Limited Feedback v3
-
Upload
kostasntougias5453 -
Category
Documents
-
view
4 -
download
0
description
Transcript of Limited Feedback v3
Limited Feedback for Single- and Multi-User MIMO389.168 Advanced Wireless Communications 1
Contents
1 Motivation
2 Codebook based Single-User MIMO Feedback for LTE
3 Multi-User MIMO Feedback for LTE
4 Multi-User MIMO with Channel Subspace Feedback
5 Multi-User MIMO with Channel Gramian Feedback
6 Feedback Overhead Reduction through Excess Antennas
Slide 2 / 89 Contents
Contents
1 Motivation
2 Codebook based Single-User MIMO Feedback for LTE
3 Multi-User MIMO Feedback for LTE
4 Multi-User MIMO with Channel Subspace Feedback
5 Multi-User MIMO with Channel Gramian Feedback
6 Feedback Overhead Reduction through Excess Antennas
Slide 3 / 89 Motivation
Limited Feedback: What and Why?
Value of channel state information at the transmitter (CSIT):
Channel quality information (e.g., SINR, achievable rate):
Improve robustness and efficiency through rate adaptation
Exploit channel and multi-user diversity through scheduling
Channel “matrix” information:
Single-user communication: mostly a power/beamforming gain(water-filling versus equal power)
Multi-user (point) communication: capacity/multiplexing gain(exploitation of degrees of freedom)
Slide 4 / 89 Motivation
Limited Feedback: What and Why?
Value of channel state information at the transmitter (CSIT):
Channel quality information (e.g., SINR, achievable rate):
Improve robustness and efficiency through rate adaptation
Exploit channel and multi-user diversity through scheduling
Channel “matrix” information:
Single-user communication: mostly a power/beamforming gain(water-filling versus equal power)
Multi-user (point) communication: capacity/multiplexing gain(exploitation of degrees of freedom)
Slide 4 / 89 Motivation
Performance of LTE’s Modulation and Coding Schemes
302520151050-5
5
4
3
2
1
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh
CQI 1
CQI 15
Frequency flat Rayleigh fading channel with 1.4 MHz bandwidth
Throughput over average SNR for transmission with fixed rate
15 MCSs with rates ranging form 0.15 bit/cu to 5.55 bit/cu
Slide 5 / 89 Motivation
Scheduling Gains through CQI Feedback [Schwarz et al., 2010]
0 5 10 15 20 250
2
4
6
8
10
UE index (= UE SNR [dB])
Thr
ough
put [
Mbi
t/s]
RRMaxMinPFBCQI
25 users with average SNRs ranging from 1 to 25 dB
Round robin: assign resources in consecutive order
MaxMin: maximize minimum user throughput – throughput equalization
Proportional fair : maxjTj
T j
Best CQI: schedule user with highest CQI
Slide 6 / 89 Motivation
Scheduling Gains through CQI Feedback (2)
RR MaxMin PF BCQI0
5
10
15
20
25
30
35
40
45
Scheduler
Thr
ough
put [
Mbi
t/s]
RR MaxMin PF BCQI0
0.2
0.4
0.6
0.8
1
Scheduler
Jain
’s F
airn
ess
Inde
x
Comparison of sum throughput and fairness of resource allocation
Jain’s fairness index :
J =
(∑Jj=1 T j
)2
J∑J
j=1 T2j
(1)
Ranges from 1 (highest fairness = equal throughput) to 1/J
Slide 7 / 89 Motivation
Scheduling with Fairness Constraint [Schwarz et al., 2011]
Throughput user 1 [bits per channel use]
Thr
ough
put u
ser 2
[bits
per
cha
nnel
use
]
0 0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
Maxmin. solution
Prop. fair solution
Max. throughput solution
Rate region J = 1
J = 0.9
J = 0.73
Two-user achievable rate region with operating points of some schedulers
A fairness constraint J ≥ J0 cuts out a convex cone
Operating points of α-fair sum-utility maximization
maximize:J∑
j=1
Uα(
T (j)), Uα(x) =
{ x1−α
1− α, α ≥ 0, α 6= 1
log(x), α = 1(2)
Slide 8 / 89 Motivation
Scheduling with Fairness Constraint [Schwarz et al., 2011]
Throughput user 1 [bits per channel use]
Thr
ough
put u
ser 2
[bits
per
cha
nnel
use
]
0 0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
J ≥ 0.85
J ≥ 0.70
Rate region
Two-user achievable rate region with operating points of some schedulers
A fairness constraint J ≥ J0 cuts out a convex cone
Operating points of α-fair sum-utility maximization
maximize:J∑
j=1
Uα(
T (j)), Uα(x) =
{ x1−α
1− α, α ≥ 0, α 6= 1
log(x), α = 1(2)
Slide 8 / 89 Motivation
Scheduling with Fairness Constraint [Schwarz et al., 2011]
Throughput user 1 [bits per channel use]
Thr
ough
put u
ser 2
[bits
per
cha
nnel
use
]
α
0 0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
= 1
αα = 1000
α = 0
J ≥ 0.85
J ≥ 0.70
Rate region
Two-user achievable rate region with operating points of some schedulers
A fairness constraint J ≥ J0 cuts out a convex cone
Operating points of α-fair sum-utility maximization
maximize:J∑
j=1
Uα(
T (j)), Uα(x) =
{ x1−α
1− α, α ≥ 0, α 6= 1
log(x), α = 1(2)
Slide 8 / 89 Motivation
Limited Feedback: What and Why? (2)
Acquisition of CSIT:
TDD or full-duplex: channel estimation at the transmitter possible
Careful calibration of uplink/downlink chains to ensure reciprocity
Timing synchronization difficult
FDD: CSIT cannot be estimated by the transmitter
Currently dominates the field
Explicit CSI feedback from the receiver required
Limited uplink overhead ⇒ quantization
Slide 9 / 89 Motivation
Limited Feedback: What and Why? (2)
Acquisition of CSIT:
TDD or full-duplex: channel estimation at the transmitter possible
Careful calibration of uplink/downlink chains to ensure reciprocity
Timing synchronization difficult
FDD: CSIT cannot be estimated by the transmitter
Currently dominates the field
Explicit CSI feedback from the receiver required
Limited uplink overhead ⇒ quantization
Slide 9 / 89 Motivation
Contents
1 Motivation
2 Codebook based Single-User MIMO Feedback for LTE
3 Multi-User MIMO Feedback for LTE
4 Multi-User MIMO with Channel Subspace Feedback
5 Multi-User MIMO with Channel Gramian Feedback
6 Feedback Overhead Reduction through Excess Antennas
Slide 10 / 89 Codebook based Single-User MIMO Feedback for LTE
Direct Channel Quantization
Remember the capacity-optimal SVD transceiver
H = UΣVH (3)
⇒ F = VP1/2, G = UH (4)
For precoder calculation the transmitter requires V and diag (Σ)
Direct quantization of H in Euclidean space [Zheng and Rao, 2008]
H = argminHq∈H
‖H− Hq‖ , H ⊂ CNr×Nt , (5)
H = UΣVH (6)
Inefficient , as the structure of the quantization problem is neglected
Slide 11 / 89 Codebook based Single-User MIMO Feedback for LTE
Direct Channel Quantization
Remember the capacity-optimal SVD transceiver
H = UΣVH (3)
⇒ F = VP1/2, G = UH (4)
For precoder calculation the transmitter requires V and diag (Σ)
Direct quantization of H in Euclidean space [Zheng and Rao, 2008]
H = argminHq∈H
‖H− Hq‖ , H ⊂ CNr×Nt , (5)
H = UΣVH (6)
Inefficient , as the structure of the quantization problem is neglected
Slide 11 / 89 Codebook based Single-User MIMO Feedback for LTE
Direct Channel Quantization (2)
Separate quantization exploiting the structure of the problem[Schwarz and Rupp, 2014b]
σ = argminsi∈S
‖σ − si‖22 , σ =
[[Σ](1,1), . . . , [Σ](Nr ,Nr )
]T, (7)
V = argminQi∈Q
d2c
([Qi ](:,j) , [V](:,j)
)= argmin
Qi∈QNr −
∣∣∣[Qi ]H(:,j) [V](:,j)
∣∣∣2 , (8)
S ⊂ RNr×1+ , Q =
{Qi ∈ CNt×Nr
∣∣QHi Qi = INr
}⊂ St(Nt ,Nr ) (9)
Notice the order of the singular-vectors is important for power allocation⇒compact Stiefel manifold St(Nt ,Nr )
Slide 12 / 89 Codebook based Single-User MIMO Feedback for LTE
Direct Channel Quantization (3)
Water-filling power allocation provides mostly only a minor gain over on-offswitching of modes⇒ employ equal power allocation over L active modes
Grassmannian quantization with rank selection
V = argminQi∈Q
d2c (Qi ,VL) = argmin
Qi∈QL− tr
(QH
i VLVHL Qi), VL = [V]:,1:L , (10)
L = number of non-zero elements (diag (P)) , (11)
Q ={
Qi ∈ CNt×L∣∣QHi Qi = IL
}⊂ G(Nt , L) (12)
Only subspace information span(VL)
required⇒ Grassmannian quantization
Unitary rotations of Qi are irrelevant
Qi ≡ Qj ⇔ Qi = Qj U, UHU = UUH = IL, (13)
log2 det(INr + HQi QH
i HH) = log2 det(
INr + HQj UUHQHj HH
)=
log2 det(
INr + HQj QHj HH
)(14)
⇒ feedback overhead reduction compared to Stiefel manifold quantization
Slide 13 / 89 Codebook based Single-User MIMO Feedback for LTE
Direct Channel Quantization (3)
Water-filling power allocation provides mostly only a minor gain over on-offswitching of modes⇒ employ equal power allocation over L active modes
Grassmannian quantization with rank selection
V = argminQi∈Q
d2c (Qi ,VL) = argmin
Qi∈QL− tr
(QH
i VLVHL Qi), VL = [V]:,1:L , (10)
L = number of non-zero elements (diag (P)) , (11)
Q ={
Qi ∈ CNt×L∣∣QHi Qi = IL
}⊂ G(Nt , L) (12)
Only subspace information span(VL)
required⇒ Grassmannian quantization
Unitary rotations of Qi are irrelevant
Qi ≡ Qj ⇔ Qi = Qj U, UHU = UUH = IL, (13)
log2 det(INr + HQi QH
i HH) = log2 det(
INr + HQj UUHQHj HH
)=
log2 det(
INr + HQj QHj HH
)(14)
⇒ feedback overhead reduction compared to Stiefel manifold quantization
Slide 13 / 89 Codebook based Single-User MIMO Feedback for LTE
Direct Channel Quantization (3)
Water-filling power allocation provides mostly only a minor gain over on-offswitching of modes⇒ employ equal power allocation over L active modes
Grassmannian quantization with rank selection
V = argminQi∈Q
d2c (Qi ,VL) = argmin
Qi∈QL− tr
(QH
i VLVHL Qi), VL = [V]:,1:L , (10)
L = number of non-zero elements (diag (P)) , (11)
Q ={
Qi ∈ CNt×L∣∣QHi Qi = IL
}⊂ G(Nt , L) (12)
Only subspace information span(VL)
required⇒ Grassmannian quantization
Unitary rotations of Qi are irrelevant
Qi ≡ Qj ⇔ Qi = Qj U, UHU = UUH = IL, (13)
log2 det(INr + HQi QH
i HH) = log2 det(
INr + HQj UUHQHj HH
)=
log2 det(
INr + HQj QHj HH
)(14)
⇒ feedback overhead reduction compared to Stiefel manifold quantization
Slide 13 / 89 Codebook based Single-User MIMO Feedback for LTE
Codebook Based Precoding
Instead of quantizing the channel, let the user directly select the precoderfrom a given codebook F
Outperforms direct channel quantization [Love and Heath, Jr., 2005]
Optimal quantization codebook constructions [Love and Heath, Jr., 2005]
Maximally spaced subspace packings on the Grassmannian
Difficult to obtain in general(some pre-calculated codebooks can be found at [Love, 2006])
Algorithm for finding good codebooks [Dhillon et al., 2007]
Slide 14 / 89 Codebook based Single-User MIMO Feedback for LTE
Codebook Based Precoding
Instead of quantizing the channel, let the user directly select the precoderfrom a given codebook F
Outperforms direct channel quantization [Love and Heath, Jr., 2005]
Optimal quantization codebook constructions [Love and Heath, Jr., 2005]
Maximally spaced subspace packings on the Grassmannian
Difficult to obtain in general(some pre-calculated codebooks can be found at [Love, 2006])
Algorithm for finding good codebooks [Dhillon et al., 2007]
Slide 14 / 89 Codebook based Single-User MIMO Feedback for LTE
Codebook Based Precoding (2)
More practical codebook constructions (low complexity implementation)
Constant modulus (phase-shifts only), nested codebooks
Codebooks based on DFT [Yang et al., 2010]
Codebooks based on QAM constellations [Ryan et al., 2007]
This approach is applied in LTE
Slide 15 / 89 Codebook based Single-User MIMO Feedback for LTE
LTE’s Signal Processing Chain
Coding
Interleaving
Segmentation
AMC scheme selection
Symbol
constellation
mapping
Spatial
layer
mapping
MIMO preprocessing
Spatial
precoding (Ws)
User data Code words c Symbols Layers (L)
Other user data
Receiver
Equalization
Demapping
DecodingTransmit
signal
Receive
signal
H
Nr
Feedback
calculation
Channel quality
indicator (CQI)
Precoding matrix
indicator (PMI)
Rank indicator (RI)Resource
element
mapping
Transmit signal generation
IFFT
CP insertion
IFFT
CP insertion
System
overhead
insertion
System
overhead
insertion
Nt
Resource
element
mapping
(Nt)
Adaptive modulation and coding (AMC)
