Documente Academic
Documente Profesional
Documente Cultură
November 4, 2015
Contents
1 Introduction
1.2 An example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Definitions
15
17
21
31
44
1
1.1
Introduction
Evolution equations
When we modelize the dynamics of cellular phenomena, we are interested into the evolution of concentrations of the products and substrates as a function of time. This variation
is described by a time derivative. Such an evolution equation typically involves synthesis,
degradation and transport terms:
dX
= synthesis of X - degradation of X + transport of X
dt
(1)
Generally, the transport and degradation of X are functions of X but also of other variables
of the system. When we modelize a biological system, we are usually faced with a system
of several coupled equations. These equations are called dierential equations because
they involve derivatives. Since the chemical interaction laws and enzyme kinetics are
generally non-linear (e.g. Michealis-Menten, Hill), these equations are said non-linear
dierential equations.
1.2
An example
Figure 1: Scheme of the model proposed for the dynamics of calcium (Dupont G, Berridge
MJ, Goldbeter A, 1991, Signal-induced Ca2+ oscillations: properties of a model based on
Ca2+ -induced Ca2+ release, Cell Calcium 12, 73-85).
4
Evolution equations for the cytosolic Calcium (variable Z) and the intravesicular Calcium
(variable Y ) are given by:
dZ
Zn
Ym
Zp
= v0 + v1 vm2 n
+
v
+ kf Y kZ
m3
dt
K2 + Z n
Krm + Y m Kap + Z p
dY
Zn
Ym
Zp
= vm2 n
v
kf Y
m3
dt
K2 + Z n
Krm + Z m Kap + Z p
(2)
The analysis of these equations shows that the model can account for the occurrence of
self-sustained oscillations of Calcium and allows to study the eect of various parameters,
such as the stimulus strength, on the period of these oscillations (fig. 2).
2
2.1
Definitions
System of dierential equations
The evolution of two interacting variables (concentrations of the species) can be described
by a system of two dierential equations:
dX
= f (X, Y, , )
dt
(3)
dY
= g(X, Y, , )
dt
The functions f and g usually depend on variables (concentrations X and Y of compounds, mRNA, proteins,...) and on some kinetics parameters (Michaelis-Menten constant, maximum rate, degree of cooperativity,..., denoted here by and ). Because
of their high non-linear degree (due to the presence of terms X 2 , XY , Michaelian, or
Hill), finding an analytical solution of this system is rarely feasable. Consequently, other
approaches have to be developped to solve such systems.
2.2
Steady state
dX
=0
dt
dY
=0
dt
(4)
the concentration of X and Y will not evolve: X and Y are constant. The system (4) is an
algebraic system of equations and we denote the solution (XS , YS ) (where the subscript
S stands for steady state):
%
f (X, Y, , ) = 0
(5)
g(X, Y, , ) = 0
The steady state is reached when the synthesis (production) and the degradation (consumption) terms are compensated (see eq. (1)). Note that the equilibrium state is a
particular steady state which is specific to equilibrium reactions.
If the initial conditions do not correspond to the steady state (X0 = XS , Y0 = YS ), we
expect that the system will evolve until it reaches a steady state satisfying the condition
4 (or, equivalently, 5), as illustrated in figure 3.
But is it always the case? Will the system always evolve monotonously towards a steady
state? Before treating this question, we will first define the notions of phase space,
trajectories and nullclines, that are very convenient for our discussion.
Variable
XS
Time
2.3
Another way to represent the dynamical behaviors than the temporal evolution is the
representation in the phase space. This is a multidimensional space whose coordinates
are the variables (concentrations) of the system.
Figure 4 examplifies the representation in the phase space (X, Y ) of the evolution toward
the steady state. The starting point, represented in white, is called the initial condition
(X0 , Y0 ), i.e. the concentations X and Y at time t = 0. The black point is the steady
state (XS , YS ) and the curve is the trajectory of the system. This trajectory is built by
calculating at each time step the values of X and Y .
