Interference-Aware Scheduling in the Multiuser MIMO … ComMag.pdf · Interference is the limiting factor in cellular ... (UMTS), that is, the same ... Note that the CQI may include

INTRODUCTION

Interference is the limiting factor in cellularradio systems. The Global System for MobileCommunications (GSM) typically reduces inter-ference by avoiding the same frequency again inthe next cell. In order to increase spectral effi-ciency, the frequency reuse factor has beenreduced down to unity in the Universal MobileTelecommunications System (UMTS), that is,the same resources are assigned in adjacentcells. The severe intercell interference iswhitened by spreading the data over the entiresystem bandwidth and scrambling them with acell-specific sequence. In high-speed downlinkpacket access (HSDPA), terminals provide regu-lar feedback about their mean channel and inter-ference situation using a channel qualityindicator (CQI). Depending on this CQI, theentire system bandwidth can be assigned tem-porarily to the particular terminal that has thebest channel. If we consider the terminals in acell as a virtual antenna array, this scheduling

approach is similar to selection combining andexploits what is called multiuser diversity [1].Note that the CQI may include informationabout the interference. Assigning the channel tothe user with the best CQI can be regarded as atechnique to actively handle the interference.

The downlink of the Third Generation Part-nership Project (3GPP) Long-Term Evolution(LTE) is based on orthogonal frequency-divisionmultiple access (OFDMA). It has the potentialfor enhanced interference reduction comparedto previous systems based on code-division mul-tiple access (CDMA). At least for stationaryusers there is no more intracell interference,since orthogonal frequency-division multiplexing(OFDM) waveforms remain orthogonal afterpassing through a multipath channel. As a novel-ty, we can exploit the multipath nature of signaland interference channels in the scheduling pro-cess. Simply speaking, one assigns those parts ofthe spectrum to a user where simultaneously thedesired signal is strong and the interferenceweak. In addition to the time-domain schedulingalready used in HSDPA, groups of subcarrierscan be assigned to users according to the fre-quency-selective signal and interference condi-tions.

Multiple-input multiple-output (MIMO)techniques use multiple antennas at both thebase station (BS) and the terminal. MIMO isexpected to contribute substantially to theenhanced capacity of LTE. Note that OFDMsimplifies the signal processing for MIMO. Sim-ple MIMO algorithms for flat fading channelsare sufficient for channel equalization [2]. Tomaximize the benefits of the new air interface,our objective is to minimize the effects of inter-ference by means of joint radio resource man-agement for multiple users in a cell exploitingthe new degrees of freedom in the frequencyand space domains. Our approach is similar to aspectral decomposition of the colored interfer-ence and adapting the transmission accordingly.

Let us first define essential requirements forthe downlink MIMO medium access control(MAC) layer. Resource assignment needs fair-ness in a cellular network in order to guaranteethe best throughput for all users. With oppor-tunistic approaches as in [1] a user at the celledge is never served. A common implementation

IEEE Communications Magazine • June 200956 0163-6804/09/$25.00 © 2009 IEEE

ABSTRACT

With the introduction of orthogonal frequen-cy-division multiplexing and multiple antennas incellular networks, there are new opportunities toadapt the transmission to propagation and inter-ference conditions. In this article we describe apractical approach using space-frequency-selectivemultiuser MIMO scheduling. Frequency-selectivefeedback is provided on achievable data rates forpreferred single- and multistream transmissionmodes. The base station selects the best modewhile providing instantaneous fairness. Weobserve that multiuser transmission increases theprobability of using multistream transmission.Besides the benefits from optimal combining atthe physical layer, there is an additional gain atthe MAC layer since the estimation of achievablerates becomes more precise. Altogether, 93 per-cent of the theoretical throughput can be realizedby synchronizing the base stations and providingcell-specific reference signals. We have imple-mented essential functions of the approach in realtime on an experimental 3GPP LTE prototype in20 MHz bandwidth. Feasibility of the key featuresis proven in laboratory and field trials.

TOPICS IN RADIO COMMUNICATIONS

Volker Jungnickel, Malte Schellmann, Lars Thiele, and Thomas Wirth, Heinrich-Hertz-Institut

Thomas Haustein, Otto Koch, Wolfgang Zirwas, and Egon Schulz, Nokia Siemens Networks

Interference-Aware Scheduling in theMultiuser MIMO-OFDM Downlink

JUNGNICKEL LAYOUT 5/14/09 10:17 PM Page 56

Authorized licensed use limited to: FhI fur Nachrichten-technik. Downloaded on September 21, 2009 at 05:21 from IEEE Xplore. Restrictions apply.

IEEE Communications Magazine • June 2009 57

is proportional fairness: In the mean, the user isassigned a constant fraction of the rate he couldrealize if he was alone in the cell. To realize lowpacket delays, which are a general requirementof LTE, the possibilities of resource assignmentare rather limited in the time domain. For trafficwith high priority a free resource shall always beavailable for a user having poor channel condi-tions. For critical real-time multimedia servicessuch as videoconferencing, instantaneous fair-ness may be desirable. The score-based sched-uler provides good heuristics to realize theseobjectives [3].

Feedback reduction is a second requirementfor cellular mobile radio systems. A terminal haslow power (e.g., 200 mW) and is expected tobridge distances of several kilometers at the celledge. This is reached by reducing the bandwidthassigned in the uplink and increasing the spec-tral power density accordingly when movingfrom cell center to cell edge. Last but not least,the mobile radio channel changes rapidly, andfeedback is needed at high repetition rates. Eachfeedback bit costs battery power and spectralresources; thus, limited feedback is paramount.Here we consider feedback on the order of sev-eral tens of kilobits per second, which may befeasible even at the cell edge. With such a lowrate it is possible to feed back a coarse charac-terization of the MIMO channel as a function offrequency based on a CQI for several spatialtransmission modes and certain groups of sub-carriers.

Efficient spatial adaptation is a third require-ment. A MIMO system can in principle selectthe operation point in the diversity gain vs. mul-tiplexing gain plane [4] by using a particulartransmission mode according to the channel con-ditions. Such MIMO mode switching is helpfulto achieve the best possible transmission rate ina mobile scenario. Diversity transmission isfavored in low-rank channels having, say, a freeline of sight (LOS) to the BS, while multiplexingis preferred when the rank is full and the effec-tive signal-to-interference-and-noise ratio(SINR) is at a sufficient level. Sometimes higherthroughput may be realized if not all the streamsare used for spatial multiplexing. Selection canbe based on the achievable rates for various spa-tial transmission modes calculated at the receiv-er side. The preferred mode and correspondingrates are fed back to the BS where the radio linkis optimized [5]. Refer to [6] for an initial pro-posal of frequency-selective MIMO mode switch-ing. Single-cell performance with two-usersupport is investigated in [7], and fair schedulingfor multiple users in [8]. Multicell performanceis analyzed in [9]. Physical and MAC layer imple-mentation and early field trials are reported in[10].