Multiple MCSsM to adapt the rate to the channel conditions
Preferred MCS signalled by means of CQI feedback
Multiple code words to account for different channel qualities of different layers
In LTE one code word c is mapped onto multiple layers Lc
The CQI feedback can be subband or wideband specific (scheduling)
Slide 16 / 89 Codebook based Single-User MIMO Feedback for LTE
LTE’s Signal Processing Chain (2)
Coding
Interleaving
Segmentation
AMC scheme selection
Symbol
constellation
mapping
Spatial
layer
mapping
MIMO preprocessing
Spatial
precoding (Ws)
User data Code words c Symbols Layers (L)
Other user data
Receiver
Equalization
Demapping
DecodingTransmit
signal
Receive
signal
H
Nr
Feedback
calculation
Channel quality
indicator (CQI)
Precoding matrix
indicator (PMI)
Rank indicator (RI)Resource
element
mapping
Transmit signal generation
IFFT
CP insertion
IFFT
CP insertion
System
overhead
insertion
System
overhead
insertion
Nt
Resource
element
mapping
(Nt)
MIMO preprocessing
Exploit the potential MIMO gains (beamforming, diversity, spatial multiplexing)
Feedback reduction: precoders are confined to a code book F (L) ⊂ CNt×L
The receiver selects and feeds back:
The preferred number of layers L (rank) - RI (OLSM, CLSM)
The best precoder from the code book - PMI (CLSM)
The same rank is used on all REs (L ≤ Lmax = rank(H))
Precoder feedback can be subband or wideband specific
Slide 17 / 89 Codebook based Single-User MIMO Feedback for LTE
LTE’s Signal Processing Chain (3)
Coding
Interleaving
Segmentation
AMC scheme selection
Symbol
constellation
mapping
Spatial
layer
mapping
MIMO preprocessing
Spatial
precoding (Ws)
User data Code words c Symbols Layers (L)
Other user data
Receiver
Equalization
Demapping
DecodingTransmit
signal
Receive
signal
H
Nr
Feedback
calculation
Channel quality
indicator (CQI)
Precoding matrix
indicator (PMI)
Rank indicator (RI)Resource
element
mapping
Transmit signal generation
IFFT
CP insertion
IFFT
CP insertion
System
overhead
insertion
System
overhead
insertion
Nt
Resource
element
mapping
(Nt)
Transmit signal generation
OFDM: the bandwidth is divided into K orthogonal subcarriers
The CP orthogonalizes OFDM symbols (in time)
Temporal frame structure: per frame R REs r are available
Consecutive REs are grouped into S ≤ R subbands
The mapping ρ maps an RE r to the corresponding subband s: ρ(r)→ s
The subband size determines the feedback granularity
Slide 18 / 89 Codebook based Single-User MIMO Feedback for LTE
Resource Elements and Subbands
Resourceelement r
Resources R
Subband s
Layers L
OFDM symbols
OFD
M s
ubca
rrie
rs
Set of REsRs
Rs . . . set of REs corresponding to subband s
Slide 19 / 89 Codebook based Single-User MIMO Feedback for LTE
Closed-Loop Precoding in LTE/LTE-A
Codebook
index
Number of layers `
1 2
0
1
2
3
1√2
11
1√2
1-1
1√2
1j
1√2
1-j
12
1 11 -1
12
1 1j -j
CLSM supports CQI, RI and PMI feedback
Codebook construction for Nt = 2 [3GPP, 2009]:
DFT matrices
Nested + constant modulus (QPSK)
Codebook construction for Nt = 4 [3GPP, 2009]:
Obtained from Householder matrices: Fi = I− 2 ui uHi , i ∈ {1, . . . , 16}
Nested + constant modulus (QPSK, some elements from 8-PSK)
Slide 20 / 89 Codebook based Single-User MIMO Feedback for LTE
Closed-Loop Precoding in LTE/LTE-A (2)
0 45 90 135 1800
2
4
6
8
Steering Angle [°]
An
ten
na
ga
in
[i1
,i2
] = [0,0]
[i1
,i2
] = [4,0]
[i1
,i2
] = [7,0]
[i1
,i2
] = [11,0]
0 45 90 135 1800
2
4
6
8
Steering Angle [°]
An
ten
na
ga
in
[i1
,i2
] = [0,0]
[i1
,i2
] = [0,1]
[i1
,i2
] = [0,2]
[i1
,i2
] = [0,3]
Wideband and Subband Beamformers of LTE-A assuming a ULA with Nt = 8
Nt = 8: product of subband and wideband precoder [3GPP, 2010]
Fs = F(1)F(2)s , (15)
F(1) ∈ F (1)L ⊂ CNt×L, F(2) ∈ F (2)
L ⊂ CNt×L (16)
Same notation for Nt ≤ 4: simply set F (1)L = I
Slide 21 / 89 Codebook based Single-User MIMO Feedback for LTE
Closed-Loop Precoding in LTE/LTE-A (2)
Rank Number of Number of Total number ofwideband precoders subband precoders precoder combinations
1 16 16 2562 16 16 2563 4 16 644 4 8 325 4 1 46 4 1 47 4 1 48 1 1 1
Nt = 8: product of subband and wideband precoder [3GPP, 2010]
Fs = F(1)F(2)s , (15)
F(1) ∈ F (1)L ⊂ CNt×L, F(2) ∈ F (2)
L ⊂ CNt×L (16)
Same notation for Nt ≤ 4: simply set F (1)L = I
Slide 21 / 89 Codebook based Single-User MIMO Feedback for LTE
Input-Output Relationship and Post-equalization SINR
Received signal vector yr ∈ CNr×1 on RE r
yr = Hr Fssr + zr , r ∈ {1, . . . ,R}, s = ρ(r)
Estimated symbol-vector sr ∈ CL×1 with linear equalizer
sr = Gr yr = Gr Hr Fs︸ ︷︷ ︸Kr
sr + Gr zr
E.g., MMSE receiver
Gr =(
(Hr Fs)HHr Fs + σ2z I)−1
(Hr Fs)H (17)
Post-equalization SINR of layer `
SINRr,` (Fs) =P`|Kr [`, `]|2∑
i 6=` Pi |Kr [`, i]|2 + σ2z |Gr [`, i]|2
(18)
Slide 22 / 89 Codebook based Single-User MIMO Feedback for LTE
Input-Output Relationship and Post-equalization SINR
Received signal vector yr ∈ CNr×1 on RE r
yr = Hr Fssr + zr , r ∈ {1, . . . ,R}, s = ρ(r)
Estimated symbol-vector sr ∈ CL×1 with linear equalizer
sr = Gr yr = Gr Hr Fs︸ ︷︷ ︸Kr
sr + Gr zr
E.g., MMSE receiver
Gr =(
(Hr Fs)HHr Fs + σ2z I)−1
(Hr Fs)H (17)
Post-equalization SINR of layer `
SINRr,` (Fs) =P`|Kr [`, `]|2∑
i 6=` Pi |Kr [`, i]|2 + σ2z |Gr [`, i]|2
(18)
Slide 22 / 89 Codebook based Single-User MIMO Feedback for LTE
Input-Output Relationship and Post-equalization SINR
Received signal vector yr ∈ CNr×1 on RE r
yr = Hr Fssr + zr , r ∈ {1, . . . ,R}, s = ρ(r)
Estimated symbol-vector sr ∈ CL×1 with linear equalizer
sr = Gr yr = Gr Hr Fs︸ ︷︷ ︸Kr
sr + Gr zr
E.g., MMSE receiver
Gr =(
(Hr Fs)HHr Fs + σ2z I)−1
(Hr Fs)H (17)
Post-equalization SINR of layer `
SINRr,` (Fs) =P`|Kr [`, `]|2∑
i 6=` Pi |Kr [`, i]|2 + σ2z |Gr [`, i]|2
(18)
Slide 22 / 89 Codebook based Single-User MIMO Feedback for LTE
Selection of Feedback Indicators [Schwarz and Rupp, 2011]
ESM SINR
averaging fmSINRr,l (Ws)
SNRs (Ws,m)c
AWGN SNR
to BLER
mapping gm Ps (Ws,m)c
Effective SNR averaging for BLER estimation
RI and PMI : maximize estimated user throughput
CQI : largest rate that ensures operation below target BLER P(t)b (common
practice)⇒ estimation of BLER for each MCS required
Static SISO-AWGN channel: BLER versus SNR is known (simulations)⇒ SNR estimation is sufficient
Fading environment: SINR changes over REs r and layers `
Linear averaging of SNRs: diversity order of the channel not represented
More accurate: effective SNR mapping (ESM) [Tsai and Soong, 2003]⇒ Mapping of SINRs onto equivalent AWGN-SNR
Slide 23 / 89 Codebook based Single-User MIMO Feedback for LTE
Selection of Feedback Indicators [Schwarz and Rupp, 2011]
ESM SINR
averaging fmSINRr,l (Ws)
SNRs (Ws,m)c
AWGN SNR
to BLER
mapping gm Ps (Ws,m)c
Effective SNR averaging for BLER estimation
RI and PMI : maximize estimated user throughput
CQI : largest rate that ensures operation below target BLER P(t)b (common
practice)⇒ estimation of BLER for each MCS required
Static SISO-AWGN channel: BLER versus SNR is known (simulations)⇒ SNR estimation is sufficient
Fading environment: SINR changes over REs r and layers `
Linear averaging of SNRs: diversity order of the channel not represented
More accurate: effective SNR mapping (ESM) [Tsai and Soong, 2003]⇒ Mapping of SINRs onto equivalent AWGN-SNR
Slide 23 / 89 Codebook based Single-User MIMO Feedback for LTE
Selection of Feedback Indicators [Schwarz and Rupp, 2011]
ESM SINR
averaging fmSINRr,l (Ws)
SNRs (Ws,m)c
AWGN SNR
to BLER
mapping gm Ps (Ws,m)c
Effective SNR averaging for BLER estimation
RI and PMI : maximize estimated user throughput
CQI : largest rate that ensures operation below target BLER P(t)b (common
practice)⇒ estimation of BLER for each MCS required
Static SISO-AWGN channel: BLER versus SNR is known (simulations)⇒ SNR estimation is sufficient
Fading environment: SINR changes over REs r and layers `
Linear averaging of SNRs: diversity order of the channel not represented
More accurate: effective SNR mapping (ESM) [Tsai and Soong, 2003]⇒ Mapping of SINRs onto equivalent AWGN-SNR
Slide 23 / 89 Codebook based Single-User MIMO Feedback for LTE
Effective SNR Mapping
Consider the SINRs of the REs corresponding to subband s and of the layersbelonging to code word c
SINRr,` (Fs) , r ∈ Rs, ` ∈ Lc (19)
To estimate the BLER of code word c using MCS m, we calculate an effectiveAWGN equivalent SNR
SNRcs (Fs,m) = f−1
m
1|Rs| |Lc |
∑r∈Rs,l∈Lc
fm(SINRr,l (Fs)
) (20)
fm(SNR). . . SINR averaging function of MCS m
Mutual information effective SNR mapping (MIESM): fm(SINR) is the(calibrated) BICM capacity of the corresponding modulation order(4/16/64 QAM) [Wan et al., 2006]
Exponential effective SNR mapping (EESM): fm(SINR) is an exponentialfunction [Sandanalakshmi et al., 2007]
Slide 24 / 89 Codebook based Single-User MIMO Feedback for LTE
Effective SNR Mapping
Consider the SINRs of the REs corresponding to subband s and of the layersbelonging to code word c
SINRr,` (Fs) , r ∈ Rs, ` ∈ Lc (19)
To estimate the BLER of code word c using MCS m, we calculate an effectiveAWGN equivalent SNR
SNRcs (Fs,m) = f−1
m
1|Rs| |Lc |
∑r∈Rs,l∈Lc
fm(SINRr,l (Fs)
) (20)
fm(SNR). . . SINR averaging function of MCS m
Mutual information effective SNR mapping (MIESM): fm(SINR) is the(calibrated) BICM capacity of the corresponding modulation order(4/16/64 QAM) [Wan et al., 2006]
Exponential effective SNR mapping (EESM): fm(SINR) is an exponentialfunction [Sandanalakshmi et al., 2007]
Slide 24 / 89 Codebook based Single-User MIMO Feedback for LTE
Effective SNR Mapping
Consider the SINRs of the REs corresponding to subband s and of the layersbelonging to code word c
SINRr,` (Fs) , r ∈ Rs, ` ∈ Lc (19)
To estimate the BLER of code word c using MCS m, we calculate an effectiveAWGN equivalent SNR
SNRcs (Fs,m) = f−1
m
1|Rs| |Lc |
∑r∈Rs,l∈Lc
fm(SINRr,l (Fs)
) (20)
fm(SNR). . . SINR averaging function of MCS m
Mutual information effective SNR mapping (MIESM): fm(SINR) is the(calibrated) BICM capacity of the corresponding modulation order(4/16/64 QAM) [Wan et al., 2006]
Exponential effective SNR mapping (EESM): fm(SINR) is an exponentialfunction [Sandanalakshmi et al., 2007]
Slide 24 / 89 Codebook based Single-User MIMO Feedback for LTE
BICM Capacity – MIESM Averaging Functions
channelcoding
modulationmapping
coded bits
bitinterleaving
modulated symbolsdata bits interleaved bits
BICM architecture
BICM architecture as applied by LTE
BICM capacity mapping functions fm(SINR) = Bm
(SINRβm
)
Slide 25 / 89 Codebook based Single-User MIMO Feedback for LTE
BICM Capacity – MIESM Averaging Functions
242220181614121086420-2-4-6
8
7
6
5
4
3
2
1
0
SNR [dB]
Spec
tral
eff
icie
ncy
[bit/
s/H
z]
SISO_AWGN
Shannon capacity
BICM 64 capacity
BICM 16 capacity
BICM 4 capacity
BICM capacity curves versus AWGN Shannon capacity
BICM architecture as applied by LTE
BICM capacity mapping functions fm(SINR) = Bm
(SINRβm
)
Slide 25 / 89 Codebook based Single-User MIMO Feedback for LTE
ESM – Calibration Sensitivity
2.52.2521.751.51.2510.750.5
100
10-1
10-2
10-3
Calibration parameter
Wei
ghte
d M
SE
SISO_1.4MHz_CQI695% confidence interval95% confidence interval
MIESM
EESM
MSE of the estimated SNR for MIESM and EESM in dependence of the calibration parameter
fm(SINR) = Bm
(SINRβm
), βm . . . calibration parameter (21)
Calibration according to [He et al., 2007, Cipriano et al., 2008]
Slide 26 / 89 Codebook based Single-User MIMO Feedback for LTE
MIESM – SNR Averaging Performance
2520151050-5-10SNR [dB]
Blo
ck e
rror
ratio
SISO_1.4MHz_AWGN_TU100
10-1
10-2
10-3
SISO AWGN BLERMIESM estimation
Comparison of the MIESM abstraction for a 1.4 MHz typical urban channel to the average BLERs of LTE’s 15 MCSsachieved over a 1.4 MHz AWGN channel.
