Professional Documents
Culture Documents
~
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
~
-
‘
IW ~~
flONUSIUMlT . ~~
I ~~
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
~ ~
FINA L REPORT
PART 1
PART 2
Analysis of Adaptive Weight Noise Covariance
R Ches tek
T Saxe B
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
-
—
- —- .—- -a -----s.
- - - ------
•
~•
‘ —
, — —-— -- —
~~~~~~~ ~~ ~~~~~~~~~~~~~ ~~~~~~~~~~~
~ ‘
u•~ ~
~~~~~
— -
~~~ !
— _iii T -—
~~~~~~~~ ~~~ —
_ _
~ ~~~~- ~~~~~~~~~~~~~~
i
2
~ ~l ~
7 N]
[ ;
PART 2
—_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
-• •— ~- ~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
—
~~~~~
--
~~~~~~
-
-
-
- -
— - - -— --
where the output power due to weight noise is less than 5% of the
output signal power. ~~~~~~~~~~~~~~~~~~
_— -
-
total Input power)
( -
it- •
A new criterion for- n has been found which guarantees jnean.squ~re
stabili ty :
!). — —- ---•-——- —
—. ----- ----•- _ _ _ _ _ ____ ~~~~_ _ ~•~~~~~
,
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
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
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 .
. . -_ _ _ -------- -
— - —-— — — ~~~ ~~~~~
-
- -
—
•
12
~~~~~~~~~~~~~~ ~~~~~~~
—-- --
- --
•- — ---
~~~~~~~~~~~~ ~~~~~~~~~~~~~ ~~~~~~~ ~~ ~~~~~~~ — ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- ----
A TRRINrN ,FILTERS
s’ ~~~~~~~
~~
-
IF2
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
.1
~
#4 - -
-
-
- - •~~~~~~
- -
I
-
- -- - -
TI
- -
~~~~~~~~~~~~~~~~~~~~~~~~~~~
- - - -
_ _ _ _
-
-
-——
- .--
~ ~~ ~~~~~
~~~
- - ----—-
~ —-~ - - - -
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 .
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 .
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 -
~~~ - . -
-, •- - - • • -
4’
~ TF I
_ s1
r
62
‘h F 2 I
I }\
ti ~~~~~~~~~~~~ II1-,~ % -
SY STEM
OVTPUT ~~
_ _ _
I
I
.
‘hTFU
~ 4TFS
SI G ~4F3 L— -
= •i
~~
+
~~~
+ 2p~~IQ WJ )
= (1-2pflW + + 2p)’WQ
~
11
‘4
-
T
— —
— — —
—
‘ -:- -- •
- . •
~~ -
--• . ~~ ~~~~~~~
‘
~~1”
- --T-- TT-~~~~.~~
-
- __ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
—- --- - ----- -
-
- - -- --- —- - - - - ---—- -
-
~~ ~ - -
‘4
-
12
~~~ — — --
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 )
~r Y 1 ~~~
z - - - :-:~
:ii2 - - 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~ -
.
~~~~
---- _— ‘_ —
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
4 ~
.
16
- •
- :- ; - •
_ _ _ _ _ _ _
~~ ~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
I
~~~3-L~~~~ -~
-
._ __i __ ___ _._
~
~~~~~ ~~~~~~~
- _ _ _ _ _ _ _ - - - _ _ _ _ _ _ _
- TIJ
I-I
LU -
H ;
LA
I. <U I — -
k ~~ ow -
<0 .
4 ~~~
-
‘7
~~ ~~
JAMMER
#2
ER #1
-6ODB 0DB +20DB
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
if
4 18
I
--
~~~~
- ----- — -- - - ---- — -
~~~
-i--’-—-
-
JAMMER #2
J MMER #1
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
:
—
-- — -- -
- .
~~~~~~ ~~~
~
20
I ____________
--- --— -- _ _
~~~~~~~~~~~
is the set of all X (i) vec tors “stacked up” to produce one vec tor.
3
Thus
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 .
1— 3
21
~~~~~~~~
- ~~~~~~~~~~~~~ - - -- - -
it by R.
when no signals (desired or o therwise ) are being rec e ived is d eno ted
the quiescent weight vec tor.
tor.
,
if
used by the sys tem to mou if y its pa r ame t ers to improve fu ture per for-
manc e.
1—4
22
J’ ITIi 1 1$.
_____________
-
T -
1T1 IY9Y~~~~ - - -
_ _ _ _ _ _ _ _ _ _ _ _ _ _ - - - -
‘ -
~~~~~~~~~~~~~~~~~~~~~~~~~~~
—
~~~~~~
minimized . -
h~~, only moving away when the output power Ely grows large.
