The Theory of Bandpass Sampling Rodney G. Vaughan, Member, IEEE, Neil L. Scott, and D. Rod White Abstract—The sampling of bandpass signals i discussed with erations, and parameter ‘the acceptable and une: cplable sample rates are presented, with specific discussion of| ‘shich are nonminimom. The minimom sam clogieal in that any imperfection in the imple ity of the aceept- Increasing guard ‘Tn applsing bandpass sampling to relocate signals to a base- band position, the signal-to-noise rato snot preserved owing tothe out-of-band noise being aliased. The degradation in sig- hnato-noise ratio i quantified ia terms ofthe postion of the bandpass signal. or the construction of a bandpass signal from second-order saves with the band being ether integer oF Toned 1. Inropucrion Ti satis ands ampting i an aps! fd ital signal processing which has relevance to a variety ‘of disciplines including, for example, opties [1], radar (2), sonar [3], communications [4], biomedical signals [5]. power measurement [6], and general instrumentation, such as sampling oscilloscopes. Historically, Cauchy seems to have been the first to hy pothesize the bandpass sampling requirement [7]. Nyquist [8] and Gabor {9} also alluded to the bandpass case. Koh lenberg [10] introduced second-onder sampling and pro= vided the basic interpolation formula, which was then re- ported in the texts by Goldman [11] and Midleton [12] “The bandpass situation is depicted in Fig. 1, where the sampling rate is expressed asf, Hz and the bandpass sig hal is located between f, Hz and f, Hz. The signal band: ‘width B is less than J — f., and the positive frequeney bband can be expressed asthe interval (i. ‘The band position refers to the fractional number of, bandwidths from the origin at which the lower band edge resides. A special case is integer band positioning, which holds when the band is located at an integral number of Manuscript esened Jl 5, 1989 revise Jane 24. 980. RG, Vanghan and 8. Soot as wih Inara Development, DSI Laer Ha, New Zealand LR. White ath Pail Sckaces, SIR, Lower Hut, New Ze TEE Lng Number 910145 Fig. The bans sation a analog signal pect, Te sampling rs ef bd ef a ig. 2, Example fa) integer band pstoning for ¢ = 3, an () bale incger bond psn for Ipandwidhs fom the origin, i... fi = elf. — fi) sti 42, sand ¢ = isthe low-pass ease. Fig. 2(2) indicates integer band positioning, fore = 3. “Another special cases alf:integer positioning, when f= (e+ 1/21, ~ fd). Fi. 20) shows an example SF halt-imegr positioning fr = 2, These cases of and positioning are useful for illustrating aspects of aliasing, nd are particulary important in the application of sam pling tothe fective relocation of signals betwecn Band- pss and low-pass positions “The classical bandpass theorem for uniform sampling states thatthe signal ean be reconstracted if the sampling Tat is at east" = 2f,/m where isthe largest integer Within f,/B. denoted by’ = Zl f/Bh “his minimum rate for enifor sampling is of interest from a theoretical viewpoint, but for practical applica- tions this ate is often presented ina misleading way, Fig 4 fom Feldman and Bennet 16} ls reproduced in se eral texts (e.g., {13]-[15]), illustrates the minimum uni- form sampling rate for bandpass signals. The theoretical rninimum rate, ~ 2B is seen to apply only for integer fund positioning. The skew lines represent the minimis Union rate required when the passband isnot located in fn integer poston, The vertical lines, which represent 3 10. 587X/91/0900197350.00 © 1991 IEEEFig. 3. Minium sampling Sequeney fr band of with 16) Liscontinuity, traverse disallowed valves ofthe sampling ‘frequency, a point elucidated in Section II, Grege’s [17] version of Fig. 3 displays more clearly the discontinuity in the function by using dotted lines for the vertical sec- tions. = 2Bisa valid uniform sampling rae only os [2atrb = ft = rx BL San S00 if there is no signal component atthe frequencies fF f This i analogous to the anomaly ofthe low-pass case (ct Shannon's theorem) where the equality inf, = 2f, can hold only if there is no signal component at f., ce. the signal is band limited to less shan B and the signal occu pies the interval (0,j,).! Kida and Kuroda {18} discuss the disallowed signal components for higher order sampling ‘The theoretical minimum sampling rate i pathological inthe sense that any engineering imperfections in an im plementation will cause aliasing. This illustrates the need to sample at above the theoretical minimum rate or, alk tematively stated, guard-bands need to be included. Gregg [17 states without elaboration that rates between /"" and 2f, (the low-pass case fate) are not, in general, valid ‘These rates correspond to the region above the function in Fig, 3. Nevertheless, some authors mistakenly pret figures such as Fig. 3 to mean that no aliasing will ‘cur as long asf, > 4B. Gaskell (1) gives an expression (ce (16)) forthe valid rates, but otherwise there is litle information regarding bandpass sampling at rates above the theoretical minimum bot still well below the frequen cies of the signal being sampled. In Section