Multi-Cell Multi-User MIMO Downlink with Partial
CSIT and Decentralized Design
Yohan Lejosne‡¶ , Atef Ben Nasser¶ , Dirk Slock‡ , Yi Yuan-Wu¶
‡ EURECOM, Sophia-Antipolis, France, Email: {lejosne,slock}@eurecom.fr
¶ Orange Labs, Issy-les-Moulineaux, France, Email: {yohan.lejosne,atef.bennasser,yi.yuan}@orange.com
Abstract—The Interfering Broadcast Channel (IBC) applies to
the downlink of multi-cell networks, which are limited by multi-
user (MU) interference. The interference alignment (IA) concept
has shown that interference does not need to be inevitable. In
particular spatial IA in the MIMO IBC allows for low latency.
However, IA requires perfect and typically global Channel State
Information at the Transmitter(s) (CSIT), whose acquisition does
not scale well with network size. Also, the design of transmitters
(Txs) and receivers (Rxs) is coupled and hence needs to be
centralized (cloud) or duplicated (distributed approach). CSIT,
which is crucial in multi-user systems, is always imperfect in
practice, especially for the intercell links. We consider mean
and covariance Gaussian partial CSIT, and the special case
of a (possibly location based) MIMO Ricean channel model.
In this paper we focus on the optimization of beamformers
for the expected weighted sum rate (EWSR) under per BS
power constraints. We apply a perfect CSI technique, based
on a difference of convex functions approach, to a number of
deterministic approximations of the EWSR, involving the Massive
MIMO (MaMIMO) limit (large number of transmit antennas),
and the large MIMO limit (both large transmit and receive
antenna numbers). We then focus on distributed techniques that
exploit local CSIT, feedback of a limited number of scalars, and
only one or few iterations.
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 transmis-
sion. 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) MIMO. 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).
Hence we focus here on the more challenging downlink (DL).
The main difficulty in realizing linear IA for MIMO I(B)C
is that the design of any BS Tx filter depends on all Rx filters
whereas in turn each Rx filter depends on all Tx filters [1].
As a result, all Tx/Rx filters are globally coupled and their
design requires global CSIT. To carry out this Tx/Rx design
in a distributed fashion, global CSIT is required at all BS
[2]. The overhead required for this global distributed CSIT
is substantial, even if done optimally, leading to substantially
reduced Net Degrees of Freedom (DoF) [3].
The recent development of Massive MIMO (MaMIMO)
[4] opens new possibilities for increased system capacity
while at the same time simplifying system design. From a
DoF point of view it may seem like a suboptimal use of
antennas. However, as shown in [5], section V, Fig. 6, the
(MaMIMO asymptotics based analytical expression for the)
optimal number of users decreases below the DoF as the
SNR decreases. Furthermore, Net DoF considerations and CSI
acquisition make the optimal number of users decrease further.
In [6], [7], MISO was considered in a single cell. Statistical
CSIT between user groups was considered and instantaneous
CSIT within user groups. The hypothesis is that some users
overlap strongly in terms of covariance subspaces but not in
terms of instantaneous CSIT. In [8] a hierarchical approach
is considered.MISO is considered and (high SNR based) user
selection also. Intercell zero-forcing (ZF) beamforming (BF)
is considered based on statistical CSIT, treating interfering
links in a binary fashion (either ZF or ignore). Intracell BF
is based on instantaneous CSIT and performs Regularized-
ZF, which is claimed to be asymptotically optimal (which is
only true for uniform user power profile). In [9], following
up on work in [10], beamspace processing is proposed, which
is the basic form of hierarchical BF. As argued in [7] also,
mmWave communications, which we target here also, facilitate
MaMIMO, and lead to a limited number of dominant paths as
they approach optics.
What is known as MaMIMO is more appropriately called
MU Massive MISO whereas here we consider actual MU
MC MaMIMO. In this paper the objective is to find the set
of beamforming (BF) vectors that maximize the Weighted
Sum Rate (WSR) of the IBC network. In [11] the alternative
problem formulation of SINR balancing is considered.
