Shie Qjan and Ue UT eur Dapang Chen TTS Cae Joint eas‘thas been well understood that a given signal can be represented in an infinite number of different ways, Different signal representations can be used for di ferent applications. For example, signals obraine MEL from most engineering applications are usualy functions of time. But when studying or designing the ss tem, we ofien lke to study signals and systems in che fre- quency domain. This is because many important features of the signal or system ate mote easily characterized in the fie- quency domain than in the time domain. Although the number of ways of describing a given sig- nal are countless, the most important and fundamental variables in nature are rime and frequaney. While the time domain function indicates how a signal's amplitude changes over time, the frequency domain function tells hhow often such changes take place, ‘The bridge berween time and frequeney is the Fourier transform ‘The fundamental idea behind Fourie’s original work 1was to decompose a signal asthe sum of weighted sinusoi- dal functions. Despite their simple interpretation of pure frequencies, the Fourier transform is not always the best tool ro analyze “real lite signals,” which are usually of fi- nite, pechaps even relatively short duration. Common ex- amples include biomedical signals [38], [50], [51], [53], [62] and seismic signals (Fig. 1). The seismic signals are not lke the sinusoichl fonctions, extending from negative infinity co positive infinity in time. For such applications, the sinusoidal functions are not good models. In addition o linear transformations, the square of the Fourier transform (power spectrum) isalso widely usedin ‘many applications, However, the power spectrum in gen- crals only suitable for those signals whose spectra do not change with time. ‘One common example isthe speech signal. The bottom plot of Fig. 2 isthe time waveform of the word “hood” spoken by 2 five-year-old boy. The plot on the right is the standard power spectrum, which reveals three frequency tones, From the spectrum alone, however, we cannot tell how those frequencies evolve overtime. The larger plocin the upper left is the time-dependent spectrum, as a fine tion of both time and frequeney, which clearly indicates the pattern ofthe formants, From the time-dependent spectrum, not only can we see hhow the frequencies change, but we also can sce the intensity ofthe frequencies as shown by the relative brightness levels of the plot. In other words, by analyzing the signal in time and frequency jointly, we ae able to better un: derstand the mechanism of human speech. Although the frequency content of the ma- jority of signals in the real world evolves over time, the classical power spectrum does notre veal Such important information. In order to overcome this problem, many alternatives, such as the Gabor expansion, wavelets [21], and time-dependent spectra, have been devel ‘oped and widely studied. In contrast to the ino) classical time and frequency analysis, we name these new techniques as joint sime frequency anal In this article, we introduce the basic concepts and ‘well-tested algorithms for joint time-frequency anal Analogous to the classical Fourier analysis, we roughly partition this article into two parts: the linear (eg, short-time Fourier transform, Gabor expansion) and the quadratic transforms (eg., Wigner-Ville distribution. Finally, we briefly introduce the so-called model-based (or parametric) time-frequency analysis method, Expansion and Inner Product Tn what follows, we will frst review ehe fundamental mathematical tools to perform the joint time-frequency analysis, thats, signal expansion and inner product. As we all understand, fora given signals from the domain , ‘where W can be of Finite dimension or infinite dimension, \wemay writea signals in terms of linear combination of the setof elementary functions (yy where Z denotes the ser of integers for the W domain, re, a ‘Second ‘4-2. "Hood" (data courtesy of ¥. Zhe, the Beckman Insite atthe Univers ofIf a physician uses a ruler whose smallest scale is the decimeter to measure a patient's height, then there is no way to tell the patient's height in terms of centimeters. ‘which is known as an expansion or srs. When the set of (¥,) gue forms frame [21], fami of vectors that charactefizes any signal from its inner products (or scalar product) a in (2) and (3) there wil biea dual set 7}. such thatthe expansion coefficients can be computed by ste ee @ for continuous-time signals, or 2, =(5,0,)= Soleo (15)= DEW oF o for discrctetime signals. “*™ in (2) and (3) denote the ‘complex conjugate operation. In electrical engineering, formolae (2) and (3) are known as transformations and W(t) is called the analysis function. Accondingly, (1) is sstown as an inverse transform and y(t) is ealled the syn thesis function, ‘The inner product has an explice physical interpreta tion, which reflets the similarity between the signals) and the dual function@., (). Inothee words, the ager the inne product a, the closer the signals) is to the dual function, (0). fa, depicts the signal's behavior in the domain The operation ofthe inner productin (2) and (3) may be thought of as using a ruler, constituted by a set of functions {G.,}, fo measure the signal under investiga tion, Each individual function can be considered as the tick mark ofthe ruler. The expansion coefficient {a,) in dicates the weight of the signal’s projection on the tick ‘mark determined by W ,. Weall know that the precision ‘of any measurement largely depends on the smallest it used. Ifa physician uses a ruler whose smallest seal is the decimeter co measure a patients height, then there is ‘no way tote the patients height in terms of centimeters ‘The goodness of the ruler is measured by the fineness of the unit. So, elementary fanetions for the signal expan: sion should be chosen such that the resulting tick mark is the finest. For a given frame {y,),_2° its corresponding dual frame is generally not unique. One particularly interest: ing scenario is, (¢}=ew ,(@), where c denotes a constant ‘number. In this Case, we name {V,} 4 a8 tight frame, For the tight frame, the expansion cOcflicients e{a,} are exactly the signal’s projection of the synthesis functions ‘up 0. constan multiplier. When the dual