A Digital Signal Processing Approach to Interpolation RONALD W. SCHAFER ano LAWRENCE R. RABINER Abstract—in many digtal signal processing systems, et, ‘ecoders, modulation systems, and digital waveform coding eyetrte, itis necessary to alter the sampling rate of «digital signal. Thus itis of considerable interest fo examine the problem of interpolation of ‘bendlimited signals from the viewpoint of digital signal processing. A frequency domain interpretation of the interpolation procees, through which iti clear thet interpolation ia fundamentally « neat ‘Altering process, is presented, ‘An examination of the relative merits of Anite duration impulse response (FIR) and infinite duration impulse response (IIR) digital ‘Alters as interpolation filters Indicates thet FIR filters are generally to be preferred for interpolation. It is shown that nea interpolation snd classical polynomial interpolation correspond to the uae of the FIR interpolation filter. The use of classical interpolation methods in signal processing applications is iluetrated by a discussion of FIR {aterpolatin filters derived from the Lagrange interpolation formula. The limitations of these Alters lead us toa consideration of optimum IR filters for interpolation that can be designed using Hnear pro- cramming techniques. Examples are presented to luatrate the tg~ ‘icant improvements that are obtained using the optimum fitere 1. Ixtropvcrion HE PROCESS of interpolation is familiar to anyone who has had occasion to *read between the lines” in a table of mathematical functions. In digital signal pro- cessing, interpolation is required whenever itis necessary to change from one sampling rate to another. For example, in speech processing systems, estimates of speech parameters are ‘often computed ata low sampling rate for low bit-rate storage oF transmission; however, for constructing a synthetic speech signal from the low bit-rate representation, the speech pa- rameters are normally required at much higher sampling rates [1], [2]. In such cases, the sampling rate must be creased by a digital interpolation process, As another example, an efficient digital realization of a frequency-multiplexed single-sideband system has been obtained (3] by performing complicated filtering functions at a low sampling rate and simpler functions at the high sampling rate required for srouping several channels into a frequency-multiplexed for- mat. In this process, there is a need for both increasing and decreasing the sampling rate. Another example where sam- pling rate reduction is required is in converting a delta modu- lation representation of a waveform to a pulse-code modula- tion (PCM) representation [4]. In these and other examples, itis important to thoroughly understand the process of interpolation from the point of view of digital signal processing rather than from a numerical anal ysis viewpoint. For example, tables of mathematical func- tions are generally constructed so that linear interpolation produces sufficiently accurate results, and for cases where linear interpolation is inadequate, there exists a great variety of higher order polynomial interpolation formulas. In signal processing applications there is a great temptati get along with linear interpolation because it is interpre- which it ie clear that terpolation process interpolation is fundamentally a [Manuscript reeived December 6, 1972; evied March 7, 1973 ‘Theatthors are with Bell Laborntorien, Murray Hil, NJ. 07974 discussion makes it abundantly clear that linear interpolatit is generally not appropriate for digital signal processing ap: cations. We discuss the advantages of finite duration impule response (FIR) over infinite duration impulse response (IIR) digital filters for use as interpolation filters and we discus te application of recently developed design techniques for FIk filters to the design of optimum interpolation filters. Thst filters are compared to filters derived from classical pi nomial interpolation formulas to illustrate the improvemet that can be achieved. II, Dictra, Samruine Rat ALrERATION A. Sampling Continuous-Time Signals Consider a continuous-time signal #(0) with Fourier tase form Re) =f ayemae, ‘The signal 4) is sampled to produce the sequence x(n) = (nT), where T is the sampling period. The # transform of thes quence #(n) is defined as -eL, there is a rmsibity of aliasing. In this ease aliasing can be avoided by tking the cutoff frequency of the