NAG Library Routine Document

Integer, Intent (In)	::	m, p, q, ldx11, ldx12, ldx21, ldx22, ldu1, ldu2, ldv1t, ldv2t, lwork, lrwork
Integer, Intent (Out)	::	iwork(m-min(p,m-p,q,m-q)), info
Real (Kind=nag_wp), Intent (Out)	::	theta(min(p,m-p,q,m-q)), rwork(max(1,lrwork))
Complex (Kind=nag_wp), Intent (Inout)	::	x11(ldx11,), x12(ldx12,), x21(ldx21,), x22(ldx22,), u1(ldu1,), u2(ldu2,), v1t(ldv1t,), v2t(ldv2t,)
Complex (Kind=nag_wp), Intent (Out)	::	work(max(1,lwork))
Character (1), Intent (In)	::	jobu1, jobu2, jobv1t, jobv2t, trans, signs

C Header Interface

#include <nagmk26.h>

void

f08rnf_ (const char *jobu1, const char *jobu2, const char *jobv1t, const char *jobv2t, const char *trans, const char *signs, const Integer *m, const Integer *p, const Integer *q, Complex x11[], const Integer *ldx11, Complex x12[], const Integer *ldx12, Complex x21[], const Integer *ldx21, Complex x22[], const Integer *ldx22, double theta[], Complex u1[], const Integer *ldu1, Complex u2[], const Integer *ldu2, Complex v1t[], const Integer *ldv1t, Complex v2t[], const Integer *ldv2t, Complex work[], const Integer *lwork, double rwork[], const Integer *lrwork, Integer iwork[], Integer *info, const Charlen length_jobu1, const Charlen length_jobu2, const Charlen length_jobv1t, const Charlen length_jobv2t, const Charlen length_trans, const Charlen length_signs)

The routine may be called by its LAPACK name zuncsd.

3

Description

The

m

m

unitary matrix

X

is partitioned as

X = (\begin{matrix} X_{11} & X_{12} \\ X_{21} & X_{22} \end{matrix})

where

X_{11}

is a

p

q

submatrix and the dimensions of the other submatrices

X_{12}

X_{21}

and

X_{22}

are such that

X

remains

m

m

The CS decomposition of

X

X = U Σ_{p} V^{T}

where

U

V

and

Σ_{p}

are

m

m

matrices, such that

U = (\begin{matrix} U_{1} & 0 \\ 0 & U_{2} \end{matrix})

is a unitary matrix containing the

p

p

unitary matrix

U_{1}

and the

(m - p)

(m - p)

unitary matrix

U_{2}

;

V = (\begin{matrix} V_{1} & 0 \\ 0 & V_{2} \end{matrix})

is a unitary matrix containing the

q

q

unitary matrix

V_{1}

and the

(m - q)

(m - q)

unitary matrix

V_{2}

; and

Σ_{p} = (\begin{matrix} I_{11} & 0 & 0 & 0 \\ C & 0 & 0 & - S \\ 0 & 0 & 0 & {- I}_{12} \\ 0 & 0 & I_{22} & 0 \\ 0 & S & C & 0 \\ 0 & I_{21} & 0 & 0 \end{matrix})

contains the

r

r

non-negative diagonal submatrices

C

and

S

satisfying

C^{2} + S^{2} = I

, where

r = \min (p, m - p, q, m - q)

and the top left partition is

p

q

The identity matrix

I_{11}

is of order

\min (p, q) - r

and vanishes if

\min (p, q) = r

The identity matrix

I_{12}

is of order

\min (p, m - q) - r

and vanishes if

\min (p, m - q) = r

The identity matrix

I_{21}

is of order

\min (m - p, q) - r

and vanishes if

\min (m - p, q) = r

The identity matrix

I_{22}

is of order

\min (m - p, m - q) - r

and vanishes if

\min (m - p, m - q) = r

In each of the four cases

r = p, q, m - p, m - q

at least two of the identity matrices vanish.