(X ,Y )
S
Variable Y
(X0,Y0)
Variable X
Variable Y
Variable Y
dX/dt=0
dY/dt=0
Variable X
Variable X
2.4
Cauchy-Lipschitz theorem
Under some continuity and regularity hypotheses (usually satisfied in biological ODE
models), the theorem of Cauchy-Lipschitz states that from a given initial condition, a
system of dierential equations has one and only one solution, i.e. there is a single trajectory that the system can follow. A direct consequence of this property is that a trajectory
can not cross itself or another one. This also implies that, once the initial condition is
specified, the behaviour of the system is completely determined. The behaviour is said
deterministic.
To study the behavior of the system around the steady state, we resort to the linearized
stability principle. We describe this method first for a one-variable and subsequently for
a two-variable system. This theory can be generalized to a N-variable system.
3.1
(6)
where f (X) is a function taking into account both the synthesis and the degradation
kinetics of X.
By definition, the steady state satisfies the condition:
dXS
= f (XS ) = 0
dt
(7)
(8)
(9)
with:
=
&
x = xet
(12)
'
(13)
df
dX
and x = x(0)
Equation (12) describes the (exponential) evolution with respect to time of the perturbation x: if > 0, x will increase and the system will leave its steady state, which is said
unstable. On the opposite, if < 0, the perturbation will damp in such a way that the
system will reach back its steady state; which is said stable.
3.2
Consider now a system of two variables described by the following coupled equations:
dX
= f (X, Y )
dt
(14)
dY
= g(X, Y )
dt
dXS
= f (XS , YS ) = 0
dt
dYS
= g(XS , YS ) = 0
dt
(15)
(16)
Y = YS + y
dXS dx
dx
dX
=
+
=
= f (X, Y )
dt
dt
dt
dt
dY
dYS dy
dy
=
+
=
= g(X, Y )
dt
dt
dt
dt
10
(17)
Since the system (14) is non-linear, we proceed to a Taylor development (we neglect the
terms of order superior or equal to 2):
&
'
&
'
&
'
&
'
df
df
df
df
x+
y + ... =
x+
y
f (X, Y ) = f (XS , YS ) + dX
dY S
dX S
dY S
S
'
&
'
&
'
&
'
&
(18)
dg
dg
dg
dg
g(X, Y ) = g(XS , YS ) +
x+
y + ... =
x+
y
dX S
dY S
dX S
dY S
We call the Jacobian matrix for the steady state (XS , YS ) the matrix:
&
'
a11 a12
J=
a21 a22
where
a11 =
&
df
dX
'
, a12 =
&
df
dY
'
, a21 =
&
dg
dX
'
, a22 =
dx
= a11 x + a12 y
dt
dy
= a21 x + a22 y
dt
&
(19)
dg
dY
'
(20)
S
(21)
x = xet
y = yet
(22)
(23)
a21 x + (a22 )
y=0
(24)
i.e.
2 (a11 + a22 ) + (a11 a22 a12 a21 ) = 0
(25)
We note that a11 + a22 = T = trace of the Jacobian matrix and a11 a22 a12 a21 = =
determinant of the Jacobian matrix
Equation (25) has in general two roots:
=
1+
T T 2 4
2
11
(26)
By replacing (26) in (22), we obtain the general form of the solution of the system (21):
%
x(t) = x1 e+ t + x2 e t
(27)
y(t) = y1 e+ t + y2 e t
The values of x1 , x2 , y1 and y2 can be found given the initial condition x(0) and y(0).
The presence of the square root in (22) implies that the (called the eigen values) can
be complex. We discuss separately both cases:
If the eigen values are dierent and real:
It can be shown that the solution (27) can be written as (see Appendix):
%
x(t) = Aept cos qt + Bept sin qt
y(t) = A ept cos qt + B ept sin qt
(28)
This is the parametric equation of a spiral. Indeed this system can be rewritten as:
%
x(t) = ept (cos qt + )
(29)
y(t) = ept (cos qt + )
where , , , and are function of A, B, A and B .
The imaginary part (q) is responsible of the oscillations around the steady state. In
this case, the steady state is called a focus. If p < 0 the oscillations will be damped,
while if p > 0 the oscillations will be amplified. Thus, if p < 0, the steady state is
stable, while if p > 0, the steady state is unstable.
