DOWNLINK SCHEDULING METHOD FOR MIMO / MISO CELLUAR SYSTEMS WITH LIMITED FEEDBACK SIGNALLING TRIGGERED BY EXCEEDING A SINR THRESHOLD INDEPENDENTLY ESTIMATED BY THE MOBILE STATIONS

FIELD OF THE INVENTION

The present invention relates to the field of telecommunication networks, and more precisely to a downlink (DL) scheduling method for MIMO / MISO cellular systems with limited feedback signalling triggered by exceeding a SINR threshold independently estimated by the mobile stations. The meaning of used acronyms are reported in APPENDIX D at the end of the description.

BACKGROUND ART

Researchers have recently focused the attention on the improvement of throughput performances on broadcast (downlink) channel in multi-user cellular systems, where the Base Stations (BS), and eventually the Mobile Stations (MS), employ multiple antennas. References [1] to [9] to the most relevant background art are reported in APPENDIX C at the end of the description.

In [Ref.1] the authors demonstrate that for a system with M transmitting antennas, N receiving antennas and K users the optimal performance is achieved by Dirty Paper Coding (DPC) described in [Ref.2], with a sum-rate scaling law of M \og\og(NK) described in [Ref.3], Anyway DPC is not practical as it requires the instantaneous full Channel State Information (CSI) at the BS. The need for partial-CSI and computationally- effective transmission schemes led to several nearly-optimal algorithms.

In [Ref.4] the authors proposed to use a random transmit beamforming beam in MISO systems, to leverage maximal diversity gain in a densely populated cell. The transmit beamforming is regularly updated, to improve multi-user diversity. During a training phase, each receiver measures its received SNR on the beamforming vector currently employed by the BS, and feeds it back to the BS on an UL signalling channel. Scheduling is performed by the BS with a proportional fair scheduler (PFS), choosing the best user with regards to MSs' feedbacks and to fairness constraints. The feedback given by each user consists of its SNR (one real value per active user per channel/beamforming realization). This scheme allows high multi-user diversity and beamforming gains, but does not employ multiplexing gain.

In [Ref.5] the authors proposed to extend the [Ref.4] approach to spatial multiplexing in MISO and MIMO systems. The BS with M antennas transmits M random orthogonal beams. Every user measures its SINR on each beam, and reports the best

value and the corresponding beam index to the BS (one real number + log2(M+1 ) bits of feedback per user). The model is analysed with regards to asymptotic capacity. Fairness in the resource assignment and scheduling are considered under very simplified conditions, which are unacceptable for real systems. An Opportunistic-SDMA (O-SDMA) scheme was proposed to asymptotically achieve for K — » ∞ the sum-rate scaling law of DPC with low required feedback, for example, lower than that of full-CSI zero-forcing SDMA schemes of [Ref.6].

In [Ref.7] and [Ref.8] the authors proposed to modify the [Ref.5] scheme, limiting the feedback to one bit per active user. Every user compares its SINR with a certain threshold, and reports logic 1 if the received SINR exceeds that threshold or 0 if the SINR is below the threshold. The system model is exactly the same as in [Ref.5]. This leads to an unrealistic solution, which is by the way useful for analytical studies. The proposed threshold is a network parameter equal for all users, it is suboptimal and holds only when asymptotically infinite users are in the cell. The scheme at [Ref.7-8] becomes also less efficient when M is high, because the received SINR is evaluated on an arbitrary transmit beam only.

In [Ref.9] the authors analyze the performance of a scheme similar to that of

[Ref.4], with a reduced feedback approach. Users give feedback (which consists of a real number) only when their SNR exceeds a certain threshold, which is evaluated in a distributed way. Anyway no explicit algorithms or closed-form are proposed in the article for the evaluation of the optimal threshold.

