RESULTS IN DATA ADAPTIVE DETECTION AND I1 …ad a129 868 a summary of results in data adaptive...
Transcript of RESULTS IN DATA ADAPTIVE DETECTION AND I1 …ad a129 868 a summary of results in data adaptive...
AD A129 868 A SUMMARY OF RESULTS IN DATA ADAPTIVE DETECTION AND I1ESTIMATION(U) RHODE ISLAND UNIV KINGSTON DEPT OF
DN AS LECTRICAL ENDINEERING S KAY APR 83 NODS 4-81-K 0144ASFIE / 21NN7DEEEomhomhEEI8
W . 8= 2.
L 13"IIL
11111 1 11 1 1.611111'__23
MICROCOPY RESOLUTION TEST CHART
NATIONAL BUREAU OF SIANDAROS-1963 A
t
00OA Summary of Results(0) in Data Adaptive DetectionC4 and Estimation
Steven Key
'Department of Electrical EngineeringUniversity of Rhode Island
Kingston, Rhode Island 02881
April 1983
SPrepad for
OFFICE OF NAVAL RESEARCH (Code 411 SP)Stadhtk and Pfbaf Dran
r Cast" N0001441 40144 LECTED. W. Tutst, Fdpd boadu LT
JUN 2 9 3
ApprwiW for pulai tr;d bt. o u aMjW i
o 21U
m 88 O0 27 oT'
t R I I C L A V I., A !O N C ) T R IS P A G E ( h o,,~ I). , -F . t .c d )R E D I S U C O NREPORT DOCUMENTATION PAGE READ INSTRUCTIONS_______________ BEFORE COMPLETING FORM
I REPORT NUMBER E2. GOVT A -SION NO 3 RECIPIENT'S CATALOG NUMBER
4 TIF L E (...d S.ubefrt.J S TYPE OF REPORT 6 PERIOD COvEREO
A Summary of Results in Data Adaptive Detectionand Estimation
6 PERFORMING ORG. REPORT NUMBER
7. AUTHORts) 8 CONTRACT OR GRANT NuMBER(a)
Steven Kay N00014-81-0144A
9 PC RIORM1-,' )RGANIZATION NAME AND ADDRESS I0 PP)GRAM ELEMENT. PRO'JECT, TASKAR A 0 WO.rK UNIT N'JM9E'RS
Department of Electrical Engineering
University of Rhode IslandKingston, Rhode Island 02881
I. CONTROLLING OPVICE NAME AND ADDRESS 12 REPORT DATE
Office of Naval Research (Code 411 SP)13, NUMBER OF PAGES
15t4 MONITORING AGENCY NAME & AOORIESS(II different from Controlin4i Office) IS. SECURITY CLASS. (or fl. f.pot.,
UnclassifiedIS.. OECLASSIFICATION, DOWNGRADING
SCHEDULE
16 DISTRIBUTION STATEMENT (of this F port)
Approved for public rell ase; distribution unlimited
17. OISTRIBUTION STATEMENT (of the obstract entered In Block 20. It different from Report)
18 SLUPPLEMENTAN' J'OTES
19 KEYf WORoiaS Cnrtn,, on rvrSl .1,1, If neeeel d Idonitt ' by hlok - sb.,I
20 V STRACT 1Conlin- an eOvees old* If ec- ery atd Identfy by block number)
This report describes a number of results which have been obtained inthe Data Adaptive Detection and Estimation project. The various paperssummarized deal with properties of autoregressive representations asthey relate to detection and estimation in partially known signal and/ornoise environments. Complete copies of the works can be obtained fromthe author.
ORMDD 1473 EDITION 0' 1 NOV 611 IS OVISOLVETIN n I n,' I ', (I Anl I
S... .I.. . . rL-. .r.. AT .. r. Ir Trr. P%-. Flo-, " l. .. ,.,I
| 1 1
SECUITy CLAsSIFICATIoNl or THOS PACE (Wheng be gEeeoe.
