ELECTRONOTES APPLICATION NOTE NO.

425
1016 Hanshaw Road
Ithaca, NY 14850 October 10, 2015

INVERTING A DTFT ?
INTRODUCTION

Three popular transforms are the z-Transform of a length-N time sequence x(n) as
given by:

the Discrete-Time Fourier Transform given by:

and the Discrete Fourier Transform (DFT or FFT) given by:

The development here is that the z-transform is evaluated on the unit circle in the z-plane
giving the DTFT, and the DTFT is then evaluated at N equally space frequencies on the unit
circle giving he DFT. Note that the DTFT is probably the least known of the three – at least
until we consider that if we substitute h for x (H for X) we are talking “FILTER” and we recognize
the DTFT as a “frequency response” of the filter. Each of the three transforms has an inversion
[1]. Here we will mention the inverse DFT:

Our reason for comparing (3) and (4) is their near-perfect symmetry and the fact that both
represent N independent linear equations in N unknowns. A matrix inversion will solve these.
In the case of the DFT, the FFT is a fast algorithm to the solution. Note that for the DFT, it is the
case (and a requirement) that the samples (for forward or inverse) are equally spaced. What
happens if we drop this requirement and agree to bring back the matrix?

AN-425 (1)
TWO VIEWS OF THE DTFT EQUATION
The DTFT equation is valid for all ω, but since this is a discrete time device, we think
in terms of values of ω between 0 and π as being of the most interest. But even though
time is discrete, ω is continuous. Let’s suppose we have in mind a particular frequency
ω1. Equation (2) is certainly valid for our choice of ω1:

This shows the computation of the DTFT based on N values of x(n). We can do this for
one value of ω or for as many value of ω as we desire. It is a computation of the
“spectrum” or frequency-domain description of the time sequence x(n). As said above,
this is much more familiar if we substitute h(n) for x(n) and H(ejω) for X(ejω). [I guess we
should always write out the complete notations H(ejω) and X(ejω), although H(ω) and
X(ω) are often seen as a shortcut, without confusion.] So X(ejω) is in a familiar sense
the frequency description of x(n) while the “filter case” has H(ejω) as the frequency
response of h(n). Mathematically exactly the same notion. Thus equation (5) is used to
compute the DTFT. In Matlab a convenient function freqz can compute the DTFT.
Often a DFT=FFT program can also be used, quite possibly involving zero-padding, as
will be discussed below. Thus the forward version of equation (5) is out first of two
views.

When we consider “inverting” equation (5), we in turn get two interpretations of the
inversion. We want x(n) from X(ejω). The approach we must choose likely depends on
how well (in what form) we know X(ejω).

For example, if we are doing filters, we might well have in mind something like a low-
pass filter where we specify a desired (ideal) passband and stopband. That is, a
continuous function on ω=0 to ω=π (quite likely both magnitude and phase). In such a
case, the direct mathematical inversion of the continuous frequency function is in order
[2].

where we have used the filter notations h and H. This is nothing more than inverting a
“Fourier Series” with the roles of time and frequency reversed.

Equation (6) (using x and X) does us no good if we don’t have an expression for
X(ejω). In such a case, perhaps we only have some isolated samples of X(ejω). That is,

AN-425 (2)
we have perhaps samples for frequencies ω0, ω1, ω2, …. ωN-1. We can then write N
versions of equation (5), each one being one equation in N unknowns [x(n) are
unknown] for a total of N equations in N unknowns. If we have fewer equations than we
need, or more, some least square fits can be considered. Solving for N unknowns from
N equations is the matrix inversion suggested. Although not required, if the N frequency
samples are equally spaced on 0 to π, the inversion can be by DFT=FFT.

TEST CASE
Here we need a test DTFT which we manipulate. We will take this to be the DTFT of
a length-12 moving average filter. Thus the impulse response is 12 values of 1/12 in
succession: h(n) = (1/12) for n=0 to 11 and is zero everywhere else. We know at least
three ways of calculating the DTFT from these h(n). First, we can use equation (2)
directly. Second, we can use Matlab’s freqz function. Thirdly we can write a closed
form expression for the result. Fig. 1a and Fig. 1b show the magnitude responses for
methods 1 and 2. We need to see the printouts of the real and imaginary parts to be
certain the results are identical (see code below).

