IEEE TRANSACTIONS ON CIRCUITS ANO SYSTEMS, VOL. CAS-30, NO.

3, MARCH 1983 125

Canonical Piecewise-Linear Analysis

LEONOCHUA, FELLOW,IEEE,AND ROBIN L.P. YING, MEMBER,IEEE

Ahsfract -Any continuous resistive nonlinear circuit can be approxi- accuracy. For n > 100, K is generally an extremely large
mated to any desired accuracy by a global piecewise-linear equation in the number. Hence for large n, it becomes extremely tedious
canonical form
and error prone for a user to input all the data necessary to
a+Bx+ ~cil(ai,x)-P,~=O.
specify (1.1) and (1.2). Moreover, this data also requires an
i=l excessive amount of computer memory. For most piece-
All conventional circuit analysis methods (nodal, mesh, cut set, loop, wise-linear functions of practical interest, the above objec-
hybrid, modified nodal, tableau) are shown to always yield an equation of tion can be overcome by representing (1.1) and (1.2) by a
this form, provided the only nonlinear elements are two-terminal resistors single piecewise-linear equation in the canonical form [8]
and controlled sources, each modeled by a one-dimensional piecewise-lin-
ear function.
The well-known Katzenelson algorithm when applied to this equation
f(x)=a+Bx+ f: cil(ai,x)-/3il=0 0.3)
yields an efficient algorithm which requires only a minimal computer i=l
storage. In the important special case when the canonical equation has a where B is a constant n n matrix, a, ci, and cri are
x
lattice structure (which always occur in the hybrid analysis), the algorithm constant n-vectors, and & is a constant. Note that in
is further refined to achieve a dramatic reduction in computation time. general, we have p -=z K.
Our first objective is to show, in Section II, that any
I. INTRODUCTION
piecewise-linear function defined by (1.1) and (1.2) which

T HE KATZENELSON algorithm [l] for solving piece-

wise-linear resistive circuits has been refined and
extended to the most general case of a system of
satisfies the linear partition assumption in [8] has an equiva-
lent canonical representation (1.3), where a, B, ci, (Y~,and pi
can be calculated via explicit formulas.
piecewise-linear equations f(x) = 0 in a series of excellent Our second objective in this paper is to show, in Section
papers during the last decade [2]-[6]. In these papers, the III, that if the only nonlinear elements in a resistive circuit
n-dimensional piecewise-linear functions are specified by are 2-terminal resistors and/or controlled sources (all 4
types) modeled by arbitrary continuous one-dimensional
f(“)=qX+s;, i=1,2;..,K 0.1) piecewise-linear functions, then any conventional circuit
where 4 is a constant n x n matrix and si is a constant analysis method (nodal, mesh, loop, cut set, hybrid, mod-
n-vector, both defined in region R, c [w”. The whole space ified nodal, tableau, etc.) always gives rise to an equation
Iwn is divided into a finite number (K) of polyhedral having the canonical form (1.3). Since the above repertoire
regions by a finite number (p) of hyperplanes of nonlinear elements are sufficient to model all resistive
i=1,2;.. XER” (1.2) n-terminal or coupled elements [9], [lo], we conclude that
(ai? x, =PiY ,P,
all resistive nonlinear circuits can be modeled, to within any
where ai is the normal vector to the i th hyperplane, pi is a prescribed accuracy, by an equation in the canonical form
constant and ( , ) denote the usual vector “dot” product (1.3).
in Iw”. For illustration, see Example I (n = 2, K = 4) in Our final objective is to apply the Katzenelson algorithm
Section 2.3. developed in [6] to the canonical form (1.3). The resulting
In order to solve (1.1) and (1.2) in a computer, it is method-called the canonical Katzenelson algorithm -given
necessary to store an n X n matrix 4, two n-vectors si and in Section IV is applicable to any system of piecewise-lin-
ai, and a scalar pi for each of the K polyhedral regions. The ear equations represented in the canonical form (1.3),
number K depends on the number of regions required to regardless of whether it comes from a circuit or not.
approximate a nonlinear function f( x) to within acceptable
II. CANONICAL FORM REPRESENTATIONOF
Manuscript received April 6, 1982; revised September 20, 1982. This CONTINUOUSPIECEWISE-LINEARFUNCTIONS
work was supported in part by the Joint Services Electromcs Program
under Contract F49620-79-C-0178, by the Office of Naval Research under In this section, we discuss the canonical form representa-
Contract NOOO14-76-C-0572, and the National Science Foundation under
Grant ECS-80-20-640 and Grant ENG-7722745. L. 0. Chua’s research tion of continuous piecewise-linear functions. First, we
was supported in part by the Humboldt Priese from the Alexander von review briefly the single-variable ( f: Iw’ + Iw ‘) case in
Humboldt Foundation of West Germany.
L. 0. Chua is with the Department of Electrical Engineering and Section 2.1 and the scalar multivariable (f: Iw” --, Iw’,
Computer Sciences and the Electronics Research Laboratory, University n > 1) case in Section 2.2. We also present a new formula
of California, Berkeley, CA 94720.
R. L. P. Ying was with the Department of Electrical Engineering and for determining the coefficients of f in the multivariable
Computer Sciences and the Electronics Research Laboratory, University case which is more general than that given in [8]. In Section
of California, Berkeley, CA. He is now with Bell Laboratories, Holmdel,
NJ 07733. 2.3 we propose a general compact form for representing an

0098-4094/83/0300-0125$01.00 01983 IEEE

126 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 3, MARCH 1983

‘f(x)

R2

.
X,’ 0 x3 xi xi+l X R3

Fig. I. A continuous one-dimensional piecewise-linear function.

(4
Fig. 2. Examples illustrating pseudo-unbounded regions and essentially.
n-dimensional piecewise-linear function ( f: [wn + Iwn, n > unbounded regions: (a) R, and Rs are the only pseudo-unbounded
1). regions. (b) R , is the only pseudo-unbounded region.

2.1. One- Dimensional Scalar Function: f: [w’ -+ Iw’

A function f: Iw’ -+ 03’is said to be continuouspiecewise- example, if three lines intersect at a common point in [w2,
linear if (1) it is continuous, and (2) it is composed of then any partition in [w2 containing those three lines is
finitely many linear segments. Points common to two seg- degenerate.
ments of different slopes are called break points. A linear partition in Iw” separates BB”into many poly-
Consider an arbitrary continuous piecewise-linear func- hedral regions. There are unbounded regions as well as
tion f with p distinct breakpoints x, < x2 < * * . < xp, as bounded regions. The unbounded regions can be divided
shown in Fig. 1. Let mi, i = 0, 1,2,. . . ,p denote the slope of into two classes. Suppose some of the hyperplanes in the
each segment. It is proved in [7] that f can be represented linear partition, or their intersections, are parallel to each
globally by the following canonical form: other. As we translate these hyperplanes (or their intersec-
tions) until they coincide into a single hyperplane (or
f(x)=a+bx+ 5 cilx-xi( intersection) some of the unbounded regions may vanish.
(2.1)
i=l
W e call them the pseudo-unbounded regions. The remaining
regions which remain unbounded in spite of the above
where x E Iw’ and the coefficients can be calculated ex-
translations are called essentially-unbounded regions. For
plicitly by
example, in Fig. 2(a), regions R, and R, are pseudo-
unbounded since they vanish upon merging L, and L,.
b=+(m,+m,) (2.2) Regions R,, R,, R,, R,, R,, and R, are essentially un-
bounded. In Fig. 2(b), R, is pseudo-unbounded since it
ei=i(m,--mi+,). i=1,2;..,p (2.3) vanishes upon merging the parallel intersections L,, L,,
and L,. Note that we can always eliminate bounded re-
a = f(O)- 5 c;lx;l. (2.4) gions as well as pseudo-unbounded regions by merging the
i=l parallel hyperplanes and/or their intersections.
A simplified derivation of these relationships are given in Consider a linear partition in 88” defined by (2.5) and an
Appendix A. It should be noted that since (2.2)-(2.4) arbitrary polyhedral region R, E Iw”. Let x be an interior
defines a unique set of coefficients for each piecewise-linear pointofR,.Thenwehave(oli,x)*Rjfori=1,2;..,p.If
function, the representation (2.1) is also unique. we define’
Wk,(X) =%n((~i, x)-Pi)9 i = 1,2;. .,p (2.6)
2.2. n-Dimensional Scalar Function: f: Iw” + Iw’
A linear partition in Iw ’ is a finite set of hyperplanes then the p x 1 vector wk(x) = [We,, wkz(x);. .,w~~(x)]~
characterized by the equation is called the sign-sequence vector of R,. W e can actually
drop the dependence of x on w, since by (2.6) w, remains
(OLiYx, = Pi, i=1,2;*.,p, XER” (2.5)
the same for all x in the interior of R,. Thus a polyhedral
where (Y~E Iw” denotes the normal vector to the i th hyper- region is uniquely identified by its sign-sequence vector.
plane. A linear partition is said to be nondegenerate if for It is shown in [8] that any continuous piecewise-linear
every set of linearly dependent aj,, OL/~,.. . , gj,, 1 < function f: Iw” -+ [w’ with a nondegenerate linear partition
j,, j,, . . . , j,,, < p, the rank of [a,,, ajj,, . . . , ajj,] is strictly less defined by (2.5) can be represented globally by the follow-
than the rank of ing canonical form:

