- -

~
rZ A D_AO71 338 STANFORD UNIV CALIF INFORMATION SYSTEMS LAD FIG ~~~~~
RESEARCH ON ADAPTIVE ANTENNA TECHNIQUES II.(U)
MAR 78 B WI DROW , R CHESTEK. T SAXE N00019—71— C—0194
UNCLASSIFIED NL

7I~~~
—
U
‘C

ii I
—‘— .
.
Il
’l

___ _ _ _ _

END
DAT E

8 79 -
OAT

a _ _ A
I OEF
~
-
‘

INFORMATION SYSTEMS LABORATORY

STANFORD ELECTRONICS LABORATORIES
RTMEN1 Of ELECTRICAL EtdGIdEERItdG
~
STANFORD UNIVERSITY • STANFORD. CA 94305
1
p~j~ j~ EiL1
k

IW ~~
flONUSIUMlT . ~~
I ~~

Research on A d ap t i v e A nt e n n a Techni ques H

O LEVEL ~~~~

by~~

i- . B: Nidrow ,
~ p “
R. Khestek
T.j Saxe
‘
I- ) /~~ ~~~~~~‘
~~

DDC
~ M arch 14 1978

B
I ~~ F I N A L REPORT.
“I

Prepared under
~~
Contract N ø 19-77-C- Q 194
~ ~
)

.1
-;
~ Li 039
w— * ----— ----
- s
..
;’ .
— - _ —:
~~~~~~~~~~~~ ~~~~~~~~ ~~~~~~

.
~~~ ~~~~~~~ ~~

- .• ~
‘
• ~~ P
~~~ -
~
~~~ ~
4 ~~~~~~~~
—
- —~~~~~

-—
_ _ _ _ _ _ _ _ _ _ _
_ _ _ _ _ _ _ _ _

TT
- ‘
- - - --- —
~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~ ~~~~~~~~~~~
 •1
~ ~

Principle Investigator: Bernard Widrow (415) 497-4949

7
Name of Contrac tor: S tanford Uni vers ity
v’
Effective Date of Contract: 15 February 1977 to 14 February 1978
Short Title of Work: Research on Adaptive Antenna Techniques
Amount o f Contract: $54,426

FINA L REPORT

PART 1

Adaptive Array Processing for Separation of Inputs by Power Level ,

Frequency, and Angle of Incidence

PART 2
Analysis of Adaptive Weight Noise Covariance

R Ches tek
T Saxe B

1. This work was supported by the Naval Air Systems

E. Coniuand of the Department of Defense under contract
• N000l9-77-C-0194 J’& ..
~

The views and conclusions contai ned in this report are those of the
authors and should not be Interpreted as necessarily representing the
official policies , ei ther expressed or implied , of the Naval A i r Systems
Coninand or the U. S. Government.

-•-‘ - - .-
~
•-
~~~ ~~~~~
.
‘.
~~
-
~
•
~~~~ ~~~~~~~~
-.‘— - —.—
~~~~~~~~~
. -,--- —

-~
—

•
,,
, •
~
-t •
Ut
,
~

£1 : - ~~~~~ :.~~~
- hi~’
~~~.— ~~~ w-~ -w
-—— --—
— - - —
~~~~ _ — ~~~~~~~~~~~ -
-—---
- ~~~
‘ - -
~~ — =~~ —-
~ ~~ ~~~~~~~
-- —

V
~~~~~ ~ *cc~sstc ’i to
W
__________

NTIS Wh
~ Section
DDC Bit Section 0
ci
1us1I -C , - .
~ ~
. - -
~

SUMMARY BY
DtSILg!; -1L ttJiJ Q(S
~ ~
01St. l s . & ! . . ‘ J/Or SPECI AL

PART 1

The concept of adaptive power separation for single channel

applications based on the lea ky IMS algorithm has now been extended
for adaptive array appl ications. A new algori thm designated as
‘
scheme 6A’ has been shown to be potentially hi ghly effective for
nullin g strong janining signals in an aircraft receiving environment.
The new scheme Is subs tan ti all y less noisy and more simple to Implement
than scheme 6’ wh i ch was anal ysed dur i ng our p revious year’s effort.

Using scheme 6A , one cou ld establ i sh an omn idi rectional quiescent

receiving pattern In order to accept incoming signals regardless
of their angles of incidence . This pattern is sustained as long as
all incom ing signals are “weak” . However, in the presence of “strong”
Incom ing signals (jammers), regardless of their ang1es of incidence ,
the quiescent receiving pattern changes as deep nul1s form In the
di rections of the strong signals (janiners )as a result of the adaptive
process.

The adaptive algorithm sustains the quiescent pattern wi th a

“soft” constraint. A conflict develops wi th the incidence of a strong

t input , wh i ch , if strong enough , causes the soft constra i n t to be

violated and a null to form. The separation between a strong (jammer)

-
—
- —- .—- -a -----s.
- - - ------
•

~•
‘ —
, — —-— -- —
~~~~~~~ ~~ ~~~~~~~~~~~~~ ~~~~~~~~~~~

~ ‘

u•~ ~
~~~~~
— -
~~~ !
— _iii T -—
~~~~~~~~ ~~~ —
_ _
~ ~~~~- ~~~~~~~~~~~~~~
 i

Input to be rejected and a weak (signal) Input to be accepted Is

-

determined by a parameter y in the algori thm which , in an operational

system, could be controlled by a panel knob. The power gain of the
receiving array in the jamnier direction is reduced by the factor:

2
~ ~l ~
7 N]
[ ;

where is the Janiner power and N Is the number of weights in the

entire antenna array processor.

PART 2

All real-time adaptive processes experience noise In the adaptive

parameters. The amount of noise depends on the nature of the adaptive
algori thm , on the number of parameters, and on the speed of convergence.
A fundamental study of parameter noise and its effects on the output
signal has been undertaken for stochastic and deterministic Inputs to
weight-controlled adaptive filters driven by the leaky IMS algori thm.
Weight noise has been determined for the case of an Input consisting
of a sinusoidal signal plus white noise, for wide ranges of SNR and
frequency relative to Nyquist. Broad operating regions have been found

—_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

-• •— ~- ~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
—
~~~~~
--
~~~~~~
-
-
-
- -
— - - -— --
 where the output power due to weight noise is less than 5% of the
output signal power. ~~~~~~~~~~~~~~~~~~
_— -

Weight noise also has a significant effect on stability of the

adaptive process. An exact analysis was performed of a special single—
weight case. It was found that leaky IMS (and LMS)filters may hot

I ~tabiUze in the mean—square sense even though they converge in the

-
mean. It is known from previous work that convergence in the mean
is Insured by:

-
total Input power)
( -

it- •
A new criterion for- n has been found which guarantees jnean.squ~re
stabili ty :

L (total input power) 6 (total input power)

>
~~~~~~

!). — —- ---•-——- —
—. ----- ----•- _ _ _ _ _ ____ ~~~~_ _ ~•~~~~~
,
a
- — -—--

- - - • ~~
-
~~ ~~~
-
II..
i

• - - ~~~ Z’t’ • -
~~~~~~~
-
j •
: 4 - -
~~~~ ~~~ ~~~ ~~
I,, •~~~~~~ ; - ~~~~~~~~~~ : / : - ? - - .-•~~~ ~~~~; t~~k ,
‘
~~~
~~~
-

~~~~~~
 ~~~~ -- - - - --- -------- - . ~~~~~
-
~~~~~~~
—‘-- - -— -
~~~~~~~
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~ ~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~

PART 1
ADAPTIVE ARRAY PROCESSING
~~~ SEPARATION OF INPUTS BY POWER LEVEL,
FREQUENCY, AND ANGLE OF INCIDEL CE
~

1—A . Introduction

Our research on ad aptive techniques for antijam systems has con-

~
centrated on the application of a modified form of the LMS algorithm
to antenna arrays , permitting discrimin ation against received signals
on the basis of their power levels , using spatial and frequency
filtering . This type of antijain system is appl icable to situations
where the desired signals are much weaker than jamming signals.
Knowledge of direction of arrival and of other specific charac teris-
tics of the desired signal is not required beforehand since the ad ap-
tation pr ocess uses only signal power aS the basis for discrimination.
~

1—B . Review of Scheme 6

The goal of this activity has been the development of an adaptiv

—
~
antenn~i array whose sensitivity is high and essentially omnidirection- .
al to weak inputs , and whose sensitivity is low to strong signals re—
gardless of their angles of incidence. The objective has been to

cause hi gh power jamming signals to be severely attenuated while lowe

powe r comgaunic~ tion signals are only slightly attenuated . We thereby
realize a substantial sigr1al—t o— jammer improvement .

Figure 1 demonstrates the type of antenna reception patto~~ s tha

I
are th.sirt.ó (anJ vjhjcfl have been achicveó with the current algorithm)

,
•— — - —.- - - — -—••- - ---—--- — - — —. — - _ - a
~ ~ 5 J . 1 s~- —
~~~~~~~~
— .—“— • -— —
-

~~ ~

I
l _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

~~~~~~~~~~~~~~~~~~~ _ _ _
~~- - - - ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -~~~~~ — ------ - -
 —

SIGNAL -
-

PowERs -
0DB .0DB +20DB
#1=0 .0 POWER —.
#2 0.02
=
~~
#3=0.05
#4 0.07
=
Y=5
#5 0.1
= FREQ. 1

(
A) WEAK SIGNALS

JAMMER
POWER S
#1 = 0.5 -60DB 0DB +20DB
#2 1,0
/
=

POWER 0.5
#3
-

5.0
_ _ _
-

= ~~~

10.0
#5 —
20.0 ~50.0
#6 = 50,0 FREQ . 1

(
B) STRONG JAf IMERS

FIG, I. DIRECTIVITY PATTERNS

SCHEMES 6 AND 6A
6 ELEMENT ARRAY
4 .

4
-~ -
______________________________
—
, -
-,.
- - ______________
- --—
——
~~
__________________

~~
‘— - ‘
~~~
—
-

--

— ___ _ _ _ _ _ _ _ _ _ _ _ _ _ _
 The patterns shown are threctivity patterns , where the antenna array
itself is located at the orig in of the pattern , and where the signal
or jammer is arriving directly from the right (along the axis shown).
Fig ur e Ia shows directivity patterns corresponding to a set of weak
incoming signals. It is evident that the patterns remain essentially
omnidirectional for all of the signal powers listed . Figure lb shows
directivity patterns for a set of strong jammers. It can be seen that
as the jammer power grows , the array places deeper and deeper nulls I
•
the direction of the jammer .

An adaptive system which behaves in the manner just described is

diagrammed in Figure 2. The behav ior of this system was examined and
reported in (1] and has been called “Scheme 6” . We sumr arize its
~
behavior here. A six element circular antenna array (as an example) i
processed by six slave filters (TF~ ) to prod uce the array ’s output.
The weights of the slave filters are taken from a corresponding set o
~
adaptive or “training ” filters. The input to each independent train—
~
ing filter (TF1) is the associated antenna element’ s signal plus a
hhite noise “pilot signal” . The impulse responses are adjusted by thE
LMS algorithm so as to best minimize (in the least squares sense) the l
difference between the summed outputs of the Six training filters and l
the su~n of the pilot signals.

This results in the following behavior: in the absence of incom

~
ing signals or jammer s , the pilot signals force the impulse responses
~
of all of the adaptive filters to become zero—delay unit impulse
res~ons~.s. In this situation, the antenna ’s pattern is essentially
omni~’trectional (due to symmetry) . Now when an external signa l is re— I

. . -_ _ _ -------- -
— - —-— — — ~~~ ~~~~~

-
- -
—
•

12
~~~~~~~~~~~~~~ ~~~~~~~

—-- --
- --
•- — ---
~~~~~~~~~~~~ ~~~~~~~~~~~~~ ~~~~~~~ ~~ ~~~~~~~ — ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 - ----

A TRRINrN ,FILTERS

s’ ~~~~~~~
~~
-

IF2

9J// N4J ., j J4r

~~ ~
__ >,

N it
-

S
El F 5 -

N
S us
E TF 6
-
~~
py ERR OR
I~EIDN T5

N3’ SIGNAL
Wa
LOCRI. 1flO!PE ~ DENT NOISE S

• A DR FT 1V E RWTENNR
SYSTEM OUTPUT
—

1r 3
V OPERRTIOMAL F~ LTER $

Ii FIG. 2. SCHEME 6

_ t r -~~ - - - —U 14.- •-~~ -’ —- .-, t;w

~~ ~~~~ __-
4
- -•-— —— -— .-— — —
— -“I’ll

.1
~

#4 - -
-
-
- - •~~~~~~
- -

I
-
- -- - -
TI
- -

~~~~~~~~~~~~~~~~~~~~~~~~~~~
- - - -
_ _ _ _
 -
-
-——
- .--
~ ~~ ~~~~~
~~~
- - ----—-
~ —-~ - - - -

ceived , the LMS algorithm attempts to reject it by creating a null in

the sensitivity pattern in the signal’ s direction at the Signal ’s fre-1
quency, because this incoming signal is uncorrelated wi th the pilot
signal. However , the pilot signals and the adaptive process have
created a “soft constraint” that attempts to maintain un iform recep-
tion in every direction. The array pattern cannot be omrii and null
simultaneously. Therefore the LMS algorithm computes a pattern that i
~
a compromise of the two . The compromise solution achieved is a func—
tion of the power of the pilot signals relative to the power of the
received signal. As we can see from Figure 1, a strong received sig-
nal causes a deepe r notch than a weak one , accomplishing the stated
beamforming goal. This intuitive arg ument has been confirmed analyti-~
cally.

