3GPP TSG RAN WG1 Meeting #62 R1-104490

Madrid, Spain, 23rd - 27th August, 2010 R1-104161

Source: HTC
Title: Autonomous Feedback Reduction for MU-MIMO
Agenda Item: 6.3.3
Document For: Discussion

1 Introduction
In a multiuser system, independent time-frequency varying channels among different users can be exploited to
provide multiuser diversity (MUD) gain and increase the system throughput. In Rel-8, DL MU-MIMO is supported with
codebook-based pre-coding to improve system throughput when the RI is equal to one. UE reports SU-MIMO CQI/PMI
which the eNB uses for selecting pre-coding matrices and performing link adaptation. This incurs a large aggregate
feedback bandwidth, especially when the number of UEs is large. Enhanced DL MU-MIMO is an important topic in
Rel-10, which may further rely on non-codebook based pre-coding for inter-user interference (IUI) mitigation. As a
result, the aggregate feedback load may be further increased. Due to the fact that the UL transmission bandwidth is
limited, reducing feedback load while maintaining good DL performance is an important issue that can also improve the
UL rate.
In order to enhance the system throughput, channel dependent scheduling algorithms (e.g., maximum sum rate
scheduling) can be used at the eNB to schedule users. If the UE has high received SNR/CQI, it will have higher
probability to be scheduled. Then feeding back its SNR/CQI to the eNB makes more sense. On the other hand, the UEs
having low SNR/CQI are less likely to be scheduled. Thus feeding back SNR/CQI to the eNB may result in waste of the
UL resource. The purpose of this contribution is to exploit this idea to reduce the aggregate feedback load of the system
while maintaining good DL throughput at the same time.
In the following illustration, we use SNR to represent the channel state information (CSI) of the UEs. The proposed
idea can be easily modified when the feedback is in the form of CQI.

2 Background
Multiple antennas can provide spatial diversity in wireless fading channels to improve the communication quality. On
the other hand, multiple antennas can also create (virtually) multiple channels to increase the system throughput. For the
DL broadcast channel employing multiple antennas, it has been shown recently that dirty paper coding (DPC) [1]
achieves the capacity [2]. However, the capacity achieving scheme of the DL MIMO broadcast channel is difficult to
derive. In addition, the vector DPC for such a scheme has a high encoding/decoding complexity. Thus, several works
resorted to more practical (but suboptimal), space division multiple access (SDMA) based designs. For example, zero-
forcing beamforming (ZF-BF) was shown in [3] to achieve the optimal sum rate growth. However, both DPC and ZF
schemes require perfect CSI feedback from the UEs to the eNB to achieve the optimal performance. This may result in
high feedback load and is not practical.
In [4, 5], a model was proposed to analyze the sum rate loss due to imperfect (quantized) CSI. In the system
considered there, each UE quantizes the channel vector to one of the N = 2 B quantization vectors and feeds back the
codebook index to the eNB to capture the spatial direction and magnitude of the channel. To reduce the feedback load,
orthogonal random beamforming (ORB) [6] can be used. In the ORB, the eNB transmits through orthogonal BF vectors
to the users, and each user only needs to feedback its received SINR on different orthogonal BF vectors for the purpose
of scheduling. In [7], it was shown that the sum rate performance of the ORB exhibits the same growth rate as the DPC
and ZF BF based schemes when the number of users is large.
In this contribution, a feedback reduction method for ORB is proposed. The basic idea is to let the UEs determine
whether their CSIs are good enough to have high probabilities to be scheduled. If not, then feeding back their CSIs (e.g.,
SNR, CQI, etc.) may be wasteful of the UL resource. Thus these UEs will not feedback to reduce the average feedback
load.

3 Multi-threshold Model
The proposed feedback load reduction method for MU-MIMO systems comprises of the following steps.
 Step 1: The eNB uses the UE distribution and channel statistics information to compute the SINR thresholds and
periodically broadcasts these information to all UEs.
 Step 2: The UEs autonomously determine the number of bits used for feeding back their SINR/CQI according to
the multi-thresholds.
 Step 3: The UEs feed back their SINR/CQI to eNB.
 Step 4: The eNB uses the feedbacks from the UEs for scheduling and performing AMC.

The cost of this multiple threshold model is using the PBCH to broadcast the multiple thresholds to the UEs
periodically. According to this model, the reduction of uplink feedback channel can be used to transmit data to utilize
the spectrum and enhance the data rate.

