Joint Space-Division and Multiplexing: How to Giuseppe Caire

Communication Theory Workshop
Joint Space-Division and Multiplexing: How to
Achieve Massive MIMO Gains in FDD Systems
Giuseppe Caire
University of Southern California, Viterbi School of Engineering, Los Angeles, CA
Phuket, Thailand, June 23-26, 2013
Channel estimation bottleneck on MU-MIMO
• High-SNR capacity of Nt ×Nr single-user MIMO with coherence block-length
T [Zheng-Tse, 2003]:
C(SNR) = M ∗(1 − M ∗/T ) log SNR + O(1),
M ∗ = min{Nt, Nr , T /2}
• Trivial cooperative bound: for large M = Nt and N = KNr , the coherence
block T is the limiting factor.
• ⇒ Disappointing theoretical performance of “CoMP” (base station
cooperation), in FDD.
18
Inter-cell Cooperation
γ=1, τ=1/32
γ=2, τ=1/32
γ=4, τ=1/32
γ=8, τ=1/32
16
cluster
controller
Cell sum rate (bps/Hz)
14
BS 3
12
10
8
6
4
BS 2
2
BS 1
0
0
5
10
15
20
25
B
Fully cooperative network MIMO/partially coordinated beamforming:
Inter-cell interference mitigation
Larger antenna array gain
H. Huh (USC)
Large System Analysis of Multi-cell MIMO Downlink
1
May 12, 2011
4 / 52
Channel model with antenna correlation
• In FDD, for large macro-cellular base stations, we have to exploit channel
dimensionality reduction while still exploiting the large number of antennas at
the BS.
• Idea: exploit the asymmetric spatial channel correlation at the BS and at the
UTs.
• Isotropic scattering, |u − u0| = λD:
1
∗
0
E [h(u)h (u )] =
2π
Z
π
e−j2πD cos(α)dα = J0(2πD)
−π
• Two users separated by a few meters (say 10 λ) are practically uncorrelated.
2
• In contrast, the base station sees user groups at different AoAs under narrow
AS ∆ ≈ arctan(r/s).
r
s
scattering ring
✓
region containing the BS antennas
• This leads to the Tx antenna correlation model
h = UΛ1/2w,
with
[R]m,p
1
=
2∆
Z
∆
e
R = UΛUH
j kT (α+θ)(um −up )
dα.
−∆
3
Joint Space Division and Multiplexing (JSDM)
• K users selected to form G groups, with ≈ same channel correlation.
H = [H1, . . . , HG], with Hg = Ug Λ1/2
g Wg .
• Two-stage precoding: V = BP.
• B ∈ CM ×bg is a pre-beamforming matrix function of {Ug , Λg } only.
• P ∈ Cbg ×Sg is a precoding matrix that depends on the effective channel.
• The effective channel matrix is given by
 H
H1 B1 HH1 B2 · · ·
 HH2 B1 HH2 B2 · · ·
H
H =
..
...
 ..
HHGB1 HHGB2 · · ·
HH1 BG
HH2 BG 


..  .
HHGBG
4
• Per-Group Processing: If estimation and feedback of the whole H is still too
costly, then each group estimates its own diagonal block Hg = BHg Hg , and
P = diag(P1, · · · , PG).
• This results in
yg =
HHg Bg Pg dg
+
X
HHg Bg0 Pg0 dg0 + zg
g 0 6=g
5
Achieving capacity with reduced CSIT
PG
• Let r = g=1 rg and suppose that the channel covariances of the G groups
are such that U = [U1, · · · , UG] is M × r tall unitary (i.e., r ≤ M and UHU =
Ir ).
• Eigen-beamforming (let bg = rg and Bg = Ug ) achieves exact block
diagonalization.
• The decoupled MU-MIMO channel takes on the form
yg = Hg HPg dg + zg = WgHΛg1/2Pg dg + zg ,
for g = 1, . . . , G,
where Wg is a rg × Kg i.i.d. matrix with elements ∼ CN (0, 1).
Theorem 1. For U tall unitary, JSDM with PGP achieves the same sum
capacity of the corresponding MU-MIMO downlink channel with full CSIT.
6
Block Diagonalization
• For given target numbers of streams per group {Sg } and dimensions {bg }
satisfying Sg ≤ bg ≤ rg , we can find the pre-beamforming matrices Bg such
that:
UHg0 Bg = 0 ∀ g 0 6= g, and rank(UHg Bg ) ≥ Sg
• Necessary condition for exact BD
Span(Bg ) ⊆ Span⊥({Ug0 : g 0 6= g}).
