Chua 86

1072 IEEE TRANSACTIONSON CIRCUITSAND SYSTEMS,VOL.CAS-33,~0.
11, NOVEMBER1986
IEEE TRANSACTIONSON CIRCUITSAND SYSTEMS,
VOL. CAS-33, NO. 11, NOVEMBER1986 1073
The Double Scroll Family

LEON 0. CHUA, FELLOW, IEEE, MOTOMASA KOMURO, AND TAKASHI MATSUMOTO, FELLOW, IEEE
Absrrucf -This paper provides a rigorous mathematical proof that the crucial way to prove the existence of a Shilnihov-type homoclinic orbit,
double scroll is indeed chaotic. Our approach is to derive a linearly thereby establishing rigorously the chaotic nature of the double scroll.
equitdent class of piecewise-linear differential equations which includes These analytical expressions are also fundamental in our in-depth analysis
the double scroll as a special case. A necessary and sufficient condition for of the birth (onset of the double scroll) and &&I (extinction of chaos) of
two such piecewise-linear vector fields to be linearly equivalent is that their the double scroll.
respective eigenvalues be a scaled version of each other. In the special case The unifying theme throughout this paper is to analyze the double scroll
where they are identical, we have exact equivalence in the sense of hear system as an unfolding of a large family of piecewise-linear vector fields in
conjugacy. R3. Using this approach, we were able to prove that the chaotic dyrumics
An explicit normalform equation in the context of global bifurcation is of the double scroll is quite common, and is robust because the associated
derived and parametrized by their eigenvalues. Analytical expressions for horseshoes predicted from Shilnikov’s theorem are structurally stable. In
various Poincur~ maps are then derived and used to characterize the birtli fact, it is exhibited by a large family (in fact, infinitely many linearly-
and the &4t/1 of the double scroll, as well as to derive an approximate equideni circuits) of vector fields whose associated piecewise-linear dif-
one-dimensional map in analytic form which is useful for further bifurca- ferential equations bear no resemblance to each other. It is therefore
tion analysis. In particular, the analytical expressions characterizing vari- remarkable that the normalized eigenvalues, which is a local concept,
ous La/f-return maps associated with the Poincare map are used in a completely determine the system’s global qualitative behavior.
Part I: Rigorous Proof of’chaos

I. INTRODUCTION Proving a circuit is chaotic is a nontrivial task. Indeed,
HE double scroll is a strange attractor recently ob- only four nonlinear circuits have so far been proved rigor-
T served from a physical electronic circuit made of four
linear circuit elements(one resistor, one inductor, and two
ously to be chaotic: the first three circuits [lo], [ll], [30]
are described by a one-dimensionaldiscretemap while the
capacitors) and a two-terminal nonlinear resistor char- fourth circuit [12] is described by a second-ordernonau-
acterized by a five-segment u-i curve [l]-[3]. The nonlin- tonomousdifferential equation.The double scroll system to
ear resistor can be realized in the laboratory by several be studied in this paper is described by a third-order
equivalent electronic circuits using two op-amps [2], one autonomous differential equation. In particular, we will
op-amp and two diodes [3], or two transistors and two choose the dimensionlessform given by (2.4) of [3], which
diodes [4]. Since its recent discovery, this rather simple we rewrite in the equivalent form
electronic circuit has been observed, both experimentally
[5], [6] and by computer simulation [6], to exhibit a surpris- a=a(y-h(x))
ingly rich variety of bifurcation phenomena[6] and routes )i=x-y+z (1.1)
to chaos [7]-[9]. Although the chaotic nature of the double , i=-py
scroll appears to be very convincing from both experimen-
tal analysis and computer simulations, there remain where
legitimate objections from some critics who demand no
less than a rigorous mathematical proof. Our main objec- h(x)hX+f(x)=mIx+f(m,-m,)[lx+!i-ix-111
tive in this paper is to supply such a proof.
0.2)
Manuscript received March 4, 1986; revised June 19, 1986. This is the canonical piecewise-linear equation [13] describing
research was supported in part by the Office of Naval Research under an odd-symmetric three-segment piecewise-linear curve’
Contract NOOO14-86-K-0351, the National Science Foundation under
Grant ECS-8313278, the Japanese Ministry of Education, the Saneyoshi having a breakpoint at x = - 1 and x = 1, a slope equal to
Foundation, the Murata Foundation, the Mazda Foundation, and the m,~a+l<Oattheinnersegment,andml~b+l~Oat
Institute of Science and Engineering of Waseda University.
L. 0. Chua is with the Electronics Research Laboratory, University of
California, Berkeley, CA 94720.
M. Komuro is with Numazu College of Technology, Shizuoka 410,
Japan. ‘We include only three segments of the five-segment piecewise-linear
T. Matsumoto is with Waseda University, Tokyo 160, Japan. u-i curve because the two outermost segments do not play any role in
IEEE Log Number 8610179. the formation of the double scroll.
009%4094/86/1100-1073$01.00 01986 IEEE

1074 IEEE TRANSACTIONSON CIRCUITSAND SYSTEMS,
VOL. CAS-33, NO. 11. NOVEMBER1986
the outer segments,respectively; namely, ear vector fields is extremely difficult if not impossible.
h(x)=m,x+(m,-m,), XZl Moreover, whereas “normal form analysis” of smooth
vector fields [9] yields only local qualitative results, our
= m,x, 1x1<l analysis applies globally.
.= mix - (m, - m,), X<-1. (l-3) Our results from Section III provide the necessary
Note that (1.1) is slightly. simpler than (2.4) in [3] because foundation in Section IV for deriving the exact parametric
h(x) includes both f(x) and x. The double scroll systemis equations describing various Poincare maps of an im-
portant class of vector fields which representsan unfolding
therefore described by four parameters {(Y,p, m,, m,},
of our double scroll equation. These results are then used
with the double scroll attractor occurring in a neighbor-
in a crucial way in Section V to prove that homoclinic
hood of {9,14(2/7), - l/7,2/7}.
Since the techniques and concepts to be used in proving orbits of the Shihnkov type [9] exist in the double scroll,
that the double scroll is chaotic are quite novel and gen- thereby providing a rigorous proof that the double scroll is
eral, we will develop our theory for a much larger class of indeed chaotic.4
piecewise-linear differential equations of which (1.1) is a The analytical formula for PoincarCmaps in Section IV
special case.Mathematically, our approach is to derive and allows us to derive the exact coordinates of the return map
analyze an unfolding of the double scroll equation (l.l), of any trajectory of the double scroll system.Thesecoordi-
which has four parameters, into a family of three-dimen- nates are used in Section VI to derive the analytical
sional continuous piecewise-linear vector fields char- expression describing the image of severalstrategic loci (to
acterized by six parameters.However, unlike the literature be defined in Section VI) which allows us to explain the
on unfoldings which considersonly differentiable functions birth (i.e., onset) and the death (i.e., extinction) of the
[14], our results are novel in the sensethat our functions double scroll attractor. Unlike the preceding five sections,
are required to be only continuous, not differentiable.2 however, where complete mathematical rigor is achieved,
Because of the nature of piecewise-linear analysis, a some reasonable ,numerical calculations are used in this
section to calculate two curves-called the birth and the
substantial amount of symbols and notations are necessary
death loci-which bound the region in the LX-P parameter
to avoid ambiguity and clutter. They are summarized in
space where the double scroll exists.
Section II ,for easeof reference.
Finally, in Section VII, we derive the analytic expression
The family of piecewise-linearvector fields which can be
of an “approximate” one-dimensionalPoincare map which
interpreted as an unfolding of the double scroll system is
can be used for further bifurcation analysis of the double
defined and characterized in Section III. The main results
scroll. In particular, this one-dimensional map is used to
in this section are summarizedin Theorems 1, 2, and 3. In
map out various regions (using different colors) in the LX-/~
particular, we have derived the necessaryand sufficient
parameter plane which exhibit different qualitative behav-
conditions for any two vector fields in this family to be
linearly conjugate, which is a strong form of equivalence iors. It is also used, in a crucial way, to generate a
“period-doubling” bifurcation tree for the double scroll
from the circuit theoretic point of view and an important
system and to calculate the associatedFeigenbaum num-
mathematical property in the theory of structural stability
ber.
of vector fields [9]. It is remarkable that while it is often
impossible to establish any topological conjugacy between
II. PIECEWISE-LINEAR GEOMETRY AND ITS
nonlinear vector fieids, we were able to prove that the
REAL JORDAN FORM
necessary and sufficient conditions for linear conjugacy
(which is a special case of topological conjugacy) between Unless otherwise stated, vectors and matrices are de-
two piecewise-linear vector fields in our family is that their noted by lower and upper case bold-face letters, respec-
eigenvaluesin corresponding regions be identical. tively. Vectors in R3 are denoted by x = (x, y, z)r. Real
This important result, which is stated in two equivalent and imaginary parts of a complex eigenvalue will be de-
forms (Theorems 1 and 2) allows us to derive the explicit noted by (I and w, respectively. Real eigenvalueswill be
form of all members of our family of piecewise-linear denoted by y. Vector fields will be denoted by 5: R3 + Iw3.
vector fields which are equivalent (i.e., linearly conjugate) Hence, t(x) denotes the vector field evaluated at x and is
to each other in terms of their eigenvaluesalone. This therefore itself a vector in R3 emanating always from the
major result, which is formulated in the form of a canoni- origin 0, unless otherwise stated.
cal piecewise-linear equation [13] parametrized by their We will now extract the essentialproperties of the vector
eigenvalues, will henceforth be called the normal form field associated with the double scroll equation (1.1) to
equationfor the doublescroll.3Again, this result is remark- define the following generalizedfamily of vector fields .Y.
able becausefinding normal forms of parametrized nonlin- Definition 2.1. Piecewise-LinearVector Field Family 9:
We define 9 to be a family of continuousvectorfields
[: [w3 + R 3 satisfying the following properties.
*Our techniques necessarily differ from those used in unfolding
“smooth” vector fields and represent, in fact, a new approach for
unfolding other continuous piecewise-linear vector fields.
3The term “normal form” is used here in the same context as that used 4The reader is referred to an interesting related work by Mees and
in global bifurcation theory of vector fields [9], and not in the circuit- Chapman [15], where they used optimization techniques to locate a
theoretic sense of a state equation. heteroclinic orbit in the double scroll system.
1075
CHUA et al.: DOUBLE SCROLLFAMILY
A=LdIL,
B=L,flL,
C=E’(O)nU*
D=E’(P+)nUv
E=LdlLz
F=(xcLz:f(x)lLzt
Fig. 1. Egenspaces of the equilibria and related sets.
(P-1) [ is symmetric with respect to the origin, i.e.,5 Notations Associatedwith Fig. 1
Q- & - y, - z> = - 5(x, Y, z). For each vector field c E 9, define7
(P.2) There are two planes U, and K, which are E ‘(0) p 2-D eigenspacecorresponding to complex eigen-
symmetric with respect to the origin (i.e.,
value gof jijo at 0,
(x, y, z) E lJ, iff (- x, - y, - z) E U-,) and they E’(O) k 1-D eigenspacecorresponding to real eigenvalue
partition R3 into three closed regions D,, Do,
and D-,, as shown in Fig. 1. Here, the reference Y. at 0,
E ‘( P+ ) d 2-D eigenspacecorresponding to complex ei-
frame for (x, y, z) is arbitrary.
genvalue 6i + jG, at P+,
(P.3) In each region Di, (i = - LO, l), the vector field < E ‘( P+ ) p 1-D eigenspacecorrespondingto real eigenvalue
is ,affine, i.e.,
$ at P+,
L$U,nEC(0), L,%Y,nE’(P+)
D~‘(x, Y, Z) =Mi, for (X7 Y, Z> E Di
L, g (x E u,: Wll~I} (2-U
where D.$ denotes the Jacobian matrix of t(x) where 11reads “is parallel to.” Here, 5(x)llU, means the
and ikfi denotesa 3 x 3 real constant matrix. vector [(CC) lies on a plane parallel to U,. That L, is a
(P-4) 5 has three equilibrium points, one at the origin straight line in Fig. 1 follows from the following.
0, one in the interior of D, (labeled P’) and ktraight Line TangencyProperty
one in the interior of D-, (labeled P- ). Let [ be a linear vector field in R3 having a pair of
(P-5) Each matrix Mi has a pair of complex conjugate complex conjugate eigenvalues a”f j3 and a real eigen-
eigenvalues (labeled I$,f jG, for M, and a”,f value y. Let U denote any plane which is not parallel to
jij, for M-r and M,, where 5, > 0 and G5,> 0) each eigenspace and which does not pass through the
and a real eigenvalue(labeled y0 for M, and T1 origin. Then
for M-i and M,, where j$, # 0 and j$ # 0).
(P-6) The eigenspaceassociatedwith either the real or LA {xH: 5”(x)~~U} (2.2)
the complex eigenvalue@ at each equilibrium
point is not parallel to Vi or U-i. is a straight line.
Proof: In Appendix I, we prove the above assump-
tions imply that there exists a suitable coordinate system
‘To avoid clutter, we will often use row vector (x, y, z) in place of
x’ g (x’, y’, z’) in R 3 such that t is transformed into the
column vector (x, y, z)r.
61n the case where the eigenvalue is complex,the eigenspace is defined
to be the vector space spanned by the real and the imaginary part of the 7Here, superscripts “c” and “r” denote “complex” and “real”, respec-
complex eigenvector. tively.
1076 IEEE TRANSACTIONSON CIRCUITSAND SYSTEMS,
VOL. CAS-33, NO. 11, NOVEMBER1986
real Jordan form [16] of the new coordinate system,we have*

4 \k,(o) = 0 (2.6)
!l$(u,)=yo~ {(x,y,z): x+z=l} (2.7)
and such that the equation for U in the new coordinate

\k,(U-,)=V,-A {(x,y,z):
-10 x+z=-1) (2.8)
[ 1
system assumesthe following simplified form:
;D$,(S(*;‘x)) =&,(x) 4 : a0 0 x (2.9)
u= {(x’, y’,z’): x’+ z’=l}. (2.4)
0 0 Yo
For each x E L, (2.2) implies that the vector dot product
(t%“(x),h) = 0, w here h P (l,O,l)r is a normal vector to U where a0p Eo/3, and y. A ~o/Oo.
in view of (2:4). Substituting [(x) from (2.3) into the
above vector dot product, and solving for y’, we find that b) \k,(P) = 0 (2.10)
L in (2.2) is a straight line defined by the equations !P1(ul)=vl~ {(x,y,z): xiz=l} (2.11)
L: y’=ax’+y(l-x’), z’=l-xx’ (2.5)
[ 1
-‘I 0
where u g I?‘/; and y A y/G7. n ;D$(E(*[‘(x)) =[,(x) p : u1 0 x (2.12)
Remark: The above straight line L intersects the line 0 0 Yl
{(x’, y’, z’): x’=l, z’= 0} at the’ point (x’, y’, z’) =
(1, u,O). where ui k a”,/~& and yi ~~,/3,. We will henceforth call
We are now ready to define the following important (2.9) and (2.12) the normalizedJordan form of MO and M,,
points in Fig. 1: respectively.
Definition 2.3. D, Unit and D, Unit of 5:
APL,nL,, BkL,nL, We define the set {to, Vo,\k,} as the Do unit of 5 and
the set of { tl, V,, qk,} as the D, unit of 5.
C~U,nE’(O), DAU,nE’(P+) Geometrically, the Do unit of 5 is simply the middle
region Do in its new referenceframe (x’, y’, z’), which we
EkL,nL,, FA {xEL~: <(x)((L,) labeled simply as (x, y, z) in Fig. 2. It is important to keep
in mind, however, that these two referenceframes involve
where <(x)11L, means the vector E(x) lies on Vi and is
different coordinate systems.
parallel to the straight line L,. These points are well
defined becauseit can be proved that no two lines among The images of the important points A, B, C, D, E, and
F in Fig. 1 will be denoted by corresponding subscripts in
L,, L,, and L, are parallel to each other.
an obvious way’: j
For simplicity, we will often suppressthe superscript +
and write P instead of P+. The following strategic points
Do: A,A\k,(A), B,A\k,(B), Co%‘o(C),
play a crucial role in Section III.
Definition 2.2. FundamentalPoints of 5: Dop’I’o(D), E,%\k,(E), Fobqo(F)
The four points A, B, E, and P defined above are
called the fundamental points of 6. D,: A,+,(A), B,+,(B), Cl%‘l(C),
Note that the continuity of the vector field 5 implies that
D+&(D), E+‘&(E), F,+,(F).
Our next goal is to derive the coordinates of each of

these points in their new referenceframes. Since A, B, C,
D, E, and F are located on various intersection lines in
Fig. 1, their images (under any affine transformation) must
lie on corresponding lines in the new reference frames.
In general, each 3 x 3 matrix Mj defining a vector field
These lines are images of intersections between various
.$E A?’in region Di requires nine nonzero parameters.Our
eigenspaces(E ‘(0) or E ‘(0)) with the plane U, in Fig. 1. In
next objective is to eliminate as many of these parameters
as possible by reducing Mi to its real Jordan form and
U,, to its simplified form. ‘Strictly speaking, we should use x’ and x” to denote vectors in the
Let q,,: D, + R3 and I,: D, + R3 denote the ap- new coordinate systems. as in (2.3) and (2.4). However, we will hence-
forth suppress the primes and double primes to avoid clutter. Since we
propriate affine transformations which reduce M, and Mi will be dealing mostly with the new coordinate systems in the following
to the real Jordan form in (2.3) while simultaneously sections, no confusion should arise.
transforming the equation describing U, 1 to the simplified ‘Note that the same symbols D,, and D, are used to denote a region in
Fig. 2(a) and a point in Fig. 2(b). There wrll be no confusion, however,
form in (2.4). It follows from (2.3) and (2.4) that in terms since its meaning will be clear from the context.
CHUA et al.: DOWti SCROLLFAMILY
unstrblo
l lgsnspsco
stsbls EC(P)
eigenvocfor z ID,-unit
E’(P)
‘1 (E’(P))
t
olrblo
l lgonspaco
EC(O)
/
unstrblo
rlgsnvrctol
E’(O) .A
(a) (b)
Fig. 2. Geometrical structure and typical trajectories of the original piecewise-linear system and their images in the Do u$t
and D, unit of the transformed system (real Jordan form). (a) Original system and typical trajectories. (b) Do and D, umts
and half-return maps.
particular, it can be shown that Fig. 2 and obtain

\k,(E”(O))= {(x,y,z): z=O},i.e.,thex-yplane E,F;=k,F,B; (2.19)
(2.13) where k, is a scaling constant.
*o@‘(O)) = {(x, Y, z): x = y = 0)) i.e., the z-axis For future reference, the explicit coordinates for the
(2.14) image of all strategic points in Fig. 1 are tabulated below.
Strategic Points in D, Unit (a, 2 ~&b/3,, y0 k ~0/~0)
qo(Lo)= {(x,y,z): x=1, z=o} (2.15)
A, = (1, po,O) (2.20)
\k,(L,)= {(x,y,z): y=uox+yo(l-x), z=l-x}.
(2.16) where”
Since C = E'(O)n U,, ii follows from (2.14) and (2.7) po~uo+~(u~+l), koAyo(po-uo)/(u;+l) (2.21)
that Co = (O,O,1).
Since E = Lo n L,, it follows from (2.15) and (2.16) that
B,=(Y~(Y,-~~-P~)/Qo~Y~~~-P~(~o-Yo)~/Qo~
Eo = (1, uo, 0).
Since F E L, and ,$(F)IIL,, it follows that F, E \k,( L2) l- ~o(~o- uo- po)/Qo) (2.22)
and Eo(Fo)((\ko(Lz). Hence, the coordinate of F. must
satisfy where
y=uox+yo(l-x), z=l-x, QoA (q,-~o)~+l
00x
-= - Y x +uoy
-=- YOZ
(2.23)
(2.17) co = (O,O,l)
1 0, - y. -1 .
E, = (1, ~08) (2.24)
Since A, lies on the line \k,(L,), we can write A, =
(1, po,O) for some p. E R.
6 = (~o(~o -2uo)/Qo, YO[~- uobo - vo>l/Qo,
Since B = L, n L, and .$(B)jIL,, the.co*ate of B, is
determined by B, E ‘k,(L,) and ~o(Bo)~~BoAo, where the (62 + We,). (2.25)
“arrow” denotes the vector from B, to A,. Since B,, E,,
and F, all he on the line Fo(L,), it follows that Strategic Points in Dl Unit (a, % &l/G,, yl p Y,//r3,)
m=k,&i$ (2.18) A, = (1, pl,O) (2.26)
where k, is a scaling constant.
Similarly, we can derive the coordinate of A,, B,, D,, “The two expressions in (2.21) (resp., (2.27)) are equivalent to each
E,, and Fl in the new reference frame for the D, unit in other. The value of k, (resp., k,) is specified in (2.33).
1078 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-33, NO. 11, NOVEMBER 1986
where other by a linear transformation. We will now define these