A Summary of Results
in Data Adaptive Detection
and Estimation
Steven KayDepartment of Electrical Engineering
University of Rhode Island
Abstract
This report describes a number of results which have been obtained in
the Data Adaptive Detection and Estimation project. The various papers
summarized deal with properties of autoregressive representations as they
relate to detection and estimation in partially known signal and/or noise
environments. Complete copies of the works can be obtained from the author.
Aooession 1o--ins ' -DTIC TABUnannounced 0justification
Distributionl/Availability Codes
.Avail and/or .
Dist Special
SUMMARY
1) "on the Statistics of the Estimated Reflection Coefficients of an
Autoregressive Process" (with John Makhoul of Bolt, Beranek, and Newman)--
submitted to IEEE Trans. on Acoustics, Speech, and Signal Processing.
This paper derives a recursive means of computing the Cramer-Rao (CR)
lower bound for the estimated reflection coefficients of an autoregressive
(AR) process. The exact algorithm which computes the bound is given in
Figure 1. Since the CR bound is attained asymptotically (for large data
records) for a Maximum Likelihood Estimate of the reflection coefficients,
one can also use the algorithm of Figure I to describe the statistics of
the MLE. It is also shown in this paper that all currently available
estimators for the reflection coefficients are MLE's for large data recc.ds.
Thus, the results can be used to statistically characterize these estima-
tors, i.e., the Burg algorithm I as an example. An example of the actual
covariance matrix (obtained via computer simulation) and the CR bound is
shown in Figure 2. The data length was 1000 points.
2) "Simple Proofs of the Minimum Phase Property of the Prediction Error
Filter" (with Louis Pakula of U.R.I.)--IEEE Trans. on Acoustics, Speech, and
Signal Processing, April 1983.
Presented are two simple proofs which assert the well known property
that all the zeros of the optimal prediction error filter lie on or within
the unit circle of the Z plane. The first proof is somewhat incomplete in
that it only shows that the zeros lie on or within the unit circle. It
gives no indication as to when the zeros must be totally within the unit
circle. However, it is extremely simple. The second proof which is slightly
more complex also determines conditions under which the zeros may be on the
ultt circle.
3) "Recursive Maximum Likelihood Estimation of Autoregressive Processes"--
IEEE Trans. on Acoustics, Speech, and Signal Processing, February 1982.
An estimation algorithm is derived which is a closer approximation to
the true MLE of the parameters of an AR process than currently existing
procedures. Specifically, the algorithm does not make the standard assump-
tion that the determinant of the filter covariance matrix can be neglected
in the maximization of the likelihood function. This common assumption is
not valid for short data records and/or highly peaked spectra. By incorpo-
rating the determinant and maximizing the likelihood function recursively
(in model order) more accurate parameter and spectral estimates are obtained.
The algorithm is summarized in Figure 3. Also, Figure 4 summarizes some
simulation results which verify the improved performance of the algorithm
for short data records. (RMLE - Recursive MLE, FB - Forward-Backward,
YW = Yule-Walker).
4) "More Accurate Autoregressive Parameter and Spectral Estimates for
Short Data Records"--IEEE Workshop on Spectral Estimation, Hamilton, Ontario,
August 1981.
This paper is a shortened version of 3.
5) "Some Results in Linear Interpolation Theory"--to be published in IEEE
Trans. on Acoustics, Speech, and Signal Processing.
The optimal finite length interpolation filter for a wide sense stationary
process is derived. It is shown that for an AR process one can significantly
reduce the power out of an FIR filter by performing interpolation rather than
the more common one sided prediction. Also, the optimal interpolation filter
for an AR process, i.e., the one which minimizes the interpolation error
power, is found to be the one which attempts to interpolate at the midpoint
of a data record. This result although intuitive is shown not to hold for
2
more general processes. Finally, some well known results in infinite
length FIR filter interpolation are simply derived using AR approximations
of infinite order.
6) "Accurate Frequency Estimation at Low Signal-to-Noise Ratios"--sub-
mitted to IEEE Trans. on Acoustics, Speech, and Signal Processing.
