Module9-0 v2

Digital Signal Processing
Module 9: Digital Communication Systems

Module Overview:
◮ Module 9.1: The analog channel

◮ Module 9.2: Meeting the bandwidth constraint
◮ Module 9.3: Meeting the power constraint
◮ Module 9.4: Modulation and demodulation
◮ Module 9.5: Receiver design
◮ Module 9.6: ADSL
9 1
Module 9.1: Digital Communication Systems

Overview:
◮ The many incarnations of a signal

◮ Analog channel constraints
◮ Satisfying the constraints
9.1 2
Overview:

9.1 2
Overview:

9.1 2
Digital data throughputs
◮ Transatlantic cable:
• 1866: 8 words per minute (≈5 bps)
• 1956: AT&T, coax, 48 voice channels (≈3Mbps)
• 2005: Alcatel Tera10, fiber, 8.4 Tbps (8.4 × 1012 bps)
• 2012: fiber, 60 Tbps

◮ Voiceband modems
• 1950s: Bell 202, 1200 bps
• 1990s: V90, 56Kbps
• 2008: ADSL2+, 24Mbps
9.1 3
Digital data throughputs
◮ Transatlantic cable:
• 1866: 8 words per minute (≈5 bps)
• 1956: AT&T, coax, 48 voice channels (≈3Mbps)
• 2005: Alcatel Tera10, fiber, 8.4 Tbps (8.4 × 1012 bps)
• 2012: fiber, 60 Tbps

◮ Voiceband modems
• 1950s: Bell 202, 1200 bps
• 1990s: V90, 56Kbps
• 2008: ADSL2+, 24Mbps
9.1 3
Success factors for digital communications
1) power of the DSP paradigm:

◮ integers are “easy” to regenerate
◮ good phase control
◮ adaptive algorithms
9.1 4
Regenerating signals
−5
x(t)
9.1 5
−5
x(t)/G + σ(t)
9.1 5
−5
G [x(t)/G + σ(t)] = x(t) + G σ(t)
9.1 5
−5
x̂1 (t) = G sgn[x(t) + σ(t)]
9.1 5
2) algorithmic nature of DSP is a perfect match with information theory:

◮ JPEG’s entropy coding
◮ CD’s and DVD’s error correction
◮ trellis-coded modulation and Viterbi decoding
9.1 6
3) hardware advancement
◮ miniaturization
◮ general-purpose platforms
◮ power efficiency
9.1 7
The many incarnations of a conversation
air copper fiber

Base Station Switch
Network
copper coax
Switch CO
9.1 8
The analog channel
unescapable “limits” of physical channels:

◮ bandwidth constraint
◮ power constraint
both constraints will affect the final capacity of the channel
9.1 9
The analog channel

9.1 9
The analog channel

9.1 9
The analog channel’s capacity
maximum amount of information that can be reliably delivered over a channel

(bits per second)
9.1 10
Bandwidth vs capacity
simple thought experiment:

◮ we want to transmit information encoded as a sequence of digital samples over a
continuous-time channel
◮ we interpolate the sequence of samples with a period Ts
◮ if we make Ts small we can send more info per unit of time...
◮ ... but the bandwidth of the signal will grow as 1/Ts
9.1 11

9.1 11

9.1 11

9.1 11
Power and capacity
another thought experiment:

◮ all channels introduce noise; at the receiver we have to “guess” what was transmitted
◮ suppose noise variance is 1
◮ suppose we are transmitting integers between 1 and 10: lots of guessing errors
◮ transmit only odd numbers: fewer errors but less information
9.1 12
Power and capacity

9.1 12
Power and capacity

9.1 12
Power and capacity

9.1 12
Example: the AM radio channel
cos(Ωc t)
x(t) ×
9.1 13
◮ from 530kHz to 1.7MHz

◮ each channel is 8KHz
◮ power limited by law:
• daytime/nighttime
• interference
• health hazards
9.1 14

• interference
• health hazards
9.1 14

• interference
• health hazards
9.1 14

• interference
• health hazards
9.1 14

• interference
• health hazards
9.1 14

• interference
• health hazards
9.1 14
Example: the telephone channel
CO Network CO
9.1 15
◮ one channel from around 300Hz to around 3000Hz

◮ power limited by law to 0.2-0.7V rms
◮ noise is rather low: SNR usually 30dB or more
9.1 16

9.1 16

9.1 16
The all-digital paradigm
keep everything digital until we hit the physical channel
s[n]
..01100
TX D/A s(t)
01010...
Fs = 1/Ts
9.1 17
Let’s look at the channel constraints
bandwidth constraint
power constraint
0 Fmin Fmax
9.1 18
Converting the specs to a digital design
0 Fmin Fmax
9.1 19
0 Fmin Fmax Fs /2
9.1 19
0 ωmin ωmax π
9.1 20
Transmitter design
some working hypotheses:

◮ convert the bitstream into a sequence of symbols a[n] via a mapper
◮ model a[n] as a white random sequence (add a scrambler on the bitstream to make sure)
◮ now we need to convert a[n] into a continuous-time signal within the constraints
9.1 21
Transmitter design

9.1 21
Transmitter design

9.1 21
Transmitter design

a[n]
..01100
Scrambler Mapper ? s(t)
01010...
9.1 21
First problem: the bandwidth constraint
Pa (e jω ) = σa2
σa2
0
−π 0 ωmin ωmax π
9.1 22
First problem: the bandwidth constraint
Pa (e jω ) = σa2
σa2
0
9.1 22
END OF MODULE 9.1
Module 9.2: Controlling the Bandwidth

