roots-multi.lisp 18.4 KB
Newer Older
1 2
;;; Multivariate roots.                
;;; Liam Healy 2008-01-12 12:49:08
3
;;; Time-stamp: <2009-01-21 22:38:26EST roots-multi.lisp>
lhealy's avatar
lhealy committed
4
;;; $Id$
5 6 7

(in-package :gsl)

8 9
;;; /usr/include/gsl/gsl_multiroots.h

10 11 12 13 14 15
;;; Currently, functions defined for root solving will be passed
;;; scalars and should return scalars as multiple values.  A possible
;;; future enhancement is to optionally pass marrays and return
;;; marrays instead.  This would allow directly manipulation of
;;; marrays by the user function.  Notes Mon Jan 19 2009.

16 17 18 19 20 21 22 23
;;;;****************************************************************************
;;;; Function definition
;;;;****************************************************************************

(cffi:defcstruct gsl-mfunction
  ;; See /usr/include/gsl/gsl_multiroots.h
  "The definition of a function for multiroot finding in GSL."
  (function :pointer)
24
  (dimensions sizet)
25 26 27 28 29
  (parameters :pointer))

(export 'def-mfunction)
(defmacro def-mfunction (name dimensions)
  "Define a function for multivariate root solving."
30 31
  `(def-single-function ,name :success-failure :double gsl-mfunction
    ,dimensions))
32

33 34 35 36 37 38 39
(cffi:defcstruct gsl-mfunction-fdf
  ;; See /usr/include/gsl/gsl_multiroots.h
  "The definition of a function and its derivatives for multiroot
   finding in GSL."
  (function :pointer)
  (df :pointer)
  (fdf :pointer)
40
  (dimensions sizet)
41 42
  (parameters :pointer))

43 44 45 46
;;;;****************************************************************************
;;;; Initialization
;;;;****************************************************************************

47 48 49 50 51 52 53 54 55 56
(defmobject multi-dimensional-root-solver-f "gsl_multiroot_fsolver"
  ((type :pointer) (dimension sizet))
  "multi-dimensional root solver with function only"			; FDL
  "Make an instance of a solver of the type specified for a system of
   the specified number of dimensions.  Optionally
   set or reset an existing solver to use the function and the
   initial guess gsl-vector."
  "set"
  ((function :pointer) ((mpointer initial) :pointer))
  (lambda (set)
57
  `((type &optional function-or-dimension (initial nil ,set))
58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73
    (:type type
     :dimension
     (if ,set (dim0 initial) function-or-dimension))
    (:function function-or-dimension :initial initial)))
  (initial))

(defmobject multi-dimensional-root-solver-fdf "gsl_multiroot_fdfsolver"
  ((type :pointer) (dimension sizet))
  "multi-dimensional root solver with function and derivative"			; FDL
  "Make an instance of a derivative solver of the type specified for
   a system of the specified number of dimensions.  Optionally
   set or reset an existing solver to use the function and derivative
   (fdf) and the initial guess."
  "set"
  ((function-derivative :pointer) ((mpointer initial) :pointer))
  (lambda (set)
74
  `((type &optional function-or-dimension (initial nil ,set))
75 76 77
    (:type type
     :dimension
     (if ,set (dim0 initial) function-or-dimension))
78
    (:function-derivative function-or-dimension :initial initial)))
79
  (initial))
80

81
(defmfun name ((solver multi-dimensional-root-solver-f))
82
  "gsl_multiroot_fsolver_name"
83 84
  (((mpointer solver) :pointer))
  :definition :method
85
  :c-return :string
86
  :documentation			; FDL
87 88
  "The name of the solver.")

89
(defmfun name ((solver multi-dimensional-root-solver-fdf))
90
  "gsl_multiroot_fdfsolver_name"
91 92
  (((mpointer solver) :pointer))
  :definition :method
93
  :c-return :string
94
  :documentation			; FDL
95 96 97 98 99 100
  "The name of the solver.")

