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
|
<HTML>
<HEAD><TITLE>SB02SD - SLICOT Library Routine Documentation</TITLE>
</HEAD>
<BODY>
<H2><A Name="SB02SD">SB02SD</A></H2>
<H3>
Estimating conditioning and forward error bound for the solution of discrete-time Riccati equation
</H3>
<A HREF ="#Specification"><B>[Specification]</B></A>
<A HREF ="#Arguments"><B>[Arguments]</B></A>
<A HREF ="#Method"><B>[Method]</B></A>
<A HREF ="#References"><B>[References]</B></A>
<A HREF ="#Comments"><B>[Comments]</B></A>
<A HREF ="#Example"><B>[Example]</B></A>
<P>
<B><FONT SIZE="+1">Purpose</FONT></B>
<PRE>
To estimate the conditioning and compute an error bound on the
solution of the real discrete-time matrix algebraic Riccati
equation (see FURTHER COMMENTS)
-1
X = op(A)'*X*(I_n + G*X) *op(A) + Q, (1)
where op(A) = A or A' (A**T) and Q, G are symmetric (Q = Q**T,
G = G**T). The matrices A, Q and G are N-by-N and the solution X
is N-by-N.
</PRE>
<A name="Specification"><B><FONT SIZE="+1">Specification</FONT></B></A>
<PRE>
SUBROUTINE SB02SD( JOB, FACT, TRANA, UPLO, LYAPUN, N, A, LDA, T,
$ LDT, U, LDU, G, LDG, Q, LDQ, X, LDX, SEPD,
$ RCOND, FERR, IWORK, DWORK, LDWORK, INFO )
C .. Scalar Arguments ..
CHARACTER FACT, JOB, LYAPUN, TRANA, UPLO
INTEGER INFO, LDA, LDG, LDQ, LDT, LDU, LDWORK, LDX, N
DOUBLE PRECISION FERR, RCOND, SEPD
C .. Array Arguments ..
INTEGER IWORK( * )
DOUBLE PRECISION A( LDA, * ), DWORK( * ), G( LDG, * ),
$ Q( LDQ, * ), T( LDT, * ), U( LDU, * ),
$ X( LDX, * )
</PRE>
<A name="Arguments"><B><FONT SIZE="+1">Arguments</FONT></B></A>
<P>
<B>Mode Parameters</B>
<PRE>
JOB CHARACTER*1
Specifies the computation to be performed, as follows:
= 'C': Compute the reciprocal condition number only;
= 'E': Compute the error bound only;
= 'B': Compute both the reciprocal condition number and
the error bound.
FACT CHARACTER*1
Specifies whether or not the real Schur factorization of
the matrix Ac = inv(I_n + G*X)*A (if TRANA = 'N'), or
Ac = A*inv(I_n + X*G) (if TRANA = 'T' or 'C'), is supplied
on entry, as follows:
= 'F': On entry, T and U (if LYAPUN = 'O') contain the
factors from the real Schur factorization of the
matrix Ac;
= 'N': The Schur factorization of Ac will be computed
and the factors will be stored in T and U (if
LYAPUN = 'O').
TRANA CHARACTER*1
Specifies the form of op(A) to be used, as follows:
= 'N': op(A) = A (No transpose);
= 'T': op(A) = A**T (Transpose);
= 'C': op(A) = A**T (Conjugate transpose = Transpose).
UPLO CHARACTER*1
Specifies which part of the symmetric matrices Q and G is
to be used, as follows:
= 'U': Upper triangular part;
= 'L': Lower triangular part.
LYAPUN CHARACTER*1
Specifies whether or not the original Lyapunov equations
should be solved in the iterative estimation process,
as follows:
= 'O': Solve the original Lyapunov equations, updating
the right-hand sides and solutions with the
matrix U, e.g., RHS <-- U'*RHS*U;
= 'R': Solve reduced Lyapunov equations only, without
updating the right-hand sides and solutions.
</PRE>
<B>Input/Output Parameters</B>
<PRE>
N (input) INTEGER
The order of the matrices A, X, Q, and G. N >= 0.
A (input) DOUBLE PRECISION array, dimension (LDA,N)
If FACT = 'N' or LYAPUN = 'O', the leading N-by-N part of
this array must contain the matrix A.
If FACT = 'F' and LYAPUN = 'R', A is not referenced.
LDA INTEGER
The leading dimension of the array A.
LDA >= max(1,N), if FACT = 'N' or LYAPUN = 'O';
LDA >= 1, if FACT = 'F' and LYAPUN = 'R'.