Partial CSIT formulations can typically be categorized as
either bounded error / worst case (relevant for quantization
error in digital feedback) or Gaussian error (relevant for analog
feedback, prediction error, second-order statistics information
etc.). The Gaussian CSIT formulation with mean and covari-
ance information was first introduced for SDMA (a Direction
of Arrival (DoA) based historical precedent of MU MIMO),
in which the channel outer product was typically replaced by
the transmit side channel correlation matrix, and worked out
in more detail for single user (SU) MIMO, e.g. [12]. The use
of covariance CSIT has recently reappeared in the context of
MaMIMO [13], where a not so rich propagation environment
leads to subspaces (slow CSIT) for the channel vectors so that
the fast CSIT can be reduced to the smaller dimension of the
subspace. Such CSIT (feedback) reduction is especially crucial
for MaMIMO.
The contributions here are significantly better partial CSIT
approaches compared to the Expected Weighted Sum MSE III. M AX WSR WITH P ERFECT CSIT
(EWSMSE) approach in [14] (which cannot even be used Consider as a starting point for the optimization the
in the zero channel mean case), and to present deterministic weighted sum rate (WSR)
alternatives to the stochastic approximation solution of [15].
K
We first treat the general Gaussian CSIT case. We also focus X 1
W SR = W SR(g) = uk ln (3)
on a location aided CSIT case with zero mean and identity ek
k=1
plus rank one Tx side covariance matrix and no Rx side
correlations. The goal here is to go beyond the extreme of where g represents the collection of BFs gk , the uk are rate
zero-forcing (ZF) and to introduce a meaningful beamforming weights, the ek = ek (g) are the Minimum Mean Squared
design at finite SNR and with partial CSIT, for e.g. a finite Errors (MMSEs) for estimating the xk :
Ricean factor when not much more than the (location based) 1 −1 −1 −1
LoS information of the intercell links is available at the BS. =1+gkH HH H H
k,bk Rk Hk,bk gk = (1−gk Hk,bk Rk Hk,bk gk )
ek
The other goal is to arrive at distributed approaches in which Rk = Hk,bk Qk HH H
k,bk + Rk , Qi = gi gi ,
global CSIT (channels from all BS) gets replaced by local X
Rk = Hk,bi Qi HH k,bi + INk .
CSIT (channels from own BS only) plus feedback of a limited
i6=k
number of scalars, as in [16], [17]. However, the design of (4)
MIMO systems, as opposed to MISO systems, complicates Rk , Rk are the total and interference plus noise Rx covariance
distributed designs if one wants to keep the feedback low. matrices resp. and ek is the MMSE obtained at the output
Another issue is to keep the number of iterations low, for bk = fkH yk of the optimal (MMSE) linear Rx fk ,
x
performance and feedback considerations.
fk = R−1 −1
k Hk,bk gk = Rk hk,k . (5)
II. S TREAMWISE IBC S IGNAL M ODEL The WSR cost function needs to be augmented with the power
constraints
In the rest of this paper we shall consider a per stream X
tr{Qk } ≤ Pj . (6)
approach (which in the perfect CSI case would be equivalent
k:bk =j
to per user). In an IBC formulation, one stream per user can be
expected to be the usual scenario. In the development below, in In a classical difference of convex functions (DC program-
the case of more than one stream per user, treat each stream as ming) approach, Kim and Giannakis [18] propose to keep the
an individual user. So, consider again an IBC with C cells with concave signal terms and to replace the convex interference
a total of K users. We shall consider a system-wide numbering terms by the linear (and hence concave) tangent approxima-
of the users. User k is served by BS bk . The Nk × 1 received tion. More specifically, consider the dependence of WSR on
signal at user k in cell bk is Qk alone. Then
X X X W SR = uk ln det(Rk−1 Rk ) + W SRk ,
yk = Hk,bk gk xk + Hk,bk gi xi + Hk,j gi xi +vk PK (7)
| {z } i6=k j6=bk i:bi =j W SRk = i=1,6=k ui ln det(R−1i
Ri )
signal bi =bk | {z }
intercell interf. where ln det(R−1 Rk ) is concave in Qk and W SRk is convex
| {z }
