Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

716 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 61, NO.

9, SEPTEMBER 2014

Stability Preserving Model Reduction Technique and


Error Bounds Using Frequency-Limited Gramians
for Discrete-Time Systems
Muhammad Imran and Abdul Ghafoor

Abstract—A frequency-limited interval Gramians-based A technique for achieving a good approximation error within
balanced model reduction technique for discrete-time systems is the desired frequency interval was proposed by Wang and
proposed. The technique provides stable models and also yields Zilouchian [28] for discrete-time systems. This technique does
frequency-response error bounds. Numerical examples are also pre- not involve explicit weights as in Enns’ technique. However,
sented. The results are comparable with other existing frequency-
limited interval Gramians-based model reduction techniques. this technique may also yield unstable reduced-order models
for stable original systems [29]. Moreover, it carries no error
Index Terms—Error bound, limited-frequency Gramians, bounds. This instability issue was addressed in [29], where a
model reduction. frequency-response error bound also appears.
In [29], two algorithms were proposed. Algorithm 1 ensures
I. I NTRODUCTION stability by taking absolute values of eigenvalues. This may
cause a large change in some eigenvalues and no effect on the
M ODEL reduction has gained significant attention in the
last few decades, particularly in modeling of complex
systems, circuit design, and very-large-scale integration circuits
remaining eigenvalues. Algorithm 2 ensures stability by taking
only positive eigenvalues and truncating negative eigenvalues.
[1]–[10]. Approximating a higher order system to a lower Therefore, Algorithms 1 and 2 in [29] do not have a similar
order system by retaining the fundamental properties of the effect on all eigenvalues.
original system is the ultimate goal of model reduction. Model In this brief, a frequency-limited balanced truncation tech-
reduction techniques tend to retain essential properties of the nique for discrete-time systems is proposed, which provides sta-
original system (including input–output behavior and stability) ble reduced-order models by working toward a similar effect on
and yield reduced-order models with less approximation error. all eigenvalues. The proposed technique provides comparable
Balanced truncation [11] is a widely used model reduction frequency-response error and yields easily computable a priori
technique, which ensures stability of reduced-order systems and error bounds. Numerical examples are given to show the useful-
also yields a priori error bound [12]. A balanced truncation ness and comparison of the proposed technique with the existing
technique for discrete-time systems was proposed in [13] and frequency-limited balanced model order reduction techniques.
[14]. Other techniques for model reduction (including Hankel
optimal approximation [15], Krylov technique [16], and Pade II. WANG AND Z ILOUCHIAN ’ S T ECHNIQUE [28]
approximation [17]) are also useful. Here, we review Wang and Zilouchian’s [28] frequency-
Model reduction techniques approximate a full-order system limited interval Gramians-based model reduction technique for
for the whole frequency range; however, several model reduc- discrete-time systems.
tion problems (such as controller reduction) draw attention to Let a discrete-time system be given as
certain frequency intervals rather than the whole frequency
range. This leads to the introduction of a frequency weighting H(z) = C(zI − A)−1 B + D (1)
concept in model reduction. Enns [12] extended the balanced
truncation technique to incorporate frequency weights for input, where A ∈ Rn×n , B ∈ Rn×m , C ∈ Rp×n , D ∈ Rp×m ,
output, and both-sided cases. This technique may provide un- {A, B, C, D} is its nth order minimal realization, m is
stable reduced-order models for stable original systems [18]. To the number of inputs, and p is the number of outputs. The
circumvent the instability problem of Enns’ technique, several controllability and observability Gramians
other techniques have been proposed in the literature [19]–[22]. π
1 −1 −1
Some other related works also appear in [23]–[27]. P = (ejω I − A) BB T (e−jω I − AT ) dω

−π

Manuscript received April 21, 2014; revised May 22, 2014; accepted June 12, 1 −1
2014. Date of publication August 8, 2014; date of current version September 1, Q= (e−jω I − AT ) C T C(ejω I − A)−1 dω

