PHYSICAL REVIEW E 108, 044605 (2023)

Noise-induced escape of a self-propelled particle from metastable orbits

*
Rubens H. Damascena and Clécio C. de Souza Silva
Departamento de Física, Centro de Ciências Exatas e da Natureza, Universidade Federal de Pernambuco, Recife–PE, 50670-901, Brasil

(Received 26 August 2023; accepted 6 October 2023; published 31 October 2023)

Active particles, like motile microorganisms and active colloids, are often found in confined environments
where they can be arrested in a persistent orbital motion. Here, we investigate noise-induced switching between
different coexisting orbits of a confined active particle as a stochastic escape problem. We show that, in the
low-noise regime, this problem can be formulated as a least-action principle, which amounts to finding the most
probable escape path from an orbit to the basin of attraction of another coexisting orbit. The corresponding action
integral coincides with the activation energy, a quantity readily accessible in experiments and simulations via
escape rate data. To illustrate how this approach can be used to tackle specific problems, we calculate optimum
escape paths and activation energies for noise-induced transitions between clockwise and counterclockwise
circular orbits of an active particle in radially symmetric confinement. We also investigated transitions between
orbits of different topologies (ovals and lemniscates) coexisting in elliptic confinement. In all worked examples,
the calculated optimum paths and minimum actions are in excellent agreement with mean-escape-time data
obtained from direct numerical integration of the Langevin equations.

DOI: 10.1103/PhysRevE.108.044605

I. INTRODUCTION is canceled out by the confining force [22–27]. At this point,

the particle stalls but thermal noise can assist it to climb
The problem of noise-induced escape from a meta-stable
further away and eventually escape the potential well. Woillez
state is a fundamental challenge in physics, chemistry, and
et al. [13] have shown that the low-noise escape time for this
biology, with implications for fields such as materials science,
problem still follows Kramers exponential form, but with the
biophysics, and neuroscience. In many cases, this problem
potential barrier replaced by a pseudo-potential that depends
can be modeled as the escape of a Brownian particle from
on both the confining potential and the particle’s propulsion
a local minimum of a potential energy landscape through
speed, leading to a range of interesting phenomena. For in-
a saddle point, with the mean escape time following the
stance, the particle may “choose” to escape through a lower
classical Kramers formula [1,2]. However, this model has
or higher barrier height, depending on the propulsion speed.
limitations in describing noise-induced escape problems in
Some self-propelled microorganisms, like motile Chlamy-
systems where the activation energy depends not only on
domonas algae [28] and Magnetospirillum bacteria [29],
the endpoints but also on the specific shape of the escape
prefer to orbit around the center of a confining potential rather
trajectory, that is, systems that cannot be reduced to an effec-
than climbing it up to the stall point. A similar orbital motion
tive potential energy. These systems include driven Josephson
is reproduced by an hexbug robot in a parabolic reflector
junctions [3,4], particles in parametrically excited traps [5,6],
[30]. Common to these systems is that their dynamics can
and excitable neuronal models [7,8], among others. The com-
be successfully modeled by adding an aligning torque to the
plexity of these problems adds considerable challenge to the
equation of motion of the particle’s orientation vector. This
task of understanding and predicting noise-induced escape
torque couples the orientation vector to the local force field
dynamics.
and induces spontaneous symmetry breaking, creating the
The escape of active Brownian particles from local minima
possibility of either clockwise or counterclockwise circular
is a particularly interesting example of an activation problem
orbits [30]. When the potential lacks radial symmetry, as in
that depends on the specific escape path [9–18]. This problem
the case of elliptic confinement, the particle can perform a
has received a lot of attention recently, in part because of the
variety of complex orbits, with distinct, well-defined topolog-
increasing interest in predicting and controlling the behavior
ical indices [31]. Interestingly, two or more of these orbits
of small self-propelled entities in various applications, from
can coexist in the same region of the parameter space. These
drug delivery and nanorobotics to bacterial motility and col-
observations pose a noise-induced escape problem fundamen-
lective behavior. These particles are endowed with persistent
tally different from the ones previously discussed: the decay
motion [19–21], typically modeled as a self-propulsion force
of a confined active particle from a metastable orbit into
of constant intensity, which allows them to climb an external
another without ever leaving the confining potential. Indeed,
potential landscape up to a point where their propulsion force
noise-induced transitions between such orbits have been re-
ported [29–31], but an in-depth study on how the escape rate
depends on the noise intensity and insights on the escape
*
[email protected] dynamics are still lacking.

2470-0045/2023/108(4)/044605(9) 044605-1 ©2023 American Physical Society

DAMASCENA AND DE SOUZA SILVA PHYSICAL REVIEW E 108, 044605 (2023)