interpolation filter equal box/T’. As in Section II-B, the resulting output sequence 49) will corespond to samples of a low-pass filtered version ‘the original continuous time signal UL, Isrenrotation Usixa FIR Drctrat. FILTERS The previous section makes it abundantly clear that the ross of changing the sampling rate requires a low-pass Ate. Since itis impossible to realize the ideal low-pass filter tit is required for exact results, we must consider digital Aas that approximate this ideal behavior. As in all filtering tmblens, there are many important considerations. A basic tunsderation is the choice between filters from the class of FR filters and fom the class of IIR filters. Given the type of Aer to be used, the problem of approximating the ideal low- raster must be solved. Finally, there are important consid- erations in how the filter is realized as software or hardware. Al of these facets of the problem are interrelated—resulting in arbitrary tradeoffs between accuracy of interpolation and ‘efficiency of realization. In this section we present some ob- servations on filter design and realization that seem to imply that FIR filters are the proper choice for most interpolation problems. A, Phase Distortion ‘The ideal interpolator has zero phase or at most a linear ‘phase corresponding to an integer number of samples of delay. TIR filters cannot have precisely linear phase [6]. In contrast, there currently exist several techniques for designing optimal FIR digital filters with precisely linear phase. These filters are ‘optimal in the sense that the width of transition band between passband and stopband is minimum for given values of pass- band and stopband ripple and specified passband and stop- band cutoff frequencies [7]-[11]. These filters can be designed with arbitrarily small values for passband ripple, stopband ripple, and transition bandwidth, at the cost of increased im- pulse response duration. Thus with FIR filters, the interpola- tion error due to phase nonlinearity can be zero and the error due to amplitude distortion can be made arbitrarily small. In the case of IIR filters, although extremely good amplitude characteristics can be achieved, there will always be an inter polation error due to phase nonlinearity, B. Filler Realization IIR filters have recursive realizations that are very eco- ‘nomical in terms of computational complexity. Leaving phase considerations aside, FIR filters in general require more com- puttation to achieve a given accuracy of approximation to the desired amplitude response than do IIR filters. However, the particular nature of the interpolation problem makes FIR filters computationally competitive with ITR filters. Consider for convenience a 2ero-phase FIR filter with im- pulse response A(x) which is nonzero in the interval -(S)sns ‘) 2 a) where WV is an odd integer.! In reducing the sampling rate by an integer factor, ic,, T’= MT, we may need a unity gain low- pass filter to insure that no aliasing occurs in retaining only every Mth sample of the sequence x(n). In this case computa- tioi is reduced because of the nature of the desired output sequence. The filtered output sequence at the original sam- pling rate, defined as 9(n), is 2 neem deen 5h) o where al sequences in (9) are astociated with sampling period T. Clearly, all values of the sequence $(n) need not be com- puted since the desired output is Lhe x(wf — &) (108) Xn) = (na) where y(n) is associated with a sampling period T”= MT. This is in contrast to a comparable IIR filter where the computa- tions required to realize the poles of the system function would have to be carried out at the original sampling rate even See Section I-C fora comment oa why should be of. 159though M1 out of Mf samples of f(x) would be discarded, A further simplification results from the fact that zero-phase FIR filters have the property An) = h(n). Thus (108) becomes conn on) = x A(R) [2(nM — 8) + x(n Mt + &)) + HO)=(naf). (100) In the case of sampling rate increase, the interpolated out- put is obtained with sampling period T’™= T/L by filtering the sequence x(n) = x(0/L), = (AC — «any Substituting for o(B) results in mega z= sn) x(k/E)Mn = 8), —-R/L an integer GENIE ey — aL) pete r-ny AL «aay where [a] means “the largest integer contained in a.” Ale though (11) suggests that computation required for each out- put sample is proportional to N, we note that only one out of ‘every L samples of o(n) is nonzero. Thus