1 On the Capacity of Distributed Antenna Systems Lin Dai City University of Hong Kong Jan. 19, 2013 JWITC 2013 2 Cellular Networks (1) Base Station (BS) Growing demand for high data rate Multiple antennas at the BS side Jan. 19, 2013 JWITC 2013 3 Cellular Networks (2) • Implementation cost Sum rate Jan. 19, 2013 Co-located BS antennas • Distributed BS antennas Lower Higher? JWITC 2013 4 A little bit of History of Distributed Antenna Systems (DAS) • Originally proposed to cover the dead spots for indoor wireless communication systems [Saleh&etc’1987]. • Implemented in cellular systems to improve cell coverage. • Recently included into the 4G LTE standard. • Multiple-input-multiple-output (MIMO) theory has motivated a series of information-theoretic studies on DAS. Jan. 19, 2013 JWITC 2013 5 Single-user SIMO Channel • Single user with a single antenna. • L>1 BS antennas. • Uplink (user->BS). • Received signal: • Channel gain : (a) Co-located BS Antennas y gs z g=γ h (b) Distributed BS Antennas s CN (0, P): Transmitted signal z C L1 : Gaussian noise zi CN (0, N0 ), i 1,..., L. γ C L1 : Large-scale fading i Log CN (di /2 , i2 ), i 1,..., L. h C L1 : Small-scale fading hi Jan. 19, 2013 CN (0,1), i 1,..., L. JWITC 2013 6 Single-user SIMO Channel • Single user with a single antenna. • L>1 BS antennas. • Uplink (user->BS). (a) Co-located BS Antennas (b) Distributed BS Antennas • Ergodic capacity without channel state information at the transmitter side (CSIT): P g Co Eh log 2 1 N0 • Ergodic capacity with CSIT: Jan. 19, 2013 2 2 P g Cw max Eh log 2 1 P: Eh [ P ] P N0 Function of large-scale fading vector JWITC 2013 7 Co-located Antennas versus Distributed Antennas • With co-located BS antennas: Ergodic Capacity without CSIT: C Eh log2 1 h C o 1 L 2 is the average received SNR: P γ / N0 2 • With distributed BS antennas: Distinct large-scale fading gains to different BS antennas. Ergodic Capacity without CSIT: CoD Eh log2 1 g Jan. 19, 2013 2 Normalized channel gain: g = β h β γ γ JWITC 2013 8 Capacity of DAS • For given large-scale fading vector : [Heliot&etc’11]: Ergodic capacity without CSIT o A single user equipped with N co-located antennas. o BS antennas are grouped into L clusters. Each cluster has M co-located antennas. o Asymptotic result as M and N go to infinity and M/N is fixed. [Aktas&etc’06]: Uplink ergodic sum capacity without CSIT o K users, each equipped with Nbk co-located antennas. o BS antennas are grouped into L clusters. Each cluster has Nll co-located antennas. o Asymptotic result as N goes to infinity and bk and ll are fixed. Jan. 19, 2013 • Implicit function of (need to solve fixed-point equations) • Computational complexity increases with L and K. JWITC 2013 9 Capacity of DAS • With random large-scale fading vector : Average ergodic capacity (i.e., averaged over ) Single user Without CSIT [Roh&Paulraj’02], [Zhang&Dai’04]: The user has identical access distances to all the BS antennas. i Log CN (1, 2 ), i 1,..., L. [Zhuang&Dai’03]: BS antennas are uniformly distributed over a /2 circular area and the user is located at the center. i i , i 1,..., L. [Choi&Andrews’07], [Wang&etc’08], [Feng&etc’09], [Lee&ect’12]: BS antennas are regularly placed in a circular cell and the user has a random location. i Log CN (di /2 , i2 ), i 1,..., L. High computational complexity! Jan. 19, 2013 JWITC 2013 10 Questions to be Answered • How to characterize the sum capacity of