An iterative algorithm for frequency estimation of sinusoids in white
noise is described and analyzed. The algorithm iteratively computes AR
parameter estimates after the original data has been filtered by an all
pole filter. The algorithm is shown to be related to the Steiglitz-McBride
algorithm 2 . It is summarized in Figure 5, where it has been termed the
Iterative Filtering Algorithm (IFA). For two sinusoids closely spaced in
frequency the algorithm performance is shown in Figure 6. Also, shown is
the CR bound (for unbiased estimators) and the principal component (PC)
3approach of Tufts and Kumaresan . The IFA outperforms the PC approach
at low SNR's. The PC approach had been claimed to offer the best per-
formance other than a direct MLE, at low SNR's. At higher SNR's a bias
is present which tends to degrade the performance of the IFA. In Figure 7
a comparison is made with a direct MLE, which amounts to a nonlinear
least squares estimator. Again at low SNR the IFA appears to perform
better.
7) "Asymptotically Optimal Detection in Unknown Colored Noise via
Autoregressive Modelin "--to be published in IEEE Trans. on Acoustics,
Speech, and Signal Processing.
The problem of detecting a known signal in Gaussian noise of unknown
covariance is addressed. The noise is assumed to be an AR process of known
order but unknown coefficients. Thus, the parameterization of the covari-
ance leads to a problem in composite hypothesis testing. Since in general
no uniformly most powerful test exists, the Generalized Likelihood Ratio
3
Test (GLRT) is applied. The GLRT detector which results is shown in Figure
8. As expected adaptive prewhiteners are included which assume either the
signal is present (upper channel) or absent (lower channel). The detector
is not a single estimated prewhitener and matched filter as might be ex-
pected. The performance of the detector is shown in Figures 9 and 10 for
an AR process of order p-I and parameter a--0.9 and for various data record
lengths N-20,100. The optimal curve is the performance assuming a known
covariance function, i.e., a prewhitener and matched filter. It is seen
that even for short data records the performance is nearly optimal. Finally,
it is proven that for large data records the GLRT performance is equal to that
of a prewhitener (assuming a known covariance) and matched filter and hence
is optimal.
8) "Detection for Active Sonars via Autoregressive Modeling,"--Workshop
on Maximum Entropy and Bayesian Methods in Applied Statistics, University
of Wyoming, Laramie, Wyoming, August 1982.
This paper is a shortened version of 7.
4
1) For n=1a. = a(l) - K1
2 1
22
- " (1-K ) I
2) n - n) a2 (1K2 ) 2 )
N 1 2)i n, n+1 2 2 2n D_ _ ~_-n-l i-2 A (I-K J) 1
1-K
C' n) -g I [c,,2 D -1 T 1LT 1_K2
n
3) If n-p, let = C(n) and exit.
4) 0~~£I n--n 2J
r
(1 [(Kn-11) an.2]An-i-
K n-1
Bn-i l n
B =
0 1n Po T n
S) Go to 2
Note that 0T is lx(n-l).
Figure 1. Summary of Cramer-Rao Bound Computationfor Reflection Coefficients.
5
TABLE 2
TRUE VALUES ESTIMATED VALUES
K K2 K3 0.570 0.570 0.570 0.569 0.568 0.564
0.541 - 0.490 0.000 0.523 - 0.479 - 0.006
NCA - 0.490 0.675 0.000 - 0.479 0.689 0.028K
0.000 0.000 0.675 - 0.006 0.028 0.665
TRUE VALUES ESTIMATED VALUES
K1 K2 K 3 0.570 -0.570 0.570 0.557 -0.563 0.563
7.210 -1.790 0.000 6.988 -1.858 -0.151
NCA -1.790 0.675 0.000 -1.858 0.716 0.016
0.000 0.000 0.675 - 0.151 0.016 0.692
Figure 2. True and estimated values of the reflectioncoefficients and their covariances for two AR(3)processes. The two processes differ by the signof K2.
