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

"""Generic interface for least-square minimization.""" 

from __future__ import division, print_function, absolute_import 

 

from warnings import warn 

 

import numpy as np 

from numpy.linalg import norm 

 

from scipy.sparse import issparse, csr_matrix 

from scipy.sparse.linalg import LinearOperator 

from scipy.optimize import _minpack, OptimizeResult 

from scipy.optimize._numdiff import approx_derivative, group_columns 

from scipy._lib.six import string_types 

 

from .trf import trf 

from .dogbox import dogbox 

from .common import EPS, in_bounds, make_strictly_feasible 

 

 

TERMINATION_MESSAGES = { 

-1: "Improper input parameters status returned from `leastsq`", 

0: "The maximum number of function evaluations is exceeded.", 

1: "`gtol` termination condition is satisfied.", 

2: "`ftol` termination condition is satisfied.", 

3: "`xtol` termination condition is satisfied.", 

4: "Both `ftol` and `xtol` termination conditions are satisfied." 

} 

 

 

FROM_MINPACK_TO_COMMON = { 

0: -1, # Improper input parameters from MINPACK. 

1: 2, 

2: 3, 

3: 4, 

4: 1, 

5: 0 

# There are 6, 7, 8 for too small tolerance parameters, 

# but we guard against it by checking ftol, xtol, gtol beforehand. 

} 

 

 

def call_minpack(fun, x0, jac, ftol, xtol, gtol, max_nfev, x_scale, diff_step): 

n = x0.size 

 

if diff_step is None: 

epsfcn = EPS 

else: 

epsfcn = diff_step**2 

 

# Compute MINPACK's `diag`, which is inverse of our `x_scale` and 

# ``x_scale='jac'`` corresponds to ``diag=None``. 

if isinstance(x_scale, string_types) and x_scale == 'jac': 

diag = None 

else: 

diag = 1 / x_scale 

 

full_output = True 

col_deriv = False 

factor = 100.0 

 

if jac is None: 

if max_nfev is None: 

# n squared to account for Jacobian evaluations. 

max_nfev = 100 * n * (n + 1) 

x, info, status = _minpack._lmdif( 

fun, x0, (), full_output, ftol, xtol, gtol, 

max_nfev, epsfcn, factor, diag) 

else: 

if max_nfev is None: 

max_nfev = 100 * n 

x, info, status = _minpack._lmder( 

fun, jac, x0, (), full_output, col_deriv, 

ftol, xtol, gtol, max_nfev, factor, diag) 

 

f = info['fvec'] 

 

if callable(jac): 

J = jac(x) 

else: 

J = np.atleast_2d(approx_derivative(fun, x)) 

 

cost = 0.5 * np.dot(f, f) 

g = J.T.dot(f) 

g_norm = norm(g, ord=np.inf) 

 

nfev = info['nfev'] 

njev = info.get('njev', None) 

 

status = FROM_MINPACK_TO_COMMON[status] 

active_mask = np.zeros_like(x0, dtype=int) 

 

return OptimizeResult( 

x=x, cost=cost, fun=f, jac=J, grad=g, optimality=g_norm, 

active_mask=active_mask, nfev=nfev, njev=njev, status=status) 

 

 

def prepare_bounds(bounds, n): 

lb, ub = [np.asarray(b, dtype=float) for b in bounds] 

if lb.ndim == 0: 

lb = np.resize(lb, n) 

 

if ub.ndim == 0: 

ub = np.resize(ub, n) 

 

return lb, ub 

 

 

def check_tolerance(ftol, xtol, gtol): 

message = "{} is too low, setting to machine epsilon {}." 

if ftol < EPS: 

warn(message.format("`ftol`", EPS)) 

ftol = EPS 

if xtol < EPS: 

warn(message.format("`xtol`", EPS)) 

xtol = EPS 

if gtol < EPS: 

warn(message.format("`gtol`", EPS)) 

gtol = EPS 

 

return ftol, xtol, gtol 

 

 

def check_x_scale(x_scale, x0): 

if isinstance(x_scale, string_types) and x_scale == 'jac': 

return x_scale 

 

try: 

x_scale = np.asarray(x_scale, dtype=float) 

valid = np.all(np.isfinite(x_scale)) and np.all(x_scale > 0) 

except (ValueError, TypeError): 

valid = False 

 

if not valid: 

raise ValueError("`x_scale` must be 'jac' or array_like with " 

"positive numbers.") 

 

if x_scale.ndim == 0: 

x_scale = np.resize(x_scale, x0.shape) 

 

if x_scale.shape != x0.shape: 

raise ValueError("Inconsistent shapes between `x_scale` and `x0`.") 