I, this sub- jet is discussed and results are depicted graphically. The existing information is clarified and the cases for band- pass sampling at above the minimum rate are reviewed ‘The discussion includes sensitivity of sampling rates and offers a practical guide to bandpass uniform sampling ‘The theoretical minimum uniform sampling rate fo = 2B can be applied only when the band is integer posi- tioned. Kohlenberg’s {10} second-order sampling theorem allows the minimum rat, in the form of an average rate, The ip ce ic encepn nha he component tf recovered because it contains trivial phase information, ie cos [infut — rx BA} Debi sin FBR 0s [2n(f, + By ~ (e+ xB] ~ cos [2xtrB ~ fit ~ or + eB) BeBe sin” + DyxBk to be applied independent of the band position. The tech- nique is second order in that two interleaved uniform sam- ple sequences are applied. Each sequence is of rate B (i.e, sampling every 1/B's), and the sequence separation is denoted by ks. Using the samples to construct signal f() at (ffs the interpolation is given by E/@)s(-§) v(Ben)s(ae2-d] o whieh p isthe sample number index and the interpotant fo StH = Sun + $0) ® where @ 4 and ris an integer given by the band position such that 24, 2 2h cy < hy ‘8 Berchet o ‘The righthand side of (5) can be expressed in terms of the center frequency Wee Baie © Geometrically, there are at least r ~ 1 bandwidths be- swoon the two-sided spectrum, i, between ~f and. The interpolant has the properties that 80) = 1, Sip /B) = Oforp + 0, and Si p/B- + k) =O, and in genera it has extrema which are larger than sity Tn the general case when the equality of (5) does not hod, the valve of Fmay not take om the ciserete vals KBr, KBr + 1) = 01,200 o for which the interpolant blows up. & is otherwise unre- stricted, but implementation difficulty using finite word- length machines is affected by the choice of &. This sen- sitivity is discussed in Section IIL ‘When the equality in (5) holds, the first term of the imerpolant, S(0, is zero, and the restrictions kBr # 0, 1,2, «++ on & are removed. The situation now corre” sponds to cither integer band positioning (r even) or half imeger band positioning (r odd) and for these conditions there are exactly r bandwidths between ~f, and +f.” Three special cases are noteworthy:i) The low-pass case isfy = 0 and r = 0 for which the interpolant reduces 10 08 (2x BI ~ Bk) ~ c0s Bk 2x Br sin x Bk Salt ® ii) The special ease of k = 1/2B corresponds to uni- form rate sampling, wit its well-known sine (2x Bi) in- terpolant, which, as mentioned above, requires integer band positioning. i) The quadrature sampling case has Bt mabe o ‘and the simplified interpolant sin xB Sol 08 2a a0) ‘Quadrature sampling is important because, as the name indicates, the in-phase and quadrature components are sampled explicitly from the bandpass signal. This is seen by denoting the bandpass signal in the usual way’ £0 ap where 1) and Q(e are quadrature components ofthe cen ter angular frequency w, rad/s. "To extract Ii) directly from f(0), ie. He) 008 wt ~ QC) sin wt cs 0 requires sampling at times a Spa pn en bay PaO bd 3) Simitarly, ro obtain Qt) age art nd ie, gt ay whence Ftp) = — U6. as) The signals 1(¢) and Q(t) are band limited to the interval (0, B/2) so that each stream requires sampling at a rate B. ‘Grace and Pitt [3] drew attention to this special case of, Kohlenberg’s sampling scheme. Rice and Wu (19) ap- plied the special ease of uniform sampling combined with the band being integer positioned, which is a form of ‘quadrature sampling in that k = 1/2B = 1/4. + m/2f. where m = 2/r(= 2 for integer band positioning). They used the Fact that 0) and Q(0) are a Hilbert transform pair to derive the required interpolation weights used for time aligning the samples. Their innovation was to "decimate by two" the samples before applying the Hilbert trans- form interpolator, Waters and Jarrett [4] implemented {quadrature sampling using the m = 0 case and also used decimation before the interpolation filter. Jackson and. Matthewson [2] also discussed implementation of quad rature sampling ‘The spectrum of a sampled signal is periodic. When: ‘ever a band-pass signal is reproduced at a baseband po- sition by sampling, the noise from all the aliased bands is Combined into the baseband, Even with an ideal antialias- ing filter, the signal-to-noise ratio (SNR) is not preserved for bandpass sampling owing to (postilter) thermal noise contributions from the aliased spectra. This has clear im- plications for communications and radar receivers where bandpass sampling is applied to relocate the signal effec- tively to a low(er) pass position. This is in contrast to analog complex mixing (.e., using an image rejecting mixer), in which the signal-to-noise ratio is ideally pre served, The degradation in signal-to-noise ratio does not seem to have been discussed (0 date, despite its ubiqui- tous presence in less critical applications, such as sam: pling oscilloscopes. Section IV describes and quantifies the signal-to-noise ratio degradation, with a simple ex- periment to demonstrate the effect. Also, the