OBJECTS OF THE INVENTION

In view of the presented state of the art, it is an object of the present invention to provide each MS with an algorithm for estimating its own SINR threshold for the aim of triggering a limited feedback to the BS in support of the BS's activity of scheduling the transmissions on the broadcast channel (downlink). The estimating algorithm should take into account throughput and fairness requirements in a realistic scenario, where single and multiple antennas terminals coexist in the cell, and where channels are unbalanced for different users, and no signalling by the base station is requested. Another object of the present invention is that to provide the BS with a new scheduling algorithm exploiting the previous concepts developed for the distributed thresholds together with additional fairness constraints to share the sum-rate among active users. In the following description terms as mobile station MS and the respective user are considered to have an equivalent meaning.

SUMMARY OF THE INVENTION

The invention achieves said objects by providing in a cellular communication system a method for scheduling downlink transmissions, subdivided in time slots, from M beamforming antennas connected to base station towards K mobile stations equipped with an arbitrary number N _k of receiving antennas, including the following steps: a) under the assumption of a probability F equal for every beam m and for every user k to transmit a feedback signalling, estimating, by every mobile station k independently to each another an individual SINR threshold a _k using the following

equation: } the average power received by the user k , and P _out a system parameter corresponding to the scheduling outage probability, that is, the probability that no user exceeds its threshold on a given beam; b) generating on every time slot, by the base station, a random set of orthogonal beamforming signal vectors S = [S ₁ -S _M ] ; c) estimating, by every mobile station its relevant SINR value γ _{k m} on every beam s _m during a downlink training step; d) calculating, by every mobile station, its best beam-index rh _k as m _k = argmax{f _{i m} }; m e) comparing, by each mobile station, its best SINR value γ _{k ή} with its own threshold

f) signalling to the base station, by every mobile station by which γ _{k ύ} > a _k is valid, the best beam-index m _k during a feedback uplink transmission, otherwise signalling a beam-index null; g) selecting, by the base station, on each beam m at most one mobile station among those reporting a non-null signalling on that beam; h) transmitting, by the base station, to the selected users with conservative rates log(\ + a _k , ) expressible in bit/s/Hz, where k _m ^* is the user eventually scheduled on beam m ; i) repeating from step a) if a remarkable change in {p _k } by user k or in the number of active users K is occurred within the last time slot, otherwise repeating from step

b), as disclosed in claim 1. Moreover, considering that:

- the BS does not have any knowledge about the instantaneous user channel realizations {ϋ _k }, and - the mean channel power ρ _k can be obtained by the BS either using explicit feedback from the MS or estimated during the uplink if the system is TDD, it descends that the base station, using the same threshold estimating algorithm previously described, is able to calculate by itself the entire set of thresholds {a _k } so as to transmit the set to the mobile stations, that are so prevented to calculate their individual threshold from themselves, simplifying the firmware and saving costs.

So another object of invention is a method for scheduling downlink transmissions by which the whole set of reporting thresholds {a _k } is calculated by the base station and transmitted to the mobile stations, as described in an independent claim 2. The calculation of the whole set of thresholds {a _k } by both the base station and the mobile stations is another possibility.

Additional features of the present invention which are believed to be novel are set forth with particularity in the appended claims.

According to some other aspects of the invention:

• The threshold set \a _k \ depends on a long-term statistic, so there is no need to often update it.

• Since {p _k } are slowly varying, their update rate can be low, so we can neglect the impact of their signalling on the overall uplink throughput.

• The first member of the equation for determining the threshold a _k increases monotonically with a _k . • We assume that each MS knows the number K of active users in the cell and has perfect knowledge of all the equivalent channel vectors [q _Km = H _k s _m }, obtained, i.e. during DL preamble equalization.

• A closed-form lower-bound a _k for the SINR threshold a _k valid for random scheduling can be calculated by the following expression:

• The user may be chosen randomly by the base station among those that report a feedback on beam m or preferably according to a scheduling metric. In the event that no user gives feedback on a certain beam, the BS does not schedule any data flow on that beam. Profitably, a Proportional Rate Scheduling (PRS) algorithm is used by which every user which report with equal probability (EP) is expected to receive on the average a given fraction R _k of the sum-rate R _SUM (corresponding the overall throughput from the BS expressible in bit/s/Hz), so that the average rate of user k in bit/s/Hz is R _k = R _k R _SUM V^ e [I ₅ A ^" ]. The PRS algorithm finds the set (k)j that maximizes the sum-rate under the EP constraints, being P ^s (k) the scheduling probability of user k .

