39
THEORY
OF
STOCHASTIC
RESONANCE
4857
than N. The SNR then falls
off
gradually
as
D~~.
Later
we
will discuss
another
regime,
in which the
signal
frequency
is
relatively
high,
which is
quite
similar
in
many
respects.
There
is
a
slight
difference in the
cooperative
mechanism,
but
the value of D
at
which
the
SNR is
a maximum is
very nearly independent
of the
sig-
nal
frequency
in
both
regimes.
A natural
simplification of the
double-well
problem
is
the
discrete
two-state
system. (As a
model
for
quantum
systems,
it
is
of
interest in
its own
right.
)
The
two-state
system
with random
hopping
between
states but no
periodic
modulation
gives
rise
to
so-called
telegraph
noise. It is
a
straightforward calculation
to
find that the
power
spectrum
of such
a
system
is a
simple
Lorentzian.
The
analysis
assumes two
hopping
rates,
one in each
direction, but
how
in
a
real
physical
system
the external
noise
actually
determines
these
rates
is
left
unspecified.
Eckmann and Thomas studied
a
two-state
system
for
which the
hopping
rates had a
small
periodic
modulation
added,
one rate
180'
out of
phase
with the other. Their
conclusion was
that
sweeping
through
different
cu,
did
not
give
rise
to a
peak
in the
output
signal.
We confirm
here that there is
only
a monotonic decrease in the
signal
and SNR ratio with
increasing
cu,
. Because
they
made no
attempt
to
present
a relationship
between
input
noise D
and the
hopping
rates 8
+,
they
could
not
study
the more
central
issue
of
output
power
as a
function of
D.
In Sec.
VI
the
rates
W+(D)
are
derived
for the
Schmitt
trigger
circuit
which
is
very
nearly
a
two-state
system,
and
we
show how their behavior
yields
stochastic
resonance.
We conclude this section with a short list of the main
issues to be addressed.
Foremost is
to
find
expressions
for
signal
and noise
output
from
bistable
systems
which
agree
with simulations and
give
added
insight
into the
physical
source of this
phenomenon.
From these
expres-
sions we can also make some
helpful
predictions
regard-
ing
stochastic resonance. The
primary
prediction
is the
value of
input
noise
at
which the SNR will be enhanced.
In
addition,
one
can use
equations
for
S
and
N to
help
ad-
just
the other
system
parameters
to
make the effect more
pronounced
or to make it
occur within the constraints
of
a
physically
realizable
experiment.
In
particular,
the
effect of the
choice of
signal frequency
can
be examined.
Finally,
by
breaking
the
problem
into
a
general theory
and a
specific
calculation of the transition rate for
a par-
ticular
system
of
interest,
one can better see the
necessary
features
present
in all
systems
displaying
stochastic
reso-
nance
as
well
as the differences
between
individual sys-
tems.
will be Ualid
only
in the adiabatic
limit, when the
signal
frequency
is much
slower than some relaxation
time
~,
'.
For
the double-well
system,
for
example,
w,
is the time
for
probability
within
one well to
equilibrate.
In the
approach
used
below the
dynamical variable is
taken to be discrete: either x
or
x
+
with
probabilities
n+
=prob(x
=x+
).
If
the true
dynamics
are
continuous,
as with the double
well,
then
we
define
n
=
I
n—
+
=
f
p(x)dx,
(3.
l)
I'll
+
dt
8n
=
W
(t)n
—
W+
(t)n+
=
W
(t)
—
[
W
(t)+
W+
(t)
jn+,
(3.
2)
where
W+(t)
is the transition rate out
of
the
+
state.
Note that
W+(t)
is time
periodic
due to
the
periodic
sig-
nal.
The reduction from
a
continuous bistable
system
whose
probability
evolves
according
to
a
Fokker-Planck
equation
to a discrete
system governed
by
a rate
equation
has been carried out
formally,
'
and the extension
to a
periodically
modulated
system
is
straightforward.
'
For
the
purpose
of
computing
moments,
the
probability
den-
sity
is
effectively
p(x,
t)=n+(t)5(x
—
x+
)+n
(t)5(x
—
x
)
.
(3.
3)
Values for
x+
can be chosen
to minimize the error in
this
reduction. For
simplicity
it will
be assumed that the sys-
tern
is
symmetrical about x
=0
so
that
x+
=
—
x
=c
and the value for
c
is
chosen
to
minimize
the error in the
variance of x. The variance
for an
unmodulated
two-
state
system
in
its
steady
state
(n+
=n
=
—,
'
)
is
just
(x
)
=
f
x
p
(x)dx
=x+n+
+x
n
=c
.
(3
4)
The solution to the linear
first-order
differential
equation
with
periodic
coefficients
Eq.
(3.
2)
is
given
by
where x
'
is the location
of the
potential
maximum
separating
the two
wells. In
general,
the
modulation of
the
potential
will
cause
the
positions
of
the
potential
minima
and maximum
to
oscillate,
and
depending
on
the
level of
approximation
used,
it
may
be
necessary to
take
this
into account. The
governing
rate
equation, then,
is
just
III. GENERAL THEORY
n+(t)=g
'(t)
n+(t, )g(t,
)+
f
W
(t')g(t')dt'
(3.
&)
In
this
section
we derive
expressions
for
the
signal
and
noise
power
from
a
bistable
system,
in
a
way
which
is
in-
dependent
of
the
precise dynamics
of the
system.
To
make
these
results
useful for
a
particular
system,
one
must
have
an
expression
for
the
transition rate
or hop-
ping
rate
as a function
of the
input
noise
and other
pa-
rameters.
Such
rates are derived
from
the
details of the
system
dynamics;
in
this section
the
rate is assumed
known.
Typically,
this
expression
for
the transition rate
g
(t)
=exp
f
[
W+
(t')+
W
(t')]dt'
W+(t)
=
f
(p+rtocosto,
t),
(3.
6)
where
p
is
a
dimensionless
parameter
formed
from the
ra-
In
general,
the form of
W'+
will
be
such that
Eq.
(3.
5)
cannot
be
integrated
in
terms of known functions. We
assume,
though,
that the rate is
of the
form