;;;;****************************************************************************
;;;; Iteration
;;;;****************************************************************************

101
(defmfun iterate ((solver multi-dimensional-root-solver-f))
102
  "gsl_multiroot_fsolver_iterate"
103
  (((mpointer solver) :pointer))
104
  :definition :method
105
  :documentation			; FDL
106 107 108 109 110 111
  "Perform a single iteration of the solver.  The following errors may
   be signalled: 'bad-function-supplied, the iteration encountered a
   singular point where the function or its derivative evaluated to
   infinity or NaN, or 'gsl-division-by-zero, the derivative of the
   function vanished at the iteration point, preventing the algorithm
   from continuing without a division by zero.")
112

113
(defmfun iterate ((solver multi-dimensional-root-solver-fdf))
114
  "gsl_multiroot_fdfsolver_iterate"
115
  (((mpointer solver) :pointer))
116
  :definition :method
117
  :documentation			; FDL
118 119 120 121 122 123
  "Perform a single iteration of the solver.  The following errors may
   be signalled: 'bad-function-supplied, the iteration encountered a
   singular point where the function or its derivative evaluated to
   infinity or NaN, or 'gsl-division-by-zero, the derivative of the
   function vanished at the iteration point, preventing the algorithm
   from continuing without a division by zero.")
124

125
(defmfun solution ((solver multi-dimensional-root-solver-f))
126
  "gsl_multiroot_fsolver_root"
127
  (((mpointer solver) :pointer))
128
  :definition :method
129 130
  :c-return (crtn :pointer)
  :return ((make-marray 'double-float :from-pointer crtn))
131
  :documentation			; FDL
132 133
  "The current estimate of the root for the solver.")

134
(defmfun solution ((solver multi-dimensional-root-solver-fdf))
135
  "gsl_multiroot_fdfsolver_root"
136
  (((mpointer solver) :pointer))
137
  :definition :method
138 139
  :c-return (crtn :pointer)
  :return ((make-marray 'double-float :from-pointer crtn))
140 141 142
  :documentation
  "The current estimate of the root for the solver.")

143
(defmfun function-value ((solver multi-dimensional-root-solver-f))
144
  "gsl_multiroot_fsolver_f"
145
  (((mpointer solver) :pointer))
146
  :definition :method
147 148
  :c-return (crtn :pointer)
  :return ((make-marray 'double-float :from-pointer crtn))
149
  :documentation			; FDL
150 151
  "The function value f(x) at the current estimate x of the root for the solver.")

152
(defmfun function-value ((solver multi-dimensional-root-solver-fdf))
153
  "gsl_multiroot_fdfsolver_f"
154
  (((mpointer solver) :pointer))
155
  :definition :method
156 157
  :c-return (crtn :pointer)
  :return ((make-marray 'double-float :from-pointer crtn))
158
  :documentation			; FDL
159 160
  "The function value f(x) at the current estimate x of the root for the solver.")

161
(defmfun last-step ((solver multi-dimensional-root-solver-f))
162
  "gsl_multiroot_fsolver_dx"
163
  (((mpointer solver) :pointer))
164
  :definition :method
165 166
  :c-return (crtn :pointer)
  :return ((make-marray 'double-float :from-pointer crtn))
167
  :documentation			; FDL
168 169
  "The last step dx taken by the solver.")

170
(defmfun last-step ((solver multi-dimensional-root-solver-fdf))
171
  "gsl_multiroot_fsolver_dx"
172
  (((mpointer solver) :pointer))
173
  :definition :method
174 175
  :c-return (crtn :pointer)
  :return ((make-marray 'double-float :from-pointer crtn))
176
  :documentation			; FDL
177 178 179 180 181 182 183
  "The last step dx taken by the solver.")

;;;;****************************************************************************
;;;; Search stopping conditions
;;;;****************************************************************************

;;; The only place we need to pick apart the gsl_multiroot_fsolver
184
;;; struct is here.  We could use last-step etc., but then we'd have
185 186 187 188 189 190 191 192 193 194 195
;;; to discriminate on mfsolver vs. mfdfsolver.
(cffi:defcstruct gsl-multiroot-fsolver
  ;; See /usr/include/gsl/gsl_multiroots.h
  (type :pointer)
  (function :pointer)
  (x :pointer)
  (f :pointer)
  (dx :pointer)
  (state :pointer))

