Common use of User Grouping Clause in Contracts

User Grouping. Enlarged cooperation areas have been introduced above to obtain sufficiently large penetration rates. Furthermore, the tortoise concept has helped to decouple adjacent CAs. Therefore, the analysis of the user grouping can be restricted here to a single CA comprising 9 cells - i.e. 3 sites a’ 3 cells - for an exemplary 4x2 MIMO system according to [3GPPTR1]. The maximum number of servable data streams per physical resource block NDS,max, is NDS,max ≤ min(Ntx,MNrx), where M is the number of UEs, so each cell might serve up to 4 data streams. This corresponds to 4*9=36 potential UEs (or UE antennas) per CA simultaneously served on each PRB. Eigenvalues in [dB] for 4 Tx eNBs and 1 Rx UE; Tx power adapted 20  [dB]0 -40 -60 -80 Mk, # of UEs in resource block Figure 5.7: Singular values of a channel matrix of dimension 36 x Mk, with 36 transmitters with fixed beams, as a function of the number of UEs Mk, that are randomly selected to co-exist in one resource block. Under these conditions, Figure 5.7 illustrates the typical singular value distribution of the total channel matrix from all 36 transmitters to M users, as being observed for an increasingly loaded system with M increasing. Serving more than k=13...18 UEs seems to be quite unreasonable due to the very high spread of singular values, with some singular values being extremely small. Additionally, the variation over frequency is small as indicated by the differently colored lines. SVD-based joint transmit- and receive beamforming would be the best transmission strategy (although it can not be implemented for the here considered non-cooperative multi-receiver case). Its performance would be severely limited by a large singularvalue spread of the total channel matrix. Beamformer design would be sensitive and difficult and only a minority of the singular values could contribute meaningfully to the sum throughput. Given these results, it could be concluded that attainable gains for JP CoMP indeed seems to have a fundamental limit. However, there is already a useful hint in Figure 5.7. Serving at most 2 instead of 3, or at most 3 instead of 4 UEs per cell would be accompanied by about 10-20dB lower spread of singular values. The number of served users per 9 cells would then drop from above 18 to max 18 and from above 27 to max 27 UEs. This would result in large reductions of the singular value spreads of the channel matrices. The conclusion is that loading one single extra UE into one single cell might already spoil the overall channel conditions for JT CoMP. A main reason for this is the rather poor orthogonality of the fixed beams within cells, due to a significant beam overlap. A further reason for large eigenvalue spreads in highly loaded systems with random user grouping or positioning, is that some users will be located far from all transmitters. In a highly loaded system, we then in effect have more users than the number of useful transmitters. This effect is clearly evident also in the evaluations of ▇▇▇▇▇▇▇▇ ▇▇-▇ in the “random user grouping” cases. Radio channels from different cells and sites are most of the time uncorrelated. Therefore, mutual orthogonality is typically spoiled between the UEs of a single cell and not between UEs of different cells. This observation is very helpful. It points to a strategy for reducing the CA-wide untractable grouping problem for e.g. 90 UEs (9 cells each scheduling 10 active UEs) to 9 low- complexity cell specific schedulings: In a first stage, cell-specific schedulers would place all users within cells on orthogonal or close to orthogonal frequency-spatial transmission resources. The users in different cells that are thereby allocated to a resource block will then mostly have differing (instantaneously) strongest eNB´s or beams. The resulting channel matrices therefore tend to be well-conditioned, with reduced spread of singular values as compared to a random user grouping. The resulting joint precoder design problem for these user groups therefore becomes easier to solve. In a second stage, sets of linear JT CoMP precoders are then designed for the whole cooperation area, one per PRB. In each PRB, the sets of UEs are those that have been scheduled in each PRB by the cell-specific scheduler. The joint precoder designs are performed for channel matrices that include the cell-specific beamformers, as outlined in Section A2.2.1 in Appendix A2-2. As investigated there, such pre-selected and pre-beamformed sets of users have much better channel properties for JT CoMP as compared to random user grouping followed by a joint CA-wide precoder design. Cooperation area CAb scheduler site s1 site s2 site s3 Cell 2 Cell 3 scheduler scheduler NBS=4 NUE=2 Cell C schedule K=10 UEs per celll In each PRB, the schedulers therefore just have to find the best groups of e.g. 3 UEs out of overall e.g. 10 UEs, which are to be served by e.g. 4 beams within one particular cell. This is precisely what multi-user MIMO schedulers are being designed to do today in LTE and LTE-A systems, even so the scheduling itself will be more complex as it has to be adapted to the IMF- A framework. As a further benefit, this leads to a smooth scheduler evolution by reusing well known - potentially enhanced - MU MIMO schedulers. The difference now is that the second stage JP CoMP processor is running over the whole CA. It uses the per-cell scheduling decisions plus the accordingly reported CSI information to perform a joint precoding design per PRB, resulting in a precoding matrix W over all cells of the CA. The design of this CA-wide joint precoding matrix is discussed in Subsection 5.2.3 below. As compared to cellular MU-MIMO transmission via separate cellular downlink beamformers, most of the UEs will then experience a further rank enhancement and only by chance will some of the UEs see bad channel conditions, requiring a final fine tuning of the CA-wide scheduler. In the evaluations of ▇▇▇▇▇▇▇▇ ▇▇-▇, this user grouping strategy has been found to be the most important effect that generates and explains the obtained CoMP gains.

