DIGITAL DATA TRANSMISSION — 2023/2024

MODULATIONS

1 MODULATIONS
Another application of the mathematical theory of signal space and orthonormal expansions are modulations. In
the previous Topics, we saw how to represent signals using coefficients (samples), namely a sequence of discrete
time values that using source coding are eventually converted to a sequence of bits. Conversely, in this topic we
will see how to construct signals from a sequence of discrete-time bits, possibly the output of a channel code. One
example is the wireless channel: data bits are mapped to electromagnetic waves that propagate through space and
are eventually measured at a receiver antenna. The process that maps bits to signals is called modulation.

We start with a rather useless but didactic example that quite resembles to the repetition code of information theory.
Example 1.1. Assume a modulation able to transmit 1 bit only using a signal that lasts forever. Let D denote the
data bit, taking values over the alphabet {0, 1} with equal probability. Then, a modulated signal could simply be
{
x0 (t) D = 0
X(t) = (1.1)
x1 (t) D = 1,
for t ≥ 0. We recall the use of capital letters, for both the data bits and the modulated signal, since both are random
quantities. For example, at t = 5 seconds, X(5) can be seen as a random variable taking values over the alphabet
{x0 (5), x1 (5)} with equal probability.

We note that Example 1.1 directly maps bits to waveforms without making use of any signal space or orthonormal
expansion. We saw in the previous Sections that a signal belonging to the signal space spanned by a set of orthonor-
mal functions is determined by its coefficients. Therefore, a previous natural step is to map bits to coefficients, and
then coefficients to signals.
Example 1.2. Assume that we want to map the sequence of bits D1 , D2 , . . . into real numbers. If we use the rule
{
+1 Dj = 0
Xj = (1.2)
−1 Dj = 1
for j = 1, 2, . . ., we will be generating one real number Xj (symbol) for every data bit Dj . Hence, the mapping (1.3)
assigns one bit per real symbol.
Example 1.3. Another example is the rule


+1 (D2j−1 , D2j ) = (0, 0)


+ 1 (D2j−1 , D2j ) = (1, 0)
Xj = 2 (1.3)

− 1
(D2j−1 , D2j ) = (0, 1)

 2
−1 (D2j−1 , D2j ) = (1, 1).

for j = 1, 2, . . .. In this case, we take tuples of two bits to generate a real number, that is, two bits per symbol.
Definition 1.1. A rate r binary-to-symbol mapping is a one-to-one assignment from bits D1 , D2 , . . . to coefficients
(symbols) X1 , X2 , . . . such that it outputs 1 symbol every r bits.

We have rate 1 and rate 2 in Examples 1.2 and 1.3, respectively, but we could have a fractional number such as rate
0.5, i.e., a modulation that outputs 2 symbols for every bit.

1
Definition 1.2. For a given binary-to-symbol mapping, the set of all possible values that X can take is called the
constellation X of the mapping. The size or cardinality of the constellation is denoted as M = |X |.

We have constellations {−1, +1} with M = 2 and {−1, −1/2, +1/2, +1} with M = 4 in Examples 1.2 and 1.3.

The next step is to construct a signal using the linear combination of orthonormal functions and the real numbers.
As the reader may predict, the symbols of the output of a binary-to-symbol mapping serve as coefficients of a signal
in a given signal space.
Definition 1.3. For a given rate r binary-to-coefficient mapping, an orthonormal modulation is the process that
produces the signal using n coefficients X1 , . . . , Xn using k = nr bits D1 , . . . , Dk as

∑
n
X(t) = A Xi ϕi (t), (1.4)
i=1

where {ϕ1 , . . . , ϕn } are real-valued orthonormal waveforms determined in advance, and A ≥ 0 is a scaling factor.
For simplicity, we will always assume that k and n = k/r are integer numbers.

We schematically summarize the modulation process from bits to signal in Figure 1.1.

