Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

Fluctuation theorems for the undriven Duffing oscillator

J. Ignacio Deza∗
Facultad de Ingenierı́a, Universidad Atlántida Argentina, Mar del Plata,
Argentina, and CABER, University of the West of England, Bristol, UK

Julián I. Peña Rosselló, Martı́n E. Giuliano, and Roberto R. Deza†


IFIMAR, Facultad de Ciencias Exactas y Naturales,
Universidad Nacional de Mar del Plata, CONICET, Mar del Plata, Argentina

Horacio S. Wio‡
IFISC, Universitat de les Illes Balears, CSIC, Palma de Mallorca, Spain
(Dated: February 24, 2019)
For nonlinear systems with one degree of freedom having Hamiltonian H(x, p) := p2 /2m + V(x)
when undamped, we prove that H(x, p)/γ is a global Lyapunov function when their damping is linear
in p. We exploit this result to obtain transient fluctuation theorems past the pitchfork bifurcation
of the undriven Duffing oscillator.

Micro- and nanoscale resonators are becoming increas- β > 0).


ingly ubiquitous in modern technology. They offer great
flexibility to design their mechanical response, as well as Method 1: Let us assume a forcing à la Langevin
their intrinsic and extrinsic dissipation mechanisms [1]. ẋ1 = x2 ,
Whereas in some applications (e.g. vibration isolation,
commutation) the motion must be quickly dampened so ẋ2 = −γ x2 − V 0 (x1 ) + γ ξ(t), (1)
0 0
that fast energy dissipation is sought, the goal in oth- hξ(t)i = 0, hξ(t)ξ(t )i = 2Dδ(t − t )
ers is small dissipation rates (i.e. efficient isolation), so (see footnote at the end of this paragraph). In qualitative
to provide sources which are more stable in frequency terms, σξ(t) provides external energy to the dissipative
[2], or allow better detection of extremely weak forces [3] autonomous system, so that a statistical energy balance
or even coherent quantum operations at room tempera- (a “stationary state”) can be eventually reached.
ture [4]. Unfortunately, as the dimensions of the vibrat- Let p := PD (xf1 , xf2 , tf | xi1 , xi2 , ti ) denote the con-
ing structures are reduced (and their resonance frequen- ditional probability density function (pdf) from initial
cies increase) their dynamic response becomes nonlinear, conditions (xi1 , xi2 ), under white noise with intensity D.
genarating large instabilities which provoke noise in their Since it obeys the Fokker–Planck equation (FPE)
frequency, so considerably degrading their performance.
The resonance frequency has a strong dependence on ∂p ∂(x2 p) ∂[γ x2 + V 0 (x1 )] ∂2p
=− + + γ2D 2 , (2)
the oscillation amplitude. Because amplitude fluctua- ∂t ∂x1 ∂x2 ∂x2
tions result in frequency fluctuations, this “a-f effect” p is a Markovian process in (x1 , x2 ) space, and thus ad-
greatly increases the frequency noise of the oscillator. mits a path-integral representation. However, Eq. (2) has
Therefore, the benefits of operating at higher amplitudes singular diffusion matrix. This difficulty can be circum-
cancel out due to the increase in noise inherent in the op- vented by performing the functional integral in the phase
eration in the nonlinear regime. Typically, the restorative space of the variables xi and their associated canonically
force acquires cubical terms, proportional to the third conjugate momenta πi [9–11],
power of amplitude. The standard classical model of Z x1 (tf )=xf1 ,x2 (tf )=xf2
nonlinear oscillations—established by Georg Duffing in
p= D[x1 ]D[x2 ]D[π1 ]D[π2 ]eS1 (D) .
1918—has been systematically studied as a source of non- x1 (ti )=xi1 ,x2 (ti )=xi2
linear phenomena ranging from hysteresis to chaos [5]. In (3)
the absence of damping and external forcing, the Duff- exp S1 (D) is the probability of a given trajectory in phase
ing model admits a Hamiltonian, which has been long space, under white noise with intensity D
conjectured to act as Lyapunov function in the damped Z tf
ds γ 2 D[iπ2 (s)]2 + iπ1 (s)[ẋ1 (s) − x2 (s)]

