Spatially Sparse Precoding in Millimeter Wave MIMO Systems

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO.
3, MARCH 2014 1499
Spatially Sparse Precoding in

Millimeter Wave MIMO Systems
Omar El Ayach, Member, IEEE, Sridhar Rajagopal, Senior Member, IEEE, Shadi Abu-Surra, Member, IEEE,
Zhouyue Pi, Senior Member, IEEE, and Robert W. Heath, Jr., Fellow, IEEE
Abstract—Millimeter wave (mmWave) signals experience achieving this efficiency [1], [2]. Since there seems to be
orders-of-magnitude more pathloss than the microwave signals little scope for further gains at the physical layer, and since
currently used in most wireless applications and all cellular the widespread deployment of heterogeneous networks is not
systems. MmWave systems must therefore leverage large antenna
arrays, made possible by the decrease in wavelength, to combat without challenges [3], these techniques alone may not be
pathloss with beamforming gain. Beamforming with multiple sufficient to meet future traffic demands. As a result, increas-
data streams, known as precoding, can be used to further ing the spectrum available for commercial wireless systems,
improve mmWave spectral efficiency. Both beamforming and potentially by exploring new less-congested spectrum bands,
precoding are done digitally at baseband in traditional multi- is a promising solution to increase network capacity.
antenna systems. The high cost and power consumption of
mixed-signal devices in mmWave systems, however, make analog Millimeter wave (mmWave) communication, for example,
processing in the RF domain more attractive. This hardware has enabled gigabit-per-second data rates in indoor wireless
limitation restricts the feasible set of precoders and combiners systems [4], [5] and fixed outdoor systems [6]. More recently,
that can be applied by practical mmWave transceivers. In this
paper, we consider transmit precoding and receiver combining advances in mmWave hardware [7] and the potential avail-
in mmWave systems with large antenna arrays. We exploit the ability of spectrum has encouraged the wireless industry to
spatial structure of mmWave channels to formulate the pre- consider mmWave for the access link in outdoor cellular sys-
coding/combining problem as a sparse reconstruction problem. tems [8], [9]. A main differentiating factor in mmWave com-
Using the principle of basis pursuit, we develop algorithms that munication is that the ten-fold increase in carrier frequency,
accurately approximate optimal unconstrained precoders and
combiners such that they can be implemented in low-cost RF compared to the current majority of wireless systems, im-
hardware. We present numerical results on the performance of plies that mmWave signals experience an orders-of-magnitude
the proposed algorithms and show that they allow mmWave increase in free-space pathloss. An interesting redeeming
systems to approach their unconstrained performance limits, feature in mmWave systems, however, is that the decrease
even when transceiver hardware constraints are considered. in wavelength enables packing a large number of antenna
Index Terms—Millimeter wave, multiple-input multiple-output elements into small form factors. Large arrays can provide the
(MIMO), antenna arrays, beamforming, precoding, cellular com- beamforming gain needed to overcome pathloss and establish
munication, sparsity, sparse reconstruction, basis pursuit, limited
links with reasonable signal-to-noise ratio (SNR). Further,
feedback.
large arrays may enable precoding multiple data streams
which could improve spectral efficiency and allow systems
I. I NTRODUCTION to approach capacity [10], [11].
T HE capacity of wireless networks has thus far scaled with

the increasing data traffic, primarily due to improved area
spectral efficiency (bits/s/Hz/m2 ) [1]. A number of physical
While the fundamentals of precoding are the same regard-
less of carrier frequency, signal processing in mmWave sys-
tems is subject to a set of non-trivial practical constraints. For
layer enhancements such as multiple antennas, channel coding, example, traditional multiple-input multiple-output (MIMO)
and interference coordination, as well as the general trend processing is often performed digitally at baseband, which
toward network densification have all been instrumental in enables controlling both the signal’s phase and amplitude.
Manuscript received May 10, 2013; revised September 20, 2013; accepted Digital processing, however, requires dedicated baseband and
November 25, 2013. The associate editor coordinating the review of this paper RF hardware for each antenna element. Unfortunately, the
and approving it for publication was X. Dong.
O. El Ayach is with Qualcomm Technologies Inc., San Diego, CA, 92122 high cost and power consumption of mmWave mixed-signal
(e-mail: oelayach@qti.qualcomm.com). hardware precludes such a transceiver architecture at present,
S. Rajagopal, S. Abu-Surra, and Z. Pi are with Samsung Research and forces mmWave systems to rely heavily on analog or RF
America - Dallas, Richardson, TX, 75082 USA (e-mail: {sasurra, srajagop,
zpi}@sta.samsung.com). processing [7], [8]. Analog precoding is often implemented
R. Heath is with The University of Texas at Austin, Austin, TX 78712 using phase shifters [7], [8], [12] which places constant
USA (e-mail: rheath@utexas.edu). modulus constraints on the elements of the RF precoder.
This work was done while the first author was with The University of Texas
at Austin and with Samsung Research America - Dallas. The authors at The Several approaches have been considered for precoding in
University of Texas at Austin were supported in part by the Army Research such low-complexity transceivers [13]–[27]. The work in [13]–
Laboratory contract W911NF-10-1-0420 and National Science Foundation [15] considers antenna (or antenna subset) selection which has
grant 1218338 and 1319556. This work has appeared in part at the 2012
IEEE International Communications Conference (ICC). the advantage of replacing phase shifters with even simpler
Digital Object Identifier 10.1109/TWC.2014.011714.130846 analog switches. Selection, however, provides limited array
1536-1276/14$31.00 © 2014 IEEE
1500 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 3, MARCH 2014
gain and performs poorly in correlated channels such as to receiver-side processing and show that designing hybrid
those experienced in mmWave [16]. To improve performance minimum mean-square error (MMSE) combiners can again
over correlated channels, the work [17]–[19] considers beam be cast as a simultaneously sparse approximation problem and
steering solutions in which phase shifters are used to optimally solved via basis pursuit [50], [51]. We argue that, in addition
orient an array’s response in space, potentially based on to providing practical near-optimal precoders, the proposed
statistical channel knowledge. The strategies in [17]–[19], framework is particularly amenable for limited feedback op-
however, are in general suboptimal since beam steering alone eration and is thus not limited to genie-aided systems with
cannot perfectly capture the channels dominant eigenmodes. perfect transmitter channel knowledge [52]. The generated
The work in [20]–[25] develops iterative precoding algorithms precoders can be efficiently compressed using simple scalar
for systems that leverage analog processing, and [26] further quantizers (for the arguments of the beam steering vectors)
proposes simple analytical solutions. Further hardware limita- and low-dimensional Grassmannian subspace quantizers (used
tions have also been considered in [27], for example, which to quantize the baseband precoder) [52]–[54]. We briefly
focuses on analog receiver processing with only quantized describe the construction of the limited feedback codebooks
phase control and finite-precision analog-to-digital converters. required, but defer the analysis of limited feedback perfor-
The work in [20]–[27], however, is not specialized to mmWave mance to future work. Finally, we present simulation results
MIMO systems with large antenna arrays. Namely, the work in on the performance of the proposed strategy and show that
[20]–[27] does not leverage the structure present in mmWave it allows mmWave systems to approach their unconstrained
MIMO channels and adopts models that do not fully capture performance limits even when practical transceiver constraints
the effect of limited mmWave scattering and large tightly- are considered.
packed arrays [28]–[31]. We use the following notation throughout this paper: A is
In this paper, we focus on the precoding insight and a matrix; a is a vector; a is a scalar; A(i) is the ith column
solutions that can be derived from jointly considering the of A; (·)T and (·)∗ denote transpose and conjugate transpose
following three factors: (i) precoding with RF hardware con- respectively; AF is the Frobenius norm of A, tr(A) is
straints, (ii) the use of large antenna arrays, and (iii) the its trace and |A| is its determinant; ap is the p-norm of
limited scattering nature of mmWave channels. We consider a; [A | B] denotes horizontal concatenation; diag(A) is a
single-user precoding for a practical transceiver architecture vector formed by the diagonal elements of A; IN is the N ×
in which a large antenna array is driven by a limited number N identity matrix; 0M×N is the M × N all-zeros matrix;
of transmit/receive chains [8], [10], [11], [32]. In such a CN (a; A) is a complex Gaussian vector with mean a and
system, transmitters have the ability to apply high-dimensional covariance matrix A. Expectation is denoted by E[·] and the
(tall) RF precoders, implemented via analog phase shifters, real part of a variable is denoted by {·}.
followed by low-dimensional (small) digital precoders that can
be implemented at baseband. We adopt a realistic clustered II. S YSTEM M ODEL
channel model that captures both the limited scattering at high In this section, we present the mmWave signal and channel
frequency and the antenna correlation present in mmWave model considered in this paper.
antenna arrays [28]–[31]. We show that the joint treatment of
practical hardware architectures and realistic channel models A. System Model
can yield simple precoding solutions with near-optimal spec- Consider the single-user mmWave system shown in Fig. 1
tral efficiency. We note that a similar observation is made in which a transmitter with Nt antennas communicates Ns
by the authors of [33]–[37] in which low-complexity hybrid data streams to a receiver with Nr antennas [32]. To enable
analog-digital transceivers are constructed by leveraging the multi-stream communication, the transmitter is equipped with
concept of beamspace MIMO in which a number of dominant NtRF transmit chains such that Ns ≤ NtRF ≤ Nt . This
orthogonal propagation paths are selected and are further hardware architecture enables the transmitter to apply an
digitally combined at baseband. NtRF × Ns baseband precoder FBB using its NtRF transmit
We exploit the sparse-scattering structure of mmWave chan- chains, followed by an Nt × NtRF RF precoder FRF using
nels to formulate the design of hybrid RF/baseband precoders analog circuitry. The discrete-time transmitted signal is there-
as a sparsity constrained matrix reconstruction problem [38]– fore given by x = FRF FBB s where s is the Ns × 1 symbol
[43]. Initial results on this precoding approach were presented vector such that E [ss∗ ] = N1s INs . Since FRF is implemented
in [44]. In this paper, we formalize the mmWave precoding using analog phase shifters, its elements are constrained to
problem and show that, instead of directly maximizing mutual (i) (i)∗
satisfy (FRF FRF ), = Nt −1 , where (·), denotes the th
information, near-optimal hybrid precoders can be found via
diagonal element of a matrix, i.e., all elements of FRF have
an optimization that resembles the problem of sparse signal
equal norm. The transmitter’s total power constraint is en-
recovery with multiple measurement vectors, also known as
forced by normalizing FBB such that FRF FBB 2F = Ns ; no
the simultaneously sparse approximation problem [45]–[48].
other hardware-related constraints are placed on the baseband
We thus provide an algorithmic precoding solution based on
precoder.
the concept of orthogonal matching pursuit [39], [41], [49].
For simplicity, we consider a narrowband block-fading
The algorithm takes an optimal unconstrained precoder as in-
propagation channel as in [10], [25], [32], [34], which yields
put and approximates it as linear combination of beam steering
a received signal
vectors that can be applied at RF (and combined digitally at
√
baseband). Further, we extend this sparse precoding approach y = ρHFRF FBB s + n, (1)
EL AYACH et al.: SPATIALLY SPARSE PRECODING IN MILLIMETER WAVE MIMO SYSTEMS 1501
Digital Analog Analog Digital
RF RF
Chain Chain
Baseband Baseband
Precoder Combiner
Ns N tRF FRF Nt Nr WRF NrRF Ns
FBB WBB
RF RF
Chain Chain
splitters
Fig. 1. Simplified hardware block diagram of mmWave single user system with digital baseband precoding followed by constrained radio frequency precoding
implemented using RF phase shifters.
where y is the Nr × 1 received