Overview:
◮ Upsampling
◮ Fitting the transmitter’s spectrum
9.2 23
Overview:
◮ Upsampling
◮ Fitting the transmitter’s spectrum
9.2 23
Shaping the bandwidth
Our problem:
◮ bandwidth constraint requires us to control the spectral support of a signal
◮ we need to be able to “shrink” the support of a full-band signal
◮ the answer is multirate techniques
9.2 24
Our problem:
9.2 24
Our problem:
9.2 24
Multirate signal processing
In a nutshell:
◮ increase or decrease the number of samples in a discrete-time signal
◮ equivalent to going to continuous time and resampling
◮ staying in the digital world is “cleaner”
9.2 25
In a nutshell:
9.2 25
In a nutshell:
9.2 25
Upsampling via continuous time
xc (t)
x[n] Int x ′ [n]
Ts Ts /K
9.2 26
Upsampling (K = 3)
x[n]
b
1 b
b
b
b
b
b
b
b
0
0
b
b
9.2 27
Upsampling (K = 3)
xc (t)
b
1 b
b
b
b
b
b
b
b
0
0
b
b
9.2 27
Upsampling (K = 3)
b b
b
1 b b
b b b b b
b
b b
b b
b
b b b b
b b
b
b
b
b b
0 b
b
0
b
b
b
b b
9.2 27
Upsampling (K = 3)
x ′ [n]
b b
b
1 b b
b b b b b
b
b b
b b
b
b b b b
b b
b
b
b
b b
0 b
b
0
b
b
b
b b
9.2 27
Upsampling
As per usual, we can choose Ts = 1...

∞
X
xc (t) = x[m] sinc(t − m)
m=−∞
x ′ [n] = xc (n/K )
∞
X n
= x[m] sinc −m
m=−∞
K
9.2 28
Upsampling
As per usual, we can choose Ts = 1...

∞
X
xc (t) = x[m] sinc(t − m)
m=−∞
x ′ [n] = xc (n/K )
∞
X n
= x[m] sinc −m
m=−∞
K
9.2 28
Upsampling (frequency domain)
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
9.2 29
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
X (jΩ)
ΩN = π/Ts
−ΩN 0 ΩN
9.2 29
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
X (jΩ)
ΩN = π/(Ts /K )
−ΩN 0 ΩN
9.2 29
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
X (jΩ)
ΩN = π/(Ts /K )
−ΩN 0 ΩN
1
X ′ (e jω )
0
−π −π/2 0 π/4 π/2 π
9.2 29
Upsampling in the digital domain
what can we do purely digitally?

◮ we need to “increase” the number of samples by K
◮ obviously xU [m] = x[n] when m multiple of K
◮ for lack of a better strategy, put zeros elsewhere
◮ example for K = 3:
xU [m] = . . . x[0], 0, 0, x[1], 0, 0, x[2], 0, 0, . . .
9.2 30

xU [m] = . . . x[0], 0, 0, x[1], 0, 0, x[2], 0, 0, . . .
9.2 30

xU [m] = . . . x[0], 0, 0, x[1], 0, 0, x[2], 0, 0, . . .
9.2 30

xU [m] = . . . x[0], 0, 0, x[1], 0, 0, x[2], 0, 0, . . .
9.2 30
Upsampling in the Time Domain
b
1 b
b
b
b
b
b
b
b
0
0
b
b
9.2 31
Upsampling in the Time Domain
b
1 b
b
b
b
b
b
b
b
0 b b b b b b b b b b b b b b b b b b b b b b
0
b
b
9.2 31
in the frequency domain

∞
X
jω
XU (e ) = xU [m]e −jωm
m=−∞
∞
X
= x[n]e −jωnK
n=−∞
= X (e jωK )
9.2 32

∞
X
jω
m=−∞
∞
X
= x[n]e −jωnK
n=−∞
= X (e jωK )
9.2 32

∞
X
jω
m=−∞
∞
X
= x[n]e −jωnK
n=−∞
= X (e jωK )
9.2 32
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
9.2 33
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
1
X (e jω )
0
−5π −4π −3π −2π −π 0 π 2π 3π 4π 5π
9.2 33
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
1
X (e jω )
0
−5π −4π −3π −2π −π 0 π 2π 3π 4π 5π
1
XU (e jω )
0
−π −π/2 0 π/2 π
9.2 33
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
1
X (e jω )
0
−5π −4π −3π −2π −π 0 π 2π 3π 4π 5π
1
XU (e jω )
0
−π −π/2 0 π/2 π
9.2 33
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
1
X (e jω )
0
−5π −4π −3π −2π −π 0 π 2π 3π 4π 5π
1
XU (e jω )
0
−π −π/2 0 π/2 π
9.2 33
X (e jω ) 1
0
−π −π/2 0 π/2 3π/4 π
1
X (e jω )
0
−5π −4π −3π −2π −π 0 π 2π 3π 4π 5π
1
XU (e jω )
0
−π −π/2 0 π/4 π/2 π
9.2 33
back in time domain...

◮ insert K − 1 zeros after every sample
◮ ideal lowpass filtering with ωc = π/K
x ′ [n] = xU (n) ∗ sinc(n/K )

∞
X n−i
= xU [i ] sinc
K
i =−∞
∞
X n
= x[m] sinc −m
m=−∞
K
9.2 34
back in time domain...

◮ insert K − 1 zeros after every sample
◮ ideal lowpass filtering with ωc = π/K
x ′ [n] = xU (n) ∗ sinc(n/K )

∞
X n−i
= xU [i ] sinc
K
i =−∞
∞
X n
= x[m] sinc −m
m=−∞
K
9.2 34
Downsampling
◮ given an upsampled signal we can always recover the original:
x[n] = xU [nK ]
◮ downsampling of generic signals more complicated (aliasing)
9.2 35
Downsampling
◮ given an upsampled signal we can always recover the original:
x[n] = xU [nK ]
◮ downsampling of generic signals more complicated (aliasing)
9.2 35
Remember the bandwidth constraint?
0 Fmin Fmax Fs /2
9.2 36
Here’s a neat trick
let W = Fmax − Fmin ; pick Fs so that:

◮ Fs > 2Fmax (obviously)
◮ Fs = KW , K ∈ N
W 2π
◮ ωmax − ωmin = 2π =
Fs K
◮ we can simply upsample by K
9.2 37