two concepts precisely.
p1 A (~1+ k,( 01’ +l)/vl, k, A ul( PI- h’(~: +I)
Definition 3.I. Linear Equivalence:
(2.27) Two vector fields E and 5’ in 08” are said to be linearly
B, = (1, q,o) (2.28) equivalent iff there exists a nonsingular linear transforma-
tion G: Iw’ + Iwn and a real number v > 0 such that”
D, = (WA 1) (2.29)
G+=v(@G). (3.1)
4 = (YI(YI - 01 -P&I, Y&- P~J, - udl/Q,~
Definition 3.2. Linear Conjugacy:
l- Y1(Y1- (J1- P&QJ Two linearly-equivalent vector fields are said to be lin-
(2.30)
where early conjugate of each other iff v =l in (3.1).
The concept of linear conjugacy is a special case of the
Q+ (u,-Y,)~+~ (2.31) well-known concept of topologicalconjugacy[9] where the
“linear transformation” is replaced by a “homeomor-
4 = (~1(~1--2d/Q,, Y&- du.1- ul)l/Ql,
phism.” In general, it is extremely difficult if not impossi-
(01’ + 1)/Q,). (2.32) ble to prove two nonlinear vector fields are topologically
Note that k, cannot be calculated directly from (2.21) conjugate, let alone linearly conjugate. It is therefore re-
since it depends on pO, which in turn depends on ‘k,. A markable that for the class of vector fields 5 E 2, we
similar situation applies to k, in (2.27). However, they can cannot only classify them into equivalenceclasses,but we
be easily calculated from the relationship can derive the explicit form of one vector field-called the
normal form -in each equivalence class which is selected
k, = l/k, = k A - j$Jyl (2.33) in accordance to a unified approach.
which will be derived in Section III. The relationship Recall from Definition 2.1 that for each vector field
< E Z’, the associatedeigenvaluesare denoted by C$f jS,
k,k, =1 (2.34) and To for M,, and ~?i+ j&i and j$ for Mi. Because.$is a
follows from the ratio between the lengths (denoted by 1.1) continuous vector field by definition, these eigenvaluesare
of the following vectors (seeFig. 2): constrained in some definite way so that arbitrarily speci-
fied eigenvaluesof the above form may not correspond to
(2.35) a vector field in P. Our main result in this section is to
derive this constraint among the eigenvalues and to use
The above explicit expressionsfor the coordinates of the them to completely characterize the class of all linearly
strategic points in the D,, and the D, units will play a conjugate vector fields.
crucial role in our derivation of PoincarCmaps in Section Theorem 3.1. Linear ConjugacyCriteria:
IV. (a) For each set of eigenvaluesdefined by the six “eigen-
value parameters”
{c+.o,~~:o,~l,~l,~l} (3.2)
III. CANONICAL PIECEWISE-LINEAR NORMAL FORM
there exists a vector field 5 E 8 having these eigenvalues
In Section II, we have defined a very large family 2 of *
continuous piecewise-linear vector fields. From the
circuit-theoretic point of view, LC representsthe family of 3, > 0, 0, > 0, and y,,j$ < 0. (3.3)
all third-order piecewise-linearcircuits whose vector fields (b) Two vector fields E E .JZ and E’ E 2 are linearly
satisfy (Pl)-(P6) of Definition 1. Our objective in this conjugate of each other e they have identical eigenvalues,
section is to partition this family into “equivalence classes” i.e.,
so that all vector fields belonging to a given equivalence 60= 64, 3,=Ly), g)=y;
class have identical qualitative behavior. We will define
two forms of equivalence; namely, linear equivalenceand Cl = a”;, 6, = 3;, y1= jq. (3.4)
linear conjugacy. Proof: We will first state and prove Theorem 3.2 and
From the circuit-theoretic point of view, two circuits are then prove that it is equivalent to Theorem 3.1. We will
said to be linearly equivalent iff, except possibly for a then prove Theorem 3.2 since it is easier. Moreover, it is
uniform change in the time scale, their respective solutions Theorem 3.2 (and not Theorem 3.1) which will be used in
are qualitatively identical. If the sameproperty holds with the following sections.
the same time scale, then the two circuits are said to be Definition 3.3. Normalized EigenvalueParameters:
linearly conjugate. For example, two first-order autono- For each set of eigenvaluesdefined by the six eigenvalue
mous RC circuits [17] with time constants 7i and r2 are parameters {c?,,,9,, &,, Ci,9,, $}, we define five nor-
linearly equivalent but not linearly conjugateunless ri = r2.
Hence, two linearly conjugate but distinct vector fields
essentially represent the .same circuit but with two differ- l1Here,“ o” denotes a “corn osition” operation. Hence, (3.i) implies
ent choices of state variables which are related to each for each x E R”, G(l(x)) = v(&Gx)).
CHUA et d.: DOUBLE SCROLL FAMILY 1079
malized eigenvalueparameters
{ @O,YO,~l>Yl, k) (3.5)
where
I
p~,yoP~,ul~~,yl~~,kp-~~ (3.6)
00
{~&i)303”1,~1,%) = 2 (~&~b,P&a”;Jw;)
Note that one more parameter must be specified before the ( i
eigenvalues associated. with (3.5) can be uniquely re-
and, hence, 6 and (a,/&)[ are linearly conjugate to each
covered.
other.
Theorem 3.2. Linear EquivalenceCriteria:
(a) There exists a continuous vector field E E B having The above proves Theorem 3.2 holds.
(3.5) as normalized eigenvalueparameters = = Suppose Theorem 3.2 holds. Then given [E $P, its
associated k = -ye/R > 0 in view of (3.7), and, hence,
y,y,<Oand k>O. (3.7) qoj$ < 0. Moreover, Go> 0 and iji > 0 by definition. Hence,
(b) Two vector fields <E A? and [‘E .Y are linearly (3.3) holds. Conversely, given any set of eigenvalueparam-
equivalent CJ they have identical normalized eigenvalue eters, (3.2) satisfies (3.3). Its associatedset of normalized
parameters. Moreover, the positive scaling constant in (3.1) eigenvalue parameters
is given by 00 To 4 --91 - To
--- w,rPr> e
v = to/q) = i&/G;. (3.8) I* 3,’ 00 a1 a1 Yl 1
Note that the eigenvaluesof two distinct vector fields clearly satisfies (3.7). It follows from Theorem 3.2 that
having identical normalized eigenvalueparametersare gen- there exists a vector field [’ E 2 having these normalized
erally not identical becauseone more parameter must be eigenvalueparameters,and (&,/&#’ E 2 is linearly con-
specified in order to identify the eigenvaluesuniquely. It jugate to 5. Hence, ((so/&#’ and 5 have identical eigen-
follows from Theorem 3.1 that two vector fields having value parameters; namely, {go, So,yo, a”,,G7,,Y,}. Hence, 5
identical normalized eigenvalue parameters are generally is the desired vector field.
not linearly conjugate to each other. Indeed, (3.8) implies To prove (b) of Theorem 3.1 holds, supposeG 0 E= E’ 0 G
that the additional condition Jo = Gb is needed for linear holds for some G. Then, 5 and 5’ have identical normal-
conjugacy. ized eigenvalue parameters
Lemma 3.1: Theorems 3.1 and 3.2 are equivalent. _ _ e w I
Proof: = Suppose Theorem 3.1 holds. Then it fol- 00 Yo 01 Yl - Yo 66 7; q y; - 7;
--~
--:>r---= -- 2 -, > N, 3 . . 3
lows from (3.3) that ~o~i/Lioijl ( 0 and, hence,(3.7) holds. i Go’ijo’ol cdl ’ 71 1 i Sb wo Wl Wl y; i
Conversely, given any { uo,yo, cri,yi, k} satisfying (3.7),
define and v A ijo/& = 1. Hence, 5 and [’ have identical eigen-
values.
Conversely, if ,$ and 5’ have identical eigenvalues,then
p (~,J,Y,, - qyo/y,k -~o/~lk, -~o/k}- (3.9)
they have identical normalized eigenvalueparameters and
v p (3,/3b = 1. It follows from Theorem 3.2 (b) that G 0 [
Since ijl 2 - yo/ylk > 0 and fo%= - y;/k < 0, (3.9) = 5’ 0 G and, hence, ,$ and 5’ are linearly conjugate to
satisfies (3.3) and, hence, Theorem 3.1 implies there exists each other.
5 E Y associated with (3.9). This proves (a) of Theorem This proves Theorem 3.1 holds. n
3.2..
Remark: Since two linearly-conjugate vector fields in 2
To prove (b) of Theorem 3.2, suppose 5 and [’ are represent the same circuit (with different choice of state
linearly equivalent and, hence, G 0 6 = ( &o/&)[’ 0 G holds variables), or two equivalent circuits, the concept of linear
for some G. Then the two vector fields 5 and (G,/sb)< conjugucy is too strong for “qualitative” analysis. Since our
are linearly conjugate and must have identical eigenvalue goal is to characterize classes of nonlinear circuits ex-
parameters {Go,3,, fob,4,&i, yi}. It follows that the eigen- hibiting similar qualitative behavior, quantitative dif-
value parameters of 5’ are given by ferences in circuit time constants are irrelevant: two series
RC circuits with different time constants ri > 0 and r2 > 0
exhibit identical qualitative behavior and belong therefore
to the same class. It is not surprising therefore that the
Using (3.6), we obtain the following normalized eigenvalue weaker concept of linear equivalenceis all that we need to
parameters of 5’: study the qualitative properties of piecewise-linearvector
fields.
_----- To Before proving Theorem 3.2, we need the following
..,,r,.,P-Y - = bo>yo,~,~y,~kl result.
00 a0 a1 a1 Yl
Lemma 3.2: Let [[u] denote the family of all vector
which are identical to those of E. fields in 2 having the same normalized eigenvalue
parameters u A (uo, yo, ui, yi, k). Let 2, ok, 3, and z E’ that
denote the four vectors from the origin 0 in Fig. I to the
M,[A,B,E]=M,[A-P,B-P,E-P] (3.14)
four fundamentalpoints P, A, B, and E (Def. 2.2), respec-
tively. Then the following properties hold. M,'[A',B',E'] =M;[A'- P',B'- P',E'- P'] (3.15)
(a) All polyhedrons whose vertices consist of the origin where [ .] denotes a 3 x 3. matrix made up of various
and the four fundamental points of vector fields belonging column vectors defined above.
to the family [[p] are similar in the sensethat Now recall that the normalized Jordan forms of MO in
ZF=lOA+mOB+nZ (2.9) and Mi in (2.12) are obtained by two appropriate
(3.10)
affine transformations \Eo and 9,. It follows from (2.6)
where I = I(u), m = m(u), and n = n(u) are real numbers and (2.10) that ‘k, and \k, can be expressedby
which depend only on P and, hence, are identical for all
vector fields in &I. \k,(x) = cpox (3.16)
(b) The numbers k,, k,, and k defined in (2.21), (2.27), and
and (3.6) are related by $(x)=@,(x-P) (3.17)
k = k, = l/k,. (3.11) where a0 and ai are 3 X 3 matrices to be determined as
(c) There exists a vector field ,$E [[u] = follows. Since q. maps {A, B, E} into {A,, B,, E,}, we
y,y,<Oand k>O. have
(3.12)
@,,[A,B,E] =k,,Bo,Eol*
Proof: SeeAppendix II. a,,= [A,,B,,,E,,][A,B,E]-l.
Proof of Theorem3.2 (3.18)
Statement (a) is equivalent to statement (c) in Lemma
3.2 and is proved in Appendix II. It remains to prove Similarly, since *i maps {A, B, E} into {A,, B,, E,}, we
statement (b). have
9 Suppose there exist a nonsingular linear transforma- (&[A - P,B- P,E- P]= [A,,B&]*
tion G and a real number v > 0 such that G 0 [ = @,=[A,,B,,E,][A-P,B-P,E-PI-'. (3.19)
vt 0 G. Then the eigenvaluesof 5 and S’.must satisfy
4 + jGj = v$ k jv6; and Ti = VT/, (i = 0,l). It fol- It follows from (2.9) and (2.12) that
lows from (3.6) that their respective normalized eigen-
value parameters are identical. (i=O,l) (3.20)
e Let ([PI be the family of all vector fields in 2
having the same P = ( uo,yo, ui, yi, k) as their normal- where
ized eigenvalue parameters.Let 4 + jLsi (cSi> 0) and
yi # 0, (i = 0,l) denote the eigenvalues of 5 E 5[ u]
and let 6: * j3; (Z,I > 0) and f; # 0 (i = 0,l) denote (3.21)
the eigenvalue of [’ E .$[l~,].Denote the fundamental
points of E and E’ by {A, B, E, P} and
{A’, B’, E’, P’}, respectively. Let the vector from the Now, by hypothesis, 5 and .$” have identical. normalized
origin to these points be denoted by {A, B, E, p} eigenvalue parameters. Hence, their respective normalized
and {A’, B', E', P'}, respectively. Jordan forms Jo and Jd of MO and M,’ are identical.
Substituting (3.18) into (3.20), we obtain
Hence, A = (A,, A,,,,A,), where (A,, A,, A,) denotes
the coordinate of the point A.
By (P.3) of Definition 2.1, there exist matrices Mj and -&I= $[&B,,r,][A, B, El-’ (
Ml (i = 0,l) such that
XED, *&[A, B, Elk,, Bo, Eel-’
Ml@- p),
t(x) = MIX, XGDo = -$a,, B,,E,I[A’, B’, ET’
Ml(X + p), xeD_,
and df,'[A',B',E'][Ao,Bo,Eo]-'=J;. (3.22)
M;(x- P'), XED; Let us define next a linear transformation G: Iw3 + R3
p(x) = iqx, x E 0; (3.13) and a real number v > 0 as follows:
M;(x + P'), XED’,
GA [A~,B~,E~][A,B,E]-l, &~~/ci3;,. (3.23)
where Di and D,! (i = 0, f 1) are the affine regions of E Premultiplying both sides of (3.22) by
and .$‘, respectively. It follows from the continuity of E and J,[A', B', E'][A,, B,, E,]-‘and postmultiplying both sides
CHUA et d.: DOUBLE SCROLL FAMILY 1081
by [A,, B,,, &,][A’, B’, I?‘]-‘, we obtain field 6 E DEP,and since there exists a one-to-one correspon-
dence between each linearly-conjugate equivalenceclass of
[A’,B’,E’][A,B,E]-l~o[A,B,E][A’,B’,E’]-’ vector fields in B and each equation in (3.32) of Theorem
I
3.3 below with a fixed set of numerical parameters,we will
= “O M,‘. (3.24)
henceforth call, (3.32) the normal form equation for the
i 3; 1
‘vector fields in 8. Although this term has already been
Substituting (3.23) into (3.24) we obtain used in circuit theory to mean “state equations,” we have
~44, = GM,G-? (3.25) adopted this terminology here at the risk of some ambigu-
ity in order to be consistent with the terminology used by
Equation (3.13) implies PoincarC,Arnold, etc. [9].
@(x)ID, = G(E(G-lx)lDo), XE 0;. (3.26) Theorem 3.3. Normal Form Equationfor 8:‘2
Every equivalenceclassof linearly conjugatevector fields
Now rewrite (3.10) from Lemma 3.2 in the following in 2 defined by { &,, L3,,y,,, a”,,5r, f,} satisfying (3.3) can
vector form: be described analytically by the following canonical piece-
l’=[A,B,E][br~,n]~, P’=[A’,B’,E’][Z,m,nlT. wise-linear equation:
(3.27)
But
GP=GIA,B,E][I,m,n]T=[A’B’E’][I,m,n]T=P’.
(3.28) (3.32)
Now solving (3.15) for Mr’ and (3.14) for MI and using The 12 parameters in (3.32) are expressed explicitly in
(3.25) and (3.23) repeatedly, we obtain terms of {I& G7,,j& a”,,G7,,y,} as follows:
vi& = vM;[ A’, B’, E’] [ A’ - P’, B’ - P’, E’ - P’] --I
= GM,,;-‘[A’, B’, E’][G(A - P), (3.33)
G(B-P),G(E-P)]-’
=GMo[A,B,E][A-P,B-P,E-P]-lG-’ a - (a:+3:)~l((~l-~o)2+3:) (3.34)
l=- ~,{(a,2+i;:)~l-(~~++~)~o}
= G&G-‘. (3.29)
Now for any x E Oil, (3.13) implies
ec4LY+, = vi&(x T P’)
= GM,G-‘(x T P’) (in view of (3.29))
= GM,(G-‘x T P) (in view of (3.28))
= GC(G-l&+l (in view of (3.13)).
(3.30)
Equations (3.26) and (3.30) together imply
vE’(x) = G[(G-lx) (3.31)
for all XE D;U Dgl. Hence, (3.1) holds and .$ and <’ are a23 =
linearly equivalent. This completes our proof of Theorem

3.2. n
Our main result (Theorem 3.1) allows us to partition all,
vector fields in 2 into linearly conjugateequivalenceclasses,
each one parametrized by the eigenvalues c$,k j3,, j& a31= a21
a”,+ j&r, and j$. Since all vector fields in 2 having the
same eigenvalues have identical qualitative behavior, it (3.40)
suffices to investigate only one member in each class. Our
next theorem provides a canonical piecewise-linearequa-
tion involving 12 parameters each of which is expressed
explicitly in terms of only six eigenvalueparameters,namely “Although only invoked indirectly in the sequel, Theorem 3.3 repre-
sents one of the main rest&s of this paper and is the basis of many future
{ c?~,G,, j$, 4, i3,, T1}. Since these are the minimum num- results yet to be published. Indeed, eq. (3.32) can be interpreted as the
ber of parameters needed to uniquely identify a vector unfoldings of the double scroll equation (1.1).
Since our main motivation in this paper is to char-

(3.41) acterize the double scroll in [3], where y0 > 0, we will
henceforth restrict our analysis to the following subset
L?,,c Z’ of vector fields, henceforth called the doublescroll
(3.42) family 5(uo, yo, ul, yl, k),
20 e { ttuo, ~o,ul, ~1, k>bo < 0, ~0 ’ 0, 01’ 0,
b _ (6:+G:)i;l-(15;+Lg)fo
2- a23 (3.43) ~~‘0, k>O} (4.1)
2@; + r$)$
where { uo,yo, uI, yl, k} are the normalized eigenvalue
parameters. Stated in word, the eigenvaluepattern of any
(3.44) member of the double scroll family at the equilibrium
point P+ (resp., P-) must be a mirror image (except for
Proof: SeeAppendix III. scales)of that at the origin 0.14
Remark: Equation (3.32) is equivalent to the following Remark: It follows from Theorem 3.3 that to study the
equation: global dynamics of the double scroll family, it suffices to
study the canonicalpiecewise-linearequation (3.32).
4(x, Y, z - sy, Z>l The eigenspaces(defined by the real and imaginary
ax> YP4 = M,(x, y, zy, IZI 6 1 (3.45) parts of the complex eigenvectors)of a typical vector field
[ E 2. are shown in Fig. 2(a) along with two typical
I MI(X, Y, z + s) T Z,<-1 trajectories. Since all trajectories occur in odd-symmetric
where pairs (property (P.l)), Fig. 2(a) shows only half of the
salient features. Note that the qualitative behavior of Figs.
-a11 9 and 11 in [3] is identical to that of Fig. 2(a).
1
a12 a13
a21 a22 a23

The upper trajectory rl in Fig. 2(a) originates from
.a31 ‘32 a33
some point on U,, moves downward, turns around (before
reaching U-J, and returns to U, after a finite amount of
(3.46) time. It continues to move upward before turning around
and returns once more to UI.15 This typical trajectory
(3.47) defines a return map, called a Poincurb map from some
subset S c U, into S.l6 We can decomposethis PoincarC ’
and aij (1~ i, j < 3) are defined by (3.33)-(3.41). map into two components: a “half-return map,” which
maps the initial point on U, to the,first return point on U,,
and a “second half-return map,” which maps the first
IV. POINCAR~ AND HALF-RETURN MAPS return point to the secondreturn point on U,.
Definition 3.1 implies that in so far as the qualitative The lower trajectory r2 in Fig. 2(a) also originates from
behavior is concerned, we only need to study one member U,, moves downward, penetrates U-,, and after some
of each ZinearZyequivalent family of vector fields < E 2. finite amount of time, turns around, and returns to U-, a
Theorem 3.2 implies that we can, without loss of gener- second time. By the odd-symmetry of the vector field,
ality, choose the simplest vector field 5 E 2 having a however, we can identify eachreturn point x in U-, by its
given set of normalized eigenvalue parameters { uO,yO, reflected image - x in U,. Similarly, the portion of r2
ul, yl, k}‘as defined in (3.6), where yOyl< 0 and k > 0. below U-, can be identified with a corresponding version
Note that a piecewise-linear vector field with an arbi- of rl above U,. Through this identification scheme,both
trary {c+,, yO,uI, yl, k} may be discontinuous at the typical types of trajectories r, and l?, actually define the
boundary planes,& and K, and, hence,is not a member same PoincarC map, which in turn is simply the composi-
of 2 even though it satisfies (P.l)-(P.6) of Definition 2.1. tion of two half-return maps.
Theorem 3.2. therefore provides the foundation for this Unfortunately, the half-return maps in Fig. 2(a) cannot
section by stipulating the additional necessaryand suffi- in general be calculated by an explicit formula or al-
cient condition (3.7) for the existence of a continuous gorithm because the coordinates of the return points can
vector field &th the given parameters.13Stated in words, only be found by solving a pair of transcendental equa-.
this eigenvalue condition asserts that the real eigenvalue tions. Since these half-return maps will be used in a crucial
associated with the equilibrium point P+ (resp., P-) must
be opposite in sign to that at 0. Hence, trajectories along 14Since the eigenvalue pattern of the feedback system in [18] satisfies
the real eigenvector at P+ (resp., P-) and those at 0 must this property, it too is a special case of the double scroll family of vector
fields to be investigated in this paper.
have opposite stability properties. “This typical trajectory can never penetrate the upper oblique plane
because this plane is an elgenspace and is therefore an invariant set.
161n the following, we will choose S to be the “infinite” wedge
13This eigenvalue condition (3.7) is not necessary for continuity of the A,BE, c U1 &-I- Fig. 2(a) representing the area bounded by the two
. _. -
vector field if we allow the piecewise-linear system to have only one stmght hnes BAA, and “Em, where A, and Em denote that these two
equilibrium point instead of three, as stipulated in (P.4). lines both originate from B and extend to co.
CHUA et al.: DOUBLE SCROLL FAMILY 1083
- way in Section V to prove the double scroll is indeed via the obvious image
chaotic in a rigorous mathematical sense,we must find a
6 (4 = cpx4 (4.7a)
new coordinate system so that these half-return maps can
easily be calculated and its errors rigorously estimated. where
That such a coordinate system always exists for any E E Z0 (4.7b)
T=T(x)kinf{t>Ol&,(x)~Vi}
constitutes one of the key contributions of this paper. Our
approach for deriving this new coordinate system is to is the time this trajectory first penetrates I’;. By identify-
work with the greatly simplified but equivalent real Jordan ing this return point in I$; with its reflected odd-symmet-
forms of the regions D, and D, in Fig. 2(a), namely, the ric” image in V,, we can define the following half-return
D, unit and the D, unit in Fig. 2(b) described earlier map
(Definition 2.3).
~~‘0:
LA,B,E, + V, (4.8)
4.1. Half-Return Map v0
by
Consider first the D,, unit at the bottom of Fig. 2(b)
representing the image of D, in Fig. 2(a) under the affine qT(x),. x E *AoBoE
transformation q0 (recall (2.6)-(2.9)). The three fundamen- fob) = - %- (4 x E LA,B,E,\AA,B,E,’
tal points A, B, and E in D, map into A,, B,, and E,,
respectively. Since L, maps into the straight line LzO (4.9)
passing through B, and E,, it follows from (2.1) and the
qualitative nature of trajectories in Do that the vector field In order to derive an algorithm for calculating ~0’(x)
s,,(x) has a downward17component for all x to the right and r; (x), let us magnify the triangular region aA,B,E,
on V, and the angular region L A,B, E, on V, as shown in
of J& and an upward component to the left. Hence, any Fig. 3(a). Since the z-coordinate of each point (x, y, z) on
trajectory originating inside the triangular region
V, is simply z = 1 - x, it suffices to specify each point on
*AoBoE, V, by its (x, y) coordinate. Our next crucial step is to
A { x E V, ]x is bounded within triangle A, B, E, } (4.2) define a “local” coordinate system (u, u) on V, so that
each point x0 = (x, y)r E LA,B,E, is uniquely specified
must move down initially. But becausethe z-axis in the D,
in terms of (u, u) such that $ (x) and 7r; (x) can be
unit is the image of an unstable eigenvector,this trajectory
expressedin terms of u and u.
must move toward V, as depicted by the upper trajectory We will define our local (u, u) coordinates2’ as a weighted
in the D, unit. This trajectory defines the map sum of the four comer points A,, B,, E,, and F,, whose
r; : aA,B& + V, (4.3) (x, y) coordinates have already been found in (2.20),(2.22)
via the obvious image (2.24), and (2.25), in terms of the normalized eigenvalue
parameters, namely,
%+(4 = &b) (4.4a)
where cp,‘<x)denotes the j7ow (in the D, unit) from x to x,(u,u)=u[uA,+(1-u)E,]+(l-u)[uB,+(l-u)F,]
the first return point where the trajectory first intersects V, (4.10)
at some time T > 0, where
where 0 < u < cc and 0 G u G 1. Here, we have abused our
T=T(x)~inf{t>Ojq&(x)~V,}. (4.4b)
notation by denoting the (x, y) coordinates of the four
Here, we assumethat q’(x) does not hit IL, before time corner points by A,, B,, E,,, and F,, respectively. Note
T. The more general caseis fully treated in [20]. that x,(1,1) = A,, x,(1,0) = E,, x,(0,1) = B, and
Consider next a typical trajectory originating from a x,(O,O)=I;,. Note also that all points along the line seg-
point in the infinite wedge(angular region) ments E,A, and FOB0have a u-coordinate equal to 1 and
L A,B,E, A { x E V,]x lies within the wedge-like o,pectively. Sir&a&y, all points along the line segments
BOA0 and FOE0 have a u-coordinate equal to 1 and 0,
extension of aA,B,E,} (4.5) respectively. A typical point H with a (uo, uo) coordinate
to the right of A,E, in the D,, unit as depicted in Fig. can be identified as the intersection between the u = u.
2(b). This trajectory must move downward (because it coordinate line and the u = u. coordinate line. All points
originates to the right of L2,) and eventually intersects V; . inside the triangular region *AoBoE have 0 < u < 1, and
This trajectory correspondsto the portion of I, within D, all’ points inside the angular region L AoBoE outside of
in Fig. 2(a) and defines the map18 the A A,B, E, have 1 < u < cc. Hence, in terms of the
r; : LA,B&,\AA,B,E, + V; (u, u) coordinate systems(4.2) and (4.5) assumethe follow-
(4.6)
“Throughout this paper, odd-symmetry in R3 means symmetry with