T (input or output) DOUBLE PRECISION array, dimension
(LDT,N)
If FACT = 'F', then T is an input argument and on entry,
the leading N-by-N upper Hessenberg part of this array
must contain the upper quasi-triangular matrix T in Schur
canonical form from a Schur factorization of Ac (see
argument FACT).
If FACT = 'N', then T is an output argument and on exit,
if INFO = 0 or INFO = N+1, the leading N-by-N upper
Hessenberg part of this array contains the upper quasi-
triangular matrix T in Schur canonical form from a Schur
factorization of Ac (see argument FACT).
LDT INTEGER
The leading dimension of the array T. LDT >= max(1,N).
U (input or output) DOUBLE PRECISION array, dimension
(LDU,N)
If LYAPUN = 'O' and FACT = 'F', then U is an input
argument and on entry, the leading N-by-N part of this
array must contain the orthogonal matrix U from a real
Schur factorization of Ac (see argument FACT).
If LYAPUN = 'O' and FACT = 'N', then U is an output
argument and on exit, if INFO = 0 or INFO = N+1, it
contains the orthogonal N-by-N matrix from a real Schur
factorization of Ac (see argument FACT).
If LYAPUN = 'R', the array U is not referenced.
LDU INTEGER
The leading dimension of the array U.
LDU >= 1, if LYAPUN = 'R';
LDU >= MAX(1,N), if LYAPUN = 'O'.
G (input) DOUBLE PRECISION array, dimension (LDG,N)
If UPLO = 'U', the leading N-by-N upper triangular part of
this array must contain the upper triangular part of the
matrix G.
If UPLO = 'L', the leading N-by-N lower triangular part of
this array must contain the lower triangular part of the
matrix G. _
Matrix G should correspond to G in the "reduced" Riccati
equation (with matrix T, instead of A), if LYAPUN = 'R'.
See METHOD.
LDG INTEGER
The leading dimension of the array G. LDG >= max(1,N).
Q (input) DOUBLE PRECISION array, dimension (LDQ,N)
If UPLO = 'U', the leading N-by-N upper triangular part of
this array must contain the upper triangular part of the
matrix Q.
If UPLO = 'L', the leading N-by-N lower triangular part of
this array must contain the lower triangular part of the
matrix Q. _
Matrix Q should correspond to Q in the "reduced" Riccati
equation (with matrix T, instead of A), if LYAPUN = 'R'.
See METHOD.
LDQ INTEGER
The leading dimension of the array Q. LDQ >= max(1,N).
X (input) DOUBLE PRECISION array, dimension (LDX,N)
The leading N-by-N part of this array must contain the
symmetric solution matrix of the original Riccati
equation (with matrix A), if LYAPUN = 'O', or of the
"reduced" Riccati equation (with matrix T), if
LYAPUN = 'R'. See METHOD.
LDX INTEGER
The leading dimension of the array X. LDX >= max(1,N).
SEPD (output) DOUBLE PRECISION
If JOB = 'C' or JOB = 'B', the estimated quantity
sepd(op(Ac),op(Ac)').
If N = 0, or X = 0, or JOB = 'E', SEPD is not referenced.
RCOND (output) DOUBLE PRECISION
If JOB = 'C' or JOB = 'B', an estimate of the reciprocal
condition number of the discrete-time Riccati equation.
If N = 0 or X = 0, RCOND is set to 1 or 0, respectively.
If JOB = 'E', RCOND is not referenced.
FERR (output) DOUBLE PRECISION
If JOB = 'E' or JOB = 'B', an estimated forward error
bound for the solution X. If XTRUE is the true solution,
FERR bounds the magnitude of the largest entry in
(X - XTRUE) divided by the magnitude of the largest entry
in X.
If N = 0 or X = 0, FERR is set to 0.
If JOB = 'C', FERR is not referenced.
</PRE>
<B>Workspace</B>
<PRE>
IWORK INTEGER array, dimension (N*N)
DWORK DOUBLE PRECISION array, dimension (LDWORK)
On exit, if INFO = 0 or INFO = N+1, DWORK(1) returns the
optimal value of LDWORK.
LDWORK INTEGER
The dimension of the array DWORK.
Let LWA = N*N, if LYAPUN = 'O';
LWA = 0, otherwise,
and LWN = N, if LYAPUN = 'R' and JOB = 'E' or 'B';
LWN = 0, otherwise.