◮ Fs = KW , K ∈ N
W 2π
Fs K
9.2 37

◮ Fs = KW , K ∈ N
W 2π
Fs K
9.2 37

◮ Fs = KW , K ∈ N
W 2π
Fs K
9.2 37
Data rates
◮ upsampling does not change the data rate

◮ we produce (and transmit) W symbols per second
◮ W is sometimes called the Baud rate of the system and is equal to the available
bandwidth
9.2 38
Data rates

bandwidth
9.2 38
Data rates

bandwidth
9.2 38
Transmitter design, continued
a[n]
..01100 K↑
Scrambler Mapper
01010...
b[n] s[n]
× D/A s(t)
cos ωc n
9.2 39
Meeting the bandwidth constraint
0
9.2 40
0
9.2 40
0
9.2 40
0
9.2 40
Raised Cosine
0
−π −π/2 0 π/2 π
9.2 41
Raised Cosine
0
−π −π/2 0 π/2 π
9.2 41
Raised Cosine
0
−π −π/2 0 π/2 π
9.2 41
Raised Cosine
0
−π −π/2 0 π/2 π
9.2 41
Raised Cosine
b b
b b
b b b b
b b b b b b b b b b b b b b b b b b b b b b
0 b b b
b b
b b b
−18 −12 −6 b 0 b 6 12 18
9.2 42
Spectral shaping with raised cosine
0
9.2 43
END OF MODULE 9.2
Module 9.3: Controlling the Power

Overview:
◮ Noise and probability of error

◮ Signaling alphabet and power
◮ QAM signaling
9.3 44
Overview:

◮ QAM signaling
9.3 44
Overview:

◮ QAM signaling
9.3 44
Transmission reliability
◮ transmitter sends a sequence of symbols a[n]

◮ receiver obtaines a sequence â[n]
◮ even if no distortion we can’t avoid noise: â[n] = a[n] + η[n]
◮ when noise is large, we make an error
9.3 45

9.3 45

9.3 45

9.3 45
Probability of error
depends on:
◮ power of the noise wrt power of the signal
◮ decoding strategy
◮ alphabet of transmission symbols
9.3 46
depends on:
9.3 46
depends on:
9.3 46
Signaling alphabets
◮ we have a (randomized) bitstream coming in

◮ we want to send some upsampled and interpolated samples over the channel
◮ how do we go from bitstream to samples?
9.3 47
Signaling alphabets

9.3 47
Signaling alphabets

9.3 47
Mappers and slicers
mapper:
◮ split incoming bitstream into chunks
◮ assign a symbol a[n] from a finite alphabet A to each chunk
slicer:
◮ receive a value â[n]
◮ decide which symbol from A is “closest” to â[n]
◮ piece back together the corresponding bitstream
9.3 48
Example: two-level signaling
mapper:
◮ split incoming bitstream into single bits
◮ a[n] = G if the bit is 1, a[n] = −G if the bit is 0
slicer: (
1 if â[n] > 0
◮ n-th bit =
0 otherwise
9.3 49
b b b b b b b b b b b b b b b b b b b
1
−1 b b b b b b b b b b b b b
−2
0 5 10 15 20 25 30
9.3 50
2 b b b
b
b b
b b b b b b b b b b b b b b b b b b bb
1 b
b b b
b b
b
b
b
0 b
b
b b
b
b
b
b
bb
−1 b b b
b
b b b b b b
b
b b b
b b
b
−2
0 5 10 15 20 25 30
9.3 50
2 b b b
b
b b
b b b b b b b b b b b b b b b b b b b b b bb
1 b
b b b
b b
b
b
b
0 b
b
b b
b
b
b
b
bb
−1 b b b b b b
b
b b b b b b
b
b b b b b
b b
b
−2
0 5 10 15 20 25 30
9.3 50
let’s look at the probability of error after making some hypotheses:

◮ â[n] = a[n] + η[n]
◮ bits in bitstream are equiprobable
◮ noise and signal are independent
◮ noise is additive white Gaussian noise with zero mean and variance σ0
9.3 51
Perr = P[ η[n] < −G | n-th bit is 1 ] P[ n-th bit is 1 ]+

P[ η[n] > G | n-th bit is 0 ] P[ n-th bit is 0 ]
= (P[ η[n] < −G ] + P[ η[n] > G ])/2
= P[ η[n] > G ]
Z ∞ 2
1 −τ2
2σ
= q e 0 dτ
G 2πσ02
1 √
= Q(G /σ0 ) = erfc((G /σ0 )/ 2)
2
9.3 52

= (P[ η[n] < −G ] + P[ η[n] > G ])/2
= P[ η[n] > G ]
Z ∞ 2
1 −τ2
2σ
= q e 0 dτ
G 2πσ02
1 √
= Q(G /σ0 ) = erfc((G /σ0 )/ 2)
2
9.3 52

= (P[ η[n] < −G ] + P[ η[n] > G ])/2
= P[ η[n] > G ]
Z ∞ 2
1 −τ2
2σ
= q e 0 dτ
G 2πσ02
1 √
= Q(G /σ0 ) = erfc((G /σ0 )/ 2)
2
9.3 52

= (P[ η[n] < −G ] + P[ η[n] > G ])/2
= P[ η[n] > G ]
Z ∞ 2
1 −τ2
2σ
= q e 0 dτ
G 2πσ02
1 √
= Q(G /σ0 ) = erfc((G /σ0 )/ 2)
2
9.3 52

= (P[ η[n] < −G ] + P[ η[n] > G ])/2
= P[ η[n] > G ]
Z ∞ 2
1 −τ2
2σ
= q e 0 dτ
G 2πσ02
1 √
= Q(G /σ0 ) = erfc((G /σ0 )/ 2)
2
9.3 52
transmitted power
σs2 = G 2 P[n-th bit is 1] + G 2 P[n-th bit is 0]

= G2
√
Perr = Q(σs /σ0 ) = Q( SNR)
9.3 53
transmitted power

= G2
√
9.3 53
transmitted power

= G2
√
9.3 53
100
10−10
Perr
10−20
0 10 20
SNR (dB)
9.3 54
Lesson learned:
◮ to reduce the probability of error increase G