aliasing noise resulting from using typical bandpass filters is presented in terms of a signal-to-aliasing) distortion ratio. This es- timates the effect of applying “acceptable” sampling rates ‘with practical, instead of ideal, rectangular bandpass fil tess TL, Untrors SamPuine ‘The conditions for acceptable uniform sampling rates can be written [1] anzi(f venei (fl a ‘The equality conditions of (16) hold only if the band is ‘confined to the interval (f;,f,): # frequency component atf =f, orf = f, will be aliased, as discussed in Section Tang by Kida and Kuroda (18. ‘Equations (16) and (17) are depicted graphically in Fig 4. The sampling frequency, normalized by B, is on the ordinate, and the abscissa represents the band position J.B. Fig. 4 is a corrected and extended version of Fig, 13. The areas inside the wedges are the allowed zones for ‘sampling without aliasing. The shaded area represents uniform sampling rates that result in aliasing. ‘The low ‘pass case f, = 2f,, given by nm = 1, corresponds to the large wedge to the left of the figure. Each wedge across the figure corresponds 10 a successive value of. “The theoretical minimum sampling rate f, = 2B, cor responding to integer band positioning, occurs at the of the wedges, and these rates are clearly dangerous points for implementation, since any engineering imperfection ‘will move the sampling rate into the disallowed area. In ‘a sampling system, the allowed sampling rates are repre sented by a vertical line within the allowed areas and above the band position given on the abscissaFig. 4 The slowed and sslowed (hades 2) wiform sampling es vere the band postion. The aurea extension ‘fig. 3, to show the paca afeas a nominate sampling tats. The OEM and oie Ye tal an and (of Sper an ower paras) fe the sampling ate, Bethe band andthe Ran ose ff. Brown [20] has pointed out that for symmetric double sideband signals, the spectra can “fold over” each other without loss of information, allowing a theoretical mini- ‘mum sampling rate of 1/8 fora band-limited signal where the total bandwidth (both sidebands) is B. This case would have an allowed operating ‘‘region” in Fig. 4 of straight lines bisecting the disallowed areas between the parallel borders. The case also represents a pathological condi tion, because any sampling rate variation will not only ‘cause aliasing, but will also cause the signal to become folded incorrectly with a likely loss of otherwise retrev able information, Sampling at nonminimum rates is equivalent to aug: ‘menting the signal band with a guard-band. In fact, any allowed operating point away from the tips of the wedges of Fig. 4 can be interpreted as having a guard:-band be- tween positions where the spectrum will be aliased. The total guard-band is given by Bor = f, ~ 2B He, which is labelied on the right-hand ordinate of Fig. 4 To find the relation between the guari-bands and the sampling frequency, the order ofthe wedge n is required This is found from the overall bandwidth (signal band- Width plus total guard-band) B+ Bor aig) and its location (fi, fi). where Si =f. ~ Bau (9) fom f+ Bow 20) and Boy, and Bo, are the lower and upper guard-bands, respectively, with Bor = Bo + Boy. The lower order wedge is then given by ew Fig. 5is an expansion of Fig. 4, showing the mth wedge with the guard-bands and the sampling frequency toler- The allowable range of sampling frequencies is 2f 2 STW Mf @ land this range is divided into values above and below a ‘nominal operating point (see Fig. 5) AM = Mle + Aha @ giving the lower and upper guard-bands as ey Bow = Mh > es) Note that symmetric guard-bands imply asymmetric sam- pling frequency tolerances. However, the asymmetry de- {reases with increasing If the operating point is the vertical midpoint of the wedge, then the sampling rate is 1(hr ML) a and the guard-bands becomeco Fig, 5. Peemetersin the mh wedge fom Fig. 4. Bo, an Bo Sheu od fc pn va ped Bo. = Mf, en Box = MG 8) [An example illustrates the use of the equations. If @ 25-KHe signal is allocated symmetric 2.5-kHz guard- bands 50 that the overall band is (10.7, 10.703 MHz). then n” = /,{10703 /30} = 356 and the sampling rate has limits of 2(10703)/356 = 60.130 kHz to 2(10700) /355 = 60.282 kHz, i.c., an allowable range of Af, = 152 Hz. ‘The theoretical minimum sampling rate (for the signal without guard-bands), corresponding to a zero allowable Sampling rate range, is on the lower edge of the, mth wedge, where m = [,{10702.5/25] = 428, and is f°" = 2(10702.5) /428 = 50.017 kHz. ‘Note, for this example, that in moving from the theo- retical minimum rate to @ rate which includes the guard- bands, a total of 428 — 356 = 72 separate ranges of dis- allowed sampling rates have been traversed. This illus: trates the pitfall of arbitrarily increasing the sampling rate from its theoretical minimum value for band-pass sam pling. “The relative precision of the minimum sampling rate requited t0 avoid aliasing clearly increases with the in “reasing separation ofthe band-pass signal and the origin, ice, increasing n. If an operating point is within the nth wedge, above (ie., infinitesimally