Design and Analysis of LDPC for MIMO-OFDM

Guosen Yue
NEC Labs Research
Princeton, NJ
Joint work with
Ben Lu
Xiaodong Wang (Columbia Univ.)
Outline
LDPC coded MIMO OFDM
Analysis & Optimization of (irregular) LDPC Coded MIMO OFDM
A few practical issues: Dierent number of antennas; dierent MIMO demodulation
schemes; dierent spatial correlation models
Large-code-length: Optimization of degree proles by density evolution with Gaussian
approximation
Short-code-length: Random construction with girth conditioning
Numerical examples and conclusions
Problem Statement
Future personal wireless communications
A popular vision: IP-based multimedia wireless services with both ubiquitous coverage
( cellular) and high speed ( Wi-Fi).
A narrow-sense engineering vision: wireless packet IP data communications with high
throughput and low latency.
Enabling techniques for high-speed wireless packet data
PHY layer: MIMO, advanced FEC, advanced DSP, adaptive transmission, ...
MAC layer: channel-aware scheduling, multi-access, fast ARQ, interference control, ...
Networking layer, cross-layer, . . .
In this work, we focus on the peak date-rate of downlink transmission
Low-Density Parity-Check (LDPC) Codes
Invented by R. Gallager in 1962; re-discovered by Mackay & Neal in 1997, by Richardson & Shokrollahi
& Urbanke in 1999.
LDPC is a linear block code dened by a very sparse parity check matrix; or equivalently by a bipartite
(Tanner) graph (variable nodes, check nodes and connecting edges).
LDPC codes subsume a class of capacity-approaching codes, e.g., turbo codes, RA codes.
Decoding complexity of LDPC codes is lower than turbo codes, and suitable for parallel processing.
Regular LDPC codes: same number of 1s in each column and row of the sparse parity check matrix.
Irregular LDPC codes: dierent number of 1s ...... Large-code-size irregular LDPC: degree proles.
Deterministic LDPC construction: array codes [Fan 99], graph theory [Lin 02], . . .
Pseudo-random LDPC construction: convergence to ensemble average theorem for large-code-size
[Gallager 63], girth conditioning for moderate/short-code-size [Campeliot & Modha & Rajagopalan
99, Yang & Ryan 02, Tian & Jones & Villasenor & Wesel 02].
LDPC Code Optimization
Previous works on LDPC optimization
for AWGN channels by density evolution [Richardson & Shokrollahi & Urbanke, 01]
for AWGN channels by density evolution with Gaussian approx [Chung & Forney & Richardson & Urbanke, 01]
for Rayleigh fading channels by density evolution with mixture Gaussian approx [Hou & Siegel & Milstein, 01]
for ISI channels by density evolution with mixture Gaussian approx [Narayanan & Wang & Yue, 02]
for MIMO channels by EXIT Chart [tenBrink & Kramer & Ashikhmin, 02, ]
...
In this work
optimization for MIMO OFDM channels by density evolution with mixture Gaussian approx.
number of antennas and bandwidth: use of MIMO technique to support the same data rate with less bandwidth
(i.e., higher spectral eciency).
low-complexity iterative receiver: use of low-complexity soft LMMSE-SIC MIMO demodulator, as opposed to
exponentially complex soft MAP MIMO demodulator.
spatially correlated MIMO: non-full-scattering scenario (due to limited antenna separation or angle spread)
LDPC Coded MIMO OFDM for 4G Downlink
MIMO: multiple-antennas at both transmit and receive sides; establish the multi-fold
virtual air-links, the spatial resource not regulated by FCC.
OFDM: low-complexity in dispersive channels; easy bond with multiuser scheduler; a highly
competitive solution for (synchronous) downlink transmission.
LDPC: capacity-approaching; low-complexity & parallizable decoder; freedom for design
and performance optimization.
.
.
.
LDPC
Encoder Modulator
MPSK
Bits Symbols
Coded
IFFT
IFFT
IFFT
Info. Coded
S/P
Bits
.
.
.
FFT
FFT
FFT

e
1
1
2
M
Turbo iterative demodulation & decoding
Demod.
Soft LDPC
Decoder
Decision
Info. Bits