◮ increasing G increases the power
◮ we can’t go above the channel’s power constraint!
9.3 55
Lesson learned:

9.3 55
Lesson learned:

9.3 55
Multilevel signaling
◮ binary signaling is not very efficient (one bit at a time)

◮ to increase the throughput we can use multilevel signaling
◮ many ways to do so, we will just scratch the surface
9.3 56

9.3 56

9.3 56
PAM
mapper:
◮ split incoming bitstream into chunks of M bits
◮ chunks define a sequence of integers k[n] ∈ {0, 1, . . . , 2M − 1}
◮ a[n] = G ((−2M + 1) + 2k[n]) (odd integers around zero)
slicer:
◮ a′ [n] = arg min[|â[n] − a|]
a∈A
9.3 57
PAM, M = 2, G = 1
b b b b
−3 −1 0 1 3
◮ distance between points is 2G

◮ using odd integers creates a zero-mean sequence
9.3 58
PAM, M = 2, G = 1
b b b b
−3 −1 0 1 3

9.3 58
PAM, M = 2, G = 1
b b b b
−3 −1 0 1 3

9.3 58
From PAM to QAM
◮ error analysis for PAM along the lines of binary signaling

◮ can we increase the throughput even further?
◮ here’s a wild idea, let’s use complex numbers
9.3 59
From PAM to QAM

9.3 59
From PAM to QAM

9.3 59
QAM
mapper:
◮ split incoming bitstream into chunks of M bits, M even
◮ use M/2 bits to define a PAM sequence ar [n]
◮ use the remaining M/2 bits to define an independent PAM sequence ai [n]
◮ a[n] = G (ar [n] + jai [n])
slicer:
◮ a′ [n] = arg min[|â[n] − a|]
a∈A
9.3 60
QAM, M = 2, G = 1
Im
b b
1
−1 1 Re
b b
−1
9.3 61
QAM, M = 4, G = 1
Im
b b b b
3
b b b b
1
−3 −1 1 3 Re
b b b b
−1
b b b b
−3
9.3 62
QAM, M = 8, G = 1
Im
b b b b b b b b b b b b b b b b
Re
9.3 63
QAM
Im
b b b b
b b b b
Re
b b b b
b b b b
9.3 64
QAM
Im
b b b b
b b b b
Re
b b b b
b b b b
9.3 64
QAM
Im
b b b b
×
b b b b
Re
b b b b
b b b b
9.3 64
QAM
Im
b b b b
b b
× b b
Re
b b b b
b b b b
9.3 64
QAM
Im
b b b b
b b b b
×
Re
b b b b
b b b b
9.3 64
QAM
Im
b b b b
b b b × b
Re
b b b b
b b b b
9.3 64
QAM, probability of error
Perr = 1 − P[| Re(η[n])| < G ∧ | Im(η[n])| < G ]

Z
=1− fη (z) dz
D
9.3 65
Perr = 1 − P[| Re(η[n])| < G ∧ | Im(η[n])| < G ]

Z
=1− fη (z) dz
D
9.3 65
b b b
b b G b
b b b
9.3 66
2
− G2
Perr ≈ e σ
0
9.3 67
transmitted power (all symbols equiprobable and independent):
1 X 2
σs2 = G 2 |a|
2M
a∈A
2
= G 2 (2M − 1)
3
2
− G2
Perr ≈ e σ
0
−(M+1)
≈ e −3·2 SNR
9.3 68
1 X 2
σs2 = G 2 |a|
2M
a∈A
2
= G 2 (2M − 1)
3
2
− G2
Perr ≈ e σ
0
−(M+1)
≈ e −3·2 SNR
9.3 68
1 X 2
σs2 = G 2 |a|
2M
a∈A
2
= G 2 (2M − 1)
3
2
− G2
Perr ≈ e σ
0
−(M+1)
≈ e −3·2 SNR
9.3 68
100
10−10
Perr
4-point QAM
16-point QAM
10−20 64-point QAM
0 10 20 30
SNR (dB)
9.3 69
100
10−10
Perr
4-point QAM
16-point QAM
0 10 20 30
SNR (dB)
9.3 69
100
10−10
Perr
4-point QAM
16-point QAM
0 10 20 30
SNR (dB)
9.3 69
QAM, the recipe
◮ pick a probability of error you can live with (e.g. 10−6 )

◮ find out the SNR imposed by the channel’s power constraint

3 SNR
◮ M = log2 1 −
2 ln(pe )
◮ final throughput will be MW
9.3 70
QAM, the recipe


3 SNR
◮ M = log2 1 −
2 ln(pe )
9.3 70
QAM, the recipe


3 SNR
◮ M = log2 1 −
2 ln(pe )
9.3 70
QAM, the recipe


3 SNR
◮ M = log2 1 −
2 ln(pe )
9.3 70
QAM
where we stand:
◮ we know how to fit the bandwidth constraint
◮ with QAM, we know how many bits per symbol we can use given the power constraint
◮ we know the theoretical throughput of the transmitter
but how do we transmit complex symbols over a real channel?
9.3 71
QAM
where we stand:
9.3 71
QAM
where we stand:
9.3 71
QAM
where we stand:
9.3 71
END OF MODULE 9.3
Module 9.4: Modulation and Demodulation

Overview:
◮ Trasmitting and recovering the complex passband signal

◮ Design example
◮ Channel capacity
9.4 72
Overview:

◮ Design example
9.4 72
Overview:

◮ Design example
9.4 72
QAM transmitter design
a[n]
..01100 K↑
Scrambler Mapper
01010...
b[n]
...
b[n] = br [n] + jbi [n] is a complex-valued baseband signal
9.4 73
QAM transmitter design
a[n]
..01100 K↑
Scrambler Mapper
01010...
b[n]
...
b[n] = br [n] + jbi [n] is a complex-valued baseband signal
9.4 73
Complex baseband signal
0
9.4 74
The passband signal
s[n] = Re{b[n] e jωc n }

