© Robert W. Heath Jr.

(2015)

Compressed-Sensing Based Multi-User Millimeter Wave

Systems: How Many Measurements Are Needed?

Ahmed Alkhateeb*, Geert Leus#, and Robert W. Heath Jr.*

* Wireless Networking and Communications Group, Department of Electrical and Computer
Engineering, The University of Texas at Austin
# Faculty of Electrical Engineering, Mathematics and Computer Science, Delft University of
Technology (TU Delft)

See more on mmWave @ School of ICASSP talk on Wednesday

www.profheath.org
 © Robert W. Heath Jr. (2015)

Why MIMO at mmWave?

millimeter wave band possible bands used for cellular several GHz of spectrum
is promising but found in
7 GHz
1.3 GHz 2.1 GHz
(unlic)
10 GHz many separate bands

28 GHz 37 / 42 GHz 60GHz E-Band … to 300 GHz

spatial multiplexing & beamforming just beamforming
mmWave multiple data streams
isotropic aperture
radiator

TX RX sub-6GHz
aperture
Beamforming for antenna gain Spatial multiplexing for spectral efficiency
Shu Sun, T. Rappapport, R. W. Heath, Jr., A. Nix, and S. Rangan, `` MIMO for Millimeter Wave Wireless Communications: Beamforming, Spatial Multiplexing,
or Both?,'' IEEE Communications Magazine, December 2014. 22
 © Robert W. Heath Jr. (2015)

Analog beamforning

Baseband RF RF Baseband
DAC ADC
Chain Chain

Phase
shifters

u  Low power consumption (1 RF chain)

u  Beamforming gain achieved using narrow beams De-facto approach in
u  MmWave specific constraints IEEE 802.11ad / WiGig
ª  Constant gains: Only phases are typically adjusted and Wireless HD
ª  Quantized phases: Fixed set of steering directions is allowed

Only provides single stream MIMO beamforming

3
 © Robert W. Heath Jr. (2015)

Hybrid analog/digital architecture

1-bit RF RF RF 1-bit
DAC Chain + RF Chain ADC
ADC Beam- ADC
Combining
forming
+
Baseband Baseband
Baseband Baseband
Precoding Combining
Precoding Precoding

RF
Beam- RF
1-bit RF + Combining RF 1-bit
DAC Chain forming Chain ADC
ADC ADC

u  Compromise on power consumption & complexity (# ADCs << # Antennas)

u  Enables spatial multiplexing and multi-user MIMO
u  Digital can correct for analog limitations [Aya’14]
Hybrid architecture is one viable approach for multi-stream MIMO
o  [Aya’14] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. Heath, “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE Transactions on
Wireless Communications, vol. 13, no. 3, pp. 1499–1513, March 2014 4
 © Robert W. Heath Jr. (2015)

MmWave channel estimation

u  MmWave channel estimation is challenging
ª  Large transmit and receive arrays
ª  Low signal-to-noise ratio before beam forming
ª  Limited number of RF chains imposes constraints on training signal design
ª  Channel is seen through the RF lens
u  Prior work (e.g., [Wan’09], [Hur’13])
ª  Avoids explicit channel estimation
ª  Relies on beam training with no prior channel knowledge
ª  Limitations: Supports only single-stream transmission

Need to design new mmWave channel estimation algorithms

o  [Wan’09] J. Wang, Z. Lan, C. Pyo, T. Baykas, C. Sum, M. Rahman, J. Gao, R. Funada, F. Kojima, H. Harada et al., “Beam codebook based beamforming protocol for
multi-Gbps millimeter-wave WPAN systems,” IEEE Journal on Selected Areas in Communications, vol. 27, no. 8, pp. 1390–1399, 2009.
o  [Hur’13] S. Hur, T. Kim, D. Love, J. Krogmeier, T. Thomas, and A. Ghosh, “Millimeter wave beamforming for wireless backhaul and access in small cell networks,”
IEEE Transactions on Communications, vol. 61, no. 10, pp. 4391–4403, 2013. 5
 © Robert W. Heath Jr. (2015)

MmWave channels are sparse

1 ✓1 2 ✓2 3 ✓3

u  MmWave channel estimation can be formulated as a sparse problem [Alk’14]

