Equalization For Discrete Multitone Transceivers

You might also like

Download as ppt, pdf, or txt
Download as ppt, pdf, or txt
You are on page 1of 36

Equalization for

Discrete Multitone Transceivers


Outline
• Introduction to high-speed wireline digital communications
• Problem: Increase ADSL transceiver bit rate by increasing
performance of the time-domain equalizer (TEQ) in the receiver
• Contribution #1: New model for equalized channel
• Contribution #2: Optimal channel capacity TEQ
• Contribution #3: Closed-form near-optimal TEQ
• Simulation results
• Summary and future work

Transmit Receive
bit stream bit stream
TEQ
Transmitter Channel 2
Receiver
Introduction
Residential Applications
Application Downstream Upstream Willingness Demand
Rate (kb/s) Rate (kb/s) to pay Potential
Internet 3,000 384 High Medium
High Definition TV 24,000 0 High Medium
Broadcast Video 6,000 0 Low High
Video Phone 1,500 1,500 High Medium
Database Access 384 9 High Medium
On-line Directory; Yellow Pages 384 9 Low High
Shop-at-Home 1,500 64 Low Medium
Video Games 1,500 1,500 Medium Medium

Business Applications
Application Downstream Upstream Willingness Demand
Rate (kb/s) Rate (kb/s) to pay Potential
Internet 3,000 384 High High
Remote Office 6,000 1,500 High Medium
LAN Interconnection 10,000 10,000 Medium Medium
Financial News 1,500 9 Medium Low
Video Phone 1,500 1,500 High Low
Video Conference 3,000 3,000 High Low
Supercomputing, CAD 45,000 45,000 High Low 3
On-line Directory; Yellow Pages 384 9 Medium High
Standard Meaning Data Rate Mode Applications
ISDN Integrated Services 144 kbps Symmetric Internet Access, Voice, Pair
Digital Network Gain (2 channels)
T1 T-Carrier One 1.544 Mbps Symmetric Enterprise, Expansion,
(requires two pairs) Internet Service
HDSL High-Speed Digital 1.544 Mbps Symmetric Pair Gain (12 channels),
Subscriber Line Internet Access, T1/E1
(requires two pairs) replacement
HDSL2 Single Line HDSL 1.544 Mbps Symmetric Same as HDSL except pair
gain is 24 channels
G.DMT Asymmetric Digital 1.5 to 8 Mbps Down Internet Access, Digital
Subscriber Line 16 to 640 kbps Up Video
G.Lite Splitterless 32 to 1500 kbps Down Internet Access, Digital
Asymmetric Digital 32 to 512 kbps Up Video
Subscriber Line
Standards for High-Speed
VDSL Very High-Speed
13 to 52 Mbps
Digital
Down Internet Access, Digital
Communications 1.5 to 6 Mbps Up
Digital Subscriber
Line (proposed)
Video

4
Courtesy of Shawn McCaslin (Cicada Semiconductor, Austin, TX)
Intersymbol Interference (ISI)
2.1
• Ideal channel
1.7
111 1 1 – Impulse response is an impulse
.7
.4 .1 – Frequency response is flat

* =
Channel Received • Non-ideal channel causes ISI
-1 signal
– Channel memory
Transmit
– Magnitude and phase variation
signal Threshold
at zero

11 1 1
• Received symbol is weighted
sum of neighboring symbols
– Weights are determined by
channel impulse response5
Detected
signal
• Problem: Channel frequency response is not flat
• Solution: Use equalizer to flatten channel frequency response
• Zero-forcing equalizer
– Inverts channel
Zero-forcing MMSE
Equalizer Equalizer – Flattens frequency response
frequency frequency
Magnitude (dB)

response response – Amplifies noise


Channel
• Minimum mean squared error
frequency (MMSE) equalizer
response
– Optimizes trade-off between
noise amplification and ISI
Combat ISI with Equalization
• Decision-feedback equalizer
– Increases complexity
Frequency – Propagates error 6
 Divide broadband channel into many narrowband subchannels
 No intersymbol interference (ISI) in subchannels if channel gain is constant
in every subchannel
 Discrete Multitone (DMT) modulation
 Multicarrier modulation based on fast Fourier transform (FFT)
 Standardized for ADSL and proposed for VDSL

