You are on page 1of 8

Systems & Control Letters 67 (2014) 32–39

Contents lists available at ScienceDirect

Systems & Control Letters


journal homepage: www.elsevier.com/locate/sysconle

Model reduction using a frequency-limited H2 -cost


Daniel Petersson ∗ , Johan Löfberg
Division of Automatic Control, Department of Electrical Engineering, Linköpings Universitet, SE-581 83, Sweden

article info abstract


Article history: We propose a method for model reduction on a given frequency range, without the use of input and output
Received 7 December 2012 filter weights. The method uses a nonlinear optimization approach to minimize a frequency limited H2
Received in revised form like cost function. An important contribution of the paper is the derivation of the gradient of the proposed
17 January 2014
cost function. The fact that we have a closed form expression for the gradient and that considerations have
Accepted 10 February 2014
Available online 12 March 2014
been taken to make the gradient computationally efficient to compute enables us to efficiently use off-
the-shelf optimization software to solve the optimization problem.
Keywords:
© 2014 Elsevier B.V. All rights reserved.
Model reduction
H2 -norm
Optimization

1. Introduction approaches. In many of the frequency-weighted methods one has


to specify input and output filter weights. In [7] they introduce
Given a linear time-invariant (lti) dynamical model, a method which does not need these weighting functions, by in-
troducing frequency-limited Gramians. This method can be inter-
ẋ(t ) = Ax(t ) + Bu(t ), preted as using ideal low-, band- or high-pass filters as weights.
y(t ) = Cx(t ) + Du(t ) However, this method has the drawback of not always producing
stable models. One approach to remedy this has been presented
where A ∈ Rn×n , B ∈ Rn×m , C ∈ Rp×n and D ∈ Rp×m , the model
in [2], where they introduce a modification of the method in [7],
reduction problem is to find a reduced order model
and additionally derive an H∞ bound for the error. Sahlan et al. [8]
x̂˙ (t ) = Âx̂(t ) + B̂u(t ), also presents a modification of the method from [7], however this
method is only applicable to siso models. Note that these methods,
ŷ(t ) = Ĉx̂(t ) + D̂u(t ), whilst having an upper bound on the H∞ error, do not minimize
an explicit measure.
with  ∈ Rn̂×n̂ , B̂ ∈ Rn̂×m , Ĉ ∈ Rp×n̂ and D̂ ∈ Rp×m with n̂ < n,
In this paper, a method that uses a variant of the H2 -norm as the
where this reduced order model describes the original model well
measure of performance will be presented. The problem of find-
in some metric. In this paper, we are interested in a reduced order
ing a reduced order model that, in H2 sense, resembles the origi-
model that describes the model well on a given frequency range.
nal model well has been a goal in many investigations. Especially
This is motivated by situations where the given model is valid only
since the work of Meier and Luenberger [9], and especially Wil-
for a certain frequency range, for example as in [1] where models
son [10], in which they derive first order optimality conditions for
coming from aerodynamical and structural mechanics computa-
minimization of the H2 -norm, see for example [11–15] and refer-
tions describing a flexible structure are only valid up to a certain
ences therein. One reason for this could be the fact that H2 criterion
frequency. provides a meaningful characterization of the error, both in deter-
For a review of balanced truncation model reduction ap- ministic and stochastic contexts.
proaches, both ordinary and frequency-weighted, see e.g. [2,3]. Finding global minimizers for the H2 -approximation problem
Some of the most commonly used frequency-weighted methods, is very difficult, it is in fact a non-linear non-convex optimiza-
according to Gugercin and Antoulas [2], are Wang et al. [4], Enns [5] tion problem, the existing methods for H2 -approximation have the
and Lin and Chiu [6], which all use different balanced truncation more modest goal of finding local minimizers and can be catego-
rized into two categories; methods using tangential interpolation
techniques or methods using gradient-flow techniques.
∗ Corresponding author. Tel.: +46 13281311. The gradient-flow algorithms use the gradients for the state
E-mail addresses: petersson@isy.liu.se (D. Petersson), johanl@isy.liu.se space matrices derived in [16] and let these evolve in time to find a
(J. Löfberg). local approximation of the given system. The different algorithms
http://dx.doi.org/10.1016/j.sysconle.2014.02.004
0167-6911/© 2014 Elsevier B.V. All rights reserved.
D. Petersson, J. Löfberg / Systems & Control Letters 67 (2014) 32–39 33

use different techniques to assure that the reduced model is stable, with
to speed up the process and to guarantee convergence. Examples
i
ln (A + iωI)(A − iωI)−1 .
 