3.1 Examples and Numerical Results

3.1.1 Transmission and Multi-threshold model

Consider a MU-MIMO system where the eNB is equipped with M antennas and there are K UEs each having NR
receive antennas. The N R receive antennas are used for ZF interference suppression. According to the ORB
strategy for multiuser transmission, the eNB uses a pre-coding matrix W = [ w1 , w2 ...wM ] , where wi �C m
M
,i=
1. 2….M, are random orthogonal vectors generated from isotropic distribution [6]. In this example, the multi-threshold
feedback scheme is implemented according to maximum sum rate scheduling. There are many ways to partition SINR
regions (shown in Fig. 1). An example is given as follows.
Let X , X � X ( K ) , be the order statistics of i.i.d. continuous random variables X 1 , X 2� X K , with the
 (1) (2) �
��
� ��
��
�
common PDF in decreasing order, i.e., X ‫������׳‬
X (2) X(K) .
(1)

When UE k has on its i-th direction SINRk ,i = sinrk ,i , the probability that UE k's SINR on the i-th direction is
ranked the p-th among all UEs is
P{ X k = X ( p ) }
P{ X k = X ( p ) | X k = sinrk ,i } =
P{ X k = sinrk ,i }
K-p
( K - 1)!{FSINR (sinrk ,i )} {1 - FSINR (sinrk ,i )} p -1
=
( K - p )!( p - 1)! (1)

where the FSINR (sinrk ,i ) is the CDF of the SINR distribution.

For example, the probability that UE k has the highest SINR among all UEs on the i-th direction is given by
P{ X k = X ( p ) | X k = sinrk ,i } = {FSINR (sinrk ,i )}K -1 . (2)

Eq. (1) satisfies

K

�P{ X
p =1
k = X ( p ) | X k = sinrk ,i } = 1 . (3)

With SINRk ,i = sinrk ,i , UE k can infer its most possible rank among the UEs on the i-th direction as
 rank (sinrk ,i ) = arg max {P{ X k = X ( p ) | X k = sinrk ,i }} (4)
p =1,2... K

To this end, the regions defined in (1) are such that in region j, rank (sinrk ,i ) = j , j=1, 2…, N. The corresponding

thresholds rth , j , j = 1, 2 L N , can thus be determined. Note that all the thresholds can be computed off-line at
the eNB as long as the number of UEs and the fading statistics are known. The values of the thresholds can be
updated and broadcasted by the eNB periodically according to the system configuration. In the following, we
show the numerical results comparing the sum rate performance of different types of feedback schemes.

3.1.2 Single-layer transmission

Fig.1: Multi-threshold model

Only one data stream is supported in Rel.8 for MU-MIMO transmission. Thus, a UE feeds back its maximum SINR
among all beam directions in multi-threshold model. Table I shows the average feedback loads for different feedback
schemes in the comparison. We will refer to the method proposed in this contribution as the t ype-I scheme where
BR = 2 bits are used to represent the region index (i.e., 4 regions). In addition, bi bits are used to quantize the SINR
value in region i. The maximum number of regions is set to 4 to avoid large feedback loads. Using numerical
optimization, the bit assignment for these 4 regions are B = [ b4 , b3 , b2 , b1 ] = [ 0, 0, 0, 4 ] . Therefore, if a UE's SINR is
higher than the threshold for the 4th region ( rth in Fig. 1), it feeds back with BR + bi = 2 or 6 bits. If the SINR is
below rth , the UE does not feedback.

The type-II scheme in Table I emphasizes the resolution in the high SINR region. If the SINR exceeds the rth,1 in Fig.1,
the UE feeds back BII ,Q = 4 bits to the eNB to exploit MUD. Otherwise, the UE does not feedback.

The type-III scheme is the traditional feedback scheme. That is, no matter what the SINR value is, the UE quantizes
SINR using BIII ,Q = 4 bits.

Finally, the type-IV scheme is the one proposed in [8] which uses BIV ,Q = 4 bits to quantize the region ( rth, N , �) in
Fig.1. If a UE's SINR exceeds rth, N in Fig.1, it feeds back BIV ,Q bits. Otherwise it does not feedback.
Fig.2 compares the sum rates of the four schemes with increasing number of UEs. It is shown that even though the
conventional (type-III) scheme has the highest sum rate performance, type-I and type-IV schemes have sum rates very
close to that of the type-III scheme. However, type-I and type-IV schemes have much lower feedback load because they
do not need to feedback all the time.

