You are on page 1of 4

Method of Bifurcation Analysis of Hybrid

Nonlinear Systems
Quentin BRANDON and Tetsushi UETA
Faculty of Engineering
The University of Tokushima
Minamijosanjima 2-1
Tokushima, 770-8506 Japan
Email: {qbrandon,tetsushi}@is.tokushima-u.ac.jp
Dani` ele FOURNIER-PRUNARET
Toulouse University
LATTIS - INSA Toulouse
135 Avenue de Rangueil
Toulouse, 31077 France
Email: daniele.fournier@insa-toulouse.fr
I. INTRODUCTION
Systems with continuous and discrete dimensions, i.e. hy-
brid systems, exist in many domains such as mechanical or
electronic systems. The solution function is usually continuous
but presents points of non-derivability where discrete changes
occur. Some of the most widely studied applications are related
to electrical engineering, such as power converters investigated
as piecewise smooth systems by di Bernardo [1], Tse [2] or
Banerjee [3][4]; or some PLL models as introduced By Acco
[5], refering to this type of model as “hybrid sequential.” Little
work has been done in the analysis of nonlinear hybrid models,
apart from Kawakami, Ueta and Kousaka [7][8]. As for hybrid
linear models, they can be analysed using rigorous analytical
methods as did Kabe [6]. But the nonlinear property of many
systems reserve this option for approximated models only.
Instead, we prefer using numerical methods.
We review the analysis method based on a Poincar´ e map and
introduce some modifications and improvements. We consider
the analysis results of an extended version of the Alpazur
oscillator, using a numerical process to approximate some
derivatives (adapted to the automated CAD tools we are
working on). But the most compelling part is to be found in our
example results, revealing an unusual bifurcation structure: the
interactions between the equilibrium point at some state and
the corresponding switching condition make appear a fractal
bifurcation structure, with an infinite number of bifurcation
curves focusing towards a limit set. As this is due to the hybrid
properties of the system, we may find such structure in many
other hybrid systems.
II. PRINCIPLES
A. Modeling the System
Let us consider a system written by a set of differential
equations defined by smooth functions piecewisely, i.e., for
the state i:
dX
dt
= f
i
(X), (1)
where,
X(t) =
_
_
_
x
0
.
.
.
x
n−1
_
_
_ ∈ R
n
. (2)
Within each state there is a solution function such as:
X
i
(t) = ϕ
i
(t, X
i0
) with X
i
(0) = X
i0
, (3)
where X
i0
is an initial value.
Now, based on the switching rules and the definition of a
period of our system, we set the Poincar´ e map, placing its
sections at the switching points. We assume for now that
the switching conditions for each state can be expressed
as a function of the system variables, e.g., for the state i:
q
i
(X) = 0.
The map is therefore expressed as:
Π
i
= {X
i
∈ R
n
| q
i
= 0}
T
i
: Π
i
→ Π
i+1
X
i
→ X
i+1
= X
i

i
) = ϕ
i

i
, X
i
(0)).
(4)
We are now able to perform a local analysis over each partial
orbit delimited by those Poincar´ e sections. Thus the Poincar´ e
mapping is defined as a differentiable map:
T =

T
i
. (5)
We finally apply a projection from R
n
to R
n−1
due to the
equation of the final state switch condition q
m−1
= 0 as seen
in Ref. [7]:
p : Π
0
→ Σ
0
X
0
→ U
0
,
(6)
which we use as a discrete definition of our system Fig.??.
B. Analysis approach
In order to compute phase portraits for a given system,
we simply integrate the differential equations. The switching
conditions previously defined are constantly verified in order
to use the variational equations matching the current state. A
proper adaptive stepsize Runge Kutta integration gave us good
performance along with precious control of the numerical error
Ref. [9].
For the computation of fixed points, we integrate the partial
derivatives of the differential equations in parallel of the
solution:
d
dt
∂ϕ
i
∂X
i−1
=
∂f
i
∂X
∂ϕ
i
∂X
i−1
where
∂ϕ
i
∂X
i−1
(0) = I. (7)
Fig. 1. Abstract representation of the Poincar´ e map
We compute the Jacobian matrix of each local map:
dX
i
dX
i−1
=
∂ϕ
i
∂X
i−1
+
dX
i
dt
∂τ
i
∂X
i−1
, (8)
thus we can express the Poincar´ e map:
dX
m
dX
0
=