t0 the left of) the tip ‘of the (x + L)th wedge, then the difference between the Jnimum and maximum allowed sampling rates is _ the @ so that the relative precision required of fis he 1 (hn) wo(2 Hemi) © represent guan-bands api sliasng ote spl ocupyng sed Gonnmend ange ofthe sampling equa fom 8 nmi i.e related to the inverse square ofthe separation of the band from the origin, (For the above-mentioned points. f,/ B~ n= 1.) For the example with n = 428, the rel- ‘tive frequeney precision required of the sampling rate is 1/1428 * 427) ~ 5 ppm. This sensitivity corresponds to the vertical Locus within the 428th wedge, above the tip fof the 429th wedge. Moving to lower order wedges (Gmaller n), while remaining above the tip of the 428th wedge (same f,/B) will relax the sampling oscillator pre~ cision requirements and increase the guard-band, but of ‘course requires higher sampling rates. “The relative precision required of the sampling oscil lator is depicted in Fig. 6 as the line 1/(n(a.~ 1)). This line gives the relative precision required of the sampling rate when operating with f,/B ~ m = 1 as a function of zn The positions of a typical crystal and RC oscillator are indicated to give a feel for practical requirements. With the inclusion of larger guard-bands by operating in the various mth wedges, the line interpolating the various Af [2B = 1/n'w’ ~ 1)Jy/B~ n'y runs above the fa/B ‘n= 1 fine. For example, for n’ =m + 10, the Line would be an onder of magnitude in 4f,/2B higher than ‘and parallel to the f,B — n = 1 line in Fig. 6. Varying. an operating point within & given wedge corresponds £0 8 positive slope locus in Fig, 6 which steepens as f,/B de Ereases. After intersecting the f,/B — n = 1 line, the Tocus approaches a vertical line at f,/B =m as Sf, ap proaches zero at the tip of the wedge. ‘The f,/B — n = I line represents operating with the ‘minimum allowed set of sampling rates and the maximum available guard-band. The guard-band size decreases with increasing m, and the line is ~O(1/n'). If the relative fuard-band is now held at a constant size, then the rela- tive precision of the sampling rate becomes, from (24) fand (25), Af,/2B = Box/B + 1/n'. This demonstrates the tradeoff between the precision required of the sam pling rate and the guard-band size. For practical situa~ ions, a modest increase in the sampling rate permitsELIE PRECSON F SAMELNG FREQUENCY Fig. 6. The eave preiion of the of minimum allowed union san Pile rc Te ine am merlin of he operating poss coespond [tet fjf) = Invemof Fig a tee operating ps re wi IME mikwedge ot postion svete up of the Ii wedge, realizable precision requirement, as shown in the example above. In addressing sampling rate variation, it must be bome Jn mind that so far the discussion has been limited to avoiding aliasing in the sampling process. Exact (fe)construction of the bandpass analog waveform re- quires exact implementation of the sampling parameters used in the interpolant computation, a point addressed within the following section, TI, InreRPOLATION IN SECOND-ORDER SAMPLING Koblenberp’s second-order sampling result gives a the- retical minimum average sampling rate which would correspond to the straight line f, = 2B on Fig. 4. The interest in second-order sampling. has been not only from the theoretical viewpoint of maintaining a constant mini- ‘mum average sampling rate independent of the band po- sition, but also from its practical application in the sim plified form of quadrature sampling a bandpass signal. ‘This section concems aspects of bandpass signal con- struction using various k: specifically, the required range (number of bits inthe coefficients) and length (number of coefficients) for realizing the interpolant as a finite im- pulse response filter, and the sampling rate sensitivity ‘The application is the generation of a band-pass signal from second-order samples. For second-order sampling, the spectra resulting from cach of the two interleaved sampling sequences applied to the bandpass signal are (21] FAS) = FU) * EBS — pB) ep FAS) = FU) Ze PF ~ pH) 2) where F(f) is the analog spectrum, # denotes convolu- tion, and p isa spectral position index. From here on, the explicit frequency dependence of Fy and Fy is dropped Equations (31) and (32) show that the sampled spectrum is a series of analog spectra located B Hz apart and scaled by B. For Fy, there is a progressive phase change of 2x Bk imposed on F(f) at successive spectral locations, caused by the sampling stream being delayed by k seconds. Fig. 7, which is an elaboration of Linden’s [21] Fig. 5, illustrates the geometric aliasing relations. There are ‘wo regions in the band in which the aliasing arses from different aliases of the negative frequencies from the ‘analog spectrum. These regions correspond t0 where the ‘So and §\2) terms of the interpolant (see (2)) are ap- plied, and are Rs fsfsB—h 6 and Ri Bf sfeheB oy, Note that from the choice of definition for r, Ry cannot tencompass the band whereas Ry can encompass the band, whence S(#) = $\(0), ie., integer or hal-integer band po- sitioning ‘The spectra