2014. This brief was recommended by Associate Editor G. J. Dolecek. −π
The authors are with the Department of Electrical Engineering, Military
College of Signal, National University of Sciences and Technology, Islamabad respectively, satisfy the following discrete-time Lyapunov
44000, Pakistan (e-mail: m.imran@mcs.edu.pk; abdulghafoor-mcs@nust.
edu.pk).
equations:
Color versions of one or more of the figures in this brief are available online
at http://ieeexplore.ieee.org.
AP AT − P + BB T = 0
Digital Object Identifier 10.1109/TCSII.2014.2346688 AT QA − Q + C T C = 0.
1549-7747 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
IMRAN AND GHAFOOR: MODEL REDUCTION TECHNIQUE AND ERROR BOUND USING FREQUENCY-LIMITED GRAMIANS 717

A model order reduction problem is to find Let the new controllability PIG and observability QIG
Gramians, respectively, be calculated by solving the following
Hl (z) = C1 (zI − A11 )−1 B1 + D1 (2) Lyapunov equations:
which approximates the original system (in the desired fre- APIG AT − PIG + BIG BIG
T
=0 (5)
quency range [ω1 , ω1 ], 0 ≤ ω1 ≤ ω2 ≤ π), where A11 ∈ Rl×l , A QIG A − QIG + CIG CIG = 0
T T
(6)
B1 ∈ Rl×m , C1 ∈ Rp×l , and D1 ∈ Rp×m with l < n.
Let the frequency-domain controllability interval Gramian where BIG and CIG are the new fictitious input and output
PWZ and the observability interval Gramian QWZ , respectively, matrices, respectively, defined as
be defined as 
 UIG (SIG − sn I)1/2 , for sn < 0
1 −1 −1 BIG =
PWZ = (ejω I − A) BB T (e−jω I − AT ) dω U S ,
1/2
for sn ≥ 0
2π  IG IG

δω (RIG − rn I)1/2 VIG
T
, for rn < 0
1 −1 −1
CIG = 1/2 T
QWZ = (e−jω I − AT ) C T C(ejω I − A) dω. RIG VIG , for rn ≥ 0.

δω The terms UIG , SIG , VIG , and RIG are calculated as
T T
These Gramians PWZ and QWZ are the solution of the follow- XWZ = UIG SIG UIG and YWZ = VIG RIG VIG , where SIG =
ing Lyapunov equations: diag(s1 , s2 , s3 , . . . , sn ), RIG = diag(r1 , r2 , r3 , . . . , rn ), s1 ≥
s2 ≥ · · · ≥ sn , and r1 ≥ r2 ≥ · · · ≥ rn .
APWZ AT − PWZ + XWZ = 0 (3) Let a contragredient transformation matrix T (used to trans-
AT QWZ A − QWZ + YWZ = 0 (4) form the original system) be obtained as

where T T QIG T = T −1 PIG T −T = diag(σ1 , σ2 , σ3 , . . . , σn )


XWZ = BB T F H + F BB T where σj ≥ σj+1 , j = 1, 2, 3, . . . , n − 1, σl > σl+1 . Reduced-
YWZ = C T CF + F H C T C  order models are calculated by partitioning the transformed
ω2 − ω 1 1 −1 realization.
F =− I+ (ejω I − A) dω.
4π 2π Remark 2: Since XWZ ≤ BIG BIG T
, YWZ ≤ CIG T
CIG ,
δω BIG BIGT
≥ 0, CIG T
CIG ≥ 0, PIG > 0, and QIG > 0, the
F H is the Hermitian of F , and δω is the integration interval realization (A, BIG , CIG ) is minimal. Moreover, the reduced-
[ω1 , ω2 ]. order systems are guaranteed to be stable.
Let Theorem 1: The following error bound for the pro-
posed technique holds  ifthe rank conditions rank[BIG B] =
T T QWZ T = T −1 PWZ T −T = diag{σ1 , σ2 , σ3 . . . , σn } rank[BIG ] and rank CCIG = rank[CIG ] (which follows from
[29]) are satisfied, i.e.,
where σj ≥ σj+1 , j = 1, 2, . . . , n − 1, σl > σl+1 , and T is a
contragredient matrix used to transform the original system 
n

realization. The reduced models are obtained by partitioning H(z) − Hl (z)∞ ≤ 2LIG KIG  σj
j=l+1
the transformed realization.
Remark 1: The symmetric matrices XWZ and YWZ obtained where
by the above procedure may not guarantee to be positive semi- 
definite; hence, the reduced-order models calculated by Wang CVIG (RIG − rn I)−1/2 , for rn<0
LIG = −1/2
and Zilouchian’s technique [28] may not be stable [29]. This CVIG RIG , for rn≥0

issue was solved in [29], and error bounds were also derived. (SIG − sn I)−1/2 UIG
T
B, for sn<0
KIG = −1/2 T
SIG UIG B, for sn≥ 0.
III. M AIN R ESULTS
 
