1.

Introduction
When one tries to measure a weak signal, a lower limit set by the spontaneous uctuation in
the current, voltage, and other physical variables of the system under test is always encountered.
These spontaneous uctuations are referred to as noise. Noise is a signicant problem in science
and engineering because it places an ultimate sensitivity on all measurements. In this class, the
mathematical methods, physical origins and characteristics of noise in various devices and systems
will be studied.
Chapter 2 provides a mathematical method for treating noise. Two complimentary averaging
methods, time averaging and ensemble averaging, and the experimental realization of the two
averaging processes are described. The Fourier analysis, specically the Wiener-Khintchin theorem,
is introduced to calculate the power spectrum and autocorrelation function of a noisy waveform.
A noisy waveform often consists of discrete random pulse trains. Various statistical distribution
functions of random pulse trains (i.e., binomial, Poisson, sub-Poisson, super-Poisson, and Gaussian)
are derived and the eect of random deletion of pulses (partition noise) is discussed using the
Burgess variance theorem. A method for calculating the power spectrum and autocorrelation
function of random pulse trains using the Carson theorem is presented.
Chapter 3 describes how to calculate the noise gure of linear networks. A noisy electrical
network can be represented by a noise-free network with external noise generators. If such a network
has two ports (input and output), the noise gure is often used as a gure of merit expressing the
inherent noisiness of the circuit. The technique is also useful to express the noise properties of
optical systems such as laser and parametric amplier systems.
Chapter 4 discusses the physical origins of two types of intrinsic noise: thermal noise and
1
quantum noise. Thermal noise of a resistor is calculated by microscopic theory using a Brownian
particle model or by macroscopic theory using the equipartition theorem. The latter approach,
developed by Nyquist, is very general and easily extended to include quantum noise (i.e., the
eect of zero-point uctuation). The expression for the power spectrum of an open circuit voltage
uctuation, including both thermal and quantum noise, is called a generalized Nyquist formula.
The noise characteristics of a mesoscopic conductor on an extremely small scale is enlightening
with regard to the physical origins of thermal and quantum noise. The transition from mesoscopic
conductor to macroscopic conductor will be discussed in some detail.
Chapter 5 treats the noise of semiconductor pn junctions diodes. Random diusive motion
and generation-recombination event of carriers contribute to the noise of a pn junction. A pn-
junction diode driven by a low-impedance, constant-voltage source has a shot-noise-limited current
uctuation, but a junction voltage uctuation is suppressed. A pn-junction diode driven by a high-
impedance, constant-current source has a shot-noise-limited voltage uctuation, but a junction
current uctuation is suppressed. A collective and single charge Coulomb blockade eect in a pn
junction is discussed as the principle of sub-Poisson light generation.
Chapter 6 discusses the noise of a bipolar transistor using the method of a noisy linear network.
Chapter 7 describes the mathematical models and physical mechanisms of 1/f noise. In prac-
tically all electronic and optical devices, the excess noise obeying the inverse frequency power law
exists in addition to intrinsic thermal and quantum noise. This 1/f noise places a serious limit on
the sensitivity of precision measurements at a very low frequency.
Chapter 8 treats the noise of a tunnel diode. While a macroscopic tunnel diode driven by
a constant voltage source has a shot-noise-limited current uctuation, a mesoscopic tunnel diode
2
features various sub-shot-noise behaviors due to single charge Coulomb blockade eect.
Chapter 9 describes the noise of negative conductance oscillators which generate a coherent
electromagnetic eld at various frequencies. A laser is also described as a negative conductance
oscillator. A negative conductance oscillator consists of a frequency selective element which provides
positive feedback and a gain element which provides signal amplication and nonlinearity. A general
electrical circuit theory (van der Pol oscillator model) is presented to calculate the amplitude and
phase noise of the coherent radiation. The dierence between the cavity internal eld noise and
output eld noise, and the dierence between a free-running oscillator and an injection-locked
oscillator, are pointed out.
Chapter 10 provides a brief overview of the noise of parametric ampliers. The fundamental
limit of phase-insensitive and phase-sensitive linear ampliers are discussed. A degenerate para-
metric amplier is introduced as a noise-free, phase-sensitive amplier.
Chapter 11 provides a brief overview of the noise in optical communication systems. The
signal-to-noise ratio and bit error rate of optical-preamplier and repeater-amplier systems are
calculated and compared with a direct detection scheme. The concept of signal regeneration for
a digital system and the advantage of using optical ampliers are discussed. The thermal and
quantum limits of communication systems are elucidated using the channel capacity.
3
2. Mathematical Methods
2.1 Time Averaging vs. Ensemble Averaging
Noise is a stochastic process of a randomly varying function of time and thus is only statis-
tically characterized. One cannot argue a single event at a certain time; one can only discuss the
averaged quantity over a certain time interval (time average) or many identical systems (ensemble
average). Let us consider N identical systems which produce noisy waveforms x
(i)
(t), as shown in
Fig. 2-1.
x
(1)
(t)
x
(2)
(t)
x
(N)
(t)
time average
ensemble average
t
t
t
Figure 2-1
One can dene the following time-averaged quantities for the i-th member of the ensemble:
x
(i)
(t)) = lim
T
1
T
_ T
2

T
2
x
(i)
(t)dt ,
(mean = rst-order time average) (2.1)
1

_
x
(i)
(t)
_
2
) = lim
T
1
T
_ T
2

T
2
_
x
(i)
(t)
_
2
dt ,
(mean square = second-order time average) (2.2)

(i)
x
() x
(i)
(t)x
(i)
(t + )) = lim
T
1
T
_ T
2

T
2
x
(i)
(t)x
(i)
(t + )dt .
(autocorrelation function) (2.3)
These time-averaged quantities directly correspond to analog measurements using a simple inte-
grator (low-pass lter) for the measurement of the mean (2.1), square-law detector followed by an
integrator for the measurement of the mean square (2.2) and delayed correlator followed by an
integrator for the measurement of the autocorrelation function. Alternatively, one can measure the
noisy waveform x
(i)
(t) using an analog-to-digital (A/D) coverter and sampling circuit and then
calculate these quantities using a computer (digital measurements).
One can also dene the following ensemble-averaged quantities for all members of the ensemble
at a certain time:
x(t
1
) = lim
N
1
N
N

i=1
x
(i)
(t
1
) =
_

x
1
p
1
(x
1
, t
1
)dx
1
,
(mean = rst-order ensemble average) (2.4)
x(t
1
)
2
= lim
N
1
N
N

i=1
_
x
(i)
(t
1
)
_
2
=
_

x
2
1
p
1
(x
1
, t
1
)dx
1
,
(mean square = second-order ensemble average) (2.5)
x(t
1
)x(t
2
) = lim
N
1
N
N

i=1
x
(i)
(t
1
)x
(i)
(t
2
) (2.6)
=
_

x
1
x
2
p
2
(x
1
, t
1
; x
2
, t
2
)dx
1
dx
2
.
(covariance = second-order joint moment)
2
Here, x
1
= x(t
1
), x
2
= x(t
2
), p
1
(x
1
, t
1
) is the rst-order probability density function, and
p
2
(x
1
, t
1
; x
2
, t
2
) is the second-order joint probability density function. p
1
(x
1
, t
1
)dx
1
is the prob-
ability that x is found in the range between x
1
and x
1
+dx
1
at a time t
1
and p
2
(x
1
, t
1
; x
2
, t
2
)dx
1
dx
2
is the probability that x is found in the range between x
1
and x
1
+dx
1
at a time t
1
and also in the
range between x
2
and x
2
+ dx
2
at a dierent time t
2
.
If x(t
1
) and x(t
1
)
2
are independent of the time t
1
and if x(t
1
)x(t
2
) is dependent only on the
time dierence = t
2
t
1
, such a noise process is called a statistically-stationary process. For a
statistically-nonstationary process, the above is not true. In such a case, the concept of ensemble
averaging is still valid, but the concept of time averaging fails. An ensemble average is a convenient
theoretical concept since it is directly related to the probability density functions. On the other
hand, time averaging is more directly related to real experiments. One cannot prepare an innite
number of identical systems in a real situation. Theoretical predictions based on ensemble averaging
are equivalent to the experimental measurement of time averaging when, and only when, the system
is a so-called ergodic ensemble. One can loosely say that ensemble averaging and time averaging
are identical for a statistically-stationary system, but are dierent for a statistically-nonstationary
system.
2.2 Fourier Analysis
When x(t) is absolutely integrable, i.e.,
_

[x(t)[dt < , (2.7)

3
the Fourier transform of x(t) exists and is dened by
X(j) =
_

x(t)e
jt
dt . (2.8)
The inverse transform is given by
x(t) =
1
2
_

X(j)e
jt
d . (2.9)
This is proven by taking the integral in (2.9) over the limits N (instead of ), substituting for
X(j) from (2.8), and interchanging the order of integration to obtain
1
2
_

X(j)e
jt
d =
1

lim
N
_

x(t

)
sinN(t

t)
(t

t)
dt

=
_

x(t

)(t

t)dt

= x(t) . (2.10)
When x(t) is a real function of time, as it always is the case for an observable waveform, the
real part of X(j) is an even function of and the imaginary part is an odd function of [i.e.,
X(j) = X

(j)].
When x(t) is a statistically-stationary process, condition (2.7) is not satised and thus the
Fourier transform does not exist. The total energy of the noisy waveform x(t) is innite, but
the noise power (energy ow per second) can be nite. In any practical noise measurement, a
measurement time interval T is nite and the energy of such a gated function x
T
(t), dened by
x
T
(t) =
_

_
x(t) [t[
T
2
0 [t[ >
T
2
, (2.11)
is also nite. The Fourier transform of such a gated function does exist.
4
If x
1
(t) and x
2
(t) have Fourier transforms X
1
(j) and X
2
(j), one obtains
_

x
1
(t)x

2
(t)dt =
1
2
_

X
1
(j)X

2
(j)d .
(Parseval theorem) (2.12)
This relation is known as the Parseval theorem and is proven by substituting the inverse transform
of x

2
(t) into the LHS of (2.12), interchanging the order of integration, and identifying the resulting
integral about t as X
1
(j). If one uses x
1
(t) = x
T
(t + ) and x
2
(t) = x
T
(t) in (2.12), one obtains
_

x
T
(t + )x
T
(t)dt =
1
2
_

[X
T
(j)[
2
e
i
d . (2.13)
When = 0, (2.13) is reduced to
_

[x
T
(t)]
2
dt =
1
2
_

[X
T
(j)[
2
d .
(Energy theorem) (2.14)
The physical interpretation of X
T
(j) and [X
T
(j)[
2
is now clear from the above relations. X
T
(j)
is the (complex) amplitude of the e
jt
component in a gated function x
T
(t) and [X
T
(j)[
2
is the
energy density of x
T
(t) with units of energy per Hz. Equation (2.14) is the total energy of x
T
(t)
and increases linearly with T for a statistically-stationary process.
The average power of x
T
(t), dened by
lim
T
1
T
_

[x
T
(t)]
2
dt = lim
T
1
2
_

0
2[X
T
(j)[
2
T
d , (2.15)
is a constant and universal quantity. If ensemble averaging is taken for a gated function x
T
(t) in
(2.15), the order of lim
T
and
_

0
d can be interchanged. The power spectral density is dened as
S
x
() = lim
T
2[X
T
(j)[
2
T
.
5
(unilateral power spectral density) (2.16)
When a statistically-stationary noisy waveform x(t) is input into an RF spectrum analyzer, the
above (unilateral) power spectral density is displayed on the screen, usually with units of power
dB
m
= 10 log
10
P(mW), as shown in Fig. 2-2. The noisy waveform x(t) beats against a strong
local oscillator wave at a frequency in a mixer and only the low-frequency part of the beat
signal is detected by using a low-pass lter with a bandwidth f, where f is termed a resolution
bandwidth. The detected noise power is thus due to the frequency component at of the noisy
waveform x(t). An RF spectrum analyzer measures the noise powers of a xed frequency for N
wavepackets with a duration of T = 1/f and displays the averaged noise power. This is essentialy
the ensemble average procedure. The averaging time TN determines a video bandwidth f/N. If
one divides this averaged noise power by a resolution bandwidth f, one can obtain the (ensemble
averaged) power spectral density S
x
().
Local Oscillator wave A sin t
Noisy waveform x(t)
f = : resolution bandwidth
1
T
= : video bandwidth
1
TN
f =
1
T
1
N
p(i)
i
T T T T T T
N
t
mixer
~
low-pass filter signal averager
t
noise
power
frequency
envelope detector
Figure 2-2
6
When ,= 0 in (2.13), one can also divide both sides of (2.13) by T, take an ensemble average,
and take a limit as T to obtain
lim
T
1
T
_

x
T
(t + )x
T
(t)dt = lim
T
1
2
_

0
2[X(j)[
2
T
cos d . (2.17)
The LHS of this expression is the (ensemble averaged) autocorrelation function
x
(). Using (2.16)
in the RHS of this expression, one obtains

x
() =
1
2
_

0
S
x
() cos d . (2.18)
The inverse relation of this expression is
S
x
() = 4
_

0

x
() cos d .
(Wiener-Khintchine theorem) (2.19)
Equation (2.19) is proven by using (2.18) for
x
() in the RHS of (2.19), interchanging the order
of the integral, and identifying the resulting integral about as a Dirac delta function (

).
Equations (2.18) and (2.19) constitute the Wiener-Khintchine theorem and indicate that 2
x
()
and S
x
() are the Fourier transform pairs.
If a noisy waveform x(t) is a statistically-stationary process, as shown in Fig. 2-3, with an
exponentially decaying autocorrelation function

x
() =
x
(0) exp
_

[[

1
_
, (2.20)
where
x
(0) = x
2
and
1
is a relaxation time constant (systems memory time). Substituting (2.20)
into (2.19), one obtains the unilateral power spectral density
S
x
() =
4
x
(0)
1
1 +
2

2
1
. (2.21)
7
statistically-stationary
noisy waveform
statistically-nonstationary
noisy waveform
x(t)
t = 0
y(t) = x(t')dt'
0
t
t = 0
y(t)
2
= D
y
t
t
t
Figure 2-3
The spectrum is Lorentzian with a cut-o frequency of 1/
1
and the low-frequency spectral density
is S
x
( = 0) = 4
x
(0)
1
. The autocorrelation function and the unilateral power spectrum are
shown in Fig. 2-4.
x( ) S
x( )
4 0
1

x() x( ) 0
1
0.5
0 1 -1 /1
10-2 10-1 10 1
10-2
10-1
1
1
Autocorrelation Function Unilateral Power Spectrum
Figure 2-4
A time-integrated function y(t) =
_
t
0
x(t

) dt

of a statistically-stationary process x(t

) goes
through a random walk diusion, as shown in Fig. 2-3. This is called a Wiener-Levy process and
8
is an example of a statistically-nonstationary process. A gated function is dened by
y(t) =
_

_
_
t
0
x(t

) dt

(0 t T)
0 (otherwise)
. (2.22)
The corresponding autocorrelation function becomes a function of and T:

y
(, T) =
1
T
_
T||
0
y(t + )y(t) dt
=
T
2
_
1
[[
T
_
2
D
y
. (2.23)
In order to derive the second line of (2.23), y(t) was used as a cumulative process and thus
y(t + )y(t) = [y(t) + y(t + )]y(t) = y(t)
2
= D
y
t, where D
y
is a diusion constant. The unilat-
eral power spectral density is given by
S
y
(, T) = 4
_
T
0

y
(, T) cos() d
=
4D
y

2
_
1
sin(T)
T
_
. (2.24)
The correlation time is now proportional to the measurement time interval T and a nite mea-
surement time T prevents the divergence of the power spectral density at = 0, as shown in
Fig. 2-5.
2.3 Random Pulse TrainsBinomial, Poisson, and Gaussian Distributions
A noisy waveform x(t) often consists of a very large number of random, discrete events and
are represented by
x(t) =
K

k=1
a
k
f(t t
k
) . (2.25)
9
0 -1 1
/
y(
, T)
/
1
2
D
y
T
1
0.5
S
x(
, T)
/
24D
y
T
2
10-2
10-1
1
10-2 10-1 1 10

Autocorrelation Function Unilateral Power Spectrum
Figure 2-5
One assumes the pulse amplitude a
k
and the time of event t
k
are random variables, but that the
pulse-shape function f(t) is a xed function, as shown in Fig. 2-6. In a real physical situation, f(t)
is determined, for example, by the relaxation time of a system or the transit time of a carrier, and
thus it is a reasonable assumption that f(t) is a xed function.
a1
a2
ak
a4
a3
t1 t2 tk t4 t3
t

measurement time interval T = N

Figure 2-6
Suppose the average rate of such random pulse events per second is and the measurement
time interval T is divided into N =
T

time slots with a duration . The probability p of pulse

emission in each time slot is , which, by making N large enough, can be much smaller than one.
In such a case, the probability of more than two pulses being emitted in the same slot is negligible.
If a time slot duration is much longer than a systems correlation time (memory time), each pulse
10
emission occurs independently (that is, the pulse emission at a certain time is not inuenced
by the previous events of the system). In such a case, the probability of nding K events in the
measurement time interval T (out of N trials) is given by the binomial distribution:
W
N
(K) =
N!
K!(N K)!
p
K
q
NK
, (2.26)
where q = 1 p is the probability of zero pulse emission. Since this is the expansion coecient of
(p + q)
N
, the binomial distribution W
N
(K) satises the normalization condition,
N

K=0
W
N
(K) = (p + q)
N
= 1 . (2.27)
The mean of the number of pulses is calculated by
K
N

K=0
KW
N
(K) =
N

K=0
N!
K!(N K)!
_
p

p
(p
K
)
_
q
NK
=
_
p

p
_
(p + q)
N
= pN . (2.28)
The mean-square is
K
2

K=0
K
2
W
N
(K) =
N

K=0
N!
K!(N K)!
_
_
p

p
_
2
(p
K
)
_
q
NK
=
_
p

p
_
2
(p + q)
N
= (pN)
2
+ pqN . (2.29)
The variance of the number of pulses is then given by
K
2
K
2
K
2
= Np(1 p) . (2.30)
11
We start with a xed (constant) number of potential events N = T/ and introduce random
deletion with the probability q = 1 p. The variance of the output event (2.30) is a rather general
result for such a stochastic process and is referred to as partition noise. In some cases, the
number of potential events N itself uctuates. In such a case, the variance of the number of output
events is given by
K
2
= N
2
p
2
+ Np(1 p) ,
(Burgess variance theorem)
(2.31)
where N
2
= N
2
N
2
. The above relation is known as the Burgess variance theorem and is
proven by replacing W
N
(K) in (2.28) and (2.29) by W(N)W
N
(K), where W(N) is the probability
distribution function for potential events N. The rst term on the RHS of (2.31) indicates that
the uctuation of the initial number of events is suppressed by a loss probability p. The second
term, on the other hand, indicates the new noise term (partition noise).
In the limit of p 1 and N 1, one obtains the following approximate relations:
(1 p)
NK
e
p(NK)
e
pN
, (2.32)
ln
N!
(N K)!

_
d
dN
lnN!
_
K lnN
K
. (2.33)
Using these relations in the binomial distribution (2.26), the approximate expression for W
N
(K)
is obtained as
W
(p)
N
(K) =
(pN)
K
K!
e
pN
=
K
K
K!
e
K
.
(Poisson distribution) (2.34)
This is the Poisson distribution. The variance of the output events K
2
can be calculated as
12
follows:
K
2
K

K=0
K(K 1)W
(p)
N
(K) = K
2

K=2
K
K2
(K 2)!
e
K
= K
2
K
2
K
2
K
2
= K = Np . (2.35)
The result is, of course, obtained directly from (2.30) by letting p 0, but keeping Np nite. The
same variance is also obtained from (2.31) in the limit p 0. Therefore, the Poisson distribution is
a quite general distribution for a random-point process. Whenever a random-deletion process with
a very large attenuation rate is imposed, the nal statistics always obey the Poisson distribution,
irrespective of an initial distribution, as shown in Fig. 2-7.
However, p cannot be always made much smaller than one for a given average rate of random
pulse events per second because any physical system has a nite memory time
m
. If the time
slot =
T
N
is made smaller than the memory time
m
of the system, each event can no longer
be considered a statistically-independent process. Therefore, there is a lower limit for which, in
turn, places a lower limit on p = for a given average rate .
The other important distribution function is obtained when Npq 1, i.e., N 1 and p and q
are not too small. In such a case, the approximate expression for W
N
(K) is obtained as
W
(G)
N
(K) =
1

2K
2
exp
_

(KK)
2
2K
2
_
.
(Gaussian distribution)
(2.36)
The mean K and variance K
2
are given by (2.28) and (2.30) for the binomial distribution. When
p 1, but the mean K = Np is much greater than one, the Poisson distribution is reduced
to the Gaussian distribution. [The Gaussian distribution (2.36) is obtained from the binomial
13
K
2
N
2
N
2
=
Poisson
1
0
0
1
total noise
attenuated original noise
partition noise
K
2
N
2
1
0
0
0
1
N
2
= 2
Super-Poisson
total noise
attenuated original noise
partition noise
K
2
N
2
1
0
0
1
N
2
=
1
2

Sub-Poisson
total noise
attenuated original noise
partition noise
T
T
T
Figure 2-7
distribution (2.26) by considering W
N
(K) as a continuous function of K, expanding ln[W
N
(K)]
around the averaged point K by a Taylor series expansion and truncating to second order.]
2.4 Random Pulse TrainsSpectrum and Autocorrelation Function
The Fourier transform of (2.25) is
X(j) = F(j)
K

k=1
a
k
e
jt
k
. (2.37)
14
The unilateral power spectral density for such random pulse trains is dened by
S
x
() lim
T
2[X(j)[
2
T
= lim
T
2[F(j)[
2
T
K

k,m=1
a
k
a
m
exp[j(t
k
t
m
)] . (2.38)
The summation in (2.38) over k and m can be split into the summation for k = m and k ,= m,
S
x
() = lim
T
2[F(j)[
2
T
_
_
_
K

k=1
a
2
k
+

k=m
a
k
a
m
exp[j(t
k
t
m
)]
_
_
_
= 2a
2
[F(j)[
2
+ 4x(t)
2
() .
(Carsons theorem) (2.39)
Here, = lim
T
K
T
is the average rate of the pulse emission and a
2
= lim
T
1
K
K

k=1
a
2
k
is the mean-
square of the pulse amplitude. The mean of the noisy waveform x(t) is
x(t) = a
_

f(t)dt , (2.40)
and a = lim
T
1
K
K

k=1
a
k
is the mean of the pulse amplitude. The term containing the delta func-
tion represents the dc contribution to the power spectral density and is obtained by calculating
exp[j(t
k
t
m
)] using completely random distributions of t
k
and t
m
, identifying F( = 0) with
_

f(t)dt, and replacing lim

T
2 sin
2
(T/2)

2
T
with (). For a symmetric distribution of a
k
about
zero, the second term of (2.39) is zero because a is zero. However, when a
k
are not symmetrically
distributed about zero, the dc term appears in the power spectral density.
From the Wiener-Khintchine theorem, the autocorrelation function is calculated as

x
() =
a
2

_

0
[F(j)[
2
cos d + 2x(t)
2
_

0
() cos d
= a
2
_

f(t)f(t + )dt + x(t)

2
, (2.41)
15
where the Parseval theorem (2.12) is used to derive the second line. When = 0, one obtains
x(t)
2
x(t)
2
= a
2
_

[f(t)]
2
dt
=
a
2

_

0
[F(j)[
2
d
(Campbells theorem of mean square) , (2.42)
where the energy theorem (2.14) is used to derive the second line.
2.5 Current Noise of a Vacuum Tube
Let us consider a uctuating current i(t) caused by random emission of electrons from the
cathode of a vacuum tube, as shown in Fig. 2-8. Emission of electrons and subsequent travel
of the electrons induce the surface charges on the cathode and the anode. There is a relaxation
current owing into a circuit to build up such a surface charge at the electrode surfaces and to
restore the steady state of the two electrodes. This relaxation current constitutes the uctuating
current. In the vacuum tube case, the pulse energy (integrated relaxation current pulse) is equal
to q (constant) and the pulse shape f(t) is determined by an electron transit velocity between the
cathode and the anode. The pulse emission event follows the Poisson distribution with an average
rate of emission = lim
T
K
T
. Therefore, the average current is i(t) = q = I.
If one assumes f(t) is an exponentially decaying function with a time constant
1
, the autocor-
relation function is calculated from (2.41) as

i
() =
qI
2
1
e

||

1
+ I
2
. (2.43)
16
space charge region
electron multiplication region
t
t
t
Poisson-point-process
sub-Poisson-process
reflected electron
super-Poisson-process
added electron
Figure 2-8
The unilateral power spectral density is calculated from (2.39) as
S
i
() =
2qI
1 +
2

2
1
+ 4I
2
() .
(Schottky formula) (2.44)
As shown in Fig. 2-9, the low-frequency power spectral density (
1

1
) is equal to the full shot
noise 2qI and the spectral shape is Lorentzian with a cut-o frequency of
1

1
due to the nite
duration of a relaxation current pulse. The dc component of the spectrum is caused by the at
response of the autocorrelation function I
2
. On the other hand, the Lorentzian noise spectrum
stems from an exponentially decaying autocorrelation function centered at = 0.
The physical origin for the full shot noise of the current spectrum (2.44) is the fact that the
emission of each electron in such a thermal-limited vacuum tube is statistically independent and thus
obeys a Poisson-point process. This is not always true. In a vacuum tube of a space-charge limited
17
0 1 2 3 4 5
0
1
1 2 0
2

I
q
S
i
( ) 2qI
1
-1 -2 -3 -4
1

1
2I /
q
i ()
qI
2
( (
1

Figure 2-9
regime, there is a strong electron-electron Coulomb interaction in the space-charge region near
the cathode. A potential minimum produced by the space charge is modulated by the uctuating
emission rate and provides negative feedback to regulate the eective electron emission rate, as
shown in Fig. 2-8. In other words, each electron emission event is no longer statistically independent
but, rather, anti-bunched. Consequently, the circuit current uctuation is less than the full shot
noise. This shot noise suppression due to negative feedback eect is a rather common feature in
various electronic systems; examples will be seen in the following chapters, in which either the
Coulomb interaction or the Pauli exclusion principle (or both) play a role in the negative feedback
mechanism to suppress shot noise.
On the other hand, if randomly emitted electrons are amplied by an electron multiplier, the
initial Poissonian electron stream is transformed to the bunched electron stream, as shown in
Fig. 2-8. Consequently, the circuit current uctuation is more than the full shot noise. This excess
noise is also a rather common feature in various amplier systems, i.e., photomultipliers, avalanche
photodiodes, and laser ampliers.
18
3. Noise gure of linear networks
A noisy electrical network can be represented by a noise-free network with external noise
generators. The magnitude of the external generator is expressed either by an equivalent noise
resistance or an equivalent noise temperature. If such a network has four terminals or two (input
and output) ports, the noise gure is often used as a gure of merit expressing the inherent noisiness
of the circuit. The technique is also useful to express the noise properties of photonic systems such
as lasers and optical parametric ampliers.
3.1 Two-terminal networks Thevenin equivalent circuit
A noisy two-terminal network with impedance Z() = R()+jX() generates the open circuit
voltage uctuation v(t) as shown in Fig. 3-1(a). The two equivalent circuits based on Thevenins
theorem
[1]
are shown in Fig. 3-1(b) and (c). One is the noise-free network with impedance Z()
v(t) Z() Z() Y() i(t)
(a)
(b) (c)
v(t)
Figure 3-1: (a) A noisy two-terminal network. (b) Thevenin equivalent circuit with an external
voltage source. (c) Thevenin equivalent circuit with an external current source.
1
in series with a voltage generator v(t). The spectral density of v(t) is expressed by
S
v
() = 4k
B
R
n
, (3.1)
where is the absolute temperature and R
n
is the equivalent noise resistance. The other is the
noise-free network with admittance Y () = G() +j B() in parallel with a current generator i(t).
The spectral density of i(t) is expressed by
S
i
() = 4k
B
G
n
, (3.2)
when G
n
is the equivalent noise conductance.
If the network is linear and passive, and there is no net energy ow, i.e. the circuit is at thermal
equilibrium condition, then R
n
= R() and G
n
= G() = R()/[R()
2
+ X()
2
]. The noise in
this case is equal to Johnson-Nyquist thermal noise. However, in nonlinear active circuits, or in
non-equilibrium condition, these equalities generally do not hold.
When the electron temperature is dierent from the lattice temperature, which is the case for
hot electron devices, it is convenient to express (3.1) and (3.2) in the alternative forms:
S
v
() = 4k
B

n
R , (3.3)
S
i
() = 4k
B

n
G , (3.4)
where
n
is the equivalent noise temperature. In circuits containing shot noise sources as primary
noise sources, it is convenient to use the expression
S
i
() = 2qI , (3.5)
2
where I is the terminal current and is the shot noise suppression factor. If some smoothing
mechanisms are dominant, for instance due to space charge eect, becomes smaller than unity.
If there is no smoothing mechanism in the system, the power spectral density is full-shot noise
( = 1).
Consider the parallel RC circuit shown in Fig. 3-2(a). The thermal noise of the register is
represented by the parallel current source i(t) with the spectral density of S
i
() = 4k
B
/R. The
series voltage source v(t) in the Thevenin equivalent circuit shown in Fig. 3-2(b) has the spectral
density of
S
v
() =
R
2
1 +
2
(CR)
2
S
i
() . (3.6)
The frequency dependent power spectral density (3.6) is due to the impedance of the capacitor.
i(t)
(a) (b)
R C R C
v
n
Figure 3-2: (a) A parallel RC circuit with thermal noise current source. (b) The Thevenin equivalent
circuit with thermal noise voltage source.
Using the Wiener-Khintchine theorem, we obtain the mean-square value of the voltage generator
v(t)
2
=
1
2

