Physica D 137 (2000) 333–352

Quasi-deterministic approximation, metastability

夽
and stochastic resonance
Mark I. Freidlin
Department of Mathematics, University of Maryland, Mathematics Building 084, College Park, MD 20742, USA
Received 18 March 1999; received in revised form 23 August 1999; accepted 30 August 1999
Communicated by C.K.R.T. Jones

Abstract
For a wide class of dynamical systems perturbed by a random noise, we describe the deterministic component of the
long-time evolution of the perturbed system. In particular, for any initial point and for a given timescale, the metastable state
can be defined. Stochastic resonance is the result of the change of the metastable state if a relatively small and slowly changing
deterministic perturbation is added to the system. If this perturbation is periodic, then under certain assumption, the system
will perform a motion which is close to a large amplitude oscillation with the same period or with a period proportional to the
period of determinisitic perturbation. All these effects are manifestations of the laws of the large deviations for the perturbed
system. ©2000 Elsevier Science B.V. All rights reserved.
Keywords: Metastable state; Stochastic resonance; Perturbed system

1. Introduction

In 1981, Benzi et al. [1,2] suggested a model explaining periodicity of the ice ages. The time evolution of the
“earth temperature” Xt was described by the equation
√
Ẋt = −B 0 (Xt ) + f (t/T ) +  Ẇt , X0 = x ∈ R1 . (1.1)

Here the potential B(x) has two wells, lim|x|→∞ B(x) = ∞, f (t) is a 1-periodic function, T > 0 is a parameter
which we specify later, and Ẇt is the standard Gaussian white noise, 0 <   1.
The periodic term f (t/T ) has a small amplitude so that the solution of (1.1) for  = 0 cannot be transferred by
f (t/T ) from one well to another. If  = 0, the trajectory may have small oscillations near the bottom of the well
containing the initial point, but it stays forever inside the well.

夽 This work was supported in part by NSF Grant DMS-9803522.

E-mail address: [email protected] (M.I. Freidlin).

0167-2789/00/$ – see front matter ©2000 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 7 - 2 7 8 9 ( 9 9 ) 0 0 1 9 1 - 8
334 M.I. Freidlin / Physica D 137 (2000) 333–352

On the other hand, if f (t) ≡ 0 but  > 0, the solution of (1.1) will make transitions between the wells. The
 and τ  are random variables and there is no periodicity in these transitions. It turns out that if
transition times τ1,2
√2,1
both terms f (t/T ) and  Ẇt , 0 <   1, are included in (1.1), the trajectory, under certain relations between 
and T = T (), will be close in an appropriate topology to a large amplitude periodic function. This effect is called
stochastic resonance.
 and τ  can be calculated explicitly in the one-dimensional case, and
The expected values of the variables τ1,2 2,1
 
2Bij
Eτij  exp , i, j ∈ 1, 2, i 6= j,  ↓ 0.

Here  is the sign of logarithmic equivalence, and Bij is the difference between the values of the potential B(x) at the
maximum separating the wells and at the bottom of ith well. Moreover, in the potential case, and one-dimensional
case which is always potential, one can write down an explicit expression for the invariant density M  (x) of the
process Xt with f (t) ≡ 0:
 
  2B(x)
M (x) = C exp − , (1.2)

C  is a normalizing factor. We conclude from (1.2) that, in the generic case, the invariant measure is concentrated
as  ↓ 0 just at one minimum of the potential — at the lowest one. Actually, exactly this property together with the
fact that the transition times have a non-random logarithmic asymptotic as  ↓ 0 are the reasons for the stochastic
resonance.
The small amplitude term f (t/T ) cannot transfer the particle from one well to another, but can shift the position
of the absolute minimum of the potential from one well to another. Then the transition of the particle governed
by (1.1) will occur due to the noise term. These transitions will follow the evolution (not necessarily periodic) of
f (t/T ) if  and T = T () satisfy certain conditions.
In the case of one-dimensional dynamical system perturbed by the white noise these asymptotic properties can
be derived from the classical Kramers’ theory. Corresponding questions for general dynamical systems and more
general random perturbations are studied in the large deviation theory for random perturbations of dynamical
systems [4]. This theory gives us the notions and the machinery for description of the stochastic resonance and other
similar effects. In particular, in a rather general situation, the notion of quasi-potential is introduced there, and the
logarithmic asymptotics of the invariant measure and the transition times are calculated.
Consider a dynamical system in Rd :

Ẋt = b(Xt ), X0 = x ∈ Rd . (1.3)

Assume for brevity that the system has a finite number l of asymptotically stable equilibrium points K1 , . . . , Kl ,
and each trajectory of (1.3), besides the trajectories belonging to the separatrix surfaces, is attracted to one of the
attractors Ki . Let i(x) be the number of the rest point such that the trajectory starting at x ∈ Rd is attracted to Ki(x) .
One can consider various types of random perturbations of system (1.3): it can be additive white noise type
perturbations
√
Ẋt = b(Xt ) + σ (Xt ) ◦ Ẇt , X0 = x, (1.4)

where 0 <   1, Wt is the Wiener process in Rd and σ (x) is the d × d matrix. It can be rapidly changing
perturbation

Ẋt = b(Xt , ζt/ ), X0 = x. (1.5)

 M.I. Freidlin / Physica D 137 (2000) 333–352 335

Here ζt is a stationary process with some mixing properties such that Eb(x, ζt ) = b(x). (Eη means expectation of
η.) As it is known (see, for example, [4]), the process Xt , 0 ≤ t ≤ T , defined by (1.5), converges to Xt as  ↓ 0
uniformly on any finite time interval in probability. Thus (1.5) can be viewed as a perturbation of (1.3). A typical
example of the process ζt in (1.5) is a Markov process with a finite state space.
Under certain additional assumptions, the qualitative behavior of the perturbed system on long-time intervals is
similar in all these cases, although the numerical characteristics are different. For each of these perturbed process Xt
one can introduce the action functional (rate function) [4] (1/)S0T (φ) : exp{(−1/)S0T (φ)} is, roughly speaking,
the main term of the probability that Xt , 0 ≤ t ≤ T , belongs to a small neighborhood of a function φ : [0, T ] → Rd
as  ↓ 0.
The action functional allows to introduce a function V (x, y):

V (x, y) = inf{S0T (φ) : φ0 = x, φT = y, T ≥ 0}.

In particular, if the system (1.3) is potential and the perturbed system is (1.4) with σ (x) equal to the unit matrix,
then V (x, y) can be expressed through the potential.
Put Vij = V (Ki , Kj ), where Ki and Kj are the stable equilibriums of the field b(x); i, j ∈ L = {1, . . . , l}. The
matrix (Vij ) defines a partition of the set L in a hierarchy of cycles. Cycles of rank 0 are the states of L = {1, . . . , l}
themselves. For each i ∈ L, define “the closest” j = J (i) ∈ L as such a j that Vij = mink∈L\{i} Vik . Such a closest
state is unique in the generic case. Since the set L is finite, the sequence, i, J (i), J 2 (i), . . . , J m (i), . . . , starting
from some m is periodic. This gives the cycles of the first rank (1-cycles). For any 1-cycle C, one can define 1-cycle
which follows after C. The 1-cycles form cycles of second rank (2-cycles). The second rank cycles form 3-cycles.
We will define a hierarchy of cycles up to rank m∗ such that the m∗ -cycle contains all the points of L. The numbers
Vij , together with the hierarchy of cycles, define for each cycle a rotation rate, an exit rate and a main state. In the
generic case, all these notions are defined in a unique way.
The cycles describe the sequence of transitions of the perturbed trajectory Xt between the basins of the attractors
K1 , K2 , . . . , Kl , as  ↓ 0. The exit rates give the asymptotics (non-random) of the logarithms of the transition
times. The rotation rate characterizes the rate of convergence to the sub-limiting distribution [3] inside the cycle.
The main state m∗ = M(C) of a cycle C defines the attractor Km∗ such that Xt spends most of the time in the basin
S
of Km∗ until it leaves the basin of i∈C Ki . We underline that the hierarchy of cycles which defines the sequence
of transitions of Xt between the attractors and logarithmic asymptotics of the transition times as well as the main
states are not random in spite of that all the transitions are due to the random perturbations.
This deterministic component of the long-time evolution of the perturbed system we call quasi-deterministic
approximation.
Let T = T () be such that lim↓0  ln T () = λ > 0. What can one say on the position of the perturbed trajectory
at the time T ()?
Let x ∈ Rd do not belong to a separatrix. Then for any λ, besides a finite number of points, there exists a cycle
S
C such that the trajectory Xt will come to the basin of i∈C Ki before the time αT (), α > 0 with the probability
close to 1 as  ↓ 0, but Xt does not have enough time to leave that basin before AT (), A < ∞. Moreover, the
rotation time for C is o(T ()) as  ↓ 0, so that XtT 
() , 0 < t ≤ A < ∞, approaches as  ↓ 0 the sub-limiting
distribution which is concentrated at Kµ , where µ = µ(x, λ) is the main state of the cycle C.
The state Kµ is called metastable state (for the timescale T ()  eλ/ and the initial point x). If λ is greater than
the rotation rate r ∗ of the cycle of the highest rank (such a cycle is unique and includes the whole set L = {1, . . . , l}),
then µ(x, λ) is independent of x and is the same for all λ > r ∗ . In general, the metastable state depends on λ and x.
Let now characteristics of the system and of its perturbations be changing slowly in time:
√
Ẋt = b(t/T , Xt ) + σ (t/T , Xt ) ◦ Ẇt . (1.6)
336 M.I. Freidlin / Physica D 137 (2000) 333–352

Here T = T ()  eλ/ , λ > 0, is a large parameter as  ↓ 0 so that the coefficients in (1.6) are changing slowly.