If FACT = 'N', then
LDWORK = MAX(LWA + 5*N, MAX(3,2*N*N) + N*N),
if JOB = 'C';
LDWORK = MAX(LWA + 5*N, MAX(3,2*N*N) + 2*N*N + LWN),
if JOB = 'E' or 'B'.
If FACT = 'F', then
LDWORK = MAX(3,2*N*N) + N*N, if JOB = 'C';
LDWORK = MAX(3,2*N*N) + 2*N*N + LWN,
if JOB = 'E' or 'B'.
For good performance, LDWORK must generally be larger.
If LDWORK = -1, then a workspace query is assumed;
the routine only calculates the optimal size of the
DWORK array, returns this value as the first entry of
the DWORK array, and no error message related to LDWORK
is issued by XERBLA.
</PRE>
<B>Error Indicator</B>
<PRE>
INFO INTEGER
= 0: successful exit;
< 0: if INFO = -i, the i-th argument had an illegal
value;
> 0: if INFO = i, i <= N, the QR algorithm failed to
complete the reduction of the matrix Ac to Schur
canonical form (see LAPACK Library routine DGEES);
on exit, the matrix T(i+1:N,i+1:N) contains the
partially converged Schur form, and DWORK(i+1:N) and
DWORK(N+i+1:2*N) contain the real and imaginary
parts, respectively, of the converged eigenvalues;
this error is unlikely to appear;
= N+1: if T has almost reciprocal eigenvalues; perturbed
values were used to solve Lyapunov equations, but
the matrix T, if given (for FACT = 'F'), is
unchanged.
</PRE>
<A name="Method"><B><FONT SIZE="+1">Method</FONT></B></A>
<PRE>
The condition number of the Riccati equation is estimated as
cond = ( norm(Theta)*norm(A) + norm(inv(Omega))*norm(Q) +
norm(Pi)*norm(G) ) / norm(X),
where Omega, Theta and Pi are linear operators defined by
Omega(W) = op(Ac)'*W*op(Ac) - W,
Theta(W) = inv(Omega(op(W)'*X*op(Ac) + op(Ac)'X*op(W))),
Pi(W) = inv(Omega(op(Ac)'*X*W*X*op(Ac))),
and Ac = inv(I_n + G*X)*A (if TRANA = 'N'), or
Ac = A*inv(I_n + X*G) (if TRANA = 'T' or 'C').
Note that the Riccati equation (1) is equivalent to
X = op(Ac)'*X*op(Ac) + op(Ac)'*X*G*X*op(Ac) + Q, (2)
and to
_ _ _ _ _ _
X = op(T)'*X*op(T) + op(T)'*X*G*X*op(T) + Q, (3)
_ _ _
where X = U'*X*U, Q = U'*Q*U, and G = U'*G*U, with U the
orthogonal matrix reducing Ac to a real Schur form, T = U'*Ac*U.
The routine estimates the quantities
sepd(op(Ac),op(Ac)') = 1 / norm(inv(Omega)),
norm(Theta) and norm(Pi) using 1-norm condition estimator.
The forward error bound is estimated using a practical error bound
similar to the one proposed in [2].
</PRE>
<A name="References"><B><FONT SIZE="+1">References</FONT></B></A>
<PRE>
[1] Ghavimi, A.R. and Laub, A.J.
Backward error, sensitivity, and refinement of computed
solutions of algebraic Riccati equations.
Numerical Linear Algebra with Applications, vol. 2, pp. 29-49,
1995.
[2] Higham, N.J.
Perturbation theory and backward error for AX-XB=C.
BIT, vol. 33, pp. 124-136, 1993.
[3] Petkov, P.Hr., Konstantinov, M.M., and Mehrmann, V.
DGRSVX and DMSRIC: Fortran 77 subroutines for solving
continuous-time matrix algebraic Riccati equations with
condition and accuracy estimates.
Preprint SFB393/98-16, Fak. f. Mathematik, Tech. Univ.
Chemnitz, May 1998.
</PRE>
<A name="Numerical Aspects"><B><FONT SIZE="+1">Numerical Aspects</FONT></B></A>
<PRE> 3
The algorithm requires 0(N ) operations.
The accuracy of the estimates obtained depends on the solution
accuracy and on the properties of the 1-norm estimator.
</PRE>
<A name="Comments"><B><FONT SIZE="+1">Further Comments</FONT></B></A>
<PRE>
The option LYAPUN = 'R' may occasionally produce slightly worse
or better estimates, and it is much faster than the option 'O'.
When SEPD is computed and it is zero, the routine returns
immediately, with RCOND and FERR (if requested) set to 0 and 1,
respectively. In this case, the equation is singular.