e
2
Turbo Iterative Demodulation and Decoding
[1] Iteration of turbo receiver: For q = 1, 2, . . . , Q
[1-a] Soft MIMO OFDM demodulation: L
q
DL
[b
i
] = g({r(t)}, {L
q1
DL
[b
j
]}
j
),
[1-b] Soft LDPC decoding: For p = 1, 2, . . . , P
Sum-product algorithm: for all variable nodes and check nodes
Variable node update: L
p,q
bc
(e
b
i,j
) = L
q
mL
[b
k
(i)] +

i
n=1,n=j
L
p1,q
bc
(e
b
i,n
).
Check node update: L
p,q
bc
(e
c
i,j
) = 2 tanh
1
_

i
n=1,n=j
tanh
_
L
p,q
bc
(e
c
i,n
)
2
__
.
[1-c] Compute extrinsic messages passed back to the multiuser detector:
L
q
DL
[b
i
] =

n=1
L
P,q
bc
(e
b
i,n
).
[2] Final hard decisions on information and parity bits:

b
i
= sign
_
L
Q
DL
[b
i
] + L
Q
DL
[b
i
]
_
.
Analysis & Optimization of LDPC Coded MIMO OFDM
Degree proles of LDPC: (x) =
d
lmax

i=1

i
x
i1
and (x) =
d
rmax

i=1

i
x
i1
Optimization problem
(

(x),

(x)) = arg min

_
L
Q
DL
[b
i
] + L
Q
DL
[b
i
]
_

.
Basic idea: track the dynamics of turbo iterative demodulation and decoding.
Major assumptions and approximations
Assume the extrinsic LLR at each variable node or check node of LDPC codes is Gaussian and
symmetric, i.e., N(m, 2m).
Assume the LLR from LDPC decoder to MIMO demodulator as mixture Gaussian f
q
DL

d
l
,max
j=2

j
N(m
j
, 2m
j
). due to sum-product algorithm
Approx the LLR from MIMO demodulator to LDPC decoder as mixture Gaussian f
q
DL

J
i=1

i
N(m
i
, 2m
i
). using EM algorithm
We then only need to track parameters of mixture Gaussians, {
i
, m
i
}
i
, rather than complete pdfs.
Analysis & Optimization of LDPC MIMO OFDM
Turbo receiver iterations: For q = 1, 2, . . . , Q
Mixture Gaussian approx of extrinsic LLR of MIMO demodulator:
f
q
DL
=
J

j=1

j
N(
j
, 2
j
)
Mixture Gaussian approx of extrinsic LLR of LDPC decoder:
2 Iterate between variable node update and check node update: For p = 1, 2, . . . , P
At a bit node of degree i:
f
p,q
bc
=
J

j=1
d
l,max

i=2

i
N
_

j
+ (i 1)m
p1,q
bc
, 2[
j
+ (i 1)m
p1,q
bc
]
_
At check node of degree j:
f
p,q
bc
=
dr,max

j=2

j
N
_
m
p,q
bc,j
, 2m
p,q
bc,j
_
2 Message passed back to the multiuser detector:
f
P,q
DL
=
d
l,max

i=2

i
N (m
q
DL
(i), 2m
q
DL
(i))
The optimized SNR threshold
(

(x),

(x)) = arg min

_
L
Q
DL
[b
i
] + L
Q
DL
[b
i
]
_

.
Performance in Ergodic Channels w/o Spatial Correlation
Within 1.0 dB from channel capacity
0 0.5 1 1.5 2 2.5 3 3.5 4
10
6
10
5
10
4
10
3
10
2
10
1
Large size LDPC code (n=880,640), 1x1 Uncorrelated MIMOOFDM
SNR (dB)
B
i
t

E
r
r
o
r

R
a
t
e

(
B
E
R
)
Capacity
MAP+reg_LDPC D.E.
MAP+reg_LDPC Simu
SIC+reg_LDPC D.E.
SIC+reg_LDPC Simu
MAP+irr_LDPC D.E.
MAP+irr_LDPC Simu
SIC+irr_LDPC D.E.
SIC+irr_LDPC Simu
Figure 1: Large-block-size LDPC in 1 1 MIMO OFDM.
Performance in Ergodic Channels w/o Spatial Correlation
Within 1.0 dB from channel capacity
0 0.5 1 1.5 2 2.5 3 3.5 4
10
6
10
5
10
4
10
3
10
2
10
1
Large size LDPC code (n=880,640), 4x4 Uncorrelated MIMOOFDM
SNR (dB)
B
i
t

