The gradient { <strong>of</strong> the unconstrained cost function(NJ = E eN Q √Ns ) }γ b with respect to ω is givenby∇J ω = N eN= −2 P TXσ 2 νßÞÐ ßÞÐVI INTERNATIONAL TELECOMMUNICATIONS SYMPOSIUM (ITS2006), SEPTEMBER 3-6, 2006, FORTALEZA-CE, BRAZILTABLE I: mBER-TD-DB Algorithm1) Initializationω =Ö10 . . .0 1 0 . . .0 1 · · ·T(19)L1M zeros M zeros2) Compute f(γ b ) and R kdQ( √ N s γ b ) dγ bdγ b dωN eN E {f(γ b )R(b)}ω ,(15)where γ b is given by (11).{ By introducing } the iteration index k and by defining R k =E f(γ b )R(b) , we can write the normalized gradient updateasω k+1 = ω k − µ ′ ∇J ω‖∇J ω ‖(16)= ω k + µ ′ R k ω k‖R k ω k ‖ ,where µ ′ is the adaptation coefficient. This update is followedby the normalization <strong>of</strong> ω k+1ω k+1 = ω k+1‖ω k+1 ‖ . (17)Equations (16) and (17) compose the core <strong>of</strong> the proposedalgorithm. We highlight that, at the update step (16), ω k wasnormalized, and ω k+1 will be normalized at the end <strong>of</strong> theR kω k‖R k ω k ‖current iteration. Hence, the update term has only thefunction <strong>of</strong> changing the direction <strong>of</strong> the precoder vector ω kbut not its norm.It is worth mentioning that R k depends on ω k , sincef(γ b ) depends on ω k , but it is considered to be constant toperform one iteration <strong>of</strong> the algorithm and then its value isupdated to match the new transmit filter ω k+1 . So, we do notwant to change too much ω k in order to slowly converge tothe optimum solution. Thus, we propose to combine the oldR kω k‖R k ω k ‖precoder ω k with the update term in order to obtainthe new precoder ω k+1 . We obtain thus the following updateequation for ω kω k+1 = (1 − µ)ω k + µ R kω k‖R k ω k ‖ , (18)where µ is the adaptation coefficient.The proposed algorithm, called mBER-TD-DB (minimumBER for Transmit Diversity and Downlink Beamforming), issummarized in Table I, where the expectation was replaced bya temporal estimation over B blocks. The initialization givenby (19) corresponds to directly wiring each virtual antenna tothe corresponding real antenna. This initialization was chosensince it represents a neutral situation, where the precoder ωdoesn’t affect the transmission.For the simulations considered in this paper, we haveempirically determined that µ = 0.25 was a good choice for afast convergence. Higher values <strong>of</strong> µ are undesirable becauseR k depends on ω k and thus if we change too much the actualsolution, it can drastically change the matrix R k causing thenon-convergence <strong>of</strong> the algorithm. On the other side, lowervalues will only slow down the convergence.The analytical pro<strong>of</strong> <strong>of</strong> convergence <strong>of</strong> the proposed algorithmis a complex task and have not yet been done, but weω H kγ b = P R(b)ω kN sTXσν2 , f(γ b ) =2 √ exp− Nsγ b2πN sγ b 2∀bR k = 1 B−1f(γ b )R(b)Bb=03) Update ω kω k+1 = (1 − µ) ω k + µ R kω k‖R k ω k ‖4) Normalization <strong>of</strong> ω k+1ω k+1 = ω k+1‖ω k+1 ‖5) Go to 2 until convergencehave not observed one single case <strong>of</strong> divergence among all thesimulations performed.IV. SIMULATION RESULTSWe consider the downlink <strong>of</strong> one cell <strong>of</strong> a wireless system.The data is transmitted in blocks <strong>of</strong> length N b and we assume,without loss <strong>of</strong> generality, that the channel realization isindependent from one block to another, i.e., path coefficientsvary from one block to another independently. The BS isequipped with a linear array <strong>of</strong> K = 4 antennas and the interelementdistance is λc2 , where λ c is the carrier wavelength. Thetransmit precoder is K×L (see Fig. 1) and we have consideredL = 2 virtual antennas. The Alamouti scheme [2] is usedas OSTBC. Furthermore, we assume that the instantaneousDCCMs R(b) for all blocks are perfectly known at the BS.In order to assess the performance <strong>of</strong> the proposed technique,we have simulated N t = 3 × 10 4 training blocks andN d = 10 5 data blocks. For each block, a different channelrealization was drawn according to the used channel model.The corresponding DCCMs R(b) <strong>of</strong> the training blocks wereused as input for the proposed algorithm, which computes theoptimum precoder ω opt , see Table I where we have B = N t .The transmit power P TX was normalized with respect to thereceiver noise variance σν 2, so that 0 dB corresponds to P TX =σν 2 (note that this normalization is equivalent to say that 0 dBcorresponds to the transmit power necessary, when using anomnidirectional antenna at the BS, to have a SNR <strong>of</strong> 0 dBat the MU). The optimum precoder is then used to evaluatethe raw BER <strong>of</strong> a 4-QAM modulation 1 at the MU for thesubsequent data blocks, using (12).In the following, we compare the performance <strong>of</strong> the proposedmBER-TD-DB, the 2D Eigen-Beamforming technique<strong>of</strong> [1], Downlink Beamforming (DB) [6] and the Alamoutischeme [2]. The Alamouti results were obtained by using theextreme real antennas to maximize the inter-antenna distanceand minimize the correlation between them. We will present1 For a 4-QAM modulation [10], we have N s = d2 min2 = 1 and NeN = 1.