channel
frequency
magnitude

response
carrier

Multicarrier Modulation subchannel

7
frequency
 Advantages
 Efficient use of bandwidth without full channel equalization
 Robust against impulsive noise and narrowband interference
 Dynamic rate adaptation
 Disadvantages
 Transmitter: High signal peak-to-average power ratio
 Receiver: Sensitive to frequency and phase offset in carriers
 Active areas of research
 Pulse shapes of subchannels (orthogonal, efficient realization)
 Channel equalizer design (increase capacity, reduce complexity)
Multicarrier
 Synchronization Modulation
(timing recovery, symbol synchronization)
 Bit loading (allocation of bits in each subchannel)
8
copy copy

CP s y m b o l i CP s y m b o l ( i+1)

v samples N samples CP: Cyclic Prefix ADSL Standard


v 32
 Convolve stream of samples with channel N 512
 Symbols are spread out in time
 No ISI if channel length is shorter than v+1 samples
 Symbols are distorted in frequency
 Cyclic prefix converts linear convolution into circular convolution
 Division in FFT domain can undo distortion if channel length is less than v+1
Eliminating ISI in DMT
samples
 Time domain equalizer shortens channel length
 Frequency domain equalizer inverts channel frequency response 9
N/2 subchannels N subchannels (N = 512 for ADSL)

mirror DAC
add
serial to QAM data parallel and
cyclic
parallel encoder and to serial transmit
prefix
N-IFFT filter

TRANSMITTER
channel
RECEIVER
N/2 subchannels N subchannels

Discrete Multitoneinvert
channel
Transmitter
N-FFT
and
and
serial remove TEQ
receive
Receiver
parallel QAM =
to serial decoder frequency
remove to cyclic time
filter
and
mirrored parallel prefix domain
domain
equalizer
ADC 10
equalizer data
 Problem:
 Find a TEQ design method that maximizes channel capacity at the TEQ
output
 Proposed solution
 Decompose equalized channel into signal, noise, and ISI paths
 Model subchannel SNR based on this decomposition
 Write channel capacity as a function of TEQ taps
 Develop design methods to maximize channel capacity
 Contributions
 A new model for subchannel SNR
 Optimal maximum channel capacity (MCC) TEQ design method
 Near-optimal minimum ISI TEQ design method

Problem Definition and Contributions


11
Minimum Mean Squared Error (MMSE) Method
Chow, Cioffi, 1992
nk
yk rk ek
xk + w
h - +

z- b zk

 Minimize mean squared error E{ek} where ek=bk- - hk*wk


 Chose length of bk to shorten length of hk*wk
 Disadvantages
 Does not consider channel capacity
 Zeros low SNR bands 12
 Deep notches in equalizer frequency response
 For each possible position of a window of +1 samples,
 energy inside window after TEQ 
max SSNR in dB  max 10 log10  
w w
 energy outside window after TEQ 

h w

• Disadvantages
– Does not consider channel capacity
Requires Cholesky decomposition

Maximum Shortening
and eigenvector calculation SNR (MSSNR)
Method
– Does not take channel noise into
Melsa, Younce, Rohrs, 1996 
account 13
 Each subchannel modeled as white Gaussian noise channel
N /2  SNR i  S x ,i H i
2
Matched Filter
bDMT   log 2 1   SNR i  
    S n ,i Bound (MFB)
i 1 
S x ,i and S n ,i : Signal and noise power in i th subchannel
H i : Channel gain in i th subchannel and  : SNR gap
• Define geometric SNR
  N / 2  SNR i  2 / N 
SNR geom     1     1
  i 1     
 
• Capacity
Channel of aofMulticarrier
capacity Channel
a multicarrier channel
N  SNR geom  14
bDMT  log 2 1  
2   
Maximum Geometric SNR Method
Al-Dhahir, Cioffi, 1996
 Maximize approximate geometric SNR
nk 2/ N
rk N /2

xk w ek SNR geom   SNR i 
h +
-+  i 1 
z- b zk 2
S x Bi
 Disadvantages Bi  H i Wi SNR i  2
 Subchannel SNR definition ignores ISI S n ,i Wi
 Objective function ignores interdependence of b and w

