Sei sulla pagina 1di 16

An Introduction to Multicarrier Modulation

Notes for ECE1520 Data Communications


Teng Joon Lim
Edwards S. Rogers Sr. Dept. of Elect. & Comp. Engineering
University of Toronto
Abstract
Multicarrier modulation helps to reduce the detrimental eects of multipath fading. Because of its
robustness to multipath, and the ease of implementating it in transmitters and receivers using the
fast Fourier transform (FFT), the MCM concept is growing rapidly in practical importance. It can
be found today in IEEE standards 802.11a, 802.15.3 and 802.16a, as well as standards for digital
broadcasting of television and radio. This document serves to derive the multicarrier modulation
method from rst principles.
1 Scope and Objectives
Multicarrier modulation is an idea which was introduced over three decades ago and is of increasing
interest today because it can now be implemented using powerful integrated circuits optimized for
performing discrete Fourier transforms. Because of its increasingly widespread acceptance as the
modulation scheme of wireless networks of the future, it attracts a lot of research attention, in
areas such as time-domain equalization, peak-to-average ratio reduction, phase noise mitigation
and pulse shaping.
Despite the number of papers which have been written about multicarrier modulation in recent
years, a comprehensive review paper which covers the fundamentals of the topic for the beginning
graduate student is hard to nd. This document hopes to set this anomaly right, and introduces the
newcomer to the fundamental operations of an OFDM system using little more than fundamental
concepts of signal processing and linear algebra.
2 Fundamental Concepts
2.1 Motivation
Consider a system in which the required symbol or baud rate 1/T
sym
symbols/s is larger than the
coherence bandwidth of the channel 1/T
m
. T
m
is also known as the channel delay spread, and
can be visualized as the maximum time extent of the channel impulse response. The number of
symbols of inter-symbol interference (ISI) in the channel is given by
L =
_
T
m
T
sym
_
, (1)
where x| is the largest integer smaller than the real number x. Clearly, the higher the required
symbol rate, the larger the value of L and the more severe ISI becomes for a given channel.
1
Introduction to Multicarrier Modulation 2
Traditionally, ISI is removed using an equalizer, which may be implemented either in the time
or frequency domains, with symbol-by-symbol or sequence estimation algorithms. However, the
complexity of an equalizer increases with the severity of the ISI introduced by the channel, and
in the modern context of wireless networks with broadband links providing several mega-bits per
second (e.g. 802.11a promises 54 Mbps), it may not be practical to implement an equalizer at all
because of overwhelming complexity.
On the other hand, if we could somehow reduce the symbol rate so that ISI becomes negligible
(ie. L = 0 or at least a very small integer) while still maintaining the required information bit
rate, equalization becomes unnecesssary. One way to do this is simply to increase the level of
modulation in an M-ary pulse modulation scheme but there is a limit on how large M can be
before modulation and demodulation complexity becomes overwhelming. For instance, suppose we
have 100 symbols of ISI (L = 100) which is realistic for transmissions at mega-bps over wireless
channels. Let each symbol carry only one bit. Then to increase the symbol interval to the extent
that L becomes close to nothing requires each symbol to carry 100 bits, or M = 2
100
, which would
be impossible to modulate or demodulate.
The other way to increase the symbol interval is through parallel transmission over many
orthogonal channels. Continuing with the previous example, if we choose 100 channels to transmit
over, each one will transport only one bit, e.g. using BPSK, per T = 100 T
sym
seconds, and ISI will
be avoided on all channels. To create these 100 orthogonal channels requires us to design a set of
100 signals g
n
(t), n = 0, . . . , 99, that are mutually orthogonal. If we constrain bandwidth usage to
be the same in the serial and parallel transmission schemes, we can use g
n
(t) = exp(j2f
n
t)w(t),
where f
n
f
n1
= 1/T and w(t) = u(t) u(t T) is a rectangular window of length T seconds.
The bandwidth occupied by these 100 pulses is approximately 100/T Hz, which is identical to the
1/T
sym
Hz required by the serial transmission scheme.
So the lesson so far: To create N orthogonal channels for transporting N symbols at a symbol
rate of 1/T
sym
, we can use N complex sinusoids with frequency separation 1/NT
sym
Hz that will in
total occupy 1/T
sym
Hz, or the same bandwidth as the serial transmission scheme... in an AWGN
channel.
Question: Are these channels still orthogonal and thus easy to demodulate (by processing each
one independently of the others) in the severe ISI channel we face in broadband transmission?
Answer: Yes, but at the cost of some decrease in spectral eciency, through the insertion of a
cyclic prex. This will be explained shortly (with the aid of some mathematics).
Multicarrier modulation can be seen as a parallel transmission scheme developed to mitigate ISI
through the lengthening of the symbol interval this removes ISI in time, i.e. symbols transmitted
in succession fusing together. However, ISI in frequency i.e. interference from other symbols being
transmitted at the same time over dierent carriers, is non-negligible unless the cyclic prex method
is used.
2.2 Signal Model for an AWGN Channel
From the description of the previous section, we can write the baseband-equivalent transmitted
signal over one symbol interval (T = NT
sym
where N is the number of orthogonal carriers used)
as
s(t) =
N/2