vector, H is the Nr × Nt scattering. Similarly, the large tightly-packed antenna arrays
channel matrix such that E H2F = Nt Nr , ρ represents the that are characteristic of mmWave transceivers lead to high
average received power, and n is the vector of i.i.d CN (0, σn2 ) levels of antenna correlation. This combination of tightly
noise. In writing (1), we implicitly assume perfect timing and packed arrays in sparse scattering environments makes many
frequency recovery. Moreover, to enable precoding, we assume of the statistical fading distributions used in traditional MIMO
that the channel H is known perfectly and instantaneously analysis inaccurate for mmWave channel modeling. For this
to both the transmitter and receiver. In practical systems, reason, we adopt a narrowband clustered channel representa-
channel state information (CSI) at the receiver can be obtained tion, based on the extended Saleh-Valenzuela model, which
via training [17], [55]–[57] and subsequently shared with allows us to accurately capture the mathematical structure
the transmitter via limited feedback [52]; an efficient limited present in mmWave channels [28], [29], [31], [59], [60].
feedback strategy is presented in Section V. Techniques for Using the clustered channel model, the matrix channel H
efficient mmWave channel estimation that potentially leverage is assumed to be a sum of the contributions of Ncl scattering
the geometric nature of mmWave channels [55]–[57], as well clusters, each of which contribute Nray propagation paths to
as the rigorous treatment of frequency selectivity, are still an the channel matrix H. Therefore, the discrete-time narrowband
ongoing topic of research. channel H can be written as
The receiver uses its Ns ≤ NrRF ≤ Nr RF chains and
analog phase shifters to obtain the processed received signal H=γ αi Λr (φri , θi
r
)Λt (φti , θi
t
)ar (φri , θi
r
)at (φti , θi
t ∗
) ,
√ ∗ ∗ ∗ ∗
i,
= ρWBB
y WRF HFRF FBB s + WBB WRF n, (2) (4)
where WRF is the Nr ×NrRF RF combining matrix and WBB where
γ is a normalization factor such that γ =
is the NrRF × Ns baseband combining matrix. Similarly to Nt Nr /Ncl Nray and αi is the complex gain of the th ray
the RF precoder, WRF is implemented using phase shifters in the ith scattering cluster, whereas φri (θi r
) and φti (θi t
)
(i) (i)∗
and therefore is such that (WRF WRF ), = Nr −1 . When are its azimuth (elevation) angles of arrival and departure
Gaussian symbols are transmitted over the mmWave channel, respectively. The functions Λt (φti , θi t
) and Λr (φri , θi
r
) rep-
the spectral efficiency achieved is given by [58] resent the transmit and receive antenna element gain at the
corresponding angles of departure and arrival. Finally, the
ρ −1 ∗
R = log2 INs + R WBB WRF ∗
HFRF FBB vectors ar (φr , θr ) and at (φti , θi
t
) represent the normalized
Ns n
receive and transmit array response vectors at an azimuth

× FBB FRF H WRF WBB ,
∗ ∗ ∗ (elevation) angle of φri (θi r
) and φti (θi t
) respectively.
2
(3) In Section VI, we assume that αi are i.i.d. CN (0, σα,i )
2 th
where σα,i represents the average power of the i cluster. The
where Rn = σn2 WBB∗
WRF∗
WRF WBB is the noise covariance
Ncl 2
average cluster powers are such that i=1 σα,i= γ where

matrix after combining. γ is a normalization constant that satisfies E H2F =
Nt Nr [29]. The Nray azimuth and elevation angles of de-
B. Channel Model parture, φti and θi
t
, within the cluster i are assumed to be
The high free-space pathloss that is a characteristic of randomly distributed with a uniformly-random mean cluster
mmWave propagation leads to limited spatial selectivity or angle of φti and θit respectively, and a constant angular spread
(standard deviation) of σφt and σθt respectively. The azimuth in mmWave beamforming since they (i) yield smaller antenna
and elevation angles of arrival, φri and θi r
, are again randomly array dimensions, (ii) facilitate packing more antenna elements
distributed with mean cluster angles of (φri , θir ) and angular in a reasonably-sized array, and (iii) enable beamforming in
spreads (σφt , σθr ). While a variety of distributions have been the elevation domain (also known as 3D beamforming).
proposed for the angles of arrival and departure in clustered
channel models, the Laplacian distribution has been found to III. S PATIALLY S PARSE P RECODING FOR THE S INGLE
be a good fit for a variety of propagation scenarios [61], U SER MM WAVE C HANNEL
and will thus be adopted in the numerical results of Sec-
tion VI. Similarly, a number of parametrized mathematical We seek to design hybrid mmWave precoders (FRF , FBB )
models have been proposed for the functions Λt (φti , θi t
) and that maximize the spectral efficiency expression in (3). Di-
r r
Λr (φi , θi ). For example, if the transmitter’s antenna elements rectly maximizing (3), however, requires a joint optimization
are modeled as being ideal sectored elements [62], Λt (φti , θi t
) over the four matrix variables (FRF , FBB , WRF , WBB ). Un-
would be given by fortunately, finding global optima for similar constrained joint
optimization problems is often found to be intractable [64],
1 ∀φti ∈ [φtmin , φtmax ], ∀θi t t
∈ [θmin t
, θmax ], [65]. In the case of mmWave precoding, the non-convex
Λt (φti , θi
t
)=
0 otherwise, constraints on FRF and WRF makes finding an exact solu-
(5) tion unlikely. To simplify transceiver design, we temporarily
where we have assumed unit gain over the sector defined decouple the joint transmitter-receiver optimization problem
by φt ∈ [φtmin , φtmax ] and θt ∈ [θmin t t
, θmax ] without loss of and focus on the design of the hybrid precoders FRF FBB .
generality. The receive antenna element gain Λr (φri , θi r
) is Therefore, in lieu of maximizing spectral efficiency, we design
defined similarly over the azimuth sector φri ∈ [φrmin , φrmax ] FRF FBB to maximize the mutual information achieved by
r r r
and elevation sector θi ∈ [θmin , θmax ]. Alternatively, instead Gaussian signaling over the mmWave channel
of considering the simplified model in (5), the functions
ρ
Λt (φti , θi
t
) and Λr (φri , θi
r
) can be replaced by the well- I(FRF , FBB ) = log2 I+ HF F F ∗
F ∗
H ∗
Ns σn2 RF BB BB RF .
known far field radiation patterns for commonly used antennas
such as patch or half-wave dipole antennas [63]. (8)
We note here that abstracting receiver operation, and focus-
The array response vectors at (φti , θi t
) and ar (φr , θr ) are a
ing on mutual information instead of the spectral efficiency
function of the transmit and receiver antenna array structure
expression in (3), effectively amounts to assuming that the
only, and are thus independent of the antenna element proper-
receiver can perform optimal nearest-neighbor decoding based
ties. While the algorithms and results derived in the remainder
on the Nr -dimensional received signal y. Unfortunately, such
of this paper can be applied to arbitrary antenna arrays, we
a decoder is impossible to realize with practical mmWave
give the following two illustrative examples of commonly-used
systems in which decoders do not have access to the Nr -
antenna arrays for completeness. For an N -element uniform
dimensional signal. In practical mmWave systems, received
linear array (ULA) on the y-axis1 , the array response vector
signals must be combined in the analog domain, and possibly
can be written as [63]
in the digital domain, before any detection or decoding is
1 T
aULAy (φ) = √ 1, ejkd sin(φ) , . . . , ej(N −1)kd sin(φ) , performed. For this reason, we revisit the problem of designing
N practical mmWave receivers in Section IV.
(6) Proceeding with the design of FRF FBB , the precoder
where k = 2π λ and d is the inter-element spacing. Note that optimization problem can be stated as
we do not include θ in the arguments of aULAy as the array’s
response is invariant in the elevation domain. In the case of a (Fopt opt
RF , FBB ) = arg max I(FRF , FBB ),
FRF , FBB
uniform planar array (UPA) in the yz-plane2 with W and H
s.t. FRF ∈ FRF , (9)
elements on the y and z axes respectively, the array response
vector is given by [63] FRF FBB 2F = Ns ,
1
aUPA (φ, θ) = √ [ 1, . . . , ejkd(m sin(φ) sin(θ)+n cos(θ)) , where FRF is the set of feasible RF precoders, i.e., the set
N of Nt × NtRF matrices with constant-magnitude entries. To
. . . , ejkd((W −1) sin(φ) sin(θ)+(H−1) cos(θ)) ] ,
T the extent of the authors’ knowledge, no general solutions to
(7) (9) are known in the presence of the non-convex feasibility
constraint FRF ∈ FRF . Therefore, we propose to solve an
where 0 ≤ m < W and 0 ≤ n < H are the y and z indices approximation of (9) in order to find practical near-optimal
of an antenna element respectively and the antenna array size precoders that can be implemented in the system of Fig. 1.
is N = W H. Considering uniform planar arrays is of interest We start by examining the mutual information achieved by
the hybrid precoders FRF FBB and rewriting (8) in terms of
1 Following the standard notation used in [63], we use the terms x, y,
the “distance” between FRF FBB and the channel’s optimal
and z-axes to refer to the axes of the standard Cartesian coordinate system
defined at the antenna array itself. Similarly, we adopt the standard notation unconstrained precoder Fopt . To do so, define the channel’s
and conventions for the polar coordinate system when dealing with the angles ordered singular value decomposition (SVD) to be H =
φr , φt , θ r , and θ t ). UΣV∗ where U is an Nr × rank(H) unitary matrix, Σ
2 Note that placing the UPA on the yz-plane (and the earlier UPA along the
y-axis) is a completely arbitrary choice that was made to simplify notation. is a rank(H) × rank(H) diagonal matrix of singular values
All results hold regardless of array orientation arranged in decreasing order, and V is a Nt ×rank(H) unitary
matrix. Using the SVD of H and standard mathematical partition of the matrix V∗ FRF FBB F∗BB F∗RF V as
manipulation, (8) can be rewritten as
V∗ FRF FBB F∗BB F∗RF V
I(FRF , FBB ) = ∗
V1 FRF FBB F∗BB F∗RF V1 , V1∗ FRF FBB F∗BB F∗RF V2
=
ρ V∗ FRF FBB F∗BB F∗RF V1 , V2∗ FRF FBB F∗BB F∗RF V2
log2 I+ 2
Σ 2 ∗
V F F F∗
F∗
V
RF BB BB RF . 2
Ns σn Q11 Q12
(10) = ,
Q21 Q22
Further, defining the following two partitions of the matrices which allows us to approximate the mutual information
Σ and V as achieved by FRF FBB as