N /2
H i : channel gain in i th subchannel
2
 ln Bi
2
L(b )  Wi : equalizer gain in i th subchannel
N i 1
th
 B : target gain in i
Requires solution of nonlinear constrainedi optimization problem
subchannel
 Based on MMSE method: same drawbacks as MMSE method
 Ad-hoc parameter MSEmax has to be tuned for different channels
15
Method MMSE MSSNR Geometric
Advantages
Maximize channel capacity 
Minimize ISI 
Bit Rate Low-medium High Low-medium
Disadvantages
Nonlinear optimization 
Computationally complexity Low Medium High
Artificial constraints  
Ad-hoc parameters 
Comparison of Existing Methods
Low-pass frequency response  
Unrealistic assumptions  16
Contribution #1
New Subchannel Model: Motivating Example
 Received signal Delay  y1  h~1a4   n~1 
~ ~
 y  xh  n  y  ~ ~  ~ 
CP  2   h~1a1  h2 a4   n2 
~ ~ ~   n~3 
 h  w h  y3  h1a2  h2 a1  h3 a4
x is transmitted   ~ ~ ~ ~  ~ 

 y4   h1a3  h2 a2  h3 a1  h4 a4   n4 
signal
 y5  h~ a 
~
h2 a3
~
 h3 a2 
~
h4 a1   n~5 
 Symbols a b    ~1 4 ~ ~ ~  ~ 
 y6   h1b4  h2 a4  h3 a3  h4 a2   n6 
 Symbol length  y    h~ b ~ ~ ~ 
CP  h2b4  h3 a4  h4 a3    n~7 
 N=4  7   ~1 1 ~ ~ ~ ~ 
~  y8   h1b2  h2b1  h3b4  h4 a4   n8 
Length of h  y  ~ ~ ~ ~   
h4b4   n~9 

 9   h~1b3  h2b2  h3b1 
 L=4 ~ ~ ~  ~
 y10   h1b4  h2b3  h3b2  h4b1 n10 
 Cyclic prefix    ~ ~ ~  ~ 
 y11   h2b4  h3b3  h4b2   n11 
~ ~
h4b3  n~12 
 v=1  y12   h3b4 
 Delay Tail  y   ~  ~ 
 13   h4b4 17n13 
 =1
ISI signal ISI noise
Contribution #1
Proposed Subchannel SNR Model
~
h signal
k h k g k
 Partition equalized channel into signal path, ~
ISI path, noise path hkISI  h k 1  g k 
nk
yk rk
hknoise  wk
xk
h + w ~
hk  hk  wk
xk w
h x Signal
1   k    
gk gk  
0 otherwise
xk w
h x
ISI
1-gk gk
1
nk
...
w noise
18
 1   1 k
Contribution #1
Subchannel SNR Definition

 SNR in i th subchannel is defined as


2
signal
signal power S x ,i H i
SNR i  
noise power  ISI power S H noise 2  S H ISI 2
n ,i i x ,i i

H isignal : gain of hksignal in subchannel i


H iISI : gain of hkISI in subchannel i
H inoise : gain of hknoise in subchannel i

S x ,i : transmitted signal power in subchannel i


S n ,i : channel noise power in subchannel i 19
Contribution #2
Optimal Maximum Channel Capacity (MCC) TEQ
 Channel capacity as a nonlinear function of equalizer taps
2
H isignal  q iH GHw H
S x ,i q GHw
i wT Ai w
SNR i  2 2 
H iISI  q iH DHw S n ,i q i Fw  S x ,i q i DHw
H H w T
Bi w
H inoise  q iH Fw GT Sx,i qiH G
Ai = HT qi H

FT Sn,i qiH F DT Sx,i qiH D


Bi = qi + HT qi H

 Maximize nonlinear function to obtain the optimal TEQ

 1 wT Ai w 
N /2
bDMT   log 2 1  
i 1   w Bi w 
T
 Good performance measure for any TEQ design method
20
 Not an efficient TEQ design method in computational sense
Contribution #2
MCC TEQ vs. Geometric TEQ
Method MCC Geometric
Advantages
Maximize channel capacity  
Minimize ISI 
Bit rate optimal Low-medium
Disadvantages
Low-pass frequency response 
Computationally complex  
Artificial constraints 
Ad-hoc parameters 
Nonlinear optimization  
Unrealistic assumptions 
21
Contribution #3
Near-optimal Minimum-ISI (min-ISI) TEQ
2
ISIi  S x ,i q DHw
H
i
 ISI power in ith subchannel is
 Minimize ISI power as a frequency weighted sum of subchannel ISI
 i  i i
2
ISI  K q H
DHw  w T
Xw
i i

