If you read an earlier article about FIR filter impulse responseinterpolation usingfrequency domain zero stuffing, it is logical toconclude that if we can interpolate time-domain impulse responses, we should beable to interpolate time-domain signals using the same frequency-domain zerostuffing method .
This is true, and the aboveinterpolation-by-M process applied to time signals is sometimes called exactinterpolation —because its performance is equivalent to using an ideal,infinite-stopband attenuation, time-domain interpolation filter—and has madeits way into DSP textbooks, journal articles, and classroom notes of famous DSPprofessors.
To establish our notation, let’s say we compute the FFTof an N -point x (n ) time sequence to produce its X (m )frequency-domain samples. Next, we stuff (M –1)N zeros in themiddle of X (m ) to yield the MN -length X int (m )frequency samples, where MN is an integer power of two. Then we performan MN -point inverse FFT on X int (m ) to obtain theinterpolated-by-Mx int (n ) times samples. Using thisfrequency-domain zero stuffing to implement time-domain signal interpolationinvolves two important issues upon which we now focus: 1) interpolating realsignals and 2) interpolating analytic signals.
ComputingInterpolated Real Signals
The first issue: to ensure the interpolated x int (n )time sequence is real only, conjugate symmetry must be maintained in thezero-stuffed X int (m ) frequency samples. If the X (m )sequence has a nonzero sample at X int (N /2), the fs /2frequency component, we must use the following steps in computing X int (m )to guarantee conjugate symmetry:
. •Perform an N -point FFT on an N -point x (n )time sequence, yielding N frequency samples, X (m ).
. •Create an MN -point spectral sequence X int (m )initially set to all zeros.
. •Assign X int (m )= X (m ),for 0 ≤ m ≤ (N /2)–1.
. •Assign both X int (N /2) and X int (MN–N /2)equal to X (N /2)/2. (Thisstep, to maintain conjugate symmetry and improve interpolation accuracy, is notso well known.)
. •Assign X (q ), where MN –(N /2)+1 ≤ m ≤ MN –1, and
X int(m )= (N /2)+1 ≤ q ≤ N –1.
•Computethe MN -point inverse FFT of X int (m ) yielding the desiredMN -length interpolated x int (n ) sequence.
. •Finally, if desired, multiply x int (n )by M to compensate for the 1/M amplitude loss induced byinterpolation.
Whew! Ourmathematical notation makes this signal interpolation scheme look complicated,but it’s really not so bad. Table 13–8 below shows the frequency-domain X int (m ) sample assignments,where 0 ≤ m ≤ 15, to interpolate an N = 8-point x (n )sequence by a factor of M =2.
Table 13–8 X INT (m)Assignments for Interpolation by Two
One of thenice properties of the above algorithm is that every M th x int (n )sample coincides with the original x (n ) samples. In practice, dueto our finite-precision computing, the imaginary parts of our final x int (n )may have small non-zero values. As such, we take x int (n )to the be real part of the inverse FFT of X int (m ).
Here’s the second issue regarding time-domain real signalinterpolation, and it’s a big deal indeed. This exact interpolationalgorithm provides correct results only if the original x (n )sequence is periodic within its full time interval.
If X (m ) exhibits any spectral leakage, like the signalswith which we usually work, the interpolated x int (n )sequence can contain noticeable amplitude errors in its beginning and endingsamples, as shown in Figure 13–74(a) below where an N = 24 sample x (n ) sequence is interpolated by M = 2.
Figure 13–74 Interpolation results, N = 24, M = 2: (a)interpolated xint (n), original x(n), and correctinterpolation; (b) interpolation error.
In that figure, squares (both white and black) represent the 48-pointinterpolated x int (n ) sequence, white squares are the original x (n )samples, and circles represent the exactly correct interpolated sample values.(In the center portion of the figure, the circles are difficult to see becausethey’re hidden behind the squares.) The interpolation error, the differencebetween the correct samples and x int (n ), is given in Figure13–74(b).
These interpolation errors exist because X int (m )is not identical to the spectrum obtained had we sampled x (n ) ata rate of Mf and performed an MN -point forward FFT. There’s noclosed-form mathematical expression enabling us to predict these errors. Theerror depends on the spectral component magnitudes and phases within x (n ),as well as N and M . Reference  reports a reduction ininterpolation errors for two-dimensional image interpolation when, in aneffort to reduce spectral leakage in X (m ), simple windowing isapplied to the first few and last few samples of x (n ).
With theadvent of fast hardware DSP chips and pipelined FFT techniques, the abovetime-domain interpolation algorithm may be viable for a number of applications,such as computing selectable sample rate time sequences of a test signal thathas a fixed spectral envelope shape; providing interpolation, by selectablefactors, of signals that were filtered in the frequency domain using the fastconvolution method; or digital image re-sampling.
One scenarioto consider is using the efficient 2N -Point Real FFT technique tocompute the forward FFT of the real-valued x (n ). Of course, theprudent engineer would conduct a literature search to see what algorithms areavailable for efficiently performing inverse FFTs when many of thefrequency-domain samples are zeros.
ComputingInterpolated Analytic Signals
We can use the frequency-domain zero stuffing scheme to generate aninterpolated-by-M analytic time signal based upon the real N -pointtime sequence x (n ), if N is even. The process is asfollows:
. • Perform an N -point FFT on an N -point real x r (n )time sequence, yielding N frequency samples, X r (m ).
. • Create an MN -point spectral sequence X int (m )initially set to all zeros, where MN is an integer power of two.
. • Assign X int (0) = X r (0),and X int (N /2)= X r (N /2).
. • Assign X int (m ) = 2X r (m ),for 1 ≤ m ≤ =(N /2)–1.
. • Compute the MN -point inverse FFT of X int (m )yielding the desired MN -length interpolated analytic (complex) x c,int (n )sequence.
. • Finally, if desired, multiply x c,int (n )by M to compensate for the 1/M amplitude loss induced byinterpolation.
The complex x c,int (n )sequence will also exhibit amplitude errors in its beginning and endingsamples.
Used with thepermission of the publisher, Prentice Hall, this on-going series of articles onEmbedded.com is based on copyrighted material from “UnderstandingDigital Signal Processing, Second Edition” by Richard G. Lyons. The book canbe purchased on line.
RichardLyons is a consultingsystems engineer and lecturer with Besser Associates. As a lecturer with Besserand an instructor for the University of California Santa Cruz Extension, Lyonshas delivered digitasl signal processing seminars and training course attechnical conferences as well at companies such as Motorola, Freescale, LockheedMartin, Texas Instruments, Conexant, Northrop Grumman, Lucent, Nokia, Qualcomm,Honeywell, National Semiconductor, General Dynamics andInfinion.