1

Performance Analysis of NOMA in Training Based

Multiuser MIMO Systems
Hei Victor Cheng, Emil Björnson, and Erik G. Larsson
Department of Electrical Engineering (ISY), Linköping University, Sweden
Email: {hei.cheng, emil.bjornson, erik.g.larsson}@liu.se

multiplexing gain for high area throughputs, as well as a large

arXiv:1711.01813v1 [cs.IT] 6 Nov 2017

Abstract—This paper considers the use of NOMA in multiuser

MIMO systems in practical scenarios where CSI is acquired array gain for improved coverage.
through pilot signaling. A new NOMA scheme that uses shared The question addressed in this paper is under what cir-
pilots is proposed. Achievable rate analysis is carried out for
different pilot signaling schemes including both uplink and cumstances the use of NOMA can provide gains in multiuser
downlink pilots. The achievable rate performance of the proposed MIMO systems. While this question per se is not new, no exist-
NOMA scheme with shared pilot within each group is compared ing study to the authors’ knowledge addressed it under realistic
with the traditional orthogonal access scheme with orthogonal assumptions on the availability of channel state information
pilots. Our proposed scheme is a generalization of the orthogonal (CSI). Specifically, previous work either assumed perfect CSI
scheme, and can be reduced to the orthogonal scheme when
appropriate power allocation parameters are chosen. Numerical [5], [14] or only statistical CSI [15]. In contrast, we consider
results show that when downlink CSI is available at the users, the use of training (pilot transmission) to acquire estimated
our proposed NOMA scheme outperforms orthogonal schemes. CSI, and we derive rigorous capacity bounds for NOMA-based
However with more groups of users present in the cell, it is access under these practical conditions. Training-based NOMA
preferable to use multi-user beamforming in stead of NOMA. schemes have been considered in [16], but only for single-
antenna systems and hence only downlink pilots are sent to the
I. I NTRODUCTION users for estimating their effective channel gains. Moreoever,
in multi-user MIMO the effective channels depend on the
Non-orthogonal-multiple-access (NOMA) is a new beamforming, which complicates the analysis. Beamforming
multiple-access concept proposed for next generation wireless with imperfect CSI also creates extra interference to the
networks [2]. The key idea behind NOMA is the use users, which has not been investigated in the literature. In
of superposition coding [3], and associated interference contrast, in this work pilots are transmitted on the uplink (UL),
cancellation techniques, to serve multiple terminals in the facilitating the base station to estimate all channels. By virtue
same time-frequency slot. This is classified as NOMA in of reciprocity and time-division-duplex (TDD) operation, the
the power domain. NOMA provides the ability to increase so-obtained estimates constitute legitimate estimates of the
capacity, especially when the number of users exceeds the downlink channel as well and can be used for coherent
dimension of the channel coherence interval, or the number beamforming. However, since the terminals do not know their
of spatial dimensions (antennas) available for multiplexing is effective channels, we consider also the possibility of sending
limited. The technology is currently attracting much attention (beamformed) pilots in the DL.
[4]–[7]. In the standardization of 3GPP-LTE-Advanced The assumptions made on availability of CSI are critical in
networks, a NOMA technique for the downlink (DL), called the analysis of wireless access performance: Perfect CSI (or
multiuser superposition transmission (MUST), was recently even high-quality CSI) is unobtainable in environments with
proposed [8]. mobility, and performance analyses conducted under perfect-
Concurrently, multiuser MIMO is becoming a cornerstone CSI assumptions often yield significantly overoptimistic re-
technology in emerging standards for wireless access. The sults. Conversely, the reliance on only statistical CSI precludes
idea is to use multiple, phase-coherently operating antennas the full exploitation of spatial multiplexing gains, rendering
at the base station to simultaneously serve many terminals any performance results overpessimistic. The quality of the
and separate them in the spatial domain. The basic multiuser channel estimates that can ultimately be obtained is dictated
MIMO concepts and the associated information theory go by the length of the channel coherence interval (CI) (product of
back a long time [9]–[11]. The most useful form of multiuser the coherence time and the coherence bandwidth): the higher
MIMO is massive MIMO, which emerged more recently [12], mobility, the less room for pilots, the lower-quality CSI – and
[13]. In massive MIMO, the base stations use hundreds of vice versa. Since the coherence time is proportional to the
antennas to serve tens of terminals – harnessing a large spatial wavelength, the use of higher carrier-frequencies accentuates
This work was supported by the Swedish Research Council (VR), the this problem. In high mobility and at high frequencies, the
Linköping University Center for Industrial Information Technology (CENIIT), channel coherence may become very short and eventually one
and the ELLIIT. Part of this work has been presented at IEEE Interna- is forced to use non-coherent communication techniques [17].
tional Workshop on Signal Processing Advances in Wireless Communications
(SPAWC) 2017 [1], however there is an error in the Fig. 2 which is corrected The specific technical contributions of this paper are:
in this paper. ‚ We propose a training scheme to obtain CSI and utilize
 2

the NOMA concept in a DL multiuser MIMO system. A. Orthogonal Access Scheme

‚ The derivation of new, rigorous, semi-closed form lower The first scheme is the traditional orthogonal access scheme
bounds on the DL capacity in multiuser MIMO with [18] that schedules K{2 users in a fraction η of time-frequency
NOMA, with and without DL pilots. resources, and then serve the others in the remaining fraction
‚ A numerical demonstration that NOMA can give gains 1 ´ η of the resources. To minimize near-far effects, we
in multiuser MIMO with estimated CSI under appropri- schedule the K{2 users at the cell center in the first fraction
ate conditions, and a discussion of relevant application η, followed by the other K{2 users at the cell edge in the
scenarios, most importantly that of rate-splitting and remaining 1 ´ η of the resources. From now on we call this
multicasting. Scheme-O.

II. S YSTEM M ODEL B. Proposed NOMA Scheme

We consider a single-cell massive MIMO system with M The second scheme is a generalization of an existing scheme
antennas at the base station (BS) and K (even number) single- in the NOMA literature [19], which creates K{2 groups, each
antenna users. Among these users, K{2 of them are located with one user at the cell edge and one at the cell center.
in the cell center, while the other K{2 users are at the cell In [19], the beamformers are selected based on the channel
edge. TDD operation is assumed and therefore the BS acquires of the cell center user, but NOMA with superposition coding
downlink channel estimates through uplink pilot signaling, by is applied within each group so that the cell edge user can
exploiting channel reciprocity. These estimates are used to get a separate data signal. The beamformers can be selected
perform downlink multiuser beamforming. These operations to mitigate the inter-group interference. For example, in [19],
have to be done within the same CI, where the channels are zero-forcing beamforming is applied to cancel inter-group
approximately constant. Therefore the more symbols spent interference. However, this existing scheme can only provide
on uplink training, the fewer symbols are available for data. the user at the cell edge with a small data rate. This is so
We consider non-line-of-sight communication and model the because the beams are directed to the stronger user in the
small-scale fading for each user as independent Rayleigh group, thus the weaker user will not have any beamforming
fading. We denote the small-scale fading realizations for the gain and this results in low received power and no interference
users at the cell center as suppression. Moreover, the existing work is based on the
gk „ CN p0, IM q, k “ 1, . . . , K{2. (1) impractical assumption of perfect CSI. In [20] a two-stage
beamformer is proposed where the outer stage aims to cancel
The corresponding large-scale fading parameters are βgk ą the inter-group interference and the inner stage beamformer is
0, k “ 1, . . . , K{2 for the b
users in the cell center; the actual optimized to enhance the rate performance for the users within
channel realization is then βgk gk . the group. However this approach needs perfect CSI at the BS
Similarly, the small-scale fading realizations for the users which is hard to obtain in practice and therefore we do not
at the cell edge are denoted as consider it here.
We propose a generalization of the NOMA scheme from
hk „ CN p0, IM q, k “ 1, . . . , K{2. (2)
[19] and devise a way to estimate the channels in practice. To
The corresponding large-scale fading parameters are βhk ą resolve the pilot-shortage problem, we propose to reuse the
0, k “ 1, . . . , K{2
b for the users at the cell edge. The actual same pilot for multiple terminals in the same cell. In particular,
channel is then βgk gk . The large-scale fading is widely the BS allocates the same pilot to the two terminals in a group,
different between the two sets of users: βgk " βhk . Note that where one is in the cell center and one is at the cell edge.1
this is the scenario of interest to us, but the formulas will Since the two users are using the same pilot and have the
actually be valid for any values of βgk and βhk . The names same small-scale fading statistics, we will later see that the
”cell edge” and ”cell center” are just descriptive, but should BS cannot distinguish their channel responses. However, the
not be interpreted literally. BS can estimate a linear combination of the channels to both
The BS is assumed to know the deterministic parameters terminals from the pilot transmission. This estimate provides
βgk and βhk . However the small-scale fading realizations are a useful description of the combined channel, particularly, if
unknown a priori and changing independently from one CI to power control is used to even out the pilot signal strengths of
another CI. To estimate the small-scale fading realizations at the two terminals. In our proposed scheme, the BS beamforms
the BS, in traditional TDD multiuser MIMO, orthogonal up- a combination of the data symbols intended for the two
link pilots are transmitted from the users in the cell. However, terminals using the estimated channels. We make use of the
the number of available orthogonal pilot sequences is limited NOMA concept for which the symbols intended for different
by the size of the CI and this effectively limits the number of users are super-imposed using super-position coding. The
users that can be scheduled simultaneously. In this study, we cell edge user performs the decoding by treating inter-user
are interested in the case when K is greater than the number of interference as noise, while the cell center user decodes the
available pilot sequences. To facilitate discussion and analysis, other user’s data first and performs interference cancellation
we assume that there are only K{2 orthogonal pilot sequences before decoding its own data. Since the beamformers are based
available. With this assumption, we compare two schemes that 1 This scheme can be extended to more than two users, and we will briefly
make use of the K{2 pilot sequences differently. discuss about this in Section V.
 3