f(X)=a+(bTx)+ i cil(aj>x)-Pil
(2.7)

i=l
Geometrically, this means that the dimension of the inter-
section among any j of the p hyperplanes must be less than
or equal to n - j where n is the dimension of the space. For ‘sgn(x) = 1 if x z 0, sgn(x) = - 1 if x < 0, and sgn(0) is undefined.
CHUA AND YING: CANONICAL PIECEWISErLINEAR ANALYSIS 127

where x, b, (Y~,i = 1,2; * * ,p are in R”, and the coefficients given explicitly by
can ,be calculated explicitly by
B = ; ,k J(x) (2.12)
J=l
b= ; ,i vf(x) W-9 xGRp

J=l
R
c,=i [J(X)I.rER,+- J(X)IxER,e]ai
1 ~T(vfcX,,,:Ivf(x)l~~_)
(2.13)
’ 2 Cai 3 ‘i>
Ci = -
2 (2.9)
Caiy Oli>
D

’ = f(O)- C (2.14)
a=fW- f: ‘iIPiI (2.10) i=l
cilPil

i=l

where J(x) is the Jacobian matrix off: Rj,, j = 1,2, * . *, k,

where R,,, j = 1,2;. . , k, denotes the essentially-unbounded
denotes the essentially unbounded regions, and R,+, Rip
regions, and Ri, and Rip denote the two adjacent regions
denote the two adjacent regions where their sign-sequence
(associated with the ith hyperplane) where their sign-
vectors differs only at the ith position.
sequence vectors differ only at the i th position. A simplified
The derivation of (2.12)-(2.14) is similar to that of
derivation in of these relationships are given in Appen-
(2.8)-(2.10). The coefficients in (2.11) clearly depends on
dix B.
the scalar factor used in representing the linear partition.
Note that for any t E LR, t * 0, (ta,, x) = t&, and (2.5)
represent the same hyperplane. Hence the coefficients in It is important to note that in generalizing from (2.7) to
(2.7) as computed by (2.8)-(2.10) are not unique but rather (2.1 l), we have assumed that each component f. of f is
depend on the scale factor t used in representing the linear associated with the same linear partition in the domain.
partition. However, if we normalize the representation (2.5) This assumption, however, entails no loss of generality
by defining the scale factor2 since we can always introduce additional hyperplanes for
each fi until their linear partitions are identical for all’
i=1,2;.. YP.
Remark: We can write (2.11) in an even more compact
matrix form. Define
then we will get a unique representation for the linear
partition as well as a unique set of coefficients for the c~f[[C,rC2,. . .,c,] EIRflXP
canonical form off(.).
Remark: Nondegeneracy is only a sufficient condition Dder[a,,a2,. . .,ap] EIRnXP
for the existence of a canonical form representation. Exam-
ples exist (see [8]) such that a piecewise-linear function f ( .) and
can be represented in canonical form (2.7) even though its
associated linear partition is degenerate. We will show in
Section III that in formulating equations for piecewise-lin-
ear circuits, this sufficient condition can be ignored. then (2.11) can be written as3
f(x)=a+Bx+Cabs(Drx-e). (2.15)
2.3. n-Dimensional Vector Function: f: R ’ + Iwn Example 1:
Definition 2. I The linear partition shown in Fig. 3 is defined by
An n-dimensional function f: !Rn + IF2 n is said to be ([1 - l]‘, x) = 0 and ([l l]‘, x) = 0. A continuous
continuous piecewise-linear if all its components fi: piecewise-linear function f: Iw2 + !R2 is defined in each
R”+R’ j=l2 . . . , n are continuous piecewise linear. region as follows:
Any n-dimensional function f: R” -+ R” with its compo-
nents f;: R” + R’ represented in the canonical form (2.7) 4% ER,
can be represented globally by the following explicit ana-
lytical expression:
‘[ I
3x2 ’ x1, ~2

x,,x,~R,
f(x)=a+Bx+ f: cil(ai,x)-&I (2.11)
i=l
37~2~4
where x, u, ci E IL!n and B E R nXn. The coefficients are

*For the case /3, = 0 where sgn( pi) is undefined, we choose x1,-3=&
dif
def
[i= &i&y 3For x = [xl, x2,.. .,x,l’~ R”, abs(x) = [IxII,I.~~ .,lx,ll’
128 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 3, MARCH 1983

istics of Rl and R2 are approximated by continuous

piecewise-linear segments and represented in the canonical
form (2.1) as follows:

Rl: i, = g(u,) = a, + b,v, + 5 cli]o, -Vu]

i=l
R
R2:v,=r(i,)=a,+b,i,+ fFJ c2ili2 - I2jl.
j-1

KVL implies 0, + r+ +2i, = E and KCL implies i, = i,.

Fig. 3. In Example I. 58’ is partitioned by two one-dimensional hyper-
Therefore, v2 = r(i2) = r(i,) = r(g(v,)) and the equi-
PAT; (straight lines) x, =x2 and x, = - x2 into 4 regions R,, R,, R,, librium equation of the circuit is v, + r(g(v,))+2g(v,) =
4 E. We note that this equation is no longer in the canonical
form (2.1) since the composition of r( *) and g( .) causes the
2il absolute sign to be nested. On the other hand, the equi-
librium equation can also be formulated as follows: i, = i,
i,
+ implies
VI

2-E
+ i2
v,+v,+2i,=E and v,+v,+2i,=E.

T v2
Writing these equations in the vector form, we get
v,+r(i,)+2g(v,)
1II E
q
I I

Fig. 4. Circuit containing two piecewise-linear resistors. I 0, + r(i,)+2i, = E’

Substituting r(e) and g( .) into the above equation and
the associated Jacobian matrices are rearranging terms, we obtain

[ 1 03’
40 x,,+=R,

[ 1 312’1 x,,xz E%
J/(x,4 =
[iiu1
02’ x,,xz~R3
+c
i=l [1
p’ 2Cli
o lv,-~,il+~,[~~jli2-l;jl=[~].

-1
3 -1
2’ 1 X,7x2E%
Indeed, this equation is in the canonical form (3.11) pro-
vided that we identify x = [v, i21T, ai = [ 1 OIT, aj =
[0 llT, /3i=V,i, bj=12j for i=1,2;..,p,,j=1,2;..,p2.
using (2.12)-(2.14) to compute the coefficients, we can The above discussion shows that arbitrary elimination of
express f in the form of either (2.11) or (2.15): variables will not, in general, give rise to a system of
equations in canonical form, even though KVL and KCL
r 11 are linear equations and the constitutive relation of all
resistors are expressed in canonical form (2.1). However,
by proper formulation, it is possible to reduce the equa-
tions in canonical form.
Consider the class of nonlinear circuits made of linear
and nonlinear two-terminal resistors, linear and nonlinear
controlled sources (all 4 types), each depending on a single
controlling variable, dc independent sources (battery and
current sources), as well as any linear multiterminal resis-
r1 1 tive elements, such as ideal transformers gyrators, etc.
=
[I [ 1
0+3
0 0
0
2
“I
Xl
x2 +
T
z1
5
-- 1
2
Through equivalent circuit transformation techniques, vir-
tually any dc resistive circuit can be reduced to this class
[9], [lo]. Our objective in this section is to show that if we
approximate the characteristics defining each nonlinear
. abs
El-M)* 1
1 - .l
1
0 resistor or nonlinear controlled source by a continuous
piecewise-linear function and represent it in the canonical
form (2.1), then, for all conventional methods of circuit
III. CANONICAL EQUATION FORMULATION FOR
formulation, the equilibrium equation of the resulting cir-
PIECEWISE-LINEAR RESISTIVE CIRCUITS
cuit will always assume the n-dimensional canonical form
Consider the simple circuit shown in Fig. 4. Rl is voltage given in (2.11). We will prove this rather remarkable and
controlled and R2 is current controlled. The n-i character- unifying result for each of the following common equation
CHUA AND YING: CANONICAL PIECEWISE-LINEAR ANALYSIS 129

