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

975

976

977

978

979

980

981

982

983

984

985

986

987

988

989

990

991

992

993

994

995

996

997

998

999

1000

1001

1002

1003

1004

1005

1006

1007

1008

1009

1010

1011

1012

1013

1014

1015

1016

1017

1018

1019

1020

1021

1022

1023

1024

1025

1026

1027

1028

1029

1030

1031

1032

1033

1034

1035

1036

1037

1038

1039

1040

1041

1042

1043

1044

1045

1046

1047

1048

1049

1050

1051

1052

1053

1054

1055

1056

1057

1058

1059

1060

1061

1062

1063

1064

1065

1066

1067

1068

1069

1070

1071

1072

1073

1074

1075

1076

1077

1078

1079

1080

1081

1082

1083

1084

1085

1086

1087

1088

1089

1090

1091

1092

1093

1094

1095

1096

1097

1098

1099

1100

1101

1102

1103

1104

1105

1106

1107

1108

1109

1110

1111

1112

1113

1114

1115

1116

1117

1118

1119

1120

1121

1122

1123

1124

1125

1126

1127

1128

1129

1130

1131

1132

1133

1134

"""Boundary value problem solver.""" 

from __future__ import division, print_function, absolute_import 

 

from warnings import warn 

 

import numpy as np 

from numpy.linalg import norm, pinv 

 

from scipy.sparse import coo_matrix, csc_matrix 

from scipy.sparse.linalg import splu 

from scipy.optimize import OptimizeResult 

 

 

EPS = np.finfo(float).eps 

 

 

def estimate_fun_jac(fun, x, y, p, f0=None): 

"""Estimate derivatives of an ODE system rhs with forward differences. 

 

Returns 

------- 

df_dy : ndarray, shape (n, n, m) 

Derivatives with respect to y. An element (i, j, q) corresponds to 

d f_i(x_q, y_q) / d (y_q)_j. 

df_dp : ndarray with shape (n, k, m) or None 

Derivatives with respect to p. An element (i, j, q) corresponds to 

d f_i(x_q, y_q, p) / d p_j. If `p` is empty, None is returned. 

""" 

n, m = y.shape 

if f0 is None: 

f0 = fun(x, y, p) 

 

dtype = y.dtype 

 

df_dy = np.empty((n, n, m), dtype=dtype) 

h = EPS**0.5 * (1 + np.abs(y)) 

for i in range(n): 

y_new = y.copy() 

y_new[i] += h[i] 

hi = y_new[i] - y[i] 

f_new = fun(x, y_new, p) 

df_dy[:, i, :] = (f_new - f0) / hi 

 

k = p.shape[0] 

if k == 0: 

df_dp = None 

else: 

df_dp = np.empty((n, k, m), dtype=dtype) 

h = EPS**0.5 * (1 + np.abs(p)) 

for i in range(k): 

p_new = p.copy() 

p_new[i] += h[i] 

hi = p_new[i] - p[i] 

f_new = fun(x, y, p_new) 

df_dp[:, i, :] = (f_new - f0) / hi 

 

return df_dy, df_dp 

 

 

def estimate_bc_jac(bc, ya, yb, p, bc0=None): 

"""Estimate derivatives of boundary conditions with forward differences. 

 

Returns 

------- 

dbc_dya : ndarray, shape (n + k, n) 

Derivatives with respect to ya. An element (i, j) corresponds to 

d bc_i / d ya_j. 

dbc_dyb : ndarray, shape (n + k, n) 

Derivatives with respect to yb. An element (i, j) corresponds to 

d bc_i / d ya_j. 

dbc_dp : ndarray with shape (n + k, k) or None 

Derivatives with respect to p. An element (i, j) corresponds to 

d bc_i / d p_j. If `p` is empty, None is returned. 

""" 

n = ya.shape[0] 

k = p.shape[0] 

 

if bc0 is None: 

bc0 = bc(ya, yb, p) 

 

dtype = ya.dtype 

 

dbc_dya = np.empty((n, n + k), dtype=dtype) 

h = EPS**0.5 * (1 + np.abs(ya)) 

for i in range(n): 

ya_new = ya.copy() 

ya_new[i] += h[i] 

hi = ya_new[i] - ya[i] 

bc_new = bc(ya_new, yb, p) 

dbc_dya[i] = (bc_new - bc0) / hi 

dbc_dya = dbc_dya.T 

 

h = EPS**0.5 * (1 + np.abs(yb)) 

dbc_dyb = np.empty((n, n + k), dtype=dtype) 

for i in range(n): 

yb_new = yb.copy() 

yb_new[i] += h[i] 

hi = yb_new[i] - yb[i] 

bc_new = bc(ya, yb_new, p) 

dbc_dyb[i] = (bc_new - bc0) / hi 

dbc_dyb = dbc_dyb.T 

 

if k == 0: 

dbc_dp = None 

else: 

h = EPS**0.5 * (1 + np.abs(p)) 

dbc_dp = np.empty((k, n + k), dtype=dtype) 

for i in range(k): 

p_new = p.copy() 

p_new[i] += h[i] 

hi = p_new[i] - p[i] 

bc_new = bc(ya, yb, p_new) 

dbc_dp[i] = (bc_new - bc0) / hi 

dbc_dp = dbc_dp.T 

 

return dbc_dya, dbc_dyb, dbc_dp 

 

 

def compute_jac_indices(n, m, k): 

"""Compute indices for the collocation system Jacobian construction. 

 

See `construct_global_jac` for the explanation. 

