Professional Documents
Culture Documents
Versatile Compressive Mmwave Hybrid Beamformer Codebook Design Framework
Versatile Compressive Mmwave Hybrid Beamformer Codebook Design Framework
Abstract—Hybrid beamforming (HB) architectures are attrac- ing [7] and digital beamforming [8]. In those publications,
tive for wireless communication systems with large antenna transmit pilot symbols [5], [8] or angles of phase shifters
arrays because the analog beamforming stage can significantly in analog beamformers [4], [6], [7] are randomly chosen
reduce the number of RF transceivers and hence power con-
from finite predefined sets under hardware constraints. This
arXiv:1909.09858v1 [cs.IT] 21 Sep 2019
…
…
. Digital . . Analog H Analog . . Digital .
xm . Precoder . . Lt Precoder Combiner . Lr . Combiner . ym
. . . Nt Nr . . .
FBB,m FRF,m WRF,m WBB,m
DAC RF Chain RF Chain ADC
ADC
DAC
Fig. 1. Block diagram of a general hybrid analog and digital beamforming architecture at both a transmitter and a receiver.
channels, the discrete time received signal in the m-th frame and Ar , [ar (θr0 ), ar (θr1 ), . . . , ar (θr(Np −1) )], the channel
(or the m-th time instant) can be written as matrix H in (2) can be expressed in matrix form as
√
ym = ρWm H
HFm xm + Wm H
nm ∈ CLr , H = Ar Hd AH
t, (3)
where ρ is the transmit power in the training phase, Wm = where Hd ∈ CNp ×Np is a square matrix with the scaled
WRF,m WBB,m ∈ CNr ×Lr is the hybrid combiner, H ∈ complex channel gains on the diagonal.
CNr ×Nt is the channel matrix, Fm = FRF,m FBB,m ∈ CNt ×Lt
is the hybrid precoder, xm ∈ CLt is the training symbols III. S PARSE F ORMULATION
vector, and nm ∈ CNr ∼ CN (0, σn2 I) is the additive noise
vector. A hybrid combiner (precoder) in the m-th frame is For application of CS algorithms to sparse CE, the received
composed of an RF combiner WRF,m ∈ CNr ×Lr (an RF signal matrix in (1) can be rewritten in vector form using the
precoder FRF,m ∈ CNt ×Lt ) and a baseband (BB) combiner matrix equality, vec(ABC) = (CT ⊗ A)vec(B), as
WBB,m ∈ CLr ×Lr (a BB precoder FBB,m ∈ CLt ×Lt ). In √
y = ρ X̄T ⊗ WH vec(H) + n,
(4)
order to keep the constant transmit power, kx̄m k2 = 1 where
x̄m , Fm xm . With Mt and Mr denoting the number of where y ∈ CM Lr collects the M received frames in vector
transmit and receive configurations, respectively, the received form, and n , vec(N). In addition, angle grids need to
signal matrix Y that contains M (= Mt Mr ) frames can be be selected based on which CS algorithms search pairs of
written as an AoD and an AoA. We choose the angles to be uni-
formly distributed in the normalized discrete spatial angle
Y = [y1T , . . . , yM
T T T T T
] , . . . , [yM −Mr +1 . . . , yM ]
√
r domain, i.e., [−1/2, 1/2), such that ϑn ∈ Θ = {ϑn |ϑn =
= ρWH HX̄ + N, (1) n 1
G − 2 , n = 0, . . . , G − 1} where G ≥ max(Nt , Nr )
is the grid size. Therefore, the higher a G value is, the
where Y ∈ CLr Mr ×Mt is the received signal matrix, W ,
finer the angle granularity becomes. The transmit array re-
[W1 , . . . , WMr ] ∈ CNr ×Lr Mr is the receive configuration
matrix, X̄ , [x̄1 , . . . , x̄Mt ] ∈ CNt ×Mt is the transmit
sponse vectors q with spatial angle ϑ can beT expressed as
1 −j2πϑ
configuration matrix, and N ∈ CLr Mr ×Mt is the noise matrix āt (ϑ) = Nt 1, e , . . . , e−j2π(Nt −1)ϑ and the re-
is expressed as N , blkdiag(W1, . . . , WMr )H × ceiver one is also similarly defined. The angle grids are
which
[n1 , . . . , nT
T T T T T defined by the transmit and receive grid array response ma-
Mr ] , . . . , [nM −Mr +1 , . . . , nM ] .