Two more requirements not currently met byLTE Release 8 (R8) are needed. Knowledgeabout the interference could be obtained fromthe covariance matrix of the received signals.However, such estimation is not precise enough.The temporal variation of the covariance canhardly be tracked in a mobile scenario, and thepotential gains are ruined [11]. As a way out, wepropose to synchronize the BSs (e.g., by GPS orover the network using the IEEE 1588v2 stan-

dard [12]) and provide cell-specific reference sig-nals. At each terminal, the channels to thestrongest BSs are estimated [13]. The covarianceis then calculated from these multicell channelestimates.

The article is organized as follows. We startwith the spatial mode switching concept anddescribe our instantaneously fair schedulingalgorithm. We highlight two recent observations:First, the more statistically independent degreesof freedom we offer to assign spatially multi-plexed streams (e.g., frequency-selective feed-back, multiple users, multiple beams), the higherthe probability of using the multistream mode.Next we show that if the channel to the interfer-ing cells is known in addition to the serving one,the estimation error for the achievable rate onthe MAC layer is significantly lower. The systemcan be loaded less conservatively, and this trans-forms into an overall throughput gain. For thesereasons, the spectral efficiency is significantlyenhanced. The MIMO capacity scaling for point-to-point links, proportional to the minimum ofthe numbers of transmit and receive antennas,can also be approximated in this way in the mul-ticell scenario. Real-time implementation andlaboratory as well as field trials are described toshow the feasibility of our approach.

THE SPATIAL MODESWITCHING CONCEPT

The concept is based on a fixed grid of beamsprovided by the BS consisting of a number ofbeamforming vectors b given in a predefinedcodebook (Fig. 1, left). The physical layer sup-ports two principal transmission modes: single-stream (ss) mode for spatial diversity where asingle user is served exclusively on one beam,and multistream (ms) mode for spatial multi-plexing where independent data streams aretransmitted in parallel on multiple beams. Foreach supported mode, a terminal determines theachievable rates per beam and conveys thisinformation to the BS. The setting of the modescan be selected individually for each availablefrequency subband, also referred to as a resourceblock (RB).

In particular, for the ss mode, the effectivepost-equalization SINR is determined for eachbeam b after optimum combining at the termi-nal. From the SINR, the achievable rate perbeam is estimated. The highest beam rate togeth-er with the corresponding beam index is then fedback to the BS.

For the ms mode, we extend the commonoptimum combining (OC) approach to separatespatially multiplexed streams at the terminalside. Optimum combining provides the best filterweights for isolating a desired signal out of theco-channel interference from all other signalslike intra- and intercell interference. At the BS,we allow Q beams to be simultaneously active.Active beams are taken from the columns of theunitary matrix B, also called a beam set in thefollowing. The codebook may contain multiplesuch sets. The rate supported on each of the Qspatially multiplexed streams is estimated. Theper-stream rates for the particular matrix B from

For each supported

mode, a terminal

determines the

achievable rates per

beam and conveys

this information to

the BS. The setting

of the modes can be

selected individually

for each available

frequency sub-band,

also referred to as a

resource block.



IEEE Communications Magazine • June 200958

the codebook achieving the highest per-streamrate and the codebook index corresponding tothis matrix are fed back to the BS.

The BS may select one out of three optionsfor spatial transmission (Fig. 1, right). In ssmode the terminal providing the best rate isselected to achieve multiuser diversity. For thems mode, there are two options. Streams may beassigned to the same terminal as in classical sin-gle-user spatial multiplexing (SU-MIMO). Alter-natively, the available beams can be assigned todifferent users (multiuser spatial multiplexing[MU-MIMO]). Note that the streams are sepa-rated at the terminal side. The striking advan-tage of this proposal is that it enablesMU-MIMO access at a very low feedback rate.MU-MIMO can be supported without havingcoherent information on the downlink channelat the BS: only the achievable rates and the pre-ferred codebook indices must be reported for ssand ms mode.

INSTANTANEOUS FAIRNESSA transmission time interval (TTI) in LTE lasts1 ms and consists of two slots. At 18 MHz band-width 100 RBs are accommodated, each havinga bandwidth of 180 kHz. The RBs can be indi-vidually assigned to users. The objective ofresource allocation is to assign each user its bestresources in a frequency-selective manner,whereby the decision on the spatial mode shouldbe made under the premise of achieving highthroughput while targeting proportional fairnessamong users. In this section we describe our fre-quency-selective scheduling algorithm. It isbased on a two-step approach.

In step 1 the terminal evaluates for each RBthe achievable rates for all spatial transmissionmodes and selects the best codebook entriesaccordingly. The achievable rates for each modeare quantized and fed back as a frequency-selec-tive CQI together with the preferred spatialmode index (PMI). This PMI contains the indexof the best beam in ss mode and the index of thebest beam set, and which subset of the beams isused in ms mode.

CQI as well as PMI are RB-specific informa-tion. At the BS, we collect the CQI and PMI vs.frequency vectors from all terminals. For eachterminal, the CQIs for all RBs are put in sepa-rate lists for each transmission mode. These listsform the basis for extending the original score-based scheduling approach [3], which assignseach transmission resource a score representinga quality rank. The key to enable a direct com-parison of ss and ms rates is the introduction ofa so-called benefit factor for the rates in themultiplexing mode. As we aim for high spectralefficiency, mode selection should favor the msmode if the user rate can be expected to be larg-er than the rate expected in ss mode.

If a user decides globally on ms mode, thetotal available spatial streams compared to ssmode are multiplied by the factor Q. According-ly, the terminal will be assigned Q times theresources it would get if it globally selected thess mode. The ss mode thus should be favoredonly if

where Rss and Rms,i are the single- and multi-stream rates reported on stream i = 1 … Q. Toinclude this as a benefit, the rate list for the msmodes is weighted by the factor Q and concate-nated with the rate list containing the ss rates.Using the concatenated list, joint sorting is per-formed. By this procedure we yield a ranking notonly of the different RBs, but implicitly also ofthe different supported transmission modes. Thisranking (i.e., the order of entries in the sortedlist) is represented by the score.