0
S
v
()d =
k
B

C
. (3.7)
3
The mean-square energy stored in the parallel RC circuit is thus equal to
1
2
Cv(t)
2
=
1
2
k
B
. (3.8)
This is an example of the equipartition theorem of statistical mechanics
[2]
. That is, if the system
energy is of the form of quadratic dependence on generalized coordinate, i.e. the voltage in this case,
the average energy of the system under thermal equilibrium condition is equal to
1
2
k
B
per degree
of freedom. Notice that the noise energy is independent of the resistance R while the magnitude
and the bandwidth of the noise spectrum are dependent on R.
3.2 Four-terminal networks
A network with two pairs of terminals, input and output ports, is known as a four-terminal or
two-port network. For a noiseless four-terminal network, the currents and voltages at the terminals
are related to each other in terms of the impedance matrix Z or the admittance matrix Y as follows:

V
1
V
2

Z
11
Z
12
Z
21
Z
22

I
1
I
2

, (3.9)

I
1
I
2

Y
11
Y
12
Y
21
Y
22

V
1
V
2

. (3.10)
The subscripts 1 and 2 refer to the input and output ports, respectively, and the sign convention is
that currents owing into the network are positive, as shown in Fig. 3-3. The upper case letters I
and V indicate the Fourier transforms of the current and voltage, which are in general dependent
on frequency.
A noisy four-terminal network is represented by an extension of Thevenins theorem
[1]
. In
Fig. 3-4(a), a series noise voltage generator appears at each port. Some degree of correlation may
4
v
1
(t)
i
1
(t) i
2
(t)
v
2
(t)
Figure 3-3: A noiseless four-terminal network.
exist between these generators since the same physical mechanism may be responsible, at least in
part, for the open circuit voltage uctuations. The dual of Fig. 3-4(a) is shown in Fig. 3-4(b), in
v
1
(t)
i
1
(t)
Z
v
n1
(t)
v
2
(t)
i
2
(t)
v
n2
(t)
v
1
(t)
i
1
(t)
Y v
2
(t)
i
2
(t)
i
n1
(t) in2
(t)
(a) (b)
Figure 3-4: (a) Thevenin equivalent circuit with two external series voltage generators. (b) Thevenin
equivalent circuit with two external parallel current generators.
which the internal noise is represented by parallel current generators.
The current-voltage relation of a noisy four-terminal network becomes

V
1
+ V
n1
V
2
+ V
n2

Z
11
Z
12
Z
21
Z
22

I
1
I
2

, (3.11)

I
1
+ I
n1
I
2
+ I
n2

Y
11
Y
12
Y
21
Y
22

V
1
V
2

. (3.12)
It is often more convenient to refer both external generators to the input port. The equivalent
circuit shown in Fig. 3-5 has a series voltage generator and a parallel current generator at the
5
input, for which the current-voltage characteristic is expressed by the relation

I
1
+ I
na
I
2

Y
11
Y
12
Y
21
Y
22

V
1
+ V
na
V
2

. (3.13)
Comparing (3.12) and (3.13), the Fourier transform of the new voltage generator v
na
(t) should be
v
1
(t)
i
1
(t)
Y v
2
(t)
i
2
(t)
i
na
(t)
v
na
(t)
Figure 3-5: The equivalent circuit of a noisy two-port with external current and voltage generators
in the input port.
related to that of the current generator i
n2
(t):
V
na
=
I
n2
Y
21
. (3.14)
The Fourier transform of the new current generator i
na
(t) is
I
na
= I
n1

Y
11
Y
21
I
n2
. (3.15)
The arrangement of Fig. 3-5 is particularly convenient for calculating the noise gure of the
two-port network. However, this equivalent circuit is valid only for calculating the noise in the
output port. It does not give the correct description of the noise in the input port. This can be
easily seen by the fact that I
na
is not equal to I
n1
.
6
3.3 Noise gure of a linear two-port
The noise gure of a two-port is dened by
F =
total output noise power per unit bandwidth
output noise power per unit bandwidth due to input termination
at a specic frequency and temperature. A signal input i
s
(t) is transferred to the output through
an input admittance Y
s
and noisy two-port network, as shown in Fig. 3-6. The noise in the input
Y i
na
(t)
v
na
(t)
i
ns
(t) Y
s
i
s
Figure 3-6: A circuit for calculating the noise gure of a two-port network.
admittance Y
s
and the noisy two-port are independent, and so the noise gure of the whole system
can be expressed as
F =
|I
ns
+ I
na
+ Y
s
V
na
|
2
|I
ns
|
2
= 1 +
S
ia
()
S
is
()
+|Y
s
|
2
S
va
()
S
is
()
+ 2Re(
iv
Y

s
)

S
ia
() S
va
()

1/2
S
is
()
, (3.16)
where S
ia
(), S
va
() and S
is
() are the power spectra of i
na
(t), v
na
(t) and i
ns
(t), respectively, and

iv
is the normalized cross-correlation spectral density between i
na
(t) and v
na
(t).
The power spectral densities can be expressed as
S
ia
() = 4k
B
G
ni
, (3.17)
S
va
() = 4k
B
/G
nv
, (3.18)
7
S
is
() = 4k
B
G
s
. (3.19)
Here G
ni
and G
nv
are the equivalent noise conductances, which are not necessarily actual conduc-
tances of the two-port network. On the other hand, G
s
= Re(Y
s
) is the actual source conductance.
The current generator i
na
(t) is split into two, one part of which is uncorrelated with v
na
(t) and the
other part is fully correlated with v
na
(t). Therefore we obtain
I
na
= I
nb
+ Y
c
V
na
, (3.20)
where Y
c
is the correlation admittance of i
na
(t) and v
na
(t). Since I
nb
V

na
= 0, we have

iv

I
na
V

na

|I
na
|
2
|V
na
|
2

1/2
= Y
c

|V
na
|
2
|I
na
|
2

1/2
=
Y
c

G
ni
G
nv
. (3.21)
The noise gure of (3.16) is now rewritten as
F = 1 +
G
ni
G
s
+
(G
s
+ G
c
)
2
+ (B
s
+ B
c
)
2
(G
2
c
+ B
2
c
)
G
nv
G
s
, (3.22)
where G
c
and B
c
are the real and imaginary parts of the correlation admittance Y
c
. This expression
is easily transformed into the form
F = F
0
+
(G
s
G
so
)
2
+ (B
s
B
so
)
2
G
nv
G
s
. (3.23)
Here, F
0
= 1 +
2
Gnv
(G
so
+ G
c
) is the minimum noise gure achieved when the source admittance
satises the following matching condition:
G
s
= G
so
= (G
nv
G
ni
B
2
c
)
1/2
, (3.24)
B
s
= B
so
= B
c
. (3.25)
8
The conditions (3.24) and (3.25) for the source conductance and susceptance are referred to as
noise tuning or noise matching. The four parameters F
o
, G
so
, B
so
and G
nv
completely characterize
the noise uctuations of the two-port network.
3.4 Noise gure of ampliers in cascade
One of the important applications of the equivalent circuit discussed in the previous section
is the overall noise gure of the system when several ampliers are connected in cascade as shown
in Fig. 3-7. Suppose each amplier in the cascade is connected to a matched load, i.e. the output
F
1
, G
1 i
na1
(t)
v
na1
(t)
i
ns
(t) Y
s
i
s
(t)
F
2
, G
2
load i
na2
(t)
v
na2
(t)
Figure 3-7: The equivalent circuit of a cascade amplier system.
and input admittances of adjoining ampliers are equal. The noise gure of the whole system in
such a case can be written as
F =
|I
sn
+ I
na1
+ Y
s
V
na1
|
2
|I
ns
|
2
+
|I
na2
+ Y
1
V
na2
|
2
G
1
|I
ns
|
2
+
|I
na3
+ Y
2
V
na3
|
2
G
1
G
2
|I
ns
|
2
+ . (3.26)
Here Y
i
and G
i
are the input admittance and power gain of the i-th amplier. If F
i
is the noise
gure of the i-th amplier dened by (3.16), the overall noise gure is
F = F
1
+ (F
2
1)/G
1
+ (F
3
1)/G
1
G
2
+ . (3.27)
9
Equation (3.27) is known as Friisss formula
[3]
. The expression indicates, that the noise gure of
the cascade amplier system is essentially determined by the noise gure of the rst stage if the
power gain of the rst stage is suciently high. It is important to use a low-noise amplier in the
rst stage in order to realize a small overall noise gure.
References
[1] L. Thevenin, Comptes Rend. Acad. Scie. Paris, 97, 159 (1883).
[2] R. Lief, Fundamentals of statistical and thermal physics (McGraw-Hill, New York, 1965).
[3] H. T. Friiss, Proc. IRE, 32, 419 (1944).
10
4. Thermal Noise and Quantum Noise
There are two types of intrinsic noise in every physical system: thermal noise and quantum
noise. These two types of noise cannot be eliminated even when a system is perfectly constructed
and operated. Thermal noise is a dominant noise source at high temperatures and/or low frequen-
cies, while quantum noise is dominant at low temperatures and/or high frequencies. A conductor is
the simplest physical system which produces these two types of intrinsic noise. The intrinsic noise
of macroscopic and mesoscopic conductors will be discussed in this chapter.
A conductor in thermal equilibrium with its surroundings shows, at its terminals, open-circuit
voltage (or short-circuit current) uctuations, as shown in Fig. 4-1. Thermal equilibrium noise
R
R
v
i
S
v
( ) = 4k
B
R S
i
( ) =
4k
B
R
Figure 4-1:
was rst observed by M. B. Johnson in 1927. He discovered that the noise power spectral density
is independent of both the material the conductor is made of and the measurement frequency,
and is determined only by the temperature and electrical resistance. The open-circuit voltage
noise spectral density is S
v
() = 4k
B
R and the short-circuit current noise spectral density is
S
i
() = 4k
B
/R. This noise is referred to as thermal noise and is the most fundamental and
1
important noise in electronic devices.
The physical origin of thermal noise in a macroscopic conductor is a Mrandom-walk of
thermally-uctuated electrons. An electron undergoes a Brownian motion via collisions with the
lattices in a conductor. The statistical properties of such a Brownian particle were studied rst
by A. Einstein twenty years before Johnsons observation of thermal noise. The electrons in a
conductor are thermally energetic via collisions with the lattice and travel randomly. The electron
velocity uctuation is a statistically-stationary process. However, the mean-square displacement
of an electron increases in proportion to the observation time. The electron position uctuation is
a statistically-nonstationary process. Such a microscopic approach can indeed explain Johnsons
observation.
Nyquist employed a completely dierent approach to the problem. He introduced the concept
of a mode for the system by using a transmission line cavity terminated by two conductors. He
then applied the equipartition theorem of thermodynamics to these transmission line modes. In
this way he could explain Johnsons observation without going into the details of a microscopic
electron transport process. Nyquists approach is very general and is easily extended to include
quantum noise, which is important in a high-frequency and low-temperature case.
In order to introduce quantum noise to a conductor, one must assume the quantization pos-
tulate. The current and voltage of a LC circuit are a pair of conjugate observables just like the
position and momentum of a particle or a mechanical harmonic oscillator. One cannot determine
these two conjugate observables simultaneously with arbitrary accuracy due to the Heisenberg
uncertainty principle. These quantum uncertainties in the current and voltage have their micro-
2
scopic origins in that the momentum (corresponding to the short circuit current) and position
(corresponding to the open circuit voltage) of an electron in a conductor cannot be determined
simultaneously.
A mesoscopic conductor does not impose any scattering on electrons and an electron travels as
a coherent wave in such a small-scale device. Nevertheless, there is still a nite electrical resistance
(quantum unit of resistance R
Q
=
h
2e
2
) due to the Pauli exclusion principle. It is truly surprising to
nd that the formula for thermal noise and quantum noise of a macroscopic conductor with many
scatterings still hold for such a mesoscopic conductor.
4.1 Microscopic Theory of Thermal Noise Einsteins Approach
4.1-1 Mean Free-Time and Mobility
u(t)
collision

f1

f2

f3
E E E E E E E
u(t) : average drift
velocity
V
collision with the lattice
t
Figure 4-2:
Consider a one-dimensional conductor under an applied voltage (Fig. 4-2). We assume that an
electron is accelerated by the uniform electric eld between collisions with the lattice and that an
3
electron velocity returns to zero at every collision event. This is not true in a real collision process
in a conductor, but the conclusion one obtains using this assumption is essentially same as that
obtained by more realistic collision models. Since the electron drift velocity in this model is given
by u(t) = t =
qE
m
t during the time between collisions
f
, the displacement between two collisions
is
x(
f
) =

2

2
f
=
qE
2m

2
f
. (4.1)
After K collisions with the lattice, the total displacement is
qE
2m

2
f
K. Therefore, the mean drift
velocity u is given by
u
total displacement
total time
=
_
qE
2m
K
_

2
f
K
f
=
q
2
f
2m
f
E . (4.2)
Here,
f
and
2
f
are the mean free time and mean-square free time. The mobility is dened by
u = E. Thus, one obtains
=
q
2
f
2m
f
. (4.3)
As shown in Fig 4-2,
f
is randomly distributed around its mean value
f
. The probability
p
i
(m, ) that the i
th
electron experiences exactly m collisions in a time interval [0, ] obeys a Poisson
distribution if the probability of electron collision with the lattice is independent of the electron
drift velocity which is a reasonable assumption for a weak eld; thus it is a statistically independent
process. Strictly speaking, the electron collision with a lattice obeys a binomial distribution with
a very small collision probability starting from a very large number of (potential) collision events,
which is well approximated by a Poisson distribution. Thus, one obtains
p
i
(m, ) =
(
i
)
m
m!
e

, (4.4)
4
where
i
is the mean rate for collision per second. The probability for a free time
f
lying between

fi
and
fi
+ d
fi
is equal to the joint probability of zero collisions in a time interval [0,
fi
] and
one collision in a time interval [
fi
,
fi
+d
fi
]. Thus,
q(
fi
)d
fi
= p
i
(0,
fi
) p
i
(1, d
fi
)
=
i
e

fi
d
fi
, (4.5)
where e

i
d
fi
1 is used and
i
e

fi
is considered a probability density function by which we
can calculate the mean free time and mean-square free time,

fi
=
i
_

0

fi
e

fi
d
fi
=
1

i
, (4.6)

2
fi
=
i
_

0

2
fi
e

fi
d
fi
=
2

2
i
= 2
fi
2
. (4.7)
As seen from (4.6) and (4.7),
fi
does not obey a Poisson distribution, but, rather, a geometrical
distribution. Figure 4-3 compares p
i
(m, ) and q
i
(
fi
). If N electrons behave in a similar but
0
0.5
1
0
1
0
0.5
1
0 1 2 3 4 5
Poisson distribution
m
Geometical distribution

i

fi

2
i
p
i
m,
q
i

fi
/
i
Figure 4-3:
independent way, the Maddition theorem of a Poisson process is applied and the total probability
of collision p(m, ) still obeys a Poisson distribution with the mean rate =
N

i=1

i
=
N

f
. The
(collective) mean free time
f
is dened by
f
=
N

.
5
Using (4.7) in (4.3), the mobility is uniquely related to the mean free time,
=
q
f
m
. (4.8)
According to the collision model, any departure of the drift velocity u(t) = u(t)u from its mean
value decays with a time constant
f
=
m
q
.
4.1-2 Velocity and Current Fluctuations
In a thermal equilibrium condition (zero applied voltage), the mean drift velocity of an electron
is zero, u = 0. However, an electron acquires a non-zero momentum when it collides with a lattice
which decays toward the mean value of u = 0 with a time constant
f
. If its momenta before and
after a collision are p
1
and p
2
, respectively, the collision imparts a momentum change (p
2
p
1
) in
an innitesimal time. The equation of motion is then
du(t)
dt
=
u(t)

f
+
(p
2
p
1
)(t)
m
. (4.9)
The drift velocity u(t) is kicked randomly by the second part of the RHS of (4.9) and is simulta-
neously damped by the rst part of the RHS of (4.9). The Fourier analysis of (4.9) requires the
introduction of a gated function u
T
(t) = u(t) for
T
2
t
T
2
and 0 for otherwise because u(t) is
a statistically-stationary process. The Fourier transform of this gated function leads to
U
T
(j) =

q
(p
2
p
1
)
1 +j
f
. (4.10)
From the Carson theorem, the power spectral density of the velocity uctuation is given by
S
u
() = 2(p
2
p
1
)
2
(/q)
2
1 +
2

2
f
, (4.11)
6
where = lim
T
K
T
=
1

f
is the mean rate of collisions, K is the number of collision events, and
(p
2
p
1
)
2
= lim
T
1
K
K

k=1
(p
2k
p
1k
)
2
is the mean square of the momentum change per collision.
Since momenta before and after a collision are statistically independent,
(p
1
p
2
)
2
= 2p
2
_
= 2 lim
T
1
K
K

k=1
p
2
k
_
. (4.12)
For a thermal equilibrium condition, the mean square of the electron momentum is determined by
the equipartition theorem:
p
2
2m
=
1
2
k
B
(one dimensional case) . (4.13)
From (4.11)(4.13), one obtains
S
u
() =
4k
B
/q
1 +
2

2
f
, (4.14)
where (4.8) is used.
From the Wiener-Khintchine theorem, one can calculate the autocorrelation function

u
() =
1
2
_

0
S
u
() cos d
=
k
B

q
f
exp
_

[[

f
_
. (4.15)
Therefore, the mean square value of u(t) is given by
u
(0) =
k
B

q
f
. Figures 4-4(a) and (b) show
S
u
() and
u
().
The velocity uctuation u(t) of the electron produces a short-circuited current uctuation i(t) =
qu(t)/L in an external circuit, where L is the length of the conductor. Since the current uctuation
of each electron is additive, the total current uctuation power spectral density is given by
S
i
() = S
u
()
q
2
L
2
ALn , (4.16)
7
where A is a cross-sectional area of the conductor and n is the electron density. If one uses (4.14)
and the expression for electrical resistance, R =
L
A
=
L
nqA
, in (4.16), one obtains
S
i
() =
4k
B
/R
1 +
2

2
f
. (4.17)
Therefore, the short-circuited current uctuation power spectral density at low frequency is 4k
B
/R,
which is exactly what Johnson observed experimentally.
1
0.5
0 1 -1 10 1
1
S
u
( ) /
4k
B
q
f
u
)
q
f
k
f
10
-1
10
-1
10
-2
10
-2
Figure 4-4:
4.1-3 Position and Charge Fluctuations
The integral of a statistically-stationary process which shows a white power spectral density
is called a Wiener-Levy process. The surface charge induced by the uctuating short-circuited
current across a conductor is
q(t) =
_
t
0
i(t

)dt

. (4.18)
Consider again the one-dimensional conductor shown in Fig. 4-2. An electron transit over a free
path
f
between collisions gives rise to a surface charge on the electrode equal to q
_

f
L
_
. If there
are m = t independent events in a time duration t , then the mean-square value of the surface
8
charge q(t) is calculated as
q(t)
2
= q
2
(t)

2
f
L
2
=
2k
B

R
t , (4.19)
where = nAL/
f
is the mean rate of collisions,
2
f
= v
2
T

2
f
= 2k
B

f
/q is the mean-square
free path, and v
2
T
= k
B
/m is the mean-square thermal velocity; therefore, the charge uctuation
is not stationary, but, rather, diuses with time. The diusion coecient, dened by q(t)
2
= 2Dt,
is given by
D
q
=
k
B

R
=
1
4
S
i
( 0) . (4.20)
Since q(t) is a cumulative process, one can write
q(t +) = q(t) + q(t, ) , (4.21)
where q(t, ) is the charge uctuation added between times t and (t + ). Since the correlation
time of the additive uctuations is very short (
f
), the two processes in the RHS of (4.21) are
uncorrelated; therefore, the covariance function is equal to the mean square value:
q(t +)q(t) = q(t)
2
+q(t)q(t, ) q(t)
2
. (4.22)
The autocorrelation function must be redened in such a manner that the observation time remains
nite rather than innite and the ordinate is set to zero outside the observation time interval. Thus,
the autocorrelation function is a function of the time delay and the observation time T:

q
(, T) =
1
T
_
T||
0
q(t +)q(t)dt
=
1
T
_
T||
0
2k
B

R
t dt
=
k
B

R
T
_
1
[[
T
_
2
. (4.23)
9
The power spectral density is obtained by the Wiener-Khintchine theorem as:
S
q
(, T) = 4
_
T
0

q
(, T) cos d
=
8k
B

2
R
_
1
sin(T)
T
_
. (4.24)
When T 1, the power spectral density is proportional to
2
, which is characteristic of a
Wiener-Levy process. Figures 4-5(a)(c) show the real-time function, autocorrelation function,
and power spectral density of the surface charge q(t). The integral of an electron velocity u(t)
t
1
1 10
q(t)
0
random walk diffusion
=
0.1
S , T)
q
0 -1 1
1
0.5
,
T)
q
.
q(t)
2
2k
B
R
t
3R
4k
B
T
2
-2
-2
-1
R
k T
Figure 4-5:
is an electron displacement x(t), which is also a Wiener-Levy process. By performing the same
10
calculation, one obtains the (position) diusion coecient,
D
x
=
k
B

q
=
1
4
S
u
( 0) .
(Einsteins relation) (4.25)
4.2 Macroscopic Theory of Thermal Noise Nyquists Approach
Nyquists treatment of thermal noise appeared very soon after Johnsons observation and the
result is often referred to as Nyquists Law. Nyquist initially considered two electrical conductors,
R
1
and R
2
, connected in parallel, as shown in Fig. 4-6. The open-circuit voltage noise source v
1
associated with conductor R
1
produces a current uctuation in the circuit, leading to an absorbed
power R
2
v
2
1
/(R
1
+ R
2
)
2
by conductor R
2
. A similar ow of absorbed power, R
1
v
2
2
/(R
1
+ R
2
)
2
,
exists from R
2
to R
1
. Since the two conductors are at the same temperature, the power ow in
each direction must be exactly the same and cancel each other out; otherwise, the second law of
thermodynamics would be violated. The second law of thermodynamics states that it is impossible
to take heat from one reservoir to another reservoir at an equilibrium temperature. In order to
satisfy this exact cancellation of power ow from R
1
to R
2
, and vice versa, the open circuit voltage
noise v
2
i
should be proportional to the electrical resistance R
i
. This exact cancellation of power ow
must hold not only for the total power, but also for the power exchanged in any frequency band;
otherwise, the second law of thermodynamics would be violated simply by inserting a frequency
lter between the resistors. In other words, the power spectrum S
v
() of the voltage uctuations
should be independent of the detailed structure and material of the conductor and should be a
universal function of R, , and (angular) frequency . Specically, if the conductors R
1
and R
2
are
11
at dierent temperatures,
1
and
2
, the net heat ow should be proportional to the temperature
dierence
1

2
and thus the power spectrum S
v
() is proportional to the temperature . Nyquist
R
1
conductor 1
conductor 2 2
v v
1
R
2
R
2
R
1
+ R
2
2
v
1
2
) (
R
R
1
+ R
2
2
v
2
) (
1
2
Figure 4-6:
then considered a long, lossless transmission line terminated at either end by conductors R
1
and
R
2
(Fig. 4-7). This lossless LC transmission line has a characteristic impedance Z
0
=
_
L
C
which
is equal to the end terminal conductance R
1
= R
2
= R and the wave velocity v =
1

LC
, where
L and C are the inductance and capacitance of the transmission line per unit length. The power
delivered to the transmission line from R
1
or R
2
in a frequency interval d/2 is given by
dP =
1
4R
S
v
()
d
2
. (4.26)
A time duration in which this noise power travels in the transmission line is t =

v
, where is the
length of the transmission line. Hence, the total stored energy in the transmission line in the same
frequency band is
dE = dP t 2 =

2Rv
S
v
()
d
2
. (4.27)
Suppose the transmission line is suddenly short-circuited by closing the two switches across R
1
and R
2
and, therefore, the energy density (4.27) on the transmission line is trapped as standing
waves. The resonant frequencies of these standing waves are
N
=
v
2
N, where N is a positive
12

1
=
v
2l

2
=
v
2l
2 ~

3
=
v
2l
3 ~

N
=
v
2l
N ~
R1
2
v
R
2
l
v
1
Figure 4-7:
integer. The number of standing wave modes in the frequency interval
d
2
is given by
m =
_
d
2
___
v
2
_
=
d
v
. (4.28)
As the transmission line length goes to innity, the number of degrees of freedom (DOF) of the
system (given by the number of the standing wave modes) also goes to innity. Therefore, it is
permissible to invoke the equipartition theorem to determine the total energy in the transmission
line at thermal equilibrium. The average energy per mode (with two DOFs) is k
B
and thus the
stored energy in the frequency band
d
2
is given by
dE = mk
B
=
d
v
k
B
. (4.29)
13
From 4.27) and 4.29), the power spectrum is obtained as
S
v
() = 4k
B
R . (4.30)
As expected from a thermodynamic argument, S
v
() is proportional to R and . If an angular
frequency becomes greater than k
B
/h, the average thermal energy per DOF is calculated as
h n
th
=
h
exp
_
h
k
B

_
1
, (4.31)
where the photon number distribution function P
n
obeys the geometrical distribution
P
n
=
n
n
th
(1 +n
th
)
n+1
= exp
_

nh
k
B

__
1 exp
_

h
k
B

__
, (4.32)
and thus the mean thermal photon number is
n
th

n=0
nP
n
=
1
exp
_
h
k
B

_
1
. (4.33)
Therefore, the total energy in the frequency band
d
2
is
dE = mhn
th
=
d
v

h
exp
_
h
k
B

_
1
. (4.34)
Comparing this result with (4.27), the power spectral density of the voltage noise is found to be
S
v
() =
4hR
exp
_
h
k
B

_
1
. (4.35)
In (4.35), the quantized energy h of a photon is taken into consideration. However, the full
quantum mechanical analysis of the problem suggests that it is still insucient since the zero-point
uctuation is not included.
14
4.3 Quantum Noise
Consider a lossless LC circuit in order to derive the zero-point uctuation. The resonant
frequency of an LC circuit is given by
0
=
1

LC
. One now models each transmission line cavity
mode in Fig. 4-7 by such a lumped LC resonant circuit. If the current owing in the inductance L
is denoted by i(t) and the voltage across the capacitance C is denoted by v(t), the Kirchho laws
are represented by
C
dv(t)
dt
= i(t) , (4.36)
L
di(t)
dt
= v(t) . (4.37)
These equations can be expressed in terms of the normalized voltage, q(t) Cv(t), and the
normalized current, p(t) Li(t):
dq(t)
dt
=
1
L
p(t) , (4.38)
dp(t)
dt
=
1
C
q(t) . (4.39)
The total energy stored in the lossless LC circuit is given by
H =
1
2
Li
2
+
1
2
Cv
2
=
p
2
2L
+
q
2
2C
. (4.40)
Suppose one interprets the above expression in terms of an analogy with a mechanical harmonic
oscillator with a mass m and oscillation frequency
0
=
1

LC
. One can then identify the inductance
L corresponding to the mass m, the capacitance C corresponding to the inverse of the spring
constant
1
k
=
1
m
2
0
, and p and q corresponding to the momentum and position, respectively. If one
considers 4.40) as a Hamiltonian function of the system, the classical Hamilton equations for the
15
position q and momentum p are written as
dq
dt

