Massive MISO IBC Beamforming - a
Multi-Antenna Stochastic Geometry Perspective
Christo Kurisummoottil Thomas, Dirk Slock,
EURECOM, Sophia-Antipolis, France, Email: {kurisumm,slock}@eurecom.fr
Abstract—This work deals with coordinated beamforming (BF)
for the Multi-Input Single-Output (MISO) Interfering Broadcast
Channel (IBC), i.e. the MISO Multi-User Multi-Cell downlink
(DL). The novel beamformers are here optimized for the Expected
Weighted Sum Rate (EWSR) for the case of Partial Channel
State Information at the Transmitters (CSIT). Gaussian (pos-
terior) partial CSIT can optimally combine channel estimate
and channel covariance information. With Gaussian partial
CSIT, the beamformers only depend on the means (estimates)
and (error) covariances of the channels. We extend a recently
introduced large system analysis for optimized beamformers with
partial CSIT, by a stochastic geometry inspired randomization
of the channel covariance eigen spaces, leading to much simpler
analytical results, which depend only on some essential channel
characteristics. In the Massive MISO (MaMISO) limit, we obtain
deterministic approximations of the signal and interference plus
noise powers at the receivers, which are tight as the number
of antennas and number of users M, K → ∞ at fixed ratio.
Simulation results exhibit the correctness of the large system
results and the performance superiority of optimal BF designs
based on both the MaMISO limit of the EWSR and using Linear
Minimum Mean Squared Error (LMMSE) channel estimates.
Index terms— Massive MIMO, stochastic geometry, partial
CSIT, ergodic weighted sum rate, optimal beamforming
I. I NTRODUCTION
In this paper, Tx may denote transmit/transmitter
/transmission and Rx may denote receive/receiver/reception.
Interference is the main limiting factor in wireless
transmission. Base stations (BSs) disposing of multiple
antennas are able to serve multiple Mobile Terminals (MTs)
simultaneously, which is called Spatial Division Multiple
Access (SDMA) or Multi-User (MU) Multi-Input Multiple-
Output (MIMO). The recent development of Massive MIMO
(MaMIMO) [1] opens new possibilities for increased system
capacity while at the same time simplifying system design.
However, MU systems have precise requirements for Channel
State Information at the Tx (CSIT) which is more difficult
to acquire than CSI at the Rx (CSIR). [2] proposes optimal
beamformers (BFs) in the case of partial CSIT in the massive
MISO limit.
Gaussian (posterior) partial CSIT can optimally combine
channel estimate and channel covariance information. Recently
a number of research works have proposed to exploit the
channel hardening in Massive MIMO (MaMIMO) to reduce
global instantaneous CSIT requirements to local instantaneous
CSIT plus global statistical CSIT. Indeed, in Massive MISO
systems, the received signal and interference powers converge
to their expected value due to the law of large numbers. We
may remark that in the case of MIMO (not considered here),
in which UEs possess a limited number of Rx antennas which
contribute actively (e.g. to Zero-Forcing (ZF)/Interference
Alignment (IA) at high SNR), the interference subspace and
hence the receiver at the UEs does not harden. A major
work on large system analysis for MaMISO systems is [3].
The authors obtain deterministic (instead of channel real-
ization dependent) expressions for various scalar quantities,
facilitating the analysis and design of wireless systems. E.g.
it may allow to evaluate beamforming performance without
computing explicit beamformers. The analysis in [3] allowed
e.g. the determination of the optimal regularization factor
in Regularized ZF (R-ZF) BF, both with perfect and partial
CSIT. In [4], the authors investigate the deterministic limits
for optimal beamformers, but only for the perfect CSIT MISO
BC (broadcast channel) case. Some other extensions appeared
recently in [5] or [6] where MISO IBC is considered with
perfect CSIT and weighted Regularized Zero-Forcing (R-ZF)
BF, with two optimized weight levels, for intracell or intercell
interference. [7] studied the energy consumption dynamics in
a MISO BC with users moving around according to a random
walk model. Recently large system analysis was applied to
investigate the BF designs for the power minimization prob-
lem with quality-of-service (QoS) targets at the users [8],
[9]. [10] applies large system analysis to reduced order ZF
beamforming for the simplified case of differently attenuated
channel covariance matrices of the users. [11], [12] proposes
the Expected Weighted Sum MSE (EWSMSE) based approach
for BF design under partial CSIT. However EWSMSE is
suboptimal and cannot even be used in the case of zero channel
mean (covariance CSIT only information). [2] proposes a
large system analysis for optimized BF with partial CSIT as
considered here. Furthermore, the channel, channel estimate
and channel error covariances can all be arbitrary and different
for all users. However, the resulting deterministic analysis is
quite cumbersome and does not allow much analytical insight.