Step 2 addresses resource scheduling, whichis performed separately for each RB. All userscores for that RB are first partitioned accord-ing to the transmission mode they refer, ss orms. User selection is at first carried out for eachtransmission mode separately, and then a finaldecision on the mode is made. In ss mode theuser is selected providing the minimum scorefor that mode. In ms mode users choosing thesame beam set B are candidates for MU-MIMOand are thus put into one group. In this group

R Q Rss i ms i> ⋅ max ,,

�� Figure 1. Left: transmission concept based on a fixed grid of beams. The terminal calculates the achievable rates in single- and multi-stream modes for preferred beams and feeds those back to the base station. Right: mode switching options; option 1 — classical mul-tiuser diversity; option 2 — SU-MIMO, both streams are assigned to one terminal; option 3 — MU-MIMO, the streams are assigned todifferent terminals.

l

Interference from other cells Multiuser diversity Multiuser multiplexing (MU-MIMO)

BS

b s

w y

Terminal 1Basestation

s

sH

j

Feedback: best-beamindices j, l and

corresponding SINRsfor single- and

multistream modes

MT mxi

ymi,j

B s

Basestation

s1s2

w y

Terminal 2

sH

W H1 y

Terminal 1s1s2

s1s2

W H2 y

Terminal 2

Streams are frequentlyassigned to different terminals




each of the Q available beams is assigned to theuser providing the minimum score for thatbeam. Obviously, this user selection alsoincludes classical SU-MIMO, as all streams maybe assigned to one user. After user selection hasbeen carried out for each of the available beamsets, we pick the set containing the user with theminimum score. Finally, we compare the mini-mum of the scores provided by the selected msuser(s) with the score of the user favored in ssmode and select the mode yielding the totalminimum.

SINGLE-CELL RESULTSAs already mentioned, an in-depth performanceanalysis of the spatial mode switching concept isreported in [7, 8]. Here we highlight a morerecent observation revealing why the spectralefficiency of this scheme is substantially higherthan SU-MIMO transmission as suggested inLTE R8. In our single-cell analysis we havegrouped 10 users on a ring around the BS; thatis, the average signal-to-noise ratio (SNR) isidentical for all users. Normalized channels areobtained from the 3GPP SCME channel modelin the urban macro scenario. Terminals and theBS are each equipped with two antennas. Themodel assumes a uniform linear array of co-polarized antennas where antenna spacing at theBS is four wavelengths, yielding a minor correla-tion between antenna signals. User channels aremodeled independently. A bandwidth of 18 MHzis used, accommodating 100 RBs of 12 subcarri-ers width each. We have used the unitary beamsets C1 and C2 defined in [14]. Potential sstransmission selects a single beam out of theavailable beam sets and allocates the entirepower to that beam. In ms mode the transmitpower is distributed equally over the Q activebeams. Potential ms modes are either MU-MIMO, where one stream is assigned to a firstuser and the second stream to a second user, orSU-MIMO, where both streams are assigned tothe same user.

In Fig. 2 the probability of ss and ms trans-mission is compared as a function of the SNRfor different scheduler constraints. Note thatbetter spectrum utilization and higher through-put is typically achieved if the probability of msmode is increased. In the first two curves (SUfixed and SU adaptive), an RB is always exclu-sively assigned to one user, which is still allowedto choose diversity or multiplexing as transmis-sion mode (SU-MIMO). For the fixed configura-tion, transmission mode as well as the beam setis fixed per user. Selection of the fixed mode andbeam set is based on the highest sum rate overthe entire frequency band. Hence, a user reportsRB-specific CQI and global PMI. For the adap-tive configuration, the user is allowed to choosethe transmission mode and beam individuallyper RB. By comparing the curves of SU fixedand SU adaptive in Fig. 2, it is observed that thecrossing point between the ss and ms probabilitymoves by 3 dB to lower SNR by allowing fre-quency-selective mode selection.

A striking shift of the crossing point isobserved if MU-MIMO is enabled. For the fixedconfiguration, the crossing point is remarkably

shifted from 13 dB down to 2 dB with 10 users.For the adaptive configuration, the crossingpoints of the MU curves shifts below 0 dB. Notethat the curves considered so far have used onlybeam set C1. If two beam sets are enabled (left-most curves in Fig. 2), the crossing point shiftsto –2 dB. Throughput gains resulting from fre-quency-selective MU-MIMO with 10 users in 20MHz in a single cell are between 24 and 30 per-cent at low and high SNR, respectively (notshown).

MULTICELL RESULTS AND THEIMPACT OF CHANNEL ESTIMATION

By introducing this scheduling approach, thelikelihood of the ms mode is remarkablyenhanced in the particular range of interest forcellular systems with full frequency reuse. Onemay expect, therefore, that users are preferablyserved in multiplexing rather than in diversitymode even at the cell edge. This would implythat cell edge users could be served in a spectral-ly more efficient way.

Let us first mention a difficulty introduced bythe above mentioned option to select variablebeam sets, which has a positive effect in a singlecell. In a multicell environment fast beam setselection would destroy the causality in thescheduling process. If the set chosen in an adja-cent cell changes rapidly, the interference is nolonger predictable in the cell of interest, andrescheduling should take place. As a conse-quence, we abstain from using this option on ashort timescale in the multicell case in favor ofmaking the interference more predictable. Beamset selection may be used on a longer timescale(e.g., for optimizing the cell geometry). Further-more, for calculating the SINR we have assumed

�� Figure 2. Probability of mode selection for ms mode (red) and ss mode(blue) vs. SNR. In the legend SU means that RBs are always exclusivelyassigned to a single user, while MU means that multiple users can share anRB. The word fixed means that the spatial mode of a user is set constant forthe whole frequency band, while adaptive means that the spatial mode of auser may change per RB.

MU-MIMO, 2 beam setsMU-MIMO, adaptiveMU-MIMO, fixedSU-MIMO, adaptiveSU-MIMO, fixed

Ps /N0 (dB)0-5

0

1

0.2Prob

abili

ty o

f m

ode

sele

ctio

n

0.4

0.6

0.8

5 10 15




that surrounding cells use the ms mode perma-nently on all RBs. Taking mode switching inadjacent cells into account would not substan-tially change our results since ms transmission isfavored, as shown above.