D1 , . . . , Dk Bit-to-symbol X1 , . . . , Xn
Bit source Modulator X(t)
mapper

Figure 1.1: Bits to signal process.

The energy of X(t) according to its definition is given by

∫∞ ∑
n
EX = X2 (t)dt = ⟨X, X⟩ = A2 X2i , (1.5)
−∞ i=1

where we used that the waveforms ϕi (t) are orthonormal. We recall once more the use of capital letters as X(t),
X1 , . . . , Xn and even the energy (1.5) are random quantities that depend on the (k, n) binary-to-reals encoder
process and on the probability distribution of the sequence of bits D1 , . . . , Dk .

Every pair of bit-to-coefficient mapping and modulator (1.4) can be described using parameters with physical mean-
ing such as the bit period Tb and the symbol period Ts both in seconds, the bit rate Rb in bits per second, the symbol
rate Rs in symbols per second, the average energy of a signal EX in Joules, the average energy per symbol Es in Joules
per symbol, the average energy per bit Eb in Joules per bit and the average power P in Watts. We will go over these
concepts through an example in the following Problem.
Problem 1.1. Consider a discrete memoryless source taking values over the alphabet {0, 1} (bits) with probability
distribution {p, 1 − p}. Assume that the source outputs bits at a rate of Rb bits per second, and that a modulated
signal X(t) is generated according to (1.4), where the symbols X1 , . . . , Xn are the output of the bit-to-coefficient
mapping described in Example 1.3. For this example,

(a) Calculate Tb and Ts .

(b) Calculate Rs .

(c) Determine the probability distribution over the constellation X .

2
 (d) Calculate the average energies EX , Es and Eb .
(e) Calculate the average power P.
Solution. One bit is sent every 1/Rb seconds. Since every symbol carries two bits, one symbol is sent every 2/Rb seconds.
The symbol rate is thus Rb /2 bits per second.


Since bits are independent, the probabilities for the constellation symbols are given by P(+1) = p2 , P 12 = p(1 −


p) = P − 12 , and P(−1) = (1 − p)2 . Therefore, the average signal energy EX is given by
 1  1
EX = nA2 E[Xi ]2 = nA2 p2 · 1 + 2p(1 − p) + (1 − p)2 · 1 = nA2 (2 − 3p + 3p2 ). (1.6)
4 2
The average symbol energy is Es = 12 A2 (2 − 3p + 3p2 ), and the average bit energy is Es = 41 A2 (2 − 3p + 3p2 ).

Assume that the signal X(t) is transmitted over an ideal channel introducing no distortion at the output of the
channel such that Y(t) = X(t). In this case, we may use the orthonormality property of the waveforms ϕi (t) to
recover every coefficient X1 , . . . , Xn from Y(t) by means of the inner product using
1
Xj = ⟨X, ϕj ⟩. (1.7)
A
The recovery process is summarized in Figure 1.2.

A ⟨X, ϕ1 ⟩
1
X1

..
.

A ⟨X, ϕj ⟩
1 Xj
X(t)

..
.

A ⟨X, ϕn ⟩
1
Xn

Figure 1.2: Recovering the coefficients form the signal.

Problem 1.2. Consider that we generate a signal X(t) as
X(t) = A (X1 ϕ1 (t) + X2 ϕ2 (t)) , (1.8)
where ϕ1 (t) and ϕ2 (t) have unit energy but are not orthogonal. Let ρ = ⟨ϕ1 , ϕ2 ⟩ = ⟨ϕ2 , ϕ1 ⟩ for 0 ≤ ρ ≤ 1.
Denote as Zj as the output of the receiver (1.7), i.e., Zj = A1 ⟨X, ϕj ⟩. Check that Zj has inter-symbol interference for
j = 1, 2. Can we always still recover X1 and X2 ?
Solution. Since the inner product is a linear operation, we have that
1 1 1
Z1 = ⟨X, ϕ1 ⟩ = ⟨A(X1 ϕ1 + X2 ϕ2 , ϕ1 ⟩ = (AX1 ⟨ϕ1 , ϕ1 ⟩ + AX2 ⟨ϕ2 , ϕ1 ⟩) = X1 + ρX2 . (1.9)
A A A
Similarly, we have Z2 = X2 + ρX1 . Therefore, we may recover the symbols from
Z1 − ρZ2
X1 = (1.10)
1 − ρ2
Z2 − ρZ1
X2 = . (1.11)
1 − ρ2

