CDRVGG
Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd..
November 2006
November 2006
Purpose
CDRVGG checks the nonsymmetric generalized eigenvalue driver
routines.
T T T
CGEGS factors A and B as Q S Z and Q T Z , where means
transpose, T is upper triangular, S is in generalized Schur form
(upper triangular), and Q and Z are unitary. It also
computes the generalized eigenvalues (alpha(1),beta(1)), ...,
(alpha(n),beta(n)), where alpha(j)=S(j,j) and beta(j)=T(j,j) --
thus, w(j) = alpha(j)/beta(j) is a root of the generalized
eigenvalue problem
det( A - w(j) B ) = 0
and m(j) = beta(j)/alpha(j) is a root of the essentially equivalent
problem
det( m(j) A - B ) = 0
CGEGV computes the generalized eigenvalues (alpha(1),beta(1)), ...,
(alpha(n),beta(n)), the matrix L whose columns contain the
generalized left eigenvectors l, and the matrix R whose columns
contain the generalized right eigenvectors r for the pair (A,B).
When CDRVGG is called, a number of matrix "sizes" ("n's") and a
number of matrix "types" are specified. For each size ("n")
and each type of matrix, one matrix will be generated and used
to test the nonsymmetric eigenroutines. For each matrix, 7
tests will be performed and compared with the threshhold THRESH:
Results from CGEGS:
H
(1) | A - Q S Z | / ( |A| n ulp )
H
(2) | B - Q T Z | / ( |B| n ulp )
H
(3) | I - QQ | / ( n ulp )
H
(4) | I - ZZ | / ( n ulp )
(5) maximum over j of D(j) where:
|alpha(j) - S(j,j)| |beta(j) - T(j,j)|
D(j) = ------------------------ + -----------------------
max(|alpha(j)|,|S(j,j)|) max(|beta(j)|,|T(j,j)|)
Results from CGEGV:
(6) max over all left eigenvalue/-vector pairs (beta/alpha,l) of
| l**H * (beta A - alpha B) | / ( ulp max( |beta A|, |alpha B| ) )
where l**H is the conjugate tranpose of l.
(7) max over all right eigenvalue/-vector pairs (beta/alpha,r) of
| (beta A - alpha B) r | / ( ulp max( |beta A|, |alpha B| ) )
Test Matrices
The sizes of the test matrices are specified by an array
NN(1:NSIZES); the value of each element NN(j) specifies one size.
The "types" are specified by a logical array DOTYPE( 1:NTYPES ); if
DOTYPE(j) is .TRUE., then matrix type "j" will be generated.
Currently, the list of possible types is:
(1) ( 0, 0 ) (a pair of zero matrices)
(2) ( I, 0 ) (an identity and a zero matrix)
(3) ( 0, I ) (an identity and a zero matrix)
(4) ( I, I ) (a pair of identity matrices)
t t
(5) ( J , J ) (a pair of transposed Jordan blocks)
t ( I 0 )
(6) ( X, Y ) where X = ( J 0 ) and Y = ( t )
( 0 I ) ( 0 J )
and I is a k x k identity and J a (k+1)x(k+1)
Jordan block; k=(N-1)/2
(7) ( D, I ) where D is diag( 0, 1,..., N-1 ) (a diagonal
matrix with those diagonal entries.)
(8) ( I, D )
(9) ( big*D, small*I ) where "big" is near overflow and small=1/big
(10) ( small*D, big*I )
(11) ( big*I, small*D )
(12) ( small*I, big*D )
(13) ( big*D, big*I )
(14) ( small*D, small*I )
(15) ( D1, D2 ) where D1 is diag( 0, 0, 1, ..., N-3, 0 ) and
D2 is diag( 0, N-3, N-4,..., 1, 0, 0 )
t t
(16) Q ( J , J ) Z where Q and Z are random unitary matrices.
(17) Q ( T1, T2 ) Z where T1 and T2 are upper triangular matrices
with random O(1) entries above the diagonal
and diagonal entries diag(T1) =
( 0, 0, 1, ..., N-3, 0 ) and diag(T2) =
( 0, N-3, N-4,..., 1, 0, 0 )
(18) Q ( T1, T2 ) Z diag(T1) = ( 0, 0, 1, 1, s, ..., s, 0 )
diag(T2) = ( 0, 1, 0, 1,..., 1, 0 )
s = machine precision.