This implies that the trajectory XtT () first approaches the metastable state for the system with frozen dependence
on time, and then evolves together with the metastable state.
The positions of the equilibrium points Ki (t) as well as their number now depend on time. The numbers Vijt and the
function µ(x, λ) also depend on time. Therefore, XtT  t
() , 0 < t ≤ A, will be close to φ(t) = φ(t, x, λ) = KM(x,λ) .

Of course, the trajectory XtT () will not be all the time near φ(t), but it will be most of the time near φ(t), and
Z A p

XtT () − φt dt → 0
0

in probability as  ↓ 0 for any p > 0, A > 0.

In particular, if the coefficients b(t, x), σ (t, x) are, say, 1-periodic in t, then φ(t) also is periodic for t ≥ n0 ,
where n0 depends on the initial point and on the number of stable compacts of the system, 0 ≤ n0 < l. If the number
of stable attractors is equal to 2, then φ(t) has period 1. If the number of attractors is greater than 2, the period may
be greater than 1, but always is an integer not greater than the number of compacts. The period, in general, depends
on λ and on the initial point. Thus, XtT 
() is close to a periodic function as 0 <   1. This phenomenon is called
stochastic resonance.
After the paper [1] appeared, hundreds of papers have been published where models with stochastic resonance
and its modifications are used in various areas of physics, chemistry, neurophysiology and engineering. A survey
of applications of the stochastic resonance is given in [5]. About 300 papers are cited there. There exists a large
number of papers where stochastic resonance is studied using the digital or analog simulation. But there are no
papers where a satisfactory mathematical theory of these phenomena is given. This can be done in the frame of the
large deviation theory for random perturbations of dynamical systems, and it is the main goal of this paper. I should
say that the term stochastic resonance is now used for a very wide class of phenomena, and of course, not all the
questions are considered here.
The general construction and the main general results are given in the next section for the white-noise type
perturbations. Some examples are considered in Section 3.
Other classes of random perturbations, in particular, of type (1.5) are treated in Section 4. We also consider in
Section 4 Markov chains with exponentially small transition probabilities. Slow change of the exponents of these
probabilities may lead to the stochastic resonance. This is, actually, a basic model for all other manifestations of the
stochastic resonance.
Finally, in Section 5, we consider some generalizations: stochastic resonance for simplest systems with a hysteresis
is studied there. The resonance can be observed for such systems even if there is just one stable equilibrium. We use
the system with hysteresis also to demonstrate that the moderate large deviations can lead to stochastic resonance
with a shorter period.
We also consider in the last section how a similar approach allows to construct a solution close to periodic one
in an autonomous system perturbed by a noise.

2. General construction: white noise type perturbations

Consider the stochastic differential equation in Rd :

√
Ẋt = b(Xt ) + σ (Xt ) ◦ Ẇt . (2.1)

Here Wt is the Wiener process in Rd , b(x) is a smooth vector field in Rd , σ (x) is a d × d-matrix with smooth
entries, 0 <   1. Put a(x) = σ (x)σ ∗ (x) and assume that det a(x) 6= 0. Let (aij (x)) = a −1 (x). The last term in
 M.I. Freidlin / Physica D 137 (2000) 333–352 337

the right-hand side of (2.1) is understood in Stratonovich sense (see, for example, [6]) so that the generator of the
diffusion process (2.1) has the form

L = div(a(x)∇) + b(x) · ∇.
2
First, we recall some notions and results from [4]. The action functional S0T for the family of processes Xt ,
 ↓ 0, in the space C0T of continuous functions φ : [0, T ] → Rd is defined as (1/)S0T (φ), where
( 1 R T Pd i i j j 
2 0 i,j =1 aij (φs )(φ̇s − b (φs ))(φ̇s − b (φs )) ds, φ − abs. cont., φ0 = X0 = x,
S0T (φ) =
+∞ for the rest of C0T .

Let

V (x, y) = inf {S0T (φ) : φ ∈ C0T , φ0 = x, φT = y, T > 0} ,

x, y ∈ Rd . The function V (x, y) is Lipschitz continuous. We say that x is equivalent to y, (x ∼ y), if and only if
V (x, y) = V (y, x) = 0.
Assumption 1. Let lim|y|→∞ V (0, y) = ∞. Assume that a finite number of compacts K1 , K2 , . . . , Kl 0 ⊂ Rr exist
such that
1. any trajectory of the system

Ẋt = b(Xt ) (2.2)

is attracted to one of the compacts Ki as t → ∞;

2. if x, y ∈ Kj , then x ∼ y;
3. if x ∈ Ki , y ∈ Kj , i 6= j , then x 6∼ y.
Define Vij = V (x, y) for x ∈ Ki , y ∈ Kj . It follows from Assumption 1 (2) that the definition of Vij is
independent of the choice x ∈ Ki , y ∈ Kj . Some of compacts Ki attract their small enough neighborhoods (stable
compacts), some do not. Just the stable attractors are essential for us, so that we throw away the unstable compacts.
Let l ≤ l 0 be the number of stable compacts Ki and they are indexed by numbers 1, 2, . . . , l.
Now, following [3] (see also [4], Section 6.6), we introduce a hierarchy of cycles related to the processes Xt as
 ↓ 0. Define J (i) = j , i, j ∈ {1, . . . , l} = L, by the equality

Vij = min Vik .

k∈L\{i}

We assume that for each i ∈ L, this minimum is achieved just for one k = J (i) = j . We say that the compact Kj
follows after Ki (or j follows after i) if j = J (i).
Let the cycles of rank 0 be the compacts K1 , . . . , Kl themselves. Starting from any i ∈ L, one can consider the
sequence i, J (i), J 2 (i), . . . , J k (i), . . . , where J k=1 (i) = J (J k (i)), k = 1, 2, . . . . Let m = min{k > 0 : J k (i) =
J n (i), n < k}. This means that the states J n (i) → J n+1 (i) → · · · → J m (i) = J n (i) form a cycle. We say that
i, J (i), . . . , J n−1 (i), J n (i) → J n+1 (i) → · · · → J m (i) = J n (i) are the cycles of rank 1 (1-cycles) generated by
the state i ∈ L. Denote by C (1) the set of 1-cycles generated by any i ∈ L. Note that some of 1-cycles consist of
just one state.
Now we define for 0-cycles and for 1-cycles notions of main state, stationary distribution rate, rotation rate and
exit rate. Later we define the cycles of order greater than 1 and introduce the same notions for the higher order
cycles.
338 M.I. Freidlin / Physica D 137 (2000) 333–352

For a 0-cycle C consisting of one point i ∈ L, we define the main state M(C) = i, stationary distribution rate
mC (i) = 0, rotation rate R(C) = 0 and the exit rate E(C) = ViJ (i) .
For a 1-cycle C, we define main state M(C) = k ∗ ∈ C, Vk ∗ J (k ∗ ) = maxi∈C ViJ (i) . We assume that the maximum
is attained just for one i = k ∗ . Stationary distribution rate mC (i), i ∈ C, is defined as mC (i) = ViJ (i) − Vk ∗ J (k ∗ ) .
Define the rotation rate R(C) = maxi∈C ViJ (i) . The exit rate E(C) for the 1-cycle is defined as follows:
E(C) = min (mC (i) + Vij ).
i∈C,j ∈C
/