Σ1 0
Σ= , V = [V1 V2 ] , (11) I(FRF , FBB )
0 Σ2
ρ
where Σ1 is of dimension Ns × Ns and V1 is of dimension
= log2 I + Σ V FRF FBB FBB FRF V
2 ∗ ∗ ∗
Ns σn2
Nt × Ns , we note that the optimal unconstrained unitary 2
ρ Σ1 , 0 Q11 Q12
precoder for H is simply given by Fopt = V1 . Further
= log2 I +
note that the precoder V1 cannot in general be expressed as Ns σn2 0 Σ22 Q21 Q22

FRF FBB with FRF ∈ FRF , and thus cannot be realized in (a) ρ
= log2 INs + Σ2 Q11
the mmWave architecture of interest. If the hybrid precoder Ns σn2 1

FRF FBB can be made sufficiently “close” to the optimal
ρ
precoder V1 , however, the mutual information resulting from + log2 I + Σ2 Q22
Ns σn2 2
Fopt and FRF FBB can be made comparable. In fact, to
−1
simplify the forthcoming treatment of I(FRF , FBB ), we make ρ2 2 ρ 2 2
− 2 4 Σ2 Q21 INs + Σ Q 11 Σ Q
1 12
the following system assumption. Ns σn Ns σn2 1

Approximation 1: We assume that the mmWave system (b) ρ
parameters (Nt , Nr , NtRF , NrRF ), as well as the parameters of ≈ log2 INs + Σ2 V∗ FRF FBB F∗BB F∗RF V1 ,
Ns σn2 1 1
the mmWave propagation channel (Ncl , Nray , . . .), are such (12)
that the hybrid precoders FRF FBB can be made sufficiently
“close” to the optimal unitary precoder Fopt = V1 . Math- where (a) is a result of using the Schur complement iden-
ematically, this “closeness” is defined by the following two tity for matrix determinants and (b) follows from invoking
equivalent approximations: Approximation 1 which implies that Q12 , Q21 and Q22 are
approximately zero. Using (12), mutual information can be
1) The eigenvalues of the matrix INs − further simplified by writing
V1∗ FRF FBB F∗BB F∗BB V1 are small. In the case of I(FRF , FBB )
mmWave precoding, this can be equivalently stated as
(a) ρ
V1∗ FRF FBB ≈ INs .3
≈ log2 INs + Σ 2
2) The singular values of the matrix V2∗ FRF FBB are small; Ns σn2 1
−1
alternatively V2∗ FRF FBB ≈ 0.
ρ 2
+ log2 INs − INs + Σ
Ns σn2 1
This approximation is similar to the high-resolution approx-

imation used to simplify the analysis of limited feedback ρ 2 ∗ ∗ ∗
× Σ (I −V F F F F V
1 RF BB BB RF 1 )
Ns σn2 1
Ns
MIMO systems by assuming that codebooks are large enough
such that they contain codewords that are sufficiently close
(b) ρ
to the optimal unquantized precoder [54]. In the case of ≈ log2 INs + Σ2
Ns σn2 1
mmWave precoding, this approximation is expected to be −1
tight in systems of interest which include: (i) a reasonably ρ 2
− tr INs + Σ
large number of antennas Nt , (ii) a number of transmit chains Ns σn2 1
Ns < NtRF ≤ Nt , and (iii) correlated channel matrices H.
ρ 2 ∗ ∗ ∗
Functionally, Approximation 1 allows us to further simplify × Σ (INs − V1 FRF FBB FBB FRF V1 )
Ns σn2 1
the mutual information I(FRF , FBB ). To do so, we use the
(c) ρ
partitions defined in (11) and further define the following
≈ log2 INs + 2
Σ1 (13)
Ns σ 2
n
− tr (INs − V1∗ FRF FBB F∗BB F∗RF V1 )

ρ
3 For the eigenvalues of I ∗ ∗ ∗
Ns − V1 FRF FBB FBB FBB V1 to be small, we = log2 INs + 2
Σ 2 ∗ 2
1 − Ns − V1 FRF FBB F ,
need V1∗ FRF FBB ≈ Ψ where Ψ is any Ns ×Ns unitary matrix (not neces- Ns σn
sarily INs ). However, if FRF FBB is a valid precoder with V1∗ FRF FBB ≈ (14)
Ψ, then so is the rotated precoder FRF F BB = FRF FBB Ψ∗ for which
we have V1∗ FRF F BB ≈ IN . Since FBB can be arbitrarily rotated, the where (a) is exact given (12) and can be obtained by defining
s
conditions V1∗ FRF FBB F∗BB F∗BB V1 ≈ INs and V1∗ FRF FBB ≈ INs can the matrices B = Nsρσ2 Σ21 and A = V1∗ FRF FBB F∗BB F∗RF V1
n
be considered equivalent in this case without loss of generality.
and noting that I + BA = (I + B)(I − (I + B)−1 B(I − 2) Structure of clustered mmWave channels: Examining the
A)). The simplification in (b) follows from Approximation channel model in (4), we note that the array response
1 which implies that the eigenvalues of the matrix X = vectors at (φti , θi t t
), ∀i, , θi ) also form a finite span-
(INs + Nsρσ2 Σ21 )−1 Nsρσ2 Σ21 (INs − V1∗ FRF FBB F∗BB F∗RF V1 ) ning set for the channel’s row space. In fact, when
n n
are small and thus allows us to use the following approxima- Ncl Nray ≤ Nt , we note that the array response vectors
tion log2 |INs − X| ≈ log2 (1 − tr(X)) ≈ −tr(X). Finally at (φti , θi
t
) will be linearly independent with probabil-
(c) follows from adopting a high effective-SNR approximation ity one and will thus form another minimal basis for the
which implies that (I + Nsρσ2 Σ21 )−1 Nsρσ2 Σ21 ≈ INs and yields channel’s row space when Ncl Nray ≤ min(Nt , Nr ).
n n
the final result in (14).4 We notice that the first term in (14) Note: To establish the linear independence of the vectors
is the mutual information achieved by the optimal precoder at (φti , θi
t
), consider the case of uniform linear arrays.
Fopt = V1 and that the dependence of I(FRF , FBB ) on the When ULAs are considered, the Nt × Ncl Nray matrix
hybrid precoder FRF FBB is now captured in the second and formed by the collection of vectors at (φti ) ∀i, will
final term of (13) and (14). be a Vandermonde matrix which has full rank whenever
In cases where FRF FBB is made exactly unitary, we note the angles φti are distinct. This event occurs with prob-
that the second term in (13) and (14) is nothing but the squared ability one when φti are generated from a continuous
chordal distance between the two points Fopt = V1 and distribution. Linear independence can be established in
FRF FBB on the Grassmann manifold. Since Approximation the case of UPAs by writing their response vectors as a
1 states the these two points are “close”, we can exploit the Kronecker product of two ULA response vectors [19].
manifold’s locally Euclidean property to replace the chordal 3) Connection between Fopt and at (φti θi t
): Regardless
distance by the Euclidean distance Fopt − FRF FBB F [66]. of whether Ncl Nray ≤ Nt or not, observation 1 implies
Therefore, near-optimal hybrid precoders that approximately that the columns of the optimal precoder Fopt = V1
maximize I(FRF , FBB ) can be found by instead minimizing are related to the vectors at (φti , θi t
) through a linear
Fopt −FRF FBB F . In fact, even without treating FRF FBB as transformation. As a result, the columns of Fopt can be
a point on the Grassmann manifold, Approximation 1 implies written as linear combinations of at (φti , θi t
), ∀i, .
that V1∗ FRF FBB 2F , and consequently (14), can be approx- 4) Vectors at (φti θi t
) as columns of FRF : Recall that the
imately maximized by instead maximizing tr (V1∗ FRF FBB ).5 vectors at (φti , θi t
) are constant-magnitude phase-only
Since maximizing tr (V1∗ FRF FBB ) is again equivalent to vectors which can be applied at RF using analog phase
minimizing Fopt −FRF FBB F , the precoder design problem shifters. Therefore, the mmWave transmitter can apply
can be rewritten as NtRF of the vectors at (φti , θi t
) at RF (via the RF
precoder FRF ), and form arbitrary linear combinations
(Fopt opt
RF , FBB ) = arg min Fopt − FRF FBB F ,
FBB ,FRF of them using its digital precoder FBB . Namely, it
s.t. FRF ∈ FRF , (15) can construct the linear combination that minimizes
Fopt − FRF FBB F .
FRF FBB 2F = Ns ,
which can now be summarized as finding the projection of Therefore, by exploiting the structure of H, we notice
Fopt onto the set of hybrid precoders of the form FRF FBB that near-optimal hybrid precoders can be found by further
with FRF ∈ FRF . Further, this projection is defined with restricting FRF to be the set of vectors of the form at (φti , θi
t
)
respect to the standard Frobenius norm · 2F . Unfortunately, and solving
the complex non-convex nature of the feasible set FRF makes
finding such a projection both analytically (in closed form) (Fopt opt
RF , FBB ) = arg min Fopt − FRF FBB F ,
(i)
and algorithmically intractable [69]–[72]. s.t. FRF ∈ at (φti , θi
t
), ∀i, , (16)
To provide near-optimal solutions to the problem in (15),
FRF FBB 2F = Ns ,
we propose to exploit the structure of the mmWave MIMO
channels generated by the clustered channel model in Section which amounts to finding the best low dimensional rep-
II-B. Namely, we leverage the following observations on resentation of Fopt using the basis vectors at (φti , θi, t
).
mmWave precoding: We note here that the set of basis vectors can be ex-
1) Structure of optimal precoder: Recall that the optimal tended to include array response vectors at (·, ·) in direc-
unitary precoder is Fopt = V1 , and that the columns of tions other than {(φti , θi t
)| 1 ≤ i ≤ Ncl , 1 ≤ ≤
the unitary matrix V form an orthonormal basis for the Nray }, though the effect of this basis extension is typically
channel’s row space. negligible. Similarly, in cases where Ncl Nray > Nt , and
{at (φti , θi
t
)| 1 ≤ i ≤ Ncl , 1 ≤ ≤ Nray } forms and over-
4 Note here that it is not the nominal SNR N ρσ 2 that is assumed to be complete representation of the channel’s right singular space,
s n
high. This would be a problematic assumption in mmWave systems. It is, it is possible to reduce the redundancy in the spanning set for
however, only the effective-SNRs in the channel’s dominant Ns subspaces example by using the orthogonal steering vectors leveraged in
that are assumed to be sufficiently high. This is a reasonable assumption since
these effective SNRs include the large array gain from mmWave beamforming. [33]–[37] for which observations 2-4 also hold. In any case,
5 This is since the magnitude of V∗ F
1 RF FBB ’s off-diagonal entries is the precoding problem consists of selecting the “best” NtRF
negligible and all V1∗ FRF FBB ’s diagonals must be made close to one. array response vectors and finding their optimal baseband
Thus V1∗ FRF FBB 2F , i.e., the 2 norm of V1∗FRF FBB ’s diagonals, can (i)
be maximized by optimizing tr V1∗ FRF FBB , i.e., the 1 norm of the combination. Finally, we note that the constraint of FRF can
diagonals [43], [67], [68]. be embedded directly into the optimization objective to obtain
Algorithm 1 Spatially Sparse Precoding via Orthogonal sparse approximation problem [45]–[48]. So, for the general
Matching Pursuit case of Ns ≥ 1, we present an algorithmic solution based on
Require: Fopt the well-known concept of orthogonal matching pursuit [39],
1: FRF = Empty Matrix [41], [49]. The pseudo-code for the precoder solution is given
2: Fres = Fopt in Algorithm 1. In summary, the precoding algorithm starts
3: for i ≤ NtRF do by finding the vector at (φti , θi
t
) along which the optimal
4: Ψ = A∗t Fres precoder has the maximum projection. It then appends the
5: k = argmax=1, ..., Ncl Nray (ΨΨ∗ ), selected column vector at (φti , θi
t
) to the RF precoder FRF .
(k) After the dominant vector is found, and the least squares
6: FRF = FRF |At
−1 solution to FBB is calculated in step 7, the contribution of the
7: FBB = (F∗RF FRF ) F∗RF Fopt selected vector is removed in step 8 and the algorithm proceeds
Fopt −FRF FBB
8: Fres = Fopt −FRF FBB F to find the column along which the “residual precoding
9: end for √ matrix” Fres has the largest projection. The process continues
10: FBB = Ns FRFFFBBBB F until all NtRF beamforming vectors have been selected. At
11: return FRF , FBB the end of the NtRF iterations, the algorithm would have (i)
constructed an Nt × NtRF RF precoding matrix FRF , and
(ii) found the optimal NtRF × Ns baseband precoder FBB
the following equivalent problem which minimizes Fopt − FRF FBB 2F . Step 10 ensures that
the transmit power constraint is exactly satisfied.
opt = arg min Fopt − At F
F BB F ,
BB
BB
To gain more intuition about the proposed precoding frame-
F
work, Fig. 2 plots the beam patterns generated by a transmitter
s.t. diag(F ∗ )0 = N RF ,
BB F (17)
BB t with a 256-element planar array for an example channel real-
2
At FBB = Ns , ization with 6 rays (or equivalently 6 clusters with an angular
F
spread of 0) using (i) the channel’s optimal unconstrained
where At = at (φt1,1 , θ1,1
t
), . . . , at (φtNcl ,Nray , θN
t
cl ,Nray
) is precoder, (ii) the proposed precoding strategy with NtRF = 4,
an Nt × Ncl Nray matrix of array response vectors and F BB is and (iii) the beam steering vector in the channel’s dominant

