Professional Documents
Culture Documents
博士論文
Graduate Institute of Communication Engineering
College of Electrical Engineering and Computer Science
National Taiwan University
Doctoral Dissertation
多輸入多輸出天線廣播通道中降低迴授量
李晉豪
Jin-Hao Li
指導教授:蘇炫榮 博士
Advisor: Hsuan-Jung Su, Ph.D.
中華民國 102 年 10 月
October, 2013
國 立 臺 灣 大 學 博 士 學 位 論 文
口試委員會客定書
多 天 線 輸入 輸 出 廣 播 通 道 下 , 降低回 授量
Feedback Reduction in MIMO Broadcast
Channel
本論文係李晉豪君(學號D
979
420
24)在國立臺灣大學電信工
程學研究所完成之博士學位論文,於民國1 02
年9月1
8日承下列考
試委員審查通過及口試及格,特此證明
口 試委員:
( 指導 教授 )
二至侈逐+
所 長 (簽名)
致謝
時間過的很快,不知不覺已經在實驗室待了五個年頭,在此,感謝在這不算
短的日子裡,曾經一同奮鬥,一起經歷研究,計畫,與幫助我的人,在此分享謝
謝你們,也想與你們分享我的喜悅。
首先,謝謝一路上支持我的家人,沒有他們一直以來的幫助,支持,讓我可以
全心全力的專注於學業上,無需為金錢上煩惱,也謝謝家人帶給我的家庭溫暖,
讓我在遇到研究瓶頸時,有一個溫暖的避風港,一個可以放鬆的家。
謝謝我的指導教授,蘇炫榮老師,在研究上給予我許多的幫助,給予我很大的
空間,可以自由的發揮,選擇自己喜愛的研究題目,並且在遇到困難時,提供許
多的建議來解決問題,此外,也謝謝老師提供一個那麼棒的研究環境,提供最好
的設備,給予最大的自由,讓大家可以很放鬆的在實驗室,一起討論,努力,很
榮幸可以加入老師的實驗室。
在這五年的時間,很開心有一群熱心的學長提供意見,一起做計畫,謝謝林秉
勳學長,對於研究的方向與內容,提供許多的想法,謝謝廖偉舜學長,一起為了
計畫努力,也給予我許多的幫助,第一次一起去舊金山,開車去 outlet 的時光,
一輩子都難忘,謝謝不承認來自交大的李崇丕學長,一起吐研究的苦水,謝謝已
經畢業很天才的徐博,跟你討論,一起去土耳其伊斯坦堡報告,承蒙您照顧了。
謝謝冠文學弟,麻煩你一肩扛起了計畫的重責,太辛苦拉。還有有一群很會搞笑
的碩班學弟妹,一起修課,一起為計畫努力,一起嘴炮聊天,跟你們在一起,都
覺得自己還很年輕,帶給我許多的歡樂。
最後,要謝謝我的老婆,在這最後的一年,陪我經歷了面試,找工作,陪我等
待論文的審核,經歷了許多的困難,謝謝你在我身邊。礙於篇幅,還有需多沒列
出來的朋友,謝謝你們這些日子的鼓勵與支持,謝謝。
李晉豪
10 月/2013
博理館 508 室
摘要
在多使用者多傳輸天線傳輸多輸入天線下行廣播通道中,資料的傳輸速率可以透
過基地台的增加天線數量與排程機制來大量的增加。當使用者落在基地台的廣播範
圍,所有的使用者都收到來自於基地台的相同的資料。針對於多使用者多傳輸天線
與多接收天線系統,基地台具有 Mt 根傳輸天線,K 總共有 K 個使用者在此基地台涵
蓋範圍內且每個使用者具有 Mr 根接收天線,所有的多工增益 Mt 可以透過空間多重存
取技術或是迫零束波成形技術來達到。此外,當使用者數目遠超過傳送天線數目
K>>Mt,由於多使用者分散增益,速率總和會呈現 MtloglogK 的成長趨勢。然而,上
述的結果都來自於傳輸端具有所有的通道消息,因此,通道迴授量會隨著使用者與
天線數量線性的增加。另一方面,透過選擇適當的調變與編碼來達到可靠的傳輸,
通道消息是非常重要的,因此,對網路系統而言,迴授量是非常大的。
此論文主要的貢獻在於考慮不同的系統應用,例如,多使用者多傳輸天線多接收
天線廣播通道,與多點廣播通道,與不同系統的排程演算法,提出降低迴授量的策
略來降低無線資源的浪費。透過分析,降低迴授量的方法,透過推導的多個閥值來
執行,數個閥值是藉由序量統計的觀念配合排成演算法來得到。模擬結果可知,在
廣播通道中,迴授量可以大量的被降低並且幾乎完美的速率總和都可以被達到。對
於多點廣播通道,迴授量也可以大量的被降低並且解碼成功的使用者的數量幾乎與
傳輸端具有完美的通道消息的情況是一樣的。此外,新的排程方法配合多閥值技術
也被提出,在異直性質瑞雷衰減通道中,不同使用者被基地台挑選到的的公平性,
迴授的量,與系統的速率和的關係也被討論。
關鍵字 – 降低迴授;多用戶分散增益;束波形成技術
Abstract
from the BS. For the multiuser MIMO system with Mt transmit antennas
and K users equipped with Mr antennas, the full multiplexing gain Mt can
grows like Mt log log K due to multiuser diversity. However, all these results
are based on the assumptions of full channel state information (CSI) at the
transmitter, thus, the feedback load linearly increases with number of users
and number of antennas. On the other hand, CSI feedback is also very
almost full sum rate performance can be achieved in broadcast channel. For
multicast scenario, the feedback load can be reduced dramatically and num-
ber of decoded user is almost the same as the full CSI case. Moreover, a new
2
Contents
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Overview of Thesis . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
gions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.3 Multiuser Diversity Using the Multi-threshold Scheme . 32
2.3 Bit Allocation and Feedback Load Analysis . . . . . . . . . . . 34
2.3.1 Optimal Bit Allocation with Given Thresholds . . . . . 35
2.3.2 Fast Bit Allocation Method . . . . . . . . . . . . . . . 37
2.3.3 Complexity Comparison . . . . . . . . . . . . . . . . . 42
2.3.4 Feedback Load Analysis . . . . . . . . . . . . . . . . . 42
2.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . 43
i
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
ii
5 Feedback Policies for Heterogeneous Rayleigh Fading Chan-
nel with Finite Feedback 85
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7 Conclusions 107
iii
C Proof of Theorem 3 113
Bibliography 120
iv
List of Figures
K = 20 users. . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Sum rate loss versus the number of users. △RP (K) = 0.25
v
2.11 Sum rate as a function of the feedback load. BQ,A = BQ,B =
BQ = BQ,D = 5, ρ = 10 dB, Mt = 2, Mr = 4. . . . . . . . . . . 50
2.12 Sum rate as a function of the feedback load for different trans-
mission powers (ρ). BQ,A = BQ,B = BQ = BQ,D = 5, Mt =
4, Mr = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
vi
6.2 Fairness comparison between different number of feedback re-
gions (P = 10W). . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.3 Sum rate vs. feedback load and fairness vs. feedback load
with different number of regions (P = 10W ). . . . . . . . . . 105
vii
List of Tables
viii
List of Abbreviations
BS Base Station
ix
MBS Multicast Broadcast Service
x
Chapter 1
Introduction
1.1 Background
the demand for the high data rate and the quality of service (Qos). The
and video telephony, each with different Qos constraint. Due to different Qos
requirements for different services, the network considers the tolerable delay,
the traffic load of the system, system fairness and channel conditions of users
to allocate the resources to the users. Through the scheduler, the multiuser
diversity is exploited through the better utilization of the spectrum [1].
1
Qos.
Recently, there has been a popular interest in the role of multiple an-
tennas for the multiuser network settings, especially in broadcast channel
ference and attain the diversity and/or multiplexing gains. Both MIMO and
scheduling techniques rely on channel state information (CSI) feedback. For
the multicast channel scenario, all users under the cell coverage receive the
same data which is sent from the BS. In this case, CSI feedback is important
been shown recently that dirty paper coding (DPC) [4] achieves the capac-
ity [5]. However, this capacity achieving scheme is difficult to derive and
the more practical (but suboptimal) space division multiple access (SDMA)
2
by the BS on the common radio resources, and hence the efficiency of radio
resources is highly improved and the quality of multimedia service is then
increased. Different from the MIMO broadcast channel scenario, the main
purpose of MBMS system is to provide the reliable transmission instead of
achieving high data rate. Due to the remarkable enhancement on efficiency
of radio resources and multimedia quality, MBMS has become one of the
important features in next generation communication standards, such as the
users in the MBMS network have to feedback its’ current channel condition
to the BS, so that the BS can select a proper modulation and coding scheme
(MCS) for starting the broadcast transmission to all the MBMS users. For
this sake, the feedback load will be increased dramatically with the num-
ber of the users in the MBMS network and hence the spectrum efficiency is
3
The dominant role of CSI at the transmitter in multiuser MIMO sys-
tems can be identified by the asymptotic sum rate growth under different
limited feedback model [9], each user feeds back B bits to represent the quan-
user selection method [10] and ZF precoder are used, the multiuser diversity
random beamforming scheme [11]. In this scheme, each user needs to feed
back the best unquantized SINR among all directions and the corresponding
beam index. The full multiuser diversity is also achieved. In [12], only one
bit feedback per user for the ORB scheme exhibits the same growth rate as
the ORB scheme with perfect SINR. However, the throughput may degrade
a lot due to large quantization error. Note that the above results are based
on the assumption that the receiver has the perfect CSI. The feedback load
of limited feedback schemes increases with number of users dramatically.
There are other previous works that sought to reduce the feedback load at
the scheduling stage. In [13], a threshold was set according to the scheduling
outage probability such that a user does not need to feedback when its CSI is
below the threshold. This method reduces the system feedback load without
4
affecting the scheduling performance much. In [14], multiple thresholds are
set, and the scheduler utilizes a polling process to select the best feedback
ZF-BF performance in the second stage. The drawback of this scheme is also
since the time varying channel may cause the unexpected transmission error.
the transmitter does not receive an acknowledgment before the timeout, the
transmitter will retransmit the data until the transmitter receives an acknowl-
many receivers are located in a cell, this ARQ mechanism can cause a large
number of ARQ feedback messages (ACKs or NACKs) to be sent to the trans-
mitter at the same time, which is referred to the feedback implosion problem.
The previous works investigate this problem as shown in [16, 17]. However,
these articles only focus on the ACK/NACK feedback instead of CQI feed-
back. Through the selection MCS and resource allocation for MBMS, the
reliable communication can be also achieved. The scalable coding method
is proposed in [18] to manage the radio resources. In [19], the resources are
5
dynamically allocated to MBMS according to the channel knowledge of users
to maximize the system rate while guaranteeing the system coverage. With
CSI of users, dynamic resource allocation is feasible strategy to make full use
of frequency-selective attenuation to achieve higher performance. In [20], an
analytical model for the coverage and capacity estimation of MBMS over a
Single Frequency Network (MBSFN) and investigation of the best cell radius
and the best MCS to reach a certain spectral efficiency target are provided.
Four different approaches for the efficient selection of the appropriate MCS
in MBSFN transmissions are presented in [21]. Due to the nature of time
varying channels, the MCS has to be adjusted to the weakest user of the
multicasting group [19]. So, each user needs to feedback the CSI to the BS.
back reduction method is proposed in [22]. In this method, the user makes
a decision whether it reports the CQI or not according to the recent frame
error rate (FER). However, the user can not meet the Qos constraint when
It is clear that the feedback load of multiuser MIMO downlink systems can
be reduced if the scheduling mechanism is taken into consideration. However,
most existing works only exploit the scheduling mechanism to moderate the
amount of feedback, but not incorporate the properties of scheduling into the
CSI quantization design. In view of this, in this thesis we propose to reduce
the feedback load by incorporating the scheduling mechanism in both the
determination of the amount to feedback and the CSI quantization. Note
that the required information for the proposed scheme to operate, such as
6
the signal-to-interference-noise ratio (SINR) or signal-to-noise ratio (SNR)
statistics and the number of users, are usually known at the BS. Thus, in
practice, the BS can compute the region thresholds and broadcast to the users
periodically, or broadcast the parameters of the SINR/SNR statistics and the
number of users periodically to the users to let them derive the thresholds.
