You are on page 1of 6
no Generalized Pencil-of-Function Method for Extracting Poles of an EM System from Its Transient Response YINGBO HUA, steer, 1x2, avo TAPAN K. SARKAR, Sesion MEMBER, IEE Abuiact—A generled pencbet-tunction (GPOF) method for ex teatng the poles of at EM system from ii tramlen tesponse ‘eveoped. The GPOF method needs the slaion of & genealied ‘genvalue problem 0 find the poles. This iin ‘comtas to the ad ing the ro 18 alo wsed to optimize the performance of GPOF method has advantages over the Prony method in both comput 1. Iermapucrion TRIS ENOWN 11-18 that intrgt ideation exacting various features of the target from its transient response is desired. The target poles that contain the information of the ‘decay factors and the resonant frequencies should be estimated with high accuracy while the residues can be computed by solving a linear least squares problem [1] after the poles are obtained. The Prony method {1}-(3] has been a very popular technique for pole retrieval. There are also many versions of the Prony method, which include the least square (LS) Prony ‘method, the total least square (TLS) Prony method, and the SVD Prony method. An alternative method is the pencil-of- function (POF) method [4], [9]. Very recently, the idea ofthe POF method has been explored along with ESPRIT {10}, and this has resulted in improved and generalized versions (11]- {14}. This paper, which isa result of this exploration, presents ‘2 generalized pencil-of-function (GPOF) method. The GPOP method finds poles by solving a generalized cigenvalue problem instead of the conventional two-step process where the first step involves the solution of a matrix equation, and the second step entails friding the roots of a polynomial, as is required by the Prony method. We develop the GPOF method and discuss it8 computational aspects in Section Il. The noise sensitivity of the GPOF method is addressed in Section Il. In Section TV, an application ofthe GPOF method toa thin-wire target is presented, Manuscript recived Jone 19,1987; revised May 10,1988, This work was supported inp by Acta Corporation an npr bythe Oe of Naval Research under Comte NOE D.C 0598, “The ahrs re wit the Deparment of Elsa nd Computer Engine: ing, Syracuse University. Syacoe, NY 120 138. TEEE Log Number 882117 1. Graerauizep Pexeit-oF-Funcrion Mento Its known that an EM transient signal can be described by = Y doexp (ik) w where = 0, 1, +++, — 1, byare the complex residues, 5 are the complex poles, and &¢ isthe sampling interval. For notational brevity, we can let z; = exp (551) which are the poles in the Z-plane, It is clear that b; and s, should, respectively, bein complex conjugate pairs for real valued yy Following the idea of the penci-of-function method, we consider the following set of “information” [9] vectors YouYn rs Ye where @ ‘The superscript T denotes transpose of a matrix. Based on these vectors, we define the matrices Y; and Y; as Y= Do Sa Hexen dl Yeelyo yn Ys yell ® @ ‘To look into the underlying structure of the two matrices, one Yin Yor Yel ‘Z,BZ; o Y= Z.Be 2: © where ea nl 8 tw 5 geen geen 0 gee [ a | Ze - ® Vow oe abet len ae “52a ° Ba diag (by, by, -7-, Dual. (10) ‘Based on the above decomposition of ¥; and Y, one can show (0018-926X/89/0200-0229801.00 © 1989 IEEE 230 that if M = L = N= M the poles (2), = 1 ==>, M} are the generalized eigenvalues [10},(17] of the matrix pencil Ys <2, Namely, it Ms Ls NM, 2 = zisarank reducing number of Yz ~ 2¥;.10L = M, ths metho is the Same asthe basic POF method as in [12] and equivalent to a ‘version of Ibeahim time-domain (FTD) metho [16]. However, Inthe GPOF method, we are more interexed in different values of L for M < L'< N-~ M, The significance ofthis vil be shown later To develop and illustrate the use of an algorithm for computing the generalized eigenvalues of the matrix pencil problem we can write Vy Yea Zz BZ i Z,BLZs #23 22s ay where the superscript + denotes the (Moore-Penrose) pseudo-inverse [17], whereas we use ““~1” for the (regular) inverse. It can be seen from (11) that there exist vectors {ps 4 Ayers, MY such that yy Yip co) and ¥; Yon= 20 (3) ‘The p; are