Stochastic Resonators
Lutz Schimansky-Geier and Udo Siewert
Humboldt-Universitat zu Berlin, Invalidenstrae 110, 10115 Berlin, Germany
Abstract. We derive analytical expressions for the signal-to-noise ratio of a periodically driven chain of coupled spins placed into a heat bath. The analysis closely
follows the pioneering work of Glauber for the stochastic Ising model. We nd optimal bath temperatures and coupling constants where the SNR maximises. Special
attention is given to a wavy excitation of the chain where we discuss the dependence
of the SNR on the wave number.
1 Introduction
Stochastic resonance (SR) occurs when bistable or threshold systems are
driven by noise and a mono-frequent 'signal'. In this case the state variable of
the system under consideration typically exhibits a continuous spectrum with
sharp peaks at multiples of the signal frequency. The point is that the spectral
power amplication (SPA), which is the height of the rst peak scaled in units
of the driving amplitude, as well as the signal-to-noise ratio (SNR) at the
rst peak, may resonance-like pass (in general not simultaneously) through a
maximumas the noise grows. Stochastic resonance thus implies that, counterintuitively, growing noise leads to enhanced SPA and SNR within some noise
range. Possible technical exploitations of this eect are being investigated
and its relevance to neural activity in biological systems is of great interest.
(For review-type references see, e.g., Moss et al. 1994, Wiesenfeld and Moss
1995, Bulsara and Gammaitoni 1996, Jung 1993.)
When stochastic resonators are coupled SPA and SNR can be further
enhanced (Jung et al. 1992, Morillo et al. 1995, Bulsara et al. 1993 and
1995, Neiman and Schimansky-Geier 1995, Lindner et al. 1995 and 1996,
Marchesoni et al. 1996). However, it was always found that both quantities
fall o again, if the coupling is too strong. Therefore, apart from the noise
strength in a single stochastic resonator, the coupling strength may be used as
a second design parameter to achieve optimal operation in potential technical
applications.
To study this phenomenon analytically we will consider a chain of stochastically
ipping spins coupled via magnetic interaction. It is one of the best
investigated coupled systems and goes back to Glauber (1963). Glauber was
able to describe the chain by deriving expressions for the rst and second
moment of the spins and their correlation function, respectively. In his pio-
neering work one also already nds the analysis of spin chains modulated by
a homogeneous periodic external magnetic eld.
In this lecture we will reconsider Glauber's results (with some modications) from a stochastic-resonance perspective. During the writing of the
manuscript we got knowledge from a similar approach by J.J. Brey and A.
Prados (1996) who could proof an non-monotonic SPA for a stochastic Ising
model. Hence, there will be some similarities in our work to that of Brey
and Prados. However, while their point was the discovery of stochastic resonance in the Glauber model, our main concern is to model coupled stochastic
resonators.
In addition to Glauber and Brey et al. we make the assumption that the
ipping rates between states = 1 and = ?1 of individual uncoupled
spins depend on the bath temperature. It is supposed that
= 0 exp[?1=T ]
(1)
holds, which is of Arrhenius-type and typical for many rate processes (Hanggi
et al. 1989). For convenience we set 0 = 1. With the Ansatz (1) our analysis
will be closer to the numerical experiments by Lindner et al. (1995 and 1996)
who introduced the notion of 'array enhanced stochastic resonance' for the
coupling-induced increase of the SNR. We later derive analytical expressions
for this enhancement in our case of a spin chain.
For a single element the SNR was already obtained analytically within the
two state theory of stochastic resonance (McNamara and Wiesenfeld 1989).
Numerical experiments on a two dimensional Ising model were previously
performed by Neda (1995). Brey and Prados (1996) implicitly also found
optimal coupling for a given temperature T with a constant rate .
As an additional feature to previous studies of stochastic resonance and
to Glauber's approach, we consider the chain of spins to be driven by a
harmonic wave, rather than a spatially homogeneous force. It will result into
a coupling-induced enhancement of SPA and SNR as well as optimal coupling
at long wave lengths.
periodic part stands for the eect of a periodic signal with frequency ! and
initial phase . The signal amplitude A typically enters via = A=T into the
rate. The signal counter-wisely changes the rates for back and forward transitions of the spin. In continuous-state bistable systems rates of the present
type result for suciently small amplitudes A within an adiabatic approach,
where the signal is suciently slow compared to relevant time scales of the
system.