In practical systems where the number of users is large, the threshold values
will be insensitive to the number of users, and may not have to be updated
frequently. Note that the scheme in [13] also has its threshold dependent on
the channel statistics and the number of users, if a certain scheduling outage
tional fair scheduling achieves strict fairness among different users but this
fairness comes at the cost of a significant sum rate penalty. The perfor-
fair scheduling and greedy SNR scheduling are discussed in [23]. In [24], a
hybrid multiuser scheduler combines the greedy SNR scheduling and nor-
malized SNR scheduling. The BS divides all active users into n groups. The
BS selects one user per group based on either the best SNR or normalized
SNR criterion and then the BS chooses one user form n selected users based
on normalized SNR or best SNR criterion. Hence, the flexible tradeoff be-
tween the sum rate and the fairness among users can be adjusted through the
number of groups n. However, the concept of feedback load reduction does
not take into account in above scheduling methods. Thus, a new scheduler
7
is motivated and proposed to jointly consider the feedback load reduction,
the fairness among users and the sum rate to achieve efficient and flexible
system.
used to divide the range of SINR into multiple regions to reduce the feed-
back load. The minimum number of regions required to meet the sum rate
ORB scheme is limited by the inter-user interference, thus the ZF-BF using
However, the exact CSI is unknown either the BS nor users, the approximated
sum rate is formulated. Then, the multiple thresholds are derived to avoid
8
by the proposed low complexity derivative algorithm. Simulations show that
the sum rate performance of proposed algorithm is close to the sum rate
6.2.1. Through simulations, the flexible tradeoff among the fairness among
users, the sum rate and the feedback load can be achieved.
1.3 Notations
Vectors and matrices are denoted by boldface lower case and capital let-
XT (xT ) stands for the transpose of matrix X (vector x), and X∗ (x∗ ) stands
for the conjugate transpose of matrix X (vector x). Moreover, X† denotes the
pseudo-inverse X∗ (XX∗ )−1 . The function ⌈x⌉ represents the smallest integer
≥ x. ⌊x⌋ denotes the largest integer equal to or less than x. log and ln are
the logarithms with base 2 and e, respectively. The function f (n) = O(g(n))
f (n)
denotes that lim ≤ α, where α is a positive constant independent of
n→∞ g(n)
n.
9
Chapter 2
the sum rate loss due to imperfect (quantized) CSI. In the system considered
there, each user quantizes the channel vector to one of the 2BD quantization
vectors and feeds back the codebook index using BD bits to the BS to capture
the scheduling stage. In [13], a threshold was set according to the scheduling
10
outage probability such that a user does not need to feedback when its CSI is
below the threshold. This method reduces the system feedback load without
the ZF-BF systems through two-stage feedback. In the first stage, each user
feeds back the coarsely quantized version of its CSI, thus the BS has some
to the scheduled users and asks them to feedback finer CSI to achieve good
ZF-BF performance in the second stage. The drawback of this scheme is also
vectors to the users, and each user only needs to feedback its received SINR or
ing. However, the aggregated feedback load increases linearly with respect
11
tized result of that region. When a user’s SNR/SINR is in the lowest region
(that is, below the lowest threshold), the user doest not feedback at all. This
ple, the proposed scheme with four regions is good enough to keep the sum
rate loss smaller than 0.25 bps/Hz when the number of antennas at both the
transmitter and the receivers is four and the number of users is less than 500.
In addition, the aggregate feedback load and the multiuser diversity using
the proposed scheme are also investigated. Our theoretical analysis shows
loads increase with the number of users, with a given number of regions,
the number of users. Moreover, while keeping the aggregate feedback load
constant, the proposed scheme almost achieves the optimal asymptotic sum
rate scaling with respect to the number of users (that is, the multiuser di-
versity). Specifically, given the number of regions, the proposed scheme can
achieve a constant portion of the optimal sum rate achievable only by the
system in which all the users always feedback, and the sum rate loss decreases
exponentially to zero as the number of regions increases. This is a strong
evidence that the proposed scheme opportunistically exploits the scheduling
gain (equivalently, the multiuser diversity) for the feedback design, unlike
the schemes in [25,26] which do not take advantage of scheduling and require
12
all users to feedback, and the scheme in [13] which does not exploit order
statistics and has the aggregate feedback load increasing with the number of
sum rate performances with the complex optimal bit allocation method and
with the fast algorithm are almost the same.
Note that the required information for the proposed scheme to operate,
such as the SNR/SINR statistics and the number of users, are usually known
at the BS. Thus, in practice, the BS can compute the region thresholds
SNR/SINR statistics and the number of users periodically to the users to let
them derive the thresholds. As will be shown later, the threshold values are
ing. In practical systems where the number of users is large, the threshold
values will be insensitive to the number of users, and may not have to be
updated frequently. Note that the scheme in [13] also has its threshold depen-
dent on the channel statistics and the number of users, if a certain scheduling
outage probability is desired. Thus periodic update of the threshold is also
necessary for the scheme in [13].
This chapter is organized as follows. Section 2.1 describes the system
model and briefs the order statistics. Section 2.2 introduces the proposed
feedback scheme and analyzes its sum rate loss and multiuser diversity. In
13
Section 2.3, the bit allocation problem is discussed along with the feedback
load analysis. We then give the simulation results in Section 2.4 and conclude
each user has Mr receive antennas. We consider a full buffer traffic model,
that is, each user always has data in the buffer to transmit. According
schedule a user on each beam [27]. The received signal at the k-th user can
yk = Hk Ws + nk , (2.1)
vector at the k-th user. The entries of Hk and nk are assumed to be in-
dependent identically distributed (i.i.d.) complex Gaussian with zero mean
and unit variance. In addition, the channel matrices for different users are
assumed to be independent.
Note that in this work we consider only identical channel distributions
for the users for the simplicity of demonstrating the idea. This assumption
14
can also be used to model the users’ instantaneous SNRs normalized by
their respective average SNRs for systems with different channel statistics
for the users, but using the SNR based proportionally fair scheduling [28]
which schedules users based on their normalized instantaneous SNRs. The
more practical situations where the users have different channel statistics or
distributions are more intricate, and are discussed in [29].
The vector s = [s1 , s2 , . . . , sMt ]T is the Mt × 1 vector of the transmitted
the equal power assumption of the ORB, each beam is equally allocated with
power ρ = Pt /Mt .
The statistical distribution of the received SINR in this situation was studied
in [11]. We define SIN Rm,k as the SINR of the m-th signal sm at the k-th
Therefore, the received SNR of the m-th signal sm at the k-th user with the
ZF receiver is given by
ρ
SN Rm,k = (2.3)
[((Hk W)∗ (H k W))
−1 ]
m
15
Feedback channel
User
1
Spatial
Scheduling User
division
Algorithm 2
multiple
.. . .. . .
.. . .. . .
(Maximum
access
sum rate)
.. . .. . .
(SDMA)
16
of freedom. For notational simplicity, we let
SIN Rm,k , Mr < Mt
Xm,k = .
SN Rm,k , Mr ≥ Mt
Then the probability density function (PDF) of Xm,k for x > 0 can be sum-
marized as [11] [30]
2Mr −1 (2Mr −1)
−x ∑ (Mr (Mt − 1) + i − 1)!
x2Mr −2 e ρ i
2Mr −i−1
i=0 (ρ) (1 + x)Mr (Mt −1)+i
fXm,k (x) = , Mr < Mt .
(2Mr − 1)!(Mr Mt − Mr − 1)!
−x
eρ
( xρ )Mr −Mt , Mr ≥ Mt
ρ(Mr − Mt )!
(2.4)
by
FXm,k (x) =
2Mr −1 ∫ (2M −1) −x
∑ x r
x2Mr −2 e ρ (Mr (Mt − 1) + i − 1)!
i
dx
i=0 0 (ρ)2Mr −i−1 (1 + x)Mr (Mt −1)+i
, Mr < M t ,
(2Mr − 1)!(Mr Mt − Mr − 1)!
− Mt + 1, xρ )
Γ(M r
1− , Mr ≥ Mt
(Mr − Mt )!
(2.5)
∫∞
where Γ(a, x) = x
ta−1 e−t dt is the upper incomplete gamma function. Ac-
cording to (2.4), in the special case when the transmitter and the receiver
have the same number of antennas, Xm,k has an exponential distribution
with parameter 1/ρ.
With the maximum sum rate scheduling, on each beam direction, the
BS selects, among the users who have fed back their CSIs, the user that
17
has the best channel to transmit to. If none of the users has fed back the
CSI, the BS randomly selects one user to transmit to. Due to the symmetric
property, we drop the beam index m of Xm,k , and let Xk represent the SNR or
SINR (depending on the antenna configuration) of user k for a certain beam.
K K K
Let X(1) , X(2) , . . . , X(K) be the order statistics of i.i.d. continuous random
variables X1 , X2 , . . . , XK with the common PDF (2.4) in decreasing order,
K
i.e., X(1) ≥ X(2)
K
≥ · · · ≥ X(K)
K K
. The CDF and PDF of X(j) , respectively, are
given by [31]:
∑
K
K!{FXm,j (x)}i {1 − FXm,j (x)}K−i
FX(j)
K (x) = , −∞ < x < ∞, (2.6)
(i)!(K − i)!
i=K−(j−1)
With the order statistics, the sum rate using the maximum sum rate schedul-
ing algorithm can be computed. As a simple example, if every user has the
BS, and the feedback events are independent of the value of the CSI and
independent from user to user and from beam to beam, the sum rate can be
obtained by
(
∑
K
K!Pf n (1 − Pf )K−n
R(K, Pf , Mt ) = Mt E n
log(1 + X(1) )
n!(K − n)!
n=1
)
+ (1 − Pf )K log(1 + Xk ) , (2.8)
where the first term accounts for the rate using the maximum sum rate
scheduling algorithm when at least one user has fed back the CSI, and the
second term is the rate when no user has fed back to the BS and the BS
randomly schedules one user k on a ceratin beam.
18
2.2 The Multi-threshold Feedback Scheme
Mr ≥ Mt case but also the SINR in the Mr < Mt case. For the scheme
in [13], if the SNR of a user is greater than the outage threshold, the user
feeds back BQ bits to represent the received SNR. Otherwise it does not
feedback. The threshold is derived according to a pre-determined scheduling
outage probability (where “scheduling outage” refers to the situation when
none of the users feeds back), but not directly related to the scheduling
mechanism. Since the maximum sum rate scheduler selects users according
The basic idea of our proposed scheme is to let a user compare its received
SNR with the thresholds derived from the order statistics. The user can
thus guess its most possible rank among all the users, and, if its rank is
high enough to make its chance to be scheduled high, it feeds back its SNR.
Otherwise the user does not feedback in order to save the reverse link resource
and avoid interfering the other users’ reverse link transmission. Note that
there might be errors in the statistical inference by the individual users about
their SNR ranks. These errors may result in the situation where the users who
actually have high SNRs do not feedback, and the BS does not have proper
users to select from. To make up for the sum rate loss due to this situation, we
allow the users with several (guessed) ranks to feedback. Therefore, for each
beam direction, a set of N thresholds Rth = {rth,1 , rth,2 , . . . , rth,N } is set (see
19
Fig. 2.2) according to the order statistics of the received SNR. Let rth,0 = ∞.
For the SNR region i bounded by the adjacent thresholds as [rth,i , rth,i−1 ), bi
additional quantization bits are used to help the BS differentiate users whose
SNRs fall in that region, and also make better link adaptation if a user in
that region is scheduled. According to the importance of the SNR regions
to the sum rate, bi , i = 1, . . . , N , (to be optimized later) are usually in non-
increasing order. When the received SNR is higher than rth,N , the user feeds
back its rank and the additional quantization bits. Otherwise the user does
not feedback at all.
value) and the users have i.i.d. SNR distributions, the probability that user
k’s SNR on the m-th beam direction is ranked the p-th among all the users
is
(
(K − 1)!
P {Xk = X(p)
K
|Xk = snrm,k } =
(K − p)!(p − 1)!