In this section we describe a second contri-bution to the enhanced throughput being obvi-ous only in the multicell scenario. Note thatbesides the channel of the desired signal, theinterference from other cells is also time andfrequency selective. LTE R8 has adopted thetraditional interference whitening approach.Based on cell-specific scrambling sequencesapplied to pilots along the frequency axis, wecan estimate the wideband averaged SINR byfrequency domain correlation with thesequences of the strongest cells. However, thisaveraging in the frequency domain has a nega-tive impact on the achievable throughput. Boththe desired signal and interference plus noiseterm remain random numbers on a given RB,and their ratio determines the achievable rateon that RB, which is not at all a constant num-ber. Since the interference is not preciselyknown for each RB, however, we have to applyempiric safety factors reducing the mean SINRand taking into account that there are in factvariations in the frequency domain. Accordingly,a conservative rate is assigned to users indepen-dent of frequency.

Frequency-selective SINR estimation is apromising way out. BS synchronization [12] andthe introduction of pilots allowing an estimationof the frequency-selective channel for adjacentBSs [13] enable both better suppression of theinterference at the physical layer and more pre-

cise SINR estimation as well. In the following wealso include the effects of channel estimationerrors due to multicell interference.

As a reference, maximum ratio combining(MRC) is used, where the interference is esti-mated in different ways. First, the interference isassumed white over both antennas and frequen-cy. This case is similar to LTE R8. Second, weobtain the colored spatial interference by esti-mating the covariance matrix of the received sig-nal vectors. These are the techniques that can berealized using asynchronous transmission.

In contrast, we consider a synchronized net-work and a more sophisticated linear receiversuch as OC having coherent knowledge aboutthe co-channel interference. OC is also referredto as the minimum mean square error (MMSE)receiver. The interference covariance is calculat-ed here from the multicell channel estimatesrather than measured from the data signals as inthe asynchronous case. In [13] we have proposedvirtual pilot sequences identifying the cells bywhich the conventional pilots are scrambled inthe time domain. The estimator uses a slidingcorrelation window over several slots. Increasingthe correlation window yields more precise iden-tification of interferers in general but limits themobility. A correlation length of one slot refersto a window size of N = 3 chips in the sequence(i.e., the sector groups can already be distin-guished). By extending the sequence over fourslots (N = 12) we can distinguish groups of fouradjacent sites. The following results include theresidual estimation errors of the multicell chan-nel if the cells use the virtual pilot sequenceassignment in [13].

�� Figure 3. Left: estimation error of the estimated SINRest in ss mode compared to the available SINRavail. Right: throughput of varioustransmission schemes including the estimation errors. μ is the mean square error of the channel.

SINRest /SINRavail (dB)-10-15

0.05

Prob

abili

ty (

rati

o ≤

absc

issa

)

0.25

0.5

0.75

0.951

-5 0 5 10 15Spectral efficiency (b/s/Hz)

10

P(sp

ectr

al e

ffic

ienc

y ≤

absc

issa

)

0.050

0.25

0.5

0.75

0.951

2 3 4 5 6 7

1.68

MRC, freq-flat i.i.d. σ2IF’μ = 0.1

MRC, freq.-sel cov.μ = 0.1MMSE, freq.-sel.corr. N = 3MMSE, freq.-sel.corr. N = 12

SISO, µ = 0.1SU-MIMO, µ = 0.1, i.i.d.σIF’Single-stream, µ = 0.1, i.i.d. σ2

IFSU-MIMO, MMSE corr. N = 12Single-stream, MMSE corr. N = 12Adaptive, MMSE corr. N = 12Adaptive, perfect CSIR including IFAdaptive, perfect CSIR including IF,CQI gran 5, 2 bit feedback




Figure 3 (left) illustrates a first result usingthe estimation error in ss mode. We comparethe ratio of the estimated SINR to the theoreti-cally achievable value by employing either MRCin an asynchronous network or MMSE in a syn-chronized one. In the asynchronous network, weassume that the SINR is estimated in widebandmode, as in 3GPP LTE R8. In the synchronousnetwork the SINR is obtained in a frequency-selective manner after channel estimation basedon virtual pilots as described in [13].

For MRC with frequency-flat knowledge ofthe interference power σIF

2 , the estimation suf-fers in two ways. First, there is a median shift of–1 dB in Fig. 3 (left), meaning that the estimat-ed SINR is systematically too low and more datacould be transmitted than expected. Second, theestimation error has a considerable variance.Only for the RBs where the estimation error is 0dB are the user rates assigned precisely. Forother RBs, the achievable rates are randomlyover- and underestimated, and we have to applya safety factor S < 1 to the estimated SINR withthe intention that the assigned rate is feasible(e.g., in 90 percent of cases). The complemen-tary 10 percent of the bits in a transport block(TB) are mapped onto bad RBs where the chan-nel is overloaded and probably cannot be recov-ered correctly. Powerful turbo codes withinterleaving across the resources assigned to auser and hybrid automatic repeat request (H-ARQ) can repair such errors. Nevertheless, thesafety factor S remains, and it implies a penaltyfor the overall system throughput. For MRC, wecan estimate S = –2.8 dB from Fig. 3, and withthe shifted median of –1 dB there is an overallpenalty of roughly 3.8 dB at the MAC layercompared to the theoretically achievable SINR.

Direct estimation of the covariance leads toan unbiased SINR estimation (see the dashedblue line in Fig. 3, left). But the S-factor is highdue to the huge variance of the estimation error.It is caused by the interference-limited channelestimation in the serving cell and the short aver-aging interval of two slots assumed for thecovariance. The overall penalty is 6.3 dB. Due tothis penalty, it makes little sense to combine fre-quency-selective covariance estimation witheither MRC or MMSE. Consequently, MRCwith white interference assumption1 fits wellwith asynchronous transmission.

But we can do better if the BSs are synchro-nized and the interference is estimated. This hasbeen done for MMSE. With a correlation win-dow of 1 slot (N = 3), interference from adja-cent sites cannot be separated. The SINR issystematically overestimated (Fig. 3, left).Already with a correlation window of four slots(N = 12) more interferers can be identified, andthe SINR is computed more precisely. Then weget S = –0.9 dB and the bias becomes negligible.

Based on multicell channel estimates, theperformance can be improved additionally at thephysical layer using the MMSE receiver. In Fig.3 (right) we have plotted the achievable rates inthe multicell system including channel estima-tion errors.2 As a lower bound, we have giventhe performance in the SISO case where theminor effect of estimation errors is included aswell. If the spatial interference is assumed white,

the performance of single-user ms transmissiongets worse than ss transmission with MRC. Thereason is that the estimation error leads to inter-stream interference in the ms case, which is notpresent with ss transmission. Although theMMSE receiver can exploit the interferenceknowledge, SU-MIMO transmission still sufferssubstantially from estimation errors compared tothe ss case. Only if the fully adaptive MAC isused, where the streams can be assigned to dif-ferent users, is there a significant throughputgain. By including the MU-MIMO gain, the msmode becomes really efficient in the multicellscenario. The gap to the adaptive system withperfect channel and interference knowledge isonly 7 percent, indicating that the proposedscheme is very robust against channel estimationerrors.

