp2f1Transform1Generator {
        /* 1 -z transformation of standard series.*/
      public:
        XSF_HOST_DEVICE Hyp2f1Transform1Generator(double a, double b, double c, std::complex<double> z)
            : factor1_(four_gammas(c, c - a - b, c - a, c - b)),
              factor2_(four_gammas(c, a + b - c, a, b) * std::pow(1.0 - z, c - a - b)),
              generator1_(HypergeometricSeriesGenerator(a, b, a + b - c + 1, 1.0 - z)),
              generator2_(HypergeometricSeriesGenerator(c - a, c - b, c - a - b + 1, 1.0 - z)) {}

        XSF_HOST_DEVICE std::complex<double> operator()() {
            return factor1_ * generator1_() + factor2_ * generator2_();
        }

      private:
        std::complex<double> factor1_, factor2_;
        HypergeometricSeriesGenerator generator1_, generator2_;
    };

    class Hyp2f1Transform1LimitSeriesGenerator {
        /* 1 - z transform in limit as c - a - b approaches an integer m. */
      public:
        XSF_HOST_DEVICE Hyp2f1Transform1LimitSeriesGenerator(double a, double b, double m, std::complex<double> z)
            : d1_(xsf::digamma(a)), d2_(xsf::digamma(b)), d3_(xsf::digamma(1 + m)),
              d4_(xsf::digamma(1.0)), a_(a), b_(b), m_(m), z_(z), log_1_z_(std::log(1.0 - z)),
              factor_(cephes::rgamma(m + 1)), k_(0) {}

        XSF_HOST_DEVICE std::complex<double> operator()() {
            std::complex<double> term_ = (d1_ + d2_ - d3_ - d4_ + log_1_z_) * factor_;
            // Use digamma(x + 1) = digamma(x) + 1/x
            d1_ += 1 / (a_ + k_);       // d1 = digamma(a + k)
            d2_ += 1 / (b_ + k_);       // d2 = digamma(b + k)
            d3_ += 1 / (1.0 + m_ + k_); // d3 = digamma(1 + m + k)
            d4_ += 1 / (1.0 + k_);      // d4 = digamma(1 + k)
            factor_ *= (a_ + k_) * (b_ + k_) / ((k_ + 1.0) * (m_ + k_ + 1)) * (1.0 - z_);
            ++k_;
            return term_;
        }

      private:
        double d1_, d2_, d3_, d4_, a_, b_, m_;
        std::complex<double> z_, log_1_z_, factor_;
        int k_;
    };

    class Hyp2f1Transform2Generator {
        /* 1/z transformation of standard series.*/
      public:
        XSF_HOST_DEVICE Hyp2f1Transform2Generator(double a, double b, double c, std::complex<double> z)
            : factor1_(four_gammas(c, b - a, b, c - a) * std::pow(-z, -a)),
              factor2_(four_gammas(c, a - b, a, c - b) * std::pow(-z, -b)),
              generator1_(HypergeometricSeriesGenerator(a, a - c + 1, a - b + 1, 1.0 / z)),
              generator2_(HypergeometricSeriesGenerator(b, b - c + 1, b - a + 1, 1.0 / z)) {}

        XSF_HOST_DEVICE std::complex<double> operator()() {
            return factor1_ * generator1_() + factor2_ * generator2_();
        }

      private:
        std::complex<double> factor1_, factor2_;
        HypergeometricSeriesGenerator generator1_, generator2_;
    };

    class Hyp2f1Transform2LimitSeriesGenerator {
        /* 1/z transform in limit as a - b approaches a non-negative integer m. (Can swap a and b to
         * handle the m a negative integer case. */
      public:
        XSF_HOST_DEVICE Hyp2f1Transform2LimitSeriesGenerator(double a, double b, double c, double m,
                                                                 std::complex<double> z)
            : d1_(xsf::digamma(1.0)), d2_(xsf::digamma(1 + m)), d3_(xsf::digamma(a)),
              d4_(xsf::digamma(c - a)), a_(a), b_(b), c_(c), m_(m), z_(z), log_neg_z_(std::log(-z)),
              factor_(xsf::cephes::poch(b, m) * xsf::cephes::poch(1 - c + b, m) *
                      xsf::cephes::rgamma(m + 1)),
              k_(0) {}