)
×{FXm,k (snrm,k )}K−p {1 − FXm,k (snrm,k )}p−1 (2.9)
and satisfies
∑
K
P {Xk = X(p)
K
|Xk = snrm,k } = 1, (2.10)
p=1
20
Region Region Region
N 2 1
0 bN
.. . ... b2 b1
SNR
rth,N rth,N-1 rth,2 rth,1
snrm,k which is the highest SNR among all the users on the m-th beam is
P {X(1)
K
= snrm,k }
P {Xk = K
X(1) |Xk = snrm,k } =
P {Xk = snrm,k }
( )K−1
= FXm,k (snrm,k ) . (2.11)
With SN Rm,k = snrm,k , user k can infer its most possible rank (i.e., the
rank with the highest probability (2.9)) among the users on the m-th beam,
denoted rank(snrm,k ), as
From (2.11), it is clear that when snrm,k is sufficiently high, user k’s most
possible rank rank(snrm,k ) will be 1. It will be verified later that as snrm,k
decreases, rank(snrm,k ) will gradually increase.
Fig. 2.3 shows the rank probabilities (2.9) for different ranks over a wide
range of snrm,k for a system with Mr = Mt and K = 20 users. It can be
seen that as as snrm,k decreases, user k’s most possible rank rank(snrm,k )
increases, and each value of rank(snrm,k ) occurs in a continuous region of
snrm,k . To this end, we define the regions in Fig. 2.2 according to the most
possible rank.
21
Mr=Mt, K=20
1
0.9 p=1
0.8 p=2
|Xk=snrm,k)
p=3
0.7
p=4
0.6
p=5
0.5 p=6
(p)
P(Xk=XK
0.4 p=7
0.3
0.2
0.1
0
0 1 2 3 4 5 6 7 8
snrm,k
Figure 2.3: The probability (2.9) that user k’s SNR on the m-th beam direc-
tion is ranked the p-th among all the users. Mr = Mt , K = 20 users.
22
Definition 1. The SNR region j denoted [rth,j , rth,j−1 ) satisfies that ∀snrm,k ∈
[rth,j , rth,j−1 ), rank(snrm,k ) = j, where rank(snrm,k ) is defined in (2.12).
Due to the symmetric assumption that the users suffer i.i.d. Rayleigh
fading processes, also due to the ORB, the same set of thresholds applies to
all users and all beam directions. By Definition 1, rank(snrm,k ) = j when
snrm,k ∈ [rth,j , rth,j−1 ), and rank(snrm,k ) = j+1 when snrm,k ∈ [rth,j+1 , rth,j ).
P {Xk = X(j)
K
|Xk = rth,j } = P {Xk = X(j+1)
K
|Xk = rth,j } (2.13)
FXm,k (rth,j ) 1−FXm,k (rth,j )
⇔ K−j
= j
⇒ FXm,k (rth,j ) = 1 − j
K
, j = 1, 2, . . . , N. (2.14)
This result shows that rth,1 > rth,2 > · · · > rth,N , which, together with Defi-
nition 1, verifies that as snrm,k decreases, rank(snrm,k ) gradually increases.
( )
K
rth,j = ρ ln . (2.15)
j
It can be observed that the value of rth,j is insensitive to the number of users
23
can still be used to numerically search for the thresholds. From (2.14) and
the nondecreasing property of a CDF, we have
j
FXm,k (x) ≥ 1 − , when x > rth,j ,
K
j
FXm,k (x) ≤ 1 − , when x < rth,j , (2.16)
K
P {Xk = X(j)
K
|Xk = x} ≥ P {Xk = X(j+1)
K
|Xk = x}, when x > rth,j ,
P {Xk = X(j)
K
|Xk = x} ≤ P {Xk = X(j+1)
K
|Xk = x}, when x < rth,j . (2.17)
each threshold rth,j found can serve as the initial upper boundary (and zero
is an obvious lower boundary) for the search of the next threshold rth,j+1 .
1
P {Xk ∈ [rth,j , rth,j−1 )} = , j = 1, 2, . . . , N. (2.18)
K
In other words, the probability for a user to infer itself as being ranked the
1
j-th place on a certain beam direction is K
, for all j = 1, 2, . . . , N , with
N ≤ K. This is very intuitive because each user has the same probability
1
K
of being ranked the j-th place, j = 1, 2, . . . , K, due to the symmetric
assumption of the users’ SNR distributions. Therefore, the probability Pf
for a user to feedback is
N
Pf = 1 − FXm,k (rth,N ) = , (2.19)
K
Note that the thresholds can be computed off-line as long as the number
of users and the channel statistics are known. In practice when the network
24
operating conditions vary with time, the values of the thresholds can be
updated periodically by the BS according to the system configuration and
channel statistics, and broadcasted to the users. Or, the users can compute
the threshold values by themselves after receiving the information of network
configuration and channel statistics broadcasted by the BS.
With the thresholds set, if a user finds its SNR on a beam lower than
rth,N , then no feedback is sent for that beam. Otherwise, the user feeds
back BR = ⌈log2 (N )⌉ bits to indicate its most possible rank on that beam.
In order to account for the situation where there are more than one users
reporting to have the same rank, each region j is further quantized with bj
bits which are also fed back together with the “rank” bits.
We now consider the sum rate loss incurred by the reduction of feedback.
The sum rate loss of the proposed scheme can be attributed to two sources:
one resulting from the situation when none of the users feeds back, and the
other being the feedback quantization error. In general, when the number
of regions increases, which results in higher feedback load, the sum rate loss
decreases. Thus, for a given tolerable sum rate loss, we should apply the
25
First, we will analyze the sum rate loss incurred when none of the users
feeds back, and determine the minimum number of regions N required to
meet the given tolerable sum rate loss of the system without considering the
quantization in each region.
Then, in Section 2.3, quantization of the regions will be considered by
allocating bits for quantizing the regions to maximize the sum rate.
Definition 2. A rate loss event occurs when all users’ SNRs on a certain
beam is smaller than the lowest threshold rth,N . (Note: In [13], the rate loss
event is called the scheduling outage event.)
When a rate loss event occurs, the BS will randomly schedule one user.
feed back is
△R(K, N, Mt )
{ }
= Mt E log(1 + X(1)
K
) − log(1 + Xk ) | X(1)
K
< rth,N PL (K) (2.21)
{ ( ( K )) }
≤ Mt E log 1 + X(1) − Xk | X(1) K
< rth,N PL (K)
( {( K ) })
≤ Mt log 1 + E X(1) − Xk | X(1)
K
< rth,N PL (K)
where the inequalities are due to the convexity of the rate function and
Jensen’s inequality. By using (2.5), (2.14), and considering the Mr ≥ Mt
26
case as an example,
{( K ) K }
E X(1) − Xk |X(1) < rth,N
∫ rth,N ∫ rth,N
K−1
xKfXm,k (x)FXm,k (x) dx xfXm,k (x)dx
= 0 ( )K − 0 ( ) (2.23)
1− K N 1− K
N
∫ rth,N [ Mr ∑
−Mt +1 ]K−1
Mr −Mt +1 −x/ρ −x/ρ (x/ρ)m
K[(Mr − Mt )!] K−2
x e 1−e dx
0 m=0
m!
= ( )K
ρMr −Mt +1 1 − N
K
[ Mr ∑
−Mt +1 ]
−rth,N /ρ (rth,N )k
ρ(Mr − Mt + 1) 1 − e
ρk k!
− ( ) k=0
.
1− N
K
Using (2.22) and (2.23), the minimum number of regions required for a given
tolerable sum rate loss △RP (K) can be approximated by the minimum N
( ) ∆RP (K)
that satisfies E (X(1) − Xk ) | X(1) < rth,N < 2 Mt PL (K) − 1.
K K
The following theorem gives the relationship between the sum rate loss
Theorem 1. Given Mt and N , the sum rate loss △R(K, N, Mt ) for the
Proof. In Appendix A.
compares the sum rate loss upper bound with the actual sum rate loss ob-
tained by simulation, and shows that the sum rate loss upper bound (2.22)
27
is tight (within 0.1 bps/Hz) when the number of regions is large. This fig-
ure also shows that the required number of thresholds is insensitive to the
number of users. Four regions are enough to keep the sum rate loss smaller
than the tolerable sum rate loss △RP (K) = 0.25 bps/Hz when the number
of users is up to 500. Using the sum rate loss upper bound to design the
system, we can see that the number of regions can be reduced to three when
the number of users is smaller than 16. When the number of users is smaller
28
0.7
∆R(K,N=3,Mt)
0.6
∆R(K,N=4,Mt)
∆RU(K,N=3,Mt)
0.5
∆RU(K,N=4,Mt)
∆RP(K)=0.25 bps/Hz)
Sum rate loss
0.4
0.3
0.2
0.1
0
0 50 100 150 200 250 300 350 400 450 500
Number of users
Figure 2.4: Sum rate loss versus the number of users. △RP (K) = 0.25
bps/Hz is the tolerable sum rate loss. ρ = 10 dB, Mt = Mr = 4.
29
Fig. 2.5 shows the simulation results of the sum rate for different numbers
of regions when the number of users increases. In this figure, the “full CSI”
case refers to the ORB system in which the users always feedback their SNRs
on all beams. It is shown that the sum rate increases with both the number
of users and the number of regions.
Under the ORB and maximum sum rate scheduling, when the number
of regions is larger than four, the sum rate achieved by the multi-threshold
scheme is very close to the sum rate with full CSI feedback (within 0.25
bps/Hz according to Fig. 2.4). An intuitive interpretation of this result is
that, on average, providing four users with the highest SNRs on each beam
30
16
15
14
Sum rate
13
Full CSI
12
N=1
N=2
N=3
11
N=4
N=5
N=6
10
0 20 40 60 80 100 120 140 160 180 200
Number of users
Figure 2.5: Sum rate performance of the multi-threshold scheme with differ-
ent numbers of regions. ρ = 10 dB, Mt = Mr = 4.
31
2.2.3 Multiuser Diversity Using the Multi-threshold
Scheme
property.
FX−1m,k (1 − N
K
). For both Mr ≥ Mt and Mr < Mt , the lower bound of achiev-
satisfies
RL (K, N, Mt )
lim = 1 − e−N .
K→∞ Mt log log K
Proof. In Appendix B.
32
Proof. In Appendix C.
The above two theorems show that with given Mt , ρ and N , when the
number of users K is large, the multi-threshold feedback scheme can achieve a
sum rate which scales like (1−e−N )Mt log log(K). In other words, this scheme
can asymptotically achieve a constant portion (1 − e−N ) of the optimal sum
rate Mt log log(K) achievable with full CSI feedback. The remaining portion,
i.e., the sum rate loss, decreases exponentially to zero as the number of
regions N increases. In fact, e−N equals the probability of the rate loss event
PL (K), given in (2.20), when the number of users is large. This result can
be observed from Fig. 2.5 where it is shown that the sum rate loss is already
The asymptotic upper bound of the sum rate loss, e−N Mt log log(K), also
wide range of the number of users as shown in Fig. 2.4. Note that asymptot-
ically the sum rate loss decreases exponentially with N and increases with
log log(K). According to Fig. 2.4, as the minimum required N increases from
3 to 4 (when K becomes larger than 16), the amount of sum rate loss this
reduces is quite large. This sum rate loss reduction will not be depleted until
K is very large.
33
2.3 Bit Allocation and Feedback Load Anal-
ysis
CSI values with finite numbers of bits. Assume that the users use BR bits to
represent the region information, and additional bj bits to quantize the SNR
when it falls in region j. On a given beam, whenever there are other users
feeding back the same rank indication and the same additional quantized bits
as the user who actually has the highest SNR, the BS will randomly schedule
one of them. As a result, the lowest possible rate due to this ambiguity
in scheduling will be the rate derived from the lower boundary of the SNR
each region. The sum rate Rq (B) with both rank and SNR quantization
where rth,j,1 , rth,j,2 , . . . , rth,j,2bj +1 are the quantization levels in rank region
j, with rth,j,1 = rth,j , rth,j,2bj +1 = rth,j−1 . In Fig. 2.6, the analytical lower
bound of the sum rate (2.25) is compared to the simulated sum rate for a
few numbers of SNR quantization bits. For each number of quantization
bits, the bit allocation to the regions is obtained by exhaustive search to
maximize the sum rate. Also shown in this figure is the “full CSI” case in
which all users always feedback perfect (unquantized) SNRs of all beams. For
the cases tested, the analytical lower bound of the sum rate almost matches
34
the simulated sum rate. Thus the lower bound (2.25) is very tight, and can
be used to analytically allocate available SNR quantization bits to quantize
bits per beam per user (and equal to the average number of SNR quantization
bits per beam for the entire system, as computed in (2.37)). Constrained on
BQ , the bit allocation problem of the N regions for a given beam m can be
formulated as
max Rq (B)
B
∑
N
s.t. bj = BQ , bj ∈ Z+ . (2.26)
j=1
maximum sum rate, because adding one more bit to any of the regions will
increase the average feedback load by the same amount.