b
a K{2
eplacements where Dh is a diagonal matrix with βh1 , . . . , βh on
UL DL DL Data: UL DL DL Data:
Pilots Pilots All Users
its diagonal. For Scheme-N, the received uplink pilot signal
Pilots All Users Pilots
YuN P CMˆK{2 is
? ?
YuN “ pu GDg Ag Φ ` pu HDh Ah Φ ` Nu , (5)
where Ab g and Ah are diagonal matrices with
UL DL UL DL DL Data:
b
DL Data: b
K{2 a K{2
Pilots Pilots Cell Center Users Pilots Pilots Cell Edge Users α1g , . . . , αg and α1h , . . . , αh on the diagonal
respectively. Nuc , Nue and Nu represent the additive noise
during pilot transmission with independent and identically
distributed (i.i.d.) CN p0, 1q entries. αkh ď 1 and αkg ď 1 are
One CI Another CI the positive power control parameters applied to the pilot to
Fig. 1. Frame structure in the considered training based multiuser MIMO
(potentially) even out the channel estimation quality between
systems. Upper figure: frame structure for the proposed Scheme-N, where all the users in the same group.
users are scheduled by sharing pilots. Bottom figure: common frame structure Without loss of generality, the k th user at the cell center is
for Scheme-O, where users are scheduled in different CIs. paired with the k th user at the cell edge to form the k th group
in Scheme-N, and they are using the same pilot sequence.
eplacements user (2,g)
user (1,g) user (2,g) From now on we call the cell edge user in the k th group
user (1,g)
“user pk, hq” and the cell center user in the k th group “user
pk, gq”.

user (2,h) A. MMSE Channel Estimation for Scheme-O

user (1,h) user (2,h) In this subsection, we consider the channel estimation for
user (1,h)
(a) (b) Scheme-O. The estimates will be used in the next section
for performance analysis. The BS first processes the received
Fig. 2. The training and the beamforming stages for Scheme-N. (a) the
transmissions during the uplink training stage where two users with the same pilots signals by multiplying with ΦH from the right. The
index share the same orthogonal pilot. (b) the beamforming transmission for processed pilot signal in (3) becomes
the data where the same beam is formed for every two users. b
O O H
ȳuc,k “ rYuc Φ sk “ pu βgk gk ` n̄uc,k , k “ 1, . . . , K{2,
(6)
on the channels of all users, the proposed scheme can deliver where n̄uc,k “ rNuc ΦH sk „ CN p0, IM q, for the users at the
good data rates to everyone. cell center, and where r¨sk denotes the kth column of a matrix
From now on we call this generalized NOMA scheme . The processed pilot signal in (4) becomes
Scheme-N. O O H
b
ȳue,k “ rYue Φ sk “ pu βhk hk ` n̄ue,k , k “ 1, . . . , K{2,
Fig. 1 shows the frame structure for the two schemes,
(7)
Scheme-O and Scheme-N, and Fig. 2 shows the training and
where n̄ue,k “ rNue ΦH sk „ CN p0, IM q, for the users at
beamforming operations for Scheme-N.
the cell edge.
Based on the processed received pilots, the BS then per-
III. U PLINK C HANNEL E STIMATION forms channel estimation. We consider MMSE channel esti-
mation here. Using classical results from [21], we obtain the
In this section we consider the uplink channel estimation for MMSE channel estimate of gk is
the two schemes that we are comparing. The channel estima- b
tion is different from in conventional systems since the number pu βgk
ĝk “ ȳuc,k , k “ 1, . . . , K{2 (8)
of users scheduled in one slot and the pilot orthogonality are pu βgk ` 1
different. We denote the pilot matrix by Φ P CK{2ˆK{2 that
for users at the cell center and the MMSE estimate of hk is
contains the K{2 orthogonal pilot sequences in its rows, i.e. b
ΦΦH “ IK{2 . pu βhk
For Scheme-O, the K{2 users at the cell center are sched- ĥk “ ȳue,k , k “ 1, . . . , K{2 (9)
O
pu βhk ` 1
uled first, and the received uplink pilot signal Yuc P CMˆK{2
is for users at the cell edge.
O ?
Yuc “ pu GDg Φ ` Nuc , (3) B. MMSE Channel Estimation for Scheme-N
b b Similar to the case of Scheme-O, the BS first processes the
K{2
where Dg is a diagonal matrix with βg1 , . . . , βg on its received pilot signal by multiplying with ΦH from the right
diagonal. Then the K{2 users at the cell edge are scheduled in (5) and obtains the processed received signals
O
in a subsequent CI and the received uplink pilot signal Yue P
b
N
MˆK{2 ȳu,k “ rYuN ΦH sk “ pu αkg βgk gk
C is (10)
O ? b
Yue “ pu HDh Φ ` Nue , (4) ` pu αkh βhk hk ` n̄u,k , k “ 1, . . . , K{2,
 4