The patterns shown in Figure 1 are the receiving patterns of the

antenna array at the convergence point of the algorithm . In reality,
the algorithm never stops exactly at the convergence point , but moves .
around it slightly in a random fashion due to weight vec tor noise

I
Figure 3 shows the effects of weight vec tor noise . We assum e
that a single sinusoid is being received by the computer—simulated at
ray. Figure 3a is the time waveform output (the “RF waveform ”) of th
array system for a strong input , i.e. a jammer. Figure 3b is the tim
waveform output of the array system for a weak input. In each case
the fi gur e shows several cycles of the sinusoids before adaptation is
allowed to begin. Without adaptation , there is no rejection. Refer—
ring to Figure 3a , we note that for the strong jammer , attenuation is
4

__________________
-~~~~~ ~~~
•. - - - - ~~~~~~~~~ - -- -

. _
- -
i
- 4_~• _ —
r•-
~~ ~• •

~a
~~~~~~
,~~~~~~~~~~~~ i •T —~~~~~~~~~
~~~~~~~~~~~~~~~~~~ _ _ _ _ _ _
 ~4~-~-
-- -
F -- ~~~~~~~~~~~~~~~~~~~~~~~~
-

—
—
- ~~~~ ~~~~~~

-
T - i
~~~~~~~~~~~~~~~~~

_ _
_
_
_ _ _
-L
_ _ _ _ _ _ _
I
II ~~~~
.

.
I __ J
—
. I-
I
I,— ‘-
~~ ..
~~~~
c
~~~~~~~~~~ ci iu
.c~~~~~~.
0, -~~~~ (
C ~ U
I-
o
S..
.
~~~~~~~~~~~ .
C,
-~~~~~~~ .
~~~

4’ Ci 4’
V) 4’
‘I- C
ow
,- ~ 4’ 4j
-.- E
..
II ci
•..
-
~~
a_ ~~~~~~~~~~~

C
~~~
C ~~~~~ •
4. ’p~~l
~~~~~~~~~~~~
43 . .~~~~~~~~~~~~
0
.0
0. (fl 0. V’, • Ci ‘
4.’ . •
~~~~~~~~~~~
43 D L. ~~ ‘ .
C
~ ~~
~ ~~~~~~~~~~
_____ 43
DL- D ’—
~~~

ci . — •
~
S.C ~~~~~~
__ _ _
4)
WE Wa~ ~~~~~~~
C E - 0.
L- ( i Jo
o
-, ~ 5-
0 ~ ~~~~~~~~~ r10 ~~~
a
_ _ _ _ _

I.-
-j- j
)c’
—
‘I— ~
~~
E
_ _ _ _
~~~~ —~~~~~~~
‘1
j—... _
a E r—i C’J D ~~~~~~~~
___ ___ _ _
— ft
ft ~~ ~ A
~~~~~~~~~~~~~~~~~~
E!
I- I ___ ___
I
~~~~
r
~..~~~
s ~~~~~~~~
.
=H i
~~ I

I -~~~~ I
I ~ I—
~~~- —
I—
--

-
I
_ _ _

U, p
_ _ _ _ _ _ _ _

U,
~~~~~~~~ _ _ _

_ _

14 _i
‘4 9

’ a
- •— _-_..
•
.
•
.
. _ _ _-.- _ _ _ - — -- a- - .. -.~~~~ . -- -

44 ~~~~~~~~~

- “
I

-- _ _ _
_ _ _ _ _ _ _
 rapid , but no isy . A portion of the time waveform is pl’tted on a mag —
nifieá scale. Afte r convergence , noise in the weight vec tor randomly
modulates the strong sinusoidal input , producing a substantial amoun t
of output noise . The jammer is thus not completely rejected . Its ef-
fects are manifested in the residual output noise . Referring to Fig-
ure 3b , we note that the low power sinuso idal input is attenuated more
slowly and to a muc h lesser extent , but that its essential cnarac—
teristics remain. However , a close examination of this time waveform
shows that its amplitud e is vary ing somewhat over time . The array pro-
cessing algorithm has added some random modulation to the signal .Be-
cause of the noise in the weights , the amount of output power result-
ing (after convergence) from the incoming strong signal exceeds the
output signal power resulting from the weak incoming signal . This is
not a satisfactory situation , even thoug h the jammer has been substan-
tially attenuated .

1—C . Introduction to Scheme 6~ and its Characteristics

The beh avior d es c r ib e d in the previous section is excessively

n o i s y an d is a r e s u l t o f the U4S al gorithm being driven by a noise pi-

lot signal. To alleviate the noise problem , a n e w sys tem was proposed
• anô testcd . This system does not require the use of a pilot signal .

It is muc h less noisy arid has the additional advantag e of being

sigiplcr in its h&rd~~ r e r eq u i r ern~ nts .

1h.~ Sy~ tdfft , an~1 adapt ation ulgor ithm , are presented in Figure 4

and labelea as “Scheme 6i~.” Using the same six element circular array

for il lu sir u tion , the array signals are processed by only a single set

4 1
~

10
-

A • ~~~~~~~~~~~
— ~~- - - - “ - - -• -- —

A’ - -
.
•
‘

.~~•
-
.- •
.
11 -
~~~ - . -
-, •- - - • • -

-.-- - -‘--- - -‘—- -- - -— - - -- - -

-J — - --- — -
-
— ~~~ ~ ~~~~~~~ ~~ ~~~~~~~~~~~~ -
—
~~~
 - -
-

4’

~ TF I
_ s1

r
62
‘h F 2 I
I }\

ti ~~~~~~~~~~~~ II1-,~ % -
SY STEM
OVTPUT ~~
_ _ _

I
I
.
‘hTFU

~ 4TFS
SI G ~4F3 L— -

USED FOR ROFIPTATLON

= •i
~~
+
~~~
+ 2p~~IQ WJ )

= (1-2pflW + + 2p)’WQ
~

WQ “QUIESCENT WEI GHT VECTOR” = [i 0

t~~~~~OR ScHEME I
A
EQUIVALENT PILOT POWER

FIG. 11, SCUEME 6.4

11
‘4
-

T
— —
— — —

—
‘ -:- -- •
- . •

~~ -
--• . ~~ ~~~~~~~
‘
~~1”

- --T-- TT-~~~~.~~
-
- __ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 —- --- - ----- -
-
- - -- --- —- - - - - ---—- -
-
~~ ~ - -

of transversal filters in order to produce the system output (Scheme 6

required two sets of filters) . In Scheme 6 the required performance
was accomplished by introduc ing a unique training signal. In this new
scheme , the effec t of the training signal is accomplished directly by
the ad aptation algorithm .A s a result , no training signal is required ,
significantly reduc ing system hardware requirements .

The adaptation algorithm of Scheme 6A is presented in Fig ure 4.

By examining this algorithm , its relationship to the standard LMS al-
gorithm is clear. However , the Scheme 6A algorithm includes one more
“driving ” term which is necessary to produce the required behavior .
To use this algorithm , one must know what the desired weight vector
would be in the absence of any signals. We call this desired we ight
vector the “quiescent weight vec tor ” ~~~~~~~ . If for each adaptive filter
we set equal to the mean of the corresponding converged wei ght vec-
tor that Scheme 6 attains in the absence of inputs , Scheme 6A will au-
tomatically produce the same mean converged weight vector solutions as
will be produced by Scheme 6, in all signal environments.

For the Scheme 6A algorithm , Y is equivalent to the power of each

pilot signal used in Scheme 6. It is possible to choose any quiescent
weight vector desired . To produce the receiving patterns shown here ,
the quiescent weight vec tor was chosen to have a unit impulse response
at zero delay for a single filter , and zero impulse response f~~ all
other filters. This results in a truly omnidirectiona l pattern ir the
absence of any signals (assumir .g omnidirectional antenna elements) .

By examining the second form of the Scheme 6A algorithm as writ—

‘4
-
12
 ~~~ — — --

ten in Figure 4, we may notice a relationship to the “leaky LZIS algo-

rithm ” described in [11. The first two terms are identical. However ,
the leaky LMS algorithm has a tendency to drive the weights to zero.
The new algorithm has a tendency to drive the we ights to the quiescent
weight vector. It is the third term in our algorithm , absent in the
leaky LMS algorithm , that causes this. Thus the new algorithm is a
generalization of the leaky LMS algorithm , since the leaky LMS algo-
rithm can be obtained from the new algorithm by choosing the quiescent
weight vector to be zero.

To demonstrate that Scheme 6A eliminates most of the noise in the

weights and the associated random modulation problems demonstrated
earlier with Scheme 6, Figure 5 presents time waveform outputs of
Scheme bA . The same signals used with Scheme 6, which resulted in the
outputs of Figure 3, have been used with Scheme 6A in generating the
waveforms of Figure 5. %~e note once again that the strong jammer is
attenuated rapidly, and does not display the noisy output seen for
Scheme 6. The weak signal is attenuated only slightly, and it does
not display the random modulation distortion that Scheme 6 induced .
Finally, we note that the output response to the strong jammer is much
weaker at the output than is the response to the weak signal. The im-
provement in output jammer to signal ratio is evident. An array sys—
tern using the new Scheme 6A algorithm and receiving both information
signals and jamminy signals would exhibit per formance superior to that
of an array system based on the old Scheme 6.

In figure 6 we present the antenna recept ion patterns attained -

when the weight vec tor is at the convergence point for the adaptive

‘4 U

•-.---—---- -- — — .
.- - -- - — *— — ---- -.•-—- — — — -*- ~~~~~~
‘—a
~~
W. - - - -
~~~~~~
—
-

~~~~~~~~~

-p.-

_ _ _

~~~~~
 r - — -—- _ _ _ _ _ _ _ _

~ ~~~~~~~~~~~~~

o
C
I
I •

.~e
I

-
II
.

..
~~~~~~~~~~~~~~~~~~~~~

~~~~~~~ .
-
-

. .
U
U,
,
W I-
C) ,-

U,
..I
-
~~~~~ W CJ
4’ C) _ _ _

~~~~~~~
—. 4.’4’
— -
C
~~
4.’4.’ 4.’ 4.’ 0
• ~1
~~ ~~~ b— .—

u_
.
— •0) -
I
~~~~ ~~

o_ u~
C C LU
4J _ - I-
0. u
•~~~~~~~ ~~~
— o ~O
~
—
~~~~~~~~
Ci
—
ai . C-U C) ..”
‘_~~~
<
~~~~~~~~~~~~~~~

— .
4

a e
~
_ _

-~~~~~~~~~
-
E ~ sJ 0 E a
— C)
_ _ _
~~~~
~ ~~~~~~~~
I— ~~ ~~~~~~ ~
_____ ~~
a
0. II
‘I —• - C,
. ~~~~~~ - —I I-
& —. •
~~~~ ~~ —•
I’ X LU 0
.
— ~~~~~~~
- ~~ C)
I— F- ~~~~

.—
~~~~~~~~~
LU
.c — LI..
~
~~~ 0. L&. ••_ U)

L -
~~~~~~~~~~~ _ _ _
_

—c~~~~~~
—=-
_ _ __ _ __ _ _

I .r
~~~~~~
I I
I a
_ _

4
14

— — -— --— — - —
A —
- - —- - --— — -- —- - — —• - --—-—-—- —
— ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
~~ ~ ~~~ ~~ - —
~~~~~~~~~~ ~~
-
-
~~~~~~~~~~~~~~~~

T1
~~~~~~~ ~~~~~~~~~~~
~~~~~~~~~~~ ~~~~T~
- ____
~~~~~~~~~~~~~ -
 - - -— — ____ ___ _
-;-- - ---

20
-

SIGNAL
10

C ~.-_---
1 r POWER = 0.0
—C
- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -

D POWER 0.1
SIG NAL
POWERS
‘In . 1 11 _ n
tt.L U.

#2 = 0.02 ‘
~~ 5 (A ) WEA K S IGNALS
#3 = 0.05 FREQ . 1
#4=0.07
#5 = 0.1

—50-
—180 —90 0 90 180
SIGNAL DIRE CT ION ( DEGREES )
20-

30- JAMMER

POWER = 0.5
0

—10 POWERS
#1= 0.5
#2 = 1.0 ~~
‘
= 5
FREQ . = 1. (B ) STRONG JAMMER
= ~
#5 = 20.0
—LO #6 50.0

~~~~~
—
POWER = 50.0

—60
—1!0 — 0 C 9D 180
JRMPIER~ DI RECTION ( DEGREES )

FIG . 6. DIRECT IVITY PATTERNS

SCHEMES 6 AND 6A
4 1 6 ELEI1ENT ARRAY
15

~r Y 1 ~~~
z - - - :-:~
:ii2 - - 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~ -
.
 ~~~~
---- _— ‘_ —
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

algorithms (Schemes 6 and 6A converge to the same expected , or aver—

ag e, solutions —— it is only their dynamic noise behavior that is dif-
ferent) . It is clear that low power signals are attenuated only
slightly, while Strong jammers receive significant attenuation. The
stronger the jammer , the greater the attenuation. The mathematical
relationships are repor ted below.

In Figure 7, we present the theoretical curves for the gain of

the antenna array to a single directional sinusoidal input as a func—
tion of the power of the input. Since the array gain is also a func-
tion of the pilot signal power for Scheme 6 and of the equivalent pi-
lot power )Ifor Scheme 6A . we have plotted the gain curve for var ious
values of pilot signal power (or equivalent pilot power).