DAS when there are a large number of BS antennas and users? Large-system analysis using random matrix theory. Bounds are desirable. • How to conduct a fair comparison with the co-located case? K randomly distributed users with a fixed total transmission power. Decouple the comparison into two parts: 1) capacity comparison and 2) transmission power comparison for given average received SNR. • What is the effect of CSIT on the comparison result? Jan. 19, 2013 JWITC 2013 11 Part I. System Model and Preliminary Analysis [1] L. Dai, “A Comparative Study on Uplink Sum Capacity with Co-located and Distributed Antennas,” IEEE J. Sel. Areas Commun., 2011. Jan. 19, 2013 JWITC 2013 12 Assumptions • K users uniformly distributed within a circular cell. Each has a single antenna. • L BS antennas. • Uplink (user->BS). R R Random BS antenna layout! *: user o: BS antenna (a) Co-located Antennas (CA) Jan. 19, 2013 (b) Distributed Antennas (DA) JWITC 2013 13 Uplink Ergodic Sum Capacity • Received signal: K y g k sk z k 1 • Uplink power control: Pk γ k 2 P0 , k 1,..., K . Ergodic capacity without CSIT 1 K † Csum _ o =E H log 2 det I L Pk g k g k N 0 k 1 P0 K † =E H log 2 det I L g g k k N 0 k 1 Jan. 19, 2013 sk CN (0, Pk ) : Transmitted signal : Gaussian noise z C L1 zi CN (0, N0 ), i 1,..., L. g k = γ k hk γ k C L1 hk C L1 : Channel gain : Large-scale fading i,k di,k /2 , i 1,..., L. : Small-scale fading hi,k CN (0,1), i 1,..., L. Ergodic capacity with CSIT 1 Csum _ w = max E H log 2 det I L Pk : E H [ Pk ] Pk N0 k 1,..., K † P g g k k k k 1 1 = max E H log 2 det I L Pk : E H [ Pk ] P0 N0 k 1,..., K † P g g k k k k 1 K K JWITC 2013 14 More about Normalized Channel Gain hk C L1: Small-scale fading γ β k k C L1 : Normalized γk Large-scale fading βk 1. • Normalized channel gain vector: g k = βk hk With CA: gk With DA: 1 1L1. L βk 2 The channel becomes deterministic with a large number of BS antennas L! 1 if L . bi,k b j ,k , i j. o With a large L, it is very likely that user k is close to some BS antenna lk* : gk Jan. 19, 2013 2 | hl* ,k | 2 k βk el* k 1 if l lk* el 0 otherwise Channel fluctuations are preserved even with a large L! JWITC 2013 15 More about Normalized Channel Gain • Theorem 1. For n=1,2,…, min Ehk g k β k : β k 1 2n max Ehk gk βk : βk 1 (n L 1)! , which is achieved when β k 1 1L1. Ln ( L 1)! L 2n n!, which is achieved when βk el* . k 1 1L1 , L maximized when βk el* . Channel fluctuations are minimized when β k k Channel fluctuations are undesirable when CSIT is absent, desirable when CSIT is available. Jan. 19, 2013 JWITC 2013 16 Single-user Capacity (1) • Without CSIT 2 CA CkC_ o 1.8 L=10 L=50 L=10 L=50 with CSIT Ck _ o / CAWGN 1 -- Fading always hurts if CSIT is absent! CkC_ w 1.6 Ck log 2 (1 0 ) DA with β k el * CkC_ o / CAWGN quickly approaches 1 as L grows. k CkD_*o CkD_*w 1.4 C D* Ck _ o Ck _ o 1.2 • With CSIT (when 0 is small) 1 Ck _ w / CAWGN 1 at low 0 --“Exploit” fading 0.8 without CSIT 0.6 -10 -5 0 N0 (dB) (average received SNR) 0 P00 /(dB) Jan. 19, 2013 5 10 D* C Ck _ w Ck _ w JWITC 2013 17 Single-user Capacity (2) 1.1 with CSIT DA with βk Single-user Capacity Ck (bit/s/Hz) CAWGN log 2 (1 0 ) • Without CSIT el* A higher capacity is achieved in the CA case thanks to better diversity gains. k 1 0.9 • With CSIT without CSIT 0.8 CA x 0.7 DA with βk 10 DA 20 25 k Analytical (Eq. (28)) Simulation Analytical (Eq. (29-30)) Simulation Analytical (Eq. (21)) Simulation