called the generalized eigenvectors of Ys ~ 2Yy To compute the pseudo-inverse Y;, one can use the singular value decomposition (SVD) [17] of Yas follows: N= D oui upv" as Yy=v-1u" as) where U = (ih, *s th V = Ui, °*°, val and D = diag Tony =>, oul. The superscript #7 denotes” the conjugate transpose of a matrix. U and ¥ are matrices of left and right singular vectors, respectively. Note that for noisy data yx one should choose o;, +++, oy to be the M largest singular values fof ¥;, and the resulting ¥; is called the truncated pseudo- inverse of Y,. Since Y/Y; = VV" and HV = 1, substituting (15) into (13) and left multiplying (13) by V# yields (Zam, ro) where / M, and Z=D“U"YY, an and an VHB. as) Note that Z is an M x M matrix, and z and 2; are, respectively, eigenvalues and eigenvectors of Z. Now we have completed the description of an algorithm of the GPOF method. vis important to mention thatthe number of poles, M, can. be estimated from the singular values, 0; = 02 = -'- = oy [EE TRANSACTIONS ON ANTENNAS AND FROPAGATION, VOL 37, NO. 2, FEBRUARY 1989 Boe B ommente SiMee Oe = 8 = Omen L ty = O for noiseless data If = M, the SVD of the ¥; is not required, and (23 7 = 1 +5, MY are the eigenvalues of the Mx. M matrix CYEY)-"YHY: which is obtained by substituting Yj" (/HY)-1 YH imo (13) for L = M. Furthermore, one can verify that with or without noise, OO -ew onvoivine|! 8 7a) aay 1 6 wich isthe companion matrix ofthe polynomial 14S oe 0) where [*] CV en ‘which is the solution ofthe least square Prony method. So for L. = M the GPOF method is equivalent to the LS Prony method. “The total least square Prony method is to compute the polynomial coefficients as follows: © 2) where ¥ = [900 Yue 75 Yul and e = Lens *2%5 615 G) A perturbation analysis [12] has shown that the LS Prony method fand the TLS Prony method ate equivalent to the first-order perturbation approximation, ‘The SVD Prony method (for L > M)is to compute ¢, lens eal” by SO Vy 1g min {c""Y "Yes fle] =1) « Yo ee and detect the M signal roots, ic. (2! 0= Te = MD. from the L roots of the L-degree polynomial 1+ Zins -i2"!. But the detection is guaranteed (without noise) to be successful only when all z;(z, 1) are inside (outside) or on the ‘nit circle [21]. Computationally, solving for the roots of an L-degree polynomial is also a disadvantage ofthe SVD Prony method, compared to solving M eigenvalues of an M x M ‘matrix forthe GPOF method. IL, Noise Sexsmivry oF GPOF Merion In this section, we illustrate the noise sensitivity of the GPOF method through some numerical examples. Specifi- cally, we let = Z Aisin wk +8) exp (ak) 24) where k = 0,1, ° Ay = Ay = 1, = 02, Wy = 0.354, 6 = by (0.02m, and as ~ 0.035, Note that 2, = exp (5.61) (ay + Jw) fori = Vand 2, and z= zi for’ “The superscript asterisk denotes the complex conjugation. Its 231 Fig levered petubatioa varancs in dBfthe estimated w, vrs the Fig. 3. Simultion ret (one by + ofthe inverted sample vance eel praneter Invent pertbaton variances in Bf the estimated cy Yr the rnc prmetr Fig 2. also important to note that a, and ware, respectively, damping factors and resonant frequencies normalized by the sampling frequency f, = 1/6¢, (There is no need to specify the sampling frequency for our numerical simulations. The sequence as in (24) isthe only sampled data sequence used in this section for illustration purposes.) The first-order perturbation analysis of For the itis assumed that the additive noise in ye is white and sufficiently weak so thatthe first-order approximation ean be carried out through the derivation. Fig. 1 shows the inverted perturbation variance in dB of w (imaginary part of the pole) versus the pencil parameter L. The Cramer-Rao bound Drovides the “absolute” best result that any technique can achieve in the present “‘noisy"* environment. The GPOF method approximately reached the Cramer-Rao bound. This implies that “‘no” other theoretical technique can do any better! Fig. 2 shows the same thing for «x (real part of the pole). The plots for w» and a are similar to the above two figures, and hence are omitted from this paper. As one observes, the optimal choice of L