n=N/2+1
a
n mod N
g
n
(t) (2)
Introduction to Multicarrier Modulation 3
E
Error
Encode
E
Modul-
ation
E
Freq.
Inter-
leave
E
Serial
to
Parallel
E i

c
e
j
N/2+1
t
E
E i

c
e
j
N/2+2
t
E
q
q
q
E i

c
e
j
N/2
t
E

E
Amp/
Filter
d
' ' E E ' E
Bit Rate
M/RT
sym
Bit Rate
M/T
sym
Symbol Rate
1/T
sym
Symbol Rate
1/T
R = Code Rate
Fig. 1: Conceptual multicarrier transmitter.
n
is dened as n, = 2/T. n ranges
from N/2 + 1 to N/2 as described in the text.
where the N pulse shapes are
g
n
(t) =
1

T
exp [j2f
n
t] w(t). (3)
The scale factor 1/

T has been added to make the energy of each pulse unity. The notation
n mod N is read n modulo N, and is the unique number within the range [0, N] given by n+iN,
where i is an integer (positive or negative). For instance, 4 mod 3 = 4 3 = 1; (2P + 3) mod P =
(2P + 3) 2P = 3; 10 mod 4 = 10 + 12 = 2. Therefore the pulse g
n
(t) is associated with a
n
when n = 0, . . . , N/2; but g
n
(t) is associated with a
n+N
when n = N/2 + 1, . . . , 1. The reason
for this convoluted notation arises from the use of the IDFT to generate the transmitted signal
(see next section).
The frequencies satisfy f
n
f
n1
= 1/T for orthogonality, so assuming that f
0
= 0 we have
f
n
=
n
T
for n = N/2 + 1, . . . , N/2. After up-conversion by the carrier frequency f
c
, the signal
spectrum ranges (approximately) from f
c

N
2T
to f
c
+
N
2T
.
The conceptual transmitter block diagram is shown in Figure 1. The information bit stream
is rst passed through an error correction encoder, and then a baseband pulse modulator, which
maps M bits at a time onto a complex symbol according to some predened signal constellation.
If necessary, symbol streams from other users or services can be multiplexed at this point, and
the resulting sequence interleaved to randomize the allocation of carriers to symbols. This ensures
that, on average, a faded carrier will aect all streams in the multiplex equally. Next, N successive
symbols are buered before each is used to modulate a complex sinusoid. Finally, these sub-carrier
signals are summed, amplied and ltered before being transmitted.
At the receiver, assuming s(t) was transmitted over an AWGN channel so that r(t) = s(t)+n(t)
where n(t) is a white Gaussian process with PSD N
0
, the optimal projection receiver consists of
a bank of lters matched to g
n
(t), since the set g
n
(t) forms an orthonormal basis for the signal
Introduction to Multicarrier Modulation 4
subspace. The output of the nth matched lter is
y
n
=
1

