1       SUBROUTINE DTBRFS( UPLO, TRANS, DIAG, N, KD, NRHS, AB, LDAB, B,
  2      $                   LDB, X, LDX, FERR, BERR, WORK, IWORK, INFO )
  3 *
  4 *  -- LAPACK routine (version 3.3.1) --
  5 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
  6 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
  7 *  -- April 2011                                                      --
  8 *
  9 *     Modified to call DLACN2 in place of DLACON, 5 Feb 03, SJH.
 10 *
 11 *     .. Scalar Arguments ..
 12       CHARACTER          DIAG, TRANS, UPLO
 13       INTEGER            INFO, KD, LDAB, LDB, LDX, N, NRHS
 14 *     ..
 15 *     .. Array Arguments ..
 16       INTEGER            IWORK( * )
 17       DOUBLE PRECISION   AB( LDAB, * ), B( LDB, * ), BERR( * ),
 18      $                   FERR( * ), WORK( * ), X( LDX, * )
 19 *     ..
 20 *
 21 *  Purpose
 22 *  =======
 23 *
 24 *  DTBRFS provides error bounds and backward error estimates for the
 25 *  solution to a system of linear equations with a triangular band
 26 *  coefficient matrix.
 27 *
 28 *  The solution matrix X must be computed by DTBTRS or some other
 29 *  means before entering this routine.  DTBRFS does not do iterative
 30 *  refinement because doing so cannot improve the backward error.
 31 *
 32 *  Arguments
 33 *  =========
 34 *
 35 *  UPLO    (input) CHARACTER*1
 36 *          = 'U':  A is upper triangular;
 37 *          = 'L':  A is lower triangular.
 38 *
 39 *  TRANS   (input) CHARACTER*1
 40 *          Specifies the form of the system of equations:
 41 *          = 'N':  A * X = B  (No transpose)
 42 *          = 'T':  A**T * X = B  (Transpose)
 43 *          = 'C':  A**H * X = B  (Conjugate transpose = Transpose)
 44 *
 45 *  DIAG    (input) CHARACTER*1
 46 *          = 'N':  A is non-unit triangular;
 47 *          = 'U':  A is unit triangular.
 48 *
 49 *  N       (input) INTEGER
 50 *          The order of the matrix A.  N >= 0.
 51 *
 52 *  KD      (input) INTEGER
 53 *          The number of superdiagonals or subdiagonals of the
 54 *          triangular band matrix A.  KD >= 0.
 55 *
 56 *  NRHS    (input) INTEGER
 57 *          The number of right hand sides, i.e., the number of columns
 58 *          of the matrices B and X.  NRHS >= 0.
 59 *
 60 *  AB      (input) DOUBLE PRECISION array, dimension (LDAB,N)
 61 *          The upper or lower triangular band matrix A, stored in the
 62 *          first kd+1 rows of the array. The j-th column of A is stored
 63 *          in the j-th column of the array AB as follows:
 64 *          if UPLO = 'U', AB(kd+1+i-j,j) = A(i,j) for max(1,j-kd)<=i<=j;
 65 *          if UPLO = 'L', AB(1+i-j,j)    = A(i,j) for j<=i<=min(n,j+kd).
 66 *          If DIAG = 'U', the diagonal elements of A are not referenced
 67 *          and are assumed to be 1.
 68 *
 69 *  LDAB    (input) INTEGER
 70 *          The leading dimension of the array AB.  LDAB >= KD+1.
 71 *
 72 *  B       (input) DOUBLE PRECISION array, dimension (LDB,NRHS)
 73 *          The right hand side matrix B.
 74 *
 75 *  LDB     (input) INTEGER
 76 *          The leading dimension of the array B.  LDB >= max(1,N).
 77 *
 78 *  X       (input) DOUBLE PRECISION array, dimension (LDX,NRHS)
 79 *          The solution matrix X.
 80 *
 81 *  LDX     (input) INTEGER
 82 *          The leading dimension of the array X.  LDX >= max(1,N).
 83 *
 84 *  FERR    (output) DOUBLE PRECISION array, dimension (NRHS)
 85 *          The estimated forward error bound for each solution vector
 86 *          X(j) (the j-th column of the solution matrix X).