3
Clearly, when ρ = 0, we end up with (1.7). However, when ρ = 1, we have Z1 = Z2 and the underdetermined
system of equations cannot be solved.

Problem 1.3. We generate a signal X(t) from two orthonormal waveforms ϕ1 (t) and ϕ2 (t), two known amplitudes
A1 and A2 , and two unknown coefficients C1 and C2 , using

X(t) = (A1 C1 + A2 C2 ) ϕ1 (t) + (A1 C1 − A2 C2 ) ϕ2 (t). (1.12)

How can we recover the unknown coefficients C1 and C2 from X(t)?

Solution. Using the inner products ⟨X, ϕ1 ⟩ and ⟨X, ϕ2 ⟩ and the orthonormality condition between ϕ1 and ϕ2 we
obtain

X1 = ⟨X, ϕ1 ⟩ = A1 C1 + A2 C2 (1.13)
X2 = ⟨X, ϕ1 ⟩ = A1 C1 − A2 C2 . (1.14)

Therefore, solving for C1 and C2 we may easily recover them calculating

X1 + X2
C1 = (1.15)
2A1
X1 − X2
C2 = . (1.16)
2A2

It is trivial to see that when ϕ1 , . . . , ϕn are not orthonormal, the recovery process (1.7) may lead to inter-symbol
interference (ISI) even without the presence of noise. Although we could still recover the symbols from any arbitrary
non-orthonormal basis, the orthonormality condition simplifies not only the analysis but also the implementation
of modulations. Therefore, preserving the orthonormality condition ϕ1 , . . . , ϕn is a key aspect that, as we will see
next, has implications in the time and frequency tradeoff of the modulation.

Note that the modulation process (1.4) has no assumptions on the time span of the waveforms ϕ1 (t), . . . , ϕn (t).
For example, all could be time-limited to T seconds, i.e., in the [−T/2, T/2] interval, so that all n symbols and all
k bits would be transmitted in T seconds. Alternatively, the waveforms ϕi (t) could spread a longer period such
that the total duration of X(t) is nT . The special case where ϕi (t) are shifted versions of a single waveform ϕ(t),
making the implementation of the receiving process (1.7) particularly simple, is called pulse amplitude modulation
(PAM).

1.1 Pulse amplitude modulation

Let ϕ(t) be a unit-energy waveform without any further assumption on time or frequency limitation. For every
i = 1, . . . , n, we let ϕi (t) be a time shifted version of ϕ(t) according to

ϕi (t) = ϕ(t − iT ), (1.17)

where T is the symbol period. Then, the transmitted signal X(t) according to (1.4) is given by
∑
n
X(t) = A Xi ϕ(t − iT ). (1.18)
i=1

In PAM, the symbols Xi are usually located evenly in the real line around the origin such that the constellation is
the set
X = {−(2ν − 1), . . . , −5, −3, −1, +1, +3, +5, . . . , +(2ν − 1)} (1.19)

4
where ν is a positive integer. This constellation has a total of M = 2ν points, hence each symbol in the constellation
can be seen as the output of a binary-to-symbol mapping of rate r = log2 (M) bits per symbol. We refer to this
modulation as M-PAM. Typical choices are ν = 1, a 2-PAM representing r = 1 bit per symbol, ν = 2 leading to a
4-PAM representing r = 2 bits per symbol, or a 8-PAM representing r = 3 bits per symbol.