An essential condition for realizing the poten-tial gains is also illustrated in Fig. 3 (right). Tar-geting simple feedback compression, one mightwant to combine CQI values (e.g., for five con-secutive RBs) and reduce the feedback quantiza-tion to 2 b/RB group. However, such a simplecompression would ruin the potential gains. TheCQI information must be available at the BSRB-wise and with 5-bit granularity [16], whichneeds smarter compression. A promising way isto exploit the channel correlations in time andfrequency domains (i.e., to apply efficient sourcecoding techniques).

Let us finally compare asynchronous and syn-chronous transmission. In the asynchronous caseone would prefer ss transmission because of thedetrimental effects of estimation errors in the mscase. In the synchronized case with multicellchannel estimation, the average throughput gainis about 68 percent of that in asynchronous SU-MIMO transmission according to the results inFig. 3 (right). Despite the estimation errors, wecan realize 93 percent of the theoretically pre-dicted performance. As shown in [9] where chan-nel estimation and feedback are assumed ideal,the throughput also scales linearly with the num-ber of antennas in the interference-limited sce-nario.

REAL-TIME IMPLEMENTATIONEssential features of the concept have beenimplemented in a real-time prototype with theaim to prove the feasibility of frequency-selectiveMU-MIMO scheduling.3 The system is operatedin frequency-division duplex mode in the 2.6GHz band identified for future LTE deploymentin Europe. Downlink and uplink are operated at2.68 GHz and 2.53 GHz, respectively. The band-width can be scaled from 1.5 to 20 MHz. BS andtest terminals called user equipment (UE) areeach equipped with two antennas. Total trans-mitter powers are +43 dBm at the BS and +23dBm at the UE. The OFDM system uses 2.048subcarriers with 4.7 μs cyclic prefix. The 10 msradio frame contains 20 slots of 0.5 ms durationeach.

In each downlink slot seven OFDM symbolsare transmitted. Representing the ss mode,antenna selection has been implemented togeth-er with MRC at the receiver where the activetransmit antenna can be selected for each RB.

1 That is, averaged overboth frequency and anten-nas.

2 Overhead has beenignored here since itdepends on the particularsystem design.

3 Development has beenstarted in the early phaseof the LTE study item. Afew technical details havebecome obsolete, and pro-prietary simplifications ofthe standard are includedin this early implementa-tion. There is minorimpact on the perfor-mance.

The CQI information

must be available at

the BS RB-wise and

with 5 bits

granularity [16],

which needs smarter

compression.

A promising way is

to exploit the

channel correlations

in time and

frequency domains,

i.e., to apply efficient

source coding

techniques.




For the ms mode two data streams are transmit-ted in parallel using polarization multiplexing,and a linear MIMO MMSE filter is implement-ed at the terminal side. The transmitter andreceiver can be switched between ss and msmode in each RB and each scheduling intervalby an adaptive data mapping unit at the BS [10],while a vector and a matrix are applied as filterweights at the receiver, respectively, according tocontrol information provided with the data.

In the uplink, contiguous subcarrier blocksare assigned to a user. Data are passed througha discrete Fourier transform (DFT) prior tomapping them onto the frequency domain. Ourtest terminal radio frontends have two transmitantenna ports. Cyclic delay diversity (CDD) isused in combination with MRC at the receiverto overcome power limitations. The terminal isremotely synchronized to the BS (i.e., the fre-quency offset is precompensated in the uplink).Timing advance is measured using a terminal-specific sequence and steered dynamically overthe downlink control channel. For more detailsof the physical layer implementation refer to[15].

At the MAC layer, and in the 2.5, 5, 10, and20 MHz modes the frequency-time grid is subdi-vided into 6, 12, 24, and 48 RBs, respectively. ARB consists of 25 subcarriers and 7 OFDM sym-bols each. 144 complex data symbols are mappedinto each RB; other resources are used as pilots.In 20 MHz mode, up to 48 RBs can be assignedto an individual user on each transmit antenna.All RBs assigned to a particular user in a slotform a variable-length TB. Going beyond LTER8 we have enabled a finer granularity of adap-tation to the frequency-selective channel by alsoallowing adaptive modulation and adaptiveMIMO mode selection within a TB: each RBmay be loaded with different modulation for-mats from quaternary phase shift keying(QPSK), 16-quadrature amplitute modulation(QAM), and 64-QAM, and operated in eitherdiversity or multiplexing mode.

Coding is performed over all RBs assigned toa user in a TB. Supported code rates are 1/2 and3/4; the minimum word length is 432 bits, corre-sponding to the smallest TB size. Convolutional

coding was chosen to reduce the hardware effort.It is critical for the performance to interleave allbits in one TB and realize frequency diversity inthis way even if adaptive modulation is usedwithin the TB.4 At higher mobility the delays inthe feedback and control loop become long com-pared to the coherence time, and resourceassignment might be outdated. Note that theinterleaved adaptive modulation enables bothsignificant throughput gains at low mobility anddiversity gains at high mobility. For furthermobility support, the SINR threshold of QPSK(denoted on-level) can be modified dependingon the bit or packet error rates. The thresholdsfor other modulation formats are coupled to thislevel. If the error rate increases (e.g., because ofhigher velocity), the on-level is increased dynam-ically. The overall rate is then reduced, and thelink is stabilized at a lower data rate.

We have implemented proprietary feedbackand control channels. For both ss and ms modesthe achievable rates on both antenna ports arequantized by 8 b/3 RBs. In 20 MHz bandwidththis gives 128 bits of frequency-selective feed-back information transmitted each 10 ms (i.e.,12.8 kb/s). The information is mapped on a shortcontrol TB in a particular slot and transmittedusing binary PSK (BPSK) modulation with rate1/2 over the shared uplink channel. The down-link control channel is transmitted in a dedicatedslot containing the complete downlink and uplinkresource map for an entire radio frame. A pro-prietary compression format called Tetris isused. In principle, the corners of the resourceareas assigned to a user in the time-frequencymap are transmitted, not the detailed informa-tion for each RB. Control information is trans-mitted with QPSK rate 1/2 in single-streammode using CDD at the transmitter and MRC atthe receiver.