formulation methods: nodal analysis, mesh analysis, cut set

analysis, loop analysis, hybrid analysis, modified nodal analy-
sis, and tableau formulation.
3.1. Nodal, Mesh, Cut Set and Loop Equations in Canonical
Form
Theorem 3.1 (Nodal and cut set analysis):
Consider a connected resistive circuit N containing
only linear two-terminal resistors, dc independent L-------l
sources, voltage-controlled piecewise-linear two-termi- Fig. 5. A composite branch.
nal resistors, and linear or piecewise-linear voltage-
controlled current sources. If each piecewise-linear Define iq = [i, ;- . . . i^,]‘, then the above equation gives
2
characteristic is represented in the canonical form
(2.1), then the nodal equation and cut set equation of N
J, + a, + i (b,juj,Eq)
always assume the canonical form (2.11). j=l
Proof Observe that the assumptions assure the ex-
istence of the nodal equation. By applying source transfor- J2 + ~2 + i (&jUj>Eq)
‘:
mation, we can assume without loss of generality that each I = j=l
4
branch in N is in the composite form (see Fig. 5) where nk
can be a linear resistor, a voltage-controlled piecewise-
linear resistor, a linear voltage-controlled current source, or
Jq+aq+ i (bqjuj>Eq)
a piecewise-linear voltage-controlled current source. The j=l
vk - i, relation of nk takes the following forms:
(1) If nk is a linear resistor, then bljui’
i, = gp,. (3.1) b2ju/’ 1
(2) If nk is a piecewise-linear voltage-controlled resistor, +iY %
j=, 1
then
bqju,T
i, = ak + b,v, + g ckilvk - V,,l. (3.2)
i=l
+ i 5 cjiujl(uj~~q)-(~~-(uj,Eq))I
(3) If nk is a linear voltage-controlled current source, j=l i=l
then
i, = gkvj (3.3)
‘fd+&q+ 2 5 ~jil(uj~s>-(~i-(uj~Eq))l.
j=l i=l
where vj, j * k is the voltage of some other branch.
(3.5)
(4) If nk is a piecewise-linear voltage-controlled current
source, then Let A be the reduced incidence matrix associated with N
relative to an arbitrary datum node, and v,,pe the node-to-
i, = uk + gkvj + E ckilvj - V,,l. I (3.4) datum voltage vector, then KCL implies Ai, = 0 and KVL
i=l implies Gq= A~U,. Applying these two equations to (3.5)
Now let q be the total number of composite branches in we get
N,u,,k=1,2;.-, q, be unit vectors in R 4, and define
iTq= [t?, 6,. . . S,]r, E4 = [E, E2. * * E,lT, then since ik
=J,+i,,andv,=G,+E,fork=1,2;..,q,wecanwrite
(3.1)-(3.4) in the following general form:
=a+Bv,+ 2 5 c~~~(Au~,u,J-/~~~~=O (3.6)
j=l i-1
i, = Jk + ak + 5 (bkjuj, i$ + E,)
j=l where
pj def def
+ 5 C cjil(uj>6q+Eq)-~il
a = Ai, B = ABAT, ,jidzffA2ji, pJizf Pji - (uj, E,).
j=l i=l

(3.7)
Note that the nodal equation (3.6) is precisely in the form of
(2.11) provided we relabel the double indexes in the last
term.
+ i $, CjiJ(“j’~q)-(~i-(uj,Eq))l. To formulate the cut set equation of N, pick a tree and
j=l i=l choose the tree branch voltages v, as independent variables.
130 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 3, MARCH 1983

The nodal equation in the canonical form is

b,+b4 -b, -b4 el

+
I -4

- b4
b,+b,+b,
- b2
-b2
b, + b4 + b, I[ e3
e2

[I
Fig. 6. Bridge circuit containing 5 voltage-controlled piecewise-linear ‘Ii
resistors.
+, E -;I; ]e,-e,-V,,]
i=l

Replacing A in (3.6) by the fundamental

we obtain the following equation:
cut set matrix Q,
+FI“‘1r,i
le2
-e3
i=l
-V,,l+
fiti I 0

i=l
0
‘ii le, - V3,]

I1
‘4i
+E 0 ]e,-e3-V4i]+ 5 O le3-Vsil

1=I
- c4i i=l CSi
i I
=a+Bu,+ 5 $ c~~I(Qu~,v,)-/~~,I=O (3.8)
j=l i=l =TO1
0.
II
0

where 0
It is remarkable to observe that the resulting nodal
a=Qti.B=QBQ’,c,,=Q~ji,Pj,=t$i-(uj,Eq).
equation is in closed analytic form. Changing the resistor
(3.9) characteristics only changes the parameters in this equa-
tion.
Equation (3.8) is precisely in the form of (2.11) provided
Theorem 3.2 (Mesh and loop analysis):
we relabel the double indexes in the last term. 0
Consider a connected resistive circuit N containing
Example 2:
linear two-terminal resistors, dc independent sources,
Consider the bridge circuit containing five voltage-con-
current-controlled piecewise-linear two-terminal resis-
trolled piecewise-linear resistors, as shown in Fig. 6. As-
tors, and linear or piecewise-linear current-controlled
sume the o-i characteristics of these resistors are expressed
voltage sources. If each piecewise-linear characteristic
in the canonical form (2.1): is represented in the canonical form (2.1), then the
mesh equation and loop equation of N always assume
Ik = ak + bkvk + ? ck,luk - vk,l, k = 1,2,3,4,5. the canonical form (2.11).
i=l Proof: Dual of Theorem 3.1. 0
Shift the current source 1, in parallel with Rl and R3 so
3.2. Canonical Equation for Modified Nodal Analysis and
that the resulting circuit contains only 5 composite branches
Tableau Formulation
and 4 nodes. The reduced incidence matrix A with respect
to datum node @ is In the case when N contains both voltage and current
controlled elements, we may resort to modified nodal
1 10 00 analysis.
A= -1 0 0. 1 1 Theorem 3.3 (Modified nodal analysis):
I 0 -1 -1
1I 0 Consider a connected resistive circuit N containing
Computing the coefficients using (3.7) we obtain only linear two-terminal resistors, dc independent
sources, current-controlled and voltage-controlled
al-Is
aI -Is
piecewise-linear two-terminal resistors, linear and

1 1
a2 a, + a4 - I,
piecewise-linear controlled sources (all 4 types), and
a=A a3--IS = - a, + a2 + a3
any linear multi-terminal resistive elements. If each
a4 - a2 - a4 + a5 piecewise-linear function is represented in the canoni-
a5 _ cal form (2.1), then the modified nodal analysis of N
B = Adiag[b,b2b,b4b,]AT always assumes the canonical form (2.15).
Proof: Partition the elements in N into the following
b,+b4 - b, - b4 groups :
=
-b, b,+b,+b,
- b2 i, = G(v,)+N(i,) (3.10)
- b4 - b2 b, + b4 + b, I uE=M(vG)+S(iR) (3.11)
cji = cji Au/, B’, = L$, j=1,2,3,4,5. v, = R(iR) (3.12)
CHUA AND YING: CANONICAL PIECEWISE-LINEAR ANALYSIS 131

Defining
where G( .) in (3.10) includes all voltage-controlled resis-
tors, voltage-controlled current sources and dc current
sources; N(a) in (3.10) includes all current-controlled cur-
49,+ AGUN

rent sources; M( .) in (3.1I) includes all voltage-controlled

voltage sources; S( .) in (3.11) includes all current-con-
a=
/ Ia,+as
aR

trolled voltage sources and dc voltage sources; R( .) in AGBGAZ A, AR+A,BN

(3.12) includes all current-controlled resistors. B= -AT,+BMA; 0 4
Using the same procedure as in the proof of Theorem
3.1, we can express (3.10)-(3.12) into the canonical form 1
-A; 0 BR I
(2.15): A& 0 0 0 AGCN
i, = [a, + BGuc + C,abs (Dzuo - e,)] c= 0 CM c, 0 0
0 0 0 CR 0 I
+ [a,,, + B,,,i, + C,abs(D$, - eN)] (3.13) [