Analytical (Eq. (22-23)) Simulation 15 A higher capacity is achieved in the DA case thanks to better waterfilling gains. el* 30 35 40 45 50 Number of BS Antennas L The average received SNR 0 0dB. Jan. 19, 2013 JWITC 2013 18 Part II. Uplink Ergodic Sum Capacity Jan. 19, 2013 JWITC 2013 19 Uplink Ergodic Sum Capacity without CSIT • Sum capacity without CSIT: P Csum _ o =E H log 2 det I L 0 N0 † g g E log det I GG k H 2 L 0 k 1 K † k G=B H • Sum capacity per antenna (with K>L): CL _ o 1 E H log 2 det I L 0GG † L 1 L E H log 2 1 0 ll L l 1 B [β1 ,..., β K ] H [h1 ,..., hK ] where {ll } denotes the eigenvalues of GG† . Jan. 19, 2013 JWITC 2013 20 More about Normalized Channel Gain • Theorem 2. As K , L and K / L , EH [l ] , and EH [l 2 ] 2 2 B , with 0 B . 1 1LK . L B 0 when B [el* ,..., el* ]. B when B 1 With CA: B f l ( x) C 1 1 1LK . L With DA and B [el* ,..., el* ]: 1 (( 1) x)( x ( 1) ) 2 2 x as K , L and K / L . [Marcenko&Pastur’1967] Jan. 19, 2013 K 2 f l ( x) e D* ( x ) K ( x )k k 0 k !(k 1)! as K , L and K / L . JWITC 2013 21 Sum Capacity without CSIT (1) Sum Capacity per Antenna CL_o (bit/s/Hz) 7 6 CLC_ o CLD_*o log 2 (1 0 ) CLC_ o CLD_*o Gap diminishes when is large -- the capacity becomes insensitive to the antenna topology when the number of users is much larger than the number of BS antennas. 5 4 =10 3 =5 2 1 =1 0 0 2 4 6 (average received SNR) 0 0P(dB) 0 / N0 (dB) Jan. 19, 2013 8 10 JWITC 2013 22 Sum Capacity without CSIT (2) 75 Sum Capacity Csum_o (bit/s/Hz) A higher capacity is achieved in the CA case thanks to better diversity gains. CA 70 65 60 DA with B [el* ,..., el* ] 55 1 50 D* Csum _ o serves as an asymptotic lower-bound D to Csum _ o. K CA Analytical (Eq. (31, 34)) 45 Simulation 40 DA Analytical (Eq. (31, 40)) 35 Simulation 30 10 15 20 25 30 35 40 45 50 Number of BS Antennas L The average received SNR 0 0dB. The number of users K=100. Jan. 19, 2013 JWITC 2013 23 Uplink Ergodic Sum Capacity with CSIT • Sum capacity with CSIT: 1 K † Csum _ w = max E H log 2 det I L Pk g k g k Pk : E H [ Pk ] P0 N0 k 1 k 1,..., K With CA: 2 With DA and B [el* ,..., el* ]: 1 The optimal power allocation policy: 1 1 Pk* N 0 g†k I L j k Pj*g j g j 1 gk where is a constant chosen to meet the power constraint E H Pk P0 , k=1,…, K. [Yu&etc’2004] Jan. 19, 2013 Pk Pk γ k K The optimal power allocation policy: 1 1 N Pk* 0 | h * |2 i , ki 0 k ki* arg max | hi ,k |2 kK i k ki* i=1,…,L, where is a constant chosen to meet the sum power constraint EH P KP . K k 1 k 0 JWITC 2013 24 Signal-to-Interference Ratio (SIR) (a) CA (b) DA with B [el1* ,..., elK* ] The received SNR 0 0dB. The number of users K=100. The number of BS antennas L=10. Jan. 19, 2013 JWITC 2013 25 Sum Capacity (1) 2 with CSIT CA 1.8 B [el* ,..., el* ] 1 K L=10 L=50 L=10 L=50 1.6 Csum L log 2 (1 KL 0 ) DA with • Without CSIT C D* Csum _ o Csum _ o C Gap between Csum _ o and D* Csum _ o is enlarged as L grows 1.4 (i.e., due to a decreasing K/L). 