In [29], two algorithms were presented to address the sta- Proof: Since rank[BIG B] = rank[BIG ] and rank CCIG =
bility issue. In algorithm 1 [29], the symmetric matrices XWZ rank[CIG ], the relationships B = BIG KIG and C = LIG CIG
and YWZ are ensured positive/semipositive definite by taking hold. By partitioning BIG = B IG1
BIG2
and CIG = [CIG1 CIG2 ]
the square root of absolute values of the eigenvalues. This may and substituting B1 = BIG1 KIG and C1 = LIG CIG1 , respec-
sometimes lead to a large change in some eigenvalues and may tively, yield
not affect other eigenvalues (of XWZ and YWZ ). On the other
end, algorithm 2 [29] ensures the positive definiteness of the H(z)
 − Hl (z)∞−1 
matrices XWZ and YWZ by retaining only positive eigenvalues =C(zI − A) B −−1C1 (zI − A11 ) B1 ∞
−1 
and truncating negative eigenvalues. This algorithm also does 
= LIG CIG (zI − A) BIG KIG 
not have a similar effect on all eigenvalues (of XWZ and YWZ ). −1 
In this brief, a technique is proposed where effort is to have  −LIG CIG1 (zI − A 11 ) BIG1 KIG ∞
= LIG CIG (zI − A) BIG
−1

a similar effect on all eigenvalues of indefinite matrices XWZ
−CIG (zI − A ) −1
B K 
and YWZ . The reduced-order model obtained is guaranteed to 1 11 IG 1 IG ∞
be stable. Moreover, it yields frequency-response error bound 
≤ LIG  CIG (zI − A) BIG −1

and improved frequency-response error. −CIG (zI − A11 )−1 BIG  KIG .
1 1 ∞
718 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 61, NO. 9, SEPTEMBER 2014

If {A11 , BIG1 , CIG1 } is a reduced-order model obtained by TABLE I


R EDUCED -O RDER S YSTEMS IN THE F REQUENCY R ANGE 0.3π TO 0.5π
partitioning a balanced realization {A, BIG , CIG }, we have
[14], [15]
  
n
 CIG (zI −A)−1 BIG −CIG (zI −A11 )−1 BIG  ≤ 2 σj .
1 1 ∞
j=l+1

Therefore

n
H(z) − Hl (z)∞ ≤ 2LIG KIG  σj .
j=l+1

Remark 3: For the case when symmetric matrices XWZ ≥ 0


and YWZ ≥ 0, then PWZ = PIG and QWZ = QIG . Other-
wise, PWZ < PIG , QWZ < QIG , and (λj [PWZ QWZ ])1/2 ≤
(λj [PIG QIG ])1/2 .
Remark 4: When XWZ ≥ 0 and YWZ ≥ 0, then
XIG T
= BIG BIG = XWZ − sn I
YIG = CIG CIG = YWZ − rn I
T