an Ncl Nray ×Ns matrix. The matrices At and FBB act as aux- physical direction. We observe that in practical mmWave
iliary variables from which we obtain Fopt opt
RF and FBB respec-
channels, optimal precoders do in fact generate spatially sparse
tively. Namely, the sparsity constraint diag(F BB F ∗ )0 = beam patterns and thus may be accurately approximated by a
BB
RF RF finite combination of array response vectors. Further, Fig. 2 in-
Nt states that FBB cannot have more than Nt non-zero
BB are non zero, only NtRF dicates that Algorithm 1 succeeds in generating beam patterns
rows. When only NtRF rows of F
which closely resemble those generated by Fopt . Therefore,
columns of the matrix At are effectively “selected”. As a
Algorithm 1 succeeds in selecting the best NtRF steering
result, the baseband precoder Fopt BB will be given by the Nt
RF
opt opt directions and forming appropriate linear combinations of the
non-zero rows of FBB and the RF precoder FRF will be given
selected response vectors. This beam pattern similarity will
by the corresponding NtRF columns of At .
ultimately result in favorable spectral efficiency performance
Essentially, we have reformulated the problem of jointly
as shown in Section VI.
designing FRF and FBB into a sparsity constrained matrix
Having presented the proposed precoding framework, we
reconstruction problem with one variable. Although the un- conclude this section with the following design remarks.
derlying motivation differs, and so does the interpretation of
Remark 2: We note that the mmWave terminals need not
the different variables involved in (17), the resulting problem
know the exact angles (φti , θit
) that make up the channel
formulation is identical to the optimization problem encoun-
matrix H, and need not use the matrix At as defined earlier.
tered in the literature on sparse signal recovery. Thus, the
We have only used this finite basis for simplicity of exposition.
extensive literature on sparse reconstruction can now be used
In general, the mmWave terminals can instead select basis
for hybrid precoder design [39], [41]. To see this more clearly,
vectors of the form at (φ, θ) using any finite set of representa-
note that in the simplest case of single stream beamforming,
tive azimuth and elevation directions (such as a set of equally
(17) simplifies to
spaced angles for example). This approach avoids having to
opt
fBB = arg min fopt − At
fBB F , decompose H into its geometric representation and is naturally

fBB (18) suited for limited feedback operation. This approach will be
s.t.
fBB 0 = NtRF , At
fBB 2F = Ns , discussed further in Section V.
Remark 3: It may be advantageous in some cases to im-
in which the sparsity constraint is now on the vector
fBB . This pose the additional constraint that FBB be unitary. Unitary
beamforming problem can be solved, for example, by relaxing precoders can be more efficiently quantized and are thus more
the sparsity constraint and using convex optimization to solve attractive in limited feedback systems. With this additional
its 2 − 1 relaxation. Alternatively, (18) can be solved using constraint, (17) can be solved again via Algorithm 1 by replac-
tools from [39]–[42], [49]. ing the least squares solution for FBB in step 7, by the solution
In the more general case of Ns > 1, the problem in (17) is to the corresponding orthogonal Procrustes problem [73]. This
equivalent to the problem of sparse signal recovery with mul- is given by FBB = ÛV̂∗ where Û and V̂ are unitary matrices
tiple measurement vectors, also known as the simultaneously defined by the singular value decomposition of F∗RF Fopt , i.e.,
(a) Beam Pattern of Optimal Beamforming Vector (b) Beam Pattern with Proposed Solution
(c) Beam Pattern of Optimal Steering Vector
Fig. 2. Beam pattern generated a 256-element square array in an example channel realization with 6 rays (or equivalently 6 clusters with 0 angular spread)
using (a) optimal unconstrained beamforming, (b) the proposed sparse precoding solution with 4 RF chains, and (c) the beam steering vector in the channel’s
dominant physical direction. The proposed algorithm is shown to result in beam patterns that closely resemble the patterns generated by optimal beamforming;
this beam pattern similarity will ultimately result in similar spectral efficiency. For illustration purposes, the channel’s angle spread is set to 0◦ in this figure.
F∗RF Fopt = ÛΣ̂V̂∗ [73]. received signals must be linearly combined in the analog
Remark 4: In the limit of large antenna arrays (Nt , Nr → domain before any detection or decoding is performed.
∞) in very poor scattering environments for which Ncl Nray =
o(min(Nt , Nr )), the results of [19] indicate that simple RF-
only beam steering becomes optimal, i.e., it becomes optimal
to simply transmit each stream along one of the Ns most In this section, we address the problem of designing linear
dominant vectors at (φti , θi
t
). For arrays of practical sizes, combiners for the mmWave receiver in Fig. 1, which uses both
however, Section VI shows that there can be significant gains analog and digital processing before detection. Assuming the
from more involved precoding strategies such as the one hybrid precoders FRF FBB are fixed, we seek to design hybrid
presented in this section. combiners WRF WBB that minimize the mean-squared-error
(MSE) between the transmitted and processed received sig-
IV. P RACTICAL M ILLIMETER WAVE R ECEIVER D ESIGN
nals. The combiner design problem can therefore be stated
In Section III, we abstracted receiver-side processing and as
focused on designing practical mmWave precoders that max-
opt opt ∗ ∗ 2
imize mutual information. Effectively, we assumed that the (WRF , WBB ) = arg min E ||s − WBB WRF y||2 ,
WRF , WBB
mmWave receiver can optimally decode data using its Nr -
s.t. WRF ∈ WRF ,
dimensional received signal. Such a decoder can be of pro-
(19)
hibitively high complexity in multi-antenna systems, making
lower-complexity receivers such as the commonly used linear where WRF is the set of feasible RF combiners, i.e., WRF is
MMSE receiver more appealing for practical implementation. the set of Nr × NrRF matrices with constant-gain phase-only
In fact, in mmWave architectures such as the one shown in entries. In the absence of any hardware limitations that restrict
Fig. 1, such optimal decoders are impossible to realize since the set of feasible linear receivers, the exact solution to (19)
is well known [74] to be equivalent to finding hybrid combiners that solve

∗ opt opt
WMMSE (WRF , WBB )=
−1
= E [sy ] E [yy∗ ]
∗ 1
arg min E [yy∗ ] 2(WMMSE −WRF WBB ) F (23)
√ −1 WRF , WBB
ρ ∗ ∗ ∗ ρ ∗ ∗ ∗ 2
= F F H HFRF FBB FBB FRF H + σn INr s.t. WRF ∈ WRF ,
Ns BB RF Ns
−1 which amounts to finding the projection of the unconstrained
(a) 1 σ 2 Ns
= √ F∗BB F∗RF H∗ HFRF FBB + n INs F∗BB F∗RF H∗ , MMSE combiner WMMSE onto the set of hybrid combiners of
ρ ρ
the form WRF WBB with WRF ∈ WRF . Thus, the design of
(20)
MMSE receivers for the mmWave system of interest closely
where (a) follows from applying the matrix inversion lemma. resembles the design of its hybrid precoders. Unlike in the
Just as in the precoding case, however, this optimal uncon- precoding case however, the projection now is not with respect
∗
strained MMSE combiner WMMSE need not be decomposable to the standard norm ·2F and is instead an E [yy∗ ]-weighted
∗ ∗
into a product of RF and baseband combiners WBB WRF with Frobenius norm. Unfortunately, as in the case of the precoding
∗
WRF ∈ WRF . Therefore WMMSE cannot be realized in the problem in (15), the non-convex constraint on WRF precludes
system of Fig. 1. Further, just as in the precoding case, the us from practically solving the projection problem in (23). The
complex non-convex constraint WRF ∈ WRF makes solving same observations that allowed us to leverage the structure of
(19) analytically impossible and algorithmically non-trivial. To mmWave channels to solve the precoding problem in Section
overcome this difficulty, we leverage the methodology used III, however, can be translated to the receiver side to solve the
in [50], [51] to find linear MMSE estimators with complex combiner problem as well. Namely, because of the structure
structural constraints. of clustered mmWave channels, near-optimal receivers can be
found by further constraining WRF to have columns of the
form ar (φ, θ) and instead solving
We start by reformulating the problem in (19) by expanding opt =
W BB
MSE as follows
∗ ∗
arg min E [yy∗ ]
1/2
WMMSE − E [yy∗ ]
1/2 BB F ,
Ar W
E s − WBB WRF y22 BB
W
∗ ∗
= E [(s − WBB WRF y)∗ (s − WBB
∗
WRF∗
y)] ∗ )0 = N RF
BB W
s.t. diag(W BB r
∗ ∗ ∗ ∗
= E [tr ((s − WBB WRF y) (s − WBB WRF y)∗ )] (21) (24)
= tr (E [ss∗ ]) − 2 {tr (E [sy∗ ] WRF WBB )}
∗ ∗ where Ar = ar (φr1,1 , θ1,1
r
), . . . , at (φrNcl ,Nray , θN
r
) is
+ tr (WBB WRF E [yy∗ ] WRF WBB ) . cl ,Nray
an Nr × Ncl Nray matrix of array response vectors and W BB