ª  Channels are sparse in the angular domain, a few paths exist [Akd’13]
u  Sparse formulation motivates using (adaptive) compressed sensing tools [Alk’14]
ª  Limitation: Training overhead scales with the # of users (repeated per user)

o  [Akd’13] M. Riza Akdeniz, Y. Liu, S. Sun, S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter Wave Channel Modeling and Cellular Capacity Evaluation,” IEEE Journal
on Selected Areas on Communications, vol. 32, no. 6, pp. 1164-1179, June 2014
o  [Alk’14] A. Alkhateeb, O. E. Ayach, G. Leus, and R. W. Heath Jr, “Channel estimation and hybrid precoding for millimeter wave cellular systems.” IEEE J. Selected
Topics in Signal Processing (JSTSP), vol. 8, no. 5, May 2014, pp. 831-846 6
 © Robert W. Heath Jr. (2015)

Contribution
u  Propose and evaluate a new mmWave system operation based on CS
ª  CS-based downlink channel training
ª  Data transmission with beamforming based on estimated channel

u  Evaluate system achievable rate as a function of the # of CS measurements

ª  Provides insights into the performance of CS-based mmWave systems
ª  Compares the performance with the exhaustive search solution

7
 which is limited by min (NRF , U ) for NBS > NRF .
plicity, we will also assume that the BS will use U o © Robert W. Heath Jr. (2015)

System
NRF availablemodel RF chains to serve the U users.
+ and
On the downlink, 1
s RF
Chain
thew1BS applies an NBS ×U RFten
+
pap
p
F = [f1 , fH 2 ,u..., fU ]. The sampled transmitted signal ass
wu
RF RF

T
is the U ×13.
Beamformers combiner RF

fore
RF
precoder
x = Fs, where NRF
s
F
= [s , s
1RF 2 ,
NBS
...,
NMS
s U ]wu
Chain

w combiner ∗ PT
transmitted symbols, s
N RF
RF
Chain
such
U that E [ss ] = U IU , and The
averageLimited totalFeedback
transmitted power. Since F is impleme
Base station
+

uth mobile station

sys
dow
1
ing quantized analog
u  U users, NBS antennas at BS and
phase
N MS
shifters,
antennas at each
[F]m,n = Ndom
MS
√ form
Fig. 1. A BS with RF beamformers and N RF chains com-
RF B
ª  Analog precoding atmunicating
where φm.n is the a base station
with theand
quantized uthanalog
angle,
MS combining
that andatRFeach
employs the user
factor of
combining. esti
ove
u  BS has # of RF chains = # of users
ma
for power normalization.
ª  Phase shifters are assumed
tiplexingto have
gain of athe
constant modulus
described & quantized
multi-user angles
precoding system, mu
which is limited by min (NRF , U ) for NBS > NRF . For sim- ove
8
 yu = wu∗ Hu fr sr + wu∗ nu , (2)
r=1
where
© Robert W. Heath Jr. (2015)
tor p
Channel model
MmWave channels are expected
✓ to have✓ limited scatter-
✓ 1 1 2 2 3 3
used
ing [2]. Therefore, and to simplify the analysis, we will as- ploys
sume a single-path geometric channel model [9, 13]. Under at MB
this model, the channel Hu can be expressed as surem
" instan
Hu = NBS NMS αu aMS (θu ) a∗BS (φu ) , (3) formi