12
In summary...
In summary, the stability describes the way in which the system reacts to a small perturbation that moves the system slightly away from the steady state. If the system is
exactly at its steady state, it will not evolve; X and Y are constant and equal to XS and
YS respectively. If we apply a small perturbation of these values, two situations can arise:
either the system comes back on its steady state, or it definitively leaves this state. In
the first case, the steady state is stable, in the second case, it is unstable.
Unstable steady state
perturbation
Variable X
Variable X
damping
Time
amplification
perturbation
Time
3.3
More generally, for a system involving N variables, the characteristic equation is of degree
N and must often be solved numerically. Except in some particular cases, such an equation
has N distinct roots that can be real or complex. These values are the eigenvalues of the
N x N Jacobian matrix. The general rule is that the steady state is stable if there is no
eigen value with a positive real part. It is sucient that one eigenvalue is positive for the
steady state to be unstable.
Remark: Very often, because of the complexity of the model (several coupled and nonlinear dierential equations), the analytical solution (for the trajectories far from the
steady state, the limit cycles or chaotic attractor) cannot be obtained. So, we resort to
numerical simulations, using a computer.
13
Eigen value
T2-4!
Steady state
Phase plot
y
Im
Re
T<0
!>0
T2-4! > 0
stable node
T>0
!>0
T2-4! > 0
unstable node
T>0
ou
T<0
!<0
T2-4! > 0
saddle
(unstable)
T<0
!>0
T2-4! < 0
stable focus
T>0
!>0
T2-4! < 0
unstable focus
T=0
!>0
T2-4! < 0
center
(marginal stability)
Figure 9: According to the nature of the eigen values, several approaches to the steady
state can be observed.
Figure 10: Depending on the sign of the trace T and of the determinant of the Jacobian
matrix and on the sign of T 2 4, several approaches to the steady state can be observed.
14
Dynamical behaviors
What happens if the steady state is unstable? How does the system behave? Where do
the trajectories go in the phase plane? Several behaviors are possible (see more examples
in section 6). We describe here some of them, which have important consequences in
biology.
4.1
If there is a single stable steady state and if the system can not explode, the only possible
behavior of the system is to evolve towards this steady state, whatever the initial condition
(provided that it is not on an unstable steady state) (fig. 11).
X, Y
X
Y
Time
4.2
Bistability
The system can have several stable steady states. Then depending on the initial condition,
the system will evolve towards one or the other steady state (see section 6.3). If two steady
states coexist, this is a particular case of multistability, refer to as bistability (fig. 12).
X starting from (X ,Y )
01
01
02
X, Y
(X01,Y01)
02
Time
15
4.3
Limit-cycle oscillations
The system can exhibit a new type of behavior, such as sustained oscillations. In the
phase space, these oscillations correspond to a close curve called a limit cycle (fig. 13).
X, Y
X
Y
Time
4.4
The system can exhibit complex (but still deterministic) behaviors, such as chaos (fig.
14). Such complex behaviors require a phase space with a dimension >2 (i.e. more than
2 variables) and multiple instability mechanisms (obtained via multiple feedback loops in
regulatory systems).
Y
Time
16
5
5.1
For a given system, the linear stability analysis allows us to determine the regions in the
phase space where the steady state is stable or unstable. We can represent these domaines
graphically in the parameter space (also called two-parameter bifurcation diagram):
parameter
coexistence between
two stable steady
states (bistability)
parameter !
5.2
A bifurcation diagram shows how a steady state, and more generally the behaviour of
the system changes, as a control parameter varies. The following figures illustrate dierent
types of bifurcations that can be encountered.
A transcritical bifurcation (fig. 16) is characterised by the passage from a stable steady
state to an unstable steady state. At the same point, an other steady state, switches from
unstable to stable.
Figure 18: Hysteresis with reversible (left) and irreversible (right) transitions.
A pitchfork bifurcation is characterised by the passage from one stable steady state to
an unstable steady state and the simultaneous appearance of two stable steady states (fig.