H
p
=
1
L
p , (4.41)
dp
dt

H
q
=
1
C
q , (4.42)
which are identical in form to the Kirchho equations (4.37) and (4.38) for the voltage q and
current p. Hence, one can conclude that the voltage and current in a lossless LC circuit
are a pair of conjugate observables, just like the position and momentum of a mechanical har-
monic oscillator. Quantum mechanically, conjugate observables q and p must satisfy the following
commutation relation,
[q, p] qp pq = ih , (4.43)
where q and p are no longer complex numbers (c-numbers); rather, they are quantum mechanical
operators (q-numbers). Using the Schwartz inequality, one can derive the following Heisenberg
uncertainty principle for the product of the variances of q and p:
q
2
p
2

h
2
4
. (4.44)
The (non-Hermitian) creation (a

) and annihilation (a) operators are dened by:

a =
1

2h
0
L
(
0
Lq +ip) , (4.45)
a

=
1

2h
0
L
(
0
Lq ip) . (4.46)
The creation operator a

is a Hermitian conjugate of the annihilation operator a. The Hamiltonian

function (4.40) and the commutation relation (4.43) are rewritten as
H = h
0
_
a

a +
1
2
_
, (4.47)
16
[a, a

] = 1 . (4.48)
If the LC circuit is not excited at all, the quantum mechanical state of such an unexcited LC circuit
is referred to as the ground state, or the vacuum state. The vacuum state is mathematically
dened by
a[0) = 0 . (4.49)
One cannot annihilate a photon from the vacuum state because there is no photon in the vacuum
state, which is the meaning of this expression. On the other hand, the operation of the creation
operator on the vacuum state creates a single-photon state:
a

[0) = 1 . (4.50)
The mean and mean-square of the voltage and current of the vacuum state are calculated by
taking the ensemble average of the respective operators:
q 0[q[0) = 0[

h
0
C
2
(a

+a)[0) = 0 , (4.51)
q
2
0[q
2
[0) = 0[
h
0
C
2
(a
2
+a
2
+a

a +aa

)[0) =
h
0
2
C , (4.52)
p 0[p[0) = 0[i

h
0
L
2
(a

a)[0) = 0 , (4.53)
p
2
0[p
2
[0) = 0[
h
0
L
2
(a
2
a
2
+a

a +aa

)[0) =
h
0
2
L . (4.54)
Here, the orthogonality relations between photon number eigenstates such as 0[1) = 0[2) = 0 are
used. The variances of the voltage and current satisfy the minimum uncertainty product,
q
2
q
2
q
2
=
h
0
2
C
p
2
p
2
p
2
=
h
0
2
L
q
2
p
2
=
h
2
4
. (4.55)
17
Notice that both voltage and current have intrinsic uctuations (variances) even at absolute zero
temperatures. The expectation value of the zero-point energy is given by
H) 0[H[0) = 0[h
0
_
a

a +
1
2
_
[0) =
h
0
2
. (4.56)
If one adds this zero-point uctuation contribution to (4.35), one obtains the full quantum
mechanical expression for an open-circuit voltage uctuation spectral density:
S
v
() = 2 hR coth
_
h
2k
B

_
. (4.57)
As shown in Fig. 4-8, S
v
() is reduced to the thermal noise value (4k
B
R) in the high-temperature
limit (k
B
h) and is reduced to the quantum noise value (2 hR) in the low-temperature limit
(k
B
h). Equation (4.57), including the quantum mechanical zero-point uctuation, is referred
to as the generalized Nyquist noise.
1
10
1 10
10
2
10
-1
10
-1
10
-2
10
2
Thermal Noise limit
S
v
() / 4k
B
R
Quantum Noise Limit
h
2k
B

Figure 4-8:
18
4.4 Thermal and Quantum Noise of a Mesoscopic System
If the size of a conducting wire becomes smaller and smaller, an electron can eventually transit
from one electrode to the other without a collision with the lattice. Such a small-scale conductor is
called a mesoscopic conductor. A collision-free electron transport in a mesoscopic conductor is
termed a ballistic transport, while a conventional electron transport accompanying many collisions
with the lattice in a macroscopic conductor is termed a dissipative transport. Since an electron
propagates as a coherent wave in such a mesoscopic conductor, various coherent electron wave
interferometric devices can be constructed. Even though such an electron wave is not scattered by
the lattice, it is known that the terminal current of such a mesoscopic conductor still features noise.
Consider a mesoscopic conductor with two terminal electrodes connected to an external constant
V
electrode electrode
mesoscopic conductor
electron
energy
coherent electron wave
k
N
=
2
L
N
1
2
1 2
qV
Figure 4-9:
voltage source, as shown in Fig. 4-9. The electron distributions in the two terminal electrodes
obey the thermal equilibrium Fermi-Dirac distributions. The Fermi-level dierence is determined
by the external voltage. An electron wavepacket with a certain energy is emitted from electrode 1
19
and absorbed by electrode 2, which contributes to an external current. As shown in Fig. 4-9, the
longitudinal wavenumber of such a coherent electron wave is quantized by the boundary condition:
k
N
=
2
L
N (N = positive integer) . (4.58)
The mode density per energy is given by
dN
dE
=
dN
dk
N
dk
N
dE
=
L
2hv
N
, (4.59)
where v
N
=
d
dk
N
is the electron group velocity. The total number of independent modes in the
energy range
1

2
is
dN
dE
(
1

2
). Each mode can accommodate a single electron due to the
Pauli exclusion principle and has a transit time t
N
=
L
v
N
between the two electrodes. Therefore,
the average current is
I = q
dN
dE
(
1

2
)
1
t
N
=
q
2
h
V , (4.60)
where
1

2
= qV (V = external bias voltage) is used. A mesoscopic conductor has a nite
conductance G
Q
=
q
2
h
(quantum unit of conductance) that is not due to scattering with the lattice
(dissipation), but due to the Pauli exclusion principle.
An alternative method to derive the quantum unit of conductance G
Q
is to count the number
of independent temporal modes (rather than Fourier modes). As shown in Fig. 4-10, the
inverse Fourier transform of a bandpass-ltered spectrum centered at a frequency
0
and bandwidth
(FWHM) =

1

2
h
is a Nyquist function f
N
=
sin(t)
(t)
. A sequence of Nyquist functions
displaced by =
1

reproduces any bandwidth-limited function, and thus one can consider the
arrival rate of independent temporal modes as
1

= . If one assigns a single electron per

20
independent temporal mode, the current is
I =
q

=
q
2
h
V . (4.61)
The same result is obtained as in 4.60). If a mesoscopic conductor supports only one transverse
mode, two spin-degenerate states can still be accommodated for the single transverse mode and
thus the total conductance is
2q
2
h
instead of
q
2
h
. If a system is at an absolute zero temperature,
Spectrum
0
E
=
1
Nyquist
function
1 2 t
Figure 4-10:
as shown in Fig. 4-11(a), the state above the Fermi energy is completely empty and the state
below the Fermi energy is completely occupied. The electron wavepackets (temporal modes) with
center frequencies lying between the two Fermi energies are emitted from electrode 1 and travel
toward electrode 2. The electron wavepackets with energies outside of this region do not contribute
to a current. This electron transport is a deterministic process and thus there is no dc current
noise in the external circuit. However, if one measures the current noise at a nite frequency, the
noise is produced by the beating between the electron wavepacket emitted from one electrode in
the energy below the Fermi energy
1
and the empty electron wavepacket (vacuum state) emitted
from the same electrode in the energy above the Fermi energy
1
. By increasing the measurement
21
frequency, additional occupied and empty electron wavepackets contribute to the beat noise and
the noise power thus increases linearly in proportion to . This is the origin of the quantum noise
which is dominant in a low-temperature/high-frequency region, as shown in Fig. 4-12.
If a system is at a nite temperature, as shown in Fig. 4-11(b), the electron occupancy near
the Fermi energy becomes probabilistic. The electron wavepackets with energies close to
1
and

2
may or may not be emitted from electrode 1 and absorbed by electrode 2. This stochasticity
introduces another current noise in the external circuit. By increasing the temperature, the partial
occupancy of electron states increases and the noise power thus increases linearly in proportion to
. This is the origin of the thermal noise which is dominant in a high-temperature/low-frequency
region, as shown in Fig. 4-12.
beating
empty packets
occupied
packets beating
(a) = 0 (Quantum Noise) (b) = 0 (Thermal Noise)
statistical occupancy
2
1
2
1
Figure 4-11:
22
1
10
1 10
10
10
10 10
10
Thermal Noise limit
Quantum Noise
h
2k
B
S ( )
R
Q
4k
B
Figure 4-12:
References
1. Mobility and diusion constant of electrons in conductors:
W. Shockley, Electrons and Holes in Semiconductors, Van Nostrand, New York (1963).
2. Langevin equation and uctuation-dissipation theory of a Brownian particle:
M. Sargent III, M. O. Scully and W. E. Lamb, Jr., Laser Physics (Addison-Wesley,
Reading, Mass. 1974).
3. Thermal noise:
M. B. Johnson, Nature 119, 50 (1927); Phys. Rev. 32, 97 (1928).
H. Nyquist, Phys. Rev. 32, 110 (1928).
4. Quantum zero-point uctuations of a loss less LC circuit:
23
W. H. Louisell, Radiation and Noise in Quantum Electronics, McGraw-Hill Book Co.,
New York (1964); Quantum Statistical Properties of Radiation, John Wiley & Sons, New
York (1973).
5. Generalized Nyquist formula:
H. B. Caller and T. A. Welton, Phys. Rev. 83, 34 (1951).
24
Contents
5. Inherent Noise of pn Junction Diodes 2
5.1 pn Junction Diodes Under Constant Voltage Operation . . . . . . . . . . . . . . . . 4
5.1.1 Current-Voltage and Capacitance-Voltage Characteristics of a pn Junction . . 4
5.1.2 Thermal Diusion Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
5.1.3 Generation-Recombination Noise . . . . . . . . . . . . . . . . . . . . . . . . . 13
5.1.4 Total Current Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.1.5 Short Diode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5.2 pn Junction Diodes Under Constant Current Operation . . . . . . . . . . . . . . . . 21
5.2.1 Eect of a Finite Source Resistance . . . . . . . . . . . . . . . . . . . . . . . 21
5.2.2 A pn Junction Diode With Negligible Backward Thermionic Emission . . . . 23
1
5. Inherent Noise of pn Junction Diodes
Shockleys 1949 paper heralded a new era in the history of semiconductor device physics.
Basic physical processes of a junction diode and a junction transistor were presented in this 1949
Bell Syst. Tech. J. paper. In the same issue, the rst report appeared on the noise of a point
contact transistor. The observed noise gure was 50-70 dB above the intrinsic noise limit! It took
almost 30 years to suppress this excess noise (mainly due to 1/f noise and surface recombination
noise) and to obtain a noise gure very close to the theoretical limit (= 0 dB). This intrinsic noise
of a transistor is determined by the thermal noise in the bulk resistive region and the shot noise in
the junction region. In this chapter we will study the inherent noise of pn junction diodes, which
set a fundamental limit on the noise performance of various semiconductor devices.
There are two distinct bias conditions for a junction diode: constant voltage operation and
constant current operation. The former is obtained when the junction (dierential) resistance R
d
is much larger than the source resistance R
S
, and the latter is obtained in the opposite limit.
There are two types of junction diodes: a macroscopic junction and a mesoscopic junction. An
electrostatic energy for a single electron q
2
/2C (where C is a junction capacitance) is much smaller
than the thermal energy k
B
in a macroscopic junction and the opposite is true for a mesoscopic
junction. A junction diode features markedly dierent noise characteristics in such dierent bias
conditions and operational regimes.
Consider a pn junction diode biased by a constant voltage source with a source resistance R
S
,
as shown in Fig. 5-1. If the source resistance R
S
is much smaller than a diode dierential resistance
2
dened by
R
d

_
dI
dV
_
1
, (5.1)
where I and V are the junction current and the junction voltage, then the junction voltage V is
p
n
R
s
v
R
d
C
dep
C
dif
R
s i
s i
Figure 5-1:
pinned by the source. There is no uctuation in the junction voltage V due to the fast relaxation
time (CR
S
) of an external circuit, but there is a uctuation in the junction current I. This bias
condition is referred to as constant voltage operation. Standard theoretical studies on the noise
characteristics of a pn junction diode have considered this mode of operation; therefore, our analysis
also starts with this bias condition (Section 5.1).
On the other hand, some experimental studies on the noise of a pn junction diode have consid-
ered the opposite bias condition; that is, the source resistance R
S
is much larger than the dierential
resistance R
d
of the diode. In such a case there is no uctuation in the junction current I due to the
slow relaxation time (CR
S
) of an external circuit, but there is a uctuation in the junction voltage
V . This bias condition is referred to as constant current operation. This mode of operation will
be analyzed in Section 5.2.
There are two types of pn junctions which feature drastically dierent noise characteristics:
macroscopic junctions and mesoscopic junctions. When a single-electron charging energy (q
2
/2C)
3
(where C is the junction capacitance) is much smaller than the thermal characteristic energy
k
B
, the behavior of each individual electron does not aect the junction dynamics. This is a
macroscopic junction limit. On the other hand, when (q
2
/2C) is much greater than k
B
, a so-called
(single-electron) Coulomb blockade eect emerges and a previous single-electron event determines
the junction dynamics at a later time. This is a mesoscopic junction limit.
5.1 pn Junction Diodes Under Constant Voltage Operation
5.1.1 Current-Voltage and Capacitance-Voltage Characteristics of a pn Junction
The noise characteristics of a p
+
-N heterojunction with a heavily p-doped narrow bandgap
material and lightly n-doped wide bandgap material will be studied, rather than a conventional
p-n homojunction, because this specic junction structure is used in various semiconductor devices
(i.e., a double-heterostructure semiconductor laser and a heterojunction bipolar transistor). The
extension of the following analysis to a pn homojunction is straightforward.
The band diagram of a p
+
-N heterojunction diode at a zero bias condition (V = 0) and a
forward bias condition (V > 0) are shown in Fig. 5-2. The built-in potential V
D
is divided into the
potentials in the p
+
- and N-layers:
V
Dp
=
V
D
K
, (5.2)
V
Dn
= V
D
_
1
1
K
_
, (5.3)
where
K = 1 +

1
(N

A1
N
+
D1
)

2
(N
+
D2
N

A2
)
. (5.4)
4
N-layer
E
c
q
V
electron gas
+
+
+
+
+
+
x
+
+
x
electron gas
+
hole gas hole gas
x = -d x = 0
(v = 0)
x = -d x = 0
(v > 0)
E
V
q
V
Dn
V
Dp
V
Dn
- V
2
V
Dp
- V
1

F
p
+
- layer

Fp

Fn
Figure 5-2:
Here,
1
, N
A1
, and N
D1
are the dielectric constant, acceptor concentration, and donar concentration
of the p
+
-layer, and
2
, N
A2
, and N
D2
are those of the N-layer. If a p
+
-N diode satises
1
>
2
and N

A1
N
+
D1
N
+
D2
N

A2
, then K 1. Consequently, the built-in potential V
D
is mainly
supported in the N-layer, i.e., V
Dn
V
D
and V
Dp
0. The transmitted electron ux from the
N-layer to the p
+
-layer across the potential barrier height V
Dn
should be equal to the transmitted
electron ux from the p
+
-layer to the N-layer across the potential barrier E
c
/q because there is
no net current at V = 0.
When a forward bias (V > 0) is applied, only the potential barrier seen by the electrons in the
N-layer decreases to V
Dn
V
2
V
D
V , where the applied voltage supported in the N-layer is
V
2
= V
_
1
1
K
_
V . The electron (minority carrier) density at the edge of the depletion layer
5
(x = 0) in the p
+
-layer is given by
n
p
= Xn
N0
exp
_

V
D
V
V
T
_
= n
p0
exp
_
V
V
T
_
, (5.5)
where V
T
=
k
B

q
is the thermal voltage and
n
p0
= Xn
N0
exp
_

V
D
V
T
_
, (5.6)
is the thermal equilibrium electron density in the p
+
-layer. X is the transmission coecient of
an electron at the heterojunction interface and n
N0
is the electron (majority carrier) density at
the edge of the depletion layer (t = d) in the N-layer which is equal to the thermal equilibrium
electron density in the N-layer.
The excess electron density 5.5) at x = 0 diuses towards x = W, where a p-side metal contact
is located. The distribution of the excess electron density n(x, t) obeys

t
n(x, t) =
n(x, t) n
p0

1
q

x
i
n
(x, t) , (5.7)
where
n
is the electron (minority-carrier) lifetime and, since there is no electric eld in the neutral
p
+
-layer, the current i
n
(x, t) is carried only by a diusion component
i
n
(x, t) = qD
n

x
n(x, t) . (5.8)
Here, D
n
is the electron diusion constant. Solving (5.7) and (5.8) with the boundary conditions,
n
p
=
_

_
n
p0
exp
_
V
V
T
_
at x = 0
n
p0
at x L
n
, (5.9)
the steady-state solution for n(x) is now given by
n
p
(x) = n
p0
+ (n
p
n
p0
)e
x/Ln
, (5.10)
6
where L
n
=

D
n

n
is the electron diusion length. The junction current density is determined by
the diusion current (5.8) at x = 0:
i i
n
(x = 0) =
qD
n
L
n
(n
p
n
p0
) =
qD
n
n
p0
L
n
_
e
V
V
T
1
_
. (5.11)
The total current I = Ai vs. the junction voltage V is plotted in Fig. 5-3, where A is a cross-
sectional area.
10
-10
-2 -1
0 1 2
5
-5
I
L
n
AqD
n
n
p0
V/V
T
Figure 5-3:
The dierential resistance R
d
, dened by
_
dI
dV
_
1
, is approximately given by V
T
/I under a
strong forward bias condition. The diusion capacitance C
dif
of the diode is dened as the voltage
derivative of the (total) minority carrier charge:
C
dif

d
dV
Q
(minority carrier)
= A
d
dV
_
q
_

0
[n
p
(x) n
p0
]dx
_
=
AqL
n
n
p0
V
T
e
V
V
T

I
V
T

n
. (5.12)
7
The CR time constant characterized by the dierential resistance R
d
and the diusion capacitance
C
dif
is equal to the electron (minority carrier) lifetime
n
.
The depletion-layer capacitance C
dep
of the diode is dened as the voltage derivative of the
(total) space charge:
C
dep

d
dV
Q
(space charge)
=

2
W
dep
A
=

q
2
N
D2
2(V
D
V )
A (5.13)
Here, W
dep
=
_
2
2
qN
D2
(V
D
V ) is the depletion layer width in the N-layer. The capacitance con-
tributed by the depletion layer in the p
+
-layer is neglected. The CR time constant characterized
by the dierential resistance R
d
and the depletion layer capacitance C
dep
is equal to the thermionic
emission time
te
, which we will discuss later in this chapter.
It will be shown that the thermionic emission time
te
= C
dep
R
d
is a key parameter for deter-
mining the noise characteristics of a pn junction diode under weak forward bias, while the minority-
carrier lifetime
n
= C
dif
R
d
is a key parameter for determining the noise characteristics of a pn
junction diode under strong forward bias. The junction capacitance of a pn junction is determined
by the depletion-layer capacitance under weak forward bias and by the diusion capacitance under
strong forward bias.
5.1.2 Thermal Diusion Noise
When a pn junction is biased by a constant voltage source, the electron (minority carrier)
densities at x = 0 (edge of the depletion layer) and x = W (p-side metal contact) are constantly
8
xed to be n
p0
e
V/V
T
and n
p0
, respectively. The electron density uctuates, however, between
x = 0 and x = W due to microscopic electron motion induced by thermal agitation and electron
generation and recombination processes. In order to keep the boundary conditions at x = 0 and
x = W and to restore the steady-state electron distribution, the relaxation current pulse ows in
the entire p
+
-region between x = 0 and x = W. This relaxation current inside the p
+
-region results
in the departure from charge neutrality of this region, which induces the external circuit current.
If an electron makes a transit over a small distance
f
between collisions with the lattice, an
instantaneous current q(t) ows at the two locations x = x

and x = x

+
f
, as shown in Fig. 5-4.
This instantaneous current creates the departure from the steady-state electron distribution (5.10)
and triggers the relaxation current in the entire p
+
-region between x = 0 and x = W, which,
after a reasonably short time, restores the steady-state electron distribution (5.10). The electron
distribution deviation n

(x, t) = n(x, t) n
p0
(x), which results in such a relaxation current in the
entire region, satises the diusion equation (5.7) and the boundary conditions n

= 0 at x = 0
and x = W. The Fourier transform of the diusion equation (5.7) is given by

2
x
2
N

(j) =
1
L
2
N

(j) , (5.14)
where N

(j) is the Fourier transform of n

(x, t) and
L
2
=
L
2
n
1 +j
n
. (5.15)
The Fourier transform of the relaxation currents i

1
(t) at x = x

and i

2
(t) at x = x

+
f
are
expressed in terms of the Fourier-transformed electron density deviations N

1
(j) at x = x

and
9
q

(t)
x = 0 x = W
x' x' + l
f
initial current
relaxation current
diffusion current direct return current
0
N' (j

)
N'
1
N'
2
W
x
Figure 5-4:
N

2
(j) at x = x

+
f
:
I

1
(j) = qD
n
N

(j)
x

x=x
= k
1
N

1
(j) , (5.16)
I

2
(j) = qD
n
N

(j)
x

x=x

+
f
= k
2
N

2
(j) , (5.17)
where
k
1
=
qD
n
L
coth
_
x

L
_
, (5.18)
k
2
=
qD
n
L
coth
_
W x

L
_
. (5.19)
There are also direct return currents i

r1
(t) and i

r2
(t) between x

and x

+
f
, as shown in Fig. 5-4.
The Fourier-transformed return currents at x = x

and x = x

+
f
are identical and are expressed
by
I

r1
(j) = I

r2
(j) =
qD
n

f
_
N

1
(j) N

2
(j)
_
. (5.20)
10
Since there can be no accumulation of charge at any point in the entire p
+
-layer, one must have
current continuity at x = x

and x = x

+
f
:
I

1
(j) +I

r1
(j) +q = 0 , (5.21)
I

2
(j) +I

r2
(j) +q = 0 . (5.22)
From (5.21) and (5.22), one obtains
N

1
(j) =

f
D
n
k
1
k
1
+k
2
, (5.23)
N

2
(j) =

f
D
n
k
2
k
1
+k
2
. (5.24)
The Fourier-transformed electron density deviation N

(j) calculated by (5.14) with the boundary

conditions (5.23) and (5.24) is plotted in Fig. 5-4. The circuit current which actually ows in the
external circuit is determined by the two relaxation currents I

0
(j) at x = 0 and I

W
(j) at x = W:
I

T
(j) = I

0
(j) I

W
(j), (5.25)
where
I

0
(j) = qD
n

x
N

(j)

x=0
=

f
D
n
k
0
k
2
k
1
+k
2
, (5.26)
I

W
(j) = qD
n

x
N

(j)

x=W
=

f
D
n
k
W
k
1
k
1
+k
2
, (5.27)
k
0
=
qD
n
L
n
cosech
_
x

L
_
, (5.28)
k
W
=
qD
n
L
n
cosech
_
W x

L
_
. (5.29)
The reason the total external circuit current I

T
(j) is given by the dierence of the two relaxation
currents I

0
(j) and I

W
(j) is that this dierence creates the departure from the charge neutrality
11
in the entire p
+
-region [0, W], which should be compensated for by the external circuit current ow
in order to restore the charge neutrality. The external circuit current at the edge of junction x = 0
is actually carried by many events of forward and backward electron thermionic emission and can
be considered a continuous charging process.
Equation (5.25) is the Fourier transform of the circuit current pulse due to a single-electron
event in the p
+
-layer. The average number of thermal diusive transit events per second in a small
volume Ax (where A is the cross-section and x is the small distance along x) is given by

T
=
n(x)Ax

f
. (5.30)
Since each thermal diusive event occurs independently, the current uctuation power spectral
density due to such Poissonian random phase trains generated in this small volume is calculated
using the Carson theorem:
S
I

T
() = 2
T
|I

T
(j)|
2
=
2n(x)Ax

2
f
D
2
n

k
0
k
2
k
W
k
1
k
1
+k
2

2
=
4A
D
n
n(x)

k
0
k
2
k
W
k
1
k
1
+k
2

2
x . (5.31)
Here,
2
f
= 2D
n

f
is used. The total current uctuation power spectral density is given by inte-
grating this equation in the entire p
+
-layer:
S
I

T
() =
4A
D
n
_
W
0
n(x)

k
0
k
2
k
W
k
1
k
1
+k
2

2
dx

4Aq
2
D
n
L
n
_
n
p
n
p0
3
+
n
p0
2
_
. (5.32)
The second equality in the above expression is derived by assuming W L
n
and
n
1; that
is, the above expression is valid only for a long diode and a low-frequency uctuation component.
12
5.1.3 Generation-Recombination Noise
The initial action of this process is the instantaneous appearance or disappearance of an
electron. If an electron is generated at x = x

, an instantaneous current q(t) ows from nowhere

to x = x

, as shown in Fig. 5-5.

Solving the Fourier-transformed diusion equation (5.14) for the boundary conditions N

(j) =
0 at x = 0 and x = W and N

(j) = N

1
at x = x

, one obtains
N

(j) =
_

_
N

1
e
x/L
e
x

/L
(e
x/L
e
x/L
) (0 x x

)
N

1
e
(Wx

)/L
e
(Wx

)/L
[e
(Wx)/L
e
(Wx)/L
] (x

x W)
. (5.33)
The counter-propagating relaxation currents I

1
(j) and I

2
(j) at x = x

are now obtained as,

-q

(t)
initial current
relaxation current
0
N" (j

)
N"
1
W
x
x'
x'
Figure 5-5:
I

1
(j) = qD
n
N

(j)
x

x=x

0
= k
1
N

1
(j) , (5.34)
I

2
(j) = qD
n
N

(j)
x

x=x

+0
= k
2
N

1
(j) . (5.35)
13
The current continuity at x = x

imposes the following relation:

I

1
(j) I

2
(j) q = 0 . (5.36)
From this condition, one can determine the value of N

1
(j) as
N

1
(j) =
q
k
1
+k
2
. (5.37)
The Fourier-transformed electron density deviation is plotted in Fig. 5-5. The relaxation currents
at x = 0 and x = W are given by
I

0
(j) = qD
n
N

(j)
x

x=0
= q
k
0
k
1
+k
2
, (5.38)
I

W
(j) = qD
n
N

(j)
x

x=W
= q
k
W
k
1
+k
2
. (5.39)
The external circuit current is the dierence between (5.38) and (5.39):
I

T
(j) = q
_
k
0
+k
W
k
1
+k
2
_
. (5.40)
The average number of recombination events in a small volume Ax is given by

R
=
n(x)Ax

n
, (5.41)
while the average number of generation events is

G
=
n
p0
Ax

n
. (5.42)
Under the zero bias condition n(x) = n
p0
, the recombination rate is equal to the generation rate, as
it should be in a thermal equilibrium condition (detailed balance). The current uctuation power
14
spectral density due to the generation and recombination events in this small volume is
S
I

T
() = 2(
G
+
R
)

T
(j)

2
= 2
[n(x

) +n
p0
]Ax

n
q
2

k
0
+k
W
k
1
+k
2

2
. (5.43)
The total current uctuation power spectral density is calculated by integrating (5.43) from x = 0
to x = W:
S
I

T
() =
2Aq
2

n
_
W
0
[n(x

) +n
p0
]

k
0
+k
W
k
1
+k
2

2
dx

2Aq
2
D
n
L
n
_
n
p
n
p0
3
+n
p0
_
, (5.44)
where W L
n
(long-diode limit) and
n
1 (low-frequency limit) are used to derive the second
equality.
5.1.4 Total Current Noise
The total current uctuation power spectral density is the simple sum of (5.32) and (5.43)
S
I
T
() =
4Aq
2
D
n
L
n
_
n
p
n
p0
3
+
n
p0
2
_
+
4Aq
2
D
n
L
n
_
n
p
n
p0
6
+
n
p0
2
_

Thermal Diusion Noise Generation Recombination Noise
. (5.45)
The following three bias regions feature dierent noise characteristics:
(1) Zero Bias (V = 0)
In this case, n
p
= n
p0
, and thus (5.45) is simplied to
S
I
T
() =
4Aq
2
D
n
n
p0
L
n
=
4k
B

