Interference Mitigation via Scheduling for the MIMO Broadcast Channel with Limited Feedback

Interference Mitigation via Scheduling for the MIMO Broadcast Channel with Limited Feedback Tae Hyun Kim The Department of Electrical and Computer Engineering The University of Illinois at Urbana-Champaign, IL 61801-2302 tkim56@illinois.edu Sunghyun Choi School of Electrical Engineering and INMC Seoul National University, Seoul 151-744, Korea schoi@snu.ac.kr Abstract The multiple-input multiple-output (MIMO broadcast channel uses multiple antennas at the transmitter to concurrently deliver information to multiple receivers. Despite the theoretical capacity gains over single user MIMO, the actual throughput of the system with limited feedback on channel state information suffers due to interspatial stream interference. On top of the orthogonal beamforming with limited feedback [1], this paper introduces two novel stream schedulers to mitigate the interference. By the proposed adaptive spatial stream schedulers, an adaptive number of spatial streams are scheduled and thus, reduces the interference, which also achieves power gains by concentrating power into fewer streams. The two proposed algorithms are evaluated and shown to effectively mitigate the interference. I. INTRODUCTION The multiple-input multiple-output (MIMO broadcast channel uses multiple antennas at the transmitter to support concurrent transmission to different receivers. The receivers in a cellular network with a MIMO broadcast-capable base station are spatially separated and served at the same time. This advantage of concurrent services has attracted much research attention with several approaches that have been proposed thus far (see [2] and the references therein. In the MIMO broadcast channel transmit signals are typically precoded with an aid of channel state information (CSI at the transmitter so as to minimize the interspatial stream interference. 1 For this purpose, CSI is quantized and fed back from the receivers to the transmitter by using a finite size codebook, which eventually introduces the interference in the system due to the imperfection of the quantized CSI. This paper focuses on the mitigation of such interference in one type of MIMO broadcast systems, namely, orthogonal beamforming with limited feedback (OBLF [1], which is also known as per user unitary rate control in standard associations [3]. The countermeasures to mitigate this interference are summarized as follows: enlarge the codebook, perform joint processing at the transmitter and receivers using multiple receive antennas [4], schedule users with small mutual interference This material is based in part upon work supported by National Science Foundation award 0831670, the DARPA IT-MANET program, Grant W911NF-07-1-0028 and the IT R&D program of MKE/IITA [2009-F-044-01, Development of cooperative operation profiles in multicell wireless systems]. 1 In this paper, we refer to interspatial stream interference as the interference. It is also called by other names such as multiuser interference and co-channel interference. and schedule a smaller number of concurrent streams and concentrate transmit power into them. Our interests in this work is the last one; interference mitigation via less streams can be achieved by a scheduler at the transmitter that can find a subset of streams that yield the largest sum throughput. Many schedulers have been proposed for systems with perfect CSI at the transmitter (CSIT and recently for quantized CSIT [1], [5], [6], but scheduling for mitigating the interference has not been considered so far. Based on OBLF, this paper proposes two spatial stream schedulers that choose a subset of streams to mitigate the interference in the MIMO broadcast channel with limited feedback. The proposed spatial stream schedulers select an adaptive number of spatial streams to maximize sum throughput by controlling the amount of the interference; they are designed to find a subset of spatial streams with small interference and concentrate transmit power to the chosen, thus achieving power gain. Their computational performance is also analyzed. Our contributions are summarized as follows. First, adaptive schedulers for practical OBLF systems are proposed and their computational performance is analyzed. Second, we have shown that adaptive scheduling is essential to have at least logarithmic throughput increase as signal-to-noise-ratio (SNR grows, thanks to the power gain as well as reduced the interference. Lastly, one of our adaptive schedulers is designed to rely on a high feedback rate to make instantaneous and adaptive decisions at the transmitter, showing that such decisions are highly expensive in terms of system resources. In what follows, Section II gives the considered system model. Adaptive schedulers and their computational performance are introduced in Section III. Section IV gives analysis on the simulations results. Finally, Section V concludes this paper. A. Signal Model II. SYSTEM MODEL We consider a MIMO broadcast system with one transmitter and multiple receivers that have a single antenna. The number of antennas at the base station (transmitter is denoted by M. Assuming narrow band signaling and flat fading, the multipath fading channel from the base station to a certain receiver is described by h C M 1. The discrete time received signal

