An Introduction to the Sampling Theorem ‘An Introduction to the Sampling Theorem With rapid advancement in data acauistion technology (Le. analog-to-digital and digital-to-analog converters) and the explosive introduction of micro-computers, selected com- plex linear and nonlinear functions currently implemented with analog circuitry are being alternately implemented with sample data systems. Though more costly than their analog counterpart, thes ‘sampled data systams feature programmability. Additionally, many of the algorithms employed are a result of develop- ments made in the area of signal processing and are in some cases capable of functions unrealizable by current analog techniques. With increased usage a proportional demand has evolved to understand the theoretical basis required in interfacing these sampled data-systems to tho analog world ‘This article attempts to address the demand by presenting the concepts of aliasing and the sampling theorem in a manner, hopetully, easily understood by those making their first attempt at signal processing. Additionally discussed are some of the unobvious hardware effects that one might en- ‘counter when applying the sampled theorem. With this ... lat us begin, |. An intuitive Development ‘The sampling theorem by C.E. Shannon in 1949 places ro- strctons on the frequency content ofthe time function sig- nal {(), and can be simply statod as folows: In ordor to recover the signal function () exactly, it necessary to sample ft) at a rate greater than twice its highest frequency component. Practically speaking for example, to sample an analog sig nal having a maximum frequonoy of 2Ke requires sampling at greater than 4K to preserve and recover the waveform exactly. ‘The consequences of sampling a signal at a rato below its highest frequency component resuits in a phenomenon known as aliasing. This concept results in a frequency mis- takenly taking on tho identiy of an entirely diferent frequer cy when recovered. In an attempt to clay this, ideal sampler of Figure 1a), with a sample period of T shown in (0), sampling the waveform f() as pictured in (c) ‘The sampled data points of f(t) are shown in (@) and can be detined as the sample set of the continuous function 1) Note in Figure 1/2) that another frequency component, a), can be found that has the same sample set of data points 28 (in (). Because ofthis itis difficult to determine which frequency a'(), is truly being observed. This effects similar to that observed in westom movies when watcting the National Semiconductor Application Note 236 tsamousaoxra «oY ® rasameuen oxtA SAMPLED ONTA) FIGURE 1. When sampling, many signals may be found to have the same set of data points. These are called allases of each other spoked wheels of a rapidly moving stagecoach rotate back- wards at a slow rate. The effect is a result of each individual {rame of fim resembling a discrete strobed sampling opera- tion fieshing at a rate sighty faster than that of the rotating ‘wheel. Each observed sample point or frame catches the spoked wheel slighty displaced from its previous position giving the effective appearance of a wheel rotating back- ‘wards. Again, aliasing is evidenced and in this example it becomes difficult to determine which is the true rotational frequency being observed. 469 9EZ-NVAN-236 EAL, ANTIAASING FER RESPONSE. ]/ acorTune aes TUE aay . te Es Frequency —> FIGURE 2. Shown in the shaded area Is an Ideal, low pass, ant/-aliasing filter response. ‘Signals passed through the fier are bandiimited to frequencies no greater than the cutoff frequency, fc. In accordance with the sampling theorem, to recover the bbandlimited signal exactly the sampling rate must be chosen to be greater than 2c. (On the surface itis easily said that anti-aliasing designs can be achieved by sampling at a rate greater than twice the ‘maximum frequency found within the signal to be sampled. In the real world, however, most signals contain the entire spectrum of frequency components; from the desired to those present in white noise. To recover such information accurately the system would require an unrealizably high ‘sample rato. ‘This difficulty can be easily