preprint

Targeted Activation in Deterministic and Stochastic Systems

Bryan Eisenhower and Igor Mezic
Department of Mechanical Engineering,
University of California Santa Barbara
(Dated: March 6, 2009)
Abstract
Metastable escape and transitions are ubiquitous in many physical systems and are becoming
a concern in engineering design as these designs become more and more inspired by dynamics of
biological, molecular, or other natural systems (eg. swarms of vehicles, coupled building energetics,
nanoengineering, etc). We study an example of a chain of coupled bistable oscillators which has
two global conformation states. Specically we investigate re-conformation between these two
states and how specialized or targeted disturbance eects the transition process. We derive a
multi-phase averaged approximation to these dynamics which illustrates the inuence of actions in
modal coordinates on the coarse behavior of this process. An activation condition that characterizes
temporal rate in relation to initial energy is then derived. This result emphasizes spatial sensitivity
and the impact of targeted activation on reconformation. We also nd an unexpected inverse energy
cascade through these actions, funneling energy to one coarse coordinate leading to the global
transition. The prediction tools are derived for deterministic dynamics while we also present
analogous behavior in the stochastic setting and an ultimate divergence of Kramers activation
behavior under targeted activation conditions.
PACS numbers: Valid PACS appear here
1
A chain of strongly coupled oscillators, each of which oscillates in a bistable potential, is a
coarse representation of many physical systems from macromolecules to Josephson junctions
which have multiple conformations. Switching between these conformations can occur by
coordinated motion of the entire chain over a global energetic barrier [1]. Characterizing this
global barrier with respect to local barrier heights is essential if this structure is to be used
for engineering design. We present such results for a deterministic harmonically coupled
oscillator chain with a strong backbone.
The class of nonlinear systems of coupled oscillators that we study are close to a coupled
chain of linear harmonic oscillators. Such near-integrable systems have been studied in [2]
where equipartition of energy and dynamic properties related to integrable instability theory
of partial dierential equations were investigated [3].
The Hamiltonian for an N oscillator system which has a strong linear backbone and local
nonlinearity takes the form
H =
N

i=1
1
2
q
2
i
+
1
2
B(q
i1
q
i
)
2
+U(q
i
) (1)
where q
i
R, B is the linear coupling strength and U is a local bistable nonlinear potential.
We have studied many such potentials, including a Morse potential acting on pendula (see [1],
[4], [5]) which is aliated with conformation change in molecular systems but its exponential
form makes analytical progress challenging. In this study we investigate a 2-4 potential
U(q
i
) =
_
1
2
q
2
i

1
4
q
4
i
_
(2)
where is a constant that separates the two nonzero equilibria (q
eq
= {0, }) to control the
curvature of the saddle to better capture dynamics of proteins, see [6]. This potential has
the same behavior as the Morse potential while being easier to analyze. The backbone has
periodic boundary (q
N+1
= q
1
), and is much stronger than the local nonlinearity (B ).
When this is the case the transition process is collective and coordinated as shown in Figure
1 (and as described in [1]).
In systems where the strength of the coupling is of the same order as the strength of
the local nonlinearity, neighboring oscillators do not restrain each other from excursions
across local barriers. These excursions or breathers initiate the transition process of the
entire chain. In [7] it was shown that the transition process occurs when combination of
small breathers merge into larger breathers with enough strength to pull the entire oscillator
2
5 0 5
0
20
40
60
80
100
T = 0
O
s
c
i
l
l
a
t
o
r

N
u
m
b
e
r
2 0 2
0
20
40
60
80
100
T = 72.5
2 0 2
0
20
40
60
80
100
T = 109
2 0 2
0
20
40
60
80
100
T = 145.5
O
s
c
i
l
l
a
t
o
r