case [6]. Our purpose in this work is to show why it is S1 (D) =
indeed so, and exploit this fact to obtain transient fluctu- ti

ation theorems along the lines of [7, 8], when the protocol +iπ2 (s)[ẋ2 (s) + γ x2 (s) + V 0 (x1 (s))]} . (4)
crosses the pitchfork bifurcation of the undriven model. Gaussian integration over π2 (s) yields
We adopt the notation x1 := x, x2 := p/m. Let γ Z x1 (tf )=xf1 ,x2 (tf )=xf2
be the damping coefficient per unit mass, and V (x1 ) := p∝ D[x1 ]D[x2 ]D[π1 ]eS2 (D) , (5)
V(x1 )/m (for the Duffing model, V (x) = 21 ω02 x21 + 14 βx41 , x1 (ti )=xi1 ,x2 (ti )=xi2
2

with Method 2 As argued in [14], the problem of meeting the


Z tf 
1 integrability conditions to find a global Lyapunov func-
S2 (D) = ds − [ẋ (s) + γ x2 (s) + V 0 (x1 (s))]2
2D 2 tion for an n–component continuous time, dissipative,
t0 4γ
autonomous dynamic flow ẋ = f (x), can be better poised
+iπ1 (s)[ẋ1 (s) − x2 (s)]} . (6) for solution by considering the deterministic limit of the
Integration over π1 (s) yields δ[ẋ1 (s) − x2 (s)]. After in- stochastic dynamic flow ẋ = f (x) + σ Ξ(t), where Ξ(t) is
tegration over x2 (s), we obtain x2 (s) = ẋ1 (s) and thus an n–component vector of independent white noises with
ẋ2 (s) = ẍ1 (s) for any ti ≤ s ≤ tf . Hence common variance D (we consider the limit D → 0) and
σ is a constant matrix. Accordingly, the nonequilibrium
x1 (tf )=xf1
potential (NEP) has been defined as
Z
p∝ D[x1 ]eS3 (D) , (7)
x1 (ti )=xi1
Φ(x) = − lim D ln P st (x; D). (14)
D→0
with
Z tf Denoting Q := σσ T = QT , the stationary n–variable
1
S3 (D) = − ds [ẍ1 + γ ẋ1 + V 0 (x1 (s))]2 . (8) FPE ∇ · [f (x)P st (x) − D Q ∇P st (x)] = 0 reduces in this
4γ 2 D t0 limit to the Hamilton–Jacobi-like equation
Equation (7) yields pr (xf1 , tf | xi1 , ti )—the reduced con- f T (x)∇Φ + (∇Φ)T Q ∇Φ = 0, (15)
ditional pdf in configuration space—under white noise
with intensity D, and exp S3 (D) is the probability of a from which Φ(x) can in principle be found. In an at-
given trajectory in this space. pr does not obey a FPE tractor’s basin, asymptotic stability imposes det Q ≥ 0.
and it is not even Markovian! Nonetheless, it can be For D → 0, it turns out that Φ̇ = ẋT ∇Φ = f T (x)∇Φ =
useful in some cases. For instance, a measure of irre- −(∇Φ)T Q ∇Φ ≤ 0. Hence, Φ(x) is a Lyapunov func-
versibility is ln(pF B
r /pr ), obtained forward and backwards tion for the deterministic dynamics. Even though Eq.
[S3B (x1 , ẋ1 ) = S3F (x1 , −ẋ1 )] on a same stochastic trajec- (15) has the structure of a Hamilton–Jacobi equation,
tory (this is known as a “detailed fluctuation theorem” the decomposition f (x) = d(x) + r(x)—with d(x) :=
[8]). From Eq. (8) it is −Q ∇Φ irrotational, but not an exact form—reduces it
 2   2  to rT (x)∇Φ = 0. Thus r(x) is the conservative part of
 F B
 ẋ1 ẋ1
γD ln(pfp ) − ln(pfp ) = +V − +V , (9) f (x), and d(x) its dissipative part.
2 tf 2 ti For n = 2 we may write r(x) = κ Ω ∇Φ, with Ω the
analog to T ∆S = ∆U . If either γ → 0 or D → 0, the N = 1 symplectic matrix. Hence f (x) = −(Q − κ Ω)∇Φ,
stochastic system must be conservative (the bracket must with det(Q − κ Ω) = D + κ2 > 0, and thus
vanish identically) in order to be reversible. ∇Φ = −(Q − κ Ω)−1 f (x). (16)
If we need a FPE in x1 space, we must perform
an approximation on pr . A suitable one is the “adia- For arbitrary real σij we can parameterize
batic elimination”—a special case the so-called “unified √ √ 
colored-noise approximation” (UCNA) [12]—consisting √ λ1 cos α1 √λ1 sin α1
σ=
in dropping ẍ1 (s). This reduces Eq. (8) to λ2 cos α2 λ2 sin α2
Z tf √
1 and define λ := λ1 λ2 cos(α1 − α2 ) (note that the con-
Sfp (D) = − ds [ẋ1 + γ −1 V 0 (x1 (s))]2 . (10)
4D t0 dition D > 0 imposes α2 6= α1 ). Then
 