To extend this approach to coupled systems we consider a chain of innitely many spins. A state of the chain, which is given by a sequence
= (:::; k?1; k; k+1; :::) ( k = 1), is occupied with probability p( ).
The transition rates of a single element retain the McNamara-Wiesenfeld
shape. In addition the rates linearly depend on the state of the element to
the left and to the right.
Glauber's starting point were transition rates
h
i
(4)
Wi (i ) = 2 1 ? i
2 (i?1 + i+1 ) :
For positive
they give enhanced rates for a parallel ordering of the spins.
Glauber required that these rates lead to a stationary state characterised by
the equilibrium distribution of an Ising model
"
X
(5)
? T1 H = Z ?1 exp T1 Jii+1
i
with Z ?1 being the normalisation constant. Assuming detailed balance this
implies that
Wi (i)peq (:::; i?1; i; i+1; :::) = Wi (?i )peq (:::; i?1; ?i; i+1; :::) (6)
which is satised by choosing
= tanh[2J=T ]:
(7)
Let us now combine both situations, a spin interaction of the type given
in Eq.(4) and a periodic modulation as in Eq.(3). We take the latter to be
caused by a harmonic wave of wave number k to address the general form of
mono-frequent driving in extended systems. The transition rates then read
Wi (i ) = 2 [1 ?
2 i(i?1 + i+1 )
(8)
? i ?
2 (i?1 + i+1 ) cos(ki ? !t + )]:
Periodic modulation of the present type would result from an external magnetic eld
H (t; i) = A cos(ki ? !t + ):
(9)
acting on the chain. One can easily check that insertion of the Hamiltonian
H = ?
X
i
H (t; i)i ? J
X
i
i i+1
(10)
into the equilibrium distribution (5) leads in the adiabatic limit under detailed balance (6) to = tanh[A=T ] or A=T for small amplitudes A.
For simplicity we set = 1.
The set of probabilities p() evolves according to
p_() =
k
X
k
(11)
Apart from the harmonic-wave forcing (9) and the temperature dependent
rates (1) this is Glauber's dynamics of the Ising model in the weak-eld
limit (Glauber 1963). For our purposes we merely consider ferromagnetictype coupling (J > 0). To avoid negative transition rates,
+ (1 +
) < 1
has to be required. Special attention will be given to the in
uence on the
chain dynamics of the wave number k scaled in the box length of a single
spin.
hi (t)j (t + )i =
X
;
X
0
;
j i0 p2( ; t + ; 0 ; t)
(12)
(13)
X 0 0
j p( ; t + j; t)
(14)
for a given random initial distribution (t). Insertion of (14) into (12) gives
hi(t)j (t + )i =
X
(15)
The averaging can be performed if knowing the second moments ri;j (t) =
hi(t)j (t)i which implicitly enter via the initial state (t) into (15).
Both expressions, the conditioned average and the second moment, in turn
are obtained via equations derived from (11). For the conditioned average one
gets
1d
dt hi(t)j i = ?hi (t)j i + 2 [hi+1 (t)ji + hi?1(t)j i] +
1 ?
2 (ri?1;i + ri;i+1) cos(ki ? !t + ): (16)
For the second moment one derives in the zero-eld limit, which will be
sucient for our purposes,
1d
(17)
dt ri;j = ?2ri;j + 2 [ri;j ?1 + ri;j +1 + ri?1;j + ri+1;j ]
for i 6= j (ri;i = 1).
Let us brie
y report the unperturbed ( = 0) case. The solution of (16)
is ( 0)
1
X
hi (t + )j(t)i = e?
m (t)Ii?m (
):
(18)
m=?1
with In being the modied Bessel function. Insertion into (15) gives
hi(t)j (t + )i = e?
1
X
m=?1
ri;m(t)Ii?m ( ):
(19)
hi(t)j (t + )i = e?j j
1
X
m=?1
(23)
be sucient to take their asymptotic expressions from the zero-eld limit.
The long-time limit solutions hi (t)ias of the conditioned averages can be
written as
hi(t)ias = q cos(ki ? !t + + )
(24)
They do not depend on the initial spin distribution, but, obviously, on the
phase of the driving eld. The amplitude q of the asymptotic oscillations
reads
p
2
(25)
q = p 2 1 ?
! + ((1 ?
cos(k)))2
and the phase shift is given by
(26)
tan = (1 ?