dX
i
dX
i−1
. (9)
We then apply the projection p and obtain the Jacobian:
DT
l
(U
0
) =
dU
m
dU
0
. (10)
For a fixed point we solve:
T
l
(U
0
) = U
m
= U
0
. (11)
We inject the approximation of the tangent to the solution DT
l
in the Newton method in order to compute an accurate value
of U
0
. If our initial value is in the domain of convergence, we
converge toward the solution within a few iterations.
Finally, concerning the critical values required to compute
bifurcation diagrams, we use the characteristic equation of our
system in order to determine an extra equation. Depending
on the desired bifurcation diagram, we choose λ, one of the
parameters from the diagram space, as an extra degree of
freedom to compute a solution of our new equations set:
χ
l
(µ) = det(DT
l
−µI
n−1
) = 0. (12)
µ is defined by the bifurcation type. So our new problem can
be written:
F(U
0
, λ) =
_
U
m
−U
0
χ
l
(µ)
_
= 0. (13)
Which means we are to compute the following Jacobian
matrix:
_
¸
¸
_
dU
m
dU
0
−I
n−1
dU
m

dDT
l
dU
0
dDT
l

_
¸
¸
_
(14)
The second row of Eq. (14) requires the second derivatives of
the solutions of variational equation (7). Computing a number
of second derivatives can be done two ways: one based on an
analytical approach, consisting in simply deriving once more
the first derivative elements; the second one is more numerical
since it consists in approaching the tangent by differentiation,
performing a multiple integration using shifted input variables.
The first approach might appear more elegant but depending on
the switching conditions, even the simplest models will involve
very elaborated equations at this step. Let us now consider the
second alternative: we compute the following elements the
same way we do in the fixed point algorithm:
U
m
(U
0
, λ), U
m
(U
0
+ ∆U, λ), U
m
(U
0
, λ + ∆λ),
DT
l
(U
0
, λ), DT
l
(U
0
+ ∆U, λ), DT
l
(U
0
, λ + ∆λ).
(15)
We derive the Jacobian matrix elements by differentiation:
dU
m
dU
0

U
m
(U
0
+ ∆U, λ) −U
m
(U
0
, λ)
∆U
dU
m


U
m
(U
0
, λ + ∆λ) −U
m
(U
0
, λ)
∆λ
dDT
l
dU
0

DT
l
(U
0
+ ∆U, λ) −DT
l
(U
0
, λ)
∆U
dDT
l


DT
l
(U
0
, λ + ∆λ) −DT
l
(U
0
, λ)
∆λ
.
(16)
The second approach introduces some tricky questions such as
how to determine a relevant ∆U or ∆λ, but it turns out to be
much easier to implement and gives satisfying performances.
Though it is not a much refined approach, we use at this point
a fraction of the correction of U and λ as our next ∆U and
∆λ.
III. EXTENDED ALPAZUR OSCILLATOR
A. Model description
We extend the orignial Alpazur oscillator [8] to 3-state for
examination of the computing method proposed in Ref[7] from
a universality point of view.
Fig. 2. Electronic implementation of 3-state Alpazur oscillator
The continuous variable is represented by: X =
_
x
y
_
According to the position of the switch, we have, for each
state (i = {1, 2, 3}), the following differential equations:
f
i
(X) =
_
f
i
(x, y)
g
i
(x, y)
_
For State i: (terminal a(i = 1),b(i = 2), or c(i = 3) in SW)
_
dx
dt
= f
i
(x, y) = −rx −y
dy
dt
= g
i
(x, y) = x + (1 −A
i
)y −
1
3
y
3
+B
i
.
(17)
where x and y are normalized variables corresponding to the
current i and voltage v (detailed in Ref[8]); and where A
i
and
B
i
are real numbers.
In order to determine the discrete behavior, we establish the
following switching rules:
q
1
= y −h
q
2
= y −b
q
3
= y −m,
(18)
combined with the switching order:
State 1 → State 2 → State 3 → State 1.
Such system exhibits chaotic orbits at particular parameter
values such as the following set:
r = 0.1 A
1
= 0.2 A
2
= 2.0 A
3
= 0.8 B
1
= −0.6
B
2
= 1.0 B
3
= −0.1 m = −0.1 b = −0.3 h = −1.0
This particular model is very convenient because the mapping
is strait forward: even within the local maps, y values are
fixed thanks to the switching conditions. We can therefore
extract the mapped variable u = x. We define the map:
T = T
0
◦ T
1
◦ T
2
: Π
0
→ Π
0
x
0
→ x
3
= ϕ(τ, x
3
).
(19)
B. Fixed points
As previously exposed, the problem of Fixed points is as
follows:
x
3
−x
0
= 0. (20)
In order to compute the appropriate correction, we need to
compute:
DT
l
=
dx
3
dx
0
=
dx
3
dx
2
dx
2
dx
1
dx
1
dx
0
. (21)
For each State i we compute:
dx
i
dx
i−1
=
∂ϕ
i
∂x
i−1
+f
i
(x
i
, y
i
)
∂τ
i
∂x
i−1
∂τ
i
∂x
i−1
=