frame is not "unique, dhe resulting expansion is redundant. ‘When the dual frame () cis unique, then the set Of {Wye forms a basis and there is no redundancy. In this ee, (we, )eeBln—'] a) where 1 3m { 0 otherwise ® If, =ey,, then (y,.W,)=c8[n—n'] and {yp are orthogonal. Otherwise (Wy) jo, and {,) ,.y are biothogonal, when {y,}yez forms a basis. For the biorthogonal expansion, th signa’ feature described by the coefficients {a,} may not have obvious relations with the set ofthe basis function {y,},,» (Fig. (a)} For the orthogonal expansion, howevery the expansion coef cients (a,} must be the signal’s projection on the basis function {y.,} .» [Fig. 3(b)]. Moreover, in al ase, the positions of y, and. are interchangeable, That is c= ther of them could be tsed for analysis or synthesis “The concepts introduced above can alsobe understood fom the matsix operation point of view. Suppose that the length ofthe signal ss L. We have the mattix G with d- mension ML and the matrix H with dimension LM, where M2 L. IF HiGiui=F 6) for all vectors # with L elements, then we say that the space constituted by the matrix His complete. Note, that ingeneral, the matrix G could not be unique. If Mf = Z(H and G are square matrices), then G must be equal to HT! In this case, H and G are said to be biorthogonal to cach other. IfG =H! = Ht, Hand Gare orthogonal. The su- perscript “H1” denotes the Hermitian conjugate. In both biorthogonal and orthogonal cases, the matrix G is unique (6), taking the transpose, with respect (a) Blormogona Expansion (Orthogonal Expansion ‘4-5 In both cases (v,}a compete for three dimensional ‘space btn orthogonal expansion expansion coefcients 0, are exoctly the signa’ projection onthe elementary functions‘petra someone The elementary functions employed in the Fourier series extend from negative infinite time to positive infinite time, and are not associated with a particular instant in time. HG sa 5=(G ay)" (Hyg) F935 ” ‘which indicates that either Hof G can be used as analysis (or synthesis). The two most fundamental questions for the expansion are 4 How to build a meaningful set of functions {ex 4 How to compute the corresponding set of dual ine tions 10 0 ye ‘One ofthe tose well-known expansions isthe Fourier se ries. Fora periodic signal 7(t) with a period T, the Fou- rice series defined a8 70 8) ‘Because the clememtary functions ¢/®"™ in (8), har- monically related complex sinusoidal functions, are corthonormal with respect to the scalar produtt, the dual function and the elementary function have the same form. Hence, we can readily obrain the expansion coefficients of the Fourier series in (8) by the regular inner product op- eration (2), eg, SAT) de . Which indicates the similarity between the signal and a set ‘of harmonically related complex sinusoidal functions {6/7}, Because e!?™ corresponds to impulses in the frequency domain, the ruler used for the Fourier trans- form possesses the fines frequency tick marks (thatis, the finest Frequency resolution). Consequently, the measure- ments, the Fourier coefficients {a,}, precisely describe the signa’s behavior at the frequency 2n1/T. For a non-periodic signal s() in Z?, we have = Lf” sieje™deo and the Fourier transform 5(@) Flanemas an Conversion the quar of Four transmis named dhe power spectrum. Aeconding othe Wiener Kinch theorem, the power spectrum can slo be considered 3 {he Fourerranform ofthe conection function, i Peo)=So}* [Rema a ‘where the correlation function R¢®) instantaneous correlation s(¢)s(¢ 2), the average of the Ro)=[_ seen “sees ye(e—ai2yde. 3) Note that the averaging process removes alltime infor- mation from (7) Finally, the reader should bear in mind that a func tion's time and frequency properties are notindependent. ‘They are linked by the Fourier transform. For example, ‘we can find a function chat has an arbitrarily shore time duration and narrow frequency bandwideh at the same time. Ifa function has a short time duration, its frequency bandwidth must be wide, or vice versa. From the uncer ‘ainty principle point of view, itis the Gaussian-type sig ral, such as )-(£) expan) aee()' ont a that achieves the optimal joint time-frequency concentra sion. The balance of the time and the frequency concen tration is controlled by the parameter et. The smaller the valucofay the narrower the fequency bandwidth (longer time duration), oF vie versa Gabor Expansion ‘As mentioned in the preceding section, the classical Fou: "er series unsuitable for characterizing signals that only last fora shore time or whose frequency contents change cover time. The reason for this is thatthe elementary fune- tions employed in the Fourier series extend from negative infinite time to positive infinite time, and are not asso ated with a particular instant in time. In other words, the tick mark adopted by the classical Fourier transform does not contain time information. 244 Gabor expansion sampling gridGabor Expansion and Short-Time Fourier Transform In 1946, Dennis Gabor, a Hungarian-bom Brits physi cist, suggested expanding a signal into a se of functions that are concentrated in both the time and frequency do rains [26]. (In 1970, Gabor was awarded the Nobe Physics Prize for his discovery of the principles underlying, the science of holography.) Then, use the coeflcints as the description ofthe signal’ local property. The resulting, sepresentation is now known as the Gabor expansion, s= EE Chl 2B as where C,.,are called the Gabor coefficients. The se ofl ementary functions {),4()} consists ofa time- and fe ‘quency shifted function b(t) i.e. bg al f)=blt—m TV as InGabors orginal paper, b() isthe normalized Gaussian functiong(#), defined in (14), because itis optimally con centrated in the join time-frequency domain, i weems 0° the uncertainty principle. The parameters T’and 2 are time and frequency sampling steps, respectively. ‘The product of 712 determines the density of the sampling, id. The smaller the product, the denser the sampling. Fig. 4 illasrates the elementary functions used inthe Gabor expansion and the Gabor expansion-sampling, grid. Because the frequency-shifted Gaussian functions have a short time duration, the Gabor expansion is more suitable for characterizing transients and signals with sharp changes (Fig 1). The time and the Frequency reso lutions