In stochastic geometry based methods [13], the location of the
users is assumed to be random, their geographic distribution
then inducing a certain probability distribution for the channel
attenuations. Whereas most stochastic geometry work focuses
on the distribution of the attenuations, here we consider an
extension to multi-antenna systems. The multipath propagation
for the various users leads to randomized angles of arrival at
the BS which can be translated into spatial channel response
contributions that depend on the antenna array response. In the
massive MIMO regime in which the number of BS antennas
gets very large, it has been observed and exploited that despite we get for the noise vk ∼ CN (0, 1). BS bk
scaled so that ∑
complex multipath propagation, the channel covariance matrix serves Kbk = i:bi =bk 1 users. The Mbk × 1 spatial Tx filter
tends to be low rank. Exploiting the randomized nature of or beamformer (BF) is gk .
the user and scatterer positions and making abstraction of the
III. C HANNEL AND CSIT M ODEL
antenna array response, we proposed to model the user channel
subspaces as isotropically randomly oriented. This allows us For simplicity, we omit all the user indices k. We start from
to assume the eigen vectors of the channel covariance matrix a deterministic Least-Squares (LS) channel estimate
to be Haar distributed, and this identically and independently b LS = h + h,
h e (2)
for all users.
where h is the true MISO channel, and the error is modeled as
A. Contributions of this paper circularly symmetric white Gaussian noise he ∼ CN (0, σ e2 I).
In this paper: Now each MISO channel is modeled according to a correlation
• We first review optimal BFs for the expected weighted structure as follows,
sum rate (EWSR) criterion in the MaMISO limit.
h = C c, c = D1/2 c′ , (3)
• We evaluate the ergodic sum rate performance for Least-
′
Squares (LS), LMMSE and subspace projection chan- where c ∼ CN (0, IL ) and D is diagonal. Here C is the M ×L
nel estimators. Numerical results suggest that there is eigen vector matrix of the reduced rank channel covariance
substantial gain by exploiting the channel covariance Rhh = CDCH . The total sum rank across all users Np =
information compared to just using the LS estimates. ∑K
• New large system analysis for various cases of BF with Lk,c is assumed to be less than Mc , where Lk,c is the
partial CSIT is proposed, with a randomized analysis of k=1
channel rank between user k and BS c. Assuming the channel
the covariance subspaces, leading to much simpler results.
covariance subspace is known, the ( LMMSE channel
) estimate
This constitutes a marriage between large system analysis b 2 −1 b
and multi-antenna stochastic geometry.
H
can be written as h = CDC CDC + σ H
e I hLS . Ap-
• Simulation results indicate that the large system approx-
plying the matrix inversion lemma and exploiting CH C = IL ,
imations are very accurate even for small system dimen- this simplifies to
( 2 −1 )−1 H
sions and reveal the deterministic dependence of the sys- b=C σ
h e D +I C hb LS = CD b 1/2b
c, (4)
tem performance on several important scalar parameters, ( 2 −1 )−1
such as the channel multipath attenuation profile, signal where D b = σ e D +I D and bc = D−1/2 (σ 2 D−1 +
powers and SNR (whereas [2] doesn’t lead to any such −1/2 H b
I) C hLS . The posterior error covariance becomes
tractable analytical solutions). ( )−1
Notation: In the following, boldface lower-case and upper-case Rhehe = CDCH − CDCH CDCH + σ e2 I CDCH , (5)
characters denote vectors and matrices respectively. The oper- which the matrix inversion lemma allows to simplify to,
ators E(·), tr(·) , (·)H , (·)T represents expectation, trace , con- [ ( 2 −1 )−1 ] H
jugate transpose and transpose respectively. diag(·) represents Rhehe = C D − σ e D +I D C = CDC e H . (6)
the diagonal matrix formed by the elements (·). A circularly ( )
So we can write for S = Eh|hb hhH = h bhb H + Re e =
complex Gaussian random vector with mean µ and covariance hh
matrix Θ is distributed as x ∼ CN (µ, Θ). Vmax (A, B) or CWCH , where W = D b 1/2b
cb b 1/2 + D.
cH D e
Vmax (A) represents (normalized) dominant generalized eigen
IV. PARTIAL CSIT BF BASED ON D IFFERENT C HANNEL
vector of A and B or (normalized) dominant eigen vector of
E STIMATES
A respectively and λmax (A) is the corresponding max eigen
value. In the MaMIMO limit, BF design ( with) partialHCSIT will
depend on the quantities S = Eh|hb hhH = h bh b + Θ.
e We
II. MISO IBC S IGNAL M ODEL shall consider three possible channel estimates.
We consider here an IBC with C cells with a total of (i) LS Channel Estimate
K single antenna users. We shall consider a system-wide We have h b LS = h + h e where h and h
e are independent. In the
numbering of the users. User k is served by BS bk . The LS case, Θ e =σ e2 I. If we want however S to be an unbiased
received signal at user k in cell bk is estimate for hhH , then we shall take Θe = −e σ 2 I.
∑ ∑ ∑ (ii) LMMSE Channel Estimate
yk = hH
k,bk gk xk + hH
k,bk gi xi + hH
k,j gi xi +vk b+h e in which h
b and h e are decorrelated
| {z } i̸=k j̸=bk i:bi =j
We have h = h
signal bi =bk | {z } and hence independent in the Gaussian case. In the LMMSE
| {z } intercell interf. e = Ce e is the posterior covariance. The resulting
case, Θ
intracell interf. hh
(1) S = hh b bH + Θe now forms an unbiased estimate of hhH :
where xk is the intended (white, unit variance) scalar signal Ehb S = Rhh .
stream, hk,bk is the Mbk × 1 channel from BS bk to user (iii) Subspace Projection based Channel Estimate
k. The Rx signal (and hence the channel) is assumed to be We also investigate the effect of limiting channel estimation
error to the covariance subspace (without the LMMSE weight- where transition (a) represents the MaMISO limit leading to
ing, this is a simplification of the LMMSE estimate). The ESEI-WSR (Expected Signal Expected Interference WSR), uk
subspace channel estimate is given as, are the rate weights, g represents the collection of BFs gk .