intracell interf. k
(1) in Qk . Since a linear function is simultaneously convex and
where xk is the intended (white, unit variance) scalar signal concave, consider the first order Taylor series expansion in Qk
stream, Hk,bk is the Nk × around Qb (i.e. all Q
b i ) with e.g. R
b i = Ri (Q),
b then
PMbk channel from BS bk to user
k. BS bk serves Kbk = i:bi =bk 1 users. We considering a b ≈ W SR (Q b − tr{(Qk − Q b k}
W SRk (Qk , Q) k
b k , Q) b k )A
noise whitened signal representation so that we get for the
noise vk ∼ CN (0, INk ). The Mbk × 1 spatial Tx filter or K
b k = − ∂W SRk (Qk , Q)
b X
= ui HH b −1 b −1
beamformer (BF) is gk . Treating interference as noise, user k A i,bk (Ri −Ri )Hi,bk
∂Qk b i6=k
will apply a linear Rx filter fk to maximize the signal power Q
b k ,Q
(8)
(diversity) while reducing any residual interference that would
Note that the linearized (tangent) expression for W SRk
not have been (sufficiently) suppressed by the BS Tx. The Rx
constitutes a lower bound for it. Now, dropping constant
bk = fkH yk
filter output is x
terms, reparameterizing the Qk = gk gkH , performing this
K
X linearization for all users, and augmenting the WSR cost
bk = fkH Hk,bk gk xk +
x fkH Hk,bi gi xi + fkH vk function with the constraints, we get the Lagrangian
i=1,6=k C
(2) X
X W SR(g, g
b, λ) = λj Pj +
= fkH hk,k xk + fkH hk,i xi + fkH vk j=1
i6=k K
(9)
X
uk ln(1 + gkH B
b k gk ) − gkH (A
bk + λbk I)gk
where hk,i = Hk,bi gi is the channel-Tx cascade vector. k=1
where the unknown Rx side LoS array response hr as a vector of
b k = HH R
B b −1
k,bk k Hk,bk . (10) i.i.d. complex Gaussian variables
The gradient (w.r.t. gk ) of this concave WSR lower bound is µ
hr i.i.d. ∼ CN (0, µ+1 ) and
actually still the same as that of the original WSR criterion! 0
1 1 (18)
H i.i.d. ∼ CN (0, µ+1 M ) , independent of hr ,
e
And it allows an interpretation as a generalized eigenvector
condition where the matrix H e represents the aggregate NLoS com-
Hb e 0 ||2 ) = µ can
ponents. Note that ( E||hr hTt (θ)||2F )/( E||H
Bb k gk = 1 + gk Bk gk (A b k + λb I)gk (11) F
uk k be considered as a Rice factor. In fact the only parameter
0 additional to the LoS AoD θ assumed in (17) is µ. So, this is
or hence gk = Vmax (B b k, A
b k + λb I) is the (normalized)
k a case of zero mean CSIT and Tx side covariance CSIT
”max” generalized eigenvector of the two indicated matrices,
with max eigenvalue σk = σmax (B b k, Ab k + λb I). Let σ (1) = µN N 1
k k Ct = E HH H = ht (θ)hH
t (θ) + IM . (19)
0
gkH B
0 0
b k g , σ (2) = g H A 0
b k g . The advantage of formulation µ+1 µ+1M
k k k k
(9) is that it allows straightforward power adaptation: intro- V. E XPECTED WSR (EWSR)
√ 0
ducing stream powers pk ≥ 0 and substituting gk = pk gk
in (9) yields For the WSR criterion, we have assumed so far that the
channel H is known. The scenario of interest however is that
C K
X X (1) (2) of partial CSIT, e.g. perfect or good partial intracell CSIT but
W SR = λ j Pj + {uk ln(1 + pk σk ) − pk (σk +λbk )}
very partial (zero mean, e.g. LoS) CSIT of the intercell links.
j k=1
(12) Once the CSIT is imperfect, various optimization criteria could
which leads to the following interference leakage aware water be considered, such as outage capacity. Here we shall consider
filling the expected weighted sum rate EH W SR(g, H) =
!+
uk 1
X
−1
pk = − (1) (13) EW SR(g) = EH uk ln(1 + gkH HHk,bk Rk Hk,bk gk )
(2)
σk + λbk σk k
(20)
where the Lagrange P multipliers are adjusted to satisfy the where we now underlign the dependence of various quantities
power constraints k:bk =j pk = Pj . This can be done by on H. The EWSR in (20) corresponds to perfect CSIR
bisection and gets executed per BS. Note that some Lagrange since the optimal Rx filters fk as a function of the ag-
multipliers could be zero. Note also that as with any alternating gregatePH have been substituted, namely W SR(g, H) =
optimization procedure, there are many updating schedules maxf k uk (− ln(ek (fk , g))). At high SNR, max EWSR at-
possible, with different impact on convergence speed. The tempts ZF.
