IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 57, NO. 11, NOVEMBER 2010 2981
A Wavelet-Collocation-Based Trajectory
Piecewise-Linear Algorithm for Time-Domain
Model-Order Reduction of Nonlinear Circuits
Ke Zong, Fan Yang*, Member, IEEE, and Xuan Zeng*, Member, IEEE
Abstract—Trajectory piecewise-linearization-based reduced-
order macromodeling methods have been proposed to charac-
terize the time-domain behaviors of large strongly nonlinear
systems. However, all these methods rely on frequency-domain
model-order-reduction (MOR) methods for linear systems. There-
fore, the accuracy of the reduced-order models in time domain
cannot always be guaranteed and controlled. In this paper, a
wavelet-collocation-based trajectory piecewise-linear approach is
proposed for time-domain MOR of strongly nonlinear circuits.
The proposed MOR method is performed in time domain and is
based on a wavelet-collocation method. Compared with nonlinear
MOR methods in frequency domain, the proposed method in
time domain maintains higher accuracy for modeling transient
characteristics of nonlinear circuits, which are very important
in macromodeling and transient analysis for nonlinear circuits.
Furthermore, a nonlinear wavelet companding technique is devel-
oped to control the modeling error in time domain, which is useful
for balancing the overall modeling error over the whole time
region and improving the simulation efficiency at higher level.
The numerical results show that the proposed method has high
macromodeling accuracy in time domain, and the modeling-error
distribution in time domain can be efficiently controlled by the
wavelet companding technique.
Index Terms—Model-order reduction (MOR), piecewise linear,
time domain, trajectory, wavelet.
I. INTRODUCTION
W
ITH the continuously increasing complexity of in-
tegrated circuits, fast simulation and verification
become an important part in design flow. In recent years,
model-order-reduction (MOR) techniques have been widely
applied in fast simulation and macromodeling to reduce the
simulation time [1]. With the maturity of the MOR techniques
for linear time-invariant systems, more research interests are
attracted to the MOR of nonlinear systems [2], [3].
Manuscript received October 26, 2009; revised January 27, 2010; accepted
March 18, 2010. First published May 27, 2010; current version published
November 10, 2010. This work was supported in part by NSFC Research
Project 60976034 and 60676018, by the National Basic Research Program of
China under Grant 2005CB321701, by the National Major Science and Tech-
nology Special Project 2008ZX01035-001-06 and 2009ZX02023-4-3 of China
during the 11th five-year plan period, by the Doctoral Program Foundation of
the Ministry of Education of China 200802460068, by the International Science
and Technology Cooperation Program Foundation of Shanghai 08510700100,
and by the Program for Outstanding Academic Leaders of Shanghai. This
paper was recommended by Associate Editor L. B. Goldgeisser.
The authors are with the State Key Laboratory of ASIC & System, Micro-
electronics Department, Fudan University, Shanghai 200433, China.
*Corresponding authors: F. Yang and X. Zeng (e-mail: yangfan@fudan.
edu.cn; xzeng@fudan.edu.cn).
Digital Object Identifier 10.1109/TCSI.2010.2048775
A variety of MOR techniques have been developed for
weakly and strongly nonlinear systems, respectively. The
MOR methods for weakly nonlinear systems are based on the
approximation of the nonlinear systems around an equilibrium
point [2], [4], [5]. Krylov subspace projection techniques are
then utilized to generate the projection matrices as well as
the reduced-order models based on the approximate systems.
Since the nonlinear systems are approximated around a single
equilibrium point, these methods only work well for the weakly
nonlinear systems.
For strongly nonlinear system, approximating the nonlinear
systems around an equilibrium point cannot efficiently capture
strongly nonlinear effects. A natural idea is to approximate
the strongly nonlinear systems around multiple expansion
points, which leads to the trajectory-based methods [3], [6],
[7]. In trajectory-based methods, the multiple expansion points
are extracted from state trajectories driven by “training”
inputs. The nonlinear system is then approximated by a piece-
wise-linear/polynomial system around those expansion points.
Afterwards, the piecewise-linear/polynomial system is reduced
to the final reduced-order model by moment-matching/trun-
cated balanced realization (TBR) [8]–[10] technique. In [3], the
nonlinear system is approximated by a piecewise-linear system
around the expansion points and the piecewise-linear system is
then reduced by moment-matching technique. The method [6]
follows the same linearization strategy as [3] while using TBR
techniques to reduce the piecewise-linear system. In [7], the
nonlinear system is approximated by a piecewise polynomial
system through which higher accuracy can be maintained. In
[11], the trajectory piecewise-linear-based MOR method [3]
was implemented in a full SPICE engine, and a clustering
method was employed to efficiently extract the expansion
points.
For all the aforementioned methods, the final reduction is per-
formed in frequency domain, no matter if it is based on mo-
ment-matching method or TBR method. The errors in the fre-
quency domain result in unpredictable accuracy loss in the time
domain. However, more concerns are raised on the system re-
sponse in time domain when a nonlinear system is analyzed.
Furthermore, it is also important to control the modeling error
in time domain because it is useful for balancing the overall
modeling error over the whole time region and thus improving
the simulation efficiency at higher level. The MOR methods for
nonlinear systems in frequency domain are unable to guarantee
the accuracy and control the modeling error in the time domain.
Therefore, achieving the MOR directly in time domain is more
desirable for nonlinear circuit macromodeling and analysis.
1549-8328/$26.00 © 2010 IEEE