 

return x_scale 

 

 

def check_jac_sparsity(jac_sparsity, m, n): 

if jac_sparsity is None: 

return None 

 

if not issparse(jac_sparsity): 

jac_sparsity = np.atleast_2d(jac_sparsity) 

 

if jac_sparsity.shape != (m, n): 

raise ValueError("`jac_sparsity` has wrong shape.") 

 

return jac_sparsity, group_columns(jac_sparsity) 

 

 

# Loss functions. 

 

 

def huber(z, rho, cost_only): 

mask = z <= 1 

rho[0, mask] = z[mask] 

rho[0, ~mask] = 2 * z[~mask]**0.5 - 1 

if cost_only: 

return 

rho[1, mask] = 1 

rho[1, ~mask] = z[~mask]**-0.5 

rho[2, mask] = 0 

rho[2, ~mask] = -0.5 * z[~mask]**-1.5 

 

 

def soft_l1(z, rho, cost_only): 

t = 1 + z 

rho[0] = 2 * (t**0.5 - 1) 

if cost_only: 

return 

rho[1] = t**-0.5 

rho[2] = -0.5 * t**-1.5 

 

 

def cauchy(z, rho, cost_only): 

rho[0] = np.log1p(z) 

if cost_only: 

return 

t = 1 + z 

rho[1] = 1 / t 

rho[2] = -1 / t**2 

 

 

def arctan(z, rho, cost_only): 

rho[0] = np.arctan(z) 

if cost_only: 

return 

t = 1 + z**2 

rho[1] = 1 / t 

rho[2] = -2 * z / t**2 

 

 

IMPLEMENTED_LOSSES = dict(linear=None, huber=huber, soft_l1=soft_l1, 

cauchy=cauchy, arctan=arctan) 

 

 

def construct_loss_function(m, loss, f_scale): 

if loss == 'linear': 

return None 

 

if not callable(loss): 

loss = IMPLEMENTED_LOSSES[loss] 

rho = np.empty((3, m)) 

 

def loss_function(f, cost_only=False): 

z = (f / f_scale) ** 2 

loss(z, rho, cost_only=cost_only) 

if cost_only: 

return 0.5 * f_scale ** 2 * np.sum(rho[0]) 

rho[0] *= f_scale ** 2 

rho[2] /= f_scale ** 2 

return rho 

else: 

def loss_function(f, cost_only=False): 

z = (f / f_scale) ** 2 

rho = loss(z) 

if cost_only: 

return 0.5 * f_scale ** 2 * np.sum(rho[0]) 

rho[0] *= f_scale ** 2 

rho[2] /= f_scale ** 2 

return rho 

 

return loss_function 

 

 

def least_squares( 

fun, x0, jac='2-point', bounds=(-np.inf, np.inf), method='trf', 

ftol=1e-8, xtol=1e-8, gtol=1e-8, x_scale=1.0, loss='linear', 

f_scale=1.0, diff_step=None, tr_solver=None, tr_options={}, 

jac_sparsity=None, max_nfev=None, verbose=0, args=(), kwargs={}): 

"""Solve a nonlinear least-squares problem with bounds on the variables. 

 

Given the residuals f(x) (an m-dimensional real function of n real 

variables) and the loss function rho(s) (a scalar function), `least_squares` 

finds a local minimum of the cost function F(x):: 

 

minimize F(x) = 0.5 * sum(rho(f_i(x)**2), i = 0, ..., m - 1) 

subject to lb <= x <= ub 

 

The purpose of the loss function rho(s) is to reduce the influence of 

outliers on the solution. 

 

Parameters 

---------- 

fun : callable 

Function which computes the vector of residuals, with the signature 

``fun(x, *args, **kwargs)``, i.e., the minimization proceeds with 

respect to its first argument. The argument ``x`` passed to this 

function is an ndarray of shape (n,) (never a scalar, even for n=1). 

It must return a 1-d array_like of shape (m,) or a scalar. If the 

argument ``x`` is complex or the function ``fun`` returns complex 

residuals, it must be wrapped in a real function of real arguments, 

as shown at the end of the Examples section. 

x0 : array_like with shape (n,) or float 

Initial guess on independent variables. If float, it will be treated 

as a 1-d array with one element. 

jac : {'2-point', '3-point', 'cs', callable}, optional 

Method of computing the Jacobian matrix (an m-by-n matrix, where 

element (i, j) is the partial derivative of f[i] with respect to 

x[j]). The keywords select a finite difference scheme for numerical 

estimation. The scheme '3-point' is more accurate, but requires 

twice as much operations compared to '2-point' (default). The 

scheme 'cs' uses complex steps, and while potentially the most 

accurate, it is applicable only when `fun` correctly handles 

complex inputs and can be analytically continued to the complex 

plane. Method 'lm' always uses the '2-point' scheme. If callable, 

it is used as ``jac(x, *args, **kwargs)`` and should return a 

good approximation (or the exact value) for the Jacobian as an 

array_like (np.atleast_2d is applied), a sparse matrix or a 

`scipy.sparse.linalg.LinearOperator`. 

bounds : 2-tuple of array_like, optional 

Lower and upper bounds on independent variables. Defaults to no bounds. 