AGDG AGDM 0 0 0
v, = [a,,, + BMuc + C,abs ( DLnG - eM)]
D= 0 0 000
+ [as + B,i, + Csabs(@ii, -es)] (3.14)
1 0 0 DS DR DN I

vR=a,+BRi,+CRabs(D~ii,-eR). (3.15)
e= [ eg e; es' eg ec I’
Let A be the reduced incidence matrix of N relative to we can recast (3.21) in the following canonical form:
some datum node, and partitioned A as A = [A,IA,IA,].
Let J denote the source current vector and u, denote the a + Bx + Cabs( DTx - e) = 0. 0
node-to-datum voltage vector, then KCL and KVL give4 Theorem 3.4 (Tableau analysis):
Consider a connected resistive circuit N containing
A,i, + A,i, + A,i, = 0 (3.16)
only linear two-terminal resistors, dc independent
A$, = v,, A& = v,, A& = v,. (3.17) sources, current-controlled and voltage-controlled
piecewise-linear two-terminal resistors, linear and
Substituting (3.13), (3.17) into (3.16) and (3.17) into (3.14) piecewise-linear controlled sources (all 4 types) and
(3.15), respectively, we get any linear multiterminal resistive elements. If each
piecewise-linear function is represented in the canoni-
A,[a, + BGAEon + C,abs(DzAEu, - ec)]
cal form (2.1), then the tableau formulation always
+ A, [ aN + BNiR + C,,,abs (S,‘i, - eN)] assumes the form of (2.15).
+A.i,+A,i,=O (3.18) Proof: Let A be the reduced incidence matrix of N
relative to some datum node, and u,, be the node-to-datum
Agu, = [a,,,, + BMAzun + C,abs (DLAgu, - eM)] voltage vector, then KCL, KVL, and element constitutive
relations give
+ [as + BsiR + Csabs(Dlii, - es)] (3.19) Ai=AJ (3.22)
A& = uR + BRiR + CRabs(Diii, - eR). (3.20) v=ATvn+E (3.23)
Let x = [tfli~li~]T be the vector of independent variables f,(i)+ fv(v) = S (3.24)
and rewrite (3.18)-(3.20) into vector form, we get
using the same procedure as in the proof of Theorem 3.1,
we can express fr( 0) and f,,( .) in the canonical form (2.15)

I
AGBGA; A, A,+ AGBN
AGO, + 49,
a,+~~ + -A;+BMA; 0 B, f,(i)=a,+B,i+C,abs(DTi-e,) (3.25)
1 I x
1 aR 11 -A; 0 BR
fr,(u)=ai,+ B,v+C,abs(DLu-er,). (3.26)
Substituting (3.25) (3.26) into (3.24) and writing (3.22)-

1 .;y-
A& 0 0 0 AGCN
(3.24) into vector form, we get

+ 00 c,0 c,0 CR
0 00

. abs II AGDG

0
AGD,,,

0
0

00
DS
0

DR
0

ox-
DN 1 es=o.
T

eR
.eNd-

(3.21) (3’.27)
4Note: Current sources have been absorbed into i,. Clearly (3.27) is in the canonical form (2.15). 0
 132 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 3, MARCH 1983

Pk

vk = ak + b,i, + c ckili, - Ikkil, k=l+l,1+2;..,m.

i=l

voltoge- current- (3.30)

control led I inear controlled
piecewise fi m-port piecewise - Definingu,=[v, v,.~.v,]~, o,,=[v,+, v,+~...v,,,]~, i,
i i I
linear
CJ
linear = [i, i, . . . i,]‘, i, = [i,,, i,+z. . . i,]‘, and recasting
resistors + + + resistors
+ (3.29)-(3.30) into vector form, we get
-Vm
GP cnl
VP -
3
Fig. 7. Extracting all nonlinear resistors we obtain a resistive m-port N
containing only linear resistive elements and dc independent sources.
The ports attached to voltage-controlled (resp. current-controlled) resis-
tors are called voltage (resp. current) ports.
where
ii = [ a,a2. . . a,a,+, . . . %lT
3.3. Canonical Equation for Hybrid Analysis
B=diag[b, 6,.*-b, b,+,e..b,,,]
Theorem 3.4 (Hybrid analysis):
Consider a resistive circuit N containing linear two- y.i j=1,2;**,1
terminal resistors, dc independent sources, voltage- Pji =
( Iji j=l+l,Z+2;.*,m
and current-controlled piecewise-linear resistors,
linear-controlled sources (all 4 types) and linear and uj is the unit vector in lRm,j = 1,2,. . . , m.
multi-terminal resistive elements. Let fi denote the From Fig. 7 we have 8,=v,,il,=i,,k=l,2;..,m.
m-port obtained by extracting all voltage-controlled Hence fiO= v,, i0 = i,, Gb = u,,, ib = i,. Equating the right-
piecewlse-hnear resistors in N across “ voltage” ports, hand side of (3.28) and (3.31) we get
and all current-controlled piecewise-linear resistors
across “current” ports (Fig. 7). If the characteristic of
each piecewise-linear resistor is represented in the
canonical form (2.1), and if & admits hybrid represen- After rearranging terms, we get
tation (3.28), then
(a) the hybrid equation of N always assumes the U+BX+ 2 5 cjil(uj,x)-~jil=o (3.32)
canonical form (2.11); j-1 i-1
(b) the linear partition in the domain always forms where
a lattice structure in the sense that the partition
hyperplanes are parallel to the coordinate axes defined
by the independent variahles.5
x=[ ;;],a=ti-[?J],B=d[l: %],cji-,iuj.
Proof: Consider the linear m-port & shown in Fig. 7. Equation (3.32) is precisely in the form of (2.11) provided
Let t?,,tJ;.., 6, be the voltages of the voltage ports and we relabel the double indexes in the last term. This com-
pletes the proof of (a).
[I+ ,, l/12, . . . 9 ‘”, be the currents of the current ports. The
assumption implies that 3 has a hybrid representation: The partition hyperplanes associated with (3.32) are
defined by (uj, x) = pji, i = 1,2, . . . , pj. Since uj defines

[;I=[;: ;:I[ ;I+[;] (3.28)

the normal direction of the jth hyperplane in the linear
partition and since uj is a unit vector in Iw”, (b) follows
immediately. 0
where 60=[i), 02-.G,]T, 6 =[8,+, 6[+2.-.6m] T ) la= Remark: Theorem 3.4 can be easily generalized such that
” lJ
[1, i2 . . . &IT, ib = [i^,+, 1,+2. . . Z,]T, and [s^,iblT is the the piecewise-linear resistors extracted across the voltage
source vector. and current ports of N can be mutually coupled; i.e.
By assumption, each voltage-controlled piecewise-linear i, = f(v,, ib) and Us = g(u,, ib), where f(e) and g(e) are
resistor is represented by represented in the canonical form (2.11).
Example 3:
i, = ak + b,v, + g ckilvk - Vkil, k=l,2;.. ,I. Consider the resistor circuit shown in Fig. S(a) where Rl
i=l is voltage controlled with its v-i characteristic shown in
(3.29) Fig. 8(b) and R2 is current-controlled with its v-i char-
Similarly, each current-controlled piecewise-linear resistor acteristic shown in Fig. 8(c). The v-i characteristics of Rl
is represented by and R2 are represented as follows:

‘Any piecewise-linear equation having this property will henceforth be Rl: i, = $ + ivi -lv,l+ :Iv, - 11 (3.33)
called an “equation with a lattice structure”.
CHUA AND YING: CANONICAL PIECEWISE-LINEAR ANALYSIS 133