For PAM, the recovery process (1.7) simplifies to

1
Xj = (X ⋆ h) (jT ), (1.20)
A
where ⋆ denotes the convolution, and h(t) = ϕ∗ (−t) is usually called the matched filter. As depicted in Figure 1.3,
we simply need to sample the output of the matched filter at every multiple of T and then rescale it by a factor of A1 .

jT
X(t) h(t) Xj

Figure 1.3: Recovering the coefficients in PAM using a matched filter.

Problem 1.4. Let Z(t) be the output of the convolution between X(t) and the matched filter h(t) ∗
∫∞ = ϕ (−t), that is,
Z(t) = (X ⋆ h)(t). Using the definition of the convolution between two signals (x ⋆ y)(t) = −∞ x(τ)y(t − τ)dτ,
check that Z(iT ) = AXj and therefore that (1.20) is satisfied.

1.2 Nyquist’s criterion

Note that for (1.20) to hold, the orthonormality condition on ϕ1 , . . . , ϕn must be preserved, otherwise inter-symbol
interference could occur (again, even without the presence of noise). The orthonormality condition of the wave-
forms in PAM is called shift-orthonormality.
Definition 1.4. For a given T > 0, a waveform ϕ(t) is shift-orthonormal if it satisfies (??) ⟨ϕi , ϕj ⟩ = δij among its
shifted versions ϕi = ϕ(t − iT ) and ϕj = ϕ(t − jT ). That is,
∫∞
ϕ(t − iT )ϕ∗ (t − jT )dt = δij (1.21)
−∞

A natural choice for ϕ(t) satisfying (1.21) is the rectangular pulse that is time-limited to T seconds:
{
√1
T
−T/2 ≤ t < T/2
ϕ(t) = (1.22)
0 elsewhere,

since clearly its shifted versions do not overlap in time and therefore the shift-orthonormality condition is satisfied.
However, this waveform has infinite frequency components. In this Section, we will see examples of waveforms that
satisfy the shift-orthonormality condition and that are frequency-limited at the same time. To study them, we first
note that (1.21) only depends on the relative difference between i and j. Then, the shift-orthonormal condition can
be written in terms of the self-similarity function defined as follows.
Definition 1.5. The self-similarity function Rϕϕ (τ) of an energy-limited waveform ϕ(t) is given by
∫∞
Rϕϕ (τ) = ϕ(t + τ)ϕ∗(t)dt. (1.23)
−∞

In the course, we will only be interested in the following properties of the self-similarity function. You make check
them using the proper change of variables in the self-similarity function:
5
 • At τ = 0, we recover the energy of the waveform, i.e., Rϕϕ (0) = Eϕ .

• It can be seen as the convolution between ϕ(t) and the matched filter h(t) = ϕ∗ (−t), i.e., Rϕϕ (τ) =
(ϕ ⋆ h) (τ).

• Its Fourier transform is given by F [Rϕϕ ] (f) = |ϕ̂(f)|2 .

Proposition 1.1. The shift-orthonormality condition (1.21) is equivalent to the condition

{
1 i=0
Rϕϕ (iT ) = δi = (1.24)
0 i ̸= 0.

Theorem 1.1 (Nyquist’s criterion). A waveform ϕ(t) satisfies the shift-orthonormality condition (1.21) if
∞
∑   2
i
ϕ̂ f − = T. (1.25)
T
i=−∞

∑
Proof. Left as exercise. As a hint, we may write the l.h.s. of (1.25) in terms of a Dirac delta comb ∞−∞ δ(t − iT )
and the r.h.s. as a Dirac delta δ(t), and treat the problem as a sampling problem by finding the Fourier transforms
in both sides.
∑ You may use ithe
fact that the Fourier transform of the Dirac delta comb is another Dirac delta comb
given by T1 ∞ i=−∞ δ f − T .

Problem 1.5. Check that the waveform ϕ(t) = √1 sinc(t/T ) satisfies the shift-orthonormality condition (1.24) and
T
the Nyquist’s criterion (1.25).