The indicated zeros represent augmentations by additional rows or columns (but not both) to the square diagonal matrices formed by

I_{i j}

and

C

S

Σ_{p}

does not need to be stored in full; it is sufficient to return only the values

θ_{i}

for

i = 1, 2, \dots, r

where

C_{i i} = \cos (θ_{i})

and

S_{i i} = \sin (θ_{i})

The algorithm used to perform the complete

C S

decomposition is described fully in Sutton (2009) including discussions of the stability and accuracy of the algorithm.

4

References

Anderson E, Bai Z, Bischof C, Blackford S, Demmel J, Dongarra J J, Du Croz J J, Greenbaum A, Hammarling S, McKenney A and Sorensen D (1999) LAPACK Users' Guide (3rd Edition) SIAM, Philadelphia http://www.netlib.org/lapack/lug

Golub G H and Van Loan C F (2012) Matrix Computations (4th Edition) Johns Hopkins University Press, Baltimore

Sutton B D (2009) Computing the complete

C S

decomposition Numerical Algorithms (Volume 50) 1017–1398 Springer US 33–65 http://dx.doi.org/10.1007/s11075-008-9215-6

5

Arguments

1: $jobu1$ – Character(1)Input

On entry:

if $jobu1 ='Y'$ , $U_{1}$ is computed;
otherwise, $U_{1}$ is not computed.

2: $jobu2$ – Character(1)Input

On entry:

if $jobu2 ='Y'$ , $U_{2}$ is computed;
otherwise, $U_{2}$ is not computed.

3: $jobv1t$ – Character(1)Input

On entry:

if $jobv1t ='Y'$ , $V_{1}^{T}$ is computed;
otherwise, $V_{1}^{T}$ is not computed.

4: $jobv2t$ – Character(1)Input

On entry:

if $jobv2t ='Y'$ , $V_{2}^{T}$ is computed;
otherwise, $V_{2}^{T}$ is not computed.

5: $trans$ – Character(1)Input

On entry:

if $trans ='T'$ , $X$ , $U_{1}$ , $U_{2}$ , $V_{1}^{T}$ and $V_{2}^{T}$ are stored in row-major order;
otherwise, $X$ , $U_{1}$ , $U_{2}$ , $V_{1}^{T}$ and $V_{2}^{T}$ are stored in column-major order.

6: $signs$ – Character(1)Input

On entry:

if $signs ='O'$ , the lower-left block is made nonpositive (the other convention);
otherwise, the upper-right block is made nonpositive (the default convention).

7: $m$ – IntegerInput

On entry:

m

, the number of rows and columns in the unitary matrix

X

Constraint:

m \geq 0

8: $p$ – IntegerInput

On entry:

p

, the number of rows in

X_{11}

and

X_{12}

Constraint:

0 \leq p \leq m

9: $q$ – IntegerInput

On entry:

q

, the number of columns in

X_{11}

and

X_{21}

Constraint:

0 \leq q \leq m

10: $x11 (ldx11 *)$ – Complex (Kind=nag_wp) arrayInput/Output

Note: the second dimension of the array x11 must be at least

\max (1, p)

trans ='T'

, and at least

q

otherwise.

On entry: the upper left partition of the unitary matrix

X

whose CSD is desired.

On exit: contains details of the unitary matrix used in a simultaneous bidiagonalization process.

11: $ldx11$ – IntegerInput

On entry: the first dimension of the array x11 as declared in the (sub)program from which f08rnf (zuncsd) is called.

Constraints:

if $trans ='T'$ , $ldx11 \geq \max (1, q)$ ;
otherwise $ldx11 \geq \max (1, p)$ .

12: $x12 (ldx12 *)$ – Complex (Kind=nag_wp) arrayInput/Output

Note: the second dimension of the array x12 must be at least

\max (1, p)

trans ='T'

, and at least

m - q

otherwise.

On entry: the upper right partition of the unitary matrix

X

whose CSD is desired.

On exit: contains details of the unitary matrix used in a simultaneous bidiagonalization process.

13: $ldx12$ – IntegerInput

On entry: the first dimension of the array x12 as declared in the (sub)program from which f08rnf (zuncsd) is called.

Constraints:

if $trans ='T'$ , $ldx12 \geq \max (1, m - q)$ ;
otherwise $ldx12 \geq \max (1, p)$ .