For the s-th single bit assigning iteration, the sum rate difference between
using bl,s−1 bits and bl,s−1 + 1 bits for region l can be expressed as
where bl,s−1 is the number of bits for quantizing region l, resulting from the
35
16
15
14
Sum Rate
13
Full CSI
Lower bound of sum rate B=(0,0,0,3)
12
Lower bound of sum rate B=(0,0,0,4)
Lower bound of sum rate B=(0,0,1,4)
Simulation B=(0,0,0,3)
11 Simulation B=(0,0,0,4)
Simulation B=(0,0,1,4)
10
9
0 10 20 30 40 50 60 70 80 90 100
Number of users
Figure 2.6: Sum rate comparison between the mathematical lower bound and
the simulation result, ρ = 10 dB, Mt = Mr = 4.
36
(s − 1)th bit assigning iteration. Rm
l
is the sum rate of region l using m
quantization bits, and can be approximated by
∑2m ∫ rth,l,j+1
Rm ≈ Mt
l
log (1 + rth,l,j ) fX(1)
K (x)dx (2.28)
j=1 rth,l,j
as in the sum rate lower bound (2.25) which has been shown to be very tight
in Fig. 2.6. The region which gives the maximum sum rate increment with
one additional bit will be assigned one more bit at the s-th iteration. The
algorithm iterates until all the available quantization bits are allocated, i.e.,
when s = BQ . The greedy algorithm is summarized in Table 2.1 where the
Despite that the sum rate lower bound (2.25) is analytical, it is difficult
37
Thus, the variance of the quantization error using bj bits can be bounded
by [34]
2
σx,j
2
σe,j ≤ ϵ 2bj ,
2
(2.30)
2
where the constant ϵ is source dependent. For example, ϵ = 1.0 for uniformly
distributed sources and ϵ = 1.4 for Gaussian sources. In our case, the SNR
PDFs for different regions are different, thus the ϵ that gives the tightest
bound (2.30) will be different for different regions. In order to simplify the
computation, we set the same ϵ for all regions such that the upper bound
is large because in that case, the probability of each region (2.18) becomes
very small, which implies that each region interval is very small, and the
SNR distribution within each region is almost uniform. We further relax the
constraint for the number of quantization bits in (2.26) from being a positive
integer to being a positive real number. Then, a new bit allocation problem
∑
N 2
σx,j
min
B=(bN ,bN −1 ,...,b1 )
j=1
22bj
∑N
bj = BQ
j=1
s.t. 0 ≤ bj ≤ BQ , j = 1, 2, . . . , N , (2.31)
bj ∈ R+
where in the objective function, the same constant ϵ and same probability
1/K for all regions are dropped for conciseness without changing the problem.
Since the optimization problem (2.31) is convex, we can apply the Karush-
38
Kuhn-Tucker (KKT) conditions [35] to solve it. To simplify the expression
in (2.31), we let
∑
N ∑
N
L(B, λ, ν1 , . . . , νN , δ1 , . . . , δN ) = f0 (B) + λf1 (B) + νj hj (B) + δj qj (B)
j=1 j=1
(2.32)
where
∑
f0 (B) = N 2
j=1 σx,j 2
−2bj
f1 (B) = ∑N bj − BQ
j=1
hj (B) = −bj
q (B) = b − B .
j j Q
Since f0 , f1 , hj , qj are differentiable, the KKT conditions for this problem are
∂L(B, λ, ν1 , . . . , νN , δ1 , · · · , δN )
= 0, j = 1, 2, . . . , N,
∂bj
∑N
b = BQ
j=1 j
λ ̸= 0 (2.33)
νj hj (B) = 0, δj qj (B) = 0, j = 1, 2, . . . , N,
νj ≥ 0, δj ≥ 0, j = 1, 2, . . . , N.
∂L(B, λ, ν1 , · · · νN , δ1 , · · · δN )
From = 0, we have
∂bj
39
( )
2
ln (ln 4)σx,j BQ ln 4
where Tj = and V = , bj can be obtained by
ln 10 ln 10
0, W > Tj
(Tj − W ) ln 10
bj = , Tj − V < W < T j (2.36)
ln 4
BQ , W < Tj − V.
40
16
15
14
Sum rate
13
12
Full CSI
11 Proposed Scheme C, BQ=3, greedy algorithm
Proposed Scheme C, BQ=5, greedy algorithm
10 Proposed Scheme C, BQ=3, fast bit allocation
Proposed Scheme C, BQ=5, fast bit allocation
9
0 10 20 30 40 50 60 70 80 90 100
Number of users
Figure 2.7: Sum rate performance comparison for different bit allocation
methods with N = 4 regions, ρ = 10 dB, Mt = Mr = 4.
Fig. 2.7 compares the simulated sum rate performance using the greedy
and the fast bit allocation methods. It is shown that the sum rate perfor-
mances for these two bit allocation methods are visually indistinguishable.
Thus, the fast bit allocation method is preferred for all practical purposes.
For the cases considered, the bit allocations obtained by these two methods
are the same as the exhaustively searched bit allocation shown previously
in Fig. 2.6. The general trend is that more bits are allocated to higher
41
SNR regions. When BQ is small, the optimal bit allocation has the form
B = (0, 0, . . . , BQ ). This is reasonable because given the same probability
of a user’s SNR falling in each region, the region [rth,1 , ∞) is the thin tail of
the PDF and has a wider range. For this wide range of SNR which results in
higher rate than the other (lower) regions, it is more important to differenti-
ate the users so the user with the highest rate is more likely to be selected.
Thus, more bits should be allocated to this region.
1)BQ + 2BQ − 2)). On the other hand, the fast bit allocation algorithm only
needs to calculate the variance in each region, and the bit allocation vector is
obtained by (2.36). Thus, the complexity of the fast bit allocation algorithm
Let BR be the number of feedback bits carrying the rank information and BQ
∑N { }
1
F b = KMt (BR + bj ) = Mt (N BR + BQ ) (2.37)
j=1
K
which does not increase with the number of users, and is a constant when the
number of transmission beams Mt and the number of regions N are fixed.
42
This is in contrast to the conventional feedback schemes whose total feedback
loads for the network increase with the number of users.
For the conventional feedback scheme, named Scheme A, each user always
feeds back to the BS the SNR values of the Mt beams. A reduced feedback
scheme was proposed in [11] [36] where each user only feeds back its largest
SNR value among all beams and the corresponding beam index. We refer
feedback scheme proposed in [13] will be called Scheme D. For that scheme,
each user feeds back the SNR value of a beam direction when the SNR
scheduling outage probability Pout which is the probability that none of the
users feeds back. In the performance comparison, we additionally introduce a
slightly modified Scheme D based on the design philosophy proposed in this
paper by setting the threshold as rth,N of the Proposed Scheme C, such
that the Pout of this scheme equals the probability of rate loss event PL of the
Proposed Scheme C in (2.20). Thus, in the comparison, we will consider
( )
N K
Scheme D with constant Pout = 10−1 , 10−4 , and Pout = PL (K) = 1 − K .
In the simulation, Scheme A and Scheme B use BQ,A and BQ,B bits, re-
43
spectively, to quantize their SNR values in the region [0, ∞). The “full CSI”
case is Scheme A with infinite quantization bits (i.e., the SNR feedbacks
∑
are unquantized). The Proposed Scheme C has BQ = N j=1 bj bits allo-
cated to N = 4 regions using the fast bit allocation method in Section 2.3.2.
The number of regions is chosen to guarantee the sum rate loss upper bound
in (2.22) smaller than the system tolerable sum rate loss △RP (K) = 0.25
bps/Hz. For Scheme D, BQ,D bits are used to optimally quantize the region
threshold values increase with number of users. As a result, the SNR ranges
of the regions change. Thus the bit allocation and quantization of each region
Fig. 2.8 compares the sum rates of different feedback schemes as the
above for different schemes are set as five. Note that for different schemes,
the relationships between the number of SNR quantization bits and the total
feedback load are different. Therefore, Fig. 2.8 is shown only to illustrate
the performance difference between similar schemes. With the same number
of SNR quantization bits, Scheme A’s total feedback load is roughly four
times that of Scheme B. Thus it is reasonable that Scheme A’s sum rate is
higher than that of Scheme B, with the sum rate difference getting smaller
as the number of users increases. This is because when the number of users is
large, feeding back only the largest SNR among all beams is good enough for
the purpose of scheduling. For Scheme D, setting the threshold such that
Pout = 10−4 results in higher sum rate compared to setting the threshold as
44
rth,4 when the number of users is large. This is because the scheduling outage
probability of the latter increases with the number of users, and is higher
than that of the former when the number of users is large. With BQ = 5, the
Proposed Scheme C’s average number of feedback bits is Mt (N BR +BQ ) =
52 which is less than Mt N BQ,D = 80 of Scheme D using the threshold rth,4
in Fig. 2.8, and confirms the above discussion on the numbers of feedback
bits for similar schemes. For example, the feedback loads of Scheme A
and Scheme B grow linearly with the number of users, and the slope of
Scheme A is four times that of Scheme B because Scheme A’s users feed-
back the SNR of every beam. The Proposed Scheme C and Scheme D
with threshold rth,4 have constant feedback loads as discussed in Section 2.3.4.
(Pout = 10−1 , 10−4 ) has its feedback load increasing with the number of users,
but saturating when the number of users is high. This is because when Pout
1/K
is fixed, Scheme D’s threshold is −ρ ln(1 − Pout ). When the number of
( )
1/K
users is large, Scheme D’s feedback load is limK→∞ Mt BQ K 1 − Pout =
Mt BQ ln(1/Pout ). Thus the feedback load behaviors of the three Scheme Ds
are similar when the number of users is large.
For fair comparison between the feedback schemes, the results of sum rate
and feedback load are combined to show the sum rate as a function of the
feedback load when Mr = Mt = 4 in Fig. 2.10. As shown in Fig. 2.10, for
Scheme A and Scheme B, the feedback load has to be increased if higher
45
sum rate is desired. The Proposed Scheme C and Scheme D with rth,4 as
the threshold (Pout = PL ), which is based on the same design philosophy as
the Proposed Scheme C, have much lower and fixed feedback loads as their
sum rates grow like (1−e−N )Mt log log(K). It can be seen that, to achieve the
same sum rate, the Proposed Scheme C requires lower feedback load than
Scheme D. Note that, based on the design philosophy in [13], Scheme D
with constant scheduling outage probability has its feedback load increasing
with the sum rate, like Scheme A and Scheme B, for smaller number of
users. In fact, if the scheduling outage probability is set to zero, Scheme D
will become exactly the same as Scheme A. When the number of users is
large, the feedback load of Scheme D saturates, making it behave more like
Proposed Scheme C. When Scheme D’s Pout is large, its sum rate loss is
also large.
Fig. 2.11 shows the sum rate as a function of the feedback load when
smaller than the sum rate in Fig. 2.10, and the feedback load in Fig. 2.11 is
46
16
15
14
Sum rate
13
Full CSI
Scheme A BQ,A=5
Scheme B BQ,B=5
12
Proposed Scheme C, BQ=5 (N=4)
Scheme D BQ,D=5, Pout=PL
11 Scheme D BQ,D=5, Pout=10-4
47
2000
1400
Scheme D, BQ,D=5, Pout=PL
1200 Scheme D, BQ,D=5, Pout=10-4
800
600
400
200
0
10 20 30 40 50 60 70 80 90 100
Number of users
Figure 2.9: Feedback load comparison for different feedback schemes. ρ = 10
dB, Mt = Mr = 4.
48
15
14.5
14
13.5
13
Scheme A, fast algorithm
Sum rate
12.5
Scheme B, fast algorithm
Proposed Scheme C (N=4), fast algorithm
12 Scheme D, Pout=PL, fast algorithm
10.5
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Number of feedback bits
Figure 2.10: Sum rate as a function of the feedback load. BQ,A = BQ,B =
BQ = BQ,D = 5, ρ = 10 dB, Mt = Mr = 4.