The closed form (method 3) begins by writing the transfer function using well
documented [2, 3] manipulations:

AN-425 (3)
and substitutine z = ejω using the Euler relationship to arrive at trig functions – in this
case the well-known “periodic sync” or “aliased sinc”:

% AN425.m
% THREE WAYS OF COMPUTING DTFT
% Direct Computation of H - Equation (1) of AN-425
h=ones(1,12)
Hdc=zeros(1,500);
for kk=1:500
w=(kk-1)*pi/500;
for n=0:11
Hdc(kk)=Hdc(kk)+h(n+1)*exp(-j*n*w);
end
end
Hdc=(1/12)*Hdc;
%
% Using freqz
Hfz=(1/12)*freqz(h,1,500).';
% Using Closed Form
w=[0:.001:.499]*(2*pi)
Hcf=(1/12)*exp(-(11/2)*j*w).*sin(6*w)./sin(w/2);
Hcf(1)=1;
%display errors
e1=max(abs(Hdc-Hfz))
e2=max(abs(Hdc-Hcf))
e3=max(abs(Hfz-Hcf))
% Plot
figure(1)
subplot(311)
plot([0:.001:.499],abs(Hdc),'r')
hold on
plot([-.05 .55],[0 0],'k:')
plot([0 0],[-.1 1.2],'k:')
hold off
axis([-.05 .55 -0.05 1.1])
title('Direct Calculation')
subplot(312)
plot([0:.001:.499],abs(Hfz),'r')
hold on
plot([-.05 .55],[0 0],'k:')
plot([0 0],[-.1 1.2],'k:')
hold off
axis([-.05 .55 -0.05 1.1])
title('Matlab freqz')
subplot(313)
plot([0:.001:.499],abs(Hcf),'r')
hold on
plot([-.05 .55],[0 0],'k:')
plot([0 0],[-.1 1.2],'k:')
hold off
axis([-.05 .55 -0.05 1.1])
title('Closed Form')
AN-425 (4)
 Fig. 1c indeed shows the magnitude plot using equation (8). The code used is printed
on page 4 for clarity. Note that the printouts list the three errors e1, e2, and e3 is being
in the 10-15 range (roundoff only). Thus we have a well defined test object and some
practice with our mathematics.

SETTING UP AND INVERTING THE EQUATIONS

Here we are pretending that we have done nothing above that involved actually
knowing that we had a length-12 moving average. We just have some data – say 12
frequency samples each of which can be written in terms of equation (5) as:

where we have set T=1 so that frequencies ω are on an interval of 0 to π. Here we will
take samples starting at π/8 and proceeding at intervals of 0.1. These samples are
found using equation (8) and appear as in Fig. 2, but keep in mind that the samples are
complex and we are disclaiming we know anything about equation (8) or Fig. 2.

We now have chosen 12 frequencies and have 12 corresponding (complex) values

for the DTFT and can write 12 equations in 12 unknown values of x(n). So we can
calculate the matrix elements and invert to equations. This is shown in the code on the
next page.
AN-425 (5)
% Form/Invert Matrix
k=1:12
w=pi/8+0.1*k
X=(1/12)*exp(-(11/2)*j*w).*sin(6*w)./sin(w/2)
figure(2)
plot([0:.001:.499],abs(Hcf),'r')
hold on
plot(w/(2*pi),abs(X),'bo')
plot([-.05 .55],[0 0],'k:')
plot([0 0],[-.1 1.2],'k:')
hold off
axis([-.05 .55 -0.05 1.1])
title('Closed Form with Samples')
%
for n=1:12
for k=1:12
M(k,n)= exp(-j*(n-1).*w(k));
end
end
M
x=inv(M)*X.'