where α#u is the complex

path gainpath array
gain, including
response theof path-loss,
angles arrival/
2 (includes path-loss)
$
with E |αu | = ᾱ. The variables vectorsθu , departure
and φu (AoA/AoD)
∈ [0, 2π]
u  are the angles
MmWave channelof arrival and departure (AoA/AoD) respec-
assumptions where
tively. Finally,
ª  Single-path channels (φusimplicity,
aBS (for ) and aitMS can(θbe
u )extended
are thetoantenna
multi-patharray
channels) ment
response vectorsvector
ª  Array response of the BS and
is known, uth MS
generally respectively.
non-uniform e.g. UPAThe BS beam
ª  Paths may be LOS/NLOS
o  [Akd’13] M. Riza Akdeniz, Y. Liu, S. Sun, S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter Wave Channel Modeling and Cellular Capacity Evaluation,” IEEE Journal
on Selected Areas on Communications, vol. 32, no. 6, pp. 1164-1179, June 2014 9
 at MBSas successive time slots, √and vectors
the MS quses MMS mea-
can be expressed surement n , n = 1, 2, ..., MMS at MMS
tor pm , and
surement we q
vectors use
n ,sparse =
nsm=instants P ,Mwith
1, nature
2, ..., atP
MS the M the
MS average
successive power
catter-
To exploit the of
to detect thechannel,
signal we ©first
transmitted vec-
over
Robert W. 3
(2015) o
Heath Jr.each
αu aMS (θuused a ∗ pertotransmission
)instants
torize
BS (φthe
u ) detect
, the
resultant(3)signalin transmitted
the training
forming
matrix Y
over
vectors,
as in
phase.
each
the
[5]
ofIfthethe
resulting
to get
BS em-
beam-
received matrix will
MS
will as-Downlink
(3) ploys
forming CS-based
Mvectors,
BS √such!the beamformingchannel
resulting vectors
received
∗
training
pm ,will
matrix m= be
√
1, ..., MBS ,
[5] G
Under at yMMS P PTtime
=successive H
Qslots, and the MS uses M H mea-
"
ath gain, including BSthe path-loss, ⊗√ (a (φ u )
Y ⊗
MS a =
MS (θ
P Q
u ))
MS α
Huu P ++v,N,
u  BSθsurement
BS c
uuse M1 training beamforming
qn ,=n =P vectors
ath-loss, H
variables , and ∈ [0,
φu vectors Y2π]
MS Q1, 2,H...,
u PM +MSN, at MMS successive (5)
[0, 2π] (6)
u  Each
d departure MS usesto M
(AoA/AoD)
instants 2 combining
respec-
detect the signalvectors
where Q for= [q
transmitted each, qbasestation
1over 2 , ...,
each qMof training
] isbeam-
MSthe the N vector
MS × MM w
drespec-
aMS where
(3)u (θuReceived
) forming
are the Q = at
Tosignal
antenna [q1user
complete
vectors, ,array
,u...,
qthe
2thecan qM
problem
ment ] matrix,
beMSwritten
resulting theasNP
isreceived
formulation, =× [p
MSmatrix M1MS ,we
p 2measure-
,assume
will ...,
bep[5] ] is the
MBSthat the BS Nw
=BS [p1are
, p2beamforming
,taken BS ] ismatrix, ×isManBS
na array
and ment matrix,
uth MS respectively.
AoDs, and ThePAoAs √..., pMHfrom the BS and
a grid NofBSN G M × M no
BS andMSGMS BS b
The BS
h-loss, beamforming matrix, YMSand =N is PQ an MH MS × MN,noise matrix. (5)
u P + BS
points, respectively. By neglecting the grid quantization er- c
[0, 2π]
u  After
ror,vectorization & neglecting y
we can approximate angle
MS in grid(6)
errors
as
espec- where Q = [q1 , q2 , ..., qMMS ] is# the NMS × M$MS measure- F
a array mentymatrix,
√
P =! [pT , p , ...,
H
"p ∗] is the BS N × M
MS = P P 1 ⊗2Q MA BS
BS ⊗ AMS zu + v,
BS BS (7)
he BS beamforming matrix, and N is an MMS × MBS noise matrix.
Vectorized receivedthe
signalN Sparse vector with a single non-
where BS× GBS matrix ABS and NMS × G AoA/AoD
zero at theMS
matrixlocation
Training
AMS beamforming/combining
are the! dictionary matricesmatrices
" thatwhose
Dictionary
(realized using analog-only or hybrid in general)! response
consist
columnsofarethe column
BS/MS array I
vectors with quantized
with φ̄grid angles
"
vectors aBS φ̄k , and aMS θ̄ℓ , respectively, k , and θ̄ℓ d
10
 h points of theP, angle
such that z[P]
grids. u isthe a Gkth,
BS
= Gand
e jφ
MS and ×the
ℓth
m,n points
where to of
get
dictionary
φ the
someangle
Ψ
is =grids.
insightsA
randomly zinto
u⊗Ais the Grelation
a MS BS, GMSintere
one ×
betw
m,n m,n networks BS [18]. dictionary Ψ)
hefficient
carries thetraining
path beamforming
gains of the
and uniformly selected from 1 and
vectorcombining
corresponding which matrices
carries
the Tsetofisof
direction the
mmWave
to path
study the
quantized gains
systems of the
conditions
angles and corresponding
compressed
onofΦ, Ψ un se
H rate
© Robert W. Heath Jr. (2015) u,
user
5]. Considering
ctions. Hence, zthe u 2πismeasurement
a∗sparse(NQBSvector
−1 2π withΦ
)matrix
quantized only=P ⊗
directions.
support in
of Q
Hence,
more
the , sparse
zu is a sparse
general cases.zvector
vector can
with only
be recove
{0, , ..., 1 }.
non-zero Each
element. MS similarly
Note that
Assumption