0
quantities to be updated are the gk , the pk and the λl . In [15] a stochastic approximation approach for maximiz-
IV. M EAN AND C OVARIANCE G AUSSIAN CSIT ing the EWSR was introduced. In this approach the statistical
average gets replaced by a sample average (samples of H
In this section we drop the user index k for simplicity. The
get generated according to its Gaussian CSIT distribution in a
separable correlation model is
Monte Carlo fashion), and one iteration of the min WSMSE
H = H + C1/2 e C1/2
H (14) (Weighted Sum MSE) approach gets executed per term added
r t
1/2 1/2 in the sample average.
where H = E H, and Cr , Ct
are Hermitian square-roots Some issues with this approach are that in this case the
of the Rx and Tx side covariance matrices number of iterations may get dictated by a sufficient size for
E(H − H)(H − H)H = tr{Ct } Cr the sample average rather than by a convergence requirement
(15)
E(H − H)H (H − H) = tr{Cr } Ct for the iterative approach. Another issue is that this approach
converges to a local maximum of the EWSR. It is not
and the elements of H e are i.i.d. ∼ CN (0, 1). It is also of
immediately clear how to combine this stochastic approxi-
interest to consider the total Tx side correlation matrix mation approach with deterministic annealing. Deterministic
H
Rt = E HH H = H H + tr{Cr } Ct . (16) annealing can be used as in [1] for a deterministic algorithm
as in Section III to track the global optimum from SNR ≈ 0
A. Location Aided Partial CSIT LoS Channel Model (where the solution is clear analytically) to the desired SNR.
Assuming that for certain (e.g. intercell) links a BS disposes This is essentially a homotopy method in which the problem
of not much more than the LoS component information, gets resolved for an SNR that increases in small steps. At
consider the following MIMO channel model each higher SNR, the global optimum will be in the region of
e0 attraction of the global optimum at the lower SNR.
H = hr hH
t (θ) + H (17)
In the rest of this paper we discuss various deterministic
where θ is the LoS Angle of Departure (AoD) and the BS side approximations for the EWSR, which can then be optimized
array response is normalized: ||ht (θ)||2 = 1. We shall model as in the full CSI case.
 VI. M A MIMO L IMIT which is now similar to the corresponding interpretation in (9).
If the number of Tx antennas M becomes very large, then Note that B̆k corresponds to a total Tx side correlation matrix
quantities of the form HR−1 HH converge to their mean (by as in (16), but now for an interference plus noise whitened
−1/2
the LLN). Hence in the MaMIMO limit, the WSR converges channel R̆k Hk,bk . The linearization of W SRk w.r.t. Qk
to a deterministic limit that depends on the distribution of the now involves
channels. The actual statistical distribution of the channel is K
X h H  
one thing. The CSIT distribution as in Section IV is another. Ăk = ui Hi,bk R̆i−1 − R̆−1
i Hi,bk
The Txs have no choice but to design their BFs according to i6=k (27)
  i
their partial CSIT. Then to get the actual resulting WSR, the +tr{ R̆i−1 − R̆−1 Cr,i }Ct,i,bk .
i
BFs designed with the partial CSIT need to be evaluated with
the actual channel distribution. The rest of the development is now completely analogous to
Now, for the design with partial CSIT, the WSR will also the case of perfect CSIT. Note that for general Gaussian partial
converge to a deterministic limit in the MaMIMO regime. We CSIT we get e.g.
get a convergence for any term of the form
M →∞ H R̆−1 − R̆−1
i = R̆−1 −1
i (R̆i − R̆i )R̆i
HQHH −→ EH HQHH = HQH + tr{QCt } Cr . (21) i
H (28)
= R̆−1 H H −1
i (gi Ct,i,bi gi I + Hi,bi gi gi Hi,bi )R̆i .