overcome by preconditioning the input signal, the means of which would be a bandsimiting or frequency fitering function performed prior to the sample data input. The prefiter, typicaly called anti-aliasing fiter ‘guarantees, for example in the low pass fer case, that the ‘sampled data system receives analog signals having a ‘spectral content no greater than those frequencies allowed by the filter. As illustrated in Figure 2, it thus becomes a ‘simple matter to sample at greater than twice the maximum frequency content of a given signal ‘A parallel analog of band-limiting can be made to the world (of perception when considering the spectrum of white light. It'can be realized that the study of violet light wavelongths, {generated from a white light source would be vastly simpli- fied it inital band-timiting were performed through the use of 2 prism or white light fiter. N.The Sampling Theorem To solidity some of the intuitive thoughts presented in the Previous section, the sampling theorem will be presented applying the rigor of mathematics supported by an ilustre- tive proof. This should hopefully leave the reader with a comfortable understanding of the sampling theorem. Theorem: If the Fourier transform F() of a signal function 1(t) is zero for all frequencies above || = we, then {() can be uniquely determined from its sampled values fy = HoT) o ‘These values are a sequence of equidistant samn- 1_Te le points spaced = 12 = T apart (is thus given by f= _ fom) Beek @ om Proot: Using the inverse Fourir transform formula 17 Fete dw = Lf" Fo 4 ® the band limited function, f(), takes the form, Figure 3a, 1 [ee 1) = zf ‘ Fla)e!** deo ny t= 1(n) ietnen given as hah fe Flere ao 6) n= 37 J a, Flee 6 See Figure Se and 0. Expressing F(a) as a Fourier series inthe intorval a < @ S te we have DY ore ©Where, C= zs. Flo}e!* 6s do ” Further manipulating ea. (7) © pcan be writen a on = 2, ® ee Substituting eq. (9) into eq. (6) gives the periodic Fourier Transform Fplo) = eo (10) d= DS Lach of Figure Sf. Using Poisson's sum formula! F(a) can be stated more clearly as Fa )) Fe- 200 on Interestingly for the interval —ae < w < we the periodic function Fp(w) and Figure Sf. equals Flw) and Figure 3b. respectively. Analogously f Fp(w) were multiplied by a rec ‘angular pulse defined, A - Ho) = 1 esse 02) ° ol 2 a 03) then as pictured in Figures 4b, 0, and f F(a) = H(o) * Fpl) = Hla) > Stne ‘Solving for (the inverse Fourier transform eq (9) is applied to0g (14) eran eo «4 ah, it “e Te wre = 2 andi et sing tiny rt fi be} ' « FIGURE 3. Fourier transform of a sampled signal. 9EZ-NVAN-236 it +t © w Fla) FIGURE 4, Recovery of a signal {(t) rom sampled data Information. wo wale) > Eq (18) is equivalent to eq (2) as is ilustrated in Figure 4e and Figure 3a respectively. ‘As observed in Figures 3 and 4, each step of the sampling theorem proof was also illustrated with its Fourier transform, pair. This was done to present alternate illustrative proofs. Recalling the convolution? theorem, the convolution of F(a), Figure 3, with a set of equidistant impulses, Figure 3d, yields the same periodic frequency function Fpl), Fig ure St, as did the Fourier transform of fy, Figure 3e, the Product of f(), Figure 3a, and its equidistant sample impul- ses, Figure 3c. In the same light the original time function 4(), Figure 4e, ‘could have been recovered from its sampled waveform by convolving fp, Figure 4a, with h(t), Figure 4c, rather than multiplying F(a), Figure 4b, by the rectangular function H(w), Figure 40, to got Fw), Figure 4f, and finally inverse transforming to achieve f(), Figure 4e, as done in the math- ‘ematic proof. Ill, Some Observations and Definitions Wf Figures 31 ot 40 are re-examined it can be noted that the original spectrum Fw), [al < we, and its images F(a), 15) |a| 2 a, are non-overlapping. On the other hand Figure 5 ilustrates spectral folding, overlapping or aliasing of the spectrum images into the original signal spectrum. This alk asing effects, in fact, a result of undersampling and further causes the information of the original signal to be incistin- uishable from its images (i.e. Figure 16). From Figure 6 one Can readily s0e that the signal is thus