designs
detecting its
the u2
column
The BS
⎛A
of
and M
NQABS ⊗A
ment. Downlink
the dictionary
Note that =
Ψ detecting
CS-basedtheMS channel
, one interesting
Ncolumn of A BS research training BS
BS BS
Q probability Consider
and with the
low system
training model
overhead. in Sectio Le
ction is to study
corresponds to the conditions
this
training non-zero
combining and
element A
Φ, Ψdirectly
onmatrix MS that
under
Q, corresponds
withwhichN MS the to this
angle work, non-zero
quantization elementlog directly
1 +pape
jective forconjugate
Q future Assumption
analog we will 3try The
beamforming/combinin 2 sizes
in
⎝
this of th
implies the detection of the AoA and AoD of the channel path.
port ofu 
ection the
of Vectorized
thesparse
AoA
bits.and signal
vector
Note AoD zcan
thatuofcan
for be
thethis written
be recovered
channel
design, aseachwith
path.
insights entry high
ofathe
tractability,
into measurement
Gwe
sufficientMSmake =(notNMS the , following
and the
necessarily angle
assum
the mi{θg
Compressed sensing measurements: Thanks to the
bability
ed sensing and with low Φ
measurements:
matrix training overhead.
will also be equal
Thanks ?toLeaving
sparse formulation
tothe jζ
e , with this
ber of measurements,
ob- anglenary
the
ofAssumption
the mmWave
ζ selected
Ψ are
and
channel the the virtualprob-
relation
estimation direct
betwee
ive foroffuture
ation randomly
work,
the mmWave we from
channel y
aestimation
will try certain
MSin “
lemthisinquantized
paper
(7), P
prob-
compressed
and the to
z
to angle
get u `
set.
some v
sensing
achievable
2πk
tools
1 All path
rate and
can
of be2πd gainsDenoting
sin
leveraged
mmWave
!
θ̄ ℓ
are consta
"
=
to
systems.de-2πℓ R
mmWave NLOS
BS λ
paths, which are domina N MS
mpressed
ghts into asensing
sufficientAoA/AoD
tools(notcan estimation:
be leveraged
necessarily the To
sign efficient to estimate
de-
minimum)training its
num- channel AoA/AoD,
beamforming and combining matrices
single-user rate
considernetworks the following [18]. measurement matrix.
for powerpoint
raining
of measurements, each and
beamforming slides
MS u needs
and
the Measurement
combiningto
relation recover
[5,15].between
T
matrix
matrices thethis
Considering
H The
Dictionary
support
number of itsmatrix
the measurement
BS will
∗
sparseUnder
design
vectorthe
matrix
its
=rate
Φ assumptions
training
T
P of ⊗ user H
Q , u,
beamfo 2-
dering zurate
. Asofthe
the measurement
the achievable measurement
matrix
mmWave Φand =systems.
P ⊗
and Q dictionary
the dictionary For, that,
Ψ=A matrices
we BS ⊗A Φ,, one
become
MS Ψ DFT are
interesting
jφm,n matrices research [19].
∗ P, such Assumption
that [P] 2 The
= eBS and MS’s
where employ
φ m,n
%
siderΨthe=following
nary ⊗AMS
ABSknown for
, oneallinteresting
measurement MS’s, eachresearch
direction
matrix. MSis tocan studyusethe sparse recovery
conditions m,n on Φ,
nel algo-
model Ψtransformation
under which R̄u ≥ theEin R
u  Random rithms compressed sensing
support measurements
of andthe yrns
uniformly
sparse (similar
vector for Pursuit
selected
zu can P
be &from Q) the with
recovered set of highqua
studyBS
The the will design(such
conditions as
Ψ LASSO
onitsΦ,training
under [16],
which
beamforming Orthogonal
the matrix
Assumption
(
Matching
BS vectors
NQ −1)
3
2π
The of sizes
both of
Different
the the grids
designs
actual (a)
in
can
and
probability and with
2π low training overhead. Leaving this ob-
sparse
such that [P](OMP)
vector zu can
m,n = [17], etc.)where
ejφrecovered
be m,n to m,n
estimate
φ2
with m,n high
xr0 isN´
{0,its channel
,G...,
BS ks
randomly MS = N
AoA/AoD.
NBS , and
dictionary
MS
In the
}. theEach
Ψ) be
angleare investigated
MS grid
columns = R̊of
similarl
points u
simulations jective for future Qwork, we will Q try in this paper to get some
d uniformly
with low training
selected fromofthe
overhead. Section
Leaving
set 5,
of we
this adopt
ob- OMP
quantized
training nary forΨlow-complexity.
angles
combining areratetheof#virtual
matrix quantized
user Q, directions
angles
u,with
R u, N can MS besatisfy
writ
angl
ure u  (AoA/AoD
2π work, we
NQ will
−1)2π tryestimation
in this paper
insights into
(equivalent
BS In this case, the support of z
to get some xr1u ´ ksuandand
a, sufficient
to detecting
supp(z 2πk ),(not
is necessarily
the non-zero
determined
2πd
sin
! the
by" minimum)
θ̄ℓlocation
=each 2πℓ of[19].
Q
z
num-
u) ≥ R̊u
BS , ..., solving
BS }. Each MSber of measurements,
similarly bits. Note
designs N that
BS its the
for relation
this
λ ⎛ between
design, N this
MS entrynumber of the
sufficient
N (not
NQ necessarily the minimum) num-
Q