""" 

i_col = np.repeat(np.arange((m - 1) * n), n) 

j_col = (np.tile(np.arange(n), n * (m - 1)) + 

np.repeat(np.arange(m - 1) * n, n**2)) 

 

i_bc = np.repeat(np.arange((m - 1) * n, m * n + k), n) 

j_bc = np.tile(np.arange(n), n + k) 

 

i_p_col = np.repeat(np.arange((m - 1) * n), k) 

j_p_col = np.tile(np.arange(m * n, m * n + k), (m - 1) * n) 

 

i_p_bc = np.repeat(np.arange((m - 1) * n, m * n + k), k) 

j_p_bc = np.tile(np.arange(m * n, m * n + k), n + k) 

 

i = np.hstack((i_col, i_col, i_bc, i_bc, i_p_col, i_p_bc)) 

j = np.hstack((j_col, j_col + n, 

j_bc, j_bc + (m - 1) * n, 

j_p_col, j_p_bc)) 

 

return i, j 

 

 

def stacked_matmul(a, b): 

"""Stacked matrix multiply: out[i,:,:] = np.dot(a[i,:,:], b[i,:,:]). 

 

In our case a[i, :, :] and b[i, :, :] are always square. 

""" 

# Empirical optimization. Use outer Python loop and BLAS for large 

# matrices, otherwise use a single einsum call. 

if a.shape[1] > 50: 

out = np.empty_like(a) 

for i in range(a.shape[0]): 

out[i] = np.dot(a[i], b[i]) 

return out 

else: 

return np.einsum('...ij,...jk->...ik', a, b) 

 

 

def construct_global_jac(n, m, k, i_jac, j_jac, h, df_dy, df_dy_middle, df_dp, 

df_dp_middle, dbc_dya, dbc_dyb, dbc_dp): 

"""Construct the Jacobian of the collocation system. 

 

There are n * m + k functions: m - 1 collocations residuals, each 

containing n components, followed by n + k boundary condition residuals. 

 

There are n * m + k variables: m vectors of y, each containing n 

components, followed by k values of vector p. 

 

For example, let m = 4, n = 2 and k = 1, then the Jacobian will have 

the following sparsity structure: 

 

1 1 2 2 0 0 0 0 5 

1 1 2 2 0 0 0 0 5 

0 0 1 1 2 2 0 0 5 

0 0 1 1 2 2 0 0 5 

0 0 0 0 1 1 2 2 5 

0 0 0 0 1 1 2 2 5 

 

3 3 0 0 0 0 4 4 6 

3 3 0 0 0 0 4 4 6 

3 3 0 0 0 0 4 4 6 

 

Zeros denote identically zero values, other values denote different kinds 

of blocks in the matrix (see below). The blank row indicates the separation 

of collocation residuals from boundary conditions. And the blank column 

indicates the separation of y values from p values. 

 

Refer to [1]_ (p. 306) for the formula of n x n blocks for derivatives 

of collocation residuals with respect to y. 

 

Parameters 

---------- 

n : int 

Number of equations in the ODE system. 

m : int 

Number of nodes in the mesh. 

k : int 

Number of the unknown parameters. 

i_jac, j_jac : ndarray 

Row and column indices returned by `compute_jac_indices`. They 

represent different blocks in the Jacobian matrix in the following 

order (see the scheme above): 

 

* 1: m - 1 diagonal n x n blocks for the collocation residuals. 

* 2: m - 1 off-diagonal n x n blocks for the collocation residuals. 

* 3 : (n + k) x n block for the dependency of the boundary 

conditions on ya. 

* 4: (n + k) x n block for the dependency of the boundary 

conditions on yb. 

* 5: (m - 1) * n x k block for the dependency of the collocation 

residuals on p. 

* 6: (n + k) x k block for the dependency of the boundary 

conditions on p. 

 

df_dy : ndarray, shape (n, n, m) 

Jacobian of f with respect to y computed at the mesh nodes. 

df_dy_middle : ndarray, shape (n, n, m - 1) 

Jacobian of f with respect to y computed at the middle between the 

mesh nodes. 

df_dp : ndarray with shape (n, k, m) or None 

Jacobian of f with respect to p computed at the mesh nodes. 

df_dp_middle: ndarray with shape (n, k, m - 1) or None 

Jacobian of f with respect to p computed at the middle between the 

mesh nodes. 

dbc_dya, dbc_dyb : ndarray, shape (n, n) 

Jacobian of bc with respect to ya and yb. 

dbc_dp: ndarray with shape (n, k) or None 

Jacobian of bc with respect to p. 

 

Returns 

------- 

J : csc_matrix, shape (n * m + k, n * m + k) 

Jacobian of the collocation system in a sparse form. 

 

References 

---------- 

.. [1] J. Kierzenka, L. F. Shampine, "A BVP Solver Based on Residual 

Control and the Maltab PSE", ACM Trans. Math. Softw., Vol. 27, 

Number 3, pp. 299-316, 2001. 