N
u
m
b
e
r
Position
2 0 2
0
20
40
60
80
100
T = 182
Position
2 0 2
0
20
40
60
80
100
T = 254.5
Position
FIG. 1: Snapshots of the evolution of 100 oscillators from one equilibrium to the other in the
potential (2) ( = 1.0e 3, = 1, B = 1). As an initial condition, two oscillators are perturbed
in the potential and because the coupling is much stronger than the nonlinearity the transition is
collective, closely following the mean (dashed line).
array over the barrier. In [8] the same authors developed an analytical calculation for the
transition process considering a single oscillator diusing through a separatrix layer and
subsequently pulling the other oscillators along. This theory does not work in our setting
since the coupling strength is much greater than the local nonlinearity and excursions of a
few oscillators and formation of breathers is discouraged. We nd that it is not excursions
of individual or groups of oscillators that drives the transition process but rather dynamics
of the mean position of all oscillators. In fact, we show that in the time-averaged sense,
the dynamics of the mean position and velocity are sucient to predict properties of the
activation process.
Since the empirical evidence of Arrhenius, noise assisted activation has been studied to a
great extent and has led to many predictive tools [9]. In this setting, external thermal energy
lifts the oscillators over the barrier. On the other hand, limited tools exist for the micro-
canonical example where energy is conserved and escape is only realized by redistribution of
initial energy. Such behavior is remarkable as energy is appropriately distributed throughout
the system leading to the activation process without any external inuence from a thermal
bath. We show that such behavior is caused by an interaction between incommensurate
linear modes due to the local nonlinearities.
3
We perform two canonical coordinate transformations to identify the coarse behavior and
energy transfer dynamics of the transition process. Because of the translational symmetry
on the backbone the fundamental normal basis is Fourier modes. We perform the canonical
transformation into these coordinates using
q
i
=
_
2
N
N
2
1

k=1
[
q
0

2
+ cos (2ik/N) q
k
(3)
+
cos (k)

2
qN
2
+ sin (2ik/N) qN
2
+k
]
where q
k
are modal amplitudes and k is the wavenumber (time dierentiation of (3)
gives transform rules for velocities q
i
to modal velocities

q
k
) . The frequencies
k
=

_
2 cos (2k/N) 2 are irrationally related and range from 0.0 to 2.0 with linear spacing
at low frequencies and decreasing spacing as they approach 2.0. The zeroth mode q
0
is the
averaged coordinate (scaled by

N) and the coordinates q

1...
N
2
1
are Fourier modes.
Projecting the nonzero modal dynamics onto canonical action-angle variables takes the
system to a series of nearly integrable oscillators which interact with one highly nonlinear
mode (we rename q
0

0
and

q
0
J
0
for notational convenience)

0
= J
0

J
0
= g
0
(

J,

, )
,

k
=
k
+ f
k
(

J,

, )

J
k
= g
k
(

J,

, )
(4)
k = 1, 2, . . . (N 1)
where
k
are angle variables and J
k
are the conjugate actions. To obtain a coarse repre-
sentation we average over the angles of M nonzero modes (M N as the analysis can be
performed on a reduced order model since the transformations are canonical). The averaged
Hamiltonian becomes

H(

J) =
1
(2)
M
_
2
0

_
2
0
H(

J,

)d
1
. . . d
M
(5)
For the specic 2-4 potential the resulting averaged ODE derived from Hamiltons equations
is

0
= J
0
(6)

J
0
=
_

3
N
M

k=1
J
k

k
_

3
0
N

J
k
= 0, k = 1, 2, . . . (M 1)
4
In an averaged sense, action in the higher modes are stationary in time and feed into
the zeroth mode dynamics. In this sense, (6) represents a parametric 2-4 potential for the
planar variables (
0
,J
0
) where higher order actions (J
k
, k = 1, 2, . . . , M) alter the existence
and location of the equilibria branches.
The rate of activation is dependent on energy in both the microcanonical and thermally
activated settings. In the deterministic setting, there exist a dependency between the ac-
tivation rate and the amount of initial energy (and in particular, the spatial character of
the initial energy), and this dependency can be captured quantitatively using the averaged
model (6). In the averaged sense, the zeroth mode has the typical gure eight phase space of
the 2-4 potential with equilibria at
0eq
=

N in modal coordinates. With an initial con-

dition with the zeroth mode at this location, and no energy in higher modes, no oscillations
will occur. As soon as energy from a higher mode is introduced the equilibrium of the zeroth
mode begins to shift, and since the oscillators are no longer on the original equilibrium they
begin to oscillate (in a state of libration). In fact, the equilibria shift toward the origin
bringing the separatrix closer to the initial location of the zeroth mode. With sucient
energy in any one of these higher order modes, the separatrix of the zeroth mode will be
driven closer to the origin and eventually the original location of the oscillators will breach
it. When this occurs the zeroth mode will be forced into the rotation state and transition
to the second equilibrium will occur. This concept is illustrated in Figure 2.
4
2
0
2
4
0.4
0.2
0
0.2
0.4
0
2
4
6
8
10
12