Of fag (t) ean be adjusted by the parameter a in (14). The smaller the value of, the bette the Frequency resolution (poorer time resolution), or vice versa. Be cause b,(¢) are concentrated in [wT 02}, Gabor sug. gested that the coefficient Cy. rellected a signals behavior inthe vicinity of fm, ni]. As a matter of face, that is not really true, because (ig(f)) in general does ‘not form a tight frame. Consequently, the Gabor coef cients, ae nota signals projetion ond, (2) Weshall claborate on tissue later. Gabor had the clear insight to see the potential ofthis ‘expansion, anit was toa great exten, his work and his claims that spurred much ofthe subsequent work on join: time-frequency analysis. However, he didnot actually contribute to the mathematical theory of the Gabor ex- pansion, Many important aspects ofthe Gabor expansion ‘were not known during his iesime (24), [29}-[31] “Moreover, Gabor limited (tobe the Gaussian func tion and 7! = 2a. ln fact, as long asthe sampling grid is dense enough, for instance, Ths2e a7 all those commonly used analysis functions, such as the ‘exponential, rectangular, and Gaussian-type functions, ny Gabor had the clear insight to see the potential of this expansion, and it was to a great extent, his work and his claims that spurred much of the subsequent work on je time-frequency analysis. can be used in the Gabor expansion. When the product T0=2x, itis considered critical sampling, When T2<2n, itisoversampling. Note tha the set of (.4()} in general ‘won't be orthogonal unless b,q(¢) are poorly concen: trated in cither the time of the frequency domain. If {iy,(f)} are not orthogonal, then the estimation of the Gator cocficients Cy, wor't be trivia. A natural ques: ‘tion at this pointis“Fiow do we compute the Gabor coef: ficients” If the Gabor coefficients are not unique, which ‘one can beter describe a signal's time and frequency be: havior jointly? Before we answer these questions, le’s re view another approach, the short-time Fourier transform (SIFT), which appeared at about the same time as the in- troduction of the Gabor expansion [32] While Gabor focused on the expansion, the STFT modified the Fourier transform (inner products of s(f) and 0%). Instead of processing the entte signal at once, STF takes the Fourier transform (FT) on a block by block basis (Fig. 5). ‘Therefore, the resulting Fourier transform can then be thought of asa signal’ frequency behavior during the time period covered by the data block. The procedure can be described by STFT [mT nO) [ser ce-mrye de as) where the analysis window function 4(¢) balances the time and frequency resolutions, ‘The smallee the time du- ration of y(F), the better the time resolution (poorer fre- quency resolution). The blocks could be overlapped or disjointed, The percentage of overlap berween each block 25 Shortie Fourier transorm.is determined by the time sampling step T and the length of the analysis window function 7). ‘Let us rewrite (18) in the regular inner produet form, SUPT WT nafs, Oe oO where Yast em em ‘Note that (20) and (16) have exactly the same form; both are time- and frequency- shifted versions ofa single pro rotype function. Let Cy = STET(m, Q], thea (15) and (19) form a pair of Gabor expansions. The relation- ship between 9(F) and b(t) is (20) Pa OP O)=BE~2 Dd bana OBE) en ‘which is known as a formal identity. Equations (19) and (18) indicate tat the STFT i, in fat, the Gabor cocti- cient. Conversely, the Gabor expansion can be thought of| as the inverse of the STET. However, these relationships ‘were not widely understood until the eaey 80s [3}-(5} “Tracitionally, we investigate signals and systems inthe time othe frequency domain separately, The Gabor ex- pansion pair, (19) and (15), now allow us to map time-domain functions into the joint time-frequency do- inain, This has been found extremely helpful when the system to be analyzed is time variant. Importantly, fora given time functions(t) and analysis window funtion (2), we are alvays abe to find the joint time-frequengy Function C,,. However, for an arbitrary 2D function, there may be ho corresponding signals) and window fanction ¥(t). The set of {C,,,} is a subspace ofthe entire set of 2D Functions (Fig. 6) ‘The key issue ofthe Gabor expansion is how to com- pte the dual function (2 fora given bie). Except for a few functions atrial sampling [4], [22], [25], the ana- 2.0 Functions {4.6 The Gabor coefcents are subset ofthe set of 20 func- Iytical solutions of (21), in general, do not exist, This fact inspired many researchers ta seek the discrete version of the Gabor expansion. Reeently, one of the major ad- vanees in the study of the Gabor expansion is the discov- cry of the discrete version [33], [34], [37], [44], [63], (72]. Although there were some discrete versions of the Inverse STFT [40] suggested before, they were not moti- vated by expansion and inner product operations. Some important aspects, such as the relationship between oversampling and perfect reconstruction were not clear. In particular, the summation of computing the dual fune- tion derived previously, was generally unbounded, which is not directly suitable for numerical implementation, By using digital computers, we can now readily compute the Gabor expansion and apply it to real-world problems. Discrete Gabor Expansion Applying the sampling theorem and the Poisson-sum for- mula [44], [63], we derive the discrete Gabor expansion sll, Yc, hema val Ca =Dslieli-maMyy” (23) where AM denotes the discrete time sampling interval. N indicates the number of frequency channels. The ratio 1N/OM is considered as the oversampling rate. Fora stable reconstruction, the oversampling rate must be more than, ‘or equal ro one. Usually, we require that Lis evenly divisi- ble by N and AM. The number of frequency bins N is a power of evo. ‘The dual fonction can be solved by AM independent linear systems [47], i, AT, £20)12,.8M-1 sy where the elements of the matrix A, and vectors 7, andy are defined as where OS p< L/ AM, and L denotes the length of fi] and ‘lil. The periodic ausiliary function i] is given by : re fa Oster, osye2 tions. An arbitrary 20 tme-trequency function may notbea (+L): : 2k, wold 0 Lsic2L-N N” (asy awe 1999 REE SIGNAL PROCESSING MAGAZINE ”When the signal length is equal to the length of iff] anel Li] [63], we