“Throughout this section, “downward component” or “moving down” respect to the origin. Hence, two points (x, y, z) and (x’, y’, z’) are odd
(resp., “upward component” or “moving up”) means the vector field symmetric iff (x’, y’, z’) = (- x, - y, - z).
enters the boundary plane V0 from above (resp., leaves Va from below). “The reason for choosing this unconventional coordinate system will
“The symbol \ denotes set difference operator throughout this paper. be obvious in Section 4.6.
1084 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-33, NO. 11, NOVEMBER1986
u-u, coordlnete line
coordlnete line
rdinete Ilne
v-0 coordinate line
, .-~--------‘-------------
u=u+(v,t)
‘k I 12
.
O4l+a-t
(b)
Fig. 3. Geometrical interpretations of the local u-v coordinate system for representing the half-return map T$ (a) Details
of the Do unit: thick arrows denote the direction of the vector field at various points along L, = qo( L2), where all vectors
lie on the V. plane. (b) Graph of a possible inverse return-time function u = u+ (v, t). Here, f+ (u) denotes the set of first
return times which is not connected whenever u+ U+ (v, t) is not a monotone function.
ing equivalent form: where

*AoBoE, = {x0( w)h u>E [OJI x [OJI} (4.11)
LAoBoEo= {-q,bv)l<~,~~ E [O,~)x[O,l]}. (4.12) cpb(x) denotes the location of the trajectory in Iw3
which originatesfrom x,
Theorem 4.1, Calculating the rO’ Return Map: A,, k x0(1, u) denotes the location in R3 of a point along
Given x0 A (x0, yo)T E aA,B,E,, the return map the line segmentE,A, “u” units from E,,
1~:(x0) is given by B,, 4 x,(0, u) denotes the location in lR3 of a point along
the,line segmentFOB0“u” units from Fo,
T: ( xo(u, u)) = coo’ ??t
[ I
-tot: x,(24, u) (4.13) h 2 (l,O,l)T denotes the normal vector from the origin
to V,, and ( , ) denotes the usual vector dot
where (u, u) is the local coordinate of (x0, yo) = product in R3.
(x,(u, u), y,(u, u)), where 0 Q u gl, 0 Q u ~1, and t is the (b) Use the first local coordinate “u” (0 < u < 1) to
“first return time” calculated explicitly as follows. calculate
(a) Use the second local coordinate “u” to calculate the
inversereturn-time function 21defined by t=inf{t>,Olu+(u,t) =u}. (4.15)
Proof: The dynamics in the Do unit is (2.9) whose
U+(U,t)4 (cpb(Bo,), h) - 1 (4.14)
flow r&(x,) from a point x0 = (x0, yo, zo)r is given by
M(BO” - AO”),h)
I[1
e(Iofcost - e”O’sint 0 X0
QGtxo) = e”O’sint eoo’cost 0 Yo . (4.16)
21Given any “return time” to, 0 Q to < cc, and anx coordinate line eYo’ ZO
v = uo, (4.14) implies that there exists a unique u = u. = tit ( vo, to) such [ 0 0
that the trajectory cpbp(x(uo, vo)) starting from x,,( uo, II,,) at t = 0 would Since A,, + &(A,,), B,,, + cpb(B,,) and since for fixed t,
hit V. at r=tO.
cpb(x0) in (4.16) is a linear transformation, the straight line along FoBowill first move upward before returning to V,,
segment A,,&, joining A,, and B,,, in Fig. 3(a) maps whereas trajectories starting from points arbitrarily close
into a straight line segment cpb(A,,)cpb( B,,) joining to FOB,,(but on the right-hand side) will first move down-
&(A,,) and cp&,). Now if we let f, A &(x0), then f, ward and return to V, after a much shorter time. Conse-
must divide the length of the vector cpb(A,,,)&( B,,) into quently, ~0’(x) is discontinuousalong FOB,,.For conveni-
the same proportion as x,, (i.e., point H in Fig. 3(a)) ence, we will define
divides the vector A,,B, into lengths u and 1 - u, respec- ~T~(x)=x forallxe FOB,,. (4.20)
tively. In particular **
In other words, we define each point x E FoBo as a fixed
point of ~0 (x) and, hence,its first return time is equal to
zero, namely,
u+(u,t) A0 at t=O. (4.21)
(P,cpb(B,,j,h) (4.17)
3) Between t = 0 and t = co, u+ (u, t) is a continuousbut
not necessarily monotonic function of t. The continuity
= ( cpb(Aov)cpbtB,,f,
h) follows from (4.13).
4) Remarks l)-3) imply that a typical inuersereturn-time
= (cpb@,,),
h)-(%, h) (4.18) function u’( u, t) has the form shown in Fig. 3(b): it starts
(cpbt4c&,),h) from the origin and approachesu = 1 asymptotically while
where (4.18) is simply the ratio between the projections making some (possibly none) oscillations in between. It
along the normal vector h of the vectors in the numerator follows from (4.15) that the set I+(u) of “first return
and the denominator in (4.17), respectively. But times” t as u changes from 0 to 1 is in general not a
connected set. For the example in Fig. 3(b), we have
(f,,h) = (&, &,o,&,),(l,o,l)) =I (4.19) I+(u)
= 2,, + ZA,, = P, t,lU(t*, co).
since f, lies on V,. Substituting (BO,h) =l into (4.18), we 5) The example in Fig. 3(b) demonstrates that, in gen-
obtain (4.14), where we have written u+ (u, t) in place of u eral, the return time t is a discontinuousfunction of u and,
to emphasize that the right-hand side of (4.14) is a well- hence, of the initial point x,,. This shows that it is, in
defined continuoussingle-ualuedfunctions of u E [0, l] and general, impossible to express the return time t as a
t E (0, co). The superscript “ + ” denotes its association continuous function of x0. Consequently, our algorithm
with YT~ to distinguish it from u- (u, t) in Theorem 4.2, for calculating t in Theorem 4.1 is the best result obtain-
which is associatedwith v,,. n able.
Remarks: Following the,same notation and proof as Theorem 4.2,
1) Since any initial point x,(1, u) lies on the stable we obtain the following theorem.
eigenspace qO(EC(0)),cpb(x,(l, u)) may not return to V, Theorem 4.2. Calculating the r,,- Return Map:
but instead converges to the origin 0 at t + 00. In this Given x0 4 (x,,, yo)r E LA,B,E,\AA,B,E,, the return
case, however, it is logical and convenient to define map r;(xO) is given by (4.13), where (u, u) is the local
1~:(~~(1, u)) A C, = \k,(C) since we have earlier identified coordinates of (x0, y,,), 1< u < 00, 0 < u d 1, and t is the
C,, and 0 as the samepoint. It follows from this definition first return time calculated explicitly as follows.
that u’(u,t)+l as t+co. (a) Use the second local coordinate “u” to calculate the
2) It can be shown that the vector field tO(E,,) is inverse return-time function
directed from E, to A,, &(B,,) is directed from A, to B,,
and .$a(&) is directed from F, to B,, as shown in Fig. (cpb@d,h)+l (4.22)
3(a). It follows from the continuity of t(x) that the vectors u-(U’t)A (cp;(B,,,-A,,),h) *
along the line segmentB,Fb are as depicted in Fig. 3(a).
Since the vector field E(x) has a downward component (b) Use the first local coordinate “a” (1~ u < co) to
for all x to the right of the line segmentEoFo in Fig. 3(a), calculate
and since t(x) is directed to the right for all x E E,F,, it t=inf{t>Olu-(u,t)=u}. (4.23)
follows that all trajectories starting on EoFo or slightly to
the right of EoFo will first move downward towards the It follows from Theorems4.1 and 4.2 that the half-return
right before returning to V,. Hence, ~0 (x) is continuous map 7r0defined in (4.9) can be explicitly calculated, i.e.,
even along the points on E,l;b. without solving any system of nonlinear equations. Here,
In contrast, the vector field t(x) has an upward compo- we assume that the inverse return time functions u+ (u, t)
nent for all x to the left of the line segmentFOB0in Fig. in (4.14) and u-(u, t) in (4.22) have been plotted and,
3(a). Moreover, since t(x) is directed to the left for all x hence, the first return times t in (4.15) and (4.23) are
E FOBO,it follows that the trajectories starting from points simply read off these curves. This operation is of course
equivalent to finding the inverse of a function of one
22A veczr from point x tzqoint y in W3 is-denoted throughout this variable-a simple reliable task compared to that of solv-
paper by xy. The length of xy IS denoted by !q~l. ing a system of transcendentalequations.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-33, NO. 11, NOVEMBER 1986
--X
tl
A
I.,, p ,,,, ,,,,,,,,
J u-u-( 1 ,t1
u2- ,.or- ,/” ..._,.,,

,,.’ .\
ut--, 3rd . ...._
%a “L ....__..
I.” - -- ._...._.
2nd - .. -,___.
__. .._.- ---------------
Wl”,”
map .<
./AJ=U+(l,t) :
I.0 - 1st
_ r.turn
m*p
(4
Fig. 4. no associated with a monotone inverse return-time functions. (a) (4
Vo @ne. (o,,,y,,,oI,yr,k)=(-0.3,1.5,0.2,-2.0,0.75). (b) Graph of
the inverse return-trme functions u = u-(1, r) and u = u+(l, t). (c) Fig. 5. rrc associated with a nomnonotone inverse return-time func-
Magnification of (b) over the region 0.90 < u < 1.10. tions. (a) V plane. (~o,y~,(1~,7~,k) = (-0.2,0.75,0.2, -1.0,0.75). The
positions o P pomts a, W, x0 ( I+, 1) and x0 ( u2, 1) are not exact but are
exaggerated to give more space. {b) Graph of the inverse return-time
For the rigorous proof and analysis in the following functions u = u- (1, r) and u = u (1, t). (c) Magnification of (b) over
the region 0.90 < u < 1.10.
sections, it is neuer necessaryto calculate the first return
time t. Instead, the image under rrOof various constant-u
lines, which is given explicitb via (4.10), (4.13), (4.14), and The images of the line’segment A,A,, and E,,E,,
(4.22), is used directly. (where A,, and E,, denote the extension of the respec-
Example 4.1. rO with Monotone Inverse Return-Time tive straight lines to + cc) in the VO-planeunder the
Function: half-return map n, = - rr; are also shown in Fig. 4(a) by
Consider the vector field 6 with (q,, yO,ui, yi, k) = the “spirals” [C&4’,,) and [C,E&), where AL, and E&
(-0.3,1.5,0.2, -2.0,0.75). The imagesof theline segments denote, respectively, the extension of the respective curves
B,A, and FOE, in the &-plane under the half-return map to +co.
7r0= n: are shown in Fig. 4(a) as two “spirals” from B, to The eraohs of the inverse return-time functions u =
C,, and from F, to C,,, respectively. We will henceforth ~‘(1, t)“along BJ, and u.= u-(1, t) along A,&,, are
denote such curves by [z] and [=I, where [ -1 denotes shown in Fig. 4(b). A magnification of these curves in Fig.
both end points are included. 4(c) shows that both functions are monotonefunctions.
CHUA et al.: DOUBLE SCROLL FAMILY
1087
Shaded area
Example 4.2. r0 with NonmonotoneInverse Return-Time donotrr the
Y
Function: fan-Ilk, roglon
qA;B,E,
Consider the vector field 5 with (ua, yO,ui, yr, k) = r.ot,l,, ,,,,,,,,, [ ,,,,
(-0.2,0.75,0.2, - 1.0,0.75). The image in the VO-planeun-
der the half-return map n,, = ~0’ of the line segmentBoAa
is shown by the spiral [a] in Fig. 5(a). Its corresponding
inverse return-time function u+ (1, t) as shown in Fig. 5(b) X
and magnified in Fig. 5(c) is a monotone function as in
Example 4.1. -1.0
However, the image in the V$plane under the half-
return map 7r0= - r; of the line segmentA,A,, consists -0.”
of the union of two disconnectedcurves (EO] and [E).
This phenomenon can be explained by looking at the -3.”
associated inverse return-time function u-(1, t) in Fig.
5(b) whose magnification in Fig. 5(c) shows a nonmono- -4.0
tonic curve with a local minimum at t,, and a local -0.0 -5.0 -1.0 0.0 I.0 2.0
maximum at t,. The image of the line seg_ment Fig. 6. V, plane. ,, yl, k) = (-0.4,0.3,0.2, -1.0,0.3).
xO(u,,l)x,(u,,l) under vO=-lTg is the spiral [ab] in F,w,D,p ?r,(F,), P q(z), aa q(e2a2), E,AIP
Fig. 5(a). w,(E,A,), and fi p q-‘(F,). The position of fi is exaggerated in this
figure for clarity. The actual position of fi is very close to aI.
If we plot the second and the third’ return maps of
xc,(ul,l)xO(uz,l), we would obtain the curves (@ dur- where cp; ‘(x) denotesthe flow (in the D, unit) from x to
ing the time interval t, < t Q t, and (a%‘) during the time the first return point where the trajectory first intersects Vr
interval t, < t < t,, where t, = inf{ t > t,lu-(1, t) = ur}. at some “reverse” time - T < 0, where
4.2. Half-Return Map VT,

Consider next the D, unit at the top of Fig. 2(b) Our next theorem shows that 7r1can be calculated by an
representing the image of D, in Fig. 2(a) under the affine explicit algorithm similar to that of r,,.
transformation \k, (recall (2.10)-(2.12)). The three funda- Theorem 4.3. Calculating the rt Return Map:
mentalpoints A, B, and E in D, map into A,, B,, and E,, Given x1 p (x1, yl)= E LA,B,E,, the half-return map
respectively. Here, we abuse our notation by using the rl(xl) is given by
same symbol D, to denote the top region in Fig. 2(a) and
a point on the z-axis in the D, unit in Fig. 2(b). We will 7r1(x1(u,v)) =e-“I’ $f& u, v) (4.28)
inherit the samenotations in the preceding section with the
exception that each subscript ‘0” corresponding to D,, unit where (u, v) is the local coordinate of (xi, yr) =
should be changed to “1” for the D, unit. Hence, we (xi(u, v), yi(u, v)), where 0 < u < cc, 0 < v ~1, and t is
define again a- local coordinate
- system (u, v) such that the the “first” return time calculated explicitly as follows.
line segmentsE,F1 and A,B, in I’, in Fig. 2 correspond to (a) Use the first local coordinate “u” to calculate the
the v = 0 and v = 1 coordinate line, respectively. Likewise, inverse return-time function
the line segmentsF,B, and E,A, correspond to the u = 0
and u = 1 coordinate line, respectively. Any point xi in-
side the wedge (angular region) bounded by B,A,, and (4.29)
B,E,, is uniquely identified by
xl(u,v)=u[vA,+(l-v)E,]+(l-u)[vB,+(l-v)F& where E,, A x,(u,O) denotes the location in R3 of a point
along the line segment F,E, “u” units from F,, and
forO<u<cc andO<vdl. (4.24) A,, p xt(u, 1) denotes the location in R3 of a point along
Under this local coordinate system, we can define the the line segment B,A, “u” units from B,.
triangular region A A,B, E, and the angular region (b) Use the second local coordinate “u” (0 < v < 1) to
LA,B,E, as follows: calculate
aA,B,E,s {~~(~,v)l(~,v)~E[0,1]X[0,1]} (4.25) t=inf{t>O]v(u,t)=v}. (4.30)
LA146 p { xdu, v)I( u,v) E [O,cc)X[O,l]}. (4.26) Proof: Follows mutatis mutandis the proof for Theo-
Finally, we define the second half-return map rem 4.1. n
TQ(X): LA,B,E, + V, (4.27a) Example 4.3. IT, with NonmonotonicInverseReturn-Time
Function:
via the obvious inverse image
Consider the vector field 5 with (u,, y,, ui,yi, k) =
dx) = cp;‘b) (4.27b) (-0.4,0.3,0.2, -1.0,0.3). Since al(x) is defined to be the
(4
(8) (h)
Fig. 7. Graphs of the inverse return-time functions u = a( U, t). The parameter values are the same as those of Fig. 6. (a)
o = ~(0 t . (b) Magrufrcanon of (a) over the region 0.995 < u < 1.005. (c) u = v( at, t), where ur = 0.570. (d) Magnification of
(c) ovei ti e regon 0 .’995 < u < 1.005. (e) u = v(u,, t), where u s = 0.786. (f) Magnification of (e) over the region 0.995 < u <
1.005. (g) u = ~(1, t). (h) Magnification of (g) over the region 0.995 < u < 1.005.
reverse flow, the vector field (i(x) on Vi becomes - (i(x) cum fla2, in addition to already being discontinuous
in following the image of x under ni(x). Hence, the along EiF,.23
direction of t(x) along the line LzO= \k,( L,) in Fig. 3(a) Let us summarize the behavior of vIT1 in Fig. 6 as follows.
must be reversedin the correspondingline Lll = \k,( L,) in
Fig. 6. Hence, TV is discontinuousalong the line seg- (1) T~(AA,B,E,) =a fan-like closed region q A;B,E,
ment m in Fig. 6, whereasit is continuous along the line (shown shaded)in Fig. 6.
segment FIB,. This is the opposite of 7ro(x), which is (2) Ir,(B,aJ = D,. [:::: 1
discontinuous along FOB0but continuous along E,F,. Note Here, ?r,(B,az) actually maps into the origin in the
that E,F, correspondsto our v = 0 coordinate line. unstable eigenspace\k,( E ‘( P)), which becomes a
The image of FIB, under rri is the spiral [Fml] in stable equilibrium under the reverseflow v;‘. It is
Fig. 6. In Appendix IV, we shall prove that this spiral is logical and convenient to identify the origin with
tangent to the line”E,B, at Fl. The image of the line D;= \k,(P+) in V,.
segment E,A, is shown in Fig. 6 as part of a large spiral (3) Since 7~~is discontinuous along E,F,, we will de-
fine (as in ro)
[m]. The continuation of this spiral to the right of A; is
the image of the extension of E,A, beyond A,. rr(x) A x for all x E E,F, . (4.33)
The inverse return-time function v = ~(0, t) in Fig. 7(a) In particular
and its magnification in Fig. 7(b) shows that it is a
monotonic increasing function of t. However, the inverse dfl) = Al = 4. (4.34)
return-time function v = ~(1, t) in Fig. 7(g) and its mag- With this definition, r1 becomes continuous at
nification in Fig. 7(h) shows that it is not monotonic and El&.
has a value larger than 1 for t, -Ct < t,, where t, p inf{ t > (4) or, is one-to-oneat all points inside the triangular
O]v(l, t) =l} is the time it takes A, to go to A;. The time region A A, B, E, and its boundary except the points
interval (t3, t4) therefore correspondsto the time where the along the line segment [B,az) and the isolated
extension of the outer spiral [m] lies to the right of the point fi, i.e., on aA,B,E,\([B,a~U {fi}).
line segmentB&(i.e., the v = 1 coordinate line). (5) 71 -’ is well-defined at all points in the fan-like
Recall that FIB, and E,A, correspond to our u = 0 and region q lA;B,E, except for the two isolated points
u = 1 coordinate lines, respectively. There exist 0 -Cufi Fl and D,.
u2 < 1 such that the corresponding coordinate lines aleI (6) The spiral (F-J is the set of discontinuous
(u = ur line) and a2e2(u = u2 line) are mapped under ,rrr points of rr;‘. The function a;’ is discontinuous
into the following two curves: at these points because r;‘(x) +fia2 from the
(a) ~i(a,e,) is a spiral [e-r], which is tangent to right as x + W, from the right, whereas m;‘(x)
E,B, at F,; + FIBI from the right as x + WI from the left.
(b) n,(a,eJ is a curve [z], which is tangent to A;A, This follows becausethe return map 7r1is discon-
at B,. tinuous along the curve fz, and because7r1&$
The graph of the inverse return-time functions v = = FIBI = 7rl-‘(Fml).
v(ui, t) is shown in Fig. 7(c) and its magnification in Fig. Using the above properties, we can now define the
7(d) shows that it is monotonic with an inflection point inverse half-return map s;’ as follows:
Tl -l: q A;,B,E,, + LA,B,E, (4.35)
2
dv
-=Oand $=O where
i dt ;i A {(x, y, z) E V,(y > UlX + yl(l - x), x 91)
W,4El,
(4.36)
at some time t,. The graph of the inverse function v =
v(u,, t) is shown in Fig. 7(e) and its magnification in Fig. is the region above the line B,E,, and to the left of AlA;,
7(f) shows that it is nonmonotonic with a maximum value in Fig. 6, and where
v=l at t=t,, where t, is the time it takes to go from a2 $-l( Dl) p B, (4.37)
to B,.
Now let fr be the inverse image of Fl in Fig. 6, i.e., q-y F,) A fl. (4.38)
7ri(f,) = Fl. Similarly, let the inverse image of FIBI be Note that ~1~ is discontinuousalong [Fml].
denoted by vx], namely, the curve fx in Fig. 6. Since
the region bounded by the closed curve ele2a2fiel is 4.3. ConnectionMap @
found to map into the region bounded by the arc a, line Since the Do unit and the D, unit in Fig. 2 have
FIB,, arc G, and line e2e1,whereas the neighboring different reference frames, let us “match” the two units by
region bounded by the closed curve jr=jr is mapped
23These additional discontinuity points occur when we choose our
into the region bounded by the closed curve Fml parameters close to those which gave us the double scroll. They may not
in Fig. 6, it follows that ITS is discontinuousalong the occur inside A&i$& for other choices of parameters.
1090 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-33, NO. 11, NOVFMBER1986
defining the affine connectionmap via the formula
@ A (QJ”(~olcJ-l (4.39)
v(x) = T;%)?T~@-~(x), if x E LA,B,E,
(4.46)
where qll, and \k,l, denote the restriction of \k, and \ko = @~~Wb;‘(x), if x E V/\LA~B,E,.
on U,. Again, since zi =l- xi, it suffices to find the
explicit formula relating (x0, yo) E Do to (x1, rl) E D,. Note that m(LA,B,E,) C LA,B,E, and r;l is well de-
Since 4, = (1, po,O) + A, = (1, pl,O), then fined for all x E V/\LA,B,E, in view of (4.36). Here, Vi’
denotes the F/,-plane to the left of x = 1.
[;:I 4[;;] =L[$-;o]+[;l]. (4.40)
4.5. V, Portrait of V,
Hence
In our study of the global dynamics of the double scroll
family in the following sections,we will often need to look
at the image via @ of the half-return map of several line
(4.41) segmentsdefined as follows:
B,C,g hoW’( A,B,) = &ro( A,B,) (4.47)
Now, since B, d (Box, B,,) + B, g (Blx, B,,) and E, A
(E,,, E,,,) + El 4 (El,, El,), it follows from the action of I;lc,g hoV( F,E,) = (avo( FOE,) (4.48)
L in (4.41) that
ma (ProW1( AlA,,) = Qro( A,A,,) (4.49)
c,E;,e aToP( E,E,, ) = @vro(E,E,, ) (4.50)
E,A;& q( E,A,). (4.51)
--
The images G, s, C,A;,, C,E;,, and a ,for a
typical set of normalized eigenvalue parameters { uo,yo,
It follows from (4.42) that ul, yl, k} for a vector field < E To are shown in Fig. 8. We
will henceforth refer to this picture as the V, portrait of V,.
Box-A,, Eo,- A,, -’
El, - 4,
El, - Al, Boy-AoyEov-A,, . I
(4.43)
Note that Cl 2 @(Co)= q,(C).
Stated in words, the V, portrait of V, consists of four
distinct sets of points:
- -
Set 1. two boundary lines B,A,, and B,E,, represent-
Substituting (2.20), (2.22), (2.24), (2.26), (2.29, and (2.30) ing the V,-coordinates
- of- points along the
for the respective components of Ai, Bi, Ei into (4.42) we boundary lines BoAOooand B,E,, of the infinite
obtain the following formula for L: wedge LA,B,E,;
(4 +G1
L = (d+l)(ko +l)Q,y,
I
-yl(k,+l)[Qo+uo<uo-vo><kl+1>1 vovdko +l)(kl +I>
X -y,(k,+l)(u,-yo)[u,(u,-y,)+lb Yo(kl+1)[Q,+y,(u,-y,)(ko+1)1 (4.44)