Hence Eq. (10) has the associated FPE λ1 λ − κ
Q − κΩ = ,
λ + κ λ2
∂pfp 1 ∂[V 0 (x1 )pfp ] ∂ 2 pfp
= +D , (11)
∂t γ ∂x1 ∂x21 and Eq. (16) reads
with stationary pdf
 
1 λ2 f1 (x) − (λ − κ)f2 (x)
∇Φ = − .

V (x1 )
 det(Q − κ Ω) −(λ + κ)f1 (x) + λ1 f2 (x)
pst
fp ∝ exp − . (12)
γD If a set {λ1 , λ2 , λ, κ} can be found such that Φ(x) ful-
Equation (11) can also be obtained by starting from the fills the integrability condition ∂2 ∂1 Φ = ∂1 ∂2 Φ (∂k is a
single Langevin equation ẋ1 = −γ −1 V 0 (x1 ) + ξ(t) [13], shorthand for ∂/∂xk ), then a NEP exists. [15]
since x2 has been “enslaved” by the adiabatic elimina- In Eq. (1), f1 (x) = x2 and f2 (x) = −γ x2 − V 0 (x1 ).
tion, to be x2 = −V 0 (x1 )/γ. Equation (9) reduces in The integrability condition yields λ1 = 0 (then det Q ≥ 0
this case to imposesλ = 0) and  λ2 = κγ.  It turns  out that
0
x2 V (x 1 )
r(x) = and ∇Φ = κ−1 , which
γD [ln(pF B
fp ) − ln(pfp )] = V (tf ) − V (ti ). (13) −V 0 (x1 ) x2
3