can also simply let - i Blernt|=mi) — osqe Lienty=i) 1 (05 In this ease, the solution of yi] almost always exists! But, the samples have to be periodically extended. Hence, iis only suitable for finite and small number of samples. On the other hand, using (25), we can impose any type of ex- tension on the samples, which enables us to perform on: line processing. Moreover, this method guarantees that the length ofthe dual window will always be the same as the length of the original window. Orthogonal-tike Gabor Expansion When N = 4.M (critical sampling), the solution of] ie (24) is unique (thereby the Gabor coeticients Cy ate unique). Fig. 7 ilusteates the dual function i] of the normalized Gaussian function gf in (14) at eiticl sara pling. Notethatin this case fi] i concentrated neither fime nor in fequency. ‘According t0 (23), the Gabor coefficients C,., rein ner products of] and aa, which reflec the similarity bereen ff] and] Efi is badly localized, then the Gabor coefiients C, will fil eo describe the signal’ property in the vicinity of fmaM, no matter ow Si] is concentrated. To ensure that the Gabor cosficiens Guy indeed characterize a signal's properties around {md}, we have wo make] concentrated in both the time and frequency domains. Onee the Function] bse lected to be optimally concentrated (ait always is), the natural soluions of 7] are those which are similar co the predetermined i], 6 7) ‘001 02 02 04 05 08 07 08 ce 4 ‘4-7. Dual function x (solid line forthe nomalred Gaussian fancion g(9 (Coted line) at extea sampling, In many applications, such as subband echo cancellation, the Gabor-expansion-based filter banks have been found to be superior to wavelet-based filter banks. where bi) is assumed tobe normalized. The dal relation ‘described by (24). lis interesting to now that the op- ‘imal solution of (27) is dhe minimum noem of (24) When bli] is a diserete version of the normalized Gaussian function in (4), the exo in (27) sa fanetion of variance o and a sampling pattern determined by AM and N. It is observed that the minimum error occurs when & = x/(NAM) [44]. Fig. 8 demonstrates i] and cortesponding yl} Note thatthe diference berween Dlé] and Yop[A] is inversely proportional to the ‘oversampling rate. For critical sampling, the solution of YE] is unique. As the oversampling rate increases, a = N/AM, Yl] becomes closer 0 bl] Iyfi]iscloser tobi], forinstance, yi] =e], then we can rewrite (22) and 23) as di-mamyw ge Cane Z Anti maeyv Although the set of {0),4{i]} may not be linearly inde- pendent, the Gabor coefficients C,,, are indeed the or- ‘thogonal projection on {H,,[#]}. $0, we name it as an orthogonal-like Gabor expansion. For the orthogo- nal-ike Gabor expansion, the Gabor coefficients C,, will well characterize the signal’s behavior in the vicinity of [mAM, 2] as longash,[f] isconcenceated in [mAM, 9] ‘The optimization in (27) can aso be considered a pro- cess to pursue the near tight frame, Finally the discrete Gabor expansion, (22) and (23), can be thought of as a set of perfect reconstruction filter banks [54], [58]. Be- ccause (22) and (23) can be realized by the FFT and in: verse FET, the implementation of the Gabor-expansion-based filter banks is very efficient. In many applications, such as subband echo cancellation, the Gabor-expansion-based filter banks have been found to be superior to waveler-based filter banks [35]. Time-Varying Filtering One of the most important applications of joint time-frequency representations such as the Gabor expan- sion, is the detection and estimation of noise-corrupted signals (70]. While random noise tends to spread evenly 10 the entire joint time-frequency domain, the signal energy is usually concentrated ina relatively sonal region. Consequently, the regional signal-to-noise ratio (SNR)canbe substantially improved inthe join ine- frequency domain [71]. Fig. 9 depicts the impale signal received by the US Department of Energy ALEXIS/BEACKBREAD satel lites Aer pasing through dispersive media, such as the ‘ionosphere, the impale signal becomes nonlinear chirp Signal While the time waveform is severely conraped by random noise, the power spectrum is dominated mail by radio carver signals that ae basally unchanged ver time (Fig 9, agsin) In this ease neither the time wave- form nor the power spectrum indicate the existence oF He impulse signa. However, when looking at the plot of Cals we can immediately identify the presence of the chitp-fype signal arching aos the joint time-frequency domain. Tis obsession sggests that we cause task to filter out the desired Gabor coefficients {C,,,} from the noise background and then reconstnct the time ‘waveform via the Gabor expansion. “The problem heres thatthe Gabor coefficients area subset of the set of 2D fentions, Usvally the mosified ‘may not bean existing signal that corresponds. it. For the sake of presentation simplicity, let us write (22) in matrix form, ie., seHe (2s) ‘where 5H, and £ denote the signal vector, synthesis ma- trix, and Gabor coefficients vector. ‘The matrix form of (23)is G5 29) where G denotes the analysis matrix, If and G satisfy the dual relation determined by (24), then HIG = I and FaHe=HGs. (30) Note that H and G, in general are not square matrices ex- cept for critical sampling, Let, denote the modified and desired Gabor coefficients, that is, {Gan} are no longer in the set of valid Gabor coelf-, =OC=OG7 (1) cient. In this case, for a given modified (C.,), there 08 os os os oa os oa 2 02 o | 7 a ° 01 03 eat ‘0 01 02 09 04 05 08 07 08 09 1 04 02 03 04 05 06 07 08 08 + (@)L=88,N=B, atta, ovetsanging 87 =0.4018 08 os = os os os oa 0a} oa o2 02 on 01 of ° ty __ 0 01 02 03 04 05 06 07 08 09 1 10 01 02 09 04 05 06 07 08 09 1 (@ L=32, N=, aM=2, ‘overaping = 2, ot = 0 0885 ‘4.8 hi (dated ine) and , (sd line) (as the oversampling rate increases, i] becomes more and mare clove to Mi).