23
- _ •—- -,
4
-
—
-
-.
—
~~~
_ _ _ _ _ ~~~~~~~— --- — — - ~~~~~~. — - -
_ _ _ _
~
~ -
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
1—6
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
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
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
-
*
= (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
— —
*
—ABWIN
-
1—15
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
4 26 -
•
- ~~
-
— —
T.
- --
- —
~~~~~~~~~~ --
‘—-
- —, -- —
— -
-7- -- --
-
~~~~~~~~~~~~~~~~~~ — - ---
—3
W.
—3
- pV J
1—19
1—20
= 2p(R W~~ + Y( )J
~~~~Q
-
•
-
1—21
-4
27
--..
-—
4
-— —- - -.5 -~~~~~~-~ - - --‘
-
~~~
= - 2J1X ~ X~ W
)
- 2PY (W~~~~)
1—22
= (l — 2 Y) ! + 2p)4
~ j+l ~
— -
~ ~~
1—24
This is the adaptation rule used in Scheme 6A .
- -
~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~~
.
.
‘ —-
-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
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
1—28
wher e we have assumed that , with W~ converging as j— oo,
H (R+ Yfl
~
* —J_
= Y (R + y I )
4
29
4 —
~~~~~~~~~
-—— — ———----—--—-- - -- — —7-- —7-——-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~
,- - ‘-- - — -
I
--
- - -: - ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~
- - --—- 7- - ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
— -- -
~~~~~~~ - - —-
~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~ ---—.
_ -.----—— ---_----
-
~~~
1—30
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
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)
wax
1—32
30
--
-
t
_ _ _ _ _ _ _ __
_ _ _ _ _ _ _ _ __
__ _ _ _ _ _ _ _ _ _ _ _ _ H
r - - -
4 31
- -
&- ~~~~
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 .
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
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
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
-
, - 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
or2
5 +
W
*
= I- W
or2
— —
~~~
+ 7-S
1—41
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
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
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
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
36
4
-- - — . - — _• - ~~r r ’ f l a -1IuuI~
.1
-- - —- -
- -
‘ ‘ -~~~~~ ~
-
- -
A
-
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
37
4 7- —7-— p_
_ _
~~~~~~~ ~~~~~~~~~ _,~~~
- s - ’- ‘- . .
-~~ -
- :
.
I~..- - - :
p ,1
l
i
-~ =r~ i
~~~~~~
T~~~~~~ i~~~~~~
~~
_ _ _ _ _ - -_~~~~~~~~~~~~~~~~~~~ - - - —— - --- --~~~~~~~~~~~-- --- -~-—7---—-—--- -7- ---~~~
r
_ _ _ _
PART 2
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 ,
38
.4 7-
- -~~ ~~~~~~~~—~~~~-— ...- — .~~~~~~~~~~~~~~~ — -- --
~~~~~~~~~~~~~~~~~~~~~~~~~~
11
- - — - -~~~~~~~~~~~ — — 7 - — -- - -
_1
— —
~~~
- -- — •— • - --
~~~~ i.L - — ~~~~~~~~~~
—- - --- - 7-
- _ _
~~~~~~~~-~~~~~-.-
- ~~~•-~~~~~~~~~~~~~~~~~~~~ ~~~
. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Figure 2—1 depicts the APS and helps to define the quantities
wea k si gnals
Input d~
~~ /
Figure 2-1
39
p -
_______________ -—- - - -
-—-
~~~~~~~~~~~~~~~~~
-
-
14 ~~~~~~~~~~~~~
- -
~~~~~~~~
-
gonithm .
The opera t ion of the ‘leaky ’ LMS (LLMS) filter is defined as follows :
Ts
y =w
]
=
2—1
=
2—2
W j.f], = V W~ +2Ii€iS~
2— 3
1-2pY
2—4
troduce scveral new quantit i~~ which define th.2 composition of the in
put signal
40
-4 - .-
-
-
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
SNR = ’
2
1
~~ ~~~~~~~~~
2— 5
4
41
4
- - -- - - --
I
_ _ _ _ _ _ _ _ _
—1
-
-
~~ ~~~ ~
ii
-
-
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _______ _
~
—ii
- -
2 SMR ’
I -
42
— - -
7- -
4 T .
.4 7--
4.’
p
7-
‘
-
- - - -
-
- —
-
-
-~~~~-- ~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~
_ ~~~~~~~~~~~ _ f~~~~~~ _
~~~~ -
-7-- --- ---- -----
---~~~~~~~~~~~~~~ ~~ -:=
‘.- - ~~~~~~~~~~~~~~~~~~ ~~ .