We now note that since the optimization problem in (19) BB
is an Ncl Nray × Ns matrix; the quantities Ar and W
is over the variables WRF and WBB , we can add any
act as auxiliary variables from which we obtain WRF and
term that is independent of WRF and WBB to its ob-
WBB in a manner similar to Section III.6 As a result, the
jective function without changing the outcome of the op-
MMSE estimation problem is again equivalent to the problem
timization. Thus, we choose to add the constant term
∗ of sparse signal recovery with multiple measurement vectors
tr (WMMSE E [yy∗ ] WMMSE ) − tr (E [ss∗ ]) and minimize the
and can thus be solved via the orthogonal matching pursuit
equivalent objective function
concept used in Section III. For completeness the pseudo code
J (WRF , WBB ) is given in Algorithm 2.
∗
= tr (WMMSE E [yy∗ ] WMMSE) Remark 5: This section relaxes the perfect-receiver as-
sumption of Section III and proposes practical methods to
− 2 {tr (E [sy∗ ] WRF WBB )}
∗ ∗
find low-complexity linear receivers. The design of precoders
+ tr (WBB WRF E [yy∗ ] WRF WBB ) and combiners, however, remains decoupled as we have as-
(a) ∗
= tr (WMMSE E [yy∗ ] WMMSE ) sumed that the precoders FRF FBB are fixed while designing
∗ WRF WBB (and that receivers are optimal while design-
− 2 {tr (WMMSE E [yy∗ ] WRF WBB )} ing FRF FBB ). This decoupled approach simplifies mmWave
∗ ∗
+ tr (WBB WRF E [yy∗ ] WRF WBB ) transceiver design, and will be shown to perform well in
∗ ∗ ∗
= tr ((WMMSE − WBB WRF ) E [yy∗ ] Section VI, however, some simple “joint decisions” may be

∗
× (WMMSE ∗
− WBB ∗
WRF )∗ both practical and beneficial. For example, consider the case
where a receiver only has a single RF chain and thus is
= E [yy∗ ]1/2 (WMMSE − WRF WBB ) 2F , (22) restricted to applying a single response vector ar (φ, θ). In
where (a) follows from rewriting the sec- such a situation, designing FRF FBB to radiate power in NtRF
ond term as tr (E [sy∗ ] WRF WBB ) = different directions may lead to a loss in actual received power
−1 (since the receiver can only form a beam in one direction). As
tr(E [sy∗ ] E [yy∗ ] E [yy∗ ] WRF WBB ) and using the
∗ −1
fact that WMMSE = E [sy∗ ] E [yy∗ ] which implies that 6 As noted in Section III the receiver need not know the exact angles
tr (E [sy ] WRF WBB ) = tr (WMMSE E [yy∗ ] WRF WBB ).
∗ ∗
(φri , θi
r ) and can instead use any set of representative azimuth and elevation
As a result of (22), the MMSE estimation problem is angles of arrival to construct the matrix of basis vectors Ar .
Algorithm 2 Spatially Sparse MMSE Combining via Orthog- A. Quantizing the RF Precoder
onal Matching Pursuit
Recall that the precoder FRF calculated Section III has
Require: WMMSE NtRF columns of the form at (φ, θ). Therefore, FRF admits
1: WRF = Empty Matrix a natural parametrization in terms of the NtRF azimuth and
2: Wres = WMMSE elevation angles that it uses. Thus, FRF can be efficiently
3: for i ≤ NrRF do encoded by quantizing its 2NtRF free variables. For simplicity,
4: Ψ = A∗r E [yy∗ ] Wres we propose to uniformly quantize the NtRF azimuth and
∗
5: k = arg max
=1, ..., N cl Nray (ΨΨ ), elevation angles using Nφ and Nθ bits respectively. Therefore,
(k)
6: WRF = WRF |Ar the quantized azimuth and elevation angles are such that
−1
7: WBB = (WRF∗
E [yy∗ ] WRF ) ∗
WRF E [yy∗ ]WMMSE t φtrange t 3φtrange t φtrange
WMMSE −WRF WBB Cφ = φmin + N +1 , φmin + N +1 , . . . , φmax − N +1
8: Wres = W 2 φ 2 φ 2 φ
MMSE −WRF WBB F t t t
9: end for t θ range t 3θ range t θ range
Cθ = θmin + N +1 , θmin + N +1 , . . . , θmax − N +1
10: return WRF , WBB 2 θ 2 θ 2 θ
(26)
where we recall that [φtmin , φtmax ] and [θmin t t
, θmax ] are
a result, it is beneficial to account for the limitations of the the sectors over which Λt (φ, θ) = 0. Further, for some
more-constrained terminal when designing either precoders or configurations of ([φtmin , φtmax ], [θmin
t t
, θmax ], Nφ , Nθ ), one
combiners. To do so, we propose to run Algorithms 1 and 2 may be able to further constrain the angles in Cφ and Cθ to
in succession according to the following rules correspond to orthogonal beams as in [33]–[36]. The receiver
⎧ can then quantize FRF by simply selecting the entries of Cφ
⎨ 1. Solve for FRF FBB using Algorithm 1. and Cθ that are closest in Euclidean distance to FRF ’s angles.
NtRF < NrRF 2. Given FRF FBB , solve for WRF WBB
⎩ Alternatively, as stated in Remark 2, Algorithm 1 can be run
using Algorithm 2. directly using the Nt × 2Nφ +Nθ matrix of “quantized response
⎧
⎪
⎪ 1. Solve for WRF WBB using Algorithm 2 vectors”
⎨
RF RF assuming FRF FBB = Fopt .
Nt > Nr Aquant = at (φt1 , θ1t ), . . . , at (φti , θt ), . . . , at (φt2Nφ , θ2t Nθ ) ,
⎪
⎪ 2. Solve for FRF FBB for the effective t
⎩ ∗ ∗ (27)
channel WBB WRF H.
and the index of the selected angles can be fed back to
(25)
the transmitter. While this latter approach has higher search
In summary, starting with the more constrained side, the complexity, it has the advantage of (i) “jointly quantizing” all
hybrid precoder or combiner is found using Algorithm 1 or 2NtRF angles, and (ii) automatically matching the baseband
2. Then, given the output, the remaining processing matrix precoder FBB to the quantized angles.
is found by appropriately updating the effective mmWave
channel. B. Quantizing the Baseband Precoder
Finally, we note that while the numerical results of Sec- To efficiently quantize FBB , we begin by highlighting
tion VI indicate that this decoupled approach to mmWave its mathematical structure in mmWave systems of interest.
transceiver design yields near-optimal spectral efficiency, a Namely, we note that for systems with large antenna arrays,
more direct joint optimization of (FRF , FBB , WRF , WBB ) is we typically have that F∗RF FRF ≈ INtRF . When coupled with
an interesting topic for future investigation. Similarly, while Approximation 1, we have that F∗BB FBB ≈ INs , i.e., FBB
we have solved the sparse formulation of the precoding and is approximately unitary. In fact, FBB can be made exactly
combining problems via orthogonal matching pursuit, the unitary as discussed in Remark 3. Further, we recall that the
problems in (17) and (24) can be solved by leveraging other spectral efficiency expression in (3) is invariant to Ns × Ns
algorithms for simultaneously sparse approximation [47]. unitary transformations of the baseband precoder. Therefore,
FBB is a subspace quantity that can be quantized on the
Grassmann manifold [52], [53]. Suitable codebooks for FBB
V. L IMITED F EEDBACK S PATIALLY S PARSE P RECODING
can be designed using Lloyd’s algorithm on a training set
Section III implicitly assumed that the transmitter has per- of baseband precoders and using the chordal distance as a
fect knowledge of H and is thus able to calculate Fopt . Since distance measure [75]. Since such codebook construction is
such transmitter channel knowledge may not be available in well-studied in the literature on limited feedback MIMO, we
practical systems, we propose to fulfill this channel knowledge omit its details for brevity and refer the reader to [76, Section
requirement via limited feedback [20], [52]–[54]. Namely, we IV] for an in-depth description of the process.
assume that the receiver (i) acquires perfect knowledge of H,
(ii) calculates Fopt and a corresponding hybrid approximation VI. S IMULATION R ESULTS
FRF FBB , then (iii) feeds back information about FRF FBB In this section, we present simulation results to demonstrate
to the transmitter. Since hybrid precoders are naturally de- the performance of the spatially sparse precoding algorithm
composed into an RF and baseband component, we propose presented in Section III when combined with the sparse
to quantize FRF and FBB separately while exploiting the MMSE combining solution presented in Section IV. We model
mathematical structure present in each of them. the propagation environment as a Ncl = 8 cluster environment

with Nray = 10 rays per cluster with Laplacian distributed Optimal Unconstrained Precoding
azimuth and elevation angles of arrival and departure [29], Beam Steering
Ns
[61]. For simplicity of exposition, we assume all clusters are Sparse Precoding & Combining
2
of equal power, i.e., σα,i = σα2 ∀i, and that the angle spread
Spectral Efficiency (bits/s/Hz)

at both the transmitter and receiver are equal in the azimuth
and elevation domain, i.e., σφt = σφr = σθt = σθr . Since
outdoor deployments are likely to use sectorized transmitters
to decrease interference and increase beamforming gain, we
consider arrays of directional antenna elements with a re-
sponse given in (5) [8], [9]. The transmitter’s sector angle is
assumed to be 60◦ -wide in the azimuth domain and 20◦ -wide
in elevation [8]. In contrast, we assume that the receivers have Ns
relatively smaller antenna arrays of omni-directional elements;
this is since receivers must be able to steer beams in any
direction since their location and orientation in real systems
is random. The inter-element spacing d is assumed to be half-
í í í í í í í í
wavelength. We compare the performance of the proposed SNR (dB)
strategy to optimal unconstrained precoding in which streams
are sent along the channel’s dominant eigenmodes. When Fig. 3. Spectral Efficiency achieved by various precoding solutions for a
perfect channel knowledge is available to the transmitter, we 64 × 16 mmWave system with planar arrays at the transmitter and receiver.
The propagation medium is a Ncl = 8 cluster environment with Nray = 10
use the channel’s actual angles of departure (φti , θi
t
) as input and an angular spread of 7.5◦ . Four RF chains are assumed to be available
to Algorithm 1, i.e., the algorithm is implemented as described for sparse precoding and MMSE combining.
in the main discussion Section III. We also compare with
a simple beam steering solution in which data streams are
steered onto the channel’s best propagation paths and received Optimal Unconstrained Precoding
Beam Steering
along the best corresponding receive paths [17], [19].7 For Sparse Precoding & Combining Ns
fairness, the same total power constraint is enforced on all
precoding solutions and signal-to-noise ratio is defined as
SNR = σρ2 . Finally, all reported results are averaged over