we see from (12) that computation required is proportional to N/L. Note that in this case the symmetry of the impulse response cannot be exploited to reduce computation If an TIR filter were used, relatively little saving could be achieved. In fact, ina cascade realization, almost no computa: tional saving could be attained by considering the zeroes in the input sequence. Changing the sampling period according to T’=(M/Z)T, requires that we first increase the sampling rate by a factor L and then reduce it by a factor M. Clearly, the savings previ ‘ously discussed could be incorporated into both steps of this process. C. Impulse Response Constraints ‘The previous discussion has presented compelling reasons for the use of linear-phase FIR filters in interpolation. To con- clude this section we discuss some constraints on the impulse response that are specific to the problem of interpolation. Recall that the output y(n) is given by (12). A reasonable requirement on the interpolation filter is that the values of the output at the original sampling times be the same as the orig nal samples.* That is, for ran integer, Wel) = x(6L/L) Equation (12) implies that 0 x cay WL) = x) (BAL — RL) (13) mt "Thin alsa property of (2), as can be cally vere. ‘PRocexDINos oF Tie tex, joxe 1 for all integer values of r. From this equation, we see that HO) =1 ‘). w Ho) = 0, Constraints of this type are rather dificult to impose on TIR filter design procedure. Some final comments on the choice of the length of & impulse response NV are inorder. First, we have assumed tt ‘Nis an odd integer; however, in general 37 can be cithe: et or odd for the FIR filter. Tecan be show [6], [12] that rd even, a incar-phase FIR filter must have a delay of at at one-half sample. This one-balf sample delay itaelf cores to interpolation between samples; thus such a filter couldst preserve the samples ofthe origina sequence. Although et may be instances where the half-sample delay may sot objectionable, oF may even be desirable, odd values of 13> pear to be appropriate for most applications. ‘A second comment regarding the choice of N conceras tx fact that we have asserted that by increasing, we at achieve a better approximation in the frequency domalab the ideal interpolating filter. In the time domaia, increas 1 implies that more values of the original sequence areit volved inthe computation of a given interpolated samples thus we should expect increasingly better results as creases. We have observed that because the sequence (hs mostly zero samples, the computation is proportional to Ni since the impulse response always spans approximately nonzero samples. Suppose that we wish to always have ( samples of the original input sequence involved in the comps tation of each interpolated sample. Using (12) it is ea showa thatthe length of the impulse response should be N=0L, if Q and L are odd = QL-1, — ifeither Q or Lareeven. (i) net If Nis slightly larger than this value; the computation ofsow interpolated samples will involve Q-+1 original samples whit the rest will involve only Q. If 1 is slightly smaller than te value given by (15), then the interpolation will involve che @ or Q~1 samples. Thus to insure that Q samples of te original sequence are always involved, N should satisy (1 ‘This constraints satisfied, as we will see, for filters dev from classical interpolation formulas. It is also importante hardware and software realizations of interpolation fs where it allows the computations to be structured so that does not have to check for zero samples in order to eliniast multiplications by zero, IV. Ciassicat Pownomrat. INTERPOLATION In this section we apply the previous discussion of interpolation process to a study of classical polynomial inte polation methods. Our aim is to give a frequency domaininte pretation of these formulas that will shed some light on tht applicability in interpolation problems arising in digital gel processing. In the course of this discussion, we shall indiat hhow to derive linear phase interpolation filters from tables Lagrange coefficients. We begin with a discussion of law interpolation. A, Linear Interpolation Linear interpolation involves only two consecutive si ples of the original sequence x(n) in the computation of 160ineplated sample. Specifically, the values interpolated be- ven wo samples (0) and x(1) lie on a straight line connect: 'g the two original samples, with the original samples of