|
50
|
1 // tc-expm.cc -*- C++ -*- |
|
|
2 /* |
|
|
3 |
|
|
4 Copyright (C) 1993 John W. Eaton |
|
|
5 |
|
|
6 This file is part of Octave. |
|
|
7 |
|
|
8 Octave is free software; you can redistribute it and/or modify it |
|
|
9 under the terms of the GNU General Public License as published by the |
|
|
10 Free Software Foundation; either version 2, or (at your option) any |
|
|
11 later version. |
|
|
12 |
|
|
13 Octave is distributed in the hope that it will be useful, but WITHOUT |
|
|
14 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
|
|
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
|
|
16 for more details. |
|
|
17 |
|
|
18 You should have received a copy of the GNU General Public License |
|
|
19 along with Octave; see the file COPYING. If not, write to the Free |
|
|
20 Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. |
|
|
21 |
|
|
22 */ |
|
|
23 |
|
|
24 // Written by A. S. Hodel <scotte@eng.auburn.edu> |
|
|
25 |
|
|
26 #ifdef __GNUG__ |
|
|
27 #pragma implementation |
|
|
28 #endif |
|
|
29 |
|
|
30 #include <math.h> |
|
|
31 |
|
|
32 #include "Matrix.h" |
|
|
33 |
|
|
34 #include "tree-const.h" |
|
|
35 #include "user-prefs.h" |
|
|
36 #include "gripes.h" |
|
|
37 #include "error.h" |
|
|
38 #include "f-expm.h" |
|
|
39 |
|
|
40 #ifdef WITH_DLD |
|
|
41 tree_constant * |
|
|
42 builtin_matrix_exp_2 (tree_constant *args, int nargin, int nargout) |
|
|
43 { |
|
|
44 tree_constant *retval = new tree_constant [2]; |
|
|
45 retval[0] = matrix_exp (args[1]); |
|
|
46 return retval; |
|
|
47 } |
|
|
48 #endif |
|
|
49 |
|
|
50 extern "C" |
|
|
51 { |
|
|
52 double F77_FCN (dlange) (const char*, const int*, const int*, |
|
|
53 const double*, const int*, double*); |
|
|
54 |
|
|
55 double F77_FCN (zlange) (const char*, const int*, const int*, |
|
|
56 const Complex*, const int*, double*); |
|
|
57 } |
|
|
58 |
|
|
59 // XXX FIXME XXX |
|
|
60 extern int empty_arg (tree_constant&); |
|
|
61 |
|
|
62 tree_constant |
|
|
63 matrix_exp (const tree_constant& a) |
|
|
64 { |
|
|
65 tree_constant retval; |
|
|
66 tree_constant tmp = a.make_numeric (); |
|
|
67 |
|
|
68 // Constants for matrix exponential calculation. |
|
|
69 |
|
|
70 static double padec[] = |
|
|
71 { |
|
|
72 5.0000000000000000e-1, |
|
|
73 1.1666666666666667e-1, |
|
|
74 1.6666666666666667e-2, |
|
|
75 1.6025641025641026e-3, |
|
|
76 1.0683760683760684e-4, |
|
|
77 4.8562548562548563e-6, |
|
|
78 1.3875013875013875e-7, |
|
|
79 1.9270852604185938e-9, |
|
|
80 }; |
|
|
81 |
|
|
82 if (empty_arg (tmp)) |
|
|
83 { |
|
|
84 int flag = user_pref.propagate_empty_matrices; |
|
|
85 if (flag != 0) |
|
|
86 { |
|
|
87 if (flag < 0) |
|
|
88 gripe_empty_arg ("expm", 0); |
|
|
89 Matrix m; |
|
|
90 retval = tree_constant (m); |
|
|
91 } |
|
|
92 else gripe_empty_arg ("expm", 1); |
|
|
93 } |
|
|
94 else if (tmp.rows () != tmp.columns ()) |
|
|
95 gripe_square_matrix_required ("expm"); |
|
|
96 else |
|
|
97 { |
|
|
98 int i, j; |
|
|
99 int n_cols = tmp.columns (); |
|
|
100 |
|
|
101 char* balance_job = "B"; // variables for balancing |
|
|
102 |
|
|
103 int sqpow; // power for scaling and squaring |
|
