Professional Documents
Culture Documents
Chapter 9: Commonly Used Models: Narrow-Band Gaussian Noise and Shot Noise
model in communication systems. For example, (t) might be the noise component in the output
where c is termed the center frequency (for example, c could be the actual center frequency of
the above-mentioned IF filter). The quantities c(t) and s(t) are termed the quadrature
components (sometimes, c(t) is known as the in-phase component and s(t) is termed the
where
S()
watts/Hz
-c c
Rad/Sec
Fig. 9-1: Example spectrum of narrow-band noise.
are on a vastly different time scale from cosct. Many periods of cosct occur before there is
notable change in c(t), s(t), R(t) or (t).
A second interpretation can be given for the term narrow-band. This is accomplished in
terms of the power spectrum of (t), denoted as S(). By the Wiener-Khinchine theorem,
S() is the Fourier transform of R(), the autocorrelation function for WSS (t). Since (t) is
real valued, the spectral density S() satisfies
S ( ) 0
(9-4)
S ( ) S ( ).
Figure 9-1 depicts an example spectrum of a narrow-band process. The narrow-band attribute
means that S() is zero except for a narrow band of frequencies around c ; process (t) has a
bandwidth (however it might be defined) that is small compared to the center frequency c.
Power spectrum S() may, or may not, have c as axes of local symmetry. If c is
an axis of local symmetry, then
S ( c ) S ( c ) (9-5)
for 0 < < c, and the process is said to be a symmetrical band-pass process (Fig. 9-1 depicts a
symmetrical band-pass process). It must be emphasized that the symmetry stated by the second
of (9-4) is always true (i.e., the power spectrum is even); however, the symmetry stated by (9-5)
may, or may not, be true. As will be shown in what follows, the analysis of narrow-band noise is
simplified if (9-5) is true.
To avoid confusion when reviewing the engineering literature on narrow-band noise, the
reader should remember that different authors use slightly different definitions for the cross-
correlation of jointly-stationary, real-valued random processes x(t) and y(t). As used here, the
cross-correlation of x and y is defined as Rxy( ) E[x(t+)y(t)]. However, when defining Rxy,
some authors shift (by ) the time variable of the function y instead of the function x.
Fortunately, this possible discrepancy is accounted for easily when comparing the work of
different authors.
(t) has Zero Mean
The mean of (t) must be zero. This conclusion follows directly from
The WSS assumption means that E[(t)] must be time invariant (constant). Inspection of (9-6)
leads to the conclusion that E[c] = E[s] = 0 so that E[] = 0.
Quadrature Components In Terms of and
Let the Hilbert transform of WSS noise (t) be denoted in the usual way by the use of a
circumflex; that is, (t) denotes the Hilbert transform of (t) (see Appendix 9A for a discussion
of the Hilbert transform). The Hilbert transform is a linear, time-invariant filtering operation
applied to (t); hence, from the results developed in Chapter 7, (t) is WSS.
In what follows, some simple properties are needed of the cross correlation of (t) and
(t) . Recall that (t) is the output of a linear, time-invariant system that is driven by (t). Also
recall that techniques are given in Chapter 7 for expressing the cross correlation between a
system input and output. Using this approach, it can be shown easily that
R ( ) E[(t ) (t)] R ( )
(t )(t)] R ( )
( ) E[
R
(9-7)
(0) R (0) 0
R
R ( ) R ( ) .
Equation (9-1) can be used to express (t) . The Hilbert transform of the noise signal can
be expressed as
(t) c (t) cos c t s (t) sin c t c (t) cos c t s (t) sin c t (9-8)
This result follows from the fact that c is much higher than any frequency component in c or
s so that the Hilbert transform is only applied to the high-frequency sinusoidal functions (see
Appendix 9A).
The quadrature components can be expressed in terms of and . This can be done by
solving (9-1) and (9-8) for
(t)sin c t
c (t) (t)cos c t
(9-9)
(t) cos c t (t)sin c t .
s (t)
R c ( ) E[c (t)c (t )]
(t )]cos c t sin c (t ) E[
E[(t) (t)
(t )]sin c t sin c (t ) .
The same procedure can be used to compute an identical result for R s ; this leads to the
conclusion that
R c ( ) R s ( ) (9-12)
for all . Note that c and s are zero mean, Gaussian, and wide-sense stationary.