14: $x21 (ldx21 *)$ – Complex (Kind=nag_wp) arrayInput/Output

Note: the second dimension of the array x21 must be at least

\max (1, m - p)

trans ='T'

, and at least

q

otherwise.

On entry: the lower left partition of the unitary matrix

X

whose CSD is desired.

On exit: contains details of the unitary matrix used in a simultaneous bidiagonalization process.

15: $ldx21$ – IntegerInput

On entry: the first dimension of the array x21 as declared in the (sub)program from which f08rnf (zuncsd) is called.

Constraints:

if $trans ='T'$ , $ldx21 \geq \max (1, q)$ ;
otherwise $ldx21 \geq \max (1, m - p)$ .

16: $x22 (ldx22 *)$ – Complex (Kind=nag_wp) arrayInput/Output

Note: the second dimension of the array x22 must be at least

\max (1, m - p)

trans ='T'

, and at least

m - q

otherwise.

On entry: the lower right partition of the unitary matrix

X

CSD is desired.

On exit: contains details of the unitary matrix used in a simultaneous bidiagonalization process.

17: $ldx22$ – IntegerInput

On entry: the first dimension of the array x22 as declared in the (sub)program from which f08rnf (zuncsd) is called.

Constraints:

if $trans ='T'$ , $ldx22 \geq \max (1, m - q)$ ;
otherwise $ldx22 \geq \max (1, m - p)$ .

18: $theta (\min (p, m - p, q, m - q))$ – Real (Kind=nag_wp) arrayOutput

On exit: the values

θ_{i}

for

i = 1, 2, \dots, r

where

r = \min (p, m - p, q, m - q)

. The diagonal submatrices

C

and

S

Σ_{p}

are constructed from these values as

$C = diag (\cos (theta (1)), \dots, \cos (theta (r)))$ and
$S = diag (\sin (theta (1)), \dots, \sin (theta (r)))$ .

19: $u1 (ldu1 *)$ – Complex (Kind=nag_wp) arrayOutput

Note: the second dimension of the array u1 must be at least

\max (1, p)

jobu1 ='Y'

, and at least

1

otherwise.

On exit: if

jobu1 ='Y'

, u1 contains the

p

p

unitary matrix

U_{1}

20: $ldu1$ – IntegerInput

On entry: the first dimension of the array u1 as declared in the (sub)program from which f08rnf (zuncsd) is called.

Constraint: if

jobu1 ='Y'

ldu1 \geq \max (1, p)

21: $u2 (ldu2 *)$ – Complex (Kind=nag_wp) arrayOutput

Note: the second dimension of the array u2 must be at least

\max (1, m - p)

jobu2 ='Y'

, and at least

1

otherwise.

On exit: if

jobu2 ='Y'

, u2 contains the

m - p

m - p

unitary matrix

U_{2}

22: $ldu2$ – IntegerInput

On entry: the first dimension of the array u2 as declared in the (sub)program from which f08rnf (zuncsd) is called.

Constraint: if

jobu2 ='Y'

ldu2 \geq \max (1, m - p)

23: $v1t (ldv1t *)$ – Complex (Kind=nag_wp) arrayOutput

Note: the second dimension of the array v1t must be at least

\max (1, q)

jobv1t ='Y'

, and at least

1

otherwise.

On exit: if

jobv1t ='Y'

, v1t contains the

q

q

unitary matrix

{V_{1}}^{H}

24: $ldv1t$ – IntegerInput

On entry: the first dimension of the array v1t as declared in the (sub)program from which f08rnf (zuncsd) is called.

Constraint: if

jobv1t ='Y'

ldv1t \geq \max (1, q)

25: $v2t (ldv2t *)$ – Complex (Kind=nag_wp) arrayOutput

Note: the second dimension of the array v2t must be at least

\max (1, m - q)

jobv2t ='Y'

, and at least

1

otherwise.

On exit: if

jobv2t ='Y'

, v2t contains the

m - q

m - q

unitary matrix

{V_{2}}^{H}

26: $ldv2t$ – IntegerInput

On entry: the first dimension of the array v2t as declared in the (sub)program from which f08rnf (zuncsd) is called.