R
d
(V = 0)
, (5.46)
15
where R
d
(V = 0) =
Lnk
B

Aq
2
Dnn
p0
is the dierential resistance
_
dI
dV
_
1
at V = 0. Equation (5.46) is
the Johnson-Nyquist formula for thermal noise. This result is expected because the junction is in
thermal equilibrium at V = 0 and thus the Nyquist argument should be applied.
However, note that only one-half of (5.46) stems from standard thermal diusion noise and the
remaining half is due to generation-recombination noise. In this sense, a simple microscopic theory
of a thermal diusion process for a metallic conductor cannot describe the thermal equilibrium noise
of a pn junction. On the other hand, the Nyquist approach to thermal noise is very general; it does
not depend on the detailed microscopic process in a resistive element, but only requires the resistive
element be in thermal equilibrium with the environments. There are two environments for a pn
junction: lattice vibration (thermal phonon) reservoirs which are responsible for thermal diusion
noise and electromagnetic eld (thermal photon) reservoirs which are responsible for generation-
recombination noise. The Johnson-Nyquist formula still holds for the thermal equilibrium noise
of a pn junction at V = 0 even though the noise is equally contributed to by phonon reservoirs
(thermal diusion noise) and photon reservoirs (generation-recombination noise).
(2) Forward Bias (V > 0)
Equation (5.45), in this case, is reduced to
S
I
T
() =
2Aq
2
D
n
L
n
(n
p
+n
p0
) = 2q(I + 2I
S
) , (5.47)
where the forward current I and the (reverse) saturation current I
S
are given by
I =
AqD
n
L
n
(n
p
n
p0
) , (5.48)
I
S
=
AqD
n
L
n
n
p0
. (5.49)
16
In a reasonably high forward bias voltage, (5.47) is reduced to the full-shot noise 2qI since I I
S
.
Two-thirds of the full-shot noise is due to thermal diusion noise and one-third is due to generation-
recombination noise. The generation-recombination noise in this case is dominated by the radiative
recombination (spontaneous emission) noise for a direct bandgap semiconductor.
(3) Reverse Bias (V < 0)
In this case n
p
n
p0
, and thus (5.45) becomes
S
I
T
() =
2Aq
2
D
n
L
n
n
p0
= 2qI
S
. (5.50)
This result is often referred to as a dark current shot noise, which is the dominant noise source
of a reverse-biased photodiode made of a narrow bandgap semiconductor material. In this case,
one-third of the full-shot noise is due to thermal diusion noise and two-thirds is due to generation-
recombination noise. The generation-recombination noise in this case is dominated by the genera-
tion (thermal photon absorption) noise.
5.1.5 Short Diode
Thus far a so-called long diode with a bulk p
+
-layer thickness W much longer than the
diusion length L
n
has been studied. However, some pn junctions have a much thinner p
+
-layer
than the electron diusion length (W L
n
) (i.e., a double-heterostructure semiconductor laser
diode and heterojunction bipolar transistor).
Consider a N-p
+
-P double-heterostructure diode, as shown in Fig. 5-6. An injected electron
from the N-layer to the p
+
-layer cannot diuse freely toward the p-side metal contact due to
the conduction band discontinuity at the p
+
-P isotype heterojunction. A junction current is not
17
carried by a thermal diusion process, but crosses an imaginary plane between the conduction
and valence bands by a recombination process. The electron and hole densities are uniform in
N-layer
P-layer
electron gas
+
+
+
+
+
+
hole gas
electron (minority carrier)

F
p
+
- layer
Figure 5-6:
the p
+
-layer since W L
n
. The electron density in a relatively small bias voltage is given by
n
p
= n
p0
exp
_
V
V
T
_
. (5.51)
The junction current is related to the total electron number in the p
+
-layer, N
e
AWn, by
I = q
N
e

n
. (5.52)
The dierential resistance R
d
and diusion capacitance C
dif
are given by
R
d

_
dI
dV
_
1
=
V
T

n
qN
e
, (5.53)
18
C
dif

dQ
(minority carrier)
dV
=
qN
e
V
T
. (5.54)
The depletion layer capacitance C
dep
is given by (5.13).
Next, the current-uctuation power spectral density due to thermal diusion noise and
generation-recombination noise will be calculated. When W L
n
, one obtains
k
0
k
2
k
1
+k
2

k
W
k
1
k
1
+k
2

qD
n
W
, (5.55)
and from (5.26) and (5.27),
I

0
(j) I

W
(j) =
q
f
W
. (5.56)
At each boundary there is a uctuating current with the power spectral density
S
I

0
() = S
I

W
() =
_
W
0
2n
p
A

f
_
q
f
W
_
2
dx
= 4qI
_
L
n
W
_
2
, (5.57)
where D
n
=
2
f
/2
f
= L
2
n
/
n
is used. This current noise is much larger than the full-shot noise
since L
n
/W 1. However, as indicated in (5.56), the uctuating currents I

0
(j) and I

W
(j) are
identical and thus cancel out completely to nullify the total external circuit current uctuation,
I

T
(j) = I

0
(j) I

W
(j) = 0 . (5.58)
The thermal diusion current does not produce any departure from the charge neutrality in the
p
+
-region [0, W] and thus does not induce any external circuit current noise.
When W L
n
, one obtains
k
0
k
1
+k
2
1
x

W
, (5.59)
19
k
W
k
1
+k
2

W
. (5.60)
Using these relations in (5.38) and (5.39), one obtains
I

0
(j) = q
_
1
x

W
_
, (5.61)
I

W
(j) = q
x

W
, (5.62)
I

T
(j) = I

0
(j) I

W
(j) = q . (5.63)
Each event of electron generation and recombination results in a current pulse with a time-
integrated area equal to q in the external circuit and thus the low-frequency power spectral density
is given by
S
I

T
() = 2q
2
(N
e
+N
e0
)/
n
. (5.64)
This expression is reduced to the Johnson-Nyquist formula of thermal noise at V = 0 and the
Schottky formula of full-shot noise at V > 0 or V < 0:
(1) Zero-Bias (V = 0)
S
I

T
() =
4q
2
N
e0

n
=
4k
B

R
d
(V = 0)
. (5.65)
One half of this thermal noise is due to thermal photon absorption and the remaining half is due
to spontaneous emission.
(2) Forward-Bias (V > 0)
S
I

T
() = 2q
2
N
c

n
= 2qI . (5.66)
This full-shot noise is due solely to the radiative recombination (spontaneous emission) process.
20
(3) Reverse-Bias (V < 0)
S
I

T
() = 2q
2
N
c0

n
= 2qI
S
. (5.67)
This full-shot noise is due solely to the generation (thermal photon absorption) process.
5.2 pn Junction Diodes Under Constant Current Operation
5.2.1 Eect of a Finite Source Resistance
The origin of current noise in a constant-voltage-driven pn junction is the thermal diusive
transit between collisions with the lattice and the generation-recombination of a minority carrier
(an electron in the p
+
-layer in the previous case). These events introduce the relaxation current
in order to restore the steady-state electron distribution and cause the departure of the electron
density n
p
at the depletion layer edge in the p
+
-layer from the steady-state value. The temporal
decrease in n
p
results in excess forward thermionic emission from the N-layer to the p
+
-layer, as
compared to backward thermionic emission from the p
+
-layer to the N-layer. This excess forward
emission leads to the reduction of the electron density n
N
at the other depletion layer edge in the
N-layer. This departure of the electron density n
N
from the steady-state value is eliminated by a
majority-carrier ow in the N-layer and in the external circuit. On the other hand, the temporal
increase in n
p
results in excess backward thermionic emission from the p
+
-layer to the N-layer,
which induces the increased electron density n
N
and, thus, an opposite-polarity, external-circuit
current ows.
When the voltage source has an innitesimally small source resistance, the above relaxation
process is completed with negligible delay. A system does not memorize a previous event of thermal
21
diusive transit or generation and recombination of a minority carrier. Therefore, each event occurs
completely independently and this independence is the physical origin for the full-shot noise of an
external circuit current.
However, if the source resistance R
S
is not negligibly small, the modulation in the density
n
N
, induced by excess forward or backward thermionic emission of an electron, cannot be instanta-
neously eliminated. The junction voltage is now allowed to uctuate by the thermal diusive transit
and generation-recombination events. If the recombination events for electrons in the p
+
-layer ex-
ceed the average value, the junction voltage decreases due to excess forward thermionic emission of
electrons. While this junction-voltage decrease is not eliminated by the circuit relaxation current,
the forward thermionic emission rate temporarily decreases and results in the lower recombination
events of electrons in the p
+
-layer. This sequence is a self-feedback (stabilization) mechanism for
regulating the recombination process. At the same time, the external circuit relaxation current
pulse is smoothed due to a long-relaxation time constant CR
S
and the external current noise is
also reduced.
A noise equivalent circuit of such a pn junction is shown in Fig. 5-1, where C
dep
is the depletion
layer capacitance, C
dif
is the diusion capacitance, R
d
is the dierential resistance, i is the current
noise source associated with R
d
, R
S
is the source resistance, and i
s
is the current noise source
associated with R
S
. The Kirchho circuit equation for this noise equivalent circuit is:
d
dt
v
n
=
v
n
CR
S
+i
s

v
n
CR
d
+i , (5.68)
where v
n
is the junction voltage-uctuation and C = C
dep
+C
dif
is the total junction capacitance.
The rst term on the RHS of (5.68) is the relaxation (dissipation) rate of v
n
due to an external
22
circuit relaxation current and the second term is the Johnson-Nyquist thermal noise associated
with R
S
and its spectral density is S
is
() = 4k
B
/R
S
. The third term on the RHS of (5.68) is
the relaxation (dissipation) rate of v
n
due to thermionic emission of electrons across the depletion
layer and/or recombination of electrons and the fourth term is the noise current associated with
thermionic emission and/or recombination.
There are two operational modes of pn junction diodes, as illustrated in Fig. 5-7.
(1) R
S
R (Constant Voltage Operation)
The junction voltage uctuation induced by thermionic emission/recombination of electrons is
instantaneously suppressed by the circuit relaxation current pulse; thus v
n
0. The system does
not have a memory eect for thermal diusion and generation-recombination events of minority
carriers, so the noise current associated with these events features full-shot noise, S
i
() = 2qI, as
demonstrated in the previous section.
(2) R
S
R (Constant Current Operation)
The circuit current uctuation is smoothed by a slow relaxation current pulse due to a large
source resistance R
S
. The thermionic emission process becomes regulated by the self-feedback
mechanism mentioned above. Therefore, S
i
() is expected to feature a sub-shot-noise character,
but the junction voltage uctuation v
n
is allowed to uctuate.
5.2.2 A pn Junction Diode With Negligible Backward Thermionic Emission
Consider a P-p
+
-I-N double-heterojunction diode, as shown in Fig. 5-8. An undoped I-layer
has a thickness of x
u
and the depletion layer of an N-layer has a varying thickness x
n
(t). The
23
t
t
t
t
t
t
electron emission electron emission
V(t) V(t)
I(t) I(t)
q q
constant voltage operation constant current operation
Figure 5-7:
forward thermionic emission of an electron from the N-layer across the depletion layer has an
average rate of
(t) =
AT
2
A

q
exp
_

V
d
V (t)
V
T
_
, (5.69)
where A is an eective cross-section, A

is the Richardson constant, V (t) is the junction voltage,

and V
d
is an eective built-in potential given by
V
d
= V
D
+
q
2C
dep
. (5.70)
24
The second term on the RHS of (5.70) represents a single-electron charging energy which is
neglected in the previous section for a macroscopic pn junction driven by a constant voltage source.
This term simply osets an eective junction voltage by a small amount (q/2C
dep
) in the constant-
voltage bias case, but plays a crucial role in the constant-current bias case. Here, one assumes that
the forward bias voltage is very small and that the backward thermionic emission of an electron
from the p
+
-layer to the N-layer is negligible due to a high barrier height E
C
/q and a low electron
density in the p
+
-layer. This assumption is valid unless a very strong forward bias is applied to the
junction and makes the analysis much simpler. The eective potential barrier from the N-layer to
the p
+
-layer, V
d
V (t), is a function of the surface charge in the depletion layer:
V
d
V (t) =
qN
D

2
x
u
x
n
(t) +
qN
D
2
2
x
2
n
(t)

potential barrier Potential Barrier
in the I layer in the depleted N layer
. (5.71)
Here, N
D
is the ionized donar concentration in the N-layer.
There are two competing processes which change x
n
(t) and V (t): discrete thermionic emission
of an electron across the depletion layer from the N-layer to the p
+
-layer and continuous charging
via a constant circuit current I. The former decreases V (t) abruptly and the latter increases V (t)
continuously.
When a single electron is emitted from the N-layer to the p
+
-layer, both electron and hole gases
must pull back from the junction in order to satisfy the charge neutrality condition in the bulk
N- and p
+
-layers, which results in an increase in the surface charge in the depletion layer by +q
25
N-layer
P-layer
electron gas
+
+
+
+
hole gas
electron (minority carrier)
x
n
(t) x
u
x
I-layer
p
+
- layer
Figure 5-8:
in the N-side and q in the p
+
-side. The shift of the electron gas front by such a single-electron
thermionic emission event is
x
n
=
1
N
D
A
. (5.72)
The corresponding decrease in the junction voltage is
V =
qN
D

2
x
u
x
n
=
q
C
dep
, (5.73)
where the depletion layer capacitance is approximated by
C
dep


2
A
x
u
. (5.74)
In order for an electron to be thermionically emitted across the depletion layer, the electron must
have an excess energy q
2
/2C
dep
above the eective barrier potential because of the increase in
depletion layer width. The depletion layer width increases during an electrons transit across it.
26
The thermionic emission rate (t) is abruptly decreased by a single-electron event:
(t
+
) = (t

) exp
_

q
C
dep
V
T
_
= (t

) exp(r) , (5.75)
where
r =
(q
2
/C
dep
)
k
B

. (5.76)
The parameter r is the ratio of the single-electron charging energy and the characteristic energy of
thermal uctuation.
An external circuit current I pushes the electron and hole gases toward the junction continu-
ously:
d
dt
x
n
(t) =
I
qN
D
A
. (5.77)
The decrease in the depletion layer width x
n
(t) is therefore a linear function of t, and thus the
decrease in the eective potential barrier is also a linear function of time:
[V
d
V (t)] =
qN
D

2
x
u
x
n
(t)
=
I
C
dep
t , (5.78)
which results in the exponential increase of the thermionic emission rate
(t) = (0) exp
_
I
C
dep
V
T
t
_
= (0) exp
_
r

t
_
, (5.79)
where the parameter r is dened by (5.76) and is the single-electron charging time by the circuit
27
current:
=
e
I
. (5.80)
The time constant /r determines how quickly the thermionic emission rate increases and is termed
a thermionic emission time
te
. When (t) reaches 1/
te
at t = 0 by continuous charging, the
thermionic emission event is likely to occur by this time because
_
0

te
e
t

/te
dt

= 1. Since
(0)
te
= 1, this thermionic emission should occur in a short time interval
te
centered at t = 0,
as shown in Fig. 5-9. The thermionic emission time
te
is represented in terms of the dierential
resistance R
d
and the depletion-layer capacitance C
dep
:

te
=

r
= C
dep
R
d
. (5.81)
0
t
(t)
1/
te

te
Figure 5-9:
The thermionic emission rate, including the above two processes, is given by
(t) = (0) exp
_
t

te
rn
e
(t)
_
, (5.82)
where n
e
(t) is the number of electrons emitted from the N-layer into the p
+
-layer in a time interval
(0, t).
28
There are four distinct operational regimes for a constant-current-driven pn junction.
(1) Coulomb Blockade Regime or Mesoscopic Regime (r > 1)
When the single-electron charging energy q
2
/C
dep
is larger than the thermal energy k
B
and the
source resistance R
S
is larger than q/IC
dep
, the three time constants satisfy the following relations:
C
dep
R
S
>
q
I
> C
dep
R
d

Circuit Relaxation Single Electron Thermionic Emission
Time Charging Time Time
te
=

r
=
k
B
C
dep
qI
. (5.83)
In such a case, the junction voltage V (t) and the thermionic emission rate (t) oscillate with a time
interval equal to the single-electron charging time = q/I, as shown in Fig. 5-10.
t
t
V(t)
(t)
I
C
dep
t
q/C
dep
q
I
Figure 5-10:
In a time interval (0, ), the junction voltage V (t) increases linearly according to
I
C
dep
t due to
the constant current I and because the circuit relaxation time C
dep
R
S
is longer than the single-
electron charging time . Such a linear increase in the junction voltage results in an exponential
29
increase in the thermionic emission rate (t) exp(t/
te
). A single-electron thermionic emission
event occurs (on average) when (t) becomes equal to 1/
te
because
_

0
1

te
exp
_
t

te
_
dt

1 (
te
) . (5.84)
The probability for thermionic emission at t = 0 is negligibly small, (0) =

te
e

te
1. On the
other hand, the probability for thermionic emission between t =
te
2
and t = +
te
2
is close to
one, ()
te
= 1. Therefore, a single-electron thermionic emission event is well regulated at a xed
time t = , 2, 3, within a small jitter of
te
, as shown in Fig. 5-10.
The junction voltage oscillates at a frequency f = I/q and each electron is thermionically
emitted with a regulated time interval = q/I. This oscillatory behavior is termed a single-
electron thermionic emission (SETE) oscillation. When r > 1 is satised,
te
< is always
satised, irrespective of the current I, which is a unique feature of SETE oscillation in a pn
junction. In SETE oscillation, the upper and lower bounds on the current are imposed by
I >
q
C
dep
R
S
(Constant Current Operation) , (5.85)
I <
q
r
f
(Quasi Equilibrium Distribution) . (5.86)
This last condition is required because the charge distributions in both sides of the depletion layer
must reach the steady-state condition by collision with the lattice much faster than the thermionic
emission time
te
.
(2) Sub-Poisson Regime (r < 1, T
meas
>
te
)
When
q
2
C
dep
is smaller than k
B
, but R
S
> R
d
, the Coulomb blockade eect by a single electron
is negligibly small, but the pn junction is still driven by a constant-current source. The three time
30
constants satisfy the following relations:
C
dep
R
S
> C
dep
R
d
>
q
I
. (5.87)
In such a case, continuous charging of n
e
(=
1
r
> 1) electrons must be completed by the current I or
the same number of electrons must be thermionically emitted in order for the thermionic emission
rate (t) to be appreciably modulated, as shown in Fig. 5-11.
single
electron
charging
electron charging
I
C
dep
t
q/C
dep
V
T
=
k
B
q
q
I

te
t
V(t) V(t)
V
T
=
k
B
q
t
electron emission n
e
n
e
Figure 5-11:
In a real situation, the continuous charging and discrete emission of an electron occurs randomly.
In such a case, the collective eect of many electrons tend to regulate the thermionic emission
events in a macroscopic level; that is, more-than-average thermionic emission events are followed
by less-than-average thermionic emission events, and vice versa. Such a self-feedback mechanism is
referred to as a collective Coulomb blockade. A well-known example of such a collective Coulomb
blockade is the sub-shot-noise behavior of a space-charge-limited vacuum tube.
The probability of n
e
electrons emitted during a time interval T is given by
p(n
e
, T) =
1
N
n
ne
e
e
ne
n
e
!
exp
_

r
2
(n
e
n
e
)
2
_
, (5.88)
31
where N is the normalization constant and n
e
= T/ is the average number of emitted electrons.
The probability (5.88) is the product of the Poisson distribution with a variance
n
2
e
= n
e
, (5.89)
and the Gaussian distribution with a variance
n
2
e
=
1
r
=
k
B
q
2
C
dep
. (5.90)
Therefore, if T is longer than
te
, then n
e
=
T

becomes larger than

1
r
=
te

, and the probability

(5.88) features a sub-Poisson distribution. The average electron number n
e
is proportional to T,
while the variance n
2
e
, given by (5.90), remains constant. Therefore, the noise suppression from
the Poisson limit is improved with increasing T.
(3) Poisson Regime (r < 1, T <
te
)
On the other hand, if T is shorter than
te
, the probability (5.88) approaches a Poisson dis-
tribution because the Gaussian distribution becomes broader than the Poisson distribution. Even
though a junction is driven by a constant current source, the electron emission process is a random
Poisson-point-process for such a short time interval.
Figure 5-12 shows the variance n
2
e
for n
e
=
I
q
T as a function of the parameter 1/r. When
1/r > n
e
, the variance approaches n
e
(Poisson limit); on the other hand, when 1/r < n
e
, the
variance decreases linearly with 1/r. This process is a collective Coulomb blockade regime. Finally,
when 1/r < 1, the variance is suppressed to below a single electron. This is a single-electron
Coulomb blockade regime.
The power spectral density S
ne
() corresponding to the two regimes of collective Coulomb
32

n
e
2
n
e
0
0 1 n
e
Coulomb blockade regime
Sub-Poissonian regime
Poissonian regime
1
r
Figure 5-12:
blockade and single-electron Coulomb blockade are shown in Fig. 5-13. The power spectral density
is reduced to below the full-shot noise level at a frequency region below f = 1/2
te
. In the case
of r > 1, a coherent oscillation peak is observed at f =
1

due to SETE oscillation, while, in the

case of r < 1, a coherent oscillation peak is absent because each individual electron event is not
regulated.
1
1
Shot Noise Limit
Sn
e
( )
r < 1

te
10
-2
10
-1
10
-1
10
10
-1
1
10
-2
10
-1
1 10
Sn
e
( )

te
Shot Noise Limit
r > 1
-3dB
Figure 5-13:
(4) Macroscopic Regime (r 1,
n
)
Thus far the backward emission of electrons from the p
+
-layer to the N-layer has been neglected.
This approximation is valid when E
c
q
n
+k
B
, where
n
is a quasi-Fermi level for electrons
33
in the p
+
-layer. However, if the average electron density n
p
=
n
AL
becomes large enough and the
temperature is high enough, the above condition is no longer satised. Here,
n
is the electron
lifetime in the p
+
-layer. In such a case, the net thermionic emission rate (forward emission rate
backward emission rate) is not only dependent on the junction voltage V (t), but is also dependent
on the electron density n
p
.
For instance, one can think of the following two relaxation processes. Suppose that more-than-
average recombination events occur in the p-layer at a certain time, resulting in the decrease in
n
p
, which increases the net thermionic emission rate and decreases the junction voltage V . This
drop in V remains for a time interval
te
=

r
. Before the steady-state junction voltage V and the
electron density n
p
are recovered, the following recombination events are less than average. One
deterministic process (charging by a constant current) and two stochastic processes (thermionic
emission and recombination of electrons) are involved in this relaxation mechanism (self-feedback
stabilization for recombinatin events).
On the other hand, if recombination events become less than average at a certain time, this
results in an increase in n
p
, which decreases the net thermionic emission rate and increases V . The
increase in V remains for a relatively long time. During this time period, recombination events
are more than average. In both cases, the light intensity (proportional to the recombination rate)
and the junction voltage (inversely proportional to the net thermionic emission rate) are negatively
correlated. Such negative correlation is indeed observed in a sub-Poissonian light emitting pn
junction.
34
Contents
6. Bipolar Transistor 2
6.1 Current-voltage relationship . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
6.2 Current gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
6.3 Input vs. output characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
6.4 Current noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
6.5 Noise gure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
References 16
1
6. Bipolar Transistor
The bipolar transistor (transfer resistor) was invented in 1947. This device turned out to
be the most important semiconductor device in the electronic industry. It is now widely used in
high speed computer and communication systems. We will study the basic properties and noise
characteristics of a bipolar transistor in this chapter.
6.1 Current-voltage relationship
The basic structure of a p-n-p bipolar transistor is shown in Fig. 6-1(a), which consists of
a forward-biased emitter junction and reverse-biased collector junction. A bipolar transistor is
usually connected in two circuit congurations. Figure 6-1(b) and (c) show the common base and
common emitter congurations for a p-n-p bipolar transistor.
I
E
P
++
n
+
P
emitter base collector
I
B
I
C
I
C
I
E
+ -
I
B
(a) (b)
I
C
I
B
-
-
I
E
(c)
Figure 6-1: The basic structure (a) and two circuit congurations, common-base (b) and common
emitter (c), of a p-n-p bipolar transistor.
Consider the p-n-p bipolar transistor with uniform doping prole and in a common base con-
guration (Fig. 6-2). The continuity and current density equations in the neutral base region are
2
given by
O =
p p
B

B
+ D
B

2
p
x
2
, (6.1)
J
p
= qD
B
p
x
, (6.2)
J
n
= J
tot
+ qD
B
p
x
. (6.3)
The excess minority carrier densities at the edge of the emitter depletion layer are
p

(0) p(0) p
B
= p
B
_
exp
_
qV
EB
k
B
T
_
1
_
, (6.4)
n

(x
E
) n(x
E
) n
E
= n
E
_
exp
_
qV
EB
k
B
T
_
1
_
. (6.5)
A similar set of equations can be found for the collector junction:
p

(W) p(W) p
B
= p
B
_
exp
_
qV
CB
k
B
T
_
1
_
, (6.6)
n

(x
C
) n(x
C
) n
C
= n
C
_
exp
_
qV
CB
k
B
T
_
1
_
. (6.7)
The solutions for the minority carrier distributions in the base, emitter and collector regions are
easily obtained,
p(x) = p
B
+
_
p

(W) p

(0)e
W/L
B
2 sinh(W/L
B
)
_
e
x/L
B

_
p

(W) p

(0)e
W/L
B
2 sinh(W/L
B
)
_
e
x/L
B
, (6.8)
n(x) = n
E
+ n

(x
E
) exp
_
x + x
E
L
E
_
, (6.9)
n(x) = n
C
+ n

(x
C
) exp
_

x x
C
L
C
_
. (6.10)
From (6.2) and (6.3) we can obtain the total dc emitter current:
I
E
= A J
p
(x = 0) +A J
n
(x = x
E
)
3
= A
_
qD
B
p
x

x=0
_
+ A
_
qD
E
n
x

x=x
E
_
= Aq
D
B
p
B
L
B
coth
_
W
L
B
__
_
e
qV
EB
/k
B
T
1
_

1
cosh(W/L
B
)
_
e
qV
CB
/k
B
T
1
_
_
+Aq
D
E
n
E
L
E
_
e
qV
EB
/k
B
T
1
_
. (6.11)
Similarly we can obtain the total dc collector current:
I
C
= A J
p
(x = W) +A J
n
(x = x
C
)
= A
_
qD
B
p
x

x=W
_
+ A
_
qD
C
n
x

x=x
C
_
= Aq
D
B
p
B
L
B

1
sinh(W/L
B
)
_
_
e
qV
EB
/k
B
T
1
_
cosh
_
W
L
B
_
_
e
qV
CB
/k
B
T
1
_
_
+Aq
D
C
n
C
L
C
_
e
qV
CB
/k
B
T
1
_
. (6.12)
Here A is the cross-sectional area of the transistor. The dierence between these two currents
appears as the base current:
I
B
= I
E
I
C
. (6.13)
6.2 Current gain
The common-base current gain
0
(= h
FB
) is dened as

0
=
I
C
I
E
=
I
PE
I
E
I
PC
I
PE
I
C
I
PC
. (6.14)
The rst term
I
PE
I
E
is called as the emitter eciency , the second term
I
PC
I
PE
the base transport
factor
T
, and the third term
I
C
I
PC
the collector multiplication factor M. Since the transistor is
4
I
E
P
++
n
+
P
I
B
I
C
input
x
C
-x
E
0 W
N
+
- N
-
D A
V
EB
V
CB
E
C
E
V
output
Figure 6-2: The common-base conguration, doping prole and band diagram of the p-n-p bipolar
transistor.
normally operated well below the avalanche breakdown voltage, the multiplication factor is M 1
and so the static common-base current gain is given by

0

T
. (6.15)
The static common-emitter current gain
0
(= h
FE
) is dened as

0
=
I
C
I
B
=

0
1
0
, (6.16)
where we used (6.13).
Under the normal operating condition of a p-n-p bipolar transistor, V
EB
> 0 and V
CB
0, so
the terms in (6.11) and (6.12) associated with V
CB
can be neglected. The emitter eciency is
calculated from (6.11) as
=
A J
P
(x = 0)
I
E
=
_
1 +
n
E
D
E
L
B
p
B
D
B
L
E
tanh
_
W
L
B
__
1
. (6.17)
5
The base transport factor
T
is obtained from (6.11) and (6.12) as

T
=
J
P
(x = W)
J
P
(x = 0)
=
1
cosh(W/L
B
)
. (6.18)
For bipolar transistors with base width of one-tenth of the diusion length,
T
1
W
2
2L
2
B
= 0.995
and the current gain is given by the emitter eciency. Under this condition,

0
=

1
=
p
B
D
B
L
E
n
E
D
E
L
B
coth
_
W
L
B
_
. (6.19)
For a given emitter doping level, the current gain
0
increases with decreasing the base doping
level.
6.3 Input vs. output characteristics
For a bipolar transistor with high emitter eciency, the dc emitter and collector currents,
(6.11) and (6.12), reduce to the terms proportional to
p
x
at x = 0 and x = W, respectively. That
is, the emitter and collector currents are determined by the hole density gradients at the edges of
the base region. The base current is the dierence between the emitter and collector currents.
Figure 6-3(a) shows the input-output characteristics of the common-base conguration. The
collector current is practically equal to the emitter current, i.e.
0
1 and is independent of
V
CB
. This means
p
x
at x = 0 is equal to
p
x
at x = W for varying V
EB
and V
CB
, as shown in
Fig. 6-4(a) and (b). To reduce the collector current to zero, a forward bias voltage must be applied
to the collector, where the hole density at x = W becomes equal to the hole density at x = 0, as
shown in Fig. 6-4(c). The collector saturation current I
CO
with the emitter circuit open (I
E
= 0)
is considerably smaller than the ordinary reverse bias current of a p-n junction, because
p
x
= 0 at
6
x = 0, corresponding to I
E
= 0, reduces
p
x
at x = W as shown in Fig. 6-4(d). At a suciently
strong bias voltage V
CB
, the collector current starts to increase rapidly. This is either by the
avalanche breakdown eect or the punch-through eect. In the latter case the collector depletion
region reaches the emitter depletion region and a large direct current ows from the emitter to the
collector.
Figure 6-3(b) shows the input-output characteristics of the common-emitter conguration. The
current gain is much greater than one (
0
1). The saturation current I