Each array must match the size of `x0` or be a scalar, in the latter 

case a bound will be the same for all variables. Use ``np.inf`` with 

an appropriate sign to disable bounds on all or some variables. 

method : {'trf', 'dogbox', 'lm'}, optional 

Algorithm to perform minimization. 

 

* 'trf' : Trust Region Reflective algorithm, particularly suitable 

for large sparse problems with bounds. Generally robust method. 

* 'dogbox' : dogleg algorithm with rectangular trust regions, 

typical use case is small problems with bounds. Not recommended 

for problems with rank-deficient Jacobian. 

* 'lm' : Levenberg-Marquardt algorithm as implemented in MINPACK. 

Doesn't handle bounds and sparse Jacobians. Usually the most 

efficient method for small unconstrained problems. 

 

Default is 'trf'. See Notes for more information. 

ftol : float, optional 

Tolerance for termination by the change of the cost function. Default 

is 1e-8. The optimization process is stopped when ``dF < ftol * F``, 

and there was an adequate agreement between a local quadratic model and 

the true model in the last step. 

xtol : float, optional 

Tolerance for termination by the change of the independent variables. 

Default is 1e-8. The exact condition depends on the `method` used: 

 

* For 'trf' and 'dogbox' : ``norm(dx) < xtol * (xtol + norm(x))`` 

* For 'lm' : ``Delta < xtol * norm(xs)``, where ``Delta`` is 

a trust-region radius and ``xs`` is the value of ``x`` 

scaled according to `x_scale` parameter (see below). 

 

gtol : float, optional 

Tolerance for termination by the norm of the gradient. Default is 1e-8. 

The exact condition depends on a `method` used: 

 

* For 'trf' : ``norm(g_scaled, ord=np.inf) < gtol``, where 

``g_scaled`` is the value of the gradient scaled to account for 

the presence of the bounds [STIR]_. 

* For 'dogbox' : ``norm(g_free, ord=np.inf) < gtol``, where 

``g_free`` is the gradient with respect to the variables which 

are not in the optimal state on the boundary. 

* For 'lm' : the maximum absolute value of the cosine of angles 

between columns of the Jacobian and the residual vector is less 

than `gtol`, or the residual vector is zero. 

 

x_scale : array_like or 'jac', optional 

Characteristic scale of each variable. Setting `x_scale` is equivalent 

to reformulating the problem in scaled variables ``xs = x / x_scale``. 

An alternative view is that the size of a trust region along j-th 

dimension is proportional to ``x_scale[j]``. Improved convergence may 

be achieved by setting `x_scale` such that a step of a given size 

along any of the scaled variables has a similar effect on the cost 

function. If set to 'jac', the scale is iteratively updated using the 

inverse norms of the columns of the Jacobian matrix (as described in 

[JJMore]_). 

loss : str or callable, optional 

Determines the loss function. The following keyword values are allowed: 

 

* 'linear' (default) : ``rho(z) = z``. Gives a standard 

least-squares problem. 

* 'soft_l1' : ``rho(z) = 2 * ((1 + z)**0.5 - 1)``. The smooth 

approximation of l1 (absolute value) loss. Usually a good 

choice for robust least squares. 

* 'huber' : ``rho(z) = z if z <= 1 else 2*z**0.5 - 1``. Works 

similarly to 'soft_l1'. 

* 'cauchy' : ``rho(z) = ln(1 + z)``. Severely weakens outliers 

influence, but may cause difficulties in optimization process. 

* 'arctan' : ``rho(z) = arctan(z)``. Limits a maximum loss on 

a single residual, has properties similar to 'cauchy'. 

 

If callable, it must take a 1-d ndarray ``z=f**2`` and return an 

array_like with shape (3, m) where row 0 contains function values, 

row 1 contains first derivatives and row 2 contains second 

derivatives. Method 'lm' supports only 'linear' loss. 

f_scale : float, optional 

Value of soft margin between inlier and outlier residuals, default 

is 1.0. The loss function is evaluated as follows 

``rho_(f**2) = C**2 * rho(f**2 / C**2)``, where ``C`` is `f_scale`, 

and ``rho`` is determined by `loss` parameter. This parameter has 

no effect with ``loss='linear'``, but for other `loss` values it is 

of crucial importance. 

max_nfev : None or int, optional 

Maximum number of function evaluations before the termination. 

If None (default), the value is chosen automatically: 

 

* For 'trf' and 'dogbox' : 100 * n. 