• The BS transmits to a scheduled user k with a conservative rate \og(l + a _k . ) lower than Shannon's limit \og(\. + χ _{k lfι} ) , but this is not a limitation and a greater rate can be used especially in case the individual threshold a _k is largely exceeded.

• The transmission rate can be adjusted to meet the variability on the set of thresholds {a _k } by selecting one among various possible Modulation and Coding Schemes

(MCS). In particular, an MCS can be selected according to the threshold a , ,

where k _m is the user eventually scheduled on beam m .

ADVANTAGES OF THE INVENTION

The method of the invention is shown to perform nearly optimally in terms of sum- rate when many users are in the cell, even if a very low feedback (signalling) rate is employed on the UL channel.

The method in subject is able to accomplish fairness in terms of predefined sum- rate fractions. Conversely, realistic fairness constraints have been neglected in opportunistic-SDMA approaches so far.

The subjected method can easily accommodate both MISO and MIMO MSs in the cell, in a totally transparent way; this is not so straightforward in some competitor algorithms.

The subjected method is distributed, so no particular knowledge about the MS receiver performances is needed at the BS.

BRIEF DESCRIPTION OF THE DRAWINGS

The features of the present invention which are considered to be novel are set forth with particularity in the appended claims. The invention and its advantages may be understood with reference to the following detailed description of an embodiment thereof taken in conjunction with the accompanying drawings given for purely non-limiting explanatory purposes and wherein:

- fig.1 shows a MIMO system model suitable for implementing the downlink scheduling method of the present invention;

- figures 2 to 6 shows some comparative curves of simulation results achievable by the downlink scheduling method of the present invention.

DETAILED DESCRIPTION OF AN EMBODIMENT OF THE INVENTION

In order to better clarify the complex arguments of the description, it is subdivided in the following Sections:

• SYSTEM MODEL - to introduce the system model. • EP ALGORITHM - to present the general solution of the Equal Probability algorithm.

Includes a subsection A - THRESHOLD CALCULATION FOR MIMO NETWORKS.

• EVALUATION OF THE OUTAGE PROBABILITY - to show that EP achieves the asymptotical scaling law of DPC and we evaluate the optimal thresholds and nearly-optimal thresholds with a simplified closed-form solution for a finite users number. Includes subsection A - SCALING LAW OF THE SUM RATE and subsection B - OPTIMAL OUTAGE PROBABILITY.

• FAIRNESS CONSTRAINTS - to introduce fairness constraints.

• SCHEDULING ALGORITHM - to propose the PRS algorithm.

• NUMERICAL RESULTS - to numerically validate the analysis through simulations. • CONCLUSIONS - to draw the conclusions.

• APPENDIX A - to calculate a lower-bound ά _k in closed-form for the EP SINR threshold a _k .

• APPENDIX B - to write a closed -form approximation of the outage probability F .

• APPENDIX C - Includes some Bibliographic references.

• APPENDIX D - Includes the used acronyms.

SYSTEM MODEL With reference to fig.1, we consider the downlink (DL) of a heterogeneous MIMO / MISO cellular network where the Base Station (BS) employs M transmit antennas and each of the K mobile stations (MS) employs N _k antennas. Every user is expected to receive (on a long term average) a fraction R _k of the sum rate R _SUM , so that the average rate of user k in bit/s/Hz is: R _k = R _k R _SUM Vk e [l,K] Eq.1

The narrowband N _k ^χ M channel H _k between the BS and the k-th MS is modelled as a Rayleigh i.i.d. matrix variable where each entry is ~ CN(O, p _k ) . The average channel power ρ _k is related to path loss and shadowing and it can be considered as slowly varying. The BS transmits M equal power independent data streams on M orthogonal beamforming vectors S = [s _j ...s _M ], with total transmit power M. The beamforming vectors are generated randomly, still preserving the orthogonality condition S ^H S = I . A new random set of beamforming vectors is generated regularly by the BS (say on every time slot). Given the beamforming vectors set S, the signal received on the N _k antennas of the k-th MS is:

M M y _k = ∑ H _k s _m d _n + w _k =∑ q _Km d _m + w _k , Eq . A m=l m=l where d _m ~ GV(O,1) are the incorrelated unit-power symbols transmitted on beam m taken from a Gaussian codebook, w _k ~ CW(O _{N 1 ?} I _{N N} ) (Eq. B) is the ambient noise received by the k-th MS, and q _{k m} = H _k s _m (Eq. C) is the equivalent channel perceived by the k-th MS when the BS transmits on beam m. When the BS schedules user k ^* on a single beam s _m , the received SINR γ _kt>m* is:

where b _k , _m , is the normalized receiver beamforming vector on beam s _m » , such that

M ₂

V |b _{k* m} = 1 . Symbol ^* might be omitted for brevity in the following. m=l

We assume that the BS is aware of the average channel power set [ρ _k } , but does not have any knowledge about the instantaneous user channel realizations (H ₄ ) . On the other side, we assume that each MS knows the number K of active users in the cell and has perfect knowledge of all the equivalent channel vectors (q _{k m} = H _k s _m |, obtained i.e. during DL preamble equalization. The proposed algorithm works as described in the SUMMARY OF THE INVENTION.

EP ALGORITHM The objective of the EP (Equal Probability) algorithm is to let each user evaluate the SINR threshold a _k in a distributed way in order to maximize sum-rate and satisfy fairness constraints. The event of a MS giving a positive feedback on a certain beam is independent from the event that another MS gives a feedback on the same beam because of channel incorrelation. We assume that the BS schedules for transmission on a certain beam one random user among those that give a positive feedback on that beam (RS), employing a conservative rate equal to 1Og(I + ^). The sum-rate is:

R _SUM

where P ^s (k,m) is the probability of scheduling user k on beam m, P _m (j \ k) is the probability that exactly j users exceed their threshold on beam m conditioned to the event that user k exceeds its threshold on beam m, and p[χ _k>m ≥ a _k ^m ) is the probability that the

SINR of user k on beam m exceeds the threshold a _k ^m . The considered system model implies that p(χ _{k m} ≥ a _k )= p(y _{k j} ≥ a _k ) Vm, j for a wide range of receivers (e.g. antenna selection, maximum ratio combining and non-adaptive beamforming receivers). We can therefore omit the beam index m from γ _{k m} and a _k ^m , and decouple the sum-rate formulation with regards to the beam indexes m:

R _SUM = ∑M [Y ₄ P(J I k)-]p{γ _k > a _k )log(l + a _k ) .

Eq .4

The Eq.4 is still difficult to be analytically treated, because P(j \ k) does not have a straightforward expression. We can simplify Eq.4 by decoupling the optimization by forcing an equal feedback probability p(γ _k > a _k ) = F to all the MSs. User channels independence implies that if a random scheduler (RS) is used each user is scheduled on a given beam with the same probability:

P ^s (k,m) = P ^s = — ^- , where: K

Eq.5 P _out = (\ -F) ^κ Eq.6 is the scheduling outage probability [9], i.e. the probability that no user exceeds its threshold on a given beam. The Eq.3 can now be written as:

R _SUM = ∑MP ^S log(l + α _k ) = ∑M^-^log(l + α _k ) . Eq.7

The sum rate maximization problem is thus reduced to finding the optimal value of P _out . We assume that every user knows its SINR CDF p(γ _k < β) and its complementary

CDF p(γ _k > β)≡ \ - p(γ _k < β). Given a target value of P _out , every user can evaluate the threshold α _k = avg[p(χ _k > α _k ) = Fj without any knowledge of other users' thresholds. We α _k >0 call this algorithm EP (Equal probability). With EP each user can locally evaluate its SINR threshold without CQI of the other users, given a target value of P _out . Subsection A - THRESHOLD CALCULATION FOR MIMO NETWORKS