n
5000 random channel realizations.
Fig. 3 shows the spectral efficiency achieved in a 64 × 16
system with square planar arrays at both transmitter and
receiver. For the proposed precoding strategy, both transmitter
and receiver are assumed to have four transceiver chains with
which they transmit Ns = 1 or 2 streams. Fig. 3 shows Ns
that the proposed framework achieves spectral efficiencies

that are essentially equal to those achieved by the optimal
unconstrained solution in the case Ns = 1 and are within a
small gap from optimality in the case of Ns = 2. This implies
that the proposed strategy can very accurately approximate
the channel’s dominant singular vectors as a combination of SNR (dB)
four steering vectors. When compared to traditional beam
Fig. 4. Spectral Efficiency achieved in a 256 × 64 mmWave system with
steering, Fig. 3 shows that there is a non-negligible improve- planar arrays at the transmitter and receiver. Channel parameters are set as in
ment to be had from more sophisticated precoding strategies Fig. 3. Six RF chains are available for sparse precoding and combining.
in mmWave systems with practical array sizes. To explore
performance in mmWave systems with larger antenna arrays,
Fig. 4 plots the performance achieved in a 256 × 64 system 4, the proposed solution still outperforms beam steering by
with NtRF = NrRF = 6 RF chains. Fig. 4 shows that approximately 5 dB in this larger mmWave system.
the proposed precoding/combining solution achieves almost- While Section III focused on the design of fixed-rank
perfect performance in both Ns = 1 and Ns = 2 cases. precoders with equal power allocation across streams, the
Further, we note that although beam steering is expected to be same framework can be applied to systems in which Ns is
optimal in the limit of large arrays, as discussed in Remark determined dynamically and streams are sent with unequal
power. This configuration allows us to compare the rates
7 Note that, when N > 1, the best propagation paths in terms of spectral
s achieved by the proposed precoding/combining framework
efficiency may not be the ones with the highest gains. This is since, with no
receiver baseband processing, different paths must be sufficiently separated to the mmWave channel’s waterfilling capacity. To do so,
so as they do not interfere. In this case, the best paths for transmission and Algorithm 1 is simply set to approximate Fopt = VΓ where
reception are chosen to maximize mutual information via a costly exhaustive Γ is a diagonal matrix resulting from the waterfilling power
search over a very fine-grain set of beam steering directions. Further, when
power allocation is considered in Fig. 5, the same waterfilling power allocation allocation. Fig. 5 demonstrates the performance achieved when
is applied to the beam steering solution. Algorithms 1 and 2 are used to approximate the channel’s
15
Waterfilling Capacity
Beam Steering 6 RF Cha
ins
Sparse Precoding & Combining
14
64×16, Ns=2
4 RF Ch
ains

13
12
256×64, Ns=1
4 RF Chain
s
11

10 Optimal Unconstrained Precoding

Sparse Precoding & Combining

9
64×16, Ns=1 4 RF Chains

SNR (dB) 8
0 5 10 15
Angle Spread (degrees)
Fig. 5. This figure compares the spectral efficiency achieved when rank
adaptation and unequal power allocation is allowed in 256 × 64 system with Fig. 6. Spectral Efficiency vs. Angle Spread in a number of different
NtRF = NrRF = 4. It is shown that sparse precoding and combining can mmWave system configuration at an SNR of 0 dB. For simplicity of ex-
approach the performance of an unconstrained capacity-achieving (waterfill- position, we assume that the angle spread is such that σφt = σr = σt = σr .
φ θ θ
ing) precoder. The figure also demonstrates large gains over a beam steering It is shown that as angle spread increased, and scattering becomes richer,
strategy in which streams are sent along different physical directions with a the performance of the proposed algorithm degrades. However, the rate gap
similar unequal power allocation. remains below 10% at a significant angle spread of 15◦ . For more reasonable
angle spreads of around 5◦ , the rate gap is negligible.
capacity-achieving precoders and combiners in a 256 × 64

mmWave system with NtRF = NrRF = 4. Fig. 5 shows 15
that the proposed framework allows systems to approach
channel capacity and provides large gains over simple beam 14
steering. Since the multiplexing gain of the mmWave system 64×16, Ns=2
is limited by Ns ≤ min{NtRF , NrRF }, capacity cannot be 13