(19) Q ( T1, T2 ) Z diag(T1)=( 0,0,1,1, 1-d, ..., 1-(N-5)*d=s, 0 )
diag(T2) = ( 0, 1, 0, 1, ..., 1, 0 )
N-5
(20) Q ( T1, T2 ) Z diag(T1)=( 0, 0, 1, 1, a, ..., a =s, 0 )
diag(T2) = ( 0, 1, 0, 1, ..., 1, 0, 0 )
(21) Q ( T1, T2 ) Z diag(T1)=( 0, 0, 1, r1, r2, ..., r(N-4), 0 )
diag(T2) = ( 0, 1, 0, 1, ..., 1, 0, 0 )
where r1,..., r(N-4) are random.
(22) Q ( big*T1, small*T2 ) Z diag(T1) = ( 0, 0, 1, ..., N-3, 0 )
diag(T2) = ( 0, 1, ..., 1, 0, 0 )
(23) Q ( small*T1, big*T2 ) Z diag(T1) = ( 0, 0, 1, ..., N-3, 0 )
diag(T2) = ( 0, 1, ..., 1, 0, 0 )
(24) Q ( small*T1, small*T2 ) Z diag(T1) = ( 0, 0, 1, ..., N-3, 0 )
diag(T2) = ( 0, 1, ..., 1, 0, 0 )
(25) Q ( big*T1, big*T2 ) Z diag(T1) = ( 0, 0, 1, ..., N-3, 0 )
diag(T2) = ( 0, 1, ..., 1, 0, 0 )
(26) Q ( T1, T2 ) Z where T1 and T2 are random upper-triangular
matrices.
routines.
T T T
CGEGS factors A and B as Q S Z and Q T Z , where means
transpose, T is upper triangular, S is in generalized Schur form
(upper triangular), and Q and Z are unitary. It also
computes the generalized eigenvalues (alpha(1),beta(1)), ...,
(alpha(n),beta(n)), where alpha(j)=S(j,j) and beta(j)=T(j,j) --
thus, w(j) = alpha(j)/beta(j) is a root of the generalized
eigenvalue problem
det( A - w(j) B ) = 0
and m(j) = beta(j)/alpha(j) is a root of the essentially equivalent
problem
det( m(j) A - B ) = 0
CGEGV computes the generalized eigenvalues (alpha(1),beta(1)), ...,
(alpha(n),beta(n)), the matrix L whose columns contain the
generalized left eigenvectors l, and the matrix R whose columns
contain the generalized right eigenvectors r for the pair (A,B).
When CDRVGG is called, a number of matrix "sizes" ("n's") and a
number of matrix "types" are specified. For each size ("n")
and each type of matrix, one matrix will be generated and used
to test the nonsymmetric eigenroutines. For each matrix, 7
tests will be performed and compared with the threshhold THRESH:
Results from CGEGS:
H
(1) | A - Q S Z | / ( |A| n ulp )
H
(2) | B - Q T Z | / ( |B| n ulp )
H
(3) | I - QQ | / ( n ulp )
H
(4) | I - ZZ | / ( n ulp )
(5) maximum over j of D(j) where:
|alpha(j) - S(j,j)| |beta(j) - T(j,j)|
D(j) = ------------------------ + -----------------------
max(|alpha(j)|,|S(j,j)|) max(|beta(j)|,|T(j,j)|)
Results from CGEGV:
(6) max over all left eigenvalue/-vector pairs (beta/alpha,l) of
| l**H * (beta A - alpha B) | / ( ulp max( |beta A|, |alpha B| ) )
where l**H is the conjugate tranpose of l.
(7) max over all right eigenvalue/-vector pairs (beta/alpha,r) of
| (beta A - alpha B) r | / ( ulp max( |beta A|, |alpha B| ) )
Test Matrices
The sizes of the test matrices are specified by an array
NN(1:NSIZES); the value of each element NN(j) specifies one size.
The "types" are specified by a logical array DOTYPE( 1:NTYPES ); if
DOTYPE(j) is .TRUE., then matrix type "j" will be generated.
Currently, the list of possible types is:
(1) ( 0, 0 ) (a pair of zero matrices)
(2) ( I, 0 ) (an identity and a zero matrix)
(3) ( 0, I ) (an identity and a zero matrix)
(4) ( I, I ) (a pair of identity matrices)
t t
(5) ( J , J ) (a pair of transposed Jordan blocks)
t ( I 0 )
(6) ( X, Y ) where X = ( J 0 ) and Y = ( t )
( 0 I ) ( 0 J )
and I is a k x k identity and J a (k+1)x(k+1)
Jordan block; k=(N-1)/2
(7) ( D, I ) where D is diag( 0, 1,..., N-1 ) (a diagonal
matrix with those diagonal entries.)