In figures 3 and 9, we present the receiving patterns for the an—

tenna array when two sinusoids of very close frequency and equal
powers are received (calculated at the converged weight vec tor). Lx—
cept for the necessary nulls , we see that the array maintains approxi—
mate omnidirectiona lity, even including the space angle between the I
two jammers , where possible. Sharper ang ular resolution could be at-
tained , but more than the six antenna elements would be required .

l—D . D e f i n i t i o n and A n a ly s i s of Scheme 6A.

In this section we define Scheme 6A and analyze some of its pro—

perti es , to confirm the statements made in the previous section.

4 ~
.
16

- •

- :- ; - •
_ _ _ _ _ _ _

~~ ~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
I

~~~3-L~~~~ -~
-
._ __i __ ___ _._
~
~~~~~ ~~~~~~~
- _ _ _ _ _ _ _ - - - _ _ _ _ _ _ _
- TIJ
 I-I
LU -

H ;

LA
I. <U I — -

k ~~ ow -

<0 .

4 ~~~
-

‘7

••- - ~~~~~~ •— ~~~-—- -- - — — - -—

A • - ---- — -— - - —
~~~~~~~~~~~~~~~ ~~
~~ ~~~ ~
.-
• -
- -
~4 - -
.- j
-- ’ - ~~~~~~ -
-j -
.— —
-
P1” ~~~~~~~~~~~~~~ -~~~~
—
—
— — — — -
-- -
~~~~~~
- -~ ~~~~~ ~~~~~~~~~~~~~~~~~~~~~~
-~ — - -~~~ —
F
--- - -- _ — - ---
-
_ _
—--- -- - - --

~~ ~~

JAMMER
#2

ER #1
-6ODB 0DB +20DB

SCHEMES 6 AND 6A, 6 ELEMENT ARRAY

20-
JAMMER JAMMER
30- 11 12
P’1.997
‘p
P2 . 0
[

0-

C ao
~
—40

~• _ 50
- 1eO -~ 0 9”O
( DEGREES)
iéo
JR 1IIIER DIRECTION

FIG. 8. DIRECT IVITY PATTERN , AT F=2.O

BOTH JA~ 1ER POWERS = 20

if

4 18

I
 --
~~~~
- ----- — -- - - ---- — -
~~~
-i--’-—-

-
JAMMER #2
J MMER #1

-60DB 0DB +20DB

SCHEMES 6 AND 6A, 6 ELEMENT ARRAY

20
JAMMERS
26 #1 #2 -

F 2.04 F4.997
4

t CD~~~~~

—90 0 -
910 iêo
J AMIIER DIRECTION (
DEGREES)

—
FIG. 9. DIRECTIVITY PATTERN, AT F=2.O
JAMMER POWERS 20

:
—
-- — -- -
- .
~~~~~~ ~~~
 ~

1—D--l . Terminology, Notation, and Definitions

In this subsection we briefly introduc e the terminology, nota—

tion , anu definitions to be used in the remainder of this section.

There are k antenna elements.

The output of each element is fed to a transversal filter. The

tapped delay line of the filte r contains n taps . The weight vec tor of
the filter therefore also contains n elements.

The letter i will indicate that the quantity is associated with

F antenna element i or transversal filter i. Thus i can take on val ues
from 1 to k inclusive.

The letter j used as a subscript is a time index , and ind icates a

sample taken at a specific time .

The output of sensor i at time j will be denoted by x (i).

3

The contents of the tapped delay line (TDL) of transversal filtei

i at time j will,
be denoted by the vec tor X (i) We see from the wa~
3
.

a TDL. operates that

X (i)
3
= (x~ (i) x~ _ 1 (i) ... x j_k+l (i))T
Tne output of the transversal filter i at time j is denoted by
Y 3 (i).

t. The output of the entire system at time 3 is denoted by y .

3

W2 will need to refer to the contents of all ‘lDL’s simultaneous—

ly. To do this , we define an augmented TDL Contents vec tor which

20
I ____________
--- --— -- _ _

~~~~~~~~~~~
is the set of all X (i) vec tors “stacked up” to produce one vec tor.
3
Thus

Similarly, we define the augmented weight vecto r as

We can now define the operatior~ of the system : X (i) is obtained

3
by doing the time shift of a tapped del ay line , using as the new input
value x (i) .Thus X (i) is just ~~~ ..l, (i) with all elements shifted down
3 )
one position (discarding the bottom element) and using x (i) for tne
3
top element. The output for transversal filter TF(i) is simply
T.
Y ) (t ) !~ (i)X (i)
.
=
-

3
1— 1
and for the entire system the output is
k
;g T~
y. ~ y.(i) = ) 3
~
1—2
Notice that we have not discussed how is determined —— this will be
described in a later section dealing with the adaptation algorithm .

We next define a covarionce ma trix for the contents of the tapped

delay lines as where
~~~
E(X ~~X~~ )

1— 3

witn F denoting tne expectation operator. We will assume that the

processes x~ (i) are stationary so is a constant matri x. Since we
will be referzing to just one co~ ariance matrix , we will denote

21

~~~~~~~~
- ~~~~~~~~~~~~~ - - -- - -

it by R.

In the application of this array scheme , one flexible aspect open

to the designer ’s discretion is the shape of the directivity pattern

that one wan ts in the absence of input signals , desire d signals or
jammers. This is the nominal pattern that will be notched by strong
incoaiing jammer signals as a result of the ad aptive process . The
weigh t vector that provides the desired quiescent directivity pattern

when no signals (desired or o therwise ) are being rec e ived is d eno ted
the quiescent weight vec tor.

Lastly, we define to be the opt imum val ue of the weight v ec-

tor.

l—D -2. Perform ance Criterion

,
if

The adaptive system Scheme 6A presen ted in the previous section

will be developed in this section from first principles. It will be
shown in a later subsection tha t this system and Scheme 6 produce the

same mean conv erged weight vec tors.

We will first define a pe r formance criterion for the adaptive

• system . In accord with this criterion , the current performance is

used by the sys tem to mou if y its pa r ame t ers to improve fu ture per for-
manc e.

Let the performance criterio n be defined as:

2
•
J = E (y J +

1—4

22

J’ ITIi 1 1$.
 _____________
-
T -
1T1 IY9Y~~~~ - - -
_ _ _ _ _ _ _ _ _ _ _ _ _ _ - - - -

‘ -
~~~~~~~~~~~~~~~~~~~~~~~~~~~
—
~~~~~~

Recall from sec t ion B tha t :

y is the sys tem ou t pu t
W is th~ wei g ht vec tor

is the quiescent weight vector

is the magnitude squared of the difference betweer

the two vectors.

It is our goal to hav e the adaptive system find W such that J is

minimized . -

To gain a clear understanding of this criterion , let us firs t ex—

am ine it in the case when no signals are received by the system . In

this case no input means no output , i.e. y = 0. So it is the task of

the system to find W such tha t $ I W— 2
~~ ll is minimize d. This is clear—
ly accomplished by setting ~~ = ~~~~~~ in other words , the pe r formance
crit erion J is minimized in the no signal case by the ad aptive process

attaining the quiescent reception pattern.

• In the c ase where an inpu t is availa b le , the following tradeoff

2
~~~I t
occurs: by making a chang e in W , the magnitude of I l W — increases ,
while E~ y 2 j decreases. If the decrease in E [y 2J is greater than the

inct ease in j i ~~_~~~,i 2 , the performance criterion .1 is decreased .What

th~ ad a pt iv e sys tem will d o then , is f i nd W such tha t any decrease

2
7 th~ t would occur due to a decrease in EL y ) is exac tly balance d by any

increase that would occur due to an increase in II ~~_~~~,I 2. What ~~•-

L heppening is that the syste.~ will attempt to have W stay close to
5~~ t.

h~~, only moving away when the output power Ely grows large.

23

- _ •—- -,
4
-
—
-

-.
—
~~~
_ _ _ _ _ ~~~~~~~— --- — — - ~~~~~~. — - -
_ _ _ _
~
~ -

To allo w the sys tem de signe r to influenc e this t radeoff , we in-

troduce a designe r controlled par ameter Y into the pe r formance cri—

ter ion:
J = E [y~ j + Y I I~~~—~~.Q II
2

1—5
2
As such , the d esi g ner can con t rol the rela t ive tradeoff b e tween E[y J
and IU ~~W H in the performance criterion.
0

l—D-3. Optimum solution to the problem

The problem as stated in the previous section is to minim ize

E[y~ J + Y i Iw— 2
~~ II
3 =

1—6

by prope r selec tion of W. Using gradients , we can determine the op—

t imum value of W —— tha t is , the W w h i c h yi el d s the mi n imum 3.

We begin by rewriting 3, using the fact that

2
I I!1 1

4
1—7

(for V a vec tor ) and using the system equation for y (1—2):
,

T T T
J = E [W X .X WJ + y (w~4~Q ) (W_ ~~ )
1—8

We see tha t 3 is quadra t ic in %~~~, so that a unique minimum exists. W~

find this minimu m by setting the gradient of 3 with respect to W to

zero. The gradient is

V J = 2E (X X~ IW + 2Y (~ -W Q )
3
1—9

24
- -.—, “.-.- ‘ ,- ~~~~~~-•——
- -‘U

- --- — - ---- —- -—
-—- _- -
—
—a ~
‘— -
~ Li ~~~~~~~ -- ~~~~~~~~~~ ~~~~~~~~~ ~~~~~~~
_________ ,
_ ~~~~~~~~~~~~~~~~~~~
r - - —.——----- —- -
‘
~1

The optimal weight vec tor that causes the gradient to be zero is

designated as W
*
, and is ob tained from

2 E (X x T * + 2y (W* ) = 0
~ !~
Jw
3
1—10
’
Recalling that E (X X~ J R , and ga thering terms in yields:
3
(R+ yI)W * - )‘W = 0
1—11
*
Now we can solve for W

=
1—12
or

W
*
= (
],
R+I)
—l
~>~ ~~~~~

1—13
So the optimum weight vector is a function of the quiescent weight

vec tor and the covariance of the inputs signals.

l—D—4 . Comparison to Scheme 6

In reference (11, Scheme 6 was also referred to as an Adaptive

Beamfo rmer with Injected Noise (ABWIN for short) . It is now our goal

to show that the ABW IN anô the scheme just presented yiel d the same

con v e r y~ d ine~ n wei g ht vector. Let us denote the converged mean weigh
-

vectcr for the A3~ lN as Then (ii showed that

*
= (R+a2 I)
— 1 P

1— 14

25

- afl’— - _- -aS-
..
.
~ ..-‘-- S-’-
—-

:_
-a

‘
I •
- ,

_ i
_ _ _ _ _ _
- --
~~~~ - ---- ~~~~~~ ~~~~~~
--
- --.
~
 —,

where the R is the same as we have defined , cr~ is the power of the in-

jec ted noise used for the pilot signal for the ABWIN , and P is the

corr elation between the contents of the tapped delay lines (X 3 J and

the pilot signal . This can be rewritten as: -

— —
*

—ABWIN
-

1—15

In Cl ), in the earlier sections , it was shown that

= ii0 ... ü i i 0 ... o i . . . i i o ... 01 T
o•
n

and that the quiescent weight vecto r for the ABWIN was P. In :

later sec tion of C l), a me thod w a s proposed for al tering the pilo t

signal forma tion so that any value for the cross—correlation vec tor P

could be €ittained~ and P would be the quiescent weight vec tor.

o•
n

Thus we showed in (11 that

1 —l
= (—ZR + )
~ AB~*JIN
~~
~~ ABW IN
1—17

where = P could be chosen beforehand by the system

AB~ IN
designer. Now if we compare our new system
1R + I ) —l
*
= (y ~~
1— 18

with the above relation for the ABW IN , we see that ~~ we

~~ BWIN

make the ussignmentS )f and So the new sys tem can

4 26 -

•
- ~~

-
— —
T.
- --

- —
~~~~~~~~~~ --
 ‘—-
- —, -- —
— -
-7- -- --
-
~~~~~~~~~~~~~~~~~~ — - ---

obtain the same solution weight vectors as the original ABWIN .

1—D—5 . Adap tation A lgorithm

It remains to present the ad aptive algorithm itself. The

development follows directly the development of the LMS algorithm ,

which is a steepest descent algorithm .

The basic idea is: assume we have a weight vector at time j.

We wan t to find a weight vec tor for time j+l that is closer to the op-

t imum wei gh t vector. To do this , we compute the gradien t of J wi th

respec t to W , and evalua te it a t the curren t we ight vec tor W . This

3
g radient (deno ted S73) defin es t he direc t ion t ha t W . shoul d be al tered

to increase J. Since we are interested in minimiz ing 3, we go the op—

~
posite direction. Mathematically, we set

—3
W.
—3
- pV J
1—19

wher e ~i g ov e rn s ho w f a r in the d irec t ion sp ecified by~~~ J we go.

-
(If ~
-
5;

p is too lar ge , we coul d overshoot our goal so much that J increases

I
again!)