T
_
T
0
r(t) exp(j2f
n
t)dt
= a
n mod N
+v
n
(4)
where v
n
(^(0, N
0
). Due to the orthogonality of g
n
(t), v
n
and v
m
are independent whenever
m ,= n. Equation (4) show that the symbols a
0
, . . . , a
N1
are transmitted over orthogonal
channels, and that the performance in every sense (BEP, spectral eciency, etc.) of the minimum-
distance detector in an AWGN channel is unchanged by the use of MCM.
2.3 A Frequency Selective Channel
MCM or OFDM is only useful when dealing with frequency-selective channels, which cannot be
used at all without equalizers using conventional single-carrier methods. If the channel impulse
response is h(t), the received signal is
r(t) =
_

h()s(t )d +n(t). (5)


If h(t) is non-zero from t = 0 to t = T
m
only, r(t) may be simplied to
r(t) =
_
T
m
0
h()s(t )d +n(t). (6)
If r(t) is processed by a bank of lters matched to g
n
(t), the mth output will be
y
m
=
N1

n=0
a
n
_
T
m
0
_
T
0
h()g
n
(t )g

m
(t)dt d. (7)
Since g
n
(t ), g
m
(t)) is not zero for all values of , it appears that y
m
is a linear combination of
all N symbols a
0
, . . . , a
N1
or in other words, the MCM system has lost its orthogonality in a
time-dispersive channel h(t).
There is no doubt that the last statement is true, but in the next section we will show that
the DFT/IDFT implementation of OFDM reveals a simple solution to the problem.
2.4 Discrete-Time Implementation Using the DFT/IDFT
If OFDM were actually to require N very precise frequency generators in each transmitter and
receiver, it would be a prohibitively expensive system. Practical implementation relies on the fact
that the transmitted signal can be generated using an inverse discrete Fourier transform (IDFT).
Referring to (2), we note that for a given set of symbols a
n
, s(t) has a spectrum which
consists of the weighted sum of a number of sinc functions, each with main-lobe bandwidth 2/T
and centered on f
n
. Clearly the bandwidth is approximately N/T, but a substantial fraction of
the energy of s(t) lies outside f (N/2T, N/2T), and s(t) will not be correctly represented
by sampling it at a rate of N/T. However, if a number of carriers at the edges of the band
(N/2T, N/2T) are unused (i.e. a
n
= 0 for n = N/2p+1, . . . , N/2+p for some integer p N),
then s(t) will approximately be given by its samples taken at the rate of N/T. This means that
s(t) can be reconstructed from its samples s(kT/N). Therefore, assume that s(t) is band-limited
Introduction to Multicarrier Modulation 5
E
E
E
r
r
r
r
r
IDFT
E
E
E
P/S
E i
+
c
E
DAC
Rate
N+P
T
E
a
N1
a
1
a
0
s
0
s
1
s
N1
s
k s(t)
Insert
Cyclic
Prex
Fig. 2: Block diagram of OFDM transmitter with cyclic prex insertion. P is the number
of samples in the cyclic prex.
to within a bandwidth of N/T Hz because the carriers near the band edge are unused. Sampling
s(t) at rate N/T yields
s
k
=
N/2

n=N/2+1
a
n mod N
g
n
_
kT
N
_
=
1

T
N/2

n=N/2+1
a
n mod N
exp (j2f
n
kT/N)
=
1

T
N1

n=0
a
n
exp
_
j
2nk
N
_
. (8)
Note the change in summation limits from N/2 + 1, . . . , N/2 to 0, . . . , N 1 compared to (2).
This is valid because exp(j2nk/N) = exp(j2(n + N)k/N), and replacing the n < 0 complex
sinusoidal terms in (2) with exp(j2(n +N)k/N) gives (8).
Now, the IDFT of a sequence of N complex numbers X
n
is another sequence of N complex
numbers x
k
given by
x
k
=
1
N
N1