-y~(ko+l)(u,-y,)[uo(uo-yo>+11 I
where Qi k (ui - yi)’ + 1, k, A k, and k, 2 l/k. Set 2. the boundary line E,A, of the triangular region
Note that L is expresseddirectly in terms of the normal- aA,B,E,;
ized eigenvalue parameters {a,, yo, ul, yl, k}. Set 3. four spirals representing the image of points in
Set 1 under the To-map (in Do unit) but
translated into the coordinates on V,;
4.4. PoincarP Map T Set 4. a partial spiral representing the image of the
We will now use the half-return maps r. and 7~~and the points in Set 2 under the al-map.
connectionmap Q, to define a Poincare map In Section V, we will consider the important casewhen
77:v;-+v; (4.45a) Set 4 includes the point C,, i.e., Cl E E,A;.
4.6. Spiral Image Property
where
The various spirals in Figs. 4(a), 5(a), 6, and 8 were
v;a {(X,j)EVrIX<l} (4.45b) calculated by computers for various specific sets of param-
CHUA et al.: DOUBLE SCROLL FAMlLY 1091
the u = us and u = u,, coordinate lines are always spirals.

It is precisely this observation that prompted us to choose
this unconventional local coordinate system.
V. PROOFOFCHAOSINTHEDOUBLESCROLL
An equilibrium point Q of a vector, field 5 is said to
have a homoclinic point if there exists a trajectory which
X
-D
tends to Q as t + + 00 and as t + - co. Such a trajectory
is called a homoclinic orbit through Q. The significance of
homoclinic orbits is given by the following important
result.
Shilnikou’s Theorem[9], [15], [19], [28], [29]25
Let 5 be a continuous piecewise-linearvector field asso-
-3.0 -2.0 -1.0 0.” 1.0 2.0 3.0 ciated with a third-order autonomous system i = f(x),
Fig. 8. VI portrait of Va for (u,,,,J,~)u~, yr, k) = (-0.4, 0.5, 0.05, -2.0,
x E R3. Assume the origin is an equilibrium point with a
pair of complex eigenvalues(I * jw (u < 0, w # 0) and a
real e&revalue y > 0 satisfying ]a] < y. If in addition, 5
eters. In general, the image under $, r;, or rIT1of any has a homoclinic orbit through the origin, then 4 can be
bounded straight line segment along a u = us or u = u0 infinitesimally perturbed into a nearby vector field 5’ with
coordinate line is always a spiral. To prove this important a countable set of horseshoes.
property, it is convenient to rewrite (4.13) and (4.28) in a Since horseshoesgive rise to extremely complicated be-
more compact form by identifying a point x = (xi, yi) in havior typically observedin chaotic systems[9], one of the
the I$-plane (i = 0,l) by a complexnumber(phasor)X = (xi few rigorous methods to prove a system is chaotic is to
+ h). 24 For example, (4.28) can be rewritten into the apply Shilnikov’s theorem. In this section, we will prove
equivalent form the double scroll family (4.1) is chaotic by showing that the
conditions of Shilnikov’s theorem is satisfied. In particular,
T,( X1(U,u)) = X,(u, u)e-(“l+jl)’
(4.52) we will prove that there exist parameters such that the
where Xi(u, u) B xia(u, u) + jx,,( u, u) and xi(u, u) p trajectory along the unstable real eigenvector E’(0) from
[%a(% u>,Xlb(W u)IT. the origin will enter the stable eigenspaceE ‘(0) in Fig.
Nbw for 1 E (0, &), X1(~a, u(uO,t)) representsone point 2(a) and, hence, return to the origin. By symmetry, the
along the u = u0 coordinate line. If u( a,,, t) increases trajectory along the other unstable real eigenvectorwould
monotonically from u = 0 to u = 1 as in Fig. 7(a) when behave in the same way. These two special trajectories are
us = 0, then Xi(u, u) moves monotonically from u = 0 to shown in Fig. 9(b) and are therefore both homoclinic
u = 1 as t increases from 0 to co. If u(ua, t) is not orbits.
monotonic but is bounded between u, and ub as in Fig. Theorem 5.1. Homo&nib Orbits in the Double Scroll
7(h), Xi(u,, ~(a,,, t)) will move back and forth along Family:
portions of the u = u0 coordinate line while moving from Let 5 be any vector field in the doublescroll family
u, to Us. In either case, since x:Ju, u)+ x&(u, u) < co,
q(uo9 u(‘Ql, t)) I+ 0 as t + 00. The loci of points under vi
along u=us is therefore a shrinking spiral whose am-
plitude is modulated in accordancewith x1(us, u( ue, t)). If
x1(2+),u(uo, t)) varies only slightly for all t E (0, co), as in Assume E satisfies the following conditions.
the cases,shown in Figs. 6 and 8, the shrinking spiral (i) Let C, 5 q,(C) mapunder rr;’ into a point on
would look almost like a “logarithmic spiral.” The same the line segmentA,E, in the D, unit (see the Vi
interpretations apply to rri and r;. portrait of V, in Fig. 9(a).26
In view of the odd symmetry of the vector field 5, spiral (ii) In the Do unit (Fig. 2(b)), no trajectory starting
images under TO’, 7r;; and rri always occur in odd-sym- from points on the line segment AoEo in the
metric pairs. This proves formally that the cross section eigenspacez = 0 intersects the boundary line x =
along the Vi and U-i boundary planes of the double -1.
scroll attractor consists of two tightly wound odd-symmet- Then c has a horrioclinicorbit through the origin.
ric spirals, thereby justifying our choice of the name “dou- If, in addition,
ble scroll.”
Since the image of rz, r;, and v1 of an arbittwy curve
25The original Shilnikov theorem requires f( .) to be an analytic
or line segmentin Vi is, in general,a curve with no special function. lhe piecewise-linear version we invoke in this uauer
_ _ is used in
properties, it is indeed remarkable that the images along [15] and [19]. -
26Recall C is the intersection of the unstable real eigenvector at the
origin with the upper boundary U, in Fig. 2(a) and n, is the half-return
map defined in Section 4.2. Condition (i) means that G= ~r(ErAr)
24We use capital letters to denote phasors. must pass through the point Cr.
cally as indicated: 27
Q
(i) co t in - 1.066296G to G - 0.906832
.
P+ 9, t in 1.382371~ 3, < 2.228686
y. 4 in 2.132590>, Toz 1.813664 (5.5)
(ii) cYl4 in 0.295297> c1 > 0.138551
G3,t in 1.879726< t, d 2.527628
-x
( GJ j$ t in - 3.590593< j$ < - 3.277103
ca :
. (5.6)
(iii) a0t in -0.771352 < a0< -0.406890
u1J in 0.157096>, ui >, 0.054814
.p-
y. J. in 1.542704>, y. > 0.813782
yi t in - 1.910168Q yi < - 1.296513 (5.7)
3
(iv) ko/yo t in 0.384997< k,/y, Q 0.680079
(a) (b)
Fig. 9. Homoclinic orbits. (a) VI portrait of V,. (b) Two odd-symmetric
k,/y, J in -0.881427 > k,/y, 5 - 1.393659. (5.8)
homoclinic orbits through the origin. Moreover, the above bounds can be calculated to any
desired accuracy.
Proof: It follows from (l.l)-(1.3) that the real eigen-
w luol< Yo (5.2) value yi corresponding to m = m i (i = 0,l) is a real root of
then 5 is chaotic in the senseof Shihrikov’s the- the characteristic polynomial equation
orem.
x3+(am+1)x2+(am-a++)x+c$m=O. (5.9)
Proof: Theorem 3.2 guarantees that the vector fieid
5 E 2. is continuous and the half-return map 7r1is well Solving (5.9) for /3, we obtain
defined.
Consider the trajectory r, through the origin and mov- 8=P(x)&x(x+l)-* (5.10)
ing upward along the unstablereal eigenuectorE’(O) in Fig. x+am’
2(a) until it hits at point C. Since C, = *i(C) and It follows from (5.10) that if a > 0 and am >l, then p:
c; A q’ (C,) EA,E,= \k,(rQ (see Fig. 9(a)) in view of (- co, - am) + Iwis an increasingbijection (i.e., one-to-one
condition (i), it follows that the trajectory I?, through C and onto), and if (Y> 0 and am < 0, then p: (- am, 00) +
must land at a point C, on segmentAE in Fig. 2(a). But AE R is a decreasing bijection. Hence, for (Y= 7 and m. =
lies on the stable eigenspuceE ‘(0) at the origin, and since -l/7 (resp., m, = 2/7), To (resp., 7,) decreases(resp.,
condition (ii) guaranteesthat the trajectory rFz through C, increases) and satisfies
will not intersect the lower boundary U-i, it follows that 1.813664< min(~o) < max(yo) Q 2.132590
I?,2 must remain on, the eigenspaceE c(O)and converge to
the origin as t + 00. Since r A I’, u I’, u I’,* tends to the (resp., - 3.5905936 min ( j$) G max ( yi) 6 - 3.277103)
origin as t + + do and as t + - 00, it is a homoclinic orbit. (5.11)
If, in addition, laoI< yo, then the hypothesesof Shilni-
as p increasesfrom 6.5 to 10.5.
kov’s theorem are satisfied and, hence, 5 is chaotic. w Now the solutions of (5.9) are related to its coefficients
Theorem 5.2. Chaosin the qouble &roll:
as follows:
The double scroll system (l.l)-(1.3) is chaotic in the
senseof Shilnikov’s theorem for some parameters m,, m,, 2tj+fi= -((cYmi+l)
(Y, and /3. In particular, if m, = -l/7, m, = 2/7, and ~~+GD?+2~i~i=cy(mi-1)+J3
(Y= 7, then there exists some /3 in the range 6.5 < p Q 10.5
such that the hypothesesof Shilnikov’s theorem are satis- yi( 6; + (3;) = - c$mi. (5.12)
fied. Solving for ci and (3; from (5.12), we obtain for i = 0,l
Before we can prove Theorem 5.2, we will need four
lemmas to be stated and proved below. To avoid rep- fYi=-i(ami+l+Ti)
etition, we make the following assumption.
Standing Assumption: The parametersfor all lemmas are a2mi
&f=-i(ami-1-yi)2- (5.13)
m, = -l/7, m,,= 2/7, a= 7, /3 E J A [6.5,10.5]. (5.3) yi+ami’
Also, we will use the abbreviated notation Combining (5.11) and (5.13), we obtain properties (i) and
X t in a G X G b (resp., X J in b > X > u)
(5 -4) (ii).
to mean the variable X = X(p) increases(resp., decreases) Property (iii) follows directly from properties (i) and (ii)
monotonically and satisfies a Q mm(X) Q max(X) < b as and the assumptions a0< 0, y,, > 0, and 3, > 0.
p increases monotonically in the range J.
27Recall the following definitions: for i = 0 or 1, q d Ei/G,, y, 2
Lemma 5.1: As p increasesmonotonically from & = 6.5 ~~/3,,k,~l/k,~k~-~~/~~,Qi~(ui-~i)2+1,pi~ui+ ($+l) X
to p2 = 10.5, the following parametersalso vary monotoni- (ki/Yi).
CHUA et al.: DOUBLE SiXtOLL FAMILY
Property (iv) follows from properties (i) and (ii) and the (iv) For each fi E J
relationships lAoI 2 IEole@O (5.21)
where 6, > 0 denotes the angle subtended by the two
vectors E, and a, on the x-y plane.
Proof
k,
-=- -71 -71 =-- 4 (5.14) (i) It suffices to show that
Yl i-)/i-
YO i a1 70 *
(5.22)
Finally, note that the bounds in properties (i)-(iv) can
WI a I4 419 O>la IJV
be calculated to be exact to any number of digits because Since x,(1, u) = Z?i + um, u E [0, 11, it follows
(5.10) and (5.13) are rational expressions. n from plane geometry that ~
Our next goal is to examine the loci of points obtained
by applying the half-return map ri to the segment &A, )x1(1,u)12=
(i.e., u = 1,O G u G 1) on Vi: they are obtained by substitut-
ing u =l and u = ~(1, t) for t E 1(l) into (4.28), where 1(l)
denotes the set of “first return times” for u E [O,l]
40 = +l(L 4 t)))
If we can show that
cost
=e -01‘
[ -sint
~$]Xl(l, u(l, c)) (5.15) (Z&g)>0 (5.24)
for t E 1(l). Using the phasor notation (4.52) (5.15) as- then (5.23) would imply (5.22) because(x,(1, u)12is
sumes the following compact form: an increasing function of u E [0, l] and since ]A,J =
X(t) = X,(1, ~(1, t))e-(al+jl)‘, t E 1(l). (5.16) lx,(Ll)l and IEd= Ixdl,W
‘To prove (5.24), we make use of the first two coordi-
Similarly, it follows from (4.13) that the loci of points nates of E, from (2.30) and A, from (2.26) to write
obtained by applying the half-return map V$ to the seg-
ment BOA0(i.e., u = 1,O G u < 1) on V, assumesthe follow- OE’1=(Yl(Y~-o,-Pl)/e,,Y,[l-P,(a,-Y,)l/e,)
ing compact form: (5.25)
x(t) = xo(u+(l,t),l)e(OO+jl)t, t E I+(l) (5.17)
E,A; = ([(J&J~- ~d+l+ YIP&Q,,
where X0 is the phasor associatedwith x0 and I+ (1) is the (5.26)
{P&J,(v YI)+~] - v&‘Q,).
set of “first return times” for u E [O,l]. We have already
identified the set of points in (5.16) and (5.17) as portions Calculating the inner product between (5.25) and (5.26)
of a shrinking spiral whose amplitude is modulated in we obtain
time. We will-now show that these two spirals are sand- (~,,EIA;)=-a,~,(p:+l)/Q,. (5.27)
wiched between two logarithmic spirals.
Remark: To simplify our notations, all vectors in the using (5.27) and Lemma 5.1 (iii) (ai > 0, y1 < 0), we
following three lemmas (Lemmas 5.2, 5.3, and 5.4) are obtain the desired inequality (5.24).
projected onto the x-y plane, and&ence, represent two- .* This is proved by the samemethod as in (i).
dimensional vecc. For example, UE, and so should be $ From (5.24), we have 0 < 6, < n/2. Hence
interpreted as D,A, and a, respectively.
Lemma 5.2: 9, < tan 6,. (5.28)
Moreover, since 0.054814Q ui < 0.157096(Lemma
(9 For each p E J, and any time t E 1(l), the magni- 5.1), it is easy to verify that
tude of x(t) of the spiral (5.16) in V, is bounded
by two exponentials 1-2~~9~ < eT2’191. (5.29)
IAlle-“l’> Ix(t)! > JElle-Olf. (5.18) Since (5.20) is equivalent to
00 For each p E J, and any time t E I+(l), the mag- lE112/lA112G ew2”1’l (5.30)
nitude of x(t) of the spiral(5.17) in V, is bounded it follows from (5.29) that to prove (iii) of Lemma
by two exponentials 5.2, it suffices to prove
IAOleOOf
> Ix(t)1 > IBOleOO’. (5.19) (E,12/IA,(2 ~1-2~~ tan6,. (5.31)
(iii) For each p E J Since A, and E, are projected onto the x-y
lEll < IAlle-“l’l (5.20) plane, we can suppressthe z- coordinate in (2.26)
and (2.30) and obtain after simplification
where 9, z 0 denotes the angle subtended by the
two vectors a, and OA, on the x-y plane. IA,12= P; +l, l&l2 = Y:( P: +1)/Q,. (5.32)
Now define the normal vector to E, as follows: Proof: It follows from Lemma 5.1 that for each /3 E J,
a0 < 0, y. > 0, u1> 0, yi < 0, and k > 0. Hence, the vector
E,I p (-Y~~-P&I - u&Q,, field E E 9 defined by (l.l)-(1.3) is a member of Z. C 2
YI(YI - UI- P&Q,> (5.33) in (5.1) for all p E J. Moreover, the rangesassumedby u.
and y. in Lemma S.l(iii) imply .luol< y. for all p E J.
then it follows from (2.30) that
Hence, we need only prove hypothesis (ii) of Theorem 5.1
(E,I=IE:Iand ?%,l%il. (5.34) holds for all /3 E J.
Suppressing the z-coordinate from (2.20) and (2.21), we
A straightforward calculation shows can write
tan 9, = (721, z$)/c( zT1,z3,) k0
A, = (1, PO), Po=~o+--(~cf+l) (5.46)
=1/h - Yl>. (5.35)
Substituting (5.32), (5.35) and (5.28) into (5.31) where - 0.771352G a0d - 0.406890 and 0.813782< y. d
and solving for yl, we obtain 1.542704.Since
(5.36) poGmax(ao)+m= (max(u,2)+1) =0.39<0.4
Hence, to prove (iii) of Lemma 5.2, it is sufficient (5.47)

to prove (5.36) holds over the parameter range we have
assumed by yi and u1 for /3 E J. To verify this,
note that the right-hand side of (5.36) decreases lAoI =l+ pi x1.16 and ‘p. 2 tan-‘(p,) E 0, z
over the range 0.054814G u1G 0.157096 with a i 1
minimum value equal to - 0.1650.Since the maxi- (5.48)
mum value assumedby yi is -1.296515 (Lemma where ‘p. is the angle between a0 and the x-axis. Now,
5.1 (iii)), it follows that (5.36) holds for all p E J. for t>~/2-9~
(iv) It follows from (2.20) and (2.24) that (5.21) is
equivalent to
(l+p,2)-(1+u~)eZUo90>0. (5.37)
To prove this, let us define the function <JiZexp[$max(o,)]
g(t)kl+tan2(cp+t)-(l+u,2)e2”of, = 0.78 cl. (5.49)
t E [o,4Il (5.38) Since 0 < ‘p. < 7r/4, it can be shown that the trajectory
x,(t) starting from A, remains in the region x > 0 for all
and
0 c t < r/2 - ‘po. Consequently, x,(t) never reaches the
cphin-luoE -z, . (5.39) line x = - 1 for t > 0, namely,
i 2O 1
{ Xo(l,l)e(“o+jl)‘(t > 0} C {(x, y)lx > -l} (5.50)
It is easy to verify that
where each phasor on the left at any time t > 0 is identi-
g(O) = 0 (5.41) fied as a point in the x-y plane. Similarly, it can be
g’(t)=2tan(cp+t)[l+tan2(cp+t)] shown that the trajectory x,(t) starting from E, never
reaches the line x = - 1 for t > 0, namely,
-2u,(l+ u,2)e2”o’ (5.42)
(Xo(l,O)e(oo+jl)t~t s- 0} C {(x, y)lx > -l}. (5.51)
g’(0) = 0 (5.43)
Since
g’(t)20 for 0<1<:-~ (5.44) Xo(Lu> = ux,(Ll) +(1- ~)&0,0), uE [OJI
where (5.44) follows from a0< 0 and - 7r/2 < ‘p <2Since (5.52)
E, = (1, uo), ‘p is the “negative” angle between OE, and and since at any time t the flow of a linear system is a
the x-axis. Hence, 0 < 9, < 1~/2- cpfalls within the range linear function of the initial state, it can be shown that
of t in (5.44). Moreozr, since A, = (1, po) and ‘p + 8, is
the angle between OA, ,and the x-axis, it follows that {X0(1, u)e(oO+il)‘lt > 0} C {(x, y)lx > -l}. (5.53)
tan(cp + ao) = po. Hence, letting t = 9, in (5.38), we ob- n
tain Lemma 5.4: Let C, p q,(C) = (xc, yo) and J’l A q,(F)
g(9,) = (1+p,2)-(1+u~)e20~90>0. W (5.45) = (xF, yF) on the x- y-plane28 in Fig. 2(b). Then for
Lemma 5.3: For each /3 E J, the double scroll system
(l.l)-(1.3) is a member of the double scroll family (5.1) 28Recall that all vectors in Lemma 5.4 are projected onto the x-y
and satisfies hypotheses(ii) and (iii) of Theorem 5.1. plane.
CHUAet al.: DOUBLE SCROLLFAMILY 1095
every/3 E J, we have
x,<x,cl and y,>O. (5.54)
Moreover, C, is a continuousfunction of j3 for all p E J.
Proof: From (2.32), we identify
xF=Y,(Yl -24/Q,, YF=Y~[~-~~(u~-Y,)I/Q,.
(5.55)
Since C, = *(Co) = Qi(O,O)when projected onto the x-y
plane, where @ is the connectionmap defined in (4.40) and
(4.44), we can calculate the exact coordinates of xc and yc
as follows:
c=l- b:+l)[(uo+yokl)2+l]
X (5.56)
ko'+l)Q,
~1b~,b,-~,)l (ul’+l)yok,
yc=
Ql - (d+l>u,Q,
~{kl~o[~,(~,-~,)+ll+2~0~,(~,-~,)}. (5.57)
From (5.55) and (5.57), we obtain
a:+1 Fig. 10. The circles bounding S, and S, on the V, plane and related
Yok,(Y&, -20,) ’ 0 (5.58) arcs.
XF
-xc= Q,(u,‘+l)
becauseyak, > 0 and a0< 0 for /3 E J (Lemma 5.1). Hence, -2min(~o)
>, - 1 = 0.0102
xc < xF. The fact that xF < 1 follows from the geometry of &Pl
the D, unit in Fig. 2(b), where A,B, lies on the line x = 1. 1 2
To prove yc in (5.57) is positive, it suffices to show becausem,=-T,m,=?
(~~+l)klYo{klYo[~~(~l-Y1)+11+2~0~~(~;-~1)~ > 0. (5.62)
‘[l-u,(a,-Y,)lY,2(u~++1) (5.59) Since yi is a continuousfunction of S in view of (5.10), it
because yi < 0 for p E J. We can rewrite (5.59) as follows: follows from (5.13) that li, (3,, and ki are also continuous
functions of /3 for i = 0,l. Since C, = (x,, yc) is given in
(5.56) and (5.57), C, is a continuous function of p. n
Proof of Theorem5.2
It follows from Lemma 5.3 that it suffices for us to
>l-q,(u,-y,). (5.60) prove that hypothesis (i) of Theorem 5.1 holds for some
Since for all /3 E J p E J, i.e., we must prove that there exists some p E J
k,you, 4
such that C, E nl(E,A1) as depicted in the Vi portrait of
-=---0 and ~,(a,-yi)>O (5.61) V, in Fig. 9(a) when this happens.
YlUO 60 To do this, let us draw two concentric circles S, and S,
we have with their centers at D, = (0,O) in the Vi-plane and a
{left side of (5.60)-right side of (5.60)) radius equal to IAll and lE11e-2n01, respectively, as shown
in Fig. 10. Let I be the horizontal line through D, (i.e., the
x-axis) and 1’ be the vertical line through F1. Clearly, 1’ is
to the left of the x =1 line in view of Lemma 5.4. Let S,
intersect I and I’ at points a and a’, respectively. Let S,
intersect 1 at a point b to the left of D,. Depending on the
value of lEll and ui, S, either intersects 1’ at two points, in
which case the upper point is labeled b’, or otherwise, let
b’ be the point where S, intersects 1 to the right of D,, as
6; + (3:
shown in Fig. 10. Let g- be the upper point where S,
=- - 1 (becausea, = &/Si, ?=O,l) intersects the y-axis. Let R denote the region enclosedby
c20 + G20
the closed contour formed by either aa’b’nba (if b’ lies on
ml70 I!) or s (if b’ lies on-l). In other words, R denotes
=--1 (becauseyi (6: + 3:) = - cY/3mi, the portion of the ring (area between S, and S,) above the
m0%
x-axis and to the left of 1’. Hence, R is a simply-connected
i=O,l) region.
Consider next the two logarithmic spirals