49
11.5
11
10.5
8.5
8
0 100 200 300 400 500 600 700 800 900 1000
Number of feedback bits
Figure 2.11: Sum rate as a function of the feedback load. BQ,A = BQ,B =
BQ = BQ,D = 5, ρ = 10 dB, Mt = 2, Mr = 4.
50
10
9
Scheme A, ρ=10
Scheme B, ρ=10
8 Proposed Scheme C (N=4), ρ=10
Scheme D, Pout=PL, ρ=10
7 Scheme D, Pout=10-4, ρ=10
Sum rate
51
2.5 Summary
proposed multi-threshold feedback scheme can reduce the feedback load and
utilize the limited feedback bandwidth more effectively than the existing
of the entire system constant regardless of the number of users, the proposed
scheme almost achieves the optimal asymptotic sum rate scaling with respect
to the number of users (i.e., the multiuser diversity). Its sum rate loss com-
pared to the system with full CSI feedback decreases exponentially to zero
52
Chapter 3
size of the mobile device, the user equipped with multiple received antennas is
not a realistic assumption. Another way to reduce interference is through the
user selection method that selects the users with minimum inter-user inter-
ference [37]. But, the high computational complexity is required to find the
optimal scheduled users. In this chapter, the ZF-BF using semi-orthogonal
user selection (SUS) algorithm is considered to improve the performance.
Each user needs to feedback the channel direction information (CDI) and
53
channel quality indicator (CQI) to the BS and the scheduler applies the SUS
algorithm to select users that satisfy the semi-orthogonal constraint to each
not have the information of the beamforming vectors of the scheduled users.
A lower bound of the SINR metric was derived in [26] as an indicator to
represent the received signal quality. However, the sum rate derived using
the SINR lower bound does not tightly match the actual sum rate when the
actual sum rate even when the semi-orthogonal parameter ϵ of the SUS algo-
to reduce the feedback load. These thresholds are derived through the car-
dinality of the candidate user sets of SUS algorithm. With these thresholds,
only when the user’s estimated SINR is greater than the smallest threshold,
the user feedbacks CDI and CQI to the BS. Through simulations, the pro-
posed scheme not only exploits full multiuser diversity but also efficiently
utilizes the feedback resources.
This chapter is organized as follows. Section 3.1 describes the system
model. The SUS scheduling algorithm is introduced and an approximated
sum rate using SUS algorithm is derived in Section 3.2. A novel multi-
threshold feedback scheme is proposed in Section 3.3. The simulation results
54
are shown in Section 3.4 and we conclude this chapter in Section 3.5.
and K users each with single antenna. The received signal of the k-th user
in the scheduled user set S = {π(1), . . . , π(|S|)} can be expressed as
∑
yk = hk x + nk = hk vk sk + hk vj sj + nk , k ∈ S, (3.1)
j̸=k
∑
where x = k∈S vk sk ∈ CMt ×1 is the transmit symbol vector related to the
and the noise nk are i.i.d. complex Gaussian random variables with zero
(P/|S|)|hk vk |2
SIN Rk = ∑ , k ∈ S. (3.2)
1 + (P/|S|) k̸=j |hk vj |2
∑
R(S) = log(1 + SIN Rk ). (3.3)
k∈S
We assume that each user has perfect channel knowledge of hk and quan-
tizes the direction of normalized channel h̃k = hk /∥hk ∥ to a vector ĥk . The
quantization vector hk⋆ is chosen from the codebook Wk = {wk,1 , . . . , wk,2BD }
55
CQI,CDI CDI
User
1
User 1 User (1)
Base Station
User
M antennas K
CQI,CDI CDI
The user feeds back the SINR and index of codebook perfectly and instan-
taneously to the BS. At the BS, the scheduled users are selected through
the SUS scheduling algorithm under semi-orthogonal constraint ϵ which is
described in [9]. The BS selects the first scheduled user π(1) from the initial
user set B0 = {1, 2, . . . , K} as
56
After selecting number of (i − 1) scheduled users, the i-th scheduled user π(i)
selected from the candidate user set Bi = {l|1 ≤ l ≤ K , l ∈
/ {π(1), . . . , π(i −
Because the SINR in (3.2) is unknown either at the BS nor the users, the
expected SINR conditioned on h̃k and ĥk is used as a CQI metric to estimate
SINR [9] [39]. Let θk be the angle between hk and ĥk . Then the h̃k can be
decomposed into
h̃k = cos(θk )ĥk + sin(θk )gk . (3.8)
Note that the gk and vj ∈ V(S) in (3.7) are unit vector on the (Mt − 1)
dimensional hyperplane orthogonal to ĥk , j ̸= k. Since vj is solely deter-
mined by ĥj , j ̸= k. Thus, the inner product |gk vj |2 is Beta distributed with
parameters (1, Mt − 1) and its expected mean value is 1/(Mt − 1). For a
57
given h̃k and ĥk , the expected SINR at user k is lower bounded by
( )
ρ∥hk ∥2 |h̃k vk |2
E{SIN Rk } = E ∑
1 + ρ∥hk ∥2 sin(θk )2 j̸=k |gk vj |2
ρ∥|hk ||2 E{|h̃k vk |2 }
≥ |S|−1
1 + ρM t −1
∥hk ∥2 sin2 θk
ρ∥hk ∥2 {cos2 (θk )E(|ĥk vk |2 ) + sin2 (θk )E{|g̃k vk |2 }
= |S|−1
1 + ρM t −1
∥hk ∥2 sin2 θk
ρ∥hk ∥2 {cos2 (θk )E(|ĥk vk |2 )}
≥ |S|−1
1 + ρM t −1
∥hk ∥2 sin2 θk
( )
(Mt −1)
ρ∥hk ∥ cos (θk ) 1 − Mt ϵ
2 2
≈ |S|−1
, γk,|S| (ϵ), (3.9)
1 + ρM t −1
∥hk ∥2 sin2 θk
( )
(Mt −1)
where E(|ĥk vk |2 ) can be approximated by 1 − Mt
ϵ [40]. Note that the
estimated SINR γk,Mt (0) is chosen as the SINR metric in [9] [41]. The CDF
B −x
|S|−1 −M D B
,T = (Γ(Mt −
x 1 ρ
where υ = ρ(1−δ−Aδx)
, A= Mt −1
, δ=2 t−1 [ 2 De
Γ(Mt −1) (1+Ax)Mt −1
∫∞
1, δ(Ax + 1)υ) − Γ(Mt − 1, υ))], Γ(a, x) = x
ta−1 e−t dt.
58
3.2.3 Sum Rate Analysis
Then, as an example, the probability that only one user among the K users
satisfies the semi-orthogonal constraint is P|S|=1 = (1 − α2 )K−1 . With the
cardinalities of the candidate sets, the probability that j users are selected
Corollary 1.
(3.11). The probability that j users are scheduled at the BS can be expressed
as follows
K−j−Q(j−1) ( j−1 )
∑
K−j
∑ ∏
P|S|=j = ··· αt+1 (1 − αt+1 )
mt
m1 =0 mj−1 =0 t=1
( )
× (1 − αj+1 )K−j−Q(j)
, j ≥ 2, (3.12)
∑
j−1
where Q(j) = mt .
t=1
Proof. In Appendix D.
According to the SUS algorithm, the user π(i) is selected in the Bi based
Bi
on π(i) = arg maxj∈Bi γj,|S| (ϵ). For notational simplicity, we define γ(1),|S| (ϵ)
59
as the largest random variable among |Bi | random variables with the common
CDF in (3.10). The sum rate with quantized CDI and SINR in (3.9) can be
approximated by
∑
Mt (∑
j )
Bi
R≈ P|S|=j E log(1 + γ(1),j )
j=1 i=1
j ∫
∑
Mt ∑ ∞
= P|S|=j log(1 + x)dFγk,|S| (ϵ) (x)|Bi | dx. (3.13)
j=1 i=1 0
Considering the maximum sum rate criterion in each candidate user set,
the users with relatively low estimated SINR are rarely selected in the can-
didate user set. Allowing these users to feed back CSI may waste feedback
the range of estimated SINR into multiple regions to reduce the feedback
load.
60
user belongs to Bi and its estimated SINR falls in the region [rth,i , ∞), the
user is more likely to be selected in the candidate set Bi . Thus, the threshold
1
= 1 − Fγk,Mt (ϵ) (rth,i ), i = 1, 2, . . . , Mt . (3.14)
|Bi |
In general, the cardinality of i-th candidate set |Bi | is larger than cardinality
of (i + 1)-th candidate user set |Bi+1 |. Because it’s easier to find the i sched-
uled users that satisfy the semi-orthogonal constraint than i + 1 scheduled
candidate sets follow |B1 | ≥ |B2 | ≥ · · · ≥ |BMt |, and these thresholds de-
that, when the user’s estimated SINR is smaller than rth,Mt , the user does
not feedback any CSI to the BS. Otherwise, the estimated SINR and CDI
are required to feed back to the BS. Besides, all thresholds can be computed
the fading statistic are known. The value of the thresholds can be updated
periodically according to the system configuration possibly broadcasted by
the BS.
In the proposed scheme, the quantization bits for CQI are BQ = log(Mt )
bits and the CDI is quantized using the codebook with the size 2BD . From
(3.14), when the |BMt | > 1, the feedback probability of the user can be
61
expressed as
1
P {γk,Mt (ϵ) > rth,Mt } = 1 − Fγk,Mt (ϵ) (rth,Mt ) = (3.15)
|BMt |
When |BMt | ≤ 1, the γth,Mt becomes to zero. All users are required to
feedback CSI to the BS in that case. Therefore, the feedback load of proposed
multi-threshold scheme is summarized as follows
K(BD + BQ ) , |BMt | ≤ 1
Lf bk = (3.16)
K (BD + BQ ) = (BD +BQ ) , |BMt | > 1
|BM | t ⌊αM ⌋ t
In this section, the simulated sum rate and approximated sum rate with
different CQI metrics without considering the quantization loss are discussed.
We also compare the different feedback schemes in terms of the sum rate and
two parts: one is the quantized SINR metric which is represented using BQ
bits, the other is the CDI which is quantized by the codebook with size
2BD . The transmit power is 10W while the additive white Gaussian noise at
receiver is 1W. For the conventional feedback type, named Type I, no matter
what the CQI value is, the user quantizes CQI in the range [0, ∞) using BQ
bits and the CDI is quantized through BD bits. The multiple thresholds
model we proposed is referred as Type II scheme, the thresholds are used to
represent the quantization levels. Only when the user’s SINR is greater than
the threshold rth,Mt , the number of bits for quantized CQI and CDI of the
user are BQ = log(Mt ) and BD bits. The single threshold feedback scheme
62
is called Type III. When the SINR falls in the region [rth,Mt , ∞), the SINR
is quantized by BQ bits, and CDI is quantized through the codebook using
BD bits. Otherwise, no CSI is required to feedback to the BS.
Fig. 3.2 shows the (simulated) actual sum rate and the approximate sum
rate using different SINR metrics. The derived approximated sum rate is
close to the actual sum rate when the semi-orthogonal constraint becomes
large, i.e., ϵ = 0.4.
In Fig.3.3, the Type II and the Type III schemes achieve higher sum
rate than the Type I scheme. Because the design of the quantization method
for Type I scheme without considering the scheduling method, thus the
quantization levels are not fine enough to represent the CQI value. Therefore,
when the number of users increases, the sum rate performance of the Type I
scheme is limited by the inaccuracy CQI. Different from the Type I, the
sum rate performance is close to the sum rate with perfect CQI because
the sum rate performance in the Type III scheme increases with number of
users, the quantization levels in SINR region (rth,Mt , ∞] can not represent the
SINR well enough. Thus, there exists the sum rate loss compared with the
Type II scheme. For the multiuser diversity, obviously, the Type II and
the Type III schemes achieve the same growth rate as the sum rate with
perfect CQI even if the CQI bits are small, i.e., BQ = 2.
The feedback load of the Type I scheme increases linearly with number
of users. Thus, when the number of users is 500, the total feedback bits
are K(BD + BQ ) = 5000. The threshold based feedback method is used
for the Type II and the Type III schemes, the feedback load is a constant
63
(BD +BQ )
regardless the number of users and can be approximated as ⌊αM ⌋
≈ 522.
In order to have a fair comparison for all schemes, the sum rate as a function
scheme without considering the scheduling algorithm not only wastes lots of
feedback resources but also limits the sum rate performance due to inaccuracy
CQI.