 87 *          If XTRUE is the true solution corresponding to X(j), FERR(j)
 88 *          is an estimated upper bound for the magnitude of the largest
 89 *          element in (X(j) - XTRUE) divided by the magnitude of the
 90 *          largest element in X(j).  The estimate is as reliable as
 91 *          the estimate for RCOND, and is almost always a slight
 92 *          overestimate of the true error.
 93 *
 94 *  BERR    (output) DOUBLE PRECISION array, dimension (NRHS)
 95 *          The componentwise relative backward error of each solution
 96 *          vector X(j) (i.e., the smallest relative change in
 97 *          any element of A or B that makes X(j) an exact solution).
 98 *
 99 *  WORK    (workspace) DOUBLE PRECISION array, dimension (3*N)
100 *
101 *  IWORK   (workspace) INTEGER array, dimension (N)
102 *
103 *  INFO    (output) INTEGER
104 *          = 0:  successful exit
105 *          < 0:  if INFO = -i, the i-th argument had an illegal value
106 *
107 *  =====================================================================
108 *
109 *     .. Parameters ..
110       DOUBLE PRECISION   ZERO
111       PARAMETER          ( ZERO = 0.0D+0 )
112       DOUBLE PRECISION   ONE
113       PARAMETER          ( ONE = 1.0D+0 )
114 *     ..
115 *     .. Local Scalars ..
116       LOGICAL            NOTRAN, NOUNIT, UPPER
117       CHARACTER          TRANST
118       INTEGER            I, J, K, KASE, NZ
119       DOUBLE PRECISION   EPS, LSTRES, S, SAFE1, SAFE2, SAFMIN, XK
120 *     ..
121 *     .. Local Arrays ..
122       INTEGER            ISAVE( 3 )
123 *     ..
124 *     .. External Subroutines ..
125       EXTERNAL           DAXPY, DCOPY, DLACN2, DTBMV, DTBSV, XERBLA
126 *     ..
127 *     .. Intrinsic Functions ..
128       INTRINSIC          ABSMAXMIN
129 *     ..
130 *     .. External Functions ..
131       LOGICAL            LSAME
132       DOUBLE PRECISION   DLAMCH
133       EXTERNAL           LSAME, DLAMCH
134 *     ..
135 *     .. Executable Statements ..
136 *
137 *     Test the input parameters.
138 *
139       INFO = 0
140       UPPER = LSAME( UPLO, 'U' )
141       NOTRAN = LSAME( TRANS, 'N' )
142       NOUNIT = LSAME( DIAG, 'N' )
143 *
144       IF.NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
145          INFO = -1
146       ELSE IF.NOT.NOTRAN .AND. .NOT.LSAME( TRANS, 'T' ) .AND. .NOT.
147      $         LSAME( TRANS, 'C' ) ) THEN
148          INFO = -2
149       ELSE IF.NOT.NOUNIT .AND. .NOT.LSAME( DIAG, 'U' ) ) THEN
150          INFO = -3
151       ELSE IF( N.LT.0 ) THEN
152          INFO = -4
153       ELSE IF( KD.LT.0 ) THEN
154          INFO = -5
155       ELSE IF( NRHS.LT.0 ) THEN
156          INFO = -6
157       ELSE IF( LDAB.LT.KD+1 ) THEN
158          INFO = -8
159       ELSE IF( LDB.LT.MAX1, N ) ) THEN
160          INFO = -10
161       ELSE IF( LDX.LT.MAX1, N ) ) THEN
162          INFO = -12
163       END IF
164       IF( INFO.NE.0 ) THEN
165          CALL XERBLA( 'DTBRFS'-INFO )
166          RETURN
167       END IF
168 *
169 *     Quick return if possible
170 *
171       IF( N.EQ.0 .OR. NRHS.EQ.0 ) THEN
172          DO 10 J = 1, NRHS
173             FERR( J ) = ZERO
174             BERR( J ) = ZERO
175    10    CONTINUE
176          RETURN
177       END IF
178 *
179       IF( NOTRAN ) THEN
180          TRANST = 'T'
181       ELSE
182          TRANST = 'N'
183       END IF
184 *
185 *     NZ = maximum number of nonzero elements in each row of A, plus 1
186 *
187       NZ = KD + 2
188       EPS = DLAMCH( 'Epsilon' )
189       SAFMIN = DLAMCH( 'Safe minimum' )
190       SAFE1 = NZ*SAFMIN
191       SAFE2 = SAFE1 / EPS
192 *
193 *     Do for each right hand side
194 *
195       DO 250 J = 1, NRHS
196 *
197 *        Compute residual R = B - op(A) * X,
198 *        where op(A) = A or A**T, depending on TRANS.