In what follows we shall go one step further in the separable
channel correlation model and assume Cr,k,bi = Cr,k , ∀bi . VII. L ARGE MIMO A SYMPTOTICS
This leads us to introduce
The large MIMO asymptotics from [19], [20], in which
1/2 e 1/2
Hk = [Hk,1 · · · Hk,C ] = Hk + Cr,k H k Ct,k both M, N → ∞ at constant ratio, tend to give more precise
X 
Qi approximations when M is not so large. For the general
C X case of Gaussian CSIT with separable (Kronecker) covariance
i:bi =1 ..  X
Q= . X = Ij Qi IH (22)
 
 j structure, [19], [20] lead to asymptotic expressions of the form
 Q  j=1
i i:bi =j
H
i:bi =C EH ln det(I(+ HQH " ) # )
Qk = Q − H
Ibi Qi Ibi I + wCr H
= max ln det H − zw .
z≥0, w≥0 −QH I + zQCt
where Ct,k = blockdiag{Ct,k,1 , . . . , Ct,k,C }, and Ij is an all
(29)
zero block vector except for an identity matrix in block j.
where the maximization over z and w should be carried
Then we get for the W SR (= EW SR), using (21),
out alternatingly (and not jointly: the joint optimization may
K
X correspond to a global maximum or a saddle point; the cost
W SR = uk ln det(R̆k−1 R̆k ) (23) function is concave however in z or w separately). We shall
k=1 assume the same fully separable correlation Gaussian channel
where model as in (22). The EWSR with large MIMO asymptotics
H now becomes
R̆k = INk + Hk QHk + tr{QCt,k } Cr,k
H (24) K
R̆k = INk + Hk Qk Hk + tr{Qk Ct,k } Cr,k X 
EW SR = uk max {ln det Sk (Q, zk , wk ) − zk wk }
This leads to zk ,wk
k=1 
 H
W SR = uk ln det(I + R̆−1

k
Hk,bk gk gkH Hk,bk − max ln det Sk (Qk , zk , wk ) − zk wk

(25) zk ,wk
+tr{gk gkH Ct,k,bk }Cr,k ) + W SRk . (30)
where
Whereas the generalized eigenvector condition for the BF gk " #
can be worked out from here, the adaptation of the stream I + wCr,k Hk
Sk (Q, z, w) = H . (31)
power pk gets a bit complex for this case of general Gaussian −QHk I + zQCt,k
CSIT. So, for what follows, consider the simplified case of Note that
a ”Ricean factor” µ in the Gaussian partial CSIT model that
behaves proportionally to the SNR, for the direct links Hk,bk ln det Sk (Q, z, w) = ln det(I+wCr,k )+ln det(I+QTk (z, w))
(only). This corresponds to the case of Hk,bk representing a H
with Tk (z, w) = zCt,k + Hk (I + wCr,k )−1 Hk
channel estimate, with properly organized (intracell) channel (32)
estimation and feedback. In that case the dependence of the where Tk plays the role of some kind of total Tx side channel
user k rate on its BF in (25) can be approximated by correlation matrix. Note that the weighting coefficients z, w
W SR = uk ln det(I + gkH B̆k gk ) + W SRk with depend on the BFs also though. The EWSR expression in (30)
H
(26) can be maximized alternatingly over the {gk }, the {zk , wk }
B̆k = Hk,bk R̆k−1 Hk,bk + tr{Cr,k R̆−1
k
}Ct,k,bk and the {zk , wk }. For the optimization of the BFs gk , for given
z, w, introduce with the value obtained in the previous iteration for g. The
corresponding optimal z is then z = N/(1 + w).
R
b
k,k = I + Qk Tk (zk , wk )
b
Rb = I +Q b Tk (z , w ) (33) VIII. C ENTRALIZED IBC D ESIGN
k k k k
Rb = I +Q b k Tk (zk , wk ) . We shall first consider centralized optimization approaches
k
(as in e.g. cloud RAN) in which CSIT information and BF
Then the generzalized eigenvector approach and the interfer- computation gets centralized. In this case we consider the
ence aware WF of the perfect CSI case can be applied with BSs in one cluster, and the users they serve, and one cluster
B b −1
b k = Tk (zk , wk ) R forms an IBC. In a centralized IBC design, the CSIT training
k,k and feedback (for the FDD case) leads to a CSIT acquisition
X  b −1 − T (z , w ) R
b −1 .

