Milcom 2015 Track 1 - Waveforms and Signal Processing

Spectrum Efficient Communications with Multiuser MIMO, Multiuser Detection and Interference Alignment Satya Ponnaluri, Sohraab Soltani, Yi Shi and Yalin Sagduyu Intelligent Automation Inc. Rockville, MD USA {sponnaluri, ssoltani, yshi, ysagduyu}@i-a-i.com

Abstract—This paper presents SECANT (Spectrum Efficient Communications and Advanced Networking Technology) as a novel multiuser communication protocol that integrates advanced physical (PHY) layer techniques for adaptive multiple access and interference management. SECANT supports concurrent multiple access (MA) transmissions to common receivers and parallel peer-to-peer (P2P) transmissions in the network. SECANT combines multiuser multiple-input multiple output (MU-MIMO), multiuser detection (MUD), and interference alignment (IA) methods at the PHY layer such that spectrum efficiency can be improved in both interference and signal spaces. An adaptive medium access control (MAC) mechanism is developed to enable spectrum efficiency gains at the PHY layer by managing concurrent transmissions and interference levels. Numerical results indicate significant spectrum efficiency gains (>14 times for fully loaded, and >56 times for 25% loading) that can be achieved over conventional single antenna systems with static MAC protocols.

domain and cannot enable spectrum efficiency gains of advanced PHY layer techniques relying on concurrent transmissions in time, space and frequency. Advanced communication schemes such as interference alignment (IA) [1], [2] and multiuser detection (MUD) [3] have been designed to approach theoretical capacity limits separately on interference and multiple access channels. However, these techniques are usually designed for a small number of nodes with certain topologies without accounting for practical implementation constraints. The theoretical analysis and realistic protocol design on how to apply both IA and MUD for a wireless network are still largely missing. In this paper, we present a protocol called SECANT that combines dynamic interference management, including interference avoidance/cancellation and IA, with spectrumaware MUD and multiuser MIMO at PHY layer to increase spectrum efficiency. Instead of using traditional MAC, a new adaptive MAC scheme is proposed in SECANT to enable PHY layer gains by allowing concurrent transmissions in a collision domain, and dynamically controlling interference and channel access with local spectrum information.

Keywords— Spectrum efficiency; MIMO; multiuser MIMO; interference alignment; multiuser detection; rate; overhead.

I.

INTRODUCTION

Wireless communication spectrum is a scarce resource that must be shared by multiple users across time, space and frequency dimensions. Thus, available spectrum resources impose fundamental limits on wireless system performance. There have been numerous efforts on increasing spectrum efficiency at all layers but a common framework with significant efficiency gains is still lacking and needs to address two main challenges: 1.

2.

We pursue a PHY layer study to evaluate the spectrum efficiency gain of three design steps: (i) multiuser multipleinput multiple-output (MU-MIMO) [4], (ii) overloaded MUMIMO with MUD (MU-MIMO-MUD), and (iii) overloaded MU-MIMO with MUD and IA (MU-MIMO-MUD-IA), over the varying network density where each node has four antennas for practical purposes (SECANT design supports an arbitrary number of antennas and spectrum efficiency will further increase with additional antenna elements). We consider concurrent multiple access (MA) and peer-to-peer (P2P) transmissions over a Wi-Fi density network with soldier walking mobility model. We use single-input single-output (SISO) TDMA quadrature phase-shift keying (QPSK) as a baseline for comparison.

Advanced PHY layer techniques are typically developed for static/small network topologies (e.g., point-to-point, MAC or interference channel) and it is not clear how to apply them to general configurations with local information and distributed control. Legacy MAC schemes, e.g., time-division multipleaccess (TDMA) and carrier-sense multiple-access (CSMA), support one transmission in a single collision

Distribution Statement A (Approved for Public Release, Distribution Unlimited). The views, opinions, and/or findings contained in this article are those of the authors and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government. This material is based upon work supported by Defense Advanced Research Projects Agency under Contract W31P4Q-13-C-0159.

