PAM Notes
PAM Notes
PAM Notes
where p(t) is any unit energy pulse such as those described in Appendix B. The notation developed
in the preceding sections is extended to denote a symbol sequence with a new symbol occurring
every Ts seconds. The resulting PAM signal is
X
s(t) = a(n)p(t − nTs ) (2.21)
n
where the subscript “0” has been dropped on the symbol value a and an index n has been added
to denote the n-th symbol in a sequence of symbols. The subscript n on the summation symbol
means that the sum is over n and may be finite (for a finite data sequence) or infinite if an infinite
data sequence is assumed. When it is important, limits on the summation will be explicit.
Some samples constellations for baseband PAM are shown in Figure 2.6. Note that except for
M = 2, the constellations points do not all have the same energy. For M = 4 the average symbol
energy is
M2 − 1
Eavg = . (2.25)
3
For each constellation, the minimum Euclidean distance is 2.
Linear Modulation 57
–1 +1
(a)
–3 –1 +1 +3
(b)
–7 –5 –3 –1 +1 +3 +5 +7
(c)
Figure 2.6: Baseband PAM constellations for M = 2 (a), M = 4 (b), M = 8 (c), and general M ,
(d).
58 2.2 M-ary Baseband Pulse Amplitude Modulation (PAM)
that drives the pulse shaping filter with impulse response p(t). The output of the pulse shaping
filter is the pulse train given by (2.21).
The matched filter PAM detector is illustrated in the lower portion of Figure 2.7. The received
waveform is the impulse train (2.21) plus noise. The matched filter is a filter whose impulse re-
sponse is a time reversed version of the pulse shape as described in Section 2.1.4. The matched
filter output, x0 (t) is sampled at Ts -spaced intervals to produce a sequence of signal space projec-
tions that are used for detection. Detection uses the decision rule (2.14). The detector presented
in Figure 2.7 is an idealized system. It is assumed that the time instant the matched filter output is
to be sampled is known precisely. In a real system, this knowledge is provided by a symbol timing
synchronization subsystem.
The mathematical expressions for these operations are as follows. Let
be the received signal where s(t) is the PAM pulse train given by (2.21) and w(t) is the additive
noise1 . The matched filter output is
Z T2 +t
x0 (t) = r(t) ∗ p(−t) = r(λ)p(λ − t)dλ. (2.28)
T1 +t
X Z T2 +t Z T2 +t
x0 (t) = a(l) p(λ − lTs )p(λ − t)dλ + w(λ)p(λ − t)dλ (2.29)
l T1 +t T1 +t
The noise term w(t) will be modeled as a zero-mean white Gaussian random process with power spectral density
1
N0 /2 Watts/Hz.
Linear Modulation 59
where v(t) is the output of the matched filter due to the additive noise at the input (the second
integral in (2.29)) and rp (τ ) is the pulse shape autocorrelation function defined by
Z T2
rp (τ ) = p(t)p(t − τ )dt (2.31)
T1
as described in Appendix B.
The matched filter output is sampled at t = kTs to produce the signal space projection
X
x0 (kTs ) = a(l)rp ((l − k)Ts ) + v(kTs ). (2.32)
l
As discussed in Appendix B, the autocorrelation function for all full response unit-energy pulse
shapes satisfies
1 m = 0
rp (mTs ) = . (2.33)
0 m 6= 0
Condition (2.33) also holds for partial response pulse shapes that satisfy the Nyquist no-ISI condi-
tion. As a consequence (2.32) may be expressed as
This shows that the signal space projection consists of the true point plus a noise term. The effect
of the noise term is to perturb the position of x0 (kT ) in the signal space. The symbol decision
follows the rule (2.14) or the equivalent form given by (??) as described in Section ??. The detector
produces an erroneous output (i.e. a symbol error) when the noise term forces x0 (kTs ) to be closer
to a point in the signal space other than the point corresponding to the waveform that was sent. The
probability of error for M-ary PAM is derived in Section 2.2.3.
An important conceptual tool is a modulo-Ts plot of the matched filter output x0 (t). Examples
of these plots for the NRZ, MAN, HS, and SRRC pulse shapes are shown in Figure 2.8. Note that
the time axis on the plots are shifted so that the optimum sampling instant occurs in the middle of
the plot. The overlay is used to present all the possible trajectories of rp (τ ) as determined by the
symbol sequence. These plots for the HS and SRRC pulse shapes resemble a human eye. for this
reason, these plots are called eye diagrams.
60 2.2 M-ary Baseband Pulse Amplitude Modulation (PAM)
i (t ) = ∑ a (k )δ (t − kTs ) t
k
s (t ) = ∑ a (k ) p (t − kTs )
pulse
LUT
bits S/P shap ing
0 k
filter
log 2 M bits h (t ) = p (t )
x0 (t ) x0 (kTs )
r (t ) â (k )
matched decision
filter
t = kTs
Figure 2.7: A PAM modulator (top) and detector (bottom) using continuous-time processing.