Constraint: if

jobv2t ='Y'

ldv2t \geq \max (1, m - q)

27: $work (\max (1, lwork))$ – Complex (Kind=nag_wp) arrayWorkspace

On exit: if

info = 0

work (1)

returns the optimal lwork.

info > 0

on exit,

work (2 : r)

contains the values

PHI (1), \dots PHI (r - 1)

that, together with

{xref}^{arg} (1), \dots {xref}^{arg} (r)

, define the matrix in intermediate bidiagonal-block form remaining after nonconvergence. info specifies the number of nonzero PHI's.

28: $lwork$ – IntegerInput

On entry: the dimension of the array work as declared in the (sub)program from which f08rnf (zuncsd) is called.

lwork = - 1

, a workspace query is assumed; the routine only calculates the optimal size of the work array, returns this value as the first entry of the work array, and no error message related to lwork is issued.

The minimum workspace required is

\max (1, p) + \max (1, m - p) + \max (1, q) + \max (1, m - q) + \max (1, p, m - p, q, m - q) + 1

; the optimal amount of workspace depends on internal block sizes and the relative problem dimensions.

Constraint:

lwork = - 1

lwork \geq \max (1, p) + \max (1, m - p) + \max (1, q) + \max (1, m - q) + \max (1, p, m - p, q, m - q) + 1

29: $rwork (\max (1, lrwork))$ – Real (Kind=nag_wp) arrayWorkspace

30: $lrwork$ – IntegerInput

On entry: the dimension of the array rwork as declared in the (sub)program from which f08rnf (zuncsd) is called.

lrwork = - 1

5 \times \max (1, q - 1) + 4 \times \max (1, q) + \max (1, 8 \times q) + 1

which equates to

11

for

q = 0

18

for

q = 1

and

17 \times q - 4

when

q > 1

Constraint:

lrwork = - 1 or lrwork \geq 5 \times \max (1, q - 1) + 4 \times \max (1, q) + \max (1, 8 \times q) + 1

31: $iwork (m - \min (p, m - p, q, m - q))$ – Integer arrayWorkspace

32: $info$ – IntegerOutput

On exit:

info = 0

unless the routine detects an error (see Section 6).

6

Error Indicators and Warnings

$info < 0$: If $info = - i$ , argument $i$ had an illegal value. An explanatory message is output, and execution of the program is terminated.

$info > 0$: The Jacobi-type procedure failed to converge during an internal reduction to bidiagonal-block form. The process requires convergence to $\min (p, m - p, q, m - q)$ values, the value of info gives the number of converged values.

7

Accuracy

The computed

C S

decomposition is nearly the exact

C S

decomposition for the nearby matrix

(X + E)

, where

{‖E‖}_{2} = O (ε),

and

ε

is the machine precision.

8

Parallelism and Performance

f08rnf (zuncsd) is threaded by NAG for parallel execution in multithreaded implementations of the NAG Library.

f08rnf (zuncsd) makes calls to BLAS and/or LAPACK routines, which may be threaded within the vendor library used by this implementation. Consult the documentation for the vendor library for further information.

Please consult the X06 Chapter Introduction for information on how to control and interrogate the OpenMP environment used within this routine. Please also consult the Users' Note for your implementation for any additional implementation-specific information.

9

Further Comments

The total number of floating-point operations required to perform the full

C S

decomposition is approximately

2 m^{3}

The real analogue of this routine is f08raf (dorcsd).

10

Example

This example finds the full CS decomposition of a unitary

6

6

matrix

X

(see Section 10.2) partitioned so that the top left block is

2

4

The decomposition is performed both on submatrices of the unitary matrix

X

and on separated partition matrices. Code is also provided to perform a recombining check if required.

NAG Library Routine Document

f08rnf (zuncsd)

▸▿ Contents

1 Purpose

2 Specification

3 Description

4 References

5 Arguments

6 Error Indicators and Warnings

7 Accuracy

8 Parallelism and Performance

9 Further Comments

10 Example

10.1 Program Text

10.2 Program Data

10.3 Program Results

1

Purpose

2

Specification

3

Description

4

References

5

Arguments

6

Error Indicators and Warnings

7

Accuracy

8

Parallelism and Performance

9

Further Comments

10

Example

10.1

Program Text

10.2

Program Data

10.3

Program Results