Equations (9-11) and (9-12) allow the computation of R c and R s given only R .
However, given only R c and/or R s , we do not have enough information to compute R
(what is missing?).
A somewhat non-intuitive result can be obtained from (9-11) and (9-12). Set = 0 in the
last two equations to conclude that
1
Sc() Ss()
2
S ( c ) S ( c )
1
2
sgn( c )S ( c ) sgn( c )S ( c ) (9-15)
1 1
2
1 sgn( c ) S ( c ) 1 sgn( c ) S ( c )
2
Since c and s are low-pass processes, Equation (9-15) can be simplified to produce
Sc( ) Ss( ) S ( c ) S ( c ), c c
(9-16)
0, otherwise,
Equation (9-16) provides an easy method for obtaining Sc and/or Ss given only S .
First, make two copies of S. Shift the first copy to the left by c, and shift the second copy
to the right by c. Add together both shifted copies, and truncate the sum to the interval c
c to get Sc . This shift and add procedure for creating Sc is illustrated by Fig. 9-2.
Given only S, it is always possible to determine Sc (which is equal to Ss ) in this manner.
The converse is not true; given only Sc , it is not always possible to create S (Why? Think
c c
Sc) , < c
c c
Sc) , < c
c c
c c
about the fact that Sc( ) must be even, but S may not satisfy (9-5);what is missing?).
Crosscorrelation R c s
It is easy to compute the cross-correlation of the quadrature components. From (9-9) it
follows that
R c s ( ) E[c (t )s ( t )]
(t )
E[ (t)]sin c (t )cos c t E[
(t )(t)]sin c (t )sin c t .
R c s () sin c cos c R ()
.
R c () cos c sin c R ()
R () sin c cos c R c s ()
, (9-19)
R () cos c sin c R c ()
S ( c ) S ( c ) (9-21)
for 0 < < c. Such a bandpass process has its center frequency c as an axis of local
symmetry. In nature, symmetry usually leads to simplifications, and this is true of Gaussian
narrow-band noise. In what follows, we show that the local symmetry stated by (9-21) is
equivalent to the condition R c s ( ) 0 for all (not just at = 0).
The desired result follows from inspecting the Fourier transform of (9-18); this transform
is the cross spectrum of the quadrature components, and it vanishes when the narrow-band
process has spectral symmetry as defined by (9-21). To compute this cross spectrum, first note
the Fourier transform pairs
R () S ()
(9-22)
R () jsgn()S () ,
where
1 for > 0
sgn() (9-23)
1 for 0
is the commonly used sign function. Now, use Equation (9-22) and the Frequency Shifting
Theorem to obtain the Fourier transform pairs
1
R ( ) sin c S ( c ) S ( c )
2j
(9-24)
1
R ( ) cos c Sgn( c ) S ( c ) Sgn( c ) S ( c ) .
2j
Finally, use this last equation and (9-18) to compute the cross spectrum
a) S()
UL L U
-2c -c c 2c
b) S(-c)
1-Sgn(-c)
UL LU
-2c -c c 2c
c) S(+c)
1+Sgn(+c)
UL LU
-2c -c c 2c
Figure 9-3: Symmetrical bandpass processes have c(t1) and s(t2)
uncorrelated for all t1 and t2.
Sc s ( ) F [ R c s ( )]
(9-25)
1
S ( c )[1 Sgn( c )] S ( c )[1 Sgn( c )] .
2j
Figure 9-3 depicts example plots useful for visualizing important properties of (9-25).
From parts b) and c) of this plot, note that the products on the right-hand side of (9-25) are low
pass processes. Then it is easily seen that
0 , c
Scs ( ) j[ S ( c ) S ( c )], c c (9-26)
0 , c .
R ( ) R c ( ) cos c . (9-27)
Also, the spectrum S of the noise is obtained easily by scaling and translating Sc F [R c ]
as shown by
1
S ( ) [Sc ( c ) Sc ( c )] . (9-28)
2
This result follows directly by taking the Fourier transform of (9-27). Hence, when the process
is symmetrical, it is possible to express S in terms of a simple translations of Sc (see the
comment after (9-20)). Finally, for a symmetrical bandpass process, Equation (9-16) simplifies
to
Sc( ) Ss( ) 2 S ( c ), c c
. (9-29)
0, otherwise
L C
+
S() = N0/2 +
watts/Hz
(WGN) R
-
Figure 9-4: A simple band-pass filter driven by white Gaussian
noise (WGN).