= Re{(br [n] + jbi [n])(cos ωc n + j sin ωc n)}
= br [n] cos ωc n − bi [n] sin ωc n
9.4 75
The passband signal

9.4 75
The passband signal

9.4 75
1
Re
Im
0
0
−1
DTFT {br [n]}
9.4 76
1
Re
Im
0
0
−1
DTFT {br [n]}

DTFT {bi [n]}
9.4 76
1
Re
Im
0
0
−1
DTFT {br [n] cos ωc n}

DTFT {bi [n]}
9.4 76
1
Re
Im
0
0
−1
DTFT {br [n] cos ωc n}

DTFT {−bi [n] sin ωc n}
9.4 76
Recovering the baseband signal
let’s try the usual method (multiplying by the carrier, see Module 5.5):
s[n] cos ωc n = br [n] cos2 ωc n − bi [n] sin ωc n cos ωc n

1 + cos 2ωc n sin 2ωc n
= br [n] − bi [n]
2 2
1 1
= br [n] + (br [n] cos 2ωc n − bi [n] sin 2ωc n)
2 2
9.4 77

= br [n] − bi [n]
2 2
1 1
2 2
9.4 77

= br [n] − bi [n]
2 2
1 1
2 2
9.4 77

= br [n] − bi [n]
2 2
1 1
2 2
9.4 77
DTFT {br [n] cos ωc n − bi [n] sin ωc n}
1
Re
Im
0
0
−1
9.4 78
DTFT {(br [n] cos ωc n − bi [n] sin ωc n) cos ωc n}
1
Re
Im
0
0
−1
9.4 78
DTFT {(br [n] cos ωc n − bi [n] sin ωc n) cos ωc n}
1
Re
Im
0
0
−1
9.4 78
DTFT {br [n]}
1
Re
Im
0
0
−1
9.4 78
◮ as a lowpass filter, you can use the same filter used in upsampling
◮ matched filter technique
9.4 79
◮ as a lowpass filter, you can use the same filter used in upsampling
◮ matched filter technique
9.4 79
similarly:
s[n] sin ωc n = br [n] cos ωc n sin ωc n − bi [n] sin2 ωc n

1 1
= − bi [n] + (br [n] sin 2ωc n − bi [n] cos 2ωc n)
2 2
9.4 80
similarly:

1 1
2 2
9.4 80
similarly:

1 1
2 2
9.4 80
QAM transmitter, final design
a[n]
..01100 K↑
Scrambler Mapper
01010...
b[n] c[n] s[n]

× Re D/A s(t)
e jωc n
9.4 81
QAM receiver, idealized design
bˆr [n]
×
ŝ[n] b̂[n]
b
cos ωc n
ŝ(t) +
sin ωc n
−j
×
b̂i [n]
â[n]
K↓ ..01100
Slicer Descrambler
01010...
9.4 82
Example: the V.32 voiceband modem
◮ analog telephone channel: Fmin = 450Hz, Fmax = 2850Hz

◮ usable bandwidth: W = 2400Hz, center frequency Fc = 1650Hz
◮ pick Fs = 3 · 2400 = 7200Hz, so that K = 3
◮ ωc = 0.458π
9.4 83

◮ ωc = 0.458π
9.4 83

◮ ωc = 0.458π
9.4 83

◮ ωc = 0.458π
9.4 83
◮ maximum SNR: 22dB

◮ pick Perr = 10−6
◮ using QAM, we find !
3 1022/10
M = log2 1− ≈ 4.1865
2 ln(10−6 )
so we pick M = 4 and use a 16-point constellation
◮ final data rate is WM = 9600 bits per second
9.4 84

3 1022/10
M = log2 1− ≈ 4.1865
2 ln(10−6 )
9.4 84

3 1022/10
M = log2 1− ≈ 4.1865
2 ln(10−6 )
9.4 84

3 1022/10
M = log2 1− ≈ 4.1865
2 ln(10−6 )
9.4 84
Theoretical channel capacity
◮ we used very specific design choices to derive the throughput

◮ what is the best one can do?
◮ Shannon’s capacity formula is the upper bound
C = W log2 (1 + SNR)
◮ for instance, for the previous example C ≈ 17500 bps

◮ the gap can be narrowed by more advanced coding techniques
9.4 85


9.4 85


9.4 85


9.4 85


9.4 85
END OF MODULE 9.4
Module 9.5: Receiver Design

Overview:
◮ Adaptive equalization
◮ Timing recovery
9.5 86
Overview:
◮ Adaptive equalization
◮ Timing recovery
9.5 86
A blast from the past
9.5 87
A blast from the past
◮ a sound familiar to anyone who’s used a modem or a fax machine

◮ what’s going on here?
9.5 87
Graphically
bˆr [n]
×
ŝ[n] b̂[n]
b
cos ωc n
ŝ(t) +
sin ωc n
−j
×
b̂i [n]
â[n]
K↓ ..01100
Slicer Descrambler
01010...
9.5 88
Pilot tones
if ŝ[n] = cos((ωc + ω0 )n):
b̂[n] = H{cos((ωc + ω0 )n) cos(ωc n) − j cos((ωc + ω0 )n) sin(ωc n)}

= H{cos(ω0 n) + cos((2ωc + ω0 )n) − j sin((2ωc + ω0 )n) + j sin(ω0 n)}
= cos(ω0 n) + j sin(ω0 n)
= e jω0 n
9.5 89
Pilot tones
if ŝ[n] = cos((ωc + ω0 )n):

= e jω0 n
9.5 89
Pilot tones
if ŝ[n] = cos((ωc + ω0 )n):

= e jω0 n
9.5 89
Pilot tones
if ŝ[n] = cos((ωc + ω0 )n):

= e jω0 n
9.5 89
In slow motion
9.5 90
It’s a dirty job...
but a receiver has to do it:

◮ interference
◮ propagation delay
◮ linear distortion
◮ clock drifts
9.5 91

◮ interference
◮ clock drifts
9.5 91

◮ interference
◮ clock drifts
9.5 91

◮ interference
◮ clock drifts
9.5 91

◮ interference → handshake and line probing
◮ clock drifts
9.5 91

◮ propagation delay → delay estimation
◮ clock drifts
9.5 91

◮ linear distortion → adaptive equalization
◮ clock drifts
9.5 91

◮ linear distortion → adaptive equalization
◮ clock drifts → timing recovery
9.5 91
The two main problems
s(t) ŝ(t)
s[n] D/A D(jΩ) A/D ŝ[n]
Ts Ts′
◮ channel distortion D(jΩ)

◮ (time-varying) discrepancies in clocks Ts′ = Ts
9.5 92
s(t) ŝ(t)
Ts Ts′

9.5 92
s(t) ŝ(t)
Ts Ts′

9.5 92
Delay compensation
Assume the channel is a simple delay: ŝ(t) = s(t − d) ⇒ D(jΩ) = e −jΩd

◮ channel introduces a delay of d seconds
◮ we can write d = (b + τ )Ts with b ∈ N and |τ | < 1/2
◮ b is called the bulk delay
◮ τ is the fractional delay
9.5 93
Delay compensation

9.5 93
Delay compensation

9.5 93
Delay compensation

9.5 93
Offsetting the bulk delay (Ts = 1)
s[n]
1 b
0 b b b b b b b b b b b
9.5 94
s(t)
1 b
9.5 94
ŝ(t)
1 b
9.5 94
ŝ[n]
1 b
b b
b b b
0 b
b
b b
b
b b b b b b
b
b b
b
0 b
b
b+τ
9.5 94
Estimating the fractional delay
◮ transmit b[n] = e jω0 n (i.e. s[n] = cos((ωc + ω0 )n))

◮ receive ŝ[n] = cos((ωc + ω0 )(n − b − τ ))
◮ after demodulation and bulk delay offset:
b̂[n] = e jω0 (n−τ )
◮ multiply by known frequency

b̂[n] e −jω0 n = e −jω0 τ
9.5 95

b̂[n] = e jω0 (n−τ )

9.5 95

b̂[n] = e jω0 (n−τ )

9.5 95

b̂[n] = e jω0 (n−τ )

9.5 95
Compensating for the fractional delay
◮ ŝ[n] = s(n − τ )Ts (after offsetting bulk delay)

◮ we need to compute subsample values
◮ in theory, compensate with a sinc fractional delay h[n] = sinc(n + τ )
◮ in practice, use local Lagrange approximation
9.5 96

9.5 96

9.5 96

9.5 96
2 b
1
b
0
n−1 n n+1
−1 b
9.5 97
2 b
1
b
0 ×
n−1 n n+τ n+1
−1 b
9.5 97
Lagrange approximation (see Module 6.2)
as per usual, choose Ts = 1

◮ we want to compute x(n + τ ), with |τ | < 1/2
◮ local Lagrange approximation around n
N
(N)
X
xL (n; t) = x[n − k]Lk (t)
k=−N
N
(N)
Y t−i
Lk (t) = k = −N, . . . , N
k −i
i =−N
i 6=k
◮ x(n + τ ) ≈ xL (n; τ )
9.5 98

N
(N)
X
xL (n; t) = x[n − k]Lk (t)
k=−N
N
(N)
Y t−i
Lk (t) = k = −N, . . . , N
k −i
i =−N
i 6=k
◮ x(n + τ ) ≈ xL (n; τ )
9.5 98

N
(N)
X
xL (n; t) = x[n − k]Lk (t)
k=−N
N
(N)
Y t−i
Lk (t) = k = −N, . . . , N
k −i
i =−N
i 6=k
◮ x(n + τ ) ≈ xL (n; τ )
9.5 98
Lagrange interpolation (N = 1)
2 b
1
b
0
n−1 n n+1
−1 b
9.5 99
2 b
1
b
0 ×
n−1 n n+τ n+1
−1 b
9.5 99
2 b
1
b
0 ×
n−1 n n+τ n+1
−1 b
9.5 99
2 b
1
b
0 ×
n−1 n n+τ n+1
−1 b
9.5 99
2 b
1
b
0 ×
n−1 n n+τ n+1
−1 b
9.5 99
2 b
1
b
0 ×
n−1 n n+τ n+1
−1 b
9.5 99
2 b
bc
1
b
0
n−1 n n+τ n+1
−1 b
9.5 99
Lagrange interpolation as an FIR
◮ x(n + τ ) ≈ xL (n; τ )
(N)
◮ define dτ [k] = Lk (τ ), k = −N, . . . , N
◮ dτ [k] form a (2N + 1)-tap FIR
◮ xL (n; τ ) = (x ∗ dτ )[n]
9.5 100
◮ x(n + τ ) ≈ xL (n; τ )
(N)
◮ xL (n; τ ) = (x ∗ dτ )[n]
9.5 100
◮ x(n + τ ) ≈ xL (n; τ )
(N)
◮ xL (n; τ ) = (x ∗ dτ )[n]
9.5 100
◮ x(n + τ ) ≈ xL (n; τ )
(N)
◮ xL (n; τ ) = (x ∗ dτ )[n]
9.5 100
Example (N = 1, second order approximation)
(1) t−1
L−1 (t) = t
2
(1)
L0 (t) = (1 − t)(1 + t)
(1) t+1
L1 (t) = t
2
9.5 101
Example (N = 1, second order approximation)



−0.08 n = −1

0.96 n=0
d0.2 [n] =


0.12 n=1

0 otherwise
9.5 102
Delay compensation algorithm
◮ estimate the delay τ

◮ compute the 2N + 1 Lagrangian coefficients
◮ filter with the resulting FIR
9.5 103

9.5 103

9.5 103
Compensating for the distortion
s(t) ŝ(t)
9.5 104
Compensating for the distortion
s[n] D(z) ŝ[n]
9.5 105
Example: adaptive equalization
ŝ[n]
s[n] D(z) E (z) sˆe [n] = s[n]
9.5 106
◮ in theory, E (z) = 1/D(z)

◮ but we don’t know D(z) in advance
◮ D(z) may change over time
9.5 107

9.5 107

9.5 107
Adaptive equalization
ŝe [n]
ŝ[n] E (z) b
e[n]
- s[n]
9.5 108
Adaptive equalization: bootstrapping via a training sequence
s[n]
at [n] TX ...
9.5 109
Adaptive equalization: bootstrapping via a training sequence
s[n]
at [n] TX ...
ŝe [n]
ŝ[n] E (z) b
e[n] s[n]
- Modulator at [n]
9.5 109
Adaptive equalization: online mode
ŝe [n] â[n]

b b
ŝ[n] E (z) Demod Slicer
e[n] s ′ [n]
- Modulator
9.5 110
So much more to do...
◮ how do we perform the adaptation of the coefficients?