If µ is varied linearly from −τ to τ , then µ̇B = −µ̇F =


1.00 1
const, and Y B = −Y F along each trajectory. For the un-
0.75 driven Duffing oscillator we may take µ := ω02 , imagining
0.50
the oscillator is realized as a Moon beam, and the mag-
0.25
nets are moved steadily in such a way that the pitchfork
bifurcation is come across. If ω02 goes from 1 to -1 in 2τ ,
0.00
Z +τ
0.25 1
F
0.50
Y =− dt x21 (t).
2γτ −τ
0.75
1.00
According to [7], when both protocol and dynamics
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 are jointly reversed, the distributions ρ(y) of Y –values
0 in an ensemble of realizations of the respective processes
(keeping the endpoints fixed) obey
ρF (y)
FIG. 1. Stochastic trajectory of the undriven Duffing oscilla- = exp(y), (19)
tor with ω0 = 1.0, β = 0.01, γ = 10−4 , starting from x1 = 1.0, ρ̂B (−y)
x2 = 0.0. Integration parameters ∆t = 0.02, 5,000 iterations,
where in ρ̂B , x2 → −x2 . Equation (19) has a Jarzynski-
D = 0.1.
like integrated version
hexp(−y)iF = 1, (20)
integrated over iany path from x = 0, yields Φ(x) =
F
h 2
x where h•i denotes average over realizations of the for-
κ−1 22 + V (x1 ) , so full consistency with Eq. (12) is
achieved for κ = γ. In sum, for any D, ward process. From Eq. (7),
Z x1 (tf )=xf1
F
H(x) F
ρ (y) = D[x1 ]eS3 (D) δ(Y F − y)
Φ(x; γ) = . (17) x1 (ti )=xi1
γ
Z x1 (tf )=xf1
B
r(x) is in fact the Hamiltonian flowin H(x), but f (x) is ρ̂B (−y) = D[x1 ]eS3 (D) δ(Y B + y)
0 x1 (ti )=xi1
not Hamiltonian, because d(x) = . r(x) 6= 0
−γ x2 We may reason trajectorywise with the aid of the canon-
is what violates detailed balance, which in Langevin ical momenta pF B
Y , pY conjugated to Y
F
and Y B .
dynamics is just an expression of Jst (x) = 0. If the In order to generate such an ensemble, we start from
turnaround mean frequency given by r(x) is high enough, the trajectory in Fig. 1 and introduce random deviations
Eq. (17) can be successfully exploited for slow enough with variance D along it.
forcing (“adiabatic approximation”).
In order to check the detailed fluctuation theorem
of Eq. (9), we must choose a stochastic trajectory in CONCLUSIONS
phase space. Figure 1 depicts one such for the un-
driven Duffing oscillator, V (x) = 12 ω02 x21 + 14 βx41 with Needless to say, global Lyapunov functions provide
β = 0.01 and γ = 10−4 (ω0−1 is set as timescale), obtained “energy” landscapes, which are an extremely useful aid
through Heun integration of Eq. (1) with ∆t = 0.02 and in the analysis of nonlinear systems. Energy landscapes
D = 0.1 from the initial condition x1 = 1.0, x2 = 0.0 not only help visualize the systems’ phase space and its
[hence (x22 /2 + V )(ti ) = 0.5025]. The last point in this structural changes as parameters are varied, but allow
trajectory is x1 = 0.897006, x2 = −0.150397 [hence predicting the rates of activated processes [15–17]. More-
(x22 /2 + V )(tf ) = 0.41523804]. S3F (x1 , ẋ1 ) and S3B (x1 , ẋ1 ) over, the knowledge of the energy landscape for elemental
have been computed from the values of x1 only, since units often allows to infer important properties of even
although the mean square deviation between ẋ1 and x2 large coupled systemas [18–20]. Some fields that bene-
was O(10−7 ), between ẍ1 and ẋ2 was O(10−3 ). The lhs fit from the energy landscape approach are optimization
of Eq. (9) yields -0.0874677 and the rhs, -0.087262. problems [21], neural networks [22], protein folding [23],
Next we wish to Robtain some transient fluctuation the- cell nets [24], gene regulatory networks [25, 26], ecology
orems for Y F := dt µ̇ ∂Φ ∂µ [7], when some control pa-
[27], and evolution [28].
rameter µ on which V (x1 ) depends is externally varied
according to some protocol µ(t). From Eq. (17), Support is acknowledged from UAA (JID, Grant ),
UNMdP (RRD, Grant EXA826/17–15/E779) and CON-
ICET (JIPR and MEG, doctoral fellowships) from Ar-
Z Z
F −1 F ∂V B −1 ∂V
Y =γ dt µ̇ , Y =γ dt µ̇B . (18) gentina. HSW thanks warmly IFISC for their hospitality.
∂µ ∂µ
4