The hardware is shown in Fig. 4. BS equip-ment for one sector is placed in an outdoorhousing (Fig. 4, left). Radio front-ends includetwo transmit-receive chains and the dual amplifi-er and combiner unit containing the duplex fil-ters. The RF unit is coupled to the LTE signalprocessing unit (LSU) via common public radiointerface (CPRI) operating at 1.2 Gb/s. The

4 For final resource map-ping, rate matching ofallocated resources andmodulation and codingschemes is implemented.

�� Figure 4. Testbed hardware. Left: base station hardware for one sector. Center: antenna setup on top of the TUB main building. Right: test terminals in the measurement van.




LSU consists of two motherboards with two Vir-tex2Pro-100 FPGAs each, four and one TI 6713DSP for the equalization and scheduling algo-rithms in the receiver and transmitter branches,respectively, and two Virtex4 FPGAs realizingthe interface to 1 Gb/s Ethernet as well as CPRI.Sector antennas with +18 dBi gain, 60° horizon-tal, and 5° vertical beam-widths and electricaldown-tilt are used.

The test mobile is shown in Fig. 4 (right). Itis fixed in a standard 19 in box with three heightunits. The same reconfigurable signal processingplatform is used as for the BS. Differences are inthe smaller RF front-ends and the Li-ion batterypack for 3 h mobile operation. Note that sophis-ticated coarse and fine synchronization tech-niques are used so that the test mobile can beunplugged and used for real-time communica-tions in the field.

LABORATORY AND FIELD TRIALSWe have tested the implementation in the labo-ratory using a wideband channel emulator. Thesetup is sketched on top of Fig. 5. Two transmitantenna signals are fed into the emulator, wherethe Pedestrian B channel model is used havingmaximal delays as large as 3.7 μs (i.e., 3/4 of thecyclic prefix). Fine timing is critical in this chan-nel. Physical noise is added after the emulator,and the signal is fed into the terminal. The SNRis set using a variable attenuator and checkedcarefully using a spectrum analyzer. Uplink anddownlink paths are separated using circulators.At the terminal side, the physical layer goodputis measured. It is given as the scheduled ratetimes the rate of correctly received packets.Results are plotted in Fig. 5, bottom. With fixedmodulation, we measure the typical blurred step-like throughput curves for QPSK, 16-QAM, and64-QAM, all with rate 1/2. The curves may besteeper when turbo coding is used instead of ourconvolutional code, and the onset appears athigher SNR. Nonetheless, the performance canbe regarded as typical for MIMO in LTE overthis channel, and it has been checked thoroughlyin the LTE/SAE Trial Initiative (LSTI).

Adaptive transmission has been measured aswell. The on-level has been set so that the pack-et loss rate is below 10 percent. With such agreedy threshold, the measured packet lossvaries between 0.2 percent at highest and 9.8percent at lowest SNR, respectively. Thethroughput curve touches the fixed mode curvesapproximately at the points where the targetedpacket error rate in adaptive mode is also real-ized with fixed modulation. The main advantageof adaptive modulation is smoother adaptationto the channel conditions; thus, higher through-put can be achieved in practice. The measuredmultiuser gain with two users over independentPedestrian B channels is 40 percent at 5 dB SNRand reduces to roughly 10 percent for SNR > 15dB. The round-trip delay has been measured asmean (maximum) ping time of 6.5 (7) ms with-out traffic, 7.3 (11) ms with 64-QAM and 25Mb/s UDP downlink traffic, and 7.5 (22) ms withQPSK and 7 Mb/s load.

In order to investigate the performance ofthe adaptive multiuser MIMO MAC under real

propagation conditions, we conducted outdoorfield trails in the city of Berlin. A single-cell sce-nario is considered. The BS is placed on top ofthe main building of the Technische UniversitätBerlin (TUB, Fig. 4, center) at a height of 45 mabove the ground and about 10 m above theaverage rooftop level to realize a typical urbanmacrocell scenario. The main azimuth lobe ofthe sector is directed toward 30° measured fromnorth over east, and the down-tilt angle is set to2°. The BS is part of a test network deployedrecently in cooperation with Deutsche Telekom(Fig. 6, bottom right). The setup emulates anelementary interference scenario with four sites,where the site at HHI serves as the sector ofinterest, and the six surrounding sectors are real-ized as shown in the inset. Sites are intercon-nected by 1 Gb/s free-space optical links; inaddition, the three sites at HHI, TUB, and T-Labs are linked by optical fibers.

The terminal is installed in a measurementvan. Omnidirectional antennas are arranged in across-polarized setup on the rooftop. The drivingroute in the city center of Berlin covers low andhigh path loss with a large dynamic range from–39 to –92 dBm of received power, respectively.The route goes through areas with dense build-ings and the large park area of Tiergarten withdense vegetation. In Fig. 6 (top), the achievablethroughput for a single user is shown on themap. Data are obtained by recording the pilotsat the terminal side. SINR calculation andscheduling has been done offline.4 At the smalldown-tilt angle, distances covered at 2.6 GHz

�� Figure 5. Top: setup for laboratory measurements. Bottom: measuredthroughput using fixed and adaptive modulation.

SNR (dB)105

0

20

Phys

ical

laye

r th

roug

hput

(M

b/s)

40

60

80

100

120

140

160

15 20 25 30 35

Fixed QPSKFixed 16-QAMFixed 64-QAMAdaptive modulation

TerminalBase station

AWGNAWGN

C8

Pedestrian B MIMOchannel model

Div




range from less than 1 km in typical urban areaswith dense buildings up to 5 km when the LOSis free.

In order to illustrate the potential of the pro-posed multiuser MIMO approach, we considerthe selection probabilities for the ss and msmodes in two different distance regions, bothvalid for realistic cellular deployments. The ter-minal positions have been sorted according totheir distance, between 300 and 1200 m andmore than 1200 m away from the BS, respective-ly. For comparison we include results obtainedfrom system-level simulations, where multicellinterference has been taken into account.

In Fig. 6 (bottom left) the distributions of themeasured SNR for the two distance groups arecompared to the statistics of the SINR in themulticell scenario for individual RBs. Our mea-surements realize situations typical for users ingood and average channel conditions in a multi-cell scenario. We have placed in each scenario

either a single user or 10 users at a constant dis-tance from each other travelling jointly along thetrack. The single-user rate as a function of posi-tion is shown in Fig. 6 (top). Note that there areonly a few blank pixels along the track where thelink is lost. All other positions are included inthe statistics.