* For 'lm' : 100 * n if `jac` is callable and 100 * n * (n + 1) 

otherwise (because 'lm' counts function calls in Jacobian 

estimation). 

 

diff_step : None or array_like, optional 

Determines the relative step size for the finite difference 

approximation of the Jacobian. The actual step is computed as 

``x * diff_step``. If None (default), then `diff_step` is taken to be 

a conventional "optimal" power of machine epsilon for the finite 

difference scheme used [NR]_. 

tr_solver : {None, 'exact', 'lsmr'}, optional 

Method for solving trust-region subproblems, relevant only for 'trf' 

and 'dogbox' methods. 

 

* 'exact' is suitable for not very large problems with dense 

Jacobian matrices. The computational complexity per iteration is 

comparable to a singular value decomposition of the Jacobian 

matrix. 

* 'lsmr' is suitable for problems with sparse and large Jacobian 

matrices. It uses the iterative procedure 

`scipy.sparse.linalg.lsmr` for finding a solution of a linear 

least-squares problem and only requires matrix-vector product 

evaluations. 

 

If None (default) the solver is chosen based on the type of Jacobian 

returned on the first iteration. 

tr_options : dict, optional 

Keyword options passed to trust-region solver. 

 

* ``tr_solver='exact'``: `tr_options` are ignored. 

* ``tr_solver='lsmr'``: options for `scipy.sparse.linalg.lsmr`. 

Additionally ``method='trf'`` supports 'regularize' option 

(bool, default is True) which adds a regularization term to the 

normal equation, which improves convergence if the Jacobian is 

rank-deficient [Byrd]_ (eq. 3.4). 

 

jac_sparsity : {None, array_like, sparse matrix}, optional 

Defines the sparsity structure of the Jacobian matrix for finite 

difference estimation, its shape must be (m, n). If the Jacobian has 

only few non-zero elements in *each* row, providing the sparsity 

structure will greatly speed up the computations [Curtis]_. A zero 

entry means that a corresponding element in the Jacobian is identically 

zero. If provided, forces the use of 'lsmr' trust-region solver. 

If None (default) then dense differencing will be used. Has no effect 

for 'lm' method. 

verbose : {0, 1, 2}, optional 

Level of algorithm's verbosity: 

 

* 0 (default) : work silently. 

* 1 : display a termination report. 

* 2 : display progress during iterations (not supported by 'lm' 

method). 

 

args, kwargs : tuple and dict, optional 

Additional arguments passed to `fun` and `jac`. Both empty by default. 

The calling signature is ``fun(x, *args, **kwargs)`` and the same for 

`jac`. 

 

Returns 

------- 

`OptimizeResult` with the following fields defined: 

x : ndarray, shape (n,) 

Solution found. 

cost : float 

Value of the cost function at the solution. 

fun : ndarray, shape (m,) 

Vector of residuals at the solution. 

jac : ndarray, sparse matrix or LinearOperator, shape (m, n) 

Modified Jacobian matrix at the solution, in the sense that J^T J 

is a Gauss-Newton approximation of the Hessian of the cost function. 

The type is the same as the one used by the algorithm. 

grad : ndarray, shape (m,) 

Gradient of the cost function at the solution. 

optimality : float 

First-order optimality measure. In unconstrained problems, it is always 

the uniform norm of the gradient. In constrained problems, it is the 

quantity which was compared with `gtol` during iterations. 

active_mask : ndarray of int, shape (n,) 

Each component shows whether a corresponding constraint is active 

(that is, whether a variable is at the bound): 

 

* 0 : a constraint is not active. 

* -1 : a lower bound is active. 

* 1 : an upper bound is active. 

 

Might be somewhat arbitrary for 'trf' method as it generates a sequence 

of strictly feasible iterates and `active_mask` is determined within a 

tolerance threshold. 

nfev : int 

Number of function evaluations done. Methods 'trf' and 'dogbox' do not 

count function calls for numerical Jacobian approximation, as opposed 

to 'lm' method. 

njev : int or None 

Number of Jacobian evaluations done. If numerical Jacobian 

approximation is used in 'lm' method, it is set to None. 

status : int 

The reason for algorithm termination: 

 

* -1 : improper input parameters status returned from MINPACK. 

* 0 : the maximum number of function evaluations is exceeded. 

* 1 : `gtol` termination condition is satisfied. 

* 2 : `ftol` termination condition is satisfied. 

* 3 : `xtol` termination condition is satisfied. 

* 4 : Both `ftol` and `xtol` termination conditions are satisfied. 

 

message : str 

Verbal description of the termination reason. 

success : bool 

True if one of the convergence criteria is satisfied (`status` > 0). 

 

See Also 

-------- 

leastsq : A legacy wrapper for the MINPACK implementation of the 

Levenberg-Marquadt algorithm. 

curve_fit : Least-squares minimization applied to a curve fitting problem. 