(8) ( I, D )
(9) ( big*D, small*I ) where "big" is near overflow and small=1/big
(10) ( small*D, big*I )
(11) ( big*I, small*D )
(12) ( small*I, big*D )
(13) ( big*D, big*I )
(14) ( small*D, small*I )
(15) ( D1, D2 ) where D1 is diag( 0, 0, 1, ..., N-3, 0 ) and
D2 is diag( 0, N-3, N-4,..., 1, 0, 0 )
t t
(16) Q ( J , J ) Z where Q and Z are random unitary matrices.
(17) Q ( T1, T2 ) Z where T1 and T2 are upper triangular matrices
with random O(1) entries above the diagonal
and diagonal entries diag(T1) =
( 0, 0, 1, ..., N-3, 0 ) and diag(T2) =
( 0, N-3, N-4,..., 1, 0, 0 )
(18) Q ( T1, T2 ) Z diag(T1) = ( 0, 0, 1, 1, s, ..., s, 0 )
diag(T2) = ( 0, 1, 0, 1,..., 1, 0 )
s = machine precision.
(19) Q ( T1, T2 ) Z diag(T1)=( 0,0,1,1, 1-d, ..., 1-(N-5)*d=s, 0 )
diag(T2) = ( 0, 1, 0, 1, ..., 1, 0 )
N-5
(20) Q ( T1, T2 ) Z diag(T1)=( 0, 0, 1, 1, a, ..., a =s, 0 )
diag(T2) = ( 0, 1, 0, 1, ..., 1, 0, 0 )
(21) Q ( T1, T2 ) Z diag(T1)=( 0, 0, 1, r1, r2, ..., r(N-4), 0 )
diag(T2) = ( 0, 1, 0, 1, ..., 1, 0, 0 )
where r1,..., r(N-4) are random.
(22) Q ( big*T1, small*T2 ) Z diag(T1) = ( 0, 0, 1, ..., N-3, 0 )
diag(T2) = ( 0, 1, ..., 1, 0, 0 )
(23) Q ( small*T1, big*T2 ) Z diag(T1) = ( 0, 0, 1, ..., N-3, 0 )
diag(T2) = ( 0, 1, ..., 1, 0, 0 )
(24) Q ( small*T1, small*T2 ) Z diag(T1) = ( 0, 0, 1, ..., N-3, 0 )
diag(T2) = ( 0, 1, ..., 1, 0, 0 )
(25) Q ( big*T1, big*T2 ) Z diag(T1) = ( 0, 0, 1, ..., N-3, 0 )
diag(T2) = ( 0, 1, ..., 1, 0, 0 )
(26) Q ( T1, T2 ) Z where T1 and T2 are random upper-triangular
matrices.
Arguments
| NSIZES |
(input) INTEGER
The number of sizes of matrices to use. If it is zero,
CDRVGG does nothing. It must be at least zero. |
| NN |
(input) INTEGER array, dimension (NSIZES)
An array containing the sizes to be used for the matrices.
Zero values will be skipped. The values must be at least zero. |
| NTYPES |
(input) INTEGER
The number of elements in DOTYPE. If it is zero, CDRVGG
does nothing. It must be at least zero. If it is MAXTYP+1 and NSIZES is 1, then an additional type, MAXTYP+1 is defined, which is to use whatever matrix is in A. This is only useful if DOTYPE(1:MAXTYP) is .FALSE. and DOTYPE(MAXTYP+1) is .TRUE. . |
| DOTYPE |
(input) LOGICAL array, dimension (NTYPES)
If DOTYPE(j) is .TRUE., then for each size in NN a
matrix of that size and of type j will be generated. If NTYPES is smaller than the maximum number of types defined (PARAMETER MAXTYP), then types NTYPES+1 through MAXTYP will not be generated. If NTYPES is larger than MAXTYP, DOTYPE(MAXTYP+1) through DOTYPE(NTYPES) will be ignored. |
| ISEED |
(input/output) INTEGER array, dimension (4)
On entry ISEED specifies the seed of the random number
generator. The array elements should be between 0 and 4095; if not they will be reduced mod 4096. Also, ISEED(4) must be odd. The random number generator uses a linear congruential sequence limited to small integers, and so should produce machine independent random numbers. The values of ISEED are changed on exit, and can be used in the next call to CDRVGG to continue the same random number sequence. |
| THRESH |
(input) REAL
A test will count as "failed" if the "error", computed as
described above, exceeds THRESH. Note that the error is scaled to be O(1), so THRESH should be a reasonably small multiple of 1, e.g., 10 or 100. In particular, it should not depend on the precision (single vs. double) or the size of the matrix. It must be at least zero. |
| THRSHN |
(input) REAL
Threshhold for reporting eigenvector normalization error.