(defun multiroot-slot (solver slot)
196
  (cffi:foreign-slot-value (mpointer solver) 'gsl-multiroot-fsolver slot))
197

liam's avatar
liam committed
198
(defmfun multiroot-test-delta (solver absolute-error relative-error)
199 200 201 202 203
  "gsl_multiroot_test_delta"
  (((multiroot-slot solver 'dx) :pointer)
   ((multiroot-slot solver 'x) :pointer)
   (absolute-error :double) (relative-error :double))
  :c-return :success-continue
204
  :documentation			; FDL
205 206 207 208 209 210 211
  "Test for the convergence of the sequence by comparing the
   last step dx with the absolute error and relative
   errors given to the current position x.  The test returns
   T if the following condition is achieved:
   |dx_i| < epsabs + epsrel |x_i|
   for each component of x and returns NIL otherwise.")

liam's avatar
liam committed
212
(defmfun multiroot-test-residual (solver absolute-error)
213 214 215
  "gsl_multiroot_test_residual"
  (((multiroot-slot solver 'f) :pointer) (absolute-error :double))
  :c-return :success-failure
216
  :documentation			; FDL
217 218 219 220 221 222 223 224 225 226 227 228
  "Test the residual value f against the absolute error,
   returning T if the following condition is achieved:
   \sum_i |f_i| < absolute_error
   and returns NIL otherwise.  This criterion is suitable
   for situations where the precise location of the root x is
   unimportant provided a value can be found where the
   residual is small enough.")

;;;;****************************************************************************
;;;; Algorithms using derivatives
;;;;****************************************************************************

liam's avatar
liam committed
229
(defmpar *powells-hybrid* "gsl_multiroot_fdfsolver_hybridsj"
230
  ;; FDL
231
  "This is a modified version of Powell's Hybrid method as implemented in
liam's avatar
liam committed
232
   the hybrj algorithm in minpack.  Minpack was written by Jorge
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
   J. More, Burton S. Garbow and Kenneth E. Hillstrom.  The Hybrid
   algorithm retains the fast convergence of Newton's method but will also
   reduce the residual when Newton's method is unreliable. 

   The algorithm uses a generalized trust region to keep each step under
   control.  In order to be accepted a proposed new position x' must
   satisfy the condition |D (x' - x)| < \delta, where D is a
   diagonal scaling matrix and \delta is the size of the trust
   region.  The components of D are computed internally, using the
   column norms of the Jacobian to estimate the sensitivity of the residual
   to each component of x.  This improves the behavior of the
   algorithm for badly scaled functions.

   On each iteration the algorithm first determines the standard Newton
   step by solving the system J dx = - f.  If this step falls inside
   the trust region it is used as a trial step in the next stage.  If not,
   the algorithm uses the linear combination of the Newton and gradient
   directions which is predicted to minimize the norm of the function while
   staying inside the trust region,
   dx = - \alpha J^{-1} f(x) - \beta \nabla |f(x)|^2.
   This combination of Newton and gradient directions is referred to as a
   dogleg step.

   The proposed step is now tested by evaluating the function at the
   resulting point, x'.  If the step reduces the norm of the function
   sufficiently then it is accepted and size of the trust region is
   increased.  If the proposed step fails to improve the solution then the
   size of the trust region is decreased and another trial step is
   computed.

   The speed of the algorithm is increased by computing the changes to the
   Jacobian approximately, using a rank-1 update.  If two successive
   attempts fail to reduce the residual then the full Jacobian is
   recomputed.  The algorithm also monitors the progress of the solution
   and returns an error if several steps fail to make any improvement,
268
   'no-progress
269 270
   the iteration is not making any progress, preventing the algorithm from
   continuing.
271
   'jacobian-not-improving
272 273 274
   re-evaluations of the Jacobian indicate that the iteration is not
   making any progress, preventing the algorithm from continuing.")

liam's avatar
liam committed
275
(defmpar *powells-hybrid-unscaled* "gsl_multiroot_fdfsolver_hybridj"
276
  ;; FDL
277 278 279 280 281
  "This algorithm is an unscaled version of *powells-hybrid*.  The steps are
   controlled by a spherical trust region |x' - x| < \delta, instead
   of a generalized region.  This can be useful if the generalized region
   estimated by *powells-hybrid* is inappropriate.")

liam's avatar
liam committed
282
(defmpar *newton-mfdfsolver* "gsl_multiroot_fdfsolver_newton"
283
  ;; FDL
284 285 286 287 288 289 290 291 292 293 294 295
  "Newton's Method is the standard root-polishing algorithm.  The algorithm
   begins with an initial guess for the location of the solution.  On each
   iteration a linear approximation to the function F is used to
   estimate the step which will zero all the components of the residual.
   The iteration is defined by the following sequence,
   x -> x' = x - J{-1} f(x)
   where the Jacobian matrix J is computed from the derivative
   functions provided by f.  The step dx is obtained by solving
   the linear system,
   J dx = - f(x)
   using LU decomposition.")