Adopting a geometric channel model [10], [12] and linear trices Āt , [āt (ϑ0 ), . . . , āt (ϑG−1 )] ∈ CNt ×G and Ār ,
antenna arrays, each scatterer contributes a channel path [ār (ϑ0 ), . . . , ār (ϑG−1 )] ∈ CNr ×G , respectively. Note that
G G
associated with its own azimuth angle of departure and Āt ĀH H
t = Nt INt and Ār Ār = Nr INr . The channel matrix
arrival (AoD and AoA) denoted by θtl and θrl , respectively. (3) can be redefined with such grid array response matrices by
G×G
Therefore, the channel matrix can be expressed as H = Ār H̄d ĀH t where H̄d ∈ C is the new channel gain
matrix. Ignoring the grid quantization errors, H̄d is a sparse
Np −1
q X matrix with Np non-zero entries being complex channel gains
H= Nt Nr /Np αl ar (θrl )aH
t (θtl ), (2) corresponding to a combination of each transmit and receive
l=0
array response vector in Āt and Ār , respectively. Unlike Hd ,
where Np is the total number of paths (equivalent to the non-zero elements in H̄d do not have to be on the diagonal.
channel rank), αl ∼ CN (0, σα2 ) is the complex channel gain of As vec(H) = (Ā∗t ⊗ Ār )vec(H̄d ), the received signal vector
the l-th path, and at (·) ∈ CNt and ar (·) ∈ CNr are the trans- in (4) can be rewritten as
mit and receive array response vectors, respectively, evaluated √
y = ρ X̄T ⊗ WH Ā∗t ⊗ Ār vec H̄d + n
for the given angles. Both θtl and θrl are random variables
√
= ρ X̄T Ā∗t ⊗ WH Ār h + n
that follow the uniform distribution U(0, 2π). Assuming the
transmit and receive antennas are in the form of uniform √ √
= ρΦΨh + n = ρΦ̄h + n,
linear array (ULA) with a half wavelength antenna spacing,
the transmit array response vectors are given as at (θ) = where h , vec(H̄d ) is the Np -sparse channel vector, Φ ,
T
X̄T ⊗WH ∈ CLr Mt Mr ×Nt Nr is the sensing matrix, Ψ , Ā∗t ⊗
p
1/Nt 1, e−jπ cos(θ) , . . . , e−jπ(Nt −1) cos(θ) , and the re-
2
ceive array response is similarly defined. Defining the array Ār ∈ CNt Nr ×G is the sparsifying dictionary, and Φ̄ , ΦΨ
response matrices At , [at (θt0 ), at (θt1 ), . . . , at (θt(Np −1) )] is the equivalent dictionary.
IV. D ETERMINISTIC S ENSING M ATRIX D ESIGN 1) Minimizing µt (X̃): By definition of total coherence,
min µt (X̃) is equivalent to minkX̃H X̃ − IG k2F . The cost
Previous efforts have carefully designed a sensing matrix function can be expressed as
rather than using a random one in order to improve the
2 2
performance of CS algorithms. Mutual coherence is one of the X̃H X̃ − IG = X̃X̃H − IMt + (G − Mt ). (7)
most popular metrics that are used for recovery guarantees of F F
CS algorithms. Finding optimal sensing matrices using mutual By definition of X̃, X̃X̃H can be rewritten as X̄T Ā∗t ĀT ∗
t X̄ =
coherence becomes intractable. Alternatives to the mutual G T ∗ 2
Nt X̄ X̄ . Due to the transmit power constraint kx̄m k = 1,
coherence, thus, are considered in [16]–[18]. the diagonal elements of X̃X̃H are fixed and cannot be
In [10], the authors considered minimizing the total coher- minimized without reducing the transmit power. Thus off-
ence for sensing matrix design. We build on [10] by replacing diagonal elements are of our interest, and each X̃X̃H (p, q)
their approximation techniques to decompose the minimiza- for p 6= q can be written as
tion problem into separate transmit and receive minimization G T T ∗ ∗ G T T
problems with a derivation. We also build on [10] by using x p Fp Fq x q = x F FT F∗ F∗ x ∗ . (8)
Nt Nt p BB,p RF,p RF,q BB,q q
greedy approach to consider all columns of the equivalent
dictionary for optimization, instead of using only a subset. Note that p, q ≤ Mt and that Mt is desired to be large to
reduce the lower bound shown in (7). It leads to a larger
A. Total Coherence matrix for which we would like to minimize the sum of
squared off-diagonal elements. Here we have three sets of
We define the total coherence of the Φ̄ as degrees of freedom to control the elements: the analog and
G2
G 2 digital precoders, and the training symbols. We define FRF and
t
X X 2 FBB as codebooks for possible analog and digital precoders,
µ (Φ̄) = Φ̄(m)H Φ̄(n)
m n,n6=m respectively.