couse beng preserved. Thus the equation relating the output 1), having sampling period 7” =T/Z, to the input sequence 21), havig sampling period T, is x(1) ~ (0) x0) +) or” — 0) a) = «0 (1-2) 4:0(2), o£, The length of the impulse response is v 1 camistent with (15) for Q=2. Fig. Sdepits linear interpolation asa convolution process Tiesequence o(b) and the triangular envelope of the impulse repose (nA) are shown for the ease of 7’ T/5. [tis clear that because the impulse response has duration N'=5(2)—1 aly two nonzero samples of o(#) are ever coincident with lis=H), Algo itis clear that an) = x(n/L), = 0, The system function corresponding to linear interpola #2, 1 fsin foz7"/2)) * me) = raat wd Tis ssem function is plotted in Fig. 6 for Le. (Curve fea chose senses x0) and x1) without leo generality since ate ig. 5, Linear tering interpretation of linear interpolation, ay Fig. 6. Comparison of Lagrange interpolation {requeney reaponae or L=S labelled Q=2.) We recall that the purpose of the interpolation filter is to remove the images of the signal spectrum that are centered at integer multiples of 2x/T, while leaving the fre- ‘quencies below x/T unaltered. It can be seen from Fig. 6, that Tinear interpolation achieves significant attenuation in only a very small region around each integer multiple of 2e/T. Spe- cifically, attenuation is 40 dB or greater in a band of width 0,35"/T centered at 2x/T, 4x/T, etc. Thus it seems reason- able to note that linear interpolation is appropriate only ifthe original sampling rate is many times the Nyquist rate B. Lagrange Interpolation Clearly linear interpolation will not be satisfactory in ‘many digital signal processing applications, Tn classical nu- merical analysis, the inadequacies of linear interpolation lead to the use of higher order polynomials: i, in contrast to con- necting two points by a straight line, one finds a polynomial of degree Q— 1 that passes through Q original samples. The inter polated values then, are samples of this polynomial. A variety of formulas have been derived for obtaining samples of the polynomial directly from the samples s(#), however since we Aare only interested in the interpolation filter corresponding to polynomial interpolation, we shall use the most convenient form; namely, the Lagrange interpolation formula (13], [14] ‘The form corresponding to interpolation with equal spacing from sampling period Tt T'=T/Lis 161wa= Yo meni coy wn) = lenin Ai%(n/L)x(),— Qeven (20) AsS(n/L)x(b), Qodd_ (206) where we have again chosen to consider interpolation around: some arbitrary time labelled 0. In (20), the quantities A.®(-) are called the Lagrange coefficients and are given by the equa- tions asty 2 ————_= 9" Gey H+ 2-2), cern EQ) H+), om am (21a) AS) = Extensive tabulations of these coeficients are available in tables of mathematical functions [18]. Tt that teresting to note A) = 0, =1, ‘tan integer, and tek This is a result of the fact that the interpolation polynomial passes through the Q original data samples. That is, (kL) = 2(), for & an integer. ‘Thus the condition of (14) is satisfied for Lagrange interpol tion filters. In general, the formulas in (20) may be evaluated for any value of such that -@22\r band and stopbands and may be solved for |A(n)} and using either linear programming techniques (11] or dist Chebyshev approximation techniques [10]. It has been shen [10] that the fiters designed by these techniques are optim in the sense of having the narrowest transition bands for gine 4; and b;, For these filters, the error in the passband and sp bands exhibits an equiripple behavior. An example of alr ‘pass filter designed by linear programming is shown in Fig. In this case L=5, N=29 (Q=6), wp =0.6r/T, ov, ~ 47, K-=10, The resulting filter has 8)~ 2% 0.00586. (The ft gain has been normalized to 0 4B for convenience in ptt ‘The equiripple behavior of the stopband is readily appare “The filter in this example does not satisfy the constr ‘on h(n) given in (14); however, these constraints are let and easily can be added to the constraints of (25) and (I ‘Thus in the optimization procedure A(0) is constrained 110 and (+2), h( #22), ete. are set to 0. The filter performant is only slightly degraded by the addition of these constrains Fig. 11 shows an example where all the fixed-design parame were the same as in the previous example. In this ease it resulting value of &)=6; was 0.00599. It can be seen from Ft 12 that the equiripple nature of the frequency response ‘with little sacrifice in performance. I signal bandwidth is much less thas sf, ‘a bandstop filter may provide superior performance toa cow parable low-pass filter. Fig. 13 shows an example for Let where N=29, wp =0.38/T, to, =1.7r/T, Bay Bue=0 67, iy, 3.7x/T, and K=1.0. 