of these algorithms are Yan and Lam [15], Fulcheri and Olivi [14] Sω = (6)
and Huang et al. [17]. 2π
The interpolation based H2 model reduction techniques try In [7] they continue by creating a balanced system and per-
to find a model whose transfer function interpolates the transfer forming a balanced truncation using the newly defined frequency-
function of the full-order system (and its derivative) at selected in- limited Gramians. A drawback of this method is that since the
terpolation points. These methods often use computationally ef- terms Sω BBT + BBT S∗ω and S∗ω CT C + CT CSω are not guaranteed to be
fective Krylov-based algorithms which makes these techniques positive semi-definite, stability of the reduced order model cannot
suitable for large-scale problems. Examples of these algorithms be guaranteed. There exists a modification to this in [2] where they
are Xu and Zeng [18], Beattie and Gugercin [12] and Poussot- propose a remedy to this.
Vassal [19]. Recently two variants of these methods have been
published. The first one is a frequency-weighted version that is
Remark 1. By using addition/subtraction of two or more different
applicable on siso systems, see [20], and the second one is a
frequency-limited Gramians it is possible to focus on one or more
frequency-limited version with the same goal as the method pre-
arbitrary frequency ranges, e.g., you can construct the frequency-
sented in this article, but using a different technique, see [21].
limited controllability Gramian, PΩ , for the interval ω ∈ Ω =
One of the main contributions in this paper is a method which
[ω1 , ω2 ] ∪ [ω3 , ω4 ] as
uses optimization, and not truncation, to find an H2 -optimal re-
duced order model, and, where this optimization is only performed APΩ + PΩ AT + SΩ BBT + BBT S∗Ω = 0, (7)
over a limited frequency interval. Another important contribution
is the derivation of the gradient for the cost function. The fact that with SΩ = Sω2 − Sω1 + Sω4 − Sω3 .
we have a closed form expression for the gradient and that con-
siderations have been taken to make the gradient computationally
3. Frequency limited model reduction using optimization
efficient to compute enables us to efficiently use off-the-shelf op-
timization software to solve the optimization problem.
The H2 -norm of G, in (1), assuming that D = 0 so that the
system becomes strictly proper, can be expressed as
2. Frequency limited Gramians
 ∞
CeAτ BBT eA τ CT dτ
T
2
The method proposed in this paper uses the idea presented ∥G∥H2 = tr (8a)
0
in [7], in which they introduce frequency-limited Gramians. Before  ∞
presenting the frequency-limited Gramians we define the standard 1
= tr G(iν)G∗ (iν)dν (8b)
Gramians, see, e.g., [22], in time and frequency domain, for 2π −∞
reference. For a system G, that is stable and described by  ∞
1
= tr CH(ν)BBT H∗ (ν)CT dν = tr CPCT (8c)
ẋ(t ) = Ax(t ) + Bu(t )

G: (1) 2π −∞
y(t ) = Cx(t ) + Du(t )  ∞
BT eA τ CT CeAτ Bdτ
T
  = tr (8d)
A B 0
and denoted as G : , the observability and
C D  ∞
1
controllability Gramians are defined as = tr BT H∗ (ν)CT CH(ν)Bdν = tr BT QB. (8e)
2π −∞
 ∞  ∞
1
eAτ BBT eA τ dτ =
T
P= H(ν)BBT H∗ (ν)dν, (2a) In this paper we introduce a new frequency-limited H2 -like
0 2π −∞ norm that uses the frequency-limited Gramians presented in the
 ∞  ∞
previous section, and we denote the new measure by ∥G∥H2 ,ω , with
1
eA τ CT CeAτ dτ =
T
Q= H∗ (ν)CT CH(ν)dν, (2b)
0 2π −∞  ω
1
where H(ν) = (iν I − A)−1 and H∗ (ν) denotes the conjugate trans-
∥G∥2H2 ,ω = tr G(iν)G∗ (iν)dν (9)
2π −ω
pose of H(ν). The controllability and observability Gramians sat-  ω
isfy, respectively, the Lyapunov equations 1
(CH(ν)B + D) BT H∗ (ν)CT + DT dν
 
= tr (10)
2π −ω
AP + PAT + BBT = 0, (3a)  ω  T
AT Q + QA + CT C = 0. (3b) = tr CPω CT + 2 tr CSω B + D D . (11)

Now we narrow the frequency band, from (−∞, ∞) to (−ω, ω) One thing that differs from the ordinary H2 -norm is that, if we do
where ω < ∞. We define the frequency-limited Gramians, see [7], not include an infinite interval in Ω , i.e., include ω = ∞ as the end
as frequency, then the system does not need to be strictly proper. This
 ω means that we can, in this case, have D ̸= 0.
1
Pω = H(ν)BBT H∗ (ν)dν, (4a) The method proposed in this paper is a model reduction method
2π −ω
 ω that, given a model G, finds a reduced order model, Ĝ, that is a good
1 approximation on a given frequency interval, e.g. [0, ω]. The ob-
Qω = H∗ (ν)CT CH(ν)dν. (4b)
2π −ω jective is to minimize the error between the given model and the
sought reduced order model in a frequency-limited H2 -norm, us-
These Gramians can be shown to satisfy the following Lyapunov
ing the frequency-limited Gramians. We formulate the optimiza-
equations, see [7],
tion problem
APω + Pω AT + Sω BBT + BBT S∗ω = 0, (5a)  2
minimize G − Ĝ = minimize ∥E ∥2H2 ,ω , (12)
 