n=0
X
n
exp
_
j
2nk
N
_
(9)
when the DFT is dened as
X
n
=
N1

k=0
x
k
exp
_
j
2nk
N
_
. (10)
Therefore, (8) says that we can generate N samples of s
k
through the DFT of a
n
. In addition,
because of our assumption that s(t) is (almost) strictly bandlimited, the continuous-time signal
s(t) to be transmitted can be constructed from s
k
using a digital-to-analog converter at the rate of
N/T Hz. The block diagram of this implementation is shown in Figure 2, where we have ignored
the scaling factor N/

T necessary for a strictly accurate implementation.


Example Consider a multi-carrier system with constant-envelope PSK pulse modulation so that
[a
n
[ =

E
s
for all n. Then the signal s(t) of (2) has a spectral density that is the sum of sinc
functions, each with mainlobe width 2/T and centered at n/T, n = N/2 + 1, . . . , N/2. This is
shown in the top half of Figure 3 where for simplicity, we assume T = 1 and N = 16.
The spectrum of the discrete-time signal s
k
of (8) is a periodic extension in frequency of the
spectrum of

N1
n=0
a
n
exp(j2nt/T)w(t), which is itself a sum of sinc functions separated by 1/T
Hz and centered on n/T Hz, where n = 0, . . . , N 1. This is shown in Figure 3(b).
Introduction to Multicarrier Modulation 6
Next, the digital-to-analog converter (DAC) output is the lowpass ltered version of Figure 3(b)
the passband of the lowpass lter is the region between the two thick dashed lines. Clearly, if the
carriers near the edge of the passband, i.e. n = N/2+1, . . . , N/2+p and n = N/2p+1, . . . , N/2,
are turned o, the spectra of the DAC output and s(t) will be nearly identical. This demonstrates
that the DFT/DAC combination is capable of producing an OFDM signal.
The receiver is designed to be implementable using the DFT it is a bank of lters matched
to the nth carrier waveform g
n
(t) rather than to h(t) g
n
(t) as it should be in theory. Since g
n
(t)
is associated with the (n mod N)th symbol, we label the matched lter outputs accordingly and
nd
y
n mod N
=
_
r(t)g

n
(t)dt =
_
r(t) exp(j2f
n
t)dt, (11)
where again, we have left out the scaling term for simplicity. Note that in this expression, n runs
from N/2 + 1 through N/2. Next, since the signal component of r(t), which is h(t) s(t), is
bandlimited by assumption to (N/2T, N/2T), we can sample r(t) at rate N/T (after lowpass
ltering) without loss of information, and express y
n mod N
as
y
n mod N
=
N1

k=0
r
k
exp
_
j
2nk
N
_
(12)
to within a scale factor
1
.
If we dene n

= n mod N, then n

goes from 0 through N 1, and substitution in (12) yields


y
n
=
N1

k=0
r
k
exp
_
j
2n

k
N
_
(13)
where the right-hand side holds because of the periodicity of the complex exponential. Therefore
y
0
to y
N1
are obtained through the DFT of the samples r
0
through r
N1
, where r
k
= r(kT/N).
This results in the receiver structure shown in Figure 4.
2.5 The Cyclic Prex
The question now is: how do we ensure that y
n
does not suer from interference from symbols
a
m
, m ,= n? The answer comes from a well-known result in digital signal processing which states
that circular convolution in the discrete-time domain is equivalent to multiplication in the discrete-
frequency domain.
To be precise, suppose x
k
is a length-N sequence. Its circular convolution with another se-
quence h
k

k=0,...,N1
is dened as
y
k
= x
k
h
k
=
N1

l=0
h
l
x
(kl) mod N
=
N1

l=0
x
l
h
(kl) mod N
. (14)
This is the same as periodically extending h
l
and x
l
to form the periodic sequences

h
l
and x
l
, and
then summing

h
l
x
kl
over one period of N samples, as illustrated in Figure 5 for k = 0 and N = 3.
Then the DFT of y
k
, k = 0, . . . , N 1, is Y
n
= H
n
X
n
where H
n
=

