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
|
SUBROUTINE MB04PB( N, ILO, A, LDA, QG, LDQG, CS, TAU, DWORK,
$ LDWORK, INFO )
C
C SLICOT RELEASE 5.0.
C
C Copyright (c) 2002-2009 NICONET e.V.
C
C This program is free software: you can redistribute it and/or
C modify it under the terms of the GNU General Public License as
C published by the Free Software Foundation, either version 2 of
C the License, or (at your option) any later version.
C
C This program is distributed in the hope that it will be useful,
C but WITHOUT ANY WARRANTY; without even the implied warranty of
C MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
C GNU General Public License for more details.
C
C You should have received a copy of the GNU General Public License
C along with this program. If not, see
C <http://www.gnu.org/licenses/>.
C
C PURPOSE
C
C To reduce a Hamiltonian matrix,
C
C [ A G ]
C H = [ T ] ,
C [ Q -A ]
C
C where A is an N-by-N matrix and G,Q are N-by-N symmetric matrices,
C to Paige/Van Loan (PVL) form. That is, an orthogonal symplectic U
C is computed so that
C
C T [ Aout Gout ]
C U H U = [ T ] ,
C [ Qout -Aout ]
C
C where Aout is upper Hessenberg and Qout is diagonal.
C Blocked version.
C
C ARGUMENTS
C
C Input/Output Parameters
C
C N (input) INTEGER
C The order of the matrix A. N >= 0.
C
C ILO (input) INTEGER
C It is assumed that A is already upper triangular and Q is
C zero in rows and columns 1:ILO-1. ILO is normally set by a
C previous call to MB04DD; otherwise it should be set to 1.
C 1 <= ILO <= N, if N > 0; ILO = 1, if N = 0.
C
C A (input/output) DOUBLE PRECISION array, dimension (LDA,N)
C On entry, the leading N-by-N part of this array must
C contain the matrix A.
C On exit, the leading N-by-N part of this array contains
C the matrix Aout and, in the zero part of Aout,
C information about the elementary reflectors used to
C compute the PVL factorization.
C
C LDA INTEGER
C The leading dimension of the array A. LDA >= MAX(1,N).
C
C QG (input/output) DOUBLE PRECISION array, dimension
C (LDQG,N+1)
C On entry, the leading N-by-N+1 part of this array must
C contain the lower triangular part of the matrix Q and
C the upper triangular part of the matrix G.
C On exit, the leading N-by-N+1 part of this array contains
C the diagonal of the matrix Qout, the upper triangular part
C of the matrix Gout and, in the zero parts of Qout,
C information about the elementary reflectors used to
C compute the PVL factorization.
C
C LDQG INTEGER
C The leading dimension of the array QG. LDQG >= MAX(1,N).
C
C CS (output) DOUBLE PRECISION array, dimension (2N-2)
C On exit, the first 2N-2 elements of this array contain the
C cosines and sines of the symplectic Givens rotations used
C to compute the PVL factorization.
C
C TAU (output) DOUBLE PRECISION array, dimension (N-1)
C On exit, the first N-1 elements of this array contain the
C scalar factors of some of the elementary reflectors.
C
C Workspace
C
C DWORK DOUBLE PRECISION array, dimension (LDWORK)
C On exit, if INFO = 0, DWORK(1) returns the optimal
C value of LDWORK, 8*N*NB + 3*NB, where NB is the optimal
C block size determined by the function UE01MD.
C On exit, if INFO = -10, DWORK(1) returns the minimum
C value of LDWORK.
C
C LDWORK INTEGER
C The length of the array DWORK. LDWORK >= MAX(1,N-1).
C
C Error Indicator
C
C INFO INTEGER
C = 0: successful exit;
C < 0: if INFO = -i, the i-th argument had an illegal
C value.
C
C METHOD
C
C The matrix U is represented as a product of symplectic reflectors
C and Givens rotators
C
C U = diag( H(1),H(1) ) G(1) diag( F(1),F(1) )
C diag( H(2),H(2) ) G(2) diag( F(2),F(2) )
C ....
C diag( H(n-1),H(n-1) ) G(n-1) diag( F(n-1),F(n-1) ).
C
C Each H(i) has the form
C
C H(i) = I - tau * v * v'
C
C where tau is a real scalar, and v is a real vector with
C v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in
C QG(i+2:n,i), and tau in QG(i+1,i).