!cos(k)) :
The correlation function can be divided into a noise part and a signal
part. Insertion of the asymptotic expression into Eq(15) gives
hi (t)j (t + )ias ? hi(t)ias hj (t + )ias
(27)
= e?j j
+1
X
m=?1
Since we know the second moments for the unperturbed spin-chain, only, we
approximate the noise part on the r.h.s. of this equation by the correlation
functions of the spin chain without external magnetic eld.
To obtain the power spectrum we remove the t-dependence of hi (t)j (t +
)ias by averaging over an uniformly distributed initial phase of the driving
wave (McNamara and Wiesenfeld 1989, Jung and Hanggi 1989) and apply
the Wiener-Khinchine theorem. For the one-sided cross spectrum we nd
Si;j (
) = 4
with
+
1
X
n=?1
+1
X
2 + jnj
Sn = p 2 1 2
+
=0
p
2 2 +
2
2 +jnj
(28)
(29)
cos (2 + jnj + 1) arctan ;
and q given by (25). For a particular element of the chain one gets a power
spectrum
Si;i (
) = S (
) = 4S0 + 8
+1
X
n Sn + q2 (
? !):
n=1
(30)
4 Stochastic Resonance
4.1 Global Response - The Signal-to-Noise Ratio
One of the interesting features of the spin chain is its global response to a
homogeneous driving (k = 0). Taking the sum over the states of N individual spins out of our spin chain of innite length one can dene the global
magnetisation
M (t) =
which asymptotically simply reads
N
X
i
i (t)
(31)
(32)
The stationary signal part arises from integration over the initial phase
2
(33)
The spectrum without external eld is obtained from the double sum
hM (t)M (t + )i =
i;j
i?j
hi (t)j (t + )i N
hi(t)j (t + )i:
(34)
To evaluate this sum one takes the summation in j over the whole chain
which is correct for large N , only. It gives the Lorentz-spectrum (Glauber
1963)
p
2
2
Sglobal (
) = N 2(1 ?
1)?2 +
2 :
(35)
In result the signal to noise ratio R per spin monotonously decreases with
increasing coupling
p1 ?
2 :
q2
1 R= 1
(36)
=
2
2
NA
N 4A Sglobal (!) 4T 2
2
J
opt
(37)
opt = tanh T = 1 + 22 ? (1 + 2
2 )2 ? 1:
= const:: The case of Brey and Prados. They considered the interesting case where is a constant and does not depend on the temperature.
Much insight can be gained from this situation. For simplicity one sets = 1.
Taking the results of section 3 one can distinguish two interesting limits (Brey
and Prados 1996):
{ Low frequency limit: For suciently small ! the SPA peaks at low T and
small T=J can be considered. The coupling parameter
can be replaced
by an exponential. Approximating 1 ?
exp ?4J=T the SPA reads
?4J=T
2
P = Aq 2 = T22 4e?e8J=T + !2 :
(38)
It has the usual bell shape. The coupling J can be seen as the origin
of some activation energy U = 2J , i.e., coupling between neighbouring
elements provides the activation barrier necessary for SR. If the spins are
parallel a barrier of height 2J has to be crossed to
ip a spin. Nevertheless,
one has to point out that the usual expression for the SPA in case of such
a barrier involves a term e?4J=T in the denominator. Our denominator
is smaller which, with growing coupling strength, leads to a more rapid
increase in the SPA, particularly at small frequencies.
= (T ): Arrhenius-Type Activation. One nds quite dierent behaviour if the rates depend on the temperature. To obtain the position
(Tabs; Jabs) of the global maximum of the SPA we determine the derivatives
of the SPA with respect to J and T and equate them to zero. Solving one of
these equations for !2 and substituting it in the other gives
Jabs = (Tabs=4)log[1=(2Tabs ? 1)]:
(42)
Substituting Jabs in the obtained expression for !2 one gets an implicit equation for Tabs at given ! reading
)2 :
(43)
!2 = exp(?2=Tabs ) T(1 ?(12?Tabs
Tabs )
abs
Finally, inserting (42) and (43) into the SPA we nd for its maximum value
at given ! ( which determines Tabs)
Pabs = T 2 1(2?TTabs? 1) :
(44)
abs abs
10
U
Q
i = +1
1i
2i
3i
i+1 = 1
1i+1 2 i+1 3 i+1
Fig. 1. The SPA of the spin chain vs temperature T for several coupling parameters
J : left to right:0, 0.25, 0.5, 1.0, 2.5 (! = 0:01, k = 0).