IV. CANONICALKATZENELSONALGORITHM
In this section, we adapt the canonical form representa-
tion to the Katzenelson algorithm. We state the conditions
which guarantee the convergence of this algorithm in terms
of the coefficients of the n-dimensional canonical represen-
tation. An alternate way of handling the familiar “comer”
problem is also discussed with an illustrative example.
Finally, we compare the bookkeeping complexity and com-
putational efficiency involved in equations with and without
1il
vz
lattice structures.
ml=* Definition 4.1
ml=-1
9=t A continuous mapping f: Iw” + [w” is said to be norm-
i2
-2 -I 0 I
-I
2 v, 2 10
-I
t 2
coercive if I] f(x)\\ + F as ]]x\] - 00.
mo= I The following theorem gives sufficient conditions for the
mo=l -2
e 7-k existence of solution in terms of the coefficients of f( -),
(b) (4 where f(e) is expressed in (2.15). The proof follows directly
from a corresponding theorem in [6].
Theorem 4.1
Let f: [w n --) Iwn be a norm-coercive continuous piece-
wise-linear function expressed in the form (2.15). If
there exists a point x,, E [w” which satisfies6
(1) DTx, - e f 0
(2) B(x - x0> f - C[abs(DTx - e)-abs(DTx, - e)]x E
R”, x =+x0
(3) det{B + C[sgn(DTx, - e)]DT} > 0
(4
then for any given y,, E Iw”, there exists an x E Iw”,
Fig. 8. Figures for Example 3. (a) Circuit containing two piecewise-
linear resistors and a controlled source. (b) u-i characteristics for such that f(x) = yO. Furthermore, the Katzenelson
piecewise-linear resistor R 1. (c) o-i characteristics for piecewise-linear algorithm starting from x0 converges to a solution in a
resistor R2. (d) N is obtained by extracting RI across the voltage port
and R2 across the current port. (e) Lattice structure defined by 3 finite number of steps.
one-dimensional hyperplanes (straight lines) o, = 0, o2 = 1 and i, = - 1,
each one parallel to either coordinate axis ot or i,. Note that all regions For complete generality, we will present the canonical
have the same regulur pattern-bounded or unbounded
rectangles- typical in a lattice structure. Katzenelson algorithm without assuming the equilibrium
equation has a lattice structure. However, remarks will be
given at each step where a lattice structure can bring
1 3 1.
R2: v2 = - 4 + Ti2 - 41~ + I]. (3.34) significant savings in bookkeeping and/or computation.

Extracting Rl across the voltage port and R2 across the Canonical Katzenelson Algorithm
current port (see Fig. S(d)), we obtain the following hybrid Consider the equation
representation for the remaining linear two-port 8:
f(x)dLfa+tx+Cabs(DTx-e)=yo. (4-l)
[:j= [ -: -.I[;;]+[;]. (3.35) Step 1: Choose an initial point x0 satisfying conditions
(l), (2) and (3) in Theorem 4.1. Let k = 1, h = 0, xck) = x0.
Substituting (3.33) and (3.34) into (3.35) and rearranging Let R, denote the region containing xck). R, can be
terms, we obtain the following hybrid equation in the identified by its sign-sequence vector (2.6). Got to Step 2.
canonical form (2.15): Step 2: Compute rick) by
3
0
dk)= [JJ’[&- f(x’k’)] (4.2)
4
where JR, is the Jacobian matrix of f in region
R,. Go to
-- Step 3.
O :
Step 3: If rick) is a zero vector, then xck) is the solution,
and the algorithm terminates here. Otherwise let Ack) be a
subset of {1,2,. . . , p} such that for i E Ack), (q, x) = pi is
def
The lattice structure of the linear partition is shown in Fig. ‘For x~lR”, sgn(x) = diag(sgn(x,),sgn(x,); .,sgn(x,)), where
8(e). 0 sgn(x,) = 1 when xi > 0 and sgn(x,) = - 1 when xi < 0.
134 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 3, MARCH 1983

a boundary hyperplane of region R,. Then for each i E A(“) Step 8: (the corner problem): 9 Since the tck)‘s are not
such that (cui, rick))) f 0, compute unique, let I ck) be a subset of A(k) containing ‘the indexes i
where the ttk)‘s are equal. Choose any i from Ick), let h = i,
I and compute x ck+‘) = xck) + t$in(k). Go to Step
t,(k) = Pi - CaiT Ick))
(4.3)
\ I
t(k)
mm = t^!k)
(cq, n(k)) 9.
and go to Step 4. Step 9: Check xck+‘) against the existing corner points
Remark: Since it is generally impossible to identify the stored in Step 10. If x(~+‘) is a new corner point then go to
boundaries associated with each region of (4.1), the num- Step 10. Otherwise go to Step 11.
ber of elements in set Ack) is always equal to p - 1 (i.e., all Step IO: Store the corner point ~(~1. Go to Step 12.
hyperplanes except the one which contains ~(~1). However, Step 11: Let N, ck) be the total number of elements in set
if the equilibrium equation has a lattice structure, then Ick). Construct a region list Ljk) which consists of all
each region has at most 2n boundaries (therefore, the size neighborhood regions of the corner point except region R,.
of A(“) is always 2n - 1 after the first iteration) and can be Remark: Since each region is uniquely identified by its
easily identified by simply comparing the coordinates of sign-sequence vector, Step II consists merely of generating
XC“) with the &‘s. Substantial computing time can be saved those sign-sequence vectors which represents the neighbor-
at this point since not only the total number of tjk) to be hood regions. Let w ck) denote the sign-sequence vector of
calculated is reduced, but also no multiplication is required region R,. We generate the sign-sequence vectors from
for the dot products in (4.3) because (Y~ is now a unit wck) by varying the sign at the jth position of wck) for
vector. Note that in general p is much greater than 2n, and j E Ick), while keeping the remaining entries identical.
p = 2n occurs only when each of the piecewise-linear char- Clearly this procedure will generate 2Njk’- 1 sign-sequence
acteristic has exactly one breakpoint. vectors.
Step 4: (a) If det JR, * 0, let A(k) be the subset of Ack) Step 12: Select a new region Rk+, from the list Lfk)
such that for all i E a(“) associated with the corner point xck+‘) in accordance with
the following precedence: (1) the sign sequence of R,, ,
sgn ( tik)) sgn (det JR,) = 1. (4.4) matches w(Z), (2) det JRk+, > 0, (3) det JR,+, = 0, (4)
If A(k) is an empty set, then xck) + rick) is the solution and detJRk+, x 0. Remove Rk+, from the list Lfk) and go to
the algorithm terminates. Otherwise, find $‘)dzf mini E A(!+ Step 13.
Remark: Step 12 and Step 13 form a loop until we
Ittk)l and go to Step 5.
identify the correct region to proceed. Here, w(a) is the
’(b) If det JR, = 0, then find sign-sequence vector computed in Step 13.
itk)
I = i Fik) Itik)l (4.5) Step 13: Repeat Step 2 and Step 3 (i.e., repeat (4.2) and
(4.3) with k replaced by k + 1) to compute .ck+‘) and
and go to Step 5. tfk+‘) for i 4 Ick+‘) (go through Step 7 first if det JR,+, * 0).
Remark: Again observe that if (4.1) has a lattice struc- Repeat Step 4 to find t^jk+‘), then compute f = xck+‘)
ture, then the size of A(k) will be smaller and searching for + ii!“’ ‘)nck+ ‘) and the sign-sequence vector w(Z). To
i!k)
I will also be easier. determine if region Rk+ , is the correct region to proceed,
Step 5: If Z(k) 2 1, then x ck) + rick) is the solution and the we compare w(i) with w ck+ ‘) (the sign-sequence vector of
algorithm terminates here. If ;I”’ < 1 and it is not unique,’ region Rk+ ,). If they are equal, (then R,, , is the correct
go to Step 8. If ti^ck) < 1 and it is unique, let h = i, t$!, = tjk) region to continue), set x(~+~) = ,6, JR,+, = J&;, and incre-
and go to Step 6. ment k by 2. Go to Step 2. If w(Z) and wck+‘) are not
Step 6: Compute equal, then go to Step 12.
x(k+ 1) = x(k) + t(k)n(k) Remark: Since most of Step 13 consists of repeating Step
mm (4.6)
2, Step 3, and Step 4, substantial savings in computing time
J Rk+,=J~k+2ChahT (4.7) will result if (4.1) has a lattice structure (recall the Remarks
where ch is the h th column in matrix C and (Ye is the h th following Step 3 and Step 4).
column in matrix D. Go to Step 7.
Remark: If equation (4.1) has a lattice structure, then OL,, Justification
in (4.7) will be a unit vector, and hence the dyad product (1) To show the Jacobian matrix can be computed from
of ch and al requires no multiplication at all. the one in the adjacent region by (4.7), let R,, R, be two
Step 7: Increment k by 1. If det JR, * 0, then go to Step adjacent regions and J,, J2 be the Jacobian matrices in RI
2. Otherwise, find a vector ti * 0 in the null space’ of JR,. and R, respectively. Then:
Let rick) = n^and go to Step 3.
J,=B+C[sgn(DTx(‘)-e)]DT (4.8)
‘We say ;tk! IS nor unique if there exist at least two indexes “WI” and J,=B+C[S~~(D~X(~)-~)]D~
‘01” such thk tCk) = ItCk)l = ItLk’l. (4.9)
‘The vector 6 is themnull space of JRk+, can be found using standard
linear system techniques. ‘Note that I(~+‘) is now a corner point.
CHUA AND YING: CANONICAL PIECEWISE-LINEAR ANALYSIS 135