If the normalization of any eigenvector differs from 1 by more than THRSHN*ulp, then a special error message will be printed. (This is handled separately from the other tests, since only a compiler or programming error should cause an error message, at least if THRSHN is at least 5--10.) |
| NOUNIT |
(input) INTEGER
The FORTRAN unit number for printing out error messages
(e.g., if a routine returns IINFO not equal to 0.) |
| A |
(input/workspace) COMPLEX array, dimension (LDA, max(NN))
Used to hold the original A matrix. Used as input only
if NTYPES=MAXTYP+1, DOTYPE(1:MAXTYP)=.FALSE., and DOTYPE(MAXTYP+1)=.TRUE. |
| LDA |
(input) INTEGER
The leading dimension of A, B, S, T, S2, and T2.
It must be at least 1 and at least max( NN ). |
| B |
(input/workspace) COMPLEX array, dimension (LDA, max(NN))
Used to hold the original B matrix. Used as input only
if NTYPES=MAXTYP+1, DOTYPE(1:MAXTYP)=.FALSE., and DOTYPE(MAXTYP+1)=.TRUE. |
| S |
(workspace) COMPLEX array, dimension (LDA, max(NN))
The upper triangular matrix computed from A by CGEGS.
|
| T |
(workspace) COMPLEX array, dimension (LDA, max(NN))
The upper triangular matrix computed from B by CGEGS.
|
| S2 |
(workspace) COMPLEX array, dimension (LDA, max(NN))
The matrix computed from A by CGEGV. This will be the
Schur (upper triangular) form of some matrix related to A, but will not, in general, be the same as S. |
| T2 |
(workspace) COMPLEX array, dimension (LDA, max(NN))
The matrix computed from B by CGEGV. This will be the
Schur form of some matrix related to B, but will not, in general, be the same as T. |
| Q |
(workspace) COMPLEX array, dimension (LDQ, max(NN))
The (left) unitary matrix computed by CGEGS.
|
| LDQ |
(input) INTEGER
The leading dimension of Q, Z, VL, and VR. It must
be at least 1 and at least max( NN ). |
| Z |
(workspace) COMPLEX array, dimension (LDQ, max(NN))
The (right) unitary matrix computed by CGEGS.
|
| ALPHA1 |
(workspace) COMPLEX array, dimension (max(NN))
|
| BETA1 |
(workspace) COMPLEX array, dimension (max(NN))
The generalized eigenvalues of (A,B) computed by CGEGS.
ALPHA1(k) / BETA1(k) is the k-th generalized eigenvalue of the matrices in A and B. |
| ALPHA2 |
(workspace) COMPLEX array, dimension (max(NN))
|
| BETA2 |
(workspace) COMPLEX array, dimension (max(NN))
The generalized eigenvalues of (A,B) computed by CGEGV.
ALPHA2(k) / BETA2(k) is the k-th generalized eigenvalue of the matrices in A and B. |
| VL |
(workspace) COMPLEX array, dimension (LDQ, max(NN))
The (lower triangular) left eigenvector matrix for the
matrices in A and B. |
| VR |
(workspace) COMPLEX array, dimension (LDQ, max(NN))
The (upper triangular) right eigenvector matrix for the
matrices in A and B. |
| WORK |
(workspace) COMPLEX array, dimension (LWORK)
|
| LWORK |
(input) INTEGER
The number of entries in WORK. This must be at least
MAX( 2*N, N*(NB+1), (k+1)*(2*k+N+1) ), where "k" is the sum of the blocksize and number-of-shifts for CHGEQZ, and NB is the greatest of the blocksizes for CGEQRF, CUNMQR, and CUNGQR. (The blocksizes and the number-of-shifts are retrieved through calls to ILAENV.) |
| RWORK |
(workspace) REAL array, dimension (8*N)
|
| RESULT |
(output) REAL array, dimension (7)
The values computed by the tests described above.
The values are currently limited to 1/ulp, to avoid overflow. |
| INFO |
(output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value. > 0: A routine returned an error code. INFO is the absolute value of the INFO value returned. |