◮ how do we compensate for differences in clocks?
◮ how do we recover from interference?
◮ how do we improve resilience to noise?
adaptive signal processing
9.5 111

9.5 111

9.5 111

9.5 111

9.5 111

9.5 111
END OF MODULE 9.5
Module 9.6: ADSL

Overview:
◮ Channel
◮ Signaling strategy
◮ Discrete Multitone Modulation (DMT)
9.6 112
Overview:
◮ Channel
9.6 112
Overview:
◮ Channel
9.6 112
The telephone network today
voice
CO
network
CO
DSLAM
internet
9.6 113
The telephone network today
voice
CO
network
CO
DSLAM
“last mile”
internet
9.6 113
The last mile
◮ copper wire (twisted pair) between home and nearest CO

◮ very large bandwidth (well over 1MHz)
◮ very uneven spectrum: noise, attenuation, interference, etc.
9.6 114
The last mile

9.6 114
The last mile

9.6 114
The ADSL channel
POTS
upstream
downstream
0 1MHz
9.6 115
The ADSL channel
0 1MHz
9.6 116
Idea: split the band into independent subchannels
0 1MHz
9.6 117
Subchannel structure
◮ allocate N subchannels over the total positive bandwidth

◮ equal subchannel bandwidth Fmax /N
◮ equally spaced subchannels with center frequency kFmax /N, k = 0, . . . , N − 1
9.6 118

9.6 118

9.6 118
The digital design
◮ pick Fs = 2Fmax (Fmax is high now!)

kFmax /N 2π
◮ center frequency for each subchannel ωk = 2π = k
Fs 2N
2π
◮ bandwidth of each subchannel
2N
◮ to send symbols over a subchannel: upsampling factor K ≥ 2N
9.6 119
The digital design

kFmax /N 2π
Fs 2N
2π
2N
9.6 119
The digital design

kFmax /N 2π
Fs 2N
2π
2N
9.6 119
The digital design

kFmax /N 2π
Fs 2N
2π
2N
9.6 119
The digital design (N = 3)
−π 0 π
9.6 120
−π −ω2 −ω1 ω0 = 0 ω1 = 2π/6 ω2 = 4π/6 π
9.6 120
−π −ω2 −ω1 ω0 = 0 ω1 = 2π/6 ω2 = 4π/6 π
9.6 120
The digital design
◮ put a QAM modem on each channel

◮ decide on constellation size independently
◮ noisy or forbidden subchannels send zeros
9.6 121
The digital design

9.6 121
The digital design

9.6 121
The subchannel modem
e j(2π/2N)kn
bk [n] ck [n]
ak [m] 2N ↑ ×
9.6 122
The bank of modems
e j(2π/2N)(0·n)
b0 [n] c0 [n]
a0 [m] 2N ↑ ×
e j(2π/2N)(1·n)
b1 [n] c1 [n]
a1 [m] 2N ↑ ×
e j(2π/2N)(2·n)
b2 [n] c2 [n] c[n]
a2 [m] 2N ↑ × + Re s[n]
...
e j(2π/2N)(N−2)n
bN−2 [n] cN−2 [n]
aN−2 [m] 2N ↑ ×
e j(2π/2N)(N−1)n
bN−1 [n] cN−1 [n]
aN−1 [m] 2N ↑ ×
9.6 123
If it looks familiar...
check back Module 4.3, the DFT reconstruction formula:
A0 ∼ 0
φ0
A1 ∼ 1
φ1
A2 ∼ 2 + x[n]
φ2
...
AN−2 ∼ N−2
φN−2
AN−1 ∼ N−1
φN−1
9.6 124
DMT via IFFT
◮ we will show that transmission can be implemented efficiently via an IFFT

◮ Discrete Multitone Modulation
9.6 125
DMT via IFFT
◮ we will show that transmission can be implemented efficiently via an IFFT

◮ Discrete Multitone Modulation
9.6 125
The great ADSL trick
instead of using a good lowpass filter, use the 2N-tap interval indicator:
(
1 for 0 ≤ n < 2N
h[n] =
0 otherwise
9.6 126
Interval indicator signal (Module 4.7)
b b b b b b b b b b
1
b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b
0
0 2N − 1
9.6 127
DTFT of interval signal (Module 4.7)
1
|H(e jω )|
0
−π 0 π/N π
9.6 128
DTFT of interval signal (Module 4.7)
1
|H(e jω )|
0
−π 0 π/N π
9.6 128
Back to the subchannel modem
e j(2π/2N)kn
bk [n] ck [n]
ak [m] 2N ↑ ×
rate: B symbols/sec 2NB samples/sec
9.6 129
e j(2π/2N)kn
bk [n] ck [n]
ak [m] 2N ↑ ×
rate: B symbols/sec 2NB samples/sec
9.6 129
by using the indicator function as a lowpass:
e j(2π/2N)nk
ak [⌊n/2N⌋] × ck [n]
9.6 130
The bank of modems, revisited
e j(2π/2N)(0·n)
a0 [⌊n/2N⌋] ×
e j(2π/2N)(1·n)
a1 [⌊n/2N⌋] ×
e j(2π/2N)(2·n)
c[n]
a2 [⌊n/2N⌋] × + Re s[n]
...
e j(2π/2N)(N−2)n
aN−2 [⌊n/2N⌋] ×
e j(2π/2N)(N−1)n
aN−1 [⌊n/2N⌋] ×
9.6 131
The complex output signal
N−1
X 2π
c[n] = ak [⌊n/2N⌋]e j 2N nk
k=0