Assume that there exist just one i = i ∗ ∈ C and just one j = j ∗ ∈ / C for which the last minimum is attained.
Now we define by induction the cycles of higher ranks, main states for them as well as the rates mC (i), R(C),
E(C).
Suppose we already introduced the cycles of rank k (k-cycles), main states m(C) and the rates mC (i), R(C), E(C)
for them. Let C (k) be the set of all k-cycles. Then for any C1 ∈ C (k) , define C2 ∈ C (k) as the k-cycle containing the
point j ∗ ∈ L such that
min [mC1 (i) + Vij ]
i∈C1 ,j ∈C
/ 1

is attained for j = j ∗ and some i = i ∗ ∈ C1 . We assume that such i ∗ and j ∗ are unique. The point j ∗ = j ∗ (C2 ) is
called the entrance point of the cycle C2 ∈ C (k) ; i ∗ = i ∗ (C1 ) is called the exit point of C1 ∈ C (k) . We say that the
k-cycle C2 follows after C1 ∈ C (k) and write C2 = J (C1 ). Consider now the sequence
C1 , J (C1 ), . . . , J m (C1 ), . . .
Let m = min{m > 0 : J m (C1 ) = J n (C1 ), n < m}. Then we say that the k-cycle C1 generates (k + 1)-cycles
{C1 }, . . . , {J n−1 (C1 )}, {J n (C1 ) → J n+1 (C1 ) → · · · → J m (C1 ) = J n (C1 )}.
Considering all the (k + 1)-cycles generated by any k-cycle C1 ∈ C (k) , we obtain the set C (k+1) of the cycles of
rank k + 1. Note that among (k + 1)-cycles, there are cycles consisting just of one k-cycle. One can define cycles
of the next order until all the points of L = {1, . . . , l} belong to one cycle.
Now I will remind the notion of i-graph for a finite set L = {1, . . . , l} (see [4]).
A system of directed edges (arrows) connecting some of the points j ∈ L is called an i-graph if one arrow starts
from any point j ∈ L \ {i}, and one can come to the point i along these arrows starting from any point j ∈ L \ {i}.
Denote by Gi (L), i ∈ L, the set of all i-graphs for the finite set L. As it is shown in [4], the i-graph is a
convenient notion to describe asymptotic behavior of the process Xt on large time intervals. In particular, the main
states, stationary distribution rate and exit rate can be described using the i-graphs.
The main state M(C) for a (k + 1)-cycle C is defined as follows: M(C) = j ∗ (C) if
X
min min Vmn
j ∈C g∈Gj (C)
(m→n)∈g

is attained for j = j ∗ . We assume that such a j ∗ is unique.

(k) (k)
Define the rotation rate R(C) for C ∈ C (k+1) : R(C) = maxi:C (k) ∈C E(Ci ), where Ci are the k-cycles which
i
(k) (k)
form the (k + 1)-cycle C and E(Ci ) is the exit rate for the k-cycle Ci .
The stationary distribution rate mC (i) for C ∈ C (k+1) , i ∈ C, is defined by the equality
X X
mC (i) = min Vmn − min Vmn ,
g∈Gi (C) g∈Gj ∗ (C)
(m→n)∈g (m→n)∈g

where j ∗ = M(C) is the main state of C defined above.

 M.I. Freidlin / Physica D 137 (2000) 333–352 339

Define the exit rate E(C) for C ∈ C (k+1) :

E(C) = min (mC (i) + Vij );

i∈C,j ∈C
/

i ∗ = i ∗ (C) and j ∗ = j ∗ (C) for which this minimum is attained is supposed to be unique; i ∗ is called the exit point
of C, and j ∗ is the entrance point for the (k + 1)-cycle which contains the state j ∗ . For the highest rank cycle C
containing all the points of L = {1, . . . , l}, put E(C) = +∞.
The rotation rate R(C) for C ∈ C (k+1) is defined as follows:
(k)
R(C) = max E(Ci ),
(k)
Ci ∈C

where the maximum is taken over all k-cycles included in C ∈ C (k+1) .

Assumption 2. We assume that each of the minima and maxima considered above are attained just at one point. It
is clear that this assumption is satisfied for a generic system.
S
Denote by D(i) ⊂ Rd the domain attracted to the compact Ki , i ∈ L = {1, . . . , l}. Let D(C) = i∈C D(i),
C ⊂ L. Let τC be the exit time from D(C):

τC = inf{t : Xt ∈

/ D(C)}.

It follows from Theorem 6.6.2 of [4] that for any cycle C,

lim  ln Ex τC = E(C), x ∈ D(C). (2.3)

↓0

Moreover, for any γ > 0,

n o
lim Px exp{ −1 (E(C) − γ )} < τC < exp{ −1 (E(C) + γ )} = 1 (2.4)
↓0

uniformly in x ∈ F for any compact F ⊂ D(C).

If the initial point x ∈ D(i), then the trajectory Xt , 0 <   1, first will come to a small neighborhood of
(1)
Ki . Then at time τi , it will leave Di for a neighborhood of KJ (i) . Then it will rotate in the 1-cycle C1 generated
(1) (1) (1)
by Ki . In times greater than exp{E(C1 )/}, the trajectory comes to 1-cycle C2 which follows after C1 . It will
(1)
rotate along 2-cycle generated by C1 , then along 3-cycle and so on up to the highest order cycle containing all the
compacts. The transition times are described by (2.3) and (2.4).
Let T = T (), 0 < , be such a function that

lim  ln T () = λ > 0. (2.5)

↓0

Consider XtT 
() , 0 ≤ t ≤ A < ∞. Let X0 = x ∈ Di and let

{i} = C (0) (x) ⊂ C (1) (x) ⊂ · · · ⊂ C (k)

be the sequence of the cycles of growing rank containing i = i(x). The cycle C (k) of the highest rank is independent
of x. Let em = em (x) = E(C (m) (x)) be the exit rates:

e0 = ViJ (i) < e1 < · · · < ek−1 < ek = ∞.

Let rm = R(C (m) (x)) be the corresponding rotation rates. It is clear that rm is an increasing sequence and rm < em .
340 M.I. Freidlin / Physica D 137 (2000) 333–352

Let m∗ be such an integer number that

em∗ < λ < em∗ +1 . (2.6)

Now we define a state Kµ , µ = µ(x, λ) ∈ C (m∗ +1)

, such that the process Xt , X0 = x, spends most of the
time, during the time interval [0, AT ()], in a neighborhood of Kµ (support of the sub-limiting distribution in the
timescale T () [3], metastable state).
∗ ∗
If λ > rm∗ +1 , then µ(x, λ) = M(C (m +1) (x)) is the main state of C (m +1) (x). If λ < rm∗ +1 , then consider
∗ ∗ ∗ ∗
m∗ -cycles following C (m ) (x) in C (m +1) (x). Let Ĉ (m ) be the first of them for which E(Ĉ (m ) ) > λ. Such a
(m∗ )
cycle always can be found since according to our assumption λ < rm∗ +1 = maxC (m∗ ) ∈C (m∗ +1) (x) E(Ci ). Now if
i
∗ ∗ ∗ ∗
r(Ĉ (m ) ) < λ, put µ(x, λ) = M(Ĉ (m ) ). If r(Ĉ (m ) ) > λ, consider (m∗ − 1)-cycles included in Ĉ (m ) . There is an
∗ ∗ ∗
(m∗ − 1)-cycle C̃ (m ) among them which contains the entrance state of Ĉ (m ) . The (m∗ − 1)-cycles in Ĉ (m ) are
∗ ∗
ordered. Start with C̃ (m +1) and take the first of those (m∗ − 1)-cycles in Ĉ (m ) for which exit rate is greater than λ.
(m∗ −1) (m∗ −1) (m∗ −1) ∗ ∗
Let it be Ĉ . If r(Ĉ ) < λ, put µ(x, λ) = M(Ĉ ) — main state of Ĉ (m −1) . If r(Ĉ (m −1) ) > λ, then
∗ ∗ ∗
go to (m − 2)-cycles included in Ĉ (m −1) and so on until we come to a cycle Ĉ (m −n) such that E(Ĉ (m −n) ) > λ
∗ ∗
and r(Ĉ (m −n) ) < λ. Then put µ(x, λ) = M(Ĉ (m −n) ). Such a number n ≥ 0 exists since the rotation rate for the
0-cycles is equal to zero.
Denote by Λ(G) the Lebesgue measure of a set G ⊂ R1 and by ρ(·, ·) the Euclidean metric in Rd .
Theorem 1. Let Assumptions 1 and 2 be satisfied and let T = T () satisfy (2.5). Then, for any δ > 0 and A > 0,

Λ{t ∈ [0, A] : ρ(XtT () , Kµ(x,λ) ) > δ} → 0 (2.7)

in probability Px as  ↓ 0. In particular, if Kµ(x,λ) consists of one point, then Λ{t ∈ [0, A] : |XtT () − Kµ(x,λ) | >
δ} → 0 in probability Px as  ↓ 0. In general, if there exists just one
R normalized invariant measure ν of the system

(2.2) with support on Kµ(x,λ) , then XtT converges to z(x, λ) = x ν(dx) weakly in probability Px :
() Kµ(x,λ)
Z A 

lim Px φ(s)(XsT () − z(x, λ)) ds > δ =0 (2.8)
↓0 0

for any C ∞ -test function φ(s) and for any δ > 0.