19). This kind of bifurcation is encountered in particular systems, usually composed of
two coupled subsystems. This bifurcation is not very robust: a slight asymetry between
the subsystems leads to a symetry breaking: the pitchfork bifurcation is lost and a saddle
node (limit point) appears (fig. 20).
19
Figure 25: Period doubling cascade leading to chaos, trajectory in the phase space.
20
Applications
6.1
Consider one compound, X, synthetised with a rate constant ks and degraded at rate kd :
k
s
d
(30)
(31)
(32)
(33)
This means that the perturbation will be damped, in an exponential way, with respect to
time. The steady state is therefore stable.
Analytical solution: Note that this system is relatively simple and the solution of equation
(30) can be found analytically2 :
X = X0 ekd t +
)
ks (
1 ekd t
kd
(34)
where X0 = X(0)
We can check that
lim x(t) =
Recall that:
1
p+qX dX
1
q
ln(p + qX)
21
ks
= XS
kd
(35)
6.2
Consider now one compound, X, synthetised with a logistic synthesis rate and linearly
degraded:
&
'
dX
X
= F (X) = ks X 1
kd X
(36)
dt
N
Steady states:
Xs1 = 0
Xs2 = N
ks kd
(exists if ks > kd )
ks
(37)
(38)
F = ks X
(39)
(40)
(41)
(42)
The following table summarizes the stability conditions for each steady state:
Xs1
Xs2
ks < kd
stable
!
ks > kd
unstable
stable
Graphical analysis:
22
6.3
Consider a system involving two compounds which activates each other according to a
Hill-type kinetics:
Evolution equations:
dX
Yn
=
v
k1 X
1 n
dt
+Yn
Xn
dY
= v2 n
k2 Y
dt
+ Xn
dX
Yn
= n
X
dt
+Yn
n
dY = X
Y
dt
n + X n
(43)
(44)
Steady state:
Lets = 1/2. We can check that XS = YS = 1/2 is a steady state. We also see that
XS = YS = 0 is also a steady state. Note that those two states may not be the only
solutions...
Representation in the phase plane:
We can represent the nullclines in the phase space. The nullclines are defined by:
dX
Yn
=0 n
=X
dt
+Yn
Xn
dY
=0 n
=Y
dt
+ Xn
(45)
(46)
23
X
Y
Y
Y
(47)
Figure 27: Phase space and nullclines corresponding to the system (44) for n = 1 (left
panel) and n = 4 (right panel).
Recall that:
&
Xn
n + X n
X
'
(X n ) (n + X n ) (X n ) (n + X n )
nn X n1
=
=
(n + X n )2
(n + X n )2
(48)
If = 1/2:
&
Xn
n + X n
X
'
=(
n 21n X n1
)
1
n 2
+
X
n
2
1
X
2
+ 21n
2n
(49)
(50)
The trace T and the determinant of the Jacobian matrix are: T = 2 and = 1n2 /4
and T 2 4 = 4 4(1 n2 /4) = n2
We can see that the trace T is always negative and that T 2 4 is always positive.
if n < 2, we have > 0 the steady state (1/2, 1/2) is thus a stable node.
if n > 2, we have < 0 the steady state (1/2, 1/2) is thus an (unstable) saddle
point.
24
Bifurcation diagram
In the previous analysis, we assumed v1 = v2 . We can now sketch the behaviour of
the system when v2 is varied. Fig. 28 shows how the nullclines are aected when v2 is
decreased. In particular we see that bistability is lost when v1 is below a critical value,
around v2 = 0.8. Fig. 29 shows the steady state of X as a function of v2 (obtained by
numerical simulation). When v2 > 0.8, the system displays bistability: two stable nodes
are separated by a (unstable) saddle point. When v2 < 0.8, the single (stable) steady
state is the trivial state (0,0). The bifurcation point, at which the upper stable state and
the middle saddle point coalesce and disappear is called Saddle-Node (SN) bifurcation.