• When Span⊥({Ug0 : g 0 6= g}) has dimension smaller than Sg , the rank
condition on the diagonal blocks cannot be satisfied.
• In this case, Sg should be reduced (reduce the number of served users per
group) or, as an alternative, approximated BD based on selecting rg? < rg
dominant eigenmodes for each group g can be implemented.
7
Performance analysis with regularized ZF
• The transformed channel matrix H has dimension b × S, with blocks Hg of
dimension bg × Sg .
• For simplicity we allocate to all users the same fraction of the total transmit
power, pgk = PS .
• For PGP, the regularized zero forcing (RZF) precoding matrix for group g is
given by
¯ g Hg ,
Pg,rzf = ζ¯g K
where
¯g =
K
and where
ζ¯g2
=
h
Hg HHg
+ bg αIbg
i−1
S0
tr(HHg KHg BHg Bg Kg Hg )
.
8
• The SINR of user gk given by
γgk ,pgp =
P ¯2 H
¯ g BHhg |2
ζ
|h
B
K
g
g
k
S g gk
P
P
P
P
H
2
¯2|hH Bg K
¯2 H
¯
¯ g BHhg |2 + P
ζ
g
j
j6=k g gk
g 0 6=g
j ζg 0 |hgk Bg 0 Kg 0 Bg 0 hgj0 |
S
S
+1
• Using the “deterministic equivalent” method of [Wagner, Couillet, Debbah,
Slock, 2011], we can calculate γgok ,pgp such that
M →∞
γgk ,pgp − γgok ,pgp −→ 0
9
Example
• M = 100, G = 6 user groups, Rank(Rg ) = 21, effective rank rg∗ = 11.
• We serve S 0 = 5 users per group with b0 = 10, r? = 6 and r? = 12.
• For rg∗ = 12: 150 bit/s/Hz at SNR = 18 dB: 5 bit/s/Hz per user, for 30 users
served simultaneously on the same time-frequency slot.
350
300
250
250
200
200
Sum Rate
Sum Rate
300
350
Capacity
ZFBF, JGP
RZFBF, JGP
ZFBF, PGP
RZFBF, PGP
150
150
100
100
50
50
0
0
5
10
15
SNR (in dBs)
20
25
30
Capacity
ZFBF, JGP
RZFBF, JGP
ZFBF, PGP
RZFBF, PGP
0
0
5
10
15
20
25
30
SNR (in dBs)
10
Training, Feedback and Computations Requirements
• Full CSI: 100 × 30 channel matrix ⇒ 3000 complex channel coefficients per
coherence block (CSI feedback), with 100 × 100 unitary “common” pilot matrix
for downlink channel estimation.
• JSDM with PGP: 6 × 10 × 5 diagonal blocks ⇒ 300 complex channel
coefficients per coherence block (CSI feedback), with 10 × 10 unitary
“dedicated” pilot matrices for downlink channel estimation, sent in parallel
to each group through the pre-beamforming matrix.
• One order of magnitude saving in both downlink training and CSI feedback.
• Computation: 6 matrix inversions of dimension 5 × 5, with respect to one
matrix inversion of dimension 30 × 30.
11
Non-ideal CSIT
• Parallel downlink training in all groups: a scaled unitary training matrix Xtr of
dimension b0 ×b0 is sent, simultaneously, to all groups in the common downlink
training phase.
• Received signal at group g receivers is given by
Yg =
HHg Xtr
+
X
Hg HBg0 Xtr + Zg .
g 0 6=g
• Multiplying from the right by XHtr and letting ρtr denote the power allocated to
training, we obtain
Yg XHtr
=
ρtrHHg
+ ρtr
X
Hg HBg0 + Zg XHtr.
g 0 6=g
12
• The relevant observation for the gk -th user effective channel is:


eg
h
k
√
√ X H 
= ρtrhgk + ρtr
zgk .
Bg0 hgk + e
g 0 6=g
• The corresponding MMSE estimator is given by
bg = E
h
k
=
√
h
eH
hg k h
gk
i
h
E

ρtr BHg Rg
i
H −1
eg h
e
eg
h
h
k gk
k
G
X
g 0 =1

Bg0  ρtr
G
X
−1
BHg0 Rg Bg00 + Ib0 
eg
h
k
g 0 ,g 00 =1
−1
1
1
T
T
eg
˜
˜
=√
Mg Rg O
ORg O +
Ib 0
h
k
ρtr
ρtr
13
where we used the fact that hgk = BHg hgk , and we introduced the b0 × b block
matrices
Mg
= [0, . . . , 0, |{z}
Ib0 , 0, . . . , 0]
block g
O = [Ib0 , Ib0 , . . . , Ib0 ].