C
C Each F(i) has the form
C
C F(i) = I - nu * w * w'
C
C where nu is a real scalar, and w is a real vector with
C w(1:i) = 0 and w(i+1) = 1; w(i+2:n) is stored on exit in
C A(i+2:n,i), and nu in TAU(i).
C
C Each G(i) is a Givens rotator acting on rows i+1 and n+i+1,
C where the cosine is stored in CS(2*i-1) and the sine in
C CS(2*i).
C
C NUMERICAL ASPECTS
C
C The algorithm requires O(N**3) floating point operations and is
C strongly backward stable.
C
C REFERENCES
C
C [1] C. F. VAN LOAN:
C A symplectic method for approximating all the eigenvalues of
C a Hamiltonian matrix.
C Linear Algebra and its Applications, 61, pp. 233-251, 1984.
C
C [2] D. KRESSNER:
C Block algorithms for orthogonal symplectic factorizations.
C BIT, 43 (4), pp. 775-790, 2003.
C
C CONTRIBUTORS
C
C D. Kressner (Technical Univ. Berlin, Germany) and
C P. Benner (Technical Univ. Chemnitz, Germany), December 2003.
C
C REVISIONS
C
C V. Sima, Nov. 2008 (SLICOT version of the HAPACK routine DHAPVB).
C
C KEYWORDS
C
C Elementary matrix operations, Hamiltonian matrix.
C
C ******************************************************************
C
C .. Parameters ..
DOUBLE PRECISION ZERO, ONE
PARAMETER ( ZERO = 0.0D0, ONE = 1.0D0 )
C .. Scalar Arguments ..
INTEGER ILO, INFO, LDA, LDQG, LDWORK, N
C .. Array Arguments ..
DOUBLE PRECISION A(LDA,*), CS(*), DWORK(*), QG(LDQG,*), TAU(*)
C .. Local Scalars ..
INTEGER I, IB, IERR, NB, NBMIN, NH, NIB, NNB, NX, PDW,
$ PXA, PXG, PXQ, PYA, WRKOPT
C .. External Functions ..
INTEGER UE01MD
EXTERNAL UE01MD
C .. External Subroutines ..
EXTERNAL DGEMM, DSYR2K, MB04PA, MB04PU, XERBLA
C .. Intrinsic Functions ..
INTRINSIC DBLE, MAX, MIN
C
C .. Executable Statements ..
C
C Check the scalar input parameters.
C
INFO = 0
IF ( N.LT.0 ) THEN
INFO = -1
ELSE IF ( ILO.LT.1 .OR. ILO.GT.MAX( 1, N ) ) THEN
INFO = -2
ELSE IF ( LDA.LT.MAX( 1, N ) ) THEN
INFO = -4
ELSE IF ( LDQG.LT.MAX( 1, N ) ) THEN
INFO = -6
ELSE IF ( LDWORK.LT.MAX( 1, N-1 ) ) THEN
DWORK(1) = DBLE( MAX( 1, N-1 ) )
INFO = -10
END IF
C
C Return if there were illegal values.
C
IF ( INFO.NE.0 ) THEN
CALL XERBLA( 'MB04PB', -INFO )
RETURN
END IF
C
C Set elements 1:ILO-1 of TAU and CS.
C
DO 10 I = 1, ILO - 1
TAU( I ) = ZERO
CS(2*I-1) = ONE
CS(2*I) = ZERO
10 CONTINUE
C
C Quick return if possible.
C
IF ( N.LE.ILO ) THEN
DWORK(1) = ONE
RETURN
END IF
C
C Determine the block size.
C
NH = N - ILO + 1
NB = UE01MD( 1, 'MB04PB', ' ', N, ILO, -1 )
NBMIN = 2
WRKOPT = N-1
IF ( NB.GT.1 .AND. NB.LT.NH ) THEN
C
C Determine when to cross over from blocked to unblocked code.
C
NX = MAX( NB, UE01MD( 3, 'MB04PB', ' ', N, ILO, -1 ) )
IF ( NX.LT.NH ) THEN
C
C Check whether workspace is large enough for blocked code.
C
WRKOPT = 8*N*NB + 3*NB
IF ( LDWORK.LT.WRKOPT ) THEN
C
C Not enough workspace available. Determine minimum value
C of NB, and reduce NB.