CO
with the base circuit
open (I
B
= 0) is much larger than (I
CO
), since
I
B
= I
E
I
C
= I
E
(I
CO
+
0
I
E
) , (6.20)
and therefore
I

CO
= I
C
(I
B
= 0) =
I
CO
1
0
I
CO
. (6.21)
As V
CE
increases, the base width decreases and the current gain
0
increases. The lack of saturation
in the common-emitter output characteristics is called Early eect. When V
CE
decreases with a
constant I
B
, the collector junction is eventually forward-biased and the collector current becomes
zero (Fig. 6-4(c)).
6.4 Current noise
In the absence of emitter-base and collector-base depletion-layer recombination processes, the
current-voltage characteristics at the two junctions are governed by Schockleys diusion theory of
a pn junction. The transistor is called ideal in such a case. The emitter and collector current noise
7
Figure 6-3: Output characteristics for a p-n-p transistor in (a) common-base conguration, (b)
common-emitter conguration.
spectral densities are obtained from the analyses presented in Chapter 5 and written as
S
i
E
() =
4A
D
I
1
+
2q
2
A

R
I
2
, (6.22)
S
i
C
() =
4A
D
I
3
+
2q
2
A

R
I
4
, (6.23)
where the rst terms in (6.22) and (6.23) represent the thermal (diusion) noise of minority carriers,
holes, in the base region, and the second terms in (6.22) and (6.23) represent the generation-
recombination noise in the base region. The integrals I
j
(j = 1 to 4) take the same forms as those
derived in Chapter 5:
I
1
=
_
W
0
p

k
0
k
2
k
1
+ k
2

2
dx , (6.24)
I
2
=
_
W
0
(p + p
n
)

k
0
k
1
+ k
2

2
dx , (6.25)
I
3
=
_
W
0
p

k
1
k
W
k
1
+ k
2

2
dx , (6.26)
I
4
=
_
W
0
(p + p
n
)

k
W
k
1
+ k
2

2
dx . (6.27)
8
0 W
x
V
EB
> 0
V
EB
= 0
P
B
P(x)
0 W
x
V
CB
(small)
P
B
P(x)
V
CB
(large)
W'
0 W
x
V
CB
> 0
P
B
P(x)
V
CB
= 0
V
CB
< 0
0 W
x
V
EB
= 0
I = 0
E
P
B
P(x)
(a) (b)
(c) (d)
Figure 6-4: Hole density in the base region of a p-n-p bipolar transistor. (a) V
CB
= constant and
V
EB
varying. (b) V
EB
= constant and V
CB
varying. (c) V
CB
= forward, zero and reverse biased.
(d) zero emitter current and zero emitter voltage.
By evaluating the integrals (6.24)-(6.27), the spectral densities (6.22) and (6.23) reduce to the
forms,
S
i
E
() = 4qI
E
_
G
E
G
EO

1
2
_
, (6.28)
S
i
C
() = 2qI
C
. (6.29)
Here G
E
is the conductance of the forward-biased emitter-base junction and G
EO
is the low fre-
quency value of G
E
. At low frequencies, G
E
G
EO
and the forward-biased emitter current shows
full shot noise. The reverse-biased collector current features full shot noise at all frequencies.
Since the same noise generation mechanisms are responsible for the emitter and collector current
noise, some degree of correlation should exist between the two uctuations. When either thermal
9
diusion or generation-recombination event occurs in the base region, the relaxation current pulses,
f
E
(t) and f
C
(t), ow in the two junctions. The cross-spectral density between i
E
and i
C
is given
by the extended Carson theorem:
S
CE
() = 2a
2
F
E
(j)F
C
(j)

, (6.30)
where is the mean rate of the pulses, a (= q) is the pulse amplitude and F
E
(j) and F
C
(j) are
the Fourier transform of the relaxation current pulse shape functions f
E
(t) and f
C
(t). Using the
expressions for the two-types of noise currents of Chapter 5, (6.30) reduces to the form
S
CE
() =
4A
D
I
5

2q
2
A

R
I
6
, (6.31)
where
I
5
=
_
W
0
p
k
1
k
W
k

0
k

2
|k
1
+ k
2
|
2
dx , (6.32)
I
6
=
_
W
0
(p + p
n
)
k

0
k
W
|k
1
+ k
2
|
2
dx . (6.33)
By evaluating the integrals in (6.32) and (6.33), the cross-spectral density is expressed as
S
CE
() = 2qI
C

0
Y
E

00
G
EO
, (6.34)
where
0
is the common-base current gain,
00
is the low-frequency value of
0
, and Y
E
is the
admittance of the emitter-base junction. For low frequencies, Y
E
becomes equal to G
EO
and (6.34)
has the frequency independent form,
S
CE
( 0) = 2qI
C
. (6.35)
10
The normalized cross-spectral density between the emitter and collector currents is dened as

CE

S
CE
()
_
S
i
E
() S
i
C
()
_
1/2
. (6.36)
For low frequencies, S
i
E
() = 2qI
E
= 2qI
C
/
00
and hence

CE
=
1/2
00
. (6.37)
Since a modern bipolar transistor has
00
1, the uctuations in the emitter and collector currents
are highly correlated.
The power spectrum of the uctuations in the base current, I
B
= I
E
I
C
, is given by
S
i
B
() = S
i
E
() +S
i
C
() 2Re
_
S
CE
()
_
= 2qI
C
_
1

0
+
2G
E
(
0
Y
E
+

0
Y

E
)

00
G
EO

2(1
00
)

00
_
, (6.38)
where
0
= I
C
/I
B
is the common-emitter current gain. At low frequencies, S
i
B
() is reduced to
2qI
B
, which features the full shot noise.
The cross-spectral density between I
C
and I
B
is given by
S
CB
() = S
CE
() S
i
C
()
= 2qI
C
_
1

0
Y
E

00
G
EO
_
. (6.39)
The term
0
Y
E
is expanded to rst order in frequency:

0
Y
E

00
G
EO
_
1
j
B
3
_
, (6.40)
11
where
B
= W
2
/2D
B
is the base-layer charging time. Using (6.40) in (6.39), the cross-spectral
density and the normalized cross-spectral density are approximated by
S
CB
() = 2qI
C
_
j
B
3
_
, (6.41)

CB
()
S
CB
()
_
S
i
C
() S
i
B
()
_
1/2
=
1/2
0
_
j
B
3
_
. (6.42)
Similarly we obtain the cross-spectral density and the normalized cross-spectral density between
the emitter and base currents:
S
EB
() = S
i
E
() S
EC
()
= 2qI
C
_
2G
E

00
G
EO

00
+
1

0
+ 1

0
Y

00
G
EO
_
2qI
C
/
0
( 0) , (6.43)

EB

S
EB
()
_
S
i
E
() S
i
B
()
_
1/2

_

0
_
1/2
( 0) . (6.44)
6.5 Noise gure
For small signal condition the bipolar transistor is essentially a linear device and it is repre-
sented by an equivalent circuit of a linear two-port, as shown in Fig. 6-5. The admittance matrix
of the intrinsic transistor, in which the base resistance is negligible, is
Y =
_

_
1
r
BE
+
1
r
BC
+ j(C
BE
+ C
BC
)
1
r
BC
+jC
BC
g
m
+
1
r
BC
+ jC
BC
1
r
BC
+jC
BC
_

_
, (6.45)
where g
m
=
I
C
V
BE
is the mutual conductance of the transistor, r
BE
=

00
gm
and C
BE
= g
m

B
. In a
normal operating condition, the reverse-biased base-collector junction has the negligible admittance,
12
that is,
1
r
BC

1
r
BE
and C
BC
C
BE
. Therefore, the two components of admittance matrix reduce
to Y
11
g
m
_
1

00
+ j
B
_
and Y
21
= g
m
, respectively.
C B
E E
g
m
V
BE
C
BC
r
BC
r
BE
C
BE
Figure 6-5: An equivalent circuit of the bipolar transistor.
The base current noise i
B
and the collector current noise i
C
are added as the external noise
sources to the noiseless equivalent circuit, as shown in Fig. 6-6(a). The spectral densities and cross-
spectral densities of these current noise are given by (6.29), (6.38) and (6.41). For calculating the
noise gure of the transistor, it is convenient to transfer the output noise generator to the input.
We obtain the two new noise generators i
na
and v
na
at the input, as shown in Fig. 6-6(b). This
circuit is valid for calculating the output noise but not for calculating the input noise as discussed
in Chapter 3. The Fourier transform of the new noise generators are
V
na
=
I
C
Y
21
, (6.46)
I
na
= I
B

Y
11
Y
21
I
C
, (6.47)
where the Y parameters are given by (6.45). The spectral densities of the two generators are
S
Va
() 2qI
C
/G
2
EO
, (6.48)
S
Ia
() 2qI
B
+
4
3
qI
C

2
B
, (6.49)
13
and the cross-spectral density between them is
S
VaIa
() 2qI
C
_
1

00

2
3
j
B
_
/G
EO
. (6.50)
C B
E E
Y
i
B
i
C
(a)
C B
E E
Y
(b)
i
na
v
na
Figure 6-6: The noise equivalent circuits of the bipolar transistor.
The noise gure of the transistor is the ratio of total noise power at the output to the noise
power resulting from thermal noise in source resistance and is calculated from the circuit shown in
Fig. 6-7. Here i
s
is the (input) signal current generator, Y
s
= Gs+jBs is the source admittance, i
ns
is the source resistance thermal noise with spectral density equal to 4k
B
TGs, i
nb
is the equivalent
noise current generator to v
na
in Fig. 6-6(b). The Fourier transform of this new noise current is
I
nb
= Y
s
V
na
. (6.51)
The noise gure is calculated from (6.48) - (6.50) as,
F = 1 +
2qI
C
4k
B
TGs
_
_
Bs
G
EO
+
2
3

B
_
2
+
2
9

2
B
+
1

0
+
G
2
S
G
2
EO
+
2Gs
G
EO

00
_
. (6.52)
The noise tuning condition is satised when the source susceptance Bs is equal to
Bs =
2
3

B
G
EO
. (6.53)
14
C
E
Y
i
s
i
ns
Y
s
i
na
i
nb
B
E
Figure 6-7: A noise equivalent circuit of the transistor including the source admittance Y
s
and
source resistance thermal noise current i
ns
.
The optimum source conductance which minimizes the noise gure is obtained by the condition,
F
Gs
= 0, and is given by
Gs = G
EO
_
1

0
+
2
9

2
B
_
1/2
. (6.54)
If we substitute (6.53) and (6.54) into (6.52), we have the minimum noise gure
F
min
= 1 +
_
1

0
+
2
9

2
B
_
1/2
, (6.55)
where G
EO
qI
C
/k
B
T is used and 1/
00
1 is neglected. At low frequencies, F
min
reduces to
1+

1
2
0
, which is 0.4 dB for = 100. At high frequencies, F
min
increases with increasing frequency.
In particular, F
min
is proportional to
2
at medium frequencies where 9/(2
0

2
B
).
The above analysis does not include the nite base resistance and the excess noise caused by
the generation-recombination process in the two depletion layers, but it is in good agreement with
the observed noise behavior of the transistor. This means that a modern bipolar transistor comes
close to this ideal model.
15
References
[1] S. M. Sze, Physics of Semiconductor Devices (Wiley, New York, 1981) Ch. 3.
[2] M. J. Buckingham, Noise in Electronic Devices and Systems (Wiley, New York, 1983) Ch. 4.
[3] A. van der Ziel, Noise in Solid State Devices and Circuits (Wiley, New York, 1986) Ch. 9.
16
Contents
7. 1/f Noise 2
7.1 Characteristics of 1/f noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
7.1.1 Scale invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
7.1.2 Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
7.2 Mathematical model of 1/f noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
7.2.1 A random pulse train model of 1/f noise . . . . . . . . . . . . . . . . . . . . . 4
7.2.2 Superposition of relaxation processes . . . . . . . . . . . . . . . . . . . . . . . 6
7.3 Physical mechanisms of 1/f noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
References 10
1
7. 1/f Noise
In practically all electronic and optical devices, the excess noise obeying the inverse frequency
power law exists in addition to intrinsic thermal and quantum noise. An enormous amount of
experimental data has been accumulated on 1/f noise in various materials and systems. However, a
physical mechanism for 1/f noise has not been identied clearly yet. We have several mathematical
models that lead to the 1/f power law but shed little light on the physical mechanism for the noise.
7.1 Characteristics of 1/f noise
7.1.1 Scale invariance
A 1/f noise form, x(t), is characterized by a power spectral density function:
S
x
() = C/ , (7.1)
where C is a constant. The integrated power in the spectrum between
1
and
2
is given by
P
x
(
1
,
2
) =
1
2
_

2

2
S
x
()d
=
C
2
ln
_

1
_
. (7.2)
This result shows that for a xed frequency ratio
2
/
1
, the integrated noise power is constant.
Thus the total noise powers in between any decade of frequency, say 0.1 Hz to 1 Hz or 1 Hz to 10
Hz or 10 Hz to 100 Hz, are identical. This property of 1/f noise is known as scale invariance.
2
7.1.2 Stationarity
Consider a 1/f noise, x(t), has the band-pass ltered power spectral density,
S
x
() =
_

_
C/ for
1

2
0 otherwise
. (7.3)
The autocorrelation function of x(t) is obtained by using the Wiener-Khintchine theorem,

x
() =
C
2
_

2

1
cos

d
=
C
2
[C
i
(
2
) C
i
(
1
)] , (7.4)
where
C
i
(z) =
_
z

cos y
y
dy , (7.5)
is the cosine integral. The series expansion of C
i
(z) is
C
i
(z) = + ln(z) +

k=1
(1)
k
z
2k
(2k)!2k
, (7.6)
where = 0.5772 is Eulers constant. Thus, in the limit of z 0, the cosine integral reduces to
C
i
(z) lnz. The mean-square of x(t) is thus given by

x
( = 0) =
C
2
ln
_

1
_
. (7.7)
It is evident from the above argument that the band-pass ltered 1/f noise is statistically stationary
because it has the second-order quantities depend only on the delay time and not on the absolute
time at which the ensemble average is performed.
However, there is no experimental evidence for the existence of the low frequency limit
1
for
1/f noise. The reason is that an observation time T is always nite in practice and so a lower
3
frequency region of the spectrum,
2
T
, cannot be observed. The autocorrelation function and
the mean square value that can be measured in actual experiments are thus given by replacing the
low frequency limit
1
with 2/T in (7.4) and (7.7).
The Wiener-Levy process, discussed in Chapter 2, is a cumulative process of random walk. The
power spectrum obeys 1/
2
law. By the very nature of the process, the Wiener-Levy process is
statistically nonstationary and there is no possibility for the low-frequency limit
1
to exist. The
corner (roll-o) frequency in the calculated spectrum (Chapter 2) is an artifact associated with
the nite gate time T. In the case of 1/f noise, however, a low-frequency limit
1
may or may not
exist. The stationarity of the process is still open to question.
7.2 Mathematical model of 1/f noise
7.2.1 A random pulse train model of 1/f noise
The power spectral density of a random pulse train, x(t), is given by Carsons theorem,
S
x
() = 2a
2
|F(j)|
2
, (7.8)
where F(j) is the Fourier transform of the pulse shape function, is the mean rate of the pulses
and a
2
is the mean-square value of the pulse height. Thus the frequency dependence of S
x
() is
entirely determined by f(t). Consider the ctitious pulse shape function,
f(t) = u(t)t
(1

2
)
e
xt
, (7.9)
4
where and
x
are positive and independent of time, and u(t) is the unit step function. The
Fourier transform of f(t) is
F(j) =
_

0
t
(1

2
)
e
(x+j)t
dt
=

_

2
_
(
x
+ j)

2
, (7.10)
where (x) is the gamma function. Using (7.10) in (7.8), we obtain
S
x
() =
2a
2

2
_

2
_
(
2
x
+
2
)

2
. (7.11)
It shows the approximate 1/f noise characteristic at
x
,
S
x
() C/

, (7.12)
when 1. Here C = 2a
2

2
_

2
_
. However small
x
may be, provided it is non-zero, S
x
() has
the at spectrum at
x
, which indicates such a random pulse train is statistically stationary.
The autocorrelation function of this random pulse train is given by

x
() =
1
2
_

0
S
x
() cos()d
=
C
2
K
0
(
x
) , (7.13)
where = 1 and K
0
(z) is the modied Bessel function of the second kind of zero order. The series
expansion of K
0
(z) is
K
0
(z) = + ln2 ln(z) + . (7.14)
For small
x
,
x
() varies as ln(
x
) and takes a nite value except at the origin = 0. This
logarithmic innity is associated with the innite extension of the 1/f noise spectrum to high
5
frequencies, which is of course unrealistic because any nite response time in a system introduces a
cut-o characteristic beyond the certain frequency and the spectrum usually rolls o with a 1/
2
dependence. Figure 7-1(a) and (b) show schematically S
x
() and
x
() of this random pulse train.
Figure 7-1: (a) S
x
(), normalized by S
x
(0), vs. /
x
. (b)
x
(), normalized by s/2, vs.
x
.
It is clear from the above argument that a random pulse train in which the pulse shape varies
as t

1
2
shows the 1/f noise behavior. However, the physical origins of such a pulse shape are not
appearent.
7.2.2 Superposition of relaxation processes
If the noisy form z(t) is a random relaxation process with a time constant
z
, the power
spectral density has the general form:
S
z
() =
g(
z
)
1 +
2

2
z
, (7.15)
where g(
z
) depends on the physical mechanism of the noise. Suppose a linear superposition, x(t),
is constructed from such relaxation processes, whose decay constants are distributed between upper
and lower limits
1
and
2
with a probability density p(
z
). The overall power spectral density is
6
then
S
x
() =
_

2

1
S
z
()p(
z
)d
z
=
_

2

1
p(
z
)g(
z
)
(1 +
2

2
z
)
d
z
. (7.16)
If the numerator p(
z
)g(
z
) is independent of
z
and equal to a constant P, the above integral
reduces to
S
x
() = P
_
tan
1
(
2
) tan
1
(
1
)
_
/ . (7.17)
When the two time constants
2
and
1
satisfy
2
1 and 0
1
1, respectively, the two
terms in the numerator are approximately equal to /2 and 0. Thus, a superposition of relaxation
processes can give rise to a spectrum
S
x
() =
_
P
2
_
/ . (7.18)
If the product p(
z
)g(
z
) is proportional to
1
z
, the power spectral density has a more general
form of

.
7.3 Physical mechanisms of 1/f noise
The most popular model of 1/f noise is the trapping model with a wide spread of time
constants. If a free carrier is immobilized by falling into a trap, it is no longer available for current
transport. The modulation of carrier numbers has the form of random telegraph signal with a
Poisson point process as shown in Fig. 7-2. The probability of observing m telegraphic signals in
the time interval T is given by
p(m, T) =
(T)
m
m!
e
T
, (7.19)
7
Figure 7-2: A random telegraph signal produced by a carrier trap.
where is the mean rate of transitions per second. If
+
and

are the average times spent in the

upper and lower states, respectively, the probability distributions of the upper t
+
and lower state
times t

are
p(t

) =
1

exp
_

_
. (7.20)
The product x(t)x(t + ) is equal to +a
2
if an even number of transitions occur in the interval
(t, t + ) and to a
2
if an odd number of transitions occur in the same interval. Therefore, the
autocorrelation function is

(x) = a
2
[p(0, ) + p(2, ) + ]
a
2
[p(1, ) + p(3, ) + ]
= a
2
e

_
1 +
()
2
2!

()
3
3!
+
_
= a
2
e
2
. (7.21)
The power spectrum is thus calculated by the Wiener-Khintchine theorem,
S
x
() = 4
_

0

x
() cos()d
=
2a
2
/
(1 +
2
/4
2
)
8
= a
2
4
z
(1 +
2

2
z
)
. (7.22)
Here
z
= 1/2 is the time constant of the trap. If
z
is distributed according to the function p(
z
),
the power spectral density of the total carrier number uctuation is
S
n
() = 4
n
( = 0)
_

0

z
p(
z
)
(1 +
2

2
z
)
d
z
. (7.23)
Here it is assumed
_

0
p(
z
)dz = 1.
Suppose the carrier trap occurs by the tunneling of carriers from semiconductors to the traps
inside the oxide layer at depth w, the time constant obeys

z
=
0
exp(w) , (7.24)
where
0
and are constants. If the traps are homogeneously distributed between the depth w
1
and w
2
, corresponding to the time constants
1
and
2
, we obtain
p(
z
)d
z
=
_

_
dz/z
ln(
2
/
1
)
(
1

z

2
)
0 (otherwise)
. (7.25)
Using (7.25) in (7.23), the power spectral density of the total number uctuation is given by
S
n
() =
4
n
(0)
ln(
2
/
1
)
_

2

1
d
z
(1 +
2

2
z
)
=
4
n
(0)
ln(
2
/
1
)

tan
1
(
2
) tan
1
(
1
)

. (7.26)
As we discussed before, (7.26) shows 1/f power law in the frequency range of
2
1 and 0

1
1.
The above argument applies also for the intrinsic bulk transport property of the hopping con-
duction. The essential requirement to obtain the 1/f power law is the Poissonian telegraphic event
with a distributed time constant which obeys 1/
z
distribution function.
9
The physical model presented above falls well short of a universal description of the 1/f noise.
The other proposed physical models include a temperature uctuation model and a quantum me-
chanical model. In the latter model, the ubiquity of 1/f noise is attributed to the fundamental
interaction between charge carriers and quantized electromagnetic eld.
References
[1] J. B. Johnson, The Shottky eect in low frequency circuits, Phys. Rev. 26, 71 (1925).
[2] D. A. Bell, A survey of 1/f noise in electrical conductors, J. Phys. C. 13, 4425 (1980).
[3] P. H. Handel, Quantum theory of 1/f noise, Phys. Lett. 53A, 438 (1975).
[4] M. B. Weissman, Survey of recent 1/f noise theories, Proc. 6th Int. Conf. on Noise in Physical
Systems (Gaithersburg, 1981) p. 133.
10
Contents
8. Tunnel Junction Diodes 2
8.1 Tunneling probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
8.2 Current-voltage characteristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
8.3 Noise of macroscopic tunnel junction diodes . . . . . . . . . . . . . . . . . . . . . . . 7
8.4 Noise of mesoscopic tunnel junction diodes . . . . . . . . . . . . . . . . . . . . . . . . 9
8.5 Coulomb staircase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
8.6 Turnstile devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1
8. Tunnel Junction Diodes
If a pn junction consists of heavily doped p and n type semiconductor (p, n 10
19
cm
3
), a
depletion layer width becomes less than 100

A and a built-in eld becomes more than 10
6
V/cm.
An electron wavefunction penetrates into such a narrow depletion layer, so a quantum mechanical
tunneling probability from the conduction band of the n-layer to the balance band of the p-layer
and that in the opposite direction are nite, as shown in Fig. 8-1.
Figure 8-1:
The current-voltage characteristic of such a tunnel junction is shown in Fig. 8-2, together with
the junction states at each bias point. At a reverse bias voltage (A), an electron tunnels from the
occupied state of the valance band of the p-layer to the empty state of the conduction band of
the n-layer (Zenar breakdown or tunneling breakdown). At a zero-bias voltege (B), the two tunnel
currents cancel out with each other (detailed balance). At a peak (forward) voltage V
P
(C), the
forward tunnel current from the occupied state of the conduction band of the n-layer to the empty
state of the valence band of the p-layer is maximum. At a valley (forward) voltage V
V
(D), there
2
is no tunnel current because no nal states are available. At an even higher forward voltage (E), a
normal thermal diusion current ows across the depletion layer.
Figure 8-2:
There are two types of tunneling processes as shown in Fig. 8-3. If p and n layers consist of direct
gap semiconductors, the conduction band electron at point of the n-layer directly tunnels into the
valance band hole at the same point of the p-layer. A wavenumber is conserved in this direct
tunneling process. On the other band, if p and n layers consist of indirect gap semiconductors, the
conduction band electron at L point of the n-layer tunnels into the valance band hole at point of
the p-layer only via phonon scattering. Otherwise, the wavenumber conservation is not satised.
This is an indirect tunneling process. A phonon assisted indirect tunneling is a much weaker
process than a direct tunneling.
3
Figure 8-3:
8.1 Tunneling probability
An electron tunneling through a forbidden band is mathematically equivalent to a particle
tunneling through a potential barrier. Suppose an electron is incident upon the forbidden band
from the valance band of the p-layer (the situation shown in Fig. 8-2 (A)). The potential barrier
seen by an electron when it traverses from the edge of the p-layer to the edge of the n-layer is shown
in Fig. 8-4. The electron wavefunction [[
2
(evanescent wave) penetrates into the forbidden gap
(x
1
< x < x
2
) and there is a nite probability to reach through the conduction band of the
n-layer. This tunneling probability T
t
is given by the WKB approximation:
T
t
= exp
_
2
_
x
2
x
1
[(x)[dx
_
, (8.1)
where (x) is the wave number in the potential barrier. In the model of Fig. 8-4, we have
(x) =

2m

h
2
E
g
_
x
2
x
x
2
+x
1
_
, (8.2)
4
Figure 8-4:
and
T
t
exp
_
_

2m

E
3
2
g
3qh
_
_
= exp
_

4
_
2m

E
g
(x
1
+x
2
)
3h
_
. (8.3)
Here c =
Eg
q(x
1
+x
2
)
is the electric eld in the depletion layer which is a function of a bias voltage. A
small eective mass m

, small bandgap E
g
and large electric eld c (or small depletion layer width
x
1
+x
2
) increase the tunneling probability.
8.2 Current-voltage characteristic
The current observed in an external circuit is the dierence of the forward and backward
tunnel currents:
I = I
CV
I
V C
, (8.4)
where
I
CV
= A
_
E
V
E
C
F
C
(E)n
C
(E)T
t
[1 F
V
(E +qV )]n
V
(E +qV )dE , (8.5)
I
V C
= A
_
E
V
E
C
F
V
(E +qV )n
V
(E +qV )T
t
[1 F
C
(E)]n
C
(E)dE . (8.6)
5
Here A is a constant, n
C
(E) and F
C
(E) are the density of states of the conduction band of the
n-layer and the Fermi-Dirac distribution function:
n
C
(E) =

2(m

)
3
2

2
h
3
_
E E
C
, (8.7)
F
C
(E) =
1
1 + exp
_
EE
F
k
b
T
_ . (8.8)
V is an applied voltage. n
V
(E) and F
V
(E) are the density of states of the valence band of the
p-layer and corresponding Fermi-Dirac distribution function. The integration of (8.5) and (8.6)
should be performed from the conduction band minimum E
C
of the n-layer to the valence band
maximum E
V
of the p-layer.
The net current I is now calculates as
I I
CV
I
V C
= A
_
E
V
E
C
[F
C
(E) F
V
(E +qV )]T
t
n
C
(E)n
V
(E +qV )dE
I
p
_
V
V
p
_
exp
_
1
V
V
P
_
. (8.9)
Here the peak voltage V
P
is given by
V
p

1
3
(V
en
+V
hp
) , (8.10)
V
en
= V
T
_
ln
_
N
D
N
C
_
+ 0.35
_
N
D
N
C
__
, (8.11)
V
hp
= V
T
_
ln
_
N
A
N
V
_
+ 0.35
_
N
A
N
V
__
, (8.12)
V
en
and V
hp
are the Fermi-level
F
for conduction electrons measured from the conduction band
minimum potential E
C
/q and the corresponding quantity for valence holes. N
D
and N
A
are the
donar concentration in the n-layer and the acceptor concentration in the p-layer. N
C
and N
V
are
6
the eective density of states in the conduction band of the n-layer and that in the valence band
of the p-layer:
N
C
= 2
_
2m

e
k
B
T
h
2
_3
2
, (8.13)
N
V
= 2
_
2m

h
k
B
T
h
2
_3
2
. (8.14)
The forward current initially increases linearly (proportional to
V
V
P
) and saturates at V = V
P
. At
a higher bias voltage, the forward current decreases exponentially ( exp
_