N1
k=0
h
k
exp(j2nk/N).
The proof of this result is straightforward and left as an exercise.
1
Calculating these scale factors is not important and is left to the interested reader.
Introduction to Multicarrier Modulation 7
15 10 5 0 5 10 15
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
a
N/2+1
Frequency (normalized so that T = 1)
a
N/2
(a)
25 20 15 10 5 0 5 10 15 20 25
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
a
0 a
N1
Frequency (normalized so that T = 1)
(b)
Fig. 3: (a) Spectrum of the desired OFDM baseband continuous-time signal s(t). (b)
Spectrum of the DFT output before (an endless repetition of the fundamental
segment shown in solid lines) and after (the part between the two dashed thick
lines) the DAC.
Introduction to Multicarrier Modulation 8
s(t)
E
h(t)
E i
+
c
n(t)
E
r(t)
ADC
Rate
(N+P)
T
E
Remove
Cyclic
Prex
E
S/P
E
E
E
DFT
E
E
E
y
0
y
1
y
N1
r
0
r
1
r
N1
r
k
Fig. 4: Block diagram of OFDM receiver based on the DFT.
d
d
d
h
l
l 0 1 2
E
d
d
d
x
l
E
d
d
d
x
l
d
d
d
d
d
d
d
d
d

h
l
d
d
d
d
d
d
0 1 2


E
h
k
x
k
[
k=0
Fig. 5: Circular convolution of x
k
and h
k
, both having three non-zero samples, so N =
3. Note that the convolution window can be moved to span any three sampling
intervals, and the result will remain the same.
Introduction to Multicarrier Modulation 9
In OFDM, the baseband received signal sampled at rate N/T is obtained by passing the
sampled transmitted signal s
k
through a linear channel with discrete-time impulse response h
k
,
and then adding receiver noise n
k
i.e.
r
k
= h
k
s
k
+n
k
=
L1

l=0
h
l
s
kl
+n
k
(15)
where denotes linear convolution. h
k
is obtained from the continuous-time channel response h(t)
by setting t = kT/N. In fact, when L N which is the usual case, linear convolution is identical
to circular convolution, except at the beginning and end of the sequence.
To make the two operations exactly identical, we can periodically extend the input x
k
, as
shown in Figure 6, by P samples where P L 1. The input sequence will now have length
N + P, so that the output sequence will have N + P + L 1 samples. The N output samples at
times P + 1 through N +P can be shown to be the output of a circular convolution operation:
y
k
=
L1

l=0
h
l
x
(kl) mod N
= h
k
x
k
, k = P + 1, . . . , N +P. (16)
It is very important to note that the equivalence between linear and circular convolution exists
only under the following conditions:
1. The cyclic prex is longer than the channel delay spread or P L 1;
2. The observation window applied to the output spans the samples P + 1 through N + P
translating the window in either direction invalidates the result.
Assuming these conditions hold, we can now state the following theorem:
Theorem 1: For a discrete-time channel h
k
of length L, and a channel input s
k
that is periodically
extended in its preamble by P L 1 samples where s
0
, . . . , s
N1
= IDFTa
0
, . . . , a
N1
, the
channel output r
k
= h
k
s
k
+n
k
has the property that
r
k
= h
k
s
k
+n
k
, k = P + 1, . . . , N +P, (17)
with representing circular convolution. The DFT of r
P+1
, . . . , r
N+P
is, by the duality between
circular convolution in the time domain and multiplication in the discrete frequency domain,
y = Fr = h
f
a +n (18)
where all vectors have N complex elements, x y is the element-wise product of vectors x and
y, F
n,k
= exp(j2nk/N) is the DFT matrix, h
f
= Fh is the Nth-order DFT of h and n is a
circularly symmetric Gaussian vector with covariance matrix N
0
I.
In scalar notation, the nth DFT output is y
n
= H
n
a
n
+ noise, where H
n
is the nth DFT
coecient of h
0
, . . . , h
L1
,
NL
..
0, . . . , 0. Since the nth matched lter output is independent of a
m
,
m ,= n, the cyclic prex together with the DFT/IDFT implementation of the transceiver creates
N orthogonal at-fading channels.
Introduction to Multicarrier Modulation 10
d
d
d
d d d d d d d d d
h
l

