Anda di halaman 1dari 3

A Note on the Approach of Narrow Band Noise after a Nonlinear Device to a

Normal Probability Density


George R. Arthur

Citation: Journal of Applied Physics 23, 1143 (1952); doi: 10.1063/1.1701998


View online: http://dx.doi.org/10.1063/1.1701998
View Table of Contents: http://scitation.aip.org/content/aip/journal/jap/23/10?ver=pdfcov
Published by the AIP Publishing

[This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded
to ] IP: 2.145.114.250 On: Fri, 20 Dec 2013 03:18:29
] 0 URN A L 0 F AP P LI ~: D P H V SIC S VOLUME 23. NUMBER 10 OCTOBER. 1952

A Note on the Approach of Narrow Band Noise after a Nonlinear Device to


a Normal Probability Density
GEORGE R. ARTHUR*
Dunham Laboratory, Yale University, Nr:w Haven, Connecticut
(Received April 30, 1952)

The output of a low pass filter preceded by a nonlinear device excited by a noise source is given in terms
of an integral equation. The first three central moments of the probability density of the output signal are
obtained from the integral equation and the approach of these moments to those of a Gaussian or normal
density is demonstrated.

C ONSIDERABLE material has been written on the


statistics of noise which bas been passed through
a nonlinear device. However, if the output of the non-
If the expansion of the log of the characteristic func-
tion is compared to the usual expansion of this func-
tion in terms of the semi-invariants of the probability
linear device is passed through a low pass filter, little density 2 corresponding to <I>(u), one has:
is known of the exact statistical nature of that filter
output, unless the output band is very narrow. In this .v
case, the probability density of the output approaches L A;=Xl=m
1
a normal law although no good general proof of this is
yet available. For special cases, the approach to a .v
normal law may be demonstrated. Often for particular L Aj= X2= (12 (5)
applications, it becomes necessary to know quantita- 1
tively just how narrow the output band must be for the
approximation by the normal law to be a good one. S
In an excellent article several years ago, Kac and 2LA/=X3=J..L3
Siegert! have shown that the characteristic function of I

the output of a low pass filter which is preceded by a and so on.


square law device, a band limiting filter and a white It is now possible to evaluate these coefficients in
noise source is terms of the iterated kernels of the integral equation (2)
1 and from these obtain a quantitative result for the
<I>(u)=----- (1) approximation to the normal law. In other words, call-
N
ing the kernel of Eq. (2), R(x, z)
II (l-iUAj)
i-I
with the A; determined from the integral equation L: dxR(x, x) = LA;

L: dtf(t)[K(s)Jlp(s-t)[K(t)J!= Al(s) (2)


i: L: dx dyR(x, y)R(y, x) = LA/ (6)
where the K(t) is the weighting function of the low pass
filter and the pet) the autocorrelation of the envelope
portion output of the band limiting filter. The number
of eigenvalues, N, mayor may now be infinite, depend-
ing on the nature of the kernel.
i: i:dx L: dy dzR(x, y)R(y, z)R(z,x) = LV

If now the log of the characteristic function is con- and so on, with the only restriction being that the
sidered, one has: kernel be symmetric.3
It is possible in several cases to evaluate the first
log<I>(u) = - [log(1-iuAl)+log(1-iuA2) three semi-invariants in terms of the various weighting
+ .. +log(l-iuAN)]. (3) functions of the system in question with the results
Using the expansion for 10g(1+z) Eq. (3) becomes dependent on the ratio of low pass band width to noise
N 1 bandwidth and thus clearly illustrate the approach to a
10g<I>(u) = L A;(iU)+- L A/Ciu)2 normal density.
1 2 1 Since for a normal law only the first two semi-
1 N 1 N invariants exist and these are also the mean and
+- L AlCiu)3+ ... +- L A;N (iU)N. (4)
3 1 N 1 2 Cramer, H., Mathematical Methods of Statistics (Princeton
University Press, Princeton, 1946).
* Now Project Engineer, Sperry Gyroscope Company, Great 3 H. Margenau and G. M. Murphy, The Mathematics of Physics
Neck, New York. and Chemistry (D. Van Nostrand Company, Inc., New York,
1 M. Kac and A. Siegert, J. App!. Phys. 18, 383 (1947). 1943).
1143
[This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded
to ] IP: 2.145.114.250 On: Fri, 20 Dec 2013 03:18:29
1144 GEORGE R. ARTHUR

variance, respectively, the effect of the third semi- The noise band width is taken as 2a and the output
invariant or third central moment which introduces a band as (3. For this case:
skewness to the density function must become negligible
to obtain a good approximation. L}..j =m=l
If the first three semi-invariants are considered as the
dominant ones, the characteristic function is (11)

The magnitude term here is normal and if the phase


term remains almost linear over the range of u where the
I<t>(u) I ~ e-2, the Fourier transform or probability den- A plot of H.us/~) vs 2a/{3 falls below 5 percent at 2a/(3
sity is nearly normal with mean m and standard devia-
equal to 50. Thus for noise to low pass band ratios
tion, 0". Thus the value of interest is u2 = 4/ ~ with m
beyond this the approach to a normal law is quite close.
normalized to one.
If Gaussian filters are assumed throughout with low
For this, the phase term is
pass band taken as O"l and noise band as 20"N and:
K(t) =7r!O"l exp( -7r20"12t2)
(12)
pet) =exp( -7r20"N2[2).
Thus if the ratio H.ua/0"2) is less than say 5 percent, This case yields:
a good approximation to the normal law is obtained.
0"1
For the case considered by Kac and Siegert 0"2---,;_-
V1O"N
K(t) = (3e- fJt t>O (13)
=0 t<O (9) .u3 O"e
and ----';---

pet) = e- al tl. 2 3V1O"N2


This means the low pass device is an integrator with with H.usl~) below 5 percent for 20"N/0"1 about twenty.
time constant 1/(3 and the effect of the band limiting The order of magnituqe is about the same as the pre-
device is characterized by a frequency function vious case, the difference probably due to slightly
different ways of defining band width. This method of
constant moments seems especially useful and has been also used
B(j) (10)
27rif with slight modification to obtain quantitative results
1+- for a frequency discriminator with random signal
a impressed.

[This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded
to ] IP: 2.145.114.250 On: Fri, 20 Dec 2013 03:18:29

Anda mungkin juga menyukai