::: ~-- ~ -7-’
.
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 .
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
4 43
- - - -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ - — --a . - -~ t -7-7- -
- ~~
-
A . - ‘ - 7-
_ _ _ _ _ _ _
-
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~ _ _ _ _
-
-
-
________________________________
(a—b) cos [9j~
-
2—10
(€ 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
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
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-
.
- ‘
_ _ _ _ _ _ _
~~~~~~~~~ i ~~~~~ - ._ - ---- - —~~~~~~~~~~~ — - - - - ---- ~~~~~~~~-- - -~~-~~~~~~~~ - -~ ~~~ ---~~~~- -
--
~ _ _ z _T --
--- - ----
~
inpu t x.
.3
AVERAGE _.__
p .3 ,~~~~~
I
_ _ _ _ _ _
_
WEIGHTS s.
‘.
.. .,
‘fII tei
,,
good ,,
Y outpul
u pda te terms SPLITTER ~
ERRORS
Figure 2-4
= V W j+2u€ jS]
W (Z)
=
2p
~~~ ( Z ) S ( Z ) Z—( /
2—1 2
- - - _
.
.. .-. - -
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
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
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
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
4
48
t _
li_Ji __ -
~~~~~~ ,*—
. _ _
- -= - -7-- ——- -~ ~~~~~~ —7--—- —— - ——7-- - _-_
~~~~~~
__ -_ --—---_-,_
~~~~~~ . .-—7---~~ ~~~~~~~~~~~~~~~~~~~~~~~~~ - — —
-7- -7-.’
curve but now the total distortion is less than 0.06% of the desired
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.
Figure 2-6
% distortion vs. p
50
4
4 ~~~~~~~~~~~
.
- -
~~~~~~
— -— —
~~~~~~~~~~~~~~~~~~~~~~~~~
. 7- 7-
~
11
- ~~~~~~~ ~~~~
- - - - _ _ _ _ _ _
_‘
--
~~~~~~~~~~
-7- — - -=- -— -
~~ ~~~~~~~~~ ~~~
I_coo —
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 —
:F
—
Numbers enclosed in ()are
power gains at input frequency,
i
~
0.010 -
O~0C8 :
0.006 -
7-
0.004 _
Figure 2-8
% distortion vs. p
s 52
~~~~ ~~~~~~
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ _ _
_ _ _ _ _ _ _
_ _ _
—-— -- -- - - - - - -~~~
10.000_.
Filter l ength 64 weights
v~~~O.9O
-
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
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 ...
-
2.000 —
I•000 :
-
•0
0.600_ .—tO (0.63) -
4-)
- -
0
4-)
0.400 _
~
NOTE: Rate of
convergence is
proportional to
p.
Figure 2—1 0
% distortion vs. p
54
~~~~~~~~ ~~~~~~ ~
-
~~~~~~~~~~~~~~~~~~ 1 L
_
~~~ _ _ _ _
—
—J
— -——----- --— -—
~~~~~~~~~~~~~~~~~~~~~~~~ -
30.000.
-
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 ~~~~~~
-- - -
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~
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
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
vARfL~7w
- ~} ~~~~~~~~~~~~~~~~~
2—4
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~
-
2 2 2 4
=
4~i a u a
L SN R ’ 2
—
) 8
—
2—5
Tb us
(VARL& ~ )
)1 = _ _ _ _ _ _ _ _ _ _ _
2
2 4
—
1
~i a
e i29 — V
-
L SNR 2
2~ 1÷ )
2— 6
considered how this affected the weight vec tor (equation 2—19) and
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 --
- -
~~=~~~ —
—-- -— —-
~~~ ~~~~~ ~~~~~ ~~~~~~~~
~ ~~~~ ~~~~~~~~~~~~~~
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
~ ~~~~ ~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~
Table 2-1
58
- - -
- - -- —~~~~ -
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -
- — ~~~~~~~~~~~~~~~~~
.4
-
- 7 - -- -
- - . - -j -.
~ ~
~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~
-
7-- - - --~~- ~~~~~~~~~~~~~~~~~~~~~~~~~~~ - I —~~~
Note that:
2—7
also
o j =n j
x .=n .
3 3— 1
where
2—8
2— 9
-4
t 59
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
2—1 2
-
-
4p 2 a4
4~io +4~~iY + 4e i Y -8ji } a
‘
=
pa2
7-i
PC.