∂φ
i
∂x
i−1
g
i
(x
i
, y
i
)
,
(22)
where we numerically integrate the required elements:
d
dt
_
x
i
y
i
_
=
_
f
i
(x, y)
g
i
(x, y)
_
State 1: from x
0
State 2: from x
1
State 3: from x
2
d
dt
_
¸
_
∂ϕ
i
∂x
i−1
∂φ
i
∂x
i−1
_
¸
_ =
_
¸
_
∂f
i
∂x
∂f
i
∂y
∂g
i
∂x
∂g
i
∂y
_
¸
_
_
¸
_
∂ϕ
i
∂x
i−1
∂φ
i
∂x
i−1
_
¸
_
where
_
¸
_
∂ϕ
i
∂x
i−1
∂φ
i
∂x
i−1
_
¸
_(0) =
_
1
0
_
.
(23)
We now use the Newton method to compute the correction to
be applied:
x

0
= x
0

x
3
−x
0
dx
3
dx
0
. (24)
C. Bifurcations
In this case the characteristic equation is fairly simple:
χ(µ) = det(DT
l
−µ) = 0, (25)
hence the Jacobian matrix:
_
¸
¸
_
∂x
3
∂x
0
− 1
∂x
3
∂λ
∂DT
l
∂x
0
∂DT
l
∂λ
_
¸
¸
_
(26)
We compute its elements using the equation (16). Through
combination with conventional methods, we obtained the
bifurcation diagram Fig.??.
−1.6 −1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0
−2
−1.5
−1
−0.5
0
0.5
1
1.5
B1


B2
B1

G I
Fig. 3. Bifurcation diagram in the B
1
/B
2
parameter plan
−1.05 −1 −0.95 −0.9 −0.85 −0.8 −0.75 −0.7 −0.65
−1.1
−1
−0.9
−0.8
−0.7
−0.6
−0.5
B
Stable At SN bif. Unstable
At PD bif.
Dbl period
At limit set
I
I
G
At PD bif.
B
B1

G
Fig. 4. Enlargement of the Fig.?? at invariant set area
An application to collision bifurcation analysis, based on this
method, is detailed in [7]. Collision bifurcations occur when
the trajectory of the solution collides with one of the switching
conditions, under the influence of a parameter change.
IV. RESULTS
We first computed the values of x
0
against B
1
for fixed
points at a constant value of B
2
. By monitoring the eigen-
values, we can find the points at which we obtain bifurcation
values. For example, we could find eigenvalues corresponding
to saddle-node bifurcation (µ = 1, marked in red on the curve)
at each point where

dB1
= 0 (note that eigenvalues corre-
sponding to period doubling bifurcation are usually very close,
hence easy to obtain as well). Once these values gathered, we
can use a tracing algorithm to compute bifurcation curves.
We obtain the bifurcation diagram in Fig.??. This method
can be illustrated by the Fig.?? and ?? where we can see a
combination of the bifurcation lines and the fixed points line.
−1.6
−1.4
−1.2
−1
−0.8
−0.6 −2
−1
0
1
2
0.6
0.8
1
1.2
1.4
1.6
1.8