Through numerical simulation it is also possible to follow the SN bifurcation point in the
(v1 , v2 ) plane (fig. 30). Such stability diagram shows the region of bistability as a function
of these two parameters.
v2=1
0.6
0.4
0.4
0.2
0.2
0.5
0.8
0.6
v2=0.7
0.8
0.8
v2=0.85
0.6
0.4
0.2
0.5
0.5
stable node
1.4
1.2
1
SaddleNode
0.8
bifurcation
0.6
unstable saddle
0.4
0.2
stable node
0
0.2
0.4
0.6
0.8
1.2
1.4
1.6
1.8
Bistability
1.4
v2
1.2
1
0.8
SN bifurcation
0.6
0.4
(X =0, Y =0)
S
0.2
0
0.2
0.4
0.6
0.8
1.2
1.4
1.6
1.8
6.4
1
A
2
B+X
Y+C
3
2X + Y
3X
4
X
Evolution equations
The concentrations of A and B are supposed to be constant and noted a and b, respectively.
The evolution equations of the concentrations X and Y are:
dX
= k1 a k2 bX + k3 X 2 Y k4 X
dt
(51)
dY
= k2 bX k3 X Y
dt
dX
= a bX + X 2 Y X
dt
(52)
dY
= bX X Y
dt
Steady state:
XS = a
(53)
YS = b/a
X
X
J =
Y
X
X
4
5
b + 2XY 1 X 2
Y
=
Y
b 2XY
X 2
Y
(54)
&
b 1 a2
b a2
26
'
(55)
The trace T and the determinant of the Jacobian matrix are: T = b 1 a2 and
= a2 .
The characteristic equation is:
2 T + = 0
(56)
(57)
The determinant is always positive. The trace T is positive if b > a2 + 1 and negative
otherwise. T 2 4 is negative if (a 1)2 < b < (a + 1)2 and positive otherwise.
The following table summarizes the dierent possible behaviors as a function of the parameters a and b:
b
T
T 2 4
type of
steady state
+
+
stable
node
(a 1)2
+
0
+
stable
focus
a2 + 1
0
+
-
+
+
unstable
focus
(a + 1)2
+
+
0
+
+
+
unstable
node
When parameter b increases, the steady state turns from a stable node to a stable focus,
then, it losts its stability and the steady state turns from an unstable focus to an unstable
node. When the steady state is unstable, the system then evolves towards a limit cycle.
These behaviours are shown in figure 31 (results obtained by numerical integration of the
dierential equations):
For a = 2 and b = 0.5: the system evolves towards a steady state (stable node)
For a = 2 and b = 4: the system evolves towards a steady state (stable focus)
For a = 2 and b = 6: the system leaves its steady state (unstable focus) to reach a
limit cycle (sustained oscillations)
For a = 2 and b = 12: the system leaves its steady state (unstable node) to reach a
limit cycle (sustained oscillations)
The stability diagram showing the stable and unstable regions as a function of a and b
is given in figure 32. The solid curves, satisfying b = a2 + 1 delimits the stability region.
The dotted curve corresponding to b = (a + 1)2 and b = (a 1)2 separate the node from
the focus in the stable and unstable regions respectively.
Figure 33 is a bifurcation diagram: it shows how the steady state of X changes as a
function of the parameter b (for a fixed to 2). Figure 34 shows how the period varies in
the oscillatory domain. These diagrams have been obtained numerically.
27
0.4
2.5
X
Y
a=2, b=0.5
0.3
1.5
X, Y
1
0.5
0
0.2
0.1
10
15
20
Time
2.5
2
2
X, Y
X
Y
a=2, b=4
1.5
1
0
1
0
10
15
20
Time
8
a=2, b=6
X
Y
X, Y
2
0
0
0
10
20
30
40
50
10
Time
25
a=2, b=12
25
X
Y
20
15
15
X, Y
20
10
10
5
0
0
0
10
20
30
40
50
20
40
Time
Figure 31: Dierent kinds of behaviors obtained for the Brusselator model with dierent
parameter (a and b) values. Left panels: time evolution. Right panels: phase space.
Note that the period at the bifurcation can be calculated. We know that the frequency
is given by the imaginary part
of the eigen values. At the bifurcation point, T = 0 and
the eigen values are 1 = 2 = = a. The period is therefore equal to 2/a. For a = 2
and b = 5, the period is thus equal to 3.14.