V V
P
V
P
_
) and has a
minimum at V
V
= V
en
+V
hp
.
8.3 Noise of macroscopic tunnel junction diodes
If a macroscopic tunnel junction diode is driven by a constant voltage source, the two currents
I
CV
and I
V C
independently carry a full shot noise. Therefore, the total current noise is given
by
S
I
() = 2q(I
CV
+I
V C
) . (8.15)
The Fermi-Dirac distribution functions in the integrals of (8.5) and (8.6) satisfy the following
relation:
F
C
(E)[1 F
V
(E +qV )] = F
V
(E +qV )[1 F
C
(E)] exp
_
qV
k
B
T
_
, (8.16)
so that the two currents are related by
I
CV
= I
V C
exp
_
qV
k
B
T
_
. (8.17)
Using (8.4) and (8.17) in (8.15), we have
S
I
() = 2qI coth
_
qV
2k
B
T
_
. (8.18)
7
At a zero-bias voltage (V = 0), (8.18) is reduced to
lim
V 0
S
I
() = 2qI
2k
B
T
qV
=
4k
B
T
R(V = 0)
, (8.19)
where R(V = 0) = lim
V 0
V
I
is a dierential resistance of the tunnel diode in the limit of V 0.
This is the Johnson-Nyquist formula of thermal noise. This noise current is carried by a quantum
mechanical tunneling of electrons across the depletion layer but still satisfy the Johnson-Nyquist
formula. This is an expected result, because the tunnel junction is in a thermal equilibrium state
with reservoirs and the above formula should be satised irrespective of a microscopic mechanism
of noise generation (Nyquist theory). A direct physical origin for the Johnson-Nyquist formula in
this case is a Fermi-Diract distribution of conduction electrons and valence holes. The noise current
increases with increasing a temperature, since the energy region of partially occupied states which
contributes to the noise current increases with a temperature. In a short pn diode, a direct physical
origin for the Johnson-Nyquist thermal noise at V = 0 is a Bose-Einstein distribution of thermal
radiation elds (photons) in a generation-recombination process. In a long pn diode, one-half of the
Johnson-Nyquist thermal noise at V = 0 is carried by the diusion process. In this case, a direct
physical origin is a thermal distribution of Brownian particles velocity (equipartition theorem).
At a forward voltage V V
T
, we have
I
CV
= I
V C
exp
_
qV
k
B
T
_
I
V C
, (8.20)
S
I
() 2qI . (8.21)
This is the Schotlkys formula of full shot noise.
8
On the other hand, when the junction is driven by a constant current source, R
S
R
d
=
dV
dI
,
a macroscopic Coulomb blockade self-regulates the forward tunneling process to below the shot
noise value in a similar way as a macroscopic normal pn junction diode driven by a constant current
source.
8.4 Noise of mesoscopic tunnel junction diodes
Let us consider a metal-insulator-metal tunnel junction diode. When an external voltage V
is applied, the forward tunneling rate is given by
r(Q) =
_
1
(E)
n
R
(E)n
L
_
E +qV
q
2
2C
_
F(E)
_
1 F
_
E +qV
q
2
2C
__
dE , (8.22)
where (E) is an elastic tunneling lifetime, n
R
and n
L
are the densities of states of the right and
left electrodes and F(E) =
1
1+exp(E/k
B
T)
is a Fermi-Dirac distribution function. The integral (8.22)
is non-zero in the energy region

qV
q
2
2C

around the Fermi energy E

F
. In such a small energy
range, (E) and n
R
(E), n
L
(E) are slowly varying functions of E and so they can be approximated
by the values at E
F
. Then we have
r(Q) =
n
R
(E
F
)n
L
(E
F
)
(E
F
)
_
F(E)
_
1 F
_
E +qV
q
2
2C
__
dE
=
n
R
(E
F
)n
L
(E
F
)
(E
F
)

qV
q
2
2C
1 exp
_

qV
q
2
2C
k
B
T
_
=
1
qCR
T

Q
q
2
1 exp
_

Q
q
2
CV
T
_ . (8.23)
Here R
T
is a tunnel resistance given by
R
T
=
(E
F
)
q
2
n
R
(E
F
)n
L
(E
F
)
. (8.24)
9
When
q
2
C
k
B
T, (8.23) is reduced to
r(Q) =
_

_
0 : Q <
q
2
Q
q
2
qCR
T
: otherwise
. (8.25)
Figure 8-5 shows r(Q) vs. Q for T = 0 and T ,= 0. A mean lifetime of tunneling at a bias voltage
0
0 0.5 1
1
T = 0
T = 0
Q/q
2CR
T
r(Q)
Figure 8-5:
V is given by
tunnel

1
r(Q)
= CR
T

1
Q
q
2
when Q >
q
2
. When Q <
q
2
, a tunneling lifetime is innite
(Coulomb blockade). The backward tunneling rate is similarly obtained as
(Q) =
1
qCR
T

Q+
q
2
exp
_
Q+
q
2
CV
T
_
1

_
q
2
C
k
B
T
_
_

_
0 : Q >
q
2

Q+
q
2
qCR
T
: otherwise
. (8.26)
A current-voltage characteristic of such a mesoscopic M-I-M tunnel junction is shown in Fig. 8-
6. When the junction is driven by a constant voltage source, i.e. R
S
R
T
, the mean lifeitme of a
single electron tunneling is given by
= CR
T
_
q
CV
q
2
+
R
S
R
T
ln
1
CV
q
2
_
. (8.27)
The rst term on the right corresponds to the Coulomb blockade eect represented by (8.25) and
the second term is the rst-order correction factor due to a nite source resistance R
S
. The average
10
V V
V V
t t
t t
(A) (B)
(D) (C)
0
0 0.5 1
1
constant current
operation
constant voltage
operation
V
q
C
2CR
T
q
I
A
B
C
D
Figure 8-6:
current is given by
I =
q

=
V
q
2C
R
T
+
_
q
C
_
R
S
R
T
ln
_
q
C
V
q
2C
_
R
T

_
R
S
R
T
0
_
V
q
2C
R
T
. (8.28)
The I-V characteristic of a constant-voltage-driven mesoscopic M-I-M tunnel junction is dierent
from that of a macroscopic M-I-M tunnel junction only by a dc voltage oset V
q
2C
.
When the same junction is driven by a constant current source, i.e. R
S
R
T
, the mean lifetime
11
is given by
=
q
I
_
1 +

2
R
T
R
S
q
CV
_
1
q
2CV
_
_
, (8.29)
where I =
Vex
R
S
is a dc current and V
ex
is an external voltage. The rst term on the right corresponds
to the single electron tunneling (SET) oscillation with a period of = q/I. The second term is the
rst-order correction factor due to a nite source resistance R
S
. The I-V characteristic in the limit
of R
T
/R
S
0 is dierent from that of a constant-voltage case (8.28). The time average voltage
waveform at each bias point is also shown in Fig. 8-6. At a point (A), the current I is small enough
to satisfy
q
I
CR
T
. In such a case, a single electron tunneling lifetime is much shorter than a
single electron charging time so that a well-dened sawtooth voltage waveform free from a jitter is
observed. The sawtooth waveform oscillates between V = +q/2C and V = q/2C, so the average
voltage is close to zero. With increasing the current I (point (B) (C)), the jitter due to a nite
tunnel lifetime cannot be neglected as compared with the single electron charging time. Due to
the random distribution of tunneling times, the time averaged voltage waveform becomes deviated
from the sawtooth waveform and the average voltage is increased. At a high current level (point
(D)), the Coulomb blockade ceases to work.
A degree of randomness in electron tunneling events is dened by

2
=

2

2
, (8.30)
where
2
=
2

2
,
2
= 1 for a completely random Poisson-point-process and
2
= 0 for a
completely regulated point-process. Under a constant-voltage-bias condition, the electron tunneling
events obey a Poisson-point-process (
2
= 1). On the other hand, under the constant-current-bias
12
condition, we have

2
= (4 )
CR
T
_
q
I
_ . (8.31)
The value of
2
goes to zero only in the limit of CR
T

q
I
.
8.5 Coulomb staircase
A single electron thermionic emission (SETE) oscillation in a mesoscopic pn junction and single
electron tunneling (SET) oscillation in a mesoscopic M-I-M tunnel junction under constant current
operation have not been observed so far. The reason is that a stray capacitance short-circuits a
mesoscopic junction and makes a high-impedance constant current operation impossible. Therefore,
practical applications using a Coulomb blockade eect must use a low-impedance constant voltage
source. A Coulomb staircase observed in two tunnel junction diodes driven by a constant voltage
source is one such example (Fig. 8-7).
n
2
n
1
C
2
, V
2
, R
2
C
1
, V
1
, R
1
V
ex
+
-
Figure 8-7:
In Fig. 8-7, n
1
and n
2
are the number of electrons tunneled across the tunnel capacitance C
1
and C
2
. If n
1
= n
2
, the central electrode is electrically neutral. The external voltage is shared by
13
the two tunnel capacitances as follows:
V
ex
= V
1
+V
2
, (8.32)
V
1
=
Q
C
1
, (8.33)
V
2
=
Q
C
2
. (8.34)
From (8.32)(8.34), the common surface charge Q is determined as
Q =
C
1
C
2
C
1
+C
2
V
ex
, (8.35)
and the junction voltages are
V
1
=
C
2
C
1
+C
2
V
ex
, (8.36)
V
2
=
C
1
C
1
+C
2
V
ex
, (8.37)
If n
1
= n
2
+ 1, the central electrode has an excess one electron and the junction voltage V
1
drops
V
1
V
2
C
2
C
1
n
1
= n
2
n
1
= n
2
+ 1
V
1
- V
V
2
+ V
C
2
C
1
Figure 8-8:
and the junction voltage V
2
jumps up by the same amount as shown in Fig. 8-8. Due to the change
in the junction voltage V in the two junctions, the surface charge in each junction is modied
as follows:
Q
1
= C
1
V , (8.38)
14
Q
2
= C
2
V . (8.39)
The dierence Q
1
Q
2
should be equal to the change of one electron q, since the bulk region
of the central electrode must be electrically neutral. Therefore, the junction voltage change is given
by
V =
q
C
1
+C
2
. (8.40)
In general, (8.36) and (8.37) have the form
V
1
=
C
2
C
1
+C
2
V
ex

q
C
1
+C
2
(n
1
n
2
) , (8.41)
V
2
=
C
1
C
1
+C
2
V
ex
+
q
C
1
+C
2
(n
1
n
2
) . (8.42)
The free energy of the whole system is given by
F =
1
2
C
1
V
2
1
+
1
2
C
2
V
2
2
Q
ex
V
ex
, (8.43)
where Q
ex
is the charge supplied by the external voltage source to restore an electro-static steady
state condition and is given by
Q
ex
=
_
C
2
C
1
+C
2
n
1
+
C
1
C
1
+C
2
n
2
_
q . (8.44)
Suppose we start with the electrically neutral central electrode (n
1
= n
2
). When the forward
bias voltage exceeds V
ex,1
= q/2C
2
, the free energy for n
1
= n + 1 and n
2
= n becomes smaller
than the initial free energy:
F(n + 1, n) < F(n, n) . (8.45)
This means that a single electron is allowed to tunnel across the capacitance C
1
. If C
1
< C
2
, this
threshold voltage lifting the Coulomb blockade is smaller than the corresponding threshold voltage
15
V
ex,1
= q/2C
1
. Therefore, an electron tunneling is always initiated via the capacitance C
1
. Once a
single electron tunnels into the central electrode, the following inequality always holds:
F(n + 1, n + 1) < F(n + 1, n) . (8.46)
Consequently, an electron tunneling via the capacitance C
2
is now allowed. If C
1
R
1
< C
2
R
2
, the
electron tunneling via C
2
is immediately followed by the electron tunneling via C
1
and so there
is always one excess electron in the central electrode. This situation remains until the forward
bias voltage exceeds V
ex,2
=
3q
2C
2
. Above this threshold voltage, the free energy for n
1
= n + 2 and
n
2
= n becomes smaller than the free energy for n
1
= n + 1 and n
2
= n,
F(n + 2, n) < F(n + 1, n) . (8.47)
This means that the central electrode has always two excess electrons above this threshold voltage.
If one electron tunnels via C
2
, a new electron immediately tunnels via C
1
and restore the steady
state number 2. In general, n excess electrons can stay in the central electrode in the forward
voltage range:
_
n
1
2
_
q
C
2
< V
ex,n
<
_
n +
1
2
_
q
C
2
. (8.48)
The circuit current is uniquely determined by the number of excess electrons in the central
electrode:
I =
q
C
2
R
2
n . (8.49)
Therefore, the I-V characteristic features a step-like increase in the current at each threshold voltage
and a plateau in between, as shown in Fig. 8-9. The above behavior is referred to as Coulomb
staircase. The principle may be utilized for a high precision electrometer.
16
2
1
0
C
2
R
2
q
I
V
C
2
q
1
2
3
2
5
2
Figure 8-9:
8.6 Turnstile devices
A turnstile device is another example of a constant-voltage-driven mesoscopic tunnel junction
device. Let us consider a double barrier p-i-n tunnel junction shown in Fig. 8-10. At a forward
Figure 8-10:
bias voltage V
0
, the conduction electron in the n-side quantum well satises the resonant tunneling
condition into the electron sub-band in the central quantum well layer:
E
Cn
< E
e,QW
+
q
2
2C
n
< E
Fn
, (8.50)
17
where E
e,QW
, E
Cn
and E
Fn
are the electron sub-band energy in the central QW, the minimum
energy in the n-side QW and the Fermi-energy and C
n
is the tunnel junction capacitance between
the two QWs. When a single electron tunnels, the Coulomb energy of the central QW increases
by q
2
/2C
n
and the subsequent electron tunneling is inhibited by the Coulomb blockade eect.
The valence hole in the p-side QW does not satisfy the resonant tunneling condition into the hole
sub-band in the central QW layer in either case, so the hole tunneling is inhibited.
Now, the forward bias voltage is increased to V
0
+V for satisfying the hole resonant tunneling
condition in the presence of a single electron in the central QW. Once a single hole tunnels into the
hole sub-band in the central QW layer, the Coulomb energy of the central QW layer decreases to
the original value, where the subsequent hole tunneling is inhibited. The conduction electron in the
n-side QW does not satisfy the resonant tunneling condition into the electron sub-band in either
case. In this way, a single electron and hole are injected into the central QW per each cycle of the
modulated voltage. The external circuit current is related to the modulation frequency, I = qf.
This principle may be used for a high precision current standard.
Suppose the radiative recombination lifetime is much shorter than the period,
recom
1/f =
q/I, a single photon is emitted during every period. This is a heralded single photon state.
18
References
1. Basic physics of a pn junction diode and tunnel junction diode:
S. M. Sze, Physics of Semiconductor Devices, John Wiley & Sons, New York (1981).
2. Noise of mesoscopic M-I-M tunnel junctions:
D. V. Averin and K. K. Likharev, J. Low Temp. Phys. 62, 345 (1986).
K. K. Likharev, IBM J. Res. Dev. 32, 144 (1988).
3. Coulomb staircase:
K. Mullen, E. Ben-Jacob, R. C. Jaklevic and Z. Schuss, Phys. Rev. B32, 98 (1988).
4. Turnstile
J. Kim, et al., Nature 397, 500 (1999).
19
Contents
9. Negative Conductance Oscillators 2
9.1 Free-Running van der Pol Oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
9.2 Amplitude and Phase Noise of a Free-Running Oscillator . . . . . . . . . . . . . . . 7
9.3 Spectral Linewidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
9.4 Amplitude and Phase Noise of an Output Wave . . . . . . . . . . . . . . . . . . . . . 15
9.5 Injection-Locked Oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1
9. Negative Conductance Oscillators
Negative conductance oscillators are employed for generating coherent electromagnetic elds
in various frequency regions. The essential components of a negative conductance oscillator are a
frequency-selective circuit, a device showing a negative dierential conductance (gain), and a device
with a nonlinear characteristic (gain-saturation or amplitude-limiting function). Various frequency-
selective circuits are used depending on operation frequencies, i.e., an LC circuit, delayed feedback
loop, Fabry-Perot cavity, ring cavity, distributed feedback cavity, and many others. Usually, the
negative dierential conductance and the nonlinear (gain saturation) characteristics are provided
by the same device. Solid-state devices such as tunnel (Esaki) diodes, IMPATT diodes, Gunn
diodes, and Josephson junctions are employed for a microwave oscillator. Accelerated electron
beams and inverted media are used as a negative dierential conductance and nonlinear element
for higher-frequency regions, including millimeter, sub-millimeter, infrared, optical, and XUV elds.
In a typical coherent communication system, a transmitter consists of an oscillator, modulator,
and post-amplier. An oscillator generates a low-noise and frequency-stabilized electromagnetic
wave. Information is encoded onto either the amplitude, frequency, or phase of the coherent carrier
wave emitted from the oscillator. A post-amplier compensates for the loss of the modulator
and/or to boost the transmitted signal power. If frequency or phase modulation is employed, an
injection-locked oscillator can be used as a post-amplier. For example, a low-noise Gunn diode
is used as an oscillator and a high-power IMPATT diode is used as an injection-locked oscillator
(post-amplier). The received signal is fed into a mixer with a coherent local oscillator wave in
order to translate the carrier wave into an intermediate frequency (IF) signal or directly into a
2
baseband signal. This mixing process has a ltering function and rejects background noise and
cross-talk from other channels. A local oscillator is frequency stabilized and often phase-locked
to the received signal. A phase-locked-loop (PLL) oscillator is employed for homodyne detection.
Negative conductance oscillators can be used as free-running oscillators, injection-locked oscillators,
and phase-locked-loop local oscillators in such a coherent communication system.
active element
-R
a
(A) + jX
a
(A)

0
=
1
LC
resonant circuit
load resistance
R
L
v
L
v
a
L C
i(t)
Figure 9-1:
A negative conductance oscillator is described by the circuit shown in Fig. 9-1. An LC series
circuit constitutes a frequency-selective element. An active element with a negative conductance
(gain) and a nonlinear characteristic is represented by a complex impedance R
a
+jX
a
, where the
negative resistance R
a
represents (saturable) gain and the reactance X
a
represents dispersion. The
active element should have an internal noise source represented by the noise voltage v
a
. A positive
resistance R
L
represents a load resistance, which accounts for the output coupling loss from the
oscillator. An external noise is fed into the oscillator through this load resistance and is represented
by the noise voltage v
L
. Finally, the oscillation eld is represented by an internal current i(t). van
der Pol was the rst to study the noise properties of such a negative conductance oscillator; thus, it is
often referred to as the van der Pol oscillator. Note that the circuit representation of the negative
3
conductance oscillator in Fig. 9-1 is quite general. This simple model covers the fundamental
performance and noise property of almost all negative conductance oscillators, including a laser
oscillator.
9.1 Free-Running van der Pol Oscillators
When the active element is pumped by an external energy source in order to produce a
negative conductance, the internal and external noise voltages, v
a
and v
L
, are selectively amplied
and the uctuation frequency component of the noise current i(t) at the LC circuit resonant
frequency,
0
=
1

LC
, grows. This process is called regenerative amplication. Once the negative
resistance of the active element balances the positive load resistance, R
a
= R
L
, the circuit becomes
unstable and the noise grows exponentially, purifying its spectral shape. The circuit starts to
oscillate and the steady-state coherent eld amplitude is established in the circuit. This steady-
state condition is established by the gain-saturation or amplitude-limiting function of the active
element. In this way, the broadband noise sources, v
a
and v
L
, are transformed into a coherent wave
with stabilized amplitude and a well-dened frequency. Frequency-selective amplication which
puries the spectral prole and gain saturation which stabilizes the amplitude are the two basic
ingredients for a negative conductance oscillator.
The circuit equation (in complex representation) for the current i is given by
_
R
L
+j
_
L
1
C
_
R
a
+jX
a
_
i = v
a
+v
L
. (9.1)
A time-dependent current i(t) is expressed as
i(t) Re(i) = Re
_
(A+ A)e
j(t+)
_
, (9.2)
4
where A and are the average amplitude and frequency of the oscillating current and A and
are slowly varying amplitude and phase uctuations. The gain saturation is represented by
R
a
=
R
0
1 +A
2
, (9.3)
where R
0
is the unsaturated negative resistance and is the saturation parameter.
If one assumes

A
A

1, one can linearlize R

a
and X
a
as follows:
R
a
= R
a
(A) +
R
a
A
A , (9.4)
X
a
= X
a
(A) +
X
a
A
A . (9.5)
Since an actual oscillation frequency is close to the LC circuit resonant frequency, one obtains
L
1
C
=
L

1
LC
_
=
L

( +
0
)(
0
) 2L(
0
) . (9.6)
Substituting (9.2), (9.4), (9.5), and (9.6) into (9.1) and taking the real part of both sides of (5.1),
one obtains
Re
__
R
L
R
a
(A)
R
a
A
A+j2L(
0
)
+j
_
X
a
(A) +
X
a
A
A
__
(A+ A)e
j(t+)
_
= v
a
(t) +v
L
(t) . (9.7)
Replacing j = j(
0
) by
d
dt
and taking the time derivative (A+A)e
j(t+)
, (9.7) is reduced
to
Re
__
R
L
R
a
(A) +j2L
_

0
+
X
a
(t)
2L
_
+ 2L
_
1
A
dA
dt

1
2L
R
a
A
A
_
+j2L
_
d
dt
+
1
2L
X
a
A
A
__
Ae
j(t+)
_
= v
a
(t) +v
L
(t) . (9.8)
5
If one ignores all uctuation terms, i.e. A = = v
a
= v
L
= 0, in (9.8), one has the following
steady-state solutions:
R
L
= R
a
(A) =
R
0
1 +A
2
, (9.9)
=
0

X
a
(A)
2L
. (9.10)
From (9.9), the squared, steady-state oscillation amplitude is given by
A
2
=
1

_
R
0
R
L
1
_
. (9.11)
As shown in Fig. 9-2(a), the squared coherent oscillation amplitude builds up at an oscillation
A
2
R
L
R
0
( pump rate)
oscillation
threshold
0
R
L
R
0
( pump rate)
0
unsaturated gain
saturated gain
R
L
R
a
(A)
Figure 9-2:
threshold R
0
= R
L
and increases linearly with the pump rate (the unsaturated negative resistance
R
0
is proportional to the pump rate). The stored coherent eld energy inside the oscillator and
6
the output power into the (external) load are expressed by
E
1
2
LA
2
, (9.12)
P
1
2
R
L
A
2
. (9.13)
The saturated negative resistance R
a
linearly increases with the pump rate below threshold and
is clumped at R
L
above threshold, as shown in Fig. 9-2(b). When the oscillation eld increases
above the steady-state value given by (9.11), the saturated gain decreases to below R
L
(the circuit
has a net loss) and the oscillation eld is attenuated. On the other hand, when the oscillation
eld decreases below the steady-state value, the saturated gain increases to above R
L
. In this case,
the circuit has a net gain and the oscillation eld is amplied. In this way, the oscillation eld
amplitude and the saturated gain are simultaneously stabilized to their steady-state values. This
nonlinear process is due to mutual coupling between the oscillation eld and the active element
and is termed relaxation oscillation.
9.2 Amplitude and Phase Noise of a Free-Running Oscillator
The small uctuating parts in (9.8) are
_
2L
d
dt
AA
R
a
A
A
_
cos(t + )
_
2LA
d
dt
+A
X
a
A
A
_
sin(t + ) =
v
a
(t) +v
L
(t) . (9.14)
Multiplying (9.14) by cos(t +) or sin(t +) and integrating over one period of oscillation,
T =
2

, one has
2L
d
dt
AA
R
a
A
A =

_
t+

(v
a
(t

) +v
L
(t

)) cos(t

+ )dt

= v
ac
+v
Lc
, (9.15)
7
2LA
d
dt
+A
X
a
A
A =

_
t+

(v
a
(t

) +v
L
(t

)) sin(t

+ )dt

= (v
as
+v
Ls
) . (9.16)
Here, the cosine and sine components of the noise voltages satisfy the following relations:
v
a
(t) = v
ac
cos(t + ) +v
as
sin(t + ) , (9.17)
v
L
(t) = v
Lc
cos(t + ) +v
Ls
sin(t + ) . (9.18)
A resistive saturation parameter s and reactive saturation parameter r are introduced and
dened by
s
A
R
a
(A)
R
a
A
, (9.19)
r
A
R
a
(A)
X
a
A
. (9.20)
If one uses (9.3) for saturated gain, the saturation parameter is given by
s =
2A
2
1 +A
2
=
_

_
0 : A
2
1 (just above threshold)
2 : A
2
1 (far above threshold)
. (9.21)
Equations (9.15) and (9.16) are rewritten using (9.19) and (9.20) as
d
dt
A+
sR
L
2L
A =
1
2L
(v
ac
+v
Lc
) , (9.22)
d
dt
+
rR
L
2LA
A =
1
2LA
(v
as
+v
Ls
) . (9.23)
The amplitude noise A is caused by the cosine components of the internal and external noise
voltages and is suppressed with the decay rate
s
2
R
L
L
=
s
2
_

Qe
_
, where Q
e
is a Q factor due to
output coupling loss and

Qe
is the associated photon decay rate. The gain saturation represented
by the resistive saturation parameter s operates as a restoring force for the amplitude. On the
8
other hand, there is no restoring force for the phase. The phase noise is caused by the sine
components of the internal and external noise voltages and driven by the amplitude noise via
the reactive saturation parameter. Therefore, the phase of a free-running oscillator diuses via
a random walk, while the amplitude is stabilized to its steady-state value. This eect is shown
schematically in Fig. 9-3.