With the AWGN equivalent SNR, the BLER of MCS m is estimated as
Pcs (Fs,m) = gm
(SNRc
s (Fs,m))
(22)
gm(SNR). . . AWGN SNR-BLER lookup table
Slide 27 / 89 Codebook based Single-User MIMO Feedback for LTE
Spectral Efficiency Estimation
ESM SINR
averaging fmSINRr,l (Ws)
SNRs (Ws,m)c
AWGN SNR
to BLER
mapping gm
Spectral
e!ciency
estimationPs (Ws,m)c
Es (Ws,m)c
Calculation steps required for spectral efficiency estimation
Define the following function
hm(P) =
{em, P ≤ P(t)
b
0, P > P(t)b
em. . . spectral efficiency of MCS m
Estimated spectral efficiency achieved with MCS m
Ecs (Fs,m) = hm(Pc
s (Fs,m)) · (1− Pcs (Fs,m)) (23)
Slide 28 / 89 Codebook based Single-User MIMO Feedback for LTE
Feedback Indicator Selection Optimization Problem
[L,{
Fs
}S, {ms}S
]= argmax
L,F(1),F(2)s ,ms
S∑s=1
CL∑c=1
Ecs (Fs,ms[c])
subject to: L ≤ Lmax
F(1) ∈ F (L)1
F(2)s ∈ F (L)
2
ms ∈MCL×1
CL. . . number of code words employed when L layers are transmittedS . . . number of feedback subsets
Non-linear combinatorial optimization problem⇒ exhaustive search
Practically infeasible within 1 ms subframe duration
Slide 29 / 89 Codebook based Single-User MIMO Feedback for LTE
Approximate Sequential Solution
1 Preselect rank and precoders from theoretical spectral efficiency (mutualinformation, BICM capacity) without considering the BLER
2 Given precoders and rank, select the highest MCS achieving the BLER target
Estimated spectral efficiency of RE r for L layers and precoder Fs
Ir (Fs, L) =
CL∑c=1
∑`∈Lc
f(SINRr,` (Fs)
), s = ρ(r)
f (SINR). . . spectral efficiency function (mutual information, BICM capacity)
Optimal subband precoder for fixed rank L and wideband precoder F(1)
F(2)s
(F(1), L
)= argmax
F(2)s ∈F
(L)2
Is (Fs, L) , (24)
Is (Fs, L) =∑
r∈Rs
Ir (Fs, L) , Fs = F(1)F(2)s (25)
Slide 30 / 89 Codebook based Single-User MIMO Feedback for LTE
Approximate Sequential Solution
1 Preselect rank and precoders from theoretical spectral efficiency (mutualinformation, BICM capacity) without considering the BLER
2 Given precoders and rank, select the highest MCS achieving the BLER target
Estimated spectral efficiency of RE r for L layers and precoder Fs
Ir (Fs, L) =
CL∑c=1
∑`∈Lc
f(SINRr,` (Fs)
), s = ρ(r)
f (SINR). . . spectral efficiency function (mutual information, BICM capacity)
Optimal subband precoder for fixed rank L and wideband precoder F(1)
F(2)s
(F(1), L
)= argmax
F(2)s ∈F
(L)2
Is (Fs, L) , (24)
Is (Fs, L) =∑
r∈Rs
Ir (Fs, L) , Fs = F(1)F(2)s (25)
Slide 30 / 89 Codebook based Single-User MIMO Feedback for LTE
Approximate Sequential Solution
1 Preselect rank and precoders from theoretical spectral efficiency (mutualinformation, BICM capacity) without considering the BLER
2 Given precoders and rank, select the highest MCS achieving the BLER target
Estimated spectral efficiency of RE r for L layers and precoder Fs
Ir (Fs, L) =
CL∑c=1
∑`∈Lc
f(SINRr,` (Fs)
), s = ρ(r)
f (SINR). . . spectral efficiency function (mutual information, BICM capacity)
Optimal subband precoder for fixed rank L and wideband precoder F(1)
F(2)s
(F(1), L
)= argmax
F(2)s ∈F
(L)2
Is (Fs, L) , (24)
Is (Fs, L) =∑
r∈Rs
Ir (Fs, L) , Fs = F(1)F(2)s (25)
Slide 30 / 89 Codebook based Single-User MIMO Feedback for LTE
Approximate Sequential Solution (2)
Optimal wideband precoder for fixed rank L
F(1)(L) = argmax
F(1)∈F(L)1
I (Fs, L) , (26)
I (Fs, L) =S∑
s=1
Is (Fs, L) , Fs = F(1)F(2)s
(F(1), L
)
Optimal rank
L = argmaxL≤Lmax
I(
Fs(L), L), Fs(L) = F(1)
(L) F(2)s
(F(1)
(L), L)
Further simplification: replace f (SINR) with the pre-equalization efficiency
I(Fs) = log2 det(
I +1σ2
zHr FsFH
s HHr
)(27)
Performs only well with maximum likelihood (ML) detection
Slide 31 / 89 Codebook based Single-User MIMO Feedback for LTE
Approximate Sequential Solution (2)
Optimal wideband precoder for fixed rank L
F(1)(L) = argmax
F(1)∈F(L)1
I (Fs, L) , (26)
I (Fs, L) =S∑
s=1
Is (Fs, L) , Fs = F(1)F(2)s
(F(1), L
)
Optimal rank
L = argmaxL≤Lmax
I(
Fs(L), L), Fs(L) = F(1)
(L) F(2)s
(F(1)
(L), L)
Further simplification: replace f (SINR) with the pre-equalization efficiency
I(Fs) = log2 det(
I +1σ2
zHr FsFH
s HHr
)(27)
Performs only well with maximum likelihood (ML) detection
Slide 31 / 89 Codebook based Single-User MIMO Feedback for LTE
Approximate Sequential Solution (2)
Optimal wideband precoder for fixed rank L
F(1)(L) = argmax
F(1)∈F(L)1
I (Fs, L) , (26)
I (Fs, L) =S∑
s=1
Is (Fs, L) , Fs = F(1)F(2)s
(F(1), L
)
Optimal rank
L = argmaxL≤Lmax
I(
Fs(L), L), Fs(L) = F(1)
(L) F(2)s
(F(1)
(L), L)
Further simplification: replace f (SINR) with the pre-equalization efficiency
I(Fs) = log2 det(
I +1σ2
zHr FsFH
s HHr
)(27)
Performs only well with maximum likelihood (ML) detection
Slide 31 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rate Adaptation
302520151050-5
5
4
3
2
1
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh
CQI 1
CQI 15
Evaluation of CQI feedback methods for SISO transmission over 1.4 MHz bandwidth
Performance of LTE’s 15 MCSs without rate adaptation
Rate adaptation based on instantaneous SINR without feedback delay
Rate adaptation based on instantaneous SINR with outdated feedback
Rate adaptation based on long term average SINR
Slide 32 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rate Adaptation
302520151050-5
5
4
3
2
1
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh
instantaneous CQI, delay 0
CQI 1
CQI 15
Evaluation of CQI feedback methods for SISO transmission over 1.4 MHz bandwidth
Performance of LTE’s 15 MCSs without rate adaptation
Rate adaptation based on instantaneous SINR without feedback delay
Rate adaptation based on instantaneous SINR with outdated feedback
Rate adaptation based on long term average SINR
Slide 32 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rate Adaptation
302520151050-5
5
4
3
2
1
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh
instantaneous CQI, delay 0
instantaneous CQI, delay 1
CQI 1
CQI 15
Evaluation of CQI feedback methods for SISO transmission over 1.4 MHz bandwidth
Performance of LTE’s 15 MCSs without rate adaptation
Rate adaptation based on instantaneous SINR without feedback delay
Rate adaptation based on instantaneous SINR with outdated feedback
Rate adaptation based on long term average SINR
Slide 32 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rate Adaptation
302520151050-5
5
4
3
2
1
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh
instantaneous CQI, delay 0
long-term CQI, delay 1
instantaneous CQI, delay 1
CQI 1
CQI 15
Evaluation of CQI feedback methods for SISO transmission over 1.4 MHz bandwidth
Performance of LTE’s 15 MCSs without rate adaptation
Rate adaptation based on instantaneous SINR without feedback delay
Rate adaptation based on instantaneous SINR with outdated feedback
Rate adaptation based on long term average SINR
Slide 32 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rate Adaptation (2)
10.10.01
4
3.75
3.5
3.25
3
2.75
2.5
2.25
2
51551.55.15
Normalized Doppler frequency
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh_20dB
User velocity [km/h] @ 2.1 GHz carrier frequency
instantaneous CQI
Impact of feedback delay on CQI adaptation methods
Rate adaptation based on instantaneous SINR
νd = fd Ts = fcvc0
Ts (28)
Rate adaptation based on instantaneous SINR with linear extrapolation
Rate adaptation based on instantaneous SINR with RLS prediction
Rate adaptation based on long term average SINR
Slide 33 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rate Adaptation (2)
10.10.01
4
3.75
3.5
3.25
3
2.75
2.5
2.25
2
51551.55.15
Normalized Doppler frequency
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh_20dB
User velocity [km/h] @ 2.1 GHz carrier frequency
instantaneous CQI
instantaneous CQIlinear extrapolation
Impact of feedback delay on CQI adaptation methods
Rate adaptation based on instantaneous SINR
νd = fd Ts = fcvc0
Ts (28)
Rate adaptation based on instantaneous SINR with linear extrapolation
Rate adaptation based on instantaneous SINR with RLS prediction
Rate adaptation based on long term average SINR
Slide 33 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rate Adaptation (2)
10.10.01
4
3.75
3.5
3.25
3
2.75
2.5
2.25
2
51551.55.15
Normalized Doppler frequency
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh_20dB
User velocity [km/h] @ 2.1 GHz carrier frequency
instantaneous CQI
instantaneous CQIlinear extrapolation
instantaneous CQIRLS prediction
Impact of feedback delay on CQI adaptation methods
Rate adaptation based on instantaneous SINR
νd = fd Ts = fcvc0
Ts (28)
Rate adaptation based on instantaneous SINR with linear extrapolation
Rate adaptation based on instantaneous SINR with RLS prediction
Rate adaptation based on long term average SINR
Slide 33 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rate Adaptation (2)
10.10.01
4
3.75
3.5
3.25
3
2.75
2.5
2.25
2
51551.55.15
Normalized Doppler frequency
Thr
ough
put [
Mbi
t/s]
SISO_1.4MHz_flatRayleigh_20dB
User velocity [km/h] @ 2.1 GHz carrier frequency
instantaneous CQI
instantaneous CQIlinear extrapolation
instantaneous CQIRLS prediction
long-term CQIadaptive averaging window size
Impact of feedback delay on CQI adaptation methods
Rate adaptation based on instantaneous SINR
νd = fd Ts = fcvc0
Ts (28)
Rate adaptation based on instantaneous SINR with linear extrapolation
Rate adaptation based on instantaneous SINR with RLS prediction
Rate adaptation based on long term average SINR
Slide 33 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rank Adaptation
4035302520151050-5-10
18
16
14
12
10
8
6
4
2
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
4x4_1.4MHz_VehA_corr0
rank 1
rank 2
rank 3rank 4
rank adaptive
4035302520151050-5-10
16
14
12
10
8
6
4
2
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
4x4_1.4MHz_VehA_corr0.9
rank adaptive
rank 4
rank 3
rank 2
rank 1
Comparison of fixed rank and rank adaptive transmission over 1.4MHz system bandwidth
Rank adaptation versus fixed rank transmission for spatially uncorrelated channel
Rank adaptation versus fixed rank transmission for spatially correlated channel
E(
vec (H) vec (H)H)
=
1 α
1/9corr α
4/9corr αcorr
α1/9corr 1 α
1/9corr α
4/9corr
α4/9corr α
1/9corr 1 α
1/9corr
αcorr α4/9corr α
1/9corr 1
⊗ INt , αcorr = 0.9
(29)
Slide 34 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Rank Adaptation
4035302520151050-5-10
18
16
14
12
10
8
6
4
2
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
4x4_1.4MHz_VehA_corr0
rank 1
rank 2
rank 3rank 4
rank adaptive
4035302520151050-5-10
16
14
12
10
8
6
4
2
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
4x4_1.4MHz_VehA_corr0.9
rank adaptive
rank 4
rank 3
rank 2
rank 1
Comparison of fixed rank and rank adaptive transmission over 1.4MHz system bandwidth
Rank adaptation versus fixed rank transmission for spatially uncorrelated channel
Rank adaptation versus fixed rank transmission for spatially correlated channel
E(
vec (H) vec (H)H)
=
1 α
1/9corr α
4/9corr αcorr
α1/9corr 1 α
1/9corr α
4/9corr
α4/9corr α
1/9corr 1 α
1/9corr
αcorr α4/9corr α
1/9corr 1
⊗ INt , αcorr = 0.9
(29)
Slide 34 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Subband Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_TU_8bit
MRTMRT
subband size = 600 subband size = 600
Impact of the subband size with Nt × Nr = 8× 1, 10 MHz bandwidth, 400 kHz coherence bandwidth and precodercodebook size of 8 bit
Performance with a single feedback subband (600 subcarriers) compared tomaximum ratio transmission (MRT) on each subcarrier
Performance with a subband size of 300 subcarriers
Performance with a subband size of 120 subcarriers
Performance with a subband size of 60 subcarriers
Performance with a subband size of 12 subcarriers (one RB = 180 kHz)
Slide 35 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Subband Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_TU_8bit
MRTMRT
subband size = 600 subband size = 600
subband size = 300 subband size = 300
Impact of the subband size with Nt × Nr = 8× 1, 10 MHz bandwidth, 400 kHz coherence bandwidth and precodercodebook size of 8 bit
Performance with a single feedback subband (600 subcarriers) compared tomaximum ratio transmission (MRT) on each subcarrier
Performance with a subband size of 300 subcarriers
Performance with a subband size of 120 subcarriers
Performance with a subband size of 60 subcarriers
Performance with a subband size of 12 subcarriers (one RB = 180 kHz)
Slide 35 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Subband Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_TU_8bit
MRTMRT
subband size = 600 subband size = 600
subband size = 300 subband size = 300
subband size = 120 subband size = 120
Impact of the subband size with Nt × Nr = 8× 1, 10 MHz bandwidth, 400 kHz coherence bandwidth and precodercodebook size of 8 bit
Performance with a single feedback subband (600 subcarriers) compared tomaximum ratio transmission (MRT) on each subcarrier
Performance with a subband size of 300 subcarriers
Performance with a subband size of 120 subcarriers
Performance with a subband size of 60 subcarriers
Performance with a subband size of 12 subcarriers (one RB = 180 kHz)
Slide 35 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Subband Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_TU_8bit
MRTMRT
subband size = 600 subband size = 600
subband size = 300 subband size = 300
subband size = 120 subband size = 120 subband size = 60 subband size = 60
Impact of the subband size with Nt × Nr = 8× 1, 10 MHz bandwidth, 400 kHz coherence bandwidth and precodercodebook size of 8 bit
Performance with a single feedback subband (600 subcarriers) compared tomaximum ratio transmission (MRT) on each subcarrier
Performance with a subband size of 300 subcarriers
Performance with a subband size of 120 subcarriers