∗
Proof. Let m∗ be defined by (2.6). Then µ(x, λ) is the main state of a (m∗ − s + 1)-cycle Ĉ (m −s+1) , s ∈
∗ ∗
{0, 1, . . . , m∗ + 1}, such that E(Ĉ (m −s+1) ) > λ and r(Ĉ (m −s+1) ) < λ. Moreover, according to the definition of
µ(x, λ) and to bound (2.4), for any t0 > 0 and any neighborhood U of the compact Kµ(x,λ) , the trajectory XtT 
()
∗ −s+1)
will hit U before time t0 with the Px -probability close to 1 if  > 0 is small enough. Since E(Ĉ (m ) > λ,
 (m∗ −s+1)
lim Px {XtT () ∈ D(Ĉ ), t ∈ [t0 , A]} = 1.
↓0

∗ 
Now, since r(Ĉ (m −s+1) ) < λ, the trajectory XtT () , starting from a small neighborhood of Kµ(x,λ) , has enough
time to approach the sub-limiting distribution (see [3] and Ch. 6 of [4]), which for the timescale tT (), 0 < t <
A < ∞, is concentrated on Kµ(x,λ) . This implies (2.7). A slight modification of the proof of Theorem 4.4.2 of [4]

allows to prove that the distribution of XtT () converges weakly to the invariant measure concentrated on Kµ(x,λ)
given that there exists just one normalized invariant measure of system (2.2) concentrated on Kµ(x,λ) . Equality (2.8)
follows from this statement. 
Corollary 1. Let Kµ(x,λ) consist of just one point. According to Assumption 1, lim|y|→∞ V (0, y) = ∞. Therefore,
the domain Fx,λ = {y ∈ Rd : V (x, y) ≤ λ} is bounded and thus XtT 
() , 0 ≤ t ≤ A , is bounded in probability:
 M.I. Freidlin / Physica D 137 (2000) 333–352 341


lim Px {XtT () ∈
/ Fx,λ for some t ∈ [0, A]} = 0.
↓0

It follows from the last statement and from (2.7) that


kXtT () − Kµ(x,λ) kLp → 0 (2.9)
in probability as  ↓ 0 for any Lp -norm, p > 0.
Corollary 2. Let δ, h > 0 and a ∈ (h, A). It follows from the definition of the action functional and of the function
V (x, y) that the trajectory XtT   d
() , X0 = x ∈ R , with probability Px close to 1 as  ↓ 0 visits δ-neighborhood
of any point of the set {y ∈ Rd : V (Kµ(x,λ) , y) ≤ λ} during the time interval ((a − h)T (), aT ()). Moreover, the
trajectory Xt will spend exponentially large (but of the logarithmic order less than λ = lim↓0  ln T () ) times in
the basins of the compacts Kj such that Vµ(x,λ),j < λ. But if additional to the conditions of Theorem 1, we assume

that λ < Vµ(x,λ),J (µ(x,λ)) , then the trajectory XtT 
() , X0 = x, 0 ≤ t ≤ A , will stay inside the basin of Kµ(x,λ) after
entering this basin with Px -probability close to 1 as  ↓ 0. More precisely: let U = {y ∈ Rd : V (Kµ(x,λ) , y) < λ}.
Then for any δ > 0,

lim Px {XtT () ∈ U for δ ≤ t ≤ A} = 1. (2.10)
↓0

Note that the set U can be a rather small neighborhood of Kµ(x,λ) .

Consider now a system with time-dependent coefficients
√
Ẋt = b(t/T , Xt ) + σ (t/T , Xt ) ◦ Ẇt , X0 = x ∈ Rd . (2.11)
Here  −1 and T are large positive parameters. Let, first, b(t, x) and σ (t, x) be step-functions in t: points 0 = t0 <
t1 < · · · < tn = A < ∞ exist such that
b(t, x) = bk (x), σ (t, x) = σk (x) for tk−1 ≤ t < tk , k ∈ {1, 2, . . . , n}. (2.12)
We assume that bk (x) and σk (x) are smooth enough and growing not too fast as |x| → ∞. Let the diffusion matrices
ak (x) = σk (x)σk∗ (x) be uniformly non-degenerate in Rd .
(k) (k)
Let the vector field bk (x) have lk stable compacts K1 , . . . , Klk . Denote by V (k) (x, y) the function V (x, y)
defined for the process (2.1) with b(x) = bk (x), σ (x) = σk (x).
Assume that for each couple (bk (x), σk (x)) Assumptions 1 and 2 are satisfied. Denote by µk (x, λ) the function
µ(x, λ) calculated for the drift bk (x) and the diffusion matrix ak (x).
(1)
Let T = T (), lim↓0  ln T () = λ > 0. Let the initial point x belong to the basin of Ki(x) for system (2.1) with
b(x) = b1 (x), σ (x) = σ1 (x).
Assume for brevity that for given λ and initial point x, Kµk (x,λ) consists of just one point. Put

π 1 (x, λ) = µ1 (x, λ), π k (x, λ) = µk (π k−1 (x, λ), λ) (2.13)

for k ∈ {2, 3, . . . , n}.
Define the function φ(t) = φ(t, x, λ), 0 ≤ t ≤ A:
(j )
φ(t) = Kπ j (x,λ) for tj −1 ≤ t < tj , j ∈ {1, . . . , n}.

Theorem 2. Let Xt be defined by Eq. (2.11) with the coefficients given by (2.12). Let the assumption on bk (x) and
σk (x) mentioned above be satisfied. Then for any δ > 0,

Λ{t ∈ [0, A] : |XtT () − φ(t)| > δ} → 0 (2.14)
342 M.I. Freidlin / Physica D 137 (2000) 333–352

in probability Px as  ↓ 0. For any p > 0 ,

Z A 
 p
lim Px |XtT () − φt | dt > δ = 0. (2.15)
↓0 0

Assume additionally that

(m)
λ < min Vπ m (x,λ),J (π m (x,λ)) .
1≤m≤n

(m)
Put Um = {y ∈ Rd : V (m) (Kπ m (x,λ) , y) < λ}. Then for any small δ > 0,

lim Px {XtT () ∈ Um for t ∈ [tm−1 + δ, tm ], m ∈ {1, . . . , n}} = 1. (2.16)
↓0

Proof. Proof of statement (2.14) follows from Theorem 1. Equality (2.15) follows from Corollary 1, and (2.16)
follows from Corollary 2.
Let now the coefficients b(t, x) and σ (t, x) in (2.11) be Lipschitz continuous in t and x. Then for any fixed
t0 ∈ [0, A], one can consider the time-homogeneous process Xt,to which is the solution of (2.11) with the drift
b(t0 , x) and the diffusion matrix a(t0 , x) = σ (t0 , x)σ ∗ (t0 , x).
Let Ki (t), i ∈ {1, . . . , l(t)}, be the stable compacts defined for the field b(t, x) with a fixed t; V t (x, y) be the
corresponding V (x, y)-function and Vijt be the corresponding l(t) × l(t)-matrix, µ(t, x, λ) is defined using Vijt .
Assume that there exists at most a finite set {t1 , . . . , tn−1 } ⊂ [0, A] such that Assumptions 1 and 2 are satisfied
for t ∈/ {t1 , . . . , tn−1 }. Moreover, assume that at any point tk ∈ {t1 , . . . , tn } a simple reconstruction of the hierarchy
of cycles occurs: just one of strict inequalities, which are assumed by Assumption 2, becomes an equality as the
parameter t goes through the point tk (in particular, just one Vijt can turn equal to zero). Between the points tk−1 and
tk , no reconstruction of the hierarchy of cycles occurs so that the function µ(t, x, λ) = µk (x, λ) does not change in
[tk−1 , tk ), in spite of the dependence of the functions b(t, x) and σ (t, x) on t. Put t0 = 0 and tn = A.
Let π k (x, λ) be defined in the same way as in (2.13). We assume for brevity that each of the sets Kπ j (x,λ) (t) for
t ∈ [tj −1 , tj ), j ∈ {1, . . . , n} consists of just one point. Put

φ(t) = Kπ j (x,λ) (t) for tj −1 ≤ t < tj , j ∈ {1, . . . , n}. (2.17)

Note that now φ(t), in general, is not a constant between tj −1 and tj . 