Im(I)
A
random-walk diffusion
amplitude restoring force
quadrature-phase noise
in-phase noise
-
s
2
R
L
L
Re(I)
Figure 9-3:
Fourier analysis of (9.22) and (9.23) results in the power spectral densities of A and :
S
A
() =
1
s
2
R
2
L

S
ac
() +S
Lc
()
1 + (/
c
)
2
, (9.24)
S

() =
_

Qe
_
2
4A
2
R
2
L

2
_
S
as
() +S
Ls
()
_
+
_
r
s
_
2
_

Qe
_
2
4A
2
R
2
L

2

S
ac
() +S
Lc
()
1 + (/
c
)
2
. (9.25)
Here, the noise bandwidth
c
is given by

c
=
s
2
_

Q
e
_
=
_

_
0 : A
2
1 (just above threshold)

Qe
: A
2
1 (far above threshold)
. (9.26)
Figure 9-4 shows the amplitude noise spectra for various pump rates. At far above threshold, the
amplitude noise spectrum is reduced to
S
A
() =
1
4R
2
L

S
ac
() +S
Lc
()
1 +
_
/

Qe
_
2
. (9.27)
9
Let us consider the two limiting cases:
1
10
10
2
1 10
10
-1
10
-1
R
0
R
L
>
(just above threshold)
R
0
R
L
(far above threshold) >>
4R
L
2
S
A
()
S
ac
() + S
Lc
()
Frequency
/

Q
e
Figure 9-4:
(1) Quantum Limit: S
ac
() = S
Lc
() = 4 hR
L
Both internal and external noise voltages are limited by a quantum mechanical zero-point
uctuation. An ideal laser oscillator is such an example. The amplitude noise spectrum for this
case is
S
A
() =
2h
R
L
1 +
_
/

Qe
_
2
. (9.28)
(2) Thermal Limit: S
ac
() = S
Lc
() = 8k
B
TR
L
Both internal and external noise voltages are limited by Johnson-Nyquist thermal noise. An
ideal microwave oscillator is such an example. The amplitude noise spectrum for this case is
S
A
() =
4k
B
T
R
L
1 +
_
/

Qe
_
2
. (9.29)
10
The stored energy inside the LC circuit is given by
1
2
LA
2
= hn , (9.30)
where n is the number of oscillator photons. Therefore, the spectrum of the photon number is
S
n
() =
_
LA
h
_
2
S
A
() =
_

_
4n
_

Qe
_
1
1+
_
/

Qe
_
2
: Quantum limit
4n
_
2k
B
T
h
_
_

Qe
_
1
1+
_
/

Qe
_
2
: Thermal limit
. (9.31)
Therefore, the variance in the photon number n
2
) is calculated by the Parseval theorem:
n
2
)
_

0
S
n
()
d
2
=
_

_
n : Quantum limit
2nn
th
: Thermal limit
. (9.32)
The van der Pol oscillator in the quantum limit has a Poissonian photon number distribution at
far above threshold. On the other hand, the variance of the photon number in the thermal limit is
larger than the Poisson limit by a factor of 2n
th
= 2 (k
B
T/h), where n
th
is the number of thermal
photons.
Figure 9-5 shows the phase noise spectra for various pump rates.
9.3 Spectral Linewidth
The oscillator circuit impedance is calculated from (9.1) as
Z() = 2L
_
1
2

Q
+j(

0
)
_
, (9.33)
where

0
=
0

Xa(A)
2L
is the actual oscillation frequency,
0
is the cold cavity resonant frequency,
and Q is the eective (or active) Q value dened by

Q


Q
e

Q
a
=
R
L
L

R
a
(A)
L
. (9.34)
11
10
5
10
4
10
3
10
2
10
1
10
-1
10
-2
10
-3
1 10
-1
10
-2
10
R
0
R
L
(just above threshold)
>
R
0
R
L
(far above threshold) >>
(1 + r
2
)
Frequency

Q
e
Figure 9-5:
The spectral prole of the oscillating eld is calculated by
[i()[
2
=
S
va
(

0
) +S
v
L
(

0
)
4L
2
_
1
4
_

Q
_
2
+ (

0
)
2
_ . (9.35)
Since S
va
(

0
) and S
v
L
(

0
) are slowly-varying functions of

0
, the spectral prole (9.35) is Lorentzian
with a full-width at half-maximum (FWHM)

1/2
=

Q
. (9.36)
(1) Below the Oscillation Threshold
In this case, gain saturation is negligible, R
a
(A) R
0
, and thus one obtains

1/2
=

Q
e
_
1
R
0
R
L
_
. (9.37)
The spectral linewidth decreases linearly with the dierence (R
L
R
0
) between the threshold pump
rate and the actual pump rate.
12
(2) Above the Oscillation Threshold
In the earlier discussion about the steady-state solution, the saturated gain R
a
(A) was made
equal to the loss R
L
. This is not exactly the case because an actual oscillator has internal and
external noise sources. The saturated gain R
a
(A) is always slightly smaller than the loss R
L
; that is,
an actual oscillator has a small net loss even above threshold, as shown in Fig. 9-6. The steady-
state oscillation eld is maintained in spite of a Mnet loss because the internal and external noise
powers are coupled into the oscillator and compensate for the net loss. By increasing the noise
powers coupled into the oscillator, the net loss, R
L
R
a
(A), increases and the spectral linewidth
becomes broader.
R
L
R
a
(A)
R
L
R
L
-R
a
(A) : net loss
R
0
0
oscillation threshold
Figure 9-6:
The total emitted power is given by
P
e
= R
L
_

0
[i(

0
)[
2
d
2
=
R
L
4L
2
_

Q
_
_
S
va
(

0
) +S
v
L
(

0
)
_
. (9.38)
Let us consider the two limiting cases:
13
(2-A) Quantum Limit: S
va
(

0
) = S
v
L
(

0
) = 2 h

0
R
L

1/2


Q
=
h

0
_

Qe
_
2
P
e
. (9.39)
If one uses
1/2
=
1
2

1/2
and
c
=
1
2
_

Qe
_
, one obtains

1/2
=
2h

0
(
c
)
2
P
e
. (9.40)
This is the Schawlow-Townes linewidth for a laser oscillator.
(2-B) Thermal Limit: S
va
(

0
) = S
v
L
(

0
) = 4k
B
TR
L

1/2
=
2k
B
T
_

Qe
_
2
P
e
, (9.41)

1/2
=
4k
B
T(
c
)
2
P
e
. (9.42)
This is the Shimoda-Takahashi-Townes linewidth for a maser oscillator.
The spectral linewidth
1/2
in the quantum limit is rewritten using (9.11) and (9.13) as

1/2
=
_

Q
e
_
2
h

0
L
_
R
0
R
L
1
_ . (9.43)
The linewidth drops by a factor of 2h

0
/L at the threshold and decreases linearly with the relative
pump rate (R
0
/R
L
1), as shown in Fig. 9-7. The physical meaning of the factor 2h

0
/L can be
elucidated in the following way. According to the saturated gain model (9.9), the gain is decreased
to one-half at the saturation intensity:
A
2
s
=
1

. (9.44)
This saturation intensity A
2
s
is converted to the saturation photon number by (9.30)
n
s
=
L
2h

. (9.45)
14
The linewidth drop at the oscillation threshold is determined by the saturation photon number.
The oscillator with a larger saturation photon number n
s
has a larger drop in linewidth.
R
L R
0

1/2
1 -
R
0
R
L
below threshold
2h
L
R
0
R
L
- 1
-1
above threshold
Figure 9-7:
9.4 Amplitude and Phase Noise of an Output Wave
The amplitude and phase noise studied thus far are those of the internal eld of an oscillator.
In most practical cases, however, one needs to know the noise of an output wave rather than that
of an internal eld, and these two are not identical.
L C
circulator
transmission
line Z = R
L
-R
a
+ jX
a
v
a
v
L
2R
L
R
L
v
L
I
e
I
Figure 9-8:
Consider a negative conductance oscillator with a circulator as the output coupling element,
15
as shown in Fig. 9-8. The circulator separates the input and output ports and serves as the load
resistance R
L
for an oscillator circuit. Therefore, as far as the internal current I is concerned, this
conguration is identical to the oscillator model shown in Fig. 9-1.
The output current I
e
consists of the reected noise current v
L
/2R
L
and the internal current I:
I
e
= I
v
L
2R
L
, (9.46)
where
I = (A+ A) cos(t + ) , (9.47)
I
e
= (A
e
+ A
e
) cos(t +
e
) , (9.48)
v
L
= v
Lc
cos(t + ) +v
Ls
sin(t + ) . (9.49)
From the above relations, the amplitude noise of the output wave is
A
e
() = A()
v
Lc
()
2R
L
=
_
1
j2L +sR
L

1
2R
L
_
v
Lc
() +
1
j2L +sR
L
v
ac
() . (9.50)
The power spectral density of A
e
is thus given by
S
Ae
() =
_
2
s
1
_
2
+ (/
c
)
2
4R
2
L
[1 + (/
c
)
2
]
S
Lc
() +
1
s
2
R
2
L

1
1 + (/
c
)
2
S
ac
() . (9.51)
At just above the oscillation threshold (s 1), the spectrum (9.51) is close to that of the internal
eld (9.24) except for the white noise in the high-frequency regime, >

Qe
, (Fig. 9-9). This white
noise comes from the reected external noise v
Lc
/2R
L
in (9.50).
At far above threshold (s 2), (9.51) is reduced to
S
Ae
() =
1
4R
2
L
_

_
_
/

Qe
_
2
1 +
_
/

Qe
_
2
S
Lc
() +
1
1 +
_
/

Qe
_
2
S
ac
()
_

_ . (9.52)
16
4R
L
2
S
A
e

S
LC
+ S
ac

10
2
10
1
10
-1
10
-2
10
-1
1 10 10
2
Frequency

Q
e
1
2
4
S
2
Figure 9-9:
Consider the following two cases. When both internal and external noise sources are quantum
limited, i.e., S
Lc
() = S
ac
() = 4 hR
L
, one has a white noise spectrum:
S
Ae
() =
h
R
L
. (9.53)
The output photon ux N is given by
N =
1
2
R
L
A
2
e
/h , (9.54)
and its spectrum is
S
N
() =
_
R
L
A
e
h
_
2
S
Ae
()
=
R
L
A
2
e
h
= 2N . (9.55)
If this photon ux uctuation is converted to photocurrent uctuation by a photodetector with
17
100% quantum eciency, the current spectrum is
S
I
() = q
2
S
N
() = 2q
2
N = 2qI , (9.56)
where I = qN is the average current. This is the full-shot noise. As shown in Fig. 9-10, the
origin of this (quantum) shot noise is the internal noise v
ac
in the low-frequency regime, <

Qe
,
and the external noise v
Lc
in the high-frequency regime, >

Qe
. The (originally white) internal
noise v
ac
is transformed into a Lorentzian spectrum due to the storage (averaging) function of the
resonator. A rapid uctuation component is averaged by the storage eect of the eld inside the
resonator. On the other hand, the (originally white) external noise v
Lc
is transformed into the
opposite spectral shape because a uctuation component near the cavity resonance
_
<

Qe
_
is
absorbed and suppressed by the gain saturation. A uctuation component far from the cavity
resonance
_
Q >

Qe
_
is simply reected. A saturated oscillator behaves as a matched load near
resonance and behaves as an innite impedance reector far from resonance.
1
1
2
10
-1
10
-2
1 10
-1
10
total noise
external noise
S
Lc
() S
ac
()
internal noise
Frequency

Q
e
Figure 9-10:
18
The phase noise of an output wave is given by

e
() = ()
v
Ls
()
2R
L
A
=
1
j2LA
_
v
as
+
_
1 +j
L
R
L
_
v
Ls
_

rR
L
j2LA

1
j2L +sR
L
(v
ac
+v
Lc
) .(9.57)
The power spectral density of
e
is
S
e
() =
1
4L
2
A
2

2
S
as
() +
_
1
4L
2
A
2

2
+
1
4R
2
L
A
2
_
S
Ls
()
+
_
r
s
_
2
4L
2
A
2

2

1
1 + (/
c
)
2
_
S
ac
() +S
Lc
()
_
. (9.58)
The phase noise spectrum S
e
() is shown in Fig. 9-11 and is dierent from the internal phase
noise spectrum S

() in the high-frequency, white-noise spectrum. The internal phase noise due

to v
Ls
and the directly reected noise wave v
Ls
are 90

out-of-phase, as indicated in (9.57), so they

are simply added. On the other hand, the internal amplitude noise due to v
Lc
and the directly
reected noise wave v
Lc
are 180

out-of-phase, as indicated in (9.50), and thus cancel each other

out.
In the special case of no internal noise, S
ac
() = S
as
() = 0, and quantum-limited external
noise, S
Lc
() = S
Ls
() = 4 hR
L
, the amplitude and phase noise spectra are
S
Ae
() =
h
R
L

_
/

Qe
_
2
1+
_
/

Qe
_
2

A
2
S
e
() =
h
R
L
1+
_
/

Qe
_
2
_
/

Qe
_
2

S
Ae
() A
2
S
e
() =
_
h
R
L
_
2
. (9.59)
This is often referred to as the spectral Heisenberg uncertainty principle. A saturated oscillator
suppresses the amplitude noise to below the shot-noise value (standard quantum limit: SQL) and
enhances the phase noise to above the SQL within the resonator bandwidth, <

Qe
. Such a
19
10
4
10
3
10
2
10
1
10
-1
1 10
-1
10
-2
10
2
Frequency

Q
e
10
4R
L
2
A
2
S

e
()
S
as
() + S
Ls
()

Q
e
2

2
1
2
Figure 9-11:
eld with reduced amplitude noise and enhanced phase noise is called an amplitude-squeezed state,
which satises the minimum uncertainty product. On the other hand, a eld with the amplitude
and phase noise equal to the SQL is called a coherent state. An ideal saturated oscillator without
internal noise produces an amplitude-squeezed state in the low-frequency regime and a coherent
state in the high-frequency regime, as shown in Fig. 9-12.
9.5 Injection-Locked Oscillators
Consider a negative conductance oscillator injection-locked by an external signal v
e
, as shown
in Fig. 9-13. The circuit equation (in complex representation) is
_
R
L
+j
_
L
1
C
_
R
a
+jX
a
_
i = v
a
+v
L
+v
e
. (9.60)
Assume that the internal current i is phase-locked by the injection signal i
e
=
ve
2R
L
, i.e.,
i
e
(t) = Re
_
v
e
2R
L
_
= Re(A
e
e
jt
) , (9.61)
20
1
10
10
3
10
2
10
4
10
-4
10
-3
10
-2
10
-1 10
-1
1 10 Frequency

Q
e
SQL
A
2
S

e
()
S
A
e
()
<<

Q
e
>>

Q
e
amplitude squeezed
state
coherent state
Im(I) Im(I)
Re(I)
Re(I)
S
h
R
L
Figure 9-12:
i(t) = Re
_
(A+ A)e
j(t++)
_
. (9.62)
The amplitude and phase noise of the injection signal are attributed to the external noise voltage
v
L
. Using (9.61) and (9.62) in (9.60), one obtains
_
R
L
R
a
(A)
R
a
A
A+ 2L
1
A
d
dt
A+j2L
_

0
+
X
a
(A)
2L
+
1
2L
X
a
A
A+
d
dt

__
(A+ A)e
j(t++)
= v
a
+v
L
+ 2R
L
A
e
e
jt
. (9.63)
The steady-state solution is obtained by neglecting all uctuating terms:
_
R
L
R
a
(A) +j2L(

0
)
_
A = 2R
L
A
e
(cos j sin) . (9.64)
21
L C
circulator
-R
a
+ jX
a
v
a
I
e
I
1
2R
L
v
L
+ v
e
Figure 9-13:
Here,

0
=
0

Xa(A)
2L
is a free-running oscillation frequency. The real part of (9.64) is
[R
L
R
a
(A)]A = 2R
L
A
e
cos . (9.65)
We expand the oscillation amplitude A in terms of a free-running oscillation amplitude A
0
and
small change of amplitude A due to the injection signal:
A = A
0
+ A , (9.66)
where A
0
satises the gain clamping condition
R
L
=
R
0
1 +A
2
0
. (9.67)
Using (9.66) and (9.67) in (9.65), one obtains
A = A
e
_
1 +A
2
0
A
2
0
_
cos . (9.68)
This increase in the internal oscillation amplitude is converted to the change in the output wave
amplitude by using the boundary condition, A
e
= AA
e
:
A
e
=
(1 +A
2
0
) cos A
2
0
A
2
0
A
e
. (9.69)
22
When the oscillator is pumped just above threshold, A
2
0
1, the conversion from the input to
output amplitude signals is given by
A
e
A
e

cos
A
2
0
(amplication) . (9.70)
On the other hand, when the oscillator is pumped far above threshold, A
2
0
1, the conversion
factor is
A
e
A
e
cos 1 (attenuation) . (9.71)
As will be shown later, if the injection signal frequency and the free-running oscillation frequency

0
are identical, the phase shift is equal to zero. In such a case, the injection-locked oscillator at
far above threshold completely suppresses the amplitude signal and, therefore, the injection-locked
oscillator operates as an amplitude limiter.
The imaginary part of (9.64) is
L(

0
)A = R
L
A
e
sin . (9.72)
In order to have a real value of the phase shift in (9.72), one has the constraint for the allowed
frequency detuning [

0
[:
[

0
[ =

R
L
L
A
e
A
sin

Q
e

A
e
A
=
L
. (9.73)
This is the Adler equation for a locking bandwidth. When the frequency detuning

0
is within
this bandwidth, the oscillator frequency is locked to the injection signal frequency . The phase
shift is now given by
sin =

0
_

Qe
_
Ae
A
=

L
. (9.74)
23
Figure 9-14 shows the oscillation frequency
a
, phase shift , internal amplitude modulation A,
and output wave amplitude modulation A
e
of the injection-locked oscillator as a function of the
frequency detuning

0
.

a
-

0
-

2
-

2

0
-

L

L

0
-
A
A
e
1 + A
0
2
A
0
2

L

L

0
-
R
0
R
L
>
R
0
R
L
>>
A
e
A
e

0
0
Figure 9-14:
Consider the amplitude and phase noise of an injection-locked oscillator. Assuming =

0
and
= 0 in (9.63), one obtains
_
2L
d
dt
AA
R
a
A
A
_
cos(t) 2LA
_
d
dt
+
1
2L
X
a
A
A
_
sin(t)
= v
a
+v
L
+ 2R
L
A
e
sin(t) . (9.75)
24
Multiplying cos t and sint and integrating over one period of oscillation, one has the following
equations for the amplitude noise A and phase noise :
2L
d
dt
AA
R
a
A
A = v
ac
+v
LC
, (9.76)
2LA
_
d
dt
+
1
2L
X
a
A
A
_
= v
as
+v
Ls
+ 2R
L
A
e
. (9.77)
The resistive saturation parameter s is now given by
s
A
R
a
(A)
R
a
A
=
A
R
L
R
a
A
_
1
2A
e
A
_
1
, (9.78)
where (9.65) with = 0 is used to derive the second equality. Equation (9.76) is rewritten as
d
dt
A+
s
2
_

Q
e
__
1
2A
e
A
_
A =
1
2L
(v
ac
+v
Lc
) . (9.79)
The Fourier-transformed internal and external amplitude noise are
A() =
1
2L
[v
ac
() +v
LC
()]
j +
a
, (9.80)
A
e
() = A()
v
Lc
()
2R
L
=

Qe

a
j
2R
L
(j +
a
)
v
Lc
() +

Qe
2R
L
(j +
a
)
v
ac
() . (9.81)
Here,
a
=
s
2
_

Qe
_ _
1
2Ae
A
_
is the amplitude noise bandwidth. At far above threshold and for
a relatively small injection signal,

Qe

a
2
L
. Since

Qe

L
in a practical situation,

a


Qe
and one thus has the amplitude noise spectrum:
S
Ae
() =
1
4R
2
L
_

2
+ 4
2
L

2
+
_

Qe
_
2
S
Lc
() +
_

Qe
_
2

2
+
_

Qe
_
2
S
ac
()
_

_ . (9.82)
25
When the internal noise voltage is negligible, S
ac
() = 0, the normalized amplitude noise spectrum
is as shown schematically in Fig. 9-15.
4R
L
2
S
A
e

S
LC

1
4
L
2

Q
e
2
2
L

Q
e
Frequency
Figure 9-15:
On the other hand, (9.77) is rewritten as
d
dt
+
L
=
r
_

Qe
_
2A
A
1
2LA
(v
as
+v
Ls
) . (9.83)
The Fourier-transformed internal and external phase noise for the negligible reactive saturation
parameter r = 0 are
() =
[v
as
() +v
Ls
()]
(j +
L
)2LA
, (9.84)

e
() = () +
v
Ls
()
2R
L
A
=
1
2LA(j +
L
)
_
_
_
v
as
() +
_
_
1
j +
L
_

Qe
_
_
_
v
Ls
()
_
_
_
. (9.85)
The phase noise spectrum is given by
S
e
() =
S
as
()
4L
2
A
2
(
2
+
2
L
)
+
S
Ls
()
4L
2
A
2

2
+
_

Qe

L
_
2
_

2
+
2
L
_
_

Qe
_
2
. (9.86)
26
When the internal noise voltage is negligible, S
as
() = 0, the normalized phase noise spectrum is
as shown schematically in Fig. 9-16.

L

Q
e
Frequency
4R
L
2
A
2
S

e
()
S
Ls
()

Q
e
2

L
2
1
Figure 9-16:
If the internal noise is negligible, S
ac
() = S
as
() = 0, and the external noise is quantum-
limited, S
Lc
() = S
Ls
() = 4 hR
L
, the amplitude and phase noise spectra are reduced to
S
Ae
() =
h
R
L
_

2
+ 4
2
L

2
+
_

Qe
_
2
_

_ , (9.87)
A
2
S
e
()
h
R
L
_

2
+
_

Qe
_
2

2
+
2
L
_

_ . (9.88)
Assume
L


Qe
and that the product of the amplitude and phase noise spectra is reduced to
the uncertainty product:
S
Ae
() A
2
S
e
()
_
h
R
L
_
2
, (9.89)
where = 4 for
L
and = 1 (minimum uncertainty product) for
L
. As shown in
Fig. 9-17, an injection-locked oscillator has a localized phase distribution due to a phase-restoring
27
force, while a free-running oscillator has a completely random phase distribution due to the absence
of a phase-restoring force.
Im(I)
Re(I)
injection-locked
oscillator
free-running
oscillator
Figure 9-17:
28
Contents
10.Parametric Ampliers and Oscillators 2
10.1 Non-degenerate parametric amplier . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
10.1.1 Principle of operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
10.1.2 Power gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
10.1.3 Noise gure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
10.2 Degenerate parametric amplier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
10.2.1 Principle of operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
10.2.2 Phase sensitive amplier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
10.3 Quantum limit of a linear amplier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
References 13
Figure Captions 13
1
10. Parametric Ampliers and Oscillators
A device exhibiting a negative conductance, such as a tunnel diode, can be utilized to construct
an amplier and oscillator. A laser is also categorized as a negative conductance oscillator as we
have seen in the previous chapter. There is another class of amplier and oscillator, which is based
on non-linear susceptances and known as a parametric amplier/oscillator.
For instance, a reverse-biased pn junction has the non-linear charge-voltage characteristic due
to the voltage-dependent capacitance. The mixing occurs between the three frequency components
of signal, idler and pump waves in such a nonlinear element and the energy ows from a strong
pump wave to weak signal and idler waves. This ow of the power from the pump to the signal
introduces the negative conductance into the signal circuit. In optical spectral domain, the atomic
dipole moment, driven by an intense pump laser, features a similar non-linearity and is capable of
amplifying weak signal and idler waves.
10.1 Non-degenerate parametric amplier
10.1.1 Principle of operation
An equivalent circuit for a non-degenerate parametric amplier is shown in Fig. 10-1. A
nonlinear capacitor is surrounded by three parallel LCR circuits, which represent the signal, idler
and pump circuits, respectively.
The charge q on the nonlinear capacitance is a function of the voltage across its terminals.
2
Figure 10-1:
Using the Taylor series expansion, the charge may be expressed in the form:
q(t) = a
1
v(t) +a
2
v
2
(t) +a
3
v
3
(t) + . (10.1)
When all the coecients except the rst and second terms are zero, the charge varies quadratically
with the voltage,
q(t) = Cv(t) +a
2
v
2
(t) , (10.2)
where a
1
is just replaced by the linear capacitance C. The current owing in the nonlinear capac-
itance is
i(t) =
dq(t)
dt
= C
dv(t)
dt
+ 2a
2
v(t)
dv(t)
dt
, (10.3)
where the voltage across the nonlinear capacitance consists of the signal, idler and pump waves at
3
angular frequencies
1
,
2
, and
3
, respectively,
v(t) = v
1
(t) +v
2
(t) +v
3
(t)
= V
1
cos(
1
t +
1
) +V
2
cos(
2
t +
2
) +V
3
cos(
3
t +
3
) . (10.4)
The angular frequencies in (10.4) satisfy

3
=
1
+
2
, (10.5)

i
= 1/
_
L
i
(C
i
+C) . (10.6)
Using (10.4) in (10.3), we obtain the expression for the current,
i(t) = i
1
(t) +i
2
(t) +i
3
(t) , (10.7)
where
i
1
(t) =
1
CV
1
sin(
1
t +
1
)
1
a
2
V
2
V
3
sin(
1
t +
3

2
) , (10.8)
i
2
(t) =
2
CV
2
sin(
2
t +
2
)
2
a
2
V
1
V
3
sin(
2
t +
3

1
) , (10.9)
i
3
(t) =
3
CV
3
sin(
3
t +
3
)
3
a
2
V
1
V
2
sin(
3
t +
1
+
2
) . (10.10)
Equations (10.8) - (10.10) can be rewritten as
i
1
(t) = C
dv
1
(t)
dt
+
a
2
V
2
V
3
V
1
_
cos(
3

2

1
)
dv
1
(t)
dt

1
v
1
(t) sin(
3

2

1
)
_
, (10.11)
i
2
(t) = C
dv
2
(t)
dt
+
a
2
V
1
V
3
V
2
_
cos(
3

2

1
)
dv
2
(t)
dt

2
v
2
(t) sin(
3

2

1
)
_
, (10.12)
i
3
(t) = C
dv
3
(t)
dt
+
a
2
V
1
V
2
V
3
_
cos(
3

2
t
1
)
dv
3
(t)
dt
+
3
v
3
(t) sin(
3

2

1
)
_
. (10.13)
4
Taking the Fourier transform of (10.11)-(10.13), we obtain the admittances Y
i
(i = 1, 2, 3) seen by
the signal, idler and pump circuits:
Y
1
=
I
1
(j)
V
1
(j)
= j
1
C +j
1
a
2
V
2
V
3
V
1
exp[j(
3

2

1
)] , (10.14)
Y
2
=
I
2
(j)
V
2
(j)
= j
2
C +j
2
a
2
V
1
V
3
V
2
exp[j(
3

2

1
)] , (10.15)
Y
3
=
I
3
(j)
V
3
(j)
= j
3
C +j
3
a
2
V
1
V
2
V
3
exp[j(
3

2

1
)] . (10.16)
The current-voltage relations for the three circuits are given by
I
s
(j) =
_
G
T
+j
1
a
2
V
2
V
3
V
1
exp[j(
3

2

1
)]
_
V
1
(j) , (10.17)
O =
_
G
2
+j
2
a
2
V
1
V
3
V
2
exp[j(
3

2

1
)]
_
V
2
(j) , (10.18)
I
P
(j) =
_
G
3
+j
3
a
2
V
1
V
2
V
3
exp[j(
3

2

1
)]
_
V
3
(j) . (10.19)
Here I
s
(j) and I
P
(j) are the Fourier transforms of the input signal and pump currents, respec-
tively, and G
T
= G
s
+G
L
+G
1
. The LC circuit resonant condition (10.6) is used.
By eliminating V
2
and V
3
from (10.17) using (10.18) and (10.19), we obtain the admittance of
the signal circuit,
Y
s
= G
T
G = G
T


1

2
a
2
2
G
2
G
2
3
[I
P
(j)[
2
_
1 +

2

3
G
2
G
3
a
2
2
V
2
1
_
2
. (10.20)
There emerges a negative conductance due to the nonlinear capacitance driven by the pump wave
at
3
. If V
1
satises the condition,

3
G
2
G
3
a
2
2
V
2
1
1 , (10.21)
the negative conductance is independent of the signal input and the linear parametric amplication
is realized.
5
10.1.2 Power gain
The power gain G of the non-degenerate parametric amplier is given by the ratio of the
power delivered to the load G
L
to the input power to the source G
s
:
G =
G
L
V
2
1
([I
s
[
2
/4G
s
)
=
4G
s
G
L
[Y
s
[
2
. (10.22)
When there is no pump ([I
P
(j)[ = 0), the amplier has no gain (Y
s
= G
T
). When the pump
current reaches the threshold:
[I
P
(j)[
2
=
G
T
G
2
G
2
3

2
a
2
2
, (10.23)
the system becomes unstable ([Y
s
[ 0) and the amplier starts to oscillate. Between these two
extreme conditions, linear amplication of the input signal is provided as far as the signal is not
too strong, i.e. (10.21) is satised.
10.1.3 Noise gure
The noise in a parametric amplier is generated by the circuit conductance G
s
, G
1
and G
2
.
The noise generated by the pump circuit conductance G
3
can be normally neglected because the
pump current i
P
(t) is usually very large and well approximated as a noise-free sinusoidal wave.
The noise from the load conductance G
L
is ignored, because it is usually taken into account in the
following state.
Equation (10.17) suggests that a voltage uctuation V
2
across the idler circuit at frequency

2
results in a current uctuation I
s
in the signal circuit at frequency
1
. The spectral density
6
of the voltage uctuation V
2
is given by
S
V
2
() =
_

_
4k
B
T/G
2
(thermal limit)
2h
2
/G
2
(quantum limit)
. (10.24)
The spectral density of the induced current uctuation I
s
is given by
S
I
s2
() =
2
1
[C

[
2
S
V
2
() , (10.25)
where
[C

[ = a
2
V
3
a
2
[I
P
(j)[ /G
3
. (10.26)
The second equality is obtained by neglecting the gain saturation eect (10.21).
The spectral densities of the current generators associated with G
s
and G
1
are
S
Iss
() =
_

_
4k
B
TG
s
2h
1
G
s
, (10.27)
S
I
s1
() =
_

_
4k
B
TG
1
2h
1
G
1
. (10.28)
Since there is no correlation between these three noise sources, the noise gure of the amplier in
the thermal limit can be written as
F =
S
Iss
() +S
I
s1
() +S
I
s2
()
S
Iss
()
= 1 +
G
1
G
s
+

2
1
[C

[
2
G
2
G
s
. (10.29)
From (10.20) and (10.26), we can express [C

[
2
in terms of the negative conductance G,
[C

[
2
=
G
2
G

2
. (10.30)
7
Using (10.30) in (10.29), the noise gure is expressed as
F = 1 +
G
1
G
s
+

1

2
G
G
s
. (10.31)
The noise gure can be reduced to one (ideal amplication) by achieving the negligible internal
loss in the signal circuit (G
1
G
s
) and the large ratio of
1
/
2
1.
The noise gure of the amplier in the quantum limit is, on the other hand, given by
F = 1 +
G
1
G
s
+
G
G
s
. (10.32)
In a high gain amplier G G
s
(G
1
, G
L
G
s
) at the quantum limit, the minimum noise gure is
F
min
= 2(3 dB) instead of F
min
= 1(0 dB) at the thermal limit.
10.2 Degenerate parametric amplier
10.2.1 Principle of operation
When the signal and idler waves have identical frequencies, such a parametric amplier is
called a degenerate parametric amplier and has a unique characteristic. Consider a swing driven
by a person (Fig. 10-2(a)). During one-half cycle (left to right) of the swing, the person makes a
full one cycle (up-down-up). The frequency of the driving person (pump) and that of the driven
swing (signal) satisfy
P
= 2
s
. Figure 10-2(b) is an equivalent LCR circuit of the swing, in which
the driving action of the person is represented by the nonlinear capacitor.
The circuit equations are given by
I =
d
dt
Q =
d
dt
CV
V = RI +L
d
dt
I (10.33)
8
R
L
C
I
V
a) ( b) (
Figure 10-2:
Eliminating the current I from (10.33), we obtain
_
d
2
dt
2
+
R
L
d
dt
+
1
LC
_
V = 0 (10.34)
The solution of a damped harmonic oscillator expressed by (10.34) is
V = V
0
exp
_