where n̄u,k “ rNu ΦH sk „ CN p0, IM q. Then the MMSE In contrast, for Scheme-N, the channel estimate at the BS will
channel estimate of gk for a user in the cell center is still be a linear combination of the channels because of the
b use of the same pilot in each group. The noise-free estimate
pu αkg βgk of wk becomes
ĝk “ ȳ N , k “ 1, . . . , K{2.
pu αkg βgk ` pu αkh βhk ` 1 u,k b
(11)
b
ŵk “ αkh βhk hk ` αkg βgk gk “ wk , k “ 1, . . . , K{2.
The MMSE channel estimate of hk for a user at the cell edge
(16)
is
b
pu αkh βhk IV. P ERFORMANCE A NALYSIS
ĥk “ ȳ N , k “ 1, . . . , K{2. (12) In this section, we analyze the ergodic achievable rates of
pu αkg βgk ` pu αkh βhk ` 1 u,k
Scheme-O and Scheme-N under imperfect channel estimation.
We observe that ĝk and ĥk are parallel, thus the BS In wireless systems with fast fading channels, channel codes
cannot distinguish between the channel “direction” of users span many realizations of the fading process. Therefore the
that share the same pilot. This effect is a consequence of ergodic achievable rate is an appropriate metric to characterize
pilot contamination. Pilot contamination is a major issue in the performance of coded systems in fast fading environment.
massive MIMO system, since it makes it hard for the BS It is commonly adopted in the multiuser MIMO literature,
from performing coherent beamforming only towards one of especially when the number of antennas is large. We make
the users that share a pilot [12]. In contrast, if the same use of the UL channel estimates from Section III for downlink
data is multicasted to multiple users, it is desirable to jointly beamforming, by assuming perfect reciprocity between UL
beamform towards all of them. Pilot contamination is then and DL. The channel estimation errors are taken into account
useful to reduce the pilot overhead [22]. In this paper, we will in the ergodic achievable rate expressions. We separate the
show how to exploit NOMA to send different data to the users analysis into three parts, namely the cases with and without
that share a pilot. instantaneous DL CSI, and the case with estimated DL channel
One alternative way to utilize the uplink pilots is to estimate gains. The case with instantaneous downlink CSI is unobtain-
the linear combination able in practice, and used only as a benchmark.
` Note that
˘ the effective ergodic rate have a prelog penalty
b b
wk “ αkg βgk gk ` αkh βhk hk 1 ´ 2TK
for the case without DL pilots, where T is the size
of the CI. This penalty accounts for the loss from spending
of the channels. The MMSE estimate of wk for group k is K
2T of every CI to estimate the ` channels.
˘ For the case with DL
pilots, the pre-log penalty is 1 ´ K T .
? ?
pu αkg βgk ` pu αkh βhk N
ŵk “ ȳ , k “ 1, . . . , K{2. (13)
pu αkg βgk ` pu αkh βhk ` 1 u,k A. Downlink Signal Model
Denote by pd the DL transmission power normalized by the
Note that ŵk is also parallel with ĝk and ĥk . The choice
noise variance. For Scheme-O, the received signal for user k
of channel estimate does not matter because in either case the
in the cell center is
channel estimates are linearly scaled versions of the processed b
pilot signal. Hence the beamforming directions suggested by yc,k “ pd βgk gkT xg ` nc,k , k “ 1, . . . , K{2, (17)
the estimates are the same by using any one of the estimators.
Since we need to normalize the beamformer to satisfy the and the received signal for user k at the cell edge is
power constraint, the scaling disappears after normalization b
and therefore does not affect the rate. ye,k “ pd βhk hTk xh ` ne,k , k “ 1, . . . , K{2, (18)
where xg (xh ) is the signal vector containing data for the cell
C. Interference-Limited Scenarios center users (cell edge users), and nc,k (ne,k ) is the normalized
We can obtain a special case by assuming there is no noise i.i.d. zero mean unit variance complex Gaussian noise at the
during the uplink training, or equivalently that the uplink k th user at the cell center (edge). Before transmission, each
power pu goes to infinity. This yields as an upper bound data symbol is multiplied with a beamforming vector as
on the performance of all the schemes. It is also a good
K{2
approximation of the interference-limited scenario with high ÿ b
xg “ bk O s
γk,g (19)
k,g
SNR, but large inter-user interference.
k“1
For Scheme-O, noise-free channel estimation implies that
the channels are perfectly known at the BS, due to the fact for the users in the cell center and
that all users use orthogonal pilots in the uplink training, i.e., K{2 b
ÿ
xh “ ak O s
γk,h (20)
ĝk “ gk , k “ 1, . . . , K{2, (14) k,h
k“1
and for the users at cell edge. In the above equations γk,h (γk,g )
ĥk “ hk , k “ 1, . . . , K{2. (15) represents the non-negative power control coefficients for user
 5

k at the cell edge (cell center), and sk,h (sk,g ) is the data cross-channel gains between different users. The achievable
symbol intended for user k at the cell edge (cell center) which rate is obtained by averaging over all sources of randomness
is zero mean and unit variance. The combined signal vectors in the channel and noise.
xh and xg need to satisfy the power constraint ErxH h xh s ď 1 For Scheme-O, every user decodes its own data symbol by
and ErxH g xg s ď 1. treating interference as noise. Since perfect CSI is available,
In this work we focus on maximum ratio transmission an ergodic achievable rate of user k with beamforming vec-
(MRT) which is simple to implement and performs close to tor a1 , . . . , aK and b1 , . . . , bK can be computed using [13,
optimality in low SNR scenarios, Section 2.3.5]
ĝ ˚
bk “ a k « ˜
k O T 2
¸ff
p β γ |g b |
ˆ ˙
Er||ĝk ||2 s O K d g k,g k k
Rc,k “ 1´ ηE log2 1 ` ř O T 2
for the cell center users and T pd βgk j γj,g |gk bj | ` 1
(24)
ĥ˚k for the users in the cell center and
ak “ b
Er||ĥk ||2 s
« ˜ ¸ff
k O T 2
p β γ |h a |
ˆ ˙
O K d h k,h k k
Re,k “ 1´ p1´ηqE log2 1 `
pd βhk j γj,h
ř O T 2
for the cell edge users. With the normalized beamforming T |hk aj | ` 1
řK{2 O
vectors, the power constraint becomes k“1 γk,g ď 1 and (25)
řK{2 O
γ ď 1. for the users at the cell edge.
k“1 k,h
For Scheme-N, the received downlink signal for users in the The ergodic achievable rates are measured in b/s/Hz, and
cell center is they can be achieved by using Gaussian signaling and code-
words that span over all channel realizations. The pre-log
K{2
factors account for the loss in achievable rate due to the fact
? that each user is only scheduled for a fraction of the CIs, in
b ÿ
yk,g “ pd βgk gkT ai γi,h si,h
i“1
time or frequency.
K{2 For Scheme-N, recall that we name the k th user at the cell
? edge as pk, hq and the k th user at the cell center as pk, gq. The
b ÿ
` pd βgk gkT bi γi,g si,g ` nk , k “ 1, . . . , K{2.
i“1 instantaneous SINR of sk,h of user pk, gq is
(21)
Similarly, the received downlink signal for users at the cell pd βgk γk,h |gkT ak |2
edge can be written as SINRk,g “
pd βg j‰k γj,h |gkT aj |2 ` pd βgk j γj,g |gkT bj |2 ` 1
k
ř ř
(26)
K{2
? and similarly the instantaneous SINR of sk,h at user pk, hq
b ÿ
yk,h “ pd βhk hTk ai γi,h si,h
can be written as
i“1
K{2 pd βhk γk,h |hTk ak |2
? SINRk,h “ .
b ÿ
` pd βhk hTk bi γi,g si,g ` nk , k “ 1, . . . , K{2. pd βhk T k T
ř 2
ř 2
j‰k γj,h |hk aj | ` pd βh j γj,g |hk bj | ` 1
i“1 (27)
(22) The condition that user pk, gq can decode the data intended
In Scheme-N, where the BS knows only the linear com- for user pk, hq is that the ergodic achievable rate of sk,h at
bination of the channels for the users in the same NOMA user pk, gq is no less than the ergodic achievable rate of sk,h
group, it regards the estimate as the true channel for both at user pk, hq, which is explicitly
users pk, gq and pk, hq since that is the best estimate available.
The combined symbols from both terminals in the same group
?
are weighted with the power control coefficients γk,h and Erlog2 p1 ` SINRk,g qs ě Erlog2 p1 ` SINRk,h qs. (28)
?
γk,g . The transmitted symbol in the k th NOMA group is
? ? When this condition does not hold, we need to lower the
ř γk,h sk,h
hence ř ` γk,g sk,g . Therefore the power constraint data rate to user pk, hq such that it can be decoded at user
is k γk,h ` k γk,g ď 1. In this case we have the MRT
pk, gq. This can be done by choosing
beamforming vector with normalization
NP
Rk,h “ min pErlog2 p1 ` SINRk,g qs, Erlog2 p1 ` SINRk,h qsq .
ŵk˚
ak “ bk “ a . (23) (29)
Er}ŵk }2 s Since Erlog2 p1 ` SINRk,h qs is an achievable rate for user
pk, hq, from an information-theoretic perspective any rate that
B. Performance With Perfect CSI at the Users is lower than that is also achievable. Therefore by transmitting
NP
In this subsection, we compute the ergodic achievable rate with the chosen Rk,h both users are able to decode the data.
for the two schemes under the assumption that the DL pilots In practice, for (28) to hold we just need to properly control
make perfect DL CSI available at the users. This assumes the pilot powers such that (28) holds. Then user pk, gq gathers
that DL pilots are sent in each CI and users perform channel all received signals over all channel realizations (coherence
estimation to obtain their own channel gain coefficients and the intervals) and decodes the data for user pk, hq. Notice that the
 6