While the rectangular pulse (1.22) is time-limited to T seconds and has infinite frequency components, the Nyquist
1
pulse in Problem 1.5 is frequency limited to W = 2T Hz while it spreads over time. In practice, digital communi-
cation systems may use waveforms satisfying the Nyquist’s criterion (hence the shift-orthonormality condition to
1
avoid ISI) that have bandwidth a little larger than 2T at the benefit of concentrating the time components around
the center of the waveform in the [−T/2, T/2] interval. This is summarized in the following result.
Corollary 1.1. Let T > 0 and let ϕ(t) a unit-energy waveform that is frequency-limited to W Hz. If ϕ(t) satisfies
the shift-orthonormality condition (1.21), then
1
W≥ , (1.26)
2T
with equality if ϕ(t) = √1 sinc(t/T ). This implies that for a given bandwidth W, we can transmit at most 2W
T
symbols per second.
Problem 1.6. In Gigabit Ethernet, every twisted cable propagates waveforms up to W = 62.5 MHz. Using adequate
isolation materials, 4 twisted cables are able to simultaneously transmit signals at the same time without causing
interference. A 5-PAM modulation is used in every twisted cable to transmit information data but also some control
data. Before the bit-to-symbols and modulation process, a channel code of rate 89 is used to encode information data.
For this communication protocol,

(a) Determine the minimum symbol period T in seconds and maximum transmission rate Rs of Cat6 Ethernet
in symbols per second.

(b) How many bits are required to simultaneously map four 5-PAM modulations?

(c) Assuming that 1 bit is reserved for control data, what is the total transmission rate Rb in bits per second?

6
One example of waveform ϕ(t) satisfying all the Nyquist conditions above that has a convenient time-frequency
tradeoff are raised-cosine waveforms, whose squared Fourier transform is cosine-shaped according to


 0 ≤ |f| ≤ 2T
1−β

T   
|ϕ̂(f)|2 = T2 1 + cos πTβ |f| − 2T
1−β
2T < |f| < 2T
1−β 1+β
(1.27)


0 elsewhere.
In the time domain, the waveform is given by
1
2β cos ((1 + β)πt/T ) + 4βt/T sin ((1 − β)πt/T )
ϕ(t) = √ . (1.28)
π T 1 + (4βt/T )2
and its self-similarity function is given by
 τ  cos(πβτ/T )
Rϕϕ (τ) = sinc . (1.29)
T 1 − 4β2 τ2 /T 2
In (1.27) and (1.28), β is a parameter in the [0, 1] interval that adjusts the bandwidth of the raised-cosine waveform,
given by W = 1+β 2T .
Problem 1.7. Use a computer mathematical tool to check that the raised-cosine waveforms satisfy the Nyquist’s
criterion (1.25) for example for β ∈ {0, 1/2, 1}. For these choices of β, plot both ϕ(t) and |ϕ̂(t)|2 .