199 *
200          CALL DCOPY( N, X( 1, J ), 1, WORK( N+1 ), 1 )
201          CALL DTBMV( UPLO, TRANS, DIAG, N, KD, AB, LDAB, WORK( N+1 ),
202      $               1 )
203          CALL DAXPY( N, -ONE, B( 1, J ), 1, WORK( N+1 ), 1 )
204 *
205 *        Compute componentwise relative backward error from formula
206 *
207 *        max(i) ( abs(R(i)) / ( abs(op(A))*abs(X) + abs(B) )(i) )
208 *
209 *        where abs(Z) is the componentwise absolute value of the matrix
210 *        or vector Z.  If the i-th component of the denominator is less
211 *        than SAFE2, then SAFE1 is added to the i-th components of the
212 *        numerator and denominator before dividing.
213 *
214          DO 20 I = 1, N
215             WORK( I ) = ABS( B( I, J ) )
216    20    CONTINUE
217 *
218          IF( NOTRAN ) THEN
219 *
220 *           Compute abs(A)*abs(X) + abs(B).
221 *
222             IF( UPPER ) THEN
223                IF( NOUNIT ) THEN
224                   DO 40 K = 1, N
225                      XK = ABS( X( K, J ) )
226                      DO 30 I = MAX1, K-KD ), K
227                         WORK( I ) = WORK( I ) +
228      $                              ABS( AB( KD+1+I-K, K ) )*XK
229    30                CONTINUE
230    40             CONTINUE
231                ELSE
232                   DO 60 K = 1, N
233                      XK = ABS( X( K, J ) )
234                      DO 50 I = MAX1, K-KD ), K - 1
235                         WORK( I ) = WORK( I ) +
236      $                              ABS( AB( KD+1+I-K, K ) )*XK
237    50                CONTINUE
238                      WORK( K ) = WORK( K ) + XK
239    60             CONTINUE
240                END IF
241             ELSE
242                IF( NOUNIT ) THEN
243                   DO 80 K = 1, N
244                      XK = ABS( X( K, J ) )
245                      DO 70 I = K, MIN( N, K+KD )
246                         WORK( I ) = WORK( I ) + ABS( AB( 1+I-K, K ) )*XK
247    70                CONTINUE
248    80             CONTINUE
249                ELSE
250                   DO 100 K = 1, N
251                      XK = ABS( X( K, J ) )
252                      DO 90 I = K + 1MIN( N, K+KD )
253                         WORK( I ) = WORK( I ) + ABS( AB( 1+I-K, K ) )*XK
254    90                CONTINUE
255                      WORK( K ) = WORK( K ) + XK
256   100             CONTINUE
257                END IF
258             END IF
259          ELSE
260 *
261 *           Compute abs(A**T)*abs(X) + abs(B).