SIC is done after the whole codeword is decoded, and not is the interference from other groups of users. Similarly, the
performed in every CI. Therefore it is not a problem if the received signal at user pk, gq is
instantaneous SINR is lower at user pk, gq, as long as (28) is
N ˚?
b
satisfied in the long term. yk,h “ ck βhk hTk ȳu,k pd γk,h sk,h
In the typical scenarios of βhk ! βgk , there is a wide (35)
N ˚?
b
range of possible choices of power control parameters on the ` ck βhk hTk ȳu,k pd γk,g sk,g ` Ik,h ` nk,h ,
pilots available to satisfy (28). With any choice of power where
N ˚?
b ÿ
control satisfying (28) we transmit with the super-position
Ik,h “ βhk cj hTk ȳu,j pd γj,h sj,h
coding scheme such that user pk, hq decodes the signal sk,h j‰k
from yk,h by treating the signal from user pk, gq as noise. (36)
N ˚?
b ÿ
Then user pk, gq performs successive interference cancellation ` βhk cj hTk ȳu,j pd γj,g sj,g
such that it first decodes sk,h from yk,g and then subtracts j‰k
?
b
pd βgk gkT ak γk,h sk,h from yk,g and decodes sk,g after- is the interference from other groups of users.
wards. Now we make use of the channel statistics to write the
With the superposition coding scheme, the achievable rate received signal at terminal pk, hq as
of user pk, gq is given in (30) and the achievable rate of user „ b 
N ˚?
pk, hq is given in (31) on top of next page. yk,h “ E ck βhk hTk ȳu,k pd γk,h sk,h ` zk,h (37)
It is worth noticing that when αkg “ γk,g “ 0 @k and
k N O where we have introduced the following effective noise term
αh “ 1, one can obtain Rk,h “ Re,k with η “ 1. Similarly
when αh “ γk,h “ 0 , @k and αkg “ 1, one can obtain
k ˆ b „ b ˙
N ˚? N ˚?
N
Rk,g “ Rc,k O
with η “ 0. By using time-sharing between zk,h “ ck βhk hTk ȳu,k pd γk,h ´ E ck βhk hTk ȳu,k pd γk,h
these two extremes, we obtain all the ergodic achievable rates
N ˚?
b
that Scheme-O can attain. This shows that Scheme-N is more sk,h ` ck βhk hTk ȳu,k pd γk,g sk,g ` Ik,h ` nk,h .
general than the traditional scheme with orthogonal access. (38)
It can be easily verified that zk,h is uncorrelated with the signal
term in (37). Therefore (37) can be regarded as an equivalent
C. Performance Without Downlink CSI
scalar channel with deterministic known gain and additive
In this section we investigate the case when instantaneous uncorrelated noise. Using the fact that additive Gaussian noise
DL CSI is not available, however we assume the channel is the worst case uncorrelated noise [13, Section 2.3.2], the
statistics are known by all parties. This corresponds to the case following rate is achievable for user pk, hq:
when no DL pilots are sent and serves as a lower bound on the
performance of all the schemes with estimated DL channels. Proposition 1. The following ergodic rate is achievable for
In this case users utilize the long term statistics as the channel user pk, hq with Scheme-N:
gain and decode the signals, that is, they take the statistical ˆ
K
˙ ˆ
pd λk,h βhk γk,h M
˙
N ip
average of the effective gain as an estimate of that gain. Then Rk,h “ 1´ log2 1 ` ,
2T pd λk,h βhk γk,g M ` pd βhk ` 1
the achievable rate is obtained by gathering all the symbols (39)
over different channel realizations and decoding the signal. where λk,h is defined as
Assume the BS uses the estimated CSI for beamforming
to all terminals. Since we are considering MRT beamforming, pu αkh βhk
λk,h “ . (40)
ak and bk are scaled versions of the channel estimate ŵk pu αkh βhk ` pu αkg βgk ` 1
N
which is a scaled version of the processed pilots ȳu,k . Then
N˚ Proof. The proof is given in Appendix A.
the beamforming vector is ak “ bk “ ck ȳu,k where the
normalizing constant ck that meets the power constraint can We define
be calculated as pu αkg βgk
1 1 λk,g “ (41)
ck “ b “b . (32) pu αkh βhk ` pu αkg βgk ` 1
N }2 s
Er}ȳu,k ppu αkh βhk ` pu αkg βgk ` 1qM
to quantify the channel estimation quality for the following
Therefore the received signal at user pk, gq is discussion. Under the condition αkh βhk ď αkg βgk , the effective
SINR of the signal sk,h at user pk, gq is greater than the
N ˚?
b
yk,g “ ck βgk gkT ȳu,k pd γk,h sk,h effective SINR of the signal sk,h at user pk, hq, i.e.,
(33)
N ˚?
b
` ck βgk gkT ȳu,k pd γk,g sk,g ` Ik,g ` nk,g , pd λk,h βhk γk,h M pd λk,g βgk γk,h M
ď .
pd λk,h βhk γk,g M ` pd βhk ` 1 pd λk,g βgk γk,g M ` pd βgk ` 1
where (42)
N ˚?
b ÿ
Ik,g “ βgk cj gkT ȳu,j pd γj,h sj,h Therefore we can use NOMA where user pk, gq decodes data
j‰k
(34) from user pk, hq and then subtracts it from the received signal
N ˚? yk,g . From the sufficient condition αkh βhk ď αkg βgk we see that it
b ÿ
` βhk cj gkT ȳu,j pd γj,g sj,g
j‰k
is better to let the user with larger large-scale fading coefficient
 7

˙ « ˜ ¸ff
pd βgk γk,h |gkT ak |2
ˆ
NP K
Rk,g “ 1´ E log2 1 ` (30)
pd βgk j‰k γj,h |gkT aj |2 ` pd βgk j‰k γj,g |gkT bj |2 ` 1
ř ř
T

˙ « ˜ ¸ff
pd βhk γk,h |hTk ak |2
ˆ
NP K
Rk,h “ 1´ E log2 1 ` (31)
pd βhk j‰k γj,h |hTk aj |2 ` pd βhk j γj,g |hTk bj |2 ` 1
ř ř
T