T
A Qω + Qω A + S∗ω CT C + C CSω = 0,
T
(5b) Ĝ H2 ,ω Ĝ
34 D. Petersson, J. Löfberg / Systems & Control Letters 67 (2014) 32–39

where Since the principal branch of the logarithm is used, Theorem 1.18
 ω in [23] can be used. Which for this case means that given a matrix
1
∥E ∥2H2 ,ω = tr E (iν)E ∗ (iν)dν. (13) C ∈ Cn×n it holds that ln C = ln C. Sω becomes
2π −ω
i i
Assume that the system E is stable and described by Sω = ln(−A − iωI) +
ln(−A − iωI)

 2π
i
 
AE BE ln(−A − iωI) . 
E: . (14) = Re
π
CE DE

Given G and Ĝ, represented as Now, we want to rewrite the cost function (17) to a more com-
  putationally tractable form. This is done by using the realization
 
A B Â B̂ given in (16) and by partitioning the Gramians PE ,ω and QE ,ω as
G: , Ĝ : , (15)
C D Ĉ D̂    
Pω Xω Qω Yω
PE ,ω = , QE ,ω = , (22)
the error system can be realized, in state space form, as XTω P̂ω YTω Q̂ω
      and SE ,ω as
  A 0 B
AE BE .
 
E: = 0 Â B̂ (16) Sω 0
CE DE   SE ,ω = . (23)
C −Ĉ D − D̂ 0 Ŝω

This realization of the error system will later prove beneficial when Pω , Qω , P̂ω , Q̂ω , Xω and Yω satisfy, due to (18), the Sylvester and
rewriting the optimization problem. Throughout the paper we will Lyapunov equations
assume that the given model is stable.
The cost function for the optimization problem (12) can be APω + Pω AT + Sω BBT + BBT S∗ω = 0, (24a)
written as
AXω + Xω Â + Sω BB̂ + BB̂ Ŝω = 0,
T T T ∗
(24b)
ω  T

∥E ∥2H2 ,ω = tr CE PE ,ω CTE + 2 tr CE SE ,ω BE + DE DE (17a) ÂP̂ω + P̂ω ÂT + Ŝω B̂B̂T + B̂B̂T Ŝ∗ω = 0, (24c)

 ω   T
A Qω + Qω A + S∗ω CT C + C CSω = 0,
T
(24d)
= tr BTE QE ,ω BE + 2 tr CE SE ,ω BE + DE DTE (17b)

AT Yω + Yω Â − S∗ω CT Ĉ − CT ĈŜω = 0, (24e)
where
T
 Q̂ω + Q̂ω  + Ŝ∗ω ĈT Ĉ + Ĉ ĈŜω = 0,
T
(24f)
AE PE ,ω + PE ,ω ATE + SE ,ω BE BTE + BE BTE S∗E ,ω = 0, (18a)
with
ATE QE ,ω + QE ,ω AE + S∗E ,ω CTE CE + CTE CE SE ,ω = 0, (18b)  
i
Sω = Re ln(−A − iωI) ,
with π
  (25)
i i
SE ,ω = ln((AE + iωI)(AE − iωI) −1
). (19) Ŝω = Re ln(−Â − iωI) .
2π π
In this paper we have also derived a simpler expression for the Note that Pω and Qω satisfy the Lyapunov equations for the
matrix SE ,ω , compared to what is presented in [7]. frequency-limited controllability and observability Gramians for
the given model, and P̂ω and Q̂ω satisfy the Lyapunov equations for
Lemma 1. For a matrix A that is Hurwitz it holds that the frequency-limited controllability and observability Gramians
for the sought model.
i
Sω = ln((A + iωI)(A − iωI)−1 ) With the partitioning of PE ,ω and QE ,ω it is possible to rewrite
2π (17) in two alternative forms
 
i
ln(−A − iωI) ,
 
= Re (20) ∥E ∥2H2 ,ω = tr BT Qω B + 2BT Yω B̂ + B̂T Q̂ω B̂
π
 ω  ω 
where ln A denotes the matrix complex logarithm using the principal + 2 tr CSω B + D − ĈŜω B̂ + D̂
branch.   2π 2π
× DT − D̂T , (26a)
Proof. Sω can be written as  
∥E ∥2H2 ,ω = tr CPω CT − 2CXω ĈT + ĈP̂ω ĈT
i
Sω = ln((−A − iωI)(−A + iωI)−1 ). (21)
 ω  ω 
2π + 2 tr CSω B + D − ĈŜω B̂ + D̂
 2π 2π
The matrices (−A − iωI) and (−A + iωI) commute and since

× DT − D̂T . (26b)
A is Hurwitz, both (−A − iωI) and (−A + iωI) will have their
eigenvalues in the right half plane. Now using Theorems 11.2 and
11.3 in [23], yields Remark 2. Note that neither the term BT Qω B nor the term CPω CT ,
i which are included in the cost function (26), depends on the
Sω = ln((−A − iωI)(−A + iωI)−1 ) optimization variables, Â, B̂, Ĉ and D̂. Hence, these terms can be

i i excluded from the optimization. These are the only terms including
= ln(−A − iωI) − ln(−A + iωI). Pω and Qω which are the most costly to compute.
2π 2π
D. Petersson, J. Löfberg / Systems & Control Letters 67 (2014) 32–39 35