Performance with a subband size of 60 subcarriers
Performance with a subband size of 12 subcarriers (one RB = 180 kHz)
Slide 35 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Subband Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_TU_8bit
MRTMRT
subband size = 600 subband size = 600
subband size = 300 subband size = 300
subband size = 120 subband size = 120 subband size = 60 subband size = 60
subband size = 12 subband size = 12
Impact of the subband size with Nt × Nr = 8× 1, 10 MHz bandwidth, 400 kHz coherence bandwidth and precodercodebook size of 8 bit
Performance with a single feedback subband (600 subcarriers) compared tomaximum ratio transmission (MRT) on each subcarrier
Performance with a subband size of 300 subcarriers
Performance with a subband size of 120 subcarriers
Performance with a subband size of 60 subcarriers
Performance with a subband size of 12 subcarriers (one RB = 180 kHz)
Slide 35 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Codebook Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_VehA_50clusters
MRTMRT
1 bit1 bit
~9 dB~9 dB
Impact of the codebook size with Nt × Nr = 8× 1, 10 MHz bandwidth, 550 kHz coherence bandwidth and subbandsize of 180 kHz
Performance with a codebook size of 1 bit compared to maximum ratiotransmission (MRT) on each subcarrier
Performance with a codebook size of 2 bit
Performance with a codebook size of 4 bit
Performance with a codebook size of 8 bit (LTE codebook)
Performance with a codebook size of 16 bit
Slide 36 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Codebook Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_VehA_50clusters
MRTMRT
2 bit2 bit
1 bit1 bit
~9 dB~9 dB
Impact of the codebook size with Nt × Nr = 8× 1, 10 MHz bandwidth, 550 kHz coherence bandwidth and subbandsize of 180 kHz
Performance with a codebook size of 1 bit compared to maximum ratiotransmission (MRT) on each subcarrier
Performance with a codebook size of 2 bit
Performance with a codebook size of 4 bit
Performance with a codebook size of 8 bit (LTE codebook)
Performance with a codebook size of 16 bit
Slide 36 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Codebook Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_VehA_50clusters
MRTMRT
4 bit4 bit
2 bit2 bit
1 bit1 bit
~9 dB~9 dB
Impact of the codebook size with Nt × Nr = 8× 1, 10 MHz bandwidth, 550 kHz coherence bandwidth and subbandsize of 180 kHz
Performance with a codebook size of 1 bit compared to maximum ratiotransmission (MRT) on each subcarrier
Performance with a codebook size of 2 bit
Performance with a codebook size of 4 bit
Performance with a codebook size of 8 bit (LTE codebook)
Performance with a codebook size of 16 bit
Slide 36 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Codebook Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_VehA_50clusters
MRTMRT
8 bit8 bit4 bit4 bit
2 bit2 bit
1 bit1 bit
~9 dB~9 dB
Impact of the codebook size with Nt × Nr = 8× 1, 10 MHz bandwidth, 550 kHz coherence bandwidth and subbandsize of 180 kHz
Performance with a codebook size of 1 bit compared to maximum ratiotransmission (MRT) on each subcarrier
Performance with a codebook size of 2 bit
Performance with a codebook size of 4 bit
Performance with a codebook size of 8 bit (LTE codebook)
Performance with a codebook size of 16 bit
Slide 36 / 89 Codebook based Single-User MIMO Feedback for LTE
Performance Investigation – Precoder Codebook Size
20151050-5-10
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
8x1_10MHz_VehA_50clusters
MRTMRT
16 bit16 bit
8 bit8 bit4 bit4 bit
2 bit2 bit
1 bit1 bit
~9 dB~9 dB
Impact of the codebook size with Nt × Nr = 8× 1, 10 MHz bandwidth, 550 kHz coherence bandwidth and subbandsize of 180 kHz
Performance with a codebook size of 1 bit compared to maximum ratiotransmission (MRT) on each subcarrier
Performance with a codebook size of 2 bit
Performance with a codebook size of 4 bit
Performance with a codebook size of 8 bit (LTE codebook)
Performance with a codebook size of 16 bit
Slide 36 / 89 Codebook based Single-User MIMO Feedback for LTE
Contents
1 Motivation
2 Codebook based Single-User MIMO Feedback for LTE
3 Multi-User MIMO Feedback for LTE
4 Multi-User MIMO with Channel Subspace Feedback
5 Multi-User MIMO with Channel Gramian Feedback
6 Feedback Overhead Reduction through Excess Antennas
Slide 37 / 89 Multi-User MIMO Feedback for LTE
Downlink Multi-User MIMO in LTE
yu = HHu fuxu + HH
u
U∑j=1j 6=u
fj xj + zu (30)
Basic multi-user MIMO already supported in Rel. 8 (transmission mode 5)
Restricted to two transmit antennas and single stream per user
Codebook based precoding using the CLSM codebook
In general, large residual multi-user interference
⇒ per user unitary rate control (PU2RC)
Extended multi-user MIMO support > Rel. 9 (modes 8, 9)
Non-codebook based precoding
Enables more sophisticated transceivers
Performance restricted by low accuracy of CSIT⇒ interference
⇒ zero forcing (ZF) beamforming
Slide 38 / 89 Multi-User MIMO Feedback for LTE
Downlink Multi-User MIMO in LTE
yu = HHu fuxu + HH
u
U∑j=1j 6=u
fj xj + zu (30)
Basic multi-user MIMO already supported in Rel. 8 (transmission mode 5)
Restricted to two transmit antennas and single stream per user
Codebook based precoding using the CLSM codebook
In general, large residual multi-user interference
⇒ per user unitary rate control (PU2RC)
Extended multi-user MIMO support > Rel. 9 (modes 8, 9)
Non-codebook based precoding
Enables more sophisticated transceivers
Performance restricted by low accuracy of CSIT⇒ interference
⇒ zero forcing (ZF) beamforming
Slide 38 / 89 Multi-User MIMO Feedback for LTE
Per User Unitary Rate Control
Precoders are restricted to be selected from a predefined codebook
Codebook for transmission of L streams from Nt antennas
Q(Nt )L =
{Fi ∈ CNt×L∣∣FH
i Fi = IL, i ∈ {1, . . . , np}}⊂ G (Nt , L) , (31)
yu = HHu
√PL
Fx + zu = HHu [f1, . . . , fL] x + zu (32)
Each column is assigned to serve a different user
E.g., column ν is assigned to user u
yu = HHu
√PL
fνxu + HHu
√PL
L∑µ=1µ6=ν
fµxµ + zu , (33)
Slide 39 / 89 Multi-User MIMO Feedback for LTE
Per User Unitary Rate Control
Precoders are restricted to be selected from a predefined codebook
Codebook for transmission of L streams from Nt antennas
Q(Nt )L =
{Fi ∈ CNt×L∣∣FH
i Fi = IL, i ∈ {1, . . . , np}}⊂ G (Nt , L) , (31)
yu = HHu
√PL
Fx + zu = HHu [f1, . . . , fL] x + zu (32)
Each column is assigned to serve a different user
E.g., column ν is assigned to user u
yu = HHu
√PL
fνxu + HHu
√PL
L∑µ=1µ6=ν
fµxµ + zu , (33)
Slide 39 / 89 Multi-User MIMO Feedback for LTE
Per User Unitary Rate Control (2)
Optimal linear receiver: interference-aware MMSE
gu =
σ2z IMu +
PL
HHu
L∑µ=1µ6=ν
fµfHµHu
−1
PL
HHu fν (34)
Can be calculated because the precoders are restricted to the codebook
Post-equalization SINR
SINRu =PL
∣∣gHu HH
u fν∣∣2
PL∑Lµ=1, µ 6=ν
∣∣gHu HH
u fµ∣∣2 + σ2
z ‖gu‖2(35)
Slide 40 / 89 Multi-User MIMO Feedback for LTE
Per User Unitary Rate Control (2)
Optimal linear receiver: interference-aware MMSE
gu =
σ2z IMu +
PL
HHu
L∑µ=1µ6=ν
fµfHµHu
−1
PL
HHu fν (34)
Can be calculated because the precoders are restricted to the codebook
Post-equalization SINR
SINRu =PL
∣∣gHu HH
u fν∣∣2
PL∑Lµ=1, µ 6=ν
∣∣gHu HH
u fµ∣∣2 + σ2
z ‖gu‖2(35)
Slide 40 / 89 Multi-User MIMO Feedback for LTE
Per User Unitary Rate Control – CSI Feedback
CSI feedback calculation at user u:
For each Fi ∈ Q(Nt )L , find the best column ν in terms of SINR
⇒ np =∣∣∣Q(Ni )
L
∣∣∣ potential beamformers
Out of the np potential beamformers, feedback the n` best performers
Feedback overhead:
n∑`=1
dlog2 (np − (`− 1))e︸ ︷︷ ︸selected precoder
+ log2 (L)︸ ︷︷ ︸selected column
(36)
Feedback the corresponding SINRs as CQI
Slide 41 / 89 Multi-User MIMO Feedback for LTE
Per User Unitary Rate Control – Scheduling
Multi-user scheduling and precoder selection:
Find compatible user sets Sj :
Same Fi ∈ Q(Ni )L selected
Different column fν from Fi chosen
Determine the best compatible user set , e.g., to maximize sum rate
Rj =∑u∈Sj
log2 (1 + SINRu) (37)
Difficulty of PU2RC: selection of feedback parameter n`, np , L
Large np : + find a good precoder; − compatible user sets shrink
Large n`: + find a good compatible user set; − linear increase in overhead
Large L: + large multiplexing gain; − large multi-user interference
Slide 42 / 89 Multi-User MIMO Feedback for LTE
Per User Unitary Rate Control – Scheduling
Multi-user scheduling and precoder selection:
Find compatible user sets Sj :
Same Fi ∈ Q(Ni )L selected
Different column fν from Fi chosen
Determine the best compatible user set , e.g., to maximize sum rate
Rj =∑u∈Sj
log2 (1 + SINRu) (37)
Difficulty of PU2RC: selection of feedback parameter n`, np , L
Large np : + find a good precoder; − compatible user sets shrink
Large n`: + find a good compatible user set; − linear increase in overhead
Large L: + large multiplexing gain; − large multi-user interference
Slide 42 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming
Consider for simplicity Mu = Nr = 1
yu =√
puhHu fuxu + hH
u
∑i∈S,i 6=u
√pi fi xi + zu , (38)
S . . . set of scheduled userspj = P
|S| ‖fj‖2 for equal power allocation
SINR of user u
SINRu =pu |hH
u fu |2
σ2z +
∑i∈S\{u} pi |hH
u fi |2(39)
Problem: cannot be estimated accurately by the users during feedbackcalculation as precoders are not yet known
Slide 43 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming
Consider for simplicity Mu = Nr = 1
yu =√
puhHu fuxu + hH
u
∑i∈S,i 6=u
√pi fi xi + zu , (38)
S . . . set of scheduled userspj = P
|S| ‖fj‖2 for equal power allocation
SINR of user u
SINRu =pu |hH
u fu |2
σ2z +
∑i∈S\{u} pi |hH
u fi |2(39)
Problem: cannot be estimated accurately by the users during feedbackcalculation as precoders are not yet known
Slide 43 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming with Limited Feedback
To calculate the ZF beamformer the transmitter requires the normalizedchannel vectors (see multi-user MIMO lecture)
hu =hu
‖hu‖(40)
Channel vector quantization
hu = q(hu) =⇒ hu = hu hHu hu + (I− hu hH
u )hu = cos θuejϕu hu + eu , (41)
eu = sin θuejψu eu
θu . . . principle angle between span(
hu
)and span
(hu
)SINR in terms of hu
SINRu =pu ‖hu‖2 ‖fu‖2
∣∣∣cos θue−jϕu hHu fu + eH
u fu
∣∣∣2σ2
z + ‖hu‖2 sin θu2∑
i∈S\{u} pi∣∣eH
u fi∣∣2 , (42)
hHu fi = 0 and
∣∣∣hHu fu
∣∣∣2 =1
‖fu‖2 due to ZF onto hu
Slide 44 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming with Limited Feedback
To calculate the ZF beamformer the transmitter requires the normalizedchannel vectors (see multi-user MIMO lecture)
hu =hu
‖hu‖(40)
Channel vector quantization
hu = q(hu) =⇒ hu = hu hHu hu + (I− hu hH
u )hu = cos θuejϕu hu + eu , (41)
eu = sin θuejψu eu
θu . . . principle angle between span(
hu
)and span
(hu
)SINR in terms of hu
SINRu =pu ‖hu‖2 ‖fu‖2
∣∣∣cos θue−jϕu hHu fu + eH
u fu
∣∣∣2σ2
z + ‖hu‖2 sin θu2∑
i∈S\{u} pi∣∣eH
u fi∣∣2 , (42)
hHu fi = 0 and
∣∣∣hHu fu
∣∣∣2 =1
‖fu‖2 due to ZF onto hu
Slide 44 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming with Limited Feedback (2)
SINR lower bound assuming eHu fu ≈ 0
SINRu ≥pu ‖hu‖2 cos θu
2
σ2z + ‖hu‖2 sin θu
2∑i∈S\{u} pi
∣∣eHu fi∣∣2 (43)
Tight when U →∞ because orthogonal users are scheduled: fu ⊥ eu
During feedback calculation fi unknown
⇒ consider the expected value of the SINR and apply Jensen’s inequality
E (SINRu) ≥pu ‖hu‖2 cos θu
2
σ2z + ‖hu‖2 sin θu
2E(∑
i∈S\{u}P
|S|‖fi‖2 ‖fi‖2∣∣∣eH
u fi
∣∣∣2) (44)
Slide 45 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming with Limited Feedback (2)
SINR lower bound assuming eHu fu ≈ 0
SINRu ≥pu ‖hu‖2 cos θu
2
σ2z + ‖hu‖2 sin θu
2∑i∈S\{u} pi
∣∣eHu fi∣∣2 (43)
Tight when U →∞ because orthogonal users are scheduled: fu ⊥ eu
During feedback calculation fi unknown
⇒ consider the expected value of the SINR and apply Jensen’s inequality
E (SINRu) ≥pu ‖hu‖2 cos θu
2
σ2z + ‖hu‖2 sin θu
2E(∑
i∈S\{u}P
|S|‖fi‖2 ‖fi‖2∣∣∣eH
u fi
∣∣∣2) (44)
Slide 45 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming with Limited Feedback (3)
Assuming fi isotropically distributed in the Nt − 1 dimensional null-space of hu
E(∣∣∣eH
u fi
∣∣∣2) =1
Nt − 1, (45)
E (SINRu) ≥pu ‖hu‖2 cos θu
2
σ2z + P
|S||S|−1Nt−1 ‖hu‖2 sin θu
2(46)
⇒ Feedback ‖hu‖2 and cos θu2 as two separate CQIs [Trivellato et al., 2007]
Further simplification: assume |S| = Nt (worst-case interference)
E (SINRu) ≥PNt‖hu‖2 cos θu
2
σ2z + P
Nt‖hu‖2 sin θu
2= CQIu , (47)
ˆSINRu = CQIuNt
|S| ‖fu‖2 (48)
Slide 46 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming with Limited Feedback (3)
Assuming fi isotropically distributed in the Nt − 1 dimensional null-space of hu
E(∣∣∣eH
u fi
∣∣∣2) =1
Nt − 1, (45)
E (SINRu) ≥pu ‖hu‖2 cos θu
2
σ2z + P
|S||S|−1Nt−1 ‖hu‖2 sin θu
2(46)
⇒ Feedback ‖hu‖2 and cos θu2 as two separate CQIs [Trivellato et al., 2007]
Further simplification: assume |S| = Nt (worst-case interference)
E (SINRu) ≥PNt‖hu‖2 cos θu
2
σ2z + P
Nt‖hu‖2 sin θu
2= CQIu , (47)
ˆSINRu = CQIuNt
|S| ‖fu‖2 (48)
Slide 46 / 89 Multi-User MIMO Feedback for LTE
ZF Beamforming – Scheduling
Exhaustive search: estimate the achievable rate of all possible usercombinations
Consider set Sk = {s1, . . . , sK }
1 Calculate the ZF beamformers for Sk (see multi-user MIMO lecture)
2 Estimate the achievable rate
ˆSINRsi = CQIsi
Nt
|Sk |∥∥fsi
∥∥2 , (49)
RSk =∑
si∈Sk
log2
(1 + ˆSINRsi
)(50)
3 Select the set with maximal estimated rate
Suboptimal greedy scheduling: see, e.g., [Trivellato et al., 2007]
Slide 47 / 89 Multi-User MIMO Feedback for LTE
Multi-User MIMO in LTE
302520151050
20
15
10
5
0
SNR [dB]
Thr
ough
put [
Mbi
t/s]
flatRayleigh_1.4MHz_multi-user_spatial_multiplexing
ZF beamformingperfect CSIT
8 × 1
8 × 4PU2RC
PU2RC
ZF beamforming
CLSM
CLSM
ZF beamforming
Comparison of ZF beamforming, PU2RC and LTE’s CLSM single-user mode.