1.3 Quadrature amplitude modulation

For the sake of completeness, we introduce here the quadrature amplitude modulation (QAM) that was studied
in the first seminar of the course. QAM is a passband modulation, that is, the QAM signal X(t) uses frequency
components located around some carrier frequency fc in contrast to PAM which is a baseband modulation whose
frequency components are located around f = 0. For both baseband and passband modulations, we define the
bandwidth as follows.
Definition 1.6. The bandwidth of a waveform ϕ(t) is the support of the Fourier transform ϕ̂(f) in the positive
frequencies.
Definition 1.7. The spectral efficiency of a modulation is the ratio between the transmission rate and the required
bandwidth. It can be expressed in bits per second per Hz, real-valued symbols per second per Hz, or complex-valued
symbols per second per Hz.
Example 1.4. A first possibility to construct
√ a passband signal is modulating a PAM signal in (1.18) by multiplying
the shift-orthonormal waveform ψ(t) by 2 · cos(2πfc t). For ψ(t) to be shift-orthonormal, we obtained in Corol-
1 1
lary 1.1 that the minimum bandwidth W = 2T , hence a PAM signal is able to transmit WT = 2 real-valued symbols
√
per second and per Hz. Let ϕ(t) = 2 · ψ(t) · cos(2πfc t). The Fourier transform of ϕ(t) is then
ψ̂(f − fc ) + ψ̂(f + fc )
ϕ̂(f) = √ . (1.30)
2
When fc > W, where W is the bandwidth of ψ(t), the Fourier transforms ψ̂(f − fc ) and ψ̂(f + fc ) do not overlap
in the frequency domain. Therefore, by Theorem (??), the energy of ϕ(t) is
∫ ∫
1 ∞ 1 ∞ 1 1
Eϕ = Eϕ̂ = |ψ̂(f − fc )|2 dt + |ψ̂(f + fc )|2 dt = + = 1, (1.31)
2 −∞ 2 −∞ 2 2
where we used that ψ̂(f − fc ) and ψ̂(f + fc ) both have unit energy. Likewise, we have that
∑∞   ∞   ∞  
i 2 1 ∑ i 2 1 ∑ i 2
ϕ̂ f − = ψ̂ f − fc − + ψ̂ f + fc − = T, (1.32)
T 2 T 2 T
i=−∞ i=−∞ i=−∞

7
where we also used that both ψ̂(f − fc ) and ψ̂(f + fc ) satisfy the Nyquist’s criterion (1.25). In conclusion, ϕ(t) is a
passband waveform that preserves shift-orthonormality.

In Example 1.4, we observe that the modulation process doubles the bandwidth, i.e., ϕ(t) has twice the bandwidth
of ψ(t). As a consequence, the spectral efficiency of the passband waveform (1.30) is half of the spectral efficiency
of ψ(f), namely only 1 real-valued symbol per second per Hz. We will how a QAM signal also preserves shift-
orthonormality while keeping the efficiency of a PAM.

Let ψ(t) be a baseband shift-orthonormal waveform of symbol period T . We define ϕ1 (t) and ϕ2 (t) as
√
ϕ1 (t) = 2 · ψ(t) · cos(2πfc t) (1.33)
√
ϕ2 (t) = − 2 · ψ(t) · sin(2πfc t). (1.34)

Using Parseval’s theorem, one can show that ϕ1 (t) and ϕ2 (t) are orthonormal. In Example 1.4 we already saw that
ϕ1 (t) is shift-orthonormal and it is straightforward that ϕ2 (t) is shift-orthonormal as well. The only condition that
remains to be checked is that all the shifts of ϕ1 (t) and ϕ2 (t), namely the waveforms ϕ1 (t − iT ) and ϕ2 (t − kT )
preserve orthonormality. We encourage the reader to check that this condition is also true. As a consequence,
{ϕ1 , ϕ2 } is a two-dimensional shift-orthonormal system that allows to simultaneously transmit two symbols B and
C free of inter-symbol interference. The QAM signal can be written as

∑
n
X(t) = A (Bi ϕ1 (t − iT ) + Ci ϕ2 (t − iT )) , (1.35)
i=1

where Bi and Ci are real-valued symbols. The QAM symbols are mathematically represented using the complex-
valued symbol Xi = Bi + jCi , where Xi belongs to some complex-valued constellation X ⊂ C. Typical QAM
constellations are phase-shift keying (PSK).
{ 1 M−1
}
X = 1, ej2π M , . . . , ej2π M (1.36)

or constellations that are product of two PAM constellations (1.19), for example the M-QAM with M = J2 complex-
valued symbols where both B and C are J-PAM symbols.

We end this Section by noting that, similarly to the PAM modulation, the QAM demodulation process can also be
implemented using matched filters, as in Figure 1.4.

iT Bi
ϕ∗1 (−t)

X(t) Xi = Bi + jCi
iT
ϕ∗2 (−t)
Ci

Figure 1.4: Recovering the coefficients in QAM using matched filters.