0
J
0
J
5
FIG. 2: Impact of energy in higher order modes (eg J
1
) on the zeroth mode illustrating how the
original equilibria are moved outside of the libration state to the rotation state.
With this in mind, the minimum activation energy becomes the amount of energy needed
5
from the higher mode(s) to pull the separatrix over the original equilibrium location
0
. The
activation condition is
U(
0
= 0.0,

J) = U(
0
=
0eq
,

J) (7)
which denes the minimum activation energy for transition. Performing the necessary cal-
culations, the resulting activation condition for the 2-4 potential becomes
M

i=1
J
i

i
=
N
6
(8)
which describes an ane hyperplane with a dimension equal to the number of nonzero
actions. It is evident in Figure 2 that with further increase of action in the higher modes,
the concept of having two wells disappears, and the zeroth mode will wildly circle both
original equilibria in a state of rotation (with dissipation the masses will settle in one of the
wells). Indeed a pitchfork bifurcation exists for these high energy orbits at

M
i=1
J
i

i
=
N
3
which is twice the minimum activation energy.
To calculate the full activation time vs. activation energy prole we note that the when
the oscillation of the zeroth mode is in its rotation state (which is needed for activation), the
activation time is the time in which the trajectory crosses the origin (ie
0
= 0.0). In fact,
since the phase space for the rotation state is symmetrical this becomes 1/4 of the period of
a rotation state. We can analytically derive the rate of activation by calculating the period
of rotation oscillations of the zeroth mode in (6) as it is parametrically altered by higher
wavelength actions. To calculate the period of rotation we generalize the Hamiltonian of the
averaged dynamics as
H =

J
2
0
2

_
a

2
0
2
+
b

4
0
4
_
(9)
a =
_

3

J
i

i
N
_
, b =

N
This system admits an analytical solution that is dierent inside and outside of the gure
eight separatrix, each of which uses Jacobi elliptic functions. On the outside of the separatrix
we calculate the period as
T
out
=
4K()

out
(10)
where =

0
(0)

b
_
2(a + b
0
(0)
2
)
,
out
=

2
2
1
6
where is the elliptic modulus and K(k) is the complete elliptic integral of the rst kind.
The variable
0
(0) is the amplitude of the initial condition of the zeroth mode.
For comparison, numerical data was generated using N = 100 oscillators, B = 1, =
0.001, and = 1. The dynamics were integrated in time using a 4
th
order symplectic method
[10] using a xed stepsize of 0.005. The analytical expression captures numerical data well
when the activation rates exceed the modal frequencies. The discrepancy in lowest mode at
high activation rates is because the modal frequency (of which we average over) approaches
the activation rate itself.
0 100 200 300 400 500
0.01
0.02
0.03
0.04
0.05
0.06
0.07
U(0) / U
A
c
t
iv
a
t
io
n

R
a
t
e

[
1
/
s
]


Mode 1
Mode 2
Mode 3
Mode 4
Mode 5
Mode 6
Mode 7
Mode 8
Mode 9
Mode 10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
1
2
3
4
5
6
7
Mode Frequency
M
in
im
u
m

A
c
t
iv
a
t
io
n

E
n
e
r
g
y


Data
Analytical
FIG. 3: Activation energy vs rate (top) and minimum activation energy (bottom). Analytical -
solid, numerical simulation - discrete points. U =
N
4
is the sum of the barrier heights and U(0)
is the initial energy.
The activation behavior discussed above is clearly impacted by selective modal energies.
The same dynamics were numerically analyzed in a stochastic setting using the Langevin
framework to determine if the phenomenon persists. In Figure 4 we illustrate the stochastic
rates using the same parameters as before with a Langevin damping value of 0.05. In this case
we impose an initial energy at a constant value of 5U in dierent spatial patterns including
a randomly selected series of positions. The data is simulated for dierent temperatures
(Boltzmanns constant k = 1.0) and compared to Kramers estimate for low damping (see
7
equation 4.49 of [9]). We nd that at higher temperatures, Kramers theory predicts the
behavior well, while at lower temperatures inertial eects dominate and the numerical data
diverges from Kramers estimates. In addition, we nd that this divergence is related to the
spatial character of the perturbation.
1 2 3 4 5 6 7 8
10
3
10
2
U/ (kT)
A
c
t
i
v
a
t
i
o
n