 

Notes 

----- 

Method 'lm' (Levenberg-Marquardt) calls a wrapper over least-squares 

algorithms implemented in MINPACK (lmder, lmdif). It runs the 

Levenberg-Marquardt algorithm formulated as a trust-region type algorithm. 

The implementation is based on paper [JJMore]_, it is very robust and 

efficient with a lot of smart tricks. It should be your first choice 

for unconstrained problems. Note that it doesn't support bounds. Also 

it doesn't work when m < n. 

 

Method 'trf' (Trust Region Reflective) is motivated by the process of 

solving a system of equations, which constitute the first-order optimality 

condition for a bound-constrained minimization problem as formulated in 

[STIR]_. The algorithm iteratively solves trust-region subproblems 

augmented by a special diagonal quadratic term and with trust-region shape 

determined by the distance from the bounds and the direction of the 

gradient. This enhancements help to avoid making steps directly into bounds 

and efficiently explore the whole space of variables. To further improve 

convergence, the algorithm considers search directions reflected from the 

bounds. To obey theoretical requirements, the algorithm keeps iterates 

strictly feasible. With dense Jacobians trust-region subproblems are 

solved by an exact method very similar to the one described in [JJMore]_ 

(and implemented in MINPACK). The difference from the MINPACK 

implementation is that a singular value decomposition of a Jacobian 

matrix is done once per iteration, instead of a QR decomposition and series 

of Givens rotation eliminations. For large sparse Jacobians a 2-d subspace 

approach of solving trust-region subproblems is used [STIR]_, [Byrd]_. 

The subspace is spanned by a scaled gradient and an approximate 

Gauss-Newton solution delivered by `scipy.sparse.linalg.lsmr`. When no 

constraints are imposed the algorithm is very similar to MINPACK and has 

generally comparable performance. The algorithm works quite robust in 

unbounded and bounded problems, thus it is chosen as a default algorithm. 

 

Method 'dogbox' operates in a trust-region framework, but considers 

rectangular trust regions as opposed to conventional ellipsoids [Voglis]_. 

The intersection of a current trust region and initial bounds is again 

rectangular, so on each iteration a quadratic minimization problem subject 

to bound constraints is solved approximately by Powell's dogleg method 

[NumOpt]_. The required Gauss-Newton step can be computed exactly for 

dense Jacobians or approximately by `scipy.sparse.linalg.lsmr` for large 

sparse Jacobians. The algorithm is likely to exhibit slow convergence when 

the rank of Jacobian is less than the number of variables. The algorithm 

often outperforms 'trf' in bounded problems with a small number of 

variables. 

 

Robust loss functions are implemented as described in [BA]_. The idea 

is to modify a residual vector and a Jacobian matrix on each iteration 

such that computed gradient and Gauss-Newton Hessian approximation match 

the true gradient and Hessian approximation of the cost function. Then 

the algorithm proceeds in a normal way, i.e. robust loss functions are 

implemented as a simple wrapper over standard least-squares algorithms. 

 

.. versionadded:: 0.17.0 

 

References 

---------- 

.. [STIR] M. A. Branch, T. F. Coleman, and Y. Li, "A Subspace, Interior, 

and Conjugate Gradient Method for Large-Scale Bound-Constrained 

Minimization Problems," SIAM Journal on Scientific Computing, 

Vol. 21, Number 1, pp 1-23, 1999. 

.. [NR] William H. Press et. al., "Numerical Recipes. The Art of Scientific 

Computing. 3rd edition", Sec. 5.7. 

.. [Byrd] R. H. Byrd, R. B. Schnabel and G. A. Shultz, "Approximate 

solution of the trust region problem by minimization over 

two-dimensional subspaces", Math. Programming, 40, pp. 247-263, 

1988. 

.. [Curtis] A. Curtis, M. J. D. Powell, and J. Reid, "On the estimation of 

sparse Jacobian matrices", Journal of the Institute of 

Mathematics and its Applications, 13, pp. 117-120, 1974. 

.. [JJMore] J. J. More, "The Levenberg-Marquardt Algorithm: Implementation 

and Theory," Numerical Analysis, ed. G. A. Watson, Lecture 

Notes in Mathematics 630, Springer Verlag, pp. 105-116, 1977. 

.. [Voglis] C. Voglis and I. E. Lagaris, "A Rectangular Trust Region 

Dogleg Approach for Unconstrained and Bound Constrained 

Nonlinear Optimization", WSEAS International Conference on 

Applied Mathematics, Corfu, Greece, 2004. 

.. [NumOpt] J. Nocedal and S. J. Wright, "Numerical optimization, 

2nd edition", Chapter 4. 