In the present work, we tackle this problem in the frame-

work of a least-action principle. We show that the low-noise
transition probability is dominated by the path that minimizes
the classical action, resulting in a Kramers-like exponential
escape time, tesc ∼ exp(S/D), where S is the action integral
along the most probable escape path (MPEP). This path is
found by solving the deterministic equations of motion of an
auxiliary Hamiltonian, derived from the least-action principle.
The details of this approach are presented in Sec. II. Then,
we apply this formalism to analyze two distinct scenarios,
presented respectively in Secs. III and IV. In the first one, we FIG. 1. Schematic zero-noise orbits of an active particle (half-
investigate transitions between clockwise and counterclock- filled dots) modeled by Eqs. (1), (2), and (5) for radially symmetric
wise circular orbits of an active particle in radially symmetric (a) and elliptic (b, c) confinement. The blank-to-filled direction in the
confinement. In the limit of strong (hard-wall) confinement, dots indicates the propulsion direction, which is typically nontangen-
we show that this problem is reduced to the original Kramers tial to the orbit. For the elliptic case, orbits of different shapes, such
escape problem, with the effective potential being a function as ovals (a) and lemniscates (b) can coexist. The labels indicate orbits
of the orbit radius and the propulsion speed of the particle, where the propulsion axis n̂ undergoes rotation (R) or libration (L)
and derive an analytical expression for the escape rate. For (see text). The +/− subscripts indicate the sense of the orbit at the
moderate or weak confinement, the problem can no longer upper lobe as indicated by the arrows.
be defined in terms of an effective potential. In this case,
we find the optimum escape path numerically by minimizing limit [32]. However, the shape and direction of the orbit of the
the action integral between the orbit and the boundary of circle swimmer are predetermined by τ . Therefore, there are
its basin of attraction. In the second scenario, we calculate no competing orbits in this case.
the optimum escape paths and corresponding escape rates for In contrast, for systems with aligning torque, the direction
noise-induced transitions between multiple orbits of different of the orbit is not decided a priori, but rather the symme-
topological indexes for the case of elliptic confinement. In try is broken spontaneously as discussed below. This torque
all cases, the minimum action approach is shown to be in typically appears when the swimmer interacts with: confining
excellent agreement with direct numerical simulations of the walls, as in the cases of geometrically confined, motile bac-
Langevin equations of motion of the active particle. Conclu- teria and algae [28,29]; phoretic gradients, as in the case of
sions and outlook of the work are given in Sec. V. self-propelled Janus particles [33]; or the potential induced by
the combination of gravity and a curved surface, as in the case
II. GENERAL FORMULATION of the hexbug crawler on a parabolic antenna [30]. Typically,
A. Model system and competing orbits the aligning torque is modeled as [29,34]
We consider an overdamped active particle in the (x, y) τ = b(n̂ × F ) · ẑ, (5)
plane subjected to a confining, conservative force field F(r)
and a generic off-plane torque ẑτ . The corresponding transla- where b is a constant with dimension of length and F is the
tional and rotational dynamics are modeled by the following local confining force. Notice that τ = 0 when the particle
Langevin equations: climbs up the confining potential in a way that n̂ and F are
parallel to each other. However, for sufficiently high angular
ṙ = v0 n̂ + μF + ζ(t ), (1)
mobility β, this situation becomes unstable and the symmetry
is spontaneously broken, as now the nonzero angle between
θ̇ = βτ + ξ (t ), (2)
n̂ and F can be either positive or negative, resulting in a
where n̂ = (cos θ , sin θ ) is the particle orientation axis, v0 torque that can make the particle follow either clockwise
is the self-propulsion speed, and μ (β) is the translational (R+ ) or counterclockwise (R− ) orbit [29,30], as illustrated
(angular) mobility. The stochastic forces ξ (t ) and ζ(T ) are in Fig. 1(a). In other words, the dynamics is now bistable.
independent Gaussian noises of zero mean and correlation Furthermore, if the confining potential lacks rotation sym-
metry, there typically appears multiple metastable orbits with
ζμ (t )ζν (t  ) = 2Dt δμν δ(t − t  ), (3)
different topological properties coexisting in the same region
of the space of parameters [31]. For example, an elliptic har-
ξ (t )ξ (t  ) = 2Dr δ(t − t  ), (4)
monic potential allows for the coexistence of ovals, where n̂
where μ, ν = x, y, Dt and Dr are, respectively, the transla- performs
 a full rotation (R), leading to a topological index
tional and the rotational diffusion constants. R dθ = ±2π , and lemniscates, where n̂ undergoes libration
The torque τ in Eq. (2) is a key ingredient for the ob- (L), that is, it swings back
 and forth without completing a full
servation of stable orbits. For example, in models of circle rotation, thus leading to L dθ = 0. Notice that, for the elliptic
swimmers, one typically consider a constant torque, induced potential, the equations of motion are invariant with respect
by some built-in asymmetry of the swimmer’s activation to reflections in the x and y axes. Therefore, each orbit is
mechanism. When such particles are confined to, e.g., a har- accompanied by a twin which is its exact mirror image, giving
monic trap, the interplay of persistent motion, confinement, a total of four coexisting orbits, as illustrated in Figs. 1(b) and
and torque does give rise to a stable orbit in the weak noise 1(c).

044605-2
NOISE-INDUCED ESCAPE OF A SELF-PROPELLED … PHYSICAL REVIEW E 108, 044605 (2023)