approached at very high SNR when the optimal Ns exceeds
min{NtRF , NrRF }. Fig. 5 indicates, however, that even at an 12
SNR of 0 dB where we observe that Ns = 3 streams are 256×64, Ns=1
sent over most channel realizations, the proposed strategy is 11
still within a small gap from capacity. Finally, we note that Optimal Unconstrained Precoding (Full CSI)
although the derivation leading up to (14) does not account for 10 Sparse Precoding & Combining (Full CSI)
unequal power allocation across streams, Fig. 5 indicates that Sparse Precoding & Combining (Limited Feedback)
Algorithm 1 is nevertheless a sensible approach to designing 9
such precoders.
The proposed precoding/combining framework leverages 64×16, Ns=1
8
the mathematical structure of large mmWave channels with 2 3 4 5 6
Quantization Bits per Angle
relatively limited scattering. To examine performance in prop-
agation environments with varying levels of scattering, Fig. Fig. 7. Spectral Efficiency vs. Quantization Bits per Angle different mmWave
6 plots spectral efficiency as a function of the channel’s system configurations, all with NtRF = NrRF = 4, at an SNR of 0 dB in a
angle spread for a number of mmWave system configurations. channel with an azimuth and elevation angular spread of 7.5◦ . For simplicity
of exposition, we assume that Nφ = Nθ and an baseband precoder codebook
Fig. 6 indicates that when the angle spread is low, i.e., the of 4 bits in the Ns = 1 case and 6 bits in the Ns = 2 case. The figure
scattering is rather limited, the performance of the proposed indicates that for the considered array sizes, 3 bits per angle is often enough
algorithm is within a small gap from the performance of to achieve almost-perfect performance.
unconstrained precoding. As angle spread increases, the rates
achieved by the proposed solutions slowly degrade. However,
Fig 6 indicates that in the two Ns = 1 cases shown, the of RF chains at the mmWave terminals which enables them to
rate gap remains below 10% at a significant angle spread generate more flexible precoders/combiners. This can be seen
of 15◦ and is negligible for more reasonable angle spreads by examining the same 64×16 system with NtRF = NrRF = 6.
of around 5◦ . In the case of Ns > 1 with smaller arrays, Finally, we examine the performance of the proposed pre-
spectral efficiency degrades more rapidly with angle spread. coding strategy in systems without channel state information
This can be seen by examining the 64 × 16 system with at the transmitter. For this performance characterization, we
NtRF = NrRF = 4 and Ns = 2. If possible, the effect of assume that the receiver calculates FRF and FBB with full
increased scattering can be mitigated by increasing the number knowledge of the channel and feeds back their parameters
as described in Section V. We assume that the receiver uses [4] S. Yong and C. Chong, “An overview of multigigabit wireless through
four and six bits to quantize FBB in the case of Ns = 1 and millimeter wave technology: potentials and technical challenges,”
EURASIP J. Wireless Commun. Netw., vol. 2007, no. 1, pp. 50–50, 2007.
Ns = 2 respectively, and constructs codebooks as described [5] R. Daniels and R. W. Heath, Jr., “60 GHz wireless communications:
in Section V-B. The receiver uses a variable number of bits emerging requirements and design recommendations,” IEEE Veh. Tech-
to quantize the azimuth and elevation angles used in FRF . nol. Mag., vol. 2, no. 3, pp. 41–50, 2007.
[6] P. B. Papazian, G. A. Hufford, R. J. Achatz, and R. Hoffman, “Study
For simplicity of exposition, we assume that Nφ = Nθ . of the local multipoint distribution service radio channel,” IEEE Trans.
Fig. 7 indicates that similar performance can be expected in Broadcasting, vol. 43, no. 2, pp. 175–184, 1997.
limited feedback systems and that the performance degradation [7] C. Doan, S. Emami, D. Sobel, A. Niknejad, and R. Brodersen, “Design
due to quantization is limited. Namely, Fig. 7 indicates that considerations for 60 GHz CMOS radios,” IEEE Commun. Mag., vol. 42,
no. 12, pp. 132–140, 2004.
no more than 3 bits are needed to quantize each steering [8] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broad-
angle in practical systems, and even 2 bits yields almost- band systems,” IEEE Commun. Mag., vol. 49, no. 6, pp. 101–107, 2011.
perfect performance for a 64 × 16 systems with Ns = 1. [9] G. Hendrantoro, R. Bultitude, and D. Falconer, “Use of cell-site diversity
in millimeter-wave fixed cellular systems to combat the effects of rain
In general the number of bits needed to properly quantize attenuation,” IEEE J. Sel. Areas Commun., vol. 20, no. 3, pp. 602–614,
the steering angles grows slowly with array size since larger 2002.
arrays generate narrower beams and require finer steering. [10] E. Torkildson, C. Sheldon, U. Madhow, and M. Rodwell, “Millimeter-
Since beam width is inversely proportional to the antenna wave spatial multiplexing in an indoor environment,” in Proc. 2009
IEEE GLOBECOM Workshops, pp. 1–6.
array dimensions, a reasonable rule-of-thumb is to add 1 bit [11] E. Torkildson, B. Ananthasubramaniam, U. Madhow, and M. Rodwell,
per azimuth (or elevation) steering angle whenever the array’s “Millimeter-wave MIMO: wireless links at optical speeds,” in Proc.
width (or height) doubles. Fig. 7 is promising as it indicates 2006 Allerton Conf. Commun., Control Comput.
[12] A. Valdes-Garcia, S. T. Nicolson, J.-W. Lai, A. Natarajan, P.-Y. Chen,
that it takes no more than 20 bits to quantize a 64×1 precoder S. K. Reynolds, J.-H. C. Zhan, D. G. Kam, D. Liu, and B. Floyd, “A fully
and about 22 bits for a 64 × 2 precoder. When considering the integrated 16-element phased-array transmitter in SiGe BiCMOS for 60-
fact that practical mmWave systems will use twenty to fifty GHz communications,” IEEE J. Solid-State Circuits, vol. 45, no. 12, pp.
2757–2773, 2010.
times more antennas compared to traditional MIMO systems,
[13] S. Sanayei and A. Nosratinia, “Antenna selection in MIMO systems,”
which use about 4 to 6 bits of feedback [52], we see that IEEE Commun. Mag., vol. 42, no. 10, pp. 68–73, 2004.
exploiting spatial sparsity in precoding helps dramatically [14] A. F. Molisch, M. Z. Win, Y.-S. Choi, and J. H. Winters, “Capacity of
compress feedback and keep its overhead manageable. MIMO systems with antenna selection,” IEEE Trans. Wireless Commun.,
vol. 4, no. 4, pp. 1759–1772, 2005.
[15] A. Gorokhov, D. A. Gore, and A. J. Paulraj, “Receive antenna selection
VII. C ONCLUSION for MIMO spatial multiplexing: theory and algorithms,” IEEE Trans.
In this paper we considered single user precoding and Signal Process., vol. 51, no. 11, pp. 2796–2807, 2003.
[16] Z. Xu, S. Sfar, and R. S. Blum, “Analysis of MIMO systems with receive
combing in mmWave systems where traditional MIMO so- antenna selection in spatially correlated Rayleigh fading channels,” IEEE
lutions are made infeasible by the heavy reliance on RF Trans. Veh. Technol., vol. 58, no. 1, pp. 251–262, 2009.
precoding. By leveraging the structure of realistic mmWave [17] J. Wang, Z. Lan, C.-W. Pyo, T. Baykas, C.-S. Sum, M. A. Rahman,
J. Gao, R. Funada, F. Kojima, H. Harada, et al., “Beam codebook based
channels, we developed a low hardware-complexity precoding beamforming protocol for multi-Gbps millimeter-wave WPAN systems,”
solution. We formulated the problem of mmWave precoder IEEE J. Sel. Areas Commun., vol. 27, no. 8, pp. 1390–1399, 2009.
design as a sparsity-constrained signal recovery problem and [18] A. F. Molisch and X. Zhang, “FFT-based hybrid antenna selection
presented an algorithmic solution using orthogonal matching schemes for spatially correlated MIMO channels,” IEEE Commun. Lett.,
vol. 8, no. 1, pp. 36–38, 2004.
pursuit. We showed that the same framework can be applied [19] O. El Ayach, R. W. Heath, Jr., S. Abu-Surra, S. Rajagopal, and Z. Pi,
to the problem of designing practical MMSE combiners for “The capacity optimality of beam steering in large millimeter wave
mmWave systems. We showed that the proposed precoders MIMO systems,” in Proc. 2012 IEEE International Workshop Signal
Process. Advances Wireless Commun., pp. 100–104.
can be efficiently quantized and that the precoding strategy [20] D. Love and R. W. Heath, Jr., “Equal gain transmission in multiple-input
is well-suited for limited feedback systems. Finally, we pre- multiple-output wireless systems,” IEEE Trans. Commun., vol. 51, no. 7,
sented numerical results on the performance of spatially sparse pp. 1102–1110, July 2003.
[21] P. Sudarshan, N. Mehta, A. Molisch, and J. Zhang, “Channel statistics-
mmWave processing and showed that it allows systems to based RF pre-processing with antenna selection,” IEEE Trans. Wireless
approach their theoretical limits on spectral efficiency. Future Commun., vol. 5, no. 12, pp. 3501–3511, Dec. 2006.
work related to such mmWave precoding includes relaxing the [22] X. Zheng, Y. Xie, J. Li, and P. Stoica, “MIMO transmit beamforming
assumptions made throughout this paper such as (i) perfect under uniform elemental power constraint,” IEEE Trans. Signal Process.,
vol. 55, no. 11, pp. 5395–5406, 2007.
channel state information at the receiver, (ii) knowledge of [23] J. Nsenga, A. Bourdoux, and F. Horlin, “Mixed analog/digital beam-
the antenna array structure, and (iii) the specialization to forming for 60 GHz MIMO frequency selective channels,” in Proc. 2010
narrowband channels. IEEE International Conf. Commun., pp. 1–6.
[24] F. Gholam, J. Vı́a, and I. Santamarı́a, “Beamforming design for simpli-
fied analog antenna combining architectures,” IEEE Trans. Veh. Technol.,
R EFERENCES vol. 60, no. 5, pp. 2373–2378, 2011.
[1] A. Ghosh, R. Ratasuk, B. Mondal, N. Mangalvedhe, and T. Thomas, [25] Z. Pi, “Optimal transmitter beamforming with per-antenna power con-
“LTE-advanced: next-generation wireless broadband technology,” IEEE straints,” in Proc. 2012 IEEE International Conf. Commun., pp. 3779–
Wireless Commun., vol. 17, no. 3, pp. 10–22, 2010. 3784.
[2] D. Lopez-Perez, I. Guvenc, G. De La Roche, M. Kountouris, T. Q. Quek, [26] X. Zhang, A. Molisch, and S.-Y. Kung, “Variable-phase-shift-based RF-
and J. Zhang, “Enhanced intercell interference coordination challenges baseband codesign for MIMO antenna selection,” IEEE Trans. Signal
in heterogeneous networks,” IEEE Wireless Commun., vol. 18, no. 3, Process., vol. 53, no. 11, pp. 4091–4103, Nov. 2005.
pp. 22–30, 2011. [27] V. Venkateswaran and A. van der Veen, “Analog beamforming in
[3] A. Damnjanovic, J. Montojo, Y. Wei, T. Ji, T. Luo, M. Vajapeyam, MIMO communications with phase shift networks and online channel
T. Yoo, O. Song, and D. Malladi, “A survey on 3GPP heterogeneous estimation,” IEEE Trans. Signal Process., vol. 58, no. 8, pp. 4131–4143,
networks,” IEEE Wireless Commun., vol. 18, no. 3, pp. 10–21, 2011. 2010.
[28] P. Smulders and L. Correia, “Characterisation of propagation in 60 GHz [54] J. C. Roh and B. D. Rao, “Design and analysis of MIMO spatial
radio channels,” Electron. Commun. Eng. J., vol. 9, no. 2, pp. 73–80, multiplexing systems with quantized feedback,” IEEE Trans. Signal
1997. Process., vol. 54, no. 8, pp. 2874–2886, 2006.
[29] H. Xu, V. Kukshya, and T. Rappaport, “Spatial and temporal character- [55] W. U. Bajwa, J. Haupt, A. M. Sayeed, and R. Nowak, “Compressed
istics of 60-GHz indoor channels,” IEEE J. Sel. Areas Commun., vol. 20, channel sensing: a new approach to estimating sparse multipath chan-
no. 3, pp. 620–630, 2002. nels,” Proc. IEEE, vol. 98, no. 6, pp. 1058–1076, 2010.
[30] Q. Spencer, B. Jeffs, M. Jensen, and A. Swindlehurst, “Modeling the [56] D. Ramasamy, S. Venkateswaran, and U. Madhow, “Compressive adap-
statistical time and angle of arrival characteristics of an indoor multipath tation of large steerable arrays,” in Proc. 2012 Inf. Theory Applications
channel,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, pp. 347–360, Mar. Workshop, pp. 234–239.
2000. [57] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Jr., “Hybrid
[31] A. M. Sayeed, “Deconstructing multiantenna fading channels,” IEEE precoding for millimeter wave cellular systems with partial channel
Trans. Signal Process., vol. 50, no. 10, pp. 2563–2579, 2002. knowledge,” in 2013 Proc. Inf. Theory Applications Workshop.
[32] P. Xia, S.-K. Yong, J. Oh, and C. Ngo, “A practical SDMA protocol [58] A. Goldsmith, S. Jafar, N. Jindal, and S. Vishwanath, “Capacity limits
for 60 GHz millimeter wave communications,” in Proc. 2008 Asilomar of MIMO channels,” IEEE J. Sel. Areas Commun., vol. 21, no. 5, pp.
Conf. Signals, Syst. Comput., pp. 2019–2023. 684–702, 2003.
[59] IEEE 802.15 WPAN Millimeter Wave Alternative PHY Task Group 3c.
[33] A. Sayeed and N. Behdad, “Continuous aperture phased MIMO: ba-
Available: www.ieee802.org/15/pub/TG3c.html,Sept.,2011.