x
N+1l
d d
d
d
d
d
d
d
d
d
d
d
E
y
N+1
= h
k
x
k
[
k=N+1
(Linear convolution)
d
d
d
d d d d d d d d d
h
l

x
N+1l
d
d
d
d
d
d
d
d
d
d
d
d
E
y
k
= h
k
x
k
k = 2, . . . , N + 1
(Linear convolution, but these
N samples identical to output
of circular convolution.)
Fig. 6: Adding a cyclic prex to the input signal makes linear convolution look like circular
convolution, and removes ISI in the OFDM signal. In this example, the input
sequence length is N = 10, the channel response length is L = 3. If the input has
a cyclic prex of length L 1 = 2, the N outputs y
2
, . . . , y
N+1
are obtained by
circular convolution of x
k
and h
k
.
Introduction to Multicarrier Modulation 11
2.6 Single-Tap Equalization
Since y
n
= H
n
a
n
+ v
n
where E[v
m
v

n
] = N
0

m,n
, to recover a
n
one can estimate H
n
(usually
through the periodic transmission of known pilot symbols) and then form the statistic
a
n
=
y
n

H
n
a
n
+
v
n

H
n
(19)
where

H
n
is the latest estimate of H
n
. A slicer is then applied to a
n
to obtain hard decisions on a
n
.
This approach comes from minimum-distance detection applied independently to each y
n
, which
will produce the minimum symbol error probability decision given that v
m
and v
n
are independent.
2.7 Loss in Spectral Eciency
When a cyclic prex of P samples is added to the block of N samples in time, the sampling rate (at
the DAC) needs to increase to (N +P)/T because the N +P samples still need to be transmitted
over T seconds. The lowpass lter at the DAC will have a cuto frequency of (N + P)/2T Hz.
So the spectrum of the transmitted signal with cyclic prex inserted occupies about (N +P)/2T,
whereas without the prex, it occupies only N/2T Hz. In that sense, the prex has wasted P/2T
Hz of bandwidth since it does not carry any information and there is a fractional loss of P/(N+P)
in spectral eciency.
So it is clear that inserting the cyclic prex to solve the ISI problem comes at a price, and in
practice we always strive to make P as small a fraction of N as possible to limit the wastage of
bandwidth.
3 Advanced Topics
3.1 Frequency Oset in a Multi-User OFDMA Uplink
OFDM can be used to provide channel access to a number of users, much like FDMA (frequency
division multi-access) except that frequency bands used by dierent users need not be separated by
a guard band. This gives OFDMA a potentially higher spectral eciency than FDMA. However,
OFDMA is very sensitive to frequency osets, which may come from mismatches in the local
oscillators at transmitter and receiver ends of the link, or Doppler shifts in the case of mobile
applications. We now quantify the eects of frequency osets.
3.1.1 Signal Model
Let T
k
=
k
(1), . . . ,
k
(N
k
) be the set of carriers allocated to user k, and let d
k,p
, p T
k
be
the N
k
symbols transmitted by user k in some OFDM symbol. We assume that the guard interval
inserted into each OFDM symbol is long enough to prevent inter-symbol interference, and hence
the OFDM symbol index will be ignored. For a frequency oset of
k
, the continuous-time signal
transmitted by user k will be
x
k
(t) =
_