""" 

df_dy = np.transpose(df_dy, (2, 0, 1)) 

df_dy_middle = np.transpose(df_dy_middle, (2, 0, 1)) 

 

h = h[:, np.newaxis, np.newaxis] 

 

dtype = df_dy.dtype 

 

# Computing diagonal n x n blocks. 

dPhi_dy_0 = np.empty((m - 1, n, n), dtype=dtype) 

dPhi_dy_0[:] = -np.identity(n) 

dPhi_dy_0 -= h / 6 * (df_dy[:-1] + 2 * df_dy_middle) 

T = stacked_matmul(df_dy_middle, df_dy[:-1]) 

dPhi_dy_0 -= h**2 / 12 * T 

 

# Computing off-diagonal n x n blocks. 

dPhi_dy_1 = np.empty((m - 1, n, n), dtype=dtype) 

dPhi_dy_1[:] = np.identity(n) 

dPhi_dy_1 -= h / 6 * (df_dy[1:] + 2 * df_dy_middle) 

T = stacked_matmul(df_dy_middle, df_dy[1:]) 

dPhi_dy_1 += h**2 / 12 * T 

 

values = np.hstack((dPhi_dy_0.ravel(), dPhi_dy_1.ravel(), dbc_dya.ravel(), 

dbc_dyb.ravel())) 

 

if k > 0: 

df_dp = np.transpose(df_dp, (2, 0, 1)) 

df_dp_middle = np.transpose(df_dp_middle, (2, 0, 1)) 

T = stacked_matmul(df_dy_middle, df_dp[:-1] - df_dp[1:]) 

df_dp_middle += 0.125 * h * T 

dPhi_dp = -h/6 * (df_dp[:-1] + df_dp[1:] + 4 * df_dp_middle) 

values = np.hstack((values, dPhi_dp.ravel(), dbc_dp.ravel())) 

 

J = coo_matrix((values, (i_jac, j_jac))) 

return csc_matrix(J) 

 

 

def collocation_fun(fun, y, p, x, h): 

"""Evaluate collocation residuals. 

 

This function lies in the core of the method. The solution is sought 

as a cubic C1 continuous spline with derivatives matching the ODE rhs 

at given nodes `x`. Collocation conditions are formed from the equality 

of the spline derivatives and rhs of the ODE system in the middle points 

between nodes. 

 

Such method is classified to Lobbato IIIA family in ODE literature. 

Refer to [1]_ for the formula and some discussion. 

 

Returns 

------- 

col_res : ndarray, shape (n, m - 1) 

Collocation residuals at the middle points of the mesh intervals. 

y_middle : ndarray, shape (n, m - 1) 

Values of the cubic spline evaluated at the middle points of the mesh 

intervals. 

f : ndarray, shape (n, m) 

RHS of the ODE system evaluated at the mesh nodes. 

f_middle : ndarray, shape (n, m - 1) 

RHS of the ODE system evaluated at the middle points of the mesh 

intervals (and using `y_middle`). 

 

References 

---------- 

.. [1] J. Kierzenka, L. F. Shampine, "A BVP Solver Based on Residual 

Control and the Maltab PSE", ACM Trans. Math. Softw., Vol. 27, 

Number 3, pp. 299-316, 2001. 

""" 

f = fun(x, y, p) 

y_middle = (0.5 * (y[:, 1:] + y[:, :-1]) - 

0.125 * h * (f[:, 1:] - f[:, :-1])) 

f_middle = fun(x[:-1] + 0.5 * h, y_middle, p) 

col_res = y[:, 1:] - y[:, :-1] - h / 6 * (f[:, :-1] + f[:, 1:] + 

4 * f_middle) 

 

return col_res, y_middle, f, f_middle 

 

 

def prepare_sys(n, m, k, fun, bc, fun_jac, bc_jac, x, h): 

"""Create the function and the Jacobian for the collocation system.""" 

x_middle = x[:-1] + 0.5 * h 

i_jac, j_jac = compute_jac_indices(n, m, k) 

 

def col_fun(y, p): 

return collocation_fun(fun, y, p, x, h) 

 

def sys_jac(y, p, y_middle, f, f_middle, bc0): 

if fun_jac is None: 

df_dy, df_dp = estimate_fun_jac(fun, x, y, p, f) 

df_dy_middle, df_dp_middle = estimate_fun_jac( 

fun, x_middle, y_middle, p, f_middle) 

else: 

df_dy, df_dp = fun_jac(x, y, p) 

df_dy_middle, df_dp_middle = fun_jac(x_middle, y_middle, p) 

 

if bc_jac is None: 

dbc_dya, dbc_dyb, dbc_dp = estimate_bc_jac(bc, y[:, 0], y[:, -1], 

p, bc0) 

else: 

dbc_dya, dbc_dyb, dbc_dp = bc_jac(y[:, 0], y[:, -1], p) 

 

return construct_global_jac(n, m, k, i_jac, j_jac, h, df_dy, 

df_dy_middle, df_dp, df_dp_middle, dbc_dya, 

dbc_dyb, dbc_dp) 

 

return col_fun, sys_jac 

 

 

def solve_newton(n, m, h, col_fun, bc, jac, y, p, B, bvp_tol): 

"""Solve the nonlinear collocation system by a Newton method. 

 

This is a simple Newton method with a backtracking line search. As 

advised in [1]_, an affine-invariant criterion function F = ||J^-1 r||^2 

is used, where J is the Jacobian matrix at the current iteration and r is 

the vector or collocation residuals (values of the system lhs). 

 

The method alters between full Newton iterations and the fixed-Jacobian 

iterations based 

 

There are other tricks proposed in [1]_, but they are not used as they 

don't seem to improve anything significantly, and even break the 

convergence on some test problems I tried. 

 

All important parameters of the algorithm are defined inside the function. 

 

Parameters 

---------- 

n : int 

Number of equations in the ODE system. 

m : int 

Number of nodes in the mesh. 

h : ndarray, shape (m-1,) 

Mesh intervals. 

col_fun : callable 

Function computing collocation residuals. 

bc : callable 

Function computing boundary condition residuals. 

jac : callable 

Function computing the Jacobian of the whole system (including 

collocation and boundary condition residuals). It is supposed to 

return csc_matrix. 

y : ndarray, shape (n, m) 

Initial guess for the function values at the mesh nodes. 

p : ndarray, shape (k,) 

Initial guess for the unknown parameters. 