PIG = PWZ + Pad


QIG = QWZ + Qad

which is shown by writing (5) and (6) as


A(PWZ + Pad )AT − (PWZ + Pad )
+ (XWZ − sn I) = 0, for sn < 0
AT (QWZ + Qad )A − (QWZ + Qad )
Fig. 1. Frequency-response error comparison of the reduced-order systems.
+ (YWZ − rn I) = 0, for rn < 0
APad AT − Pad − sn I = 0, for sn < 0
AT Qad A − Qad − rn I = 0, for rn < 0. Table I shows reduced-order systems obtained by using
Wang and Zilouchian’s [28], Ghafoor and Sreeram’s algorithms
Remark 5: To ensure efficient computation, Hammarling’s (1 and 2) [29], and proposed techniques in the frequency range
technique [30] is used to obtain Cholesky factors of Gramian 0.3π to 0.5π, respectively. Note that the reduced-order models
matrices from the original system realization without actually obtained for the first and second orders are unstable for Wang
computing controllability and observability Gramian matrices, and Zilouchian’s [28] technique, whereas the Ghafoor and
respectively. Note that, in the case of Wang and Zilouch- Sreeram’s algorithms (1 and 2) [29] and proposed technique
ian’s technique, Hammarling’s technique is not applicable (due yields stable reduced-order systems. The reduced third-order
to possibility of having XWZ ≥ 0 and YWZ ≥ 0). However, model (not shown in Table I) obtained by these techniques
Hammarling’s technique is applicable to Ghafoor and is stable.
Sreeram’s algorithms (1 and 2) and proposed techniques. For Example 2: Consider a sixth-order stable discrete-time sys-
example, in the proposed technique, we can apply Ham- tem represented by the equation at the bottom of the page.
marling’s technique on the realization {A, BIG , CIG }. More- Fig. 1 represents the frequency-response errors, σ[H(z) −
over, the computation cost using Hammarling’s technique for Hl (z)], where Hl (z) is the fourth-order reduced model pro-
the proposed and Ghafoor and Sreeram’s algorithms remains duced by balanced truncation [11], Wang and Zilouchian’s [28],
the same. Ghafoor and Sreeram’s algorithms (1 and 2) [29], and proposed
techniques, respectively. The frequency interval for computing
the error response is 0.35π to 0.45π. Fig. 2 represents a magni-
IV. N UMERICAL E XAMPLES fied view of the frequency-response error.
Example 1: Consider a fourth-order stable discrete-time sys- Example 3: Consider a sixth-order stable discrete-time sys-
tem [18] represented by tem represented by

z3 z 5 + 5z 4 + 7z 3 + 8z 2 + 10z + 1
H(z) = . H(z) = .
z4 + 1.1z 3 − 0.01z 2 − 0.275z − 0.06 4z 6 + 3.5z 5 + 3z 4 + 2z 3 + 1.2z 2 + z + 0.4

0.011z 6 + 0.0635z 5 + 0.1653z 4 + 0.2169z 3 + 0.1565z 2 + 0.7124z + 0.0882


H(z) =
z 6 − 1.5z 5 + 2.3z 4 − 2.1z 3 + 1.5z 2 − 0.69z + 0.2
IMRAN AND GHAFOOR: MODEL REDUCTION TECHNIQUE AND ERROR BOUND USING FREQUENCY-LIMITED GRAMIANS 719

Fig. 2. Frequency-response error comparison—magnified view.


Fig. 4. Frequency-response error comparison—magnified view.

Fig. 3. Frequency-response error comparison of the reduced-order systems.


Fig. 5. Frequency response with magnitude and phase error comparison of the
Fig. 3 represents the frequency-response errors σ[H(z) − reduced-order systems.
Hl (z)], where Hl (z) is the second order reduced model pro-
duced by balanced truncation [11], Wang and Zilouchian’s [28], Fig. 6 represents a magnified view of the magnitude and phase
and Ghafoor and Sreeram’s algorithms (1 and 2) [29] and of frequency-response errors. It is observed that the proposed
proposed techniques, respectively. The frequency interval for technique compares well with other existing techniques when
computing error response is 0.5π to 0.7π. Fig. 4 represents both the magnitude and the phase of frequency-response errors
a magnified view of the frequency-response error, where the are considered in the desired frequency interval. Note that the
unweighted balanced truncation plot is not shown because of reduced-order models obtained using Wang and Zilouchian’s
its large value. technique is unstable, having a pole at z = 1.5699.
Example 4: Consider a sixth-order stable discrete-time sys- Discussion: It is observed (from Examples 2 to 4) that in the
tem represented by the equation at the bottom of the page. desired frequency interval, Wang and Zilouchian’s technique
Fig. 5 represents the magnitude and phase of the frequency gives better approximation error, but it sometimes yields un-
response errors, i.e., σ[H(z) − Hl (z)], where Hl (z) is the stable reduced-order models as given in Example 1. The pro-
fourth-order reduced model produced by using Wang and posed technique provides comparatively good approximation
Zilouchian’s [28], Ghafoor and Sreeram’s algorithms (1 and 2) as compared to Ghafoor and Sreeram’s algorithms (1 and 2)
[29], and proposed techniques, respectively. The frequency in the desired frequency interval. Ghafoor and Sreeram’s tech-
interval for computing the error response is 0.37π to 0.42π. nique algorithms (1 and 2) and proposed techniques have the