978-1-5090-0073-9/15/$31.00 ©2015 IEEE

1

By accounting for protocol overhead, our results show that MU-MIMO-MUD-IA achieves the spectrum efficiency improvement reaching from × 14 to × 20 with the coverage area varying from 100 × 100 to 400 × 400 and with node density of 0.3 per square meters. SECANT MAC

1494

Milcom 2015 Track 1 - Waveforms and Signal Processing

adapts to dynamic traffic compared to fixed traffic load balancing via TDMA. In particular, the improvement factor achieved by SECANT under dynamic traffic conditions (25% loading) is from ×56 to ×80.

4.

The rest of the paper is organized as follows. Section II provides the multi-layer system architecture for SECANT. MU-MIMO, overloaded MU-MIMO and IA are combined gradually in Sections III, IV, and V, respectively. Section VI studies constrained alphabet signaling effect. SECANT MAC is described in Section VII along with the overhead analysis. Finally, Section VIII concludes the paper. II.

interference from P2P transmissions is aligned within a single dimension. MUD: Each HN uses MUD to decode the information simultaneously transmitted by potentially multiple transmitters.

SYSTEM ARCHITECTURE FOR SECANT

The network is dynamically divided into clusters (with a distributed clustering algorithm for ad hoc networks such as the one in [5]), each with a head node (HN) responsible for coordinating local transmissions and managing interference. SECANT supports multiple simultaneous MA transmissions to HNs in parallel with multiple P2P transmissions. Fig. 1 shows the SECANT network reference model, which is a logical representation of the network architecture.

Fig. 2: The SECANT protocol.

SECANT system model allows nodes in a network, each with antennas, and supports simultaneous transmissions up to transmitters to each HN, along with up to simultaneous P2P direct transmissions. The P2P direct transmissions do not interfere with each other, but may interfere with the transmissions received at a HN. The × 1 received signal vector at a HN in a given cluster is

,

∈

,

!∈"

!

!,

!,

! !

# ,

where , is the × channel matrix between the $-th node and 0-th node (i.e., a HN); %, is the received power from the &-th node; % is the power control parameter at the &-th node; ⊂ (1, 2, … , *, | | , , , is the set of users intending to communicate with the HN; " ⊂ (1,2, … , *, with " ∩ . and |"| , is the set of users intending to communicate with other users in the network; is the × 1 beam-forming vector used by the $-th user; ∈ / is the symbol transmitted by the $ -th user; and # is the complex additive Gaussian noise vector with zero mean and covariance matrix 0 1 2. Similarly, for the P2P communications, the received signal 5 ⊂ (1,2, … , *, " 5 ∩ " ., " 5∩ at the node 34 ∈ " 5 5 ) is ., " ↔ " (there is a one-to-one map between " and " 4 !

Fig. 1: SECANT-MA and SECANT-P2P.

III.

The SECANT protocol is designed based on this network architecture with the following characteristics: 1.

2.

3.

!∈"

4 ! !,!

4 ! ! !,!

#!4 .

SECANT PHY: MU-MIMO

Suppose ∈ transmitters in a given cluster communicate with a HN simultaneously. Then, the × 1 received signal vector at the HN is

Coordination: Nodes are frame-synchronized, and at the beginning of every frame (shown in Fig. 2), nodes requiring channel access request the current HN for channel resources. Joint MA & P2P communication: Based on the number of antennas available, each HN provides channel access to nodes that require the MA and P2P communications. Beamforming and IA: Each HN determines the beamforming vectors for transmitters to ensure that the



∈

,

,

#

7

# ,

where in the second representation, is a matrix of size × whose columns are 8 9 , and 7 is a column , , ∈ vector containing the symbols transmitted by all the users.