5), and is not repeated here. However, we will illustrate the

method in Example 4.
Example 4 (Step-by-step illustration of the canonical
Katzenelson algorithm):
Consider the circuit shown in Fig. 9(a) where both
resistors Rl and R2 are voltage controlled. Their v-i
characteristics in Fig. 9(b) and 9(c) are represented in
canonical form:
(4
Rl: i, = f + iv, -Iv,l+ ZIu, -I]

ml=-Im2=+ I rn,=i
m: i,=-- ; + iv2 - flv2 + I].
-2 -I 0 I 2 VI -2 -i 0 I 2 vp
-I The canonical form equation for this circuit is found to be
mo=l -I

+I= ir-- mo= I

(4

.
VI
(4
;x”’ vp=-I Let EO = 1 so that y, = [ - $ g]‘. The linear partition in
/ this case consists of 3 straight lines 0, = 0, v7 = 1, and
f
,/l-i.-+,
x”‘zx 0
v,=o v,=I
v2 = - 1, as shown in Fig. 9(d). Define x-=
Step 1: Choose an arbitrary p
shown in Fig. 9(d).
, as I-01-v,IT.
Fig. 9. Figures for Example 4. (a) Circuit containin two piecewise- Step 2 & 3: (k = 1).
linear resistors and a controlled source. (b) o-i c a aracteristics for

I I II
piecewise-linear resistor RI. (c) u-i characteristics for piecewise-linear 1 -- 1 -20
-I
resistor R2. (d) Lattice structure in the et - oa plane. The dotted line y -1
path indicates the iteration goes from x(I) = xc, to xc2), xt3), and finally 8 13

ll I
,(*j =
converges to the true solution at (l/6, l/2). 7 13 33
1 - -
where xc’) and xc2) are arbitrary interior points in R, and
3 = I 12 26
R,, respectively. For continuous piecewise-linear functions, Since the region R, containing x0 is bounded only by the
first (v, = 0) and third ( v2 = - 1) hyperplane, we set A(‘) =
J,= J2+yaT
wo { 1,3) and compute ti’) and t$‘):
where y is a constant vector to be determined [ 11. Substitut-
ing (4.8) and (4.9) into (4.10), we get o-([l ol’,[-1 --;I=) 13
J,- J2=C’[sgn(DTx(‘)-e)-sgn(DTx(2)-e)]DT
t(‘)=
’ ([l o]=,[$ g]‘) =!%
(4.11)
since the sign-sequence vectors differ only at the ith posi-
tion for R, and R,, we have -1-([o l]‘.,[-1 -;I=) 13
sgn( DTx(‘) - e)-sgn(DTxc2)-e) “‘)= (i. l]T,[; $1’) =%
=diag(O,O;*. ,0,2,0;.. ,O)
where the number “2” is at the ith diagonal position. Step 4 & 5: (k = 1). Since det JR, = 13/6 > 0, we have
Therefore, (4.11) reduces to a(‘)={1,3}, ts=13/33, and h =3.

hence (4.7) is proved.

J, - J2 = 2ci4-.
Comparing (4.10) with (4.12), we identify y = 2ci and

(2) The method we used to solve the corner problem

(Steps 8-13) is fully justified in [6] (Theorem 3-Theorem
(4.12)

x(2) =I
Step 6: (k = 1). (4.6) mplie

-1

-- 3
2
-20

-
13
33
26
-- 13

-1
33
136 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 3, MARCH 1983

(4.7) implies Consider the same Example as in Example 4. However,

instead of x,, = [ - 1 - +I’, let us choose x0

42=[;
-:I+;[
j. ll=!f-g.
Step 7: (k = l).. det JR, = 4/3 > 0. Set k = 2 and go to
Step 2.
Step 2 & 3: (k = 2). (4.2) implies rz(*) = [fi $1 r. Set
A(*) = (1,2} and compute t{*) and t$*): A(‘) = (1 23)

o-(v oK[-g -11’) 52 o-([l o]r,[-1 +]r) 13

t(2) = t,(l) =
’ ([l o]=,[g ;I=) =85’ ([1 o]r,[Z F]‘) =20

1-p oK[-g -11’) 1*4 -I-([0 l]r,[-1 -J+l’) 13

tp =
([1 o]r,[% {I’) =“. t6’)= ([o llT,[2+ 2f+]‘) =20.

Step 4 & 5: (k = 2). Since det JR, = 4/3 > 0, we have Step 5: (k = 1). Since t,‘(‘)= ;j” = $j, they are not unique.
a(*‘={1,2}, t(*) =l, andh=l. Go to Step 8.

=[--1I-l[ 1
Step 6: (km: 2). (4.6) implies xc3) = [0 - (6)lT, and Step 8: (k = 1). I (‘) = { 1,3}. Set h = 1 and compute
(4.7) implies

-J*) -l 13;
O ,gotoStep9.
:’ +20 200 = -1
91
Step 9: (k = 1). This is the first time we hit the corner
Step 7: (k = 2). det JR, = 2 > 0. Set k = 3 and go to [0 - l]r. Go to Step IO.
Step 2. Step 10 & 11: (k = 1). IV/*’ = 2. Sign-sequence vector for
Step 2 & 3: (k = 3). (4.2) implies rrc3)= [i $1 r. Set region R, is rvJ”= [- 1 - 1 - l]T; construct Lf” =
Ac3) = {2,3} and compute ti3) and ti3): (wf’), wp, w,“)}, where wj” = [ 1 - 1 - 11’3 w2(‘ =) [ - 1
- 1 l]T, wj’)= [l -1 l]?
Step 12: (k = 1). Since det JW,l,=$, det JWil, = 4, and
l-((1 01r,[O -$]J
det J,,,,u= 2 are all positive, and since w(Z) is still unde-
p =
fined at this moment, pick an arbitrary region (say wf’)) to
([l o]r$ g]q =6 continue. Go to Step 13.
Step 13: (k = 1). Repeat Step 2. n(*)= [& $1’. Since
I(‘) =, (1,3), we only need to compute t$*) in Step 3.
-1-p 11qo -+lq 26
tp = -- l-(P OIT,[O -11’) 15
tp =
([O l]r,[; $]r) = 25’
([1 o]r,[$ $1’) =7

Step 4 & 5: (k = 3). det JR, = 2 > 0 implies A^c3)= (2). compute
Since ;i3’ = ti3) = 6 > 1 solution of (4.13) is given by

w(a)=[1 -1 11=.
The iteration procedure is now terminated. The solution Since wI(‘)=[-1 -1 - l]r is not equal to w(Z), go
path is shown in Fig. 9(d). 0 back to step 12 to select another region.
Example 5 (Step-by-step illustration of Step 8- 13: corner Step 12: (k = 1). Since w$” = w(Z), choose the region
__ renresented bv w!” and go to SteD 13.
problem):
CHUA AND YING: CANONICAL PIECEWISE-LINEAR ANALYSIS 137