delivered by beamformer k is 2 E s y k = K hh( w k x }{{ k } + information k K\{k} w k x k } {{ } interference + n, (1 where n C is a zero-mean additive complex Gaussian noise with variance N 0. E s is the mean symbol energy of the received signal. After scheduling M or less spatial streams by the transmitter, w k C M 1 are chosen as transmit beamformers. K is a set of beamformer indices that create spatial streams to the receivers. The term x k represents the transmitted symbol associated with beamformer w k. We assume that mean energy of the symbol x k is statistically normalized (E{ x k 2 } =1. From (1, the SINR for a spatial stream conveyed on beamformer w k is E s h H w k 2 φ k = E s k K\{k} hh w k 2 + N 0 K, (2 which depends upon our choices of the set of beamformers (K and transmit beamformers themselves (w k. B. Orthogonal Beamforming with Limited Feedback (OBLF Our proposals against the interference are on top of OBLF. The needed details of OBLF for exposition are given in this section. Let us decompose CSI into channel direction information (CDI and magnitude. Limited feedback is an index feedback of the codeword which is desirable for the transmitter to know as CDI. In OBLF, M orthogonal codewords form a subset of the codebook, which will be referred to as a (orthogonal code set. Multiple code sets constitute the entire codebook, which is defined as F := {w k : k V i=1v i } = {w k : k =1,,M V }, (3 where w k is the k-th codeword and V i = {(i 1M +1, (i 1M +2,,i M} for i = 1,, V denotes the set of codeword indices in i-th orthogonal code set. The codewords w k,k V i are thus orthogonal to each other in the same set. A code set index for a certain codeword is easily obtained by i = (k 1/M +1. In OBLF, the codeword for feedback is found by the criterion k = arg max k h H w k h, (4 which is to select the codeword whose CDI has the least inner angle to the channel direction. As implicitly shown in (3, in OBLF the codewords in a code set are used for transmit beamforming as beam weights for M spatial streams. If a receiver reports that its CDI is the k-th codeword, the other codewords in the code set of k- th s are close to orthogonal to the channel. Therefore, if the reported channel direction is indeed identical to the k-th s, the 2 We use superscript T for transpose, superscript H for conjugate transpose, and superscript for the matrix pseudo inverse. The expression denotes the usual Euclidean vector norm. spatial stream destined to that direction will not interfere with other streams. Provided that the transmitter chooses code set V i for beamforming, a receiver can decompose an incoming signal to individual SNRs corresponding to the beamformers in V i by applying different transmit beamformers. Define individual SNR by γ n := Es N 0 h H w n 2 where n is the beamformer index of the considered signal. The transmitter can collect this set of information from the receivers and reconstruct SINRs of any combinations of spatial streams by φ TX γ k k = n K\{k} γ n + K, (5 which may be slightly different from φ k at the receiver due to quantization for feedback. If such a set of individual SNR information is not necessary, the receivers may feedback SINR itself only. For later discussion, the numerator of (5, γ k, is named as an individual information SNR while individual SNRs refer to all γ n,n K. III. SPATIAL STREAM SCHEDULERS In this section, previous schedulers for OBLF are explained, and then, schedulers that choose an adaptive number of spatial streams are proposed to improve sum throughput. A. Previous Schedulers for OBLF The purpose of spatial stream scheduling is to maximize sum throughput by selecting a subset of available spatial streams, which is formulated as ( K = arg max max log 1+φ TX k,j k (K, (6 K V i :1 i V 1 j k J k :k K k K where φ TX k is from (5 or explicit SINR feedback. J k is the number of spatial streams using codeword k, which have been reported by receivers to the transmitter. An index j k is used to indicate a certain spatial stream out of J k s. The term j k is also used to refer to the individual information SNR of a certain spatial stream; the term γ k,jk denotes γ k for the j k -th spatial stream among all streams that use the beamformer k, i.e., the numerator of φ TX k,j k in (5. Without loss of generality, it is assumed for the following discussion that γ k,j1 γ k,j2 for all 0 <j 1 <j 2 J k, k, which is done by sorting spatial streams in a descending order with respect to their individual information SNRs γ k,jk. Let us denote the number of receivers by Z. The brute-force search for an optimal K takes O(M2 Z time, assuming M< Z. Therefore, spatial streams that maximize sum throughput are found only when Z is small or moderate. Let us call this brute-force search as OptSch for discussion. For OBLF, it is introduced in [1] a user scheduler that always chooses M streams, which is described in Algorithm 1. Let us name this scheduler as M-FixSch. The running time of this heuristic is easily proven as follows. Proposition 1. Algorithm 1, which is named as M-FixSch, has a complexity of O(Z + V M.