2

1495

Milcom 2015 Track 1 - Waveforms and Signal Processing

For linear receivers with single-user decoders to decode all the transmissions, the underlying matrix must be invertible, and the following conditions must be satisfied: 1. there should be at most transmitters, otherwise there will be fewer dimensions to support all the users, and 2. the channel matrices and beamforming vectors should be such that the resultant vectors from all the users should be linearly independent. In general, if the channel matrices are drawn from a continuous distribution such as zero-mean complex Gaussian (the amplitude is Rayleigh distributed), then the channels themselves are invertible, and therefore :

;< , =

/ ?@

;< , =

@?A

∈( ,<,…,B;<*

A simple mechanism for the receiver design is to assign one receive antenna to each transmitter. The beamforming are chosen such that vectors Y <, 1, b, cZ 8 < < , 1 1 , b b , c c 9 are ideally, mutually orthogonal, ;< or minimally, linearly independent. For example, % % a% , where R% is i-th column of identity matrix. Then we have, < < <

DdeEf

,

N

I JD K

where DY Z ≜ log 1 Y1

?@

<

_`

,

;< , =

@?

1 LM O&PQ

+

c c c

#.

I gDY

%,

1

@U%, @ Zh O&PQ RS TUVWWR$ XQR,

where %, is the received power from node & to the HN, and the channel gain between the node and the HN is U%, . A similar rate expression follows for P2P communications. 60 SECANT SISO TDMA 50 Sum rate in bits/v.c.u

B;<

b b b

+

The average sum-rate as a function of V using MU-MIMO and TDMA (wherein each node is provided with equal timeslot for transmission) is shown in the Fig. 4. The expression for the achievable rate using TDMA is

where 8= 9 form a set of orthogonal (unit-norm) vectors, is sufficient to ensure that all the users are separable (there exists a one-to-one map between the sets and (0,1, … , C 1* so that for every $ ∈ (0,1, … , C 1*, there exists a unique $ ∈ ). In general, there exists no closed-form solution to find the beamforming vectors that maximize the channel capacity; nevertheless, iterative algorithms have been proposed to determine the appropriate set of beamforming vectors under SINR constraints [6], [7]. The overall average sum-rate under such conditions, assuming Gaussian codebooks (not necessarily independent), is DEF;EGEH

1 1 1

RS TUVWWR$ XQR,

40

30

20

10

Z, and the expected value is taken

over all the channel matrices. The choice of 8= 9 is arbitrary, therefore we let = a , where a is the $-th column of the identity matrix of size .For example, consider a network of four nodes distributed randomly in a square of side V, and communicating with a HN located at the center of the square (shown in Fig. 3). All the nodes in the network, including the HN, have 4 antennas. MU-MIMO uses the dimensions from multiple antennas to support multiple simultaneous transmissions to the A. For example, in a 4x4 MIMO setting, potentially 4 users can communicate simultaneously with a single HN. The mutual interference is avoided by careful selection of beamforming vectors for each transmitter.

0 0

100

200 300 400 side of the square area in meters

500

Fig. 4: Average sum-rate in bits/v.c.u of MU-MIMO and TDMA with Gaussian signaling.

We assume the following transmitter, channel and receiver conditions: each node transmits with an average power of 10 dBm; the bandwidth is set to 1.28 MHz; the individual links of MIMO channels are Rayleigh flat-fading; the operating frequency is around 2.4 GHz; the path loss exponent is 3.5 (i.e., path loss in dB 40 dB 10 log i, where i is distance in meters); and the receiver noise figure is set to 10 dB. Both approaches assume Gaussian signaling, as opposed to constrained alphabet such as QPSK, 16-QAM or 64-QAM. We observe that the X-factor performance improvement of MU-MIMO over TDMA is 4 for different values of a. IV.