When optimizing the frequency-limited H2 -norm using the where


system matrices as optimization variables we have the freedom to  T
i 
choose which elements we optimize over, i.e., we can introduce W = Re L −Â − iωI, V , (30a)
structure in Â, B̂, Ĉ and D̂, as long as we can find an  that is π
Hurwitz. Let us introduce the matrices SÂ , SB̂ , SĈ and SD̂ which hold
 
V = ĈT ĈP̂ − ĈT CX − ĈT D − D̂ B̂T , (30b)
the structure of the sought matrices, i.e.,
   0 = ÂP̂ + P̂ÂT + B̂B̂T , (30c)
  1, if  is a free variable;
SÂ = (27)
ij
ij
0 = AX + XÂ + BB̂ ,T T
(30d)
0, otherwise.
with the function L(·, ·) being the Frechét derivative of the matrix
We will see in the next section, that due to the element-wise logarithm, see [23].
differentiation, this structure will be inherited in the gradient.
The parametrization of the sought system using the full system Remark 3. Remember that if ω = ∞ is included as the end
matrices is of course redundant, which leads to a non-unique frequency we need to have DE = D − D̂ = 0, i.e., D̂ constant and
minimum of the cost function in the parameter space. This leads, with D = D̂.
in theory, to a singular Hessian matrix. However, this is taken care
of in most quasi-Newton solvers to ensure that the minimum is 3.2. Numerical remarks
reached in a numerically stable way. The over-parametrization can
also help the algorithm numerically, see [24]. The most computationally heavy parts in the cost function are
solving the Lyapunov equations in (24a) and (24d) and computing
Sω which cost O (n3 ) to compute. However, these entities do
3.1. Gradient of the cost function not depend on the optimization variables and can be computed
ones, before the optimization starts and during the iterations
An appealing feature of the proposed nonlinear optimization in the optimization the cost is O (n2 nr ) to compute the cost
approach, using our proposed H2 -like measure to solve the prob- function. To compute the gradient we also need to compute the
lem, is that the Eqs. (26) are differentiable in the system matri- Frechét derivative of a matrix of size nr by nr which costs about
ces, Â, B̂, Ĉ and D̂. In addition, the closed form expression obtained
O (n3r ) to compute, see [25] for more details. This means that the
computational cost to compute the cost function and its gradient
when differentiating the cost function is expressed in the given
is, first an overhead of O (n3 ) and then O (nr n2 ) per iteration.
data (A, B, C and D), the optimization variables (Â, B̂, Ĉ and D̂) and Looking at the matrices used in this article to evaluate the ma-
solutions to the equations in (24). trix logarithm and its Frechét derivative one realizes that com-
To show this we start by differentiating with respect to B̂, Ĉ and plex arithmetic is needed to compute these entities. However, the
D̂. First we note that neither Qω , Yω nor Q̂ω in Eq. (26a) depends on resulting matrices used, i.e., Sω , Ŝω and W, are always real since
B̂ which means that the equation is quadratic in B̂. Analogous ob- we are only interested in the real part of the computations. This
servations can be made with (26b) and the variable Ĉ and similarly means that complex arithmetic is needed to compute parts of the
cost function and its gradient. However, it is not needed in the op-
with D̂. Hence, the derivative of the cost function with respect B̂, Ĉ
timization solver to perform a step. By supplying the cost func-
and D̂ becomes tion and its gradient in these computationally efficient forms the
presented method can be used in any off-the-shelf Quasi-Newton
∂ ∥E ∥2H2 ,ω   
= 2 Q̂ω B̂ + YTω B − ŜTω ĈT D − D̂ ⊙ SB̂ , (28a) solver.
∂ B̂ The optimization problem described in this article is a non-
convex problem and we cannot guarantee to find any global min-
∂ ∥E ∥2H2 ,ω    
= 2 ĈP̂ω − CXω − D − D̂ B̂T ŜTω ⊙ SĈ , (28b) ima, hence, it is very important to have a good starting point when
∂ Ĉ performing the optimization. Additionally, to keep the cost func-
tion bounded, the initial point needs to be a stable model. The
∂ ∥E ∥2H2 ,ω  ω ω 
= −2 CSω B + D − ĈŜω B̂ − D̂ ⊙ SD̂ , (28c) method described in this article can, hence, be thought of as a two
∂ D̂ 2π 2π step method, first use your favorite model reduction method to
find a stable model of right order, e.g. balanced truncation, and then
where ⊙ represents the Hadamard product of matrices, i.e.,
use this model to initialize the optimization problem described
element-wise multiplication.
herein.
For the more complicated case of differentiating with respect to In the optimization problem described in this article we try to
 we observe that Q̂ω and Yω do depend on Â, see the equations in optimize over Hurwitz matrices of a specific order, which is a well
(24). The calculations of this part of the gradient are lengthy and known hard problem. As can be expected we do not have any solu-
can be found in the Appendix. tion to this, however, by starting with a stable model (Â Hurwitz)
The complete gradient becomes we can check the eigenvalues of the new  when a step is proposed
to assure that the next step keeps the stability of the model.
∂ ∥E ∥2H2 ,ω  
= 2 YTω X + Q̂ω P̂ ⊙ SÂ − 2W ⊙ SÂ , (29a)
∂ Â 4. Numerical examples
∂ ∥E ∥2H2 ,ω   
In this section we will present three examples that will show
= 2 Q̂ω B̂ + YTω B − ŜTω ĈT D − D̂ ⊙ SB̂ , (29b)
∂ B̂ the applicability of the proposed method, and compare with other
relevant methods. To be able to measure how well the different
∂ ∥E ∥2H2 ,ω    
= 2 ĈP̂ω − CXω − D − D̂ B̂T ŜTω ⊙ SĈ , (29c) methods perform the relative error of the norm is used, i.e.
∂ Ĉ  
G − Ĝ
 