perform successive interference cancellation as the condition where

is easier to satisfy. We have the new received signal pu βgk
λO
k,g “ . (47)
pu βgk ` 1
N˚ ?
” b ı
ȳk,g “ yk,g ´ E ck βgk gkT ȳu,k pd γk,h sk,h
” b ı? For users at the cell edge, we have
N˚
“ E ck βgk gkT ȳu,k pd γk,g sk,g
˜ ¸
λO O
k,h γk,h M
ˆ ˙
Oip K
´ b ” b ı¯ ? Re,k “ 1 ´ p1 ´ ηq log2 1 ` ,
N˚
` ck βgk gkT ȳu,k ´ E ck βgk gkT ȳu,kN˚
pd γk,h sk,h 2T pd βhk ` 1 (48)
´ b
N˚
” b
N˚
ı¯ ? k “ 1, . . . , K{2
` c βgk gkT ȳu,k ´ E ck βgk gkT ȳu,k pd γk,g sk,g
where
` Ik,g ` nk,g . pu βhk
(43) λO
k,h “ . (49)
pu βhk ` 1
We can similarly write the effective noise as
´ b ” b ı¯ ? As in the case with perfect CSI at the users, when we set
N˚
zk,g “ ck βgk gkT ȳu,k N˚
´ E ck βgk gkT ȳu,k pd γk,h sk,h αkh “ γk,h “ 0, @k and αkg “ 1 in Scheme-N we get the
´ b ” b ı¯ ? achievable rate of the users in the cell center in Scheme-O
N˚ N˚
` c βgk gkT ȳu,k ´ E ck βgk gkT ȳu,k pd γk,g sk,g with η “ 1. Setting αkh “ γk,h “ 0, @k and αkh “ 1 we get the
` Ik,g ` nk,g . achievable rate of the users at the cell edge in Scheme-O with
(44) η “ 0. By using time-sharing between these two extremes,
we obtain all the ergodic achievable rates that Scheme-O can
Proposition 2. The following ergodic rate is achievable for attain.
user pk, gq with Scheme-N:
˜ ¸
N ip
´ τ ¯ λk,g βgk γk,g M D. Performance With Estimated Downlink CSI
Rk,g “ 1 ´ log2 1 ` . (45)
2T pd βgk ` 1 DL CSI does not come for free. In practice some form of
estimation of the beamformed channel gain is usually needed.
Proof. The proof is given in Appendix B. In this subsection, we investigate the performance of the two
From the ergodic rate expressions, we can observe that the schemes when we send DL (beamformed) pilots [23] for the
signal terms are proportional to M , which is the array gain channel estimation. For Scheme-O every user receives its own
from coherent beamforming. Moreover, we observe that the orthogonal pilot. For Scheme-N, since we are using the same
total interference from other groups of users is a constant beamformer for the pair of users in every group k, only one
that only depends the user’s own large-scale fading, but not downlink pilot is needed for every pair of users. In this case
on the number of antennas or channel estimation quality. the users estimate their effective channel gain and perform a
Therefore the only parameters that affect the rate are the form of “equalization” using the estimated channel gain (see
power control parameters and the uplink channel estimation below for the details).
quality. Adding more groups of users in Scheme-N will only We denote the channel gain at user pk, hq as fk,h fi hTk ak .
change the amount of power that is allocated to each group, Then the received pilot at each of these users is
but not the total interference. Each user at the cell edge is
b
ydpk,h “ fk,h pd βhk ` ndpk,h , k “ 1, . . . , K{2. (50)
affected by coherent interference from the signal intended
for the cell center user in its group. However, for the users Assuming LMMSE estimation [21] at the user, we obtain the
in the cell center, coherent interference disappears in the estimate
successive interference cancellation and the only effect of the
fˆk,h “ Erfk,h s
pilot contamination is the degraded channel estimation quality. b
Using similar calculations, we obtain the ergodic achievable βhk pd Varrfk,h s ˆ b ˙
rate expressions for Scheme-O. For users in the cell center, we ` k ydpk,h ´ βhk pd Erfk,h s ,
βh pd Varrfk,h s ` 1
have (51)
˜ ¸
λO k O
k,g βg γk,g M
of the channel gain where
ˆ ˙
Oip K
Rc,k “ 1 ´ η log2 1 ` ,
2T pd βgk ` 1
a
(46) Erfk,h s “ M λk,h ,
(52)
k “ 1, . . . , K{2 Varrfk,h s “ 1.
 8

The estimation quality will improve with M as the mean of A. User Pairing
the channel gain is increasing with M while the variance is In this paper we are investigating the effects of imperfect
constant. CSI obtained through uplink training. The channels are not
Similarly, denote the channel gain at user pk, gq as fk,g fi known a priori; the only information available at the BS
gkT bk . The received pilot at each of these user is regarding the channel strength is the large scale fading co-
efficients of the users. As a result the user pairing has to be
b
ydpk,g “ fk,g pd βgk ` ndpk,g , k “ 1, . . . , K{2. (53)
done based on the large-scale fading coefficients tβk u. This
Applying LMMSE estimation yields the estimate can also be observed from the achievable rate expressions.
This is the same condition that has been discussed in [24].
fˆk,g “ Erfk,g s
b However the differences are that first, in our case there is
βgk pd Varrfk,g s ´ b ¯ (54) a beamforming gain of order M which effectively increases
` k ydpk,g ´ βgk pd Erfk,g s , the SNR and second, the existence of self-interference caused
βg pd Varrfk,g s ` 1
by channel estimation errors. A detailed analysis would be
where a interesting, but has to be left for future work.
Erfk,g s “ M λk,g ,
(55)
Varrfk,g s “ 1.
B. More than Two Users Per Group
With these estimates of the channel gains, we first divide the The proposed Scheme-N can be extended to include more
received signal at user pk, hq by the channel estimate. This than two users per group. Suppose there are L users in each
can be seen as a form of equalization, and ideally the ratio group k and each user is labeled as user pk, 1q to user pk, Lq.
fk,g
fˆk,g
is one. Then we use the same method as above to obtain In the channel estimation phase they are assigned the same
the achievable rate of user pk, hq in (56) on top of next page. pilot. The BS estimates a linear combination of the channels
Similarly for user pk, gq, an achievable rate is given in (57) from all L users in the group. Then the BS uses this for MRT
on top of next page. beamforming. Without loss of generality, assume they have
For Scheme-O, similar techniques can be applied to obtain large-scale fading coefficients ordered as β1k ď β2k ď . . . ď
the achievable rate for the users in the cell center given in (58) βLk . The required condition such that NOMA can be applied
on top of next page. The corresponding achievable rate for the is that user pk, iq can decode all messages intended for user
users at the cell edge is given in (59) on top of next page. pk, jq for all j ď i. The condition can be written as
As in the case with perfect CSI at the users, when we set
αkh “ γk,h “ 0, @k and αkg “ 1 we get the achievable rate of Erlog2 p1 ` SINRk,i qs ě Erlog2 p1 ` SINRk,j qs @ i ě j,
the users at the cell center in Scheme-O with η “ 1. Setting (60)
αkh “ γk,h “ 0, @k and αkh “ 1 we get the achievable rate of where SINRk,i is the effective SINR of user pk, iq which
the users at the cell edge in Scheme-O with η “ 0. By using has different forms according to the availability of CSI. This
time sharing between these two extremes, we obtain all the condition can be met by controlling the pilot power of the
ergodic achievable rates that Scheme-O can attain. users. Detailed analysis of this extension is out of scope and
Table I summarizes all the ergodic rate expressions we have has to be left for future work due to the limit of space.
obtained, they are all listed in Table I with reference to the
equation numbers. C. Users with Multiple Antennas
Comparing the achievable rates of the different schemes
In the case when users are equipped with more than one
under different CSI assumptions, we observe that the main
antenna, adding more antennas can be viewed as adding users
difference among them is that imperfect CSI at the users is
at the same distance. Thus the same analysis and results can
causing self-interference. Without any downlink pilots, this
be applied by putting the different antennas of the same user
self-interference is proportional to the received power (pd β),
in different groups in Scheme-N. This argument does not
which fundamentally limits the achievable rate of the user.
consider the possibility of receive beamforming at the users as
Therefore we can conclude that neither increasing the DL
it requires accurate channel estimation at the users. Since the
power nor putting the user closer to the BS would help much.
scenario we considered is when the pilot resources are scarce,
This would not create a large SINR difference at the user, and
the consideration of receive beamforming at the user side is
thus we expect that Scheme-N would not provide much gain.
out of scope.
However with DL pilots, the self-interference can be reduced
substantially if we increase the DL SNR. This creates a larger
SINR difference at the users and thus we expect that Scheme- D. Power Control
N would provide higher gains. Power control in any communication systems is crucial. We
have considered both power control in the UL for the pilots
V. P RACTICAL I SSUES AND E XTENSIONS and in the DL for the data. They are optimized according to the
In this section we discuss various issues when implementing requirement of the users. In Section VII we will look at the rate
the proposed Scheme-N in practical systems and some possible region and a particular operating point on the Pareto boundary
extensions. Due to space limitations, these issues are discussed of the rate region which is obtained by performing power
briefly and in-depth investigations are left for future work. control on both UL pilots and DL data. However these are
 9

¨ ˇ ” ıˇ2 ˛
ˇ f
pd βhk γk,h ˇE fˆk,h ˇ
ˆ ˙ ˇ
N dp K ˚ k,h
‹
Rk,h “ 1´ ˝1 `
log2 ˚ ıˇ2 „ˇ ˇ2  „ˇ ˇ2  ‹ (56)
T ”
f
ı ˇ ”
ˇ f ˇI ˇ ˇ ‚
pd βhk γk,h Var fˆk,h ` pd βhk γk,g ˇE fˆk,h ˇ ` E ˇ fˆk,h ˇ ` E ˇ fˆ1 ˇ
ˇ ˇ
k,h k,h k,h k,h