sic theory and applications,” in Proc. 2010 Allerton Conf. Commun.,
[60] V. Raghavan and A. M. Sayeed, “Sublinear capacity scaling laws for
Control, Comput., pp. 1196–1203.
sparse MIMO channels,” IEEE Trans. Inf. Theory, vol. 57, no. 1, pp.
[34] J. Brady, N. Behdad, and A. Sayeed, “Beamspace MIMO for millimeter- 345–364, 2011.
wave communications: system architecture, modeling, analysis, and [61] A. Forenza, D. Love, and R. W. Heath, Jr., “Simplified spatial correlation
measurements,” IEEE Trans. Antennas Propag., vol. 61, no. 7, pp. 3814– models for clustered MIMO channels with different array configura-
3827, 2013. tions,” IEEE Trans. Veh. Technol., vol. 56, no. 4, pp. 1924–1934, 2007.
[35] A. M. Sayeed and N. Behdad, “Continuous aperture phased MIMO: a [62] S. Singh, R. Mudumbai, and U. Madhow, “Interference analysis for
new architecture for optimum line-of-sight links,” in Proc. 2011 IEEE highly directional 60-GHz mesh networks: the case for rethinking
International Symp. Antennas Propagation, pp. 293–296. medium access control,” IEEE/ACM Trans. Netw., vol. 19, no. 5, pp.
[36] G. H. Song, J. Brady, and A. Sayeed, “Beamspace MIMO transceivers 1513–1527, 2011.
for low-complexity and near-optimal communication at mm-wave fre- [63] C. Balanis, Antenna Theory. Wiley, 1997.
quencies,” in Proc. 2013 IEEE International Conf. Acoustics, Speech, [64] D. P. Palomar, J. M. Cioffi, and M. A. Lagunas, “Joint Tx-Rx beam-
Signal Process. forming design for multicarrier MIMO channels: a unified framework
[37] A. Sayeed and J. Brady, “Beamspace MIMO for high-dimensional for convex optimization,” IEEE Trans. Signal Process., vol. 51, no. 9,
multiuser communication at millimeter-wave frequencies,” to appear in pp. 2381–2401, 2003.
Proc. 2013 IEEE Global Telecommun. Conf. [65] D. P. Palomar and M. Chiang, “A tutorial on decomposition methods
[38] I. Markovsky and S. Van Huffel, “Overview of total least-squares for network utility maximization,” IEEE J. Sel. Areas Commun., vol. 24,
methods,” Signal Process., vol. 87, no. 10, pp. 2283–2302, 2007. no. 8, pp. 1439–1451, 2006.
[39] J. Tropp and A. Gilbert, “Signal recovery from random measurements [66] J. M. Lee, Introduction to Smooth Manifolds. Springer, 2012, vol. 218.
via orthogonal matching pursuit,” IEEE Trans. Inf. Theory, vol. 53, [67] T. Figiel, J. Lindenstrauss, and V. D. Milman, “The dimension of almost
no. 12, pp. 4655–4666, Dec. 2007. spherical sections of convex bodies,” Acta Mathematica, vol. 139, no. 1,
[40] J. Tropp, “Greed is good: algorithmic results for sparse approximation,” pp. 53–94, 1977.
IEEE Trans. Inf. Theory, vol. 50, no. 10, pp. 2231–2242, 2004. [68] N. Kwak, “Principal component analysis based on l1-norm maximiza-
[41] L. Rebollo-Neira and D. Lowe, “Optimized orthogonal matching pursuit tion,” IEEE Trans. Pattern Analysis Machine Intelligence, vol. 30, no. 9,
approach,” IEEE Signal Process. Lett., vol. 9, no. 4, pp. 137–140, 2002. pp. 1672–1680, 2008.
[42] R. Tibshirani, “Regression shrinkage and selection via the lasso,” J. [69] J. A. Tropp, I. S. Dhillon, R. W. Heath, Jr., and T. Strohmer, “Designing
Royal Statistical Society. Series B (Methodological), vol. 58, no. 1, pp. structured tight frames via an alternating projection method,” IEEE
267–288, 1996. Trans. Inf. Theory, vol. 51, no. 1, pp. 188–209, 2005.
[43] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge Uni- [70] A. S. Lewis and J. Malick, “Alternating projections on manifolds,”
versity Press, 2004. Mathematics Operations Research, vol. 33, no. 1, pp. 216–234, 2008.
[71] R. Escalante and M. Raydan, Alternating Projection Methods. Society
[44] O. El Ayach, R. W. Heath, Jr., S. Abu-Surra, S. Rajagopal, and Z. Pi,
for Industrial and Applied Mathematics, 2011, vol. 8.
“Low complexity precoding for large millimeter wave MIMO systems,”
[72] H. H. Bauschke and J. M. Borwein, “On projection algorithms for
in Proc. 2012 IEEE International Conf. Commun., pp. 3724–3729.
solving convex feasibility problems,” SIAM Rev., vol. 38, no. 3, pp.
[45] S. F. Cotter, B. D. Rao, K. Engan, and K. Kreutz-Delgado, “Sparse 367–426, 1996.
solutions to linear inverse problems with multiple measurement vectors,” [73] J. C. Gower and G. B. Dijksterhuis, Procrustes Problems. Oxford
IEEE Trans. Signal Process., vol. 53, no. 7, pp. 2477–2488, 2005. University Press, 2004, vol. 3.
[46] J. A. Tropp, A. C. Gilbert, and M. J. Strauss, “Algorithms for simul- [74] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estimation. Prentice
taneous sparse approximation—part I: greedy pursuit,” Signal Process., Hall, 2000, vol. 1.
vol. 86, no. 3, pp. 572–588, 2006. [75] R. Gray, “Vector quantization,” IEEE ASSP Mag., vol. 1, no. 2, pp.
[47] J. A. Tropp, “Algorithms for simultaneous sparse approximation—part 4–29, Apr. 1984.
II: convex relaxation,” Signal Process., vol. 86, no. 3, pp. 589–602, [76] S. Zhou and B. Li, “BER criterion and codebook construction for finite-
2006. rate precoded spatial multiplexing with linear receivers,” IEEE Trans.
[48] J. Chen and X. Huo, “Theoretical results on sparse representations of Signal Process., vol. 54, no. 5, pp. 1653–1665, 2006.
multiple-measurement vectors,” IEEE Trans. Signal Process., vol. 54,
no. 12, pp. 4634–4643, 2006.
[49] S. Wright, R. Nowak, and M. Figueiredo, “Sparse reconstruction by Omar El Ayach (S’08, M’13) is currently a Senior
separable approximation,” IEEE Trans. Signal Process., vol. 57, no. 7, Systems Engineer at Qualcomm Research in San
pp. 2479–2493, 2009. Diego, CA. He received his M.S. and Ph.D. in
[50] T. Michaeli and Y. Eldar, “Constrained linear minimum MSE estima- Electrical and Computer Engineering degree from
tion,” Oct. 2007. Available: http://webee.technion.ac.il/publication-link/ The University of Texas at Austin in 2010 and
index/id/439 2013, respectively. Before joining the University of
Texas, he received his B.E. degree in Computer and
[51] T. Michaeli and Y. Eldar, “Constrained nonlinear minimum MSE es-
Communications Engineering from the American
timation,” in Proc. 2008 IEEE International Conf. Acoustics, Speech,
University of Beirut, Lebanon, in 2008. He was an
Signal Process., pp. 3681–3684.
intern at the University of California, Berkeley in the
[52] D. Love and R. W. Heath, Jr., “Limited feedback unitary precoding for Summer of 2007 and an intern at Samsung Research
spatial multiplexing systems,” IEEE Trans. Inf. Theory, vol. 51, no. 8, America - Dallas in the Summers of 2011 and 2012. His research interests
pp. 2967–2976, Aug. 2005. are in the broad area of network sciences, signal processing and information
[53] D. Love, R. W. Heath, Jr., and T. Strohmer, “Grassmannian beamforming theory. In the context of wireless communication, his interests are in MIMO
for multiple-input multiple-output wireless systems,” IEEE Trans. Inf. systems, interference management, and mmWave communication.
Theory, vol. 49, no. 10, pp. 2735–2747, Oct. 2003.
Sridhar Rajagopal (M98, SM09) is currently a Robert W. Heath, Jr. (S’96– M’01–SM’06–F’11)
Sr. Staff Engineer at Samsung Research America received the B.S. and M.S. degrees from the Uni-
Dallas. He received his M.S. and Ph.D. degrees in versity of Virginia, Charlottesville, VA, in 1996
Electrical and Computer Eng. from Rice Univer- and 1997 respectively, and the Ph.D. from Stanford
sity. He has previously worked at Nokia Research University, Stanford, CA, in 2002, all in electrical
Center and at WiQuest Communications, and has engineering. From 1998 to 2001, he was a Senior
contributed to multiple communication standards. Member of the Technical Staff then a Senior Con-
His research interests are in algorithms and archi- sultant at Iospan Wireless Inc, San Jose, CA where
tectures for short-range, high throughput and low he worked on the design and implementation of
power technologies, mmWave and optical wireless the physical and link layers of the first commercial
communication. MIMO-OFDM communication system. Since Jan-
uary 2002, he has been with the Department of Electrical and Computer
Shadi Abu-Surra is currently a Staff Engineer at Engineering at The University of Texas at Austin where he is a Cullen Trust
Samsung Research America Dallas. He received for Higher Education Endowed Professor, and is Director of the Wireless
M.S. in Electrical and Computer Engineering from Networking and Communications Group. He is also President and CEO of
the Jordan University of Science and Technology in MIMO Wireless Inc. and Chief Innovation Officer at Kuma Signals LLC. His
2004 and his Ph.D. degree in Electrical and Com- research interests include several aspects of wireless communication and sig-
puter Engineering from the University of Arizona nal processing: limited feedback techniques, multihop networking, multiuser
in 2009. His research interests are in coding theory, and multicell MIMO, interference alignment, adaptive video transmission,
LDPC code design and decoder architectures, cel- manifold signal processing, and millimeter wave communication techniques.
lular communication, millimeter wave and 60 GHz Dr. Heath has been an Editor for the IEEE T RANSACTIONS ON C OMMUNI -
communication. CATION , an Associate Editor for the IEEE T RANSACTIONS ON V EHICULAR
T ECHNOLOGY, lead guest editor for an IEEE J OURNAL ON S ELECTED A R -
EAS IN C OMMUNICATIONS special issue on limited feedback communication,
Zhouyue Pi (SM’13) is a Senior Director at Sam- and lead guest editor for an IEEE J OURNAL ON S ELECTED T OPICS IN
sung Research America in Dallas, Texas, where he S IGNAL P ROCESSING special issue on Heterogenous Networks. He currently
leads the Emerging Technology Lab doing research serves on the steering committee for the IEEE T RANSACTIONS ON W IRE -
in next generation mobile devices, smart home so- LESS C OMMUNICATIONS . He was a member of the Signal Processing for
lutions, and medical diagnostic technologies. Before Communications Technical Committee in the IEEE Signal Processing Society
joining Samsung in 2006, he was with Nokia Re- and is a former Chair of the IEEE COMSOC Communications Technical
search Center in Dallas and San Diego, where he Theory Committee. He was a technical co-chair for the 2007 Fall Vehicular
worked on 3G wireless standardization and modem Technology Conference, general chair of the 2008 Communication Theory
development for systems such as 3GPP2 1xEV-DV Workshop, general co-chair, technical co-chair and co-organizer of the 2009
Revision C and D, 1xEV-DO Revisions A and B, IEEE Signal Processing for Wireless Communications Workshop, local co-
and UMB. In2006 2009, he was a main contributor organizer for the 2009 IEEE CAMSAP Conference, technical co-chair for
to Samsung’s 4G standardization efforts in 3GPP LTE and IEEE 802.16m the 2010 IEEE International Symposium on Information Theory, the technical
(mobile WiMAX evolution), and to IEEE 802.11ad for 60 GHz communica- chair for the 2011 Asilomar Conference on Signals, Systems, and Computers,
tion also commonly referred to as Wireless Gigabit (WiGig). In 2009 2012, general chair for the 2013 Asilomar Conference on Signals, Systems, and
he pioneered mm-wave mobile communication and led the development of Computers, founding general co-chair for the 2013 IEEE GlobalSIP confer-
the world’s first baseband and RF system prototype that demonstrated the ence, and is technical co-chair for the 2014 IEEE GLOBECOM conference.
feasibility of mobile communication in frequency as high as 28 GHz. He has Dr. Heath was a co-author of best student paper awards at IEEE VTC 2006
authored more than 20 technical papers in internationally circulated journals Spring, WPMC 2006, IEEE GLOBECOM 2006, IEEE VTC 2007 Spring,
and premier conferences and is the inventor of more than 140 patents and and IEEE RWS 2009, as well as co-recipient of the Grand Prize in the 2008
applications. He holds a B.E. degree from Tsinghua University (with honor), WinTech WinCool Demo Contest. He was co-recipient of the 2010 and 2013
a M.S. degree from the Ohio State University, and an MBA degree from EURASIP Journal on Wireless Communications and Networking best paper
Cornell University (with distinction). He is a Senior Member of IEEE. awards and the 2012 Signal Processing Magazine best paper award. He was
a 2003 Frontiers in Education New Faculty Fellow. He is also a licensed
Amateur Radio Operator and is a registered Professional Engineer in Texas.

Spatially Sparse Precoding in Millimeter Wave MIMO Systems

Uploaded by

Copyright:

Available Formats

Spatially Sparse Precoding in Millimeter Wave MIMO Systems

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Spatially Sparse Precoding in Millimeter Wave MIMO Systems

Uploaded by

Copyright:

Available Formats

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO.

3, MARCH 2014 1499