ning combining
ements, supp(z
relationQ,between
and the matrix
andMS
NuQ )
with consider = arg
this angle
numberxr2
the achievable
matrix
max ´Ψ Φ
Hks
rate
quantizationΦ Hof also
will y
mmWave
Under the, be
log
systems.
equal
Different
1 (8)
+
to e
support
jζ
For that, the
,recovery
with we an {θ̂
therandomly
following from MS
measurement 2assumptions
matrix. can be where 2-3, the (a)set. matr
Ufollo
⎝
a algorithms
certain quantized investigated
angle
)
vable
. NoteratethatofformmWave
this design,
which systems.
directly each entry
determinesFor
The ofthat,
the
BS
the xr3wewill´design
measurement
estimated ksbecome
AoA/AoDits DFT ˆu , φ̂beamforming
,θmatrices
training [19].
u .To estimate
=θFurther,
{θ̂ution u} matrix
success r=1us
11ev
AoA/AoD estimation: its chann
 which is limited by min (NRF , U ) for NB
plicity, we will also assume that the BS w © Robert W. Heath Jr. (2015)

Downlink data NRFtransmission

available RF chains to serve the U u
On the downlink, the w1BS applies an N
nel, we first vec- 3.2. Conjugate Analog Beamforming
get F = [f1 , fH 2 ,u..., fU ]. The sampled transm
Given the estimated AoA, each MS u will w
# design
$ u its analog T
S (θu )) αu + v, fore xvector
RF
combining
precoder
= Fs, where s = [s , s2 , ...,
such that wu = aMS θ̂u1RF. Each
s ]
MS uU
(6) w ∗
will also feed the index of its estimated AoD back to the BS ] =
[ss
combiner
transmitted symbols, such
U that E
e assume that the which needs log 2 GBSFeedback
Limited bits. Finally, the BS power.
designs its analog
f GBS and GMS
average total transmitted
beamforming matrix F to match the effective channels (in-
Since F
u  BS uses conjugate beamforming based on the estimated AoDs at the MS’s
quantization er- ing
cluding
% thequantized
#effect
$ of the analog
# combining
$ #phase
vectors),
$& i.e., the BS sets [F]m
shifters,
= aBSBeamforming
$ 3.2. Conjugate F whereφ̂φ
Analog 1 , aBS φ̂2 , ..., aBS φ̂U
m.n is a quantized .
angle, and th
zu +u v,
MS’s (7) maximum ratio combining (matched to the estimated AoA’s)
Given the use
estimatedforAoA, each MS
power u will design
normalization. its analog
4. ACHIEVABLE AND EFFECTIVE RATES
# $
combining
MS × GMS matrix vector such that wu = aMS θ̂u . Each MS u
st of the column
For simplicity, we adopt a narrowband
will also feed the index of its estimated AoD back to the BS
In this section, we evaluate the achievable rate of the proposed
y, with φ̄ , and θ̄ 12
 r which the Assumption 3 The sizes of the grids in (7) are GBS = NBS ,
d with high GMS = NMS , and the angle grid points φ̄k , θ̄ℓ , of the!© dictio- Robert
2πd
" W. Heath Jr. (2015)
ing this ob- nary Ψ are the virtual directions satisfying λ sin φ̄k =
o get Calculating
some 2πk the
2πd achievable 2πℓ rate
! "
NBS and λ sin θ̄ ℓ = NMS [19].
mum) num-
u  Key assumptions
his number Under the assumptions 2-3, the matrices AMS and AMS
or that, weª  All channels are line-of-sight (generalization to NLOS is possible)
become DFT matrices [19]. Further, using the virtual chan-
ª  Arrays are uniform (ULAs/UPAs)
nel model transformation in [19], we note that the steering
ª  AoAs/AoDs are taken from a grid with critical quantization
ing matrix vectors of both the actual and estimated AoAs/AoDs (of the
s randomly dictionary Ψ) are columns of these matrices. Therefore, the
u  The achievable rate of user u can be written as
zed angles rate of user u, Ru , can be written as
designs its ⎛ ⎞
{θ̂u =θu ,φu =φ̂u }
Ru = log2 ⎝1 +
uantization )U ⎠.
1
easurement {θ̂u =θu } r=1 {φu =φ̂r } + SNR 2
U NBS NMS |αu |