Example 9-1: Figure 9-4 depicts a simple RLC bandpass filter that is driven by white Gaussian
noise with a double sided spectral density of N0/2 watts/Hz. The spectral density of the output is
given by
2
N 2 N 2 0 ( j )
S ( ) 0 H bp ( j ) 0 , (9-30)
2 2 ( 0 j )2 c2
where 0 = R/2L, c = (n2 - 02)1/2 and n = 1/(LC)1/2. In this result, frequency can be
normalized, and (9-30) can be written as
2
N0 2 0 ( j )
S ( ) , (9-31)
2 ( 0 j )2 1
where 0 = 0/c and = /c. Figure 9-5 illustrates a plot of the output spectrum for o = .5;
note that the output process is not symmetrical. Figure 9-6 depicts the spectrum for o = .1 (a
much sharper filter than the o = .5 case). As the circuit Q becomes large (i.e., o becomes
small), the filter approximates a symmetrical filter, and the output process approximates a
symmetrical bandpass process.
Envelope and Phase of Narrow-Band Noise
Zero-mean quadrature components c(t) and s(t) are jointly Gaussian, and they have the
same variance 2 = R ( 0) R c ( 0) R s ( 0) . Also, taken at the same time t, they are
S() S()
1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
-2 -1 0 1 2 -2 -1 0 1 2
(radians/second) (radians/second)
Figure 9-5: Output Spectrum for o = .5 Figure 9-6: Output Spectrum for o = .1
independent. Hence, taken at the same time, processes c(t) and s(t) are described by the joint
density
f ( c , s )
1
exp
LM
c2 s2
.
OP (9-32)
2 2 MN
2 2 PQ
We are guilty of a common abuse of notation. Here, symbols c and s are used to denote
random processes, and sometimes they are used as algebraic variables, as in (9-32). However,
always, it should be clear from context the intended use of c and s.
The narrow-band noise signal can be represented as
where
are the envelope and phase, respectively. Note that (9-34) describes a transformation of c(t)
and s(t). The inverse is given by
c 1 cos(1 )
(9-35)
s 1 sin(1 )
The joint density of 1 and can be found by using standard techniques. Since (9-35) is
the inverse of (9-34), we can write
(c , s )
f (1, 1 ) f (c , s ) (9-36)
(1, 1 ) c 1 cos 1
s 1 sin1
(c , s ) cos 1 1sin1
det 1
(1, 1 ) sin1 1 cos 1
(again, the notation is abusive). Finally, substitute (9-32) into (9-36) to obtain
1 1
f ( 1, 1 ) exp 2 12 (sin 2 1 cos2 1 )
2
2 2
. (9-37)
1 1 2
exp 1 ,
22 22
Fig. 9-7: A hypothetical sample function of narrow-band Gaussian noise. The envelope is
Rayleigh and the phase is uniform.
f ( 1, 1 ) f ( 1 )f(1 ) , (9-38)
where
LM
1
f ( 1 ) 12 exp 2 12 U( 1 )
OP (9-39)
2N Q
describes a Rayleigh distributed envelope, and
1
f(1 ) , - < 1 (9-40)
2
describes a uniformly distributed phase. Finally, note that the envelope and phase are
independent. Figure 9-7 depicts a hypothetical sample function of narrow-band Gaussian noise.
Envelope and Phase of a Sinusoidal Signal Plus Noise - the Rice Density Function
Many communication problems involve deterministic signals embedded in random noise.
The simplest such combination of signal and noise is that of a constant frequency sinusoid added
to narrow-band Gaussian noise. In the 1940s, Steven Rice analyzed this combination and
published his results in the paper Statistical Properties of a Sine-wave Plus Random Noise, Bell
System Technical Journal, 27, pp. 109-157, January 1948. His work is outlined in this section.
where A0, c, and 0 are known constants. To signal s(t) we add noise (t) given by (9-1), a
zero-mean WSS band-pass process with power 2 = E[2] = E[c2] = E[s2]. This sum of signal
and noise can be written as
where
are the envelope and phase, respectively, of the signal+noise process. Note that the quantity
(A0 / 2 )2 / 2 is the signal-to-noise ratio, a ratio of powers.