Let B be an N-by-M matrix (if TRANA = 'N') or an M-by-N matrix
(if TRANA = 'T' or 'C'), let R be an M-by-M symmetric positive
definite matrix (R = R**T), and denote G = op(B)*inv(R)*op(B)'.
Then, the Riccati equation (1) is equivalent to the standard
discrete-time matrix algebraic Riccati equation
X = op(A)'*X*op(A) - (4)
-1
op(A)'*X*op(B)*(R + op(B)'*X*op(B)) *op(B)'*X*op(A) + Q.
By symmetry, the equation (1) is also equivalent to
-1
X = op(A)'*(I_n + X*G) *X*op(A) + Q.
</PRE>
<A name="Example"><B><FONT SIZE="+1">Example</FONT></B></A>
<P>
<B>Program Text</B>
<PRE>
* SB02SD EXAMPLE PROGRAM TEXT
*
* .. Parameters ..
INTEGER NIN, NOUT
PARAMETER ( NIN = 5, NOUT = 6 )
INTEGER NMAX
PARAMETER ( NMAX = 20 )
INTEGER LDA, LDG, LDQ, LDT, LDU, LDX
PARAMETER ( LDA = NMAX, LDG = NMAX, LDQ = NMAX, LDT = NMAX,
$ LDU = NMAX, LDX = NMAX )
INTEGER LIWORK
PARAMETER ( LIWORK = NMAX*NMAX )
INTEGER LDWORK
PARAMETER ( LDWORK = 8*NMAX*NMAX + 10*NMAX )
DOUBLE PRECISION ZERO, ONE
PARAMETER ( ZERO = 0.0D+0, ONE = 1.0D+0 )
* .. Local Scalars ..
DOUBLE PRECISION FERR, RCND, RCOND, SEPD
INTEGER I, INFO1, INFO2, INFO3, IS, IU, IW, J, N, N2,
$ SDIM
CHARACTER*1 FACT, JOB, JOBS, LYAPUN, TRANA, TRANAT, UPLO
* .. Local Arrays ..
LOGICAL BWORK(2*NMAX)
INTEGER IWORK(LIWORK)
DOUBLE PRECISION A(LDA,NMAX), AS(LDA,NMAX), DWORK(LDWORK),
$ G(LDG,NMAX), Q(LDQ,NMAX), T(LDT,NMAX),
$ U(LDU,NMAX), X(LDX,NMAX)
* .. External Functions ..
LOGICAL LSAME, SELECT
EXTERNAL LSAME, SELECT
* .. External Subroutines ..
EXTERNAL DGEES, DGESV, DLACPY, DLASET, DSWAP, DSYMM,
$ MA02AD, MA02ED, MB01RU, SB02MD, SB02SD
* .. Executable Statements ..
*
WRITE ( NOUT, FMT = 99999 )
* Skip the heading in the data file and read the data.
READ ( NIN, FMT = '()' )
READ ( NIN, FMT = * ) N, JOB, FACT, TRANA, UPLO, LYAPUN
IF ( N.LT.0 .OR. N.GT.NMAX ) THEN
WRITE ( NOUT, FMT = 99993 ) N
ELSE
READ ( NIN, FMT = * ) ( ( A(I,J), J = 1,N ), I = 1,N )
READ ( NIN, FMT = * ) ( ( Q(I,J), J = 1,N ), I = 1,N )
READ ( NIN, FMT = * ) ( ( G(I,J), J = 1,N ), I = 1,N )
CALL DLACPY( 'Full', N, N, A, LDA, AS, LDA )
CALL DLACPY( UPLO, N, N, Q, LDQ, X, LDX )
N2 = 2*N
IS = 2*N2 + 1
IU = IS + N2*N2
IW = IU + N2*N2
* Solve the discrete-time Riccati equation.
CALL SB02MD( 'discrete', 'direct', UPLO, 'no scaling',
$ 'stable', N, AS, LDA, G, LDG, X, LDX, RCND,
$ DWORK(1), DWORK(N2+1), DWORK(IS), N2, DWORK(IU),
$ N2, IWORK, DWORK(IW), LDWORK-IW+1, BWORK, INFO1 )
*
IF ( INFO1.EQ.0 ) THEN
WRITE ( NOUT, FMT = 99995 )
DO 10 I = 1, N
WRITE ( NOUT, FMT = 99994 ) ( X(I,J), J = 1,N )
10 CONTINUE
IF ( LSAME( FACT, 'F' ) .OR. LSAME( LYAPUN, 'R' ) ) THEN
CALL DLASET( 'Full', N, N, ZERO, ONE, DWORK, N )
CALL DSYMM( 'Left', UPLO, N, N, ONE, G, LDG, X, LDX,
$ ONE, DWORK, N )
IF ( LSAME( TRANA, 'N' ) ) THEN
* Compute Ac = inv(I_n + G*X)*A.