= 2N · IDFT2N a0 [m] a1 [m] . . . aN−1 [m] 0 0 . . . 0 [n]
(m = ⌊n/2N⌋)
9.6 132
The complex output signal
N−1
X 2π
c[n] = ak [⌊n/2N⌋]e j 2N nk
k=0

= 2N · IDFT2N a0 [m] a1 [m] . . . aN−1 [m] 0 0 . . . 0 [n]
(m = ⌊n/2N⌋)
9.6 132
We can do even better!
◮ we are interested in s[n] = Re{c[n]} = (c[n] + c ∗ [n])/2

◮ it is easy to prove (exercise) that:
∗ n ∗ o
IDFT x0 x1 x2 . . . xN−2 xN−1 = IDFT x0 xN−1 xN−2 . . . x2 x1

◮ c[n] = 2N · IDFT a0 [m] a1 [m] . . . aN−1 [m] 0 0 . . . 0 [n]
◮ therefore
∗ ∗ [m] . . . a1∗ [m] [n]

s[n] = N · IDFT 2a0 [m] a1 [m] . . . aN−1 [m] aN−1 [m] aN−2
9.6 133

∗ n ∗ o

◮ therefore
∗ ∗ [m] . . . a1∗ [m] [n]

9.6 133

∗ n ∗ o

◮ therefore
∗ ∗ [m] . . . a1∗ [m] [n]

9.6 133

∗ n ∗ o

◮ therefore
∗ ∗ [m] . . . a1∗ [m] [n]

9.6 133
ADSL transmitter
s[2Nm]
2a0 [m] b
s[2Nm + 1]
a1 [m] b
s[2Nm + 2]
a2 [m] b
s[2Nm + 3]
s[2Nm + 4]
s[2Nm + 5]
parallel to serial
aN−2[m] b
IFFT
aN−1[m] b
b
∗ s[n]
∗
∗
s[2Nm + N − 4]
s[2Nm + N − 3]
∗ s[2Nm + N − 2]
∗ s[2Nm + N − 1]
9.6 134
ADSL specs
◮ Fmax = 1104KHz
◮ N = 256
◮ each QAM can send from 0 to 15 bits per symbol
◮ forbidden channels: 0 to 7 (voice)
◮ channels 7 to 31: upstream data
◮ max theoretical throughput: 14.9Mbps (downstream)
9.6 135
ADSL specs
◮ Fmax = 1104KHz
◮ N = 256
9.6 135
ADSL specs
◮ Fmax = 1104KHz
◮ N = 256
9.6 135
ADSL specs
◮ Fmax = 1104KHz
◮ N = 256
9.6 135
ADSL specs
◮ Fmax = 1104KHz
◮ N = 256
9.6 135
ADSL specs
◮ Fmax = 1104KHz
◮ N = 256
9.6 135
END OF MODULE 9.6
END OF MODULE 9

Module9-0 v2

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Module9-0 v2

Uploaded by

Copyright:

Available Formats

Digital Signal Processing

Module 9: Digital Communication Systems

◮ Module 9.1: The analog channel

Module 9.1: Digital Communication Systems

◮ The many incarnations of a signal

◮ The many incarnations of a signal

◮ The many incarnations of a signal

• 1956: AT&T, coax, 48 voice channels (≈3Mbps)

• 2005: Alcatel Tera10, fiber, 8.4 Tbps (8.4 × 1012 bps)

• 2012: fiber, 60 Tbps

• 1990s: V90, 56Kbps

• 2008: ADSL2+, 24Mbps

• 1956: AT&T, coax, 48 voice channels (≈3Mbps)

• 2005: Alcatel Tera10, fiber, 8.4 Tbps (8.4 × 1012 bps)

• 2012: fiber, 60 Tbps

• 1990s: V90, 56Kbps

• 2008: ADSL2+, 24Mbps

1) power of the DSP paradigm:

G [x(t)/G + σ(t)] = x(t) + G σ(t)

x̂1 (t) = G sgn[x(t) + σ(t)]

2) algorithmic nature of DSP is a perfect match with information theory:

air copper fiber

unescapable “limits” of physical channels:

both constraints will affect the final capacity of the channel

unescapable “limits” of physical channels:

both constraints will affect the final capacity of the channel

unescapable “limits” of physical channels:

both constraints will affect the final capacity of the channel

maximum amount of information that can be reliably delivered over a channel

simple thought experiment:

simple thought experiment:

simple thought experiment:

simple thought experiment:

another thought experiment:

another thought experiment:

another thought experiment:

another thought experiment:

◮ from 530kHz to 1.7MHz

◮ from 530kHz to 1.7MHz

◮ from 530kHz to 1.7MHz

◮ from 530kHz to 1.7MHz

◮ from 530kHz to 1.7MHz

◮ from 530kHz to 1.7MHz

◮ one channel from around 300Hz to around 3000Hz

◮ one channel from around 300Hz to around 3000Hz

◮ one channel from around 300Hz to around 3000Hz

keep everything digital until we hit the physical channel

some working hypotheses:

some working hypotheses:

some working hypotheses:

some working hypotheses:

Module 9.2: Controlling the Bandwidth

As per usual, we can choose Ts = 1...

As per usual, we can choose Ts = 1...

what can we do purely digitally?

xU [m] = . . . x[0], 0, 0, x[1], 0, 0, x[2], 0, 0, . . .

what can we do purely digitally?

xU [m] = . . . x[0], 0, 0, x[1], 0, 0, x[2], 0, 0, . . .

what can we do purely digitally?

xU [m] = . . . x[0], 0, 0, x[1], 0, 0, x[2], 0, 0, . . .

what can we do purely digitally?

xU [m] = . . . x[0], 0, 0, x[1], 0, 0, x[2], 0, 0, . . .