Rt
2L
_
exp
_
_
j

2
0

R
2
4L
2
t
_
_
. (10.35)
If the capacitance is modulated at the pump frequency
P
as
C = C
0
[1 C sin(
P
t +)] , (10.36)
(10.34) is modied to
_
d
dt
2
+
R
L
d
dt
+
2
0
_
1 +
C
C
0
sin(
P
t +)
__
V 0 , (10.37)
where
0
= 1/

LC
0
and it is assumed C C
0
. If we assume the solution of (10.37) has the
form,
V = Re 2V
0
exp(t) exp(jt) , (10.38)
9
we obtain
2V
0
exp(t)
__

2
+
R
L
+
2
0
_
cos(t)
_
2 +
R
L

_
sin(t)
_
= 2V
0
exp(t)

2
0
C
2C
0
[sint cos + cos t sin] . (10.39)
By comparing the cos t and sint terms in both sides of (10.39), we have the equations which
determine the new oscillation frequency and amplication/attenuation coecient :

2
=
2
0
+
2
+
R
L
+

2
0
C
2C
0
sin , (10.40)
2 =

0
C
2C
0
cos
R
L
. (10.41)
If = 0 and C >
2RC
0

0
L
, we have a growing solution ( > 0). The energy is provided to the
signal from the pump. If =

2
or
3
2
, there is no energy exchange between the pump and signal
waves. If = , we have an attenuating solution
_
<
R
L
_
. The energy is extracted from the
signal and transferred to the pump.
10.2.2 Phase sensitive amplier
Changing the pump phase from = 0 to = in (10.36) corresponds to shifting the
capacitance modulation by half a pump period, which is equivalent to one-quarter signal period.
That is, one quadrature amplitude of the signal wave corresponding to the = 0 solution is
amplied by a gain coecient 2 =

0
C
2C
0

R
L
but the other quadrature amplitude corresponding
to the = solution is deamplied by an attenuation coecient 2

0
C
2C
0

R
L
. This type of
operation is called a phase sensitive amplier. If the signal wave is expressed by the two quadrature
10
amplitudes a
1
and a
2
as
E
s
= a
1
cos
s
t +a
2
sin
s
t , (10.42)
and the pump phase is set to amplify the cos
s
t component and deamplify the sin
s
t compo-
nent, the two-kinds of input signals with isotropic (phase insensitive) noise are transformed to the
squeezed state as shown in Fig. 10-3. When the input noise is dominated by thermal noise, the
process is called thermal noise squeezing or classical squeezing. When the input noise is dominated
by quantum mechanical zero-point noise, the process is called quantum noise squeezing.
output
a) ( b) (
output
input
input
a
1
a
2
a
1
a
2
Figure 10-3:
10.3 Quantum limit of a linear amplier
The simplied input-output relations for a degenerate parametric amplier are given by
b
s1
=

Ga
s1
, (10.43)
b
s2
=
1

G
a
s2
, (10.44)
where a
s1
(b
s1
) and a
s2
(b
s2
) are the cos
s
t and sin
s
t components of the input (output) signal
waves. Equation (10.43) indicates that the amplication of one quadrature component does not
11
introduce any additional noise. The noise gure of this amplier is
F =
b
2
s1
)
Ga
2
s1
)
= 1 (0dB) . (10.45)
The sacrice of the noise-free amplication is the loss of the signal information stored in the other
quadrature a
s2
since that quadrature component is deamplied.
The input-output relations for a nondegenerate parametric amplier are given by
b
s1
=

Ga
s1
+

G 1a
i1
, (10.46)
b
s2
=

Ga
s2

G 1a
i2
, (10.47)
where a
i1
and a
i2
are the cos
i
t and sin
i
t components of the input idler wave. The nondegenerate
parametric amplier allows the extraction of the two quadrature information simultaneously, but
the amplier introduces the additional noise. The minimum noise gure in this case is
F =
b
2
s1
)
Ga
2
s1
)
= 2 (3 dB) , (10.48)
where it is assumed that the signal and idler carries the identical noise, i.e. a
2
s1
) = a
2
i1
). This
typical situation corresponds to the case that the input signal wave is in a coherent state and the
input idler wave is in a vacuum state (no input), where a
2
s1
) = a
2
i1
) =
1
4
.
A microwave nondegenerate parametric amplier is dominated by thermal noise rather than
quantum noise. In such a case, cooling the idler input port to below the noise equivalent temperature
of the signal channel is eective to reduce the noise gure. Indeed, the noise gure of close to 0 dB
is achieved in a microwave nondegenerate parametric amplier by this technique.
12
References
[1] A. von der Ziel, J. Appl. Phys. 19, 999 (1948).
[2] J. M. Manley and H. E. Rowe, Proc. IRE 47, 2115 (1959).
[3] M. Uenohara, Low Noise Amplication, Handbuch der Physik, vol. 23 (Springer, Berlin, 1962).
[4] J. A. Giordmaine and R. C. Miller, Phys. Rev. Lett. 14, 973 (1965).
[5] S. E. Harris, M. K. Oshman and R. L. Byer, Phys. Rev. Lett. 18, 732 (1967).
[6] H. A. Haus and J. A. Mullen, Phys. Rev. 128, 2407 (1962).
[7] C. M. Caves, Phys. Rev. D23, 1693 (1981).
Figure Captions
Fig. 10-1: An equivalent circuit of a non-degenerate parametric amplier.
Fig. 10-2: A swing driven by a person and an equivalent circuit.
Fig. 10-3: The input and output signals of a degenerate parametric amplier.
13
Contents
11.Optical Communication Systems 2
11.1 Regeneration in Digital Communication Systems . . . . . . . . . . . . . . . . . . . . 2
11.2 Direct Detection and Optical Preamplier in PCM-IM Communication Systems . . . 3
11.3 Optical Repeater Ampliers in PCM-IM Systems . . . . . . . . . . . . . . . . . . . . 10
11.4 Fundamental Limits of Communication Systems . . . . . . . . . . . . . . . . . . . . 12
11.4.1 Channel Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
11.4.2 Quantum Limit of Communication Systems . . . . . . . . . . . . . . . . . . . 15
11.4.3 Thermal Limit of Communication Systems . . . . . . . . . . . . . . . . . . . 18
1
11. Optical Communication Systems
In optical communication systems [both pulse code modulation-intensity modulation (PCM-
IM) and coherent modulation-demodulation systems], a weak optical signal attenuated by ber
loss is amplied by an optical linear amplier before the signal is detected by a photodetector
and electronically regenerated. An optical amplier adds extra noise on the signal and thus the
signal-to-noise (S/N) ratio of the optical signal wave is always degraded. However, an electronic
regeneration circuit has much higher (thermal) noise than noise generated in an optical amplifer;
thus, the use of an optical amplier improves the overall S/N ratio if the thermal noise of the
receiver circuit is taken into account.
11.1 Regeneration in Digital Communication Systems
In digital communication systems, the information is carried by the binary states of the
amplitude, phase, and frequency of the carrier wave. A transmission line such as an optical ber
attenuates the signal power level through its loss and broadens the signal pulse duration through its
dispersion. Once the signal is distorted and buried in the background noise, information cannot be
retrieved. One of the most important advantages of digital communication systems is that, before
information is completely lost due to transmission line loss and dispersion, the signal is detected
and regenerated.
How often it is necessary to regenerate the signal is an important factor for evaluating the
system because the cost and maintenance of the system critically depend on the number of such
regenerative repeaters. The signal-to-noise (S/N) ratio is gradually degraded by the transmission
2
loss but the original S/N ratio can be recovered after regeneration. The penalty for the recovery
of the original S/N ratio is a nite rate of erroneous regeneration. Even though the transmitted
signal was state 0, a decision circuit may report state 1 and so the signal pulse of state 1 is
transmitted to the next section. A denite advantage of the regeneration in digital communication
systems is that the error rate increases only linearly with the number of regenerative repeaters. For
instance, if each regenerative repeater introduces a bit error rate of P
e
= 10
9
while the total bit
error rate of the system should be less than P
e
= 10
6
, one can install 10
3
regenerative repeaters
in the whole system. The total system length for P
e
= 10
6
can thus be increased 10
3
times longer
than the transmission length for P
e
= 10
9
. Without a regenerative repeater, the bit error rate
increases exponentially with transmission loss, and thus there is no substantial dierence between
the system length for P
e
= 10
9
and that for P
e
= 10
6
.
11.2 Direct Detection and Optical Preamplier in PCM-IM Communication
Systems
Consider a communication system in which information is carried by the presence or absence
of photons per pulse [pulse code modulation-intensity modulation (PCM-IM)]. As shown already
in Chapter 2 (Mathematical Methods) and Chapter 9 (Lasers), the optical pulse from a laser
transmitter does not necessarily contain the xed number of photons, but, rather, the photon
number uctuates from pulse to pulse due to a random deletion process in a ber. The on-pulse
has Poissonian photon statistics with an average photon number n,
p
on
(n) =
e
n
n
n
n!
, (11.1)
3
but the o-pulse always has zero photons, as shown in Fig. 11-1. If no photons are received, assume
an o-pulse was sent; if more than one photon is received, assume an on-pulse was sent. In such
an ideal noiseless communication system, the optimum decision level is 0.5 photon. There is
still a nite probability of error expected in this decision process because the on-pulse has a nite
probability of having no photons due to Poissonian photon statistics. If one desires an error rate
P
e
smaller than 10
9
(which is a required bit error rate in a commercial system), one needs the
average photon number of the on-pulse to be larger than
n > 21 P
e
p
on
(n = 0) = e
n
10
9
. (11.2)
This is the ultimate limit on the receiver sensitivity in PCM-IM optical communication systems.
P(n)
0 1
photon number n
P(n=0)
2 3
Poisson distribution
Figure 11-1:
However, a real electronic receiver circuit cannot achieve this ultimate sensitivity because there
is huge thermal noise generated in the electronic amplier following a photodiode. An optical
amplier which adds extra noise to the amplied signal can still improve the overall S/N ratio
because of the thermal noise of the electronic amplier.
4
Optical
Preamlifier
photodetector
amlified
signal
baseband amlifier
equalizer
decision circuit
received
signal
+
V
ant
>
<
D
R
L
Figure 11-2:
A typical optical receiver conguration with an optical preamplier is shown in Fig. 11-2.
A photodiode converts a photon ux to an electrical current and generates an electrical voltage
across the load resistance R
L
. The voltage across the load resistance is amplied by the baseband
electronic amplier and is ltered by the equalizer to suppress the spurious noise outside the signal
band. The intersymbol interference (overlap of the pulses) must be simultaneously minimized by
the optimum design of the equalizer. Both the load resistance and baseband electronic amplier
add thermal noise to the signal voltage; thus the voltage pulse input to the decision circuit carries
not only the optical noise (Poissonian photoelectron statistics), but also the electrical noise, which
is well approximated by the Gaussian electron statistics, as shown in Chapter 2. It is convenient
to introduce the input-equivalent variance for the noise electron counts to evaluate the eect of
thermal noise:
n
2

th
=
2k
B
F
t
q
2
R
L
T , (11.3)
where is the temperature, F
t
is the noise gure of the baseband electronic amplier (see Chapter
6), R
L
is the load resistance, and T is the pulse duration which is inverse of the bit rate. A higher
5
load resistance R
L
results in lower thermal noise; however, there is an upper limit on R
L
. If the
photodiode has a junction capacitance C, the response time CR
L
of the photodiode circuit must
be shorter than the pulse duration T (i.e., CR
L
T). Otherwise, the voltage pulse broadens to
overlap with each other and introduce an undesired interference eect between consecutive pulses.
Therefore, one obtains
n
2

th

2k
B
C
q
2
F
t
. (11.4)
The RHS of (11.4) for F
t
= 1 (0 dB) is the parameter which determines the collective Coulomb
blockade in a pn junction photodiode (see Chapter 5). If the electrostatic energy q
2
/2C created by
a single photoelectron is greater than the thermal noise k
B
, the input equivalent variance for the
thermal noise electron count can be less than one and the ultimate sensitivity in (11.2) is achieved.
Unfortunately, an ordinary photodiode has a large junction capacitance, C 1 10 pF. At room
temperature ( 300
o
K) and for C 3 pF, the input-equivalent variance for the thermal noise
electron counts (n
2

th
) is on the order of 10
6
, which is much larger than the ultimate sensitivity
of the received photon number n = 21 required to achieve the bit error rate P
e
= 10
9
.
If one does not use an optical preamplier, one needs many more photons ( 10
4
) per pulse to
achieve the S/N ratio of 20 dB
_
n
2
n
2

th
=
10
8
10
6
= 10
2
_
, for which the bit error rate is approximately
equal to P
e
= 10
9
. As shown in Fig. 11-3, the voltage distributions for the on- and o-pulses are
completely determined by Gaussian electrical noise rather than Poissonian optical noise. The bit
error rate is given by
P
e
=
1
2
_
_
_
1
_
2
2
0
_

v
T
exp
_

(v v
0
)
2
2
2
0
_
dv +
1
_
2
2
1
_
v
T

exp
_

(v v
1
)
2
2
2
1
_
dv
_
_
_
. (11.5)
6
Since
2
0
=
2
1
= n
2

th
10
6
and v
0
= 0, the optimum threshold voltage for the decision is equal
to v
T
=
1
2
v
1
and one obtains
P
e
=
1
_
2
2
0
_

v
T
exp
_

v
2
2
2
0
_
dv
1 erf
_
v
T

0
_
, (11.6)
where
erf(x) =
1

2
_
x

exp
_

2
2
_
dx

. (11.7)
In order to achieve P
e
= 10
9
, the S/N ratio
v
2
1

2
1
=
n
2
n
2

th
= 10
2
is required, which corresponds to
the minimum detectable photon number as high as n 10
4
(500 times larger than the theoretical
limit!).
10
4
decision
level
error
probability
output voltage
(in units of photon number)
10
3 ~
0
n
2
2
1
/
on-pulse off-pulse
voltage distribution P(n)
Figure 11-3:
If one uses an ideal optical preamplier in front of the photodetector, as shown in Fig. 11-2,
the average and variance of the amplied signal photon number for the on- and o-pulses are given
7
by
n
out
=
_

_
Gn
in
+ G1 : on pulse
G1 : o pulse
(11.8)
n
2
out
=
_

_
G
2
n
2
in
+G(G1) (n
in
+ 1) : on pulse
G(G1) : o pulse
(11.9)
Equations (11.8) and (11.9) are derived from the relation for the input and output of an ideal
laser amplier which was discussed in Chapter 10. The average photon number for the on-pulse is
the sum of the amplied signal Gn
in
and the spontaneous emission (G 1), while that for the
o-pulse is only the spontaneous emission (G1). The variance for the on-pulse can be rewritten
to represent the physical origins of the various noise:
n
2
out
=
2G(G1)n
in
+ (G1)
2
+ G
2
n
2
in

ex
+ Gn
in
+ G1

signal spont. spont. spont. amplied signal spont. emission
emission beat noise emission beat noise excess noise shot noise shot noise
(11.10)
Here, the input variance n
2
in
is split into the (intrinsic) Poisson noise part n
in
and the excess
noise part n
2
in

ex
= n
2
in
n
in
. The classical interpretation for each term of n
2
out
is given
in (11.10). The output S/N ratio for the on-pulse is
(S/N)
out

n
out

2
n
2
out

=
G
2
n
in

2
G
2
n
2
in
+G(G1)(n
in
+ 1)
= (S/N)
in
_
1 +
_
1
1
G
_
n
in
+ 1
n
2
in

_
1
. (11.11)
8
When the input signal does not have excess noise (Poisson limit), n
2
in
= n
in
, and the average
photon number is much larger than one, n
in
1, one obtains
(S/N)
out

1
2
(S/N)
in
; (11.12)
that is, the S/N ratio is degraded by 3 dB, which is the theoretical limit for the noise gure of
an optical amplier. Even though the S/N ratio is degraded by 3 dB in the optical domain, the
preamplier still improves the overall S/N ratio if the thermal noise generated in the electronic
amplier is taken into account.
The bit error rate of the optical preamplier receiver is calculated by (11.13), where the average
and variance of the normalized voltages are
v
1
= Gn
in
, (11.13)

2
1
2G
2
n
in
+
2k
B
CF
t
q
2
, (11.14)
v
0
= (G1) , (11.15)

2
0
G
2
+
2k
B
CF
t
q
2
. (11.16)
Figure 11-4 shows the output voltage statistics for the on- and o-pulse at the input to the decision
circuit for n
in
= 10
2
and G 10
3
. Figure 11-5 shows the required photon number per pulse to
achieve P
e
= 10
9
vs. preamplier gain. In the limit of high gain, G 1, the required photon
number n
in
is reduced to 10
2
, which is still higher than the ultimate limit of (11.2), but much
smaller than the previous case without optical preampliers, n
in
10
4
.
9
10
5
P(n)
decision
level
error
probability
output voltage n
(in units of photon number)

0
10
3
~

1
10
4
~
Figure 11-4:
0 10 20 30
10
2
10
3
10
4
n
in
for P
e
= 10
-9
optical amplifier gain G(dB)
Figure 11-5:
11.3 Optical Repeater Ampliers in PCM-IM Systems
Next, consider an optical communication system in which ber loss is compensated for by
optical amplier gain, as shown in Fig. 11-6. The S/N ratios at the transmitter output, the rst
ber output, and the rst amplier output are given, respectively, by
(S/N)
0
=
n
2
n
2

= n (Poisson statistics) , (11.17)

(S/N)
1
=
L
2
n
2
L
2
n
2
+L(1 L)n
= Ln (random deletion noise) , (11.18)
10
transmitter Amplifier
fiber
Amplifier
fiber
photon number
per pulse
<n>
L<n>
loss gain
<n>
loss
L<n>
gain
Figure 11-6:
(S/N)
2
=
G
2
L
2
n
2
G
2
Ln +G(G1)(Ln + 1)

L
2
n (3 dB noise gure) . (11.19)
Equation (11.18) accounts for the random deletion noise due to ber loss (Verguess variance the-
orem), which was derived in Chapter 2. The second equality in (11.17) and (11.18) stems from
the assumption that the signal does not have excess noise (i.e., n
2
= n). The S/N ratio is
degraded by the ber loss L and further degraded by 3 dB by the amplier noise. The S/N ratio
at the k-th amplier output is calculated similarly:
(S/N)
2k
=
L
2k
n . (11.20)
In order to achieve the bit error rate of P
e
= 10
9
, the S/N ratio must be larger than 20 dB. If
one assumes n = 10
8
and L =
1
G
= 10
3
(30 dB loss/gain), the number of amplier repeaters
which can be put in the system is k 500. If the total input/output coupling loss of the amplier
and the isolator is 5 dB per section, the net gain is 25 dB, which corresponds to a total ber loss
of 12,500 dB or a total ber length of 62,500 Km if the ber loss is assumed to be 0.2
_
dB
Km
_
. The
transmitter and optical amplier output photon number of 10
8
per pulse corresponds to the optical
power of 10 mW (10 dB
m
) if the bit rate is 1 Gbit/s (T = 10
9
s) and the wavelength is 1.5 m.
11
If the same output power of 10 mW is employed in a 10-Gbit/s system (T = 10
10
s), the average
photon number output per pulse is n = 10
7
. In this case, the maximum number of amplier
repeaters is k 50 and the total ber length is 6250 Km.
If the optical amplier repeater is not used and only the optical preamplier is used, the total
ber loss and ber length necessary to achieve P
e
= 10
9
are only 60 dB and 300 Km for a 1-
Gbit/s system and 50 dB and 250 Km for a 10-Gbit/s system. Without the optical preamplier,
the total ber loss and ber length are, respectively, 40 dB and 200 Km for a 1 Gbit/s system and
30 dB and 150 Km for a 10 Gbit/s system. Therefore, one can see that the linear optical amplier
repeater can enormously expand the electronic terminal repeater spacing and thus is very useful
for intercontinental under-sea communication systems.
11.4 Fundamental Limits of Communication Systems
11.4.1 Channel Capacity
Quantum and thermal noise place fundamental limits on communication systems. Shannons
channel capacity is one way to elucidate such fundamental limits. The channel capacity formulates
how many bits of information can be transmitted over a given channel bandwidth B and S/N ratio:
C = Blog
2
(1 +S/N) (bits/s) . (11.21)
The basis of the Shannon formula is the counting of degrees of freedom (DOF) which arrive
at the receiver through the channel with a bandwidth B. This calculation was performed earlier
by Nyquist. Consider the ltered spectral density with a bandwidth = 2B centered at an
12
Spectral density
0
3
B
-
t
0
2B
F
N
(t)
2
B
-
1
B
-
1
B
2
B
3
B
Figure 11-7:
angular frequency
0
, as shown in Fig. 11-7. The Fourier transform of this rectangular spectrum
results in a Nyquist function, as shown in Fig. 11-7. A sequence of Nyquist functions,
F
N
_
t
k
B
_
=
sin
_
B
_
t
k
B
__
B
_
t
k
B
_ (k = 0, 1, 2, ) , (11.22)
are orthogonal to each other and fully reproduce all bandwidth-limited functions. Each Nyquist
function carries independent information and thus the arrival rate of the DOF in a communication
channel with a bandwidth B is equal to B.
In the intensity modulation scheme, one can assign a dierent number of photons n for each
Nyquist function (DOF) to transmit information. The (information theoretic) entropy of the signal
is described by the probability p(n) of having n photons:
H =

n
p(n) lnp(n) . (11.23)
One can maximize (11.23) under the constraints

n
p(n) = 1 (11.24)
13
and

n
np(n) = n , (11.25)
Equation (11.25) indicates that the average photon number is xed in the entropy maximization
procedure and the result is given by
H
max
= n ln
_
1 +
1
n
_
+ ln(1 +n) , (11.26)
where p(n) satises the so-called geometrical (thermal) distribution:
p(n) =
n
n
(1 +n)
n+1
. (11.27)
The channel capacity is simply the product of the arrival rate B of the DOF and the maximum
entropy H
max
per DOF:
C = B
_
n ln
_
1 +
1
n
_
+ ln(1 +n)
_
. (11.28)
The channel capacity in (11.28) is expressed in terms of natural digits (nats) per second and, in
order to translate it into ordinary bits per second, one can divide (11.28) by n 2. The rst term on
the RHS is the product of the arrival rate of photon Bn (photons/s) and the information carried
by the number of DOFs per photon,
1
n
, i.e., H = log
_
1 +
1
n
_
; this term is called photon
entropy or particle entropy. The second term on the RHS is the product of the arrival rate
of DOF, B (DOF/s), and the information carried by the number of photons per DOF n, i.e.,
H = log(1 +n); this term is called wave entropy. As shown in Fig. 11-8, the channel capacity
C is dominated by the particle entropy when n 1 and is dominated by the wave entropy when
n 1.
14
The implicit assumption for the channel capacity (11.28) is that each pulse has a denite number
of photons without uctuation and the statistics of the photon number over many pulses obey the
geometric distribution (11.27). It is also assumed that the photon counter is free from any noise and
thus the photon number of each pulse is determined without error. The channel capacity provides
the ultimate upper bound for such an ideal communication system.
1
total capacity
n
C/B
10
10
-1
10
-2
10
-3
10 10
-2
10
-1
10
3
10
2
1
wave entropy
ln (1+ n )
1
n
particle entropy
<n> ln (1+ )
Figure 11-8:
11.4.2 Quantum Limit of Communication Systems
The minimum energy cost to transmit one bit of information is calculated by
E
h
0
Bn
_
C
ln 2
_ =
h
0
ln2
1
n
ln(1 +n) + ln
_
1 +
1
n
_ . (11.29)
The second term in the denominator goes to innity when n goes to zero, which means that
the minimum energy cost to transmit one bit of information is reduced to zero in the limit of
n 0. In other words, in principle, one photon can transmit innite bits of information. The
information is carried by the particle entropy of a photon in such a case.
Consider a pulse position modulation (PPM), as illustrated in Fig. 11-9. Each word with a
15

time slot
word T = M x
Figure 11-9:
time duration T is divided into M slots with a time interval = T/M. A single photon arbitrarily
occupies one slot out of M slots in each word; thus the information per word (i.e., per photon)
is given by log
2
M (bits). When M becomes much larger than one, the average photon number
per DOF, which corresponds to each slot, n =
1
M
, becomes much smaller than unity and the full
particle entropy is approached by the PPM signal:
C
PPM
=
1
T
lnM Bn ln
_
1 +
1
n
_
. (11.30)
Obviously, C
PPM
/Bn (nats/photon) goes to innity in the limit of M (n 0).
2B
2
T

1

2

3

4

5

6

7

word
t

1

2

3

6
Figure 11-10:
16
Another example of a modulation scheme for achieving more than one bit of information per
photon is multiple-frequency shift keying (MFSK), as shown in Fig. 11-10. Each word with a
time duration T is occupied by a photon of a dierent color. If the total channel bandwidth is
2B, there are M =
2B
(
2
T
)
= BT independent colors that can be accommodated in this
channel. A single-color photon occupies one word out of M dierent colors, so that the information
per word (i.e., per photon) is again given by log
2
M (bits). The channel capacity is the same as
the PPM case:
C
MFSK
=
1
T
lnM Bn ln
_
1 +
1
n
_
. (11.31)
The above examples illustrate that there is no fundamental quantum limit as far as the minimum
energy cost per bit is concerned and a nite photon energy h
0
does not impose a minimum energy
cost per bit of information.
Next, consider the minimum time-energy product necessary to transmit one bit of information.
The time duration required to transmit one bit of information is given by the inverse of the bit
rate:
T
ln2
C
=
ln2
B
_
n ln
_
1 +
1
n
_
+ ln(1 +n)
_ . (11.32)
Since the maximum available bandwidth B is on the order of the carrier frequency f
0
=
0
/2, the
time-energy product is written as
ET
2(ln2)
2
h
n
_
ln
_
1 +
1
n
_
+
1
n
ln(1 +n)
_
2
. (11.33)
As shown in Fig. 11-11, the time-energy product takes the minimum value ET
h
2
when n
is equal to one. This relation is often referred to as the Bohrs time-energy uncertainty product;
17
that is, when one tries to minimize the product of the energy cost and time duration per one bit of
information, the quantum limit emerges and places a fundamental limit on communication systems
equal to the Planck constant.
ET
2
h
10
2
1
10
10
-2
10
-1
10
2
1 10
N
Figure 11-11:
11.4.3 Thermal Limit of Communication Systems
The above argument holds when the temperature of the communication system is absolute
zero. When the temperature is nite, there are nite (thermal) noise photons n
th
in each DOF
as well as signal photons n
s
. Suppose the total photon number n = n
s
+ n
th
obeys the thermal
distribution (11.27). In this case, the maximum entropy is also equal to (11.26), but this maximum
entropy cannot be fully extracted as useful information due to the presence of (thermal) noise
photons. If the photon counter reports a photon number n
0
, the signal photon number is not
18
necessarily equal to n
0
. The probability density for n
s
has the following distribution:
P
meas
(n
s
) =
_

_
0 : n
s
> n
0
n
th

n
0
ns
(1+n
th
)
n
0
ns+1
: n
s
< n
0
. (11.34)
This means that, even after the detection of the received photon number n
0
, there is still uncertainty
with respect to the signal photon number n
s
. This residual (noise) entropy is given by
H
noise
=

ns
P
meas
(n
s
) lnP
meas
(n
s
)
= n
th
ln
_
1 +
1
n
th

_
+ ln(1 +n
th
) . (11.35)
The useful information that can be extracted under thermal background photons is thus given
by
I H
max
H
noise
= (n
s
+n
th
) ln
_
1 +
1
n
s
+n
th

_
+ ln(1 +n
s
+n
th
)
n
th
ln
_
1 +
1
n
th

_
ln(1 +n
th
) . (11.36)
One can now calculate the minimum energy cost necessary to transmit one bit of information:
E =
h
0
Bn
s

_
C
ln 2
_

n
s
1
k
B
T ln2 . (11.37)
This is a thermal limit on the minimum energy cost per bit; for instance, at T = 300
o
K and
= 1.5 m, a single photon can transmit
h
0
k
B
T ln2
= 46 bits , (11.38)
or one bit of information costs 0.022 photon.
19
References
Regeneration in Digital Communication Systems
M. Schwartz, Information, Transmission, Modulation and Noise (McGraw-Hill, New York,
1990).
Optical Preampliers and Repeater Ampliers
Y. Yamamoto, IEEE J. Quantum Electron. QE-16, 1073 (1980).
Y. Yamamoto, ed., Coherence, Amplication and Quantum Eects in Semiconductor
Lasers (John Wiley & Sons, New York, 1990).
Channel Capacity
H. Hyquist, AIEE Trans. 47, 617 (1928).
C. E. Shannon, Bell Syst. Tech. J. 27, 379 (1948).
Entropy and Information
L. Brillouin, Science and Information Theory (Academic, New York, 1956).
Fundamental Limits of Optical Communication Systems
Y. Yamamoto and H. A. Haus, Rev. Mod. Phys. 58, 1001 (1986).
20