• Notice that in the case of perfect BD we have that Rg Bg0 = 0 for g 0 6= g.
Therefore, the MMSE estimator reduces to
bg
h
k
−1
1
1 ¯ ¯
eg
h
Ib0
= √ Rg Rg +
k
ρtr
ρtr
¯ g = BHRg Bg .
where R
g
14
• Also in this case, the deterministic equivalent approximations of the SINR
terms for RZFBF and ZFBF precoding can be be computed.
• Eventually, the achievable rate of user gk is given by
Rgk ,pgp,csit
b
= max 1 − , 0 × log 1 + γ
bgok ,pgp,csit .
T
0
15
Tradeoff parameter b0
• b0 large yields better conditioned matrices, but it “costs” more in terms of
training phase dimension.
SNR = 30 dB
SNR = 10 dB
230
220
80
210
RZFBF, PGP, ICSI
ZFBF, PGP, ICSI
RZFBF, PGP
ZFBF, PGP
70
RZFBF, PGP, ICSI
ZFBF, PGP, ICSI
RZFBF, PGP
ZFBF, PGP
200
Sum Rates
Sum Rates
190
60
50
180
170
160
40
150
140
30
130
4
6
8
10
12
14
b‘
(a) S’ = 4, SNR = 10dB
16
4
6
8
10
12
14
16
b‘
(b) S’ = 8, SNR = 30dB
16
Impact of non-ideal CSIT
300
250
400
Full CSI, RZFBF
Full CSI, ZFBF
JGP, RZFBF
JGP, ZFBF
PGP, RZFBF
PGP, ZFBF
PGP ICSI, RZFBF
PGP ICSI, ZFBF
350
Full CSI, RZFBF
Full CSI, ZFBF
JGP, RZFBF
JGP, ZFBF
PGP, RZFBF
PGP, ZFBF
PGP ICSI, RZFBF
PGP ICSI, ZFBF
300
200
Sum Rate
Sum Rate
250
150
200
150
100
100
50
50
0
0
5
10
15
20
25
30
0
0
5
10
15
SNR (in dBs)
SNR (in dBs)
(c) S’ = 4
(d) S’ = 8
20
25
30
17
Discussion: is the tall unitary realistic?
• For a Uniform Linear Array (ULA), R is Toeplitz, with elements
[R]m,p
1
=
2∆
Z
∆
e−j2πD(m−p) sin(α+θ)dα,
m, p ∈ {0, 1, . . . , M − 1}
−∆
• We are interested in calculating the asymptotic rank, eigenvalue CDF and
structure of the eigenvectors, for M large, for given geometry parameters
D, θ, ∆.
• Correlation function
rm
1
=
2∆
Z
∆
e−j2πDm sin(α+θ)dα.
−∆
18
• As M → ∞, the eigenvalues of R tend to the “power spectral density” (i.e.,
the DT Fourier transform of rm),
S(ξ) =
∞
X
rme−j2πξm
m=−∞
sampled at ξ = k/M , for k = 0, . . . , M − 1.
• After some algebra, we arrive at
1
S(ξ) =
2∆
X
1
p
.
2
2
D − (m − ξ)
m∈[D sin(−∆+θ)+ξ,D sin(∆+θ)+ξ]
19
Szego’s Theorem: eigenvalues
Theorem 2. The empirical spectral distribution of the eigenvalues of R,
M
X
1
(M )
FR (λ) =
1{λm(R) ≤ λ},
M m=1
converges weakly to the limiting spectral distribution
(M )
lim FR
M →∞
(λ) = F (λ) =
Z
dξ.
S(ξ)≤λ
20
Example: M = 400, θ = π/6, D = 1, ∆ = π/10. Exact empirical eigenvalue cdf
of R (red), its approximation the circulant matrix C (dashed blue) and its
approximation from the samples of S(ξ) (dashed green).
1
0.9
Toeplitz
Circulant, M finite
Circulant, M ∞
0.8
0.7
CDF
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5
Eigen Values
21
A less well-known Szego’s Theorem: eigenvectors
Theorem 3. Let λ0(R) ≤ . . . , ≤ λM −1(R) and λ0(C) ≤ . . . , ≤ λM −1(C)
denote the set of ordered eigenvalues of R and C, and let U = [u0, . . . , uM −1]
and F = [f0, . . . , fM −1] denote the corresponding eigenvectors. For any interval
[a, b] ⊆ [κ1, κ2] such that F (λ) is continuous on [a, b], consider the eigenvalues
index sets I[a,b] = {m : λm(R) ∈ [a, b]} and J[a,b] = {m : λm(C) ∈ [a, b]},
and define U[a,b] = (um : m ∈ I[a,b]) and F[a,b] = (fm : m ∈ J[a,b]) be the
submatrices of U and F formed by the columns whose indices belong to the
sets I[a,b] and J[a,b], respectively. Then, the eigenvectors of C approximate the
eigenvectors of R in the sense that
2
1 H H
lim
U[a,b]U[a,b] − F[a,b]F[a,b] = 0.