∂ ∥E ∥2H2 ,ω  ω ω  H2 ,ω
= −2 CSω B + D − ĈŜω B̂ − D̂ ⊙ SD̂ , (29d) . (31)
∂ D̂ 2π 2π ∥G∥H2 ,ω
36 D. Petersson, J. Löfberg / Systems & Control Letters 67 (2014) 32–39

Table 1
Numerical results for Example 1.
   
G−Ĝ
 
G−Ĝ
 
H2 ,ω H∞ ,ω
∥G∥H2 ,ω ∥G∥H∞ ,ω

wbt 3.01e−01 2.91e−01


flistia 6.38e−02 3.96e−02
mflbt 1.00e+00 1.00e+00
flbt 6.31e−02 4.00e−02
spanos 5.97e−02 3.95e−02
flh2 nl 1.02e−02 1.15e−02

To shorten the names and make the figures more readable we will
denote our proposed method as flh2 nl. The methods that will be
used for comparison, in the different examples, are
• bt—ordinary balanced truncation, the implementation used is
the function schurmr in Robust Control Toolbox in Matlab
• wbt—weighted balanced truncation, an implementation of the
method in [5]
Fig. 1. The true and reduced order models for Example 1. The dashed vertical line
• flbt—frequency limited balanced truncation, an implementa- denotes ω = 2. The reader is referred to the digital version of this paper, for a color
tion of the method in [7] plot.
• mflbt—modified frequency limited balanced truncation, an
implementation of the method in [2]
• flistia—frequency limited iterative tangential interpolation
algorithm (see [21]), the implementation in the more-toolbox
is used
• spanos—weighted H2 -model reduction, an implementation of
the method in [26].
When solving the examples in this section we use the function
fminunc in Matlab, that uses a quasi-Newton algorithm, to
perform the optimization. Since the optimization problem is a non-
convex problem the initialization of the optimization is important.
In the presented examples we have used the model from balanced
truncation to initialize the optimization.
We will start with a small example to illustrate how the fre-
quency limited methods behave and show some bode plots. We
will then continue with two larger examples to show the applica-
bility for more realistic systems.

Example 1 (Small Illustrative Example). This example addresses a


small model with four states. The model is composed of two second Fig. 2. The error models for the different methods for Example 1. The dashed
vertical line denotes ω = 2. The reader is referred to the digital version of this
order models in series, one with a resonance frequency at ω = 1 paper, for a color plot.
and the other at ω = 3. We will limit the frequency range to ω ∈
[0, 2] and find a model with only two states to try to only capture Table 2
the first model. Numerical results for Example 2.
   