liam's avatar
liam committed
296
(defmpar *gnewton-mfdfsolver* "gsl_multiroot_fdfsolver_gnewton"
297
  ;; FDL
298 299 300 301 302 303 304 305 306 307 308 309 310
  "A modified version of Newton's method which attempts to improve
   global convergence by requiring every step to reduce the Euclidean norm
   of the residual, |f(x)|.  If the Newton step leads to an increase
   in the norm then a reduced step of relative size,
   t = (\sqrt(1 + 6 r) - 1) / (3 r)
   is proposed, with r being the ratio of norms
   |f(x')|^2/|f(x)|^2.  This procedure is repeated until a suitable step
   size is found.")

;;;;****************************************************************************
;;;; Algorithms without derivatives
;;;;****************************************************************************

liam's avatar
liam committed
311
(defmpar *hybrid-scaled* "gsl_multiroot_fsolver_hybrids"
312 313
  ;; FDL
  "This is a version of the Hybrid algorithm which replaces calls to the
314 315 316 317 318 319
     Jacobian function by its finite difference approximation.  The finite
     difference approximation is computed using gsl_multiroots_fdjac
     with a relative step size of GSL_SQRT_DBL_EPSILON.")
;; Where is this function and parameter?  Only thing that shows in the
;; library is gsl_multiroot_fdjacobian.
 
liam's avatar
liam committed
320
(defmpar *hybrid-unscaled* "gsl_multiroot_fsolver_hybrid"
321
  ;; FDL
322 323 324
  "A finite difference version of the Hybrid algorithm without
   internal scaling.")

liam's avatar
liam committed
325
(defmpar *discrete-newton* "gsl_multiroot_fsolver_dnewton"
326
  ;; FDL
327 328 329 330 331 332 333 334 335 336 337 338 339 340 341
  "The discrete Newton algorithm is the simplest method of solving a
   multidimensional system.  It uses the Newton iteration
   x -> x - J^{-1} f(x)
   where the Jacobian matrix J is approximated by taking finite
   differences of the function f.  The approximation scheme used by
   this implementation is
   J_{ij} = (f_i(x + \delta_j) - f_i(x)) /  \delta_j
   where \delta_j is a step of size \sqrt\epsilon |x_j| with
   \epsilon being the machine precision 
   (\epsilon \approx 2.22 \times 10^-16}).
   The order of convergence of Newton's algorithm is quadratic, but the
   finite differences require n^2 function evaluations on each
   iteration.  The algorithm may become unstable if the finite differences
   are not a good approximation to the true derivatives.")

liam's avatar
liam committed
342
(defmpar *broyden* "gsl_multiroot_fsolver_broyden"
343
  ;; FDL
344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366
  "The Broyden algorithm is a version of the discrete Newton
   algorithm which attempts to avoids the expensive update of the Jacobian
   matrix on each iteration.  The changes to the Jacobian are also
   approximated, using a rank-1 update,
   J^{-1} \to J^{-1} - (J^{-1} df - dx) dx^T J^{-1} / dx^T J^{-1} df
   where the vectors dx and df are the changes in x
   and f.  On the first iteration the inverse Jacobian is estimated
   using finite differences, as in the discrete Newton algorithm.
    
   This approximation gives a fast update but is unreliable if the changes
   are not small, and the estimate of the inverse Jacobian becomes worse as
   time passes.  The algorithm has a tendency to become unstable unless it
   starts close to the root.  The Jacobian is refreshed if this instability
   is detected (consult the source for details).

   This algorithm is included only for demonstration purposes, and is not
   recommended for serious use.")

;;;;****************************************************************************
;;;; Examples
;;;;****************************************************************************

(defparameter *powell-A* 1.0d4)
367
(defun powell (arg0 arg1)
368
  "Powell's test function."
369 370 371
  (values
   (- (* *powell-A* arg0 arg1) 1)
   (+ (exp (- arg0)) (exp (- arg1)) (- (1+ (/ *powell-A*))))))
372

373
(def-mfunction powell 2)
374 375 376 377 378 379

;;; This is the example given in Sec. 34.8.