Equation (9-43) represents a transformation from the components c and s into the
envelope 2 and phase 2. The inverse of this transformation is given by
Note that constants A0cos0 and A0sin0 only influence the means of 2cos( 2) and 2sin( 2)
(remember that both c and s have zero mean!). In the remainder of this section, we describe
the statistical properties of envelope 2 and phase 2.
At the same time t, processes c(t) and s(t) are statistically independent (however, for
0, c(t) and s(t+) may be dependent). Hence, for c(t) and s(t) we can write the joint
density
exp[( c2 s2 ) / 2 2 ]
f ( c , s ) (9-45)
2 2
(we choose to abuse notation for our convenience: c and s are used to denote both random
processes and, as in (9-45), algebraic variables).
The joint density f( 2) can be found by transforming (9-45). To accomplish this, the
Jacobian
(c , s ) cos 2 2sin2
det 2 (9-46)
( 2 , 2 ) sin2
2 cos 2
(c , s )
f ( 2 , 2 ) f (c , s ) (9-47)
( 2 , 2 ) c 2 cos 2 A0 cos 0
s 2 sin2 A 0 sin 0
f ( 2 , 2 )
2
2 2 2
exp 1 2 [ 22 2A0 2 cos( 2 0 ) A02 ] U( 2 ) .
Now, the marginal density f() can be found by integrating out the 2 variable to obtain
2
f ( 2 ) f ( 2 , 2 ) d 2
0
(9-48)
A 2
22 exp 1 2 [ 22 A02 ] U( 2 ) 21 exp{ 0 2 2 cos( 2 0 )}d 2 .
2 0
I0 ( ) 21
2
z
exp{ cos( )}d ,
0
(9-49)
the modified Bessel function of order zero. Now, use definition (9-49) in (9-48) to write
A FH IK RS
f ( 2 ) 22 I0 2 2 0 exp 1 2 [22 A 02 ] U( 2 ) , UV
2 T W (9-50)
a result known as the Rice probability density. As expected, 0 does not enter into f(2).
Equation (9-50) is an important result. It is the density function that statistically
describes the envelope 2 at time t; for various values of A0/, the function f(2) is plotted on
Figure 9-8 (the quantity (A0 / 2)2 / 2 is the signal-to-noise ratio). For A0/ = 0, the case of no
sinusoid, only noise, the density is Rayleigh. For large A0/ the density becomes Gaussian. To
observe this asymptotic behavior, note that for large the approximation
e
I 0 ( ) , >> 1, (9-51)
2
becomes valid. Hence, for large 2A0/ Equation (9-50) can be approximated by
f ( 2 )
2 RS UV
exp 1 2 [2 A 0 ]2 U ( 2 ) .
2A 0 2 T
2 W (9-52)
0.8
0.7
A0/ = 0
0.6
f (2)
0.5 A0/ = 1
A0/ = 2
A0/ = 3
0.4 A0/ = 4
0.3
0.2
0.1
0.0
0 1 2 3 4 5 6
Figure 9-8: Rice density function for sinusoid plus noise. Plots
are given for several values of A0/. Note that f is approximately
Rayleigh for small, positive A0/ density f is approximately
Gaussian for large A0/.
For A0 >> , this function has a very sharp peak at 2 = A0, and it falls off rapidly from its peak
value. Under these conditions, the approximation
f ( 2 )
1 RS
exp 1 2 [2 A 0 ]2 UV
2 2 T2 W (9-53)
holds for values of 2 near A0 (i.e., 2 A0) where f(2) is significant. Hence, for large A0/,
envelope 2 is approximately Gaussian distributed.
The marginal density f(2) can be found by integrating 2 out of (9-47). Before
integrating, complete the square in 2 , and express (9-47) as
2
{ } RST UV
2
f ( 2 , 2 ) exp 1 2 [2 A 0 cos( 2 0 )]2 exp A 02 sin 2 ( 2 0 ) U( 2 ) . (9-54)
2 2 2 2 W
f ( 2 ) f ( 2 , 2 )d2
0
(9-55)
2
2
2
2
exp A0 sin 2 ( 2 0 )
0
2 2
exp 12 [ 2 A0 cos( 2 0 )]2 d 2 .