¨ ˛
ˇ fk,g ˇ2
ˇ ” ıˇ
ˆ
K
˙ pd βgk γk,g
ˇE fˆ ˇ
N dp
˚ k,g
‹
Rk,g “ 1´ ˝1 `
log2 ˚ „ˇ ˇ2  „ˇ ˇ2  ‹ (57)
T ”
f
ı
ˇI ˇ ˇ ‚
pd βgk Var fˆk,g ` E ˇ fˆk,g ˇ ` E ˇ fˆ1 ˇ
ˇ
k,g k,g k,g

¨ ˇ „ ˇ2 ˛
O
ˇ f ˇ
ˆ ˙ ˚ pd βgk γk,g
O
ˇE O ˇ k,g
fˆk,g ˇ
‹
Odp K ˚ ˇ ‹
Rc,k “ 1´ η log2 ˚1 `
˚ « ˇ2 ff « ˇ2 ‹
ff (58)
T „
fO
 ˇ
ˇ IO ˇ
ˇ
ˇ ˇ ‹
˝
pd βgk Var fˆk,g ` E ˇˇ fˆk,g ˇ ` E ˇ O1 ˇ ‚
O O ˇ ˆ
ˇ fk,g ˇ
k,g k,g

¨ ˇ „ ˇ2 ˛
O
ˇ f ˇ
ˆ ˙ ˚ pd βhk γk,h
O ˇ
ˇE fˆk,h
k,h ˇ
O
‹
O K ˚ ˇ ‹
Re,k “ 1´ p1 ´ ηq log2 ˚ 1 ` «
2
ff «
2
ff ‹ (59)
T ˚ „
O
fk,h
 ˇ
O ˇ
ˇ Ik,h
ˇ ˇ ˇ
ˇ 1 ˇ ‹
˝ k
pd βh Var fˆO ` E ˇˇ fˆO ˇˇ ` E ˇˇ fˆO ˇˇ
‚
k,h k,h k,h

done by a grid search over different power control coefficients. rate-splitting approach where there is no private part for the
More efficient algorithms for this purpose would be useful but user pk, hq and all message to user pk, hq is contained in the
have to be left for future work. common part. Our proposed Scheme-N can be adapted for the
rate-splitting scheme to handle the problem of pilot shortage
VI. OTHER A PPLICATIONS by decomposing the message of user pk, hq into two parts and
A. Application in Multicasting the analysis can be carried out using similar techniques.
A specific application of the techniques in Scheme-N is to
multiresolution multicasting [20]. In multiresoultion multicast- VII. N UMERICAL R ESULTS
ing, signals of different resolutions are multicasted to multiple In this section we compare the performance of the two
users requesting the same data. Users with low SINR decode schemes in different settings. The comparison is done by
only the low resolution signal treating the high resolution comparing the complete achievable rate regions. The achiev-
signal as noise, while users with high received SINR decode able rate region is obtained by considering a grid of pilot
both the low and high resolution signals. It is natural to apply power control and data power control coefficients to obtain
NOMA here since the low resolution signal is wanted by all the rate pairs for each set of power control parameters, and
users in the cell. In this setup we only need to use one uplink then take the convex hull of all the rate pairs. This assumes
pilot for channel training and the same beamforming vector is the use of time-sharing between different sets of power control
applied to all users in the cell. This can be viewed as a special parameters. This gives an approximate rate region which is a
case of Scheme-N where the data intended for all users pk, hq lower bound on the actual rate region.
are the same and data intended for all users pk, gq are the
same. A. Small-Scale Antenna Systems
The first setup that we are looking into is the case with a
B. Rate-Splitting for Improving Sum Degree of Freedom small number of antennas at the BS. In the simulations we
Recently a rate-splitting approach was proposed to improve choose M “ 10, K “ 2, βh “ 1, βg “ 100, pu “ pd “ 1.
the sum degree of freedom in broadcast channels [25] which Since we compare schemes that use the same number of pilots,
is an approach that was first used for interference channels we omit the pre-log penalty caused by the use of pilots for
and then called the the Han-Kobayashi scheme [26]. In the acquiring CSI. For the case without downlink CSI it has fewer
rate-splitting scheme, one selected user’s message is split into pilots than the other cases.
a common part and a private part where the common part can Fig. 3 shows the performance with noise free uplink estima-
be decoded by all users. The common part is super-imposed tion and perfect CSI at the users. This case represents an upper
on the private part and sent with a different beamformer. bound on the performance for practically realizable schemes.
All NOMA schemes can be viewed as a special case of the From this figure we observe that with perfect CSI available,
 10

TABLE I
S UMMARY OF A CHIEVABLE R ATE R ESULTS

Schemes (users) Estimated CSIT, Perfect CSIR Estimated CSIT, no CSIR Estimated CSIT, CSIR
Scheme-O (cell center) (24) (46) (58)
Scheme-O (cell edge) (25) (48) (59)
Scheme-N (cell center) (30) (45) (57)
Scheme-N (cell edge) (31) (39) (56)

12 3.5
Scheme−N
Scheme−N Scheme−O
Scheme−O 3 Orthogonal UL pilots
10

2.5
Rate of User (k,g) (b/s/Hz)

Rate of User 2 (b/s/Hz)

8

2
6
1.5

4
1

2
0.5

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2
Rate of User (k,h) (b/s/Hz) Rate of User 1 (b/s/Hz)

Fig. 3. Achievable rate region with noise free uplink channel estimation and Fig. 5. Achievable rate region with noisy uplink channel estimation and no
perfect CSI at the users. M “ 10, K “ 2 βh “ 1, βg “ 100, pu “ pd “ 1. CSI at the users. M “ 10, K “ 2, βh “ 1, βg “ 100, pu “ pd “ 1.

10 Fig. 5 shows the achievable rate region with noisy uplink

Scheme−N
9 Scheme−O estimation and no CSI at the users. Comparing to Fig. 4 we see
8
that CSI at the users is critical as Scheme-N and Scheme-O are
overlapping. Without CSI, Scheme-N is performing the same
Rate of User (k,g) (b/s/Hz)

7
as Scheme-O which means there is no gain from using NOMA.
6 We also plot the performance with orthogonal UL pilots for
5 all K users as reference. In Scheme-N we send K{2 uplink
4
pilots, while with the ‘Orthogonal UL Pilots’ scheme we send
K uplink pilots. In this comparison all schemes do not require
3
downlink pilots. This shows that without taking the penalty of
2 using more pilots, it is better to use orthogonal pilots when
1 DL CSI is not available. When the number of pilot symbols
is limited and sending K orthogonal pilots is not possible, we
0
0 0.5 1 1.5 2 2.5 3 can only compare Scheme-O and Scheme-N. There are still
Rate of User (k,h) (b/s/Hz)
some gains from using NOMA with other sets of parameters
(when M is of the order of thousands) than the one considered
Fig. 4. Achievable rate region with noisy uplink channel estimation and
perfect CSI at the users. M “ 10, K “ 2, βh “ 1, βg “ 100, pu “ pd “ 1. in this figure, but they are marginal and applying NOMA may
not be worth it since it increases the complexity and delays at
the user.
the performance gained by using NOMA is quite significant. Fig. 6 shows the achievable rate region with noisy uplink
For example, when the rate of user pk, hq is 2.5 b/s/Hz, the estimation and estimated channel gains at the users, which is
rate of user pk, gq can be increased by almost 2 b/s/Hz. the most practical scenario. Comparing to Fig. 5 we see that
Fig. 4 shows the performance with noisy uplink estimation with the estimated channel gains, we see some gains from
and perfect CSI at the users. Comparing with Fig. 3 we observe using NOMA. We also plot the performance with orthogonal
that the uplink channel estimation errors do not lower the UL pilots for all K users as reference. In Scheme-N we send
performance much for the user in the cell center. However K{2 uplink pilots and K{2 downlink pilots, while with the
the rate of the user at the cell edge loses more than 20%, due ‘Orthogonal UL Pilots’ scheme we send K uplink pilots and
to the poor quality of the uplink channel estimate. Never the no downlink pilots. Comparing the rate regions we see that our
less, the gain from using NOMA is still large. proposed Scheme-N outperforms both traditional schemes.
 11