Nt × Nr ∈ {8× 4, 8× 1} with B = 8 bit/TTI of feedback and U = 20 users
Interference-aware MMSE receiver
Subspace selection for ZF feedback [Schwarz and Rupp, 2014a]
Similar performance of ZF beamforming and CLSM based single-user MIMO
Slide 48 / 89 Multi-User MIMO Feedback for LTE
Contents
1 Motivation
2 Codebook based Single-User MIMO Feedback for LTE
3 Multi-User MIMO Feedback for LTE
4 Multi-User MIMO with Channel Subspace Feedback
5 Multi-User MIMO with Channel Gramian Feedback
6 Feedback Overhead Reduction through Excess Antennas
Slide 49 / 89 Multi-User MIMO with Channel Subspace Feedback
Recap: Downlink Multi-User MIMO
Remember the downlink multi-user MIMO input-output relationship
yu = GHu HH
u Fu xu︸ ︷︷ ︸intended signal
+ GHu HH
u
∑s∈Ss 6=u
Fs xs
︸ ︷︷ ︸interference
+ GHu zu︸ ︷︷ ︸
noise
(51)
Channel matrix Hu ∈ CNt×Nr ,
Linear transceivers Gu ∈ CNr×L, Fu ∈ CNt×L
Effective channel Heffu = HuGu ∈ CNt×L
We consider the case Mu = Nr , ∀u and Nr ≤ Nt
Slide 50 / 89 Multi-User MIMO with Channel Subspace Feedback
Recap: Transceiver Design
Assume the schedule S to be given
Optimal transceiver: dirty paper coding [Costa, 1983]
Vector-perturbation precoding [Hochwald et al., 2005]
Tomlinson-Harashima precoding [Mezghani et al., 2006]
Disadvantage: complexity
Practically more relevant: linear transceivers
Block-diagonalization precoding [Spencer et al., 2004]
Iterative joint optimization, e.g., based on MMSEcriteria [Shi et al., 2008]
Slide 51 / 89 Multi-User MIMO with Channel Subspace Feedback
Recap: Transceiver Design
Assume the schedule S to be given
Optimal transceiver: dirty paper coding [Costa, 1983]
Vector-perturbation precoding [Hochwald et al., 2005]
Tomlinson-Harashima precoding [Mezghani et al., 2006]
Disadvantage: complexity
Practically more relevant: linear transceivers
Block-diagonalization precoding [Spencer et al., 2004]
Iterative joint optimization, e.g., based on MMSEcriteria [Shi et al., 2008]
Slide 51 / 89 Multi-User MIMO with Channel Subspace Feedback
Considered Transceiver Architecture
Problems of iterative approaches:
Large signalling overhead
Slow convergence
We consider non-iterative linear transceiver designs:
Selfish selection of Gu [Schwarz and Rupp, 2013]
Block-diagonalization precoding at base station
Selection of S based on achievable rate estimate[Schwarz and Rupp, 2014a]
Advantages of this approach:
Reduced computational complexity (closed-form solutions)
Decreased signalling overhead when L < Nr
Heffu = HuGu ∈ CNt×L versus Hu ∈ CNt×Nr (52)
Slide 52 / 89 Multi-User MIMO with Channel Subspace Feedback
Considered Transceiver Architecture
Problems of iterative approaches:
Large signalling overhead
Slow convergence
We consider non-iterative linear transceiver designs:
Selfish selection of Gu [Schwarz and Rupp, 2013]
Block-diagonalization precoding at base station
Selection of S based on achievable rate estimate[Schwarz and Rupp, 2014a]
Advantages of this approach:
Reduced computational complexity (closed-form solutions)
Decreased signalling overhead when L < Nr
Heffu = HuGu ∈ CNt×L versus Hu ∈ CNt×Nr (52)
Slide 52 / 89 Multi-User MIMO with Channel Subspace Feedback
Recap: Block-Diagonalization (BD) Precoding
Assume for now Gu as given and S = {1, . . . ,S}
yu =(Heff
u)H Fu xu +
(Heff
u)H S∑
s=1s 6=u
Fs xs + GHu zu
Goal of BD precoding: eliminate multi-user interference(Heff
s)HFu = 0, ∀s, u ∈ S and s 6= u, (53)
rank((
Heffu)HFu
)= L, ∀u ∈ S (54)
This can be achieved by selecting the precoders as follows ∀u ∈ S
Hu =[Heff
1 , . . . ,Heffu−1,H
effu+1, . . . ,H
effS
]H∈ C(S−1)L×Nt ,
Fu ∈ null(Hu), rank (Fu) = L (55)
Slide 53 / 89 Multi-User MIMO with Channel Subspace Feedback
Recap: Block-Diagonalization (BD) Precoding
Assume for now Gu as given and S = {1, . . . ,S}
yu =(Heff
u)H Fu xu +
(Heff
u)H S∑
s=1s 6=u
Fs xs + GHu zu
Goal of BD precoding: eliminate multi-user interference(Heff
s)HFu = 0, ∀s, u ∈ S and s 6= u, (53)
rank((
Heffu)HFu
)= L, ∀u ∈ S (54)
This can be achieved by selecting the precoders as follows ∀u ∈ S
Hu =[Heff
1 , . . . ,Heffu−1,H
effu+1, . . . ,H
effS
]H∈ C(S−1)L×Nt ,
Fu ∈ null(Hu), rank (Fu) = L (55)
Slide 53 / 89 Multi-User MIMO with Channel Subspace Feedback
Channel Subspace Quantization – Grassmannian Feedback
Notice, Heffj can be replaced with any matrix spanning the same subspace
Heffj ≡ Hj ∈ CNt×L ⇐⇒ span
(Heff
j
)= span
(Hj
), (56)(
Heffj)HFu = 0⇐⇒ HH
j Fu = 0 (57)
⇒ the users have to convey span(
Heffj
)∈ G (Nt , L) to the base station
Precoding is based on channel subspace information
Grassmannian quantization for limited feedback operation
Hj = argminQi∈Q
d2c
(Heff
j ,Qi
)= argmin
Qi∈QL− tr
(HH
j Qi QHi Hj
), (58)
Q ={
Qi ∈ CNt×L ∣∣ QHi Qi = IL, i ∈ {1, . . . , 2B}
}
Slide 54 / 89 Multi-User MIMO with Channel Subspace Feedback
Channel Subspace Quantization – Grassmannian Feedback
Notice, Heffj can be replaced with any matrix spanning the same subspace
Heffj ≡ Hj ∈ CNt×L ⇐⇒ span
(Heff
j
)= span
(Hj
), (56)(
Heffj)HFu = 0⇐⇒ HH
j Fu = 0 (57)
⇒ the users have to convey span(
Heffj
)∈ G (Nt , L) to the base station
Precoding is based on channel subspace information
Grassmannian quantization for limited feedback operation
Hj = argminQi∈Q
d2c
(Heff
j ,Qi
)= argmin
Qi∈QL− tr
(HH
j Qi QHi Hj
), (58)
Q ={
Qi ∈ CNt×L ∣∣ QHi Qi = IL, i ∈ {1, . . . , 2B}
}
Slide 54 / 89 Multi-User MIMO with Channel Subspace Feedback
Why Chordal Distance Quantization?
Achievable rate with perfect CSIT
RBD = E log2
∣∣∣IL + ρ(
Heffu
)HFuFH
u Heffu
∣∣∣ , ρ =P
σ2z S L
. (59)
assuming Gu to be semi-unitary GHu Gu = IL
and equal power allocation FHu Fu = IL
Achievable rate with limited feedback
RBD-Quant = E log2
∣∣∣∣∣∣IL + ρS∑`=1
(Heff
u
)HF`FH
`Heffu
∣∣∣∣∣∣−E log2
∣∣∣∣∣∣IL + ρS∑
`=1, 6=u
(Heff
u
)HF`FH
`Heffu
∣∣∣∣∣∣ (60)
Interference only over the null-space component due to BD
Heffu = HuHH
u Heffu +
(INt − HuHH
u
)Heff
u = HuHHu Heff
u + H⊥u(
H⊥u)
H Heffu , (61)
Slide 55 / 89 Multi-User MIMO with Channel Subspace Feedback
Why Chordal Distance Quantization?
Achievable rate with perfect CSIT
RBD = E log2
∣∣∣IL + ρ(
Heffu
)HFuFH
u Heffu
∣∣∣ , ρ =P
σ2z S L
. (59)
assuming Gu to be semi-unitary GHu Gu = IL
and equal power allocation FHu Fu = IL
Achievable rate with limited feedback
RBD-Quant = E log2
∣∣∣∣∣∣IL + ρS∑`=1
(Heff
u
)HF`FH
`Heffu
∣∣∣∣∣∣−E log2
∣∣∣∣∣∣IL + ρS∑
`=1, 6=u
(Heff
u
)HF`FH
`Heffu
∣∣∣∣∣∣ (60)
Interference only over the null-space component due to BD
Heffu = HuHH
u Heffu +
(INt − HuHH
u
)Heff
u = HuHHu Heff
u + H⊥u(
H⊥u)
H Heffu , (61)
Slide 55 / 89 Multi-User MIMO with Channel Subspace Feedback
Why Chordal Distance Quantization?
Achievable rate with perfect CSIT
RBD = E log2
∣∣∣IL + ρ(
Heffu
)HFuFH
u Heffu
∣∣∣ , ρ =P
σ2z S L
. (59)
assuming Gu to be semi-unitary GHu Gu = IL
and equal power allocation FHu Fu = IL
Achievable rate with limited feedback
RBD-Quant = E log2
∣∣∣∣∣∣IL + ρS∑`=1
(Heff
u
)HF`FH
`Heffu
∣∣∣∣∣∣−E log2
∣∣∣∣∣∣IL + ρS∑
`=1, 6=u
(Heff
u
)HF`FH
`Heffu
∣∣∣∣∣∣ (60)
Interference only over the null-space component due to BD
Heffu = HuHH
u Heffu +
(INt − HuHH
u
)Heff
u = HuHHu Heff
u + H⊥u(
H⊥u)
H Heffu , (61)
Slide 55 / 89 Multi-User MIMO with Channel Subspace Feedback
Why Chordal Distance Quantization? (2)
The rate loss RBD − RBD-Quant is determined by d2c
(Heff
u , Hu
)Upper bounds on the expected rate loss for iid Rayleigh fading
Rate loss of ZF beamforming with L = Nr = 1 [Jindal, 2006]
RZF − RZF-Quant ≤ log2
(1 +
Pσ2
nD), D = 2
− BNt−1 , (62)
Rate loss of BD precoding with L = Nr [Ravindran and Jindal, 2008]
RBD − RBD-Quant ≤ Nr log2
(1 +
Pσ2
n NrD), (63)
D = CBD 2− B
Nr (Nt−Nr ) .
D average chordal distance distortion with random isotropic codebooks
Slide 56 / 89 Multi-User MIMO with Channel Subspace Feedback
Why Chordal Distance Quantization? (2)
The rate loss RBD − RBD-Quant is determined by d2c
(Heff
u , Hu
)Upper bounds on the expected rate loss for iid Rayleigh fading
Rate loss of ZF beamforming with L = Nr = 1 [Jindal, 2006]
RZF − RZF-Quant ≤ log2
(1 +
Pσ2
nD), D = 2
− BNt−1 , (62)
Rate loss of BD precoding with L = Nr [Ravindran and Jindal, 2008]
RBD − RBD-Quant ≤ Nr log2
(1 +
Pσ2
n NrD), (63)
D = CBD 2− B
Nr (Nt−Nr ) .