SECANT PHY: OVERLOADED MU-MIMO WITH MUD (MU-MIMO-MUD)

When the number of users exceeds the available dimensions (i.e., j , see Fig. 5), we refer to the scenario as overloaded – a term that is used mostly in the context of code-division multiple-access (CDMA). Overloaded MU-MIMO affects the transmitter and receiver design in two ways: (1) The beam-forming vectors

Fig. 3: SECANT MU-MIMO. 3

1496

Milcom 2015 Track 1 - Waveforms and Signal Processing

8k 9 cannot be mutually orthogonal and (2) the matrix longer invertible. The resulting received vector is

where

is an

×

# ,

matrix whose columns are :

,

/

@? k A . The channel now resembles an overloaded ?@ CDMA channel (cf. [8], [9]). Therefore the beamforming vector design is identical to the spreading sequence design in overloaded CDMA, and the sequences that maximize sumcapacity were described in [9], and is summarized as follows: users whose power is large relative to the other users are assigned orthogonal sequences, and the remaining users are assigned so-called generalized Welch-bound equality (WBE) sequences. In order to limit the receiver complexity, we adopt a different approach to the beamforming vector design. Instead of assigning WBE sequences, we instead divide the users into groups, each group consisting of / users (for simplicity, let be a multiple of ). A similar approach was considered for coding in multiple-access channels in [10].

2.

3.

The users communicating with the HN align their signals within C 1 degrees of freedom (DoF) available, and provide the remaining DoF for IA. Therefore the number of users communicating with the base-station decreases. (In general, the number of DoFs provided for IA can be increased from a practical standpoint for easier and better alignment.) The P2P nodes are carefully selected so that the transmitter-receiver pairs are physically close to each other and sufficiently far from the other transmitterreceiver pairs. Any residual interference from other transmitters is treated as noise at each receiver. The beamforming vectors for the P2P transmitters are such that the P2P transmissions arrive in a dimension orthogonal to the intended signals received by the HN. 8 7.5 Ratio of sum rate

;< , k

7

1.

is no

7 6.5 6 5.5 5 4.5 0

100 200 300 400 side of the square area in meters

500

Fig. 6: Performance gain of overloaded MU-MIMO relative to TDMA.

Continuing with the example described earlier with eight nodes communicating with a HN, in Fig. 7 we randomly distribute nodes within a network with a density of l nodes per square meter. The HN selects pairs of nodes for P2P communication based on the following conditions: i!,! ̅ n io%p and i! q ,!4 j iors , i!,! ̅q j iors .

Fig. 5: Network topology with overloaded MU-MIMO communication.

The performance improvement, relative to TDMA, for an eight-user network with four antennas is shown in Fig. 6. Unlike MU-MIMO, the X-factor improves gradually to a factor of 8. This gradual increase may be attributed to the fact that the power disparity between two users increases as the area of the network grows, and the difference between Gaussian multiple-access sum-capacity and TDMA is more when there is a power disparity between users. Both approaches assume Gaussian signaling (and independent codebooks for users within a group), as opposed to constrained alphabet such as QPSK, 16-QAM or 64-QAM. V.

SECANT PHY: OVERLOADED MU-MIMO WITH MUD AND IA (MU-MIMO-MUD-IA)

IA refers to the idea of aligning multiple interfering transmissions at each receiver by exploiting the knowledge of the channel between the various users [1]. There are, however, several practical challenges with IA [11]-[12], such as the complexity associated with knowing all the channel matrices at all the users. In order to tradeoff between complexity and performance gain promised by IA (at least asymptotically, at large SNR), we simplified the application of IA:

Fig. 7: Overloaded MU-MIMO with IA.

4

1497

Milcom 2015 Track 1 - Waveforms and Signal Processing

The performance gain using a combination of overloaded MUMIMO and IA is shown Fig. 8. We set io%p V/4 and iors V/√8.

We assume that the largest alphabet available to SECANT transmitters is 256-QAM. The primary difference we observe relative to unconstrained Gaussian signaling is that the Xfactor is better for smaller coverage areas also. We observe that in area ranging from 100 × 100 to 200 × 200 , SECANT is able to use higher order modulation scheme such as 64-QAM and 256-QAM due to shorter range to the HN, and thereby improve spectrum efficiency relative to the baseline. Beyond 150 , SECANT and TDMA may use QPSK or lower order, thereby retaining the same performance as with Gaussian signaling (equivalent to less than 2 bps/Hz).

Ratio of sum rate

25 20 15 10

VII.