Algorithm 1 Fixed M stream scheduler (M-FixSch [1] 2: jk = arg max 0 j Jk φ TX k,j for k F j=0means no stream for that beamformer 3: 4: Execution: 5: i = arg max 1 i V log k V ( 1+φ TX i k,j k 6: K = V i 7: return K and jk jk are for Algorithm 2 Proof: It first searches the stream with the largest SINR for each beamformer ( O(Z on Line 2 and then finds the code set that yields the largest sum throughput ( O(MV on Line 5. The computational complexity of M-FixSch is thus exponentially less than OptSch. Moreover, it has a desirable feature that it actually achieves the optimality in terms of sum throughput if K = M. However, if sum throughput is mainly limited by nonzero interference, it is possible that K < M. In this case, it is probable that a fewer scheduled streams would experience less sum interference and per stream transmit power would be larger, which together results in higher SINRs of served streams. Algorithm 2 Adaptive scheduler with individual SNRs (Adp- Sch 2: R =0, K = 3: for k F do j k = 1 if stream with beamformer k and otherwise j k =0 4: find jk by M-FixSch 5: 6: Execution: 7: for all code set i =1,,V do 8: K = V i \{k : j k =0} 9: while j k jk for k V i do Mstream-selection begins 10: R =0, K = 11: while K do Stream deselection begins 12: R = log( 1+φ TX k K k,j k (K 13: if R>R then 14: R = R, K = K 15: end if 16: remove k = arg max k K γn from K n K\{k} 17: end while 18: if R >R then 19: R = R, K = K 20: end if 21: K = V i \{k : j k =0} 22: increase j k : k = arg max k K:j k j k n K\{k } γn 23: end while 24: end for 25: return K B. Adaptive Scheduler with Individual SNRs (AdpSch Observe that the sum throughput in (6 is mainly contributed by two factors: (i the SINRs of the selected spatial streams and (ii the number of the streams. To obtain the largest SINR, a stream has to have strong signal power γ k as well as small sum interference n K\{k} γ n where K is a set of streams. Instead of seeking a set of streams simultaneously satisfying both factors, AdpSch, which is described in Algorithm 2, performs a two step procedure: for each code set, it does (i selection of M stream-combination and (ii subset selection of that stream combination. The two step procedure is iterated until a search is performed on a sufficiently large space. 1 M stream-selection: The first step is to select M or less spatial streams for a given set if the number of streams in the set is less than M. Without loss of generality, we assume there are always M streams in the rest of the discussion. M stream-selection is from the streams with the largest individual SNRs to those with smaller ones. For each beamformer k in a code set, the spatial stream with the largest individual information SNR γ k is found, forming M stream-selection. After performing stream subset selection, the stream of the beamformer with the largest sum interference is replaced with the stream of the next largest information SNR, which has the same beamformer, and so forth. Then the central question in M stream-selection is when the algorithm should cease evolving to a different set of spatial streams. The scheduler stops evolution of stream-selection if it has considered all streams of which individual information SNRs are larger than or equal to those of streams selected by M-FixSch. This rule is based on the fact that the streams of which j k is less than or equal to jk have stronger signal power and larger sum interference than those that would have been scheduled by M-FixSch. 2 Stream subset selection: The second step is to choose a subset of the M streams given by the M stream-selection. Instead of computing all possible subsets, the one with the largest sum interference is iteratively deselected from the streams, which is done on Line 16. The subset that yields the largest sum throughput is chosen after all streams are deselected. In sum, the algorithm reveals a low complexity which can be proven as follows: Proposition 2. AdpSch described in Algorithm 2 has a complexity of O(Z max{m 2, log Z}. Proof: The adaptive scheduler performs as follows: M stream-selection (S c, stream deselection (D c for each code set (V sets and summation to obtain sum throughput (M. Let us denote the complexity as (S c V D c M. First of all, all spatial streams should be sorted in a descending order and classified to corresponding codewords, which takes Z log Z time. For M stream-selection, it is not difficult to observe that S c V depends on the distribution of spatial streams over a codebook and is maximized when all spatial streams are mapped to the codewords in a single code set. This can be proven by the fact that the first M stream-selection on a code set takes M while subsequent selection takes only one spatial stream for replacement. Thus, S c V Z M +1. In addition, D c M where the equality holds when all codewords in a code set have at least one spatial stream.