.. [BA] B. Triggs et. al., "Bundle Adjustment - A Modern Synthesis", 

Proceedings of the International Workshop on Vision Algorithms: 

Theory and Practice, pp. 298-372, 1999. 

 

Examples 

-------- 

In this example we find a minimum of the Rosenbrock function without bounds 

on independent variables. 

 

>>> def fun_rosenbrock(x): 

... return np.array([10 * (x[1] - x[0]**2), (1 - x[0])]) 

 

Notice that we only provide the vector of the residuals. The algorithm 

constructs the cost function as a sum of squares of the residuals, which 

gives the Rosenbrock function. The exact minimum is at ``x = [1.0, 1.0]``. 

 

>>> from scipy.optimize import least_squares 

>>> x0_rosenbrock = np.array([2, 2]) 

>>> res_1 = least_squares(fun_rosenbrock, x0_rosenbrock) 

>>> res_1.x 

array([ 1., 1.]) 

>>> res_1.cost 

9.8669242910846867e-30 

>>> res_1.optimality 

8.8928864934219529e-14 

 

We now constrain the variables, in such a way that the previous solution 

becomes infeasible. Specifically, we require that ``x[1] >= 1.5``, and 

``x[0]`` left unconstrained. To this end, we specify the `bounds` parameter 

to `least_squares` in the form ``bounds=([-np.inf, 1.5], np.inf)``. 

 

We also provide the analytic Jacobian: 

 

>>> def jac_rosenbrock(x): 

... return np.array([ 

... [-20 * x[0], 10], 

... [-1, 0]]) 

 

Putting this all together, we see that the new solution lies on the bound: 

 

>>> res_2 = least_squares(fun_rosenbrock, x0_rosenbrock, jac_rosenbrock, 

... bounds=([-np.inf, 1.5], np.inf)) 

>>> res_2.x 

array([ 1.22437075, 1.5 ]) 

>>> res_2.cost 

0.025213093946805685 

>>> res_2.optimality 

1.5885401433157753e-07 

 

Now we solve a system of equations (i.e., the cost function should be zero 

at a minimum) for a Broyden tridiagonal vector-valued function of 100000 

variables: 

 

>>> def fun_broyden(x): 

... f = (3 - x) * x + 1 

... f[1:] -= x[:-1] 

... f[:-1] -= 2 * x[1:] 

... return f 

 

The corresponding Jacobian matrix is sparse. We tell the algorithm to 

estimate it by finite differences and provide the sparsity structure of 

Jacobian to significantly speed up this process. 

 

>>> from scipy.sparse import lil_matrix 

>>> def sparsity_broyden(n): 

... sparsity = lil_matrix((n, n), dtype=int) 

... i = np.arange(n) 

... sparsity[i, i] = 1 

... i = np.arange(1, n) 

... sparsity[i, i - 1] = 1 

... i = np.arange(n - 1) 

... sparsity[i, i + 1] = 1 

... return sparsity 

... 

>>> n = 100000 

>>> x0_broyden = -np.ones(n) 

... 

>>> res_3 = least_squares(fun_broyden, x0_broyden, 

... jac_sparsity=sparsity_broyden(n)) 

>>> res_3.cost 

4.5687069299604613e-23 

>>> res_3.optimality 

1.1650454296851518e-11 

 

Let's also solve a curve fitting problem using robust loss function to 

take care of outliers in the data. Define the model function as 

``y = a + b * exp(c * t)``, where t is a predictor variable, y is an 

observation and a, b, c are parameters to estimate. 

 

First, define the function which generates the data with noise and 

outliers, define the model parameters, and generate data: 

 

>>> def gen_data(t, a, b, c, noise=0, n_outliers=0, random_state=0): 

... y = a + b * np.exp(t * c) 

... 

... rnd = np.random.RandomState(random_state) 

... error = noise * rnd.randn(t.size) 

... outliers = rnd.randint(0, t.size, n_outliers) 

... error[outliers] *= 10 

... 

... return y + error 

... 

>>> a = 0.5 

>>> b = 2.0 

>>> c = -1 

>>> t_min = 0 

>>> t_max = 10 

>>> n_points = 15 

... 

>>> t_train = np.linspace(t_min, t_max, n_points) 

>>> y_train = gen_data(t_train, a, b, c, noise=0.1, n_outliers=3) 

 

Define function for computing residuals and initial estimate of 

parameters. 

 

>>> def fun(x, t, y): 

... return x[0] + x[1] * np.exp(x[2] * t) - y 

... 

>>> x0 = np.array([1.0, 1.0, 0.0]) 

 

Compute a standard least-squares solution: 

 

>>> res_lsq = least_squares(fun, x0, args=(t_train, y_train)) 

 

Now compute two solutions with two different robust loss functions. The 

parameter `f_scale` is set to 0.1, meaning that inlier residuals should 

not significantly exceed 0.1 (the noise level used). 