x,(t>=E,exp[-(u,+jl)t], t>O (5.63)
and
X,(t)=~,exp[-(u,+~l)t], t ” 0. (5.64)
Note that XA(t) and XE(t) correspond to the two shrink-
ing spirals Am’ (starting from A, at t = 0) and E,cc’
(starting from E, at t = 0), respectively, as shown in Fig.
10. It follows from Lemma 5.2(iii) that d” lies on the
extension of the line D,E,.
Since both 1X,( t ) 1and 1X, ( t ) 1shrink exponentially with
-3.0 -P.” -,.o 0.0 I.0 2.0 3.0
the same rate ui, the time tElc it takes XE(t) to go from (a)
E, to c (where it first intersects i) is equal to the time t,,,,
it takes X,(t) to go from d” to d (where it first intersects
r). Note that tElc = t,,,, = L E, D,dn radians), where
L E,D,d is the angle between D,E, and D,d. Since
LE,D,d < 2~7,it follows that d must lie to the left of c,
which in turn must lie to the left of b.
Depending on ul, the continuation of the shrinking
spiral from points d and c may either intersect I’ or 1. Let
this point of intersection be d’ and c’, respectively. Let
t,,,,, denote the time it takes to go from d” to d’ and let
tE1,, denote the time it takes to go from E, to c’. Since
t,,,,, < 2m and tEic, < 2a, both d’ and c’ must lie outside
of S, in Fig. 10, and c’ must be below d’ in view of
Lemma 5.2(i). Hence, d must lie between a and c, whereas
d’ must lie between a’ and c’ in Fig. 10.
Recall next the image under 7~~of the line segment
E,A, 7 {(x(u,u), y(u,u))lu=1,O~u~1} (5.65)
and its extension beyond A,( u > 1) is given by 29
x(t) = x,(1, ~(1, t))exp [-(q + jlbl, t 2 0.
(5.66)
A part of this image is shown by the bold spiral E,ee’ in
Fig. 10 (it corresponds to a part of a in Fig. 6 and
(4.51)). Here, e 2 X(t,) is the point at which X(t) first
intersects I at some time t, and e’ $ X( t2) is the point at
which X(t) first interesectseither I’ or I to the right of D, -3.u -2.0 -1.0 0.0 1.0 2.0 3.0
(if it does not intersect I’) at some time t,. Since both e (4
and e’ lie to the left of x =l, its associatedstarting point Fig. 11. Vt portrait of V and the two bounding circles S, and S,
(which appear as ellipses &e to unequal horizontal and vertical scales).
X,(1, ~(1, t)) must lie to the left of the u =1 line. Hence, The parameters (a, /3, m,, mr) are (a) (10.5,7, -l/7,2/7), (b)
we must have 0 < ~(1, ti) < 1, i = 1, 2, and (8.6,7, -l/7,2/7), (c) (6.5,7, -l/7,2/7).
x1(1> ~(1, ti)) E ‘IE, 7 i =1,2. (5.67)

located inside of S,. Lemma 5.4 guarantees that C,(P)
It follows that e must lie between c and d, and e’ must lie must lie in the simply-connected region
between c’ and d’ in Fig. 10 for all /3 E J.
If we can show that there exists some p E J such that ffb {(x,y)lx~x,, yzo}. (5.68)
C, p q,(C) lies on the bold spiral 2, we will be done. Since C,(p) is a continuous function (Lemma 5.4), the
Since C, is a function of j3 (assuming (Y,m,, and m, are set (assuming without loss of generality 8, < P2>
fixed), we will denote this function by C,(p). Now sup- (5.69)
La {G03)I&a3a32} =H
pose it is possible to find a & E J such that C,(&) is
located outside of S,, and a p2 E J such that C,(&) is is a plane curve (parametrized by p) starting from a point
(p = j?J outside S, and ending at a point (j3 = p2) inside
S,. Since this curve must lie within H, I, ‘must cross the ee’
29Recall from Fig. 6 that the image under ?I of the extension of the spiral at some point &,, & < & < p2. Hence, hypothesis (i)
line segment to the right of A, corresponds to th e extension of the outer
spiral beyond Ai to the right and, hence, must he in the region with u > 1. of Theorem 5.1 is satisfied when B = &,.
It remains for us to show there exist & and p2 with the sponding to /3 =10.5, 8.6, and 6.5 are shown in Fig. 11(a),
above stipulated properties. When /3 = 10.5, we calculate (b), and (c), respectively. It follows from Theorem 5.2 that
(xc, yc) using (5.56) and (5.57) and obtain the double scroll system (l.l)-(1.3) has a homoclinic orbit
when m, = - l/7, m, = 2/7, (Y= 7, and p = 8.6.
]C,(10.5)] = 0.7064 < 0.8 < (E1]e-2no,= 0.9151. (5.70)
2) Using the parameters((Y,/3, m,, m,) = (7,8,6, - $, f),
Similarly, when Jl = 6.5, we obtain we have confirmed by computer simulation the existence
]C,(6.5)] ~1.4155 >1.3> IA,1~1.2477. (5.71) of a double scroll attractor similar to those reported in
HI-[51*
Hence, & = 6.5 and p2 = 10.5 represent one (out of many) 3) Mees and Chapman [15] have also carefully analyzed
valid choice. n the dynamics of the double scroll system (l.l)-(1.3) and
Remarks: confirmed the existencealso of heteroclinic orbits.
1) By computer simulation, we have found the ap- 4) Additional insights and conditions for the appearance
proximate value of PO= 8.6. The V, portrait of V, corre- of the double scroll attractor are given in [20].
Part II: Rigorous Analysis of Bifurcation

Phenomena
VI. BIFURCATION ANALYSIS the locations of the Rossler screw-type attractor and the
By extensive and systematiccomputer simulations of the double scroll attractor. This in-depth analysis in turn leads
double scroll system (l.l)-(1.3) over a wide range of to an algorithm for actually calculating the bifurcation
parameters ((Y,/3, m,, m,), which include those cited previ- boundaries (see Fig. 17(a) and the frontispiece)-hence-
ously in [l]-[6], we have observed two distinct types of forth called the birth and death boundaries-in the LX-~
chaotic attractors, in addition to various stable periodic plane which separatethe double scroll attractors and their
orbits (both period-doubling types and periodic window periodic windows from the other attractors (both chaotic
types). The first type of chaotic attractor is sandwiched and periodic).
between the eigenspacethrough P+ and the eigenspace Before getting into the formal details, examine the typi-
through 0, see Fig. 2(a), and is henceforth referred to as a cal trajectories I’, and I, in Fig. 2(a) again. Note that Ii
Riissler screw-type attractor 3o because it bears a strong and I2 originate from a point on U, to the right, and the
resemblance to a screw-like structure first reported by Zeft, respectiuely,of the boundaryline Lo passing throughA
Rossler [21]. An odd-symmetric image of this attractor has and E. This line therefore bifurcates the set of all trajecto-
also been observed between the eigenspacesthrough P- ries which return to D, from those which continue down-
and 0, as expected. These two Rijssler screw-type attrac- ward to De,. Recall next that all trajectories originating
tors are separatedby the eigenspacethrough 0. The secbnd from U, to the left of L, (passingthrough E and B) must
type of chaotic attractor is the double scroll, which has move down while those on the right of L, must move up.
already been extensively reported [l]-[6] and which spans Finally, note that if ]yi] is large, as is the case when the
all three regions D- 1, Do, and D, in Fig. 2(a). As we Riissler screw-type attractor and the double scroll have
increase the value of a for fixed /3, m,, and m,, we been observed, all trajectories originating on either side of
observe that the two disjoint Rossler screw-type attractors the top eigenspace E’(P) get sucked in rapidly toward
grow in size until eventually they collide and give birth to E’(P) and eventually cross Vi along an infinitesimally
the double scroll [6]. As we increase (Yfurther, the double thin “slit” centered at the line L, passing through A
scroll grows while the co-existing unstable saddle-typeperi- and B.
odic orbit shrinks in size until eventually they too collide We will show shortly that the triangle AABE bounded
with each other and the double scroll disappearsthereafter by the three lines Lo, L,, and L, is crucial in predicting
[6]. This evolution scenario-henceforth called the birth the asymptotic behavior of the trajectories. As before, we
and death of the double scroll-has been found to be quite will switch back and forth into the new reference frames
typical over wide rangesof /3, m,, and m,. corresponding to the D, unit and Do unit in Fig. 2(b) in
Our objective in this section is to use the analytical tools order to take advantage of the analytical equations char-
we have developed in the previous sections to carry out a acterizing the Poincare map m in (4.46) and its associated
rigorous analysis of the above bifurcation phenomena. half-return maps rro in (4.9) and ri in (4.27). Moreover,
Among other things, we will give a rigorous derivation of since it is essential to follow the dynamics originating from
aAoBoEo A \k,(aABE), and taking place in the Do unit
but viewed from the reference frame in the D, unit, the
30For simplicity, we will refer to both “spiral” and “screw” attractors “vi portrait of vo” defined in Section 4.5 (recall Fig. 8)
reported in [6] as a Rijssler screw-type attractor. will play a crucial role in our analysis. In particular, the
dynamics taking place within the Do unit can be “trans-

lated” into the D, unit via the “pull-up map”
vr2e &ro@-‘:~A,B,E, + V,. (6.1)
6.I. Trapping Region

The Vi portrait of V, corresponding to the parameters
(cy,p, mot ml) = (4.0,4.53, -l/7,2/7), which corresponds
to (uo,yo,ul,yl, k) =(-0.721,1.075,0.074, -1.600,0.530),
is shown in Fig. 12. Note that in terms of the local
coordinates (u, u), u =l along A,E, and u =1.53 along
AluE1,, respectively. Recalling Fig. 8, we can identify the
following images under the above pull-up map ~~‘2:
B,CI=7r2(B,AI),~=7r2(qq) (6.2) to Al,
Fig. 12. VI portrait of V. with trapping region .T A AA,,B,E,,.
C,A;:= rz( AA,), a= 772( -WI,) (6.3)
A;:E;:= v2( AI”E1,), Fml= q( FIB,). (6.4)
(see Appendix V) that L
Recall that C, = r2(A1) = r2(E1) and any point on B,F, is B : 9 + Y is a continuous map
defined to be a fixed point we of rz. Let S,,- denote the (1)
“snake-like” area bounded by B,C,,
--
F&, and B,Fl and let (2) a(Y) is a compact (i.e., bounded and closed) subset
S, denote the area bounded by CIA;:, C,E&, and A;:E;:. of .7. (6.9)
We will often refer to S, and S, as “snakes” and call It follows from (6.9) that31
S, p S, U S, = v~(AA,,B,E,,) (6.5) AA n 7r”(.9-) (6.10)
II>0
as the double-snakearea.
Let q lA;,B,E,, denote the fan-like region bounded by is a-invariant in the sensethat
4,B1,B1E1,, and +)=A (6.11)
Gz= T( J%4, >. (6.6) because
Note that the double-snake area S, is bounded within IT( n n”(Y)) =77(.Tn7?(.7)n7r2(.F)n ..-)
q A;,B,E,,. Had we chosenE,,A,, nearer to EIA,, where ?I>0
u is closer to 1, the corresponding fan-like region
q A;,B,E,, could actually cross the double-snake area S,. c~(.T)n~~(~)n~~(.F)n ...
Since a key assumption in our following analysis is that = n 7?(F) =.Tn( n 1r”(.9-))
S, = W,B,E,,, we must choose u to be sufficiently nbl n>l
large. However, as we will seein Section 6.3, u should not
‘be chosen too large either. For the parameters associated = n Irn(.9-) (6.12)
n>O
with Fig. 12, u = 1.53 is a satisfactory choice.
Translating the above definitions back into U, in Fig. and because a(A) 1 A is proved in Appendix V.
2(a), we can interpret the corresponding snake-like area If we define
S 4 \k; l( S,) as the set of all points q; ‘( S,) where retum- AikAulr,(A) (6.13a)
ing trajectories of the type I, originating from AABE
and
intersect the l-7, plane, and the set \k;‘(S,) represent-
ing the odd-symmetric image of the set of all points dPclosureof( u ‘p’[\k;‘(A~)“(-~;‘(A~))]}
where returning trajectories of the type I, originating t>o
from LABE \A ABE intersect the U-, plane. Since (6.13b)
?r,‘(OA;,B,E,,) = AA~,,B~E~,, and since S, c OA;,B,E,,,
it follows that s;‘(S,) c AA~~B~E~,,. Consequently, if we where q’(x) is the flow associatedwith (l.l)-(1.3), then A
restrict our Poincark map n: Vi + Vi to the region can be interpreted as the cZosure32 of the Rbssler screw-type
attractors, or the double scroll, depending on the parame-
9-k AA~~B~E~,, (6.7) ters. We will henceforth call A an attractor of the double
henceforth called the trapping region, then a(Y) C Y. scroll system (l.l)-(1.3).
Hence, we have isolated a small area on Vi’ where the
Poincare map 7~maps into itself. 31We denote the nth iterate of B by T”, e.g., n’(5) 2 .T, a’(Y) d
a(F); 7?(F) P a(fl(Y)), etc.
Since the double-snake area S, does not intersect with 321t is traditional to define an attractor asa closedset. Zf wedo not take
the spiral Fml = I~,(F,B,) except Fl, it can be proved the closure, ii would exclude the origin and, hence, would not be closed.
CHUA et a/.: DOUBLE SCROLL FAMILY 1099
Since 7r”(Y) C interior Y for all n > 2, it follows (see henceforth called the upper entrance gate, or at
Appendix V) that there exists an open neighborhood N(i) \k;l(A,A,,) c L,, henceforth called the lower entrance
of ii which satisfies gate. Likewise, returning trajectories exiting from D, to D,
can do so only through the infinitesimally-thin gate centered
d= n r&v(A)). (6.14)
t,o at \ki-l(&?& henceforth called the upper exit gate, and
(by symmetry of the vectorfield 5) returning trajectories
Hence, ii possessesthe properties of an attractor defined
exiting from Do to D- 1 can do so only through the
by several researchersincluding Hurley [22].
Observe that the region 7 in (6.7) is called a trapping infinitesimally-thin gate - *k; l(m), henceforth called
region of A becauseY is a neighborhood of A and every the lower exit gate.
trajectory originating from Y tends to A under the Poin- We will often abuseour terminology by also calling A,B,,
care map IT. Although there exists some attractor ii in the A,A,,,-- B,C,, and C,A;, the, upper entrance gate, lower
literature which contains no denseorbits,33our computer entrance gate, upper exit gate, and lower exist gate, respec-
simulations strongly suggestthat both the Rijssler screw- tively. Their union As will henceforth be called i-gates.
type and the double scroll attractors contain at least one These gates will play a crucial role in our following bifur-
denseorbit. cation analysis.
The macroscopic structure of ii associated with
(l.l)-(1.3) has been carefully analyzed by computer simu-
lations in [3], where we have discovered that each x = 6.2. Birth of the Double Scroll
constant cross section of d consists of two tightly-wound Our computer simulations in [6] consistently show that
spirals-hence, the name double scroll-for someparame- as cyincreases(for fixed /3, m,, and m,), the two Rijssler
ter values. For example, the double-snake area S, u S, screw-type attractors eventually collide with each other,
defined in (6.5) and shown in Fig. 12 (see also the upper and that the double scroll suddenly emerges after any
snakes S, and S, in Fig. 14(b)) corresponds to the x = 1 further infinitesimal increasein (Y.We will henceforth refer
cross section. to this collision processas the ‘birth of the doublescroll. Our
The microscopic (local) structure of A, however,is much objective in this section is to derive the bifurcation value (Y
more complicated. Indeed since li contains infinitely many which heralds this event.
horseshoesat least for some parameters (recall Theorem A qualitative picture of the structure of a RGssler
5.2), we can expect that the local structure of ii consists of screw-type attractor corresponding to the value of (Yat the
a product between a manifold and a Cantor set similar to collision point is shown in Fig. 13(a). Note that the attrac-
that described in [23]. tor “funnels through” the upperentrancegate AB where its
Observe, however, that if the magnitude of the real extreme left point on U, coincides with A in Fig. 13(a).
eigenvalue j$ at P+ ($ < 0) is very large compared to the Any further increase in (Y would cause this attractor to
real part of the other eigenvalues,then the set A, must be expand with its extreme left point on U, appearing to the
tightly squeezednear the curve34 left of A, thereby causing this trajectory to move down-
- - ward and eventually link up with its twin from the D-,
As A ( A,B, u A,&, ) u (Ku C,A;,). (6.15) region.
The reason responsible for this important property is due Translating this picture into the Vr-plane, we obtain the
. to the strong rate of contraction of the trajectory compo- Vi portrait of V, in Fig. 13(b), where we have assumed35
nent along the real eigenvector E’(P) in Fig. 2(a) on the that m= ?r,(A,E,) intersects the line \k,(L,) = {(x, y)jx
one hand, and the fact that trajectories passing through =l} at A; -- as shown in Fig. 13(b). The snake area S,
points on A, represent the, asymptotic behavior, i.e., long bounded by B,C,, FJ,, and B,F, is tangent to EIQIA;=
after the trajectory component along E’(P) has shrunk to T~(E,A,) at Q,. Since the Rijssler screw-type attractor
an infinitesimal value,.*therebyensuring that the trajecto- above the eigenspaceE ‘(0) is not connected to its twin
ries through A, are literally coasting on the surface of below E’(O), only one snakeS, is shown in Fig. 13(b)36
E’(P) in Fig. 2(a). This mechanism explains why the The ~1~ image of the upper snake S, gives rise to
double scroll in [3] must cross the Vi and U-r plane along another snake-like region S:,A r;‘(S,) in Fig. 13(b). Since
a very thin contour. S, = T;%~(AA,B,E,) = a(aA,B,E,), the lower snake Sa
The above analysis shows that, in so far as computer is the image of the triangular region aA,B,E, under the
simulation is concerned, all trajectories originating from PoincarCmap 7. Consequently, Sl, must be tangent to E,A,
the attractor A can enter Do from D, only through the at Qi = n;r( Q,).
infinitesimally-thin gate centered at 9;l(A1B1) c L,, It follows from the above analysis that the birth of the
double scroll must occur at such a parametervalue that the
upper snake S, is tangent to rl(E,A,). A computer-calcu-
33“A has a dense orbit” means that so-me
trajectory originating from k
visits a neighborhood of every point of A. Roughly speaking, this implies
that numerical errors in computer simulation are sufficient to
guarantee that the entire attractor A will be observed by integrating from 35For some parameter values, vrl(AIE1) may clear the x =l line and
a single initial point. spiral toward F, as in z in Fig. 6.
34For the parameter assumed in Fig. 12, we can replace A,, by A,,. 36Recall two snakes S, and S, are present in Fig. 12 and (6.5).
(b)
Fig. 13. Geometrical structure at the birth of the double SC x011. (a) Macrosconic
.-. nicture of the oriainal svstem. (b) En-
largement of the VI portrait of 5. Ggvr2(B,Al) is tangent to GAiI(m) at Q,. “$ kic’(S,)‘ is an
“infinitesimally” thin set (infinitely many layers compressed into a sheet) whose actual location is very close to A,B,. (c) VI
portrait of b for (a,/~?, q, ml) = (8.8,14.3, -l/7,2/7). GA n2(B,A,) is tangent to Gg q(AIE1) at Q,.
lated example of such a situation is shown in Fig. 13(c), collision event as the deathof the doublescroll and our goal
which corresponds to the parameter values (a, p, m,, m,) is to derive’the parameter (Ywhen this occurs.
= (8,8,14.3, - l/7,2/7). Fig. 14(a) shows the double scroll at the verge of collid-
ing with the periodic orbit r* (shown dotted). Let I?*
intersect U, at point H - in its downward swing and at
6.3. Death of the Double Scroll point H+ in its return upward swing. Note that H- must
Using a “shooting method” [24], we have discovered [3] lie to the right of the line L, becauseas r* moves down
an unstable (saddle-type) periodic orbit actually co-exists through H- in Fig. 14(a), it will first hit U-, and turn
with the double scroll. As we increase CI while fixing /?, around without hitting E ‘( P- ), and eventually hit U- 1 in
m,, and m,, we observe the periodic orbit shrinks while its upward swing at a point & A - H- to the left of e,
the double scroll grows in size. At the parameter a0 (or (odd symmetric image of L,). Hence, H- E LABE.
just below to be precise) where they collide with each Let H; k \k,(H-‘) and Hc k qI,(H+). Since Hc and
other, the double scroll suddenly disappears while the, H; are fixed points of g, we have
unstable periodic orbit continues to exist. We refer to this H; = q(H,-) = n2(H;) (6.16)
1101
Snake Sa
Fig. 14. Geometrical structure at the death of the double scroll. (a) Macroscopic picture of the orii system.(b)
Enlargement of the Vt portrait of V,. H: and H; denote the position of the saddle-type periodic orbit. BrCr d q(BIA1) is
tangent to Exe wl(Elu,Alu,) at Q,. $a U$ is an “infinitesimally” thin set (infinitely many layers compressed into a
sheet) whose actual location is very close to Aluo B,. (c) Vt portrait of V, for (a,/?, ma, ml) = (10.73,14.3, -l/7,2/7).
as shown in the Vi portrait of V, in Fig. 14(b). Note that a EIuOas follows:

double-snake area S, g S, U S, now appearsin Fig. 14(b) Al,O=uoAl+(l-uo)B,, E,,O=uoE,+(l-uo)FI.
because the double scroll in Fig. 14(a) intersects U, on
both sides of the line Lo. The ~1~ image of S, and SJ is (6.18)
shown in Fig. 14(b) by another double-snake area S, d Since El,OA1,Op assesthrough the point H<, q(EluOAluJ
ai ‘( So) and-$;, e k;- l<s,). passes through the point H[ as shown in Fig. 14(b). In
Now, given the cqordinates of H- as obtained by the Appendix VI, we will show that EluOAluOis an excellent
shooting method, we can identify the corresponding local approximation of the stable manifold
coordinates ( uo, vo) of H;, namely
W”( H;) = {x e LA,B,E,(T”(x) + H; as n + 00)
H; = xl(uo, v,). (6.17) (6.19)
From this, we can define the local coordinates of AIuOand that is, W’( H; ) = EluOAIuO.
1102 IEEE TRANSACTIONS ON CIRCUITSAND SYSTEMS,VOL. CAS-33,N0. ll,NOVEMBER1986
Now let A,,,, denote the intersection of the double scroll

attractor with U, and define A, = *,(A,). By definition,
the death of the double scroll occurs when A, intersects
the points Hc (and H;). This condition is equivalent to
the condition that A, touchesthe stable manifolds W”( H: )
because x E W’( H: ) n A, implies H: = lim, + mv“( x)
belongs to W”(H:)n A1 c A,. Since the upper exit gate
G approximates a portion of A, as stated in Section
6.1, the parameter value where m touches fix,=
7ri(Alu,EluJ = W’( H[ ) gives an excellent approximation
of the value at which the double scroll disappears.
The preceding analysis shows that the V, portrait of V,
corresponding to the death of the double scroll must be as
shown in Fig. 14(b). Observethat the upper snake S, must
be tangent to Q, and, correspondingly, the lower snake $
must be tangent to Q;.
To show that the double scroll would disappear if the
oarameter is further tuned so that 0: crosses the stable
manifold W”(H;) = E,, A,, and m%es below E,, A,, ,
we note that in this caie 370theiterates of 0: under i
would eventually leave the trapping region k-and fail to
converge to an attractor within Y.
A computer-calculated Vi portrait of V, corresponding
to the death of the double scroll is shown in Fig. 14(c),
where (a, /3, m,, m,) = (10.73,14.3,-l/7,2/7). The point
Hc is not identified in Fig. 14(c) becauseit is located very
close to the point AiuO.
6.4. Hole-Filling and Heteroclinic Orbits
All the double scrolls given in [llr[5], have a hole
centered at P+ and P- becausethe parameterswere such
that no trajectory in i passes through the point D in
Fig. 2(a), where the real eigenvector E’(P+) hits U,. It
is possible, however, to choose parameters such that D
lies on A. For example, when ((Y,p, m,, m,) = (b)
(9.85,14.3, - l/7,2/7), the corresponding Vi portrait of Fig. 15. A hole-filling double scroll appears when (a,/?, mc, mt) =
V, is as shown in Fig. 15(a). Note that D, = 9,(D) lies on (9.85,14.3, -l/7,2/7). (a) Vt portrait of Ve. (b) The double scroll
with hole-filling orbits.
the lower exit gate a= I~,(A,A,,). Now, assuming38
that the set A has a dense orbit under the “discrete”
Poincart map 8: .? + Y defined in (6.7), then, since hole and is henceforth called a hole-filling orbit. The dou-
C,A;, converges(under a) rapidly to a point in A, p A u ble scroll in Fig. 15(b) is a casein point.
rz( A), it follows that we can make an infinitesimally-small Clearly, another hole-filling orbit exists when D, lies on
perturbation on p so that D, lies on A, A A U rz(A). the upper exit gate G= vi(A,B,).
Under this condition, there exists a trajectory originating Suppose, in addition to D, E C,A;, in Fig. 15(a), the
from D, in Fig. 2(a) which exits U, at exactly the point D. point B; p m;‘(B1) lies on the lower entrancegate A,A,,
Such a trajectory would then follow the real eigenvector in Fig. 15(a).39This implies that B;’ 2 q(Bi) lies on the
E’(P+) and convergesrapidly toward P+. Since P+ is an lower exit gate C,A;,. Now assuming D, lies between B;
“ unstable focus” when restricted to the eigenspaceE “( P+ ), and C, on a, then the hole-filling orbit starting from
it follows that the resulting double scroll will not have a P+ would, after entering D, from above, continue to move
downward and eventually hit U-i at D- = - \k;‘( Dl),
37The unstable manifold W’(H; ) in this case must be a subset of $ where the lower eigenvectorE’( P- ) intersects U-i. By the
because W”(H; ) is an invariant set and the only i_nvariant set in Fig. odd symmetry of [, the return orbit would be a symmetric
14(b) other than W’(H; ) which contains H- is S,,. A more detailed
discussion of the stable and unstable manifol d s of H; and H: is given image and, hence, must exit U, at D. Such a hole-filling
in Appendix VI. orbit is called a heteroclinicorbit.
38This assumption is consistent with all computer simulations of the
double scroll observed so far. Note that the dense orbit here differs from
that associated with A in (6.14): the dense orbit in x pertains to a
“continuous flow,” whereas the dense orbit in A refers to a “discrete 39Note that B; corresponds to the point a2 in Fig. 6 (except that, for
map.” the parameter used in Fig. 6, a2 lies on the upper entrance gate).
CIWA et a/.: DOUBLE SCROLL FAMILY 1103
Since Shilnikov’s theorem also applies when the “homo-

clinic orbit” in the hypotheses is replaced by a “hetero-
clinic” orbit [15], [25], any rigorous demonstration of the
existence of a heteroclinic orbit would also prove the
existence of chaos in the double scroll system(l.l)-(1.3) in
the sense of Shilnikov. Such a demonstration has been
given recently in [15], where a computer-calculated hole-
filling heteroclinic orbit is shown.
6.5. Homoclinic Orbits [26]
We have already proved the existence of at least one
homoclinic orbit through the equilibrium point 0 in Sec-
tion IV. To complete our bifurcation analysis, Fig. 16
shows the Vi portrait on V, associatedwith such a homo-
clinic orbit, where ((u,j?, m,, ml) = (4.1,4.7,-l/7,2/7).
Note that the point C, lies on E,A; as required by Fig. 16. The Vt portrait of Va which give rise to two odd-symmetric
hypothesis (i) of Theorem 5.1. homoclinic orbits through the origin when (a, p, m,,, ml) =
(4.1,4.7, -l/7,2/7).
Homoclinic orbits through the other two equilibrium
points Pi and P- can also occur under appropriate
parameter values. In particular, they occur when one of the
following two conditions is satisfied:
Hopf blfurcatlon occurs @
along this curve, ,/
1. (a) e~Ir;l(B,) lies on the upper entrance gate
A,B,, as shown in Fig. 15(a). Y birth
boundary -$,.’
P
(b) D, lies between 3;’ A ~r,(fi;) and B, on the ,/f’
upper exit gate B,C,.

2. (a) 8’; k rIT-’( B 1) lies on the lower entrance gate
AlAW
(b) D, lies on the upper exit gate G (between
B, and C,).
6.6. Bifurcation Diagram

Using the conditions derived in Sections6.4 and 6.5 for
the birth and death of the double scroll, we carry out a
detailed (double-precision) computer bifurcation analysis
of the a-j? parameter plane (with m, = -l/7 and ml = (a)
2/7). First, we derive the set of all (cy,p) for which the
eigenvalue at P + is pure imaginary, i.e., when a”,= 0. It
turns out that by fixing m, = l/7, this set can be derived
explicitly, namely,
P=(l-m,)a(m,a+l). (6.20)
Substituting m, = 2/7 into (6.21) we obtain curve 0 in
Fig. 17(a). It follows from the Hopf bifurcation theorem
that any parameter (OL,fi) where P+ and P- are sinks (i.e.,
4 < 0 and & < 0) lie above curve 0, henceforth called
the Hopf bifurcation curve, and that for (cy,p) in a small
band to the right of this Hopf bifurcation curve, we can
expect nearly sinusoidal oscillations.
The sets of (a, @) which give rise to the birth and the
death of the double scroll are given by curve @ and curve
0, respectively. It is natural to call curves 0 and @ the
birth boundary and the death boundary,respectively. (b)
It follows from our preceding analysis that those param- Fig. 17. (a) The exact bifurcation diagram on the n-p plane (drawn
eters (a, p) associated with the period-doubling and the with (m , m,) = (-l/7 2/7)). (b) Detailed a-8 bifurcation diagram
derived Prom the 1-D .Po!mcarCmap from Section VII shows a self-simi-
Rossler screw-type attractor must all lie between the Hopf lar structure, where the same pattern (see enlargement in Fig. 18(f))
bifurcation, curve 0 and the birth boundarycurve 0. All appears repeatedly in increasingly smaller clones arbitrarily close to the
origin. The isolated thin color streak represents an artifact of the
parameters associated with the double scroll must lie be- graphics software and is not therefore a part of this figure.
(b)
(4
Fig. 18. The a-/3 bifurcation diagram based on 1-D Poincare map with (ma, m,) = (- l/7,2/7). Each region with the
same color has identical qualitative behavior:
equilibrium point = light blue or black;
period-l = red;
period-2 = orange;
period-4 = yellow;
period-8 = green;
period-16 = blue;
chaos or period greater than 32 = purple;
periodic window with period # 2” = white.
Due to printing imperfection, the “printed” colors may differ from those specified in the above le end. For example, “red”
may appear to be “crimson” and “orange” may appear to be somewhat “yellowish.” (a) a: 0 - 18, ,8: 0-35.(b) a: 6-18, /I:
7-35. (c) a: 9-18, p: 14-35. (d) a: 12-18,/3: 21-35. (e) a: 12-18,/3: 28-35. (f) a: 15-18,p: 28.-35.
CHUA et a/.: DOUBLE SCROLL FAMILY
tween this birth and death boundary. Of course,there exist

numerous periodic windows within the region bounded by
these two boundaries.
Fig. 17(b) shows the a-p bifurcation diagram derived
from an “approximate” one-dimensional Poincare map to
be derived in Section VII. Such a one-dimensional map
yields results virtually indistinguishable from those derived
using the preceding exact but much more time-consuming
analysis. By “zooming” in various regions of Fig. 17(b), we
obtain the five magnified pictures in Fig. 18(b)-(f). The
colors in these pictures correspond to parameter regions lo N,,-
having identical qualitative behaviors, as defined in the
figure caption.
Whereas the a-/I bifurcation diagram in the fron- /
/ ‘v=l
A-7
tispiece gives only boundaries separating regions where (4
certain qualitative behavior could (but need not) occur,
those in Fig. 18 give a much finer structure showing the
chaotic region (in purple) is not a contiguous area, but
rather is interspersed by countless periodic windows, as
predicted in Section VII. This picture also reveals a re-
markable “self-similar” structure at different scalesin the
a-P plane.
VII. ONE-DIMENSIONAL-P• INCAI& MAP
Our analysis in Section VI shows that the qualitative
behavior of the double scroll system (l.l)-(1.3) is de-
termined essentially by the two-dimensionalPoincare map -X
7~ of points on an infinitesimally-narrow “corridor”
centered along the two entrance gates A,& and A,A,, Fig. 19. Geometrical interpretation of the definition of the 1-D
which correspond to the semi-infinite line L: c L, defined Poincare map T*. (a) WI plane in the D, unit. (b) Graph of Q* for
(u,,,y,,,aIryI, k) = (-0.42,0.5,0.15, -l&0.2).
in Fig. 1 to be that part of L, to the left of point B. Since
this “corridor” is “numerically” indistinguishable from L;
when Iv11is relatively large compared to the other eigen- we will define the 1-D Poincart map
values, it is natural therefore to define a one-dimensional
T*: P+N, + p+N,: (7.1)
approximation r* of the Poincare map r by restricting its
domain to L;, and compare its qualitative behavior with In order for r* to be well defined, we must make the
those of 7~.By brute-force computer integration of the following two assumptions:
system (l.l)-(1.3), we have constructed such a 1-D Poin-
care map for many parameter values. Our “numerical” (1) The spiraL= (i.e., the ‘kc1 image of the lower
results show that inspite of the inevitable local truncation exit gate CIA;,) on Vi of Fig. 1 does not intersect
and round-off errors, this 1-D PoincarCmap predicted all the line L, through points E, F, and B.
of the qualitative behavior that we have so far observedby (2) The point D (where the real eigenvectorhits Vi) on
computer simulation (including period-doubling, periodic Vi in Fig. 1 is located on the left-handside of CA’,.
windows) and by rigorous analysis in the preceding sec- To prove that r* in (7.1) is well defined under the
tions (e.g., Rijssler screw-type attractors and the double above assumptions, it is more convenient to translate our
scroll). analysis into the D, unit in Fig. 2(b) via the coordinate
This remarkable observation motivates a more rigorous transformation qk,, which we redraw in Fig. 19(a). Con-
analysis of this 1-D discrete map. In order to do this, it is sider the rectangular region
necessary to describe this 1-D map in analytic form. Our
main objective in this final section is to derive this 1-D w,fi
{(x, y,z) ER31X<0, y=o} (7.2)
map r* and analyze its qualitative behavior. It turns out -
that a much simpler analytical expressionfor YT*is possi- passing through the line segments ON, and G,. Since
ble if we choose the domain of the function v* to be 0 = \k,(P+ ), D, = qk,(D), and Ni = q,(N), it follows that
another semi-infinite line segment-P+N and its extension W, corresponds to the plane W in Fig. 1 passing through
beyond N to N, at infinity as shown in Fig. 1. This line is the two line segmentsp+D and ND.
constructed by connecting the point A4 2 \k;‘(l, 0,O) and Now, in terms of the local coordinates (u, v), points
point P+ by a straight line and extending it beyond N to along the line B,A,, are uniquely identified by a single
oq and deleting the portion P+M in Fig. 1. In other words, coordinate u since v = 1 on this line. In particular, any
point x(u) on this line is describedby Now, if ITi1is relatively large, which is the case in the
double scroll, then the double spiral on W, in reality is
x(u) =q(u,l), O,<u<lifr(u)E B,A,
squeezed into a thin line sitting infinitesimally close to
l<u<ccifx(u)EAiAi,. N,,O. Consequently, for all computation purposes,we can
approximate Y’(u) as the point Y(u) on N,,O. Note that
(7.3) Y(u) is a positive real number given by
Since B,A,, lies on the eigenspace\k,(E’(P+ )), all trajec-
tories originating from B,A,, must remain on the x-y Y(u)=Iy(u)lexp[a,(v-argy(u)], 09~~00.
plane in Fig. 19(a) while spiraling inwards (in backward
time), and must eventually hit ON, (on the negative x-axis) 0.7)
at some point a distance4 X(u) from 0 after a time Since u = u + (1, t ) for 0 < u < 1 is given explicitly by
interval v - 6, where 19A - arg x( u) = -tan-’ [x,(u)/ (4.14) and since u = u- (1, t) for 1~ u < 00 is given ex-
x,.(u)]. Here, x,(u) and x,(u) denote the x and y compo- plicitly by (4.22), we can specify the graph of the Poincare
nent of n(u), respectively. Clearly map ?T* for X(u) > X(0) by the following explicit para-
X(u)=lx(u)lexp[-a,(m+argx(u))] 20. (7.4) metric equations:
(X(u+(l,t)), Y(u+(l,t))), O<t<oo forOdu<l

(7.8)
(x(u)“(u))={
_I (X(24-(l,t)), Y(u-(l,t))), O<t<ocforl<u<cc 1.
Now, Assumption 1 is equivalent to the condition that the Equation (7.8) defines the 1-D Poincare map ?T* for all
lower exit gate C,A;, does not touch or intersect the line X(u) between X(0) and N,,. For points X(u) between
through B,, I;;, E; in Fig. 19(a). It follows from our analy- X(0) and 0, where u < 0,41we simply make use of (7.5),
sis of Figs. 4 and 5 that both inverse-returnfunctions namely,
~‘(1, t) in (4.14) and u-(1, t) in (4.22) are strictly mono- Y(u) = e2”“1-X(u), u<o.
tone functions and, hence, have a unique inverse. Hence, (7.9)
any point X(u) > 0 on Nix(O) maps uniquely into a point We will henceforth call (7.1), (74, and (7.9) the 1-D double
x(u) on B,A,, via the flow cpi, where X(0) is the limiting scroll PoincarP map.
point which maps (under cpi) into B,. Note that any point A typical graph of r* corresponding to the parameters
d, between X(0) and 0 in Fig. 19(a) must map (under cp:) (uo,yo,ul,yl, k) = (-0.42,0.50,0.15, -1.5,0.20) is shown
into a point d,, where in Fig. 19(b). Note that since ui is a constant, the graph
from X= 0 to X= X(0) is always a straight line with a
d, = e*nol. d, (7.5) slope equal to e2nol.Note also that to emphasizethat the
because the expanding logarithmic spiral from dl cannot one-dimensional Poincare map rr* as defined by (7.1),
touch B,A,,. (7.8), and (7.9) is valid not only for system (l.l)-(1.3), but
The upper exit gate G= T~(B,A,) and the lower exit also for the entire double-scroll family of vector fields
< E PO, we use the normalized eigenvalue parameters in-
gate C,A;,= ?r,(A,A,,) are shown in Fig. 19(a). Note that
stead of the usual (cy,/3, m,, m,) in Fig. 19(b).
each point x(u) on B,A,, map under ITSuniquely into a
Translating the Vi portrait of V, in Fig. 19(a) back into
point y(u) with coordinates (y,(u), I,, y,(u)). Now
Fig. 1, we can identify the above 1-D double scroll Poin-
Assumption 2 is equivalent to the condition that the point
care map as
D, in Fig. 19(a) is located below (relative to Vi plane) the
lower exit gate C,A;,. It follows from this condition that VT*: P+N, -p+N,. (7.10)
the flow ‘pi from y(u) must intersect the W, rectangle at The point B’ on P+N is identified with the point X(0).
Y’(u). This translates into Fig.Lto meaet trajectories For each point x EP+B’, n’* is a linear map from m
starting from the exit gates BC and CA’, will always onto P+77*( B’). For points x EB’N,, 77* is a continuous
intersect the plane W= qi-i( W,). Hence, the exit gates nonlinear map from B’N, into p+N,.
B,C, and CIA;, in Fig. 19(a) must map into another We close this paper by exhibiting severaldifferent graphs
double spiral on W, as shown in Fig. 19(a), where each of the 1-D double scroll Poincare map 7r*, which illustrate
point y(u) maps into the various qualitative behavior analyzed in Section VI.
Y’(u) = (-- lu(u>lexp[a,(a-~gy(~))l, 7.1. 1-D PoincarP Map T* for Birth of the Double Scroll
0, yzbbw hb - arg y(u))]) (7.6) The graph of 7r* for the parameter (a, p, m,, ml) =
where arg y(u) 2 tar-i [ y,,( u)/y,( u)]. (8.8,14.3, - l/7,2/7) is shown in Fig. 20(a). Note that the
a We define X(u) as the distance from 0 since we want the domain of 4’For convenience we extend our local coordinate u > 0 to include
R* to be partof the positive real axis. negufive u in order d parametrize the points between X(0) and 0.
maximum value of Y on the interval [0, X(l)] is equal to