‘where @ is a diagonal matrix whose diagonal elements are cither one or zero, Based on £, we compute a time wave form 5, Hecause the modified Gabor coefficients are generally not valid Gabor coefficients, oo) which says that the joint time-frequency properties of F are not the same as the desired Gabor coefficients ¢, ‘A.commion approach i to seek a. 5. such thatthe dis- tance between the Gabor coefficients Of 5g, and the de- sired Gabor coetficients Z,is minimum, that is, min, Gig) which sknown asthe ast square cro (LSE) method. is well known that the solution of (32) is the pscudoinverse of G, that is, (32) J G@TGN'GT, a) Fig, 10 illustrates the LSE algorithm, “The LSE has been widely used and extensively stud- ad for many years [7], [23], [27], [65]. Bue it may not be the best solution for many applications. In particular, to solve the LSE, we nced to compute th pseuidoinverse, which is computationally intensive and memory demanding. When the data includes over 10,000 samples, ie would be difficule o apply an LSE al gorithm with conventional PCs. A more efficient method developed recently isan iterative algorithm ia which we continuously apply the same maskeand recon structthe time waveform. The procedure ean be summa- rized as follows: 2, =0G5, =0GHz, (GH)? =, =0G4,=(@CH)'z, ‘The sufficient condition for the iteration above 10 con- verge can be found in [69]. Under thae condition, AZ, converges to inside of @; 14 The result of the first itesation 5, is equivalent to F, computed by the LSE metho. ‘Two trivial cases, Hi] = Yi] or critical sampling ((H,,) forms a basis), satisfy the sufficient condition (69]. For critical sampling, the modified Gabor coefficients ate al- \ways valid Gabor coefficients, that i, Gs, (34) cy IEEE SIGNAL PROCESSING MAGAZINE Innthiscase, the iteration converges atthe first step. How ver, critical sampling often leads to bad localization of the dul unetion 7, and, as pointed our earlier, iis less Interesting in applications. On the other hang i] = i] usually implies high oversampling and thereby heavice ‘computation. To reduce the computational burden in real applications, we employ the orthogonal-like Gabor expansion introduced in the preceding section. Although, “fis hardly exactly equal toh[], the performance ofthe orthogonal-lke, Gabor-expansion-based algorithm has been found t0 be good enough for most applications. Fig. 11 plots the simulation results for the example i Justated in Fig. 9. Fig. 11(a) depicts the error vs. the numberof iterations, which shows the iteration exponen- tilly converging. Note that there is a substantial im- provement between the first and second iterations. The result obtained by the firs iteration i loser to that com- puted by the LSE method. Bur, they are generally not ‘equal unless the convergence condition [69] is truly met. Jn many applications, the LSE solution often is not the best one. Fig. 11(b) depits the reconstructed signal afeer 30 iterations. In this example, the number of samples is ‘9000, whichis impossible to'be solved by the LSE de: ‘cabo Coorsens fATNSea EES] l Fu ssegesé © 0018 0.027 0.040 0.590.080 (ms) ‘4.3. Due 0 the low SNR, the fnized imple signal connot be recognized in either tne or requency domains. However, from Cx we can readily distinguish (Data courtesy ofthe Nanproiteratin and internavonal security division, Los ‘Alamos National Laboratory) 2210. Map of SE eringscribed by (33) due to the computational complexity and memory limitations. The iterative algorithm not only yields a better SNR, but also is practical for online imple- Time-Dependent Spectrum Besides the linear transforms, such as the Gabor expan- sion and wavelets, another important method for the joint time-frequency analysis isthe time-dependent spec: trum. The goal here is to seek a representation thar de- scribes a signal's power spectrum changing over time. Analogous to the conventional power spectrum, a straightforward approach of computing the time-dependent spectrum is to take the square of the STFT, the counterpart of the classical Fourier transform, Pera” scone the af oS We call ic the STT: based spectrogram to distinguish it from the Gabor-expansion-based spectrogram intto- duced later Fig. 12 illatrates the STFT-based spectrogram fora linea chirp signal. While the time waveforms (bottom plot) and classical power spectrum only tella partof the Signal’s behavior, the STFT-based spectrogram indi cates how the linear chip signals frequencies change ‘over time. The main problem of the STPT-based spec ‘wogeam i that i suffers from the so-called window ef fect. As shown in Fig. 12, the width of (¢) governs the resulting imeand frequency resolutions Inthe left plo, 2 short time duration window function is used, whereas the right plot uses along time dusation window. Obv ously the left plot has better time resolution (poor fre «quency resolution). On the other hand, che ight one has Poor time resolution (better frequency resolution). Al- though the STFT-basedspectrogeam is simple and cas ily implemented, it has been found inadequate forthe applications where both high time and frequency resolu- tions are required, Wigner-Ville Distribution [As mentioned in the preceding section (12), the conventional correlation, in fact, isthe av- erage of instantaneous correlations SUE#t/2)5*(0-1)2), 5, Rior[ sere apnee/2ue=[ Ret By averaging the instantaneous correlations, | we lose the time information. How about us Obviously, there is no window effect, Compared to the STET-based spectrogram, as we shal se later, (37) can berter characterize a signal's properties in the joint time-frequency domain, Formula (37) is known as the Wigner-Ville distribution (WVD), which was originally developed in the area of quantum mechanics by a Hun soe | | | | | | | | | | | | { | | | | (6) Reconstruction (After 20 erations) | = = zs 4 2212 The STF-bosed spectrogram for linear chirp signal (The resus ore subject fo the selection ofthe analysis window funcin. The short ime dura- tion window (le) leads to god tie resolution and poor requency resol tion, or vice versa). ing R(r,2) to replace R(x)? For instance, WvDIe,w)=[_ Reed aIt is the crossterm interference that prevents the Wigner-Ville distribution from being used for real applications, though it possesses many desirable properties for signal analysi: garian-born American physicist, Eugene P. Wigner (64), in 1932, Fifteen years ater, ie was introduced into the sigy nal processing area by a French scientist J. Ville [59] (Wigner’s pioneering application of group theory to an atomic nucleus established a method for discovering and applying the principles of symmetry to the behavior of physical phenomena, In 1963, he was awarded the Nobel Physics Prize.) Let s(t}=Ait}e"", where both