Theorem 3. Assume that the coefficients b(t, x) and σ (t, x) in (2.11) are Lipschitz continuous. Let there exist
at most finite number of points 0 < t1 < · · · < tn−1 < A such that Assumptions 1 and 2 are satisfied for any
t∈/ {t1 , . . . , tn−1 } and simple reconstructions occur at the points tj . Let lim↓0  ln T () = λ > 0. Then statements
(2.14) and (2.15) hold if φ(t) is defined by (2.17), t0 = 0 and tn = A.
If additionally for some h > 0 ,

λ < min min Vπ j (x,λ),J (π j (x,λ)) , (2.18)

j h+tj −1 ≤t≤tj −h

then

lim Px {XtT () ∈ Um (t) for t ∈ (tm−1 + h, tm − h), m ∈ {1, . . . , n}} = 1, (2.19)
↓0

where Um (t) = {y ∈ Rd : V t (Kπ m (x,λ) )(t) , y) < λ}.

This theorem follows from Theorem 2 if one takes into account the comparison results for Markov chains proven

in Section 6.2 of [4] and the fact that the process XtT () , with probability Px tending to 1 as  ↓ 0, spends most
 M.I. Freidlin / Physica D 137 (2000) 333–352 343

of the time in the time interval (tk − h, tk + h) in the union of neighborhoods of Kπ k (x,λ) (t) and Kπ k+1 (x,λ) (t).
The proof of the last statements uses the assumption that the system has at tk a simple reconstruction. We omit the
detailed proof.
Let now the functions b(t, x) and σ (t, x) be 1-periodic in t:

b(t + 1, x) ≡ b(t, x), σ (t + 1, x) ≡ σ (t, x).

Let 0 = t0 < t1 < · · · < tn = 1 be the reconstruction points on [0, 1] (discontinuity points also can be treated
as reconstruction points). The other reconstruction points can be obtained from t0 , t1 , . . . , tn−1 by a shift on an
integer. Since b(t, x) and σ (t, x) are 1-periodic, the number l(t) of stable compacts at time t and the positions of
the compacts Ki (t) are also 1-periodic.
For a given λ > 0, define a map Π of the set L = {1, 2, . . . , l(0)} in itself:

Π(i) = π n (x, λ), x ∈ Ki (0), i ∈ L.

This definition is correct since π n (x, λ) depends just on the number of the compact attracting the point x, but not on
x itself. Since the set L has a finite number l(0) of points, the sequence i, Π(i), Π 2 (i), . . . , Π n (i), . . . , is periodic
starting from some n = m(i) ≤ l(0). The period N = N (i) of this sequence depends on i but it is an integer not
greater than l(0). This implies that in the case of 1-periodic coefficients, the function φ(t, x, λ) = φ(t) is periodic
in t for t ≥ m(i) with the period N(i):

φ(m(i) + t) = φ̃(t/N(i)), t > 0,

φ̃(t) is 1-periodic; m(i), N(i) ≤ l(0). This together with Theorem 2 implies the following result.
Theorem 4. Let the functions b(t, x) and σ (t, x) in (2.11) be 1-periodic in t, l(0) be the number of stable com-
pacts for the field b(0, x). Let the conditions of Theorem 3 be satisfied. Then there exist non-negative integers
m(i), N (i) ≤ l(0), 1 ≤ i ≤ l(0), such that for x from the basin of the compact Ki (0) and any A > m(i) and
p > 0:
Z A 
 p
lim Px |XtT () − φ̃(t/N (i))| dt > δ = 0. (2.20)
↓0 m(i)

If additionally (2.18) is satisfied, then equality (2.19) holds.

Corollary 3. Equality (2.20) means that Xt is close to φ̃(t/N (i)T ()) as  ↓ 0 , 0 ≤ t ≤ AT (). The last function
is periodic with period T ()N(i). If the transformation Π is such that Π k (i) = j for any i, j ∈ L and some
t − V t changes the
k = k(i, j ), then N(i) = N is independent of i. For example, if l(t) = 2 and the difference V12 21
sign in the interval 0 ≤ t ≤ 1, then the period is equal to T ().
A counterpart of Theorem 4 also holds for periodic b(t, x) and σ (t, x) which are step-functions in t as in
Theorem 2.

3. Examples

1. Consider Eq. (2.11) in the one-dimensional case with σ (t, x) ≡ 1:

Ẋt = b(t/T , Xt ) +  Ẇt . (3.1)

344 M.I. Freidlin / Physica D 137 (2000) 333–352

Fig. 1.

Let b(t, x) be 1-periodic in t and satisfy condition (2.12) with n = 2:

(
−B10 (x) for 0 ≤ t ≤ t1 < 1,
b(t, x) =
−B20 (x) for t1 ≤ t < 1.

The potentials B1 (x) and B2 (x) are given in Fig. 1a,b.

There are three stable attractors for each of the fields −B10 (x) and −B20 (x) : K1 = {x1 }, K2 = {x3 }, and
K3 = {x5 }. As it follows from Section 4.3 of [4], the function V i (x, y), i = 1, 2, for each of these fields can be
expressed through the potential: for x and y from the same well, V i (x, y) = 2(Bi (y) − Bi (x)) ∨ 0. This gives the
following values for Vij1 and Vij2 — the numbers Vij calculated for the potentials B1 (x) and B2 (x), respectively:
1 = 16,
V12 V13
1 = 16, V21
1 = 8, V23
1 = 6, V31
1 = 12, V32
1 = 4,

2 = 16,
V12 V13
2 = 16, V21
2 = 6, V23
2 = 4, V31
2 = 12, V32
2 = 6.

There are two 1-cycles for each of the fields and they are the same for −B10 (x) and for −B20 (x):

C11 = {1}, C21 = {2 → 3 → 2}.

The 2-cycle contains already all the states. But the main states of the cycle C21 are different for the fields −B10 (x)
and −B20 (x). The main state of C21 for −B10 (x) is M1 (C21 ) = 2, and M2 (C21 ) = 3 for −B20 (x).
Let λ = lim↓0  ln T () > 0, X0 = x. If x belongs to the basin D(1) of x1 , then µi (x, λ) = 1, i = 1, 2, for any
λ > 0. If x ∈ D(2),

(
2 if λ < 8,  2 if λ < 4,

µ1 (x, λ) = µ2 (x, λ) = 3 if 4 < λ < 8,
1 if λ > 8, 

4 if λ > 8.

If x ∈ D(3), then

 if λ < 4, (
3 3 if λ < 8,
µ (x, λ) = 2
1
if 4 < λ < 8, µ2 (x, λ) =

 1, if λ > 8.
1 if λ > 8,
 M.I. Freidlin / Physica D 137 (2000) 333–352 345

This means that if x ∈ D(1), XtT  

() , 0 ≤ t ≤ A < ∞, X0 = x, will be close to x1 for any λ > 0. No periodic
oscillations will be observed in this case. If x ∈ D(2) ∪ D(3) and 4 < λ < 8, then XtT 
() , 0 ≤ t ≤ A < ∞,
0 <   1, will be close to the 1-periodic function φ(t):
(
x3 if t ∈ [0, t1 ),
φ(t) =
x5 if t ∈ [t1 , 1).

If λ < 4, then XtT () will stay near the attractor of the initial point for 0 ≤ t < A < ∞ and 0 <   1 with the

probability close to 1. If λ > 8, then XtT () will be close to x1 for any initial point x ∈ R and  > 0 small enough.
1

The closeness is understood in the sense of (2.14) or (2.15). If λ ∈ (4, 6), then condition (2.16) is satisfied. In this
case, for any δ > 0, and any integer n < A − 1,

N3− (λ) < XtT () − x3 < N3+ (λ) for n + δ < t < n + t1 ,
N5− (λ) < XtT () − x5 < N5+ (λ) for n + t1 + δ < t < n + 1,

with Px probability close to 1 for x ∈ D(2) ∪ D(3) if  > 0 is small enough. The numbers N3∓ (λ) satisfy the
equation 2B1 (N3∓ (λ)) = λ + 10, and N5∓ (λ) are the solutions of the equation 2B2 (N5∓ (λ)) = λ + 10 (see Fig.
1a,b).
2. Consider now a potential (gradient) system in Rd perturbed by the white noise:
√
Ẋt,z = −∇B(z, Xt ) +  Ẇt . (3.2)

Here ∇B(z, x) is the gradient in x ∈ Rd for a fixed parameter z ∈ R1 . Let K1 (z), . . . , Kl(z) (z) be the minima
of B(z, x) in x for a given z. For any z, one can introduce the function V z (x, y). In the potential case V z (x, y) =
2(B(z, x) − B(z, y)) ∨ 0 if x and y belong to the same potential well. Taking into account that V z (x, y) =
minu∈Rd (V z (x, u) + V z (u, y)) for any x, y ∈ Rv d , one can now express Vijz through the potential B(z, x) and
describe the hierarchy of cycles, calculate main states, exit rates, rotations rates and the function µz (x, y) in an
explicit form. Let the parameter z in (3.2) be replaced by t/T :
√
Ẋt = −∇B(t/T , Xt ) +  Ẇt , X0 = x.