B2
B1
X
Period
Doubling
Saddle-
node
Limit
set
Fixed pts
@ B2=1.92
Fig. 5. Bifurcation diagram put against the fixed points set at B
2
= 1.92
This actually corresponds to the solution wrapping itself
around the equilibrium point of the current state. The switch-
ing characteristic of the system makes it possible for the
solution to return to such unstable regions after a period, a
little bit like a Rossler type oscillator.
In the neighborhood of a particular value B
1
= B

1
, we can
find a self-repeating bifurcation pattern Fig.??.
Due to precision limitations inherent to numerical methods, it
gets more and more difficult to compute the bifurcation lines
as we get closer to the limit value, but we can conjecture this
structure is infinite.
One can notice that this limit corresponds to the parameter
values where the equlibrium point of State 1 and the initial
point X
0
merge. The precise limit value of B
1
can be
easily obtained by injecting the previous information into the
equation set of the system: for state 1 (hence in eq.(1)), the
equilibrium point (
dx
dt
= 0 and
dy
dt
= 0) is on the switching line
(y = m). Given this, we can extract x
0
and B
1
corresponding
to the limit set. With our parameter values, we obtain x
0
= 1
and B

1
=
−2761
3000
. The limit set itself is represented by the
X
B
1
B
2
B
1
Fixed points
SN bifurcation
Limit set
Fig. 6. Schematic view of Fig.??.
green line in the bifurcation diagram Fig.?? and ??. There
are an infinity of bifurcation curves because for each one that
we find, there exists a closer one on the other side of the
line, which corresponds to the piece of the solution at state
1 turning an extra 180 degrees around the equilibrium point
(making the initial point closer to this latter).
REFERENCES
[1] di Bernardo, M.; Tse, C., “Chaos in Power Electronics: An Overview,”
Chaos in Circuits and Systems, Series B, Vol. 11, pp. 317–340, 2002.
[2] Tse, C., “Modeling of Switching Power Converters for Nonlinear Dy-
namical Analysis,” Complex Behavior of Switching Power Converters,
pp. 57–80, 2004.
[3] Banerjee, S.; Chakrabarty, K., “Nonlinear Modeling and Bifurcations in
the Boost Converter,” IEEE Trans. Power Electronics, Vol. 13, No. 2, pp.
252–260, Mar. 1998.
[4] Banerjee, S.; Karthik, M.S.; Yuan, G.H.; Yorke, J.A., “Bifurcations in
one-dimensional piecewise smooth maps - theory and applications in
switching circuits,” IEEE Trans. Circuits and Systems I, Vol. 47, No.
3, pp. 389–394, Mar. 2000.
[5] Acco, P., “Etude de la boucle ` a verrouillage de phase par impulsions de
charge,” PhD. thesis INSAT, No. 707, 2003.
[6] Kabe, T.; Parui, S; Torikai, H.; Banerjee, S; Saito, T., “Analysis of Piece-
wise Constant Models of Current Mode Controlled DC-DC Converters,”
IEICE Trans. Fundamentals of Electronics, Communications and Comp.
Sci., E90-A, 2, pp. 448–456, 2007.
[7] Kousaka, T.; Ueta, T.; Kawakami, H., “Bifurcation of switched nonlinear
dynamical systems,” IEEE Trans. Circuits and Systems II: Analog and
Digital Signal Processing, Vol. 46, No. 7, pp. 878–885, Jul. 1999.
[8] H. Kawakami and R. Lozi, “Switched Dynamical Systems – dynamical
of a class of circuits with switch –,” Proc. RIMS Conf. “Structure and
Bifurcations of Dynamical Systems,” ed. by S. Ushiki, World Scientific,
pp.39–58, 1992
[9] Brandon, Q.; Ueta, T.; Kousaka, T.; Fournier-Prunaret, D., “Local Bi-
furcations of Nonlinear Hybrid Systems,” Nonlinear Theory and its
Applications (NOLTA), pp. 160–163, 2007
[10] Kitajima, H.; Kawakami, H., “An algorithm tracing out the tangent
bifurcation curves and its application to Duffings equation,” IEICE
technical report. Nonlinear problems,Vol. 94, No. 44, pp. 1–7, May 1994