9
Scheme−N 18

8 Scheme−O Scheme−N
Orthogonal UL Pilots Scheme−O
16

Sum Ergodic Achievable Rate (b/s/Hz)

7 Orthogonal UL Pilots
Rate of User (k,g) (b/s/Hz)

6 14

5
12

4
10
3

8
2

1 6
5 10 15 20 25 30 35 40
Path Loss Difference (dB)
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Rate of User (k,h) (b/s/Hz)
Fig. 8. Sum rate with noisy uplink channel estimation and estimated channel
gains at the users with path loss differences (large-scale fading of user pk, hq
Fig. 6. Achievable rate region with noisy uplink channel estimation and is fixed while large-scale fading of user pk, gq is varying). M “ 100, K “ 2,
estimated CSI at the users. M “ 10, K “ 2, βh “ 1, βg “ 100, pu “ βh1 “ 1, pu “ pd “ 1. The rate of the user pk, hq is constrained to be the
pd “ 1. rate it would get when using Scheme-O with η “ 0.5.

11 numbers of antennas M at the BS with βh1 “ 1, βg1 “ 100,

Scheme−N
Scheme−O
and pu “ pd “ 1. From the plot we see that the sum rate
10
Orthogonal UL Pilots difference between Scheme-O and Scheme-N is increasing
Sum Ergodic Achievable Rate (b/s/Hz)

9 when M increases. This contradicts the common notion that

NOMA is only useful when the number of antennas at the BS
8 is less than the total number of antennas at the users [5]. The
7
reason for this is that CSI at the users is very important in
NOMA, and when M is small, the estimation quality is not
6 good enough, resulting in a lower rate. When M increases,
the estimation quality at the users increases (due to the array
5
gain that increases the SNR with M in the DL estimation)
4 and hence the gain from NOMA is more significant. We
also observe that the performance gap between Scheme-O and
3
10 20 30 40 50 60 70 80 90 100
the ‘Orthogonal UL Pilots’ decreases with M and eventually
Number of Antennas at the BS (M) Scheme-O performs worse than the latter. This is due to the
channel hardening effect. The more antennas at the BS, the
Fig. 7. Sum rate with noisy uplink channel estimation and estimated channel less fluctuation in the norm of the channel vector (normalized
gains at the users for different number of antennas at the BS (M). K “ 2,
βh1 “ 1, βg1 “ 100, pu “ pd “ 1. The rate of the user pk, hq is constrained
by the number of antennas): the norm of the realization of the
to be the rate it would get when using Scheme-O with η “ 0.5. channel vector is almost equal to its statistical mean.
In Fig. 8 we compare the constrained sum rate with different
large-scale fading coefficients between the paired users, with
B. Constrained Sum Rate Comparison M “ 100, pu “ pd “ 1, βh1 is fixed to be 1 while βg1 varies.
From the plot we see that the sum rate difference between
In this subsection we compare a specific operating point Scheme-O and Scheme-N is increasing with the large-scale
on the achievable rate region. We choose the point such that fading difference. This is expected and matches the results
users at the cell edge get the same rate as in Scheme-O with for single antenna NOMA systems [2]. When the large-scale
η “ 0.5. This means that users at the cell edge do not lose any fading difference is small, the orthogonal UL pilots scheme
rate by using NOMA. We compare the sum rate of the whole gives the best performance because both users have low SNR
cell under this constraint and vary the number of antennas, and therefore DL estimates are of poor quality. This verifies
large-scale fading parameters. In all plots we choose K “ 2 the importance of user pairing in NOMA.
with 1 user at the cell edge and 1 user in the cell center. For
Scheme-O, 1 user is scheduled in one slot, thus full power is
O
used with γ1,g O
“ 1 and γ1,h “ 1. For Scheme-N, we vary the C. Effect of Number of Users or Number of Antennas at the
power between the two users to find the optimal constrained User
sum rate. In this subsection, we look into the effect of increasing the
In Fig. 7 we compare the constrained sum rate with different number of users in the cell, or equivalently, the number of
 12

VIII. C ONCLUSION
26
Scheme-N
24 Scheme-O
Sum Ergodic Achievable Rate (b/s/Hz)

Orthogonal UL Pilots
22
In this work we analyzed the performance of NOMA in
multiuser MIMO under practical scenarios where the CSI
20 was obtained through pilot signaling. The performance anal-
18 ysis was done for a conventional orthogonal scheme and a
NOMA scheme under this setup. Extensive simulations were
16
done using the derived achievable rate expressions. From the
14 simulation results we draw the following conclusions:
12
1) NOMA works well only when high quality CSI is avail-
10 able at the user and there is no inter-group interference;
2) When there is more than one group, it is preferable to
8
2 3 4 5 6 7 8 9 10 use multiuser beamforming instead of NOMA. In this
Number of Users in the Cell (K) case, we need a higher beamforming gain to enhance
the SINR;
Fig. 9. Sum rate with noisy uplink channel estimation and estimated channel
gains at the users with different number of users K. M “ 100, βh “ 1, 3) The gain of NOMA increases with the path loss dif-
βg “ 100 and pu “ pd “ 1. The rate of the user pk, hq is constrained to be ference between the users in the same NOMA group.
the rate it would get when using Scheme-O with η “ 0.5. When the difference is small, multiuser beamforming is
preferable.

The above conclusions hold when the BS precoding is re-

stricted to MR. For other more advanced precoding methods,
antennas at the users. We compare the same operating point most importantly zero-forcing, the observations may change
as in the previous subsection. In the simulation we have the because accurate channel estimates are required by these
same number of users at the cell edge and in the cell center. methods. Some initial simulations have shown that the pro-
The users at the cell edge have the same large-scale fading posed shared-pilot scheme only provides little gain with zero-
βhk “ βh , k “ 1, . . . , K{2 and the users at the cell center forcing precoding. More exploration is needed to find out the
have the same large-scale fading βgk “ βg , k “ 1, . . . , K{2. strategy of applying NOMA in training based systems with
For Scheme-O, all users that are scheduled in one slot have zero-forcing type precoding and it is left for future work.
the same large-scale fading, thus equal power allocation with Moreover, from our simulation results we see that CSI at the
O
γk,g “ 2{K, k “ 1, . . . , K{2 and γk,hO
“ 2{K, k “ user is critical for the NOMA scheme. Instead of sending DL
1, . . . , K{2 is optimal in terms of achievable sum rate. For pilots, blind channel estimation methods designed for NOMA
Scheme-N, we allocate equal power to each group which is can help to reduce pilot overhead and therefore is worthy of
also optimal for the sum rate due to the symmetry in the K{2 exploration.
groups. The length of the coherence interval T is chosen to be
200 which is corresponding to a typical fast fading scenario.