Then one can express the limiting function φ(t) = φ(t, x, λ) in a more or less explicit form through the functions
µt (x, λ) and Ki (t).
In particular, if a “signal” z = f (t/T ) is added to a potential system,
√
Ẋt = −∇U (Xt ) + f (t/T ) +  Ẇt ,

then B(z, x) = U (x) − z · x, and the function φ(t) is defined by the time evolution of the critical points of
U (x) − x · f (t).
3. Let U (x), x ∈ Rd , be a smooth function with a finite number of non-degenerate critical points, lim|x|→∞ U (x) =
∞. Let l(x), x ∈ Rd , be a smooth vector field in Rd such that l(x) · ∇U (x) = 0 for x ∈ Rd . Consider system (2.11)
with a non-potential vector field b(x) = −∇U (x) + l(x):
√
Ẋt = −∇U (Xt ) + l(Xt ) +  Ẇt , X0 = x ∈ Rd . (3.3)

The function U (x) is called the quasi-potential of the field b(x) [4]. The quasi-potential U (x) satisfies the equation

b(x) · ∇U (x) + |∇U (x)|2 = 0, x ∈ Rd ,

which is equivalent to the orthogonality condition l(x) · ∇U (x) = 0.

346 M.I. Freidlin / Physica D 137 (2000) 333–352

Fig. 2.

Consider the graph Γ homeomorphic to the set of all connected components of the level sets of the function U (x)
(Fig. 2).
The critical points correspond to the vertices of the graph. The vertices corresponding to the saddle points are
called interior; the extrema of U (x) correspond to the exterior vertices. Let the edges of Γ be indexed by number
1, . . . , n. The pairs (U, k) form a global coordinate system on Γ : a point y ∈ Γ has coordinates (H, k) if it
belongs to the edge indexed by k and corresponds to a component of {x ∈ Rd : U (x) = H }. Put Y : Rd → Γ ,
Y (x) = (U (x), k(x)) is the point of Γ corresponding to the level set component containing x. The graph Γ has the
structure of a tree.
Let A(z, y) = A(z; U, k), y = (U, k) ∈ Γ , z ≥ 0, be a twice continuously differentiable function on
[0, ∞) × Γ . The derivatives of A(z, H, k) calculated along the edges attached to a vertex 0 coincide at 0. Let
0 < ∂A(z; U, k)/∂U < M < ∞. Let β(z, x), x ∈ Rd , z ≥ 0, be a continuously differentiable scalar function.
Consider the system
√
Ẋt = −∇A(t/T ; Y (Xt )) + β(t/T , Xt )l(Xt ) +  Ẇt . (3.4)

Since ∇A(z; U (x), k(x)) = (∂A/∂U )(z; U (x), k(x))∇U (x), ∇A(z; U (x), k(x)) · l(x) = 0, and the function
A(z; U (x), k(x)) is the quasi-potential for the field b(z, x) = −∇A(z; U (x), k(x)) + β(z, x)l(x) for any
z ≥ 0. The field b(z, x) has the same equilibrium points as ∇U (x); just the value of the quasi-potential
A(z, x) = A(z; U (x), k(x)) is changing together with z. The numbers Vijz can be expressed through the
values of A(z, x) at the critical points. These numbers define the hierarchy of cycles and the function
µz (x, λ).
Let, for example, U (x) have two minima at x = x1 and x = x2 separated by a saddle point at x = x3 ,
lim|x|→∞ U (x) = ∞. The trajectories of the vector field b(z, x) are shown in Fig. 3.

Let lim↓0  ln T () = λ > 0. Then the trajectory XtT 
() , X0 = x is close to φ(t) = φ(t, x, λ),
(
x1 if A(t, x2 ) > A(t, x1 ),
φ(t) =
x2 if A(t, x2 ) < A(t, x1 ).

The vector field b(z, x) = −∇A(z; U (x), k(x)) + β(z, x)l(x) may have attractors different from equilibrium
points. It may have, for example, limit cycles. Transition between such attractors can also be described
explicitily.
 M.I. Freidlin / Physica D 137 (2000) 333–352 347

Fig. 3.

I will mention one more non-potential model where Vijz and the limiting behavior can be calculated explicitly: let

b(z, x) = −∇U (x) + z + l(x)

with ∇U (x) ⊥ l(x) and z ⊥ l(x), x ∈ Rd . Then the field b(z, x) has the quasi-potential U (x) − z · x.
If the diffusion matrix a(x) = σ (x)σ ∗ (x) is not unit, the gradient and the orthogonality should be understood in
P
the Riemannian metric corresponding to the form ds 2 = i,j aij (x) dx i dx j , (aij (x)) = a −1 (x).
4. Consider the one-dimensional equation
√
Ẋt = −B 0 (Xt ) + σ (t/T , Xt ) ◦ Ẇt . (3.5)

Let B(x) = B(−x), lim|x|→∞ B(x) = ∞, and let B(x) have two minima at x = 1 and at x = −1. Let σ 2 (t, x),
x ∈ R1 , t > 0, be strictly positive, bounded, 1-periodic in t function such that the difference σ 2 (t, x) − σ 2 (t, −x)
is positive for x > 0 and 0 < t < t1 < 1, and negative for x > 0, t1 < t < 1. Then, as one can derive from the
t < V t for 0 < t < t and V t > V t for t < t < 1. Assume that
definition of V (x, y), V12 21 1 12 21 1
   
t t
lim  ln T () = λ > min V12 ∨ min V21 .
↓0 0≤t≤1 0≤t≤1

Then, as follows from Theorem 4, one derives from Theorem 4 that XtT () , 0 ≤ t ≤ A < ∞, converges as  ↓ 0
to the 1-periodic step-function φ(t):

1, 0 ≤ t < t1 ,
φ(t) =
−1, t1 ≤ t < 1.

This example shows that the oscillations of XtT () can be enforced by the oscillations of the noise intensity.

4. Other classes of random perturbations

1. Consider an oscillator with one degree of freedom with a friction term and white noise perturbations:
√
q̈t + f (qt ) = −β q̇t +  Ẇt , β > 0, 0 <   1. (4.1)
Rq
Assume that the potential F (q) = 0 f (s) ds has a finite number of non-degenerate critical points and lim|q|→∞
F (q) = ∞. The process qt is not Markovian. But the pair (pt , qt ), pt = q̇t , form a Markov diffusion process:
√
ṗt = −f (qt ) − βpt +  Ẇt , q̇t = pt . (4.2)

Degeneration of this process does not allow immediate application of the results of Section 2.
348 M.I. Freidlin / Physica D 137 (2000) 333–352

As it follows from Section 3.4 of [4], the action functional for the family (pv , qt ) in the space C0T of continuous
functions (ψ(t), φ(t)) : [0, T ] → R2 is equal to
( RT
1 (1/2) 0 |ψ̇t + βψt + f (φt )|2 dt if ψ0 =p0 , φ0 =q0 and ψ=φ̇ is absolutely continuous,
S0T (ψ, φ) =
 +∞ for the rest of C0T .

This allows the introduction of the function V (x, y), x, y ∈ R2 in the same way as in Section 2. The set of stable
attractors for this system consists of the points (0, qk ) ∈ R2 , where qk are the zeros of f (q) where it changes sign
from “+” to “−” as q grows. The numbers Vij are defined through V (x, y). If now f = f (t/T , q), β = β(t/T ),
T = T ()  eλ/ , then the asymptotic behavior of (ptT  
() , qtT () ) as  ↓ 0 is described by the evolution of the
stable zeros of f (t, q) and of the numbers Vij = Vijt , 0 ≤ t ≤ A. In particular, if f (t, q) and β(t) are periodic, the
process (ptT 
() , qtT () ) is close to a periodic function.
System (4.2) is a special case of the system
√
Ẋt = b(Xt ) + ζt , X0 = x ∈ Rd ,

where ζt is a mean zero Gaussian process with the correlation operator R acting in L20T . Then the family Xt ,
0 ≤ t ≤ T , has the action functional (1/)S0T (φ),

S0T (φ) = 21 kR −1/2 (φ̇ − b(φ))k2L2 .