        XSF_HOST_DEVICE std::complex<double> operator()() {
            std::complex<double> term = (d1_ + d2_ - d3_ - d4_ + log_neg_z_) * factor_;
            // Use digamma(x + 1) = digamma(x) + 1/x
            d1_ += 1 / (1.0 + k_);         // d1 = digamma(1 + k)
            d2_ += 1 / (1.0 + m_ + k_);    // d2 = digamma(1 + m + k)
            d3_ += 1 / (a_ + k_);          // d3 = digamma(a + k)
            d4_ -= 1 / (c_ - a_ - k_ - 1); // d4 = digamma(c - a - k)
            factor_ *= (b_ + m_ + k_) * (1 - c_ + b_ + m_ + k_) / ((k_ + 1) * (m_ + k_ + 1)) / z_;
            ++k_;
            return term;
        }

      private:
        double d1_, d2_, d3_, d4_, a_, b_, c_, m_;
        std::complex<double> z_, log_neg_z_, factor_;
        std::uint64_t k_;
    };

    class Hyp2f1Transform2LimitSeriesCminusAIntGenerator {
        /* 1/z transform in limit as a - b approaches a non-negative integer m, and c - a approaches
         * a positive integer n. */
      public:
        XSF_HOST_DEVICE Hyp2f1Transform2LimitSeriesCminusAIntGenerator(double a, double b, double c, double m,
                                                                           double n, std::complex<double> z)
            : d1_(xsf::digamma(1.0)), d2_(xsf::digamma(1 + m)), d3_(xsf::digamma(a)),
              d4_(xsf::digamma(n)), a_(a), b_(b), c_(c), m_(m), n_(n), z_(z), log_neg_z_(std::log(-z)),
              factor_(xsf::cephes::poch(b, m) * xsf::cephes::poch(1 - c + b, m) *
                      xsf::cephes::rgamma(m + 1)),
              k_(0) {}

        XSF_HOST_DEVICE std::complex<double> operator()() {
            std::complex<double> term;
            if (k_ < n_) {
                term = (d1_ + d2_ - d3_ - d4_ + log_neg_z_) * factor_;
                // Use digamma(x + 1) = digamma(x) + 1/x
                d1_ += 1 / (1.0 + k_);    // d1 = digamma(1 + k)
                d2_ += 1 / (1 + m_ + k_); // d2 = digamma(1 + m + k)
                d3_ += 1 / (a_ + k_);     // d3 = digamma(a + k)
                d4_ -= 1 / (n_ - k_ - 1); // d4 = digamma(c - a - k)
                factor_ *= (b_ + m_ + k_) * (1 - c_ + b_ + m_ + k_) / ((k_ + 1) * (m_ + k_ + 1)) / z_;
                ++k_;
                return term;
            }
            if (k_ == n_) {
                /* When c - a approaches a positive integer and k_ >= c - a = n then
                 * poch(1 - c + b + m + k) = poch(1 - c + a + k) = approaches zero and
                 * digamma(c - a - k) approaches a pole. However we can use the limit
                 * digamma(-n + epsilon) / gamma(-n + epsilon) -> (-1)**(n + 1) * (n+1)! as epsilon -> 0
                 * to continue the series.
                 *
                 * poch(1 - c + b, m + k) = gamma(1 - c + b + m + k)/gamma(1 - c + b)
                 *
                 * If a - b is an integer and c - a is an integer, then a and b must both be integers, so assume
                 * a and b are integers and take the limit as c approaches an integer.
                 *
                 * gamma(1 - c + epsilon + a + k)/gamma(1 - c - epsilon + b) =
                 * (gamma(c + epsilon - b) / gamma(c + epsilon - a - k)) *
                 * (sin(pi * (c + epsilon - b)) / sin(pi * (c + epsilon - a - k))) (reflection principle)
                 *
                 * In the limit as epsilon goes to zero, the ratio of sines will approach
                 * (-1)**(a - b + k) = (-1)**(m + k)
                 *
                 * We may then replace
                 *
                 * poch(1 - c - epsilon + b, m + k)*digamma(c + epsilon - a - k)
                 *
                 * with
                 *
                 * (-1)**(a - b + k)*gamma(c + epsilon - b) * digamma(c + epsilon - a - k) / gamma(c + epsilon - a - k)
                 *
                 * and taking the limit epsilon -> 0 gives
                 *
                 * (-1)**(a - b + k) * gamma(c - b) * (-1)**(k + a - c + 1)(k + a - c)!
                 * = (-1)**(c - b - 1)*Gamma(k + a - c + 1)
                 */
                factor_ = std::pow(-1, m_ + n_) * xsf::binom(c_ - 1, b_ - 1) *
                          xsf::cephes::poch(c_ - a_ + 1, m_ - 1) / std::pow(z_, static_cast<double>(k_));
            }
            term = factor_;
            factor_ *= (b_ + m_ + k_) * (k_ + a_ - c_ + 1) / ((k_ + 1) * (m_ + k_ + 1)) / z_;
            ++k_;
            return term;
        }