D average chordal distance distortion with random isotropic codebooks
Slide 56 / 89 Multi-User MIMO with Channel Subspace Feedback
Why Chordal Distance Quantization? (2)
The rate loss RBD − RBD-Quant is determined by d2c
(Heff
u , Hu
)Upper bounds on the expected rate loss for iid Rayleigh fading
Rate loss of ZF beamforming with L = Nr = 1 [Jindal, 2006]
RZF − RZF-Quant ≤ log2
(1 +
Pσ2
nD), D = 2
− BNt−1 , (62)
Rate loss of BD precoding with L = Nr [Ravindran and Jindal, 2008]
RBD − RBD-Quant ≤ Nr log2
(1 +
Pσ2
n NrD), (63)
D = CBD 2− B
Nr (Nt−Nr ) .
D average chordal distance distortion with random isotropic codebooks
Slide 56 / 89 Multi-User MIMO with Channel Subspace Feedback
Memoryless Grassmannian Quantization Codebooks
Codebooks for quantization of the full Nr dimensional subspace (Gu = I)
Random isotropic codebook for isotropic channels (e.g., iid Rayleigh fading)
[Hu ]m,n ∼ NC (0, 1) ∈ CNt×Nr ,
Q(iso)u =
{Q(iso)
j
∣∣∣∣Q(iso)j ΣVH = Qj , Qj ∈ CNt×Nr , [Qj ]m,n ∼ NC (0, 1)
}, (64)
Optimal codebook: maximally spaced subspace packing
Random codebook: asymptotically optimal in the codebook size⇒ random vector quantization (RVQ)
Random correlated codebook for spatially correlated channels
Hu = Γ1/2u Hu , [Hu ]m,n ∼ NC (0, 1) ∈ CNt×Nr ,
Q(corr)u =
{Q(corr)
j
∣∣∣∣Q(corr)j ΣVH = Γ
1/2u Qj , Qj ∈ CNt×Mu , [Qj ]m,n ∼ NC (0, 1)
},
The codebook has the same distribution as the channel
Receive-side correlation does not impact the subspace distribution!
Slide 57 / 89 Multi-User MIMO with Channel Subspace Feedback
Memoryless Grassmannian Quantization Codebooks
Codebooks for quantization of the full Nr dimensional subspace (Gu = I)
Random isotropic codebook for isotropic channels (e.g., iid Rayleigh fading)
[Hu ]m,n ∼ NC (0, 1) ∈ CNt×Nr ,
Q(iso)u =
{Q(iso)
j
∣∣∣∣Q(iso)j ΣVH = Qj , Qj ∈ CNt×Nr , [Qj ]m,n ∼ NC (0, 1)
}, (64)
Optimal codebook: maximally spaced subspace packing
Random codebook: asymptotically optimal in the codebook size⇒ random vector quantization (RVQ)
Random correlated codebook for spatially correlated channels
Hu = Γ1/2u Hu , [Hu ]m,n ∼ NC (0, 1) ∈ CNt×Nr ,
Q(corr)u =
{Q(corr)
j
∣∣∣∣Q(corr)j ΣVH = Γ
1/2u Qj , Qj ∈ CNt×Mu , [Qj ]m,n ∼ NC (0, 1)
},
The codebook has the same distribution as the channel
Receive-side correlation does not impact the subspace distribution!
Slide 57 / 89 Multi-User MIMO with Channel Subspace Feedback
Memoryless Grassmannian Quantization – Performance
10.80.60.40.20
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Correlation coefficient
Cho
rdal
dis
tanc
e M
SE
Memoryless_Grassmannian_quantization
4 × 2 (6 bit)
correlated codebook
isotropic codebook
8 × 2 (10 bit)
isotropic codebook
correlated codebook
Comparison of isotropic and correlated codebooks in dependence of the channel correlation.
Consider Nt × Nr = 8× 2 with B = 10 bit and Nt × Nr = 4× 2 with B = 6 bit
Kronecker correlation model
Γu =
1 αcorr . . . αcorr
αcorr 1 . . . αcorr...
. . ....
αcorr . . . αcorr 1
(65)
Slide 58 / 89 Multi-User MIMO with Channel Subspace Feedback
Predictive Grassmannian Quantization [Schwarz et al., 2013]
hk-1hk-2
hk
1D Grassmannian
Illustration of Grassmannian predictive quantization.
Exploit temporal correlation of the channel
Predict current channel subspace from previously quantized subspaces⇒ Grassmannian prediction exploiting geodesics
Generate a local codebook around the prediction
Volume covered by local codebook depends on prediction accuracy
Slide 59 / 89 Multi-User MIMO with Channel Subspace Feedback
Predictive Grassmannian Quantization [Schwarz et al., 2013]
hk-1hk-2
hk-2
hk-1
hk
^
^
1D Grassmannian
Illustration of Grassmannian predictive quantization.
Exploit temporal correlation of the channel
Predict current channel subspace from previously quantized subspaces⇒ Grassmannian prediction exploiting geodesics
Generate a local codebook around the prediction
Volume covered by local codebook depends on prediction accuracy
Slide 59 / 89 Multi-User MIMO with Channel Subspace Feedback
Predictive Grassmannian Quantization [Schwarz et al., 2013]
hk-1hk-2
hk-2
hk-1
hp,k
hk
^
^
1D Grassmannian
Illustration of Grassmannian predictive quantization.
Exploit temporal correlation of the channel
Predict current channel subspace from previously quantized subspaces⇒ Grassmannian prediction exploiting geodesics
Generate a local codebook around the prediction
Volume covered by local codebook depends on prediction accuracy
Slide 59 / 89 Multi-User MIMO with Channel Subspace Feedback
Predictive Grassmannian Quantization [Schwarz et al., 2013]
hk-1hk-2
hk-2
hk-1
hp,k
hk
^
^
1D Grassmannian
Illustration of Grassmannian predictive quantization.
Exploit temporal correlation of the channel
Predict current channel subspace from previously quantized subspaces⇒ Grassmannian prediction exploiting geodesics
Generate a local codebook around the prediction
Volume covered by local codebook depends on prediction accuracy
Slide 59 / 89 Multi-User MIMO with Channel Subspace Feedback
Predictive Grassmannian Quantization [Schwarz et al., 2013]
minimumchordal distance
quantization
Quantization
H[n,k] H[n,k]^ codebookindex extraction
codebook index
scale index
codebookgeneration
Adaptive codebook generation
H(p)[n,k]
subspaceprediction
~
codebooks
Encoder Decoder
feedbackchannel
subspacereconstruction
adaptivecodebookgeneration
H[n,k]^
codebooks
Structure of predictive quantization.
Exploit temporal correlation of the channel
Predict current channel subspace from previously quantized subspaces⇒ Grassmannian prediction exploiting geodesics
Generate a local codebook around the prediction
Volume covered by local codebook depends on prediction accuracy
Slide 59 / 89 Multi-User MIMO with Channel Subspace Feedback
Predictive Grassmannian Quantization – MSE Performance
100
10-1
10-2
10-3
10-4
10-2
10-5
10-6
5·10-3
Normalized Doppler frequency
Cho
rdal
dis
tanc
e M
SE
95% confidence interval95% confidence interval 4x1_1stream_flat
10-1 0.5
4 bit
8 bit
robust prediction (Zhang)adaptive prediction (Schwarz)
differential (El Ayach)
10-110-210-3
100
10-1
10-2
10-3
10-4
10-6
10-5
Normalized Doppler frequency
Cho
rdal
dis
tanc
e M
SE
95% confidence interval95% confidence interval 8x2_2stream_flat
7 bit
11 bit
predictive quant. (vector)predictive quant. (matrix)
differential quant.
Performance of differential and predictive Grassmannian quantization (4× 1 and 8× 2).
Nt × Nr ∈ {4× 2, 8× 2} with varying speed (Doppler frequency)
Differential quantization [Ayach and Heath, Jr., 2011]
Robust predictive quantization [Zhang and Lei, 2012]
Adaptive predictive quantization [Schwarz et al., 2013]
Slide 60 / 89 Multi-User MIMO with Channel Subspace Feedback
Predictive Grassmannian Quantization – Throughput Performance
302724211815129630
20
18
16
14
12
10
8
6
4
2
0
SNR [dB]
Sum
thro
ughp
ut [M
bit/s
]
2UE_10Hz_AR_4x2_2streams
RSQ 8 bit
SU-MIMO 4 bit
Perfect CSI
ACSQ 8 bit
ACSQ 3 bit
ACSQ 2 bit
302724211815129630
35
30
25
20
15
10
5
0
SNR [dB]
Sum
thro
ughp
ut [M
bit/s
]
Perfect CSI
4UE_10Hz_8x2_2streams
ACSQ 11 bit
ACSQ 9 bit
ACSQ 7 bit
SU-MIMO 8 bit
RSQ 11 bitACSQ 5 bit
Throughput with predictive Grassmannian quantization (4× 2 and 8× 2).
Nt × Nr ∈ {4× 2, 8× 2} at low mobility νd = 0.01 (walking speed at 1 GHz)
Memoryless quantization (RSQ) versus predictive quantization (ACSQ)
Notice: no scheduling applied→ always 2 (resp. 4) users served in parallel
Single-user MIMO using LTE’s CLSM mode with best CQI scheduler
Slide 61 / 89 Multi-User MIMO with Channel Subspace Feedback
Other Subspace Transceivers
Single-user MIMO with unitary precoding (equal power allocation)
Interference alignment [Cadambe and Jafar, 2008, Maddah-Ali et al., 2008]
Rate-loss of interference alignment with quantized CSIT is also determined bythe chordal distance quantization error [Rezaee and Guillaud, 2012]
Slide 62 / 89 Multi-User MIMO with Channel Subspace Feedback
Contents
1 Motivation
2 Codebook based Single-User MIMO Feedback for LTE
3 Multi-User MIMO Feedback for LTE
4 Multi-User MIMO with Channel Subspace Feedback
5 Multi-User MIMO with Channel Gramian Feedback
6 Feedback Overhead Reduction through Excess Antennas
Slide 63 / 89 Multi-User MIMO with Channel Gramian Feedback
Recap: Regularized Block Diagonalization (RBD) Precoding
Precoder calculation consist of two parts (see multi-user MIMO lecture):
Trade-off residual interference and noise through MMSE precoding
Optimize transmission over effective single-user channel treating residualinterference as noise through SVD precoding
Necessary CSI at the transmitter [Schwarz and Rupp, 2014b]
Hu = UuΣuVHu ∈ CNt×Nr , (66)
Individual columns of Uu ,
Singular values in Σu
Both can be obtained from an eigen-decomposition of the channel Gramian
Ru = HuHHu = UuΣ2
uUHu (67)
Slide 64 / 89 Multi-User MIMO with Channel Gramian Feedback
Recap: Regularized Block Diagonalization (RBD) Precoding
Precoder calculation consist of two parts (see multi-user MIMO lecture):
Trade-off residual interference and noise through MMSE precoding
Optimize transmission over effective single-user channel treating residualinterference as noise through SVD precoding
Necessary CSI at the transmitter [Schwarz and Rupp, 2014b]
Hu = UuΣuVHu ∈ CNt×Nr , (66)
Individual columns of Uu ,
Singular values in Σu
Both can be obtained from an eigen-decomposition of the channel Gramian
Ru = HuHHu = UuΣ2
uUHu (67)
Slide 64 / 89 Multi-User MIMO with Channel Gramian Feedback
Recap: Regularized Block Diagonalization (RBD) Precoding
Precoder calculation consist of two parts (see multi-user MIMO lecture):
Trade-off residual interference and noise through MMSE precoding
Optimize transmission over effective single-user channel treating residualinterference as noise through SVD precoding
Necessary CSI at the transmitter [Schwarz and Rupp, 2014b]
Hu = UuΣuVHu ∈ CNt×Nr , (66)
Individual columns of Uu ,
Singular values in Σu
Both can be obtained from an eigen-decomposition of the channel Gramian
Ru = HuHHu = UuΣ2
uUHu (67)
Slide 64 / 89 Multi-User MIMO with Channel Gramian Feedback
Channel Gramian Quantization – Stiefel Manifold Feedback
Two possibilities for CSI feedback
Direct quantization of the Gramian Ru [Sacristan-Murga et al., 2012]
Separate quantization of Uu and diag (Σu) [Schwarz and Rupp, 2014b]
Separate quantization:
Quantization of Uu ∈ St (Nt ,Nr ) on the Stiefel manifold
d2s (U,Qi ) =
Nr∑j=1
d2c
(uj ,q
(j)i
)∈ [0,Nr ], q(j)
i = [Qi ]:,j (68)
d2c (A,B) = Nr − tr
(AHBBHA
), AHA = BHB = INr ,
Q ={
Qi ∈ St (Nt ,Nr )∣∣ i ∈ {1, . . . , 2B}
}(69)
Quantization of singular values
σu = argminsi∈Su
‖σu − si‖22 , (70)
σu =[[Σu ](1,1), . . . , [Σu ](Nr ,Nr )
]T (71)
Slide 65 / 89 Multi-User MIMO with Channel Gramian Feedback
Channel Gramian Quantization – Stiefel Manifold Feedback
Two possibilities for CSI feedback
Direct quantization of the Gramian Ru [Sacristan-Murga et al., 2012]
Separate quantization of Uu and diag (Σu) [Schwarz and Rupp, 2014b]
Separate quantization:
Quantization of Uu ∈ St (Nt ,Nr ) on the Stiefel manifold
d2s (U,Qi ) =
Nr∑j=1
d2c
(uj ,q
(j)i
)∈ [0,Nr ], q(j)
i = [Qi ]:,j (68)
d2c (A,B) = Nr − tr
(AHBBHA
), AHA = BHB = INr ,
Q ={
Qi ∈ St (Nt ,Nr )∣∣ i ∈ {1, . . . , 2B}
}(69)
Quantization of singular values
σu = argminsi∈Su
‖σu − si‖22 , (70)
σu =[[Σu ](1,1), . . . , [Σu ](Nr ,Nr )
]T (71)
Slide 65 / 89 Multi-User MIMO with Channel Gramian Feedback
Singular Value Codebook Su
7654321
5
4
3
2
1
0
First singular value
Seco
nd s
ingu
lar v
alue
8x2_uncorrelated_SV_CB
7654321
5
4
3
2
1
0
First singular value
Seco
nd s
ingu
lar v
alue
8x2_correlated_SV_CB
Training set
Codebook
Training sets and Lloyd codebooks for singular value quantization with uncorrelated and correlated receive antennas.