considered nor-recov- erable. ‘The frequency |fc| of Figure 3f and 4b is exactly one half the sampling frequency, fo=fs/2, and is defined as the Nyquist frequency (after Harry Nyquist of Bell Laboratories). It is ‘also often called the aliasing frequency or folding frequency {or the reasons discussed above. From this we can say that in order to prevent aliasing in a sampled-data systom the sampling frequency should be chosen to be greater than twice the highest frequency component feof the signal be- ing sampled. By definition te2 Qe (16) Note, however, that no mention has been made to sample at procisely tho Nyquist rate since in actual practice itis 2 The convolution theorem slows one te mathematcaly conte in he time domain by simly mang nthe equency domain. That Mas {he Four vansiom Fa). and (has the Power Wanatorm Xe), then 6 omen ix has the Fourerwanstorm Fla). 40) x(t) <—> Flo) # X(0} f(t) xt) <> Flo) *X(o) 472Fate) Te © 1 Fi) o o o FIGURE 5. Spectral folding or aliasing caused by: FIGURE. Allased spectral envelope (a) and (b) of () under sampling Figures Sa and Sb respectively. (b) exaggerated under sampling. Cel S| porter) sare Pn smoorana [obreur gow mnt ne sounce fie Pe nist ceren ‘aren’ | Sonat FIGURE 7. Generalized single channel sample data system. impossible to sample at fe ~ 2te unless one can gueranteo thore are absolutely no signal components above fe. This can oniy be achieved by fitering the signal prior to sampling witha fiter having infinite rolioft ... a physical impossibility, 00 Figure 2. IV. The Sampling Theorem and Its Hardware Implications ‘Though there are numerous sophisticated techniques of im- plementation, it is appropriate to re-emphasize that the in- tent ofthis article is to give the first time user a basic and fundamental approach toward the design of a sampled-data, system. The method with which to achieve this goal will be to introduce a few of the common pers encountered when implementing such a system. We begin by considering the generalized block diagram of Figure 7. ‘As shown in Figure 7, prior to any signal processing manipu- lation the analog input signal must be preconditioned to pre- vent aliasing and thereafter digitized to logic signals usable by the logic function block. The antialiasing and digitizing functions are performed by an input iter and analog-to-dig- tal converter respectively. Once digitized the signal can then be altered or processed and upon completion, reconstruct- fed back to a continuous analog signal via a digtal-to-enalog converter followed by a smoothing fiter. To this point no mention has been made concerning the ‘sample and hold circult block depicted in Figure 7. In gener al the analog-to-digtal converter can operate as a stand alone unit. In many high speed operations however, the converter speed is insufficient and thus requires the assist- ‘ance of a sample and hold circuit. This willbe discussed in detail furthor in the article ‘A. The Antialiasing Input Filter {As indicated earlier in the toxt, the antialiasing fiter should bandit the input signa's spectrum to frequencies no {greater than the Nyquist frequency. In the real world howev- 6%, filters are non-ideal and have typical attenuation or band- limiting and phase characteristics as shown in Figure 8.3 It ‘must also be realized that true band-limiting of a specific frequency spectrum is not possible. n the sample data sys- tem band-liiting is achieved by attenuating those frequen- cies greeter than the Nyquist frequency to a level undetect- ‘able or invisible to the system analog-to-cigital (A/D) con- verter. This tevel would typically be less than the rms quant- zation noise level defined by the speciic converter being sed. 2 order not to erupt tho flow ofthe dscussion a sto tor terms has ‘been presented in Append: A. ‘Foran explanation of quantzatin ort secon WV. 8. ofthis rtte. 473 ‘9€2-NVAN-236 ‘As an example of how an antialiasing filter would be applied, ‘assume a sample data system having within it an 8-bit A/D ‘converter. Eight bits translates to 2 56 levels of resolution. It a 2.56 volt reference were used each quantiza- tion level, q, would represent the equivalent of 2.56 volts/ 256 