From section ]—D—3 we already have an expression for V 3:

w
_
V 3 = 2R W~ + 2Y (W~ ~~ )

1—20

Using this in the expression above (1—19) we obtain

= 2p(R W~~ + Y( )J
~~~~Q
-
•
-

1—21

-4
27

--..
-—
4
-— —- - -.5 -~~~~~~-~ - - --‘
-

5 — — - —— — -— - - - -•-, --7 - - - ~~~~~~~~~~~~~ ~

-
I •
4
‘- ~L~•~ --

- -~~ -~~--_ ~~~~ T~~ -~ — - . i 7 - T T~~~__ ~~~~ _ _ _ _ _ _

 -

~~~

Now in general we do no t know R. Instead we use an instantaneous but

unbiased estimate of The estima te is

~~~.
~~ X~ • Using this in the above-
expression we obtain

= - 2J1X ~ X~ W
)
- 2PY (W~~~~)
1—22

Now since y = X~~W~ (1—2) we have

3
= !- 2PY) 2PY( _
~ j+l
— —
~j ~~ ~~ ~&
1—23

which may also be rewritten as -

= (l — 2 Y) ! + 2p)4
~ j+l ~
— -

~ ~~
1—24
This is the adaptation rule used in Scheme 6A .

l— D— 6. Convergence of the Adaptation Rule

We mus t now demonstrate that the adaptation rule presented in the

previous section converges , and that it converges to the optimum

weight vector . The quantity we will study to indicate convergence is
‘1 the mean of the weight vec tor. Other criteria are possible , as in sto-.~
cnastic approximation techniques. However , many algorithms based on
r
stochastic approximation have a tendency to “turn themselves off”
af ter a t ime span , ignoring la ter data. While this may be suitable

r for a truly stationary environment , it does not allow any capability

— for following changes in a nonstationary environm ent.

- -
~i1l show that with a stationary input environment, if
First we .

- the algori thm converges , the mean of the wei gh t vector has only one

28

—
-
— -— - -
• .- - _____________

4 - -
• —,
i
l ~~~~~~
:~~~~~~~~~
-

--
- -- -
~~~~~
--=
~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~
T1TT T~~
.

.
 ‘ —-
-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

point to converge to——the only stationary point of t’ e performance

~
criterion 3, which is therefore the optimum .

- , To find the stationary point , take the expec tation of bo th sidel

of the adaptation rule (1—23) :
E(W~ E (W ) —
÷1 I
= J — 2P Y E [W ~ —~~~J
J
2 P E [Y X~~~
)

1— 25
Now , at convergence , (which has yet to be demonstrated ) we would hav4
*
u r n E (W. J W
j—>oo ~
-

1—26
*
-
Recalling that W denotes the optimal weight vec tor. Thus we write :
-

W* W* - 2p u rn E (X.X T
.W. J - 2 p y (w
*- )
~~~~
)— >oo —V

1—2 7

where we have reexpand ed y ’

as X~ W (1—2). Continuing ,
) )
—

2pR %~i = -2pyW

1—28
wher e we have assumed that , with W~ converging as j— oo,

u r n E (X .X~ W .J = l int (E [X .X T)E (W .J)

~ ~ j—>oo 3 3 3
*
RW
1-2 ~

Fjflally , after grouping term s , we c:n ~~~

H (R+ Yfl
~
* —J_
= Y (R + y I )

4
29

4 —
~~~~~~~~~
-—— — ———----—--—-- - -- — —7-- —7-——-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~
,- - ‘-- - — -

I
--
- - -: - ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-

~~~~ ~~~ ~~~

~~~~~~~~~~~~ -
--

~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~
- - --—- 7- - ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 — -- -
~~~~~~~ - - —-
~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~ ---—.
_ -.----—— ---_----

-
~~~

1—30

which is the optimum solution found in section D.

We must next show that the adaptation rule causes E(W~ J to con-
ver ge to a single value , which mus t necessarily be the stationary

point found above. Since the proof of convergence is directly paral-

lel to that of the LMS algorithm , and is rather leng thy , we will not
pr oduce it here , bu t will point out the modifi cations needed to the

proof of convergence for the LMS algorithm . (A proof of convergence

of the LMS algorithm is contained in references 5 and 6). The major

differ ence is tha t the t erm which appears as

tI+2k s/\
-

in reference 6, must be modified to be

(I + 2k 5y1 + 2k/\) C (l+2k 5Y)I + 2k g~\
1—31

Notice that the eigenvec tors in ma trix Q have not chang ed . (Also not~
that k8 in 16) corresponds to p in this report)

The other chang e required is the replacement of the term Ø (k,d)

by 2k 5YW~ .

— with these modificati ons , the proof of convergence follows the

same steps leading to the conclusion that the adaptation rule causes
LIW ) to converge to the opt imum solution so long as
)

wax
1—32

where Is the maximum elgenvalue of the m atrix R.

30

--

-
t

_ _ _ _ _ _ _ __
_ _ _ _ _ _ _ _ __
__ _ _ _ _ _ _ _ _ _ _ _ _ H
r - - -

1—D— 7 . Response of Scheme 6A to single sinusoids

In this subsection we analyze the gain that a single sinusoid

impinging on a Scheme 6A array would encounter.

We represent the sinusoid being received at the ind ividual anten—

na element as a phasor . Let us denote the phase of the sinusoid at
antenna element i as Let the phase difference of the sinuso id at
two ad jacent elements of a tapped delay line be 0. If the sinuso id has

power cr~ then may be wtitten as:

-

4 31

—7-— - 7 - — — - — - - ———- 7- —— 7--— — — - - —‘-— -—--—-- — —

—
- - - .- - - -
.
~~~
-
- 5-
~~~~~~~~~~~~~~~ ~~~~~~~~~~~~

- -
&- ~~~~
 7-
~~~ lJIllU
- - — ------ - -
~~~~~~~~~~~~ ~~~ ~~~~ -7-,

e~~~

j ( Ø —29)
1
S

j(01— (n--l)O)
e

e
j ( 0 2 9)
= °
e
~~~~ ~~~~~ ~~

e~

1—33
Notice that we hav e defined the vec tor !to be the above column ma-
trix .

Nex t we note tha t t he wei gh t vector W consists of n Keal weight

How ever , using complex notation , we replace the n reQl we ights by n/
complex weights which prod uce the same output. Henceforth , we

4 ~~
-
32

--
A
- -
—
--- - - — — — p.-—- - -- — -
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- — —N
~~
7-
‘ -
- -
-
a1 ’ - -
-~ I
.,
~~ 7-

—a—.- - —~~
_ _ _ _ _ _ _

- —~~~~ —-
 - --

represent all we ight vec tors as vec tors of n/2 complex weights , and X
as a vector of n/2 complex samples , with autocorrelation matrix
2 +

1—3 4

(where denotes the complex conjugate of V).

Using this no ta t ion , we can express the expected ou tput power of

~
the array for a single sinusoidal inpu t as:

2 2 7-
Ely ) = ci W ”V V~W

1—3 5
for any s teering vec tor V , signal power Ø~~ , and weight vec tor W.

Now , if adap tation is not allowed and the weight vec tor is set t
~
an ini tial v€ ~iue t h en the ex pec ted ou t put power is:
~~~~~~~~,

= a2W~ V V~ W
--Q
1—3 6

Consider next the expected output power at convergence. We have~

shown t ha t the wei g ht vec tor at convergence is:
* R —
- y-+I) ~~
1—37

hus , we have
T

= ( — ~- V V + I ) -l ~~ +

H - h

1—3 8

By apply ing the matri x inversio n l erama LThaorem ~ .22, refe r ence 7) w
c~ n ob ta in

33

— --— 7-— - -- -- — ---- _ .

-- _ - - 5- -- -

-
, - p
‘~~ -
—- —‘. - --
- ---~ 4
-~~~ } 7-
-~~~~~~ - . -
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ -

-
r ——------ - ----———

~~
---
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--- ‘
- -
~~~~~~

2
a’
(
S
v ~, + I ) 1
+
~~ ~~~
- — Y- —
a
V I -1V
S
1+ +

-
-

1—3 9
Now , from the defini t ion of V , we can see tha t
7-”
2 =!
!!= II !I l ~~ -

1—4 0

(r ecalling th a t k is the number of an tenna elem ents , and n/2 is the

numbe r of taps in each tapped delay line (complex samples)). Let us

define N = nk/2. Then we can use these formulas to give us a con-
verged weight vec tor of

or2
5 +

W
*
= I- W
or2
— —

~~~
+ 7-S
1—41

Now we may compute the expected power output at convergence (using W *

in 1—35):

E(y 2) a a 2W*+ V V W+ *
*
5
h W
 ___ _____
r
- --- -.7--—
_ ___i___i—=-- —- — —--7-—- -
~~ * I —
~~~~--~~~~ ~~~~ —

or~ o-~
+
+

=
1. -
!V~ ! - _ _ _ _ _ _ _

.4
+ + _v!v +v v +V V +
7-’ ” ~~~~
~~~V V - 2 ~~~~~~~
1 + - ~~ N (1

1—42

Using (1—40) again gives us:

2
as
Ely 2) ’ y
*
~~~~V V ~ - 2
a’
~~— !
!
~~
- 2 v v +~~
I + —
~~ N (1 + — N)~
~~

2
4
0’
0
5
= 1 - 2— +

1 + - ~~ N (1+ - ~~ zg)2

= a ’~~~~~~ WV V W

= a’

1—43

27-his c..n be rewritten (using l—3b) as:

4
35

I
- -
-

-.7-
_ _ _ _

- - _ _ _ _
-
- 1]
 2

2
Ely ) = E(y 21
2
* w=w
y
1—44

Thus we see that the expected output power of the array at convergence

is the expec ted ou tput power of the array ini t ially, mul t iplie d by an -

attenuation factor. This factor is always less than (or at most equal

to ) one since -~~~ N > 0. ‘This attenuation fac tor is independ ent of the

antenna geometry, arrival direction of the sinusoid , and of the quies—

cen t weight vec tor. The attenuation factor depend s only on the

sinusoid ’ s power (o~ ), the to tal number of taps in the array filters
(N), and the equivalen t pilo t noise po we r (Y) . Thus , if we know the

response of the q uiesc en t una d a pt e d a rray to a single sinusoid , we can

easily calcul ate its response at convergence of the adaptive process.

36
4
-- - — . - — _• - ~~r r ’ f l a -1IuuI~

.1
-- - —- -

- -
‘ ‘ -~~~~~ ~
-
- -
A

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ _ ~~~~~~~~ -~~~~~~~~~~~~ _ -~~~~ .~~~~ s--

 TJ I
~~~~~~ ~~

Conc1u~ ions of Par t 1

-
1— E.

-
We have proposed and studied a new ad aptive algorithm which has
been designated Schem e 6~ . It is related to the leaky LMS algorithm
previously studied . When applied to an antenna array,. it yields a

method of antijainming based on attenuation of received signals on the

basis of their input power levels. No a priori knowled ge of the signal
characteristics is required . This algorithm replaces a prev iously
studied algori thm , Scheme 6. It exhibits improved noise behav ior and

reauires less har dware for implementation.

-

37

4 7- —7-— p_
_ _
~~~~~~~ ~~~~~~~~~ _,~~~
- s - ’- ‘- . .
-~~ -
- :
.
I~..- - - :

p ,1
l
i

-~ =r~ i
~~~~~~
T~~~~~~ i~~~~~~
~~
_ _ _ _ _ - -_~~~~~~~~~~~~~~~~~~~ - - - —— - --- --~~~~~~~~~~~-- --- -~-—7---—-—--- -7- ---~~~

r
_ _ _ _

PART 2

ANALY SIS OF ADAPTIVE W EIGHT NOISE COVA RIANCE

2—A . Introduction

Previous work [1] int rod uced the con c ep t of using an a d aptive
lin e enhancer [8, 9) to separate signals by power level . Althou g h the
or iginal proposal require d injected noise and ‘slav e ’ fil ters , a re-

fined version was developed that eliminated the need for injected

noise by replacing the LMS adaptive filter with a ‘leaky ’ LMS filter .
Because an adaptive line enhancer using the ‘leaky ’ LMS algorithm can
separa te sig nals b y po w er level it is c alled an ad ap t ive power se p ar a-
tor (APS).

During the pas t year we have b een eng ag e d in anal ysing the pe r-
formance of the APS . Per formance in the mean has alread y been

descri bed in [1) ; however , performance in the mean does not complete—
ly characterise an adaptive filter. Specifically it is important to

know abou t the noise in the weights since , b y a mo d ula t ion process ,

weight noise caus e s nois e componen ts in t he fi lte r ou t put [ 9 , 10)

which deg r ades the performance of the APS. Tne first step is to charac—
ten se the variance of the i~oise in the weights.