(34) overhead that remains linear in C, the number of BS [2]. Also
Ak =
b ui Ti (zi , wi ) R i i i i i
i6=k in the TDD case, though in this case only uplink training is
required. The proper partial CSIT model instance in this case
Note that in spite of their appearance, matrices of the form
will have a strong non-zero mean H which represents the
TR−1 are symmetric. Indeed, if e.g. T is invertible then
channel estimate, and the weaker covariance part represents
T(I + QT)−1 = (T−1 + Q)−1 . For the optimization
estimation (and feedback) noise. In practice however, the
max {ln det S(Q, z, w) − zw}, we get from the extremum
z≥0,w≥0 cluster will not be isolated, and the intercluster interference
conditions can be handled similarly to intercell interference. However,
w = f (Q, z, w) = tr{QCt [I + QT(z, w)]−1 } the intercluster CSIT may be more degraded, and is possibly
H
z = g(Q, z, w) = tr{Cr [I +wCr H(I +zQCt )−1 QH ]−1 } only composed of Tx side covariance information. Note that
(35) in the MaMIMO regime, strong direct link CSIT with zero-
which can be iterated until a fixed point. mean cross link CSIT induces quite some structure e.g. in the
For the simpler case of zero channel means Hk = 0 and computation of the matrices Ăk (see (28)) and leads to even
no Rx side correlations Cr = I, and with per user Tx side more simplification in the Interference Channel (not IBC) case.
correlations Ck , the EWSR can be rewritten with large MIMO The MaMIMO IBC with zero mean and Cr = I intercell
asymptotics as CSIT scenario leads to the same rate expressions as if the
intercell interference is modeled as Gaussian while preserving
EW SR =
K  the covariance matrix:
X P
uk max [ln det(I +zk QCk )+N ln(1+wk )−zk wk ] yk = Hk,bk gk xk + i:bi =bkHk,bk gi xi + v ek
zk ,wk (40)
k=1
Cvek vek = (1 + i:bi 6=bk giH Ct,k,bi gi ) I .
 P
 
−uk max ln det(I +zk Qk Ck )+N ln(1+wk )−zk wk .
zk ,wk In general, modeling the intercell interference plus noise as a
(36) Gaussian noise v
ek would lead to a mutual information lower
This criterion can be used to evaluate the EWSR for given bound. Also the large MIMO formulation of Section VII can
Q. It can also be used to optimize Q, in which case we can be used.
apply again the algorithm of Section III, with the following
conventions: IX. D ISTRIBUTED IBC D ESIGN
R
b (z) = I + z Q
k
b Ck , R
b (z) = I + z Q b Ck In this case, global intracluster CSIT can also be gathered
k k
−1 X b −1(z ) − z R
b −1 but it takes an overhead that evolves with C 2 [2]. Hence, high
B
b k = zk Ck Rb (z ) , A
k k
bk = ui C i (z R
i i i i i (zi )) quality (high Ricean factor) intracell CSIT and Tx covariance
i6=k only intercell CSIT may be a more appropriate setting. For
(37)
what follows we shall assume the LoS Tx intercell CSIT. We
where zk , zk are obtained from
shall focus on a MaMIMO setting.
max g(zk , wk , Q, Ck ) , max g(zk , wk , Qk , C) where
zk ,wk zk ,wk A. Initialization Phase
g(z, w, Q, C) = ln det(I + zQC) + N ln(1 + w) − z w .
The approach considered here is non-iterative, or could be
(38)
taken as initialization for further iterations.
For these optimizations, we get from ∂g/∂w = 0 that z =
1) Iteration 0: To properly gauge the intercell interference
N/(1 + w). From this and ∂g/∂z = 0 we get
caused by the other BS, we shall start with a per cell design. In
1+w 1+w the case of multiple Tx and Rx antennas, different WSR local
w = f (w) = tr{( IM + QC)−1 QC} (39)
N N optima correspond at high SNR to different distributions of
The curves y = w and y = f (w) have a unique intersection the zero forcing (ZF) roles between the various Txs and Rxs.