In this subsection we show how to calculate the set of thresholds {α _k } when heterogeneous MSs are active in the same cell (e.g. both single antenna and multiple antennas MSs). For simplicity reasons we consider multiple antenna receivers where each receiving antenna is processed independently and there is no adaptive beamforming at the receiver side. The considered Rayleigh channel model implies that each receiving antenna is subject to independent fading. We define γ\ as the SINR experienced on the n-th receiving antenna of the k-th user on a given transmit beam, and γ _k = max[χ _k }. The n complementary CDF of γ _k " is [5]:

J_

Pk p{γ _k > β) _≡ \ - p{γ _k < β)= . Eq.8

The statistical independence of γ _k " with regards to n implies that:

p{γ _k < a _k ) = p{γ _k ⁿ < a _k ) ^Nk . Eq.9

The EP algorithm forces every user to give feedbacks with probability F . The equation that gives the threshold a _k , to be solved by every MS, is

P _out = F ^K = p(r _k ⁿ < a _k Y ^Nt → p{γ _k ⁿ < a _k )= {P _0Ut ≠r _t . Eq.10

Every user can compute its threshold by solving Eq.10 with regards to a _k as:

Eq.1 1 the first member of Eq.1 1 is monotone in a _k , so the single solution of Eq.1 1 can be easily computed. Eq.1 1 holds also for single-antenna MSs, since these users simply set N _k = 1.

EVALUATION OF THE OUTAGE PROBABILITY.

In the former section we showed that every user can locally evaluate its threshold a _k according to a certain scheduling outage probability P _out , so that all the users have the same feedback probability F regardless of their average channel power p _k or their number of receiving antennas N _k . The problem is now reduced to finding the scheduling outage probability P _out which maximizes the sum-rate. The main focus here is to have P _out as a function of K and M only, to allow a fully distributed algorithm where both P _out and a _k are evaluated locally by every MS. We will first find a simple condition on P _out that enables an optimal scaling law of the sum-rate when K — > ∞ (Subsection A), and then we will evaluate the scheduling outage probability P _0Ut (κ,M) that maximizes the sum-rate when K is large, but finite (Subsection B).

Subsection A. SCALING LAW OF THE SUM RATE.

Dirty Paper Coding (DPC) is known to achieve the optimal sum rate in a Gaussian broadcast channel [1]. For a high number of users with N receive antennas and random

scheduling, the sum- rate of DPC scales as M\og\og(NK) [3]. We suppose to have K users, M >\ beams and N>\ antennas per user.

Property 1: The average sum-rate R _SUM of the EP algorithm in conjunction with a random scheduler (RS) satisfies:

Eq.12 and is therefore asymptotically optimal.

Proof: The threshold of user k given by the EP algorithm is the solution of Eq.11, which cannot be expressed in closed form. We lower-bound p{γ _k >a _k ) with the function c _k {a _k ) (see appendix A for details):

-M +1

P(r _k ≥a _k )≥c _k (a _k )=e Pt Eq.13

It is easy to see that if F=I- P _0Ut ^~ κN = p(y _k >a _k )=c _k {a _k )→a _k <a _k . We can solve c(a _k )=F and write the lower-bound a _k in closed-form:

Eq.14

If we put Eq.14 into Eq.7 we obtain a lower-bound of the sum-rate:

R _suM =∑M- ^λ ^\og{\ + a _k )>R _S ^L _UM =∑M- ^l ^ ^L \og(l + ά _k (P _out )) i=l i=l

Eq.15

Let P. where φ > 0 is a non infinitesimal real number. It can be shown that (see

K ^φ APPENDIX B)

and

The asymptotic ratio between the sum-rate lower-bound and the scaling law of DPC sum- capacity with full channel information at the transmitter is:

lim *— - κ→ ^∞ M loglog(KN) Eq.18

It is therefore shown that letting P _out (κ) = , with ^ > 0 any real number, is a sufficient condition to enable the asymptotical scaling law of DPC.