B : ndarray with shape (n, n) or None 

Matrix to force the S y(a) = 0 condition for a problems with the 

singular term. If None, the singular term is assumed to be absent. 

bvp_tol : float 

Tolerance to which we want to solve a BVP. 

 

Returns 

------- 

y : ndarray, shape (n, m) 

Final iterate for the function values at the mesh nodes. 

p : ndarray, shape (k,) 

Final iterate for the unknown parameters. 

singular : bool 

True, if the LU decomposition failed because Jacobian turned out 

to be singular. 

 

References 

---------- 

.. [1] U. Ascher, R. Mattheij and R. Russell "Numerical Solution of 

Boundary Value Problems for Ordinary Differential Equations" 

""" 

# We know that the solution residuals at the middle points of the mesh 

# are connected with collocation residuals r_middle = 1.5 * col_res / h. 

# As our BVP solver tries to decrease relative residuals below a certain 

# tolerance it seems reasonable to terminated Newton iterations by 

# comparison of r_middle / (1 + np.abs(f_middle)) with a certain threshold, 

# which we choose to be 1.5 orders lower than the BVP tolerance. We rewrite 

# the condition as col_res < tol_r * (1 + np.abs(f_middle)), then tol_r 

# should be computed as follows: 

tol_r = 2/3 * h * 5e-2 * bvp_tol 

 

# We also need to control residuals of the boundary conditions. But it 

# seems that they become very small eventually as the solver progresses, 

# i. e. the tolerance for BC are not very important. We set it 1.5 orders 

# lower than the BVP tolerance as well. 

tol_bc = 5e-2 * bvp_tol 

 

# Maximum allowed number of Jacobian evaluation and factorization, in 

# other words the maximum number of full Newton iterations. A small value 

# is recommended in the literature. 

max_njev = 4 

 

# Maximum number of iterations, considering that some of them can be 

# performed with the fixed Jacobian. In theory such iterations are cheap, 

# but it's not that simple in Python. 

max_iter = 8 

 

# Minimum relative improvement of the criterion function to accept the 

# step (Armijo constant). 

sigma = 0.2 

 

# Step size decrease factor for backtracking. 

tau = 0.5 

 

# Maximum number of backtracking steps, the minimum step is then 

# tau ** n_trial. 

n_trial = 4 

 

col_res, y_middle, f, f_middle = col_fun(y, p) 

bc_res = bc(y[:, 0], y[:, -1], p) 

res = np.hstack((col_res.ravel(order='F'), bc_res)) 

 

njev = 0 

singular = False 

recompute_jac = True 

for iteration in range(max_iter): 

if recompute_jac: 

J = jac(y, p, y_middle, f, f_middle, bc_res) 

njev += 1 

try: 

LU = splu(J) 

except RuntimeError: 

singular = True 

break 

 

step = LU.solve(res) 

cost = np.dot(step, step) 

 

y_step = step[:m * n].reshape((n, m), order='F') 

p_step = step[m * n:] 

 

alpha = 1 

for trial in range(n_trial + 1): 

y_new = y - alpha * y_step 

if B is not None: 

y_new[:, 0] = np.dot(B, y_new[:, 0]) 

p_new = p - alpha * p_step 

 

col_res, y_middle, f, f_middle = col_fun(y_new, p_new) 

bc_res = bc(y_new[:, 0], y_new[:, -1], p_new) 

res = np.hstack((col_res.ravel(order='F'), bc_res)) 

 

step_new = LU.solve(res) 

cost_new = np.dot(step_new, step_new) 

if cost_new < (1 - 2 * alpha * sigma) * cost: 

break 

 

if trial < n_trial: 

alpha *= tau 

 

y = y_new 

p = p_new 

 

if njev == max_njev: 

break 

 

if (np.all(np.abs(col_res) < tol_r * (1 + np.abs(f_middle))) and 

np.all(bc_res < tol_bc)): 

break 

 

# If the full step was taken, then we are going to continue with 

# the same Jacobian. This is the approach of BVP_SOLVER. 

if alpha == 1: 

step = step_new 

cost = cost_new 

recompute_jac = False 

else: 

recompute_jac = True 

 

return y, p, singular 

 

 

def print_iteration_header(): 

print("{:^15}{:^15}{:^15}{:^15}".format( 

"Iteration", "Max residual", "Total nodes", "Nodes added")) 

 

 

def print_iteration_progress(iteration, residual, total_nodes, nodes_added): 

print("{:^15}{:^15.2e}{:^15}{:^15}".format( 

iteration, residual, total_nodes, nodes_added)) 

 

 

class BVPResult(OptimizeResult): 

pass 

 

 

TERMINATION_MESSAGES = { 

0: "The algorithm converged to the desired accuracy.", 

1: "The maximum number of mesh nodes is exceeded.", 

2: "A singular Jacobian encountered when solving the collocation system." 

} 

 

 

def estimate_rms_residuals(fun, sol, x, h, p, r_middle, f_middle): 

"""Estimate rms values of collocation residuals using Lobatto quadrature. 

 

The residuals are defined as the difference between the derivatives of 

our solution and rhs of the ODE system. We use relative residuals, i.e. 

normalized by 1 + np.abs(f). RMS values are computed as sqrt from the 

normalized integrals of the squared relative residuals over each interval. 

Integrals are estimated using 5-point Lobatto quadrature [1]_, we use the 

fact that residuals at the mesh nodes are identically zero. 

 

In [2] they don't normalize integrals by interval lengths, which gives 

a higher rate of convergence of the residuals by the factor of h**0.5. 