Linear Modulation 61
(a) (b)
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−0.5 0 0.5 −0.5 0 0.5
t/T t/T
s s
(c) (d)
1.5
1
1
0.5
0.5
0 0
−0.5 −0.5
−1
−1
−1.5
−0.5 0 0.5 −0.5 0 0.5
t/T t/T
s s
Figure 2.8: Eye diagrams for (a) the NRZ pulse shape, (b) the MAN pulse shape, (c) the HS pulse
shape, and (d) the SRRC pulse shape with a 0.5 roll-off factor.
62 2.2 M-ary Baseband Pulse Amplitude Modulation (PAM)
where T1 and T2 define the time support of the pulse shape p(t). The matched filter output is
downsampled at n = k TTs to produce the sequence of signal space projections x0 (kTs ). The k-th
symbol estimate is made using the decision rule (2.14).
This detector is an idealized system in two ways. First, it is assumed that the ADC produces an
integer number of samples/symbol. In other words, that Ts /T is an integer. Since this assumption
is rarely true in practice, a resampling filter is often included in the system block diagram. The
second assumption is the exact symbol timing instant is known and coincides with a sample. In
reality, this will not be the case and a discrete-time symbol synchronizer must be used to estimate
the phase and frequency of the downsampling operation.
Linear Modulation 63
There are at least two clock rates in this system. The first clock rate is the sampling rate
used by the ADC to sample the received waveform. The sample rate must satisfy the sampling
theorem. Unlike the modulator, however, there is little need to sample the signal any higher than
the minimum rate given by the sampling theorem2 . The matched filtering occurs at the high clock
rate and the output of the matched filter is downsampled by N to one sample/symbol. The clock
rate here is the symbol rate. Symbol decisions are made at the clock rate. There is a third clock
rate, the bit clock, if the symbol decisions are mapped to log2 M bits.
The mathematical expressions for the receiver operations mimic those for the continuous-time
case. In this text, signal energies and correlations are referenced to continuous-time since they
were defined for the continuous-time signals3 . For example, the pulse correlation function can be
expressed in terms of T -spaced samples of p(t) using the backward difference approximation for
integration:
Z T2
rp (τ ) = p(t)p(t − τ )dt
T1
T2 /T
X
≈T p(nT )p(nT − τ ). (2.37)
n=T1 /T
From this the relationship between the signal energy in continuous-time and discrete-time is
Z T2
Em = a2m p2 (t)dt (2.38)
T1
X
r(nT ) = s(nT ) + w(nT ) = a(l)p(nT − lTs ) + w(nT ). (2.41)
l
2
The complexity of symbol timing synchronization can reduced somewhat in certain cases by sampling at 2 or 4
times the minimum sampling rate given by the sampling theorem.
3
Preference is given the the continuous-time definitions since the continuous-time signals are what is transmitted
across the waveform channel.
64 2.2 M-ary Baseband Pulse Amplitude Modulation (PAM)
n+
T2 Ã ! n+
T2
XT X XT
x0 (nT ) = a(l)p(mT − lTs ) p(mT − nT ) + w(mT )p(mT − nT )
T
m=n+ T1 l T
m=n+ T1
T2 T2
n+ n+
X XT XT
= a(l) p(mT − lTs )p(mT − nT ) + w(mT )p(mT − nT ). (2.42)
l T
m=n+ T1
T
m=n+ T1
1
= rp (lTs − nT ) (2.44)
T
where the last step follows from (2.37). Using v(nT ) to represent the second term in (2.42), the
matched filter output may be expressed as
1X
x0 (nT ) = a(l)rp (lTs − nT ) + v(nT ). (2.45)
T l
The signal space projection is produced by taking every k TTs -th sample of the matched filter
output. Evaluating (2.45) at n = k TTs gives
1X
x0 (kTs ) = a(l)rp ((l − k)Ts ) + v (kTs ) . (2.46)
T l
When the pulse shape satisfies the Nyquist no-ISI condition, the signal space projection is
a(k)
x(kTs ) = + v (kTs ) . (2.47)
T
The additive noise term moves the signal space projection x(kTs ) away from the true value a(k)/T
as was the case with continuous-time processing. A decision error occurs when the noise moves
x(kTs ) closer to a signal space point other than the one that was sent. The probability that this
occurs is derived in Section 2.2.3.
Linear Modulation
t
pulse
LUT
bits S/P ↑N shaping DAC s (t ) = ∑ a (k ) p (t − kTs )
0 k
filter
T
log 2 M bits N = s h (n ) = p (nT )
T T
Figure 2.9: A PAM modulator using discrete-time processing and a DAC (top) and a PAM matched filter detector using discrete-
time processing with an ADC (bottom).