(Most of the above code just plots Fig. 2.) The frequency samples are “stolen” from
equation (8) and the matrix elements M(k,n) are the evaluation of on the
frequency vector ωk for n=0 to 11. The result is that x is 12 values all equal to 1/12:

x =
0.0833+ 0.0000i
0.0833- 0.0000i
0.0833
0.0833+ 0.0000i
0.0833+ 0.0000i
0.0833+ 0.0000i
0.0833- 0.0000i
0.0833+ 0.0000i
0.0833+ 0.0000i
0.0833- 0.0000i
0.0833+ 0.0000i
0.0833+ 0.0000i

So we see that we can ”invert” from the frequency domain to the time domain in
several was. If we have a closed form for H(ejω), or if we have a data file or even a
digitizable graph, we can possibly invert by integration as in equation (6), either
analytically or numerically. In the case where we have a limited number of samples,
perhaps from an isolated region of frequency, we may be able to solve N equations in N
unknowns, or use a least squares procedure.

AN-425 (6)
GETTING MORE VALUES OF X(ejω)

At this point, we assume we already have some length-N time sequence, x(n) so we
can thus also easily obtain the DFT=FFT, also length-N, X(k), which is a corresponding
frequency-domain description of x(n) as evaluated at frequencies kfs/N where fs is the
sampling frequency associated with x(n). It may be the case that we want a frequency
description at frequencies other than the available values kfs/N. In such instances, we
might think about a general interpolation of X(k) to a somewhat larger number of points
by using FFT interpolation in frequency. Then from this larger set we select or reject
points we want to use. Or perhaps, we might just think of evaluating the DTFT
[equation (2)] at just one frequency, or at relatively few frequencies of interest. Both
approaches should (must) give the same correct answers. This we will show.

Of importance here is the question of whether or not the new spectral point or points
obtained are meaningful. If we are looking for increased spectral “resolution”, there will
be no improvement because there is no information added. Sometimes we think we are
better off, because a plot is more comfortable to understand (smoother). So there can
be an issue of viewing comfort. But we don’t find a sharper “bump” of spectral energy
suddenly appearing inside a more gradual existing bump. Moreover, there can be
cases where information other than that in x(n) can be used. For example, if we were
specifically told that the samples are, perhaps, those of a single sinusoidal waveform.

On the other hand, perhaps x(n) is an impulse response like h(n) in which case the
new frequency point is meaningful. It tells us how the filter will respond to the frequency
of interest, even though the frequency is not a DFT frequency. (Typically we measure a
frequency response with a function generator which has a continuous frequency dial.)

The conventional wisdom as to how an “FFT-Based” interpolation is achieved may

well be confusing [4]. It is well established that in order to interpolate in time, we take
the FFT, zero-pad it in the middle, and then take the inverse FFT of the larger sequence
[ 5 ]. On the other hand, it is less known perhaps but still conventional wisdom to
understand that to interpolate in frequency (erroneously sometimes said to increase
resolution) you zero pad by adding a string of zeros to the end of x(n) and then take the
forward FFT. As we have mentioned [ 4 ], while there are differences between the way
time and frequency are “appreciated”, we really can’t make the case that these two, as
mathematical “duals”, are mathematically different [6 - note]. In fact, the zero padding is
always in the middle in both the frequency and the time cases.

How can this be that putting zeros on the end is actually the same as putting them in
the middle? Well – because the sequences here are periodic, any insertion of zeros will
be in the middle (somewhere). Any portion we think of as above the middle can be
thought of as below zero on the negative side.
AN-425 (7)
AN EXAMPLE

We are badly in need of an example. Let’s suppose we make up a length-12

sequence as:
x=[1 -2 3 -2 -1 5 5 -4 4 3 -3 2 ]

This we suppose is some time indexed data. It might be some actual data sequence of
length exactly 12. Of it might be the length-12 impulse response of some filter. Or it
might be 12 samples snipped out of a much longer sequence. Note that it is real and
has no apparent patterns. We are most familiar with time sequences that are purely
real. It is plotted in Fig. 3a.

Now, at the point where we propose to analyze it by a DFT(=FFT), we have given it a

particular property: it is periodic with length 12. It might even have been a cycle of
some longer sequence that was actually length-12 periodic. In any event, it is now! It
overlaps periodically at the ends.

The sequence x(n) has a DFT X(k) as defined by equation (3). This DFT is also
length-12. Unlike x(n), X(k) is complex (in general, we expect a complex result – see
equation), and it has symmetry about k=0 (constant) and k=6 (half the sampling
frequency). In fact, we show in Fig. 3b the real part (red), the imaginary part (blue) and
the magnitude (black). It too is periodic. The dashed lines are just to avoid confusion.