6
60 -Soo
fornl:C,$=So, d, =S,
to find KI solve
N-2C, E.0 +Nd N C,
N- I d, (N-I)dl N- I d,
and choose toot within [- !1.'(' )
. g l
I, -So0 + 2KISO, + KISI,
2 ( l) 6 1
Fornm2,3,- .p:
I Sn-2 Pn-tJ
= b fT Sn- n-l bT
to find Kn solve (for d. > 0)*
K -2n K n.n-, + Ndn N C,,
N-n d, (N-n)d, N-n d,
and choose root within 1- i, I1.
a().fa(n-i .n n-i)
K() .' +n, n- , i = n,2(nI
rbn l - :.I+Cnd I
i= an-i + 2CnKn + dngKn
02(n) =I IV 1
where(n-i) (n-i) (,-I) T
a.Il-I a3 a "a,,-(.-i) (n-I) ("-i) 7bn-l [an-, a.- '. . at
N-i-l I = 0, 1.
{SMn1=S#= / x(.ir1
rTSOO qn-I SO"I X lt n- I I Ix
Sn qT-1 S$-, P'I-I\t I) I i -- ) x (of - II I
,p0, iN,,
"or do -. 0 see Sclin III for dis cusin .t 'roots.
Figure 3. SumMary of Recursive ILE Algorithm.
7
TABI.E ICOMPARIS,s oF AR PARAMFTIR FSPN.,ITIw.N MITI('
N I I Data Point%
a 2 a 1 04
True AR Parameters -2.76070 3.81060 -2.65350 0 92388;Mean-YW -1.01527 061949 0.04466 0 05735Mean-FB -2.5S333 3.42185 2.273o3 0.77262Mean- RMLE -2.6^975 351086 -2.37156 '.81412Variance - YW 3.13780 10.29o27 7,33884 0.7584tVariance 11B 0.11598 0.52574 0 4930 1 ( 08462Variance RMLE 0.10960 033371 0.26227 0,04148
N = 50 Data Points
a, a'2 0"1 d4
True AR Parameters 2.76271 3.8 1 60 2.65 Is() 0I7 2,
Mean YW 1.30552 0.92378 0.0515 (1, 7 19Mean -FB 2.7236H 3.71024 -2.54984 11.87_6Mean -RMLI-. 2.71515 3.69876 -2.54345 h.h?,765Variance-YW 2.22544 8.56804 6.90835 ().745:6)Variance- 1.B 0.00758 0.14138 0.04219 0.I ,,s 4Variance RMLI* 0.01076 0.0448. 0.04069 7.:)4 1
Figure 4. Comparison of AR Parameter Lstimation Methods
" .. ... . II I II II-- . .. . .. .
'0
boC:
4-3
cj'%
/0 LoG C-1. ~iA~ 7/
60-
S-01CRAPS-AA0
-30 2-LZA~A4
/0
0 /0 Zo
Figure 6. Comparison of Frequency Estimators,S77~t .r. 0
10
/0e 001
SA/
Figue 7.Comarisn /o Frequency j Estim ars, 0
0 w
Mi i0 0c C,A Vc-
+
z
ac
zl Zi IL
0
LUI
N zj< 0)
<0. I-I-- +4
wd 4)
o
W (A
zL
- -12
z00
0mw
Ho Nz z 44U, 0
00
:.
000w- w
L N40d
00
0 .. . ......
C;l
loll
00
U-4
LLI w
w In
~>1
0ow
144
REFERENCES
1) J. Burg, "Maximum Entropy Spectral Analysis," Ph.D. Dissertation,Stanford University, Stanford, CA, 1975.
2) K. Steiglitz, L.E. McBride, "A Technique for Identification of
Linear Systems," IEEE Trans. on Aut. Control, Vol. AC-10, pp. 461-
464, October 1965.
3) D. Tufts, R. Kumaresan, "Estimation of Frequencies of Multiple
Sinusoids: Making Linear Prediction Perform Like Maximum Likeli-
hood," IEEE Proc., Vol. 70, pp. 975-989, September 1982.
1
15!