R
a
t
e

[
1
/
s
]


0th Mode
1st Mode
2nd Mode
3rd Mode
Random
Kramers
FIG. 4: Stochastic simulation with targeted perturbation in the rst 4 Fourier modes compared to
a random initial condition. These rates are compared to Kramers estimate.
Above we have been able to predict a static relationship between initial energy and
activation time. However, the transfer of energy from modes which contain initial energy
(higher modes) to the zeroth mode is a dynamical process. The equipartition of energy
throughout modes in high dimensional systems is expected due to their ergodic nature but
is not always the case. As shown initially by Fermi, Pasta, and Ulam (FPU) [11] this
balancing of energy may not occur for even relatively simple dynamics. In fact, Ford later
illustrated that this lack of energy balance is due to a lack of commensurate temporal
frequencies in the set of spatial modes [12]. As we have mentioned, the frequencies of the
linear normal modes in our model are incommensurate. Without any nonlinear perturbation
to the backbone, we expect there to be no energy transfer at all. However as we have shown,
there is no doubt that energy transfer occurs because activation can be observed by global
motions of the zeroth mode and this can be realized by imposing initial energy on modes of
high wavelength. In fact, the energy does not lead to equipartition immediately, it follows
a unidirectional route to the lowest mode.
By calculating the Jacobian of Hamiltons equations of (4) we obtain a matrix that reveals
the change in action variables as inuenced by other actions and angles. It turns out that
8
the change in action is mostly dependent on other actions (as opposed to angles) and so
we present this submatrix in Figure 5. We see in this gure that the change in action in
FIG. 5: Absolute value of the time averaged Jacobian of action variables illustrating the funneling
of energy towards the lowest mode.
any particular mode is dominated by inuences of modes of equal or higher wavenumber
only. This natural funneling of energy is remarkable and strengthens the robustness of the
transition process. In fact, the funneling concept is associated with the horizontal-vertical
graph theoretic structure of the dynamics which has been shown to be a way to characterize
the robustness of a dynamical system (see reference [13]).
We have found that even though this high dimensional system uses a complex nonlinear
process to achieve global transition between dierent states, the qualitative nature is easily
understood as a funneling of energy from higher modes to the lowest mode whos phase
space is manipulated by these higher modes. In fact the dynamics in the averaged sense
are enough to predict the activation process. These results were compared in the stochastic
setting and it is shown that the eect of targeted perturbation is an accelerant with respect
to Kramers theory and the spatial character of the perturbation eects the divergence from
Kramers estimate.
This work was supported in part by DARPA DSO under AFOSR contract FA9550-07-C-
9
0024 and AFOSR Grant FA9550-06-1-0088
[1] I. Mezic, PNAS 103 (2006).
[2] M. G. Forest, C. G. Goedde, and A. Sinha, Physical Review Letters pp. 27225 (1992).
[3] N. Ercolani, M. G. Forest, and D. W. McLaughlin, Physica D Nonlinear Phenomena pp.
34984 (1990).
[4] B. Eisenhower and I. Mezic, Proc. of the 46th IEEE CDC (2007).
[5] P. DuToit, I. Mezic, and J. Marsden, Physica D (in print) (2009).
[6] S. Lee and W. Sung, Physical Review E 6302 (2001), ISSN 1063-651X.
[7] S. Fugmann, D. Hennig, L. Schimansky-Geier, and P. Hanggi, Physical Review E 77 (2008),
ISSN 1539-3755.
[8] D. Hennig, S. Fugmann, L. Schimansky-Geier, and P. Hanggi, Acta Physica Polonica B 39
(2008).
[9] P. Hanggi, P. Talkner, and M. Borkovec, Reviews of Modern Physics 62, 251 (1990).
[10] H. Yoshida, Physics Letters A 150 (1990).
[11] E. Fermi, J. Pasta, and S. Ulam, Los Alamos report LA-1940 (1955).
[12] J. Ford, Journal of Mathematical Physics 2 (1961).
[13] I. Mezic, 43rd IEEE Conference on Decision and Control (2004).
10