      private:
        double d1_, d2_, d3_, d4_, a_, b_, c_, m_, n_;
        std::complex<double> z_, log_neg_z_, factor_;
        std::uint64_t k_;
    };

    class Hyp2f1Transform2LimitFinitePartGenerator {
        /* Initial finite sum in limit as a - b approaches a non-negative integer m. The limiting series
         * for the 1 - z transform also has an initial finite sum, but it is a standard hypergeometric
         * series. */
      public:
        XSF_HOST_DEVICE Hyp2f1Transform2LimitFinitePartGenerator(double b, double c, double m,
                                                                     std::complex<double> z)
            : b_(b), c_(c), m_(m), z_(z), term_(cephes::Gamma(m) * cephes::rgamma(c - b)), k_(0) {}

        XSF_HOST_DEVICE std::complex<double> operator()() {
            std::complex<double> output = term_;
            term_ = term_ * (b_ + k_) * (c_ - b_ - k_ - 1) / ((k_ + 1) * (m_ - k_ - 1)) / z_;
            ++k_;
            return output;
        }

      private:
        double b_, c_, m_;
        std::complex<double> z_, term_;
        std::uint64_t k_;
    };

    class LopezTemmeSeriesGenerator {
        /* Lopez-Temme Series for Gaussian hypergeometric function [4].
         *
         * Converges for all z with real(z) < 1, including in the regions surrounding
         * the points exp(+- i*pi/3) that are not covered by any of the standard
         * transformations.
         */
      public:
        XSF_HOST_DEVICE LopezTemmeSeriesGenerator(double a, double b, double c, std::complex<double> z)
            : n_(0), a_(a), b_(b), c_(c), phi_previous_(1.0), phi_(1 - 2 * b / c), z_(z), Z_(a * z / (z - 2.0)) {}

        XSF_HOST_DEVICE std::complex<double> operator()() {
            if (n_ == 0) {
                ++n_;
                return 1.0;
            }
            if (n_ > 1) { // Update phi and Z for n>=2
                double new_phi = ((n_ - 1) * phi_previous_ - (2.0 * b_ - c_) * phi_) / (c_ + (n_ - 1));
                phi_previous_ = phi_;
                phi_ = new_phi;
                Z_ = Z_ * z_ / (z_ - 2.0) * ((a_ + (n_ - 1)) / n_);
            }
            ++n_;
            return Z_ * phi_;
        }

      private:
        std::uint64_t n_;
        double a_, b_, c_, phi_previous_, phi_;
        std::complex<double> z_, Z_;
    };

    XSF_HOST_DEVICE std::complex<double> hyp2f1_transform1_limiting_case(double a, double b, double c, double m,
                                                                             std::complex<double> z) {
        /* 1 - z transform in limiting case where c - a - b approaches an integer m. */
        std::complex<double> result = 0.0;
        if (m >= 0) {
            if (m != 0) {
                auto series_generator = HypergeometricSeriesGenerator(a, b, 1 - m, 1.0 - z);
                result += four_gammas(m, c, a + m, b + m) * series_eval_fixed_length(series_generator,
                                                                                     std::complex<double>{0.0, 0.0},
                                                                                     static_cast<std::uint64_t>(m));
            }
            std::complex<double> prefactor = std::pow(-1.0, m + 1) * xsf::cephes::Gamma(c) /
                                             (xsf::cephes::Gamma(a) * xsf::cephes::Gamma(b)) *
                                             std::pow(1.0 - z, m);
            auto series_generator = Hyp2f1Transform1LimitSeriesGenerator(a + m, b + m, m, z);
            result += prefactor * series_eval(series_generator, std::complex<double>{0.0, 0.0}, hyp2f1_EPS,
                                              hyp2f1_MAXITER, "hyp2f1");
            return result;
        } else {
            result = four_gammas(-m, c, a, b) * std::pow(1.0 - z, m);
            auto series_generator1 = HypergeometricSeriesGenerator(a + m, b + m, 1 + m, 1.0 - z);
            result *= series_eval_fixed_length(series_generator1, std::complex<double>{0.0, 0.0},
                                               static_cast<std::uint64_t>(-m));
            double prefactor = std::pow(-1.0, m + 1) * xsf::cephes::Gamma(c) *
                               (xsf::cephes::rgamma(a + m) * xsf::cephes::rgamma(b + m));
            auto series_generator2 = Hyp2f1Transform1LimitSeriesGenerator(a, b, -m, z);
            result += prefactor * series_eval(series_generator2, std::complex<double>{0.0, 0.0}, hyp2f1_EPS,
                                              hyp2f1_MAXITER, "hyp2f1");
            return result;
        }
    }