OFFICE OF NAVAL RESEARCHSTATISTICS AND PROBABILITY PROGRAM
BASIC DISTRIBUTION LISTFOR
UNCLASSIFIED TECHNICAL REPORTS
FEBRUARY 1982
Copies Copies
Statistics and Probability Navy LibraryProgram (Code 411(SP)) National Space Technology Laboratory
Office of Naval Research Attn: Navy LibrarianArlington, VA 22217 3 Bay St. Louis, MS 39522 1
Defense Technical Information U. S. Army Research OfficeCenter P.O. Box 12211
Cameron Station Attn: Dr. J. ChandraAlexandria, VA 22314 12 Research Triangle Park, NC
27706Commanding OfficerOffice of Naval Research DirectorEastern/Central Regional Office National Security AgencyAttn: Director for Science Attn: R51, Dr. MaarBarnes Bijilding Fort Meade, MD 20755495 Summer StreetBoston, MA 02210 1 ATAA-SL, Library
U.S. Army TRADOC SystemsCommanding Officer Analysis ActivityOffice of Naval Research Department of the ArmyWestern Regional Office White Sands Missile Range, NMAttn- Dr. Richard Lau 88002lOC East Green StreetPasadena, CA 91101 1 ARI Field Unit-USAREUR
Attn: LibraryU. S. ONP Liaison Office - Far East c/o ODCSPERAttn: Scientific Director HQ USAEREUR & 7th Army' C.r. r -ico 9U03 1 .M w v', qArN
Applied "athematics Laboratory Library, Code 1424David Taylor Naval Ship Research Naval Postgraduate Schoolan; Development Center Monterey, CA 93940
Attn: Mr. G. H. GleissnerBethesda, Maryland 20084 1 Technical Information Division
Naval Research LaboratoryCommandant of the Marine Coprs Washington, DC 20375
(Code AX)Attn: Dr. A. L. Slafkosky OASD (I&L), Pentagon
Scientific Advisor Attn: Mr. Charles S. Smith
Washington, DC 20380 Washington, DC 20301
Copies Copies
Director Reliability Analysis Center (RAC)AMSAA RADC/RBRACAttn: DRXSY-MP, H. Cohen Attn: I. L. KrulacAberdeen Proving Ground, MD 1 Data Coordinator/
21005 Government ProgramsGriffiss AFB, New York 13441
Dr. Gerhard HeicheNaval Air Systems Command Technical Library
(NAIR 03) Naval Ordnance StationJefferson Plaza No. 1 Indian Head, MD 20640Arlington, VA 20360
LibraryDr. Barbara Bailar Naval Ocean Systems CenterAssociate Director, Statistical San Diego, CA 92152Standards
Bureau of Census Technical LibraryWashington, DC 20233 1 Bureau of Naval Personnel
Department of the NavyLeon Slavin Washington, DC 20370Naval Sea Systems Command
(NSEA 05H) Mr. Dan LeonardCrystal Mall #4, Rm. 129 Code 8105Washington, DC 20036 1 Naval Ocean Systems Center
San Diego, CA 92152B. E. ClarkRR #2, Box 647-B Dr. Alan F. PettyGraham, NC 27253 1 Code 7930
Naval Research LaboratoryNaval Underwater Systems Center Washington, DC 20375Attn: Dr. Derrill J. Bordelon
Code 601 Dr. M. J. FischerNewport, Rhode Island 02840 1 Defense Communications Agency
Defense Communications EngineeringNaval Coastal Systems Center CenterCode 741 1860 Wiehle AvenueAttn: Mr. C. M. Bennett leston, VA 22090Panama City, FL 32401 1
Mr. Jim Gates
Naval Electronic Systems Command Code 9211'N E L E ,,? ,r ', l e t Y a t er i a l -u o o r l- # c
Attn: J nn .. 1nu 'ste, ,
National Center No. 1 Mecnanicsburg. A 17.t55Arlington, VA 20360 1 Mr. 'led Tupoer
Defense Logistics Studies Code M-311C
Information Exchange Military Sealift CommandArmy Logistics Management Center Department of the NavvAttn: Mr. J. Dowling Washington, DC 20390
Fort Lee, VA 23801 1
Copies Copies
Mr. F. R. Del PrioriCodie 224Operational Test and EvaluationForce (OPTEVFOR)
Norfolk, VA 23511 1
a!
IAE