B. OPTIMAL OUTAGE PROBABILITY.

When K is large but finite, EP does not reach the optimal sum-rate of DPC. Anyway, from a practical standpoint, we are still interested in finding the value of P _0Ut (κ,M) which maximizes the sum-rate lower-bound given by Eq.15. Without loss of generality, for analytical convenience, we can still let P _out = and rely on the approximation given by

K

Eq.16. Joining Eq.15, Eq.14 and Eq.16, and the following similar steps to Eq.18 we obtain:

Rsu _M = ∑M— ^λ ^log(l + ά _k (P _out )) k=\ Eq.19

log log K

Every user can numerically maximize Eq.19 and find an estimate of the optimal outage probability P _0Mr .

Fig.2 shows the average sum-rate in Eq.19 as a function of P _0Mr for different values of K when the EP algorithm is employed with a random scheduler (RS), M = N _k = 4 and {p _k } ~ £/[θ;2θ] (dB) . The optimal sum-rate is evaluated as

^ar] d ' ^s displayed with the "D" marker. Eq.19 is shown to give a good

estimate of the optimal P _out even when the number of users is moderate. When K > 50 the sum-rate penalty provided by the use of Eq.19 with regards to the optimal sum-rate is negligible. We also compare the optimal sum-rate with the sum-rate obtained using the closed-form asymptotically optimal value:

5 P _out = ^ Eq.20

The rate loss of Eq.20 with regards to the optimal sum-rate grows with K, but it is still lower than OJb / s / Hz _w h _en «=1000. When K=100 the rate loss of Eq.20 is equal to 0.36b/s/Hz ( _{4 5} o _{/o of the} optimal sum-rate).

FAIRNESS CONSTRAINTS

10 According to the EP algorithm, every user has the same probability of giving a feedback on a certain beam. If a random scheduler (RS) is used together with EP, the user selected for transmission is chosen randomly among those that reported a positive feedback, while the beam remains unscheduled if no user exceeds its SINR threshold on that beam. We

1 - P recall from section 3 that the probability of scheduling of user k is P ^s = — when a

K

15 random scheduler is used in conjunction with EP (cfr. Eq.5). Under the assumptions for the EP algorithm the average throughput experienced by every user is

which is an unfair split of rate that privileges users with higher thresholds \a _k ) (i.e. users with higher N _k or p _k ). We instead want to split the sum rate according to the proportional 20 rate constraints R _k , as stated in Eq.1. Joining Eq.21 and Eq.1 we obtain a set of K constraints on P ^s (k)

P ^s (k)= f ^k R _SUM \/k . log(l + «J

Eq .22

The EP criterion provides other K+1 constraints on \P ^S (k))'-

Eq.23

The scheduler objective is to find the set \P ^s (k)j that maximizes the sum-rate still satisfying the constraints of Eq.22 and Eq.23:

P ^s (k)} =

Eq.24 is a non-linear maximization with non-linear constraints that requires an iterative numerical solution.

SCHEDULING ALGORITHM

Herein a suboptimal scheduler which evaluates the scheduling probabilities [P ^s (k)j in closed-form is presented. The difficulty of the numerical solution of Eq.24 is due to the non linear constraint P ^s (k)≤ F . Anyway, if channel powers {p _k } and rate coefficients \R _k ) are not excessively unbalanced, constraint P ^s (k)< F is implicitly satisfied, as the optimal scheduler tends to allocate homogeneous values of P ^s (k) to all the users and F > (\ - P _out )/ K for the values of K of practical interest. Constraint P ^s (k)≤ F can thus be omitted from Eq.24 if a simplified scenario where {p _k } and \R _k \ are moderately unbalanced is assumed, and a closed-form linear solution of Eq.24 with regards to is obtained after some simple algebraic passages:

Eq.25

1 — P where P \k) is the optimal scheduling probability of user k, and R _SUM = — — is y ^κ 1 L έ? log(l + αj the optimal sum-rate. Let P _k be the actual scheduling probability of user k averaged over an equivalent time window of N _w time slots with an exponential weight function:

Eq.26 where T is the time-index of the present time slot and ^(t) is a variable equal to 1 if user k is scheduled in time slot t, and equal to 0 if user k is not scheduled in time slot t. The scheduling probability can be iteratively updated:

P _k ^s (t) = _k (t) .