Codebook optimized using Lloyd’s algorithm (k-means clustering) [Lloyd, 1982]
Nt × Nr = 8× 2 with uncorrelated and correlated receive antennas (αcorr = 0.9)
More accurate quantization with increasing correlation
⇒ Exploit correlation!
Slide 66 / 89 Multi-User MIMO with Channel Gramian Feedback
Quantization on the Stiefel Manifold
Memoryless quantization:
Optimal: maximally spaced subspace packings on the Stiefel manifold
Random: same codebooks as with Grassmannian quantization
Different quantization metrics!
⇒ seamless transition between subspace and Gramian quantization
Predictive quantization:
Same principle applicable as with subspace quantization
Just replace manifold calculations appropriately
[Schwarz and Rupp, 2015b]
Slide 67 / 89 Multi-User MIMO with Channel Gramian Feedback
Quantization on the Stiefel Manifold
Memoryless quantization:
Optimal: maximally spaced subspace packings on the Stiefel manifold
Random: same codebooks as with Grassmannian quantization
Different quantization metrics!
⇒ seamless transition between subspace and Gramian quantization
Predictive quantization:
Same principle applicable as with subspace quantization
Just replace manifold calculations appropriately
[Schwarz and Rupp, 2015b]
Slide 67 / 89 Multi-User MIMO with Channel Gramian Feedback
Predictive Quantization on the Stiefel Manifold
Normalized Doppler frequency
Cho
rdal
dis
tanc
e M
SE8x2_2stream_flat_corr0
11 bit
7 bit
predictive quant. differential quant.memoryless quant.
100
10-1
10-2
10-3
10-4
10-6
10-5
10-110-210-3
Normalized Doppler frequency
Cho
rdal
dis
tanc
e M
SE
8x2_2stream_flat_corr0.9
100
10-1
10-2
10-3
10-4
10-6
10-5
10-110-210-3
11 bit
7 bit
predictive quant. differential quant.memoryless quant.
Performance of Stiefel manifold quantization with uncorrelated and correlated receive antennas.
Nt × Nr = 8× 2 with varying speed (Doppler frequency)
Substantial gain of differential/predictive quantization over memoryless scheme
Larger slope of MSE curve of predictive quantization
MSE improvement with correlated receive antennas (αcorr = 0.9)
Notice: the prediction order needs to be adapted to the speed
Slide 68 / 89 Multi-User MIMO with Channel Gramian Feedback
BD versus RBD Precoding with Limited Feedback
302520151050
40
35
30
25
20
15
10
5
0
SNR [dB]
6x2_10Hz_MU-MIMO_RXcorr0.995% confidence interval95% confidence interval
BD quantized CSITBD perfect CSIT
RBD quantized CSITRBD perfect CSIT
302520151050
40
35
30
25
20
15
10
5
0
SNR [dB]
6x2_10Hz_MU-MIMO_RXcorr0
6 bit
6 bit
10 bit
10 bit A
chie
vabl
e su
m ra
te [b
its/c
hann
el u
se]
6 bit 6 bit
10 bit 10 bit
BD quantized CSITBD perfect CSIT
RBD quantized CSITRBD perfect CSIT
95% confidence interval95% confidence interval
Comparison of the achievable rate of block-diagonalization and regularized block-diagonalization precoding.
Nt × Nr = 6× 2 at low mobility νd = 0.01 (walking speed at 1 GHz)
Negligible overhead for singular value quantization: 4 bit/10 TTI
Uncorrelated and strongly correlated receive antennas: αcorr ∈ {0, 0.9}
⇒ Switch between the two schemes depending on SNR
Slide 69 / 89 Multi-User MIMO with Channel Gramian Feedback
BD versus RBD Precoding with Limited Feedback
302520151050
100
90
80
70
60
50
40
30
SNR [dB]
Rel
ativ
e su
m ra
te [%
]6x2_10Hz_MU-MIMO_RXcorr0
302520151050
100
90
80
70
60
50
40
30
SNR [dB]
6x2_10Hz_MU-MIMO_RXcorr0.9
6 bit
6 bit
10 bit
10 bit
BD quantized CSITBD perfect CSIT
RBD quantized CSIT
6 bit
6 bit
10 bit
10 bit
Comparison of the rate relative to regularized block-diagonalization precoding with perfect CSIT.
Nt × Nr = 6× 2 at low mobility νd = 0.01 (walking speed at 1 GHz)
Negligible overhead for singular value quantization: 4 bit/10 TTI
Uncorrelated and strongly correlated receive antennas: αcorr ∈ {0, 0.9}
⇒ Switch between the two schemes depending on SNR
Slide 69 / 89 Multi-User MIMO with Channel Gramian Feedback
Other Gramian-Based Transceivers
Single-user MIMO with water-filling power allocation
MMSE based precoding schemes (iterative)
Interference leakage based schemes, e.g., signal to leakage and noise ratio(SLNR) beamforming
SLNRu =
∥∥HHu fu∥∥2
F
Nr σ2z +
∑j 6=u
∥∥∥HHj fu
∥∥∥2
F
(72)
Slide 70 / 89 Multi-User MIMO with Channel Gramian Feedback
Contents
1 Motivation
2 Codebook based Single-User MIMO Feedback for LTE
3 Multi-User MIMO Feedback for LTE
4 Multi-User MIMO with Channel Subspace Feedback
5 Multi-User MIMO with Channel Gramian Feedback
6 Feedback Overhead Reduction through Excess Antennas
Slide 71 / 89 Feedback Overhead Reduction through Excess Antennas
Receive Antenna Combining – Maximum Eigenmode Transmission
Excess receive antennas: Nr > L
Heffu = HuGu ∈ CNt×L (73)
Minimize feedback overhead: selfishly select Gu and quantize Heffu
⇒ How to select Gu?
Maximum eigenmode transmission (MET)
G(MET)u = [Vu ]:,1:L , (74)
Heffu = [Uu ]:,1:L [Σu ]1:L,1:L
Achieves maximum rate in the absence of interference
Slide 72 / 89 Feedback Overhead Reduction through Excess Antennas
Receive Antenna Combining – Maximum Eigenmode Transmission
Excess receive antennas: Nr > L
Heffu = HuGu ∈ CNt×L (73)
Minimize feedback overhead: selfishly select Gu and quantize Heffu
⇒ How to select Gu?
Maximum eigenmode transmission (MET)
G(MET)u = [Vu ]:,1:L , (74)
Heffu = [Uu ]:,1:L [Σu ]1:L,1:L
Achieves maximum rate in the absence of interference
Slide 72 / 89 Feedback Overhead Reduction through Excess Antennas
MET with Quantized CSIT [Schwarz and Rupp, 2013]
Performance of MET combining with BD precoding, L streams per user, S usersand B bit Grassmannian quantization (iid Rayleigh fading and RVQ))
RMET − RMET-Quant ≤L∑`=1
log2
(1 + ρ σ2
`,uS − 1Nt − L
D), ρ =
Pσ2
z S L
σ2`,u expected value of `th eigenvalue ofWC
Nr
(Nt , INr
)D = CMET 2
− BL (Nt−L) average quantization distortion with RVQ
With fixed D the rate loss grows with SNR⇒ interference limitation
Avoid interference limitation:
B ∝ log (ρ) L(Nt − L) =⇒ 2− B
L (Nt−L) ∝1ρ
(75)
Slide 73 / 89 Feedback Overhead Reduction through Excess Antennas
MET with Quantized CSIT [Schwarz and Rupp, 2013]
Performance of MET combining with BD precoding, L streams per user, S usersand B bit Grassmannian quantization (iid Rayleigh fading and RVQ))
RMET − RMET-Quant ≤L∑`=1
log2
(1 + ρ σ2
`,uS − 1Nt − L
D), ρ =
Pσ2
z S L
σ2`,u expected value of `th eigenvalue ofWC
Nr
(Nt , INr
)D = CMET 2
− BL (Nt−L) average quantization distortion with RVQ
With fixed D the rate loss grows with SNR⇒ interference limitation
Avoid interference limitation:
B ∝ log (ρ) L(Nt − L) =⇒ 2− B
L (Nt−L) ∝1ρ
(75)
Slide 73 / 89 Feedback Overhead Reduction through Excess Antennas
Subspace Quantization Based Combining [Schwarz and Rupp, 2013]
Select the effective channel to minimize the quantization error{G(SQBC)
u , H(SQBC)u
}= argmin
G,Qj
d2c
(Heff
u ,Qj
)= argmin
G,Qj
d2c(HuG,Qj
)H(SQBC)
u = argminQj∈Q
(Nt )
L
d2c(Hu ,Qj
), (76)
G(SQBC)u =
(HH
u Hu)−1 HH
u H(SQBC)u (77)
Rate loss with respect to perfect CSIT (same assumptions as before)
R(L)BD − R(L,Nr )
SQBC ≤ L log2
(1 + ρ
Nt − Nr + L
Nt − L(S − 1) D
)+ log2 (e)
L−1∑k=0
Nt−1∑`=Nt−Nr +L
1
`− k
D = CSQBC 2− B
L (Nt−Nr ) average quantization distortion with RVQ
Bit-scaling law: B ∝ log (ρ) L(Nt − Nr )
Slide 74 / 89 Feedback Overhead Reduction through Excess Antennas
Subspace Quantization Based Combining [Schwarz and Rupp, 2013]
Select the effective channel to minimize the quantization error{G(SQBC)
u , H(SQBC)u
}= argmin
G,Qj
d2c
(Heff
u ,Qj
)= argmin
G,Qj
d2c(HuG,Qj
)H(SQBC)
u = argminQj∈Q
(Nt )
L
d2c(Hu ,Qj
), (76)
G(SQBC)u =
(HH
u Hu)−1 HH
u H(SQBC)u (77)
Rate loss with respect to perfect CSIT (same assumptions as before)
R(L)BD − R(L,Nr )
SQBC ≤ L log2
(1 + ρ
Nt − Nr + L
Nt − L(S − 1) D
)+ log2 (e)
L−1∑k=0
Nt−1∑`=Nt−Nr +L
1
`− k
D = CSQBC 2− B
L (Nt−Nr ) average quantization distortion with RVQ
Bit-scaling law: B ∝ log (ρ) L(Nt − Nr )
Slide 74 / 89 Feedback Overhead Reduction through Excess Antennas
MET versus SQBC
302724211815129630
50
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bits
/s/H
z]
6xN_flat_2_streams
SQBC MET
Nr = 2
95% confidence interval95% confidence interval
Nr = 5
Nr = 5
Achievable rate with perfect CSIT
dSQBC(L,Nr)
3028262422201816141210
90
80
70
60
50
40
30
20
10
0
SNR [dB]
Suff
icie
nt n
umbe
r of f
eedb
ack
bits
6xN_flat_2_streams
SQBC MET Nr = 2
Nr = 5
Nr = 5
L (Nt-L)
L (Nt-Nr)
Feedback bit-scaling to achieve a loss of 1 bit/s/Hz
Nt = 6 transmit antennas, Nr ∈ {2, . . . , 5} receive antennas, L = 2 data streams
Throughput loss of SQBC with perfect CSI at the base station
Significant reduction of feedback overhead with SQBC for moderate SNR loss
Slide 75 / 89 Feedback Overhead Reduction through Excess Antennas
MET versus SQBC
302724211815129630
50
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bits
/s/H
z]
6xN_flat_2_streams
SQBC MET
Nr = 2
95% confidence interval95% confidence interval
Nr = 5
Nr = 5
Achievable rate with perfect CSIT
dSQBC(L,Nr)
3028262422201816141210
90
80
70
60
50
40
30
20
10
0
SNR [dB]
Suff
icie
nt n
umbe
r of f
eedb
ack
bits
6xN_flat_2_streams
SQBC MET Nr = 2
Nr = 5
Nr = 5
L (Nt-L)
L (Nt-Nr)
Feedback bit-scaling to achieve a loss of 1 bit/s/Hz
Nt = 6 transmit antennas, Nr ∈ {2, . . . , 5} receive antennas, L = 2 data streams
Throughput loss of SQBC with perfect CSI at the base station
Significant reduction of feedback overhead with SQBC for moderate SNR loss
Slide 75 / 89 Feedback Overhead Reduction through Excess Antennas
SQBC Performance
302724211815129630
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bits
/s/H
z]
6xN_flat_2_streams
SQBC quant. CSIT SQBC perfect CSIT
95% confidence interval95% confidence interval
Nr = 2
Nr = 4
Nr = 2
Nr = 3
Nr = 4
Nr = 5
Achievable rate with feedback bits scaled to achieve a loss of 1 bit/s/Hz with Nr = 5.
Nt = 6 transmit antennas, Nr ∈ {2, . . . , 5} receive antennas, L = 2 data streams
Feedback overhead growing from 0 to 17 bits per TTI
Significant throughput gain with growing Nr at same feedback overhead
Substantial gain over MET at same overhead
Slide 76 / 89 Feedback Overhead Reduction through Excess Antennas
SQBC Performance
302724211815129630
45
40
35
30
25
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bits
/s/H
z]
6xN_flat_2_streams
SQBC quant. CSIT SQBC perfect CSIT
MET quant. CSIT Nr = 5
95% confidence interval95% confidence interval
Nr = 2
Nr = 4
Nr = 2
Nr = 3
Nr = 4
Nr = 5
Achievable rate with feedback bits scaled to achieve a loss of 1 bit/s/Hz with Nr = 5.
Nt = 6 transmit antennas, Nr ∈ {2, . . . , 5} receive antennas, L = 2 data streams
Feedback overhead growing from 0 to 17 bits per TTI
Significant throughput gain with growing Nr at same feedback overhead
Substantial gain over MET at same overhead
Slide 76 / 89 Feedback Overhead Reduction through Excess Antennas
SQBC Performance (2)
302724211815129630
40
35
30
25
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bits
/s/H
z]
6x5_flat_N_streams
SQBC quant. CSIT (scaled bits)SQBC perfect CSIT
3 streams per user
2 streams per user
1 stream per user
95% confidence interval95% confidence interval
Achievable rate with different number of streams per user.