In sum, the complexity is S c V D c M + Z log Z (Z M +1D c M + Z log Z (Z M +1M 2 + Z log Z Z M 2 + Z log = O ( Z max{m 2, log Z}, (7 which completes the proof. In terms of the number of receivers, the complexity of the proposed adaptive scheduler is exponentially less than O ( M2 Z required by the exhaustive search, and comparative to that of M-FixSch, which is O(Z + V M (note that M is usually a small number. Specifically, for M = 2, the complexity is O(Z log Z unless Z 16. In practical scenarios where M > 2 and Z 2 9, AdpSch runs in O(Z M 2, which is linear to the number of receivers. It also turns out by Proposition 2 that AdpSch has a desirable feature that the complexity stays the same even with the increase of the codebook size; in fact, it decreases as V increases due to the first M stream-selection for each set, which always takes M streams if any. Algorithm 3 Any fixed m stream scheduler (Any-FixSch 2: jk = arg max 0 j Jk φ TX k,j for k F 3: W i = predetermined subset of beamformers such that W i V i and W i = m 4: 5: Execution: 6: i = arg max 1 i V k W i log ( 1+φ TX k,j k over a subset W i 7: K = W i 8: return K Summation is C. Any-FixSch: Generalized M-FixSch The adaptive scheduler in the previous subsection demonstrates that even suboptimal adaptive scheduler requires substantial knowledge on γ n to make decisions on the number of streams and choices on which streams will be served. This tells us that as long as such decisions are instantaneously and adaptively made at the transmitter, the information should be fed back to the transmitter, which is a prohibitively large amount. Inspired by this fact and simplicity of M-FixSch, Any- FixSch is proposed, which generalizes M-FixSch to arbitrary m streams cases and achieves part of sum throughput gains by adaptive scheduling. Since Any-FixSch is nothing but to schedule m instead of M streams, the complexity decreases from that of M-FixSch, O(Z + V M to O(Z + V m. Itis described in Algorithm 3. By comparison between Algorithms 1 and 3, two differences are observed. On Line 3 of Algorithm 3, predetermined subsets W i V i are used as a filter to strain off M m beamformers. Another is the summation on Line 6. It sums up throughput of a subset of spatial streams in a code set. How can Any-FixSch avoid making the expensive adaptive decisions at the transmitter? Any-FixSch presumes that a TABLE I SUMMARY OF SPATIAL STREAM SCHEDULERS FOR OBLF Running time OptSch O(M 2 Z AdpSch O(Z max{m 2, log Z} Any-FixSch O(Z + V m M-FixSch O(Z + V M Feedback rate ( B M + log 2 V M B + log 2 ( V M certain number of streams are sufficient to be scheduled, and the number is m. This requires only the same amount of feedback as M-FixSch and other user schedulers in the literature do [5], [6]. As a result, the performance of Any-FixSch heavily depends on the choice of m. Ideally, it should be m = K for every instance, which is possible only when the transmitter is fed more information on receiver channels. Instead, it is assumed that the transmitter in our system keeps tracking E [ K ] by intermittently running either OptSch or AdpSch. Thus, Any- FixSch is not a stand-alone scheduler. It is also considerable to track K by increase and decrease of m over time according to resulting sum throughput (see similar type of tracking for a MAC protocol rate adaptation in [7], which is left as future work. In Any-FixSch, the m codewords that will be used at the transmitter are predetermined by W i on Line 3, which does not necessarily yield the largest SINR. However, fixing them does not require a receiver to send back either information on which codewords are used for the SINR or multiple SINR values for different combinations of codewords in use. As the SINR reconstruction by (5 in AdpSch, this adaptation also needs a substantial amount of feedback, which should be avoided to keep the feedback rate low. D. Feedback Rate for Spatial Stream Schedulers In Table I, the computational performance of all schedulers discussed in this section are summarized. In addition, feedback rates for the schedulers are also listed. When OptSch and AdpSch are used, the transmitter needs to have information on SINRs of all possible combinations of spatial streams. The amount of this information for feedback is substantially reduced by SINR decomposition in (5; the rate drops from B (2 M 1 to B M bits where B is the number of bits to represent a (real scalar. Compared to high feedback rates for OptSch and AdpSch, M-FixSch and Any-FixSch have the same, low feedback rate; the codeword index is sent back by log(v M bits. Additional B bits are for the SINR. If Any-FixSch is used in a combination with either OptSch or AdpSch, the frequency of running a high feedback rate scheduler determines the feedback rate of Any-FixSch. In Table I, it is basically assumed that the frequency is small enough to be ignored for the overall feedback rate. IV. SIMULATION RESULTS We carry out Monte-Carlo simulations on the sum throughput. A Rayleigh fading channel is assumed. The scalar feed-