M →∞ M
F
Consequence 1: Ug is well approximated by a “slice” of the DFT matrix.
Consequence 2: DFT pre-beamforming is near optimal for large M .
22
Theorem 4. The asymptotic normalized rank of the channel covariance matrix
R, with antenna separation λD, AoA θ and AS ∆, is given by
ρ = min{1, B(D, θ, ∆)},
with B(D, θ, ∆) = |D sin(−∆ + θ) − D sin(∆ + θ)|.
Theorem 5. Groups g and g 0 with angle of arrival θg and θg0 and common
angular spread ∆ have spectra with disjoint support if their AoA intervals
[θg − ∆, θg + ∆] and [θg0 − ∆, θg0 + ∆] are disjoint.
23
DFT Pre-Beamforming
1500
8
θ = −45
θ=0
θ = 45
7
RZFBF, Full
ZFBF, Full
RZFBF, DFT
ZFBF, DFT
6
Sum Rate
Eigen Values
1000
5
4
3
500
2
1
0
−0.5
0
ξ
• ULA with M = 400, G = 3, θ1 =
0.5
0
0
5
10
15
20
25
30
SNR
−π
4 , θ2
= 0, θ3 = π4 , D = 1/2 and ∆ = 15 deg.
24
Super-Massive MIMO
25
• Idea: produce a 3D pre-beamforming by Kronecker product of a “vertical”
beamforming, separating the sector into L concentric regions, and a
“horizontal” beamforming, separating each `-th region into G` groups.
• Horizontal beam forming is as before.
• For vertical beam forming we just need to find one dominating eigenmode
per region, and use the BD approach.
• A set of simultaneously served groups forms a “pattern”.
• Patterns need not cover the whole sector.
• Different intertwined patterns can be multiplexed in the time-frequency
domain in order to guarantee a fair coverage.
26
An example
• Cell radius 600m, group ring radius 30m, array height 50m, M = 200
columns, N = 300 rows.
• Pathloss g(x) =
1
1+( dx )δ
with δ = 3.8 and d0 = 30m.
0
• Same color regions are served simultaneously. Each ring is given equal
power.
0
120 degree sector
1000
BD, RZFBF
BD, ZFBF
DFT, RZFBF
DFT, ZFBF
600 m
50 m
Sum rate of annular regions
900
800
700
600
500
400
1
2
3
4
5
6
7
8
Annular Region Index l
27
Sum throughput (bit/s/Hz) under PFS and Max-min Fairness
Scheme
PFS, RZFBF
PFS, ZFBF
MAXMIN, RZFBF
MAXMIN, ZFBF
Approximate BD
1304.4611
1298.7944
1273.7203
1267.2368
DFT based
1067.9604
1064.2678
1042.1833
1037.2915
1000 bit/s/Hz × 40 MHz of bandwidth = 40 Gb/s per sector.
28
Our on-going work
• Compatibility with an in-band Small-Cell tier: eICIC in the spatial domain:
turn on and off the “spotbeams”.
• Multi-cell strategies: activate mutually compatible patterns of groups in
adjacent sectors.
• User grouping: we developed a very efficient way to cluster users according
to their dominant subspaces (quantization according to chordal distance).
See [Adhikary, Caire, arXiv:1305.7252].
• Hybrid Beamforming and mm-wave application: the DFT pre-beamforming
can be implemented by phase shifters in analog domain.
• Estimation of the long-term channel statistics: revamped interest in superresolution methods (MUSIC, ESPRIT) especially for the mm-wave case.
29
Conclusions
• Exploiting transmit antenna correlation reduces the channel to a simpler ≈
block diagonal structure.
• This is generalized sectorization! with MU-MIMO independently in each
“sector” (group).
• We need only very coarse information on AoA and AS for the users .... DFT
pre-beamforming.
• The idea can be easily extended to 3D beamforming (introducing elevation
direction, Kronecker product structure).
• Downlink training, CSIT feedback and computation are greatly reduced
(suitable for FDD).
• JSDM lends itself naturally to spatial-domain eICIC, simple inter-cell
coordination, hybrid beamforming for mm-wave applications.
30
Thank You
31