 Constrain signal path gain to one to prevent all-zero solution


|h | | GHw |  w Yw  1
signal 2 2 T

 Solution is a generalized eigenvector of X and Y


K i  S x ,i K i  S x ,i | H i |2 / S n ,i
 Possible weightings
 Performance virtually equal to that of the optimal method
 Generalizes MSSNR method by weighting in frequency domain 22
Contribution #3
Min-ISI TEQ vs. MSSNR TEQ
Method Min-ISI MSSNR
Advantages
Maximize channel capacity
Minimize ISI  
Frequency domain weighting 
Bit rate high high
Disadvantages
Computationally complex very high high

 Min-ISI weights ISI power with the SNR


 Residual ISI power should be placed in high noise frequency bands
1 1
SNR 50 
 0.1 SNR 50   0.09
signal power 10 10  1
SNR i  23
noise power  ISI power 1 1
SNR 2   10 SNR 2   0.9
0.1 0.1  1
Contribution #3
Efficient Implementations of Min-ISI Method

Xw k 1  Yw k

Method Capacity MACs


Original 99.6% 132,896
 Generalized eigenvalue problem can solved with generalized power iteration:
Recursive 99.5% 44,432
 Recursively calculate diagonal elements of X and Y from first column [Wu,
Row-rotation 99.5% 25,872
Arslan, Evans, 2000]
No-weighting 97.8% 10,064

24
Bit Rate vs. Number of TEQ Taps

 Min-ISI, MCC, and MSSNR perform


close to Matched Filter Bound (MFB)
even with small TEQ sizes

 Geometric and MMSE TEQ require 20


taps to achieve 90% of MFB
performance
TEQ taps
 Geometric TEQ gives little cyclic prefix () 32
improvement over MMSE TEQ FFT size (N) 512
coding gain 4.2 dB
 Two-tap min-ISI TEQ outperforms 21- margin 6 dB
tap MMSE TEQ input power 14 dBm
noise power -113 dBm/Hz,25
crosstalk noise 10 ADSL disturbers
Bit Rate vs. Number of TEQ Taps

 Min-ISI and MCC give virtually same


performance

 Min-ISI and MCC outperform MSSNR


by 2%

TEQ taps
 9 taps is enough for best performance
for min-ISI, MCC, and MSSNR TEQs cyclic prefix () 32
FFT size (N) 512
 No performance gain for more than 9 coding gain 4.2 dB
taps margin 6 dB
input power 14 dBm
noise power -113 dBm/Hz,26
crosstalk noise 10 ADSL disturbers
Bit Rate vs. Cyclic Prefix Size

 Min-ISI, MCC, and MSSNR perform


close to MFB

 Geometric and MMSE TEQ require


cyclic prefix of 30 samples

 Geometric TEQ gives worse


performance for short cyclic prefix

 Performance drops because cyclic TEQ taps 17


prefix does not carry new information FFT size (N) 512
coding gain 4.2 dB
 MSSNR does not work for cyclic margin 6 dB
prefix smaller than the number of input power 14 dBm
TEQ taps noise power -113 dBm/Hz,27
crosstalk noise 10 ADSL disturbers
Simulation Results