Analysis of weight noise covariance has proceeded in four main

phases:
1) a determinis t ic anal ysis w it h the input cons is t ing of a sinusoid

2) a statistical analysis with the input Consisting of a sinusiod

38

.4 7-
- -~~ ~~~~~~~~—~~~~-— ...- — .~~~~~~~~~~~~~~~ — -- --
~~~~~~~~~~~~~~~~~~~~~~~~~~

11
- - — - -~~~~~~~~~~~ — — 7 - — -- - -

_1
— —
~~~
- -- — •— • - --
~~~~ i.L - — ~~~~~~~~~~
 —- - --- - 7-
- _ _
~~~~~~~~-~~~~~-.-
- ~~~•-~~~~~~~~~~~~~~~~~~~~ ~~~
. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

3) statistical analysis with the input Consisting of only noise and

then approximate extensions to sinusoids in noise

4) exact analysis of a sing le weight APS for comparison with the

previous approxima te analyses.

Figure 2—1 depicts the APS and helps to define the quantities

wea k si gnals

Input d~

~~~ ;trong signals

~~ /

ADAPTIVE POWER SEPARATOR

Figure 2-1

used throug hout this section :

is the system input at time j

is the filt er input at time j
is the filter output at time j
-

is t ne ‘e r r o r ’ signal ~ t t ime j. This si gnal is used to modify,

or upJ~ te , the fill e r wei ghts according to the ‘leaky ’ LM S al—

39
p -

_______________ -—- - - -
-—-
~~~~~~~~~~~~~~~~~
-

-
14 ~~~~~~~~~~~~~
- -

~~~~~~~~
-

~~~~~~ ~ —7- E~~L~~ - --. —~~ ___________

j
~~~~ ~~~~ p-~~
 -
-7--- - ---- -- - --.- - —— — — — - — - - -- —- - - -7--
~~~~~ ~~~

gonithm .

The opera t ion of the ‘leaky ’ LMS (LLMS) filter is defined as follows :

If L leng th of filter (number of weights)

W= weig ht vec tor = (w l,w 2,...,w J T

L
S= state vec tor of filter = L X X ...l?...S X j. T
j~ j L+l J
then

Ts
y =w
]
=

2—1
=

2—2

W j.f], = V W~ +2Ii€iS~

2— 3

p = a constant controlling rate of adaption

V = the tleak’ facto r (generally less than 1)

1-2pY

2—4

= the equivalent injected noise power. That is , the effec t of

t he le a k is the sam e as a dd ing noise of power Y to the inpu t

and tnen using a conventional LMS filter. (Note : LMS is a Spe

cial case of LLMS where V =1)

To ch aracterise the averag e performance of the APS we have to in

troduce scveral new quantit i~~ which define th.2 composition of the in

put signal

power o~ inpu t nois -_

40

-4 - .-
-
-

- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ —7--- --~~~~~~~~~~~~~~~~~~~~ - - -7-- --7---

r ~~~~~~
- --- - - -- -
-
-
-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

2
0 +), e f f e c t ive powe r of inpu t noise
p = power of input sinuso id
SNR’ = —E~. = e f f e c t i v e signal to noise ratio
a’

Witn these definitions we can draw a gain curve from input ,d , to th~
- 3

gaIn 1. _

’ ’7- —

d t~ ~~
3
-

~
SNR

APS use d to reject low powered inputs

Figure 2-2

filter output , versus SNR’ (see figure 2—2). Alternatively we c~

draw the gain curve from the inpu t , d , to the error output ,
3
-

versus SNR’ (see figure 2—3).

Thus to d isc r imin at e ag ains t weak signals w e use the y ou t put ,

and to discriminate against strong signals we use the € output. Not~

that SNR’ c a n b e v a ri ed by sele ct ion of (I sinc e:

-

SNR = ’
2

1
~~ ~~~~~~~~~

2— 5

4
41

4
- - -- - - --
I
_ _ _ _ _ _ _ _ _

—1
-

-
~~ ~~~ ~

ii
-
-

_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _______ _
~
—ii
 - -

gain 1 — —— Ideal filter

d~~to c~
\ APS
PS
1 ~~~~~~~~~~~~~~~ J, ~

2 SMR ’

APS used to reject high powered Inputs

FIgure 2-3

I -

42

— - -
7- -
4 T .

.4 7--

4.’
p
7-
‘
-
- - - -
-

- —
-

-
-~~~~-- ~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~
_ ~~~~~~~~~~~ _ f~~~~~~ _
~~~~ -
 -7-- --- ---- -----
---~~~~~~~~~~~~~~ ~~ -:=
‘.- - ~~~~~~~~~~~~~~~~~~ ~~ .
::: ~-- ~ -7-’
.

2—B . Deterministic Analysis of Sinusoidal Input

One case of interest is the behavio r of the APS when the input
consists of a sinusoid and little or no noise . This might occur in
pr actice if an APS were used in a jam resis tan t mobile communication
system and the base station were close to the mobile receiver , or the
mobile receiver were very close to a powe r ful narrowband jammer . By
assumption the sinuso idal component of the input is either the desired
signal (say narrowband FM or AM) or a narrowband jammer —— the differ-
ence is power level .

For thi~ case we let tne delay , ~~~~~~

be one unit and so:

d = a cos [Gj)
)
2—6

= a cosfej-O)

2—7
8 = wT

2—8
w = frequency of sinusoid
T = sampl ing interval

From previous analysis of the mean performance [1] we know that the
ou tpu t will be approxima tely :

y = b cos [ejJ
~
2-9

Substituting for d~ and y . in equation 2—2 yields:

4 43

- - - -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ - — --a . - -~ t -7-7- -

- ~~
-
A . - ‘ - 7-

_ _ _ _ _ _ _
-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~ _ _ _ _
-
 -
-
________________________________
(a—b) cos [9j~
-

2—10

Using this we compute:

(€ S ) . =
~~~ (cos[eiI +cos [20j—ei)}
))

-
2—11

Notice that the first term of equation 2—li is not a function of time
j, only of weight index i. The second term is a function of time and
weight index . In previous analysis of the APS , when the mean perfor-
mance was desired , the second term was ignored on the basis that it
“averages Out ” over a period . Hence only the first term contributes
to the mean wei ght vector. Indeed the mean value of the weight vec tor

can be found by using on ly the first term of equation 2—li.

-

We see
tha t the secon d term causes t ime chang e s in the w e igh t vec tor and
therefore is undesirable.

We can mod el the LLMS fil ter , which is the heart of the APS , as

an “id eal” filter whose weights are fixed at the averag e weight vec—
A tor , in parallel wi th a “perturbation ” or “ error ” fil ter (Figure 2—4 )

whose wei ghts have zero mean and vary according to the fluctuations
abou t the averag e of the actual adapting weights.

The “good” update term is ~j-~cos [9iI and the “bad” term is

~j-~cos[2ej—GiI . Both “good” and “bad” update terms are present in a

filter usiny the LL~-13 al gorithm . The “ good” term drives the filter

towards the m in int u ~n rne~ n square error solution; the “bad” term is a t

tWj cC the input frequenc y and c~ u~.es th~ w eig ht s to churn , which is

4 44

__________ - -7-— — — - ~~~~~~ ~~~~~~~~~~~~~~~~~ -

7- 7-
.
- ‘

_ _ _ _ _ _ _
~~~~~~~~~ i ~~~~~ - ._ - ---- - —~~~~~~~~~~~ — - - - - ---- ~~~~~~~~-- - -~~-~~~~~~~~ - -~ ~~~ ---~~~~- -
--
~ _ _ z _T --
--- - ----
~

inpu t x.
.3
AVERAGE _.__
p .3 ,~~~~~
I
_ _ _ _ _ _
_

WEIGHTS s.
‘.
.. .,
‘fII tei
,,
good ,,
Y outpul
u pda te terms SPLITTER ~

“bad” update terms

÷
WEIGHT
—

ERRORS

Weight perturbation model of an ILMS filter

Figure 2-4

nonpr oductive and causes output distortion . To determine the weight

variations of the weight error filter we appl y the “bad ” update term

to the LLMS update scheme (equation 2-3).

To make the problem solvable , note that equation 2-3 can be

Z— transforrned to give :

= V W j+2u€ jS]

(Z—~ )W (Z) = 2pe (Z)S (Z)

W (Z)
=
2p
~~~ ( Z ) S ( Z ) Z—( /
2—1 2

Since the Input to the system det1n~d by equation 2-12 is a sinusoid ,

we know that the output will be a sinusoid of the same frequency.
The complex gain of the difference equation at this frequency is:
4 .
—
45

- - - _
.
.. .-. - -
4 ~~~~~~~~~~~~~~~~ - - -

.4
- -.
. ~~
- -
-
.
7-
‘
.

-- --
---- ~~ -_-7-- - _~~~ -
—- _ _
-
 -7-’

g 29 =
2~i

The amplitude of the weights , W~ , of the weight error filte r will be:

a-b
—2— I g 28I
2— 13

Fig ur e 2—5 shows 1g 29 1 vs. 0 for a variety of V’s. Using equation 2— 1

we can compute , in the time domain , the output of the weight error
filter as :

=
i=l
=
~ ~j_Ig 29 I COS[29j_ei+ø )1 (aCOS (ei_ 9iJ 1
(
i= 1
b) L
a(;
~
=
~ ~ tCOS1aj-ø )+COS [39j-29i+~~~)
~~20’
a ( _ b) a ( _ b)
Ig 29 IL c O S t ej ØJ+ Ig c o $ [ 3 a j_ 2 ej + Ø J
~~ ~~ 29 I~~~~
=

2—14

Where 0 = ARG(g ) tan _ l [Im~~ ~~~~~~~~].

29 e

Th e f i r s t t erm of y~~ is at the same frequency as the desired

a(a—b )
ou tpu t an d has am pi l t ud e amp 1 = Ig 29 IL. We note that b < a , so
2
a—b < a which implies amp 1 < bu t this can be made arbi— ,

tr ari ly small by selection of p which controls y unless 0 is very

~ 2,~J
small. If we restrict our attent ion to a reasonable rang e of frequen—
ci~~ s , say 2u—& U% of Nyquist (see also 18) where Treichler wakes this

same ~.ssun ~ tion in analysing the mean performance) then : ,

2
ar.tp1 - - iy 2, i ~ 3a~ L VV
~~ ~ ~
< <

2—15

46

4
Lii
‘ - - . - -- -
~~~~~~ ~~~~~~ I
—
‘
‘

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Th —~~~~~~~~~
~~~
~~~~~~~
-
~~~ —~~~~ I T 1 ~~~ —— _- - - TJ
 - 7-- ---
- - —-
~~~~~
-- -
~~~~~~~~~----- -

0.9
\ .8
“-= 1. 0 ~ 0.7

6 .

9 .6
20

3~~~~~~~~~~~

0 50 100

fre q uenc y (as % Nyqu i st )

Ga i n vs. Fre quenc y

as a func ti on o f leak
Figure 2-5

The second term of 2—1 4 will cause an output at three times the
original frequency. ihi1 the expression for the am p l i t u d e is ra ther
~
complicated we can easil :- 5ound it:

‘a
(L
~ cos(39j—2 0i+Ø )
i=l
-4
47

A - - - -- - ~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~
-

1 1 7-
~~~~~~
-
_ _ _ _
-
- - - 7---- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~— - - ~~- —~~~~~~~~~~~~~~~~~~~~~~ - - - - -~~~~
r -
~~~~~~~~
---— ---
~
- -- -
~~~~~~
-
~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~

2—16

from which we see that the amplitude of this third harmonic distortion
is also bounded by equation 2—15.

Looking at the filter model (figure 2—4), we notice that the out-
put of the weight error filter , which is part of the total filter out-
put , was not subtracted from the desired signal d . If y were sub-
) 0
tracted from d as well as y , it woul d cause a ne w componen t in the
3 1
error signal , C , equal to —y e. To account for this , another weight
3
error filter should be added . This filter is updated using
= W 2PY S an d c auses a new ou tpu t y
V which will have terms
~~ ~ i ~2
at the fund amen tal , third harmonic and fifth harmonic . These outputs

con tribute to and so on. However , note that y can be mad e arbi.—
~
ay
trarily small , and y 2 < — 4-—(Ig ~ ( - —fI g
~ 29 I+ I g 4~~l ) ~~ 29 t+I g 49). The
we a kn e ss of th i s an al y s i s te c h n i q u e becomes apparent here. Previously
we restricted 1g 20 I by selec ting a reasonable rang e of value s for 0.
Simi l iarly we could selec t a rang e of 0 to restrict 1 4a 1 • But , i.t is
~
evióent that wi l l g ive rise to other error terms which will cas—
~ e2
cade to create an infinite number of error terms of all different fre-

quencies. Some of these error terms must surely have very large Ig i

sinc e they will be near some multipl e of the sampling frequency. How—
ever , e x p e r i m e n ta l r esul t s in d ica t e tha t i f 0 is between 20% and 80%

-
of Ny~ uis t frequency, t h en the on l y ou tp u t compo n e n t of any si gnifi—

c~..nce is the third harmonic. Ik~nce for pr actical purpo ses only the

f i rs t “wei gh t e r r o r ” filter need s to be analysed .