In Fig. 9, we compare the constrained sum rates with

different numbers of users, with M “ 100, pu “ pd “ 1, A PPENDIX A
βh “ 1, and βg “ 100. From the figure we see that Scheme- P ROOF OF P ROPOSITION 1
N outperforms the other schemes only when there is one group
of users. As soon as there are more than one group, Scheme-N
and Scheme-O are the same (Scheme-O is a special case of Using results from [13, Section 2.3.2], we have the capacity
Scheme-N) and they are both worse than the ‘Orthogonal UL lower bound:
Pilots’ scheme. This is because the inter-group interference
lowers the SINR difference between the cell center user and ˇ2 ˛
k hT ȳ N ˚ ?p γ
¨ ˇ ” b ı
the cell edge user, and thus NOMA does not provide any ˇ
ˇE c k β h k u,k
ˇ
d k,h ˇ ‹
N ip
gain. This shows that the SINR difference is the key factor Rk,h “ log2 ˝1 ` ‚. (61)
˚
Varpzk,h q
for NOMA to outperform the orthogonal scheme, but not the
SNR difference. Moreover, from Fig. 9 we also observe that
when we have more users in the cell, it is better to user
multiuser beamforming instead of NOMA. That is because the The numerator can be calculated as
inter-group interference levels are the same for all schemes
and that is the major factor that lowers the SINR. In this
ˇ2
case, increasing the beamforming gain is more effective than
ˇ „ b 
ˇE ck β k hTk ȳ N ˚ ?pd γk,h ˇ “ pd λk,h βhk γk,h M,
ˇ ˇ
removing the intra-group interferences. ˇ h u,k ˇ (62)
 13

and the denominator can be calculated as [6] Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, C. L. I, and H. V.
„ˆ b Poor, “Application of non-orthogonal multiple access in LTE and 5G
N ˚?
Varrzk,h s “ Var ck βhk hTk ȳu,k pd γk,h networks,” IEEE Communications Magazine, vol. 55, no. 2, pp. 185–
191, February 2017.
„ b ˙ [7] V. W. Wong, R. Schober, D. W. K. Ng, and L. Wang, Key Technologies
k T N ˚? for 5G Wireless Systems. Cambridge, U. K.: Cambridge University
´ E ck βh hk ȳu,k pd γk,h sk,h
Press, 2017.
 [8] 3GPP TD RP-150496, “Study on downlink multiuser superposition
N ˚?
b
` ck βhk hTk ȳu,k pd γk,g sk,g ` Ik,h ` nk,h transmission.”
[9] S. C. Swales, M. A. Beach, D. J. Edwards, and J. P. McGeehan, “The
„ b  performance enhancement of multibeam adaptive base-station antennas
N˚
“ Var ck βhk hTk ȳu,k ` VarrIk,h s ` Varrnk,h s for cellular land mobile radio systems,” vol. 39, no. 1, pp. 56–67, 1990.
[10] S. Anderson, M. Millnert, M. Viberg, and B. Wahlberg, “An adaptive
ˇ „ b  ˇ2 array for mobile communication systems,” vol. 40, no. 1, pp. 230–236,
N˚ ? 1991.
` ˇˇE ck βhk hTk ȳu,k
ˇ ˇ
pd γk,g ˇˇ [11] G. Caire and S. Shamai, “On the achievable throughput of a multiantenna
Gaussian broadcast channel,” IEEE Transactions on Information Theory,
“ pd βhk ` 1 ` pd λh,k βhk γk,g M. vol. 49, no. 7, pp. 1691–1706, July 2003.
(63) [12] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
bers of base station antennas,” IEEE Transactions on Wireless Commu-
nications, vol. 9, no. 11, pp. 3590–3600, November 2010.
A PPENDIX B [13] T. Marzetta, E. G. Larsson, H. Yang, and H. Ngo, Fundamentals of
P ROOF OF P ROPOSITION 2 Massive MIMO. Cambridge, U.K.: Cambridge University Press, 2016.
[14] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance of
Using the results from [13, Section 2.3.2], we have the non-orthogonal multiple access in 5G systems with randomly deployed
capacity lower bound: users,” IEEE Signal Processing Letters, vol. 21, no. 12, pp. 1501–1505,
ˇ2 ˛ Dec 2014.
N˚ ?
¨
[15] Q. Sun, S. Han, C. L. I, and Z. Pan, “On the ergodic capacity of MIMO
ˇ ” b ı
ck βgk gkT ȳu,k pd γk,g ˇ
ˇ ˇ
N ip
ˇE NOMA systems,” IEEE Wireless Communications Letters, vol. 4, no. 4,
Rk,g “ log2 ˝1 ` ‚. (64)
˚ ‹
pp. 405–408, Aug 2015.
Varpzk,g q [16] Z. Yang, Z. Ding, P. Fan, and G. K. Karagiannidis, “On the perfor-
mance of non-orthogonal multiple access systems with partial channel
The numerator can be calculated as information,” IEEE Transactions on Communications, vol. 64, no. 2, pp.
654–667, Feb 2016.
ˇ2
N˚ ? [17] L. Jing, E. D. Carvalho, P. Popovski, and A. O. Martinez, “Design and
ˇ ” b ı
ˇE ck βgk gkT ȳu,k pd γk,h ˇ “ pd λk,g βgk γk,g M, (65)
ˇ ˇ
performance analysis of noncoherent detection systems with massive
receiver arrays,” IEEE Transactions on Signal Processing, vol. 64,
and the denominator can be calculated as no. 19, pp. 5000–5010, Oct 2016.
„ˆ b [18] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
N ˚?
Varrzk,g s “ Var ck βgk gkT ȳu,k pd γk,h Cambridge, U. K.: Cambridge University Press, 2005.
[19] B. Kim, S. Lim, H. Kim, S. Suh, J. Kwun, S. Choi, C. Lee, S. Lee,
ı˙ and D. Hong, “Non-orthogonal multiple access in a downlink multiuser
N ˚?
” b
´ E ck βgk gkT ȳu,k pd γk,h sk,h beamforming system,” in IEEE Military Communications Conference,
Nov 2013, pp. 1278–1283.
ˆ b [20] J. Choi, “Minimum power multicast beamforming with superposition
N ˚?
` ck βgk gkT ȳu,k pd γk,g coding for multiresolution broadcast and application to NOMA systems,”
IEEE Transactions on Communications, vol. 63, no. 3, pp. 791–800,
ı˙  March 2015.
T N ˚?
” b
k
´ E ck βg gk ȳu,k pd γk,g sk,g ` Ik,g ` nk,g [21] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Upper Saddle River, NJ, USA: Prentice-Hall, Inc., 1993.
” b ı [22] H. Yang, T. L. Marzetta, and A. Ashikhmin, “Multicast performance of
N˚
“ Var ck βgk gkT ȳu,k ` VarrIk,g s ` Varrnk,g s large-scale antenna systems,” in 2013 IEEE 14th Workshop on Signal
Processing Advances in Wireless Communications (SPAWC), June 2013,
“ pd βgk ` 1. pp. 604–608.
(66) [23] H. Q. Ngo and E. G. Larsson, “No downlink pilots are needed in
TDD massive MIMO,” IEEE Transactions on Wireless Communications,
vol. 16, no. 5, pp. 2921–2935, 2017.
[24] Z. Ding, P. Fan, and H. V. Poor, “Impact of user pairing on 5G
R EFERENCES nonorthogonal multiple-access downlink transmissions,” IEEE Transac-
tions on Vehicular Technology, vol. 65, no. 8, pp. 6010–6023, Aug 2016.
[1] H. V. Cheng, E. Björnson, and E. G. Larsson, “NOMA in multiuser [25] S. Yang, M. Kobayashi, D. Gesbert, and X. Yi, “Degrees of freedom
MIMO systems with imperfect CSI,” in 2017 IEEE International of time correlated MISO broadcast channel with delayed CSIT,” IEEE
Workshop on Signal Processing Advances in Wireless Communications Transactions on Information Theory, vol. 59, no. 1, pp. 315–328, Jan
(SPAWC). 2013.
[2] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and [26] A. E. Gamal and Y. H. Kim, Network Information Theory. Cambridge,
K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular fu- U. K.: Cambridge University Press, 2012.
ture radio access,” in 2013 IEEE 77th Vehicular Technology Conference
(VTC Spring), June 2013, pp. 1–5.
[3] T. Cover, “Broadcast channels,” IEEE Transactions on Information
Theory, vol. 18, no. 1, pp. 2–14, Jan 1972.
[4] L. Dai, B. Wang, Y. Yuan, S. Han, C. l. I, and Z. Wang, “Non-orthogonal
multiple access for 5G: solutions, challenges, opportunities, and future
research trends,” IEEE Communications Magazine, vol. 53, no. 9, pp.
74–81, September 2015.
[5] Z. Ding, F. Adachi, and H. V. Poor, “The application of MIMO
to non-orthogonal multiple access,” IEEE Transactions on Wireless
Communications, vol. 15, no. 1, pp. 537–552, Jan 2016.