[15] H. S. Wio, R. R. Deza, and J. M. López, An Introduc-


tion to Stochastic Processes and Nonequilibrium Statisti-
cal Physics, revised edition (World Scientific, Singapore,

ignacio.deza@atlantida.edu.ar 2012).

deza@mdp.edu.ar [16] H. S. Wio, in 4th Granada Seminar in Computational

horacio@ifisc.uib-csic.es Physics, edited by G. P and J. Marro (Springer, Dor-
[1] C. Y. Chen, D. H. Zanette, D. A. Czaplewski, S. Shaw, drecht, 1997) pp. 135–195.
and D. López, Nature Comm 8, 15523 (2017). [17] H. S. Wio and R. R. Deza, Eur. Phys. J. Special Topics
[2] J. T. M. Van Beek and R. Puers, J. Micromech. Micro- 146, 111 (2007).
eng. 22, 013001 (2011). [18] G. G. Izús, A. D. Sánchez, and R. R. Deza, Physica A
[3] B. P. Abbott et al., Phys. Rev. Lett. 116, 061102 (2016). 388, 967 (2009).
[4] R. A. Norte, J. P. Moura, and S. Grossblacher, Phys. [19] A. D. Sánchez and G. G. Izús, Physica A 389, 1931
Rev. Lett. 116, 147202 (2016). (2010).
[5] I. Kovacic and M. J. Brennan, eds., The Duffing Equa- [20] A. D. Sánchez, G. G. Izús, M. G. dell’Erba, and R. R.
tion: Nonlinear Oscillators and their Behaviour (Wiley, Deza, Phys. Lett. A 378, 1579 (2014).
Chichester, 2012). [21] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, Science
[6] T. Kanamaru, Scholarpedia, 3(, 6327 (2008). 220, 671 (1983).
[7] V. Chernyak, M. Chertkov, and C. Jarzynski, J. Stat. [22] H. Yan, L. Zhao, L. Hu, X. Wang, E. Wang, and J. Wang,
Mech.: Theor. Exp. 2006, P08001 (2006). Proc. Nat. Acad. Sci. USA 110, E4185 (2013).
[8] U. Seifert, Rep. Prog. Phys. 75, 126001 (2012). [23] J. Wang, R. J. Oliveira, X. Chu, P. C. Whitford,
[9] H. S. Wio, Path Integrals for Stochastic Processes: An J. Chahine, W. Han, E. Wang, J. N. Onuchic, and V. B.
Introduction (World Scientific, Singapore, 2013). Leite, Proc. Nat. Acad. Sci. USA 109, 15763 (2012).
[10] F. Langouche, D. Roekaerts, and E. Tirapegui, Func- [24] J. Wang, B. Huang, X. Xia, and Z. Sun, PLoS Comp.
tional Integration and Semiclassical Expansions (Reidel, Biol. 2, e147 (2006).
Dordrecht, 1982). [25] K.-Y. Kim and J. Wang, PLoS Comp. Biol. 3, e60 (2007).
[11] H. S. Wio, P. Colet, L. Pesquera, M. A. Rodrı́guez, and [26] J. Wang, L. Xu, E. Wang, and S. Huang, Biophys. J.
M. San Miguel, Phys. Rev. A 40, 7312 (1989). 99, 29 (2010).
[12] P. Jung and P. Hänggi, Phys. Rev. A 35, 4464 (1987). [27] C. Li, E. Wang, and J. Wang, PLoS Comp. Biol. 6,
[13] Had we chosen cγ (c 6= 1) as noise factor, it would have e17888 (2011).
been ẋ1 = −γ −1 V 0 (x1 ) + c ξ(t). [28] F. Zhang, L. Xu, K. Zhang, E. Wang, and J. Wang, J.
[14] R. Graham, in Instabilities and Nonequilibrium Struc- Chem. Phys. 137, 065102 (2012).
tures, edited by E. Tirapegui and D. Villarroel (Reidel,
Dordrecht, 1987) pp. 271–290.

You might also like