X(l), i.e., the point Y( u,J = maxOGu<iY( u) coincides with
the point X(1). Hence, rr*(X(u,)) = X(1) maps precisely
through point A,, where u =l. All other trajectories have
Y(U) < X(1) and, hence, can only enter D, through the
upper gate B,A,. Hence, by definition, the graph in Fig.
20(a) heralds the birth of the double scroll.
7.2. 1-D Poincart?Map v * for Death of the Double Scroll
The graph of r * for the parameter (a, /3, m,, m,) =
(10.73,14.3, - l/7,2/7) is shown in Fig. 20(b). Note that
X, is an unstablefixed point of or* and the maximum
(4 value max Y(U) on the interval [0, X(l)] is equal to X,.
Since X, > X(l), X, corresponds to u > 1. This situation
corresponds to the case where the unstable (saddle-type)
periodic orbit through X, collides with the double scroll. It
follows that the graph in Fig. 20(b) heralds the death of
the double scroll.
7.3. 1-D Poincar6 Map T* for a Hole-Filling Orbit
The graph of rr* for the parameter (OL,p, mO,ml) =
(9.85,14.3, - l/7,2/7) is shown in Fig. 20(c). Note that on
the interval [X(l), ao], the minimum value of Y(u) is zero,
namely, tin, c u< m Y(u) = 0. Since maxgGUGlY(u) >
X(l), the attractor A is a double scroll. Now min Y(u) = 0 .
implies that the spiral through Y’(u) associatedwith this
(b)
point is tangent to the z-axis. This situation correspondsto
the casewhere CA’, in Fig. 1 passesthrough D. Hence, the
graph in Fig. 20(c) is associatedwith a hole-filling orbit..
7.4. 1-D Poincart?Map T * for a Homoclinic Orbit
The graph of n’ for the parameter ((u,p, m,, mi) =
(4.1,4.7, - l/7,2/7) is shown in Fig. 20(d). Note that X(1)
is a fixed point and, hence, Y(1) = rr*( X(1)) = X(1). Since
u = 1 at point A,, this implies that the trajectory originat-
ing from X(1) would enter D, through A, on the stable
eigenspace through 0 and, hence, converges to 0. This
trajectory continues along the unstable eigenvectorthrough
0 until it hits U, at C, which is identified with C, in Fig.
19(a). Since Y(1) = X(l), the trajectory continuing from C,
must intersect II’, at a point Y’(1) whoseprojection Y(1) is
precisely equal to X(1). Hence, this trajectory is a homo-
clinic orbit of the origin, and the graph in Fig. 20(d)
therefore predicts the existence of the homoclinic orbit
proved earlier in Section V.
7.5. Periodic Points of the 1-D PoincarJ Map ?T*
In this section, we will describe the correspondence
between the periodic points of the 1-D Poincare map r*
and the periodic orbits in the double scroll system. The
1-D PoincarC map IT* gives an excellent approximation
under the condition that ]yJ is relatively large compared to
the other eigenvalues,and that A is infinitesimally thin.
(4 This condition implies that each periodic orbit of the
Fig. 20. 1-D Poincark maps corresponding to (a) the birth of the double
scroll when ((Y, /3, mO, m,) = (8.8,14.3, -l/7,2/7); (b) the death of the double scroll system has at least one stable direction (i.e.,
double scroll when (cu,B, m,, m,) = (10.73.14.3, -l/7,2/7); (c) a the magnitude of at least one characteristic exponent is
hole-filling double scroll ahen “(a, p, md, ml) = (9.85,14.3, 1 lj7, ij7); less than one). In particular, a stable periodic point of rr*
(d) the existence of two odd-symmetnc homochnic orbits when
(a,/% mo, m,) = (4.1,4.7, -l/7,2/7). corresponds to a stable periodic orbit and an unstable
1108 IEEE TRANSACTIONS ON CIRCUITS Ati SYSTEMS, VOL. CAS-33, NO. 11, NOVEMBER 1986
‘(4
Do
P-
X;--X,
D-I
d--x(1)
?
n-
(b) Cc)
Fig. 21. Fixed point X1 of B* with 0 < X, < a. (a) Graphof 1-D Poincari map r*. (b) Corresponding periodic orbits in the
original double-scroll system. (c) Abstraction of the main features of (b).
periodic point of rr* correspondsto a saddle-type periodic are summarized in the “abstract sketch” shown in Fig.
orbit of the double scroll. Since Y,, A maxa ( u.,Y( u) 21(c), where N- = - N, a’ = - X(1) and Xr’ = - X,.
corresponds to the outermost orbit of A, if the period-n Case (ii): a < X, < 00.
points {X=(a*)“(X),n*(X),...,(a*)n-l(X)} satisfy Fig. 22(a) shows a fixed point X, and X, = X(u) for
some u > 1. The trajectory originating from X, would
(~r*)‘(x)gY,,, O<i<n-1 (7.11) enter D, through a point on the lower entrance gate AA,,
then the periodic orbit of the double scroll system corre- continue its downward motion until it hits X; = - X,.
sponding to X is located in the attractor A. Define Therefore, we have a period-l orbit as shown in the
a P X(1). abstract sketch in Fig. 22(b).
c7.g (2) Period-2 Point {X, = B *(XI), X, = ?T*(X1)}
As shown later, the type of periodic orbit of the double Case (i): 0 -c X, < X, < a.
scroll system is determined by the position of the point a, Two period-2 points XI and X, satisfying (i) are shown
relative to the periodic points of rr*. in Fig. 23(a). The trajectory originating from X, would
(1) Fixed Point XI = T*(X,) enter D, through the upper entrancegate, return to D, and
Case (i): 0 < XI < a. hit X,. The trajectory continuing from X, would enter D,
Fig. 21(a) shows a fixed point XI of rr * with XI = X(u) again through the upper entrance gate, and eventually
for some 0 < u < 1. The corresponding period-l orbit in return to X,. Therefore, we have a pair of period-2 orbits
the double scroll system is depicted in Fig. 21(b). The as depicted in Fig. 23(b).
trajectory originating from X, would enter D, through a Case (ii): a < X, < X,.
point on the upper entrance gate AB, return to D, and hit Two period-2 points satisfying (ii) are shown in Fig.
X,. By symmetry, we have a pair of periodic orbits as 24(a). The trajectory originating from XI would enter D,
shown in Fig. 21(b). The essentialfeatures of this situation through the lower entrance gate, continue its downward
X(l) x(1)
(a) (a)
N n
Xl II
12
I x
P+
u
P+
m a,,
P-
A
N-
(b)
O’
Xl
Fig. 22. One period-l fixed point XI of n* with a < XI < co. (a) Graph
P l’ 5
P-
6
n-
(b)
x;
Fig. 23. Two period-2 points X, and X, with 0 < XI < X, < a. (a)
of 1-D Poincari map r*. (b) Abstraction of the corresponding periodic Graph of 1-D Poincark map r*. (b) Abstraction of the corresponding
orbits in the original double-scroll system. periodic orbits in the original double-scroll system.
motion through D,, and hit X,l= - X,. Note that X1 and through the upper entrance gate, return to D, and hit X,.
Xi’ (resp. X, and Xi) of the double scroll system are The trajectory continuing from X, would then enter D,,
“identified” as one point X, (resp., X,) in the graph of through the lower entrance gate, pass D,, and hit. Xi =
rr*. The trajectory continuing from Xi would enter and - Xi. The portion of the trajectory from X; to Xi must be
continue its upward motion through DO before returning “symmetric” to the portion of the trajectory from Xi to
to Xi. Therefore, we have a pair of period-l orbits as Xi’ with respect to the origin. Therefore, this situation
depicted in Fig. 24(b), even though V* in Fig. 24(a) seems corresponds to a period-3 orbit in the double scroll system
to suggest that we have a period-2 orbit. It follows from as depicted in Fig. 25(b).
this analysis that the period-doubling of a fixed point X of (3) Period-n Point {X = (n.*)“(X), s*(X); * *,
T* with a < X < co (as in Fig. 22(a)) & Fig. 24(a) corre- c@Y-lN~~
sponds to the splitting of the single “odd-symmetric” Let the above period-n point be ordered as follows:
period-l orbit in Fig. 22(b) into two period-l orbits in Fig. 0 < x, -c x, c * f * -c x, < 00 (7.13)
24(b). Note that each of the orbits in Fig. 24(b) is not odd
symmetric, but the two orbits are odd-symmetric imagesof where we assume X= Xi without loss of generality. Then
each other in view of the symmetry of the vector field. The the type of period-n orbit of ‘T* is uniquely characterized
orbit in Fig. 22(b) exists by itself because it already by a permutation of the indices {2,3,. * *, n } following the
exhibits odd symmetry. index 1. For example, the permutation (1,4,2,3,5) corre-
Case (iii): Xl < a < X2. sponds to the following periodic points:
Two period-2 points satisfying (iii) are shown in Fig. o< (7.14)
25(a). The trajectory originating from Xi would enter DO
(a) (b)
Fig. 24. Two period-2 points XI and X, with a < X, < X, < co. (a) Graph of 1-D Poincark map r*. (b) Abstraction of
corresponding periodic orbits in the original double-scroll system.
(a) (b)
Fig. 25. Two period-2 points X, and X, with $, < a < X,. (a) Graph of 1-D Poincak map ?I*. (b) Abstraction of the
corresponding periodic orbit in the original double-scroll system.
The type of periodic orbit of the double scroll system is 7.6. Feigenbaum’sNumber
therefore determined by the position of the symbol a In Fig. 29, two “period-doubling” bifurcation trees are
among the symbols (0, Xi, X,; . ., X,,, co} along the’half- shown. One is derived by a “brute force” method (in-
line P+N, where P+ may be 0 and N may be 00. Hence, tegrating (1.1) numerically); the other is derived using the
the total number NT of distinct types of periodic orbits of preceding 1-D map. Note that they agree qualitatively.
the double scroll system is equal to Consequently, for computational efficiency, the 1-D map
N,=.(n-l)!X(n+l)=(n+l)!/n. (7.15) was used to compute the associatedFeigenbaum number
[8]. The result ‘was 4.6933.
For example, in the case of n = 3, we have eight different
types of periodic orbits in the double scroll system. Figs. VIII. CONCLUDING REMARKS
26(b) and 27(b) show two periodic orbits corresponding to
the following two “dynamic routes”: We have developed a novel ‘and rigorous method for
analyzing a large family of third-order piecewise-linear
6) 0 < x2-3 < cc (7.16) ordinary differential equations.By an unfolding of the dou-
(ii) O<X~X~~<co. (7.17) ble scroll equation, we have derived an extremely general
(a) (b)
Fig. 26. Three period-3 points XI, X,, and X, with 0 -CX, < X, < a < x3. (a) Graph of 1-D Poincark map T*. (b)
Abstraction of the corresponding periodic orbit in the ongmal double-scroll system.
normal form equation representing a 6-parameter family 9

of continuous piecewise-lineardifferential equations. This
normal form equation is given explicitly and characterizes
completely the qualitative behaviors of a large class of
equivalentpiecewise-linearthird-order circuits.
This paper focuseson an in-depth bifurcation analysisof
a large subclass ~3’~c Y of piecewise-linear equations
which includes the double scroll equation as a special case.
In particular, a rigorous proof of chaosin this equation is
given. The results of this analysis allows us to derive
rigorously the various bifurcation boundaries in the 2-
parameter a-P plane summarizedin the frontispiece.These
boundaries were drawn with m, = -l/7 and m, = 2/7,
and partition the a--/3 plane into regions where different
qualitative behavior could occur. A typical trajectory
simulated at a point in each region (black dots along the
horizontal line /3 = (14)(2/7)) is given in the insets. Each
bifurcation curve is identified by the new phenomenonit is
associated with. Hence, the curve labeled Hopf at P *
means that at any point on this curve, the complex ei-
genvalues at P+ and P- are purely imaginary. Similarly,
as one crosses the bifurcation curve labeled “period 2,”
one observes the onset of the familiar “period-doubling
phenomenon.” The “spiral” curve heralds the onset of the
chaotic Rossler attractor. The curve labeled “window”
corresponds to the onset of one type of (out of infinitely
many) “periodic windows.” Finally, the curves labeled
birth, hetero, death, and homo correspond to the “birth of
the double scroll;” the “birth of the heteroclinic orbit,” the
“death of the double scroll,” and the birth of the “homo-
clinic orbit,” respectively. Observe that the “death” and
the “homo” curves intersect each other and there are
points on the “homo” curve for which the double scroll
attractor can be observed.
0)) Perhaps the most important contribution of this paper is
Fig. 27. Three period-3 points XI, X,, and X, with 0 < XI < n < X, <
Xs. (a) Graph of 1-D Poincark map n*. (b) Abstraction of the corre- the new method given for studying piecewise-linear dy-
sponding periodic orbits in the original double-scroll system. namic circuits. The same approach, for example, can be
used to systematically study other types of chaotic attrac- Note that the neti coordinate system x’ is related to x” by
tors that could arise from other members of the piecewise- x’ = Q;‘x”, where Q, h [ei, e2,eJ]. In the x’-coordinate
linear family 8 not covered in this paper. Two such system, the expression of 5 and U will assume the form
attractors that have already been observedare a “ toroidal” given in (2.3) and (2.4). To seethis, define
attractor, and an attractor similar to the one observedby
Sparrow [31]. Q1p [el~e2~e31
1
APPENDIX I -dm/(12+m2) 0
DERIVATION FORM
OF REAL JORDAN dl/(12 + m2) 0 641.6)
Choose vectors e,, eb, and e, in R3 such that
x”
y,,
0 I d/n
[I [I,,
1) e, is the real part of the complex eigenvectorcorre-
X’
sponding to a”f jG?,
&Q;l
2) eb is the negative imaginary part of the complex Y’
eigenvector corresponding to a”f j3,
=
Z’ Z
3) e, is the eigenvectorcorresponding to 7.
I[1
If we choose Q p [e,, eb, e,], then J = Q-%fQ l/d(12 + m2) m/d(12 + m2) 0 X”
transforms an arbitrary 3 x 3 matrix with eigenvaluesGk - m/d(12 + m2) l/d(12+m2) 0 y” .

jG and 7 into its real Jordan form (see [16, theorem 3, p. 0 0 n/d z”
681).Hence, under this new coordinate system x” = Q-lx, $
assumesthe following real Jordan form: (Al .7)
Then we haveAZ
(Al .l)
where x” = (x”, y”, z”). Moreover, U is representedby

lx” + my” + nz” = d (Al .2)
where l2 + m 2 # 0, n # 0, and d + 0 because U is not (A1.8)
parallel to either eigenspaceand does not pass through the
origin. and
In the new x” coordinate system, the three vectors e,, rn
eb, and e, are transformed into three orthonormal axes, U:(l,m,n) Y: =d (Al .9)
the eigenspacespanned by e, and eb is transformed into 1 1
the x”--y/l plane, and the real eigenvector e, is
[$1=d
transformed into the z”-axis. X’
The U-plane is of course transformed into another plane (l,m,n)Ql (Al .lO)
U” not passing through the origin and is not parallel to
the x”- y” plane. Our next goal is to rotate U” so that it I I
makes a 45” angle with the x”- y” plane, and intersecting
it at x” = l.A1 This can be achieved by choosing yet
another coordinate system x’= (x’, y’, z’) such that the
(d,O,d) ;; =d
[1
(Al.ll)
three orthonormal vectors e:, A [l,O,O], ei A [0, LO], and x’+ z’=l. (Al .12)
e: A [O,O,l] in the x’-coordinate system are transformed
from ei, e2, and e3 with the geometrical property which
achieves the above transformation; namely, (i) make e2 APPENDIX II
parallel to U”; (ii) make e, perpendicular to e2, and such PROOF OF LEMMA 3.2
that the tip of e, lies on U”; (iii) make e, and e2 lie on the
To prove Lemma 3.2, we nes the following lemma.
x”- y” plane; (iv) make ]e,] = ]e,]; (v) make e3= [O,O,d,],
Lemma AZ. 1:a, OB, and OE are linearly independent.
where d, is chosen so that the tip of e3 lies on U”. The
Proof:
above requirements define e,, e2, and e3 uniquely as
Case (i): E # B. Assume that ok can be written as a
follows:
linear combination of 88 and a. Then, since B, E E L,,
e, ~2(d/(12 + m2))[l, m,O] (Al .3) we have A E L,, and so A, E ‘k,( L2). SinceA, = (1, p,,,O),
from the equation of sb(L,) in (2.16), it follows that
e2 A (d/(12 + m2))[ - m, l,O] (Al .4) p0 = uO.Therefore, A ,, = E,. Similarly, from A, E qI(L2),
e3A (d/n)[O,O,l]. (Al .5)

A2Note that Q;?IQ1 = J because by choosing lell = le 1 and el -( e2,
the first two rows of Qi are a product of a scalar and a p;ianar rotation,
A’The choice of 45” and x” = 1 is strictly for convenience. and since the first two rows of J define a planarrotation.
we have p1 = ul, and hence, A, = B,. Therefore, we obtain q,(x) = @1(x - p),
E = A = B, a contradiction.
@l=[A,,B,,El][A-P,B-P,E-P]-‘,
Casei): E = B. Choose a point K on Ul defined by (T
OK= OE+ E(E). Since E E L, and t(E)]] L,, we have (x E Dl). (A2.4)
K E L,. Since E, = (l,u,,O) and B, = (l,ur,O), from the By (2.9) and (2.12), since for i = 0,l
expression of ti (i = 0,l) in (2.9) and (2.12) it follows that
;D9i([(qr,1(x))) = Jx ’ (A2.5)
&,A \k,(K) = \k,(E)+ ‘K&(E))
‘i Yi
1
=E,+~,~0(E,)=(1,u0+9,(u;+1),0). where
[ 1
Hence -1 0
JAl ui 0
~,(K,)=(-~,(a,Z+1),(u~+1)(1+~o~g),O).
O O
Since B = E
we obtain
K,%‘,(K)=\k,(B+5(B))
~I&> = ~o%‘Jo~ox (~2.6)
=B1+d1~,(B)=(1,u,+3,(u~+1),0).
&,(x) =cjl@;lJIQ;(x- P). (A2.7)
Hence
The continuity of 5 is equivalent to the condition
Cl(&) = (-~,(~:+l),(u:+l)(l+u,~,),O).
b,(4 = El&) (~2.8)
Defining the normal vector h k (LO, 1) of K (i = 0, l), we
obtain for all x E U, = D, n D,. Since each x E U, is a linear
combination of A, B, and E in view of Lemma A2.1, the
Wo(Ko)>=-~o(d+l) <o (A2.1) continuity of t is equivalent to the condition that (A2.8)
holds for x = A, B, and E. Substituting x = A, B, E in
(h, &(K,)) = - Gl( u: +1) < 0. (A2.2)
(A2.6)-(A2.8), we obtain
Now (A2.1) implies that the vector &(K,) at the point @,,‘J@,[A, B, E] = M’,;‘J,Q,[A - P, B - P, E - P]
K, E \k,(U,) must point towards the origin of the eigen- (A2.9)
space \k,( E ‘(0)) in the I&, unit in Fig. 2(b), i.e., below V,.
This implies that 5(K) at K E U, must point toward the where X 2 3,/G,. Defining
interior of the D, region, i.e., downwards. However, (A2.2) w,& [A,,B,,E,]-‘J,[A,,B,,E,] 642.10)
implies that the vector tl( K,) at the point K, E \k,(U,)
and
must point towards the origin of ‘the eigenspace‘3!r(E ‘( P))
in the D, unit in Fig. 2(b), i.e., below V,. This implies that W, A [A,, 4, El]-‘Jl [A,, B,, E,] 642.10
t(K) at K E U, must point towards the interior of the D, and using (A2.3) and (A2.4), we can rewrite (A2.9) as
region, i.e., upwards. This is a contradiction and Lemma [A, B, E]W, = X[A - P, B - P, E - P]W,, and hence
A2.1 is proved.
We are now ready to prove Lemma 3.2. Given u = [A,B,E](XW,-W,)=X[P,P,P]W,. (A2.12)
((I,,, y,,, ul,yl, k), choose any [E &.L]. Denote the ei- Substituting the coordinate of Ai, B,, and E,(i = 0,l) in
genvalue parameters of [ by (I& 3,, j$, &3,, yr). Let the (2.20)-(2.24) and (2.26)-(2.31) into (A2.10) and (A2.11),
vector from the origin to the fundamental points we obtain, after some algebraic simplification, the follow-
{A, B, E, P} be denoted by {A, B, E, P}, respectively. By ingi
-l
1 hl-Pcl)‘-(P,2+q -Yoh-PO) -(d+l)
w, = - 0 YOGJO
- PO) 0 (A2.13)
00- PO
i Po2‘tl 0 u,2+1 I
1 kPdZ-(P:+l) +1’+1> -Ylh-Pl)

WI = - P:+l a;+1 0 (A2.14)
a,-Pl
i 0 0 Ylbl- Pl) J
Lemma A2.1, the matrices [A, B, E] and [A - P, B - P, where pi = ui +(uf +i)ki/yi (i = 0,l). Note that y. is
E - P] are invertible. Since the affine maps ‘ki carry determined by only ui, yi, and ki (i = 0,l). Defmmg
{A, B, E} into { Ai, Bi, Ei}, i = 0,1, respectively, ‘ki can ci % ui - pi (i = 0, l), we obtain (1, 1,l) Wi = ( ci, 0,O). Since
be written as follows: [P, P, P] = P(l,l, l), by (A2.12)
%,(x) = Qox, a0 = [A,, B,,, &][A, B, El-‘, [AB,E]@W,-W,)
(x E D,) (A2.3) = XP(l,l,l) W,= XP(c,,O,O) = Xc,[P,O,O]. (A2.15)
The column vectors in (A2.15) can be written as follows: To prove the converse,let p = (a,,, y,,, q, yi, k) be given
such that y,,yi -C0 and k > 0. Using (A2.21), define A,
P=$+4,B,E](Awl-wo)(i,o,o)T (~2.16) ci, W,, W,, (I, m, n) and s by (A2.20), (A2.24), (A2.13),
1 (A2.14), (A2.22), and (A2.25), respectively. Define the
0 = (XW,- W,)(O,l,O)T (A2.17) following four vectors:
A = (l,l,l), B = (1, -(I+ n)/m,l)
o= (AW,-W,)(O,O,1)r. (~2.18)
E= (-(I+m)/n,l,l), P = (O,O,s). (A2.27)
It follows from (A2.13), (A2.14), (A2.17), and (A2.18) that Using (A2.27), we obtain
%-PI u;+1 [A,B,E]W’&[&P,B-P,E-P]W,. (~2.28)
A=y,(I,= (A2.19)
1 Yl(“0 - PO) . This guaranteesthat the vector field 6 defined by
++%[A-P,B-P,E-P]W,[A-P,B-P,E-PI-&P), z>l
9 [A,B,E]w,[A,B,E]-‘x, (ZIG1 (A2.29)
%[A-PJ-P,E-P]W,[A-P,B-P,E-PI-‘(x+P), zg-1
Since k, p y,,( p0 - ~~)/(a~ + 1) in (2.21), (A2.19) implies for x = (x, y, z)‘, is continuou.s.Moreover, we can verify
that the piecewise-linearvector field 5 as defined in (A2.29)
j+-yo (A2.20) satisfies (P.l)-(P.6) in Definition 2.1. Therefore, 5 E 2.
Go’
This proves statement (c). ,
Since h = d,/G,, and since k, =1/k, as stated in (2.34),
APPENDIX III
we obtain
PROOF OF THEOREM 3.3
1
-=ko=-??f!=?!=k. (A2.21) Let {&a, c&,,~~,d1,3i,y1} be given such that 3, > 0,
kl YA f1 3, > 0, and Topi -C0. Put u = (q,, yo, ui, yl, k) A
This proves statement (b) of Lemma 3.2. (60/Oo, To/G70,IF~/L&,~,/Li,, - To/$). As shown in
To prove statement (a), define (A2.22), I = Z(u), m = m(p), and n = n (cl) are given by
(Z,m,n)T= &Vl - W,)(l,O,O)r. (A2.22) (I, m, n) = &(hW,- Wo)(l,O,O)T. (A3 .l)

1
1
This is determined by a,, yo, ui, yi, and k in view of Using c1= ui - p1 = - k,(u; + 1)/y, (by (A2.24)), X =
(A2.13), (A2.14), (A2.20), and (A2.21). It follows from - yo/yIko (by (A2.20)), and k, = k (by (A2.21)), and
(A2.16) that substituting (A2.13) and (A2.14) for Wi, we obtain after
simplification
P=[A,B,E](Z,m,n)T=lA+mB+nE. (A2.23)
I= - kyoy,{2(uoy,k + 01~0)
Hence, (3.10) holds.
To prove statement (c) of Lemma 3.2, note that (2.21) +uoul(k+1)}/{y~(u:+l)} (A3.2)
and (2.27) imply
Ci p lJi - pi = -
Yi
k,(u,‘+l)
(i = OJ).
Using (A2.13), (A2.14), (A2.20), and (A2.24), we obtain