I chose to do such normalization for an ease of interpretation of return 

values as RMS estimates. 

 

Returns 

------- 

rms_res : ndarray, shape (m - 1,) 

Estimated rms values of the relative residuals over each interval. 

 

References 

---------- 

.. [1] http://mathworld.wolfram.com/LobattoQuadrature.html 

.. [2] J. Kierzenka, L. F. Shampine, "A BVP Solver Based on Residual 

Control and the Maltab PSE", ACM Trans. Math. Softw., Vol. 27, 

Number 3, pp. 299-316, 2001. 

""" 

x_middle = x[:-1] + 0.5 * h 

s = 0.5 * h * (3/7)**0.5 

x1 = x_middle + s 

x2 = x_middle - s 

y1 = sol(x1) 

y2 = sol(x2) 

y1_prime = sol(x1, 1) 

y2_prime = sol(x2, 1) 

f1 = fun(x1, y1, p) 

f2 = fun(x2, y2, p) 

r1 = y1_prime - f1 

r2 = y2_prime - f2 

 

r_middle /= 1 + np.abs(f_middle) 

r1 /= 1 + np.abs(f1) 

r2 /= 1 + np.abs(f2) 

 

r1 = np.sum(np.real(r1 * np.conj(r1)), axis=0) 

r2 = np.sum(np.real(r2 * np.conj(r2)), axis=0) 

r_middle = np.sum(np.real(r_middle * np.conj(r_middle)), axis=0) 

 

return (0.5 * (32 / 45 * r_middle + 49 / 90 * (r1 + r2))) ** 0.5 

 

 

def create_spline(y, yp, x, h): 

"""Create a cubic spline given values and derivatives. 

 

Formulas for the coefficients are taken from interpolate.CubicSpline. 

 

Returns 

------- 

sol : PPoly 

Constructed spline as a PPoly instance. 

""" 

from scipy.interpolate import PPoly 

 

n, m = y.shape 

c = np.empty((4, n, m - 1), dtype=y.dtype) 

slope = (y[:, 1:] - y[:, :-1]) / h 

t = (yp[:, :-1] + yp[:, 1:] - 2 * slope) / h 

c[0] = t / h 

c[1] = (slope - yp[:, :-1]) / h - t 

c[2] = yp[:, :-1] 

c[3] = y[:, :-1] 

c = np.rollaxis(c, 1) 

 

return PPoly(c, x, extrapolate=True, axis=1) 

 

 

def modify_mesh(x, insert_1, insert_2): 

"""Insert nodes into a mesh. 

 

Nodes removal logic is not established, its impact on the solver is 

presumably negligible. So only insertion is done in this function. 

 

Parameters 

---------- 

x : ndarray, shape (m,) 

Mesh nodes. 

insert_1 : ndarray 

Intervals to each insert 1 new node in the middle. 

insert_2 : ndarray 

Intervals to each insert 2 new nodes, such that divide an interval 

into 3 equal parts. 

 

Returns 

------- 

x_new : ndarray 

New mesh nodes. 

 

Notes 

----- 

`insert_1` and `insert_2` should not have common values. 

""" 

# Because np.insert implementation apparently varies with a version of 

# numpy, we use a simple and reliable approach with sorting. 

return np.sort(np.hstack(( 

x, 

0.5 * (x[insert_1] + x[insert_1 + 1]), 

(2 * x[insert_2] + x[insert_2 + 1]) / 3, 

(x[insert_2] + 2 * x[insert_2 + 1]) / 3 

))) 

 

 

def wrap_functions(fun, bc, fun_jac, bc_jac, k, a, S, D, dtype): 

"""Wrap functions for unified usage in the solver.""" 

if fun_jac is None: 

fun_jac_wrapped = None 

 

if bc_jac is None: 

bc_jac_wrapped = None 

 

if k == 0: 

def fun_p(x, y, _): 

return np.asarray(fun(x, y), dtype) 

 

def bc_wrapped(ya, yb, _): 

return np.asarray(bc(ya, yb), dtype) 

 

if fun_jac is not None: 

def fun_jac_p(x, y, _): 

return np.asarray(fun_jac(x, y), dtype), None 

 

if bc_jac is not None: 

def bc_jac_wrapped(ya, yb, _): 

dbc_dya, dbc_dyb = bc_jac(ya, yb) 

return (np.asarray(dbc_dya, dtype), 

np.asarray(dbc_dyb, dtype), None) 

else: 

def fun_p(x, y, p): 

return np.asarray(fun(x, y, p), dtype) 

 

def bc_wrapped(x, y, p): 

return np.asarray(bc(x, y, p), dtype) 

 

if fun_jac is not None: 

def fun_jac_p(x, y, p): 

df_dy, df_dp = fun_jac(x, y, p) 

return np.asarray(df_dy, dtype), np.asarray(df_dp, dtype) 

 

if bc_jac is not None: 

def bc_jac_wrapped(ya, yb, p): 

dbc_dya, dbc_dyb, dbc_dp = bc_jac(ya, yb, p) 

return (np.asarray(dbc_dya, dtype), np.asarray(dbc_dyb, dtype), 

np.asarray(dbc_dp, dtype)) 

 

if S is None: 

fun_wrapped = fun_p 

else: 

def fun_wrapped(x, y, p): 

f = fun_p(x, y, p) 

if x[0] == a: 

f[:, 0] = np.dot(D, f[:, 0]) 

f[:, 1:] += np.dot(S, y[:, 1:]) / (x[1:] - a) 

else: 

f += np.dot(S, y) / (x - a) 

return f 

 

if fun_jac is not None: 

if S is None: 

fun_jac_wrapped = fun_jac_p 

else: 

Sr = S[:, :, np.newaxis] 

 

def fun_jac_wrapped(x, y, p): 

df_dy, df_dp = fun_jac_p(x, y, p) 

if x[0] == a: 

df_dy[:, :, 0] = np.dot(D, df_dy[:, :, 0]) 

df_dy[:, :, 1:] += Sr / (x[1:] - a) 

else: 

df_dy += Sr / (x - a) 

 

return df_dy, df_dp 

 

return fun_wrapped, bc_wrapped, fun_jac_wrapped, bc_jac_wrapped 

 

 

def solve_bvp(fun, bc, x, y, p=None, S=None, fun_jac=None, bc_jac=None, 

tol=1e-3, max_nodes=1000, verbose=0): 

"""Solve a boundary-value problem for a system of ODEs. 