10 milivolts. Realizing this the antialiasing fiter would bbe designed such that frequencies in the stopband were attenuated to less than the rms quantization noise level of (9/218 and thus appoaring invisible to the system. More spe- cifically Viullscale ~20 0910 nag = 5908 = Aun a It can be seen, for example in the Butterworth fiter case (characterized as having a maximally flat pass-band) of Fig- ture 9a that any order of fiter may be used to achieve the 89 dB attenuation level, however, the higher the order, the faster the roll off rate and the closer the fiter magnitude response will approach the ideal. Referring back to Figure 8 it is observed that those frequen- cies greater than we are not recognized by the A/D convert- ‘rand thus the sampling frequency of the sample data sys- tom would be defined as w > 209. Additionally, the fr ‘quencies present within the fitered input signal would be those less than «4. Note however, that the portion of the signal frequencies least distorted are those betwoon «=O ‘and wp and those within the transition band are distorted to ‘a substantial degree, though it was originally desired to limit the signal to frequencies less than the cutoff a»p, because of the non-ideal frequency response the true Nyquist frequer ‘oy occurred at «wa. We see then that the sampled-data sys- tom could at most be accurate for those frequencies within the antialiasing fiter passband, From the above example, the design of an antialiasing fiter, ‘appears to be quite straight forward. Recall howaver, that all waveforms are composed of the sums and differences. of various frequency components and as a result, i the re- ‘sponse of the filter passband were not fat for the desired signal frequency spectrum, the recovered signal would be an inaccurate summation of all frequency components al tered by their relative attenuations in the pass-band. ‘Additionally the antialiasing fiter design should not neglect the effects of delay. As illustrated in Figure 8 and 96, delay time corresponds to @ specific phase shift at a particular Teaysron a a rassoana mn ||] ell g z 5 | en) QUENCY = {roquency. Similar to the flat pass-band consideration, i the phase shift of the fiter is not exactly proportional to the frequency, the output of the fiter wil be a waveform in \hich the summation ofall requency components has been altered by shits in their relative phase. Figure 9b further indicates that contrary to the roll off rate, the higher the fitter Corder the more non-ideal the delay becomes (increased de- lay) and the result is a distorted output signal A final and complex consideration to understand is the fects of sampling. When a signal is sampled the end effect is the muttipication of the signal by a unit sampling pulse train as recalled from Figure Ja, c and e. The resultant waveform has a spectrum that isthe convolution of the sig- nal spectrum and the spectrum of the unit sample pulse train, 2. Figure 3b, d, and f I the unit sample pulse has the classical sin X/X spectrum of a rectangular pulse, see Fig- ture 13, hen the convolution of the pulse spectrum with the signal spectrum would produce the non-ideal sampled sig- ral spectrum shown in Figure 10a, b, and ¢. It should be realized that because of the bandlimiting or fitering and delay response of the Sin X/X function com- bined with the effects of the non-ideal antialiasing filter (. rron-tat pass-band and phase shift) certain of the sum and difference frequency components may fall within the de- sired signal spectrum thereby creating aliasing errors, Fig- ure 106. ‘When designing antialiasing fiters it will be found thet the ‘loser the fiter response approaches the ideal the more ‘complex the filter becomes. Along with this an increase in delay and pass-band ripple combine to distort and alias the input signal. Inthe final analysis the design wil involve trade offs made between fiter complexity, sampling speed and thus system bandwidth B. The Analog-to-Digital Converter Following the antialiasing iter is the A/D converter which performs the operations of quantizing and coding the input ‘signal in some finite amount of time. Figure 17 shows the ‘quantization process of converting a continuous analog in- Dut signal into a set of discroto output levels. A quantization, 4, is thus defined as the smallest step used in the digital Thi wil be explained more clay in Section V. ofthis arte. PUASE ANGLE, © op Frequency = FIGURE 8. Typical filter magnitude and phase versus frequency response. 