But the process Xt , in general, is not Markovian, and the corresponding function V (x, y) = inf{S0T (φ), φ0 =
x, φT y, T > 0} does not give the quasi-deterministic approximation for long-time behavior of Xt . If the process
ζt is the stationary solution of the equation P (d/dt)ζt = Q(d/dt)Wt , where P (λ) and Q(λ) are polynomials and
Wt is the Wiener process, then Xt is a component of a multidimensional Markov process. Then the construction
described in Section 2 can be applied. Some comments concerning the case of a general Gaussian process ζt can
be found in Section 4.5 of [4].
2. Let ηth be a one-dimensional process such that ηth has a jump during a short time interval dt with probability
(dt/ h) + o(dt). Let the jump be equal to ±h with equal probabilities, independent of the past behavior of the
process.
Consider the equation

Ẋth = b(Xth ) + η̇th . (4.3)

It is easy to see that Xth converges in probability as h ↓ 0 uniformly on any finite time interval [0, T ] to the trajectory
of the dynamical system

Ẋt = b(Xt )

with the same initial condition as Xth . A more precise approximation of Xth is given by the diffusion process

˙ h = b(X̃h ) + √hẆ .
X̃ (4.4)
t t t

This is a result of the central limit theorem. But the long-time behavior is defined by the large deviations and they
are different for the process Xt and X̃t . The action functional for the family (4.3) is equal to (1/ h)S0T (φ), where
(RT
h
0 L(φs , φ̇s ) ds, φ is absolutely continuous, φ0 = X0 ;
S0T (φ) = (4.5)
+∞ for the rest of C0T ,
 M.I. Freidlin / Physica D 137 (2000) 333–352 349

and
 q  q
L(x, β) = (β − b(x)) ln β − b(x) + (β − b(x)) + 1 − (β − b(x))2 + 1 + 1.
2

The proof of this statement can be found in Ch. 5 of [4]. Note that the action functional for process (4.4) differs
from (4.5).
z
If now the function b(x) in (4.3) depends on time: b = b(z, x), z = t/T , then one can introduce S0T (φ) by (4.5)
z z
with b depending on z. Using this action functional, the function V (x, y) and the numbers Vij are defined. Then
the long-term behavior of the process defined by (4.3) can be described as in Section 2.
Action functionals for a general class of perturbations, leading to Markov processes with both, continuous and
jumping, components, are calculated in ([4], Ch. 5).
3. Let ζt be a stationary process with regular enough trajectories and some mixing properties. Consider the system
 
t
Ẋt = b , Xt , ζt/ , X0 = x ∈ Rd . (4.6)
T

The field b(z, x, ζ ) is assumed smooth in x. Together with (4.6), consider the following system depending on z ∈ R1
as a parameter:

Ẋtz, = b(z, Xtz, , ζt/ ), X0z, = x ∈ Rd . (4.7)

Put b̄(z, x) = Eb(z, x, ζt ) and consider the averaged system

z z z
Ẋ t = b̄(z, X t ), X 0 = x ∈ Rd . (4.8)

One can prove (see, for example, Ch. 7 in [4]) that under mild assumptions concerning the mixing properties of ζt ,
for any T < ∞,
z
max |Xtz, − X t | → 0
0≤t≤T

in probability as  ↓ 0. This means that (4.7) can be viewed as a random perturbation of dynamical system (4.8).
Under certain assumptions concerning the process ζt , the normal and the large deviations of Xtz, from X̄tz can be
described.
Let us assume that ζt is a continuous time Markov process with a finite phase space {1, . . . , N} = E. Let pij (t) be
the transition probabilities, qij = dPij (t)/dt|t=0 . Assume that qij 6= 0, and that the convex envelope of the vectors
b(z, x, i), i ∈ E, has the origin of Rd as an interior point for any x ∈ Rd , z ∈ R1 . This assumption guarantees a
kind of non-degeneration of the random perturbations.
Consider the matrix Qz,α (x) = qijz,α (x), α ∈ Rd ,

qijz,α (x) = qij + α · b(z, x, i)δij ,

where δij is the Kronecker symbol. Let λ = λ(z, x, α) be the eigenvalue of Qz,α (x) having the maximal real part.
Such an eigenvalue exists, is simple, and is convex in α ∈ Rd (see [4], Section 7.4). Let L(z, x, β) be the Legendre
transformation of λ(z, x, α) in α:

L(z, x, β) = sup (α · β − λ(z, x, α)), β ∈ Rd .

α∈Rd

Then (see [4], Theorems 7.4.1 and 7.4.2) the action functional for the family Xtz, , 0 ≤ t ≤ T ,  ↓ 0, in C0T is
 (φ),
equal to (1/)S0T
350 M.I. Freidlin / Physica D 137 (2000) 333–352
(RT
L(z, φs , φ̇s ) ds, φ is absolutely continuous and φ0 = x,
S0T (φ) = 0
(4.9)
+∞ for the rest of C0T .

Now we can define, as in Section 2, function V z (x, y), x, y ∈ Rd , and the matrix Vijz and describe asymptotic
behavior of the process Xt defined by (4.6) as  ↓ 0.
Let now ζt be the process which changes just at integer values of t: ζt = ηi for i ≤ t < i + 1, where
η0 , η1 , . . . , ηn , . . . , is a sequence of independent identically distributed random variables. Put λ(z, x, α) =
ln E exp{b(z, x, η1 ) · α}, α ∈ Rd . Assume that λ(z, x, α) < ∞. The function λ(z, x, α) is convex in α ∈ Rd .
Denote by L(z, x, β) the Legendre transformation of λ(x, z, α) : L(z, x, β) = sup(α · β − λ(z, x, α)). Then the
action functional for Xtz, with such a process ζt is given by (4.9).
4. Construction of the cycles and actually the whole quasi-deterministic approximation for the long-time behavior
of the perturbed system is based on the approximation of the perturbed system by a Markov chain with exponentially
small transition probabilities as  ↓ 0. The states of this chain are, roughly speaking, the stable compacts K1 , . . . , Kl .
One can consider a Markov chain with a finite state space E and exponentially small transition probabilities
z, P
depending on a parameter z ∈ Rd : pij = exp{−Vijz /}, i 6= j ,i, j ∈ E, piiz, = 1 − j :j 6=i pij
z,
. Here Vijz > 0.
For any z ∈ R1 , one can define the cycles and their characteristics through the numbers Vijz as we did in Section
2. One can calculate µ(z, i, λ), i ∈ E, z ∈ R1 , for any λ > 0.
t/T , t/T
Consider now the time-non-homogeneous chain with one-step-transition probabilities pij = exp{−Vij /};
t = 0, 1, 2, 3, . . . . Then if lim↓0  ln T () = λ > 0, the trajectory of the chain at time [tT ()] will be situated at
the state µ(t, i, λ) with probability close to 1 as  ↓ 0. In particular, if Vijz are periodic in z, the long-time evolution
of the chain will be close to a periodic evolution.

5. Remarks and generalizations

1. Consider a system with hysteresis. The output of the system at time t is equal to 1 or −1 depending on which
threshold α or β, α < β, was hit last by the input Xt : put τ0 = min{s : Xs = α or Xs = β}; for t ≥ τ0 , define
τ1 (t) = max{s ≤ t : Xs = α or Xs = β}. Then the output i(t) is defined as follows:

 i(0) for t < τ0 ,

i(t) = γ [Xs , 0 ≤ s ≤ t] = 1 if t ≥ t0 , Xτ1 (t) = β,


−1 if t ≥ t0 , Xτ1 (t) = α.