    XSF_HOST_DEVICE std::complex<double> hyp2f1_transform2_limiting_case(double a, double b, double c, double m,
                                                                             std::complex<double> z) {
        /* 1 / z transform in limiting case where a - b approaches a non-negative integer m. Negative integer case
         * can be handled by swapping a and b. */
        auto series_generator1 = Hyp2f1Transform2LimitFinitePartGenerator(b, c, m, z);
        std::complex<double> result = cephes::Gamma(c) * cephes::rgamma(a) * std::pow(-z, -b);
        result *=
            series_eval_fixed_length(series_generator1, std::complex<double>{0.0, 0.0}, static_cast<std::uint64_t>(m));
        std::complex<double> prefactor = cephes::Gamma(c) * (cephes::rgamma(a) * cephes::rgamma(c - b) * std::pow(-z, -a));
        double n = c - a;
        if (abs(n - std::round(n)) < hyp2f1_EPS) {
            auto series_generator2 = Hyp2f1Transform2LimitSeriesCminusAIntGenerator(a, b, c, m, n, z);
            result += prefactor * series_eval(series_generator2, std::complex<double>{0.0, 0.0}, hyp2f1_EPS,
                                              hyp2f1_MAXITER, "hyp2f1");
            return result;
        }
        auto series_generator2 = Hyp2f1Transform2LimitSeriesGenerator(a, b, c, m, z);
        result += prefactor *
                  series_eval(series_generator2, std::complex<double>{0.0, 0.0}, hyp2f1_EPS, hyp2f1_MAXITER, "hyp2f1");
        return result;
    }