0.0107z 6 + 0.0642z 5 + 0.1595z 4 + 0.2168z 3 + 0.1525z 2 + 0.0704z + 0.0009


H(z) =
z 6 − 1.4637z 5 + 2.2838z 4 − 2.0587z 3 + 1.4467z 2 − 0.6746 + 0.1825
720 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 61, NO. 9, SEPTEMBER 2014

[9] B. Yan, S. X. Tan, and B. McGaughy, “Second order balanced truncation


for passive order reduction of RLC K Circuits,” IEEE Trans. Circuits Syst.
II, Exp. Briefs, vol. 55, no. 9, pp. 942–946, Sep. 2008.
[10] A. Davoudi, J. Jatskevich, P. L. Chapman, and A. Bidram, “Multiresolu-
tion modeling of power electronics circuits using model-order reduction
techniques,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 60, no. 3,
pp. 810–823, Mar. 2013.
[11] B. C. Moore, “Principal component analysis in linear systems: Control-
lability, observability, model reduction,” IEEE Trans. Autom. Control,
vol. AC-26, no. 1, pp. 17–32, Feb. 1981.
[12] D. F. Enns, “Model reduction with balanced realizations: An error bound
and a frequency weighted generalization,” in Proc. IEEE CDC, Las Vegas,
NV, USA, Dec. 1984, pp. 127–132.
[13] L. Pernebo and L. M. Silverman, “Model reduction via balanced state
space representation,” IEEE Trans. Autom. Control, vol. AC-27, no. 2,
pp. 382–387, Apr. 1982.
[14] U. M. Al Saggaf and G. F. Franklin, “An error bound for discrete re-
duced order model of a linear multivariable system,” IEEE Trans. Autom.
Control, vol. 32, no. 9, pp. 815–819, Sep. 1987.
[15] K. Glover, “All optimal Hankel-norm approximation of linear multivari-
able systems and their L∞ —Error bounds,” Int. J. Control, vol. 39, no. 6,
pp. 1115–1193, Jan. 1984.
Fig. 6. Frequency response with magnitude and phase error comparison—
[16] P. Benner, T. Breite, and T. Damm, “Krylov subspace methods for model
magnified view.
order reduction of bilinear discrete-time control systems,” Proc. Appl.
advantage of providing stable reduced-order models and carry Math. Mech., vol. 1, no. 10, pp. 601–602, Dec. 2010.
[17] O. Ismail, B. Bandyopadhyay, and R. Gorez, “Discrete interval system re-
error bounds also. duction using Pade approximation to allow retention of dominant poles,”
IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 44, no. 11,
V. C ONCLUSION pp. 1075–1078, Nov. 1997.
[18] V. Sreeram, B. D. O. Anderson, and A. G. Madievski, “New results on
In this brief, a frequency-limited Gramians-based model frequency weighted balanced reduction technique,” in Proc. IEEE ACC,
Seattle, WA, USA, Jun. 1995, pp. 4004–4009.
order reduction technique for discrete-time systems has been [19] C. A. Lin and T. Y. Chiu, “Model reduction via frequency weighted bal-
presented. Simulation results show that the proposed technique anced realization,” Control–Theory Adv. Technol., vol. 8, no. 2, pp. 341–
compares well with other existing frequency-limited Gramians- 451, Mar. 1992.
based model order reduction techniques in the desired fre- [20] K. Campbell, V. Sreeram, and G. Wang, “A frequency weighted discrete
system balanced truncation method and an error bound,” in Proc. ACC,
quency range. Chicago, IL, USA, Jun. 2000, pp. 2403–2404.
[21] A. Varga and B. D. O. Anderson, “Accuracy enhancing methods for the
R EFERENCES frequency-weighted balancing related model reduction,” in Proc. IEEE
CDC, Orlando, FL, USA, Dec. 2001, pp. 3659–3664.
[1] X. Chen and J. T. Wen, “Positive realness preserving model reduction with [22] M. Imran, A. Ghafoor, and V. Sreeram, “Frequency weighted model order