_
e
j
k
t

pP
k
d
k,p
e
jpt
0 < t T
e
j
k
t

pP
k
d
k,p
e
jp(t+T)
T
g
< t 0
(20)
where T is the OFDM symbol interval, and T
g
is the guard interval.
p
denotes the reference
frequency for the pth carrier, which in baseband-equivalent form would be
p
= 2p/T. (With
Introduction to Multicarrier Modulation 12
pulse shaping to limit the bandwidth of the transmitted signal, the actual transmitted signal will
be x
k
(t) p(t) where p(t) is the impulse response of the pulse-shaping lter. However, since p(t)
can be absorbed into the channel, we lose no generality in treating x
k
(t) as the transmitted signal
and gain notational simplicity.)
The signal received from user k will be the convolution of x
k
(t) with the kth channel impulse
response h
k
(t), or
y
k
(t) = x
k
(t) h
k
(t). (21)
If we sample the received signal above its Nyquist rate and discard the cyclic prex, we nd that
the linear convolution becomes equivalent to a circular or periodic convolution, or
y
k
[n] = x
k
[n] h
k
[n] (22)
with n = 0, . . . , N 1 where N is the number of samples in T seconds, and x
k
[n] = x
k
(nT/N).
Note that the sampling rate is N/T if the carriers near the band edges are not used, then the one-
sided bandwidth of the OFDM signal is N
c
/2T, and hence N = N
c
will yield sucient statistics
for the detection of all transmitted symbols. We therefore assume that N = N
c
.
The discrete-time Fourier transform (DTFT) of y
k
[n] is simply the product of the DTFTs of
x
k
[n] and h
k
[n], or
Y
k
() = X
k
()H
k
(). (23)
We can show from rst principles that X
k
(F) is a sum of frequency-shifted Dirichlet functions
2
:
X
k
() =

pP
k
d
k,p
drcl ( (
k
+
p
)T/N
c
, N
c
) (24)
where
drcl(x, N)

=
sin(Nx/2)
N sin(x/2)
. (25)
To be more concise, we can write X
k
() =

pP
k
d
k,p
S
k,p
().
The complete received signal in the frequency domain is therefore
Y () =
K

k=1
Y
k
() +W() =
K

k=1
H
k
()X
k
() +W() (26)
=
K

k=1
H
k
()

pP
k
d
k,p
S
k,p
() +W(), (27)
where W() is complex AWGN. In vector notation, treating Y (), < as a vector y and
similarly converting all other frequency functions to vectors, we have
y =
K