E
r
r
o
r

R
a
t
e

(
B
E
R
)
Capacity
MAP+reg_LDPC D.E.
MAP+reg_LDPC Simu
SIC+reg_LDPC D.E.
SIC+reg_LDPC Simu
MAP+irr_LDPC D.E.
MAP+irr_LDPC Simu
SIC+irr_LDPC D.E.
SIC+irr_LDPC Simu
Figure 2: Large-block-size LDPC in 4 4 MIMO OFDM.
Performance in Ergodic Channels with Spatial Correlation
LMMSE-SIC demodulator suers extra loss due to spatial correlation
2 2.5 3 3.5 4 4.5 5 5.5 6
10
6
10
5
10
4
10
3
10
2
10
1
Large size LDPC code (n=880,640), 4x4 Correlated MIMOOFDM
SNR (dB)
B
i
t

E
r
r
o
r

R
a
t
e

(
B
E
R
)
Capacity
MAP+reg_LDPC D.E.
MAP+reg_LDPC Simu
SIC+reg_LDPC D.E.
SIC+reg_LDPC Simu
MAP+irr_LDPC D.E.
MAP+irr_LDPC Simu
SIC+irr_LDPC D.E.
SIC+irr_LDPC Simu
Figure 3: Large-block-size LDPC in 4 4 MIMO OFDM.
Performance in Outage Channels
Within 1.5 dB from channel capacity
0 1 2 3 4 5 6 7 8 9
10
3
10
2
10
1
10
0
Small size LDPC code (n=2048), 4x4 Uncorrelated MIMOOFDM
F
r
a
m
e

E
r
r
o
r

R
a
t
e
SNR (dB)
Capacity
MAP+reg_LDPC
SIC+reg_LDPC
MAP+irr_LDPC
SIC+irr_LDPC
Figure 4: Short-block-size LDPC in 4 4 MIMO OFDM, target FER of 10
2
.
Performance in Outage Channels:
Convergence of Turbo Iterative Receiver
Irregular LDPC expedites the convergence of overall turbo receiver
1.5 2 2.5 3 3.5 4 4.5 5 5.5
4.5
5
5.5
6
6.5
7
Small size LDPC code (n=2048), 4x4 Uncorrelated MIMOOFDM
R
e
q
u
i
r
e
d

S
N
R

t
o

a
c
h
i
e
v
e

F
E
R

o
f

1
0

2
(
d
B
)
Number of turbo receiver iteration
MAP+reg_LDPC
SIC+reg_LDPC
MAP+irr_LDPC
SIC+irr_LDPC
Figure 5: Short-block-size LDPC in 4 4 MIMO OFDM, target FER of 10
2
.
Gain of Channel-Specic LDPC Design
Design gain of MIMO-OFDM-optimized LDPC increases for larger number of antennas, as
compared to AWGN-optimized LDPC.
Large Block Irregular LDPC Small Block Irregular LDPC
SNR (dB) LDPC.I LDPC.II Channel-specic Design LDPC.I LDPC.II Channel-specic Design
Gain (LDPC.II - LDPC.I) Gain (LDPC.II - LDPC.I)
MAP (1 1) 2.57 2.57 0.00 7.08 7.08 0.00
MAP (2 2) 2.56 2.61 0.05 5.57 5.72 0.15
MAP (4 4) 2.46 2.65 0.19 4.48 4.81 0.33
SIC (1 1) 2.52 2.52 0.00 7.06 7.06 0.00
SIC (2 2) 2.75 2.92 0.17 6.32 6.44 0.12
SIC (4 4) 2.82 3.17 0.35 5.33 5.70 0.37
LDPC.I: Performance of MIMO-OFDM-optimized LDPC in MIMO-OFDM channels.
LDPC.II: Performance of AWGN-optimized LDPC in MIMO-OFDM channels.
Summary
LDPC coded MIMO OFDM is capable of supporting 4G wireless packet data transmission
with higher spectral eciency which translates into either bandwidth saving or further
data rate increase.
In ergodic channels, channel-specic (irregular) LDPC optimization results in larger SNR
gain in systems with larger number of antennas.
In outage channels, irregular LDPC codes lead to faster receiver convergence.
LMMSE-SIC based receiver performs near-optimal in spatially uncorrelated MIMO OFDM
channels; but suers additional loss in MIMO channels with severe spatial correlation.