Suppose the input Xt = Xt is defined by the equation

√
Ẋt = b(Xt ) +  Ẇt , X0 = x ∈ R1 . (5.1)

The pair (Xt , i  (t)), i  (t) = γ [Xs , 0R≤ s ≤ t], is a Markov process. Let b(0) = 0, b(x) > 0 for x < 0, b(x) < 0
x
for x > 0, α < 0 < β. Put B(x) = − 0 b(y) dy and assume that B(α) = B(β).
Let now the threshold α be perturbed by a small slowly changing function: the new threshold

α̃ = α + θ(t/T ),

where θ (t) is a Lipschitz continuous function, 1  T = T (), θ (t) < β − α. Consider an interval [0, A],
0 < A < ∞, and denote by 0 < t1 < · · · < tn < A the zeros of the function θ (t) on [0, A]. Assume that θ (0) > 0
and let θ (t) change the sign at each tk .
 M.I. Freidlin / Physica D 137 (2000) 333–352 351

Let lim↓0  ln T () = λ > 0. If λ < mint∈[0,A] B(α + θ (t)), then XtT
() , 0 ≤ t ≤ A, will stay inside (α̃, β) with
 
probability Px close to 1 as  ↓ 0. This means that i (t) = i (0) for t ∈ [0, AT ()]. If λ > mint∈[0,A] B(α + θ (t)),
but λ < B(β), then i  (tT ()) = −1 for t ∈ (inf t∈[0,A] {t > 0 : B(α+θ (t)) < λ}, A]. Finally, if λ > (mint∈[0,A] {t >
0 : B(α + θ (t))} ∨ B(β)), then for any δ > 0,

lim Px { max |i  (tT ()) − φ(t)| = 0} = 1,

↓0 t∈[0,A]\∪ni=0 Eδ (ti )

where Eδ (ti ) = {t : |t − ti | < δ}, t0 = 0,

(
1 if B(α + θ(t)) > B(β),
φ(t) =
−1 if B(α + θ(t)) < B(β).

In particular, if θ(t) is periodic, then φ(t) is periodic with the same period.
Of course, one can consider other types of random perturbations for systems with hysteresis. For instance, let the
input be defined by the equation

Ẋt = B(Xt , ζt/ ), X0 = x ∈ R1 , (5.2)

where ζt is the stationary Markov process with a finite state space considered in the previous section. Suppose
additionally that b(x, ζ ) = −b(−x, ζ ). Consider the averaged system

Ẋ t = b̄(X̄t ), X̄0 = x ∈ R1 ,
P
b̄(x) = Eb(x, ζt ) = i b(x, i)pi ; {pi } is the stationary distribution of the chain ζt . Assume that X t has just one
rest point at x = 0, and this point is asymptotically stable with b̄0 (0) 6= 0. Then, using the action functional for the
family of processes Xt defined by (5.2), we calculate the corresponding function V (x) = V (0, x). Let α = −β.
Assume that the threshold α is replaced by α̃ = α + θ (t/T ()). Then using the function V (x), one can describe
the asymptotic behavior of i  (t) = γ [Xs , 0 ≤ s ≤ t] on large time intervals as  ↓ 0.
2. We will use model (5.2) to mention one more effect. Let β = β() =  κ , α̃  (t) = − κ +  κ θ (t), where
θ (t) was introduced earlier, κ ∈ (0, 21 ). Then the exit from (α̃  (t/T ()), β()) is defined by moderate deviations
of order  κ . To write down the action functional for such deviations ([4], Section 7.7), put C = ∂ 2 λ(α)/∂α 2 |α=0 ,
where λ(α) is the eigenvalue of the matrix Qα = (qijα ), qijα = qij + δij αb(0, i) (qij were introduce in Section 4),
with the maximal real part. The action functional for the family  −κ Xt , where Xt is defined by (5.2), is equal to
(1/ 1−2κ )S0T (φ), φ ∈ C0T ,
( RT
1
c−1 (φ̇s − b̄0 (0)φs )2 ds, φ0 = 0, φs is absolutely continuous;
S0T (φ) = 2 0
+∞ for the rest of C0T .

Then V (x) = V (0, x) = inf{S0T (φ), φ0 = 0, φT = x, T ≥ 0} = (b̄0 (0)/c)x 2 . Let lim↓0  1−2κ ln T () = λ > 0.

Let θ (t) be 1-periodic. Then i  (tT ()) = γ [XsT () , 0 ≤ s ≤ t] will be close to the 1-periodic function
(
+1 if V (−1 + θ(t)) < V (1),
φ(t) =
−1 if V (−1 + θ(t)) > V (1),
if  > 0 is small enough.
3. Consider an autonomous system depending on a parameter T  1:
1
ẊtT = b1 (XtT , YtT ), X0T = x ∈ Rd ; ẎtT = b2 (XtT , YtT ), Y0T = y ∈ R1 . (5.3)
T
352 M.I. Freidlin / Physica D 137 (2000) 333–352

On a time interval [0, A], 0 < A < ∞, maxt∈[0,T ] |YtT − y| → 0 as T → ∞. Therefore, the evolution of XtT can
be described by the first equation with the “frozen” second component:
˙ y = b (X̃ y , y),
X̃
y
X̃0 = x. (5.4)
t 1 t

Assume that for any y ∈ R1 , system (5.4) has a finite number of equilibriums K1 (y), . . . , Kl 0 (y) ∈ Rd . Assume
that any trajectory of (5.4) is attracted to one of Ki (y). Let K1 (y), . . . , Kl (y) be the asymptotically stable attractors
and let l be independent of y ∈ R1 . Assume that the curves Ki (y), y ∈ R1 , have no intersections for i = 1, . . . , l.
Denote by y T ,i (t) the solution of the equation
1
Ẏ T ,i (t) = b2 (Ki (y T ,i (t)), y T ,i (t)). (5.5)
T
System (5.3) has no periodic solution under these assumptions, at least if T is large enough.
Consider now small random perturbations of (5.3):
√
ẊtT , = b1 (XtT , , YtT , ) +  Ẇt , X0T , = x ∈ Rd ;
1
ẎtT , = b2 (XtT , , YtT , ), YtT , = y ∈ R1 . (5.6)
T
The component XtT , , for 0 < T −1 ,   1, first of all approaches the equilibrium Ki(x) (y), where i(x) is the
number of equilibrium attracting the initial point x. Then, as t grows, XtT , follows the evolution of the attractor
Ki(x) (YtT , ) making time to time excursions to the outskirts of the basin of Ki(x) (YtT , ).
Let T = T (), lim  ln T () = λ > 0. Let V y (x1 , x2 ), x1 , x2 ∈ Rd , y ∈ R1 , be the V -function for the system
˙ y, = b (X̃ y, , y) + √ Ẇ ,
X̃ t 1 t t
R T y
V y (x1 , x2 ) = 21 inf{ 0 |φ̇s − b1 (φs , y)|2 ds, φ0 = x1 , φT = x2 , T > 0}. Put Vij = V (Ki (y), Kj (y)). One can
y
define the hierarchy of cycles, exit and rotation rates and the function µy (x, λ) using these Vij . It turns out that the
trajectory (XtT  
() , YtT () ), under certain additional assumptions, will be close to a periodic function.
Let, for example, l = 2. Assume that b2 (K1 (y), y) > 0, b2 (K2 (y), y) < 0, y ∈ R1 . Let y1 , y2 ∈ R1 , y1 < y2 ,
y y y y
exist such that V12 < λ for y > y2 and V12 > λ for y < y2 , V21 < λ for y < y1 and V21 > λ for y > y1 .
T (), T (), 
Let Zt = (Xt , Yt ) be the solution of (5.6) with the initial position (K1 (y1 ), y1 ). Then ZtT () will be
T (),1 T (),1
closeR to (K1 (Yt ), Yt ), where YtT ,1 is the solution of (5.5) with initial condition Y0T ,1 = y1 , until time
y2 
tˆ = y1 dy/b2 (K1 (y), y). At time tˆ, trajectory ZtT () as  ↓ 0 R“instantaneously” jumps to the point (K2 (y2 ), y2 )
y
and starts to move down along the curve K2 (y). It takes time y21 dy/b2 (K2 (y), y) to come to the neighborhood
of the point (K2 (y1 ), y1 ). From this point, the trajectory “jumps” as  ↓ 0 to a neighborhood of (K1 (y1 ), y1 ) and
the cycle is repeated. Thus, under the conditions mentioned above, the evolution of ZtT 
R y2 R y1 () is close as  ↓ 0 to the
periodic motion with period y1 dy/b2 (K1 (y), y) + y2 dy/b2 (K2 (y), y).

References

[1] R. Benzi, A. Sutera, A. Vulpiani, J. Phys. A 14 (1981) 1453.

[2] R. Benzi, A. Sutera, G. Parisi, A. Vulpiani, SIAM J. Appl. Math. 43 (1983) 563.
[3] M. Freidlin, Sov. Math. Dokl. 18 (1977) 1114.
[4] M. Freidlin, A. Wentzell, Random Perturbations of Dynamical Systems, 2nd ed., Springer, Berlin, 1998.
[5] L. Gammaitoni, P. Hänggi, P. Jung, F. Marchesoni, Rev. Mod. Phys. 70 (1998) 1.
[6] I. Karatzas, S. Shreve, Brownian Motion and Stochastic Calculus, Springer, Berlin, 1988.