Transcript IEEE C802.16m-09/0927r2

Differential Feedback Scheme for Closed-Loop Beamforming
IEEE 802.16 Presentation Submission Template (Rev. 9)
Document Number:
IEEE C80216m-09_0927
Date Submitted:
2009-03-07
Source:
Qinghua Li, Yuan Zhu, Eddie Lin, Shanshan Zheng,
Jiacheng Wang, Xiaofeng Liu, Feng Zhou, Guangjie Li,
Alexei Davydov, Huaning Niu, and Yang-seok Choi
Intel Corporation
Venue:
Session #61, Cairo, Egypt
Re:
TGm AWD
Base Contribution:
None
Purpose:
Discussion and adoption by TGm AWD
Notice:
E-mail:
[email protected]
[email protected]
This document does not represent the agreed views of the IEEE 802.16 Working Group or any of its subgroups. It represents only the views of the participants listed in
the “Source(s)” field above. It is offered as a basis for discussion. It is not binding on the contributor(s), who reserve(s) the right to add, amend or withdraw material
contained herein.
Release:
The contributor grants a free, irrevocable license to the IEEE to incorporate material contained in this contribution, and any modifications thereof, in the creation of an
IEEE Standards publication; to copyright in the IEEE’s name any IEEE Standards publication even though it may include portions of this contribution; and at the IEEE’s
sole discretion to permit others to reproduce in whole or in part the resulting IEEE Standards publication. The contributor also acknowledges and accepts that this
contribution may be made public by IEEE 802.16.
Patent Policy:
The contributor is familiar with the IEEE-SA Patent Policy and Procedures:
<http://standards.ieee.org/guides/bylaws/sect6-7.html#6> and <http://standards.ieee.org/guides/opman/sect6.html#6.3>.
Further information is located at <http://standards.ieee.org/board/pat/pat-material.html> and <http://standards.ieee.org/board/pat >.
1
Outline
• Introduction
• Differential feedback
– Scheme I: C80216m-09_0528r4, Qinghua Li, et al.
– Scheme II: C80216m-08_1187, Bruno Clerckx, et
al.
• Comparison of throughput, reliability,
overhead, and complexity
• Conclusions
• Proposed text
2
Signal model —matrix dimensions
y
=
H
V̂
• H is channel matrix of dimension
s
+ n
N r  Nt .
•
V̂
is beamforming matrix of dimension N t  N s.
•
s
is transmitted signal vector of dimension N s 1.
3
One-shot reset and differential
feedbacks
SS:
Initial feedback V̂ 1

Differential
feedback D̂
2
...
Differential
feedbackD̂ 10
 
Initial feedback V̂ 11
 
Differential
feedback D̂ 12
 
...
Time
Reset period
BS:
Beamforming
using V̂ 1

...
Beamforming
using V̂ 9

...
Time
4
There is always correlation between adjacent
precoders that can be utilized.
Set of ideal
beamforming matrixes
Vt 1
Vt 
• Adjacent beamforming matrixes are
not independent. Otherwise, the
beamforming gain vanishes before
the next feedback arrives.
Beamforming accuracy
Feedback
t-1 arrival
Feedback
t arrival
Time
5
Differential codebook — polar cap
Polar cap
Cd
Set of ideal
beamforming matrixes
ˆ t  1
V
ˆ t  1
V
V̂ t 
V t 
6
Scheme I
• Differentiation at SS:
D  Q t  1 Vt 
• Quantization at SS:
ˆ  arg max D H D
D
i
Beamforming
matrix needs to
be fed back.
H
DiCd
F
• Beamforming matrix reconstruction at BS:
ˆ t   Qt  1 D
ˆ
V
• Beamforming at BS:
Sanity check:
ˆ t s  n
y  HV
ˆ t   Qt  1 D
ˆ  Qt  1Q H t  1Vt   Vt 
V

I
7
Actual implementation
• Actual quantization at SS:


 H
H
H
ˆ
D  arg max det I 
Di Qt  1 H HQt  1 Di 
Ns
Di Cd


• Beamforming matrix reconstruction at BS:
ˆ t   Qt  1 D
ˆ
V
• Beamforming at BS:
ˆ t s  n
y  HV
Note that quantization criterion here is better than that in C80216m-09_0058r4.
8
Computation of Q(t-1)
Qt  1
Ω Vˆ t 1
Ns  1
ˆ t  1
V
ˆ t  1
V
ˆ t  1
V
M
e1  b b  b b 2
1
N s  2, N t  4
b1
b2
ReB   ImB 
b1
Qt  1
M
*
1,1 1
*
1, 2
m3
m3
b2
e4  b b  b b 2  b b3
m4
m4
*
4 ,1 1
b1
1
b2
b3
*
4, 2
*
4,3
b1
b2
b3
b4
1
• Low computational complexity
– Householder matrix for rank 1
– Gram-Schmidt for rank 2
• Add no hardware
– Reuse hardware of the mandatory, transformed codebook
9
Scheme II
• Actual quantization at SS:


 ˆH
H
H
ˆ
ˆ
D  arg max det I 
V t  1Di H HDi Vt  1
Ns
Di Cd


• Beamforming matrix reconstruction at BS:
ˆ t   D
ˆV
ˆ t  1
V
•Beamforming at BS:
ˆ t s  n
y  HV
Note that quantization criterion here is better than that in C80216m-08_1187.
10
Updates of Scheme II
• Identity matrix is included into the codebook
lately.
• Polar cap size is reduced through increased
correlation 0.95 for highly correlated channels,
and remains at 0.9 for uncorrelated channels.
11
Scheme I and Scheme II track perturbations of e1 and [e1
e2 e3 e4], respectively.
• Scheme I’s codewords quantize perturbations of e1 on unit circle with 6
degrees of freedom (DOF), while Scheme II’s codewords quantize
perturbations of whole identity matrix on Stiefel manifold with 12 DOF.
Scheme I
Scheme II
1
 0.9 
0 
0.05
   

 Di  
0 
 0.01
 


0
0
.
00
 


1

 0.9 0.02 0.00 0.00
 1

0.06 0.9 0.03 0.01

  

 Di  

 0.01 0.05 0.9 0.07
1 




1
0
.
00
0
.
00
0
.
04
0.9




12
Scheme I codebook matches to precoder delta
distribution while Scheme II’s doesn’t.
• Differential matrix is symmetric about center [e1] or identity matrix.
• Scheme I’s codebook is symmetric about [e1], while Scheme II’s codebook
is asymmetric about identity matrix with uneven quantization errors.
Scheme II
Scheme I
V(t-1)
V(t-1)
Rotation angle
Rotation angle
13
Grassmainian manifold
Grassmainian manifold
V(t)
V(t)
Other differences
• Correlation adaptation
– Scheme I has two codebooks for small and high correlation
scenarios, respectively. The two codebooks are pre-defined and
stored.
– Scheme II varies codebook using measured correlation matrix
and costly online SVD computation.
• Rank adaptation
– Scheme I can change precoder rank anytime.
– Scheme II can not change precoder rank during differential
feedbacks and has to wait until next reset.
14
The complexity of Scheme II’s 4-bit version is more
than triple of Scheme I's 3-bit because Scheme II uses
4x4 matrix operation rather than 4x1 or 4x2.
15
Scheme II's 4-bit complexity is more than 50% of
Scheme I's 4-bit because Scheme II uses 4x4 matrix
operation rather than 4x1 or 4x2.
16
Summary of Differences
Scheme I
Scheme II
Principle
Track perturbation of e1 or
[e1 e2].
Track perturbation of [e1 e2 e3 e4].
Codebook dimension
4x1 and 4x2
4x4
Computational Complexity
Low
High
No. of codebooks
One small codebook for
each rank
One large codebook for all ranks
Codeword distribution
Even distribution on
Grassmannian manifold
Uneven distribution on Grassmannian
manifold
Rank adaptation
Rank can be changed at
anytime.
Rank changes only after reset.
Correlation adaptation
Two predefined codebooks
for small and large
correlation scenarios,
respectively.
Adaptive codebooks with online SVD
computation for different correlation
scenarios.
Support of 8 antennas
Easy
Difficult because of wasted codewords
and high complexity.
17
Comparison of throughput and reliability
• System level simulation
• Single-user MIMO
• Implementation losses are included
– Feedback error and error propagation
– Quantized reset feedback
18
General SLS parameters
Parameter Names
Parameter Values
Network Topology
57 sectors wrap around, 10 MS/sector
MS channel
ITU PB3km/h
Frame structure
TDD, 5DL, 3 UL
Feedback delay
5ms
Inter cell interference modeling
Channel is modeled as one tap wide band
Antenna configuration
4Tx, 2 Rx
Codebook configuration
Baseline 6 bits, Diff 4 bits or 3 bits
Q matrix reset frequency
Once every 4 frames
PMI error
free
PMI calculation
Maximize post SINR
System bandwidth
10MHz, 864 data subcarriers
Permutation type
AMC, 48 LRU
CQI feedback
1Subband=4 LRU, ideal feedback
19
Codebook related parameters
Samsung diff.
codebook
Intel diff. 4-bit
codebook
Intel diff. 3-bit
codebook
Codebook size
4 bits i.e. 16
codewords