} // namespace detail

XSF_HOST_DEVICE inline std::complex<double> hyp2f1(double a, double b, double c, std::complex<double> z) {
    /* Special Cases
     * -----------------------------------------------------------------------
     * Takes constant value 1 when a = 0 or b = 0, even if c is a non-positive
     * integer. This follows mpmath. */
    if (a == 0 || b == 0) {
        return 1.0;
    }
    double z_abs = std::abs(z);
    // Equals 1 when z i 0, unless c is 0.
    if (z_abs == 0) {
        if (c != 0) {
            return 1.0;
        } else {
            // Returning real part NAN and imaginary part 0 follows mpmath.
            return std::complex<double>{std::numeric_limits<double>::quiet_NaN(), 0};
        }
    }
    bool a_neg_int = a == std::trunc(a) && a < 0;
    bool b_neg_int = b == std::trunc(b) && b < 0;
    bool c_non_pos_int = c == std::trunc(c) and c <= 0;
    /* Diverges when c is a non-positive integer unless a is an integer with
     * c <= a <= 0 or b is an integer with c <= b <= 0, (or z equals 0 with
     * c != 0) Cases z = 0, a = 0, or b = 0 have already been handled. We follow
     * mpmath in handling the degenerate cases where any of a, b, c are
     * non-positive integers. See [3] for a treatment of degenerate cases. */
    if (c_non_pos_int && !((a_neg_int && c <= a && a < 0) || (b_neg_int && c <= b && b < 0))) {
        return std::complex<double>{std::numeric_limits<double>::infinity(), 0};
    }
    /* Reduces to a polynomial when a or b is a negative integer.
     * If a and b are both negative integers, we take care to terminate
     * the series at a or b of smaller magnitude. This is to ensure proper
     * handling of situations like a < c < b <= 0, a, b, c all non-positive
     * integers, where terminating at a would lead to a term of the form 0 / 0. */
    std::uint64_t max_degree;
    if (a_neg_int || b_neg_int) {
        if (a_neg_int && b_neg_int) {
            max_degree = a > b ? std::abs(a) : std::abs(b);
        } else if (a_neg_int) {
            max_degree = std::abs(a);
        } else {
            max_degree = std::abs(b);
        }
        if (max_degree <= UINT64_MAX) {
            auto series_generator = detail::HypergeometricSeriesGenerator(a, b, c, z);
            return detail::series_eval_fixed_length(series_generator, std::complex<double>{0.0, 0.0}, max_degree + 1);
        } else {
            set_error("hyp2f1", SF_ERROR_NO_RESULT, NULL);
            return std::complex<double>{std::numeric_limits<double>::quiet_NaN(),
                                        std::numeric_limits<double>::quiet_NaN()};
        }
    }
    // Kummer's Theorem for z = -1; c = 1 + a - b (DLMF 15.4.26)
    if (std::abs(z + 1.0) < detail::hyp2f1_EPS && std::abs(1 + a - b - c) < detail::hyp2f1_EPS && !c_non_pos_int) {
        return detail::four_gammas(a - b + 1, 0.5 * a + 1, a + 1, 0.5 * a - b + 1);
    }
    std::complex<double> result;
    bool c_minus_a_neg_int = c - a == std::trunc(c - a) && c - a < 0;
    bool c_minus_b_neg_int = c - b == std::trunc(c - b) && c - b < 0;
    /* If one of c - a or c - b is a negative integer, reduces to evaluating
     * a polynomial through an Euler hypergeometric transformation.
     * (DLMF 15.8.1) */
    if (c_minus_a_neg_int || c_minus_b_neg_int) {
        max_degree = c_minus_b_neg_int ? std::abs(c - b) : std::abs(c - a);
        if (max_degree <= UINT64_MAX) {
            result = std::pow(1.0 - z, c - a - b);
            auto series_generator = detail::HypergeometricSeriesGenerator(c - a, c - b, c, z);
            result *=
                detail::series_eval_fixed_length(series_generator, std::complex<double>{0.0, 0.0}, max_degree + 2);
            return result;
        } else {
            set_error("hyp2f1", SF_ERROR_NO_RESULT, NULL);
            return std::complex<double>{std::numeric_limits<double>::quiet_NaN(),
                                        std::numeric_limits<double>::quiet_NaN()};
        }
    }
    /* Diverges as real(z) -> 1 when c <= a + b.
     * Todo: Actually check for overflow instead of using a fixed tolerance for
     * all parameter combinations like in the Fortran original. */
    if (std::abs(1 - z.real()) < detail::hyp2f1_EPS && z.imag() == 0 && c - a - b <= 0 && !c_non_pos_int) {
        return std::complex<double>{std::numeric_limits<double>::infinity(), 0};
    }
    // Gauss's Summation Theorem for z = 1; c - a - b > 0 (DLMF 15.4.20).
    if (z == 1.0 && c - a - b > 0 && !c_non_pos_int) {
        return detail::four_gammas(c, c - a - b, c - a, c - b);
    }
    /* |z| < 0, z.real() >= 0. Use the Maclaurin Series.
     * -----------------------------------------------------------------------
     * Apply Euler Hypergeometric Transformation (DLMF 15.8.1) to reduce
     * size of a and b if possible. We follow Zhang and Jin's