0 this case By= 8, 0,0000681, Ths this filter does a much better job of attenuating the imapsé the signal spectrum around the frequencies 2r/T and 47 ‘but only if the original signal bandwidth was much les the /T. It can be seen that the extra attenuation around 2 ‘and 4x/T is obtained at the expense of the regions arvud 164oe PES & Fig. Frequency response of a typeal low-pass I ‘eed by linear programming. L=' vn29. ate/T, bse 0 00886" (No impulse reaponse coments) F Fe 12 Froqency response for the example of Fig. 11, with imple esyonue contrnnts Ta this cane 8 =e. 00598, a1 Frequency response of a typical bandsto interpolation filter ‘signed by tinea programming. LS, Ore. N=29. wp 0.8e/T, vt Fel Tg Pel St ay. T a = om, 0000081 Fig. 14. Comparison of tinear-phase Lagrange interplators and opti- ‘mum Bandatop fiers (Lo) Pots show minimum seoptband sttemae tion av a function of hall the topband width, For the optimum Tinuntop ter (W=29), the curve eof tae or uy CO.2r/T. Sn/T and Sr/T where the attenuation is much less. Thus ‘optimum filters of this type (and classical filters) should only be used if one is certain of the bandwidth of the input se quence, C. Comparison to Classical Interpolators Itis of considerable interest to compare the filters derived from classical interpolation formulas and those designed by linear programming. In the design of the optimum filters, the parameters of the tolerance scheme of Fig. 10 were set so that 4)=h with symmetrical stopbands of width Bu =Aus= + ~ Day centered at integer multiples of 2x/T. This is reasonable since in interpolation, preservation of the passband is gen- rally as important as rejection of stopbands. In order to com: pare the Lagrange filters tothe optimal filters, values of 6; and 5, were measured at the edges of the passband and the first stopband, respectively. Since for the Lagrange filters, 8 is always greater than & for the above definition of passband land stopbands, the comparison has been made on the basis of passband error Fig. 14 shows such a comparison for the case L=5.* The solid curves are for linear-phase filters derived from the La range interpolation formula where N=QL~I and Q=2, 4, land 6, These curves show the error in decibels at the edge of the passband, ie., 20 logio dy, 88 a function of wy. As an ex- ample, for linear interpolation (Q=2 and N=9) we see that the passhand error is ~42 4B for wy =0.1x/T. Furthermore, since )>4: we can say that the attenuation js at least 42 4B in the region (2r/T—O.1n/T) Sw <(2r/T+O0.1x/T). For wider bandwidths, the performance is worse: however, for higher order interpolators the performance becomes appreci ably better as is expected. “The dotted curves in Fig. 14 show passband error for band stop filters designed by linear programming. By comparing corresponding curves, itean be seen that the optimum designs are always significantly better than the corresponding cla cal interpolator, with the improvement being most striking for harrow bandwidths and for the higher order filters. ‘Clearly, there area variety of optimum designs correspond- ing to situations in which passband and stopband approxima: tion errors are not treated as being of equal importance. For ‘example if5,— Kés, then the case K > I corresponds to placing more importance on stopband attenuation than on passband, + Te comparison are similar for ager value of E 1650 .0¢e/8e anator!-20 064082 Lomas) © 81 02 03 04 Os Fig. 15. Compas Delation Ste of optimum bandetop an optimum Low: pas inter 3) The curve fot N'=20 tell seal for oy 02x. error. This situation would be a more favorable situation for the classical filters, although it is always possible to desiga a better filter using the optimum design procedures, Another interesting comparison is between optimum low- pass filters and optimum bandstop filters. To compare these filters, we set an ae + Feo z-w, =F and Aun Aus: =2ap for the bandstop filters and Dn and Au =#/T'—a,, for the low-pass filters. That is, both filters were designed to accommodate the same input signal band= width, Fig. 15 shows the difference between stopband attenua tions for the bandstop filters