2 Assuming Nt is a multiple of Lt , we define the analog
G2 X
G
precoder codebook FRF as
X 2 2
= Φ̄(m)H Φ̄(n) − Φ̄(m)H Φ̄(m) .
FRF = {FRF,p ∈ CNt ×Lt : ∀p, q ∈ {1, . . . , Nt /Lt },
m n
B. Transmitter Design ILt . It leads x̃p to be the same standard basis that is used
for Σp with a reciprocal of the scaler if scaled. It shows that
the digital precoder and the training symbols have one-to-one
correspondence, that the degrees of freedom we actually have Lt +
…
is two, and that |FBB | = Lt . Thus, without loss of generality,
…
Nt
…
…
we use the standard bases without scaling. Consequently the .
. . .
. . .
Nt Nt
digital precoder codebook FBB can be expressed as .
.
Lt . .
Lt . Lt Nt
…
…
Nt
+
…
H
FBB,p = UΣp V , Σp = diag(ep ) ,
(a) (b) (c)
where ep denotes the Lt -dimensional standard basis with one Fig. 2. Three types of analog precoders: (a) phase shifting network, (b)
in the p-th entry. For the given FBB,p , the corresponding switching network, and (c) CAP-MIMO.
training symbol vector is given as xp = Vep . p
2) Minimizing ν(X̃): We now focus on minimizing ν(X̃). Nr /GILr for all p. Then the digital combiner codebook
As all terms in summation in ν(X̃) are non-negative, the can be defined as
PG
problem min ν(X̃) is equivalent to min m |X̃(m)H X̃(m)|,
WBB = WBB,p ∈ CLr ×Lr : ∀p ∈ {1, . . . , Lr },
and the cost function can be rewritten as Tr(X̃X̃H ). It is a sum
of the diagonal elements in X̃X̃H which, as shown previously,
p H
are constrained by the transmit power. It means that ν(X̃) WBB,p = Nr /GŨṼ . (11)
cannot be minimized without reducing the transmit power.
2) Minimizing ν(W̃):
PG As with Section IV-B2, min ν(W̃)
is equivalent to min m |W̃(m)H W̃(m)|, and its cost func-
C. Receiver Design
tion can be expressed as Tr(W̃W̃H ). Since the analog com-
H
1) Minimizing µt (W̃): As with the minimization of µt (F̃), biners have a condition WRF,p WRF,p = ILr and the digital
min µt (W̃) is equivalent to minkW̃H W̃−IG k2F . As with (7), combiner codebook has a single element, Tr(W̃W̃H ) reduces
we have the cost function given by kW̃W̃H − IMr Lr k2F + (G to G/Nr Tr(WpH Wp ) which is G/Nr Tr(Σ̃H p Σ̃p ) assuming
− Mr Lr ), and W̃W̃H can be expressed as WBB,p = ŨΣ̃p ṼH . A trivial solution to min Tr(Σ̃Hp Σ̃p ) is
Σ̃p = 0Lr , which does not make sense. We thus pose a
W1H W1 W1H WMr
··· temporary constraint kWBB,p k2F = α > 0. Then the problem
G .. .. ..