Fig.2: Comparison of the sum rate performances of different feedback schemes (M=4, N R =1).

Fig.3 compares the aggregate feedback loads of the four schemes with increasing number of UEs. It is shown that the
aggregate feedback load of the conventional (type-III) scheme increases linearly with the number of UEs. On the other
hand, type-I, type-II and type-IV schemes have their aggregate feedback loads quickly saturated. Therefore, for these
schemes, the UL will have much more bandwidth left for data transmission.
 Fig.3: Comparison of the aggregate feedback load of different feedback schemes (M=4, N R =1).

Fig. 4. Sum rate vs. feedback load (M=4, N R =1).

In Fig.4, the results of Fig.2 and Fig.3 are put together to have a clearer comparison. It is shown that overall, the type-I
scheme has a better sum rate vs. feedback load performance.
4 Conclusion
In this contribution we proposed a MU-MIMO feedback reduction scheme. The basic idea is to let the UEs
autonomously determine whether they need to feed back their CSI (SINR, CQI, etc.) according to the probabilities of
them being scheduled. If a UE is less likely to be scheduled, it will not feed back its CSI. Thus the feedback load can be
significantly reduced. Even though the numerical results shown are for maximum rate scheduling, the proposed method
can be easily modified to accommodate other types of scheduling schemes. In addition, as long as the feedback is
derived according to the CSI, the proposed method can be used. The only necessary change is to let the eNB to compute
and broadcast the threshold parameters periodically. Since these parameters are broadcasted, the overhead they need in
the DL should be very limited. However, the UL overhead saved by the feedback reduction is significant.

References
[1] M. H. M. Costa, “Writing on dirty paper” IEEE Trans. Inf. Theory, vol. 29, no. 3, pp. 439-441, 1983.
[2] H. Weingarten, Y. Steinberg, and S. Shamai, “The capacity region of the Gaussian multiple-input multiple-output
broadcast channel,” IEEE Trans. Inf. Theory vol. 52, no. 9, pp. 3936 -3964, Sept. 2006.
[3] T. Yoo and A. Goldsmith, “On the optimality of multi-antenna broadcast scheduling using zero-forcing
beamforming,” IEEE J. Select. Areas Commun., vol. 24, pp. 528-541, Mar.2006.
[4] N. Jindal, “MIMO Broadcast Channels with Finite-Rate Feedback,” IEEE Trans. Inf. Theory, vol. 52, no. 11, pp.
5045-5060, Nov.2006.
[5] N. Jindal T. Yoo and A. Goldsmith, “Multi-Antenna Downlink Channels with Limited Feedback and User
Selection,” IEEE J. Select. Areas Commun., vol. 25, no. 7, pp. 1478-1491, Sep.2007.
[6] T.L. Marzetta and B.M. Hochwald, “Capacity of a mobile multiple-antenna communication link in Rayleigh at
fading," IEEE Trans. Inf. Theory, vol. 45, pp. 139-157, 1999.
[7] M. Sharif and B.Hassibi, “On the capacity of MIMO broadcast channel with partial side information,” IEEE Trans.
Inf. Theory, vol. 51, no. 2, pp. 506-522, Feb. 2005.
[8] D. Gesbert and M.S. Alouini, “How much feedback is multi-user diversity really worth?,” IEEE Int. Conf. on
Commun. (ICC), vol. 1, pp. 234-238, June 2004.
[9] D. Gesbert V. Hassel, M.S. Alouini and G.E. Oien, “Exploiting multi-user diversity using multiple feedback
thresholds,” IEEE VTC., vol. 2, pp. 1302-1306, May 2005.
[10] R. Zakhour and D. Gesbert, “A Two-Stage Approach to Feedback Design in Multi-User MIMO Channels with
Limited Channel State Information," IEEE Int. Symp. on Pers., Indoor and Mobile Radio Commun., Sep. 2007.
[11] S. Ghahramani, “Fundamentals of probability with stochastic processes,” Prentice-Hall, third edition, 2005.
[12] R1-094535, “Consideration on MU-MIMO feedback reduction”, HTC, 3GPP TSG RAN WG1 #59, Jeju, Korea,
November 9 – 13, 2009
[13] R1-104161, “Feedback reduction mechanism for MU MIMO”, HTC, 3GPP TSG RAN WG1 #61bis, Dresden,
Germany, June 28 – July 2, 2010