28
12
unstable
node
parameter b
10
unstable
focus
stable
focus
stable
node
0.5
1.5
2.5
3.5
4.5
parameter a
maximum of X
(oscillations)
X or X
max
or X
min
14
12
10
8
Hopf
bifurcation
6
4
minimum of X
0
10
12
Parameter b
Figure 33: Bifurcation diagram for the Brusselator as a function of parameter b (with
a=2). The steady state looses its stability and limit cycle oscillations emerge at b = 5
(Hopf bifurcation).
14
12
Period
10
10
12
Parameter b
Figure 34: Period for the Brusselator as a function of parameter b (with a=2). At the
bifurcation point (b = 5), the period is = 3.14.
29
(X-nullcline)
(58)
(Y-nullcline)
are shown in figure 35. They delimit regions in the phase space where the vector field has
a particular direction (figure 36).
9
X nullcline
Y nullcline
8
7
6
5
4
stable steady
state (focus)
3
2
1
0
0.5
1.5
2.5
3.5
9
8
7
6
5
4
3
2
1
0
0.5
1.5
2.5
3.5
30
7
7.1
Appendix
Detailed study of the perturbation evolution in a 2-variable
system
The evolution of the perturbation around the steady state in the case of a 2-variable
system is given by equation (27):
6
x(t) = x1 e+ t + x2 e t
y(t) = y1 e+ t + y2 e t
(59)
Note that + and are related to the trace and the determinant of the Jacobian matrix
(see Fig. 9).
As explained in the text, two cases should be distinguished: (a) The eigen values +
and are real, or (b) they are complex conjugated. We describe the evolution of the
perturbation for these both cases in a bit more details.
(a) + and are real
In order to sketch the trajectories of x and y in the plane (x, y), it is convenient to calculate
the ratio dy/dx:
dy
y1 + e+ t + y2 e t
=
(60)
dx
x1 + e+ t + x2 e t
we can rearrange eq (60) as
dy
y1 + + y2 e( + )t
=
dx
x1 + + x2 e( + )t
(61)
(62)
This is the equation of a straight line crossing the origin (0,0) (see line d1 on fig. 37) that
describes the asymptotic approach of the evolution of the perturbation.
We can also calculate the limit
y2
dy
=
t dx
x2
lim
(63)
This is the equation of a straight line crossing the origin (see line d2 on fig. 37) that
describes the slope of the trajectories of the evolution of the perturbation initially.
Similarly, if > + > 0, the trajectories will leave the steady state, initially being
dy
y1
asymptotic to the line d1 :=
=
dx
x1
When and + have opposite signs, the trajectory approaches the point (0,0) along the
line d1 but then goes away following the line d1 (see fig. 38).
31
Figure 37: Evolution of the perturbation in the case + < 0 and < 0 (stable node).
Figure 38: Evolution of the perturbation in the case + > 0 and < 0 (saddle).
(b) + and are complex conjugates
When the eigen values + and are complex conjugates, lets say +
= p iq, equations
59 can be rearranged as
6
x(t) = Aept cos qt + Bept sin qt
(64)
y(t) = A ept cos qt + B ept sin qt
which can then be rewritten as
6
(65)
(66)
(67)
(68)
A2 + B 2 = K 2 K = A2 + B 2
tan =
32
(69)
(70)
K = A2 + B 2
(71)
Equation (65) is the equation of a spiral in the plane (x, y). If p < 0, the spiral converges
to the origin (0,0) and the trajectory corresponds to damped oscillations of x and y with
the time (fig. 40). If p > 0, the spiral diverge from the origin (0,0) and the trajectory
corresponds to amplified oscillations of x and y (not shown).
If + and are purely imaginary (p = 0), then the oscillations are not damped, neither
amplified. Their amplitude remains constant and depends on the perturbation applied to
the system. In this particular case, Eq. (64) becomes
6
x(t) = A cos(qt) + B sin(qt)
(72)
y(t) = A cos(qt) + B sin(qt)
which can be written as:
x(t) = K cos(qt + )
y(t) = K cos(qt + )
(73)
Equation (73) is the equation of a close curve in the pane (x, y) and corresponds to
self-sustained oscillations of x and y. Note that its amplitude and period, respectively
governed by K and are depending on A, B, A and B , i.e. on the initial conditions of
the perturbation x(0) and y(0).