W̃W̃H = . (10) now becomes
Nr
. . .
H H
WM W1 ··· WMr WMr min Tr(Σ̃H 2
r
p Σ̃p ) subject to kΣ̃p kF = α > 0.
NMSE [dB]
angle set. Columns of the normalized Nt -point DFT matrix Oracle Random
-15 LSE Codebook
can be shown to be in F1 . In this case, the cardinality |F1 | [Unfilled] Proposed
[Filled] Random Proposed
is Nt , and FRF with columns being all f ’s ∈ F1 satisfies -20
Phase shifter Codebook
Switch
FHRF FRF = INt .
Lens
OMP [13]
BPDN [14]
-25
B. Switching Network -10 -5 0 5 10 15
SNR [dB]
The switching network architecture connects each RF chain
(a) Grid size G is 64 points
to one of the antennas via a switch as shown in Fig. 2(b).
At a given moment, the number of active antennas is Lt .
0
The feasible set of precoding vectors is given as F2 =
f ∈ B Nt : kf k0 = 1 . Due to the absence of splitters and
-5
combiners, the L2 norm of f ’s ∈ F2 is fixed at one. Without
compensation from baseband processing, the condition is -10
NMSE [dB]
satisfied with a set of standard bases of RNt considered for
Oracle
the precoding vectors. -15 LSE
[Unfilled] Proposed
[Filled] Random
C. Continuous Aperture Phased MIMO -20
Phase shifter
Switch
CAP-MIMO that directly exploits beamspace MIMO com- Lens
OMP [13]
munications is enabled by high resolution DLAs [19]. For -25 BPDN [14]
-10 -5 0 5 10 15
simulation, the DLA with adaptive selecting network pro- SNR [dB]
posed in [5] is used in the following section. A transmitter (b) Grid size G is 180 points
architecture of a 1D DLA is illustrated in Fig. 2(c). Feed
Fig. 3. NMSE for channel estimation vs. SNR. The Oracle estimator knows
antennas can ideally generate mutually orthogonal beams, and the AoA and AoD. The LSE uses the proposed deterministic sensing matrix.
Lt antennas out of Nt are activated by selecting desired The other results are for the possible combinations of deterministic/random
antennas and feeding input streams. To this end, DLAs hybrid beamforming designs, three RF beamforming architectures, and two
estimation algorithms.
are designed in order to make the analog precoding matrix
approximate the DFT matrix. Considering the ideal precoding
to lie on the angle grids of the dictionary. In the simulations,
matrix,the set of analog precoding vectors can be expressed as
500 channel realizations are used for each point. The complete
F3 = f ∈ CNt : f = Udf t (i), i = 1, 2, . . . Nt where Udf t
source code is available [20].
denotes the Nt -point DFT matrix. Since the DFT matrix
is unitary, the analog precoding matrix obviously satisfies Fig. 3 shows NMSE of channel estimates as a function
FH of SNR with combinations of codebooks, architectures and
RF FRF = INt .
algorithms. The oracle estimator outperforms the others and
VI. N UMERICAL R ESULTS scales well with SNR serving as the lower bound. Both figures
In this section, performance of CE based on the proposed show that the proposed deterministic design outperforms the
sensing matrix is evaluated using OMP and BPDN algorithms random design for all considered combinations across the
as representatives of CS algorithms with coherence-based entire SNR range including the low SNR regime. Performance
recovery guarantees [15]. Both the normalized mean squared of the OMP and BPDN with the random sensing matrix, in
error (NMSE) and the achievable SE are used as performance general, is between the proposed design and the LSE. It is
metrics where NMSE is defined as E[kH − Ĥk2F /kHk2F ]. expected because the equivalent dictionary has higher total
SNR is defined as ρ/σ 2 . We provide results obtained with the coherence than the proposed one does. Compared with the
random sensing matrix in addition to the one we propose. For random design, the proposed deterministic design makes the
comparison purposes, the least squares estimator (LSE) and various HB architectures achieve very similar performance
the oracle estimator are evaluated as well. The LSE evaluated since they generate dictionaries with the identical total co-
in this section is based on the proposed sensing matrix. The herence as shown in Section V.