3.5 Summary
A multi-user MIMO broadcast channel with ZF-BF using SUS scheduling was
sum rate was derived which is close to actual sum rate. Moreover, when
the proposed feedback scheme was used, the feedback resources were utilized
well.
64
13.5
13
Approximated sum rate (ε=0.2, CQI=γk,|S|(0))
12.5
Simulated sum rate (ε=0.2)
Approximated sum rate (ε=0.4, CQI=γk,|S|(0))
12
Sum rate (bps/Hz)
10.5
10
9.5 ε=0.2
9
100 150 200 250 300 350 400 450 500
Number of users
Figure 3.2: Comparison of simulated sum rate and approximated sum rate
with different CQI metrics, P =10W, BD =8 bits.
65
11.5
11
10.5
Sum rate (bps/Hz)
8.5
100 150 200 250 300 350 400 450 500
Number of users
Figure 3.3: Comparison of sum rate for different feedback types Mt =4,
P =10W, BD =8 bits, BQ =2 bits and ε=0.4
66
11
10.5
Type-I (CQI=γk,M (ε=0.4), BQ=2, BD=8,ε=0.4)
t
t
10
Type-III (CQI=γk,M (ε=0.4), BQ=2, BD=8,ε=0.4)
t
9.5
8.5
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Number of bits
Figure 3.4: Comparison of sum rate as a function of feedback load for different
feedback types, Mt =4, P =10W, BD =8 bits, BQ =2 bits, ε=0.4 and K =
100 ∼ 500.
67
Chapter 4
systems, the MBMS users just have to receive data broadcasted on the com-
mon radio resources, and hence the efficiency of radio resources is highly im-
proved and the quality of multimedia service is then increased. Therefore, due
dia quality, MBMS has become one of the important features in next genera-
However, to enable MBMS functions, all the users in the MBMS network
have to feedback its current channel condition to the BS, so that the BS can
select a proper modulation and coding scheme (MCS) for starting the broad-
cast transmission to all the MBMS users. For this sake, the feedback load
will be increased dramatically with the number of the users in the MBMS
68
network and hence the spectrum efficiency is degraded due to the accordingly
increased feedback bandwidth. Therefore, for the reason mentioned above,
the reason that statistically there must be some users with SNR lower than
the threshold, the feedback load is hence effectively reduced. In [44], a feed-
is proposed. In this method, the range of user’s SNR is divided into multi-
ple classes with multiple thresholds. When the user’s SNR is greater than
the smallest threshold and belongs to a certain class j, the user randomly
The idea of the method is that only when user’s SINR belongs to the region
[0, threshold), the user feeds back its SINR to the BS. The threshold is de-
rived based on the rankings of the users’ received SINRs among all users, i.e.,
the order statistics of the users. Because of the fact that the users with lager
SINRs should not affect the MCS selection of the MBMS network, the users
whose SINRs larger than the selected threshold should not feedback their
channel condition to reduce the system feedback load. From the simulation
results, it can be seen clearly that the feedback load is reduced significantly
while the performance using proposed feedback model is kept closely to the
69
Figure 4.1: MBMS system model
This chapter is organized as follows. Section 4.1 describes the MBMS sys-
70
orthogonal vectors generated from isotropic distribution [45]. The received
signal yi for user i is given by
yi = h∗i Wx + ni , (4.1)
CMt ×1 represents the channel fading vector seen by user i with i.i.d. complex
Gaussian random variables with zero mean and unit variance, ni ∈ C is
AWGN seen by user i with zero mean and variance σ 2 . It is assumed that
the feedback channel is error-free and the feedback delay time is negligible.
is also assumed that the total power is equally allocated at each beam, i.e.,
The beamformer wj is served for the broadcast signal xj , then the received
∑
Mt
yi,j = h∗i wj xj + h∗i wt xt + ni , j = 1, 2, . . . , Mt . (4.2)
t=1,t̸=j
The last two terms in (4.2) are interference and noise terms, and the corre-
|h∗i wj |2 Zj
γi,j = ∑M t = , (4.3)
Mt
ρ
+ t=1,t̸=j |h∗i wt |2 Mt
ρ
+ Wj
∑Mt
where ρ = Pt
σ2
, Zj = |h∗i wj |2 , and Wj = t=1,t̸=j |h∗i wt |2 . Since the channels
between the BS and users are all assumed to be i.i.d. Rayleigh distribu-
tion and the precoder W is an unitary matrix, the terms Zj and Wj are
independent chi-square distribution, i.e., Zj ∼ χ(2) and Wj ∼ χ(2Mt − 2)
71
respectively [11]. The PDF and CDF of γi,j , ∀i, j can be expressed as
Mt γi,j { }
e− ρ Mt
f (γi,j ) = (1 + γi,j ) + Mt − 1 , (4.4)
(1 + γi,j )Mt ρ
and Mt γi,j
e− ρ
F (γi,j ) = 1 − . (4.5)
(1 + γi,j )Mt −1
Because of the nature of MBMS, the system needs to select a user according
to the criterion as follows
After the user i⋆j is selected, the SINR of user i⋆j can be used to choose a
proper MCS such that most of the users for the broadcast signal xj of the
MBMS network can decode the data successfully. The sum rate of MBMS
∑
Mt
R= E{log(1 + ri⋆ ,j )}. (4.7)
j=1
To achieve the purpose of MCS selection, all users in the MBMS system need
to feed back their CSIs, i.e., the SINRs of all beams, to the BS resulting in
72
Due to symmetric property, the index j is removed in (4.3) to simplify
the representation of i-th user’s SINR γi . Let γ(1) ≥ γ(2) ≥ · · · ≥ γ(K) be
∑
K
P {γk = γ(p) |γk = sinrk,j } = 1. (4.9)
p=1
Due to the symmetric assumption that users suffer i.i.d. Rayleigh fading
process, also due to ORB, the same set of thresholds applies to all users and
l
F (rth,l ) = 1 − , l = 1, 2, . . . , K. (4.10)
K
73
feeds back the SINR to the BS. Otherwise, the user no needs to feedback any
CSI to the BS. The set of feedback users is defined as Sf (rth,l ) = {j | γj ≤
Once no one’s SINR is smaller than the threshold rth,l , no one needs to
feedback any CSI to the BS. The BS randomly selects one user among all
users. Applying the proposed feedback scheme, the sum rate can be expressed
as follows
[( ( )K )( )
l
RF (rth,l ) = Mt E 1 − log(1 + min γj )
K γj ∈Sf
( )K ( )]
l
+ log(1 + γi ) . (4.12)
K
1, 2, . . . , K.
74
4.3.2 Performance Loss Analysis
is used. All users can perfectly decode the data in the conventional feedback
scheme because the threshold value is infinity. Then, the number of success-
fully decoding users for the conventional scheme is ND (∞) = K. However,
decoding users is
[ ( )K ] ( )K ∑K
K −n K −n j
ND (rth,K−n ) = 1 − K+ (4.13)
K K j=1
K
( )K
n (K − 1)
=K − 1− . (4.14)
K 2
The first term in (4.13) describes that the user with the worst SINR is fed
back to the BS and is selected at the BS. In this case, all users can decode
the data correctly. Another term explains the event occurred that when no
one feedbacks the CSI to the BS, the BS randomly selects one user among
all users. Once the user that its received sinrk,j on the j-th beam is ranked
p-th among all users is randomly selected by the BS, then p users can decode
the data correctly.
75
Lemma 1. When the proposed feedback method is applied and the threshold
rth,K−n is used, the number of successfully decoding users satisfies
ND (rth,K−n ) 1
lim = 1 − n.
K→∞ K 2e
n K
Proof. Since lim (1 − ) = e−n , the lemma holds immediately.
K→∞ K
The benefit of reduction of feedback load comes at the cost of the per-
formance loss which is defined by the difference of number of successfully
decoding users between conventional scheme and the scheme with proposed
feedback method. Thus, the performance loss compared with the conven-
Note that when no one feeds back CSI to the BS, the performance loss is at
most about fifty percent. In general, more users feed back CSI to the BS,
as follows
( )[( )K ( )K ]
K −1 K −n K − (n + 1)
∆ND (rth,K−(n+1) , rth,K−n ) = − .
2 K K
(4.16)
76
all users can decode the data correctly. This hints that when the threshold
used in the MBMS network is greater than rth,K−5 , the feedback bandwidth
is wasted.
77
60
K=10
50
K=100
K=500
∆ ND(rth,K-(n+1), rth,K-n)
40
30
20
10
0
1 2 3 4 5 6 7 8 9
n
Figure 4.2: Increment of number of successfully decoding users increasing
threshold from rth,K−n to rth,K−(n+1) .
loss ∆ND (∞, rth,K−n ) is an increasing function with respect to the number of
users K shown in lemma 2. For a given n, an intuitive explanation of lemma
78
threshold should be adjusted with number of users K to keep the performance
loss be allowed. Thus, for a given the tolerable performance loss ∆NL (K), the
Lemma 2. When the threshold is selected as rth,K−n , the feedback load be-
are K users, each having single antenna. The transmitter power P is 20W
and the power of AWGN at the receiver is 1W. For the conventional feed-
back scheme, the user always feeds back the CSI to the BS on each beam.
The threshold based feedback scheme we proposed is referred as proposed
scheme. For a given threshold rth,K−n ,only when user’s SINR locates in the
region (0, rth,K−n ], the user feeds back the SINR to the BS. Otherwise, the
user no needs to feedback any CSI to the BS. Besides, in order to keep the
79
performance loss be allowed, the n is selected according to (4.17) to meet the
performance loss requirement. The tolerable performance loss of number of
back scheme and proposed scheme becomes large when the number of users
increases. For a given tolerable performance loss constraint ∆NL (K) = 0.1K,
the proposed scheme using the threshold rth,K−2 can meet the performance
Fig. 4.4. Not only constant feedback load is achieved but also the perfor-
fully decoding users and the feedback load are combined to show the number
our proposed feedback scheme is used, the feedback load can be reduced dra-
4.5 Summary
80
100
Full CSI
90
Simulated ND(rth,K-1)
Number of successfully decoding users
80 Analytical ND(rth,K-1)
Simulated ND(rth,K-2)
70
Analytical ND(rth,K-2)
60
Simulated ND(rth,K-3)
50 Analytical ND(rth,K-3)
K-∆NL(K)=0.9K
40
30
20
10
0
0 10 20 30 40 50 60 70 80 90 100
Number of users
81
100
90
Full CSI
80 Proposed scheme using rth,K-1
Proposed scheme using rth,K-2
70
Proposed scheme using rth,K-3
Feedback load
60
50
40
30
20
10
0
0 10 20 30 40 50 60 70 80 90 100
Number of users
82
100
Full CSI
90
Proposed scheme using rth,K-1
Number of successfully decoding users
70
Proposed scheme using rth,K-3
60
50
40
30
20
10
0
0 10 20 30 40 50 60 70 80 90 100
Feedback load
83
quirement.
84
Chapter 5
sider the optimization problem with objective that maximizes the sum rate
two-user case with limited feedback under the Rayleigh fading channel was
85
Simulation results show that our proposed low complexity algorithm can
almost achieve the performance of the greedy algorithm.
The user j feeds back its SNR value to the BS when the SNR value is greater
than its pre-determined threshold τj which is broadcasted from the BS. Then
the BS performs the maximum sum rate algorithm to schedule the users.
K (∏
∑ i−1 )∫ τi−1 (∏
K )
= Fj (τj ) log(1 + x)d Fj (x) , (5.2)
i=1 j=1 τi j=i
86
∏
0
where we define Fj (τj ) = 1. We set a total feedback load C as a constraint,
j=1
and formulate the problem as
max R(τ1 , . . . , τK )
(τ1 ,...,τK )
∑
K K (
∑ )
s.t. Pj = 1 − Fj (τj ) ≤ C. (5.3)
j=1 j=1
timization problem (5.3) does not exist. Thus we turn to developing subop-
timal algorithms that assign feedback probabilities to users. Note that the
each user is set to zero. At the l-th assignment step, the sum rate increment
where Pl−1 = [P1l−1 , P2l−1 , . . . , PKl−1 ] with element Pml−1 being the feedback
probability for user m at the (l − 1)-th assignment step, and Vj = [0, . . . ,
δp , . . . , 0] which has only one non-zero element δp at the j-th location. The
probability increment δp is assigned to user j ⋆ at the l-th assignment step
87
according to
j⋆ = max ∆RIl (j). (5.5)
j=1,2,...,K.
met.