TABLE I multidimensional function f: Iw” -+ Iw” can be approxi-

SUMMARY OF COMPUTATION FOR EXAMPLES 4,6,7, AND 8
mated to any desired accuracy over any compact region in
T”ld CPUtime (m neconds)‘~ aB” by a composition of functions of only one variable. This
in turn will allow us to construct a circuit model whose
only nonlinear elements are two-terminal resistors [lo]. It
follows, therefore, that any resistive circuit can be modeled,
to within any desired accuracy, by a system of piecewise-lin-
‘*Since UNIX is a time-sharing system, the actual time depends on the ear equations in canonical form.
system at that time. Hence we give only a range of the total CPU time. This result is particularly appealing from a circuit-theo-
retic point of view because all methods of circuit analysis
Step 13: (k = 1). Repeat Step 2 and Step 3: n(*) given rise to an equilibrium equation of the same form. In
=[i ?I’, tJ*) = 6, and f = [ 4 $1’. Since w(Z) = fact, for small circuits, such as the bridge circuit presented
[l - 1 l]r is equal to wj’), set xc3) = [4 s]‘, JR, = J+, in Example 2, it is now possible to derive highly efficient
increment k by 2 and go to Step 2. “canned computer programs” where the user need only
Step 2 & 3: (k = 3). (4.2) implies rzc3)= [ - 4 - 31 r. Set specify the breakpoints and slopes of the nonlinear resis-
Ac3) = { 1,2,3) and (4.3) implies tt3) = 3/2, ti3) = - 3/2, ti3) tors in the circuit. Such a program are ideally suited for
= 3/2. implementation in personal microcomputers. For example,
Step 4 & 5: (k = 3). det JRIR,= 2 > 0 implies AC31= { 1,3}. one can easily store in a single 5;” floppy disk, the canned
Since ;I” = 3/2 > 1, solution of (4.13) is given by program for analyzing piecewise-linear resistive circuits in
all standard configurations, e.g., ladder, lattice, bridge-T,
twin-T, etc.
In the case where the circuit has a hybrid equation, the
canonical equation always possesses a lattice structure. This
remarkable property allows a substantial saving in comput-
ing time in the canonical Katzenelson algorithm. In the
Remark: Two programs have been written to implement case where the circuit has multiple solutions, this property
the canonical Katzenelson algorithm on a VAX-11/780 allows the development of a special algorithm which
computer running a UNIX time-sharing operating system.” guarantees that all solutions are found.
These programs were written in the C programming lan- From an analytical point of view, the canonical form
guage in order to take the advantage of the UNIX environ- allows us to carry out algebraic manipulations on piece-
ment. Dynamic memory allocation has been used so that wise-linear circuits. This makes it possible to derive closed
the programs can adapt to any size of circuits. form solutions and sensitivity formulas for many prototype
One program exploits the special properties possessed by circuits.
equations having a lattice structure. The other program Finally, it would be desirable to compare the computa-
does not exploit this structure and is therefore more gen- tional efficiency between the canonical Katzenelson algo-
eral, but less efficient. Examples 4, 6, 7, and 8” are used to rithm and the recent piecewise-linear approach described
compare these two programs and the result is shown in in [12] and [13].
Table I. Note that every example in Table I has lattice
structure. By exploiting this structure, one can expect the APPENDIX
saving in CPU time to increase with the size of the circuit. A. Justification of Coefficients for Canonical Equation (2.1)
V. CONCLUDING REMARKS Forx*xi, i=1,2;..,p, (2.1)implies that
Since the dc circuit model of most electronic devices are
$j(x)=b+ 5 cisgn(x-xi) (A.1)
made of two-terminal nonlinear resistors and nonlinear i=l
controlled sources depending on a single controlling varia-
since mj = d/dxf(x)l,,,,,,,+,, (A.l) implies that
ble, the equilibrium equation of most dc electronic circuits
P
are automatically in the canonical form, upon approximat-
m,=b- c ci 64.2)
ing each nonlinear function by a one-dimensional piece-
i=l
wise-linear function.
In the more general situation when the nonlinearity
mi=b+ ici- i ci,
consists of coupled equations, they can often be approxi- i=l i=j+l
mated by combinations of a few functions of one variable.
In fact, from a theoretical point of view, we can invoke j=1,2;*.,n-1 (A.3)
Kolmogorov’s theorem which asserts that any continuous P

mP =b+ c ci. (A.41

“VAX is a trademark of the Digital Equipment Co., UNIX is a i=l
trademark of Bell Laboratories.
’‘Examples 6, 7, and 8 are listed in Appendix C. It is clear that (A.2) and (A.4) imply (2.2), (A.3) implies
138 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 3, MARCH 1983

(2.3), and finally (2.4) follows from (2.1) upon substituting

x = 0.
B. Justification of Coefficients for Equation (2.7)
We derive (2.8) first. For x E Iw” and (ai, x) * pi, i = I I
1,2;. . ,p, (2.7) implies Fig. 10. Figure for Example 6. Circuit containing three piecewise-linear
resistors.

vf(x)=b+ i ci[sgn((cxi,x)-&)]a, (AS) Similarly, let x E Rj- in (A.5), we get

i=l
(AS) implies vfb)I,ER,-= b+i~,ci[sgn((a,,x)-8i)]a,-cja,.
i*j
5 {vf(4L,,}= i b ’ (A-8)
j=l j=l
Multiplying both side of (A.7) and (A.8) by $ and sub-
+ J$, z~,C;[s@((ai,
X)-Bi)lailxtRj,
tracting, we obtain (2.9).
i Finally, substituting x = 0 in (2.7), we get (2.10). 0

which can- be simplified to C. Examples

Example 6:
Consider the circuit shown in Fig. 10. Piecewise-linear
resistors Rl and R3 are voltage controlled, piecewise-linear
resistor R2 is current controlled. Their canonical form
5 Ci[Sgn((aiv
X)-~i)lailxtRjm representation are given by
i=l

Therefore, to derive (2.8), it suffices to show Rl: i, = RIO, +61- $0, -61

j=l
f ( ,~,ci[sEn((ni,~)-~i)lail,~,,~)=O* CAe6) R2: t+ = kli, + II- $i, -51

Since we are only interested, in x E Rj,, for j = 1,2,. . . , k, R3:i,=o,-~lo,-ll-2lu,-2l-lu,-31.

the procedure described below will eliminate all bounded
regions as well as pseudo-unbounded regions. The associated circuit equation is represented in the canon-
Consider a very general case such that there are g groups ical form as follows:
of hyperplanes in the linear partition, each containing ng
parallel hyperplanes. For each group, we merge the ng
def
hyperplanes into one. (Note: we have k, = p -C;= ,nj + g
hyperplanes left.) Then move each of the k, hyperplanes in where
parallel to itself until all pseudo-unbounded and bounded
regions vanish. This is always possible since no two of the 11 1
k, hyperplanes are parallel to each other. Clearly there are
kdzf 2k, regions left and all of them are essentially un- B=Ol 1
bounded. Since each hyperplane ( (Y~,x) = pi, i = 1,2,. . . ,p,
divides R” into two regions whose sign-sequence vectors 0 0 -2
differ only at the i th position, we have
I. r-5 5
e
sgn((ai,xj))=O,
j=l
fori=l,L---,P 6 6
0
t 1
0

c= 0 02;
where xi E Rj, for j = 1,2, + . . , k. Equation (4.6) follows
upon interchanging the summation sign. 00 00
I
i sgn((ai, xj)-Pi)] =O, fori=1,2;‘-.,p.
110000
j=l

To derive (2.9), let x E Rj+ in (A.5), we get 0 110 0

VfbLR,+ =b+i$,ci[sgn((ai,x)-B,)]ai+cjaj. 000011

i-j I
(A-7) e=[-6 6 -1 5 1 2 31T.
CHUA AND YING: CANONICAL PIECEWISE-LINEAR ANALYSIS 139

Fig. 11. Figure for Example 7. Circuit containing two piecewise-linear

resistors.

The initial point is x0 = [7 6 - llT and the solution is

[O.l 2.2 2.871T. 0
+
+111:-:..
Example 7: IOV IOK
(4
Consider the circuit shown in Fig. 11. Rl and R2 are
voltage controlled. Their characteristics are given in the “5
canonical form as follows:
m4
i
Rl: i, = - y + iv, + :Iu, +I]- S]U, -2]i :]u, -51

- $1~~ - ll]- f]~, - 13]+2]v, - 151 (,)I bEmitter

(b)
Fig. 12. Figures for Example 8. (a) A 4-transistor circuit. (b) Simplified
Ebers-Moll model of an n-p-n transistor. (c) Piecewise-linear ap-
proximation of diode o-i characteristic: m. = 0, WI, = 2.153 X lo-‘,
+ ;Io2 + II- $Io2 - 3]- ;Io2 - 8]+ $10~ - 101 m2 = 2.666 x IO-‘, m3 = 3.765~ IO-*, mq = 8.603 X IO-*, m5 = 1.865
x lo-‘; V, = 0.306, V2 = 0.321, V, = 0.336, V4 = 0.351, Vs = 0.376.
I, = -3.33570322x lO-2 +9.32400146x IO-*u,
+lu, - 13]- $I~ - 16]+ $]02 - 181.
+ 1.25666608x 10-2]u, - V,I
The associated circuit equation is represented in the canon-
+2.23537270x 10-3]o, - V21+8.49354618x 10-3]o, - V3]
ical form as follows:
+2.41900658x 10-2]u, - 1/,]+5.02251145x lo-*]u, - V,I.