474PASSRAND ATTENUATION M8 nour veLay wm scones 5 1 nay : t ry | Wri | 2 | « | 5 | | I 1 4s s 2 t as | x 1 2 1 : | A / i “sy . . o a2 03 4 05 08 a7esoat eS 8) Attenuation characteristics of a normalized Butterworth filter as a function of degree n. ar 02 03 04 05 08 07 Of O8 UO 41 V2 a 4a as te a7 ae aS ) Group delay performances of normalized Butterworth lowpass fiters as a function of degree n. FIGURE 9 ‘9€2-NV 478AN-236 © @ EWELOPE OF SH sauna waveronm veo OF saMPLNs yg MAVEFORM SPECTRUM. sunvse20=11 FIGURE 10. (c) equals the convolution of (a) with (b). representation of fa(n) where f(n) is the sample set of an input signal {() and is expressed by a finite number of bits ‘ving the sequence fan). Digitally speaking qis the value of the least significant code bit. The diference signal «(n) shown in Figure 17s called quantization noise or error and ‘can be detined as ¢(n) = f(r) ~ fa(n). This eror isan iredu- ible one and is a function of the quantizing process. Its error amplitude is dependent on the number of quantization levels or quantizer resolution and as shown, the maximum ‘quantization error is |q/2]. Generally ¢() is treated as a random error when described in terms ofits probability density function, that i, all value of e(n) between q/2 and —q/2 are equally probable, then for the average value e(nlayg=0 and for the rms value (nhs = 9/28. ‘As a side note itis appropriate at this point to emphasize that all analog signals have some form of noise corruption. I for example an input signal has a finite signal-to-noise ratio of 40dB it would be supertivous to select an A/D con- verter with a high numberof bits. it may be realized that the Use of a large number of bits does not give the digitized signal a higher signal-to-noise ratio than that of the original analog input signal. As a supportive argument one may say that though the quantization steps q are very smal with re- spect to the peak input signal the lower order bits of the [A/D converter merely provide a more accurate representa- tion of the noise inherent in the analog input signa. Returning to our discussion, we define the conversion time ‘as the time taken by the A/D converter to convert the ana- log input signal to its equivalent quantization or digital code. ‘The conversion speed required in any particular application depends upon the time variation ofthe signal to be convert- ‘ed and the amount of resolution or bits, n, required. Though ‘the antialiasing fier helps to control the input signal time rate of change by band-limiting Its frequency spectrum, a finite amount of time is stil required to make a measure ‘ment or conversion. This time is generally called the aper- ture time and as illustrated in Figure 12 produces amplitude measurement uncertainty errors. The maximum rate of ‘change detectable by an A/D converter can simply be stat- V full scale BT conversion time 7 tl cum resolvable rato of change It for example V full scale ~ 10.24 volts, T conversion time = 10 ms, and n = 10 or 1024 bits of resolution then the maximum rate of change resolvable by the A/D converter would be 1 vott/sec. I the input signal has a faster rato of change than 1 volt/sec, 1 LSB changes cannot be resolved within the sampling period. In many instances a sample-and-hold circuit may be used to reduce the amplitude uncertainty error by measuring the in- put signal with a smaller aperture time than the conversion time aperture of the A/D converter. In this case the 476bara ont nit — orie 8 oer Es oro oor aare soor sooo = VE D/P /R eA WIE « eu 0 =A sun “ _— nw FIGURE 11. Quantization error. t } g . g g 3 wll ‘ 49 AVPLTUDE UNCERTAINTY ERROR ~ ‘APERTURE TNE au APERTURE TME UNCERTANTY FIGURE 12. Amplitude uncertainty asa function of (2) nonvarying aperture and (b) aperture time uncertainty. maximum rate of change resolvable by the sample-ancshold It 3 appropiate to