In this section, we consider the model reduction method based on cross Gramians for square descriptor systems. First, we present the definition of the proper cross Gramian and improper cross Gramian and show that the proper cross Gramian of the square descriptor system in (
1) is the unique solution of a projected generalized continuous-time Sylvester equation (PGCTSE), whereas the improper cross Gramian is the unique solution of a projected generalized discrete-time Sylvester equation (PGDTSE). Then, a balancing transformation for the square descriptor system in (
1) is established by exploiting the proper cross Gramian and improper cross Gramian. Finally, we develop a cross-Gramian-based BFSR model reduction method. Moreover, we also consider the numerical solution of the PGCTSE and PGDTSE and generalize the LR-ADI and LR-Smith methods to these two matrix equations.
3.1. Cross-Gramian-Based Balanced Realization
Cross Gramians [
17,
29] are another kind of Gramian and are defined only for standard square systems. We can extend them to a square descriptor system as follows.
Definition 5. For a stable square descriptor system , its proper cross Gramian is defined asand its improper cross Gramian is defined as It is easy to show that the proper cross Gramian
X is the unique solution of the PGCTSE
whereas the improper cross Gramian
Y is the unique solution of the PGDTSE
Theorem 4. Let (
1)
be a SISO descriptor system or a square descriptor system with for all , where Assume that this system is completely controllable and completely observable. Then, we have Proof. Define
where
,
,
, and
.
Partition
and
appropriately as
We can show that the solution
X of the PGCTSE (
13) can be formulated as
where
is the solution of
Correspondingly, the solutions
and
of the PGCTLEs ((
6) and (
8)) can be expressed as
where
and
are, respectively, the solutions of
Since
is completely controllable and completely observable, it follows from Theorems 1 and 2 that the standard stable system
is controllable and observable. On the one hand, if
is a SISO system, then
is also a SISO system. On the other hand, if
is a system with
for all
, it is easy to verify that
for all
, i.e.,
is a symmetric system. Thus, from [
17], it follows that
, and
can be decomposed into
with
being a real diagonal nonsingular matrix.
It follows from (
15) and (
3) that
By (
17), (
18), and (
3), we obtain
□
Theorem 5. Let (
1)
be a SISO descriptor system or a square descriptor system with for all , where Assume that is completely controllable and completely observable. Then, we have Proof. Define
where
,
,
, and
.
Similarly, it is easy to show that the solution
Y of the PGDTSE (
14) can be formulated as
where
is the solution of
That is,
can be formulated as
We now consider the improper Gramians
and
, which are, respectively, the solutions of these two PGDTLEs ((7) and (9)). We have
where
and
are the solutions of
i.e.,
Since
is a SISO descriptor system or a square descriptor system with
for all
, it follows that for any
,
Therefore, , i.e., . Correspondingly, can be decomposed into , with being a real diagonal nonsingular matrix. □
It is shown that has positive eigenvalues and has positive eigenvalues. The square roots of the largest eigenvalues of , denoted by , are called the proper Hankel singular values of , whereas the square roots of the largest eigenvalues of , denoted by , are called the improper Hankel singular values of .
From (
21), it follows that the largest
eigenvalues of
are also the eigenvalues of
. Thus, from [
17], we have
where
. Similarly, we have
where
.
Partition
and
as
where
and
.
By (
25) and (
26), we obtain
Thus,
defined as in (
27) is a balancing transformation of the descriptor system
.
Assume that the proper cross Gramian
X and the improper cross Gramian
Y are formulated as follows:
Let
and
, respectively, be the Jordan decompositions of
and
. Define
Similarly, we can show that as defined above is also a balancing transformation of the descriptor system .
For more details on the balancing transformation of a descriptor system, the interested reader is referred to [
18].
3.2. Cross-Gramian-Based Model Reduction
The cross-Gramian-based version of the BFSR method for standard state-space systems was proposed by Baur and Benner [
16]. In this subsection, we extend the idea to obtain a cross-Gramian-based BFSR method for square descriptor systems.
The balancing-free square-root method is more stable numerically than the square-root method when the system has poor balance (see, for example, [
20]). The BFSR model reduction method for the descriptor system in (
1) is described in Algorithm 4.
As we know, the reduced system for the standard system generated by the BFSR model reduction method is stable. Note that the transfer function
can be written as
, where
and
are the strictly proper rational part and the polynomial part of
, respectively. Following [
18], we can prove that the reduced transfer function is
, that is, that
and
have the same polynomial part. Then, under the conditions in Theorems 4 and 5, the stability and error bounds
can be proved similarly to [
16].