B. Least-action principle and the most probable escape path and their conjugate momenta p and pθ . Here, H represents
Each metastable orbit discussed above is associated with the (Wentzel-Freidlin) Hamiltonian and S the classical action
a basin of attraction, which is a region in phase space con- [36]. The corresponding Hamilton equations of motion are
taining all the initial conditions that will lead to the orbit in ṙ = v0 n̂ + μF + 2up, (10)
the absence of noise. The boundary of the basin of attrac-
tion defines the separatrix that separates it from neighboring θ̇ = βτ + 2pθ , (11)
basins. During a noise-induced escape, the noise provides
energy to the particle to abandon its current orbit and escape ṗ = −μ∇(F · p) − β(∇τ )pθ , (12)
the corresponding basin of attraction through some point at
the separatrix, where the escape process ends. This process ∂τ
ṗθ = −v0 (ẑ × n̂) · p − β pθ . (13)
can be viewed as a stochastic trajectory, the escape path, ∂θ
that connects the starting point of the escape to the point Therefore, given a starting point in the attraction basin ar-
where the particle reaches the separatrix. When this process bitrarily close to the orbit, the escape trajectory is given
is finished, the particle enters a new basin of attraction and deterministically by the solution of Eqs. (10)–(13). From
relaxes towards the new orbit. Eq. (9), the stationarity of S requires that H = 0 (eikonal con-
The probability density of a particular escape path can be dition), that is, the escape path from a given initial condition
estimated as follows [35]. Let dt be the time interval elapsed is also a zero energy solution.
as the particle advances from point A at to another point By comparing Eqs. (10) and (11) with Eqs. (1) and (2),
B infinitesimally close. The probability that B also lies at one identifies the generalized momenta as p = ζ/2u and pθ =
is equal to the probability that the values of the Langevin ξ /2. This establishes the formal correspondence between the
forces are exactly those necessary for that end, that is, P ∝ solutions obtained for p and pθ and the exact sequence of
exp[− 4D ( u ζ + ξ 2 )dt], where, from here on, we parameter-
1 1 2
fluctuations necessary to assist the particle through the op-
ized the noise intensities as Dr = D and Dt = uD. Therefore, timum path. Furthermore, the corresponding action can be
the probability density of the specific escape path is obtained by integrating the Lagrangian L = p · ṙ + pθ θ̇ − H.
ρ( ) = Ae−S( )/D
, (6) Using Eqs. (10), (11), and (9), the Lagrangian can be reduced
to L = up2 + p2θ , which leads to the same form of the action
where integral given by Eq. (7). Therefore, the MPEP is the path
  
1 1 2 corresponding to the minimum action integral,
S( ) = ζ + ξ dt
2
(7)  
4 u
S = min [S( )] = min (up + pθ )dt .
2 2
(14)
can be interpreted as the action integral and A is a normalizing
factor. As D → 0, the stochastic escape process is dominated
by the most probable escape path (MPEP), that is, by the path III. TRANSITIONS BETWEEN METASTABLE
that minimizes the action integral and, thereby, maximizes CIRCULAR ORBITS
the probability. The MPEP can be obtained in two steps: (i)
finding the path o→s between two given endpoints, one at the For radially symmetric confinement, V (r), it is convenient
orbit (o) and the other one at the separatrix (s), that minimizes to express the deterministic version of Eqs. (1) and (2) in
So→s ; (ii) finding which one of the possible pairs of endpoints terms of the polar coordinates, r and φ, and the tilt angle
minimizes the action integral globally. χ = ∠(n̂, r) = θ − φ. This way, r and χ form an independent
The minimization of S between two given endpoints can subspace, described by the equations [31]
be performed by solving the corresponding Euler-Lagrange ṙ = v0 cos χ − μV  (r), (15)
equations subjected to the constraints given by Eqs. (1) and  
(2). Here we choose the alternative Hamiltonian approach, v0
χ̇ = βbF − sin χ , (16)
which consists in substituting the probability density function r
in the form of Eq. (6) into the Fokker-Planck equation corre-
sponding to the Langevin Eqs. (1) and (2):  φ can be obtained straightforwardly from φ =
while
v0 dt r −1 sin χ . These equations admit as steady state solu-
∂ρ ∂ 2ρ ∂ (ρτ ) tions two circular orbits of radius R, tilt angle χ = ±χ0 , and
= uD∇ 2 ρ + D 2 − ∇ · [ρ(v0 n̂ + μF )] − β . constant angular speed φ̇ = ±ω, where
∂t ∂θ ∂θ
(8) v0 1
In the low-noise (D → 0) limit, we can expand Eq. (8) in R= , (17)
bβ F (R)
powers of D and obtain, to lowest order, a partial differential
equation for S, χ0 = arccos (α), (18)
  v0 √
∂S ∂S ω= 1 − α,
− = H r, θ , p = ∇S, pθ = (19)
∂t ∂θ R
= (v0 n̂ + μF ) · p + βτ pθ + up2 + p2θ . (9) and α ≡ μ/(bRβ ). The orbits are stable for any attractive
confining potential, that is V  (r)  0, ∀ r, as long as α < 1
Equation (9) is formally equivalent to the Hamilton-Jacobi [31]. Notice that in the limit α → 1− one has χ0 = 0, ω = 0,
equation of a system described by the coordinates r and θ and V  (R) = v0 /μ, so that R coincides with the stall point and

044605-3
DAMASCENA AND DE SOUZA SILVA PHYSICAL REVIEW E 108, 044605 (2023)

the period of the orbit diverges. This corresponds precisely to

the climbing phase. Therefore, by decreasing α from α > 1,
the climbing state (χ = 0) gives place continuously to one
of two equally stable orbiting states (χ = ±χ0 ), similar to
spontaneous symmetry-breaking phase transitions observed in
thermodynamic systems like an uniaxial ferromagnet.