Lemma 3. The partial derivatives of the sum rate with respect to the thresh-
olds satisfy
(
∂R(τ1 , . . . , τK ) ∏
K−1
= − log(1 + τK )fj (τj ) Fm (τm ) − fj (τj )
∂τj m=1,m̸=j
{ K−1
∑ ∏ m ∫ τm
∏K })
F v (x)
× Ft (τt ) v=m+1
dx < 0.
m=j t=1,t̸=j τm+1 (1 + x) ln 2
is reduced which hurts the scheduling and the sum rate. Based on the result
of Lemma 3, the sum rate loss due to decrease of the feedback probability of
the l-th assignment can be approximated by
88
Then, at each assignment step, user k ⋆ will be selected for reducing the
feedback probability by δp if
∑K
The algorithm iterates until the condition j=1 Pj ≤ C is met. However, the
computational burden of integration in Lemma 3 is significant. Therefore,
we replace the integration by an upper bound as shown below
∫ ∏K ( )
τm
Fv (x) ∏K
1 + τm
v=m+1
dx ≤ Fv (τm ) log .
τm+1 (1 + x) ln 2 v=m+1
1 + τm+1
posed algorithm with that of the greedy algorithm under the finite feedback
constraint. The fading channel coefficients for users are assumed to be in-
dependent and complex Gaussian distributed with zero mean and variances
σj2 , where σj2 are drawn uniformly from the interval [0, 1] to model heteroge-
89
3.8
Greedy algorithm
3.6 Partial deravative algorithm: low complexity
Partial deravative algorithm
3.4
Sum rate (bps/Hz)
3.2
2.8
2.6
0 5 10 15 20
Number of users
Figure 5.1: Rate comparison for different algorithm. ρ=10dB.
90
5.4 Summary
91
Chapter 6
users suffer channel fluctuations [48]. However, the feedback load increases
linearly with the number of users and the average system fairness is damaged
The hybrid scheduling is proposed in [24] to achieve the flexible tradeoff be-
tween the sum rate and the fairness among all users. In [24], a hybrid mul-
tiuser scheduler combines the greedy SNR scheduling and normalized SNR
scheduling. The users are grouped into n groups and the grouping method
can be done through sector by sector or ring by ring. One user is selected in
each group based on either the ”best SNR” or ”normalized SNR” criterion
and then the scheduled user is chosen from n selected users based on ”nor-
92
malized SNR” or ”best SNR” criterion. Thus, the flexible tradeoff between
the sum rate and the fairness among all users can be achieved through the
parameter n. However, the concept of feedback load reduction does not take
into account in above scheduling methods.
In this chapter, the relations among the sum rate, average system fairness
and the feedback load are established through the multiple thresholds and
proposed scheduling algorithm. The range of SNR is divided into N regions
according to the multiple thresholds. Only when user’s SNR is greater than
the smallest threshold, the user feeds back SNR to the BS. Thus, the reduc-
selected randomly among N regions, and then the scheduled user with the
highest SNR is chosen from that region which is selected in previous step.
reduction scheme can achieve the flexible tradeoff among the sum rate, the
93
6.1 System Model
The multiuser downlink system included the BS with single antenna and
K users is considered. Each user is equipped with single antenna. The
mean and variance σi2 . The noise for user i, denoted ni , is modeled as i.i.d
complex Gaussian distributed with zero mean and unit variance. We assume
that the total transmit power at the BS is P . The received SNR for user i,
is shown as follows
94
BS to enhance the accuracy of received SNRs at the BS. Especially to note
that only when the user’s received SNR is greater than the pre-determined
threshold, the user feedbacks CSI to the BS. Otherwise, the user no needs to
feed back any CSI to the BS.
To derive the thresholds in the training mode, the statistical mean of the
the notation, we let random variable Ymn represent the m-th statistic in the
95
users in group i when its instantaneous SNR is r.
i
Pm,n (r) = P {Ymi = Y(n) i
|Ymi = r}
∑
= P {Yti1 ≥ Yti2 ≥ · · · ≥ Ytin−1 ≥ r ≥ Ytin+1 · · · ≥ YtiL }
S
| {z } | {z }
(n − 1) variables (L − n) variables
{ n−1
∑∏
1
= (n−1)!(L−n)! exp (−λ(K−(i−1)L−(ta −1)) r)
S a=1 }
∏
L
× (1 − exp−λ(K−L(i−1)−(tb −1)) r ) ,
b=n+1
∑
where S denotes the summation over all permutations (t1 , . . . , tn−1 , tn+1 ,
the instantaneous SNR of user t in group one is infinity, the rank of user t
1
among the L users in group one is one with probability one, i.e., Pt,1 (∞) =
P {Yt1 = Y(1)
1
|Yt1 = ∞} = 1. In order to reduce computational complexity,
The most probable rank of user m in group i when its instantaneous SNR is
r can be obtained by
n⋆ = arg max i
Pm,n (r). (6.5)
n∈{1,2,...,L}
i
Let Qin be defined such that Pm,n i
(Qin ) = Pm,n+1 (Qin ). It can be seen that
[ )
when the instantaneous SNR of user m falls in the range Qin , Qin−1 , the
96
most probable rank of user m in group i is n. To reduce the feedback load
, we let only the users who are most likely to be rank one in each group to
1
rth,i = Qi1 = ln L, i = 1, 2, . . . , N. (6.6)
µi
In order to achieve the flexible tradeoff among the sum rate, the average
system fairness and the feedback load, the proposed scheduling method in
1. The BS uses multiple thresholds rth,1 , rth,2 , . . . , rth,N to divide the range
the BS in the training mode. Only when the user’s received SNR is
greater than the smallest threshold rth,N , the user feeds back the SNR
to the BS. Otherwise, the user no needs to feedback any CIS to the BS.
After the BS collecting the CSI, the candidate user group i, denoted
as Gi = {j | Yj ∈ [rth,i+1 , rth,i ), j = 1, 2, . . . , K} can be obtained at
the BS. Then the BS randomly selects a candidate user group with
non-zero cardinality.
3. When no one feeds back SNR to the BS, the BS randomly selects one
user among K users.
97
6.3 Performance Analysis
6.3.1 Sum Rate Analysis
In this section, we investigate the sum rate performance using the pro-
Thus, the actual number of regions that the BS can randomly select is Q =
∑N ∑N
l=1 I(|Sl |) ≤ N and the number of feedback users is m = l=1 |Sl |.
The sum rate is composed of two parts, one is the sum rate obtained in
the case when the number of candidate user groups is greater than or equal
to one, the other is the sum rate contributed by a user that is randomly
98
selected at the BS. The sum rate of proposed scheme is shown below
[ K (
∑ ∑ K−m ∏ P (Ytm+j < rth,N ) ([
R= P (Yt1 ∈ S1 | Yt1 ≥ rth,N )
m=1 G j=1
m!(K − m)!
∏ (1 − e−λi rth,N ) ∑ [ ]
K K
+ E log(1 + Yq | Yq < rth,N ) , (6.8)
i=1
K q=1
∑
where G denotes the summation over all permutations (t1 , t2 , . . . , tm , . . . ,
tK ) of (1, 2, . . . , m, . . . , K). The sum rate of selected user when its SNR
99
probability that the user i is selected randomly at the BS. Then, the Pi can
be expressed as follows:
[ K−1 ( (
∑ 1 ∑ m−1 ∏
Pi = P (Ytm+j < rth,N )
m=1
(m − 1)!(K − m)! ′ i=1
G
[
P (Yt1 ∈ S1 | Yt1 ≥ rth,N )P (Yt2 ∈ S1 | Yt2 ≥ rth,N ) × · · ·
[ 1
× P (Ytm−1 ∈ S1 | Ytm−1 ≥ rth,N ) P (Ym ∈ S1 | Ym ≥ rth,N )
|S1 |
∑N
] ] [
+ P (Ym ∈ Sl | Ym ≥ rth,N ) + P (Yt1 ∈ S1 | Yt1 ≥ rth,N )
l=2
rth,N , the user no needs to feed back any CSI to the BS. Otherwise, the user
feeds back CSI to the BS with probability e−λk rth,N , k = 1, 2, . . . , K. The
100
average total feedback load can be expressed as
K (
∑ )
−λk rth,N
Fb = e (6.10)
k=1
From the equation (6.10), the average system feedback load can be reduced
In this simulation, we show the sum rate, the average fairness of the
system and the feedback load for the proposed scheduler with multi-threshold
model. The fading channel coefficients for users are assumed to be i.i.d.
complex Gaussian distribution with zero and variance σi2 , where σi2 are drawn
transmit power P is 10W, while the additive white Gaussian noise power
2
at the receivers σN is 1W. The proposed scheduling scheme with feedback
fair scheduling (normalized SNR) and the scheduling using random selection
tradeoff among the average system fairness, the sum rate and the feedback
load. When number of region N equals to one, our scheme becomes to
the scheme with single threshold using maximum SNR scheduling algorithm.
Once the number of regions becomes to K, the threshold rth,N turns to zero.
Thus, our scheme in that case is identical with the conventional feedback
scheme which means no feedback load can be reduced. The full fairness can
be achieved but the gain of multiuser diversity is damaged.
In Fig. 6.1, when the number of regions N increases, then the sum rate
101
performance is decreased. Because the BS randomly selects a region among
N regions first, the probability that the scheduled user has lower SNR is
increased. When the number of regions is equal to one, the sum rate perfor-
mance of our proposed scheme closes to the greedy SNR algorithm. The sum
rate performance using random selection at the BS is also matched with our
proposed scheme when the number of regions is K. The random selection
phase in our proposed scheduling scheme selecting one region among N = K
regions just likes random selection of one user among K users. In that case,
the cardinality of candidate user set of our proposed scheme is too small to
As shown in Fig. 6.2, the average system fairness increases with the num-
ber of regions. When the number of regions increases, the number of candi-
date user sets is increased. The more users with lower SNR covered in the
candidate user sets, the more average fairness is obtained. Therefore, the
flexible tradeoff between the sum rate and the fairness among users can be
In Fig. 6.3, we plot sum rate as the function of the feedback load as an
102
5.5
4.5
Sum rate (bps/Hz)
4
Max SNR scheduling
Normalized random selection
3.5 Random selection
Proposed scheme (N=K)
3 Proposed scheme (N=4)
Proposed scheme (N=3)
Proposed scheme (N=2)
2.5
Proposed scheme (N=1)
1.5
10 20 30 40 50 60 70 80 90 100
Number of users
Figure 6.1: Sum rate comparison between different number of feedback re-
gions (P= 10W).
103
1
0.9
0.8
Average system fairness
0.7
0.6
0
0 10 20 30 40 50 60 70 80 90 100
Number of users
104
Sum rate (bps/Hz)
6
1
0 10 20 30 40 50 60 70 80 90 100
Feedback load
1
System fairness
0.95
0.9
Feedback load
Figure 6.3: Sum rate vs. feedback load and fairness vs. feedback load with
different number of regions (P = 10W ).
6.5 Summary
105
fairness among users and the feedback load is achieved through the number
of regions N .
106
Chapter 7
Conclusions
gorithms are proposed for different transmission scheme such as ORB scheme,
ZF-BF using SUS algorithm and MBMS system using ORB approach. In the
ORB scheme with proposed feedback scheme, the sum rate growth is 1−e−N ,
almost achieves the full multiuser diversity, where N is the minimum number
the feedback load is a constant regardless the number of users. For the ZF-
scheduling algorithm and thus the feedback method is more efficient than
the existing methods. However, the sum rate performance is not the desired
purpose of MBMS system. Thus, the number of successfully decoding users
with proposed feedback scheme is investigated. Simulation shows that the
number of successfully decoding users is almost the same with traditional
feedback scheme while the feedback load keeps in a small constant. For the
heterogeneous scenario, each user has its own threshold to reduce feedback
load. For a given feedback load constraint, the threshold is derived through
the proposed low complexity derivative algorithm. Simulations show that the
107
sum rate performance using proposed algorithm is close to the performance
using the greedy algorithm.