recall the earlr discussion that the would be spectrum of a sampled signal is one in which the resuitant Py Val sate spectrum fs the product obtain by convolving the input ig- s wee (18) fal spectrum wih the sin X/X specttum of the sampling Pees wavelorm. Figur 13 lustrates the frequency spectrum plot- ‘mocrenange tod from th Fourier vanstorm Noto also thatthe actual calculated rata of change may be ein limitod by the slow rato specticatio fo tho samplo-and:nol 2 inthe wack move. Addtonaly tie very mporam tweety FO) = AT 0) that this does not imply violating the sampling theorem in ei liou of the increased abilty to more accurately sample sig- nals having a fast time rate of change. ‘of a rectangular pulse. The sin X/X form occurs frequently in Anidelsarloandnold tect tates asampein xo ‘osm commuricaon thar and i common called the time and with perfect accuracy holds the value of the sam- ad ” sa eters tects er, anogsn andes yase0 ort am ‘order hold circuit and its effect on a sample data system pler spectrum ute =af2222 Beane] wo a7 9€Z-NVAN-236 o =a q wan atu wan oan FIGURE 13, The Fourier transform of the rectangular pulse (a) Is shown in (b). Is shown in Figure 14 and Figure 15 illustrates the spectra of, various sampler pulse-widths. The purpose of presenting this illustrative information is to give insight at to what of- fects cause the aliasing described in Figure 10. From Figure 16 its realized that the main lobe of the sin X/X function ‘varies inversely proportional with the sampler pulse-width. In other words a wide pulse-width, or in this case the apor- ture window, acts as a low pass filtering function and 0 3 a ve ‘ ” g i a runvaeeo-18 FIGURE 14. Sampling Pulse (a, Its Magnitude (b) and Phase Responee (c). limits the amount of information resolvable by the sample data system. On the other hand a narrow sampler pulse- width or aperture window has a broader main lobe or band- width and thus when convolved with the analog input signal Produces the least amount of distortion. Understandably then the effect of the sampler’s spectral phase and main lobe width must be considered when developing a sampling system so that no unexpected aliasing occurs from its con- ‘olution with the input signal spectrum. ry 0) Is oy o 0 Fe allies « FIGURE 15. Pulse width and how it effects the sin X/X envelop spectrum (normalized amplitudes). . The Digital-to-Analog Converter and ‘Smoothing Filter ‘After a signal has beon digitally conditioned by the signal processing unit of Figure 7, a D/A converter is used to con- vert the sampled binary information back in to an analog signal. The conversion is called a zero order hold type ‘where each output sample level is a function of its binary ‘weight value and is held until the next sample arrives, see Figure 16. As a result of the D/A converter step function response itis apparent that a large amount of undesirable high frequency energy is present. To oliminats this the D/A Converter is usually followed by a sinoothing filter, having a Cutoff frequency no greater than haif the sampling frequen- ‘oy. AS its name suggests the filter output produces a ‘smoothed version of the D/A converter output which in fact, is @ convolved function, More simply said, the spectrum of the resulting signal is the product of a step function sin X/X. ‘spectrum and the band-imited analog fiter spectrum. Anal- ‘ogous to the input sampling problem, the smoothed output may have aliasing offects resulting from the phaso and at tenuation relations ofthe signal recovery system (defined as the D/A converter and smoothing filter combination). 