in the first quadrant, with y = f (w) lying initially above y = To simplify design, we shall assume here that Rx antennas
w. Hence the optimal w can be found by iterating w(i) = are used to handle intracell interference. Hence all intercell
f (w(i−1) ). The first time, one can initialize with w(0) = 0. In interference needs to be handled by Tx (BS) antennas. In
the iterative algorithm from Section III, w can be initialized that case, the crosslinks (cascades of channel and Rx) can
be considered as independent from the intracell channels. ACKNOWLEDGMENTS
In a MaMIMO setting, the ZF by BS j towards K − Kj EURECOM’s research is partially supported by its industrial
crosslink channels (or LoS components in fact) will tend to members: ORANGE, BMW Group, SFR, ST Microelectron-
have a deterministic effect of reducing the effective number ics, Symantec, SAP, Monaco Telecom, iABG, and also by the
of Tx antennas by this amount and hence of reducing the Tx EU FP7 projects ADEL and NEWCOM#.
Mj
power by a factor Mj −(K−K j)
. Hence a per BS design can R EFERENCES
be carried out with (partial) intracell CSIT, with BS Tx power [1] F. Negro, I. Ghauri, and D.T.M. Slock, “Deterministic Annealing Design
Mj
Pj replaced by Mj −(K−K j)
Pj , and with all intercell links and Analysis of the Noisy MIMO Interference Channel,” in Proc. IEEE
Hk,bi = 0, bi 6= bk . The power reduction factor considered Information Theory and Applications workshop (ITA), San Diego, CA,
USA, Feb. 2011.
corresponds to ZF or hence a high SNR assumption. At a [2] F. Negro, D.T.M. Slock, and I. Ghari, “On the Noisy MIMO Interference
finite SNR, the user k SINR reduction will be less because Channel with CSI through Analog Feedback,” in Int’l Symp. Comm’s,
the optimal BF would do some regularized ZF of intercell Control and Sig. Proc. (ISCCSP), Rome, Italy, May 2012.
[3] Y. Lejosne, D. Slock, and Y. Yuan-Wu, “Foresighted Delayed CSIT
interference caused, but the SINR would on the other hand be Feedback for Finite Rate of Innovation Channel Models and Attainable
smaller because the intercell interference is not ZF’d. We shall NetDoFs of the MIMO Interference Channel,” in Proc. Wireless Days,
assume that these two opposite effects roughly compensate Valencia, Spain, Nov. 2013.
[4] E.G. Larsson, O. Edfors, F. Tufvesson, and T.L. Marzetta, “Massive
each other. This first step (which is itself an iterative design MIMO for Next Generation Wireless Systems,” IEEE Comm’s Mag.,
for the scenario considered with reduced Tx power and no Feb. 2014.
intercell links) leads to BFs g(0) which, from (24), lead to [5] S. Wagner, R. Couillet, M. Debbah, and D. Slock, “Large System
Analysis of Linear Precoding in Correlated MISO Broadcast Channels
(0) (0) H Under Limited Feedback,” IEEE Trans. Info. Theory, Jul. 2012.
R̆k = (1 + tr{Qbk Ct,k,bk })INk + Hk,bk Qbk Hk,bk , [6] A. Adhikary, J. Nam, J.-Y. Ahn, and G. Caire, “Joint Spatial Division
(0)
X (0) (0)H and Multiplexing: the Large-Scale Array Regime,” IEEE Trans. Info.
where Qbk = gi gi Theory, Oct. 2013.
i:bi =bk [7] A. Adhikary, E. El Sadafi, M. Samimi, R. Wang, G. Caire, T. Rappaport,
(41) and A. Molisch, “Joint Spatial Division and Multiplexing for mm-Wave
Channels,” [Online]. Available: http://arxiv.org/abs/1312.2045.
and similarly for R̆k . [8] A. Liu and V.K.N. Lau, “Hierarchical Interference Mitigation for Large
2) Iteration 1: The idea is now to do one iteration in order MIMO Cellular Networks,” IEEE Trans. Signal Processing, Apr. 2013.
to adjust the Tx filters for the intercell interference. So, with [9] J. Brady and A. Sayeed, “Beamspace MU-MIMO for High-Density
Gigabit Smallm Cell Access at Millimeter-Wave Frequencies,” in Proc.
the initial BFs g(0) , the local intercell CSIT Ct,i,bk also, the IEEE Workshop on Signal Processing Advances in Wireless Communi-
correct power constraints, and R̆k , R̆k as in (41), we get B̆k cations (SPAWC), Toronto, Canada, Jun. 2014.
as in (26) and Ăk from (27) becomes [10] A. Sayeed, “Deconstructing Multiantenna Fading Channels,” IEEE
Trans. Sig. Proc., Oct. 2002.