The following statements can be made now. For Jabs (42) to be real-valued,
Tabs 1=2 has to hold. Together with (43) one then nds that tuning ! from
zero to innity steadily shifts Tabs from 1=2 to 1. At the same time, Jabs and
Pabs decrease from innity to zero, since both are monotonous functions of
Tabs within its specied range. Hence, contrary to the former case = 1, the
global maximum Pabs is located at nite J and T for any nite modulation
frequency !. For instance, at ! = 0:01 the SPA has a maximum for Tabs
0:51 and Jabs 0:44. Fig. 1 shows this situation. Since Jabs is never zero,
the maximum SPA at given ! always exceeds that of the uncoupled case.
Furthermore, one nds that Tabs exceeds Topt (J = 0) of the uncoupled chain,
if the respective conditions
2(Topt (J = 0))=!2 = 1=Topt (J = 0) ? 1;
2(Tabs )=!2(1 ?
abs )2 = 1=Tabs ? 1
(45)
are solved geometrically.
11
1.5
0.5
SNR
0.25
1.0
1.0
1.5
0.05
2.0
0.5
0
0.0
0.0
1.0
2.0
3.0
4.0
5.0
Fig. 2. The local SNR (scaled in A2 ) of a single spin in a spin chain vs temperature.
The curves show its behaviour for various couplings J (values attached), when the
chain is driven homogeneously (k = 0) with frequency ! = 0:01.
the
P global response, the noise part is now taken locally. It implies that from
i;j hi j i the sum over j is dropped. Obviously, the value of the noise part
is lower than in the former case (34);
X
X
(46)
hi(t)i (t + )i < hi (t)j (t + )i:
i
R=
i;j
q2
P 1 n :
4S0 + 8 +n=1
Sn
(47)
12
The SNR at medium frequencies will show HF-behaviour at low temperatures and LF-behaviour at high temperatures, since the temperature enters
into the frequency limits given above. At medium temperatures a transition
between both limits takes place. At medium frequencies the SNR shows optimal coupling as is shown in Fig. 2, where the noise part of the SNR was
evaluated numerically. It proofs the existence of array enhanced SR (Lindner
et al. 1995 and 1996) for the considered spin chain.
Now we turn to the interesting case where the spin chain is driven by a
wavy excitation. The wave number k, scaled with respect to the box length
of a single spin, enters into the SPA via cos(k). Hence it is sucient to
consider wave numbers 0 k , which re
ects the isotropy of the chain
(right-travelling waves cause the same eects as left-travelling ones) and its
lattice character (Brillouin zones). We basicly nd that the SPA decreases
monotonously with growing k, whereby this eect is the stronger the smaller
the wave frequency. For suciently short wave lengths the SPA even falls
below its value of the uncoupled case, i.e., here coupling decreases the SPA.
A typical situation is shown in Fig.3.
This behaviour can easily be traced back. The numerator of the SPA
always leads to its decrease under coupling, which is a direct consequence of
the coupling-dependent amplitude of the modulation of the transition rates
(8). The denominator of the SPA, however, whose coupling dependence arises
from the sum [hi+1(t)j i+hi?1 (t)j i] in (16), behaves qualitatively dierent
for long wave lengths 0 k < =2 and short wave lengths =2 < k
, respectively. For long waves it increases the SPA whereas at short wave
length it causes the SPA to fall. Both eects are the stronger the longer or,
respectively, shorter the wave length.
To understand this eect one has to realize that, coupled or not, the
spins asymptotically follow the wave imposed on the chain. Hence the sum
[hi+1(t)ias + hi?1(t)ias ] in the asymptotic version of (16) is given by the
term 2cos(k)hi (t)ias , which under ferromagnetic-type coupling (
> 0)
leads to the behaviour just described. A special situation arises at k = =2.
Here the neighbours of a given spin i are always driven anti-phase and the sum
[hi+1(t)ias + hi?1(t)ias ] vanishes. Subsequently, in this case the coupling
has no eect, except for the decrease of the SPAs numerator.
In the end the in
uence of the wave number k is almost trivial; if an
uncoupled chain is modulated by a wave of such a short length that adjacent
spins tend to group in a nearly anti-parallel order, one cannot enhance the
chains response with ferromagnetic-type coupling.