2
On the right-hand-side of (9-55), the integral can be expressed as the two integrals
z0
2
2 2 {
exp 1 2 [2 A 0 cos( 2 0 )]2 d2
2 }
z 2{2
0
A 0 cos( 2 0 )}
4 2 { 2 }
exp 1 2 [2 A 0 cos( 2 0 )]2 d2 (9-56)
A 0 cos( 2 0 )
2 2 z0 { 2 }
exp 1 2 [2 A 0 cos( 2 0 )]2 d2
After a change of variable = [2 - A0cos(2 - 0)]2 , the first integral on the right-hand-side of
(9-56) can be expressed as
0
2{ 2 A0 cos( 2 0 )}
4 2
exp 1 2 [ 2 A0 cos( 2 0 )]2 d 2
2
1
2 2 exp[ 2 ]d (9-57)
42 A0 cos ( 2 0 ) 2
1 A2 cos2 ( 2 0 )
exp 0 .
2 2 2
After a change of variable = [2 - A0cos(2 - 0)]/, the second integral on the right-hand-side
of (9-56) can be expressed as
1
2 0
exp 12 [ 2 A0 cos( 2 0 )]2 d2
2
1
2 (A0 / ) cos[ 2 0 ]
exp 12 2 d
(9-58)
1
1 (A0 / ) cos[ 2 0 ]
2
exp 12 2 d
A
F( 0 cos[ 2 0 ]),
where
F( x ) 1 z
2
x
exp d
2
2
is the distribution function for a zero-mean, unit variance Gaussian random variable (the identity
F(-x) = 1 - F(x) was used to obtain (9-58)).
Finally, we are in a position to write f( ), the density function for the instantaneous
phase. This density can be written by using (9-57) and (9-58) in (9-55) to write
1 A2
f ( 2 ) exp 0
2 2 2
, (9-59)
A cos( 2 0 ) 2 A
0 exp A0 sin 2 ( 2 0 ) F( 0 cos[ 2 0 ])
2 2 2
the density function for the phase of a sinusoid embedded in narrow-band noise. For various
values of SNR and for 0 = 0, density f() is plotted on Fig. 9-9. For a SNR of zero (i.e., A0 =
0), the phase is uniform. As SNR A02/2 increases, the density becomes more sharply peaked (in
general, the density will peak at 0, the phase of the sinusoid). As SNR A02/2 approaches
infinity, the density of the phase approaches a delta function at 0.
1.8
1.6
1.4
f (2)
1.2
A0/ = 4
1.0
0.8
0.6
A0/ = 2
0.4
A0/ = 0 A0/ = 1
0.2
0.0
- -/2 0 /2
Phase Angle 2
Figure 9-9: Density function for phase of signal plus noise A0cos(0t+0) +
{c(t)cos(0t) - s(t)sin(0t)} for the case 0 = 0.
Shot Noise
Shot noise results from filtering a large number of independent and randomly-occuring-
in-time impulses. For example, in a temperature-limited vacuum diode, independent electrons
reach the anode at independent times to produce a shot noise process in the diode output circuit.
A similar phenomenon occurs in diffusion-limited pn junctions. To understand shot noise, you
must first understand Poisson point processes and Poisson impulses.
Recall the definition and properties of the Poissson point process that was discussed in
Chapters 2 and 7 (also, see Appendix 9-B). The Poisson points occur at times ti with an average
density of d points per unit length. In an interval of length , the number of points is distributed
with a Poisson density with parameter d.
Use this Poisson process to form a sequence of Poisson Impulses, a sequence of impulses
located at the Poisson points and expressed as
x(t) z(t)
... ... d/dt ... ...
z(t) (t t i ) , (9-60)
i
where the ti are the Poisson points. Note that z(t) is a generalized random process; like the delta
function, it can only be characterized by its behavior under an integral sign. When z(t) is
integrated, the result is the Poisson random process
n(0, t), t 0
t
x(t) z()d 0, t0 (9-61)
0
-n(0, t) t 0,
where n(t1,t2) is the number of Poisson points in the interval (t1,t2]. Likewise, by passing the
Poisson process x(t) through a generalized differentiator (as illustrated by Fig. 9-10), it is
possible to obtain z(t).