A. Transitions in the hard-wall potential

Here we consider an active particle of radius R p geomet-
rically confined by a circular wall of radius Rw , similar to
the experiments of Refs. [28] and [29]. In this kind of con-
finement, the particle interacts with the wall mainly via steric
forces, so that V  (r) is essentially zero for r < R Rw − R p
and very large for r > R. This results in a large V  (r) in the
close vicinity of the orbit radius, which makes this problem
essentially 1D, as shown below.
For small deviations (δr, δχ ) from the circular orbit and for
V  (R) v0 /μR, Eqs. (15) and (16) can be expressed as the
following linear system:
    
δ ṙ −μV  (R) −ωR δr
ω 
μV (R) . (20)
δ χ̇ αv0
0 δχ

The eigenvalues, λ1 = −μV  (R), associated with a mixed

change of r and χ , and λ2 = −v0 α(1 − α 2 )/R, associated
with displacements of χ without changing the orbit radius, are
both negative, thereby guaranteeing the stability of the orbit. FIG. 2. (a) Effective potential ψ for an active particle in hard-
Furthermore, since |λ1 | |λ2 |, changing the orbit radius is wall confinement as a function of tilt angle χ for different values of
much harder than changing the tilt angle. Therefore, δ ṙ 0, the angular mobility β [Eq. (22), with α = μ/(bRβ )]. (b) Minimum
so that μV  = v0 cos χ at all times, and the dynamics can be action for the switching between clockwise and counterclockwise or-
described by a single approximate equation, bits as a function of β [Eq. (26)]. Symbols are numerical estimates of
activation energies obtained from direct simulations of the Langevin
v0 equations. Both S and ψ are in units of v0 /R and β is in units
χ̇ = f (χ ) + ξ = sin χ (cos χ − α) + ξ , (21)
αR of μ/bR. Inset: semilog plot of the mean escape time, τesc = 1/k,
as a function of 1/D calculated from simulations (symbols) and
where now the deterministic term f (χ ) depends only on χ
analytically using Eq. (27) (dashes) for selected values of β. Open
and we add the orientational Langevin force ξ (the effect of
symbols correspond to the linear part of the data used for extracting
translational fluctuations will be discussed later on).
the activation energy by an Arrhenius fitting (solid lines).
Within the simplifications discussed above, the switching
of the orbit direction can be seen as the noise-induced escape
of a particle confined in a one-dimensional effective potential As discussed in the previous section, the MPEP must sat-
ψ (χ ) = − f (χ )dχ or isfy the eikonal condition, H = 0, which leads to a simple
v0 nontrivial solution for the conjugate momentum, pχ = − f .
ψ (χ ) = cos χ (cos χ − 2α). (22) Substituting this result in Eq. (23), we obtain
2αR
As shown in Fig. 2, the behavior of ψ (χ ) nicely illustrates pχ = χ̇ = − f (χ ), (25)
the spontaneous symmetry breaking taking place when β >
μ/bR and giving rise to two degenerate local minima located that is, the MPEP corresponds to the time-reversed version of
at ±χ0 . The minimum at negative (positive) χ corresponds the deterministic motion, as expected for a 1d escape problem.
to the clockwise (counterclockwise) orbit. The decay of an From Eqs. (14) and (25), the action integral along this path can
orbit into the other occurs by noise-induced activation over be expressed as the difference in effective potential between
the barrier located at χ = 0. Therefore, the escape rate can be the local minimum (χ = ±χ0 ) and the saddle point (χ = 0),
determined using Kramers formula for the effective potential 
ψ (χ ). However, it is instructive to find the escape rate using
0
v0 (1 − α)2
S=− f (χ )dχ = ψ (0) − ψ (χ0 ) = . (26)
the theory presented in Sec. II to check for its consistency. χ0 2αR
For this 1d problem, the auxiliary Hamiltonian is H =
f pχ + p2χ and the corresponding Hamilton equations are This demonstrates the consistency of the MPEP approach, ap-
plied to the problem of an active particle in hard-wall circular
χ̇ = f (χ ) + 2pχ , (23) confinement, with the classic Kramers theory for a Brownian
particle in a 1d potential. In this case, the prefactor of the
ṗχ = − f  (χ )pχ . (24) escape rate can be calculated using Kramers formula [37],

044605-4
NOISE-INDUCED ESCAPE OF A SELF-PROPELLED … PHYSICAL REVIEW E 108, 044605 (2023)

leading to
 
v0 (1 − α)(1 − α 2 ) v0 (1 − α)2
k= exp − . (27)
2π αR 2αRD
To validate these theoretical results, we performed numer-
ical simulations of Eqs. (1) and (2) for a circular particle of
radius R p confined geometrically by a circular wall of radius
Rw for different values of noise intensity D. The particle-wall
interaction is modeled by a truncated Lennard-Jones (WCA)
potential, U (d ) = 0, for d > 21/6 , and U (d ) = 4(d −12 −
d −6 ) − , for d < 21/6 , where d = (Rw − r)/R p is the re-
duced radial distance between the particle center and the wall.
For each value of D, the equations of motion were integrated
following a stochastic second-order Runge-Kutta algorithm
starting from a perfect counterclockwise circular orbit all the
way until the particle first crosses the separatrix, which in
this case corresponds to the surface χ = 0. The mean escape
time, τesc = k −1 , was then estimated by averaging the crossing
time over 104 realizations of the stochastic force. Once D falls
below a certain threshold, the τesc (1/D) data approaches the
Kramers law, as illustrated in the inset of Fig. 2. The slope
of the fit to the data points in the linear regime (unfilled
points) is taken as an estimate of the activation energy. Finally,
in Fig. 2(b), we show the agreement between the activation
energy estimated from simulations and the analytical result
Eq. (26).