[11] F. Negro, M. Cardone, I. Ghauri, and D.T.M. Slock, “SINR Balancing
X h H  
Ăk = ui Hi,bk R̆−1i
− R̆ −1
i Hi,bk and Beamforming for the MISO Interference Channel,” in Proc.
i6=k:bi =bk   i IEEE Int’l Symp. Personal Indoor and Mobile Radio Communications
(PIMRC), Sept. 2011.
+tr{ R̆−1i
− R̆ −1
i C r,i }C t,i,bk (42) [12] R. de Francisco and D.T.M. Slock, “Spatial Transmit Prefiltering for
X
−1 −1 Frequency-Flat MIMO Transmission with Mean and Covariance Infor-
+ ui tr{R̆i − R̆i } Ct,i,bk . mation,” in IEEE Asilomar Conf, on Signals, Systems, and Computers,
i:bi 6=bk | {z }
Pacific Grove, US, Oct. 2005.
= µi
[13] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach
Hence the only information that needs to be fed back from user to channel estimation in large-scale multiple-antenna systems,” IEEE
J. on Selected Areas in Comm’s (JSAC), Special Issue on: Large Scale
i in another cell is the positive scalar µi . This is related to the Antenna Systems, Jan. 2013.
interference pricing in game theory [17]. The normalized BFs [14] F. Negro, I. Ghauri, and D.T.M. Slock, “Sum Rate Maximization in the
0
are then computed as gk = Vmax (B̆k , Ăk + λbk I) where the Noisy MIMO Interfering Broadcast Channel with Partial CSIT via the
Expected Weighted MSE,” in IEEE Int’l Symp. on Wireless Commun.
λbk are taken from the previous iteration. The stream powers Systems (ISWCS), Paris, France, Aug. 2012.
are obtained from (13). [15] M. Razaviyayn, M.S. Boroujeni, and Z.-Q. Luo, “A Stochastic Weighted
MMSE Approach to Sum Rate Maximization for a MIMO Interference
B. Continuous Adaptation Channel,” in IEEE workshop on Signal Processing Advances in Wireless
Communications (SPAWC), Darmstadt, Germany, June 2013.
The solution discussed above should be adequate in many [16] Y. Huang, G. Zheng, M. Bengtsson, K.K. Wong, L. Yang, B. Ottersten,
“Distributed Multicell Beamforming Design Approaching Pareto Bound-
scenarios. Further refinements can be considered. For instance, ary with Max-Min Fairness,” IEEE T. Wireless Comm., Aug. 2012.
staying with covariance intercell CSIT but going beyond LoS, [17] W. Xu and X. Wang, “Pricing-Based Distributed Downlink Beamform-
the subspaces of the Ct to be accounted for in the initialization ing in Multi-Cell OFDMA Networks,” IEEE J. Sel. Areas Comm., Oct.
2012.
stage would be more than one dimensional, leading to a [18] S.-J. Kim and G.B. Giannakis, “Optimal Resource Allocation for MIMO
further power reduction in the initial per cell design. Further Ad Hoc Cognitive Radio Networks,” IEEE T. Info. Theory, May 2011.
iterations beyond the first iteration discussed above require [19] J. Dumont, W. Hachem, S. Lasaulce, P. Loubaton, and J. Najim, “On
the Capacity Achieving Covariance Matrix for Rician MIMO Channels:
for each iteration the feedback of the scalars µi characterizing An Asymptotic Approach,” IEEE Trans. Info. Theory, Mar. 2010.
the intercell links, and within each cell, the feedback of the [20] G. Taricco, “Asymptotic mutual information statistics of separately
−1/2 correlated Rician fading MIMO channels,” IEEE Trans. Info. Theory,
interference plus noise whitened channels R̆k Hk,bk and
Aug. 2008.
the scalars tr{Cr,k R̆−1
k
}.