The mean of z(t) is simply the derivative of the mean value of x(t). Since E[x(t)]=dt, we
can write
d
z E[z(t)] E[x(t)] d . (9-62)
dt
This formal result needs a physical interpretation. One possible interpretation is to view z as
t/2
z limit 1t z( )d limit 1t d t random fluctuation with increasing t d . (9-63)
t t / 2 t
For large t, the integral in (9-63) fluctuates around mean dt with a variance of dt (both the
mean and variance of the number of Poisson points in (-t/2, t/2] is dt). But, the integral is
multiplied by 1/t; the product has a mean of d and a variance like d/t. Hence, as t becomes
large, the random temporal fluctuations become insignificant compared to d, the infinite-time-
interval average z.
Important correlations involving z(t) can be calculated easily. Because Rx(t1,t2) = d2 t1t2
+ dmin(t1,t2) (see Chapter 7), we obtain
R xz (t1, t 2 ) R x (t1, t 2 ) d2 t1 d U(t1 t 2 )
t 2
(9-64)
R z (t1, t 2 ) R xz (t1, t 2 ) d2 d (t1 t 2 ) .
t1
Note that z(t) is wide-sense stationary but x(t) is not. The Fourier transform of Rz() yields
Sz ( ) d 2 d2 ( ) , (9-65)
a sum known as shot noise. The basic idea here is illustrated by Fig. 9-11. A sequence of
functions described by (9-60) (i.e., process z(t)) is input to system h(t) to form output shot noise
process s(t). The idea is simple: process s(t) is the output of a system activated by a sequence of
h(t)
s E s(t) E z(t) h(t) h(t) E z(t) d h(t)dt d H(0) . (9-67)
0
2 2 2
Ss ( ) H() Sz ( ) 2 d2 H 2 (0)( ) d H( ) 2s2 ( ) d H( ) . (9-68)
R s ( ) = F -1 Ss ( ) d2 H 2 (0) d H( ) e jd d2 H 2 (0) d ( ) ,
2
(9-69)
2
where
1
H() e jd h(t)h(t )dt .
2
()
2
(9-70)
From (9-67) and (9-69), shot noise has a mean and variance of
s = d H(0)
(9-71)
2
s2 = [ d2 H 2 (0) + d (0)] - [ d H(0)]2 d (0) d H() d ,
2
2
d
s E[s(t)] d R s ( ) d e
2
(9-72)
2
s2 d Ss () 2 d ( ) 2 d 2
2
Because of (9-73), shot noise s at time t depends only on the Poisson impulses in the
interval (t - T, t]. To see this, note that
s(t) h(t ) ( t i )d h(t t i ) , (9-74)
i t T t i t
so that only the impulses in the moving window (t - T, t] influence the output at time t. Let
random variable nT denote the number of Poisson impulses during (t - T, t]. From Chapter 1, we
know that
( dT) k
P[n T k] e d T . (9-75)
k!
Now, the Law of Total Probability (see Ch. 1 and Ch. 2 of these notes) can be applied to write
the first-order density function of the shot noise process s(t) as
( d T) k
fs (x) fs (xnT k)P[nT k] fs (xnT k)edT k!
(9-76)
k 0 k 0
(note that fs(x) is independent of absolute time t). We must find fs(xnT = k), the density of shot
noise s(t) conditioned on there being exactly k Poisson impulses in the interval (t - T, t].
When nT = k = 0, there are no Poisson points in (t - T, t], and we have
At fixed time t, the random variable x2(t) is described by the density g2(x) fs(xnT = 2) = g1g1
(i.e., the convolution of g1 with itself) since h(t - t1) and h(t - t2) are independent and identically
distributed with density g1.
The general case fs(xnT = k) is similar. Let ti, 1 i k, denote independent random
variables, each uniform on (t - T, t]. At fixed time t, the density that describes
is
( d T) k
fs (x) e d T g k (x) . (9-82)
k 0
k!
Convergence is fast, and (9-82) is useful for computing the density fs when dT is small (the case
for low density shot noise), say on the order of 1, so that, on the average, there are only a few
Poisson impulses in the interval (t - T, t]. For the case of low density shot noise, (9-82) cannot
be approximated by a Gaussian density.
fs(x) For An Infinite Duration h(t)
The first-order density function fs(x) is much more difficult to calculate for the general
case where h(t) is of infinite duration (but with finite time constants). In what follows, assume
that infinite duration h(t) is significant only over 0 t < 5max, where max is the filters dominant
time constant. We show that shot noise is approximately Gaussian distributed when d is large
compared to the time interval over which h(t) is significant (so that, on the average, many
Poisson impulses are filtered to form s(t)).