For the SNR we merely consider the local version. At xed J it inherits
the qualitative behaviour of the SPA (Fig.3), since its noise part is unaected
by k. On the other hand, at xed k, looking at (48) one nds optimal coupling
for suciently long wave lengths even at low frequencies. The relative growth
13
0.001
N = 2
0.003
BarrierPot.
k = 0.2
0.005
J0
correlated
0.007
0.009
uncorrelated
0.011
anticorrelated
0.013
0.0
5.0
10.0
15.0
Fig. 3. The SPA vs temperature for dierent values of the wave number k; top down
in units of : 0, 0.25, 0.5, 0.75, 1.0; (J = 0:5, ! = 0:01). Dashed curve: J = 0.
of the signal part can now be out-paced by that of the noise part, which was
not the case at k = 0. Hence, the qualitative behaviour of the SNR in this
case is the same as shown in Fig. 2, even at small frequencies.
5 Conclusion
We investigated a periodically modulated chain of stochastically
ipping spins
closely following the work of Glauber (1963). We interpreted his results within
the theory of stochastic resonance and found an optimal coupling where the
SPA and the local SNR exhibit maximal values. Special attention was paid to
a spatially inhomogeneous driving of the chain which is a general situation.
Here we found that the response of the chain is maximum at vanishing wave
vector k = 0. We also found that for wave lengths larger than approximately
four spin-boxes (k < =2) the coupling-induced enhancement of SPA and
SNR still works. Hence, a chain of stochastic resonators is still suitable for
the detection of small wavy signals.
We found analytically that coupling increases the SPA and the SNR of
a spin chain. To gain a global maximum for nite T and J the introduction
of Arrhenius-like rates (1) was necessary which is a novel feature compared
to Brey et al. (1996). Nevertheless, in the considered case of a spin chain
14
of innite length, the increase of the SNR was found in the local response,
only. Further investigation of nite chains should provide insight, whether
or not a small number of coupled spins can enhance the SNR (Neiman and
Schimansky-Geier 1995).
Our ndings are in good qualitative agreement with the results of the
simulation by Lindner et al. (1995 and 1996) of a chain of continuous bistable
elements, although these simulations were carried out with rather strong
forcing, whereas we are bound to weak forcing. Furthermore, some of the
results of the simulation have been conrmed by a model based on kinkantikink nucleation (Marchesoni et al. 1996). This model essentially involves
coupling-induced transitions, which do not occur in our chain. It is therefore
interesting that we, nevertheless, nd qualitatively similar eects; optimal
coupling and SNR-peaks moving to higher noise strengths when the coupling
increases. It suggests that the qualitative behaviour of our simple model is
likely to carry over to more complex situations.
We thank P. Hanggi, A. Neiman, F. Marchesoni, P. Jung and A. Bulsara
for several fruitful discussions on this topic.
References
Brey J.J. and Prados A. (1996): Phys. Lett. A216, 240.
Bulsara A. and Schmera G. (1993): Phys. Rev. E47, 3734; Bulsara A., Maren A.,
and Schmera G. (1993): Biol. Cybern. 70, 145; Inchiosa M.E. and Bulsara A.R.
(1995): Phys. Lett.A200, 283 and Phys. Rev. E52, 327.
Bulsara A. and Gammaitoni L. (1996): Physics Today, March 1996, 39.
Glauber R.J. (1963): J. Math. Phys. 4, 294.
Hanggi P., Talkner P., and Borkovec M. (1990): Rev. Mod. Phys. 62, 251.
Hanggi P. and Thomas, H. (1982): Phys. Rep. 88, 207.
Jung P. and Hanggi P. (1989): Europhys. Lett. 8, 505.
Jung P., Behn U., Pantazelou E., and Moss F. (1992): Phys. Rev. A46, R1709.
Jung P. (1993): Phys. Rep. 234, 175.
Lindner J., Meadows B., Ditto W., Inchiosa M., and Bulsara A. (1995): Phys. Rev.Lett. 75, 3; (1996): Phys. Rev. E53, 2081.
Marchesoni F., Gammaitoni L., and Bulsara A. (1996): Phys. Rev. Lett. 76, 2609.
Marchesoni F. (1996): contribution to this issue.
McNamara B. and Wiesenfeld K. (1989): Phys. Rev. A39, 4854.
Morillo M., Gomez-Ordo~nez J., and Casado J.M. (1995): Phys. Rev. E52, 316.
Moss F., Pierson D., and O'Gorman D. (1994): Int. J. Bif. Chaos 4(6), 1383.
Neda Z. (1995): Phys. Rev. E51, 5315.
Neiman A. and Schimansky-Geier L. (1995): Phys. Lett. A197, 379.
Wiesenfeld K. and Moss F. (1995): Nature 373, 5 January, 33.