LAPACK 3.3.0
|
00001 SUBROUTINE SLARRR( N, D, E, INFO ) 00002 * 00003 * -- LAPACK auxiliary routine (version 3.2.2) -- 00004 * -- LAPACK is a software package provided by Univ. of Tennessee, -- 00005 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- 00006 * June 2010 00007 * 00008 * .. Scalar Arguments .. 00009 INTEGER N, INFO 00010 * .. 00011 * .. Array Arguments .. 00012 REAL D( * ), E( * ) 00013 * .. 00014 * 00015 * 00016 * Purpose 00017 * ======= 00018 * 00019 * Perform tests to decide whether the symmetric tridiagonal matrix T 00020 * warrants expensive computations which guarantee high relative accuracy 00021 * in the eigenvalues. 00022 * 00023 * Arguments 00024 * ========= 00025 * 00026 * N (input) INTEGER 00027 * The order of the matrix. N > 0. 00028 * 00029 * D (input) REAL array, dimension (N) 00030 * The N diagonal elements of the tridiagonal matrix T. 00031 * 00032 * E (input/output) REAL array, dimension (N) 00033 * On entry, the first (N-1) entries contain the subdiagonal 00034 * elements of the tridiagonal matrix T; E(N) is set to ZERO. 00035 * 00036 * INFO (output) INTEGER 00037 * INFO = 0(default) : the matrix warrants computations preserving 00038 * relative accuracy. 00039 * INFO = 1 : the matrix warrants computations guaranteeing 00040 * only absolute accuracy. 00041 * 00042 * Further Details 00043 * =============== 00044 * 00045 * Based on contributions by 00046 * Beresford Parlett, University of California, Berkeley, USA 00047 * Jim Demmel, University of California, Berkeley, USA 00048 * Inderjit Dhillon, University of Texas, Austin, USA 00049 * Osni Marques, LBNL/NERSC, USA 00050 * Christof Voemel, University of California, Berkeley, USA 00051 * 00052 * ===================================================================== 00053 * 00054 * .. Parameters .. 00055 REAL ZERO, RELCOND 00056 PARAMETER ( ZERO = 0.0E0, 00057 $ RELCOND = 0.999E0 ) 00058 * .. 00059 * .. Local Scalars .. 00060 INTEGER I 00061 LOGICAL YESREL 00062 REAL EPS, SAFMIN, SMLNUM, RMIN, TMP, TMP2, 00063 $ OFFDIG, OFFDIG2 00064 00065 * .. 00066 * .. External Functions .. 00067 REAL SLAMCH 00068 EXTERNAL SLAMCH 00069 * .. 00070 * .. Intrinsic Functions .. 00071 INTRINSIC ABS 00072 * .. 00073 * .. Executable Statements .. 00074 * 00075 * As a default, do NOT go for relative-accuracy preserving computations. 00076 INFO = 1 00077 00078 SAFMIN = SLAMCH( 'Safe minimum' ) 00079 EPS = SLAMCH( 'Precision' ) 00080 SMLNUM = SAFMIN / EPS 00081 RMIN = SQRT( SMLNUM ) 00082 00083 * Tests for relative accuracy 00084 * 00085 * Test for scaled diagonal dominance 00086 * Scale the diagonal entries to one and check whether the sum of the 00087 * off-diagonals is less than one 00088 * 00089 * The sdd relative error bounds have a 1/(1- 2*x) factor in them, 00090 * x = max(OFFDIG + OFFDIG2), so when x is close to 1/2, no relative 00091 * accuracy is promised. In the notation of the code fragment below, 00092 * 1/(1 - (OFFDIG + OFFDIG2)) is the condition number. 00093 * We don't think it is worth going into "sdd mode" unless the relative 00094 * condition number is reasonable, not 1/macheps. 00095 * The threshold should be compatible with other thresholds used in the 00096 * code. We set OFFDIG + OFFDIG2 <= .999 =: RELCOND, it corresponds 00097 * to losing at most 3 decimal digits: 1 / (1 - (OFFDIG + OFFDIG2)) <= 1000 00098 * instead of the current OFFDIG + OFFDIG2 < 1 00099 * 00100 YESREL = .TRUE. 00101 OFFDIG = ZERO 00102 TMP = SQRT(ABS(D(1))) 00103 IF (TMP.LT.RMIN) YESREL = .FALSE. 00104 IF(.NOT.YESREL) GOTO 11 00105 DO 10 I = 2, N 00106 TMP2 = SQRT(ABS(D(I))) 00107 IF (TMP2.LT.RMIN) YESREL = .FALSE. 00108 IF(.NOT.YESREL) GOTO 11 00109 OFFDIG2 = ABS(E(I-1))/(TMP*TMP2) 00110 IF(OFFDIG+OFFDIG2.GE.RELCOND) YESREL = .FALSE. 00111 IF(.NOT.YESREL) GOTO 11 00112 TMP = TMP2 00113 OFFDIG = OFFDIG2 00114 10 CONTINUE 00115 11 CONTINUE 00116 00117 IF( YESREL ) THEN 00118 INFO = 0 00119 RETURN 00120 ELSE 00121 ENDIF 00122 * 00123 00124 * 00125 * *** MORE TO BE IMPLEMENTED *** 00126 * 00127 00128 * 00129 * Test if the lower bidiagonal matrix L from T = L D L^T 00130 * (zero shift facto) is well conditioned 00131 * 00132 00133 * 00134 * Test if the upper bidiagonal matrix U from T = U D U^T 00135 * (zero shift facto) is well conditioned. 00136 * In this case, the matrix needs to be flipped and, at the end 00137 * of the eigenvector computation, the flip needs to be applied 00138 * to the computed eigenvectors (and the support) 00139 * 00140 00141 * 00142 RETURN 00143 * 00144 * END OF SLARRR 00145 * 00146 END