AN-425 (8)
AN-425 (9)
 At this point, we are going to do some very simple things and if we are careful, we
should get good results quite promptly. First we will add zeros to the end of x(n) (the
conventional wisdom). In fact, we add 84 zeros making it length 96 in all. Then we will
take the FFT of the padded sequence. We anticipate a frequency-interpolated version
of Fig. 3b. It is composed of sinusoidal components, so should be smooth and not very
much like the dashed lines of Fig. 3b. Fig. 4 shows the results. Fig. 5 shows the same
results with dashed lines between the interpolated set of points. As a check, we will
also compute the intermediate spectral values (one at a time) using equation (3). Fig. 6
is just a re-plot of Fig. 4 with the calculated values shown as green stars. The
agreement is perfect. The methods are equivalent. (No doubt from the equations, but
the surprise is always when you do get it work! )

So Fig. 4, Fig. 5, and Fig. 6 are pretty much the same graphs and constitute the
elements of a “slide show” presentation. Fig. 4 relates best to the original length-12 plot
while Fig. 6 shows essentially a verification that the calculation works (all dots turn
green). This leaves Fig. 5 as the plot that is easiest to follow. Note that we see that
interpolation adds a lot to the understanding of the sinusoidal components, and we
understand how much the dashed lines help us follow the swings. For example, the red
dashed line between k=24 and k=32 as seen in Fig. 4 (straight!) and then in reality, Fig.
5! In going from 12 points to 96 points (Fig. 4), a lot of definition is added by the
additional circles. However, the dashed line linear interpolations (Fig. 5) clear up a lot
of uncertainty. A better interpolation, perhaps to 1200 points using the DFT, would
likely be better than the dashed lines. A lot to examine.

AN-425 (10)
 One final variation here is shown in Fig. 7. In as much as we did interpolation by
taking a zero-padded time sequence, thus adding no additional information, we might
suppose that we could pad with something other than zeros, and still argue that no new
information is added. In particular, we might simply repeat the original 12 samples
seven additional times and look at the length 96 FFT. One new feature is that we are
adding energy in this case, so we need to divide by 8 at some point. The result is the
DFT of the original length-12 sequence with 7 zeros between each of the original
frequency points. This is no surprise. The DFT harmonics of the length 96 sequence
are just 0, 8, 16, … . 96.

CODE FOR FIG. 4 TO FIG. 7

%
% Now zero pad x to xe
xe=[x zeros(1,84)];
Xe=fft(xe) ;
figure(4)
plot([0:95],real(Xe),'ro')
hold on
plot([0:95],imag(Xe),'bo')
plot([0:95],abs(Xe),'ko')
% Plot originals with *
plot([0:8:95],real(X),'r*')
plot([0:8:95],imag(X),'b*')
plot([0:8:95],abs(X),'k*')

AN-425 (11)
% Plot dashed lines between orig
plot([0:8:95],real(X),'r:')
plot([0:8:95],imag(X),'b:')
plot([0:8:95],abs(X),'k:')
% fudge in dashed lines to starting values for clarity
plot([88 97],[real(X(12)),real(X(1))],'r:')
plot([88 97],[imag(X(12)),imag(X(1))],'b:')
plot([88 97],[abs(X(12)),abs(X(1))],'k:')
%
plot([0 0],[-25 25],'k:')
plot([-5 105],[0 0],'k:')
axis([-5 102 -24 24])
hold off
%
%
% replot Fig. 4 with dashed lines
figure(5)
plot([0:95],real(Xe),'ro')
hold on
plot([0:95],imag(Xe),'bo')
plot([0:95],abs(Xe),'ko')
%Plot originals with *
plot([0:8:95],real(X),'r*')
plot([0:8:95],imag(X),'b*')
plot([0:8:95],abs(X),'k*')
% Plot dashed lines between all
plot([0:95],real(Xe),'r:')
plot([0:95],imag(Xe),'b:')
plot([0:95],abs(Xe),'k:')
%
plot([0 0],[-25 25],'k:')
plot([-5 105],[0 0],'k:')
axis([-5 102 -24 24])
hold off
%
%
% Calculate with equation (2)
Xdc=zeros(1,96);
for kk=1:96
w=(kk-1)*2*pi/96;
for n=0:11
Xdc(kk)=Xdc(kk)+xe(n+1)*exp(-j*n*w);
end
end
figure(6)
% Interpolated by FFT
plot([0:95],real(Xe),'ro')
hold on
plot([0:95],imag(Xe),'bo')
plot([0:95],abs(Xe),'ko')