CALL DLACPY( 'Full', N, N, A, LDA, T, LDT )
CALL DGESV( N, N, DWORK, N, IWORK, T, LDT, INFO3 )
ELSE
* Compute Ac = A*inv(I_n + X*G)
CALL MA02AD( 'Full', N, N, A, LDA, T, LDT )
CALL DGESV( N, N, DWORK, N, IWORK, T, LDT, INFO3 )
DO 20 J = 2, N
CALL DSWAP( J-1, T(1,J), 1, T(J,1), LDT )
20 CONTINUE
END IF
* Compute the Schur factorization of Ac.
JOBS = 'V'
CALL DGEES( JOBS, 'Not ordered', SELECT, N, T, LDT, SDIM,
$ DWORK(1), DWORK(N+1), U, LDU, DWORK(2*N+1),
$ LDWORK-2*N, BWORK, INFO3 )
IF( INFO3.NE.0 ) THEN
WRITE ( NOUT, FMT = 99996 ) INFO3
STOP
END IF
END IF
*
IF ( LSAME( LYAPUN, 'R' ) ) THEN
IF( LSAME( TRANA, 'N' ) ) THEN
TRANAT = 'T'
ELSE
TRANAT = 'N'
END IF
*
CALL MB01RU( UPLO, TRANAT, N, N, ZERO, ONE, X, LDX,
$ U, LDU, X, LDX, DWORK, N*N, INFO2 )
CALL MA02ED( UPLO, N, X, LDX )
CALL MB01RU( UPLO, TRANAT, N, N, ZERO, ONE, G, LDG,
$ U, LDU, G, LDG, DWORK, N*N, INFO2 )
CALL MB01RU( UPLO, TRANAT, N, N, ZERO, ONE, Q, LDQ,
$ U, LDU, Q, LDQ, DWORK, N*N, INFO2 )
END IF
* Estimate the condition and error bound on the solution.
CALL SB02SD( JOB, FACT, TRANA, UPLO, LYAPUN, N, A, LDA, T,
$ LDT, U, LDU, G, LDG, Q, LDQ, X, LDX, SEPD,
$ RCOND, FERR, IWORK, DWORK, LDWORK, INFO2 )
*
IF ( INFO2.NE.0 ) THEN
WRITE ( NOUT, FMT = 99997 ) INFO2
END IF
IF ( INFO2.EQ.0 .OR. INFO2.EQ.N+1 ) THEN
WRITE ( NOUT, FMT = 99992 ) SEPD
WRITE ( NOUT, FMT = 99991 ) RCOND
WRITE ( NOUT, FMT = 99990 ) FERR
END IF
ELSE
WRITE ( NOUT, FMT = 99998 ) INFO1
END IF
END IF
STOP
*
99999 FORMAT (' SB02SD EXAMPLE PROGRAM RESULTS',/1X)
99998 FORMAT (' INFO on exit from SB02MD =',I2)
99997 FORMAT (' INFO on exit from SB02SD =',I2)
99996 FORMAT (' INFO on exit from DGEES =',I2)
99995 FORMAT (' The solution matrix X is')
99994 FORMAT (20(1X,F8.4))
99993 FORMAT (/' N is out of range.',/' N = ',I5)
99992 FORMAT (/' Estimated separation = ',F8.4)
99991 FORMAT (/' Estimated reciprocal condition number = ',F8.4)
99990 FORMAT (/' Estimated error bound = ',F8.4)
END
</PRE>
<B>Program Data</B>
<PRE>
SB02SD EXAMPLE PROGRAM DATA
2 B N N U O
2.0 -1.0
1.0 0.0
0.0 0.0
0.0 1.0
1.0 0.0
0.0 0.0
</PRE>
<B>Program Results</B>
<PRE>
SB02SD EXAMPLE PROGRAM RESULTS
The solution matrix X is
-0.7691 1.2496
1.2496 -2.3306
Estimated separation = 0.4456
Estimated reciprocal condition number = 0.1445
Estimated error bound = 0.0000
</PRE>
<HR>
<p>
<A HREF=..\libindex.html><B>Return to index</B></A></BODY>
</HTML>
|