We design new MAC layer protocol using control information exchange and negotiation among potential network access and P2P transmitters, and the HN. According to the SECANT network topology, after the traffic flows are configured by the routing layer at each node in the system, then a sequence of events occurs among wireless end users and the HN in the system so the SECANT MAC has complete information to select and allow concurrent transmissions in the network. This sequence of events is as follows: 1. Each transmitter sends ‘SYN’ message to the HN. 2. The HN selects SECANT-P2P transmitter and receiver pairs and SECANT-HN transmitters according to the MU-MIMO HN optimization. 3. The HN sends Clear-to-Send (CTS) messages to the selected SECANT-P2P transmitters. 4. The HN sends Request-to-Send (RTS) to the SECANTHN transmitters. We simulate the SECANT network topology in Fig. 7 and incorporate the SECANT network management in conjunction with the PHY layer overloaded MU-MIMO with IA to evaluate the impact of the overhead on the overall spectrum efficiency improvement. Based on this topology, potential transmitters send SYN message to the HN. The HN then selects a set of transmitters. Since the SECANT HN is aware of PHY MIMO with IA and MUD capabilities, the HN selects a set of coding vectors to solve the optimization problem that satisfies the PHY MUD and IA requirements.

5 0 0

100 200 300 side of the square a, in meters

400

Fig. 8: Performance gain of overloaded MU-MIMO with IA, relative to TDMA.

VI.

CONSTRAINED ALPHABET SIGNALING

Thus far, we focused on Gaussian signaling that lends to simple expressions of capacity. However, in practice, there are limits to the modulation scheme (e.g., QPSK, 16-QAM, or 256-QAM). There are two methods to find the expression for capacity for constrained alphabet signaling as a function of SNR: 1. deriving the capacity expression based on the modified channel with discrete inputs and continuous outputs, 2. using the Gaussian signaling expression for capacity and limiting the achievable data rate to the corresponding value attained by finite alphabet. We compare the capacities using both approaches in Fig. 9 for QPSK signaling. The gap between the two curves is very small suggesting that using the Gaussian signaling capacity with appropriate limits is a reasonable approximation, at least for QPSK. Based on the above observation, the performance of SECANT (i.e., overloaded MU-MIMO with IA) relative to TDMA SISO QPSK is shown in Fig. 10 without overhead.

After finding the optimal solution (or a suboptimal one through fast heuristics), the HN will inform each SECANTP2P transmitter through a set of CTS messages to code their transmission with a particular beamforming vector so their interference with the HN could be aligned. This would satisfy the IA requirement and allow simultaneous P2P transmissions. The HN also informs the receivers for each P2P transmissions using CTR message about their intended transmitter. The next step is for the HN to select the set of transmitters in the SECANT-HN network (the ones that intend to transmit their data packets to the HN). The HN utilizes the MUD techniques to separate each transmission. Consequently, the HN selects its transmitters and sends RTS messages to them. This information includes the beamforming vectors and the index of the code used for the multiple access techniques. Table I shows the overhead packet for the network simulation.

2

Capacity in bits/channel use

1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 QPSK approximation using Gaussian

0.2 0 -20

-15

-10

-5

0 SNR in dB

5

10

15

SECANT MAC AND OVERHEAD

20

Fig. 9: Comparison between QPSK channel capacity and approximation based on Gaussian signaling. 5

1498

Milcom 2015 Track 1 - Waveforms and Signal Processing

TABLE I.

Message Length

THE OVERHEAD PACKET LENGHTS (IN BITS)

SYN 16

Pilot Symbols 128

CTS 128

CTR 16

In conclusion, SECANT introduces a paradigm shift in wireless communication protocol design to achieve spectrum utilization beyond the legacy “single-link single-user” PHYMAC operational rule. SECANT presents adaptive MAC design that supports PHY layer spectrum efficiency gains to improve system throughput on limited frequency resources. We evaluated the spectrum efficiency gain of three design steps: (i) MU-MIMO, (ii) MU-MIMO-MUD and (iii) MUMIMO-MUD-IA over the SECANT network topology with varying network density where each node has four antennas. We also added the capability to increase the modulation rate supported by the combined PHY layer solution. We considered concurrent HN and P2P transmissions over a Wi-Fi density network with soldier walking mobility model. The numerical results demonstrate major spectrum efficiency improvements over the baseline SISO QPSK TDMA.