FIG. 3. The most probable escape path (solid line) from counter-
B. Transitions in the harmonic potential and effect
clockwise orbit and the relaxation path (dashes) towards clockwise
of translational noise
orbit for an active particle in parabolic confinement for (a) β = 1.6
Here we consider a much softer kind of confinement, the and (b) β = 1.2 and 2.0. Grayscale lines in panel (a) are 50 randomly
parabolic well V (r) = 21 κr 2 . For this problem, circular orbits chosen nonoptimum escape attempts, which minimize the action S
are the only stable attractors of the dynamics for β > βc ≡ only locally; the lighter the line, the larger the action along the
μ2 κ/bv0 and their radius and tilt angle are given by corresponding path. The transparent noisy lines in panel (b) depict
a small subset of 20 escape trajectories obtained by numerically
βc βc integrating the original stochastic problem for β = 1.2 (red) and 2.0
R = Rstall and χ0 = arccos , (28) (blue). Here, r and β are expressed in units of v0 /κμ and μ2 κ/bv0 ,
β β
respectively.
where Rstall = v0 /κμ is the stall point of the particle in the
harmonic potential. In contrast to the hard-wall case, here the
escape problem cannot be reduced to one dimension, requiring βc = μ2 κ/bv0 ) and u = 0 (negligible translational noise). No-
one to consider the full set of Eqs. (1) and (2) to correctly tice that, since this problem is symmetric with respect to
model the problem. Moreover, the right hand side of these the variable φ, the escape problem can be fully described in
equations cannot be expressed as the gradient of a potential, the rχ plane, with the orbit being represented by a single
so Kramers rate formula does not work here. Therefore, it is point located at (R, χ0 ). A few remarkable points are readily
necessary to find the MPEP by minimizing the action inte- noted: (i) for larger values of β, the MPEP presents a spiral
gral. We perform this task by solving Eqs. (10)–(13) starting shape, meaning that the escape trajectory typically wobbles
from 105 random initial conditions in the close vicinity of around the circular orbit before reaching the boundary; (ii)
the source orbit all the way up to the separatrix, which, by the MPEP ends at the stall point, (r, χ ) = (Rstall , 0), which
symmetry, corresponds to the φ = θ (χ = 0) plane in the rφθ thereby can be viewed as the saddle point for the escape;
phase space. For the initial conditions we fixed a point at (iii) the relaxation from the saddle to the new orbit or back
the orbit, (r, φ, θ ) = (R, 0, χ0 ), and chose 105 different values to the original one (dashed line) is not anti-parallel to the
of the momenta uniformly distributed on the surface of a optimum escape path, which is a consequence of the time
sphere of radius 10−6 and centered at (pr , pφ , pθ ) = (0, 0, 0). irreversibility of the system. In contrast, nonoptimal escape
This shooting method allows for sampling different take-off attempts [shown as gray-shaded lines in Fig. 3(a)] end up
directions, each one generating an escape path that minimizes at a radial position r < Rstall . These paths have higher action
the action locally. The MPEP is chosen as the one path that integral (lighter shade of gray) and therefore are less probable
minimizes S globally. routes.
In Fig. 3 we show the MPEP obtained by this method To validate the MPEP approach applied to this particular
for different values of β (1.2, 1.6, and 2.0, in units of example, we performed simulations of the Langevin equa-

044605-5
DAMASCENA AND DE SOUZA SILVA PHYSICAL REVIEW E 108, 044605 (2023)

the activation energy from numerical simulation data of the

Langevin equations for different values of u. As shown in
Fig. 4, the simulation data points are in excellent agreement
with the action integrated along the MPEP (line) in the full
range of u values investigated. The results evidence a strong
dependence of the MPEP and the corresponding minimum
action on u. Overall, as u increases, the MPEP oscillates
more violently around the circular orbit before reaching the
boundary (inset), while the corresponding action integral de-
creases. In particular, for u = 0.33Rstall
2
, which corresponds to
a particle of radius Rstall /2 in thermal equilibrium, the action
is approximately half the value observed for the u = 0 case.