1 9 G−Ĝ
 
G−Ĝ
 
H2 ,ω
G = G1 G2 = . (32)
H∞ ,ω

s2 + 0.2s + 1 s2 + 0.003s + 9
∥G∥H2 ,ω ∥G∥H∞ ,ω

wbt 1.24e−03 9.50e−04


We compare the methods flh2 nl, flistia, flbt and mflbt, we also mflbt 1.25e−03 9.43e−04
compare these methods with the methods spanos and wbt using flbt 1.24e−03 9.41e−04
a low pass Butterworth filter of order 10 with a cutoff frequency flh2 nl 6.95e−04 6.83e−04
of 2. The results from the different methods can be seen in Figs. 1
and 2 and Table 1. As can be seen in the result we are successful We will compare the frequency limited methods flbt, mflbt and
in finding a good model for the relevant frequencies with flh2 nl, flh2 nl and the weighted method wbt with a tenth order Butter-
flistia, flbt, spanos and wbt. However, mflbt captures the wrong worth band pass filter with cutoff frequencies equal to ω. Looking
resonance mode (from our perspective) and fails completely in the at the results in Figs. 3, 4 and Table 2 we see that in this case all the
lower frequency region. Interesting to note is how the models, that methods do a good job, and again flh2 nl finds the best model.
do a good job, sacrifice the model fit at higher frequencies for the
lower.
Example 3 (Clamped Beam Model, Varying Frequency Interval). In
this example we will use the model of a clamped beam, a
Example 2 (CD Player). This example uses a slightly larger model, a siso model with 348 states which can be found in [27]. In this ex-
model of a compact disk player with 120 states and two inputs and ample, we will try to focus on certain frequency intervals of the
two outputs, see [27]. In this example we will only look at the trans- model. We will focus on different frequency intervals, [0, ω], ω ∈
fer function from the second input to the first output. Here we will [2, 40] and fix the reduced order model to have 12 states, nr = 12.
focus on a banded frequency interval, ω ∈ [10, 1000] where the We will use the method flh2 nl and compare to the frequency lim-
main peak gain is and reduce the model to a model with 12 states. ited methods flistia, flbt and mflbt. Additionally, we will use the
D. Petersson, J. Löfberg / Systems & Control Letters 67 (2014) 32–39 37

5. Conclusions

In this paper we have proposed a method, based on nonlin-


ear optimization, that uses the frequency-limited Gramians in-
troduced in [7], and the method does not have the drawback of
finding unstable models, as explained in the previous section. We
use these Gramians to construct a frequency-limited H2 -norm
which describes the cost function of the optimization problem. We
derive a gradient of the proposed cost function which enables us to
use off-the-shelve optimization software to solve the problem ef-
ficiently. The derivation of the method also enables us to impose
structural constraints, e.g., upper triangular A-matrix, in the sys-
tem matrices. The derivation follows closely the technique in [28]
and it is easy to extend the method to identify lpv-models. We have
also presented three examples of different sizes and characteristics
to show the applicability of the method. The method uses a non-
linear optimization approach on a nonconvex problem, hence the
unavoidable problem of local minima is present. However, we con-
Fig. 3. The true and reduced order models for Example 2. The dashed vertical lines sider this as a two step approach where you use your favorite re-
denote ω = 10 and ω = 10 000. The reader is referred to the digital version of this duction method as an initial point to this method which will refine
paper, for a color plot. the solution, as we have shown in the examples.

Appendix. Derivation of the gradient w.r.t. Â

In this appendix we present the differentiation of the cost


function with respect to Â.
The cost function is
 
∥E ∥2H2 ,ω = tr BT Qω B + 2BT Yω B̂ + B̂T Q̂ω B̂
 ω  ω 
+ 2 tr CSω B + D − ĈŜω B̂ + D̂
2π 2π
 
T T
× D − D̂ (A.1)

and by looking at the equations

AT Yω + Yω Â − S∗ω CT Ĉ − CT ĈŜω = 0, (A.2a)


T
 Q̂ω + Q̂ω  + Ŝ∗ω ĈT Ĉ + Ĉ ĈŜω = 0,
T
(A.2b)

we observe that Q̂ω and Yω depend on  which we need to

Fig. 4. The error models for the different methods for Example 2. The dashed   when differentiating (A.1) with respect to Â. Hence,
keep in mind
∂∥E ∥H ,ω
2
vertical lines denote ω = 10 and ω = 10 000. The reader is referred to the digital 2
becomes
version of this paper, for a color plot. ∂ Â
ij

∂ ∥E ∥2H2 ,ω
 
method wbt with a tenth order Butterworth low-pass filter, with
the cutoff frequency equal to ω. Looking at the left plot of Fig. 5 we ∂ Â ij
see that for small ω, the H2 optimal methods do very well. How-  
ever, for ω > 7 we see that flh2 nl, even though optimizing the ∂ Yω T ∂ Q̂ω ∂ Ŝω  T 
= tr 2B̂B T
+ B̂B̂ − 2Ĉ B̂ D − D̂T
, (A.3)
same cost function, does better than flistia. As in the previous ex- ∂ âij ∂ âij ∂ âij
ample, we see that the relative H∞ -norm still keeps low. Almost
∂ Q̂
for all ω, the H2 optimal methods have better relative H∞ -error where ∂∂Yâω and ∂ âω depend on  via the differentiated versions of
ij ij
than the methods using balanced truncation. the equations in (A.2),
In all three of the above examples we observe that the proposed
method finds an at least as good model as the other methods, but ∂ YTω ∂ YT ∂ ÂT T ∂ Ŝ∗ω T
with the proposed method we can guarantee that the reduced or- ÂT + ωA+ Yω − Ĉ C = 0, (A.4a)
∂ âij ∂ âij ∂ âij ∂ âij
der model is stable and with the possibility to impose structure in
the system matrices. The way that is used to guarantee that the ∂ Q̂ω ∂ Q̂ω ∂ ÂT ∂ Â
ÂT + Â + Q̂ω + Q̂ω
stability of the reduced order model is to use a standard bracket- ∂ âij ∂ âij ∂ âij ∂ âij
ing line-search method and check the eigenvalues of the matrix
 for all the proposed steps and reject steps that render the model ∂ Ŝ∗ω T ∂ Ŝω
+ Ĉ Ĉ + ĈT Ĉ = 0. (A.4b)
unstable (effectively using an indicator function as a penalty on the ∂ âij ∂ âij
objective function). Since the algorithm always starts in an initially ∂ Q̂
strictly stable point, there will always exist a non-zero step length To be able to substitute ∂ âω and ∂∂Yâω to something more computa-
ij ij
preserving stability. tionally tractable we use the following lemma.
38 D. Petersson, J. Löfberg / Systems & Control Letters 67 (2014) 32–39