Eq.27

The objective of the PRS is to schedule users in time slot T according to:

k _m ^* (T) = (τ)- P _k Eq.28 where k _m ^* (T) is the user scheduled on beam m at time slot T and S _m (τ) is the set of users that give feedback on beam m at time slot T.

The PRS algorithm works as follows:

1. Evaluate the set of target scheduling probabilities \P _k j according to Eq.25.

2. Schedule on each beam the user

3. Update the effective scheduling probabilities P ^aS (T) ' according to Eq.27.

4. Repeat from step 2. on next time slot. If a significant change in ^ ^k > occurred within last time slot, repeat from step 1.

Under an algorithmic point of view, PRS is similar to the PFS proposed in [4]. By the way PRS is optimized to work in coordination with EP and enables fairness in terms of rate splitting among the active users while PFS, as it is proposed in [4], enables only scheduling probability constraints.

NUMERICAL RESULTS

In this section we compare the performance of the proposed algorithms with other significant schemes taken from recent literature by numerical simulations. We consider a

DL cellular system with an M-antennas BS and K active users with N _k antennas each, where the number of receiving antennas is an i.i.d. integer variable N _k ~ t/[l;4] . Every

user is assigned an equal share of the sum rate R _k = — to be accomplished by the

K scheduling algorithm. In our first simulations all the users have equal power Rayleigh channels with gain p _k = \0dB .

Performances are evaluated in terms of average sum-rate R _SUM and average normalized rate deviation (NRD), which is measured as:

The performance of the proposed EP algorithm is evaluated in conjunction with a random scheduler (RS) and with the PRS described in Section 6, with a time window N _w equal to the number of simulated time slots. We also consider the following schemes from current literature: i. Single antenna opportunistic transmission, ii. Opportunistic beamforming with full CQI [4]. iii. Opportunistic-SDMA with full CQI [5].

Details about the implemented schemes can be found in Appendix C.

Fig.3 shows the sum-rate performance of the aforementioned algorithms versus the number of active users K. Even if the scaling law of EP is similar to that of [iii], the reduced feedback causes an almost constant performance gap of about3δ/s/Hz with regards to [iii]. As the number of users in the cell grows, EP outperforms [i]. and [ii]. In

Fig.4 we take into account a more realistic model, where the channel power p _k of every user is uniformly distributed (in dB) between OdB and 2OdB. Fig.4 shows Average Sum- Rate with heterogeneous channel gains. Fig.5 shows Normalized Rate Deviation (NRD) with heterogeneous channel gains. The Rate gap paid by EP with regards to [iii] is slightly greater than in Fig.3, even if the scaling laws of the sum-rate are still the same. Fig.5 shows the Normalized Rate Deviation (NRD) as it was defined in Eq.29 for all the considered schemes, in the same scenario used to generate Fig.4. EP with the PRS algorithm is the only scheme that guarantees rate fairness in this scenario among the ones considered in this description. The rate gap paid by EP with regards to [iii] is thus partly due to the fairness constraints accomplished by the EP algorithm. The analysis of Fig.3 and Fig.4 confirms the small sum-rate penalty given by the use of Eq.20 instead of the near-optimal P _out provided by Eq.19.

CONCLUSIONS A novel practical algorithm beyond the teaching of the DL SDMA scheduling algorithm considered in [7]-[8] for heterogeneous MIMO systems has been proposed in the present invention. A very low signalling overhead of Iog ₂ (i/ + l) bits per user per scheduling slot is needed to allow asymptotically optimal sum-rate, since the parameters required to run the algorithm are locally available to every MS. A simplified version of the algorithm (Eq.20) is shown to achieve the nearly-optimal sum-rate. Fairness requirements are accomplished by the use of the PRS in conjunction with the EP algorithm. Although the invention has been described with particular reference to a preferred embodiment, it will be evident to those skilled in the art, that the present invention is not limited thereto, but further variations and modifications may be applied without departing from the scope of the claims.