 Min-ISI, MCC, and MSSNR require


cyclic prefix of 17 samples to hit
matched filter bound performance

 Geometric and MMSE TEQs do not


work with 2 taps even with a cyclic
prefix of 32 samples

 Geometric TEQ gives lower


performance for small cyclic prefix TEQ taps 2
length FFT size (N) 512
coding gain 4.2 dB
 Min-ISI TEQ with 3-sample cyclic margin 6 dB
prefix outperforms MMSE TEQ with input power 14 dBm
32-sample cyclic prefix noise power -113 dBm/Hz,28
crosstalk noise 10 ADSL disturbers
Simulation Results for 17-tap TEQ
Achievable percentage of MFB bit rate Mbps
Loop MMSE Geo MSSNR Min-ISI MCC MFB
1 93% 93% 99% 100% 100% 1.07
2 97% 97% 98% 100% 100% 1.15
3 94% 95% 98% 100% 100% 1.12
4 91% 92% 100% 100% 100% 0.92
5 93% 93% 100% 100% 100% 1.09
6 93% 93% 99% 100% 100% 0.99
7 93% 94% 100% 100% 100% 1.06
8 93% 94% 100% 100% 100% 1.06

Cyclic prefix length of 32 Input power 14 dBm


FFT size (N) 512 Noise power -113 dBm/Hz
Coding gain 4.2 dB Crosstalk noise 10 ADSL disturbers
Margin 6 dB 29
Simulation Results for Two-Tap TEQ
Achievable percentage of MFB bit rate Mbps
Loop MMSE Geo MSSNR Min-ISI MCC MFB
1 29% 30% 99% 99% 99% 1.07
2 31% 30% 98% 99% 99% 1.15
3 33% 30% 99% 100% 100% 1.12
4 22% 23% 99% 99% 99% 0.92
5 29% 29% 99% 100% 100% 1.09
6 29% 29% 99% 99% 99% 0.99
7 27% 26% 99% 99% 99% 1.06
8 28% 29% 98% 99% 99% 1.06
Cyclic prefix length of 32 Input power 14 dBm
FFT size (N) 512 Noise power -113 dBm/Hz
Coding gain 4.2 dB Crosstalk noise 10 ADSL disturbers
Margin 6 dB 30
 Design TEQ to maximize channel capacity
 No previous method truly maximizes channel capacity
 New subchannel SNR model
 Partitions the equalized channel into signal, noise, and ISI paths
 Enables to write channel capacity as a function of equalizer taps
 New maximum channel capacity TEQ design method
 Good benchmark for all design methods
 Requires nonlinear optimization
 New minimum-ISI design method
Summary
 Virtually same performance as the optimal method

 Fast implementation using recursive calculations


31
MATLAB DMTTEQ Toolbox
 Toolbox features ten TEQ design methods

32
 Available at http://signal.ece.utexas.edu/~arslan/dmtteq/
 End-to-end optimization of channel capacity
 Joint optimization of bit loading and TEQ
 On-line adaptation of TEQ taps to track changes in channel
 Analysis of TEQ design methods
 Effect of analog transmit/receive filters and A/D and D/A converters
 Analyze performance under channel estimation errors
 Fixed-point analysis
 Extension to MCC and min-ISI methods
 Taking into account the noise floor
Future Research
 Modifications to subchannel SNR model
 Optimal frequency domain weighting in min-ISI method
33
Capacity of Additive White Gaussian Noise Channel
 Maximum theoretical capacity of an additive white Gaussian noise channel
(no inter-symbol interference) is

C  log 2 1  SNR  bits/symbol


• Maximum achievable capacity can be defined as
 SNR 
m  log 2 1   bits/symbol
  
 : SNR gap between theoretical and practical capacity
– Modulation method
– Coding gain
– Probability of error
– Margin for unaccounted distortions 34
Publications
35
 ADC: Analog digital converter  LAN: Local area network
 ADSL: Asymmetric DSL
 MCC: Maximum channel capacity
 CAD: Computer aided design
 MFB: Matched filter bound
 min-ISI: Minimum ISI
 CP: Cyclic prefix
 MMSE: Minimum MSE
 DAC: Digital-analog converter
 MSE: Mean squared error
 DMT: Discrete multitone
 MSSNR: Maximum SSNR
 DSL: Digital subscriber line
 QAM: Quadrature amplitude modulation
 FFT: Fast Fourier transform
 SNR: Signal-to-noise ratio
 HDSL: High-speed DSL
 SSNR: shortening SNR
 IFFT: Inverse FFT
 TEQ: Time domain equalizer
Acronyms
ISDN: Integrated service digital
network
 VDSL: Very-high-speed DSL

 ISI: Intersymbol interference


36

You might also like