H∞ norm error bounds,” IEEE Trans. Circuits Syst. I, Fundam. Theory reduction technique and error bounds for discrete time systems,” Math.
Appl., vol. 42, no. 1, pp. 23–29, Jan. 1995. Probl. Eng., vol. 2014, pp. 498 453-1–498 453-8, Mar. 2014.
[2] J. T. Hsu and L. Vu-Quoc, “A rational formulation of thermal circuit mod- [23] W. M Muda, V. Sreeram, and H. H. C. Lu, “Passivity-preserving frequency
els for electrothermal simulation. II. Model reduction techniques [power weighted model order reduction techniques for general large-scale RLC
electronic systems],” IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., systems,” in Proc. IEEE CCARV, Singapore, Dec. 2010, pp. 1310–1315.
vol. 43, no. 9, pp. 721–732, Sep. 1996. [24] T. Breiten, C. Beattie, and S. Gugercin, Near Optimal Frequency
[3] T. Auba and Y. Funahashi, “Interpolation approach to Hankel-norm model Weighted Interpolatory Model Reduction, Preprint, Max Planck Institute
reduction for rational multi-input multi-output systems,” IEEE Trans. Magdeburg Preprint MPIMD/13-15, 2013.
Circuits Syst. I, Fundam. Theory Appl., vol. 43, no. 12, pp. 987–995, [25] H. R. Shaker, “Upper and lower bounds of frequency interval Gramians
Dec. 1996. for a class of perturbed linear systems,” in Proc. IFAC Symp. Robost
[4] Y. Dolgin and E. Zehab, “Model reduction of uncertain FIR discrete- Control Design, Denmark, Jun. 2012, pp. 713–716.
time systems,” IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 51, no. 8, [26] S. Sahlan, A. Ghafoor, and V. Sreeram, “A new method for the model
pp. 406–411, Aug. 2004. reduction technique via a limited frequency interval impulse response
[5] J. Cullum, A. Ruehli, and T. Zhang, “A method for reduced-order mod- Gramian,” Math. Comput. Model., vol. 55, no. 3/4, pp. 1034–1040,
eling and simulation of large interconnect circuits and its application to Feb. 2012.
PEEC models with retardation,” IEEE Trans. Circuits Syst. II, Analog [27] P. Vuillemin, C. P. Vassal, and D. Alazard, A Frequency-Limited H2
Digit. Signal Process., vol. 47, no. 4, pp. 261–273, Apr. 2000. Model Approximation Method With Application to a Medium-Scale Flex-
[6] J. M. Wang, C. C. Chu, Q. Yu, and E. S. Kuh, “On projection based algo- ible Aircraft, Advances in Aerospace Guidance, Navigation and Control.
rithms for model order reduction of interconnects,” IEEE Trans. Circuits Berlin, Germany: Springer-Verlag, 2013.
Syst. I, Fundam. Theory Appl., vol. 49, no. 11, pp. 1563–1585, Nov. 2002. [28] D. Wang and A. Zilouchian, “Model reduction of discrete linear system
[7] Q. Yu, J. M. L. Wang, and E. S. Kuh, “Passive multipoint moment match- via frequency domain balanced realization,” IEEE Trans. Circuits Syst. I,
ing model order reduction algorithm on multiport distributed interconnect Fundam. Theory Appl., vol. 47, no. 6, pp. 830–837, Jun. 2000.
networks,” IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 46, [29] A. Ghafoor and V. Sreeram, “Model reduction via limited frequency
no. 1, pp. 140–160, Jan. 1999. interval Gramians,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 55,
[8] P. Julian, A. Desages, and B. D’Amico, “Orthonormal high-level canon- no. 9, pp. 2806–2812, Oct. 2008.
ical PWL functions with applications to model reduction,” IEEE Trans. [30] S. J. Hammarling, “Numerical solution of the stable, non-negative definite
Circuits Syst. I, Fundam. Theory Appl., vol. 47, no. 5, pp. 702–712, Lyapunov equation,” IMA J. Numer. Anal., vol. 2, no. 3, pp. 303–323,
May 2000. Jul. 1982.

You might also like