|
104 double inf_norm; // norm of preconditioned matrix |
|
|
105 int minus_one_j; // used in computing pade approx |
|
|
106 |
|
|
107 switch (tmp.const_type ()) |
|
|
108 { |
|
|
109 case tree_constant_rep::complex_matrix_constant: |
|
|
110 { |
|
|
111 ComplexMatrix m = tmp.complex_matrix_value (); |
|
|
112 Complex trshift = 0.0; // trace shift value |
|
|
113 |
|
|
114 // Preconditioning step 1: trace normalization. |
|
|
115 |
|
|
116 for (i = 0; i < n_cols; i++) |
|
|
117 trshift += m.elem (i, i); |
|
|
118 trshift /= n_cols; |
|
|
119 for (i = 0; i < n_cols; i++) |
|
|
120 m.elem (i, i) -= trshift; |
|
|
121 |
|
|
122 // Preconditioning step 2: eigenvalue balancing. |
|
|
123 |
|
|
124 ComplexAEPBALANCE mbal (m, balance_job); |
|
|
125 m = mbal.balanced_matrix (); |
|
|
126 ComplexMatrix d = mbal.balancing_matrix (); |
|
|
127 |
|
|
128 // Preconditioning step 3: scaling. |
|
|
129 |
|
|
130 ColumnVector work (n_cols); |
|
|
131 inf_norm = F77_FCN (zlange) ("I", &n_cols, &n_cols, m. |
|
|
132 fortran_vec (), &n_cols, |
|
|
133 work.fortran_vec ()); |
|
|
134 |
|
|
135 sqpow = 1.0 + log (inf_norm) / log (2.0); |
|
|
136 |
|
|
137 // Check whether we need to square at all. |
|
|
138 |
|
|
139 if (sqpow <= 0.0) |
|
|
140 sqpow = 0.0; |
|
|
141 else |
|
|
142 { |
|
|
143 for (inf_norm = 1.0, i = 0; i < sqpow; i++) |
|
|
144 inf_norm *= 2.0; |
|
|
145 |
|
|
146 m = m / inf_norm; |
|
|
147 } |
|
|
148 |
|
|
149 // npp, dpp: pade' approx polynomial matrices. |
|
|
150 |
|
|
151 ComplexMatrix npp (n_cols, n_cols, 0.0); |
|
|
152 ComplexMatrix dpp = npp; |
|
|
153 |
|
|
154 // Now powers a^8 ... a^1. |
|
|
155 |
|
|
156 minus_one_j = -1; |
|
|
157 for (j = 7; j >= 0; j--) |
|
|
158 { |
|
|
159 npp = m * npp + m * padec[j]; |
|
|
160 dpp = m * dpp + m * (minus_one_j * padec[j]); |
|
|
161 minus_one_j *= -1; |
|
|
162 } |
|
|
163 |
|
|
164 // Zero power. |
|
|
165 |
|
|
166 dpp = -dpp; |
|
|
167 for (j = 0; j < n_cols; j++) |
|
|
168 { |
|
|
169 npp.elem (j, j) += 1.0; |
|
|
170 dpp.elem (j, j) += 1.0; |
|
|
171 } |
|
|
172 |
|
|
173 // Compute pade approximation = inverse (dpp) * npp. |
|
|
174 |
|
|
175 ComplexMatrix result = dpp.solve (npp); |
|
|
176 |
|
|
177 // Reverse preconditioning step 3: repeated squaring. |
|
|
178 |
|
|
179 while (sqpow) |
|
|
180 { |
|
|
181 result = result * result; |
|
|
182 sqpow--; |
|
|
183 } |
|
|
184 |
|
|
185 // reverse preconditioning step 2: inverse balancing XXX FIXME XXX: |
|
|
186 // should probably do this with lapack calls instead of a complete |
|
|
187 // matrix inversion. |
|
|
188 |
|
|
189 result = result.transpose (); |
|
|
190 d = d.transpose (); |
|
|
191 result = result * d; |
|
|
192 result = d.solve (result); |
|
|
193 result = result.transpose (); |
|
|
194 |
|
|
195 // Reverse preconditioning step 1: fix trace normalization. |
|
|
196 |
|
|
197 result = result * exp (trshift); |
|
|
198 |
|
|
199 retval = tree_constant (result); |
|
|
200 } |
|
|
201 break; |
|
|
202 case tree_constant_rep::complex_scalar_constant: |
|
|
203 { |
|
|
204 Complex c = tmp.complex_value (); |
|
|
205 retval = tree_constant (exp (c)); |
|
|
206 } |