Throughput Gain over OBLF with M FixSch 3 2.5 2 1.5 1 0.5 M=8, Total SNR=10 db V=64, OptSch V=64, AdpSch Aggr V=64, AdpSch V=64, Any FixSch V=4, OptSch V=4, AdpSch Aggr V=4, AdpSch V=4, Any FixSch 0 0 20 40 60 80 100 120 140 Number of Receivers Fig. 1. Throughput gain over OBLF with respect to the number of receivers where M =8and SNR=10 db. Throughput Gain over OBLF with M FixSch 5 4.5 4 3.5 3 2.5 2 1.5 1 V=64, OptSch V=64, AdpSch Aggr V=64, AdpSch V=64, Any FixSch V=4, OptSch V=4, AdpSch Aggr V=4, AdpSch V=4, Any FixSch M=8, Total SNR=10 db 0.5 0 5 10 15 20 25 30 SNR (db Fig. 2. Throughput gain over OBLF with respect to SNR where M =8 and Z =8. back corresponding to γ n is not quantized. Any-FixSch is assumed to obtain E [ K ] by infrequently running OptSch. Throughput gain is computed by (R new R/R where R new is the throughput of a new scheme and R is that of OBLF with M-FixSch. Figs. 1 and 2 show the sum throughput gain by using spatial stream schedulers. One major trend in Fig. 1 is that the gain exponentially diminishes as the number of receivers increases. With a large number of receivers, the streams chosen by OBLF with FixSch do not interfere with each other much, which results in K M. Interestingly, the gain increases as feedback rates increase from V =4to V =64; recall that the interference mitigation by spatial stream schedulers are from the subtraction of highly interfering spatial streams, which is done independent of a feedback rate. The logarithmic gain observed in Fig. 2 over SNR shows that the system comes out of the interference-limited regime thanks to the adaptive schedulers. When the number of receivers is small, the gain by AdpSch depicted in Fig. 2 is close to the exhaustive search. This gap becomes larger as the number of receivers increases as shown in Fig. 1, which is the result of the design of AdpSch; the diversity of spatial streams makes the ignored cases in two steps of AdpSch more probable, thus worsening the suboptimality. In designing AdpSch, it has been conjectured that the evolution of M stream-selection beyond jk would be rarely helpful. To verify this, the performance of AdpSch with the evolution to all spatial streams in the M stream-combination step is depicted by the name of AdpSch-Aggr. It is noticeable that the conjecture is indeed true in all cases in Figs. 1 and 2. While AdpSch requires a high feedback rate as discussed in Section III-D, Any-FixSch, which requires the same feedback rate as M-FixSch performs close to OptSch and AdpSch in many cases, which are observed in Fig. 2. Moreover, in Fig. 1, it is seen that its performance even exceeds that of OBLF- AdpSch when multiuser diversity is largely induced and the feedback rate is high (V =64. This is, however, more due to the worsened sub-optimality of AdpSch with large multiuser diversity. Nevertheless, in a practical system, it is expected to have feedback from only a subset of receivers, thus a few receivers regime being more important. V. CONCLUSION In this paper, we observe the significance of the interference in a limited-feedback based MIMO broadcast channel, proposing one way of countermeasure, adaptive scheduling. It chooses only a subset of all streams that it may schedule together, considering mutual interference caused by limited feedback. AdpSch exploits a large set of information to make instantaneous and adaptive decisions at the transmitter and Any-FixSch is designed to avoid such decision makings. Both show a linear complexity to the number of receivers. As future work, the tradeoff between the overall feedback rate and frequency of running either OptSch or AdpSch for estimating E [ K ] will be studied. REFERENCES [1] K. Huang et al., Performance of orthogonal beamforming for SDMA with limited feedback, IEEE Trans. Veh. Tech., vol. 58, no. 1, pp. 152 164, Jan. 2009. [2] D. Gesbert et al., From single user to multiuser communications: Shifting the MIMO paradigm, IEEE Sig. Proc. Mag., vol. 24, no. 5, pp. 36 46, Oct. 2007. [3] Samsung Electronics, Downlink MIMO for EUTRA, 3GPP TSG RAN WG1 44/R1-060335. [4] C. B. Chae et al., Coordinated beamforming with limited feedback in the MIMO broadcast channel, IEEE Jr. Sel. Ar. Comm., vol. 26, no. 8, pp. 1505 1515, Oct. 2008. [5] T. Yoo et al., Multi-antenna broadcast channels with limited feedback and user selection, IEEE Jr. Sel. Ar. in Comm., vol. 25, no. 7, pp. 1478 1491, Sept. 2007. [6] A. Bayesteh and A. K. Khandani, On the user selection for MIMO broadcast channels, IEEE Trans. Info. Theory, vol. 54, no. 3, pp. 1086 1107, Mar. 2008. [7] A. Kamerman and L. Monteban, WaveLAN-II: A high-performance wireless LAN for the unlicensed band, Bell Labs Technical Journal, pp. 118 133, Summer 1997.