 

>>> res_soft_l1 = least_squares(fun, x0, loss='soft_l1', f_scale=0.1, 

... args=(t_train, y_train)) 

>>> res_log = least_squares(fun, x0, loss='cauchy', f_scale=0.1, 

... args=(t_train, y_train)) 

 

And finally plot all the curves. We see that by selecting an appropriate 

`loss` we can get estimates close to optimal even in the presence of 

strong outliers. But keep in mind that generally it is recommended to try 

'soft_l1' or 'huber' losses first (if at all necessary) as the other two 

options may cause difficulties in optimization process. 

 

>>> t_test = np.linspace(t_min, t_max, n_points * 10) 

>>> y_true = gen_data(t_test, a, b, c) 

>>> y_lsq = gen_data(t_test, *res_lsq.x) 

>>> y_soft_l1 = gen_data(t_test, *res_soft_l1.x) 

>>> y_log = gen_data(t_test, *res_log.x) 

... 

>>> import matplotlib.pyplot as plt 

>>> plt.plot(t_train, y_train, 'o') 

>>> plt.plot(t_test, y_true, 'k', linewidth=2, label='true') 

>>> plt.plot(t_test, y_lsq, label='linear loss') 

>>> plt.plot(t_test, y_soft_l1, label='soft_l1 loss') 

>>> plt.plot(t_test, y_log, label='cauchy loss') 

>>> plt.xlabel("t") 

>>> plt.ylabel("y") 

>>> plt.legend() 

>>> plt.show() 

 

In the next example, we show how complex-valued residual functions of 

complex variables can be optimized with ``least_squares()``. Consider the 

following function: 

 

>>> def f(z): 

... return z - (0.5 + 0.5j) 

 

We wrap it into a function of real variables that returns real residuals 

by simply handling the real and imaginary parts as independent variables: 

 

>>> def f_wrap(x): 

... fx = f(x[0] + 1j*x[1]) 

... return np.array([fx.real, fx.imag]) 

 

Thus, instead of the original m-dimensional complex function of n complex 

variables we optimize a 2m-dimensional real function of 2n real variables: 

 

>>> from scipy.optimize import least_squares 

>>> res_wrapped = least_squares(f_wrap, (0.1, 0.1), bounds=([0, 0], [1, 1])) 

>>> z = res_wrapped.x[0] + res_wrapped.x[1]*1j 

>>> z 

(0.49999999999925893+0.49999999999925893j) 

 

""" 

if method not in ['trf', 'dogbox', 'lm']: 

raise ValueError("`method` must be 'trf', 'dogbox' or 'lm'.") 

 

if jac not in ['2-point', '3-point', 'cs'] and not callable(jac): 

raise ValueError("`jac` must be '2-point', '3-point', 'cs' or " 

"callable.") 

 

if tr_solver not in [None, 'exact', 'lsmr']: 

raise ValueError("`tr_solver` must be None, 'exact' or 'lsmr'.") 

 

if loss not in IMPLEMENTED_LOSSES and not callable(loss): 

raise ValueError("`loss` must be one of {0} or a callable." 

.format(IMPLEMENTED_LOSSES.keys())) 

 

if method == 'lm' and loss != 'linear': 

raise ValueError("method='lm' supports only 'linear' loss function.") 

 

if verbose not in [0, 1, 2]: 

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

 

if len(bounds) != 2: 

raise ValueError("`bounds` must contain 2 elements.") 

 

if max_nfev is not None and max_nfev <= 0: 

raise ValueError("`max_nfev` must be None or positive integer.") 

 

if np.iscomplexobj(x0): 

raise ValueError("`x0` must be real.") 

 

x0 = np.atleast_1d(x0).astype(float) 

 

if x0.ndim > 1: 

raise ValueError("`x0` must have at most 1 dimension.") 

 

lb, ub = prepare_bounds(bounds, x0.shape[0]) 

 

if method == 'lm' and not np.all((lb == -np.inf) & (ub == np.inf)): 

raise ValueError("Method 'lm' doesn't support bounds.") 

 

if lb.shape != x0.shape or ub.shape != x0.shape: 

raise ValueError("Inconsistent shapes between bounds and `x0`.") 

 

if np.any(lb >= ub): 

raise ValueError("Each lower bound must be strictly less than each " 

"upper bound.") 

 

if not in_bounds(x0, lb, ub): 

raise ValueError("`x0` is infeasible.") 

 

x_scale = check_x_scale(x_scale, x0) 

 

ftol, xtol, gtol = check_tolerance(ftol, xtol, gtol) 

 

def fun_wrapped(x): 

return np.atleast_1d(fun(x, *args, **kwargs)) 

 

if method == 'trf': 

x0 = make_strictly_feasible(x0, lb, ub) 

 

f0 = fun_wrapped(x0) 

 

if f0.ndim != 1: 

raise ValueError("`fun` must return at most 1-d array_like.") 