We also discussed the relation among the fairness across users, the feed-
back load and the sum rate in the heterogeneous Rayleigh fading channel. A
new scheduling method with multi-threshold feedback scheme is proposed to
achieve the flexible fairness of the system. The fairness is increased when the
number of regions of feedback scheme increases. Thus, the system fairness
108
Appendix A
Proof of Theorem 1
K = N , the users always feedback and there is no sum rate loss. Therefore,
∂△R(K,N,Mt )
we will prove this theorem by showing that > 0 for the nontriv-
(
∂K
can be expressed
{ }
△R(K, N, Mt ) = Mt E log(1 + X(1)
K
) − log(1 + Xk ) | X(1)
K
< rth,N PL (K)
where PL (K) is the probability of rate loss event defined in (2.20). The
109
According to the inequality ln(x) ≥ 1 − 1/x, we have
( )K (
∂PL (K) ∂ N N )( N )K N ( N )K−1
= 1− = ln 1 − 1− + 1−
∂K ∂K K K K K K
( ) [ ( ) ( ) ]
N N K−1 K N N
= 1− 1− ln 1 − +1
K K N K K
[ ]
N( N )K−1 ( K )( 1 )
≥ 1− −1 1− + 1 = 0. (A.3)
K K N 1− K
N
∂rth,N
where ∂K
> 0 can be observed from (2.14), and fXm,k (x) and FXm,k (x) are,
respectively, the PDF and CDF of Xm,k . Since f1 (Xk , K) > 0, PL (K) > 0,
∂PL (K)
∂K
≥ 0 and ∂f1 (Xk ,K)
∂K
> 0, we conclude that ∂△R(K,N,Mt )
∂K
> 0.
110
Appendix B
Proof of Theorem 2
{ }
R(K, N, Mt ) ≥ Mt P {rth,N ≤ X(1)
K
< ∞}E log(1 + X(1)
K
) | rth,N ≤ X(1)
K
<∞
where PL (K) is the probability of the rate loss event defined in (2.20). Be-
cuss these two cases separately. For Mt > Mr , since rth,N has a similar form as
N
the variable un in the proof of [11, Theorem 4] (that is, FXm,x (rth,N ) = 1− K ),
we can follow the steps in the proof of [11, Theorem 4] to compute the lower
bound of rth,N by using the lower bound of fXm,k (x) for x > 1. Therefore,
L
when the number of users is large, the lower bound of rth,N , denoted as rth,N ,
can be expressed as
( )
L
rth,N = ρ ln(K/N ) − ρ Mr (Mt + 1) − 3 ln ln(K/N ) + O(ln ln ln(K/N )).
(B.2)
111
Then,
For Mr ≥ Mt , rth,N can be found using the asymptotic expression of the in-
complete Gamma function [49]. For large x, the incomplete Gamma function
We have
rth,N
= ln(K/N ) + (Mr − Mt ) ln ln(K/N ) + O(ln ln ln(K/N )). (B.6)
ρ
log(1 + rth,N )
Then lim = 1 and we have the same result as (B.3).
K→∞ log log K
112
Appendix C
Proof of Theorem 3
Proof. With the property of the sum rate lower bound already shown in
Theorem 2, we now focus on the sum rate upper bound. Using Jensen’s
CDF FX (x),
∫ ∫
1 { K} 1
FX−1 (u)du ≤ E X(1) ≤K FX−1 (u)du.
1
0 1− K
{ }
Thus, E X(1) | rth,N ≤ X(1) < ∞ can be upper bounded by
K K
{ K } { K }
E X(1) | rth,N ≤ X(1)
K
< ∞ ≤ E X(1) | rth,N ≤ Xk < ∞, k = 1, 2, . . . , K
∫ 1 ∫ 1 ( )
−1
≤K FX|X≥rth,N (u)du = K −ρ ln (1 − u)e−rth,N /ρ du
1 1
1− K 1− K
∫ 1 ( ) ( )
=K FX−1m,k (u) + rth,N du = ρ ln(K) + 1 + rth,N , (C.2)
1
1− K
113
where FX|X≥rth,N (x) is the conditional CDF of X, which is distributed like
FXm,k (x) defined in (2.5) with Mr = Mt , given that X ≥ rth,N . That is,
e−x/ρ
FX|X≥rth,N (x) = 1 − .
e−rth,N /ρ
Substituting (C.2) and E {Xk } = ρ into (C.1), an upper bound of the sum
rate RU (K, N, Mt ) can be found as
(( )
R(K, N, Mt ) ≤ Mt 1 − PL (K) log(1 + ρ ln K + ρ + rth,N )
)
+ PL (K) log(1 + ρ)
, RU (K, N, Mt ).
114
Appendix D
Proof of corollary
Proof. With the cardinalities of the candidate sets, the probability that j
users are selected in the SUS algorithm with constraint ϵ is denoted by P|S|=j .
Only one scheduled user selected at the BS can be interpreted that the other
(K−1) users not satisfy the semi-orthogonal constraint with the selected user.
SINR criterion with α1 = 1. The probability that only one scheduled user at
For the case of only two scheduled users, the second user is selected from
/ {π(1)}, : |ĥl ĥ∗π(1) | ≤ ϵ} according to maximum
B2 = {l|1 ≤ l ≤ K , l ∈
SINR criterion. Due to only two scheduled users, one belongs to B1 and the
other one belongs to B2 . We assume that m users not satisfy the constraint
with π(1) with probability (1 − α2 ) and number of (K − 2 − m) users can
not satisfy the constraint with π(1) and π(2) with probability (1 − α3 ). In
other words, the cardinality of B3 is zero. Thus,
∑
K−2
P|S|=2 = α2 (1 − α2 )m (1 − α3 )K−2−m . (D.2)
m=0
115
Similarity, the probability that only j users are selected at the BS is
K−j−Q(j−1) ( j−1 )
∑
K−j
∑ ∏
P|S|=j = ··· αt+1 (1 − αt+1 )
mt
m1 =0 mj−1 =0 t=1
( )
× (1 − αj+1 )K−j−Q(j)
, j ≥ 2, (D.3)
∑
j−1
where Q(j) = mt .
t=1
116
Appendix E
Proof of Lemma 2
Proof.
117
Appendix F
Proof of Lemma 3
Without losing the generality, we assume that the thresholds of the system
∂R(τ1 , . . . , τK )
< 0, ∀j
∂τj
Proof.
[ j−1 ) ∫ τj−1
∂ (∏ (∏ )
K
∂R(τ1 , . . . , τK )
= Fm (τm ) log2 (1 + x)d Fm (x)
∂τj ∂τj m=1 τj m=j
(∏
j )∫ τj ( ∏
K )
+ Fm (τm ) log2 (1 + x)d Fm (x)
m=1 τj+1 m=j+1
(∏
j+1 )∫ τj+1 ( ∏
K )
+ Fm (τm ) log2 (1 + x)d Fm (x) + · · ·
m=1 τj+2 m=j+2
∏
K−1 ∫ τK−1 ]
+ Fm (τm ) log2 (1 + x)dFK (x)
m=1 τK
118
According to Leibniz’s rule for differentiation under the integral form and
integration by parts, we have
[{
∂R(τ1 , . . . , τK ) (∏
j−1 ) ( ∏
K )
= − Fm (τm ) log2 (1 + τj )fj (τj ) Fm (τj )
∂τj m=1 m=j+1
(∏ j ) ( ∏K ) }
− Fm (τm ) log2 (1 + τj )d Fm (τj )
m=1 m=j+1
{( ∏
j−1 )[ ( ∏
K )
+ Fm (τm )fj (τj ) log2 (1 + τj ) Fm (τj )
m=1 m=j+1
( ∏
K )] ∫ τj
∏K }
Fm (x)
m=j+1
− log2 (1 + τj+1 ) Fm (τj+1 ) − dx
m=j+1 τj+1 (1 + x) ln 2
{ ∏
j+1 [ ∏
K
+ fj (τj ) Fm (τm ) log2 (1 + τj+1 ) Fm (τj+1 )
m=1,m̸=j m=j+2
] ∫ ∏K }
∏
K τj+1
m=j+2 Fm (x)
− log2 (1 + τj+2 ) Fm (τj+2 ) − dx + · · ·
m=j+2 τj+2 (1 + x) ln 2
{ ∏
K−1 [ ]
+ fj (τj ) Fm (τm ) log2 (1 + τK−1 )FK (τK−1 ) − log2 (1 + τK )FK (τK )
m=1,m̸=j
∫ τK−1 }]
FK (x)
− dx
τK (1 + x) ln 2
(
∏
K−1
= − log2 (1 + τK )fj (τj ) Fm (τm )
m=1,m̸=j
{ K−1
∑ ∏ m ∫ τm
∏K })
Fv (x)
− fj (τj ) Ft (τt ) v=m+1
dx <0
m=j t=1,t̸=j τm+1 (1 + x) ln 2
119
Bibliography
[2] 3GPP, “Long term evolution, evolved universal terrestrial radio access
March 2007.
[3] Air Interface for Fixed Broadband Wireless Access Systems. IEEE Std.
802.16-2004, 2004.
[4] M. Costa, “Writing on dirty paper,” IEEE Trans. on Inf. Theory, vol. 29,
120
[7] Evolved Universal Terrestrial Radio Access (E-UTRA); LTE Physical
Layer-General Description (Release 8). 3GPP TSG RAN TS 36.201
beamforming for mimo broadcast channel with many users,” vol. 55,
no. 1, pp. 11–15, Jan. 2007.
nels with limited feedback and user selection,” IEEE J. Select. Areas
Commun., vol. 25, no. 7, pp. 1478–1491, Sept. 2007.
2006.
[11] M. Sharif and B. Hassibi, “On the capacity of MIMO broadcast channel
with partial side information,” IEEE Trans. on Inf. Theory, vol. 51,
121
[14] V. Hassel, M. S. Alouini, D. Gesbert, and G. E. Oien, “Exploiting mul-
tiuser diversity using multiple feedback thresholds,” IEEE VTC, vol. 2,
[18] S. Yu, M. Peng, and W. Wang, “Adaptive traffic multiplexing for mbms
[19] S. Lu, Y. Cai, L. Zhang, J. Li, P. Skov, C. Wang, and Z. He, “Channel-
aware frequency domain packet scheduling for mbms in lte,” IEEE VTC,
pp. 1–5, April 2009.
2008.
122
[21] A. Alexiou, C. Bouras, V. Kokkinos, A. Papazois, and G. Tsichritzis,
“Efficient mcs selection for mbsfn transmissions over lte networks,” IFIP
[22] S. Sohn, H.-S. Kim, and Y. Y. Kim, “A novel multicast scheme for
[24] L. Yang, M. Kang, and M.-S. Alouini, “On the capacity-fairness tradeoff
Trans. on Inf. Theory, vol. 52, no. 11, pp. 5045–5060, Nov. 2006.
123
[28] L. Yang, M. Kang, and M.-S. Alouini, “On the capacity-fairness tradeoff
in multiuser diversity systems,” IEEE Trans. on Vehicular Technology,
[29] J.-H. Li and H.-J. Su, “Feedback reduction for MIMO broadcast chan-
of distortion measures,” IEEE Trans. on Inf. Theory, vol. 24, no. 6, pp.
[34] H.-M. Hang and J.-J. Chen, “Source model for transform video coder
and its application−Part I: Fundamental theory,” IEEE Trans. Circuits
Syst. Video Technol., vol. 7, no. 2, pp. 287–298, April 1997.
124
[36] M. Pugh and B. D. Rao, “Reduce feedback schemes using random beam-
forming in MIMO broadcast channels,” IEEE Trans. on Signal Process-
[37] T. Kang and H. Kim, “Optimal beam subset and user selection for
Trans. on Inf. Theory, vol. 52, no. 11, pp. 5045–5060, Nov. 2006.
for MIMO broadcast channels with limited feedback,” IEEE VTC, pp.
[40] K. Ko, S. Jung, and J. Lee, “Hybrid MU-MISO scheduling with limited
[42] Y. Shao and J. Yuan, “A lower bound to the sum-rate of mimo broadcast
channels with limited-rate feedback,” IEEE Int. Conf. on Commun., pp.
3674–3678, May 2008.
125
[44] J. So, “Opportunistic feedback with multiple classes in wireless sys-
tems,” IEEE Commun. Lett., vol. 13, no. 5, pp. 384–386, June 2009.
http://arxiv.org/abs/1107.4822, 2011.
dumb antennas,” IEEE Trans. on Inf. Theory, vol. 48, no. 6, pp. 1277–
126