VI INTERNATIONAL TELECOMMUNICATIONS SYMPOSIUM (ITS2006), SEPTEMBER 3-6, 2006, FORTALEZA-CE, BRAZIL10 0 P TX[dB]10 −1Downlink Beamforming (DB)Alamouti (TD)Eigen−BeamformingmBER−TD−DB10 −210 −2BERBER10 −310 −310 −410 −50 5 10 15 20 25Fig. 2: Performance for ∆ = 5 ◦ NLOS channels and 4-QAMmodulation as a function <strong>of</strong> the transmit power P TX .10 −42 4 6 8 10 12 14 16 18 20IterationFig. 3: Convergence <strong>of</strong> the mBER-TD-DB algorithm forP TX =20 dB.simulations for a flat non-line-<strong>of</strong>-sight (NLOS) scenario, whichcorresponds to a Rayleigh channel, and for a flat line-<strong>of</strong>-sight(LOS) scenario, corresponding to a Rician channel.100P TX=5 dBl=1l=2A. NLOS scenarioThe NLOS scenario is the same as the one simulated in [1]and corresponds to a flat-fading Rayleigh channel, i.e., allthe channel coefficients are Rayleigh distributed. This channelwas defined in [12] and has only one path perpendicular tothe multiple-transmit antennas (”broadside” as in [12]) withan angle spread <strong>of</strong> ∆. The mean DCCM can be obtainedin closed form as in [1, eq. (57)]. We have used this meanDCCM to obtain the instantaneous DCCM for each block b byconsidering the instantaneous channel generated as describedin [12, eq. (7)]. We have considered an angle spread <strong>of</strong> 5 ◦ .Fig. 2 shows the BER for all simulated techniques as afunction <strong>of</strong> the transmit power P TX . We can see that theproposed mBER-TD-DB has the same performance as theoptimum Eigen-Beamforming and they outperform the othertwo techniques. The mBER-TD-DB follows the DB up to10 dB since the other diversity branches are too weak andtheir use would only waste power. After 10 dB, the mBER-TD-DB begins to use the second diversity branch to exploitthe channel diversity. This makes the BER curve to changeits slope due to diversity. At high P TX regime (equivalent tohigh SNR regime), the mBER-TD-DB attains the same slopeas the Alamouti scheme, but with a gain in transmit power <strong>of</strong>about 1.75 dB in this scenario.The evolution <strong>of</strong> the BER during the convergence <strong>of</strong> themBER-TD-DB algorithm is shown in Fig. 3, for a transmitpower <strong>of</strong> 20 dB. It can be seen that the algorithm presents afast convergence, typically between 10 and 20 iterations.The radiation patterns for each precoder layer for themBER-TD-DB for a transmit power <strong>of</strong> 5 dB and 20 dB areshown in Fig. 4. For P TX = 5 dB, we see that the first layer(l = 1) radiates in the direction <strong>of</strong> the user, i.e., in the 5 ◦cone around 0 ◦ , while the second layer (l = 2) is switched<strong>of</strong>f.This is exactly the DB solution, since there is only oneGain [dB]Gain [dB]−10−20−30−100 −80 −60 −40 −20 0 20 40 60 80 100Angle100−10−20P TX=20 dB l=1l=2−30−80 −60 −40 −20 0 20 40 60 80AngleFig. 4: Radiation pattern <strong>of</strong> the precoder layers formBER-TD-DB for a transmit power <strong>of</strong> 5 dBand 20 dB in the NLOS scenario.significant diversity branch at such small transmit power. Onthe other hand, for P TX = 20 dB, the first layer continuesto radiate mostly in the direction <strong>of</strong> 0 ◦ , while the secondlayer radiates avoiding this region, in order to create a seconddiversity branch uncorrelated to the first one. The dashed curveshows the total radiated power from the two branches, whichbecomes less directive. The total radiated power tends to beasymptotically omni-directional for high SNR, since the arraygain becomes negligible when compared to the diversity gain.We have shown that the proposed mBER-TD-DB is able tojointly achieve TD and perform DB, performing as well asthe, also optimum, Eigen-Beamforming in a NLOS Rayleighscenario. In the sequel, we will show that mBER-TD-DBoutperforms Eigen-Beamforming in other scenarios, being anattractive solution for the general application.