k=1
H
k
S
k
d
k
+w (28)
where H
k
= diag(h
k
), h
k
is the vector representing H
k
(), S
k
= [s
k,
k
(1)
, . . . , s
k,
k
(N
k
)
] and
d
k
= [d
k,
k
(1)
, . . . , d
k,
k
(N
k
)
]. Finally, we can also write
y = HSd +w (29)
with H = [H
1
, . . . , H
K
], S = diag[S
1
, . . . , S
K
] and d = [d
T
1
, . . . , d
T
K
]
T
.
2
Complex scaling term ignored.
Introduction to Multicarrier Modulation 13
4 3 2 1 0 1 2 3 4
0.4
0.2
0
0.2
0.4
0.6
0.8
1
(rad/s)
Fig. 7: 15 instances of S
k,p
(), with zero frequency osets. N
c
= 16.
3.1.2 Interpretations
The Dirichlet function dened in (25) is central to our analysis so we should study it more closely.
The following properties are important:
1. drcl(x, N) has zero crossings at x = 2p/N for integer p;
2. drcl(0, N) = 1.
In this sense, the Dirichlet function is similar to the sinc function sinc(Nx/2) where sinc(x) =
sin x/x. In fact, a plot of the two functions would be virtually indistinguishable.
If all users transmit over AWGN channels with unit gain, i.e. h
k
= 1, then the frequency-
domain received signal will be a linear combination of Dirichlet functions, plotted in Figure 7 where
we assumed 16 carriers. This is the spectrum or discrete-time Fourier transform (DTFT) of the
sequence y[n].
When the channels are frequency selective, the received signal will be a linear combination
of distorted versions of these Dirichlet functions, namely H
k
()S
k,p
(), such as those plotted in
Figure 8. In this gure, the vertical lines indicate the frequencies = 2p/N
c
notice that at
these frequencies, all but one of the functions are zero. This is due to these frequencies being at
the zero crossings of the shifted Dirichlet functions of Figure 7.
In the presence of frequency osets, the received signal becomes a linear combination of shifted
Dirichlet functions, which no longer all have zero crossings at = 2p/N
c
. For instance, with four
users each having its own frequency oset, S
k,p
() will look like the curves in Figure 9.
Introduction to Multicarrier Modulation 14
4 3 2 1 0 1 2 3 4
0.5
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
(rad/s)
Fig. 8: 15 instances of H
k
()S
k,p
(), with zero frequency osets.
3.1.3 Interference Introduced by Frequency Oset
Given that the desired component of y(t) is bandlimited to [N
c
/2T, N
c
/2T], sampling at the rate
of N
c
/T over one OFDM symbol interval will yield sucient statistics for detecting d. So the N
c
samples y(nT
c
/N), n = 0, . . . , N
c
1, are sucient statistics, and any invertible function applied
to them will yield sucient statistics too. In particular, the continuous-frequency function Y ()
must be sucient for detecting d since the DTFT is invertible. But since there are a total of N
c
independent symbols transmitted using N
c
linearly independent vectors, the set y(nT
c
/N), n =
0, . . . , N
c
1 is a minimal set of sucient statistics, whereas the function Y () is not minimal.
A minimal set of sucient statistics in the frequency domain is obtained by sampling Y () at
N
c
points in [, ]
3
. The most convenient N
c
points to take are of course those evenly spaced
over [, ], in other words Y (2p/N
c
), p = N
c
/2, . . . , N
c
/2. These are directly found from the
discrete Fourier transform (DFT) of y[n].
We already established that in the absence of frequency oset, the frequency-domain received
signal is
y = HSd +w (30)
where the columns of HS have the property that their zero crossings match. Sampling y at intervals
of 2/N
c
will give
y = Ad + w C
Nc
(31)
where A is a diagonal matrix, made up of the respective channel frequency responses at frequencies
2p/N
c
.
3
This is guaranteed since any set of N
c
points in Y () is generated from a multiplication of the time-
domain received-signal vector with an invertible matrix.
Introduction to Multicarrier Modulation 15
2 1.5 1 0.5 0 0.5 1 1.5 2
0.2
0
0.2
0.4
0.6
0.8
1
(rad/s)
Fig. 9: Some of the frequency-domain basis functions for x
k
[n], with frequency oset. Ob-
serve that at frequencies = 2p/N
c
, marked by the dashed vertical lines, several
of these functions will be non-zero, in contrast to the zero oset case.
Introduction to Multicarrier Modulation 16
Since Ais diagonal, there is no interference in the minimal frequency-domain sucient statistics
y, and optimal detection proceeds on a per-user basis.
When frequency oset is present, then the columns of HS do not have the same zero crossings
(except those belonging to the same user), as we can see from Figure 9. Uniform sampling of Y ()
will then give (31) but with a non-diagonal A, indicating non-zero multi-user interference.
The most important point proven here is that frequency oset in an OFDMA system creates
multi-user interference. Whether the osets can be estimated and then used in multi-user detection,
or whether practical frequency osets are small enough to avoid serious performance issues are
topics for future discussion.

Potrebbero piacerti anche