4
48

t _
li_Ji __ -

~~~~~~ ,*—
. _ _
- -= - -7-- ——- -~ ~~~~~~ —7--—- —— - ——7-- - _-_
~~~~~~
__ -_ --—---_-,_
~~~~~~ . .-—7---~~ ~~~~~~~~~~~~~~~~~~~~~~~~~ - — —
 -7- -7-.’

Computer simulations were per formed to verify these analytic

pr edictions. The frequency of the input sinuso id was set at 4% of Ny—
quist. The results are graphed in figures 2—6 throug h 2—li. Each
plot shows percent power not at the fund amental versus p . Thus each
plot relates total distortion power to p . From 2—6 we see that in all
cases the total distortion power was less than 0.2% of the output
power at the fund amental frequency. From 2—7 we get the same shape of

curve but now the total distortion is less than 0.06% of the desired

ou tput power. In 2— 8 the fund amental frequency has been increased to

81 of Nyguist. The error power is substantially lower than for the

prev ious cases , as is to be expected .

In an a tt emp t to s t r e s s t he AP S to cause si g n i f i c a n t dis tor tion ,

th r e e simula t ions were run in wh i c h the fun d amen tal f r e q u e n cy was 1%

of Nyguis t as opposed to being in the 20—80 % range. The results are

plotted in figures 2-9 ,2—10 and 2— il. No te tha t the g rea tes t observed
distortion power is 4% of the power of the desired signal.

A final point to consider is at what frequencies the distortion

-

power occurred . An alytic efforts indicate that the 3~~ h a r m o n i c

should be dominant. In fac t , the simulations supported this. Even I

when the fund amental frequency was 1% of Nyquist , over 9 % of the dis-
~ ~
tortion powe r was in th e 3~~ harmonic .

49

4 — -i
- -- 7- - - 7- — - .-
— -- --7- - - --

~ ~~~~~~~~~
.4 -
7-
.. - : 7- -
•1
~
- -
— ---
~~~~~ ~~~~~~~~~~~~~~~~~~ .’—~~ m-*
-
~~~~~~ ~~~~~ — -7 - - — - - — -7- 7- -- —
 - -—

3 .000
Ffl ter l eng th 64 we i ghts
-
In pu t frequenc y 4% of Nyquist
0.800 _ Numb ers enclose d i n ( )are
— power gains a t in pu t frequency
0.400 _ ~u 41
0
41
-

0.200 -

• , (0.03)
• ° °8
X~~ ~
X (0.24)
0400 -

X
0.000

0.040 -

0.020

o.o1o _ p (xlO 3)
7-i
I
l ’ t 0 - ‘
1 ITt0.

NOTE: Rate of convergence

is propor ti onal to ~i

Figure 2-6
% distortion vs. p

50
4

4 ~~~~~~~~~~~
.
- -
~~~~~~
— -— —
~~~~~~~~~~~~~~~~~~~~~~~~~
. 7- 7-
~
11

- ~~~~~~~ ~~~~
- - - - _ _ _ _ _ _
 _‘
--

~~~~~~~~~~
-7- — - -=- -— -
~~ ~~~~~~~~~ ~~~

I_coo —

a.eoo _ Fi lter length 64 weig hts

- v=O.95
°~~~~ °°:
-
Inpu t frequenc y 4% o f Nyquis t

0.400 -
~1
~~
•0

- ‘0
+1
0
4’

0.200 -

0.300 -

t —

0.080

0.080 _
x
-
x
x
0.0*0 -

0.020 -

x 7 -3
p(xlo )
0.CI0 1
~
I $
o %o ~ $30 $3O
~
%ca
~
NOTE : Rate of convergenct
is proportional to

Fi gure 2-7
% distortion vs. p

51
I

- -- —.- w wr--
~~~~~~~~~~~~~ ~
-

- - ‘•. •L? -
7-
7-7-

~ ~~
ii- ,,
- --
- - —
- — — - I . i:
i
Tii
-

~~~ ~~~~ ~~~~~~ ~~~~~~~ ~~~~~ -

-~~~~~~~~
~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~
_ _ _

: —-
~~

0.100 —

Filter l ength 64 weights-

—
v 0.90
0.080 _
-

Input frequency 8% of Nyquist

:F
—
Numbers enclosed in ()are
power gains at input frequency,

i
~

0.010 -

O~0C8 :

0.006 -

7-
0.004 _

,(O.02) (0.05) (0.21 ) (0.39)

0.002 X X x
7-7-
p (xlO 3)
0 ~~
$3
-
I I I I
6 L I
$3 I

NOTE: Rate of convergenc

is proportional to
p.

Figure 2-8
% distortion vs. p

s 52

~~~~ ~~~~~~
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ _ _
 _ _ _ _ _ _ _
_ _ _

—-— -- -- - - - - - -~~~

10.000_.
Filter l ength 64 weights
v~~~O.9O
-

Input frequency 1% of Nyquist

-

6.000 .. Numbers In ()denote power- gain

— at input frequency.
4.000 .~ (0.05)

2.003 ..

C -
0
1.000 _
~~
-

X
0
0.800.. 4’
U,
-
~0
0.000 - —‘0
4-)
— 0
4-)

0.400 -

(0.68)
•
-
x
0.200 -

p (x10 3)
0.100 ...
I I I I I I I I I I I I I I F1~ i
1 2 4 6 6 10 20 40 CD 80 100

NOTE: Rate of convergence

is proportional to
p.

Figure 2-9
% distortion vs. p

I
53

4 — - - .- 7- 7 --
~~~~~~~~~~~~~~~~~~~~~~
._ rE ~~ 7-••7-~~~~
-

,%
~~~~~~ •

- 7--- ___ _ _ _ -- -—
- - - - - -~~~~~~~ -- --- -- - ---

-— ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~ ~~~~~~~~~~~~~~~~~~~ ~~~~~~~~ ,_

~
—
: - ~- = ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~ ~~~~~~~~~~~~~~~~~
--

10.000..

0.000 :
Filter l ength 64 weights
v = 0.95
—
Input frequency 1% of Nyq uis t
8.000 ...
-

Numbers i n ()deno te p ower

gain at input frequency .
-
4.000
(0.10)
-
,
~

2.000 —

I•000 :

0.000 ... 4-’

-
•0
0.600_ .—tO (0.63) -

4-)
- -
0
4-)

0.400 _
~

0.200 ... (0.78)

x
3
p (x10 )
0.,00 _
I I I I I I I I I I I I I I I I I I
1 2 4 C 8 10 20 40 bo eo ~I

NOTE: Rate of
convergence is
proportional to
p.

Figure 2—1 0
% distortion vs. p

54

-4 - - — - ________ - - “ -- - -7- ’ -•--—•~~~~~~~ ——,- - .•- - _ “

-~1
-

~~~~~~~~ ~~~~~~ ~

-
~~~~~~~~~~~~~~~~~~ 1 L
_
~~~ _ _ _ _
—
—J
 — -——----- --— -—
~~~~~~~~~~~~~~~~~~~~~~~~ -

30.000.
-

Filter length 64 weights

e.oco _
-

v = 0.99
- Input frequency 1% of Nyquist
0.000 - ~ Numbers in ( ) denote power
gain at input frequency .
I
-
0
-
4.000 .
. ~~~~

‘0
7- 4-)
0
4.)

2.000 _

3 .000 -

o.voo . , (0.31 )

0.600 -

0.400 -

(0.73)
x
- -

0.2C0

(0.86) p (xlO~3)
0.1c0 _
I I I
Ia I $3~I ~
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

I I
~
$3 $3 ‘
-

Figure 2-11
% distortion vs. p

55

i _ _ _ _ _ _ _ _ _ _ _ _
- -- -- -
—
_ _ _ _ _ _

~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~

- - ‘: - -- - - - -
1
4 ~~~~ ‘ -: 7-
-- —-
~~ 1
.7- _ _ _ _ — -

-- —7- ___ - -
_J__4- _ * - ~~~~t~~~~~4 - - • ~~~~~~~~~~ r~~~~~~~~~~~~~~~~~~~ -~~-~~~ ——t ~~~~~~
 -- - -
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~

2—C . Statistical Analysis of Sinusoidal Input

From pr evious work El) the general form of tha APS output is
known . If we use equations 2—6 ,7 & 8 to define the input , then we

know that the output will be approximatel y:

y. = bCOS(9j) =
a
COS (9j1
L SNR’
2—1
If we d e f i n e ~~q . as W~ _E (W

2—2
and then notice that (i— (/)E{W) = E (2ji€S J,
so by taking the expectation of 2—3 and 2—18 we get:

- 2— 3

hence
~ W(Z) —
—
1

but if we know V1~R (2~i€. S .} then we know

vARfL~7w
- ~} ~~~~~~~~~~~~~~~~~
2—4

VM (211€ .S .} = VM~{2ei[aCOS (aj)- a cOs ( a j ) J s . )

3 3 1+ 2 )
L SNR’

4p 2VAb(- a
cos(9j)S .}
,
=

(vAR{2 ~ E~ S~ ))
~ = 4
d
~~~~~~~~~~~~~~~~
VAR ((CCS [~~jJ) (uCOS(9j-6iJ ))
L S

-4 -
56

i_I
. — .— —— — —— — - - -— - —7--— .7-— - 7-— - — - - - - -
-‘- —— ~~~~~~~~~~~~~~~~~~~~~~~~
4 ‘:—
-

~ ~
•
ci.
7-

~ i~
-

7- T~~~~ J T - ~~~-~~-- ~ _ ___J_ ___ 7- - _ _ _ _ _ _ _ _

 -
_ _ _ _ _ -- J L_ ~~~~~~~~~~~~~~~ ã~IL -~~~~~~~

2 2 2 4
=
4~i a u a
L SN R ’ 2
—

) 8
—

~~1~~~ ~~~~ ~~~~ L SNR’ 2

2 1 )

2—5
Tb us

(VARL& ~ )
)1 = _ _ _ _ _ _ _ _ _ _ _

2
2 4
—
1
~i a
e i29 — V
-
L SNR 2
2~ 1÷ )

2— 6

These formulae are in close agreement with experimental results

as Table 2— 1 shows . In the cases where w is 12.5% of Nyquist , t h e r e

is a 15% e r r o r in det e r m i n i n g VA R ( 2PE ~ S }. althoug h VAR (.
~~W1 is still
)
given by equation 2—19 if we use the ob ser v ed v a l u e for VA R (2 P€ ~ S~ }.

To suminarise we used a simple statistical analysis to determine

th e v a r i a n c e of the d r i v i n g f u n c t ion , 2~ €~ S~ (equation 2—21). Then we

considered how this affected the weight vec tor (equation 2—19) and

derived a relationship (equation 2-22) for the weight variance in

terms of the input quantities. These equations (2—21 & 22) will give
a good idea of the weig ht variance (note VAR (W } = VAR ( provid—
3
• I
in g the inpu t f r e quency is not to near DC. A reasonable range of in—
pu t f r e q u e n c i e s appe a rs to be from 20% to 80% of Nyquist.

57
I

7-4 .

i.1•
— —7-- - - - - 7- 7 - 7-.- 7--• — 7- -— - -.
. 7-’-- -
~ ~~~ ~~~ ~~~~~~~~~~~~~~~~~ ~~~~~~

7-.
7-

-—
.

- -- - --
-

= -- - --- — — - — - -
- .
~~~~~~~~~~~~~~ ~~~~ t_c —_ ~~~~~~~~~
-
- •
~~~ ~~ ~~~~~~ ~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 -7 - - — ----— - - _ _ _
i --
- -

~~=~~~ —
—-- -— —-
~~~ ~~~~~ ~~~~~ ~~~~~~~~
~ ~~~~ ~~~~~~~~~~~~~~

v w CVAR MVAR % error

lv ~~~~ .‘)
- . J j~~
)4. •~~ . 2
~~~ ~~~~~~ ~~ .

7-
~~~~~~~~
Th.
~~~ ~~~~~~~~~~~~~~~~~~~ 4.55xl~., 11 o.:
L4 ) -i.5- ;l 7- 1. li ~~
~~~
IU i J .. J

7-
j l~~.5 4.
~~~:tlu~~ ~~~~~~~~~
i~~j
~~~~~~

.LU
- —4
,
ii.
— —
i2 .~ - 7- i t —Il
~~~~~~~ ~~~~~~~
j —
1~,
)
l~.. — 1~ l.jj•,.j.
. — 12
-~.
~~-*x 1u
j
S -
~~~ ~~~ - ~~~~~~ -

~
_ _ _ _ _I
lu 4..~.)
;.lu
—3 _I
Au
_i 7-- — - - - -
2 v .~~
~~~~~~
~

1,_i
~ ~~~~ ~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~

*Th i s va l ue almos t cer ta i nly i ncorrec t d ue to numer ic al

inaccuracies in the computer simulation .

w Is measured in % of Nyquist frequency

CVAR is the calculated variance
MV A R -is the measured variance
MVAR-CVAR
% error = x ioo ~

Table 2-1

58

- - -
- - -- —~~~~ -
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -
- — ~~~~~~~~~~~~~~~~~

.4
-
- 7 - -- -
- - . - -j -.
~ ~
~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~
 -
7-- - - --~~- ~~~~~~~~~~~~~~~~~~~~~~~~~~~ - I —~~~

2-D. Statistical Aalysis of the A d a p t i v e Power Se para tor (

AP S) wi th
Noise Input

In a real communication environment the input to an APS might

well consist of both sinusoidal components (either signal or nar—
rowband jammers) and broadband noise (thermal noise or wideband jam—
mets)
. Thus we wish to understand the behavior of the APS to a

sinuso id in broadband noise. The first step is to characterise the

wei ght vec tor var iance when the input consists only of noise .

Note that:

VAR (X} = E{XX

T
}_ E { X) E ( x }
T

2—7
also
o j =n j

x .=n .
3 3— 1

where

is from a whi te, zero—mean Gaussian process with variance cr2.

2—8

From equations 2—2 ,3 & 4

W~~1= (I_2IJ YI_2~JSS T)W ~ + 2e.zd~ S~

2— 9

If we assume E (SW TI = 0, which is a common assumption (well support

by expt~rier .ce If ~ Is small enough) , then we find :

-4
t 59

4 1. - - _ __ - -7- - — —- -- ---- 5- — -~~~~~~~ — —-

S
~~~~~~~~~ -~~~~~~~~~~~~~~~~~~ -~~ - - -
~~

11 — — — -- _ _ _ _ __
__ _ _
j
~~~~~ ~~~~~~
- ~~~~~ _ _ _ _ _ _ _ _ _
 _ _ _ _ _ _ _ _ _ _ _ _ - 7--—-
_
-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

W 2
= ~~ ( I— 2~~Y I— 2~icr I) ~~ + 0

2—1 0

ELW ~ W ) —
÷1 ~ +1
V A R tW ~~~~ ) = W ~~~ W ~ ~~~
1 j 1 j 1

2 lR(W 2 2 4
= (1—21i Y) V !
~ ~ )—4Ii (l-2J1Y)a VAR (W~ 1+4II a I
2 T
+ 4il E(SS VAR (W~ }SST}

2 2 2 2 2 2 4
( l— 4 ~1Y+4P Y —4I.I c7 +8M Yo7- ) VA R (W ~~J +4 P cT I+O

2—1 1

Whe re we have assumed :

T T 2
i1E (SS VAR { W~ )SS }<<O VAR( W~ }

2—1 2

If we no w assume s te a d y s t a te so t ha t VAR ( W ~ ) = V A R(W +i 1P t h e n :

~

-
-
4p 2 a4
4~io +4~~iY + 4e i Y -8ji } a
‘

=
pa2

7-i

PC.