Figure 39: Evolution of the perturbation when + and are complex conjugates ( =
p iq) with real part p > 0 (stable focus).
33
Figure 40: Evolution of the perturbation when + and are purely imaginary (center,
marginal stability).
7.2
A 3-variable system has three eigen values. Again, the type of behavior can be characterized as a function of the position of the these eigen values in the Re/Im plane. Five
non-degenerated cases can be distinguished: (1) the three eigen values are real and negative (stable steady state), (2) the three eigen values are real, Two of them are negative
(unstable steady state), (3-4) two eigen values are complex conjugates with a negative
real part and the third one real is negative (stable steady state) (two cases can be distinguished depending on the relative value of the real part of the complex eigen values
and of the real one, and (5) two eigen values are complex conjugates with a negative real
part and the third one real is positive (unstable steady state). The cases (3) and (5) are
illustraed on fig. 41.
Figure 41: Evolution of the perturbation for a 3-variable system (a) when two eigen values
are complex conjugates with a negative real part and the third one is real and negative
(stable steady state) (left panel) and (b) when two eigen values are complex conjugates
with a negative real part and the third one is real and positive (unstable steady state)
(right panel).
34
7.3
dX1
= f1 (X1 , X2 , ...XN )
dt
dX2
= f2 (X1 , X2 , ...XN )
dt
...
(74)
dXi
= fi (X1 , X2 , ...XN )
dt
(75)
with i = 1, ...N
or in the vectorial form:
dX
= f(X)
dt
(76)
(77)
This a N x N matrix.
J=
f1
X1
f2
X1
...
fN
X1
f1
f1
...
X2
XN
f2
f2
...
X2
XN
fN
fN
...
X2
XN
(78)
(79)
X=XS
(80)
(81)
If N = 2, we obtain the quadratic equation discussed in details for the case of a 2-variable
system (remember that in this case, a1 = T race(J) and a2 = Det(J)). When N > 2, it
is dicult (usually impossible) to find all the roots of equations (81). We can nevertheless
obtain informations about their signs (and hence about the stability of the steady state).
Suppose 1 , 2 , ...N all (known) eigenvalues of the linearized system:
dX
= JX
dt
(82)
(83)
If one eigen value has a real part > 0, the peturbation will ultimately increase and the
steady state is thus unstable. How to determine if all the eigen value have a negative
real part? This can be done by checking some conditions, known as the Routh-Hurwitz
criteria.
We can define N matrices as follows:
H1 =
Hj =
a1
a3
a5
...
1
a2
a4
a1
, H2 =
0
a1
a3
0
1
a2
&
a1 1
a3 a2
...
...
...
0
0
0
a1 1 0
, H3 = a3 a2 a1 , ...
a5 a4 a3
a1 1 0 ... 0
a3 a2 a1 ... 0
, ..., HN =
...
0 0 0 ... aN
'
(84)
Then, all eigen values have negative real parts (steady steady stable) if and only if the
determinants of all Hurwitz marices are positive:
det(Hj ) > 0
(85)
Robert May (1973) summarized the stability criteria in the cases N = 2, N = 3, and
N = 4 as follows:
N
Conditions for stability
N = 2 a1 > 0, a2 > 0
N = 3 a1 > 0, a3 > 0, a1 a2 > a3
N = 4 a1 > 0, a1 a2 > a3 , a4 > 0, a1 a2 a3 > a23 + a21 a4
An application of the Routh-Hurwitz criteria to a 3-variable system is given in EdelsteinKeshet, pp.234-235.
36
7.4
The cusp bifurcation (also called cusp catastrophe) is a bifurcation, observed in 2-parameter
space, at which two branches of saddle-node bifurcations vanish. The cusp bifurcation is
associated to the presence of bistability and to hysteresis phenomenon.