 

if not np.all(np.isfinite(f0)): 

raise ValueError("Residuals are not finite in the initial point.") 

 

n = x0.size 

m = f0.size 

 

if method == 'lm' and m < n: 

raise ValueError("Method 'lm' doesn't work when the number of " 

"residuals is less than the number of variables.") 

 

loss_function = construct_loss_function(m, loss, f_scale) 

if callable(loss): 

rho = loss_function(f0) 

if rho.shape != (3, m): 

raise ValueError("The return value of `loss` callable has wrong " 

"shape.") 

initial_cost = 0.5 * np.sum(rho[0]) 

elif loss_function is not None: 

initial_cost = loss_function(f0, cost_only=True) 

else: 

initial_cost = 0.5 * np.dot(f0, f0) 

 

if callable(jac): 

J0 = jac(x0, *args, **kwargs) 

 

if issparse(J0): 

J0 = csr_matrix(J0) 

 

def jac_wrapped(x, _=None): 

return csr_matrix(jac(x, *args, **kwargs)) 

 

elif isinstance(J0, LinearOperator): 

def jac_wrapped(x, _=None): 

return jac(x, *args, **kwargs) 

 

else: 

J0 = np.atleast_2d(J0) 

 

def jac_wrapped(x, _=None): 

return np.atleast_2d(jac(x, *args, **kwargs)) 

 

else: # Estimate Jacobian by finite differences. 

if method == 'lm': 

if jac_sparsity is not None: 

raise ValueError("method='lm' does not support " 

"`jac_sparsity`.") 

 

if jac != '2-point': 

warn("jac='{0}' works equivalently to '2-point' " 

"for method='lm'.".format(jac)) 

 

J0 = jac_wrapped = None 

else: 

if jac_sparsity is not None and tr_solver == 'exact': 

raise ValueError("tr_solver='exact' is incompatible " 

"with `jac_sparsity`.") 

 

jac_sparsity = check_jac_sparsity(jac_sparsity, m, n) 

 

def jac_wrapped(x, f): 

J = approx_derivative(fun, x, rel_step=diff_step, method=jac, 

f0=f, bounds=bounds, args=args, 

kwargs=kwargs, sparsity=jac_sparsity) 

if J.ndim != 2: # J is guaranteed not sparse. 

J = np.atleast_2d(J) 

 

return J 

 

J0 = jac_wrapped(x0, f0) 

 

if J0 is not None: 

if J0.shape != (m, n): 

raise ValueError( 

"The return value of `jac` has wrong shape: expected {0}, " 

"actual {1}.".format((m, n), J0.shape)) 

 

if not isinstance(J0, np.ndarray): 

if method == 'lm': 

raise ValueError("method='lm' works only with dense " 

"Jacobian matrices.") 

 

if tr_solver == 'exact': 

raise ValueError( 

"tr_solver='exact' works only with dense " 

"Jacobian matrices.") 

 

jac_scale = isinstance(x_scale, string_types) and x_scale == 'jac' 

if isinstance(J0, LinearOperator) and jac_scale: 

raise ValueError("x_scale='jac' can't be used when `jac` " 

"returns LinearOperator.") 

 

if tr_solver is None: 

if isinstance(J0, np.ndarray): 

tr_solver = 'exact' 

else: 

tr_solver = 'lsmr' 

 

if method == 'lm': 

result = call_minpack(fun_wrapped, x0, jac_wrapped, ftol, xtol, gtol, 

max_nfev, x_scale, diff_step) 

 

elif method == 'trf': 

result = trf(fun_wrapped, jac_wrapped, x0, f0, J0, lb, ub, ftol, xtol, 

gtol, max_nfev, x_scale, loss_function, tr_solver, 

tr_options.copy(), verbose) 

 

elif method == 'dogbox': 

if tr_solver == 'lsmr' and 'regularize' in tr_options: 

warn("The keyword 'regularize' in `tr_options` is not relevant " 

"for 'dogbox' method.") 

tr_options = tr_options.copy() 

del tr_options['regularize'] 

 

result = dogbox(fun_wrapped, jac_wrapped, x0, f0, J0, lb, ub, ftol, 

xtol, gtol, max_nfev, x_scale, loss_function, 

tr_solver, tr_options, verbose) 

 

result.message = TERMINATION_MESSAGES[result.status] 

result.success = result.status > 0 

 

if verbose >= 1: 

print(result.message) 

print("Function evaluations {0}, initial cost {1:.4e}, final cost " 

"{2:.4e}, first-order optimality {3:.2e}." 

.format(result.nfev, initial_cost, result.cost, 

result.optimality)) 

 

return result