Example 8:
Consider the four-transistor circuit shown in Fig. 12(a).
Each transistor is modeled by a controlled source in series
where (Y~= [l OIT for i=1,2;..,6 and CQ= [0 llT for with a p-n junction diode, as shown in Fig. 12(b). The

c, =
II I
i = 7,8,. . . ,16.
07
s,c,=
-- 03
2
diode ID - VD characteristic is represented by a continuous
piecewise-linear function with 6 segments as shown in Fig.
12(c). The reader is referred to [l l] for the canonical
representation of the equilibrium equation. The initial point

I 1 [01,
is x0 = [l 0 1 1lT and the solution is [3.290404x 10-l
9 3.652363 x-lo- ’ 3.376266x 10-l 3.602442x lo-‘IT. q
-- 2
cg = 8 >+i=
2
REFERENCES
[ 1] J. Katzenelson, “An algorithm for solving nonlinear resistive net-
works,” Bell System Tech. J., vol. 44, pp. 1605-1620, Oct. 1965.
[2] T. Fujisawa and E. S. Kuh, “Piecewise-linear theory of nonlinear
networks,” SIAM J. Appl. Math., vol. 22, no 2, pp. 307-328, Mar.
.,I977.-.
[3] T. Fujisawa, E. S. Kuh,, and T. Ohtsuki, “A sparse matrix method
for analysis of piecewise-linear resistive networks,” IEEE Trans.
Circuit Theoty, vol. CT-19, pp. 571-584? Nov. 1972.
[4] M. J. Chien and E. S. Kuh, “Solving precewise-linear equation for
p,=-l,a=2,p,=5,P,=ll,Ps=l3,P,=15, resistive networks,” Circuit Theory and Applications, vol. 4, pp.
3-24, 1916.
pl=-8,~8=-5,~s=-3,~,o=-1,P,,=3,~,~=~~ [5] M. J. Chien and E. S. Kuh, “Solving nonlinear resistive networks
using piecewise-linear analysis and simplicial subdivision,” IEEE
& = 10, ,fJ4 = 13, & = 16, & = 18 Trans. Circuifs Sysi., vol. CAS-24, pp. 305-317, Jan. 1977.
[6] T. Ohtsuki, T. Fujisawa, and S. Kumagai, “Existence theorems and
The initial point is x0 = [ 16 201T and the solution is [I .5 a solution algorithm for piecewise-linear resistor networks,” SIAM
J. Math. Annl., vol. 8, no. 1, pp. 69-99, Feb. 1977.
1.S]T. Cl [7] L. 0. Chua and S. M. Kang, “Section-wise piecewise-linear func-
 140 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 3, MARCH 1983

tions: canonical representation, properties, and applications,” Proc. Network Theory (New York: McGraw-Hill, 1969) and coauthor of the
IEEE, vol. 65, pp. 915-929, June 1977. book Computer-Aided Analysis of Electronic Circuits: Algorithms and Com-
‘[8] S. M. Kang and L. 0. Chua, “A global representation of multidi- putational Techniques (Englewood Cliffs, NJ: Prentice-Hall, 1975). He has
mensional piecewise-linear functions with linear nartitions.” IEEE
Trans. Cir&its Syst., vol. CAS-25, pp. 938-940, Nov. 1978: also published many research papers in the area of nonlinear networks
[9] L. 0. Chua, “Device modeling via basic nonlinear circuit elements,” and systems. He was the Guest Editor of the November 1971 Special
IEEE Truns. Circuits Sys.,, vol. CAS-27, pp. 1014-1044, Nov. 1980. Issue of the IEEE TRANSACTIONS ON EDUCATION on “Applications of
[lo] L. 0. Chua and P. M. Lm, Computer Aided Analysis of Electronic Computers to Electrical Engineering Education,” and the Editor of the
Circuits: Algorithms and Computational Techniques. Englewood IEEE TRANSACTIONSON CIRCUITS AND SYSTEMSfrom 1973 to 1975.
Cliffs, NJ: Prentice-Hall, 1975. Dr. Chua is a member of Eta Kappa Nu, Tau Beta Pi, Sigma Xi. He
[I I] L. 0. Chua and R. Ying, “Finding all solutions of piecewise-linear was a member of the Administrative Committee of the IEEE Society of
equations”, U.C. Berkeley ERL Memo UCB/ERL M81/54, July Circuits and Systems from 1971 to 1974, and the past President of the
23; 1981.
[ 121 S. N. Stevens and P. M. Lin, “Analysis of piecewise-linear resistive IEEE Society on Circuits and Systems. He has been awarded four patents
networks using complementary pivot theory,” IEEE Trans. Circuits and is a recipient of the 1967 IEEE Browder J. Thompson Memorial Prize
Sys~, vol. CAS-28, pp. 429-441, May 1981. Award, the 1973 IEEE W.R.G. Baker Prize Award, the 1973 Best Paper
[ 131 W. M. G. van Bokhoven, “Macromodelling and simulation of Award of the IEEE Society on Circuits and Systems, the Outstanding
mixed analog-digital networks by a piecewise-linear system ap- Paper Award at the 1974 Asilomar Conference on Circuits, Systems, and
proach,” m Proc. 1980 IEEE Int. Conf. Circuits and Computers. Computers, the 1974 Frederick Emmons Terman Award and the 1976
Miller Research Professorship from the Miller Institute.

+
Leon 0. Clma (S’60-M’62-Sh4’70-F’74) was
born in the Philippines on June 28, 1936. He
received the B.S.E.E. degree from Mapua In- Robin L. P. Ying (S’75-M’82) was born in Taipei,
stitute of Technology, Manila, the Philippines, Taiwan, The Republic of China in 1951. He
the S.M. degree from Massachusetts Institute of received the B.S. degree in electrical engineering
Technology, Cambridge, in 1961, and the Ph.D. from National Taiwan University, Taipei, Tai-
degree from the University of Illinois, Urbana, in wan, the Republic of China; the M.S. degree in
1964. engineering and applied science from Yale Uni-
He worked for the IBM Corporation, versity, CT, in 1976 and the Ph.D. degree in
Poughkeepsi, NY, from 1961 to 1962. He joined electrical engineering and computer sciences from
the Department of Electrical Engineering, Purdue University of California, Berkeley, in 1982.
University, Lafayett, IN, in 1964, as an Assistant Professor. Subsequently, He received the Yale Fellowship at Yale Uni-
he was promoted to Associate Professor in 1967, and to Professor in 1971. versity from 1975 to 1976 and the University
Immediately following this, he joined the Department of Electrical En- Fellowship at U.C. Berkeley from 1978 to 1979. Since June 1976, he has
gineering and Computer Sciences, University of California, Berkeley, been Research Assistant and Teaching Assistant at the EECS Depart-
where he is currently Professor of Electrical Engineering and Computer ment, U.C. Berkeley. He is now with Bell Telephone Laboratories in
Science. His research interests are in the areas of general nonlinear Holmdel, NJ. His academic research interests, include computer-aided
network and system theory. He has been a Consultant to various elec- circuit analysis, design optimization, and nonlinear circuit and system
tronic industries in the areas of nonlinear network analysis, modeling, and theory. At Bell Labs, his work involves developing analysis tools for
computer-aided design. He is the author of Introduction to Nonlinear design process of facility networks.

Coherency Identification for Power System

Dynamic Equivalents
FELIX F. wu, MEMBER, IEEE, AND NATARAJAN NARASIMHAMURTHI, MEMBER, IEEE

Ahstroct -Coherency-based approach to. the problem of constructing tion, to identify coherent groups directly from system data is developed. A
power system dynamic equivalents has been very successful in applications. physical interpretation of the algebraic characterization of coherency is
The key step in this approach is to identify groups of coherent generators. presented, where the condition for coherency is described in terms of
An analytic study of coherency is conducted. An algebraic characterization generator inertia constants and their equivalent admittances to the bus of
of coherency is given. An algorithm, based on the algebraic characteriza- the disturbance.

Manuscript received August 30, 1977; revised September 27, 1982. I. INTRODUCTION
F. F. Wu is with the Department of Electrical Engineering and Com-
puter Sciences and the Electronics Research Laboratory, University of N THE ANALYSIS of an interconnected power sys-
California, Berkeley, CA 94720.
N. Narasimhamurthy is with the Department of Electrical and Com- I tem, normally one is interested only in the responses in
puter Engineering, University of Michigan, Ann Arbor, MI 48109. a portion of the system, called the study system. The rest of
0098-4094/83,‘0300-0140$01.00 01983 IEEE