RTS 128

In this simulation, we set the sensing communication rate Suvws at 20Mbps, and the data rate duration varies from 50ms to 150ms. Fig. 10 shows the impact of the information exchange on the spectrum efficiency over the SISO TDMA. 24

Ratio of sum rate

22

w/o overhead w/ overhead, frame duration = 150 ms frame duration = 75 ms frame duration = 50 ms

20 18

REFERENCES [1]

S.A. Jafar, “IA: A new look at signal dimensions in a communication Network,” Foundations and Trends in Communications and Information Theory, vol. 7, no. 1, pp. 1–134, 2010. [2] V. R. Cadambe, S. A. Jafar, “IA and Degrees of Freedom of the K -User Interference Channel,” IEEE Transactions on Information Theory, vol. 54, no. 8, pp. 3425-3441, Aug. 2008. [3] S. Verdu, Multiuser Detection, Cambridge, UK, Cambridge University Press, 1998. [4] M. Jiang, L. Hanzo, “Multiuser MIMO-OFDM for Next-Generation Wireless Systems,” Proceedings of the IEEE, vol.95, no.7, pp.1430,1469, Jul. 2007. [5] S. Basagni, “Distributed clustering for ad hoc networks.” Parallel Architectures, Algorithms, and Networks,” IEEE I-SPAN, 1999. [6] G. Zheng, K.-K. Wong, and T.-S. Ng, “Transmit Beamforming Optimization for Multiuser MIMO Uplink Using Successive Interference Cancellation,” IEEE TENCON 2006, pp. 1-4, 14-17, Nov. 2006. [7] M. Schubert, and H. Boche, “Iterative multiuser uplink and downlink beamforming under SINR constraints,” IEEE Trans. On Sig. Proc., vol. 53, no. 7, pp. 2324-2334, Jul. 2005. [8] A. Kapur and M. K. Varanasi, “Multiuser detection for overloaded CDMA systems,” IEEE Transactions on Information Theory, vol. 49, no. 7, pp. 1728-1742, Jul. 2003. [9] P. Viswanath, V. Anantharam,, “Optimal sequences and sum capacity of synchronous CDMA systems,” IEEE Transactions on Information Theory, vol. 45, no. 6, pp. 1984-1991, Sep. 1999. [10] J. L. Massey, “Coding for multiple access communications.” ITG FACHBERICHT (1994): 11-11. [11] O. El Ayach, S.W. Peters, and R. W. Heath, “The practical challenges of IA,” IEEE Wireless Communications, Feb. 2013. [12] H. Zeng, Y. Shi, Y.T. Hou, W. Lou, S. Kompella, and S.F. Midkiff, “On interference alignment for multi-hop MIMO networks,” Proc. IEEE INFOCOM, 2013.

16 14

12 0

100 200 300 side of the square a, in meters

400

Fig. 10. The impact of overhead in spectrum efficiency gain.

VIII. SUMMARY OF RESULTS AND CONCLUSIONS The following spectrum efficiency improvements are obtained over the baseline (SISO TDMA QPSK) via theoretical analysis and simulation for WiFi density wireless environment with soldier walking mobility: 1. MU-MIMO improves spectrum efficiency by a factor of four (× 4) relative to SISO TDMA QPSK. 2. MU-MIMO-MUD improves spectrum efficiency by a factor of eight (× 8). 3. MU-MIMO-MUD-IA achieves spectrum efficiency improvement reaching from × 14 to × 20 (for coverage area varying between 100 × 100 to 400 × 400 , and with node density of 0.3 per square meters). 4. The improvement factor ranges from × 56 to × 80 under dynamic traffic conditions (25% loading).

6

1499