and the low-pass filters as a func tion of bandwidth, From these curves we see that for narraw bandwidths the bandstop filters have significantly greater attenuation; however as the bandwidth approaches half the original sampling frequency, there is no difference between the two types of filters. VI. Conewusions In this paper we have discussed the process of interpolation as a problem in digital filtering. Most of our discussion has involved frequency-domain representations of the interpola- tion process and desiga criteria for digital interpolation filters. ‘We have taken this approach because itis the most reasonable for digital signal processing applications where it is necessary to either raise or lower the sampling rate of a signal. This point of view is in contrast to that of interpolation in tables, ‘where one is concerned primarily with minimizing the error in, 2 particular interpolated sample. Because of the variety of factors involved in the design of an interpolation filter, we hhave not tried to give design formulas and error bounds that would have limited value, but rather we have chosen to at- tempt to illuminate the important factors involved in the PRoceRDINGS OF THE IEEE, ox interpolation process and to discuss general design procsin that can be adapted to a variety of situations. In particular, we have argued that linear-phase FIR ins hhave many attractive features for discrete-time interpis and have shown how they may be eficiently utilized, Csi polynomial interpolation has been discussed in the costa digital signal processing. Interpolation filters derived im polynomial interpolation formulas are attractive because impulse response can be easily computed or looked up at table, However, we have seen that the frequency respned such systems leaves much to be desired in digital signa cessing applications where the original sampling rate may only slightly above the Nyquist rate ‘As an alternative to filters based on classical interplia formulas, we discussed optimum low-pass and bandstpfil filters that were designed by linear programming. Tbe bal stop filters have frequency responses that are very sinilew the classical interpolators, but are always superior. The bak stop designs appear to be most important for cases whe ‘original sampling rate is several times the Nyquist rate vb the low-pass designs are appropriate when the orginal se pling rate is close to the Nyquist rate AckNownepowent The authors would like to acknowledge the helpful a ments and criticisms provided by Dr, O. Herrmans af Dr. J, F. Kaiser, RerERences [t) A. K, Rosenberg, RW. Sehater, and L. R. Raber, "end ‘meooihine and quanlatag the palameters of formant code ra tech," J Acoust Soe Amer wa. SO. pp. 1532-1838, Dee [2] Howe Schaterand Lo Re Relnen "Desigusnd simulationt net Enalyeieyathess sjwtem based on shor-ume Fourier ami TREE Trans. Audio Blecroocous, vol. AU, 9p. 165-11, inns [91 SIL. Proeney, Rw, Kieburts, KV. Mina, and S.K. Tein incign of sgt tere fora all digs requeney divin mab plextime division maltplex translstoy” ZEEE Tren. Create Srp val C118, pp. 102-71, Now. 1971 [61 Bry. Goodman, tive application of dea modulation to aman PCM encoding.” Bell Spt. Tech J, Val 48, 9p. 321-34, Fea [sy H. Breeman, Disrete-Pime Systema. New Vork: Wik 198 [o] A.J. 'Gitbe "Om the frequency-domain reeponaes of cel I Sera" Phi. digertation, University of Wincor, Mal (7) Le’ Rabiner, “Teenglquee for designing finite duratin spe (8) 0. Herrman, “On the design of nontecursive dig fen wi Near phase" Eledrom. Let, wal 6, pp. 928-32, 1070, [p) E. Holster, A. V-Oppenteim, and J Siegel, "A new sce fhedesin of nonrecoate data fiers,” Proc Sh An Prat information Scteces ond Systems, O4-12, 1 sand JH. McClellan, “Cheb yay approxima’ ith ear phase,” TEEE Tre Theory, vol. C19, 9p. 189-194, Mat. 1972, (11) LI Tabines, "The doug of finite pulse reponse dg fan ‘Ging Linene programing techniques” Bell Syst Tech J. vl pp. ti07-1198, July-Aug. (072 (12) E2R Rabiner and RW Sehaler, “Recursive and ronecusinn sligatiogs of digital eee designed by Irequency sung lauen" FREE Trews. Audio Mlecvoacous, val AUD, 9.202% See i971 (13) RW. Hamming, Numerical Methods for Scents and Ensen New York: McCraw: Hil, 1962 Uis] FB, Hildebrand, Twvoducton fo Numerical Amelysis, New Yt MeGraw-Hil 1986 (15) Handbook of Methematial Functions, M. Abramowite aad Stegun, Eds. (Nat. Bur: Stand. Appl Math, Ser=88) Wate, D.C.'U- 8. Gov. Pristing Ofer, 1964, 166