 

This function numerically solves a first order system of ODEs subject to 

two-point boundary conditions:: 

 

dy / dx = f(x, y, p) + S * y / (x - a), a <= x <= b 

bc(y(a), y(b), p) = 0 

 

Here x is a 1-dimensional independent variable, y(x) is a n-dimensional 

vector-valued function and p is a k-dimensional vector of unknown 

parameters which is to be found along with y(x). For the problem to be 

determined there must be n + k boundary conditions, i.e. bc must be 

(n + k)-dimensional function. 

 

The last singular term in the right-hand side of the system is optional. 

It is defined by an n-by-n matrix S, such that the solution must satisfy 

S y(a) = 0. This condition will be forced during iterations, so it must not 

contradict boundary conditions. See [2]_ for the explanation how this term 

is handled when solving BVPs numerically. 

 

Problems in a complex domain can be solved as well. In this case y and p 

are considered to be complex, and f and bc are assumed to be complex-valued 

functions, but x stays real. Note that f and bc must be complex 

differentiable (satisfy Cauchy-Riemann equations [4]_), otherwise you 

should rewrite your problem for real and imaginary parts separately. To 

solve a problem in a complex domain, pass an initial guess for y with a 

complex data type (see below). 

 

Parameters 

---------- 

fun : callable 

Right-hand side of the system. The calling signature is ``fun(x, y)``, 

or ``fun(x, y, p)`` if parameters are present. All arguments are 

ndarray: ``x`` with shape (m,), ``y`` with shape (n, m), meaning that 

``y[:, i]`` corresponds to ``x[i]``, and ``p`` with shape (k,). The 

return value must be an array with shape (n, m) and with the same 

layout as ``y``. 

bc : callable 

Function evaluating residuals of the boundary conditions. The calling 

signature is ``bc(ya, yb)``, or ``bc(ya, yb, p)`` if parameters are 

present. All arguments are ndarray: ``ya`` and ``yb`` with shape (n,), 

and ``p`` with shape (k,). The return value must be an array with 

shape (n + k,). 

x : array_like, shape (m,) 

Initial mesh. Must be a strictly increasing sequence of real numbers 

with ``x[0]=a`` and ``x[-1]=b``. 

y : array_like, shape (n, m) 

Initial guess for the function values at the mesh nodes, i-th column 

corresponds to ``x[i]``. For problems in a complex domain pass `y` 

with a complex data type (even if the initial guess is purely real). 

p : array_like with shape (k,) or None, optional 

Initial guess for the unknown parameters. If None (default), it is 

assumed that the problem doesn't depend on any parameters. 

S : array_like with shape (n, n) or None 

Matrix defining the singular term. If None (default), the problem is 

solved without the singular term. 

fun_jac : callable or None, optional 

Function computing derivatives of f with respect to y and p. The 

calling signature is ``fun_jac(x, y)``, or ``fun_jac(x, y, p)`` if 

parameters are present. The return must contain 1 or 2 elements in the 

following order: 

 

* df_dy : array_like with shape (n, n, m) where an element 

(i, j, q) equals to d f_i(x_q, y_q, p) / d (y_q)_j. 

* df_dp : array_like with shape (n, k, m) where an element 

(i, j, q) equals to d f_i(x_q, y_q, p) / d p_j. 

 

Here q numbers nodes at which x and y are defined, whereas i and j 

number vector components. If the problem is solved without unknown 

parameters df_dp should not be returned. 

 

If `fun_jac` is None (default), the derivatives will be estimated 

by the forward finite differences. 

bc_jac : callable or None, optional 

Function computing derivatives of bc with respect to ya, yb and p. 

The calling signature is ``bc_jac(ya, yb)``, or ``bc_jac(ya, yb, p)`` 

if parameters are present. The return must contain 2 or 3 elements in 

the following order: 

 

* dbc_dya : array_like with shape (n, n) where an element (i, j) 

equals to d bc_i(ya, yb, p) / d ya_j. 

* dbc_dyb : array_like with shape (n, n) where an element (i, j) 

equals to d bc_i(ya, yb, p) / d yb_j. 

* dbc_dp : array_like with shape (n, k) where an element (i, j) 

equals to d bc_i(ya, yb, p) / d p_j. 

 

If the problem is solved without unknown parameters dbc_dp should not 

be returned. 

 

If `bc_jac` is None (default), the derivatives will be estimated by 

the forward finite differences. 

tol : float, optional 

Desired tolerance of the solution. If we define ``r = y' - f(x, y)`` 

where y is the found solution, then the solver tries to achieve on each 

mesh interval ``norm(r / (1 + abs(f)) < tol``, where ``norm`` is 

estimated in a root mean squared sense (using a numerical quadrature 

formula). Default is 1e-3. 

max_nodes : int, optional 

Maximum allowed number of the mesh nodes. If exceeded, the algorithm 

terminates. Default is 1000. 

verbose : {0, 1, 2}, optional 

Level of algorithm's verbosity: 

 

* 0 (default) : work silently. 