A(t) and 6(1) are real-valued time functions. Then, the first derivative of the phase 6() represents the weighted average of the in- stantaneous frequency. Traditionally, the first derivative ‘of the phase is called the instantaneous frequency, which is actually incorrect for several reasons. For example, 6) isa single-valued function, whereas at each time instant, generally there are mulkple frequencies. We can prove thar the mean frequency of time waveform and the power spectrum, we can easily distinguish those two components. Essentially, there is nothing inthe vicinity of 2 ky. during the time interval from 0.01 ~ 0.015 s, However, the WVD plot shows a strongly oscillated term in that area. Because this term re flects the correlation of two signal components [47] itis named the crossterm. Although che average of the crossterm in this example is near to zero (in other words, it does not contribute energy to the signal), the magni tnide ofthe crossterm can be proven to be ewice as large as thar of the signal terms. It isthe crossterm interference that prevents the Wignet-Ville distribution from being ‘used for real applications, though it possesses many desi able properties for signal analysis. Tn what follows, we shall introduce two alternative methods, the Cohen's class and the Gabor spectrogram, to reduce the crossterm interference with limited affects fon the useful properties. While the Cohen's clas can be shoughe of as 2D linear filtering of the Wigner-Ville dis- ribution, the Gabor spectrogram is a truncated Wigner-Ville distribution, Cohen's Class [As mentioned previously, the crossterm is almost always strongly oscillated. So, an intuitive approach of removing the WVD at time tis truly equal to the sig- ‘nal’s weighted average instantaneous fre- quency, i, [Lowrey “H'(r) Flare oo . ‘Therefore, the WVD indeed describes the sig sal joint time equeney behavior. More over the energy ofthe WVD ithe sane asthe ener coment in he signal) ie, [weve oydedo={~ perp ae 1 [sey do (39) Asa result ofthis property, the WVD is often thought of as a signal’s energy distribution in the joint time-frequency domain, Ie is interesting to note that all useful properties of the WVD [i.., (38) and (39)] are as- sociated with certain rypes of averages. The STFT-based spectrogram possesses neither (38) nor (39). Fig. 13 il Iustrates the WVD of the linear chigp signal. Compared to the STFT-based spectrogram in Fig. 12, the WD has much beter time and frequency resolutions, ‘The problem ofthe WVD isthe so-called crossterm in rerference. Asan example, le’s look at two time- and fre- «quency shifted Gaussian functions in Fig. 14. From the ar G00 00s 10 adr a0 oes Soon subject to the selection ofthe analysis window function. The Shor time dura- tion window (of) leads to good time rescltion ond poor frequency resol tion, or vce verso) | 5 | | : | _— 2} | eee camo 0osabi0 gts 01a) tas ‘Soon 215. The WD ofa lncor chirp signalthe crossterm sto apply a 2D low-pass filter (0,2) tothe Wigner-Ville distribution, ie., lee 7 ES, [Low amDe-10- nye 0) Expanding theterm of WVD in the above equation yields Coeff an ong b ern Fdure (41) where the function @(, x) denotes the Fourier transform ‘of (1,1). Formula (41) was first developed inthe field of the quantum mechanics by Leon Cohen (17] and, itis named as Cohen’s class, Because the kernel function (F, 1) [or 6(t, )] determines the property of (41), Cohen's class greatly facilitates the Selection of our desired trans- formation. It is interesting to note that almost all previ ‘ously known time-dependent spectra (38), [49] can be ‘written in the form (41). The prominent members of Co- hen’s class include the STFT-based spectrogram, the CChoi-Williams distribution [14], the cone-shaped distr: bbution [74], the adaprive kemel representation [6], and the radial Butterworth kernel representation [67] and [68]. Comprehensive discussions of Cohen's class ean be found in [16], [18], [19], [28], [47], and [66] ‘soe = 406.2 00840 | i 0060 506-2 106 ro SEI a0E1 20E1 £74, WUD of he sum of 0 Gaussian functions. {815 The sgnol energy in the jit tine requency domain can ‘be represented in terms of on infinite numberof elementary energy tems Al those atoms are concerted, symmetrical ‘nd osclate. The emount of energy contained in each ind vival atom is inversely proportional othe rte of esilaton nnci 1998 Gabor Spectrogram Another way of separating a signal's components in time and Frequency i to apply the Gabor expansion [42], [46], as introduced in the preceding section, For example, we first expand the signal into, EF Cobaall ‘Then take the Wigner- Ville distribution on both sides to brain WrDE.o) FTC. Chae WUD, ir XL BouCrae Pruett) 9 where WD, ,. (42) denotes the WVD of evo time and Frequency sifted Gaussian functions defined in (16). As ilustrated in Fig, 13, 77D, ,.(¢.0)is concentrated, sym- metrical, and osilaged, fe can be shown that the energy Of WD, ,.(¢,0) i inversely proportional tothe rae of coseillation. IF we consider WVD,,, (640) a8 an energy ftom, then (42) indicates that a signals energy can be thought of as the sum of an infinite numberof energy t- ‘ms. The contribution ofeach individual energy atom the signals energy not only depends on is weight, Cu aC as but abo on the osilation rate. The highly ‘ostllated stom ae directly associated with ceossteem in terference, but have negligible influence tothe signal en- erey [46], [47] Thased on the closeness of fgg) an ys (it te time and frequency domains, we rewrite (42) as GS, (¢.0)= Cass ary: WVD, lbs) patente (aa) which is known as the Gabor spectrogram (GS), because itisa Gabor expansion-based-spectrogram, The parame- ter D in (43) denotes the order of the GS. When D = 0, the GS only contains those terms in which wv = mand» = In other words, in this case, we only consider the correlation of two identical components. As the order D increases, inereasingly more les-correlated components are included, When D goes to infinity, the GS converges to the WVD. Fig. 16 depicts the WVD of an echo-location pulse emitted by the large brown bat, Eprescus fascus. In add tion to the true signals the WD also exhibits strong, ‘rossterm interference. Fig, 17 illustrates the STFT-based, spectrogeam and Gabor Spectrogram for the same bat sound, Obviously, as che order increases, the