To establish this fact, consider first a finite duration interval (-T/2, T/2), and let random
variable nT, described by (9-75), denote the number of Poisson impulses that are contained in the
interval. Located in (-T/2, T/2), denote the nT Poison Impulse locations as tk, 1 k nT. Use
these nT impulse locations to define time-limited shot noise
nT
sT (t) h(t t k ), t . (9-83)
k 1
In our analysis of shot noise s(t), we first consider the characteristic function
s () E e js limit E e j sT . (9-84)
T
E e j sT E e jsTnT k P nT k , (9-85)
k 0
where P[nT = k] is given by (9-75). In the conditional expectation used on the right-hand-side
of (9-85), output sT results from filtering exactly k impulses (this is different from the not-
conditioned-on-anything sT that appears on the left-hand-side of (9-85)). We model the impulse
locations as k independent, identically distributed (iid they are uniform on (-T/2, T/2)) random
variables. As a result, the terms h(t - ti) in sT(t) are independent so that
,
k
E e j sT nT k E e j sT nT 1 (9-86)
where
1 T/2 jh(t x)
E e j sT nT 1 e dx, (9-87)
T T/2
since each ti is uniformly distributed on (-T/2, T/2). Finally, by using (9-84) through (9-87), we
can write
s () limit E e j sT limit E e j sTnT k P nT k
T T
k 0
k
1 T/2 jh(t x) k d T ( d T)
limit e dx e (9-88)
T k 0 T T/2 k!
k
T/2 e jh(t x) dx
d T
d T/ 2
limit e
T
k 0 k!
T/2
For finite T, the quantity T / 2 e jh(t x) dx depends on t. However, the quantity e jh(t x) dx
does not depend on t. Recalling the Taylor series of the exponential function, we can write
(9-88) as
a general formula for the characteristic function of the shot noise process.
In general, Equation (9-89) is impossible to evaluate in closed form. However, this
formula can be used to show that shot noise is approximately Gaussian distributed when d is
large compared to the time constants in h(t) (i.e., compared to the time duration where h(t) is
significant).
This task will be made simpler if we center and normalize s(t). Define
s(t)- d H(0)
s(t) , (9-90)
d
so that
E s 0
. (9-91)
R s () () h(t)h(t )dt
(see (9-67) and (9-69)). The characteristic functions of s and s are related by
s - d H(0)
s ( ) E e js E exp j exp j d H(0) s ( d ) . (9-92)
d
Use (9-89) and H(0) h(t x)dx on the right-hand-side of (9-92) to write
j j
s ( ) exp d exp h(t x) 1 h(t x) dx . (9-93)
d d
Now, in the integrand of (9-93), expand the exponential in a power series, and cancel out the
zero and first-order terms to obtain
k k
( j) k h(t x)
h(x)
( j) k
s ( ) exp d k! dx exp d k!
dx
d d
k 2 k 2
(9-94)
k
h (x)dx
1 ( j) k
exp 2 h 2 (x)dx exp d k! k
.
2 k 3
d
As d approaches infinity, (i.e., becomes large compared to the time constants in low-pass filter
h(t)), Equation (9-94) approaches the characteristic function of a Gaussian process.
To see this, let max be the dominant (i.e., largest) time constant in low-pass filter h(t).
There exists constant M such that (note that h is causal)
h(x) dx M k e kx / max dx M k max .
k
h k (x)dx (9-96)
0 k
Consider the second exponent on the right-hand-side of (9-94). Using (9-96), it is easily
seen that
h k (x)dx k
( j) k
( j) k M
max
d k! k
. (9-97)
k!
k k 2
k 3 d k 3 d
Now, as d (so that max / d 0) in (9-97), each right-hand-side term approaches zero
so that
max
h k (x)dx 0
k d
( j)
d k! k
0 . (9-98)
k 3
d
Use this last result with Equation (9-94). As max / d 0 with increasing d , we have
( j)2 2
s () exp h (x)dx exp 1 s2 2 , (9-99)
2 2
where
s2 R s (0) (9-100)
is the variance of standardized shot noise s(t) (see (9-91)). Note that Equation (9-99) is the
characteristic function of a zero-mean, Gaussian random variable with variance (9-100). Hence,
shot noise is approximately Gaussian distributed when d is large compared to the dominant
time constant in low-pass filter h(t) (so that, on the average, a large number of Poisson impulses
temperature. Under this condition, between electrons, space charge effects are minimal, and
individual electrons are, more or less, independent of each other.