e ζ selected # $ (9)
SNR 2
Denoting R̊u = log2 1 +Estimated U N AoAs/AoDs
BS N MS |α uusing
| as the
AoA/AoD, the recovery algorithm
single-user rate (without interference), the average achievable 13
 {φu =φ̂r }(M+
NBS λ ℓ NMS U
measurement
ngle ζ selected {θ̂u =θuN}umberr=1 of Measurements
# xSNR
BS UMS BS$ MS (9)
M N ) N |αu |2
e ζ selected
Under the assumptions 2-3, the
Denoting = log2A1MS
R̊u matrices #+ and SNRA
N MS
BS N MS |α u |
2
as $ (9) W. Heath Jr. (2015)
the © Robert
U
2
nnel AoA/AoD,
become DFT matrices [19]. Further,
Denoting
single-user u using
rateR̊(without = log the2 virtual1 + SNR
interference), chan-
theN N
average
BS MS |α |
achievable
u as the
Achievable rate
Fig. 2. Achievable rates using the Uproposed system operation Fig. 3. A
ls nel
sparse
AoA/AoD, vector
model transformation
rate of
single-user
in
user [19],
u,
rateR̄ we
u , can
(without
note be that
then the
lower
interference),
steering
bounded as
the average achievable
rices Φ,
vectors
parse vectorΨof are
both thefor different
actual and numbers
estimated of
AoAs/AoDs compressed (of the sensing measurements for differ
u 
recovery
dictionary The
algo-
Ψ) rate
are ratecan
columnsofbeuser
of u, R̄umatrices.
bounded
%
these , !can
as be Therefore,
then lowerthebounded as &
es Φ, Ψ are R̄ ≥ E R̊ , (10)
r̸=u (φu ̸=φ̂r )} {θ̂u =θu ,φu =φ̂u }
u u {
atching user u, Ru , can be written as
rate of Pursuit % &
covery algo- u ≥
rateR̄given (a) E R̊u! {! 1.(φNow,
assumption we {note Pφ̂θ̂uu}=θ(11)
,u ,φu =(10)
A/AoD. ⎛ In the = R̊u P{ u ̸=φ̂ P r )} ⎞ θ̂u =θthat
u ,φu, ={ φ̂u } sensing
hing Pursuit is the probability (r̸φ=u
of correct
r̸=u u =
̸ φ̂ r ) }) {θ u =θ u ,φ u = φ̂ }
support recovery of the sparse
u
ow-complexity. {θ̂u =θ,
(a) u ,φu =φ̂u } 330 mea
log 1 + .
-
AoD. 2 In the )U z=
vector R̊u P1directly! U relates the P{θachievable Probability
, to(12) of support
(11) recovery
u .R̊This rate the com-
⎝ ⎠
determined by ≥ u {−+ r̸ = u ( φ u =
̸ P 1
φ̂ r ) }) u =θ u,,φ u = φ̂ u } with perf
-complexity. { θ̂ =θ } r=1 {φ = φ̂ } SNR { θ̂ =θ 2 u =φ̂u }
,φ
u u
pressed sensing u r NBS
, literature. U N BS N u |α u|
-For a general relation, assume the
MS u
U (9) large, it i
termined
S, by(8) channel≥fading # R̊uSNR1coherence− the independence P2{θ̂u =θLu ,φ
equals $
symbols, , the effective(12)
where
Denoting R̊u where= log(a) follows
1 + from
N BS MS N |α u | as Cbetween
the u =φ̂ u } the estima-
Single-user rate quired b
2 BS
AoD ,θˆu , φ̂u . achievable
tion success rate U
eventof{θ̂user u = uθu(considering
, φu = φ̂u } and thethe training
single-useroverhead) NBS NM
u  Accounting
single-user(8)rate (without for the channel
interference), the coherence
average time,
achievable the rate is bounded by
rate of user u, R̄uwhere,can
can be written
be(a)
thenfollows
lower asbounded
from theasindependence between the estima- head doe
ˆ
D ,θu , φ̂u . % tion success event {θ̂u = U ! θu&, φu = φ̂uM} and the single-user
" ! "
compres
ϵ
R̄u ≥ E R̊u { ! R̄u,eff ≥ R̊u 1 − φ̂u } , 1(10) − (1 − ϵ), (13) systems.
r̸=u (φu ̸=φ̂r )} {θ̂u =θu ,φu = NBS LC
(a) niques [5
= R̊u P{!where P
M )})equals {θu =θuthe ,φu =number , (11)
of measurements needed to
r̸=u (φu ̸=φ̂rϵ φ̂u } channel,
Single-user
, rate with - no interference Channel coherence # of CS measurements with
and perfectguarantee
channel
U the support recovery
knowledge time (# symbols)of the sparse vector support zrecovery
u with errorscales✏ w
≥ R̊u 1 − P{θ̂u =θu ,φu =φ̂u } , (12)
probability
N at least 1 − ϵ. Given this relation in (13), it is 14
 © Robert W. Heath Jr. (2015)