resulting from the two sampling sequences Ri R= FN + Pf) 38) Fy = Ff) + eo F-fi) 60) and Rs R= A+ Ff) o Fea Rpt ePr-fy G8) where Ry Oy = wBkr 6 Re 8 = x Bk + 1) 49) and Re aft 2B ap Re flash f+ DB. a) ‘The scaling factor B in (31) and (32) is omitted in (35)- (38) and hereon. The aliased contribution (the term containing F(—J")) is the same for F, and Fp, except for a phase shift which is constant throughout each region. The phase shifts are, however, different for each of the Ry and each of the Ry in the aliased spectrawipe Fig. 7. The geometric relations fo aiasng in secomd-nderbaps sampling and eect Bnd poston. FL) he analog ‘pEsnam sl sf) an al) te he pets mca othe meen oy 4) norm sample wghenes The sae ‘Sadr ar protests ps sede to the analog hand ax india the maton = 2” For ge an tl Incest stoning he Teion Ay Psomes 20 ‘The bandpass signal construction is (ef. (19) Ry PEF) 43) Ry a) so the disallowed values fork are Re: Be m= O41, 22, as Ro ke m=O, £1 820 be Oe ae UO) Which are unique excopt when & = 0, 1/B,2/B, => * + ‘whence the sampling sequences coincide In Fig. 7 the phase factor notation ¥ an is used for brevity. Considerable insight into the inter- polant can be gained by expressing it in terms of the fe- fqencies in Fig, 7. Using the angular frequencies «= Zales, = Defi op = 2a Band wy = 2x(7B ~ fe which is the common border of Ry and Ry inthe analog spec: trum, and noting that Ro: arg (y’) = 20, Ry: ake." *") £08 (aut — 4) — 608 (wit = 0) ‘oat in By as) 20s (at = 0) + 608 ut — 8 iggy “oat in 8, Which shows the composition of the interpolant in the same Form as the low-pass case of (8) [As already noted for integer and half-integer band po- sitioning, 2f,/B = rand Ry becomes zero. Further sin plification in the signal construction occurs if e 7 e., for quadrature sampling, in which ease Ff) = s(Fa + Fa). (30) This occurs when D2akBr +1) = 2jOm+ er SD an +4 ‘ 2B + 1) oa Lim ate 3) singe f. = (7 + 1/2)B, so that these values of k indeed correspond to quadrature samples. 'As approaches its disallowed values, construction of the bandpass signal becomes more dificult to implement sinee the interpolant becomes larger. Both the dynamic range (number of bits required for the filter coeficiens) ‘and the length (number of filter taps) required for the in- ‘erpolant contribute to the cost of an implementation. The ‘general interpolant form (2) predominantly resembles a ‘modulated (by .) sine function, so the dynamic range and the length are related, although the relation is not simple. "The dynamic range is & minimum for quadrature sampling. ‘when the interpolant assumes its maximum valve at (0, as evident in (10), In this case, te filter length is also clearly defined for a given coefficient size as long as f S> B, ie., for the narrow-band situation. When & de>viates from the values corresponding to quadrature sam- pling, both the dynamic range and the length of the inter- polant increase, ‘The low-pass case illustrates the increase in interpolant length with a poor choice of k. Assuming integer band positioning, the low-pass interpolant can be written, from, (8) im the form sin? = Br sin 22 Sul) = eo a Bk a 4) ‘The first term decays as 1/Br whereas the second term decays as 1/2" Br. For a given dynamic range and coet- ficient word size, the magnitude ofthe first term should be minimized, of atleast kept small with respect to the second term, for minimum biter length, Consequently uniform sampling where k = 1/2B (corresponding to quadrature sampling since f- = 8/2), is optimum in this case. For baseband interpolation of & sampled bandpass Signal, configuring the ¥stem such thatthe signal is in- teger positioned and applying uniform sampling is the simplest arrangement. In an implementation, i should be bbome in mind that half-integer band positions alias to be spectrally inverted a the low-pass position Ina nonuniformly sampled lowpass system, the alter: native to applying the interpolant of (S4) directly is to generate uniform samples and interpolate with sinc 2B ‘The required resampling is shift of one of the sample sequences using the sinc 2 Br interpolant on alternate samples Further expense inthe implementation of bandpass sig ral constriction may arise because ofthe need to include ‘guard bits” in the sampled signal and the interpol coeficients. This is because errors inthe measurement oF generation process of F, and Fp, such as those due t0 (quantization, may in tum create errors in the recon structed signal. The effect of the choice of & on the re sulting uncertainty in the constricted signal is of interest, and is estimated here using a sensitivity analysis in the frequency domain. The specific aim isto get an idea of the cost of poor choice ok Progress can be made by assuming that the variations of corresponding (same analog frequency) components in FF, and Fy are uncorrelated. The variance of component in Fcan then be writen in terms of its magnitude: arf iF, where 03, and of, are the variances ofthe components of FF, and Fp. For correlated variations, caused for example by k becoming very small, #3 will increase, so (55) rep- resents a best case. The units of of are waits /her and the noise power i the constructed signal is found by in- tegrating over the band of F. Letting o? = of, = of, and using (43) and (44) in (55) and integrating the nose power in F is ae) of 69) 66) where 2h, B denotes the extent of the Ry portion of the band and 0 = Ar < 1, The minimum noise power is ¢°B/2, and from the choice of definition ofr, occurs for Ar = O and x/4 + mr,m=0, £1, +2, «++ 5 be., for imeger or half-integer band positioning snd quadrature sampling Por general values of Ar and @, 8; the noise power relative to its minimum value can’be expressed as the umber of bits of resolution irrecoverably lost in ealeu- lating F; viz., ar on 1, = bog (3) (58) a 1-4? nme (gt a)" om In terms of an implementation, (59) indicates for poor choices of k, how the errors in the sampled data magnify during the construction of an analog bandpass signal. Ny is plotted against & normalized by B in Fig. 8(a) for the example where r = 9 and Ar = 0 (half-integer band po- sitioning); and in Fig. 8(0) for r = 10 and Ar = 0.5. The functions are symmetric about k = Qand & = 1/2B. The asymptotes correspond to the disallowed values if; viz.. k = m/10B, m = 010 10, for Fig. 8(a); and both 1m/10B and k = m/11B for Fig, 8(b). For the half-integer band positioning case, the quadrature sampling values eae cin, iy, as CE 3 SS ste tern ieee haley me ea re Sl fa peo se eat abot ey She tae mi tm secon i Gp imatg a Reb oa see ee sores Sei nee es san ie St sd 7 ea Cease sy cosy emegceen ge a Bearers ence cara wasn ear 1/218, 3/218, j and these do not occur at the func- iS onan ee is ee se Em ego Secu ean epi T UVV [* o “) Acro) ieee 1 fi Hea cn Li eed bres, where Oy = —akir + DB. For small Ak fy = Fy + RD in 28; which approximates how the degree to which the alias is Suppressed depends on the difference of kand k,, The rel lative difference between the constructed spectrum and the wanted spectrum can be expressed as ® Fig. $. Anestinat ofthe equivalent number af bits of elton iow ‘a om in nerpoatings bandpass sing uoeenain sean rae ‘Stents unsuon of The grap repent cos ofa por ehoke Sits sson over bandas sumpbing process. The average sample heir (fa = 3a) dy ar =O hlatege band poutonn). The ‘nim concspond i aes guaratue sampling.) r= 10.2 = 05 (ee ger nr lege nd ostonng) The rinima So (Shor, conespod quarreling inthe signal. However, the first quadrature sampling value is close tothe frst minimum of the function, which isthe global minimum. In this case of general band positioning, fptimum quadrature sampling requires the minimum k, fe. k = 1/4f.. The best general value of k occurs at the first funetion minimum (or the last minimum, which amounts tothe same sampling sequence interleaving). and £0 the configuration of the optimum general sampling sys- tem requires finding this minimum. This disadvantage, a8 ‘well as the need here for a complicated interpolant, make it worthwhile arranging the band to be integer or half= ieger positioned. “The accurate construction of the bandpass signal from its samples also depends on the accuracy of the sampling, rate and k. Ifthe sample sequences are interleaved by ky ‘and the interpolant implementation Features k = &, + ky then the constructed spectrum of interest F(f) and its al~ is F(-f') in terms of the wanted spectrum is, expressed here for R only, Auf) = Ff) are = Bes pak aan Le adit 4 ret which i also readily obtained fom time domain donsidertons MA inoreases, the timing err Ak becomes increas: ingly etal Fr lage values ofr pertinent to commu tintin or radar, ete demands on 3k become con ‘Mrae For the previous example i Seton 1, wing Suintbandsy with r= 336 and B= 30 KH, and f= {OT Mite and an alias suppression of 40 dB, there Guired aogarney of the sample sequence interleaving ace eee = 357 | 730(10)° ies Ak ~ 0.318, Forloner intermediate frequencies in communications, such as 485 kHz, the requirement is correspondingly simpler, viz., Ak ~ 7ns (r= 13) ak IV. Notse is BANDPASS SAMPLED SIGNALS [Noise considerations are important not only where the signal-to-noise ratio is important, such as in communi tation receivers, but also in the measurement of nose it~ Self, such as in noise power measurements. In a sampling System, the periodicity of the spectrum means that wide- band noise, such as the thermal noise introduced by the relevant hardware, is all combined into each of the fs/2 bands. ‘In applying bandpass sampling to relocate a bandpass signal to a low-pass position, the esulting signal-to-noise ratio will be poorer than that from an equivalent analog, system (ie., an ideal image-tejecting mixes), in which the signal-to-noise ratio is preserved. ‘Consider a system with a bandpass signal of spectral power density 5, in-band noise power density of Np, and Satvofsband noise power density N,. The analog signal- to-noise ratio is thus S/N,. The signal-to-noise ratio for the sampled signal becomes degraded by atleast the noise aliased from the bands between de and the passband, and. is thus SNR, = 6) pe DN,Often, Np >> Ny and the signal-to-noise ratio is estab lished