b S = PC h
b LS = h + PC h
e LS , Ce e = σ sk is the (channel dependent) interference plus noise power
h hS hS e2 PC , (7)
and sk is the signal plus interference plus noise power. Their
where PC = C(CH C)−1 CH represents the projection onto conditional expectations are
the covariance subspace. Estimates for hhH : ∑ ∑
bS h
bH = rk = 1 + Eh|hb |hH
k,bi gi | = 1 +
2
giH Sk,bi gi ,
(a) Naive Subspace Channel Estimator, S = h S (10)
i̸=k i̸=k
Cbcb
cH CH ,
bS h
b H + Ce e = rk = rk + gkH Sk,bk gk , Sk,bk = Ck,bk Wk,bk Ck,bk .
(b) Subspace Channel Estimator, S = h S hS hS
C(bcb
cH + σ e2 I)CH , Further we split gk = gk′ pk , where pk is the power allocated
1/2
bS h
(c) Unbiased (S) Subspace Channel Estimator, S = h bH − to user k, and ||gk′ || = 1. By adding the Lagrange terms for the
S
CehS ehS = C(bcb
c −σ
H
e I)C .
2 H
∑C ∑
BS power constraints, µc (Pc − ||gk ||2 ), to the EWSR
A. BF with Partial CSIT c=1 k:bk =c
Three types of BF design with partial CSIT can be analyzed. in (9), we get the gradient (with αk = uk
rk , βk = uk ( r1 − 1
rk ))
k
In the case of partial CSIT we get for the Rx signal, ∑
∂EW SR
yk = hb H g k xk + h e H gk xk ∗ = αk Sk,bk gk − ( βi Si,bk + µbk I)gk = 0,
k,bk k,b ∂gk
| k{z } i̸=k
sig. ch. error (11)
∑K (8) with Si,bk = Ci,bk Wi,bk CH i,bk . This leads to the generalized
+ b H gi xi + h
(h e H gi xi ) + vk . eigen vector,
k,bi k,bi
i=1,̸=k
| {z } ∑
interf. ch. error gk′ = Vmax (Sk,bk , βi Si,bk +µbk I) . (12)
1) Naive BF EWSR: just replace h by h b in a perfect CSIT i̸=k

approach. Ignore h e everywhere. 2) Optimal BF EWSR: ac-
counts for covariance CSIT in the signal and interference
terms. 3) EWSMSE BF (Expected Weighted Sum MSE) [11]:
accounts for covariance CSIT in the interference terms, but
e term with the interference !
also associates the signal h
B. Max EWSR BF in the MaMISO limit (ESEI-WSR)
The scenario of interest here is to design optimal beam-
formers when there is only partial CSIT. Once the CSIT is
imperfect, various optimization criteria such as outage capacity
can be considered. Here the design is based on expected
weighted sum rate (EWSR) (and in a first instance with
LMMSE channel estimates). The actual EWSR represents two
rounds of averaging. In a first stage, the WSR is averaged
over the channels given the channel estimates and covariance
information (i.e. the partial CSIT), leading to a cost function
that can be optimized by the Tx. The optimized result then
needs to be averaged over the channel estimates to obtain
the final ergodic WSR. In the MaMISO limit, due to the
law of large numbers, a number of scalars converge to their
expected value, facilitating averaging the WSR. From the law
of total expectation and motivated from the ergodic capacity
formulations [14] (point to point MIMO systems), [15] (multi
user MISO systems),
EW SR = Ebh maxg EW SR(g),
∑
K
EW SR(g) = Eh|hb W SR(g) = uk Eh|hb ln(sk /sk )
While (11) can be interpreted many ways, (12) comes from
the following DC programming. Introducing the Tx covariance
H
∑ Qi = gi gi , the power constraints can be written
matrices
as k:bk =c tr {Qk } ≤ Pc . The EWSR problem is non-
concave in the Qk due to the interference terms. Therefore
finding the global optimum is challenging. In order to find at
least a local optimum, we consider the difference of convex
functions programming (DCP) approach as in [16]. Whereas
[16] however solves the Lagrange multipliers by Lagrangean
duality, here we solve them together with the powers (as in
standard water filling) in an alternating optimization approach
(alternating with optimizing the gk′ ). In DCP one keeps the
concave signal term and linearizes the convex term, leading to
a concave cost function in the Qi (or a minorizer actually in
the gi′ and pi ), which can be optimized iteratively.
EW SR = uk ln det(rk−1 rk ) + EW SRk ,
∑ K
(13)
EW SRk = ui ln(ri−1 ri ),
i=1,̸=k
where ln(rk−1 rk ) is concave in Qk and W SRk is convex
in Qk . Since a linear function is simultaneously convex and
concave, consider the first order Taylor series expansion of
W SRk in Qk around Q b (i.e. all Q
b i ).
b b b
EW{ SRk (Qk , Q) ≈} EW SRk (Qk , Q)− (14)
tr (Qk − Qb k )Abk ,
k=1  ∑
∑ ∑ b)
K K K
∂EW SRk (Qk ,Q
bk = −  βbi Si,bk .