478‘Asa final note, the attenuation due to the D/A convertor sin, X/X spectrum shape may in some cases be compensated for in the signal processing unit by pre-processing using a digital fter with an inverse response X/sin X prior to D/A. conversion. This allows an overall flat magnitude signal re- ‘sponse to be smoothed by the fina fier. 1) o FIGURE 16. (a) Processed signal data points (©) output of D/A converter {6) output of smoothing titer. V.AFinal Note ‘This article began by presenting an intultive development of the sampling theorem supported by a mathematical and i- lustrative proot. Following the theoretical development were ‘afew of the unobvious and troublesome results that devel- ‘op when trying to put the sampling theorem into practice, ‘The purpose of presenting these thought provoking perils was to perhaps give the beginning designer some insight or ‘guidelines for consideration when developing a sample data system's interface. VI. Acknowledgements ‘The author wishes to thank James Moyer and Barry Sogo! for their encouragement and the time they allocated for the wring of this article. APPENDIX A Basic Filter Concepts A filter is a network used for separating signal waves on the ‘basis oftheir frequency and is usually composed of passive, reactive and active elements such as resistors, capacitors, inductors, and amplitirs, or combinations thereot. ‘There are basicaly five types of fiters used to pass or eject such signals and they are defined as follows: 1. A low-pass iter passes a band of frequencies called the ‘passband, ranging trom zero frequoncy oF DC toa certain Cutoff troquency, we", and in addition has @ maximum attenuation or ripple level of Aqax within the passband. Soe Figure ‘acl a eran toqueney = 2 FIGURE 1. Common Low Pass Filter Response Frequencies beyond the w, may have an attenuation greater than Ayax but beyond a specific frequency os defined as the stopband frequency, a minimum attenua- tion of Ayiny must prevail, The band of frequencies higher than «ws and maintaining attenuation greater than or equal ‘to Amin is called the stopband. The transition region or transition band is that band of frequencies between «- and ws. 2. A high-pass filter allows frequencies above the passband, frequency, w., fo pass and rejects frequencies below this, ppoint. Ajax must be maintained in the passband and fre- ‘quencies equal to and below the stopband frequency, as, ‘must have a minimum attenuation of Ayn. See Figure 2. FIGURE 2. Common High Pass Filter Response 479 9EZ-NVAN-236 3. A bandpass fiter performs the function of passing a spe- Ciffc band of frequencies while rejecting those frequen- cies above and below aa and lower, «ict Cutoff frequen- cy limits. See Figure 3. Figure 3. Common Band-pass Filter Response ious two cases the passband is required to Asin the sustain an attenuation of Ayax, and the stopband of fre: quencies above and below w32 and wsz respectively, ‘must have a minimum attenuation of Ayan. Figure 4. Common Band-Reject Filter Response 4.A band-reect iter or notch fiter allows all but a specific band of frequencies to pass. As shown in Figure 4, those frequencies between ws; and aso are fitered out and those frequencies above and below ica and «ct fespeC- tively are passed. The attenuation requirements of the stopband Ayn and passband Ayax must still hold. 5. An ali-pass or phase shit fiter allows all frequencies to pass without any appreciable attenuation. It further intro- duces a predictable phase shit to all frequencies passed, though not restricting the entir range of frequencies to a specific phase shift (.e., a phase shift may be imposed upon a selected band of frequencies and appear invisible to all others). APPENDIX B ARTICLE REFERENCES SD. Stoams, Digital Signal Analysis, Hayden, 1975. S.A, Trotter, ntroduction to Discrete-Time Signal Process- ing, Wiley, 1976. WD. Stanley, Digital Signal Processing, Reston, 1975. |A. Papoulis, The Fourier Integral and its Applications, Mc~ Graw-Hil, 1962. E.A, Robinson, M. T. Sivia, Digital Signal Processing and Time Series Analysis, Hoiden-Day, 1978. CE, Shannon, “Communication in the Presence of Noise,”” Proceedings IRE, Vol. 37, pp. 10-21, Jan. 1949. M, Schwartz, L. Shaw, Signal Processing: Discrete Spectral Analysis, Detection and Estimation, McGraw-Hil, 1975. LR. Rabiner, B. Gold, Theory and Application of Digital Sig- ral Processing, Prentice-Hall, 1975. WH. Hayt, JE. Kemmerly, Engineering Circuit Analysis, McGraw-Hil, 3rd edition, 1978. E.0. Brigham, The Fast Fourier Transform, Prentice-Hall, 1074, J. Sherwin, Specitying A/D and D/A converters, National ‘Semiconductor Corp., Application Note AN-156, Analog-Digital Conversion Notes, Analog Devices Inc., 1974, Al. Zvorev, Handbook of Filter Synthesis, Wiley, 1967. 480