The mode selection probabilities are given inthe table embedded in Fig. 6. With only oneuser at large distances of d > 1.200 m, 18.7 per-cent of RBs are assigned in ss mode, while for80 percent ms mode is used. Compared to therelations illustrated in Fig. 2, this is surprising atfirst glance, since in 60 percent of the locationsthe SNR is lower than 15 dB here. However, thehigh ms probability even at lower SNR may beattributed to the cross-polarized antennas usedin our field trials. In fact, they help the terminalsrealize a higher rank of the MIMO channelcompared to the co-polarized antennas used inour simulations.

�� Figure 6. Top: achievable data rate with one user in a single sector with 20 MHz bandwidth. Bottom left: cumulative distributions ofthe measured SNR in two distance ranges and the SINR in a simulated multicell scenario. Bottom right: multisite test network in Berlin.

BS

TU Berlin Main Building

300 m 600 m 1200 m

Spree

Bellevue

Siegessäule

Probabilities of ss and ms mode selection in percentage1 user

Scenario ss msSU-

MIMO

No alloc.

ss msSU-

MIMO

msMU-

MIMO

Noalloc.

300m<d<1200m 5.0 95.0 0.0 0.0 12.3 87.8 0.0 d>1200m 18.7 80.0 1.3 0.0 11.1 88.9 0.0 Multi-cell 83.8 14 2.2 11.0 5.9 83.1 0.0

10 users

Bahnhof Zoo

Tiergarten

Straße des 17. Juni Heinrich-Hertz Institute

N

Ernst-Reuter- Platz

Capacity [Mb/s] [b/s/Hz] 173.0 9.61

156.9 8.72

140.8 7.82

124.7 6.93

108.6 6.03

92.5 5.14

76.4 4.24

60.3 3.35

44.2 2.46

28.1 1.56

12.0 0.67

Legend Base station Trees Buildings

500m

SINR [dB] -10 -20

0

0.1

Prob

abili

ty (

SIN

R<ab

scis

sa)

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30

300-1200m 1200-2860m SCME, multi-cell

4

Optical Throughput

(Mbps)

135 105 75 45 15

150 120 90 60 30 0

TEL

HHI ZRZ

S.HOF

H

EB

Airlaser

Airlaser

fiber network

1

2 3

Serving cell

Interfering cell

Interfering cell

4 Such offline processingis more realistic to illus-trate the potential of thetechnique. The measuredreal-time throughputs aresmaller due to the simpli-fied feedback compres-sion, as indicatedsimilarly in Fig. 3, right.




For 10 users at d > 1200 m, the probabilityof ss transmission is reduced to zero. SU-MIMOmode is only selected in 11 percent of cases,while MU-MIMO mode is used for 89 percentof the RBs. This remarkable preference forresource sharing can be observed in the multi-cell scenario as well. With a single user, the ssmode is clearly favored by about 84 percent ofthe RBs, due to the low SINR in the interfer-ence-limited scenario. Only 14 percent of theRBs are assigned in ms mode, as only those ter-minals with excellent channel conditions canbenefit from spatial multiplexing. The situationis now completely reversed if MU-MIMO isenabled. With 10 users, the ss probability col-lapses down to 11 percent while as much as 89percent of the RBs are assigned in ms mode:spatial multiplexing becomes suddenly dominantif resource sharing is enabled in the multicellscenario.

CONCLUSIONSWe have outlined the potential of using fre-quency-selective multiuser MIMO scheduling ina further evolved cellular network in which BSsare synchronized and terminals are enabled toestimate not only their own channel but alsothe channels of interference signals from othercells. We observe that the more statisticallyindependent degrees of freedom the BS gets toschedule multiple users, the higher is the prob-ability of multistream transmission in general.Even a significant fraction of users close to thecell edge can be served using spectrally efficientmulti-stream transmission. We have investigat-ed the performance in an interference-limitedenvironment, and shown that frequency-selec-tive interference knowledge improves the per-formance by means of optimum combining atthe terminal side and yields a more precise esti-mation of the achievable rates as well. By shar-ing the same resource among multiple users ina cell, the overall throughput can be enhancedby 68 percent in a cellular 2 × 2 MIMO linkcompared to the traditional interference-whitening approach. Our real-time implementa-tion shows that the approach is easilyintroduced in the LTE signal processing chain.We have also tested the scheme in field trials ina typical urban macrocell deployment. Ourresults confirm that there is a significantly high-er benefit of multiple antennas in a cellularnetwork if the mobile terminals are aware ofthe frequency-selective interference and sharingof resources among multiple users in a cell isenabled.

ACKNOWLEDGMENTSThe authors wish to thank A. Forck, H. Gae-bler, S. Jaeckel, L. Jiang, S. Schiffermüller, S.Schubert (HHI), E. Costa, J. Eichinger, R. Half-mann (NSN), C. Juchems, F. Luhn, R. Zavrtak(IAF GmbH), K. Kojucharow (KMDC), H.Droste, W. Kreher, J. Mueller, G. Kadel (T-Labs), and W. Stoermer (T-Mobile) for stimu-lating discussions, assistance during theimplementation, and cooperation in deployingthe testbed in Berlin. The work described in thisarticle has been funded partly by the German

Ministry of Education and Research (BMBF) inthe collaborative projects ScaleNet and EASY-C and by the European Union (EU) in theWINNER II project.

REFERENCES[1] R. Knopp and P. Humblet, “Information Capacity and

Power Control in Single-Cell Multiuser Communica-tions,” Proc. IEEE ICC, vol. 1, June 1995, pp. 331–35.

[2] G. Raleigh and J. Cioffi, “Spatio-Temporal Coding forWireless Communication,” IEEE Trans. Commun., vol.46, no. 3, Mar. 1998, pp. 357–66.

[3] T. Bonald, “A Score-based Opportunistic Scheduler forFading Radio Channels,” Proc. 5th Euro. Wireless Conf.,Feb. 2004.

[4] L. Zheng and D. Tse, “Diversity and Multiplexing: A Fun-damental Trade-Off between in Multiple Antenna Chan-nels,” IEEE Trans. Info. Theory, vol. 49, no. 5, May2003, pp. 1073–96.

[5] V. Jungnickel et al., “Link Adaptation in a Multi-Anten-na System,” Proc. 57th IEEE VTC 2003-Spring, vol. 2,2003, pp. 862–66.

[6] 3GPP R1-051470, “Iterative Table-Driven (ITA) MIMOConcept for E-UTRA,” 2005.