- -
a
2—1 3

Pro vidi n g p~ << l +-~ and

a a

>

2—14
-4
60

-_——
-.
--- . -
I
- .c ‘~1
4

Lu
~~ ~~~~~~~~~~~~~~~~ ~~~~~~~~

~~ -

- -
 Notice th a t for Lt’IS Y = 0, and so VA R (W} = po 21 which correspond s to

earlier analysis L 91 of the LMS filter. Also note that the ‘leaky ’
algorithm never increases the weight variance , in fact it actually
reduces the variance. This supports the conjecture (11 that algo-

rithmically simulated noise (via ‘leaky ’ LMS) is more desirable than

actual noise injection , even thoug h the converg ed mean weight vec tor
solution is the same in both cases.

In equation 2—28 we assumed that

T 2
PE{SSTVA R (W } SS }<<a vA R (w )

—
To check this assumption we no t e t ha t :

E{ SSTV A R ( W ) SST} = O 4VA R{W) + o-4Di a g [V A R{W} ) + a4TracetvA R (w 1JI

2—1 5
where Diag [ . I means the m at r i x consis t ing only of the diagonal ele-
men ts of the operand .

Assuming that 2—29 is substantially correct lead s to the following

conclusion :

E{SSTVAR (WISST ) =
6 6 6
2pa ~~ Lp a p (L + 2 )a
1 1

a a a
2— 16

Therefore 2-28 requires:

t
_ _ _ _ _ _ _ _ _ _ _ _

a a

4 -
-

61

5 - -
- ~~~--— ~~~ —
5 S
_ _ _ _ _ _ _
~~~~~~~~~~~~~ ~~~~~~

L
~~~~~~~~ I ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 -- —- — -

-- 7
.-— —

which implies

2—17
-

This condition is often assumed in ad aptive filtering and it im—

plies slow , low—noise adaptation.

Extension to the case of a sinusoid in noise is unfortunately not

obvious. Intuitively it is appealing to argue that the correct formu—
la should be:

VAR (W} = ~~ min ~~ -

mm -
- -

by analogy wi th the conventional. Lf4S forn~u1a of :

VAR (W ) =

Experiments ind icate that this is not the case , in fact it is

easy to see that this would not be true since in the no noise case the
f o r m u l a is su b s t an t i al l y incorrect.

Ano the r approx imation assumes that . the noise and sinusoid affect
the weights independently, which results in the curve of figure 2—12.

~hi s approach is not strictly correct either . However , if the

—
weight. variance due to noise alone is substantially greater than the

vari ance due to the sinusoid alone , then tri e formula is indicative of

the ~~tu~.1 vari ance . This result has been checked expe r imentally , and
some results are presented in Table 2—2. An important point to con—
sider is that cor~putin g we ight var iance by using the formula:

4 5

62

~~~~~~~~~~~~~~~ ;:IT.
J.:
 , ,
I ~
weight ?
/ ~~~~~~

varl ancel
‘

combined
variance

/ variance due
,
to noise only
#
,
/
-‘ variance due to
sinusoid only

noise power

Weight variance obtained by assuming independent effects of

noise and sinusoid

Figure 2-12

VAR (W ) = p (error— output—power )

or

p (error— output—power )
VAR ( W )
1+
(error—output—power)

(as was done In (it)) for example) is apparently very conservative.

The power due to sinusoiual components does not contribute to we ight
variance as muc h , proportionally, as power due to r ndoin components.
~
he are currently wor k ing to quantify this effect.

63

4 -
-~~-- — -•-
~~ . -..-
- -

‘1
— —
~~~~~~~~~~~~~~~~~~
-
~~~~~~~~~~~~~~~~~~~~
~~~~~~~ : ~~~ _ i~~ ~~~
-
-- ~~~~~~~~~~~~ -- ~~~~
 7- - - -— ~~~~~ --~~ — S _ _ _ _

noise 1-y p weight approx. #1 approx. #2

power variance

1.l7xlO 3 10 10 3.27xl0 8 9.31x 10 6 1.37x10 9

~~ ~~
3 4
1.0 2x10 5x10
4 1.75x10 4 3.27x 10 4 2.5x 10
7- 1.0 io~~ 5x10
5 5.67x10 6 3.27x10 4 2.5x10 ’5
~
1.0 l0 5x10
4 2.56xl0 4 3.27xl0 4 2.5x10 4
~~
.6 7-7
1.0 10~~ 10 5. 8lxlO 9 6.55x10 5.0x10 7
5x10
4 10~ 1.29x10 8 9.18x10 6 2.5x10~~°
~~~

Frequency : 25% of Nyquist

Power of s i nuso i d: 1.0
Number of weights : 20

approx. #1 VAR = -—

1+ —

“men
Ii

approx. #2 VAR =

1+
7-r
‘1
Table 2-2
Comparison of measured weight variance and
two approximate expressions

64

*
-
. _ ..
. -
-
___________

4 ~~~ ~~

_~~~~~~~~
±_ *
_ -
7-
-- -- -
_ _
- — ~~~~~~~~~ -~~~~~~~~~~~ ~~~~~~~
- - —- -7-- -5- 7-— -
~~~~~~~~ — 7- — -~~~~v - r ~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ : i --.
~~~~~~~~~ ~~~~~~~~~~_
 -
--
~~~~~— - -- -5 -- - - 5 - - - — --~~ -5-
-
- — - -5- -—- 5--
~~~~~~~~~~~~~~~~~~ L~
—

2—E. The Single height Power Separator

Analysis of a single weight adaptive power—separator is of great

in terest for many reasons. Among them is that a complete analysis can

be performed with no need for approx imations. In this report the mean
weigh t val ue and weigh t v a r i a n c e a r e analysed , which lead s to some
surprisin g results that further illuminate the performance of an LNS

filter. The results show that previous approximate analysis of

mul ti—weigh t filters by Widrow (9, 10) , Senne [111 , Brown [1 2) , an d
Davisson [131 to name a few , ar e accura te enoug h for useful resul ts in
mos t c ases , wi th the greatest errors occurring during fast adaptation.

The configura tion of the adaptive power—separator is a leaky—LMS

filter in a line—enhancer as shown in figure 2—1 . Note that in this

case W~ =w~ has a single element; and S =x~~, a sing le element. The in-
3
put , d =n~~ is composed of a DC value , a , w h i c h is the ‘s i g n a l ’ ; and
3
white , zero—mean Gaussian noise with variance cr2 . Ph y s i c a l l y th i s may
be thought of as a degenerate (zero—frequency) sinusoid of power a 2 in
noise of power or2 .

By substituting Eq. 2— 2 into Eq. 2-3 and then expanding Eq . 2—3 , we

find :

N-i N
W N-f = Z~ n N n s,_ 1 + 2~i ~ (n .n .
1 1_ 1
fi
~ j=i+1
N
+ W
0
f l (V —2pn~~ 1 )
3
2—18

Using this relationship we can derive the expected (or average) weight
-I
65

4 —- - — — 5- 7-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ W — fl~~
— 5- -
.4
1
— - —- - -—-—- - - - -- -
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~ 7- — _ _ _

1 - — - -
~ ~~~~~~~~~ ! — ~~~~~~~~~ ±~~~~~ 57- --— —a- -- -- - —— -
~~~~~~~~~ --
 .7- ,; —7-
_5-_ _ --- - - -5-- - - — -5

value as:

E (W
N+l
) = E{ 2 1n N:
N...l J + 2ji~~~ E (n . n .
1 fl ((/—2~in 2 )J
~ 1

+ W QE ( II (V —2pn .
1
)}
j=0
2—19

Usin g the f o l l o w i n g prope r t ies of whi te , G a u s s i a n noise wi th the d e-

fined parameters:

2 2
1)
E{n .n .} =
2
1
~ a ij
~
E ( t (n~ )f(n~ )} = E{f(n )}E{f(n.)}
1
if i~ j
we get:
N-i N
E {W = 2~ia 2 2p ~ [aE (n .((/ —2pn~ )) fl E-( (/, -2j~n~~ ))
N
+

j 1-f- 2 3 1
N
+ W
a fi E {V —2 ~in~~ 1}
j=O

N—l N
= 2pa 2 + 2p ~ [ a ( a (/ -2~ia 3—6~.ia2a) fl f
(( ,
2 2
—2iia -2~ia ) J
i=O j=i+2
N
f l (V
.-~

+ W~ —2~ a ’
— 2~icT’)
j =o

2 4 2 2
+ 2p~~ ((a V -2 ~ a —6 ~ cr a ) ( V

+ ~~~~~~~~~~~~~ ~~~~~~~~~~~~~~

2~~(a 2 V -2 4-b 2a 2 [i_ cv

2~ n 2
2 : N]

)
_ 2 ~~~~ Cr 2~~~~~~~~ )
= +
~~ /
I l— (, +2p (o- +a

-4
66

~~~~~L
 - - - 7- -5 -- —-— --5- -- - — - - -5 -~~~~~ _ - - - - 5--- ~
5 ~~

_ 2jJa 2_ 2po.2 N 7-]

+ W
0
( (/ )~

2—20

This equation (2—36) expresses the mean val ue of the weight as a

fun aPion of the amount of data , or time , used in adaption. An in-

teresting result is the final mean value of the weight. We see that

if IV — 2~.zcr2—2~ia 2 I<l then the adaptive process is stable (convergent) ,

and so we can find :

2)
u r n E{W } = 2pa 2 2~i (a 2~ — 2pa 4-6èici2a
N->oo N +

1—V +2p~ +2j a

2 2
a — 4pcr a 2
—
2
a + (a21-~~~”
2~i
2—21
The Wiener solution (minimum mean—square—error) for the converged
weight v a l u e can be shown to be;

2
W
*
=
a
-)
a +0

2—22

The adaptive powe r separator based on the leaky LMS algorithm gen-

erates a converged solution sim iliar to a Wiener solution for a prob-

lem in which the variance of the noise is ~ 2+1~~~ . However , the adap-
tive solution g iven by equation 2—36 is not quite equal to the Wiener

solution since there is an extra term of Senne (11)

2 21— ” .

dcrnori~~tr~~ted by exp€ ~r itnera t that LMS filt e rs do not converge , in gen-

eral , to a %~iener solution if the inputs are correlated , and since the

inp ut has a DC component it is correlated over time. Still , we note

a
6/

.4
-
- - 7- - 5- - 5- -
. V-.-
~~~~~~~~~~~~~~~~~~~~~
t - -

i -
~~~~ ~~~~~~~~~~~~~~~~~ - - ~~~~~~~~~~~~~~~~ ~~ ~~~~ -~~ -
- - -
-
 — 7- - -

~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~

that if p<<—.i the adaptive solution will be essentially the Wiener

~
solu tion.

To reconsider the conditions under which we can find the l i m i t in g

v a l u e of the wei g h t , we note that V is chose n in the rang e 0 V ( 1

for normal use , an d h ence f rom th e conver ge nce c ond i t ions tha t:
2 2
IV — 2~ior — 2pa I<l

2 (a +a
2—23
will guarantee that the mean weight value converges. Note also that

transien ts die g eometrically as

While the average , or expe ct ed , solution is very useful in under-

stand i ng the behavior and utility of the filter (see for example pre-
vious final re port [11 ) , i t i s n ot t he compl e te s to r y . In t hi s sec-
tion we shall analyse the variance of the single weight . From this
informa tion we can dete~rmine how muc h extra noise (misadjustment (1))
appears in the output due to the adaptive process. Also , we will find
a new convergence criterion.

) , the mean—square value of the

The first step is to find E {W
~ +i
weigh t. Knowing this , we can determine the varianc e of the wei ght by

using the following formula:

4 . VAR {
~~ ,~~1
} E(
~~~41 }
— (E{W N+u }) 2

2— 24

For N)3 wc ~~ n show that:

I

68

————_ — --- 5 -- -7- - —--s--—-— - --— - - r SSfl.t11 14

t !~~~~ -
~

~1 211~ j
7- -
-7-— -
- -7-—- - ~~~ 7-~ 7-_~ .j ~~~~
5- __5-_ __ i
~ ~ ~~~~~ -~~ 5—
 ___________________ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~

2 2
E (W~~ }
1
= 4i p
~ + 8~i 2a ’ 2 2
~c + 8p a q 2[1js ]