IV. TRANSITIONS BETWEEN METASTABLE

NONCIRCULAR ORBITS
Here, we turn our attention to transitions between noncir-
cular orbits of an active particle trapped in the elliptic potential
V (x, y) = 21 κ ( 1−e
1
2 x + y ), where e is the eccentricity. As
2 2
FIG. 4. Minimum action (in units of 1/κμ) for the switching
shown in Ref. [31], this potential allows for the coexistence
between clockwise and counterclockwise orbits of an harmonically
of a variety of orbits, of different shapes and topological
trapped active particle as a function u = Dt /Dr (in units of Rstall
2
).
indexes. To be specific, we focus on the case e = 0.731 and
Symbols correspond to activation energies estimated by fitting the
low-noise data of direct Langevin simulations. Inset: MPEPs in the
β = 2.5 µ2 κ/bv0 , where two reflection-symmetric oval orbits
rχ plane calculated for a few selected values of u. (R+ and R− ) and two reflection-symmetric lemniscates (L+
and L− ) coexist with each other, as shown in Fig. 5. Further,
we assume negligible translational noise (u = 0).
tions of motion and calculated the mean escape time τesc The task of finding the MPEP in this case is considerably
(averaged over 104 realizations) as a function of noise in- more challenging than in the previous examples. First, the
tensity D. As in the previous example, we estimated the basins of attraction are highly nontrivial, making it necessary
low-noise activation energy of the escape by fitting the linear to map them numerically. For that, we solved the noiseless
part of the log τesc versus (1/D) data. Once again, the numeri- version of Eqs. (1) and (2) for ∼2 × 108 initial conditions de-
cally estimated activation energies are in excellent agreement fined in a regular grid in the region −1.5 < x < 1.5, −1.5 <
with the action integral along the MPEP, as discussed in more y < 1.5, −π < θ < π of phase space and classified each grid
detail further below. In Fig. 3(b), we compare the MPEP point according to which orbit the trajectory ends up. Second,
with 20 randomly chosen realizations of escape trajectories the lack of rotational symmetry of the orbits makes it neces-
obtained by integrating the Langevin equations for β = 1.2βc sary to sample escape attempts taking off from different places
(D = 0.001) and β = 2.0βc (D = 0.01). These values of D of the orbit to find the MPEP. Accordingly, we integrated
were chosen within the range where the log τesc simulation the Hamilton’s Eqs. (10)–(13) from 107 takeoff points chosen
data scales linearly with 1/D. Although the escape trajectories randomly in the close neighborhood of the whole extension of
are considerably noisy, they follow in average the theoretical the source orbit. For that, we chose 100 different points at the
MPEP and the escape through the boundary of the attraction source orbit and for each of them we chose 105 different val-
basin takes place at points close to the stall point, as predicted ues of the momenta uniformly distributed on the surface of a
by the minimum action principle. sphere of radius 10−6 and centered at (px , py , pθ ) = (0, 0, 0).
An example of MPEP calculated following this procedure is
shown in Fig. 5 for the case where R+ is the source orbit.
C. Effect of translational noise The color background indicates projections of the attraction
In thermal equilibrium, the translational and rotational basins.
noises acting on a spherical active particle are related by In principle, a full comprehension of the noise-induced
u = Dt /D = σ 2 /3. Since in most cases of interest the particle transitions between these orbits requires investigating all 12
diameter σ is small compared to other relevant length scales, possible transitions. This complexity can be mitigated by con-
the contribution of translational noise is often neglected, spe- sidering the symmetry properties of the noiseless version of
cially considering that other nonthermal noise sources usually Eqs. (1) and (2). For the elliptic potential, these equations are
contribute to decrease u even further [38]. For the case of invariant under π rotations, (x, y, θ )→(−x, −y, θ + π ), and
confined active particles, the particle size is typically consid- reflections in the x axis, (x, y, θ )→(x, −y, −θ ). Therefore,
erably smaller than the trapping region (σ < Rstall . Therefore, to each escape path, there are three other paths that can be
one typically has u  Rstall
2
. Nevertheless, in what follows, we mapped exactly to the first one by some combination of these
shall also consider u > Rstall
2
to have a broader picture of the symmetry operations. For example, under a π rotation R+
effect of translational noise. transforms to itself whereas L− transforms to L+ , as can be
In Fig. 4, we present the action integrated along the MPEP inferred from Fig. 5(a). Therefore, there exists a path from R+
as a function of u for fixed β = 2βc . We also estimated to L+ that is identical to the MPEP from R+ to L− shown

044605-6
NOISE-INDUCED ESCAPE OF A SELF-PROPELLED … PHYSICAL REVIEW E 108, 044605 (2023)

and (px , py , pθ )→(px , −py , −pθ ) under x reflections. This

way, Hamilton’s equations (10)–(13) are fully invariant and
so is the action integral, since it is insensitive to the sign
of the conjugate momenta. Therefore, all four transitions are
degenerate MPEPs, that is, they produce the same least-action
integral. This was indeed verified within the numerical accu-
racy of our least-action calculations, where we found SR→L =
3.4 × 10−4 (κµ)−1 . Analogously, all four escapes from libra-
tion to rotation can also be recovered from a single one, say
L− to R+ . In this case, our MPEP calculation leads to SL→R =
1.9 × 10−4 (κµ)−1 , revealing that, in general, escapes from R
to L are not symmetrical to escapes from L to R.
These observations were confirmed by direct simula-
tions of the Langevin equations, which resulted in similar
number of escapes from R+ to L− and from R+ to L+ in
the full range of investigated noise values as well as approxi-
mately the same value of activation energy estimated from the
low noise data [see Fig. 5(d)]. The same applies for escapes
from L orbits. The activation energies for both L → R and
R → L escapes were estimated from Arrhenius fitting the
log τesc versus 1/D data, averaged over 5 × 103 independent
realizations of the problem. As in the previous examples,
the activation energies are in good agreement with the action
integral over the corresponding MPEPs.
Finally, we also calculated optimum escape paths between
orbits of the same species. These escapes are obviously sym-
FIG. 5. (a) Most probable escape path (black dashes) from orbit
R+ (orange) and subsequent relaxation path (red dashes) towards L−
metrical, in the sense that R+ → R− has the same action
(blue). Also shown are orbits R− and L+ (gray transparent lines), integral as R− → R+ . However, for the chosen values of β
which are not involved in this particular escape. The yellow, orange, and e, direct transitions between orbits of same species have
blue and light blue backgrounds depict the attraction basins of R+ , considerably higher action integral: SR→R 350SR→L and
R− , L− , and L+ respectively. The basins are projected on the xy SL→L 10SL→R . Therefore, at the low noise limit, a full
plane at θ = 2.26 rad, corresponding to the propulsion direction at reversal of the R+ (L+ ) orbit towards its twin R− (L− ) is only
the point where the escape trajectory first enters the attraction basin possible by first visiting the attraction basin of one of the lem-
of the target orbit in all three projections, xy (b), xθ (c), and yθ niscate (oval) orbits, that is R+ → L± → R− (L+ → R± →
(not shown). (d) Mean escape time τesc versus 1/D obtained from L− ). Indeed, in our simulations of the Langevin equations,
direct Langevin simulations of R+ → L± and L− → R± escapes escapes between orbits of the same species are never observed
(symbols). The open symbols were selected for Arrhenius fits (lines), for D < 10−3 . Of course, these results are not general, as the
which allowed us to estimate the activation energies ER→L a
= 3.5 × hierarchy of the most probable target orbits might change
−4 −4
10 and EL→R = 2.0 × 10 , remarkably close to the corresponding
a
completely for other values of the system parameters β and
least actions, SR→L = 3.4 × 10−4 and SL→R = 1.9 × 10−4 . e.