65
66 2.2 M-ary Baseband Pulse Amplitude Modulation (PAM)
2.2.3 Performance
Bandwidth
For independent and equally likely symbols, the power spectral density of the PAM pulse train
(2.21) is given by
σ2
Ps (f ) = a |P (f )|2 (2.48)
Ts
where P (f ) is the continuous-time Fourier transform in f = ω/2π of the pulse shape p(t). The
power spectral density is determined by the properties of the Fourier transform of the pulse shape.
The bandwidth efficiency is a measure of how well a modulation uses the spectrum it occupies.
The bandwidth efficiency is the ratio of the bit rate to the occupied bandwidth. Thus the bandwidth
efficiency has units bits/second/Hz. The bandwidth efficiency is strongly dependent on the defini-
tion of bandwidth. For example, if the 90% bandwidth B90 is the measure of bandwidth, then the
bandwidth efficiency for the NRZ, MAN, HS, and SRRC pulse shapes are
If the 99% bandwidth B99 is the measure of bandwidth, then the bandwidth efficiencies for the
same pulse shapes is drastically reduced:
Probability of Error
The decision rule (2.14) partitions the signal space into decision regions as illustrated for the case
of 4-ary baseband PAM in Figure 2.10. The probability of error is derived by first computing the
conditional probabilities of error and applying the total probability theorem to give the desired
result. In mathematical terms let s0 , s1 , . . . , sM −1 be the M constellation points and let P (E|a =
sm ) be the probability of error given the waveform corresponding to symbol sm was sent for m =
0, 1, . . . , M − 1. Applying the total probability theorem, the probability of error is
M
X −1
P (E) = P (E|a = sm )P (a = sm ) (2.49)
m=0
where P (a = sm ) is the probability that symbol sm was transmitted in the first place. Usually, it
is assumed that the transmitted symbols are equally likely (i.e., neither the data nor the transmitter
prefers sending one symbol over another). In this case
1
P (a = sm ) = m = 0, 1, . . . , M − 1 (2.50)
M
and (2.49) reduces to
M −1
1 X
P (E) = P (E|a = sm ) (2.51)
M m=0
Linear Modulation 67
which is the same as P (E|a(k) = +A) since the decision regions are the same shape. Assuming
a(k) = +3A, x(kTs ) is a random variable whose probability density function is illustrated in
Figure 2.11 (b). Following the same procedure as before
µ ¶
2A
P (E|a(k) = +3A) = 1 − P ≤ x(kTs )
T
µ ¶
2A 3A
=1−P ≤ + v (kTs )
T T
µ ¶
A
= 1 − P − ≤ v (kTs ) . (2.54)
T
P (E|a(k) = +3A) is different from P (E|a(k) = +A) since the decision regions have different
shapes. The decision region for a(k) = +3A is “open ended” while the decision region for a(k) =
+A is not. Following the same procedure once more, it can be shown that
µ ¶
A
P (E|a(k) = −3A) = 1 − P − ≤ v (kTs ) (2.55)
T
which is the same as P (E|a(k) = +3A) since the decision regions are the same shape.
68 2.2 M-ary Baseband Pulse Amplitude Modulation (PAM)
The forgoing shows that the conditional probabilities of error reduce a computation involving
the random variable v (kTs ) and the constant A. The constant A is related to the received symbol
energy. For 4-ary PAM with the four constellation points shown in Figure 2.11, the average symbol
energy
Once the probability density function of v (kTs ) is known, evaluation of (2.52) – (2.55) is straight
forward. The probability of density function of v (kTs ) is derived next.
It was shown in Chapter ?? that when w(t) is white Gaussian random process with zero mean
and power spectral density N0 /2 Watts/Hz, T -spaced samples of the ideally bandlimited w(t) are
zero mean Gaussian random variables with autocorrelation function
N0
σ2 = . (2.58)
2T
T2
n+
XT
v(nT ) = w(mT )p(mT − nT ). (2.59)
T
m=n+ T1
Since v(nT ) is a linear combination of the Gaussian random variables w(mT ), v(nT ) is also a
Gaussian random variable. The joint probability density function is determined completely from
the mean and auto-covariance of the sequence. The mean of each v(nT ) is
n+
T2
T
n+ T2
XT X
E {v(nT )} = E w(mT )p(mT − nT ) = E {w(mT )} p(mT − nT ) = 0
m=n+ T1
m=n+ T1
T T
(2.60)
Linear Modulation 69
σ2
= rp (−(n − l)T ). (2.61)
T
The noise component in the signal space projection is a downsampled version of v(nT ). Clearly
v (kTs ) has zero mean. The correlation is
σ2
E {v (kTs ) v (lTs )} = rp (−(k − l)Ts ) (2.62)
T
σ2
= δ(k − l). (2.63)
T
where the last line is true only if the pulse shape satisfies the Nyquist no-ISI condition. In summary,
the random variables v (kTs ) are independent and identically distributed random variables with
zero mean and variance
σ2 N0
σv2 = = . (2.64)
T 2T 2
Now back to the the probability of error for 4-ary PAM. The probabilities (2.52) – (2.55) can
be expressed in terms of the area under the upper tail of the standard Gaussian random variable,
denoted by the Q-function as described in Chapter ??:
µ ¶ Ãs ¶ ! µ
A/T A 2A2
P (E|a(k) = +A) = P = 2Q
E|a(k) = − = 2Q (2.65)
σv T N0
µ ¶ µ ¶ Ãs !