(a)
= uk ln((Eh|hb sk )/(Eh|hb sk )) = uk ln(rk−1 rk ), where, A ∂Qk b b =
k=1 k=1 Qk ,Q i=1,̸=k
(9) (15)
Note that the linearized tangent expression for EW SRk iid with variance 1/N and independent of AN , and similarly
N →∞
N AN yN − −−−→ 0.
constitutes a lower bound for it and hence the DC approach when yN is independent of xN , that xH
is also a minorization approach. Now dropping the constant Hence the expression above goes to,
terms and reparameterizing the Qk in terms of the gk , we can ∑
write the original WSR as the Lagrangian, CHk,bk [
H
βi Ci,bk Wi,bk Ci,b k
+µbk I]−1 Ck,bk =
K [
∑ ( ) ∑
i̸=k
(19)
−1
EW SR(g) = uk ln 1 + b k gk
gkH B
1
Mbk tr{[ βi Ci,bk Wi,bk CH
i,bk +µbk I] }ILk,bk .
i̸=k
k=1
{ ( ) }] ∑C
Further we apply Lemma 6 from [3] which states
−tr gkH Ab k + µb I gk + b k = rb−1 Sk,b .
µ j Pj , B N →∞
k k k
that N1 tr{A−1
N } −
1 H −1
∑ N tr{(AN + vv H) } −−−−−1 → 0
j=1
to approximate [ i̸=k βi Ci,bk Wi,bk Ci,bk + µbk I] =
(16) ∑
H −1
(16) leads again to (11) and esp. (12). The advantage of formu- [ βi Ci,bk Wi,bk Ci,bk + µbk I] . Now using Theorem 1
lation (16) is that it allows straightforward power adaptation: i=1
√ from [3] that any terms of the form N1 tr{(AN − zIN )−1 },
substituting gk = pk gk′ in (16) and optimizing leads to the
(2)
following interference leakage (σk ) aware water filling where AN is the summation of independent rank one ma-
( )+ trices with covariance matrix Θi is equal to the unique
uk 1 ∑K
Θi
pk = − (1) , (17) positive solution of ej = N1 tr{( − zIN )−1 }. Thus
(2)
σk + µbk σk 1 + e i
∑ i=1
−1
tr{[ βi Ci,bk Wi,bk CH +µ bk I] } can be simplified as ebk ,
where (x)+ = max{0, x} and the Lagrange multipliers µc are i,bk
i̸=k
adjusted (e.g. by bisection) to satisfy the power constraints. which is defined as the solution to the following fixed point
Also, σk = gk′ H B
(1) b k g′ , σ (2) = g′ H A
b k g′ . With σ (2) = 0
k k k k k equation,
this would be standard waterfilling.  −1
1 ∑K ∑Li,c
β ζ
(r)
ec =  + µc  .
V. S TOCHASTIC G EOMETRY M A MIMO R EGIME i i,c
(20)
As argued earlier, an appropriate model for the covariance Mc i=1 r=1 1 + βi ζ (r) ec
i,c
subspaces C is a Haar distribution (randomly oriented semi-
Now (18) gets simplified as, bk = αk ebk Wk,bk bk and thus
unitary matrices). However, as we shall consider that the
bk will be the max eigen vector of Wk,bk . Finally we write
rank L remains finite, whereas M grows unboundedly, for
the optimized BF w.r.t. partial CSIT as,
the large system analysis we may equivalently consider the ∑
elements of C as i.i.d. with zero mean and variance 1/M gk = [ βi Si,bk +µbk I]−1 Ck,bk vk,bk , where,
so that the expected squared norm (and asymptotically the i̸=k (21)
actual squared norm) of the columns of C is normalized vk,bk = Vmax (Wk,bk ).
to 1. The subspaces C of different channels are considered
(r)
independent. For the large system analysis, we use Theorem Computation of eigen values ζk,bi of Wk,bi : from Section III,
1, Lemma 1, 4, 6 from ∑ [3]. From (12), we can see that gk′
e b 1/2
−1
will be of the form [ i̸=k βi Si,bk +µbk I] Ck,bk bk , where k,bi + Dk,bi , c̆k,bi = Dk,bi b
Wk,bi = c̆k,bi c̆H ck,bi , ∀i, k
′ (22)
bk = αk Wk,bk CH k,bk gk is of size Lk,bk ×1. bk can be written
as, In (3), we assume that all the eigen values are equal and
∑ e k,b = ηek,b I. Thus the eigen
positive, i.e Dk,bi = ηk,bi I, D i i
bk = αk Wk,bk CH k,bk [ βi Si,bk +µbk I]−1 Ck,bk bk . (18) (1)
values of Wk,bi can be shown to be ζk,bi = λmax (Wk,bi ) =
i̸=k 2 (2) (Lk,bi )
∥c̆k,bi ∥ + ηek,bi and ζk,bi = .... = ζk,bi = ηek,bi , where
Hence, bk∑ is the max eigen vector of ek,b
σ 2
ηk,bi
e k,b from (6).