V. CONCLUSIONS
in Fig. 5, but rotated by π . Similarly, other paths can also be
found by performing symmetry operations as illustrated in the In this study, we introduced a method rooted in the
diagram below. least-action principle to rigorously investigate the noise-
R →L π rotation R → L induced escape of active particles from metastable orbits in
+ − + + the low-noise limit. We methodically analyzed transitions
between various dynamical orbits spanning different confine-
x reflection x reflection ment scenarios. From transitions between basic clockwise and
counterclockwise circular orbits in radially symmetric traps to
R − → L− π rotation R− → L + more complex orbits in elliptic confinements, our theoretical
results align consistently with direct numerical simulations
A crucial question is whether all the paths generated by based on the Langevin equations of motion.
these symmetry operations also minimize the action, that is, The formalism we presented not only simplifies the task
are all of them MPEPs? To be MPEP, each of them must be an of accurately determining the activation energy tied to orbital
equivalent solution of the least-action problem. Notice that the transitions but also offers pivotal insights. One salient reve-
Hamilton’s equations are not invariant under the rotation and lation is that under low-noise conditions, particle adhere to a
reflection operations defined above. However, we can extend well-defined trajectory: the most probable escape path. This
them to the six-dimensional Hamiltonian phase space by also path is intricately influenced by various system parameters,
requiring that (px , py , pθ )→(−px , −py , pθ ) under π rotations such as angular and translational mobilities, self-propulsion

044605-7
DAMASCENA AND DE SOUZA SILVA PHYSICAL REVIEW E 108, 044605 (2023)

velocity, and the specific nature of the confining poten- specific behaviors or trajectories are desired. Unveiling the
tial. Furthermore, the derived Hamiltonian equations possess role of noise on switching between different dynamical states
symmetry properties essential for pinpointing degenerate es- can shed light on new strategies for manipulating particle
cape paths and for establishing a hierarchy amidst multiple behaviors for specific tasks.
competing orbits.
Our findings open avenues for expanded research on the
dynamical properties of confined active matter. Potential di-
ACKNOWLEDGMENTS
rections include exploring noise-induced escapes from chaotic
orbits and investigating synchronization phenomena. Beyond We thank L. R. E. Cabral for useful discussions. This work
these theoretical prospects, our results hold potential signifi- was financed in part by Coordenação de Aperfeiçoamento
cance for practical applications. Active particles, due to their de Pessoal de Nível Superior-Brasil (CAPES), Finance Code
persistent motion, can get ensnared in specific dynamical 001. C.C.S.S. is funded by Conselho Nacional de Desenvolvi-
states. Such patterns, while intriguing, can introduce com- mento Científico e Tecnológico-Brasil (CNPq), Grant No.
plications in real-world scenarios, like drug delivery, where 312240/2021-0.