User Grouping. Enlarged cooperation areas have been introduced above to obtain sufficiently large penetration rates. Furthermore, the tortoise concept has helped to decouple adjacent CAs. Therefore, the analysis of the user grouping can be restricted here to a single CA comprising 9 cells - i.e. 3 sites a’ 3 cells - for an exemplary 4x2 MIMO system according to [3GPPTR1]. The maximum number of servable data streams per physical resource block NDS,max, is NDS,max ≤ min(Ntx,MNrx), where M is the number of UEs, so each cell might serve up to 4 data streams. This corresponds to 4*9=36 potential UEs (or UE antennas) per CA simultaneously served on each PRB. Eigenvalues in [dB] for 4 Tx eNBs and 1 Rx UE; Tx power adapted 20  λ [dB]0 -20 -40 -60 -80 Mk, # of UEs in resource block Figure 5.7: Singular values of a channel matrix of dimension 36 x Mk, with 36 transmitters with fixed beams, as a function of the number of UEs Mk, that are randomly selected to co-exist in one resource block. Under these conditions, Figure 5.7 illustrates the typical singular value distribution of the total channel matrix from all 36 transmitters to M users, as being observed for an increasingly loaded system with M increasing. Serving more than k=13...18 UEs seems to be quite unreasonable due to the very high spread of singular values, with some singular values being extremely small. Additionally, the variation over frequency is small as indicated by the differently colored lines. SVD-based joint transmit- and receive beamforming would be the best transmission strategy (although it can not be implemented for the here considered non-cooperative multi-receiver case). Its performance would be severely limited by a large singularvalue spread of the total channel matrix. Beamformer design would be sensitive and difficult and only a minority of the singular values could contribute meaningfully to the sum throughput. Given these results, it could be concluded that attainable gains for JP CoMP indeed seems to have a fundamental limit. However, there is already a useful hint in Figure 5.7. Serving at most 2 instead of 3, or at most 3 instead of 4 UEs per cell would be accompanied by about 10-20dB lower spread of singular values. The number of served users per 9 cells would then drop from above 18 to max 18 and from above 27 to max 27 UEs. This would result in large reductions of the singular value spreads of the channel matrices. The conclusion is that loading one single extra UE into one single cell might already spoil the overall channel conditions for JT CoMP. A main reason for this is the rather poor orthogonality of the fixed beams within cells, due to a significant beam overlap. A further reason for large eigenvalue spreads in highly loaded systems with random user grouping or positioning, is that some users will be located far from all transmitters. In a highly loaded system, we then in effect have more users than the number of useful transmitters. This effect is clearly evident also in the evaluations of ▇▇▇▇▇▇▇▇ ▇▇-▇ in the “random user grouping” cases. Radio channels from different cells and sites are most of the time uncorrelated. Therefore, mutual orthogonality is typically spoiled between the UEs of a single cell and not between UEs of different cells. This observation is very helpful. It points to a strategy for reducing the CA-wide untractable grouping problem for e.g. 90 UEs (9 cells each scheduling 10 active UEs) to 9 low- complexity cell specific schedulings: In a first stage, cell-specific schedulers would place all users within cells on orthogonal or close to orthogonal frequency-spatial transmission resources. The users in different cells that are thereby allocated to a resource block will then mostly have differing (instantaneously) strongest eNB´s or beams. The resulting channel matrices therefore tend to be well-conditioned, with reduced spread of singular values as compared to a random user grouping. The resulting joint precoder design problem for these user groups therefore becomes easier to solve. In a second stage, sets of linear JT CoMP precoders are then designed for the whole cooperation area, one per PRB. In each PRB, the sets of UEs are those that have been scheduled in each PRB by the cell-specific scheduler. The joint precoder designs are performed for channel matrices that include the cell-specific beamformers, as outlined in Section A2.2.1 in Appendix A2-2. As investigated there, such pre-selected and pre-beamformed sets of users have much better channel properties for JT CoMP as compared to random user grouping followed by a joint CA-wide precoder design. Cooperation area CAb scheduler site s1 site s2 site s3 Cell 2 Cell 3 scheduler scheduler NBS=4 NUE=2 Cell C schedule K=10 UEs per celll In each PRB, the schedulers therefore just have to find the best groups of e.g. 3 UEs out of overall e.g. 10 UEs, which are to be served by e.g. 4 beams within one particular cell. This is precisely what multi-user MIMO schedulers are being designed to do today in LTE and LTE-A systems, even so the scheduling itself will be more complex as it has to be adapted to the IMF- A framework. As a further benefit, this leads to a smooth scheduler evolution by reusing well known - potentially enhanced - MU MIMO schedulers. The difference now is that the second stage JP CoMP processor is running over the whole CA. It uses the per-cell scheduling decisions plus the accordingly reported CSI information to perform a joint precoding design per PRB, resulting in a precoding matrix W over all cells of the CA. The design of this CA-wide joint precoding matrix is discussed in Subsection 5.2.3 below. As compared to cellular MU-MIMO transmission via separate cellular downlink beamformers, most of the UEs will then experience a further rank enhancement and only by chance will some of the UEs see bad channel conditions, requiring a final fine tuning of the CA-wide scheduler. In the evaluations of ▇▇▇▇▇▇▇▇ ▇▇-▇, this user grouping strategy has been found to be the most important effect that generates and explains the obtained CoMP gains.