−1 ηek,bi = i
, using the definition of D
Wk,bk CH ( β S + µ I) C . Asymptotically ek,b
σ 2 +ηk,bi i
∑k,bk i̸=k i i,b k bk
−1
k,b k i
λmax (Wk,bi ) is random since c̆k,bi is random. By the law
CHk,bk ( i̸=k βi Si,bk + µbk I) Ck,bk converges to a
deterministic limit which is a multiple of identity, ebk I, where of large numbers (assuming Lk,bi is large but finite) we
ebk is defined below. This leads to bk = Vmax (Wk,bk ). replace it by the expectation which can be computed as
b
Now we derive follows. E(λmax (Wk,bi )) = E(b k,bi Dk,bi b
cH ck,bi ) + ηek,bi . This
∑ the deterministic equivalents for terms of the
form CH k,bk [ i̸=k β i S i,b k
+ µ bk
I]−1
C k,b k
. Considering the gets simplified as, E (λmax (Wk,bi )) = Lk,bi dbk,bi + ηek,bi ,
2
where dbk,b = b k,b = dbk,b I) and
H ηk,b
eigen decomposition of Wk,bk = Vk,bk Λk,bk Vk,b k
, where i
2 from (4) (D
i ηk,bi +e
σk,b i i
(Lk,b )
Λk,bk = diag(ζk,bk , ..., ζk,bk k ) and let Ck,bk Vk,bk = C′k,bk ,
(1) i
E(bk,bi b
cH ck,bi ) = Lk,bi from (4).
with C′k,bk still being Haar matrix. Considering the
∑
term CH k,bk [
H
i̸=k βi Ci,bk Wi,bk Ci,bk + µbk I]−1 Ck,bk , VI. L ARGE S YSTEM A NALYSIS OF SINR AND P OWER
we can use Lemma 4 in Appendix VI of [3], that In this section, we derive under large system limit, with
N →∞
N AN xN − −−−→ (1/N )trAN when the elements of xN are Mc , Kc → ∞ at a fixed ratio M < 1, ∀c, approximations to
Kc
xH c
the scalar quantities involved in the rate expression, which we among the users. Substituting these values, leads to the fol-
denote as the deterministic equivalent. lowing simplified expressions for sum rate.
(1)
Theorem 1: In the large system limit, the quantities σk − Optimal ZF BF with LMMSE:
(1) Mbk →∞ (2) (2) Mbk →∞ Mb →∞
σk −−−−−−→ 0, σk − σ k −−−−−−→ 0, rk − rk −−−k−−−→
a.s a.s a.s ( )
Mbk →∞ (K−1)L
0 and rk − rk −−−−−−→ 0,
(1) (2)
where σ k , σ k , rk , rk are the RLM M SE = K ln 1 + (1 − Mbk )(Lλ1 + λ2 ) K P
.
a.s
Mbk →∞ (26)
deterministic equivalents. Here −−−−−−→ denotes almost sure Naive BF with LMMSE: D e k,b = 0 leading to ζ (1) = Lλ1
a.s k k,bk
convergence. Further we can show that, since the logarithm (r)
and rest of the eigen values ζk,bk = 0, ∀r = 2, ..., L.
is a continuous function, by applying the continuous mapping
theorem [17], it follows from the almost sure convergence of
Mb →∞ ( )
rk and rk that, Rk − Rk −−−k−−−→ 0, where Rk is the rate RnaiveLM M SE = K ln 1 + (1 − (K−1)L P
a.s Mbk )Lλ1 K . (27)
of user k, with Rk = ln( rrk ). By using similar argument, we
k
Mbk →∞ Mb →∞
state that βk − β k −−−−−−→ 0 and αk − αk −−−k−−−→ 0. The (1)
a.s a.s For naive LMMSE, maximum eigen value for Wk,bk (ζk,bk =
deterministic limits are obtained as, Lλ1 ) is lesser compared to the optima ZF BF case and hence
(1) e2b λmax (Wk,bk ) the performance degrades compared to the optimal BF case.
σk = k
e′b
, EWSMSE BF with LMMSE: In this case, the error covari-
k
(1+Υk )
Li,bk ance for the intended user get moved to the interference part,
∑
K ∑ (r)
ζi,bk
σk =
(2) 1
βi[ ], thus the sum rate expression will be,
Mbk (r) (23)
i=1,i̸=k r=1 (1 + β i ζi,bk ebk )2
r k = 1 + Υk , ( (K−1)L
)
(1− )Lλ1
e2 λmax (Wk,bk ) Mb
P
r k = 1 + Υk + pk bk e′ , REW SM SE = K ln 1 + (K−1)L
k
K . (28)
b (1− M )Lλ2 +1
k bk

where,
Lk,b Compared to naive LMMSE, there is scaling factor in the
∑
K
1 ∑i
(r)
ζk,bi
Υk = pi [ ], denominator factor accounting for the channel estimation error
Mbi r=1 (1 + βk ζ (r) ebi )2 power being moved to the noise part (8) and this explains the
i=1, k,bi (24)
i̸=k( ) performance degradation for EWSMSE.
β k = uk r1 − r1k , αk = urkk , Optimal BF with Subspace Estimator: For subspace esti-
k
mator, the sum rate expression RSubspace is similar to the
Proof: Main steps leading to this using standard results from optimal BF with LMMSE case, (29), but with different eigen
(1)
random matrix theory [3] are outlined in Appendix A. values. The eigen values can be derived as ζk,bk = L + σ e2
All the deterministic equivalents described above depend just (r)
e2 , ∀r = 2, ..., L.
and ζk,bk = σ
on the scalar parameters such as eigen values of the channel
covariance matrices, transmit powers and channel estimation
error variances. Note that the BF computation algorithm based ( )
(K−1)L
RSubspace = K ln 1 + (1 − Mbk )(L e2 ) K
+σ P
. (29)
on EWSR still remains iterative in pi , βi and µbi .