C
NBMIN = MAX( 2, UE01MD( 2, 'MB04PB', ' ', N, ILO, -1 ) )
NB = LDWORK / ( 8*N + 3 )
END IF
END IF
END IF
C
NNB = N*NB
PXA = 1
PYA = PXA + 2*NNB
PXQ = PYA + 2*NNB
PXG = PXQ + 2*NNB
PDW = PXG + 2*NNB
C
IF ( NB.LT.NBMIN .OR. NB.GE.NH ) THEN
C
C Use unblocked code.
C
I = ILO
C
ELSE
DO 20 I = ILO, N-NX-1, NB
IB = MIN( NB, N-I )
NIB = N*IB
C
C Reduce rows and columns i:i+nb-1 to PVL form and return the
C matrices XA, XG, XQ, and YA which are needed to update the
C unreduced parts of the matrices.
C
CALL MB04PA( .TRUE., N-I+1, I-1, IB, A(1,I), LDA, QG(1,I),
$ LDQG, DWORK(PXA), N, DWORK(PXG), N,
$ DWORK(PXQ), N, DWORK(PYA), N, CS(2*I-1),
$ TAU(I), DWORK(PDW) )
IF ( N.GT.I+IB ) THEN
C
C Update the submatrix A(1:n,i+ib+1:n).
C
CALL DGEMM( 'No transpose', 'Transpose', N-I-IB, N-I-IB,
$ IB, ONE, QG(I+IB+1,I), LDQG, DWORK(PXA+IB+1),
$ N, ONE, A(I+IB+1,I+IB+1), LDA )
CALL DGEMM( 'No transpose', 'Transpose', N-I-IB, N-I-IB,
$ IB, ONE, A(I+IB+1,I), LDA,
$ DWORK(PXA+NIB+IB+1), N, ONE,
$ A(I+IB+1,I+IB+1), LDA )
CALL DGEMM( 'No transpose', 'Transpose', N, N-I-IB, IB,
$ ONE, DWORK(PYA), N, QG(I+IB+1,I), LDQG, ONE,
$ A(1,I+IB+1), LDA )
CALL DGEMM( 'No transpose', 'Transpose', N, N-I-IB, IB,
$ ONE, DWORK(PYA+NIB), N, A(I+IB+1,I), LDA,
$ ONE, A(1,I+IB+1), LDA )
C
C Update the submatrix Q(i+ib+1:n,i+ib+1:n).
C
CALL DSYR2K( 'Lower', 'No Transpose', N-I-IB, IB, ONE,
$ DWORK(PXQ+IB+1), N, QG(I+IB+1,I), LDQG, ONE,
$ QG(I+IB+1,I+IB+1), LDQG )
CALL DSYR2K( 'Lower', 'No Transpose', N-I-IB, IB, ONE,
$ DWORK(PXQ+NIB+IB+1), N, A(I+IB+1,I), LDA,
$ ONE, QG(I+IB+1,I+IB+1), LDQG )
C
C Update the submatrix G(1:n,1:n).
C
CALL DGEMM( 'No transpose', 'Transpose', I+IB, N-I-IB,
$ IB, ONE, DWORK(PXG), N, QG(I+IB+1,I), LDQG,
$ ONE, QG(1,I+IB+2), LDQG )
CALL DGEMM( 'No transpose', 'Transpose', I+IB, N-I-IB,
$ IB, ONE, DWORK(PXG+NIB), N, A(I+IB+1,I), LDA,
$ ONE, QG(1,I+IB+2), LDQG )
CALL DSYR2K( 'Upper', 'No Transpose', N-I-IB, IB, ONE,
$ DWORK(PXG+IB+I), N, QG(I+IB+1,I), LDQG, ONE,
$ QG(I+IB+1,I+IB+2), LDQG )
CALL DSYR2K( 'Upper', 'No Transpose', N-I-IB, IB, ONE,
$ DWORK(PXG+NIB+IB+I), N, A(I+IB+1,I), LDA,
$ ONE, QG(I+IB+1,I+IB+2), LDQG )
END IF
20 CONTINUE
END IF
C
C Unblocked code to reduce the rest of the matrices.
C
CALL MB04PU( N, I, A, LDA, QG, LDQG, CS, TAU, DWORK, LDWORK,
$ IERR )
C
DWORK( 1 ) = DBLE( WRKOPT )
C
RETURN
C *** Last line of MB04PB ***
END
|