AN-425 (12)
% Originals from length 12
plot([0:8:95],real(X),'r*')
plot([0:8:95],imag(X),'b*')
plot([0:8:95],abs(X),'k*')
% Direct Calculations as green *
plot([0:95],real(Xdc),'g*')
plot([0:95],imag(Xdc),'g*')
plot([0:95],abs(Xdc),'g*')
% dashed lines for original length 12
plot([0:8:95],real(X),'r:')
plot([0:8:95],imag(X),'b:')
plot([0:8:95],abs(X),'k:')
% fudge connection
plot([88 97],[real(X(12)),real(X(1))],'r:')
plot([88 97],[imag(X(12)),imag(X(1))],'b:')
plot([88 97],[abs(X(12)),abs(X(1))],'k:')
%
plot([0 0],[-25 25],'k:')
plot([-5 105],[0 0],'k:')
axis([-5 102 -24 24])
hold off
%
%
%
% Repeat the length-12 sequence 8 times
xe=[x x x x x x x x ]/8;
Xe=fft(xe) ;
figure(7)
plot([0:95],real(Xe),'ro')
hold on
plot([0:95],imag(Xe),'bo')
plot([0:95],abs(Xe),'ko')
% Plot originals with *
plot([0:8:95],real(X),'r*')
plot([0:8:95],imag(X),'b*')
plot([0:8:95],abs(X),'k*')
% Plot dashed lines between orig
plot([0:8:95],real(X),'r:')
plot([0:8:95],imag(X),'b:')
plot([0:8:95],abs(X),'k:')
plot([88 97],[real(X(12)),real(X(1))],'r:')
plot([88 97],[imag(X(12)),imag(X(1))],'b:')
plot([88 97],[abs(X(12)),abs(X(1))],'k:')
%
plot([0 0],[-25 25],'k:')
plot([-5 105],[0 0],'k:')
axis([-5 102 -24 24])
hold off

AN-425 (13)
REFERENCES (All from EN Publications)

[1] “Fourier Map”, Electronotes Application Note No. 410, May 6, 2014
http://electronotes.netfirms.com/AN410.pdf

[2] “Basic Elements of Digital Signal Processing: Filter Elements – Part 1”,
Electronotes, Vol. 20, No. 197, pp 3-6; http://electronotes.netfirms.com/EN197.pdf
“Basic Elements of Digital Signal Processing: Filter Elements – Part 2”, Electronotes,
Vol. 20, No. 198 , pp 12-13 http://electronotes.netfirms.com/EN198.pdf

[3] “Linear Phase - The Term Taken Outside”, Electronotes Application Note No. 350,
November 1998 http://electronotes.netfirms.com/AN350.PDF

[4] “FFT – Interpolation In Time And Frequency “, Electronotes, Volume 23, Number
222 June 2015 http://electronotes.netfirms.com/EN222.pdf

[5] “A Short Presentation Of FFT (DFT) Interpolation”, Electronotes Application Note

No. 398, July 20, 2013 http://electronotes.netfirms.com/AN398.pdf
See also: “Interpolating Samples with the DFT/FFT,” Electronotes, Vol. 16, No. 172,
July 1988. See also: Interpretation of DFT as One Frequency Sinusoidal Waveforms”,
Electronotes Application Note No. 373, February 2009
http://electronotes.netfirms.com/AN373.pdf and “Calculating the DFT and the Fourier
Series: Each with the Other”, Electronotes, Vol. 19, No. 188, February 1997

[6] Note: This relates to the “distress” admitted to in [4] as to why there is a well-
defined zero in frequency (zero frequency = direct current) with no corresponding zero
in time. This, and the tendency to make time functions and sequence be purely real
(instilling certain symmetries in the transformed frequency description) tends to
prejudice our expectations about how time and frequency relate.

AN-425 (14)