oracle estimator refers to the LSE with actual AoAs and AoDs Spectral efficiencies computed based on SVD beamform-
known at the receiver. ing for various combinations of architectures and estimation
The system is equipped with Nt = 64, Nr = 16, Lt = 8, algorithms are plotted in Fig. 4. The SE using the perfect
Lr = 8 (hence Mt = 64 and Mr = 2), G = 64, Np = 4 and channel state information (CSI) is provided in the figure as
bPS = 6 for simulation unless otherwise specified. The AoDs the performance upper bound. In both figures, the oracle
and AoAs of the multipath components are not constrained estimator yields the indistinguishable SE from the perfect CSI
45
optimization problems improve CE performance of the CS
40
23 algorithms that rely on coherence to guarantee sparse recovery
Spectral Efficiency [bits/s/Hz] 22 and were shown by simulation to outperform, in terms of both
35
21
estimation error and SE.
30 20
25 -1 -0.5 0 0.5 1
Perfect CSI
Oracle
R EFERENCES
20
LSE
15
[Unfilled] Proposed [1] 3GPP, TS 38.104 V15.0.0 Base Station (BS) radio transmission and
[Filled] Random
Phase shifter reception (Release 15), Dec. 2017.
10 Switch
Lens
[2] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun,
5 OMP [13] “Wideband millimeter-wave propagation measurements and channel
BPDN [14]
0
models for future wireless communication system design,” IEEE Trans.
-10 -5 0 5 10 15 Commun., vol. 63, no. 9, pp. 3029–3056, Sep. 2015.
SNR [dB]
[3] T. S. Rappaport, F. Gutierrez, E. Ben-Dor, J. N. Murdock, Y. Qiao,
(a) Grid size G is 64 points and J. I. Tamir, “Broadband millimeter-wave propagation measurements
and models using adaptive-beam antennas for outdoor urban cellular
communications,” IEEE Trans. Antennas Propag., vol. 61, no. 4, pp.
45 1850–1859, Apr. 2013.
40
24
[4] K. Venugopal, A. Alkhateeb, N. González-Prelcic, and R. W. Heath,
23 “Channel estimation for hybrid architecture-based wideband millimeter
Spectral Efficiency [bits/s/Hz]
35
22
wave systems,” IEEE J. Sel. Areas in Commun., vol. 35, no. 9, pp.
30 1996–2009, Sep. 2017.
25
21
-1 -0.5 0 0.5 1 [5] X. Gao, L. Dai, S. Han, C. L. I, and X. Wang, “Reliable beamspace
Perfect CSI channel estimation for millimeter-wave massive MIMO systems with
Oracle
20 lens antenna array,” IEEE Trans. Wireless Commun., vol. 16, no. 9, pp.
LSE
[Unfilled] Proposed 6010–6021, Sep. 2017.
15 [Filled] Random
Phase shifter [6] J. Rodriguez-Fernandez, N. Gonzalez-Prelcic, K. Venugopal, and
10 Switch
Lens
R. W. Heath, “Frequency-domain compressive channel estimation for
5 OMP [13] frequency-selective hybrid mmWave MIMO systems,” IEEE Trans.
BPDN [14]
0
Wireless Commun., vol. 17, no. 5, pp. 2946–2960, May 2018.
-10 -5 0 5 10 15 [7] D. Ramasamy, S. Venkateswaran, and U. Madhow, “Compressive
SNR [dB]
adaptation of large steerable arrays,” in Proc. Information Theory and
(b) Grid size G is 180 points Applications Workshop, Feb. 2012, pp. 234–239.
[8] P. Schniter and A. Sayeed, “Channel estimation and precoder design for
Fig. 4. SE for channel estimation vs. SNR. The Oracle estimator knows
millimeter-wave communications: The sparse way,” in Proc. Asilomar
the AoA and AoD. The LSE uses the proposed deterministic sensing matrix.
Conf. Sign., Sys. and Comp., Nov. 2014, pp. 273–277.