     * implementation [1] although there is very likely a better heuristic
     * to determine when this transformation should be applied. As it
     * stands, this hurts precision in some cases. */
    if (z_abs < 0.9 && z.real() >= 0) {
        if (c - a < a && c - b < b) {
            result = std::pow(1.0 - z, c - a - b);
            auto series_generator = detail::HypergeometricSeriesGenerator(c - a, c - b, c, z);
            result *= detail::series_eval(series_generator, std::complex<double>{0.0, 0.0}, detail::hyp2f1_EPS,
                                          detail::hyp2f1_MAXITER, "hyp2f1");
            return result;
        }
        auto series_generator = detail::HypergeometricSeriesGenerator(a, b, c, z);
        return detail::series_eval(series_generator, std::complex<double>{0.0, 0.0}, detail::hyp2f1_EPS,
                                   detail::hyp2f1_MAXITER, "hyp2f1");
    }
    /* Points near exp(iπ/3), exp(-iπ/3) not handled by any of the standard
     * transformations. Use series of López and Temme [5]. These regions
     * were not correctly handled by Zhang and Jin's implementation.
     * -------------------------------------------------------------------------*/
    if (0.9 <= z_abs && z_abs < 1.1 && std::abs(1.0 - z) >= 0.9 && z.real() >= 0) {
        /* This condition for applying Euler Transformation (DLMF 15.8.1)
         * was determined empirically to work better for this case than that
         * used in Zhang and Jin's implementation for |z| < 0.9,
         *  real(z) >= 0. */
        if ((c - a <= a && c - b < b) || (c - a < a && c - b <= b)) {
            auto series_generator = detail::LopezTemmeSeriesGenerator(c - a, c - b, c, z);
            result = std::pow(1.0 - 0.5 * z, a - c); // Lopez-Temme prefactor
            result *= detail::series_eval(series_generator, std::complex<double>{0.0, 0.0}, detail::hyp2f1_EPS,
                                          detail::hyp2f1_MAXITER, "hyp2f1");
            return std::pow(1.0 - z, c - a - b) * result; // Euler transform prefactor.
        }
        auto series_generator = detail::LopezTemmeSeriesGenerator(a, b, c, z);
        result = detail::series_eval(series_generator, std::complex<double>{0.0, 0.0}, detail::hyp2f1_EPS,
                                     detail::hyp2f1_MAXITER, "hyp2f1");
        return std::pow(1.0 - 0.5 * z, -a) * result; // Lopez-Temme prefactor.
    }
    /* z/(z - 1) transformation (DLMF 15.8.1). Avoids cancellation issues that
     * occur with Maclaurin series for real(z) < 0.
     * -------------------------------------------------------------------------*/
    if (z_abs < 1.1 && z.real() < 0) {
        if (0 < b && b < a && a < c) {
            std::swap(a, b);
        }
        auto series_generator = detail::HypergeometricSeriesGenerator(a, c - b, c, z / (z - 1.0));
        return std::pow(1.0 - z, -a) * detail::series_eval(series_generator, std::complex<double>{0.0, 0.0},
                                                           detail::hyp2f1_EPS, detail::hyp2f1_MAXITER, "hyp2f1");
    }
    /* 1 - z transformation (DLMF 15.8.4). */
    if (0.9 <= z_abs && z_abs < 1.1) {
        if (std::abs(c - a - b - std::round(c - a - b)) < detail::hyp2f1_EPS) {
            // Removable singularity when c - a - b is an integer. Need to use limiting formula.
            double m = std::round(c - a - b);
            return detail::hyp2f1_transform1_limiting_case(a, b, c, m, z);
        }
        auto series_generator = detail::Hyp2f1Transform1Generator(a, b, c, z);
        return detail::series_eval(series_generator, std::complex<double>{0.0, 0.0}, detail::hyp2f1_EPS,
                                   detail::hyp2f1_MAXITER, "hyp2f1");
    }
    /* 1/z transformation (DLMF 15.8.2). */
    if (std::abs(a - b - std::round(a - b)) < detail::hyp2f1_EPS) {
        if (b > a) {
            std::swap(a, b);
        }
        double m = std::round(a - b);
        return detail::hyp2f1_transform2_limiting_case(a, b, c, m, z);
    }
    auto series_generator = detail::Hyp2f1Transform2Generator(a, b, c, z);
    return detail::series_eval(series_generator, std::complex<double>{0.0, 0.0}, detail::hyp2f1_EPS,
                               detail::hyp2f1_MAXITER, "hyp2f1");
}

XSF_HOST_DEVICE inline std::complex<float> hyp2f1(float a, float b, float c, std::complex<float> x) {
    return static_cast<std::complex<float>>(hyp2f1(static_cast<double>(a), static_cast<double>(b),
                                                   static_cast<double>(c), static_cast<std::complex<double>>(x)));
}

XSF_HOST_DEVICE inline double hyp2f1(double a, double b, double c, double x) { return cephes::hyp2f1(a, b, c, x); }

XSF_HOST_DEVICE inline float hyp2f1(float a, float b, float c, float x) {
    return hyp2f1(static_cast<double>(a), static_cast<double>(b), static_cast<double>(c), static_cast<double>(x));
}

} // namespace xsf