Algorithm 4The BFSR method based on cross Gramians |
- Input:
. - Output:
.
Compute the low-rank matrix of the proper cross Gramian X. Compute the low-rank factors of the improper cross Gramian Y. Compute the real Schur decomposition
where with . Compute the real Schur decomposition
where with . Compute the real Schur decomposition
Compute the real Schur decomposition
Compute the skinny QR decompositions
where the columns of are orthonormal, and the matrices are nonsingular. Construct the reduced system matrices
|
3.3. Low-Rank Iterative Methods for PGCTSE and PGDTSE
We now study how to compute the low-rank factors of the solutions of the PGCTSE (
13) and the PGDTSE (
14).
By multiplying Equation (
13) by
on the left and right, respectively, we obtain the following projected matrix equation
Following the idea of the ADI iteration, we can produce the iterates
for (
28) by solving the following matrix equations
The real parts of the shift parameters are negative, that is,
belong to
. The initial iterate is
. These two iteration steps can be rewritten into one single iteration step as follows:
We can rewrite the iteration (
29) as
After some simple calculations, we can obtain the expression of the error matrix
Let
X denote the exact solution of (
13).
where
X is the solution of (
13), and
and
By using (
3) and (
4), we obtain
where
From Equation (
31), it is obvious that we should choose the shift parameters
so that
is as small as possible.
From Equations (
31)–(
34), it follows that for the ADI approximate solution
, the following error estimate holds.
Theorem 6. Suppose that J in (
3)
is a diagonal matrix. Then, we have Now, we consider constructing a low-rank ADI iteration. Assume that the iteration
has a low-rank form as
We point out that since the initial iteration
is set to the zero matrix, the previous assumption always holds. By using
, the low-rank ADI iteration step in (
30) can be rewritten as follows:
where
Since both
and
are zero matrices, it follows that
and
are
. So, the rank of the approximation solution
is less than or equal to
. By reversing the order of the ADI parameters
as in [
30], we obtain the following iteration step
where
Summarily, we obtain LR-ADI for solving the PGCTSE (
13), which is described in Algorithm 5.
Algorithm 5LR-ADI for PGCTSE |
- Input:
; the ADI shifts . - Output:
such that is an approximate solution of the PGCTSE ( 13).
Compute , ; Compute , ; For ; ; ; ; End For
|
We note that we reuse these shift parameters circularly, as the number of parameters in Algorithm 5 is less than the number of iterations required to obtain an approximation solution, which has an error below a specified tolerance.
We now consider how to choose the shift parameters. These shift parameters are extremely important to the success of LR-ADI. In Theorem 6, we can see that the spectral radii of two matrices
and
in (
34) determine the rate of convergence of the ADI iteration. Thus, we choose the shift parameters
so that
and
have as small spectral radii as possible. This leads to a generalized minimax problem as follows:
where
denotes a set of finite eigenvalues of the pencil
. In practice, we do not know the exact eigenvalues of the pencil
. Often, it is expensive to compute these eigenvalues. So, we will replace
with a domain, which contains the eigenvalues of
. Since
A is a nonsingular matrix, the minimax problem can be reformulated equivalently as
where
denotes the domains containing the spectra of the matrices
.
In [
19], Stykel extended the idea in [
28] to propose a heuristic algorithm for choosing the ADI parameters. By some simple calculation, we have
Note that the largest eigenvalues of are the reciprocals of the smallest non-zero eigenvalues of . So, in order to obtain the smallest non-zero eigenvalues of , we apply the Arnoldi process to the matrix . We point out that the matrices P, , , can be computed easily from some special block structures of in some applications.
Now, we consider how to solve the PGDTSE (
14) numerically. It is easy to verify that the solution
Y of the PGDTSE (
14) can be expressed as
So, we can rewrite the matrix
Y in the low-rank form as
where
Z and
are given by
From the low-rank expression of
Y, we can propose the following low-rank Smith method for the PGDTSE, which is described in Algorithm 6.
Algorithm 6LR-Smith for PGDTSE |
- Input:
. - Output:
low-rank factors Z and of the solution Y of the PGDTSE ( 14).
Compute Compute ; ; For ; ; ; ; End For
|
We point out that the main cost of balanced truncation methods is in solving matrix equations. For the cross-Gramian-based method and the Gramian-based method, two systems of linear equations with coefficient matrices or A are required to be solved in every iteration step of the low-rank ADI and the Smith methods. In the numerical test, the low-rank ADI and the Smith methods have the same number of iterations. So, the cross-Gramian-based approach has approximately the same computational complexity as the Gramian-based BFSR method.