resolution improves, The best selection usually is D = 3 ~ 4. When D gets larger, the GS converges to the WD. While the instantancous Frequencies computed by the STFT-based, spectrogram are subject to the selection of the analysis, ‘window function, they are much closer to the real value in ‘the Gabor spectrogram cases. eis worth noting that the energy atom, IVD, (40) in (43), has a closed form thar ean be saved as a table. {EEE SIGNAL PROCESSING MAGALINE oyConsequently, once we ob: tain the Gabor coefficients C4 (which can be effi citntly computed by the windowed FFT), the rest of the process is nothing more than repeatedly using the look-up table. In principle, we can also decompose other trans forms, such as the Choi-Williams distribution, via a signal’s Gabor expan: sion. Irhas been found, how- ever, that the Wigner- Villedistribution-based de- composition (42) has the best performance—a result of the energy atom PVD, (2.0) in (42) being optimally concentrated in the joint time-frequency domain, Moreover, as previ- ‘ously mentioned, the energy atom WVD, (0) bas 3 closed form. Consequently, we don’t need fo compute it online, resulting in a tremendously improved computa tion speed, Parametric Time-Frequency Analysis ‘The joint time-frequency analysis methods presented so far are essentially nonparamettic in that no model is as- sumed to bea priori. When the analyzed signal is discrete, the time and frequency resolutions ofall those methods are subject to the number of data samples. In contrast with the nonparametric methods, there are several para metric approaches, such as the matching pursuit and adaptive spectrogram that have been proposed recently 136], [41], [43], [45]. In whacfollows, we shal briefs trodluce the signal adaptive approximation and adaptive spectrogram, ‘Assume that the given dara are samples of the signal s@rie, st) F Ad te 2 Abst a lO) “ where by(f) denotes a desired signal model, such as 2 chirpiet, 1,0)-(% expat sito, +B o-rye-1) 4s) Which has unit energy, that is, BCP =fpaceP dea 46) ‘4-16. WWD of echolocation pulse (Bat data provided by Cuts Condon, Ken White, and Al Feng of ‘the Beckman Insitute a te Univesity of lings). |< | —— | = Se | || ae ———— | { & TT 117. Comparison of Sand SIFT based spectrogram 4(¢) in (448) denotes the residue after the kth approxima: tion. For example, s,(¢)=s00) «ay (ES (EI{5 (FL Dg (FI ho (F)=5 (6) Agha lt) (48) Sys B= (5 Phy hy ()=8y CA hy (F) a9) where the coefficient A indicates the signal projection on the elementary function y(t). The elementary function 1b, {t) is selected such that the residue is minimum, i.c., 2 ba OP =e fsa nb MME (60)It is easy to see that adding (47), (48), and (49) yields (44), Based on (46) and (48), we have b@P oo? 44)? (sh) BOP BOP AP bo OP 42 HAP gay 4nd Finally, bot SAF Ak a Ircan be shown that the residue sy. (€)[° converges as the number of terms N increases. In real applications, we "usually neglect the residue when i is small enough, to re- write (44) a5 Abt) os (54) and (83) 38 bor Zar 2 (55) Taking the WYD on the both sides of (4) yields WD, 0)=S)A,P WD, (0) TY DACA WD, 60) es : (56) ‘The first summation represents the set of autoterms, whereas the second indicates the set of crossterms, Be cause the Wignet-Ville distribution conserves energy’ (39) and isa using relationship (55), we have the adaptive spectrogram as AD, (t,0)=3|4,|' WYD,, (¢,0) RAtwn, «7 Formula (54) is a signal adaptive approximation, which was independently developed by the authors ofthis article and Mallat et l (36], [41], [43], [45]. When the elemen- tary function /y(¢) has a form (45), we name (57) as a chinpler-based adaptive spectrogram, ‘The chirplet-based adaptive spectrogram has very high resolution without crossterm interference. Compated to the nonparametric methods, the parametric algorichms are computationally intensive. ‘The key issue of signal adaptive approximation and chigplet-based adaptive spectrogram [73] isthe estimation of the optimal lemen- tary function (2) in (50). Usually, they are only suitable forcertain types of signals. However, they could perform extremely well fa parametric model indeed fits the ana- lyzed signal [57], [60] While the STFT and Gabor expansion provide vehicles to map a signal between the time domain and time-frequency domain, the time-dependent spectrum is a powerful tool for understanding the nature of those signals whose power spectra change with time. Summary In this article, we briefly introduced the concepes and methods of joint time-frequency analysis. Like the classical Fourier analysis, the mathematical tools of joint time-frequency analysis are also inner product and expan- sion, Moreover, the joint time-frequency algorithms also fall into two categories: linear and quadratic. While the 'STFT and Gabor expansion provide vehicles to map asig- nal berween the time domain and time-frequency domain, ‘the time-dependent spectrum isa powerful too for under- standing the nature of those signals whose power spectra change with time, The material presented inthis article ap pears ro have reached a level of maturity for real applica tions 9), (11}:(13), [15], [52], [55], [56], [60], [61] Note that all methods introduced in this section are mainly designed for signals whose feequeney contents change rapidly overtime. But their high performance is at the cost of computational complexity. Many applications may stare with the STFT-based spectrogram when a sig- nal’ frequencies do not change dramatically because the STFT-based spectrogram is simple and easily imple ‘mented. When high resolution isa must, the Gabor spec: ‘rogeam is usually a favorite because itis relatively robust and computationally efficient ‘Acknowledgments ‘The authors wish to express their deep appreciation 10 Professors Piotr J. Durka, Hans G. Feichringer, Shidong Li, Flemming Pedersen, and XangGen Xia for their many constructive comments as well as advice, The authors ‘would also like to thank their colleague, Dr. Mahesh CChugani, for his carefal reading of manuscripts and valu able suggestions Shie Qian isa Senior Research Scientist in the Advanced, ‘Signal Processing Department at National Instruments Co. ([email protected]). Dr. Dapang Chen is Manager ofthe Advanced Signal Processing Department at National Instruments Co. (dapang-chen(@natinst. com).References [1] Ay LA bier. i peach 0 shrine Fore nmi’ [21 Amin “Time cry pcr aa and ion ‘sont poses” Time ren Sea ‘Mata patios, B Rost pe 20828, Asa: ey [a1 Rosana epuio sign fo Ga clerty i al Poe HE, vol pp 558539, 18D 4. asians, ampli ho rte ples pcg ‘Gabor eprson of sgn Guin Geeniy ah” Opt Ea uo, 9120.0, 9 S598 aA 198 [51M asin, 0 th lingo rpcoion in dal ign po cong IEEE Tran dot Spc Sal Prin, a ASSP $34, 9 866873, Aap 1988. 