Simulation results
9

Setup: 7

•  BS has ULA with 64 antennas

Achievable Rate (bps/ Hz)

•  MS has ULA’s with 32 antennas
6

•  Operating at 28 GHz 5 for a given ✏, best M✏

•  Bandwidth 50 MHz 4
will be less than this
•  TX-RX separation is 500 m
•  Average transmit power of 37 dBm
3

•  2 bits for phase shifter angle quantization 2

•  Random phase shifted P, Q 1

Analog Beamforming − Perfect Channel Knowledge
Analog Beamforming − CS−Based Channel Estimation
•  OMP recovery Lower Bound in (12) with 0.95 recovery success probability
0
0 50 100 150 200 250 300 350 400 450 500
Number of Measurements (MBS x MMS)
u  Need for 200-300 random measurements to approach optimal rate
u  Order of magnitude less than exhaustive search ~ 64x32
u  Optimizing the measurement matrices design should improve the performance
u  With N RF chains at the receiver, overhead will be (200-300)/N
15
 © Robert W. Heath Jr. (2015)

Simulation results
6
Training that maximizes
the achievable rate
5
Setup:

Effective Achievable Rate (bps/ Hz)

•  BS has ULA with 64 antennas 4
•  MS’s have ULA’s with 32 antennas
•  Operating at 28 GHz 3
•  Bandwidth 50 MHz
•  TX-RX separation is 500 m 2
•  Average transmit power of 37 dBm LC = 600 symbols
•  2 bits for phase shifter angle quantization 1
LC = 400 symbols
LC = 200 symbols

0
0 50 100 150 200 250 300 350 400
Number of Measurements (MBS x MMS)

u  More CS measurements does not necessarily means better performance

u  CS measurements need to be adapted with the channel coherence time
u  Having more RF chains at the receivers will increases the efficiency
16
 © Robert W. Heath Jr. (2015)

Conclusion
u  Compressed sensing based mmWave systems are promising
ª  An order of magnitude less training overhead vs exhaustive search solutions
ª  All users can simultaneously estimate their channels (similar codebooks)
ª  With hybrid receivers, RF chains can be used to reduce training overhead

u  Future work

ª  Investigate different CS beamforming/measurement matrices design
ª  Evaluating the performance for different sparse recovery algorithms (OMP,…)
ª  Extend to multi-path & wideband channels
ª  Include the impact of AoAs/AoDs quantization errors

17
 © Robert W. Heath Jr. (2015)

Questions?

Professor Robert W. Heath Jr.

Wireless Networking and Communications Group
Department of Electrical and Computer Engineering
The University of Texas at Austin

www.profheath.org 18