before sampling. However, if Ny = N. and assum- ing n >> 1, then the degradation of the signal-to-noise ratio in decibels is simply Deyn = 10 log 6) Equivalently, the effective noise temperature is increased by at least. Note that this analysis assumes a flat analog noise spectrum, ideal filters, and an infinitesimally small sampling aperture. For the example with B = 30 kHz located atthe integer position n = r/2 = 30(f, ~ 455 kHz), orn = 356(f, ~ 10.7 MHL), then the increase in the noise power is at least 15 and 25 dB, respectively. In the case where the noise spectrum is not uniform, 2 conservative estimate of the noise power density in exch of the bands in the sampled spectrum can be obtained by ‘equating the noise power before and after sampling. De- noting Neq as the equivalent noise spectral density (ENSD) and Bey as the equivalent noise bandwidth (ENBW) of the analog signal; and Ngs as the ENSD of the sampled signal with 2 the sampling rate (itis known [22] that the ENBW of a sampled system cannot exceed B), then NeaBea = NesB. 63) ‘The degradation in a signal-to-noise ratio is Nes/Nea which, in decibels, is then at least A simple experiment using a noise source and a sam pling oscilloscope demonstrates the effect. The test signal ‘was 10.95-kHz carrieramplitude-modulated by a200-Hz sine wave. A Bruel and Kjaer 0-20 kHz noise source sig- nal was added to the test signal. The resultant signal was sampled at 125 kHz, windowed using a minimum 3-term Blackman-Hartis function, and the spectra calculated by FFT on a LeCroy oscilloscope. The average of 10 power spectra is shown in Fig. 9{a), The signal-to-noise ratio is about 38 dB, The signal was then sampled at 1250 Hz giving a Nyquist bandwidth of 625 Hz, and m = 20 000/625 ~ 42. The average spectra is shown in Fig, 9¢b), with the signal-to-noise ratio now being ~21 dB. ‘The observed degradation is thus about 17 dB. ‘The minimum degradation given by (64) is 10 log 32, i.e., about 15 dB. Recall that this minimum is given by considering only the aliased noise from the noise source, and assumes that Bg, is 20 kHz, while in practice itis @ little greater. In general, the potential for signal-to-noise degradation in bandpass Sampling systems is considerable, even when the analog spectrum contains only thermal noise. In most communications systems, for example, the out-of-band power (adjacent band signals, etc.) has similar, or even higher levels, than that in the band of interest, so am anti- aliasing filter is necessary To get a feel forthe filter requirements, the signal dis- (66) o waa oar a ae sw » Ines by 200 sno a he presence of white noise The Sampng fai 0) 125 Al, gvng a supe pec SHR of aout 38 a) tortion ratio (SDR) [24] is 2 useful metric. It is the ratio ‘of the mean-square aliasing error and the total signal ‘power passed through the filter and can be interpreted as ‘a signal-to-noise ratio, It is here defined for the bandpass case [23] as [mere SDR = =a 5 \H(w)? do hein later d+ f where H(a) is the filter transfer function and fis the arith- metic center frequency ofthe upper and lower 3-dB point, ‘The limits ate chosen so that the wanted signal band oc cupies a Nyquist bandwidth, In practice, this requires se- Tecting f, such that the band is integer or half-integer po- sitioned ‘This signal distortion ratio offers only a rough guide for trading off the filter requirements against sampling rate and is only one of many factors contributing to the effec~ tive signal-to-noise ratio [23]. This is because of the as sumption in (67) that the signal power at the filter input is constant across the spectrum. In practice, the presence of guard-bands will normally ensure this will not be the «ase atthe edges of the filter. On the other hand, adjacent channel power levels may be tens of decibels above the power level in the channel of interest, In radio commu nications cases, the filter H() not only represents the antialiasing filter, but also the receiver front end, includ- ing the antenna. Nevertheless, the antialiasing filter will ronmnally be the dominant effect on the signal distortion Fig. 10 [23] shows the signal distonion ratio for two filter types against the sampling rate normalized to theFig. 10, The sina sintoion ratio SDR) as dtd forth ands cast Forscson-oder und eiphvorder Baers nd Cheyeney ites. The Envelope edges forech cure preven @ = and Q= 200 (3) 3.