1.2 • With CSIT 1 CA 0.8 DA with B [el* ,..., el* ] 1 without CSIT 0.6 -10 -5 K even at high SNR L=10 L=50 L=10 L=50 0 / N0 (dB) (average received SNR) 0 P00(dB) The number of users K=100. Jan. 19, 2013 D* C Csum _ o Csum _ o 5 10 (i.e., thanks to better multiuser diversity gains) JWITC 2013 26 Sum Capacity (2) 90 • Without CSIT DA with B [el* ,..., el* ] 1 A higher capacity is achieved in the CA case thanks to better diversity gains. K 80 Sum Capacity Csum (bit/s/Hz) with CSIT CA 70 60 • With CSIT CA 50 x DA 40 Analytical (Eq. (45-46)) Analytical (Eq. (31, 40)) Simulation without CSIT 30 A higher capacity is achieved in the DA case thanks to better waterfilling gains and multiuser diversity gains. Analytical (Eq. (41-42)) Analytical (Eq. (31, 34)) Simulation 10 15 20 25 30 35 40 45 50 Number of BS Antennas L The average received SNR 0 0dB. The number of users K=100. Jan. 19, 2013 JWITC 2013 27 Part III. Average Transmission Power per User Jan. 19, 2013 JWITC 2013 28 Average Transmission Power per User • Transmission power of user k: Pk P0 γk 2 , k 1,..., K . • Average transmission power per user: K P 1 K 0 f γ 2 ( x)dx P k 0 x k K k 1 P With CA: With DA: o Users are uniformly distributed in the circular cell. BS antennas are co-located at cell center. γk 2 L k 2x f k ( x ) 2 R Jan. 19, 2013 P C 2 P0 R 2 L o Both users and BS antennas are uniformly distributed in the circular cell. What is the distribution of γ k JWITC 2013 2 ? 29 Minimum Access Distance • With DA, each user has different access distances to different BS antennas. Let dk(1) dk(2) dk( L) denote the order statistics obtained by arranging the access distances d1,k,…, dL,k. • γk 2 dl , k L l 1 d k(1) for L>1. • An upper-bound for average transmission power per user with DA: R R y P D DU 0 P P f k ( y ) f d (1) | ( x | y )dxdy 0 0 k k x 2y f k ( y ) 2 R Jan. 19, 2013 f d (1) | ( x | y ) L(1 Fdl ,k |k ( x | y)) L 1 f dl ,k |k ( x | y ) k k 2x 0 x R y R2 f dl ,k |k ( x | y) x2 y 2 R2 2x R y x R y arccos 2 xy R2 JWITC 2013 30 Average Transmission Power per User -1 10 Analytical Average Transmission Power Per User C L1 Eq. (53) Eq. (56) -2 10 C DA: P DU O L /2 (path-loss factor >2) For given received SNR, a lower total transmission power is required in the DA case thanks to the reduction of minimum access distance. DU -3 10 C L L2 -4 Simulation 10 CA with P0C 1 DA with P0D P0C -5 10 O L1 CA: P 10 D 0 DA with P P L 30 40 50 20 C 0 With =4, P 1 5 60 70 80 90 100 C P D if P0D P0C 15 L. Number of BS Antennas L Path-loss factor =4. Jan. 19, 2013 JWITC 2013 31 Sum Capacity without CSIT 350 C • For fixed K and 0 D C (0 0 15 L such that P C P D) CA Analytical (Eq. (31, 34)) Simulation Sum Capacity Csum_o (bit/s/Hz) 300 DA Analytical (Eq. (31, 40)) Simulation 250 200 C C Csum _ o L log2 (1 0 K / L) L 0C K log 2 e C 0 =10dB D Csum _ o O( L) 150 100 50 0C=0dB 0 10 15 20 25 30 35 40 45 Number of BS Antennas L 50 Given the total transmission power, a higher capacity is achieved in the DA case. Gains increase as the number of BS antennas grows. 0D 0C 15 L. The number of users K=100. Jan. 19, 2013 JWITC 2013 32 Conclusions • A comparative study on the uplink ergodic sum capacity with colocated and distributed BS antennas is presented by using largesystem analysis. – A higher sum capacity is achieved in the DA case. Gains increase with the number of BS antennas L. – Gains come from 1) reduced minimum access distance of each user; and 2) enhanced channel fluctuations which enable better multiuser diversity gains and waterfilling gains when CSIT is available. • Implications to cellular systems: – With cell cooperation: capacity gains achieved by a DAS over a cellular system increase with the number of BS antennas per cell thanks to better power efficiency. – Without cell cooperation: lower inter-cell interference with DA? Jan. 19, 2013 JWITC 2013 33 Thank you! Any Questions? Jan. 19, 2013 JWITC 2013