VII. VARIOUS BF S UM R ATE E XPRESSIONS FOR M ULTI

C ELL WITH LMMSE/S UBSPACE C HANNEL E STIMATOR VIII. S IMULATION R ESULTS

In this section, we consider the simplified sum rate ex-

pressions for naive, EWSMSE and EWSR ZF BFs for In this section, we present the Ergodic Sum Rate Evalua-
LMMSE/Subspace channel estimators under multi cell (C tions for BF design for the various channel estimates. Monte
ek,c
cells), with identical parameters, σ 2
= σ e2 , Lk,c = L, Carlo evaluations of ergodic sum rates are done with the
Dk,c = ηI and Mc = M, ∀k, c. Number of users in cell c following parameters: C, number of cells.∑ Kc , number of
is denoted as Kc = K/C, ∀c. We denote, (single-antenna) users in cell c and K = Kc . M , number
c
(1) η2 e2 η
σ of transmit antennas in each cell. We consider a path-wise or
ζk,bk = L + , (25)
e2 + η σ
σ e2 + η low rank channel model as in section III, with L = number
(r) e2 η
of paths = channel covariance rank. d : scale factor in the LS
e2 +η , ∀r = 2, ..., L. For
σ
and rest of the eigen values ζk,bk = σ channel estimation error variance σ e2 = d/SN R. Notations:
η2 e2 η
σ
convenience, we define the terms λ1 = η+e σ 2 and λ2 = σ e2 +η . in the figures, iCSIT refers to the optimal BF design for the
Further at high SNR, BS power becomes equally distributed instantaneous CSIT case [18].
 In Figure 3, we plot the EWSMSE beamforming perfor-
160
iCSIT
mance also and it is evident from the figure that ESEI-WSR
140 Optimal EWSR BF: LMMSE Channel Estimate based beamformers (i.e. MaMISO limit based) perform better
Optimal EWSR BF: Large System Approximation
EWSMSE BF: LMMSE Channel Estimate EWSR approximations than a EWSMSE design. From the
Sum Spectral Efficiency

120 Naive BF: LMMSE Channel Estimate

100
80
60
40
20
0 the direct link in the BF design.
0 10 20 30
Transmit SNR in dB
e2 = 0.1.
Fig. 1. EWSR for C = 1 cell, K = 10 users, M = 64, L = 4, σ
In Figure 1, we plot the optimal BF performance with
LMMSE channel estimator comparing to the optimal BF
performance for the case of large system approximation. It is
evident that the deterministic approximations are accurate even
for finite M, K. Further, we have fixed the channel estimator
error variance σe2 to be 0.1 and thus it is evident from the
figure that exploiting the channel estimation error covariance
information has significant performance gain compared to the
sub-optimal methods such as EWSMSE and naive BFs.
140
iCSIT
Optimal EWSR BF: Subspace Channel Estimate
120 Optimal EWSR BF: LMMSE Channel Estimate
Optimal EWSR BF: Unbiased Subspace Channel Estimate
naive BF: Subspace Channel Estimate
Sum Spectral Efficiency
numerical simulations in both Figures 2,3, it is quite evident
that just using LS channel estimates may lead to substantial
EWSR loss. In Massive MIMO, the exploitation of channel
subspaces (reduced rank covariances) in channel estimates
may lead to substantial reductions in SNR loss due to partial
CSIT. Moreover, there is significant gain from exploiting
(error) channel covariances in addition to (LMMSE) channel
estimates and proper handling of channel error covariance in
40 50 60 70 80
IX. C ONCLUSION
This paper investigated the optimal linear precoder based
on partial CSIT in the multi-cell MU-MISO downlink. We
introduced a stochastic geometry inspired randomization of the
channel covariance eigen spaces and analyzed the large system
behavior. This leads to simpler analytical results with SINR
or the user rates depending only on some scalar quantities
such as eigen value profile, channel rank, the number of
antennas M or users K and the channel estimation error
variance. Moreover, we show the improvement in performance
by using an LMMSE channel estimate compared to just
having LS estimates, and by furthermore properly exploiting
all covariance information. Numerical simulations suggest that
the large system approximations are accurate even for finite

100 naive BF: LMMSE Channel Estimate

naive BF: LS Channel Estimate values of M, K.
Optimal EWSR BF: LS Channel Estimate
80 Optimal EWSR BF: Unbiased LS Channel Estimate
A PPENDIX A
60
P ROOF OF T HEOREM 1
40 (1)
First we compute the deterministic equivalent for σk ,
20 (1) b k,b g′ = rb−1 g′ H Ck,b Wk,b CH g′ .
σk = rbk−1 gk′ H S k k k k k k k,bk k
0 (30)
-20 -10 0 10 20 30 40 50 60 H
Transmit SNR in dB
Using the eigen decomposition of Wk,bk = Vk,bk Λk,bk Vk,b k
,
Fig. 2. EWSR for C = 1 cell, K1 = K = 10 users, M = 64, L = 2,
e2 = 1/SN R.