The basic circuit is illustrated by Figure 9-12. In a random manner, electrons are emitted
by the cathode, and they flow a distance d to the plate to form current i(t). If emitted at t = 0, an
independent electron contributes a current h(t), and the aggregate plate current is given by
where tk are the Poisson-distributed independent times at which electrons are emitted by the
cathode (see Equation (9-66)). In what follows, we approximate h(t).
i(t)
d
RL
+ - - +
Vf Vp
Filament Plate
Figure 9-12: Temperature-limited vacuum
diode used as a shot noise generator.
As discussed above, space charge effects are negligible and the electrons are
independent. Since there is no space charge between the cathode and plate, the potential
distribution V in this region satisfies Laplaces equation
2V
0. (9-102)
2x
The potential must satisfy the boundary conditions V(0) = 0 and V(d) = Vp. Hence, simple
integration yields
Vp
V= x, 0xd. (9-103)
d
As an electron flows from the cathode to the plate, its velocity and energy increase. At
point x between the cathode and plate, the energy increase is given by
Vp
E n (x) eV (x) = e x, (9-104)
d
dE n dE n dx Vp dx
e Vp h , (9-105)
dt dx dt d dt
where h(t) is current due to the flow of a single electron (note that d -1dx/dt has units of sec-1 so
that (e/d ) dx/dt has units of charge/sec, or current). Equation (9-105) can be solved for current to
obtain
e dx e
h vx , (9-106)
d dt d
e Vp
ax = . (9-107)
m d
As it is emitted by the cathode, an electron has an initial velocity that is Maxwellian distributed.
However, to simplify this example we will assume that the initial velocity is zero. With this
assumption, electron velocity can be obtained by integrating (9-107) to obtain
e Vp
vx = t. (9-108)
m d
1 tT e Vp d
vx
tT 0
v x dt =
2m d
tT .
tT
(9-109)
h(t)
2e/tT
tT t
Figure 9-13: Current due to a single electron emitted by
the cathode at t = 0.
2d
vx = t, 0 t tT . (9-110)
t2
T
With the aid of this last relationship, we can determine current as a function of time.
Simply combine (9-106) and (9-110) to obtain
2e
h(t) t, 0 t tT , (9-111)
t2
T
the current pulse generated by a single electron as it travels from the cathode to the plate. This
current pulse is depicted by Figure 9-13.
The bandwidth of shot noise s(t) is of interest. For example, we may use the noise
generator to make relative measurements on a communication receiver, and we may require the
noise spectrum to be flat (or white) over the receiver bandwidth (the noise spectrum
amplitude is not important since we are making relative measurements). To a certain flatness,
we can compute and examine the power spectrum of standardized s(t) described by (9-90). As
given by (9-91), the autocorrelation of s(t) is
R s ( ) 2
e 2 tT
tT 0
t(t + )dt =
4 e2
3 tT
1
2
tT
1
2tT
, 0 tT
R s (), tT 0 . (9-112)
0, otherwise
The power spectrum of s(t) is the Fourier transform of (9-112), a result given by
0
S () 2 R s () cos()d
4
(tT ) 4 (tT )2 2(1 cos tT tT sin tT ) . (9-113)
Plots of the autocorrelation and relative power spectrum (plotted in dB relative to peak power at
= 0) are given by Figures 9-14 and 9-15, respectively.
To within 3dB, the power spectrum is flat from DC to a little over = /tT. For the
Sylvania 5722 noise generator diode, the cathode-to-plate spacing is .0375 inches and the transit
time is about 310-10 seconds. For this diode, the 3dB cutoff would be about 1/2tT = 1600Mhz.
In practical application, where electrode/circuit stray capacitance/inductance limits frequency
range, the Sylvania 5722 has been used in commercial noise generators operating at over
400Mhz.
10Log{S()/S(0)}
Rs() 2
0
Relative Power (dB)
4e 2
3tT -2
-4
-6
-8
-10
-tT tT /tT (Rad/Sec)
Figure 9-14: Autocorrelation function of Figure 9-15: Relative power spectrum of nor-
normalized shot noise process. malized shot noise process.