- -
a
2—1 3
>
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-
T 2
PE{SSTVA R (W } SS }<<a vA R (w )
—
To check this assumption we no t e t ha t :
conclusion :
E{SSTVAR (WISST ) =
6 6 6
2pa ~~ Lp a p (L + 2 )a
1 1
a a a
2— 16
t
_ _ _ _ _ _ _ _ _ _ _ _
a a
4 -
-
61
5 - -
- ~~~--— ~~~ —
5 S
_ _ _ _ _ _ _
~~~~~~~~~~~~~ ~~~~~~
L
~~~~~~~~ I ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-- —- — -
-- 7
.-— —
which implies
2—17
-
mm -
- -
VAR (W ) =
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.
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
Figure 2-12
or
p (error— output—power )
VAR ( W )
1+
(error—output—power)
63
4 -
-~~-- — -•-
~~ . -..-
- -
‘1
— —
~~~~~~~~~~~~~~~~~~
-
~~~~~~~~~~~~~~~~~~~~
~~~~~~~ : ~~~ _ i~~ ~~~
-
-- ~~~~~~~~~~~~ -- ~~~~
7- - - -— ~~~~~ --~~ — S _ _ _ _
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~
—
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
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
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 ~~~~ Cr 2~~~~~~~~ )
= +
~~ /
I l— (, +2p (o- +a
-4
66
~~~~~L
- - - 7- -5 -- —-— --5- -- - — - - -5 -~~~~~ _ - - - - 5--- ~
5 ~~
2—20
2)
u r n E{W } = 2pa 2 2~i (a 2~ — 2pa 4-6èici2a
N->oo N +
2
W
*
=
a
-)
a +0
2—22
The adaptive powe r separator based on the leaky LMS algorithm gen-
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
eral , to a %~iener solution if the inputs are correlated , and since the
a
6/
.4
-
- - 7- - 5- - 5- -
. V-.-
~~~~~~~~~~~~~~~~~~~~~
t - -
i -
~~~~ ~~~~~~~~~~~~~~~~~ - - ~~~~~~~~~~~~~~~~ ~~ ~~~~ -~~ -
- - -
-
— 7- - -
~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~
2 (a +a
2—23
will guarantee that the mean weight value converges. Note also that
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.
4 . VAR {
~~ ,~~1
} E(
~~~41 }
— (E{W N+u }) 2
2— 24
68
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)
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~
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
2 2 2 4
=uo [l_8Mor +4~~ a
l-3jior
2—27
4
2(a +0
Which is the same condition as for convergence of the mean weight
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
100,000
-~
Fi gure 2— 13
vari ..n~~ ot the wei g h t vector). From this we see that there are two
.
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
‘<
~~
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. -—
.:— ;. ~
—
~~~~~
-
~
- - — -- - -- -
— -
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
_ _
_ _ _ _ _
_ _ - -_ _
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-
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
( 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 .
— the LLMS driven APS was per formed . This analysis confirms Senne ’s
Ill ) observation that LMS does not converge to the Wiener solution if
LLMS (and LMS) filters may not stabaliz~ in a m ean—square sense even
73
4_p51 -
— 7 - -- . - —7-—-- — — —- - —
~~~~~~~~~~~~~~~~~
, -..---
~
- - - -
~~
—
-: -- —
~ - -
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
~
- -5-- ~~~~ - - - -— - - - — -5~~— - -- - -- -—- .--
— ~~ — - -——
-4 74 -
p
1
1
_ __ • -. •-5-5-5 -
•• - —~~~~~~~~~~ ---~~~-~~~~~ , -.~~~
~ ~ ~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- - - - —--—- -- -— ~~~~~~~~ --~~~~~ ~~~~~ — - —-- _ -~~~~~~~~~~~ - - 5-- — - -
REFERE NCES
Un i v . , S t a n f o r d C a l i f ., J u n e . 1977.
4 -
- 5 - -- - -— -— - —- —
~~~~~~~~~
-
—— —~ 7- — - .
‘ -~.-
‘_ _— .
.~~~ .
~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~
-- ~~~~~~~~~~~~~~~~~~~~~~~
.7- p
i
— - .
‘
-
—
a- --
- -
- - -
~~~~~~ .
~~~~~~
-- - -‘- - - . -- - -I -
-
- - - 5 -- -— - - - - - — --5-- 5- -5- -5—~~~~~~~~~~~~
~~~~~~~~~ ~~~~~~~~~~~~~
- - . -- 5-5-5-
~ ~
a. .
4 76
i
i
_
7- - 5 ‘5
- --
~~~~~~~
- —
~
- _ _4; I
&~~_
~~
i~. —
~~~~~~~~~~~~~~~~~ _ _ _ _