[1] H. A. Kramers, Brownian motion in a field of force and the [16] A. Militaru, M. Innerbichler, M. Frimmer, F. Tebbenjohanns, L.
diffusion model of chemical reactions, Physica 7, 284 (1940). Novotny, and C. Dellago, Escape dynamics of active particles
[2] L. I. McCann, M. Dykman, and B. Golding, Thermally acti- in multistable potentials, Nat. Commun. 12, 2446 (2021).
vated transitions in a bistable three-dimensional optical trap, [17] Y. Li, F. Marchesoni, T. Debnath, and P. K. Ghosh, Two-
Nature (London) 402, 785 (1999). dimensional dynamics of a trapped active Brownian particle in
[3] R. L. Kautz, Activation energy for thermally induced escape a shear flow, Phys. Rev. E 96, 062138 (2017).
from a basin of attraction, Phys. Lett. A 125, 315 (1987). [18] T. Debnath, Y. Li, P. K. Ghosh, and F. Marchesoni, Active
[4] J. A. Blackburn, H. J. T. Smith, and N. Grønbech-Jensen, Chaos microswimmers in a finite two-dimensional trap: The role of
and thermal noise in a Josephson junction coupled to a resonant hydrodynamic interaction, J. Chem. Phys. 150, 104102 (2019).
tank, Phys. Rev. B 53, 14546 (1996). [19] B. ten Hagen, S. van Teeffelen, and H. Löwen, Brownian mo-
[5] L. J. Lapidus, D. Enzer, and G. Gabrielse, Stochastic phase tion of a self-propelled particle, J. Phys.: Condens. Matter 23,
switching of a parametrically driven electron in a penning trap, 194119 (2011).
Phys. Rev. Lett. 83, 899 (1999). [20] M. C. Marchetti, J. F. Joanny, S. Ramaswamy, T. B. Liverpool,
[6] K. Kim, M.-S. Heo, K.-H. Lee, H.-J. Ha, K. Jang, H.-R. Noh, J. Prost, M. Rao, and R. A. Simha, Hydrodynamics of soft active
and W. Jhe, Noise-induced transition of atoms between dynamic matter, Rev. Mod. Phys. 85, 1143 (2013).
phase-space attractors in a parametrically excited atomic trap, [21] C. Bechinger, R. Di Leonardo, H. Löwen, C. Reichhardt, G.
Phys. Rev. A 72, 053402 (2005). Volpe, and G. Volpe, Active particles in complex and crowded
[7] A. Longtin, A. Bulsara, and F. Moss, Time-interval sequences environments, Rev. Mod. Phys. 88, 045006 (2016).
in bistable systems and the noise-induced transmission of infor- [22] A. Pototsky and H. Stark, Active Brownian particles in two-
mation by sensory neurons, Phys. Rev. Lett. 67, 656 (1991). dimensional traps, Europhys. Lett. 98, 50004 (2012).
[8] I. A. Khovanov, A. V. Polovinkin, D. G. Luchinsky, and P. V. E. [23] Y. Fily, A. Baskaran, and M. F. Hagan, Dynamics of self-
McClintock, Noise-induced escape in an excitable system, propelled particles under strong confinement, Soft Matter 10,
Phys. Rev. E 87, 032116 (2013). 5609 (2014).
[9] P. S. Burada and B. Lindner, Escape rate of an active Brown- [24] S. C. Takatori, R. De Dier, J. Vermant, and J. F. Brady, Acoustic
ian particle over a potential barrier, Phys. Rev. E 85, 032102 trapping of active matter, Nat. Commun. 7, 10694 (2016).
(2012). [25] K. Malakar, A. Das, A. Kundu, K. V. Kumar, and A. Dhar,
[10] A. Geiseler, P. Hänggi, and G. Schmid, Kramers escape of a Steady state of an active Brownian particle in a two-dimensional
self-propelled particle, Eur. Phys. J. B 89, 175 (2016). harmonic trap, Phys. Rev. E 101, 022610 (2020).
[11] A. Sharma, R. Wittmann, and J. M. Brader, Escape rate of active [26] M. Caraglio and T. Franosch, Analytic solution of an active
particles in the effective equilibrium approach, Phys. Rev. E 95, Brownian particle in a harmonic well, Phys. Rev. Lett. 129,
012115 (2017). 158001 (2022).
[12] L. Caprini, U. Marini Bettolo Marconi, A. Puglisi, and A. [27] U. Nakul and M. Gopalakrishnan, Stationary states of an active
Vulpiani, Active escape dynamics: The effect of persistence on Brownian particle in a harmonic trap, Phys. Rev. E 108, 024121
barrier crossing, J. Chem. Phys. 150, 024902 (2019). (2023).
[13] E. Woillez, Y. Zhao, Y. Kafri, V. Lecomte, and J. Tailleur, [28] T. Ostapenko, F. J. Schwarzendahl, T. J. Böddeker, C. T. Kreis,
Activated escape of a self-propelled particle from a metastable J. Cammann, M. G. Mazza, and O. Bäumchen, Curvature-
state, Phys. Rev. Lett. 122, 258001 (2019). guided motility of microalgae in geometric confinement,
[14] K. S. Olsen, L. Angheluta, and E. G. Flekkøy, Escape problem Phys. Rev. Lett. 120, 068002 (2018).
for active particles confined to a disk, Phys. Rev. Res. 2, 043314 [29] A. Codutti, M. A. Charsooghi, E. Cerdá-Doñate, H. M. Taïeb,
(2020). T. Robinson, D. Faivre, and S. Klumpp, Interplay of surface
[15] D. Wexler, N. Gov, K. O. Rasmussen, and G. Bel, Dynamics interaction and magnetic torque in single-cell motion of magne-
and escape of active particles in a harmonic trap, Phys. Rev. totactic bacteria in microfluidic confinement, Elife 11, e71527
Res. 2, 013003 (2020). (2022).

044605-8
NOISE-INDUCED ESCAPE OF A SELF-PROPELLED … PHYSICAL REVIEW E 108, 044605 (2023)

[30] O. Dauchot and V. Démery, Dynamics of a self-propelled parti- [35] R. Feynman, A. Hibbs, and D. Styer, Quantum Mechanics and
cle in a harmonic trap, Phys. Rev. Lett. 122, 068002 (2019). Path Integrals, Dover Books on Physics (Dover Publications,
[31] R. H. Damascena, L. R. E. Cabral, and Clecio C. de Souza Mineola, NY, 2010).
Silva, Coexisting orbits and chaotic dynamics of a confined [36] M. Freidlin, J. Szucs, and A. Wentzell, Random Perturbations
self-propelled particle, Phys. Rev. E 105, 064608 (2022). of Dynamical Systems, Grundlehren der mathematischen Wis-
[32] S. Jahanshahi, H. Löwen, and B. ten Hagen, Brownian motion senschaften (Springer, New York, NY, 2012).
of a circle swimmer in a harmonic trap, Phys. Rev. E 95, 022606 [37] H. Risken and H. Haken, The Fokker-Planck Equation: Methods
(2017). of Solution and Applications Second Edition (Springer, Berlin,
[33] B. Liebchen and H. Löwen, Which interactions dominate in 1989).
active colloids? J. Chem. Phys. 150, 061102 (2019). [38] J. Elgeti, R. G. Winkler, and G. Gompper, Physics of
[34] I. S. Aranson and A. Pikovsky, Confinement and collective microswimmers-single particle motion and collective behavior:
escape of active particles, Phys. Rev. Lett. 128, 108001 (2022). A review, Rep. Prog. Phys. 78, 056601 (2015).

044605-9