Fig. 5. Reduction of a clamped beam model to 12 states with focus on the frequency interval [0, ω], ω ∈ [2, 40] using flh2 nl, flistia, flbt, mflbt and wbt. The filter used
for the weighted method is a tenth order Butterworth low-pass filter with cutoff frequency ω. To the left we see the relative H2 error and to the right the relative H∞ error.
The reader is referred to the digital version of this paper, for a color plot.

Lemma 2. If M and N satisfy the Sylvester equations where L(A, E) is the Frechét derivative of the matrix logarithm,
see [23], with
AM + MB + C = 0, NA + BN + D = 0, (A.5)
 1
then tr CN = tr DM. L(A, E) = (t (A − I) + I)−1 E (t (A − I) + I)−1 dt , (A.10)
0
Proof. Multiplying the left equation in (A.5) with N from the left,
the right equation with M from the right and taking the trace of r (Â) = −Â − iωI. (A.11)
both equations entail
The function L(A, E) can be efficiently evaluated using an algorithm
tr NAM + tr MBN = −tr NC, tr NAM + tr MBN = −tr DM. described in [23].
By substituting (A.9) into (A.7) and using (A.10) with the fact
Hence, tr CN = tr DM.  that we can interchange the tr-operator and the integral we obtain
∂ Q̂
Studying Lemma 2, the two factors in front of ∂∂Yâω and ∂ âω ∂ ∥E ∥2H2 ,ω
 
ij ij
in (A.3) and the structure of the Lyapunov/Sylvester equations in
∂ Â
(A.4), ÂT · + · Â + ⋆ = 0 and ÂT · + · A + ⋆ = 0, brings us to 
ij

the conclusion that, to do the substitution, we need to solve two ∂ Â  T T  ∂ Ŝ∗ 
ω T T

additional Lyapunov/Sylvester equations, namely = 2 tr Yω X + Q̂ω P̂ + Ĉ ĈP̂ − Ĉ CX
∂ âij ∂ âij
AX + XÂT + BB̂T = 0, (A.6a)
 
∂ Ŝω   T T
 
− 2 tr B̂ D − D̂ Ĉ
ÂP̂ + P̂Â + B̂B̂ = 0.
T T
(A.6b) ∂ âij
 
Note that P̂ in (A.6b) is the controllability Gramian for the reduced ∂ ÂT  T 
order model. = 2 tr Yω X + Q̂ω P̂
∂ âij
Rewriting (A.3) using Lemma 2, (A.4) and (A.6) leads to
∂ Â T 
 T 
i 
∂ ∥E ∥2H2 ,ω
 
− 2 tr Re L r (Â), V
∂ âij π
∂ Â ij  
∂ ÂT    
,
  T
∂ ÂT  T  ∂ Ŝ∗ 
ω
 = tr 2 Yω X + Q̂ω P̂ − 2W (A.12)
= 2 tr Yω X + Q̂ω P̂ + T T
Ĉ ĈP̂ − Ĉ CX ∂ âij
∂ âij ∂ âij
  where
∂ Ŝω   T  
− 2 tr B̂ D − D̂ Ĉ .
T
(A.7) T

i 
∂ âij W = Re L −Â − iωI, V , (A.13)
π
What remains is to rewrite the two last terms in (A.7), which  
∂ Ŝ∗ V = ĈT ĈP̂ − ĈT CX − ĈT D − D̂ B̂T . (A.14)
include ∂∂Ŝâω and ∂ âω . Recall the definition of Ŝω ,
ij ij

   
i i
Ŝω = Re ln(−Â − iωI) = Re ln(r (Â)) (A.8) References
π π
[1] A. Varga, A. Hansson, G. Puyou (Eds.), Optimization Based Clearance of Flight
and differentiate with respect to an element in Â, i.e., aij . This yields Control Laws, in: Lecture Notes in Control and Information Science, Springer,
   2012.
∂ Ŝω i ∂ r (Â) [2] S. Gugercin, A.C. Antoulas, A survey of model reduction by balanced truncation
= Re L r (Â), and some new results, Internat. J. Control 77 (8) (2004) 748–766.
∂ aij 2π ∂ aij [3] A. Ghafoor, V. Sreeram, A survey/review of frequency-weighted balanced
   model reduction techniques, J. Dyn. Syst. Meas. Control 130 (2008) 061004.
i ∂ Â [4] G. Wang, V. Sreeram, W.Q. Liu, A new frequency-weighted balanced truncation
= Re L r (Â), − (A.9) method and an error bound, IEEE Trans. Automat. Control 44 (9) (1999)
2π ∂ aij 1734–1737.
D. Petersson, J. Löfberg / Systems & Control Letters 67 (2014) 32–39 39