is around L = N/2. Intuitively, this phenomenon can be explained as follows, ‘The noiseless Y; (or Y;) has @ column subspace of dimension M, (Note thatthe GPOF method requires L. = M and N ~ L = M.) This subspace is called the signal subspace, denoted by $5. But the noisy ¥; (or Ya), which consists of tn of w, eine bythe GPOF method nd te Prony mah. The Ties re obi rom he perttaton ala Fig. 4. Simulation rests (dened by +) of be evened sample variance {nd of a extn bythe GPOF method andthe LS Prony muha. The Sup ies are obtained rom he pertain al signal and noise, spans a column subspace, denoted by Ss. of dimeasion equal min (N ~ LL), 1s cleat that we can wte Sy. = Ss + Sq, where Sy contains all signl component and the nose components projected ont the signal subspace, and Sy conains noise only. Asi seen inthe GPOF method, we perform subspace decomposition of Ya in (14) and throw away the noise component in Sy which is spanned by u fori > M. Iseems, therefore, that the larger the noise subspace 5, the more nose can be filtered out by the GPOF method. The dimension of Syis the largest when N ~ L = L, ite, L = N/2. This is consistent with our perturbation analysis. Note that around L. = 1N/2 the performance ofthe GPOF method is very close to the optimal bound, ic, the CCramer-Rao bound [14], {19}. A different incrpretation of similar phenomenon with the ITD method was made in (16 With the choice L = N/2 = 13, some simulation results forthe GPOF method are shown in Figs, 3and 4, The (Monte Carlo simulation was conducted with 200 runs. During each ru, we computed the estimated a; and, from the data contaminated by (pseudo) white Gaussian noise. (Note that a =Re {log )} and w; = Im {log (2]}.) The noise used in cach rum is independent ofthat used in thes. Figs. 3 and 4 show the inverted sample variances (denoted by the pis signs) in ofthe estimated w, and c, versus SNR which s defined Fig. 5. Core reipnse ate cone of the 2-m dpe uniauted by 2 ‘oot ue by SNR |yel*/No? es) and SNR (4B) =10 log (SNR. 25) «isthe noise variance. The straight lines ae obtained from the perturbation analysis. As one observes, for high SNR, the simulation results agree with the analytical results. As @ reference, the simulation results of the LS Prony method are also shown in the plots. The detailed perturbation analysis of the Prony method is available in {14}, [15]. The noisy data used forthe LS Prony method were the same as those used for the GPOF method, We should mention that the SVD Prony method as represented by (21) performs better than the LS Prony method and the TLS Prony method (21). In fact, the SVD Prony method performs almost as well as the GPOF ‘method for this particular example, However, itcan be shown, {in general, [12}-[14 that the GPOF method is less sensitive 10 ‘noise than the SVD Prony method. IV. AN Armucarion Consider that « 2-m dipole antenna of radius 0.001m is illuminated by a short EM pulse of the form Es = nin'c0) exp (—(¢—f)°/0"] from broadside where » = 377 ohms, ¢ = 3.x 10% mis, 0 = 0.5/c, and fp = 66. The current response at the centerpoint of the dipole was computed by the approach in (20) with sampling time 8¢ = 0.5 ns, and is shown in Fig. 5. To get the intrinsic poles ofthe dipole itself, we considered a segment of the current for ¢ = 8.4 through 25.5 lightmeters, which ‘consists of 114 samples. Note that for ¢ > 8.4 lightmeters the EM pulse Eis almost zero, Fig. 6 shows the fast Fourier transform (FFT) amplitude spectrum ofthe current. From the spectrum, four resonant components are detected at 0.0684, (0.203, 0.391, and 0.414 GHz. The spectrum was computed with resolution equal to f,/1024 = 0.00195 GHz. Applying the GPOF method with L = N/2 = S7 tothe 114 sampled data, we observed thatthe ten largest singular values of the data matrix Y; are 9.3, 7.7, 0.45, 0.42, 0.057, 0.056, Ws ~| tie st i AN Fig. 6. FFT amplitude specram ofthe curren 0.039, 0.0388, 0.0052, and 0.0049. With Mf = 8 (since there isa large drop from the eighth singular vale to the ninth), the GPOF method yielded the following poles (i... 61 = log [zl = a + J: 0.02042 0.218 -0.0266+ 0.642 =0,0046-4/1.17 0.00394 /1.27. By least squares fiting [1], the corresponding residues {absolute values) were computed t0 be 0.3835 x 10° 0.2486 x 10-1 0.9379% 10-4 0.7506 10° From the first six stable poles, three estimated resonant frequencies (J; = W)/281) are obtained to be 0.0694, 0.204, and 0.372 GHz, The unstable estimated poles with the frequency 0.404 GHz appear to correspond to the “fourth” resonant component as is shown in the FFT spectrum. “Applying the LS Prony method to the same data samples and ‘withthe assumption M = 8, we found the following poles: = 0.0203 + 0.218 ~0.0261 +/0.645 03x 10-/1.251 0.611 4/2.244. It is seen that the first two pairs of poles are close to the corresponding pairs obtained by the GPOF method while the next two pairs differ from those obtained by the GPOF ‘method, In the following table, the resonant frequencies estimated by different approaches are compared for the identical data records Freueney (OH) hook hh By fi ~ ei ~ Wa oor 025 ons By GroF owe 020 082 40s by Fer Bows 02m O31 aie BY LS Prony Boos 0203 OH OIE Lis the length of the dipole and c isthe light velocity. It is ‘lear that for the same data, the GPOF method provides 3 stable solution, V. Concuusios We have presented the GPOF method which solves a generalized cigenvalue problem to estimate poles of EM system. In this new approach, one finds the poles directly by solving a generalized eigenvalue problem. This isin contrast to the contemporary methods like Prony and pencil-of- function methods, which are generally two step processes. In the frst step one solves a matrix equation followed by finding the roots of a polynomial. Compared to the SVD Prony method, the GPOF method i less restricted because it does not require that al system poles must be either stable or nonstable, and computationally more efficient hecause it does not solve ‘an Lth-degree polynomial. Compared to the LS or TLS Prony ‘method, the GPOF method is more robust to noise as has been shown. An application of the GPOF method to a synthesized dipole has been shown, with comparison to the FFT and the LS Prony method. Arresoos In this section, we give an outline of the first order perturbation analysis of the GPOF method. For detailed discussion, see [14]. First, itcan be verified thatthe eigenvalues of Z are same as the nonzero eigenvalues of YY, whether or not the data are ‘noisy. Then, itis known [18] that the first-order variation in the poles is given by bax as(Y} Yp./a"'p, @ ‘where 8 isthe first-order differential operator and py and 4 ae, respectively, the right and the left eigenvectors of the noiseless YY. So py (see (11) is the ith column of 2+, and is the ith column of 24. Therefore, a", Secondly, it can be shown [13]-[14] that 8) Note that ¥ is truncated pseudoinverse of Y; in the noisy data case. From (27) and (28), one can obtain the following after some algebraic manipulations: APBY) im ~a"¥; O11) v. ee (oye HBY 28%), @ wre isthe fh ow of 2; and m= [ : il 0) [ ] ey Finally, it ean be shown from (29)-(31) that if {mis / = 0, +N ~ 1} are white with variance equal 10 0, then the first-order perturbation variances are E(\be*} =107/16171 YD Ce eCuee el? @ E(\seul2 © Re MC 4-8C ye. bz} GD FY bm? (Ce 2S ye z} (34) (lb 233, where Diam — 2sk=min(L, NL) DZ Sism LSN-L and . L+1sksN-L+1 Ce G5) " N-LsL and NoL+isksL+l max (L, N-L)+2=ksN Sin Dim Cc) in which ris the mth element of the row vector #f and Py is the mth element ofthe column vector p; Renenences 10} ML, an Barca and R. Mins, “A eigen he Pols al resides of «tem dc from hansen IEEE Troms. Antennas Propagat v0. AP-23, nop. PPT, Non is 12) AV} Poggi eat, Evan of peeing ech forascat hig” IEEE Trams. Arenas Propogt, wl. AP, 0. I. PP 5682178 Jan. 178 (0) MCL en Bacon and R. Mita, “Problem nd olin sociated ‘ih Pon's meio Tor poring Want aa" MEEE. Trans “nternas Propagat ‘AP... te 171, fan. 1978 a) TK Sar cf al, "Scb-opinal appinatonldnttcaton of “ansiat wanetors from electromag sate pene a con Itbod.”" IEEE Trans. Antennas Propapt. vl. APB, M0. 