262 *
263             IF( UPPER ) THEN
264                IF( NOUNIT ) THEN
265                   DO 120 K = 1, N
266                      S = ZERO
267                      DO 110 I = MAX1, K-KD ), K
268                         S = S + ABS( AB( KD+1+I-K, K ) )*
269      $                      ABS( X( I, J ) )
270   110                CONTINUE
271                      WORK( K ) = WORK( K ) + S
272   120             CONTINUE
273                ELSE
274                   DO 140 K = 1, N
275                      S = ABS( X( K, J ) )
276                      DO 130 I = MAX1, K-KD ), K - 1
277                         S = S + ABS( AB( KD+1+I-K, K ) )*
278      $                      ABS( X( I, J ) )
279   130                CONTINUE
280                      WORK( K ) = WORK( K ) + S
281   140             CONTINUE
282                END IF
283             ELSE
284                IF( NOUNIT ) THEN
285                   DO 160 K = 1, N
286                      S = ZERO
287                      DO 150 I = K, MIN( N, K+KD )
288                         S = S + ABS( AB( 1+I-K, K ) )*ABS( X( I, J ) )
289   150                CONTINUE
290                      WORK( K ) = WORK( K ) + S
291   160             CONTINUE
292                ELSE
293                   DO 180 K = 1, N
294                      S = ABS( X( K, J ) )
295                      DO 170 I = K + 1MIN( N, K+KD )
296                         S = S + ABS( AB( 1+I-K, K ) )*ABS( X( I, J ) )
297   170                CONTINUE
298                      WORK( K ) = WORK( K ) + S
299   180             CONTINUE
300                END IF
301             END IF
302          END IF
303          S = ZERO
304          DO 190 I = 1, N
305             IF( WORK( I ).GT.SAFE2 ) THEN
306                S = MAX( S, ABS( WORK( N+I ) ) / WORK( I ) )
307             ELSE
308                S = MAX( S, ( ABS( WORK( N+I ) )+SAFE1 ) /
309      $             ( WORK( I )+SAFE1 ) )
310             END IF
311   190    CONTINUE
312          BERR( J ) = S
313 *
314 *        Bound error from formula
315 *
316 *        norm(X - XTRUE) / norm(X) .le. FERR =
317 *        norm( abs(inv(op(A)))*
318 *           ( abs(R) + NZ*EPS*( abs(op(A))*abs(X)+abs(B) ))) / norm(X)
319 *
320 *        where
321 *          norm(Z) is the magnitude of the largest component of Z
322 *          inv(op(A)) is the inverse of op(A)
323 *          abs(Z) is the componentwise absolute value of the matrix or
324 *             vector Z
325 *          NZ is the maximum number of nonzeros in any row of A, plus 1
326 *          EPS is machine epsilon
327 *
328 *        The i-th component of abs(R)+NZ*EPS*(abs(op(A))*abs(X)+abs(B))
329 *        is incremented by SAFE1 if the i-th component of
330 *        abs(op(A))*abs(X) + abs(B) is less than SAFE2.
331 *
332 *        Use DLACN2 to estimate the infinity-norm of the matrix
333 *           inv(op(A)) * diag(W),
334 *        where W = abs(R) + NZ*EPS*( abs(op(A))*abs(X)+abs(B) )))
335 *
336          DO 200 I = 1, N
337             IF( WORK( I ).GT.SAFE2 ) THEN
338                WORK( I ) = ABS( WORK( N+I ) ) + NZ*EPS*WORK( I )
339             ELSE
340                WORK( I ) = ABS( WORK( N+I ) ) + NZ*EPS*WORK( I ) + SAFE1
341             END IF
342   200    CONTINUE
343 *
344          KASE = 0
345   210    CONTINUE
346          CALL DLACN2( N, WORK( 2*N+1 ), WORK( N+1 ), IWORK, FERR( J ),
347      $                KASE, ISAVE )
348          IF( KASE.NE.0 ) THEN
349             IF( KASE.EQ.1 ) THEN
350 *
351 *              Multiply by diag(W)*inv(op(A)**T).
352 *
353                CALL DTBSV( UPLO, TRANST, DIAG, N, KD, AB, LDAB,
354      $                     WORK( N+1 ), 1 )
355                DO 220 I = 1, N
356                   WORK( N+I ) = WORK( I )*WORK( N+I )
357   220          CONTINUE
358             ELSE
359 *
360 *              Multiply by inv(op(A))*diag(W).
361 *
362                DO 230 I = 1, N
363                   WORK( N+I ) = WORK( I )*WORK( N+I )
364   230          CONTINUE
365                CALL DTBSV( UPLO, TRANS, DIAG, N, KD, AB, LDAB,
366      $                     WORK( N+1 ), 1 )
367             END IF
368             GO TO 210
369          END IF
370 *
371 *        Normalize error.
372 *
373          LSTRES = ZERO
374          DO 240 I = 1, N
375             LSTRES = MAX( LSTRES, ABS( X( I, J ) ) )
376   240    CONTINUE
377          IF( LSTRES.NE.ZERO )
378      $      FERR( J ) = FERR( J ) / LSTRES
379 *
380   250 CONTINUE
381 *
382       RETURN
383 *
384 *     End of DTBRFS
385 *
386       END