[5] D.F. Enns, Model reduction with balanced realizations: an error bound [18] Y. Xu, T. Zeng, Optimal H2 model reduction for large scale mimo systems via
and a frequency weighted generalization, in: Proceedings of the 23rd IEEE tangential interpolation, Int. J. Numer. Anal. Model. 8 (1) (2011) 174–188.
Conference on Decision and Control, Las Vegas, USA, 1984, pp. 127–132. [19] C. Poussot-Vassal, An iterative SVD-tangential interpolation method for
[6] C.-A. Lin, T.-Y. Chiu, Model reduction via frequency weighted balanced medium-scale mimo systems approximation with application on flexible
realization, Control Theory Adv. Technol. 8 (1992) 341–351. aircraft, in: Proceedings of the 50th IEEE Conference on Decision and Control
[7] W. Gawronski, J.-N. Juang, Model reduction in limited time and frequency and the European Control Conference, Orlando, FL, USA, 2011, pp. 7117–7122.
intervals, Int. J. Syst. Sci. 21 (2) (1990) 349–376. [20] B. Anić, C. Beattie, S. Gugercin, A.C. Antoulas, Interpolatory weighted-H2
[8] S. Sahlan, A. Ghafoor, V. Sreeram, A new method for the model reduction model reduction, Automatica 49 (5) (2013) 1275–1280.
technique via a limited frequency interval impulse response Gramian, Math. [21] P. Vuillemin, C. Poussot-Vassal, D. Alazard, H2 optimal and frequency limited
Comput. Modelling 55 (3–4) (2012) 1034–1040. approximation methods for large-scale lti dynamical systems, in: Proceedings
[9] L. Meier, D.G. Luenberger, Approximation of linear constant systems, IEEE of the 5th IFAC Symposium on System Structure and Control, Grenoble, France,
Trans. Automat. Control 12 (5) (1967) 585–588. 2013, pp. 719–724.
[10] D.A. Wilson, Optimum solution of model reduction problem, Proc. Inst. Electr. [22] K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control, Prentice-Hall, Inc.,
Eng. 117 (6) (1970) 1161–1165. 1996.
[23] N.J. Higham, Functions of Matrices: Theory and Computation, SIAM, 2008.
[11] A. Lepschy, G.A. Mian, G. Pinato, U. Viaro, Rational L2 approximation: a non-
[24] T. McKelvey, A. Helmersson, System identification using an over-parametrized
gradient algorithm, in: Proceedings of the 30th IEEE Conference on Decision
model class-improving the optimization algorithm, in: Proceedings of the 36th
and Control, Vol. 3. Brighton, UK, 1991, pp. 2321–2323.
IEEE Conference on Decision and Control, Vol. 3, 1997, pp. 2984–2989. URL
[12] C.A. Beattie, S. Gugercin, Krylov-based minimization for optimal H2 model
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=657905.
reduction, in: Proceedings of the 47th IEEE Conference on Decision and
[25] A.H. Al-Mohy, N.J. Higham, S.D. Relton, Computing the Fréchet derivative of the
Control, New Orleans, USA, 2007, pp. 4385–4390.
matrix logarithm and estimating the condition number. Manchester Institute
[13] S. Gugercin, A.C. Antoulas, C. Beattie, H2 model reduction for large-scale linear for Mathematical Sciences. The University of Manchester, UK MIMS EPrint
dynamical systems, SIAM J. Math. Anal. 30 (2) (2008) 609–638. 2012.72. 2012. URL: http://eprints.ma.man.ac.uk/1852/.
[14] P. Fulcheri, M. Olivi, Matrix rational H2 approximation: a gradient algorithm [26] J. Spanos, M. Milman, D. Mingori, A new algorithm for L2 optimal model
based on schur analysis, SIAM J. Control Optim. 36 (6) (1998) 2103–2127. reduction, Automatica 28 (5) (1992) 897–909.
[15] W.-Y. Yan, J. Lam, An approximate approach to H2 optimal model reduction, [27] F. Leibfritz, W. Lipinski, Description of the Benchmark Examples in COMPLeib
IEEE Trans. Automat. Control 44 (7) (1999) 1341–1358. 1.0, Tech. Rep., University of Trier, Department of Mathematics, Germany,
[16] D.A. Wilson, Model reduction for mutivariable systems, Internat. J. Control 20 2003.
(1) (1974) 57–64. [28] D. Petersson, Nonlinear optimization approaches to H2 -norm based lpv
[17] X.-X. Huang, W.-Y. Yan, K.L. Teo, H2 near-optimal model reduction, IEEE Trans. modelling and control. Licentiate Thesis No. 1453, Department of Electrical
Automat. Control 46 (8) (2001) 1279–1284. Engineering, Linköping University, 2010.

You might also like