(defparameter *rosenbrock-a* 1.0d0)
(defparameter *rosenbrock-b* 10.0d0)

380
(defun rosenbrock (arg0 arg1)
381
  "Rosenbrock test function."
382 383 384
  (values
   (* *rosenbrock-a* (- 1 arg0))
   (* *rosenbrock-b* (- arg1 (expt arg0 2)))))
385

386 387
(def-mfunction rosenbrock 2)

388
(defun roots-multi-example (&optional (print-steps t))
389 390
  "Solving Rosenbrock, the example given in Sec. 34.8 of the GSL manual."
  (let ((max-iter 1000))
391
    (let* ((vect #m(-10.0d0 -5.0d0))
392 393
	   (solver (make-multi-dimensional-root-solver-f
		    *hybrid-scaled* rosenbrock vect)))
394 395 396 397 398
      (loop for iter from 0
	 with fnval and argval
	 while (and (< iter max-iter)
		    (not (multiroot-test-residual solver 1.0d-7)))
	 do
399 400 401
	 (iterate solver)
	 (setf fnval (cl-array (function-value solver))
	       argval (cl-array (solution solver)))
402 403 404 405 406 407 408
	 (when print-steps
	   (format t "iter=~d~8tx0=~12,8g~24tx1=~12,8g~38tf0=~12,8g~52tf1=~12,8g~&"
		   iter
		   (aref argval 0)
		   (aref argval 1)
		   (aref fnval 0)
		   (aref fnval 1)))
409 410 411 412 413
	 finally (return
		   (values (aref argval 0)
			   (aref argval 1)
			   (aref fnval 0)
			   (aref fnval 1)))))))
414

415
(defun rosenbrock-df (arg0 arg1)
416
  "The partial derivatives of the Rosenbrock functions."
417 418 419 420 421 422 423 424 425 426 427 428 429
  (declare (ignore arg1))
  (values (- *rosenbrock-a*)
	  0.0d0
	  (* -2 *rosenbrock-b* arg0)
	  *rosenbrock-b*))

;;; Why is it necessary to define a function that calls the two other functions?
(defun rosenbrock-fdf (arg0 arg1)
  (multiple-value-bind (v0 v1)
      (rosenbrock arg0 arg1)
    (multiple-value-bind (j0 j1 j2 j3)
	(rosenbrock-df arg0 arg1)
      (values v0 v1 j0 j1 j2 j3))))
430

431
;;; Because def-solver-functions and def-single-function bind a symbol
432 433
;;; of the same name as the first function, and we want both to run,
;;; we'll make an alias function so we can use both.  
434 435
(defun rosenbrock-f (argument return)
  (rosenbrock argument return))
436 437 438

(def-solver-functions rosenbrock-f rosenbrock-df rosenbrock-fdf 2)

439
(defun roots-multi-example-df (&optional (print-steps t))
440 441 442
  "Solving Rosenbrock with derivatives, the example given in Sec. 34.8
   of the GSL manual."
  (flet ((print-state (iter argval fnval)
443 444 445 446 447 448 449
	   (when print-steps
	     (format t "iter=~d~8tx0=~12,8g~24tx1=~12,8g~38tf0=~12,8g~52tf1=~12,8g~&"
		     iter
		     (maref argval 0)
		     (maref argval 1)
		     (maref fnval 0)
		     (maref fnval 1)))))
450
    (let ((max-iter 1000))
451
      (let* ((vect #m(-10.0d0 -5.0d0))
452 453 454
	     (solver (make-multi-dimensional-root-solver-fdf
		      *gnewton-mfdfsolver* rosenbrock-f vect)))
	(loop for iter from 0
455 456
	   with fnval = (function-value solver)
	   and argval = (solution solver)
457 458 459 460
	   while (and (< iter max-iter)
		      (not (multiroot-test-residual solver 1.0d-7)))
	   initially (print-state iter argval fnval)
	   do
461 462 463
	   (iterate solver)
	   (setf fnval (function-value solver)
		 argval (solution solver))
464 465 466 467 468 469
	   (print-state iter argval fnval)
	   finally (return
		     (values (maref argval 0)
			     (maref argval 1)
			     (maref fnval 0)
			     (maref fnval 1))))))))
470 471 472 473

(save-test roots-multi
 (roots-multi-example nil)
 (roots-multi-example-df nil))