(A2.24)
m= ((y,k+o,)2fl)/(u~+l)
s=l+m+n=1+k3y-f
(A3.3)
(A3.4)
(~;+l)/{~,f(u:+l)}~ (A3.5)
s.4 I+ m + n = $-(l,l,l)(yW,- W,)(l,O,O,)r Defining A i = (5; + 3?)yi and T = 2ti + Tit (i = 0, l), we
1
can rewrite (A3.2)-(A3.5) as follows:
=l-$=‘l+k’y&,;+l),y,2(u;+l). (x42.25) (~3.6)
1
l=?o%(T,-T,)/%
Since by (P.4) in Definition 2.1, P = P+ must be locatedin m=-(~o~l(~l-(~o++l))-~~)/~~ 643.7)
the interior of D,, it follows from (A2.23) that s = I + m + n = {~o~~(~-(~o+~~))-Ao}/A~ (~3.8)
n > 1, that is s =l-~,/a,. (A3,9)
s-1=k3yi+++1)/y;(u;+1) >o. (~2.26)
The vector field .$ defined by (A2.25) has eigenvalues
Therefore, k > 0 holds. Since - yo/ylk =A = 3,/G. > 0, a0* jl, y. (in Do region) ad (&Po)(ul f jl), (4/Go)~l
we have yoyl -C0. This proves that [ E E[p] * yoyl < 0 and (in D, region), becausematrix Wi is similar to 4 (i = 0,l)
k > 0. in (A2.5) and h = &/Go. Hence, the piecewise-linearvec-
tor field
[(x)“cZ,[A-P,B-P,E-P]w,.[A-P,B-P,E-PI-’(X-P), z>l
%,[A,B,E]W,[A,B,E]-‘x, ItI <l (A3.10)
%30,[A-P,B-P,E-P]W,[A-P,B-P,E-P]-’(~+P), Z<-1
where x = (x, y, z)r must have eigenvaluesI& f jL$ and yi region near F,, it follows that
in the Di region (i = 0,l). Substituting (A3.6)-(A3.9) into
~l-‘ls, : S, --f ~4AElu (A5 .l)
(A2.23), (A2.13) and (A2.14) into (A3.10), and expressing
A, B, E, and P in terms of ci, di, and yi (i = O,l), we can is a homeomorphismfrom the compact domain S, into
recast t(x) in (A3.10) in terms of only the six eigenvalue AA~,,B~E~,,.Since .“iJ4A’,BLE,.:~A,.B,Elu --, S, is continu-
parameters. Finally, we can verify, after some involved ous, we have
algebraic manipulations, that (3.32)is equivalent to (A3.10). ~1~: 9- = A&B~&, --, .? (A5.2)
is continuous. Since the image of a compact set under a
APPENDIX IV continuous map is compact, a(Y) is a compact subset of
T~(F~B~)is tangent to B,E, at Fl. Y. This proved (6.8) and (6.9).
Proof: By Theorem 4.3, the spiral Fml= vl(FIB1) (2) Equation (6.9) implies r(h) c A. Hence, to prove
is defined explicitly by (6.11), we only need to prove r(A) 1 A. Take x E A k
Gn;](u(f)B1 + (1 _ u(t))Fl) ;;;;;;;;; ,y x E n”+‘v-‘)y and since T”(y) is
x(t) = e-“I’ Ty:&
[
(A4.1) Y, = 57-1(x)n7r~(s) (A5.3)
where u(t) A ~(0, t), 0 Q t < 00 (U = 0, seeFig. 6) is non-empty and compact. Since
B,= (Ld (A4.2) Y,+i = rr -yx)n7r”+‘(s) c Y, = T-‘(x)(IT”(q

and (A5.4)
r=n n20Yn=nn.or”(Y) is non-empty and
1;1= (Y~(Y~-~~~)/Q,,Y,[~-~,(~,-Y,)~/Q~)=~ 644.3) ;;yr;. Therefore
\ ,
Note that
x=n(Y)~n( npo~n(Y)) =a(A) (A5.5)
x’(t) = ix(t) = e-‘l’ - ul cos t sint
i [ -.Sillt cost I
that is, A c r(A).
+
-sint
[ -cosr
:zt
I) (u(t)4 + (l- dt>)r;,)
(3) In Fig. 12, we can observe that ?r;’ maps S,\{ B,}
into the interior of AA~~B~E~~.However, the point B,
+e-“l’
[ cost
-sint ~~$‘(rm - 4).
maps into the point a2 on B,A,, in Fig. 6. From this we
have
(A4.4) 7r(.F) C { a2} Uinterior .T. (~5.6)
Substituting t = 0 and u(O)= 0 in (A4.4), and making use Since a2 # B,, we have
of (A4.2) and (A4.3), we obtain
7r(a2) = 7rc17r2( a2) E interior Y. (A5.7)
x’(O) = (ul+ u’(O))(B, - r;,>- (A4.5)
Therefore, it follows that
Since y1+ u’(O)is a scalar, x’(0) is a vector in the direction
of B, - Fi, i.e., along the line segmentB,E,. Since x(O) = 7r2(T) C 7r(a2) U r(interior 7) (~5.8)
Fl, it follows that x(t) <istangent to B,E, at Fl when c interior 7 (A5.9)
t = 0.
because a(interior Y) c interior 7. It follows from.lr(Y)
C 7 that, for n 2 2
APPENDIX V
PROPERTIES OF TRAPPING REGION r”(T) C interior 9. (A5 .lO)
(1) In Fig. 6, if x tends to Fl from the inside of the To prove the existence of an open neighborhood N(A)
“curvilinear wedge” region bounded by m and I;,e,, which satisfies (6.14), take a small open ball B(C,) at Cl
n;l( x) tends to F,, and so lim n; ‘( x) = Fl # fi = R; ‘( I;;). such that
However, if x tends to I;; from the outside of this “curvi- a;‘(B(C,)) c interior 7. (A5 .ll)
linear wedge” region, s;‘(x) tends to fi, and so
Cm+(x) =fl = a;’ ( F J. Since the double-snakearea S, Then the set
4 S, U S, in Fig. 12 lies outside of this “curvilinear wedge” IV1p B (C,) U interior S, U interior 9 (A5.12)
R.”
(4
(a)
-+C bc)
(b)
Fig. 29. Bifurcation tree for /I = 15, and OLE [8.43, 8.81. (a) Using the
(b) “brute force” method. (b) Using the 1-D Poincare map.
Fig. 28. A general trapping region corresponding to (LX,/I, ma, ml) =

(4,4.85, -l/7,2/7) and u = 2. (a) VI portrait of V,. ThesnakeA$Z,E[;
intersects the spiral mA q(m), which coincides with the set of
discontinuous.points of rr;‘. (b) Illustration of n(AA,,B,Et,). The S,, where (cq p, mo, ml) = (4,4.85, -l/7,2/7) and u = 2.
snake-like area a(hA,,B,E,,) is actually an “infinitesimal” thin set Note that the spiral Fml intersects the spiral
e.. AZ=
located very near a. rr2(A,A,,) at two points a and b, and the-spiraIE$Y, = r2(
E,E,,) at two points d and c. Since Fml is the set of
is an open neighborhood of A, in the Vi plane which discontinuous points of r;’ (see (6) in Example 4.3), it
satisfies follows that the set r;l(Si) = m(aA,,B,E,,) must be as
depicted in Fig. 28(b),A3 where Ci = or;‘, n( A,,) =
A,= n #(A$). q-‘(A;:), n(E,,) = q’(E[;), vr(F,) = +(Fl) =fi, and
Il>O
a(B,) = vrl~‘(B,) = a, (see Fig. 6), where C,, A;:, E.&‘,are
Choose a small neighborhood N(A) of d in the double indicated in Fig. 12(a). In this case, we expect that Y =
scroll system such that any trajectory originating in N(A) A4444u to be a trapping region and that A =
intersects Vi U K, only at points belonging to the set n n> or”(Y) is a a-invariant compact subset of Y. The
‘k,l( ZVi)u (- \k;‘(N,)), where iVi is defined in (A5.12). proof of this statement, however, is complicated because
Then N(A) satisfies (6.14). we must consider the discontinuity of the map n;‘.
In the more general situation, the double-snake area S,
may intersect the spiral Fml= sl(FIB1). Fig. 28(a) A3The symbol q ( ) .m .Fig. 28(b) denotes a curvilinear region with
shows the Vi portrait of V, with such a double-snakearea boundary points listed inside the parentheses.
CHUA et al.: DOUEXE SCROLL FAMILY 1117
APPENDIX VI [61 T. Matsumoto, L. 0. Chua, and M. Komuro, “The double scroll

bifurcations,” Int. J. Circuit Theory Appl., vol. 14, pp. 117-146,
EluoAluoAPPROXIMATES
Ws(H; ) Apr. 1986.
[71 J. P. Eckmann, “Roads to turbulence in dissipative dynamical
Suppose that the magnitude of the real eigenvque y1 at systems,” Reu. Mod. Phys., vol. 53, no. 4, pt 1, pp. 643-654, Oct.
P+ (yl < 0) is very large compared to the real part of the 1981.
PI fG8:ytanovic, Universality in Chaos. Bristol: Adam Hilger Ltd.,
other eigenvalue.This is equivalent to considering the limit
as yl + - 00. Hence, upon substituting y1= g,/O, and [91 J. Guckenheimer and P. Holmes, Nonlinear Oscillations, Dynamical
Systems, and Bifurcations of Vector Fields. New York: Springer
uI = I?~/& into the coordinates for Fl and E,, and then Verlag, 1983.
taking the limit as p1+ - co, we obtain WI Y. S. Tang, A. I. Mees, and L. 0. Chua, “Synchronization and
(~6.1)
and
E,=(~l(~,-~l-~l)/[(~,-~,)Zfl],y,[l-~,(~,-~~)l/[(yl-~~~~+1])
= (~;(R-~~-pl~l)/[(X-~l)2+~:]~~l~~l-p,(~l-9l)1/[(~l-~l~2+~:l)
+ (1, pl) = A,, as j$ * - 00. (~6.2)
It follows from (A6.1) and (A6.2)‘that chaos,” IEEE Trans. Circuits Syst., vol. CAS-30, pp. 620-626, Sept.
1983.
Elt+,= @, + (1 - uo)F, + uoAl + (1- u,,)B, = AluO. WI A. Rodriguez-Vasquez, J. L. Huertas, and L. 0. Chua, “Chaos in a
switched-capacitor circuit,” IEEE Trans. Circuits Syst., vol. CAS-32,
(~6.3) pp. 1083-1085, Oct. 1985.
WI F. M. A. Salam and S. S. Sastry, “Dynamics of the forced Joseph-
son junction circuit: The regions of chaos,” IEEE Trans. Circuits
Under this condition, the arc Emu, shrinks to one Syst., vol. CAS-32, pp. 784-796, Aug. 1985.
point Eluo= H< = AluO under 7r;l, and therefore also [I31 L. 0. Chua and R. Ying, “Canonical piecewise-linear analysis,”
IEEE Trans. Circuits Syst., vol. CAS-30, pp. 125-140, Mar. 1983.
under 7~.Therefore, the arc l&A;, may be considered as 1141Th. Briicker Differential Germs and Catastrophes. Cambridge,
England: Cambridge University Press, 1975.
the stable manifpld W”(H,+) as $ : - 00, i.e., Ex,= WI A. I. Mees and P. B. Chapman, “Homoclinic and heteroclinic
Ws( H: ). This implies that orbits in the double scroll attractor,” IEEE Trans. Circuifs Syst., to
WI %P”w”: Hirsh and S. Smale, Differential Equations, Dynamical

Eluo4uo = Tl -‘(ExJ ‘- a,l(W”(q+)) = W”(H,-). Systems, and Linear Algebra. New York: Academic Press, 1974.
(~6.4) PIJI L. 0. Chua, C. A. Desoer, and E. S. Kuh, Linear and Nonlinear
Circuits. New York: McGraw-Hill, 1969.
[W R. W. Brockett, “On conditions leading to chaos in feedback
systems,” in Proc. I982 CDC, Dec. 1982.
1191A. Ameodq, P. Coulett, and C. Tresser, “Possible new strange
ACKNOWLEDGMENT attractors with spiral structure,” Commun. Math. Phys., vol. 79, pp.
S73-579,198l.
The authors would like to thank Y. Takahashi from t201 C. Kahlert and L. 0 Chua, “Transfer maps and return maps for
piecewise-linear three-region dynamical systems,” Znt. J. Circuit
Tokyo University for ,any valuable discussions. They are Theory Appt., in press.
also grateful to R. Tokunaga, K. Tokumasu, and K. Ayaki 12110. E. Rbssler, “Continuous chaos-Four prototype equations,”
Ann. N.Y. Acad. Sci., vol. 31, pp. 376-392, 1979.
from Waseda University for their assistance.The bifurca- P21 M. Hurley, “Attractors: Persistence and density of their basins,”
tion diagram in Fig. 18 was pbtaomed bu Yang Lin and Trans. Amer. Math. Sot., vol. 269, pp. 247-271.
1231 R. F. Williams, “Expanding attractors,” IHES Publication, Math.,
Youlin Liao. The bifurcation trees in Fig. 29 and the vol. 43, p. 169-203, 1974.
associated feigenbaum number were calculated by M. ~241 L. 0. C Kua and P. M. Lin, Computer-Aided Analysis of Electronic
Circuits: Algorithms and Computational Techniques. Englewood,
Broucke. NJ: Prentice Hall, 1975.
[251 C. T. Sparrow, “Chaos in a three-dimensional single loop feedback
system with a piecewise-linear feedback function,” J. Math. Analy.
Appl., vol. 83, pp. 275-291, 1981.
REFERENCES
WI P. Glendinning and C. Sparrow, “Local and global behavior near
Dl T,.Matsumoto, “A chaotic attractor from Chua’s circuit,” IEEE homoclinic orbits,” J. Stat. Phys., vol. 35, pp. 645-697, 1984.
Trans. Circuits Syst.; vol. CAS-31, pp. 1055-1058, Dec. 1984. 1271 M. Yuri, “The existence of an invariant stable foliation and the
PI G.-Q. Zhong and F. Ayrom, “Experimental confirmation of chaos problem of reducing to a one-dimensional map,” Tokyo J. Math.,
from Chua’s circuit,” Int. J. Circuit Theory Appl., vol. 13, pp. vol. 6, pp. 247-266, 1983.
93-98, Jan. 1985. WI L. P. Shilnikov, “A case of the existence of a denumerable set of
I31 T. Matsumoto, L. 0. Chua, and M. Komuro, “The double scroll,” periodic motions,” Sow. Math. Dokl., vol. 6, pp. 163-166, 1965.
IEEE Trans. Circuits Syst., vol. CAS-32, pp. 797-818, Aug. 1985. [291 A. Ameodo, P. Coulett, and C. Tresser, “Oscillators with chaotic
[41 T. Matsumoto, L. 0. Chua, and K. Tokumasu, “Double scroll via a behavior: An illustration of a theorem by Shilnikov,” J. Stat.
two-transistor circuit,” IEEE Trans. Circuits Syst., vol. CAS-33, pp. Phys., vol. 27, pp. 171-182, 1982.
828-835, Aug. 1986. 1301 T. Saito, “A chaos generator based on a quasi-harmonic oscillator,”
151 G.-Q. Zhong and F. Ayrom, “Periodicity and chaos in Chua’s IEEE Trans. Circuits Syst., vol. CAS-32, pp. 320-331, Apr. 1985.
circuits,” IEEE Trans. Circuits Syst., vol. CAS-32, pp. 501-503, [311 T. S. Parker and L. 0. Chua, “The dual double scroll equation,”
May 1985. submitted to IEEE Trans. Circuits Syst.
1118 IEEE TRANSACTIONSON CIRCUITSAND SYSTEMS, VOL. CAS-33, NO. II, NOVEMBER1986
Leon 0. Chua (S’60-M’62-SM’70-F’74) was Most recently, he was awarded a Professeur invite at the Ecole Poly-
born in the Phi&pines on June 28, 1936, of technique Federal-Lausanne (Summer 1986), and a Professeur at the
Chinese nationality. He received the B.S.E.E. University of Paris (Fall 1986).
degree from Mapua Institute of Technology,
Manila, the Phillipines, in 1959, the SM. degree
from the Massachusetts Institute of Technology,
Cambridge, in 1961, and the Ph.D. degree from
the University of Illinois, Urbana, in 1964. m
He worked for the IBM Corporation,
Poughkeepsie, NY, from 1961 to 1962. He joined
the Department of Electrical Engineering, Purdue
University, Lafayette, IN, in 1964, as an Assistant Professor. Subse-
quently, he was promoted to Associate Professor in 1967, and to Profes-
sor in 1971. Immediatelv followina this, he ioined the Denartment of Motomasa Komk was born in Tokyo, Japan,
Electrical Engineering and Cornput& Sciences, University of California, on December 15, 1954. He received the B.S.
Berkeley, where he is currently Professor of Electrical Engineering and degree in 1979, the M.S. degree in 1981, and the
Computer Sciences. His research interests are in the areas of general Dr. Sci. degree in 1985, all in mathematics from
nonlinear network and system theory. He has been a consultant to Tokyo Metropolitan Universitv. Tokvo. Janan.
various electronic industries in the areas of nonlinear network analysis, He joined -the Department ‘of Mathematics,
modeling, and computer-aided design. He is the author of Introduction to Tokyo Metropolitan University, as a Postdoc-
Nonlinear Network Theoty (New York: McGraw-Hill, 1969), coauthor of toral Fellow, Japan Society for the Promotion of
the book Computer-Aided Analysis of Electronic Circuits: Algorithms and Science in 1985. From September 1985 to March
Computational Techniques (Englewood Cliffs, NJ: Prentice,;Hall, 1975), 1986, he was with the Department of Electrical
and a coauthor of the forthcoming book Linear and Nonlmear Circuits Engineering and Computer Sciences, University
(New York: McGraw-Hill, 1987). He has also published many research of California, Berkeley, as a Visiting Scholar. Presently, he is on the
papers in the area of nonlinear networks and systems. He was the Guest Faculty of Mathematics, Numazu College of Technology, Shizuoka,
Editor of the November 1971 Special Issue of the IEEE TRANSACTIONS Japan. He is a member of the Mathematical Society of Japan. His
ON EDUCATION on “Applications of Computers to Electrical Education,” research interests are in the areas of topology, dynamical systems, and
the Editor of the IEEE TRANSACTIONSON CIRCUITS AND SYSTEMSfrom ergodic theory.
1973 to 1975, and the Guest Editor of the August and September Special
Issues of the IEEE TRANSACTIONSON CIRCUITS AND SYSTEMSon “Nbn-
linear Phenomena, Modeling, and Mathematics.” He is presently a Deputy
Editor of the International Journal of Circuit Theory and Applications.
Dr. Chua is a member of Eta Kappa Nu, Tau Beta Pi, -and Sigma Xi.
He was a member of the Administrative Committee of the IEEE Circuits 9
and Systems Society from 1971 to 1974, and is a past President of the
IEEE-Circuits and Systems Society. He has been awarded four patents
and is the recinient of the 1967 IEEE Browder J. Thomnson Memorial
Prize Award, the 1973 IEEE W. R. G. Baker Prize Award, the 1973 Best
Paper Award of the IEEE Circuits and Systems Society, the Outstanding
Paper Award at the 1974 Asilomar Conference on Circuits, Systems, and Takashi Matsumoto (M’71-F’85) was born in
Computers, the 1974 Frederick Emmons Terman Award, the 1976 Miller Tokyo, Japan, on March 30, 1944. He received
Research Professorship from the Miller Institute, the 1982 Senior Visiting the B. Eng. degree in electrical engineering from
Fellowship at Cambridge University, England, the 1982/83 Alexander Waseda University, Tokyo, Japan, the M.S. de-
Humboldt Senior U.S. Scientist Award at the Technical University of gree in applied mathematics from Harvard Uni-
Munich, W. Germany, and the 1983/84 Visiting U.S. Scientist Award at versity, Cambridge, MA, and the Dr. Eng. degree
Waseda University, Tokyo, from the Japan Society for Promotion of in electrical engineering from Waseda Univer-
Science. In May 1983, he received an Honorary Doctorate (Doctor sity, Tokyo, in 1966, 1969, and 1973, respec-
honoris causa) from the Ecole Polytechnique Federal-Lausanne, tively.
Switzerland. In July 1983, he was awarded an Honorary Professorship at Presently, he is Professor of Electrical En-
the Chengdu Institute of Radio Engineering in Sichuan, the People’s gineering, Waseda University, Tokyo, Japan.
Republic of China. In 1984, he was awarded an Honorary Doctorate at From 1977 to 1979, he was on leave at the Department of Electrical
the University of Tokushima, Japan, and the IEEE Centennial Medal. In Engineering and Computer Sciences, University of California, Berkeley,
1985. he was awarded the Wasedsa Universitv International Fellowship CA. His interest is in nonlinear networks and fault diagnosis problems.
from’ Japan, the IEEE Guillemin-Cauer Prize, and the Myril B. Reed Best He was an Overseas Associate Editor of the IEEE TRANSACTIONSON
Paper Prize. CIRCUITS AND SYSTEMS.

Chua 86

Uploaded by

Copyright:

Available Formats

Chua 86

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chua 86

Uploaded by

Copyright:

Available Formats

1072 IEEE TRANSACTIONSON CIRCUITSAND SYSTEMS,VOL.CAS-33,~0.

The Double Scroll Family

Part I: Rigorous Proof of’chaos

009%4094/86/1100-1073$01.00 01986 IEEE

Fig. 1. Egenspaces of the equilibria and related sets.

real Jordan form [16] of the new coordinate system,we have*

and such that the equation for U in the new coordinate

Our next goal is to derive the coordinates of each of

particular, it can be shown that Fig. 2 and obtain

where other by a linear transformation. We will now define these

linearly equivalent. This completes our proof of Theorem

Since our main motivation in this paper is to char-

a21 a22 a23

“Throughout this paper, odd-symmetry in R3 means symmetry with

u-u, coordlnete line

ing equivalent form: where

u2- ,.or- ,/” ..._,.,,

4.2. Half-Return Map VT,

defining the affine connectionmap via the formula

X -y,(k,+l)(u,-yo)[u,(u,-y,)+lb Yo(kl+1)[Q,+y,(u,-y,)(ko+1)1 (4.44)

the u = us and u = u,, coordinate lines are always spirals.

(5.36) poGmax(ao)+m= (max(u,2)+1) =0.39<0.4

Hence, to prove (iii) of Lemma 5.2, it is sufficient (5.47)

Consider next the two logarithmic spirals

x1(1> ~(1, ti)) E ‘IE, 7 i =1,2. (5.67)

Part II: Rigorous Analysis of Bifurcation

dynamics taking place within the Do unit can be “trans-

6.I. Trapping Region

as shown in the Vi portrait of V, in Fig. 14(b). Note that a EIuOas follows:

Now let A,,,, denote the intersection of the double scroll

Since Shilnikov’s theorem also applies when the “homo-

upper exit gate B,C,.

6.6. Bifurcation Diagram

tween this birth and death boundary. Of course,there exist

(X(u+(l,t)), Y(u+(l,t))), O<t<oo forOdu<l

maximum value of Y on the interval [0, X(l)] is equal to

normal form equation representing a 6-parameter family 9

transforms an arbitrary 3 x 3 matrix with eigenvaluesGk - m/d(12 + m2) l/d(12+m2) 0 y” .

where x” = (x”, y”, z”). Moreover, U is representedby

e3A (d/n)[O,O,l]. (Al .5)

1 kPdZ-(P:+l) +1’+1> -Ylh-Pl)

(Z,m,n)T= &Vl - W,)(l,O,O)r. (A2.22) (I, m, n) = &(hW,- Wo)(l,O,O)T. (A3 .l)

Using (A2.13), (A2.14), (A2.20), and (A2.24), we obtain

B,= (Ld (A4.2) Y,+i = rr -yx)n7r”+‘(s) c Y, = T-‘(x)(IT”(q

Fig. 28. A general trapping region corresponding to (LX,/I, ma, ml) =

APPENDIX VI [61 T. Matsumoto, L. 0. Chua, and M. Komuro, “The double scroll

+ (1, pl) = A,, as j$ * - 00. (~6.2)

WI %P”w”: Hirsh and S. Smale, Differential Equations, Dynamical

You might also like