APPENDIX A

The optimal thresholds a _k given by the EP algorithm are the solutions of:

We need to find an invertible expression c _k (a _k ) which is a lower-bound of p{γ _k ≥ a _k ) . It can be shown that:

p(r _k ≥ a _k ) = + a _k j- ^M ≥ e™ = c _k (a _k ) , Eq.31

where q _k = lim ^ — . We can now write: α _t →0 da.

Hm + 1. Eq.32

We can finally write the lower-bound of Eq.30 as:

-M— — +1

Eq.33 and write the closed-form lower-bound approximation of a _k

Pk

APPENDIX B

We need to write an analytically convenient approximation of F, to be put into Eq.15, that has to be asymptotically tight for large K values:

F = p

Eq.35

Eq.35 is asymptotically tight, since i ^ O when K is high. After some algebra it can be shown that:

Eq.36

Therefore Eq.35 can be written as:

= 1 + — x ^N logx « 1 +— xlogx = l -— — ² —

N ^{6 χ} →o TV N K which is the closed form approximation we were looking for. Fig.6 shows comparison curves of Eq.35 and Eq.37.

APPENDIX C (BIBLIOGRAPHY)

[1] G. Caire, S. Shamai (Shitz). "On the Achievable Throughput of a Multiantenna Gaussian Broadcast Channel", IEEE Trans. Inform. Theory, vol. 49, pp. 1691 — 1706, July 2003 [2] M. Costa, "Writing on Dirty Paper", IEEE Trans. Inform. Theory, vol. IT-29, pp.

439-441 , May 1983

[3] M. Sharif and B. Hassibi, "Scaling laws of sum rate using time-sharing, DPC, and beamforming for MIMO broadcast channels", Information Theory, 2004. ISIT 2004. Proceedings International Symposium on, 27 June-2 July 2004 Page(s):175P

[4] Viswanath, D. N. Tse, and R. Laroia, "Opportunistic beamforming using dumb antennas," IEEE Trans. Inform., vol. 48, no. 6, pp. 1277-1294, June 2002

[5] M. Sharif and B. Hassibi, "On the capacity of MIMO broadcast channels with partial side information", Information Theory, IEEE Transactions on, Volume 51 , Issue 2, Feb. 2005 Page(s):506 - 522

[6] Q. H. Spencer, A. Lee Swindlehurst and M. Haardt, "Zero-forcing methods for downlink spatial multiplexing in multiuser MIMO channels", IEEE Trans. Signal Processing, vol. 52, no. 2 , pp. 461-471 , Feb. 2004

[7] J. Diaz, O. Somekh, O. Simeone, Y. Bar-Ness, "Scaling law of the sum-rate for multi-antenna broadcast channels with deterministic or selective binary feedback", in Proc. ITW'06

[8] J. Diaz, O. Simeone, Y. Bar-Ness, "Sum-rate of MIMO broadcast channels with one bit feedback", In Proc. ISIT'06

[9] D. Gesbert and M. Alouini, "How Much Feedback is Multi-User Diversity Really Worth?", in Proc. IEEE ICC, 2004.

APPENDIX D (ACRONYMS)

BS Base Station

CDF Cumulative Distribution Function CN(O, p _k ) Complex Normal (mean, standard deviation)

CQI Channel Quality Information

CSI Channel State Information

DL Downlink

DPC Dirty Paper Coding EP Equal Probability

NRD Normalized Rate Deviation

MCS Modulation and Coding Scheme

MIMO Multiple Input Multiple Output

MISO Multiple Input Single Output MS Mobile Station

NRD Normalized Rate Deviation

O-SDMA Opportunistic-SDMA

PFS Proportional Fair Scheduler

RS Random Scheduler SDMA Space Division Multiple Access

SNR Signal-to-Noise Ratio

SINR Signal to Noise plus Interference Ratio

TDD Time Division Duplexing

UL Uplink