€B bandwidth of te filter Bs yy, Because ofthe selection off; mentioned above, there are only discrete points on the curves which are valid. The envelopes enclose the curves for filters with Q values from | to 200. The param- eter Q = f./Byyn is related to the band position parameter r= 2f,/B (for imeger band positioning) by Q = 4(r + 1), In the example with a 25-kHz band ut 455 KHz, ic., 1F-= 36, a signal distortion ratio of 37 dB requires a sam- pling rate of f./Byey = 2 (if, ~ 50 ksamples/s) with a filter equivalent t0 an eighth-order Chebychev with Q = 16. ‘The definition of the signal distortion ratio in (67) perhaps be more easily related to a signal power to" tortion power" ratio ifthe integration in the numerator is ‘over the 3-dB bandwidth only, -e., the denominator cor- responds tothe signal power only rather than total power. For such a definition, however, the curves of Fig. 10 ‘change by less than a couple of decibels. Strictly speak- ing, only interleaved portions of the filter ‘'stop band” contribute to the aliasing distortion, but again, this defi- nition does not appreciably shift the curves of Fig. 10. This matter is discussed in [23]. In applying Fig. 10, it should be borne in mind that the signal distonion ratio is ‘only an approximation to the effective signal-to-noise ra tio caused by the aliasing V. Conctusion For practical bandpass sampling, there is a need to sample at above the theoretical minimum rate. Care must bbe taken when increasing the sampling rate from its min- imum value because there are bands of rates for which aliasing will oceur. Operating at a nonminimum sampling rate is equivalent tothe introduction of guand-bands or a corresponding tolerance to variations inthe sampling rate ‘The relations between the sampling rate tolerance, the guard-band size, and the various allowed nonminimum sampling rates give practical design choices. For sampling an analog bandpass signal in order to pro: ‘ess digitally the signal, the simplest system is 10 com- figure the band to be integer positioned and apply uniform (hich is also quadrature) sampling. In the sampling process, the signal becomes degraded ‘owing to aliasing of the noise introduced after any band- pass antialiasing filter) berween de and at least the original bandpass spectral postion. This degradation is unavoid- able and a simple estimate for the minimum degradation is available in terms ofthe band position For constructing an analog bandpass signal from its second-order samples, itis wise to configure the system. such that the band is integer or half-integer positioned. ‘The advantages are a simplified interpolant and the poten- tial to apply quadrature sampling with the flexibility to tse several values for the sequence spacing. Quadrature sampling represents the optimum sample sequence inter- Teaving inthe sense that it fequires the shortest filter length (least number of taps) for interpolant implementation using a given coefficient word size, and displays the least ‘sensitivity to error propagation in the construction of bandpass signal from its samples For cases where the band is not integer of half-integer positioned, the optimum sampling sequence interleaving for bandpass signal construction is not trivial to find Here, quadrature sampling spacing should not be applied in general, owing to possible error propagation from im- perfect samples, although the minimum quadrature spac- ing k = 1/4f, i usually close tothe optimum. REFERENCES 111 J. Gaskl Lice htm, Fore Tramfoms and Opis. New Yorks Wiley 1978 {2] Me lucia and P, Mathewson, “Digi processing of andpuss Niemi” GEC esol fr m1 1986 19) 0D Grace and Pi “Quadrature sampling of high equency Steveionmee= acon Soe Amer vol App 82-1636, 1968 ay We Wai and Bt Jaret Bandpass ial sampling ar oe Trem deo, EEE Trans teri. Elec, st v0: AES-TR for Now 1 Is] By Del Re, "Bundpss sgn tenes and reconsiction though Pinimamsanplingae dpa Process.” ala Frequnca, sl Xivin ne 9. Sep 1978 lo) Fels J. Gat and 1 Re Stockton, “Ponies and thon of wat fst pet nh nel ps Ph 7) HS. Black, Modtonon Peors New You: Van Non, 1983. {al Nyquist, "Carin topice ler Ussing A0EE Tr el 47.198 19) Bi hoya smmen (10) A Rotenberg. “Eas iteration of bandied fonetions. "J 111) $°Golaman,Pjomarion Theory Peace Hal Elric Engines ing Seren” London: Conable and Co Li. 1983, 112) Be dion, oducion to Sista! Communication Theo New York: Mistral 160 11a) PTF Pamor, Medan, Noe and petal Anal, McGraw 1063 {14} Stank, ConmmicatonSputemt, New York Wiley. 1978 {13} K. 8 Shanmugam, ig and Analogue Communication semi New Yor: Wiley. 199, 16) O° Fatman aad WR. Benne “Bandwiih and aniston feifomanaee Bel St Tech sel 28, p. 490-95, 198, tun WD" Gaeg Arlog wad Diptal Commantcrion Systems, New York Wigs a le. Eng. 98[U8]. Kila andT. Kura, “Relations tmuen hepsi of re (9) BWR a Wa. naming wth ih pani (20) PB noe for somping apne mdse ig [211 DLA. Linke, "8 discussion of sping eve. Pro 1,0. 22) D-H White, "The ae nsf spl date pts," EEE Pri Iran, Meat Sh mo pp We 1083 98, 1 NL so ad Rs Vaughan: the sal dsonion Stes {ie wo alining Pr. tt Spinal Provesing a AP ‘om: HSSPA WO) (Geld Coast Rasta), Ase. 180. (2s) NIU"SSecand RG. Vase. The signal dation ratio fen plcaions (S$P490)(Ootd Cons Austen Aug. 19900 Rade, Vaughan 8) cee eB PRD. ep rom Aube Uneniy.Denmar Puerco sramen for ibortnyand ‘as, arrays. multiple path propagatsn, and digital signal processing. He {Xion ijt Leet Communications Techagiog) at DSI Neil, Scot ese be BE and MF. (le {rel depres om he Unicity of Comer Itsnimtaton, He ton moved he Depa ‘roel on image prcenng. nieactve sean ‘el sig procesing Hieron wok hasbeen tine Svs smal and te appaton of DSP ted White ese Se deena Wala Hamion, New Zand a 190, Standards Section ofthe Depainent of Seenic Shu adsl Reteaeh i ower Ht New Zam ian is inser ane nthe Beds of tee