4 bits i.e. 16
codewords
3 bits i.e. 8
codewords
Feedback
overhead
18 bits / 4 frames
/ Subband
including 6-bit
reset
18 bits / 4 frames /
Subband including
6-bit reset
15 bits / 3 frames
/ Subband
including 6-bit
reset
CQI erasure rate
10%
20
3-bit Scheme I vs. 4-bit Scheme II
21
4 Tx, 2 Rx, 1 stream
0.5λScheme I
5o, Samsung
0.9
4λScheme I
20o, Samsung
0.9
Uncorrelated,
Scheme I 20o,
Samsung 0.9
SE gain 5%-ile SE
over 16e gain over
16e
SE gain
over 16e
5%-ile SE
gain over
16e
SE gain
over 16e
5%-ile SE gain
over 16e
Scheme I: 3bit
11.6%
36.3%
3.7%
21.9%
2.5%
16.2%
Scheme II:
4-bit
10.7%
35.3%
1.3%
19%
-1%
7.7%
Scheme I
over
Scheme II
0.8%
0.7%
2.7%
2.4%
3.5%
7.9%
22
Remarks
• Scheme I's 3-bit has higher throughput and
reliability than Scheme II's 4-bit. In addition,
Scheme I's feedback overhead and complexity
are lower than Scheme II's.
• Scheme II's codebook is optimized for highly
correlated channels and scarifies
uncorrelated/lowly correlated channels.
• Scheme II's codebook can not track channel
variation in uncorrelated channel and performs
even poorer than 16e codebook.
23
Scheme I's 4-bit vs. Scheme II's 4-bit
24
4 Tx (0.5λ), 2 Rx, 1 stream
4Tx 0.5L, 2Rx, rank 1 user Rx SE CDF
SE gain
over
16e
Scheme I: 11.7%
4-bit, 5o
5%-ile
SE gain
over 16e
37.7%
1
0.9
0.8
0.7
0.6
0.5
Samsung: 10.6%
4-bit,
0.95 ρ
35.4%
0.4
0.3
0.2
Scheme I
over
Samsung
1.06%
1.37%
16e413
16m416 Intel Diff414,5, 4F reset
16m416 Samsung Diff414,0.95, 4F reset
0.1
0
0
1
2
3
4
user RX SE (b/s/Hz)
5
6
7
25
4 Tx(4λ), 2Rx, 1 stream
4Tx 4L, 2Rx, rank 1 user Rx SE CDF
SE gain
over
16e
Scheme I: 3.9%
4-bit, 20o
5%-ile
SE gain
over 16e
21.5%
1
0.9
0.8
0.7
0.6
Samsung: 1.36%
4-bit, 0.9
ρ
19%
0.5
0.4
0.3
Scheme I
over
Samsung
2.5%
2.1%
16e413
16m416 Intel Diff414,20, 4F reset
16m416 Samsung Diff414,0.9, 4F reset
0.2
0.1
0
0
1
2
3
4
user RX SE (b/s/Hz)
5
6
7
26
4 Tx (uncorrelated), 2 Rx, 1 stream
4Tx U, 2Rx, rank 1 user Rx SE CDF
SE gain
over 16e
Scheme I:
4-bit, 20o
0.2%
5%-ile SE
gain over
16e
17.4%
1
0.9
0.8
0.7
0.6
Samsung:
4-bit, 0.9
-2.9%
13%
0.5
0.4
ρ
0.3
Scheme I
over
Samsung
3.1%
3.9%
16e413
16m416 Intel Diff41-24,20, 4F reset
16m416 Samsung Diff41-24,0.9, 4F reset
0.2
0.1
0
0
1
2
3
4
user RX SE (b/s/Hz)
5
6
7
27
Comparison on rank-2
codebooks
28
4 Tx (4λ), 2 Rx, 2 streams
SE gain
over 16e
4Tx 4L, 2Rx, rank 2 user Rx SE CDF, Round Robin
1
0.9
Scheme I:
4-bit, 20o
10.29%
0.8
0.7
Samsung:
4-bit, 0.9
10%
0.5
ρ
Scheme I
over
Samsung
0.6
0.4
0.27%
16e41-23
16m41-26 Intel Diff41-24,20, 4F reset
16m41-26 Samsung Diff41-24,0.9, 4F reset
0.3
0.2
0.1
0
1
1.5
2
2.5
3
3.5
user RX SE (b/s/Hz)
4
4.5
5
29
4 Tx (uncorrelated), 2 Rx, 2 streams
4Tx U, 2Rx, rank 2 user Rx SE CDF, Round Robin
1
SE gain
over 16e
0.9
0.8
Scheme I:
4-bit, 20o
9.45%
0.7
0.6
Samsung:
4-bit, 0.9
9.28%
ρ
Scheme I
over
Samsung
0.5
0.4
0.15%
0.3
0.2
16e41-23
16m41-26 Intel Diff41-24,20, 4F reset
16m41-26 Samsung Diff41-24,0.9, 4F reset
0.1
0
1
1.5
2
2.5
3
3.5
user RX SE (b/s/Hz)
4
4.5
5
30
Remarks
• Scheme I's 4-bit has higher throughput
and reliability than Scheme II's 4-bit.
• Scheme I's complexity is lower than
Scheme II's.
• Scheme I's 4-bit has 0.3% higher
throughput than Scheme I's 3-bit.
31
Conclusions
• Scheme I's 3-bit outperforms Scheme II's 4-bit in
all cases in terms of throughput and reliability.
• Scheme I's 3-bit scheme has feedback overhead
and computational complexity lower than
Scheme II's 4-bit by 17% and 60%, respectively.
• Scheme I's 4-bit has even higher throughput
than Scheme I's 3-bit.
• Scheme II's new design solves vibration problem
by adding identity matrix but it can not track
channel variation in uncorrelated channels.
32
Proposed text
Add proposed text to line 63, page 91, section
15.3.7.2.6.6.4.
33
34
35
36
37
38
39
40
41
42
43
44
45