4pW aqs t~ 2
4~~ pv [1_ rN] 81 2aut [~7-j~
0
+ +

r]
+

2 N_ 2]
8J ~~a~~ u [l_ r~~~2 s 1 _ sr
+ 4~ w uqs~~~1
0
+

4 N+l
+
~~~~~~~_ ] w~ r
0qu [ +

2— 2 5

Where

q (I a—2Ja (a3+3aa2 )

r 2
— 4 .&V ~~2~~ .2) + 4 .1 2 (a 4+6a 2ci2+3o- )
(
~ ~ ~ ~
s -
~~
2~i (a 2+a2)
(I —

2+cr2
2 p(a 4 +6a 2a2 +3cr 4 )
~ V (a ) —

u
~ V
2
a 4 i V (a 3+3acr2 ) 4 12 ( a 5 +l0a 3 a2 +l5acr4 )
—

~ ~
+

2 (a 2+cr2 — 4 2 2 4
V V ) 4
~-’V
(a +6a a +3o )
+ 4p 2 (a 6+15a 4a2+45a 2a4+l5c16)

NOTE: this result was reported earlier in Quarterly Report 3 f o r t h i s

~
contract. However , d ue to a n a l g eb r a i c e r r o r th e o r i g in a l equa t ion ,
-
-

equa tion i& , is incorrect. Thus Eq. 2—41 and 2—42 of this repor t sup-
q.. 8 and 9 of Quarterly Report ~3, “Exac t Analysi s of a Special
plant E~

Adaptive Power Sejarutor ” .

This is a formidable expression ! However , i f I s I< l and Ir I<l ,

then u r n E (W~ } = E ( W 2 } will converge to:

N->oo
2
E(W 2 1 2 2 4~.i pv
8p 2 1~ 2 t
a c1
= 4~i p + +
~ -~~1 S +

69

.5-- ,
5- -.-~~~ .. -

4
- —

—
- .
5,
~~~~~~~
— - 7- - - - .5-
—~~~ -5 _
5_ 5- -- - - - ~~~~~~~~~~ -
-

7- _”. -
~~~~~
—
~~~~~~ ~~
_ ._l
~~~~~~~~~~~~~~~~~~ — ~~.
 2 2
8i.i aut 8~i ag 2 u
l—r (1—s) (1—r)
+ +

2—26

Fur thermore , if we take a to be zero and V to be 1, then

VAR E W} 4p 2cr4 4p 2a2 (or2_ l 2pcr4+6W .i2or 6)

= + — E{W J
4por —l2p a

2 2 2 4
=uo [l_8Mor +4~~ a
l-3jior
2—27
4

This case is that of a conventional rIMS filter , and was analysed

by W idrow in [5 ) Th e approxima te analy sis used by Widrow ind icated
that the variance of the weight should be pa2 . This agrees very well

wi th the exac t a n a l y s i s it w e r e c a l l Wi d ro w ’s stipulation that

<< 1, whi c h correspon d s to slow , low—noise ad aptation.

Recall that convergence of E (W~ } required that Is I<1 and r I< l .

The first condition becomes :

I s I< l =>I V —2p (a2+cr 2)I < 1

2(a +0
Which is the same condition as for convergence of the mean weight

value . The second condition becomes:

IrI < 1 => IV 2 4pV p+4p 2 (a4+6a 2a2+3a4)I < 1

2— l—4p (/p+4p 2
= a (a4+óa 2o~2+3o 4)
~
>

2+a2) +v + a 2 2 4 2 2 2 2 4
V c ~~ ~ or +3a _ 4y a o — 2V ~
-
~ 2(a +6a a +3o-
-

2— 28

4
70

__
4 -- -
-
—-5-.
~~~~ 5- .M~~~-—.*..~~~1
*-r . - -
-
.4 ‘7-
-
Ill
~~~ ~ ,
-
- --
~~~~
- A-- - -

- -
~~~~
-=

~~~~~~~~
~~~~~~~~~~ - — --
~~~~~~ ~~~~~ ~~~~~ ~ . - - ~~~~~~~-~~~~~~~ - ~~~~~~
 2
if SNR
a
(
/ (SNR+1 ) 2 2
~/~ NR +6SNR÷3— 4~~ SNR— 2~~~
+

~
2 (a2SNR+6a 2+3cy2)
2—29

Again , this is not a particularly simple expression. Figure 2—13

is a plot of the max imum possible ~i vs (
I for various SNR ’s (holding
total inpu t power = 1, max imum p assuring convergence of mean and

100,000

-~

Maximum p for convergence of variance vs. V , as a function of SNR

(holding total input power at 1).

Fi gure 2— 13
vari ..n~~ ot the wei g h t vector). From this we see that there are two

cases : a) hi gh St~R and b) low SNR . In the high SNR case

cqu~.tt osi 2—44 simplifies to:
(
V +l
I x tota1—ir ~put--powe r
4
71

.
4 - - - - - - - -
~~~~~~~~~~~~~~~~~~~~~~~~~
 r -~~~~~~ -- — ~~~ --
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ - ~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -- .-—

-~
2—30

w h ich v a r i e s from 2 x to
total—input—power input—powe r as V varies
from 0 to 1. In the low SNR case equation 2— 44 simplifies to:

~~‘
6 x total—input—powe r
2—31

which varies from 6 x

total— input—powe r to 3 x total—input—po wer as V
varies from U to 1. The most stringent requirement on p occurs when
V 0 (although in actual practice V is rarely less than 0.99) and
SNR is low. For this case ~ é7-
~ ~ tot ~pu t—pow~ ~ inp ower ’
If p is selected by this criterion , the variance will always converge
and remain finite. Note that this condition is more stringent than
the condition for guaranteed convergence of the mean which required
that:

‘<
~~
total—input—power
in the worst case . Also , no te that the conditions for convergence in

the mean presented in [9 , 51 for an LMS filter I< input powe r) where
~
(1 = 1 is not s u f f i c i e n t to guarantee convergence of the variance in
low SNR regions. Of course , in normal prac tice p is very much smaller
than the bound so this problem does not arise . However , if a p is
selec ted so that the filter converges in the mean but not in variance ,

then we would expect the weight to oscillate randomly about the mean
value with ever larger oscillations. This analysis of a single—weight
filter corroborates Senne ’s earlier expe r imental work with multi—
wei ght filters: that to guarantee convergence of the weight noise
variance requires p to be several times smaller than the val ue needed
to assure convergence of the m ean weight value .
‘4 72

- - 7- — — -- - , - 5- 7 - I~,- -
. - - —- 5 5— — ~~~~~~~~~ -‘~~~~~ ~~~~ ~ ~~~~
7- - - -
- - -
_
-

--

~-~
ui. -—
.:— ;. ~
—

~~~~~
-
~
- - — -- - -- -
— -
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
_ _
_ _ _ _ _
_ _ - -_ _

2—F. Conclusions of Part 2

We have stud ied the second order effects of the Leaky LMS (LLMS)
algorithm . The LLMS algorithm causes the weights to var y about the
mean or expec ted solution. This will modul ate the output of the
filter , causing added undesirable noise components in the filter ’s
output. For sinusoidal inpu t s w i th f r e q u enci es be tween 10% and 9 0% of
Nyquis t f r e q u ency w e fo un d tha t t he d is tor t ion power was less than
0.2% of the des~ .able output power. For very low frequency i n p u t s the
dis tortion power increased , bu t even a t 1% of Nyguist the distortion

power was still less than 5%. Furthermore , most (96%) of the distor-

tion was in the third harmonic .

For whi te noise inputs to a LLMS based Adaptive Power Separator

CAP S) we have derived an equation for the variance of the noise in the
weight vector. This equation (Eq . 2—29) agrees with previou s analysis

(5, 9, 1]) of weight vec tor noise in [1M S filters , a special c a se of

LLM S in which V = 1. This analysis also confirms an ealier conjecture

( 1 ) tha t the LLM S a l g o r i thm woul d have less wei gh t noise than an
equivalent injected noise scheme .

Finally, an exac t analysis of a special case (single—weight) of

— the LLMS driven APS was per formed . This analysis confirms Senne ’s
Ill ) observation that LMS does not converge to the Wiener solution if

the inpu t is correlated . However , the b ias from th e h i en er solu t ion

can be mad e arb itrarily small by decreasing p. Also , we found that

LLMS (and LMS) filters may not stabaliz~ in a m ean—square sense even

thoug h they converge in the me an. A new criterion for p is presented

73
4_p51 -
— 7 - -- . - —7-—-- — — —- - —
~~~~~~~~~~~~~~~~~
, -..---
~
- - - -
~~
—

-: -- —

~ - -
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
~
 - -5-- ~~~~ - - - -— - - - — -5~~— - -- - -- -—- .--
— ~~ — - -——

which guaran tees mean—square stability.

-4 74 -

(4 -7-—— — - -5— —5- ---_

~~~~~~~~~~~~~
-.- -
~
._..
- -— — s—— - ~~~~~~~~~~~~~~~~

p
1
1

_ __ • -. •-5-5-5 -
•• - —~~~~~~~~~~ ---~~~-~~~~~ , -.~~~
~ ~ ~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 - - - - —--—- -- -— ~~~~~~~~ --~~~~~ ~~~~~ — - —-- _ -~~~~~~~~~~~ - - 5-- — - -

REFERE NCES

1. B. Widrow , R. Chestek , J. R Treichler , “Final Report on Adaptive

Antenna Techniques” , March 1977, Naval Ai r System Command Contrac t
No. N000l9— 76—C—0250.
2. 8. Widrow , J. M. McCool , M. G. Larimore , C. R. Johnson , Jr., “Sta-
tionary and Nonstationary Learning Charac teristics of the LMS
Ad a p t i ve Fil ter ,” Proc. IEEE, Va]. 64, No. 8, August 1976, pp.
1151— 11 62.

3. J. K. Kim and L. D. Davisson , “Adaptive Linear Estimation for Sta-

tionary M—dependent Processes ,” IEEE Trans. Informa tion Theory,
Vol. IT2l , pp. 23—31 , January 1975

4. L. E. Brennan , E. L. Pug h , and I. S. Reed , “Contro1— 1oo Noise in

-‘
Ad aptive Array Antennas ,” J. Acoust. Soc . Am., Vol. 34,~ pp. 239—
297, March 1962

5. B. Widrow , “Adap tive Filters I: Fundamentals ,” Stanford Electron-

ic Labs Technical Repo r t No. 6764—6, Decem ber 1966

6. B. Widrow , “Adap tive Filters ,” from Aspec ts of Network and System -

Theory, edited by R . E. Kalman and N. D e C l a r i s ; H o lt , R e i n h a r t ,
and Winston, I n c . , N. y . , 1970
(Al so r e p r i n ted in “Proceeding s of the Ad aptive Antenna Systems
Wor k shop, ” March 11—1 3 , 1974 , Volume I ; W. F. Gabriel , w o r k i n g
c h a i r m a n , published by Naval Research Laboratory, Washing ton , D.
C. as NRL Report 7803 , September 27 , 1974)

7. 8. Noble , Ap plied Linear Algebra , Prentice Hall , Inc., Engle wood

Cliffs , N.J ., 1969

8. J . R. T r e i c h l e r, The spectral line enhancer: the concept, an Un—

p l e m e n t a t i o n and an ap p lica t ion Ph. D. d isser ta t ion , Stanford ,

Un i v . , S t a n f o r d C a l i f ., J u n e . 1977.

9. B. W i d r o w et a l ., “ A d a p t i v e noise c a n c e l l i n g : P r i n c i p l e s and Ap-

plica t ions ” , Proc. of IEEE, vol 63 No. 12 , pp 1692—17 16, Dec.
1975.

10. S. T. Alexander , The ALE Out~ ut Covariance Function For a Sinusoi

in Uncorr e la ted N~ T~ e7~~ echnical report 1b2 , Naval Ocean Systems
~~ n ter , San Dieg o Calif., 1977.
11. K. Senne , Adap tive linear discrete—time estimation , Stanford aec
tronics Lab., Stanford Univ., Rep. SU SEL 6~ — 090, June l96d (Ph .
D. dissertation) .
12. J. E. Brown , Ada ptive Estimation in Nonstationary Enviroments ,
Stanford Electronics Lab., Stanford Univ., Rep. SU— SEL—7u— 056 ,
Aug . 1970 (Ph . D. dissertation) .
4
75 —

4 -
- 5 - -- - -— -— - —- —
~~~~~~~~~
-
—— —~ 7- — - .

‘ -~.-
‘_ _— .

.~~~ .
~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~
-- ~~~~~~~~~~~~~~~~~~~~~~~

.7- p
i
— - .
‘
-

—
a- --
- -

- - -
~~~~~~ .
~~~~~~
-- - -‘- - - . -- - -I -
-
 - - - 5 -- -— - - - - - — --5-- 5- -5- -5—~~~~~~~~~~~~

~~~~~~~~~ ~~~~~~~~~~~~~
- - . -- 5-5-5-
~ ~

13. L. D. Davisson, Steady-state error in adaptive mean—square

minimisation” , IEEE Transactions on Information Theory, vol IT—1 6
No. 4 , pp 382—385 , July 1970.

a. .

4 76

4 -7- -— — —-5- — —7-

~~
—
-
- —— - -
~~~~~~~~ L
-

i
i
_
7- - 5 ‘5
- --

~~~~~~~
- —

~
- _ _4; I

&~~_
~~
i~. —

~~~~~~~~~~~~~~~~~ _ _ _ _