6: Stns, Nov. 190, 13) Fo Teng and TK Sitar, “Experient determin of ronant fgets by wait catering om conurng sper aad ‘Sins, ZEEE Trans Antenna Propogt va APS e0 9. Sutais, So. 10 (6) "beconolton ofthe impulse espn of conducting sphere by te conte gralem metho,” IEEE Tron Amemnas Prope {at vol APS, 1p 105-10, Jn OT (71 E:A Kenaugh and DE® ora, “Tansent and impulse sponse Iprinaoes* Pro 1EBE- vol 53: p.38-901 Aug. 196. (8) UL Moti aa K Mais." Dean dissin ear tarsi" IEEE Trans. Antennas Propagat AP, 90 9. 358-367, May 18 (9) V.K tain TK Sarkar, and DD. Wainer, “Raion! modsting by renciattontion meta TEEE Tron cons Spee, Sal ‘Procesing,vOlASSPI, m0 3 pp 34-57, Jun 88 AC Pla Roy and T. Kali Esato of nal practers ‘atti inarance hie ESPRIT Proc 90 Alomar {Con Cie St Compat Aso. CA. Nov. 18 ‘Act. Mackay and'A. MeConen, "An inprved pre ion ‘etd and cnptsos wih adolf Cuan TELE Tran. Antennas Prope. a APS, 90.4 9. 5 a, fon ‘Pita an, Saar, “Funerals hee mates techies {or pole rev rom ds sere" Pro. 70m Midwest Sym, (Grits Sr. yeas NY. Aap 1987. ali pene atid tnd We perfomance, TEASSP.8H New York, RY. pr 158 {tan "On technics fo iain arms of exponent io. 1} us in Proe us anpediundanped sini in nose PhD. scrtom, Syracuse Un, 198 ‘YM and 7. K. Sata, “Perwtation ays of TK method for harmonic real problem," IEEE Trane Acoust. Speoh Signal Processing, vol. ASSP 6, no. 2. pp. 28-240. Feb 1988 STR Theat and RS. Papa, age mel sure eng ang he abi me drain ieifeation xu". Spacer. vl 1, fr 8. 430-465 Sept-Oct 198 GH Gow and C.F Van Lowe, Mair Compudations Bahar, MD: Jos Hopkins Uni 1983, ue. un 234 U8) 4. H. Wikinon, The Algebroie Eigenvalue Problem. Otic, UKs Chendoo, 98s. Lup] ©. R. Ran Lior Stott! Inference and fs Applications, 2n8 (is New York, Wie 1973 (20) SM fan TK. Sarkar, and, A Dien, “The aplicin ofthe cohen ain med toe slo of wast cletomapaeie akehe Hom tin wes.” Rado Se wl. 19.80. 3, pp 119 1ae, Sep-Oct. 1984, Kumars, simating he parame of exponentially dumped or Ainmped sisi sige aoe,” PRD” dsseraton, Un Rhode ibe 1982 en ‘Vingbo Hea (S'86-M'5) was tor in Siang, Ghnn, on November 26, 100, He receved he SBS" epee in ctl agineting hom Nang Irate of Techoolony, Nang Jags Cin TORz, ad he MS, and PhD. epross n lecea ngcring tom Syacre Univers. Syacse NY, in 198 and 198, respectively Tie wasa Graduate Teatng Ast om 1984 4085 Grunt Fellow ro 1988 1986 and 1 Grate Research Assistant frm 1986101988, {ata Syracee Unser, He waranIacter of franc course at Syease University in the summer of 1988 Sie Srpembes (98, he hs been a Research Asoite t Syracuse Universi Hatch ces na spo pl ran a ‘upan K, Sarkar ('9-M'76-SM°1) wasbor in Cea, indi, on August 2, 198, He eee he Bh Tech depes fom the nda Tso of Te tology, Khargpur nd 196, the MLSE, ‘rere fom he Unive” of New: Bras Frderon, Canada i197 and he M.S. nd PhD, degies rom Syracuse Univers, Syricte, DY, in 5, om 1075-1976 e was wih be TACO division ‘o€ Gena struments Corporation. He wa i) the Recher Inte of Technology. Roches, NY. from 1976-1985. He wr a Resear Fellow av the Gordon Mek) ‘Shorty, Harvard iver, Cambie. MA, om 19771978, He tow 2 Profesor the Deprmeat of Eta Compatr Engieeting, Syracuse Unversity Syracuse, NY. Hs current escarch interes del wh ‘sincncal sain of operator eqotons aleg In eleromagretes and ‘gma! processing ih aplication fo rye desig. He obtained oe ofthe ‘et slain saris May 1977 the Rome At Developmen Cet (ADC) Spal Eaumation Workshop. He hat sated and coaubored Sher 54 joral aces an conference pps He fas writen chet n iat books, ‘Dr. Sarkar is a epitere Profesional Engine in the Sine of New Yor He recrved the Best Paper Ward of the TEE TRANSACTIONS Of ELECTROMAGNETIC COMPATIIITY in 1979, He was an Associ Eto for cane aris of the IEEE Antennar and. Propagation Society ‘Newser, He was the Technical Program Chatman fr the 1988 IEEE ‘Amon ad Propagation SocetyIetrational Symposium and URS Radio ‘Sent Mating Heat Assvite tor fore IEEE TRANSACTIONS ON ELECTROMAGNETIC COMPATIILITY and forthe Jounal of Electroma: metic Waves and sApplcanons We fas ben appointed US. Resch Cont Represetne 10 many URS Geer Asemblis. He ao the ‘Vice Chirean of fe Intecoomiston Working Group of ltrationat URSI ton Tine Dorin Meircogy: Hels member f Sia Xian Intaona ‘non of Rado Science Commision A and 8.

You might also like