σ gk′ H Sk,bk gk′ = gk′ H Ck,bk Wk,bk CH k,bk gk ,
′
′ ′′ ′′ ′′ −1
gk = gk / ∥gk ∥ , gk = Γk Ck,bk vk,bk , (31)
′ −1 ′′
||,
140
iCSIT CH g
k,bk k = C H
Γ
k,bk k C k,b k
vk,b k
/||g k
Optimal EWSR BF: Subspace Channel Estimate

120
Optimal EWSR BF: Unbiased Subspace Channel Estimate
naive BF: Subspace Channel Estimate ∑
Optimal EWSR BF: LMMSE Channel Estimate
naive BF: LMMSE Channel Estimate
where Γk = βi Si,bk +µbk I. Using large system analysis
100 EWSMSE BF: LMMSE Channel Estimate
naive BF: LS Channel Estimate i̸=k
Sum Spectral Efficiency

−1
Optimal EWSR BF: LS Channel Estimate
Optimal EWSR BF: Unbiased LS Channel Estimate simplifications shown in (20), CH
k,bk Γk Ck,bk = ebk I,
80

e2b vk,b
H
Wk,bk vk,bk e2b λmax (Wk,bk )
60 gk′ H Sk,bk gk′ = k k
= k
,
∥ ∥ ∥gk′′ ∥
′′ 2 2
gk
40
(32)
We define Γbk = Γk + βk Sk,bk . Further we consider simpli-
fying ∥gk′′ ∥ = vk,b Ck,bk Γ−2
2 H
k Ck,bk vk,bk . By using Lemma
20
k
4 from [3] leads to ∥gk ∥ = M1b tr{Γ−2
′′ 2 2
0 k } ∥vk,bk ∥
k
=
−2
Mbk tr{Γk }. Further, by using Lemma 6 we
-20 -10 0 10 20 30 40 50 60 1
Transmit SNR in dB approximate
Fig. 3. EWSR for C = 2 cells, K1 = K2 = 5 users, M = 32, L = 2,
−1
Γk ≈ (Γk + βk Sk,bk )−1 = Γ−1 bk . From [3], in the large
e2 = 1/SN R. −2
σ system limit, for (1/Mbk )tr{Γbk }, we have an almost sure
convergence value as e′bk , where e′bk is the derivative of ebk ACKNOWLEDGMENTS
w.r.t. µbk , and thus ∥gk′′ ∥ = e′bk ,
2
EURECOM’s research is partially supported by its industrial
members: ORANGE, BMW, SFR, ST Microelectronics, Sy-
K L∑
∑ βi2 ζi,bk e′bk
i,bk (r), 2 mantec, SAP, Monaco Telecom, iABG, and by the projects
e′bk = e2bk ( M1b (r)
+ 1) MASS-START (French FUI) and DUPLEX (French ANR).
k
i=1 r=1 (1 + βi ζi,bk ebk )2
e2b (33) R EFERENCES
=⇒ e′bk = k
.
K L∑
∑ i,bk (r), 2 [1] E.G. Larsson, O. Edfors, F. Tufvesson, and T.L. Marzetta, “Massive
e2
b
βi2 ζi,bk MIMO for Next Generation Wireless Systems,” IEEE Comm’s Mag.,
1− M k
bk (r) Feb. 2014.
i=1 r=1 (1 + βi ζi,bk ebk )2 [2] Wassim Tabikh, Dirk Slock, and Yi Yuan-Wu, “MIMO IBC beamform-
ing with combined channel estimate and covariance CSIT,” in IEEE
Deterministic limit for rk , rk : Each term in rk is of International Symposium on Information Theory, 2017.
the form pi gi′′ Sk,bi gi′′ / ∥gi′′ ∥ , where gi′′ H Sk,bi gi′′ =
2
[3] S. Wagner, R. Couillet, M. Debbah, and D. T. M. Slock, “Large system
′ H −1 −1 ′ ′ analysis of linear precoding in MISO broadcast channels with limited
vi,b i
Γi Ck,bi Wk,bi CHk,bi Γi vi,bi and we defined vi,bi = feedback,” IEEE Trans. Inf. Theory, vol. 58, no. 7, pp. 4509–4538, July
′
Ci,bi vi,bi . Since vi,bi is independent of all other ran- 2012.
dom quantities in this expression, we apply Lemma 4 and [4] S. Wagner and Dirk Slock, “Weighted sum rate maximization of
′ H −1 −1 ′ correlated MISO broadcast channels under linear precoding : a large
then Lemma 6 to get, vi,b i
Γi Ck,bi Wk,bi CH k,bi Γi vi,bi = system analysis,” in Proc. IEEE Int’l Workshop on Sig. Proc. Advances
−1 −1
Mbi tr{Γbi Ck,bi Wk,bi Ck,bi Γbi }. Applying Lemma 1 to
1 H
in Wireless Comm’s (SPAWC), San Fransisco, USA, 2011.
H −1 [5] S. Lakshminarayana, M. Assaad, and M. Debbah, “Coordinated multicell
each of the rows of Vk,b i
CH k,bi Γi , then Lemma 4 and 6, beamforming for massive MIMO: A random matrix approach,” IEEE
we obtain the following simplified expression, Trans. Information Theory, vol. 61, no. 6, pp. 3387 – 3412, June 2015.