* 1 : display a termination report. 

* 2 : display progress during iterations. 

 

Returns 

------- 

Bunch object with the following fields defined: 

sol : PPoly 

Found solution for y as `scipy.interpolate.PPoly` instance, a C1 

continuous cubic spline. 

p : ndarray or None, shape (k,) 

Found parameters. None, if the parameters were not present in the 

problem. 

x : ndarray, shape (m,) 

Nodes of the final mesh. 

y : ndarray, shape (n, m) 

Solution values at the mesh nodes. 

yp : ndarray, shape (n, m) 

Solution derivatives at the mesh nodes. 

rms_residuals : ndarray, shape (m - 1,) 

RMS values of the relative residuals over each mesh interval (see the 

description of `tol` parameter). 

niter : int 

Number of completed iterations. 

status : int 

Reason for algorithm termination: 

 

* 0: The algorithm converged to the desired accuracy. 

* 1: The maximum number of mesh nodes is exceeded. 

* 2: A singular Jacobian encountered when solving the collocation 

system. 

 

message : string 

Verbal description of the termination reason. 

success : bool 

True if the algorithm converged to the desired accuracy (``status=0``). 

 

Notes 

----- 

This function implements a 4-th order collocation algorithm with the 

control of residuals similar to [1]_. A collocation system is solved 

by a damped Newton method with an affine-invariant criterion function as 

described in [3]_. 

 

Note that in [1]_ integral residuals are defined without normalization 

by interval lengths. So their definition is different by a multiplier of 

h**0.5 (h is an interval length) from the definition used here. 

 

.. versionadded:: 0.18.0 

 

References 

---------- 

.. [1] J. Kierzenka, L. F. Shampine, "A BVP Solver Based on Residual 

Control and the Maltab PSE", ACM Trans. Math. Softw., Vol. 27, 

Number 3, pp. 299-316, 2001. 

.. [2] L.F. Shampine, P. H. Muir and H. Xu, "A User-Friendly Fortran BVP 

Solver". 

.. [3] U. Ascher, R. Mattheij and R. Russell "Numerical Solution of 

Boundary Value Problems for Ordinary Differential Equations". 

.. [4] `Cauchy-Riemann equations 

<https://en.wikipedia.org/wiki/Cauchy-Riemann_equations>`_ on 

Wikipedia. 

 

Examples 

-------- 

In the first example we solve Bratu's problem:: 

 

y'' + k * exp(y) = 0 

y(0) = y(1) = 0 

 

for k = 1. 

 

We rewrite the equation as a first order system and implement its 

right-hand side evaluation:: 

 

y1' = y2 

y2' = -exp(y1) 

 

>>> def fun(x, y): 

... return np.vstack((y[1], -np.exp(y[0]))) 

 

Implement evaluation of the boundary condition residuals: 

 

>>> def bc(ya, yb): 

... return np.array([ya[0], yb[0]]) 

 

Define the initial mesh with 5 nodes: 

 

>>> x = np.linspace(0, 1, 5) 

 

This problem is known to have two solutions. To obtain both of them we 

use two different initial guesses for y. We denote them by subscripts 

a and b. 

 

>>> y_a = np.zeros((2, x.size)) 

>>> y_b = np.zeros((2, x.size)) 

>>> y_b[0] = 3 

 

Now we are ready to run the solver. 

 

>>> from scipy.integrate import solve_bvp 

>>> res_a = solve_bvp(fun, bc, x, y_a) 

>>> res_b = solve_bvp(fun, bc, x, y_b) 

 

Let's plot the two found solutions. We take an advantage of having the 

solution in a spline form to produce a smooth plot. 

 

>>> x_plot = np.linspace(0, 1, 100) 

>>> y_plot_a = res_a.sol(x_plot)[0] 

>>> y_plot_b = res_b.sol(x_plot)[0] 

>>> import matplotlib.pyplot as plt 

>>> plt.plot(x_plot, y_plot_a, label='y_a') 

>>> plt.plot(x_plot, y_plot_b, label='y_b') 

>>> plt.legend() 

>>> plt.xlabel("x") 

>>> plt.ylabel("y") 

>>> plt.show() 

 

We see that the two solutions have similar shape, but differ in scale 

significantly. 

 

In the second example we solve a simple Sturm-Liouville problem:: 

 

y'' + k**2 * y = 0 

y(0) = y(1) = 0 

 

It is known that a non-trivial solution y = A * sin(k * x) is possible for 

k = pi * n, where n is an integer. To establish the normalization constant 

A = 1 we add a boundary condition:: 

 

y'(0) = k 

 

Again we rewrite our equation as a first order system and implement its 

right-hand side evaluation:: 

 

y1' = y2 

y2' = -k**2 * y1 

 

>>> def fun(x, y, p): 

... k = p[0] 

... return np.vstack((y[1], -k**2 * y[0])) 

 

Note that parameters p are passed as a vector (with one element in our 

case). 

 

Implement the boundary conditions: 

 

>>> def bc(ya, yb, p): 

... k = p[0] 

... return np.array([ya[0], yb[0], ya[1] - k]) 

 

Setup the initial mesh and guess for y. We aim to find the solution for 

k = 2 * pi, to achieve that we set values of y to approximately follow 

sin(2 * pi * x): 

 

>>> x = np.linspace(0, 1, 5) 

>>> y = np.zeros((2, x.size)) 

>>> y[0, 1] = 1 

>>> y[0, 3] = -1 

 

Run the solver with 6 as an initial guess for k. 