Nt = 6 transmit antennas, Nr = 5 receive antennas, L ∈ {1, 2, 3} data streams
Feedback overhead per user
L = 1, U = 6: B ∈ [0, 8] bits per TTI
L = 2, U = 3: B ∈ [0, 16.1] bits per TTI
L = 3, U = 2: B ∈ [0, 18.3] bits per TTI
Slide 77 / 89 Feedback Overhead Reduction through Excess Antennas
Maximum Expected Achievable Rate Combining (MERC)
Exploit the BD construction to further improve performance⇒ Precoding-specific combiner [Schwarz and Rupp, 2015a]:
Estimate the expected achievable rate with BD precoding, antennacombiner G and quantized channel subspace Qi : Ru(Qi ,G)
Expectation over input covariance matrices
Interference lies in the null-space (INt − Qi QHi )
MERC optimization and quantization problem:{Hu ,G
(MERC)u
}= argmax
Qi∈Q(Nt )
L ,G∈CNr ×L
Ru(Qi ,G), (78)
Hu = argminQi∈Q
(Nt )
L
log2 det(σ2
z INr +1Nt
HHu (INt − Qi QH
i )Hu
), (79)
G(MERC,1)u =
(σ2
z INr +1Nt
HHu (INt−HuHH
u )Hu
)−1HH
u Hu (80)
MMSE solution versus ZF as with SQBC
Slide 78 / 89 Feedback Overhead Reduction through Excess Antennas
Maximum Expected Achievable Rate Combining (MERC)
Exploit the BD construction to further improve performance⇒ Precoding-specific combiner [Schwarz and Rupp, 2015a]:
Estimate the expected achievable rate with BD precoding, antennacombiner G and quantized channel subspace Qi : Ru(Qi ,G)
Expectation over input covariance matrices
Interference lies in the null-space (INt − Qi QHi )
MERC optimization and quantization problem:{Hu ,G
(MERC)u
}= argmax
Qi∈Q(Nt )
L ,G∈CNr ×L
Ru(Qi ,G), (78)
Hu = argminQi∈Q
(Nt )
L
log2 det(σ2
z INr +1Nt
HHu (INt − Qi QH
i )Hu
), (79)
G(MERC,1)u =
(σ2
z INr +1Nt
HHu (INt−HuHH
u )Hu
)−1HH
u Hu (80)
MMSE solution versus ZF as with SQBC
Slide 78 / 89 Feedback Overhead Reduction through Excess Antennas
Maximum Expected Achievable Rate Combining (MERC)
Exploit the BD construction to further improve performance⇒ Precoding-specific combiner [Schwarz and Rupp, 2015a]:
Estimate the expected achievable rate with BD precoding, antennacombiner G and quantized channel subspace Qi : Ru(Qi ,G)
Expectation over input covariance matrices
Interference lies in the null-space (INt − Qi QHi )
MERC optimization and quantization problem:{Hu ,G
(MERC)u
}= argmax
Qi∈Q(Nt )
L ,G∈CNr ×L
Ru(Qi ,G), (78)
Hu = argminQi∈Q
(Nt )
L
log2 det(σ2
z INr +1Nt
HHu (INt − Qi QH
i )Hu
), (79)
G(MERC,1)u =
(σ2
z INr +1Nt
HHu (INt−HuHH
u )Hu
)−1HH
u Hu (80)
MMSE solution versus ZF as with SQBC
Slide 78 / 89 Feedback Overhead Reduction through Excess Antennas
MERC Performance (1)
4035302520151050-5-10
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bit/
s/H
z]
4x2_4UE_1streams
MET
MERCMET perfect CSITSQBC
Achievable rate with interference unaware antenna combining.
Nt × Nr = 4× 2, L = 1 and B = 10 bits per TTI with αcorr = 0.9
Comparison of MET, SQBC and MERC feedback and antenna combining⇒ MERC performs equal to MET and SQBC at low and high SNR, resp.
MET, SQBC and MERC feedback with interference-aware MMSE combining⇒ Overall better performance; difference reduces
Slide 79 / 89 Feedback Overhead Reduction through Excess Antennas
MERC Performance (1)
4035302520151050-5-10
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bit/
s/H
z]
4x2_4UE_1streams
MET perfect CSIT
MET
MERCSQBC
Achievable rate with interference aware antenna combining.
Nt × Nr = 4× 2, L = 1 and B = 10 bits per TTI with αcorr = 0.9
Comparison of MET, SQBC and MERC feedback and antenna combining⇒ MERC performs equal to MET and SQBC at low and high SNR, resp.
MET, SQBC and MERC feedback with interference-aware MMSE combining⇒ Overall better performance; difference reduces
Slide 79 / 89 Feedback Overhead Reduction through Excess Antennas
MERC Performance (2)
4035302520151050-5-10
25
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bit/
s/H
z]
8x4_4UE_2streams
METSQBCMERC
MET perfect CSIT
Achievable rate with interference unaware antenna combining.
Nt × Nr = 8× 4, L = 2 and B = 14 bits per TTI with αcorr = 0.9
Comparison of MET, SQBC and MERC feedback and antenna combining⇒ MERC strictly outperforms MET and SQBC
MET, SQBC and MERC feedback with interference-aware MMSE combining⇒ Overall better performance; difference reduces
Slide 80 / 89 Feedback Overhead Reduction through Excess Antennas
MERC Performance (2)
4035302520151050-5-10
25
20
15
10
5
0
SNR [dB]
Ach
ieva
ble
sum
rate
[bit/
s/H
z]
8x4_4UE_2streams
METSQBCMERC
MET perfect CSIT
Achievable rate with interference aware antenna combining.
Nt × Nr = 8× 4, L = 2 and B = 14 bits per TTI with αcorr = 0.9
Comparison of MET, SQBC and MERC feedback and antenna combining⇒ MERC strictly outperforms MET and SQBC
MET, SQBC and MERC feedback with interference-aware MMSE combining⇒ Overall better performance; difference reduces
Slide 80 / 89 Feedback Overhead Reduction through Excess Antennas
Limited Feedback for Single- and Multi-User MIMO389.168 Advanced Wireless Communications 1
Abbreviations I
AMC adaptive modulation and coding
AWGN additive white Gaussian noise
BD block diagonalization
BICM bit-interleaved coded-modulation
BLER block error ratio
CLSM closed loop spatial multiplexing
CoMP coordinated multipoint transmission/reception
CP cyclic prefix
CQI channel quality indicator
CSI channel state information
CSIT channel state information at the transmitter
EESM exponential effective SNR mapping
ESM effective SNR mapping
FDD frequency division duplex
LTE long term evolution
MCS modulation and coding scheme
MERC maximum expected achievable rate combining
MET maximum eigenmode transmission
Slide 82 / 89 Abbreviations
Abbreviations IIMIESM mutual information effective SNR mappingMIMO multiple-input multiple-output
ML maximum likelihoodMMSE minimum mean squared error
MRT maximum ratio transmissionOFDM orthogonal frequency division multiplexingOLSM open loop spatial multiplexing
PMI precoding matrix indicatorPU2RC per user unitary rate control
RB resource blockRBD regularized block diagonalization
RE resource elementRI rank indicator
RVQ random vector quantizationSINR signal to interference and noise ratioSISO single-input single-outputSLNR signal to leakage and noise ratio
SNR signal to noise ratioSQBC subspace quantization based combining
SVD singular value decompositionTDD time division duplex
ZF zero forcing
Slide 83 / 89 Abbreviations
References I
3GPP (2009).Technical Specification Group Radio Access Network; Evolved Universal Terrestrial Radio Access(E-UTRA); Physical Channels and Modulation (Release 8).[Online]. Available: http://www.3gpp.org/ftp/Specs/html-info/36211.htm.
3GPP (2010).Technical Specification Group Radio Access Network; Evolved Universal Terrestrial Radio Access(E-UTRA); Physical Channels and Modulation (Release 10).[Online]. Available: http://www.3gpp.org/ftp/Specs/html-info/36211.htm.
Ayach, O. E. and Heath, Jr., R. (2011).Grassmannian differential limited feedback for interference alignment.CoRR, abs/1111.4596.
Cadambe, V. and Jafar, S. (2008).Interference alignment and degrees of freedom of the K-user interference channel.IEEE Transactions on Information Theory, 54(8):3425 –3441.
Cipriano, A., Visoz, R., and Salzer, T. (2008).Calibration issues of PHY layer abstractions for wireless broadband systems.In IEEE 68th Vehicular Technology Conference, pages 1–5, Calgary, Alberta.
Costa, M. (1983).Writing on dirty paper (corresp.).IEEE Transactions on Information Theory, 29(3):439 – 441.
Dhillon, I. S., Heath, Jr., R., Strohmer, T., and Tropp, J. A. (2007).Constructing packings in Grassmannian manifolds via alternating projection.ArXiv e-prints.
Slide 84 / 89 References
References II
He, X., Niu, K., He, Z., and Lin, J. (2007).Link layer abstraction in MIMO-OFDM system.In International Workshop on Cross Layer Design, pages 41–44.
Hochwald, B., Peel, C., and Swindlehurst, A. (2005).A vector-perturbation technique for near-capacity multiantenna multiuser communication-part II:perturbation.IEEE Transactions on Communications, 53(3):537–544.
Jindal, N. (2006).MIMO broadcast channels with finite-rate feedback.IEEE Transactions on Information Theory, 52(11):5.
Lloyd, S. (1982).Least squares quantization in PCM.IEEE Transactions on Information Theory, 28(2):129–137.
Love, D. (2006).Grassmannian subspace packing.https://engineering.purdue.edu/˜djlove/grass.html.
Love, D. and Heath, Jr., R. (2005).Limited feedback unitary precoding for spatial multiplexing systems.IEEE Transactions on Information Theory, 51(8):2967–2976.
Maddah-Ali, M., Motahari, A., and Khandani, A. (2008).Communication over MIMO X channels: Interference alignment, decomposition, and performance analysis.IEEE Transactions on Information Theory, 54(8):3457 –3470.
Slide 85 / 89 References
References III
Mezghani, A., Hunger, R., Joham, M., and Utschick, W. (2006).Iterative THP transceiver optimization for multi-user MIMOsystems based on weighted sum-MSEminimization.In IEEE 7th Workshop on Signal Processing Advances in Wireless Communications, pages 1–5.
Ravindran, N. and Jindal, N. (2008).Limited feedback-based block diagonalization for the MIMO broadcast channel.IEEE Journal on Selected Areas in Communications, 26(8):1473 –1482.
Rezaee, M. and Guillaud, M. (2012).Interference alignment with quantized Grassmannian feedback in the K-user MIMO interference channel.CoRR, abs/1207.6902.
Ryan, D., Vaughan, I., Clarkson, L., Collings, I., Guo, D., and Honig, M. (2007).QAM codebooks for low-complexity limited feedback MIMO beamforming.In IEEE International Conference on Communications, pages 4162–4167, Glasgow, Scotland.
Sacristan-Murga, D., Payaro, M., and Pascual-Iserte, A. (2012).Transceiver design framework for multiuser MIMO-OFDM broadcast systems with channel Gram matrixfeedback.IEEE Transactions on Wireless Communications, 11(5):1774–1787.
Sandanalakshmi, R., Palanivelu, T. G., and Manivannan, K. (2007).Effective SNR mapping for link error prediction in OFDM based systems.In International Conference on Information and Communication Technology in Electrical Sciences, pages684–687.
Schwarz, S., Heath, Jr., R., and Rupp, M. (2013).Adaptive quantization on a Grassmann-manifold for limited feedback beamforming systems.IEEE Transactions on Signal Processing, 61(18):4450–4462.
Slide 86 / 89 References
References IV
Schwarz, S., Mehlfuhrer, C., and Rupp, M. (2010).Low complexity approximate maximum throughput scheduling for LTE.In Conference Record of the Forty Fourth Asilomar Conference on Signals, Systems, and Computers, pages1563–1569, Pacific Grove, California.
Schwarz, S., Mehlfuhrer, C., and Rupp, M. (2011).Throughput maximizing multiuser scheduling with adjustable fairness.In International Conference on Communications ICC 2011, Kyoto, Japan.
Schwarz, S. and Rupp, M. (2011).Throughput maximizing feedback for MIMO OFDM based wireless communication systems.In Signal Processing Advances in Wireless Communications SPAWC 2011, pages 316–320, San Francisco,CA.
Schwarz, S. and Rupp, M. (2013).Subspace quantization based combining for limited feedback block-diagonalization.IEEE Transactions on Wireless Communications, 12(11):5868–5879.
Schwarz, S. and Rupp, M. (2014a).Evaluation of distributed multi-user MIMO-OFDM with limited feedback.IEEE Transactions on Wireless Communications, 13(11):6081–6094.
Schwarz, S. and Rupp, M. (2014b).Subspace versus eigenmode quantization for limited feedback block-diagonalization.In 6th International Symposium on Communications, control and signal processing, pages 1–4.
Schwarz, S. and Rupp, M. (2015a).Maximum expected achievable rate combining for limited feedback block-diagonalization.Submitted to ICASSP 2015.
Slide 87 / 89 References
References V
Schwarz, S. and Rupp, M. (2015b).Predictive quantization on the Stiefel manifold.IEEE Signal Processing Letter, 22(2):234–238.
Shi, S., Schubert, M., and Boche, H. (2008).Downlink MMSE transceiver optimization for multiuser MIMO systems: MMSE balancing.IEEE Transactions on Signal Processing, 56(8):3702–3712.
Spencer, Q., Swindlehurst, A., and Haardt, M. (2004).Zero-forcing methods for downlink spatial multiplexing in multiuser MIMO channels.IEEE Trans. on Signal Processing, 52(2):461 – 471.
Trivellato, M., Boccardi, F., and Tosate, F. (2007).User selection schemes for MIMO broadcast channels with limited feedback.In 65th IEEE Vehicular Technology Conference, Spring 2007, Dublin.
Tsai, S. and Soong, A. (2003).Effective-SNR mapping for modeling frame error rates in multiple-state channels.Technical Report 3GPP2-C30-20030429-010, 3GPP2.
Wan, L., Tsai, S., and Almgren, M. (2006).A Fading-Insensitve Performance Metric for a Unified Link Quality Model.In Proc. IEEE Wireless Communications & Networking Conference WCNC, volume 4, pages 2110–2114.
Yang, D., Yang, L.-L., and Hanzo, L. (2010).DFT-based beamforming weight-vector codebook design for spatially correlated channels in the unitaryprecoding aided multiuser downlink.In IEEE International Conference on Communications, pages 1–5, Cape Town, South Africa.
Slide 88 / 89 References
References VI
Zhang, Y. and Lei, M. (2012).Robust Grassmannian prediction for limited feedback multiuser MIMO systems.In IEEE Wireless Communications and Networking Conference (WCNC), pages 863 –867.
Zheng, J. and Rao, B. (2008).Capacity analysis of MIMO systems using limited feedback transmit precoding schemes.IEEE Transactions on Signal Processing, 56(7):2886–2901.
Slide 89 / 89 References