[7] M. Schellmann et al., “Rate-Maximized Switching betweenSpatial Transmission Modes,” Proc. IEEE 40th AsilomarConf. Signals, Sys., Comp., Nov. 2006, pp. 1635–39.

[8] M. Schellmann et al., “A Fair Score-Based Scheduler forSpatial Transmission Mode Selection,” Proc. IEEE 41st Asilo-mar Conf. Signals, Sys., Comp., Nov. 2007, pp. 1961–66.

[9] L. Thiele et al., “Capacity Scaling of Multi-User MIMOwith Limited Feedback in a Multi-Cell Environment,”Proc. IEEE 41st Asilomar Conf. Signals, Sys., Comp.,Nov. 2007, pp. 93–100.

[10] T. Wirth et al., “Realtime Multi-User Multi-AntennaDownlink Measurements,” Proc. IEEE WCNC, Mar.2008, pp. 1328–33.

[11] L. Thiele et al., “On the Value of Synchronous Down-link MIMO-OFDMA Systems with Linear Equalizers,”Proc. IEEE ISWCS ‘08, Oct. 2008, pp. 428–32.

[12] V. Jungnickel et al., “Synchronization of CooperativeBase Stations,” IEEE ISWCS ‘08, Oct. 2008, pp. 329–34.

[13] L. Thiele et al., “Multi-Cell Channel Estimation usingVirtual Pilots,” Proc. IEEE 67th VTC 2008-Spring, May2008, pp. 1211–15.

[14] 3GPP TS 36.211 V8.0.0, “E-UTRA — Physical Channelsand Modulation (Release 8),” Sept. 2007.

[15] V. Jungnickel et al., “Demonstration of Virtual MIMOin the Uplink,” Proc. IET Smart Antennas CooperativeCommun. Seminar, London, U.K., Oct. 2007.

[16] V. Jungnickel et al., “Feedback Design for Multi-UserMIMO Systems,” Proc. 13th Int’l. OFDM Wksp., 2008,pp. 188–92.

BIOGRAPHIESVOLKER JUNGNICKEL [M‘99] ([email protected]) received a Dr.rer. nat. (Ph.D.) degree in physics from Humboldt Univer-sität zu Berlin, Germany, in 1995. He worked on semicon-ductor quantum dots and laser medicine before joiningHHI in 1997. He has worked on high-speed indoor wirelessinfrared links, 1 Gb/s MIMO-OFDM radio transmission, andinitial LTE trials. He is a lecturer at Technische UniversitätBerlin and project leader at Heinrich-Heinz-Institut (HHI).His current research concerns interference reduction in cel-lular networks.

MALTE SCHELLMANN [S‘05] received a Dip l .-Ing. (M.S.)degree in information technology from Technische Uni-versität München, Germany, in 2003. For his diploma the-sis, he worked on equalization in MIMO systems atAdvanced Micro Devices (AMD), Dresden, Germany. In2004 he joined HHI, where he is currently pursuing a Dr.-Ing. (Ph.D.) degree. His current research concerns practi-cal aspects of multiuser MIMO-OFDM communicationsand particularly focuses on transmission via time-varyingchannels.

LARS THIELE [S‘05] received a Dipl.-Ing. (M.S.) degree in elec-trical engineering from Technische Universität Berlin, Ger-many, in 2005. He worked in the vision research laboratoryat the University of California, Santa Barbara (UCSB). In2005 he joined HHI where he is currently pursuing his Dr.-Ing. (Ph.D.) degree. His main research interests focus onfair resource allocation algorithms in combination with

Our results confirm

that there is a

significantly higher

benefit of multiple

antennas in a cellular

network if the

mobile terminals are

aware of the

frequency-selective

interference and

sharing of resources

among multiple

users in a cell is

enabled.




physical layer optimization at the receiver and/or transmit-ter and its assessment in cellular OFDM systems.

THOMAS WIRTH received a Dipl.-Inform. (M.S.) degree incomputer science from the Universität Würzburg, Ger-many, in 2004. In 2004 he joined Universität Bremen,Germany, where he worked in the field of robotics. In2006 he joined HHI where he is pursuing his Dr.-Ing.(Ph.D.) degree. His research interests are in the field ofQoS-aware multiuser resource allocation algorithms forMIMO-OFDMA systems, including real-time implementa-tion and field trials.

THOMAS HAUSTEIN received a Dr.-Ing. (Ph.D.) degree inmobile communications in 2006 from the Technische Uni-versität Berlin. In 1997 he joined HHI working on wirelessinfrared systems and radio communications with multipleantennas and OFDM. He focused on real-time algorithmsfor baseband processing and advanced multiuser resourceallocation. In 2006 he joined Nokia Siemens Networks con-ducting research for LTE and LTE-Advanced. Recently, hereturned to HHI as head of the Broadband Mobile Commu-nications Department.

OTTO KOCH has received a Dipl.-Ing. degree (M.S.) in RFengineering from Universitaet Ulm, Germany, in 1996. Hejoined Bosch Telecom and developed RF modules for space

applications. In 1999 he joined Siemens and developedhardware for mobile communications. From 2003 to 2006he worked in a joint venture of NEC and Siemens in theUnited Kingdom. Since 2006 he has been with NokiaSiemens Networks, Munich, Germany, and is responsiblefor the LTE/SAE field trial initiative (LSTI).

WOLFGANG ZIRWAS received a Dipl.-Ing (M.S.) in communica-tion technologies from Technische Universität München in1987. He joined Siemens working on RF communicationsystems. He researched broadband transmission over opti-cal fiber, coax-cable, twisted pair, and radio. Since 1999 hehas focused on multihop, MIMO, and distributed coopera-tive antennas. In 2006 he contributed to the LTE MIMOstandardization. He has filed more than 200 patents andreceived the inventor of the year award from Siemens in1997.

EGON SCHULZ received a Dr.-Ing. (Ph.D.) degree from Tech-nische Universität Darmstadt, Germany, in 1988. In 1988he joined Siemens, Munich. He developed radio link proto-cols and contributed to the ETSI standardization of GSM.In 1992 he became a professor at the Universität Darm-stadt and returned to Siemens in 1993 as head of DECTand WCDMA system engineering. In 1998 he contributedto RAN simulation in the UMTS standardization. Since 2000has been serving as head of future radio concepts.



Documents

Interference-Aware Scheduling in the Multiuser MIMO … ComMag.pdf · Interference is the limiting factor in cellular ... (UMTS), that is, the same ... Note that the CQI may include