For the possible combinations of hybrid beamforming designs, RF precoders,
[9] Y. Xiao, Y. Wang, and W. Xiang, “Dimension-deficient channel esti-
and estimation algorithms, the SE is computed using SVD beamforming.
mation of hybrid beamforming based on compressive sensing,” IEEE
since NMSE is very low across the SNR range as shown Access, vol. 7, pp. 13 791–13 798, 2019.
[10] J. Lee, G. T. Gil, and Y. H. Lee, “Channel estimation via orthogonal
in Fig. 3. The proposed deterministic design outperforms the matching pursuit for hybrid MIMO systems in millimeter wave com-
random from the perspective of SE as well, and makes spectral munications,” IEEE Trans. Commun., vol. 64, no. 6, pp. 2370–2386,
efficiencies of the various architectures tend to converge. Jun. 2016.
[11] V. Raghavan, J. Cezanne, S. Subramanian, A. Sampath, and O. Koymen,
In both Fig. 3 and 4, an increase in the grid size is favorable “Beamforming tradeoffs for initial ue discovery in millimeter-wave
for both deterministic and random approaches since larger mimo systems,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3,
grids mean finer resolution in the angle search spaces. By pp. 543–559, Apr. 2016.
[12] A. Alkhateeb, O. E. Ayach, G. Leus, and R. W. Heath, “Channel
increasing G from 64 to 180, NMSE curves are shifted down estimation and hybrid precoding for millimeter wave cellular systems,”
and SE curves are lifted up for both OMP and BPDN. At IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831–846, Oct.
the same time, however, gaps between the proposed and the 2014.
[13] J. A. Tropp and A. C. Gilbert, “Signal recovery from random mea-
random become smaller. It implies that the proposed codebook surements via orthogonal matching pursuit,” IEEE Trans. Inf. Theory,
is more efficient when with a smaller grid size which could vol. 53, no. 12, pp. 4655–4666, Dec. 2007.
be preferred for computation reduction. [14] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decomposition
by basis pursuit,” SIAM review, vol. 43, no. 1, pp. 129–159, 2001.
VII. C ONCLUSION [15] M. F. Duarte and Y. C. Eldar, “Structured compressed sensing: From
theory to applications,” IEEE Trans. Signal Process., vol. 59, no. 9, pp.
In this paper, we proposed a versatile deterministic HB 4053–4085, Sep. 2011.
design framework for CS based CE in narrowband mmWave [16] L. Zelnik-Manor, K. Rosenblum, and Y. C. Eldar, “Sensing matrix
communication systems. The deterministic sensing matrix optimization for block-sparse decoding,” IEEE Trans. Signal Process.,
vol. 59, no. 9, pp. 4300–4312, Sep. 2011.
design we proposed works for a variety of hybrid beam- [17] M. Elad, “Optimized projections for compressed sensing,” IEEE Trans.
forming architectures that are implemented with variable Signal Process., vol. 55, no. 12, pp. 5695–5702, Dec. 2007.
phase shifters, switches or a DLA. Our design approach is [18] J. M. Duarte-Carvajalino and G. Sapiro, “Learning to sense sparse
signals: Simultaneous sensing matrix and sparsifying dictionary opti-
to configure analog and digital beamformers by minimizing mization,” IEEE Trans. Image Process., vol. 18, no. 7, pp. 1395–1408,
the total coherence of the equivalent sparsifying dictionary. Jul. 2009.
We decoupled the joint transmitter and receiver optimization [19] J. Brady, N. Behdad, and A. M. Sayeed, “Beamspace MIMO for
millimeter-wave communications: System architecture, modeling, anal-
problem into two disjoint problems. The analog and digi- ysis, and measurements,” IEEE Trans. Antennas Propag., vol. 61, no. 7,
tal beamformer codebooks that are obtained by solving the pp. 3814–3827, Jul. 2013.
[20] J. Sung and B. L. Evans, “Versatile compressive mmWave hybrid
beamformer codebook design framework,” Software Release, Jul. 26,
2019, https://github.com/junmo-sung/versatile-cb-design.