[6] A. Muller, R. Couillet, E. Bjornson, S. Wagner, and M. Debbah,
1
Mbi tr{Γ−1 −1
bi Sk,bi Γbi } =
1
Mb2
tr{Γ−2 −2
bi }tr{Λk,bi Bk,bi },
“Interference-Aware RZF Precoding for Multi Cell Downlink Systems,”
i IEEE Trans. Signal Proc., vol. 63, no. 15, pp. 3959 – 3973, August
(1) (L)
where, Bk,bi = diag(1 + βk ζk,bi ebi , ..., 1 + βk ζk,bi ebi ). 2015.
(34) [7] Luca Sanguinetti, Aris L. Moustakas, Emil Bjornson, and Merouane
Debbah, “Large system analysis of the energy consumption distribution
Finally we obtain, in multi-user MIMO systems with mobility,” IEEE Trans. on Wirel.
Commun., vol. 14, no. 3, pp. 1730–1745, March 2015.
∑
K [8] R. Zakhour and S. V. Hanly, “Base station cooperation on the downlink:
pi gi′ H Sk,bk gi′ = Large system analysis,” IEEE Trans. on Info. Theo., vol. 58, no. 4, pp.
i=1, 2079–2106, April 2012.
i̸=k, [9] Y. Huang, C. W. Tan, and B. D. Rao, “Large system analysis of power
Lk,b (35) minimization in multiuser MISO downlink with transmit-side channel
∑ K
1 ∑i
(r)
ζk,bi correlation,” in Intl. Sympo. on Info. Theo. and its Applic. (ISIT), October
pi [ ] = Υk , 2012.
i=1,
Mbi r=1 (1 + βk ζ (r) ebi )2
k,bi [10] C. K. Thomas and Dirk Slock, “Reduced-order zero-forcing beamform-
i̸=k ing vs optimal beamforming and dirty Paper coding and massive MIMO
Analysis,” in 10th IEEE Sens. Arr. and Mul.chnl. Sig. Process. Wkshp.,
Thus we can write rk and rk , Sheffield, UK, 2018.
[11] F. Negro, I. Ghauri, and Dirk T. M. Slock, “Sum rate maximization
e2 λmax (Wk,bk ) in the noisy MIMO interfering broadcast channel with partial CSIT via
rk = 1 + Υk , rk = 1 + Υk + pk bk e′ , the expected weighted MSE,” in Proc. Int’l Symposium on Wireless
( ) bk
(36) Communication Systems, ISWCS, Paris, France, 2012.
Also, β k = uk r1 − r1k , αk = urkk . [12] X. Li, E. Bjornson, E.G. Larsson, S. Zhou, and J. Wang, “Multi-cell
k
MMSE precoder for massive MIMO systems and new large system
Finally, combining (32), (33), (36), we can write the deter- analysis,” in IEEE Globecom, San Diego, CA, USA, 2015.
(1) (1) e2b λmax (Wk,bk ) [13] G. George, R. K. Mungara, A. Lozano, and M. Haenggi, “Ergodic
ministic equivalent for σk as, σ k = . Eachk
e′b
Spectral Efficiency in MIMO Cellular Networks,” IEEE Trans. on
(1+Υk ) k Wireless Communications, May 2017.
term in σk is of the form βbi gk′ ′ H Si,bk gk′ ′ H / ∥gk′ ′ ∥ , which
(2) 2
[14] G. Caire and K. R. Kumar, “Information theoretic foundations of
gets simplified as follows: adaptive coded modulation,” Proc. IEEE, vol. 95, no. 12, pp. 2274–
2298, December 2007.
[15] Hamdi Joudeh and Bruno Clerckx, “Sum-rate maximization for linearly
′H
gk′ ′ H Si,bk gk′ ′ H = vk,b k
Γ−1 H −1 ′
k Ci,bk Wi,bk Ci,bk Γk vk,bk precoded downlink multiuser MISO systems with partial CSIT: A rate-
(a)
= 1
M bk tr{Γ−1 H −1
k Ci,bk Wi,bk Ci,bk Γk },
splitting approach,” IEEE Transactions on Communications, vol. 64, no.
11, pp. 4847–4861, November 2016.
(37) [16] S. J. Kim and G. B. Giannakis, “Optimal resource allocation for MIMO
′ ad-hoc cognitive radio networks,” in IEEE Trans. on Inf. Theory, May
where (a) follows from Lemma 4 (since vk,b k
is independent
2011, vol. 57, pp. 3117–3131.
of all other matrices involved). By following the same steps [17] P. Billingsley, “Probability and measure.,” in Hoboken, NJ: Wiley, 1995.
(2)
as in (34)-(35), this gets simplified and we write σ k as, [18] S. S. Christensen, R. Agarwal, E. de Carvalho, and J. Cioffi, “Weighted
sum-rate maximization using weighted MMSE for MIMO-BC beam-
∑
K
forming design,” IEEE Trans. on Wireless Commun., pp. 4792–4799,
β i gk′ ′ H Si,bk gk′ ′ H / ∥gk′ ′ ∥ =
(2) 2
σk = December 2008.
i=1,i̸=k
Li,bk (38)
∑
K ∑ ζi,bk
(r)
1
Mbk βi[ (r)
].
i=1,i̸=k r=1 (1 + β i ζi,bk ebk )2