3A A/T 2A2
P (E|a(k) = +3A) = P E|a(k) = − =Q =Q . (2.66)
T σv N0
70 2.2 M-ary Baseband Pulse Amplitude Modulation (PAM)
Note that the last equality follows from the relationship (2.64) Applying the total probability theo-
rem assuming equally likely symbols gives
Ãs !
3 2A 2
P (E) = Q . (2.67)
2 N0
This shows that when a matched filter detector is used with perfect timing synchronization, the
probability of symbol error is a function only of the ratio of the average symbol energy and the
power spectral density level of the white noise.
Two normalizations are commonly used to facilitate comparisons between modulations with
different symbol sizes. The first is to reduce the symbol error probability to a bit error probability.
This is due to the notion that the fundamental quantity being transmitted is a bit, no matter how
the bits are mapped into symbols. As a consequence, the fundamental measure of reliability is the
bit error rate4 . The translation from symbol error probability to bit error probability follows from
the observation that the most likely symbol errors are those involving adjacent symbols. Using a
gray code, the bit patterns associated with adjacent symbols differ in only one position. Hence, the
most likely symbol errors produce one erroneous bit and log2 (M ) − 1 correct bits. Thus
1
Pb = P (E). (2.69)
log2 M
The translation from symbol error probability to bit error probability can be thought of as a
scaling of the “y-axis” in a probability of error versus signal-to-noise ratio plot. The scaling of the
“x-axis” in such a plot involves scaling the signal-to-noise ratio. Since the bit is the fundamental
quantity being transmitted, the bit energy is the fundamental unit of energy. The average bit energy,
Eb , is related to the average symbol energy, Eavg , by
Generalizing the forgoing to M -ary PAM, the probability of symbol error and bit error are
Ãr !
M −1 6 Eavg
P (E) = 2 Q (2.72)
M M 2 − 1 N0
Ãr !
2(M − 1) log2 M Eb
Pb = Q 6 2 . (2.73)
M log2 M M − 1 N0
The probability of bit error (2.73) for M = 2, 4, and 8 is plotted in Figure 2.12. A vertical
slice through the plot is the line of constant bit energy. The probability of bit error increases as
M increases since the average symbol energy must be shared with more bits. A horizontal line
through the plot is the line of constant reliability. For a fixed probability of bit error, more bit
energy is needed as M increases. For example, at Pb = 10−6 binary PAM requires Eb /N0 = 10.6
dB, 4-ary PAM requires Eb /N0 = 14.6 dB, and 8-ary PAM requires Eb /N0 = 19.2 dB. 4-ary PAM
is said to be 4 dB worse than binary PAM at Pb = 10−6 and 8-ary PAM is 8.6 dB worse than binary
PAM at Pb = 10−6 .
However, these conclusions ignore bandwidth. 4-ary PAM is 2 times more bandwidth efficient
than binary PAM and 8-ary PAM is 3 times more bandwidth efficient than binary PAM. This can
be seen in one of two ways: assuming the bit rate is constant so that the bandwidth varies with
M , or by assuming the bandwidth is constant so that the bit rate varies with M . Either assumption
leads to the same conclusion. The bandwidth efficiency is measured in bits/second/Hz as described
above.
72 2.2 M-ary Baseband Pulse Amplitude Modulation (PAM)
–3A –A +A +3A
p (x0 (kTs ) a (k ) = + A )
–3A –A +A +3A
(a)
p (x0 (kTs ) a (k ) = + 3 A )
–3A –A +A +3A
(b)
Figure 2.11: Probability density function for matched filter output (a) when a/T = +A and (b)
when a/T = +3A.
Linear Modulation 73
0
10
M=2
M=4
−1
10 M=8
−2
10
−3
Pb
10
−4
10
−5
10
−6
10
0 5 10 15 20
Eb/N0 (dB)
Figure 2.12: Probability of bit error versus Eb /N0 for M-ary PAM for M = 2, 4, and 8.