 

>>> sol = solve_bvp(fun, bc, x, y, p=[6]) 

 

We see that the found k is approximately correct: 

 

>>> sol.p[0] 

6.28329460046 

 

And finally plot the solution to see the anticipated sinusoid: 

 

>>> x_plot = np.linspace(0, 1, 100) 

>>> y_plot = sol.sol(x_plot)[0] 

>>> plt.plot(x_plot, y_plot) 

>>> plt.xlabel("x") 

>>> plt.ylabel("y") 

>>> plt.show() 

""" 

x = np.asarray(x, dtype=float) 

if x.ndim != 1: 

raise ValueError("`x` must be 1 dimensional.") 

h = np.diff(x) 

if np.any(h <= 0): 

raise ValueError("`x` must be strictly increasing.") 

a = x[0] 

 

y = np.asarray(y) 

if np.issubdtype(y.dtype, np.complexfloating): 

dtype = complex 

else: 

dtype = float 

y = y.astype(dtype, copy=False) 

 

if y.ndim != 2: 

raise ValueError("`y` must be 2 dimensional.") 

if y.shape[1] != x.shape[0]: 

raise ValueError("`y` is expected to have {} columns, but actually " 

"has {}.".format(x.shape[0], y.shape[1])) 

 

if p is None: 

p = np.array([]) 

else: 

p = np.asarray(p, dtype=dtype) 

if p.ndim != 1: 

raise ValueError("`p` must be 1 dimensional.") 

 

if tol < 100 * EPS: 

warn("`tol` is too low, setting to {:.2e}".format(100 * EPS)) 

tol = 100 * EPS 

 

if verbose not in [0, 1, 2]: 

raise ValueError("`verbose` must be in [0, 1, 2].") 

 

n = y.shape[0] 

k = p.shape[0] 

 

if S is not None: 

S = np.asarray(S, dtype=dtype) 

if S.shape != (n, n): 

raise ValueError("`S` is expected to have shape {}, " 

"but actually has {}".format((n, n), S.shape)) 

 

# Compute I - S^+ S to impose necessary boundary conditions. 

B = np.identity(n) - np.dot(pinv(S), S) 

 

y[:, 0] = np.dot(B, y[:, 0]) 

 

# Compute (I - S)^+ to correct derivatives at x=a. 

D = pinv(np.identity(n) - S) 

else: 

B = None 

D = None 

 

fun_wrapped, bc_wrapped, fun_jac_wrapped, bc_jac_wrapped = wrap_functions( 

fun, bc, fun_jac, bc_jac, k, a, S, D, dtype) 

 

f = fun_wrapped(x, y, p) 

if f.shape != y.shape: 

raise ValueError("`fun` return is expected to have shape {}, " 

"but actually has {}.".format(y.shape, f.shape)) 

 

bc_res = bc_wrapped(y[:, 0], y[:, -1], p) 

if bc_res.shape != (n + k,): 

raise ValueError("`bc` return is expected to have shape {}, " 

"but actually has {}.".format((n + k,), bc_res.shape)) 

 

status = 0 

iteration = 0 

if verbose == 2: 

print_iteration_header() 

 

while True: 

m = x.shape[0] 

 

col_fun, jac_sys = prepare_sys(n, m, k, fun_wrapped, bc_wrapped, 

fun_jac_wrapped, bc_jac_wrapped, x, h) 

y, p, singular = solve_newton(n, m, h, col_fun, bc_wrapped, jac_sys, 

y, p, B, tol) 

iteration += 1 

 

col_res, y_middle, f, f_middle = collocation_fun(fun_wrapped, y, 

p, x, h) 

# This relation is not trivial, but can be verified. 

r_middle = 1.5 * col_res / h 

sol = create_spline(y, f, x, h) 

rms_res = estimate_rms_residuals(fun_wrapped, sol, x, h, p, 

r_middle, f_middle) 

max_rms_res = np.max(rms_res) 

 

if singular: 

status = 2 

break 

 

insert_1, = np.nonzero((rms_res > tol) & (rms_res < 100 * tol)) 

insert_2, = np.nonzero(rms_res >= 100 * tol) 

nodes_added = insert_1.shape[0] + 2 * insert_2.shape[0] 

 

if m + nodes_added > max_nodes: 

status = 1 

if verbose == 2: 

nodes_added = "({})".format(nodes_added) 

print_iteration_progress(iteration, max_rms_res, m, 

nodes_added) 

break 

 

if verbose == 2: 

print_iteration_progress(iteration, max_rms_res, m, nodes_added) 

 

if nodes_added > 0: 

x = modify_mesh(x, insert_1, insert_2) 

h = np.diff(x) 

y = sol(x) 

else: 

status = 0 

break 

 

if verbose > 0: 

if status == 0: 

print("Solved in {} iterations, number of nodes {}, " 

"maximum relative residual {:.2e}." 

.format(iteration, x.shape[0], max_rms_res)) 

elif status == 1: 

print("Number of nodes is exceeded after iteration {}, " 

"maximum relative residual {:.2e}." 

.format(iteration, max_rms_res)) 

elif status == 2: 

print("Singular Jacobian encountered when solving the collocation " 

"system on iteration {}, maximum relative residual {:.2e}." 

.format(iteration, max_rms_res)) 

 

if p.size == 0: 

p = None 

 

return BVPResult(sol=sol, p=p, x=x, y=y, yp=f, rms_residuals=rms_res, 

niter=iteration, status=status, 

message=TERMINATION_MESSAGES[status], success=status == 0)