Because the cusp bifurcation can only be seen in a 2-parameter plane, it is called a codimension 2 bifurcation. Another example of co-dimension 2 bifurcation is the degenerate
Hopf bifurcation at which a super-critical Hopf bifurcation becomes sub-critical. In a
way these co-dimension 2 bifurcations show how a bifurcation changes when a second
parameter is changed. Further examples of co-dimension 2 bifurcations are described
in [Guckenheimer and P. Holmes (1983) Nonlinear Oscillations, Dynamical systems and
Bifurcations of Vector Fields. Springer] and briefly reviewed in the appendix of the paper
by Tyson et al, Biophys J (2006).
37
7.5
Poincar
e-Bendixson theorem
It is dicult (usually impossible) to proof the existence of a limit cycle for any N-variable
non-linear dynamical system. In the simple case of a 2-variable system, the PoincareBendixson theorem, based on the topology of the phase space, provides however a way
to determine if a limit cycle is present. The Bendixson criteria can sometimes also be
applied to rule out the existence of a limit cycle.
Poincar
e-Bendixson theorem
If, for t t0 , a trajectory is bounded and does not approach any singular point (steady
state), then the trajectory is either a closed periodic orbit or approaches a closed periodic
orbit, i.e. a limit cycle, for t (Fig. (43)).
(86)
The procedure is in two steps. First we need to find a trapping region for the system.
Then we need to find conditions on parameters a and b under which the steady state is
unstable.
Finding a trapping region requires an analysis in the phase space. First we find the
nullclines:
x
x = 0 y =
(87)
a + x2
b
y = 0 y =
(88)
a + x2
These nullclines are sketched in Fig. 44. They delimit various regions of the phase space
distinguished by the direction of the vectors.
38
39
The get the right intuition, consider x and y in the limit of very large x. Then
x = x2 y
y = x2 y
so,
x
=1
y
(89)
(90)
(91)
along trajectories.
Hence the vector field is roughly parallel to the diagonal line. This is confirmed by a more
rigorous calculation, which consists of comparing the size of x and y:
x (y)
= x + ay + x2 y + (b ax x2 y) = b x
(92)
y > x if x > b
(93)
Said otherwise,
This inequality implies that the vector field point inward on the diagonal line, as illustrated
in Fig. 45, because dy/dx is more negative than -1, and therefore the vectors are steeper
than the diagonal line. Thus the region is the trapping region.
(94)
We then find the steady state (i.e. the intersection between the two nullclines):
xs = b
b
ys =
a + b2
(95)
(96)
&
1 + 2xy
a + x2
2xy
(a + x2 )
'
(97)
b4 + (2a 1)b2 + (a + a2 )
a + b2
Hence the fixed point is unstable if T > 0, i.e. when b2 = 21 (1 2a 1 8a). This
defines a curve in the parameter space (a, b), as shown in Fig. 47.
The determinant is = a + b2 > 0 and the trace is T =
Remark: The Poincare-Bendixon theorem is one of the central results in nonlinear dynamics. It says that the dynamical possibilities in the phase space are very limited: if a
trajectory is confined to a close, bounded region that contains no fixed points then the
trajectory must eventually approach a close orbit. Nothing more complicated is possible.
This result depends crucially on the 2-dimensionality of the plane. In higher dimensions,
The Poincare-Bendixon theorem no longer applies. Something new can happen: trajectories may wander around forever in a bounded region without setting down to a fixed point
or a close orbit. They converge to more complex attractors (e.g. chaos or quasi-periodic
torus).
Bendixson criteria:
dx
dy
F G
Lets call
= F and
= G. If
+
= 0 for all x and y and does not change sign
dt
dt
x
y
in D, there is no close orbit (no limit cycle)
The proof - which requires advanced mathematics - can be found in Edelstein-Keshet p.
379-380.
42
7.6
Poincare map
20
18
16
14
12
10
10
11
0.144
0.146
0.148
12
Poincare map
0.17
0.16
0.15
x2
0.14
0.13
0.12
0.11
0.1
0.13
0.132
0.134
0.136
0.138
0.14
y1
43
0.142
0.15
References
44