161 RG. Brak a DL. Joes“ sil pnt ine essen ep senso IEEE Tan Sal Prag, wo yp. 189-1683, Fan 1994 (7G Boa at Pa Toei gn our, Sh Sal Pcs, 3, m6, 9p. 4-4 fe 1986 [8] V. Chena aii Fart, ene ring mat fle, nd al wate coneltor® Opt xpi sl 33 7, 39122917, [9]. Crean i. “Time ey rr Foe ran hing” Ps IEEESP stration Sppron Tone rgeny Tek Ana, Pace, ramp. 39.92, fae 196 [001 V, Chen “CFAR dettio destin fukn sil in in ‘isin ey Cahora” Pe SHE oe Ware Apo, So 2763, pe 285-29, Ape 198, 110] ¥- her, “Apion of ee src resin aiming" Osa gig, 35,04 pp. 1152161, pe 187 112], Chen “Tine een itd ISAR ings nti eg” Pe. Stim ti pre A Rar mae 07 [15] ¥. Chena. “on dine aguey an foreee ‘ange Spey” EEE Tra Aa. Hs Se, 34,0 2, #819, Apl [U8] H.C at Wii, Ingrove ine sey erection ‘lcmpene signing esncri kermch IEEE Tis ‘Seth Sa Poss, 37, pp BSD, ne 98 [1S], Chagn, "Reems of Bape gain ‘rom mana eri” PD Test, Rel Plyce I, age 196, 11617. Canon ad W. Mesa, “The Wig lr ‘insignia Cosine iat Pa I here ig Pre Reston wth ar een ia semtomstns” Pi Ra, v.85, 9217280, 18D [1711 Cake, “Gener phe pce ibn ution”) Aah Poe 7p 781-808, 1968 118} 1. Cae, “Tine equities” Pe EEE 0. 7, ‘2 7p. SL SL 1 [1911 Ca, Trg Anas, Egeoos CH, Nh Pee Hal [20] LF Coste an LR. Haines, late Dia Sia Pig, glo CE, NJ Prec Hal 19S, (211 Dateien, The wae eaten, gue eon, “ln IEE Tr fn Ty 1, Spt [22] PD. Baier" expr of a aperite ilin cxponctl sees” EE Ha Lat 24,7656, 1988 (2315. Ran, “Tee arin ising a he Gabor xpi,” ‘Spel rcs Dr ad pins, pp S012, Nee Yor Bae, 190. (24) Hinge an. Sta, ar Anand Ait: Ty nd ‘Ais, Ror, 98, [25] Bodner and Pon, Deion of wae ily he Gabor ‘pce ERE Tn Aen, Src, gal Prwenany. 37 0, 2ipp. 168-180, Fey 198. [26] D Gabor, “Teor fom HE, a 9,0 [271 Hiss an W. Ket "ica pal ples” BBE ‘me Sm Presa, vl 400.2, 88286, Ferny 1992 [28] Hhosech at G8 Hacer a ques see rege ial ern" TREE Sal Prin ageing, [29] ATES. amsea, Gator mprnzon of geen aos” ‘Math Aaa. vo 3p 377395, 198 [30] ATE, Janse, Openaly pope of he ats window spect am IEEE Trane Sal Ping, 39,0, pp 292208, ay [31] ALES. ans, Danity a bonboporly fr Wey seta ran wap aa Ap [32] R Konig LK, Dan nL. Lag “Te wn teh A Sener, 18, p19, 1986, [3315.14 °8 genera hey of cee Gad eapaion Pe SPEDE ‘Macon igen: Wav gating, Sn Diego, CA ly 19 [34] 5. tian DM. Bey "A parame of crete Gabor xa ‘SHEETS Pgh 2.202, ey [38]. Laat LM, Moms Gate Expansion or Eso Cael Sout [8] Min 2 Zhang, “Maing rw ime quency one SIRE Tne Sg Prcin, oA, 9p 39073418, son [B71}.M Monin. a Dit Gat epson of dren signa Saye aE Sa Pang gna 2 [38] CH. Page “toerancns pomepecran” AP Pl we 2, [39] Fase The Gabor ess Sn fae spect Ps ICASSP, Sexe, WA May 1998 [s0] 48 rol, Tincegery epson of gl sia anda ‘econ sie Fe an TELE Tra, pS rl Pci, Yel 28, 0p. 5-49, bear 1980. [#1]. D. Chena Khe Sign spprosimiion vi da ade ermal Gaus enone and appr fps cen.” Ir HCASSP92, Sen Fann, CA, pp. 48 Mach 1992 [42)S. Qn an). Mori "Wier dsntondosoposson a stm eed een, Spa Pca, 280.2, 9 125-144, ay 192. 143] 5.Qiun nD. Chn, “Spa ersten a ape Gain ne ons ant opie peroneal Ce mai Stand Son. 5-65, Baior, MD, Mach 1993 {8} $a nD, Ch, Diet Ca an IEEE Tom Sal Presing ol 41,807, 2829280, hy 1993 {48} 8. Qin and D, Chen, *Sgnlqreseaton wing ae noma Sen pl Pang 3 1 ah uxsci 1999|{46} Qa D, Cher, “Dssipernn fhe WigeViledsaion seine equ toner FEE Tom Sigal Pring, $2, 0, pp 2885-2841, Ooter (8715. Qin and D. Chen Time Foe Aas Epon Ci, Dp Mee Ha 196, [a8] 5. Quan 16. Fda “Dice Gor saan pil ep rent” EES Tan Signaling, 48, 10, 2258 2268, ee 15, 99). Rock Sg eng deo ine a hen TEEE ‘rn Ra ago, Tp. 38-274, 1068, 150]. Sai M Cag: Mash, M, Tag and A. Hain "Deseo of set pagina” 1c Ana nition he IEEE Eyeing it in nd BS, Son Dos, 317-31, be 1993, [51] Sivig ML Chga K Peso, A. Hain, M Tagan Mac, “Dipti procenn sleep deesom,”r CASSIS, Do, ML pp. 1781195 May 1998 [82] Ry. Sao MSDN. Keg Fas a MS Ser, Tite iso th EG ip” Pre ISSP90, Sal ering, Tatil en Apne, Go Coo, ea (58) Sab, tS DLN. Kg Jit nd MLS. Ss, “Tine een domain ren ine eons” Tine gy Sioa drt Nidan pti, Boss EL gman Chests, Eagan Wy sd romp 5419, 1952, [54}6.Seangant.@ Spe, Wir ad ier ane, Wee, MA Weta Canteen, 195 [88]. Sin, San X Yan, Baym, X-6, Xia De Nya at St “Lang aoa aye an tine ey ali ad pte of HEC," Pr TEE, 5, 90, 2p. ISI, Nose 186 180). Sin ALL Se, S. Qn, Basan PD. Adon abt Selb Taney sr of gh pry ioe a of pic cnrs” Pe TREE EAB? Cog Oe 197 (8711 Tian Ling. “hin sme sey SAR ing aie eosin” IEEE Tram Arena Pp ol 86, 02, 20127, 158] PP Vsthon, lie Sym amd Bay, glenn Ci, [iPr Hat 198. 180. Vite, “Teer plato de non gal ante in ech ie Taos 3,9. 61.7, 1H, {oo}. Wang a. Zh, "cipal apron fons ‘gine pound moon ts ah ep. (61) ang ing. nV Che, ISAK nn compen vi ap jel ine tury ecg IRE Tne AES, 9 ape [6217 Wang, Son CC Lian A agra “Chin of abo ‘lel per by etre om Wire pecr” inf Hn tr ‘sal Cnr ox arn Regn, atc Cy Np 22829, me 90 149), Wes and. Raz, "Dacee Caer pansion Sil ring, vo [64] Wie, "Os ee qr iron fr enya ei hp Ro, 8p. 709,982 [65] Wea, "Te pce re ar ambiguity on” Tk Seay Rip 157, Mah Res. Cen, Unive of Wasnt at Main Ape 18 46). Wits Reta natn du wd iterrenions Pn EFF 8,909, 12641260, eee [671 D. Waa 1M. Nos “ine eqns epee wing ra woth nel” P, EEE Sai Spee Ta gy a Tne Sale asp 868, Pdphi, PA, Oo erie [68] D.We,“Tieouc sila a Ch ca of buco so inplemeasn agro TD. Te, Unity of Maran {691 X-G Xian. Qian “An rina frien ring in ‘he dee Gabor weston da, IEEE Tam on Sigal Ping epubined 170] XG." deco wing hy sgna an eying ‘eine ie fey don IELE Te Sil Pras. {Bono pp 20722008, Ag 197 (71) £6. Xs, *A uneven o SNR in he ssi Ror ae {erm ori ape pl” ESE Dr Sal Desig so 4p. 200208, mary 98. 17211 Yo, "CanplceGabr somata eps” BE “hae ae Pg, 2p. 1559, pe 1988, 1731 Q Yi, ZN. Qian ant. Che, Adapter ome po eon decompo") Ban (Chines) 3, 49 258, ont 97 (74). 21 Ala nd. Maks, The eof cone stp eri ovine ne rape peers of nema a TELE ‘rt Aamo. Sehgal Presn,so Mo7p. HSK 1, oy