|
|
207 break; |
|
|
208 case tree_constant_rep::matrix_constant: |
|
|
209 { |
|
|
210 |
|
|
211 // Compute the exponential. |
|
|
212 |
|
|
213 Matrix m = tmp.matrix_value (); |
|
|
214 |
|
|
215 double trshift = 0; // trace shift value |
|
|
216 |
|
|
217 // Preconditioning step 1: trace normalization. |
|
|
218 |
|
|
219 for (i = 0; i < n_cols; i++) |
|
|
220 trshift += m.elem (i, i); |
|
|
221 trshift /= n_cols; |
|
|
222 for (i = 0; i < n_cols; i++) |
|
|
223 m.elem (i, i) -= trshift; |
|
|
224 |
|
|
225 // Preconditioning step 2: balancing. |
|
|
226 |
|
|
227 AEPBALANCE mbal (m, balance_job); |
|
|
228 m = mbal.balanced_matrix (); |
|
|
229 Matrix d = mbal.balancing_matrix (); |
|
|
230 |
|
|
231 // Preconditioning step 3: scaling. |
|
|
232 |
|
|
233 ColumnVector work(n_cols); |
|
|
234 inf_norm = F77_FCN (dlange) ("I", &n_cols, &n_cols, |
|
|
235 m.fortran_vec (), &n_cols, |
|
|
236 work.fortran_vec ()); |
|
|
237 |
|
|
238 sqpow = 1.0 + log (inf_norm) / log (2.0); |
|
|
239 |
|
|
240 // Check whether we need to square at all. |
|
|
241 |
|
|
242 if (sqpow <= 0.0) |
|
|
243 sqpow = 0.0; |
|
|
244 else |
|
|
245 { |
|
|
246 for (inf_norm = 1.0, i = 0; i < sqpow; i++) |
|
|
247 inf_norm *= 2.0; |
|
|
248 |
|
|
249 m = m / inf_norm; |
|
|
250 } |
|
|
251 |
|
|
252 // npp, dpp: pade' approx polynomial matrices. |
|
|
253 |
|
|
254 Matrix npp (n_cols, n_cols, 0.0); |
|
|
255 Matrix dpp = npp; |
|
|
256 |
|
|
257 // now powers a^8 ... a^1. |
|
|
258 |
|
|
259 minus_one_j = -1; |
|
|
260 for (j = 7; j >= 0; j--) |
|
|
261 { |
|
|
262 npp = m * npp + m * padec[j]; |
|
|
263 dpp = m * dpp + m * (minus_one_j * padec[j]); |
|
|
264 minus_one_j *= -1; |
|
|
265 } |
|
|
266 // Zero power. |
|
|
267 |
|
|
268 dpp = -dpp; |
|
|
269 for(j = 0; j < n_cols; j++) |
|
|
270 { |
|
|
271 npp.elem (j, j) += 1.0; |
|
|
272 dpp.elem (j, j) += 1.0; |
|
|
273 } |
|
|
274 |
|
|
275 // Compute pade approximation = inverse (dpp) * npp. |
|
|
276 |
|
|
277 Matrix result = dpp.solve (npp); |
|
|
278 |
|
|
279 // Reverse preconditioning step 3: repeated squaring. |
|
|
280 |
|
|
281 while(sqpow) |
|
|
282 { |
|
|
283 result = result * result; |
|
|
284 sqpow--; |
|
|
285 } |
|
|
286 |
|
|
287 // Reverse preconditioning step 2: inverse balancing. |
|
|
288 |
|
|
289 result = result.transpose(); |
|
|
290 d = d.transpose (); |
|
|
291 result = result * d; |
|
|
292 result = d.solve (result); |
|
|
293 result = result.transpose (); |
|
|
294 |
|
|
295 // Reverse preconditioning step 1: fix trace normalization. |
|
|
296 |
|
|
297 result = result * exp (trshift); |
|
|
298 |
|
|
299 retval = tree_constant (result); |
|
|
300 } |
|
|
301 break; |
|
|
302 case tree_constant_rep::scalar_constant: |
|
|
303 { |
|
|
304 double d = tmp.double_value (); |
|
|
305 retval = tree_constant (exp (d)); |
|
|
306 } |
|
|
307 break; |
|
|
308 default: |
|
|
309 panic_impossible(); |
|
|
310 break; |
|
|
311 } |
|
|
312 } |
|
|
313 return retval; |
|
|
314 } |
|
|
315 |
|
|
316 /* |
|
|
317 ;;; Local Variables: *** |
|
|
318 ;;; mode: C++ *** |
|
|
319 ;;; page-delimiter: "^/\\*" *** |
|
|
320 ;;; End: *** |
|
|
321 */ |
