Received: Added at production Revised: Added at production Accepted: Added at production

DOI: xxx/xxxx

RESEARCH ARTICLE

Channel Estimation Based on Low-Complexity Hierarchical

Codebook Design for Millimeter-Wave MIMO Systems

Baghdad Hadji*1 | Lamya Fergani1 | Mustapha Djeddou2

1 LISIC Laboratory , Université des Sciences

et de la Technologie Houari Boumédiène, Summary
B.P. 32, El Alia, Bab Ezzouar, 16111
Algiers, Algeria
Millimeter-wave bands (mmWave) are considered as a strong candidate for achieving
2 Communication System Laboratory, Ecole high-quality communication links for the future outdoor cellular systems to overcome
Militaire Polytechnique, BP 17 Bordj El the spectrum congestion problem. Due to the extremely high path-loss in mmWave
Bahri 16111 Algiers, Algeria
band, large antenna arrays at both the transmitter and receiver are necessary. Hybrid
Correspondence beamforming architectures are used to exploit the potential array gain with several RF
*Baghdad HADJI, LISIC
Laboratory,Université des Sciences et de la chains, which poses a problem of complexity when estimating the mmWave channel.
Technologie Houari Boumédiène,B.P. 32, El To address the challenge of this hardware complexity, we propose in this paper, an
Alia, Bab Ezzouar, 16111 Algiers,Algeria.
approach to design a multi-resolution hierarchical codebook to meet parallel multi-
Email: [email protected]
stream data based on the physical design via hybrid analog/digital architecture with
low complexity, i.e., 2-bit phase state and few numbers of RF chains. The simulation
results verify that our proposed method to design the codebook has better design
performance of beams and can achieve higher average spectral efficiency gains of
channel estimation compared to the one based on high hardware complexity.

KEYWORDS:
Millimeter wave communications, MIMO, parallel multi-stream data, Hybrid beamforming, multi-
resolution hierarchical codebook design, Orthogonal Matching Pursuit

1 INTRODUCTION
The rapid increase in mobile data growth and the use of smartphones is creating unprecedented challenges for wireless service
providers to overcome a global bandwidth shortage 1 . As the mobile data demand grows, the sub-6 GHz spectrum is becoming
increasingly crowded leading to spectrum congestion. Because, the physical layer (i.e., electronic circuit transmission) technol-
ogy has reached Shannon’s capacity 2 as an alternative to increasing the data traffic at these bands. Therefore, fifth-generation
(5G) is the next generation of cellular networks 3 and is expected to be facing the growth of global mobile data traffic in the range
of 10-100 times from 2020 to 2030 4 . It is widely believed that 5G is not just an incremental version of the fourth-generation
(4G) communication systems due to not only the increasing demand of data traffic explosion 5 but due to the vast evolution of
the internet and the connectivity of objects and devices have evolved 3 .
Accordingly, 5G and beyond will be enabling the collection and exchange of information and data globally between people,
communities, and more importantly, businesses 3 , thanks to leading applications, such as internet-of-things (IoT) 5 , ultra-high-
definition (UHD)3D video streaming, cloud-based services and augmented reality 6 . However, to enhance the communication
of the new wireless transmission systems and to adapt the various emerging applications and services that are mentioned above,
5G cellular networks must address six challenges that are not effectively addressed by 4G, i.e., higher capacity, higher data rate,
lower end to end latency, massive device connectivity, reduced cost, and consistent Quality of Experience provisioning 7 . To
2 B.Hadji ET AL

accommodate these challenging requirements, mmWave communications is a promising technology for 5G-and-beyond mobile
cellular networks, as well as for emerging Gbps-speed Wi-Fi networks based on the IEEE 802.11ad and draft IEEE 802.11ay
standards 8 , due to the tens of GigaHertz of available licensed and unlicensed spectrum bands in the range of 20-100 GHz 9 .
Besides, the use of mmWave bands for wireless communications offers channel bandwidths that will be more than ten times
greater than today’s 4G Long-Term Evolution (LTE) 20 MHz cellular channels 10 . Therefore, mmWave communications provide
an opportunity to meet such explosive capacity demand for 5G 4 . On the other hand, propagation aspects are unique at mmWave
due to the very small wavelength compared to the size of most of the objects in the environment 11 . Thus, the development
of communications systems in the mmWave band is a challenging task. MmWave signals suffer from severe propagation loss,
penetration loss, and rain fading compared to signals in lower frequencies 12 . Hence, one of the challenges of using mmWave
frequencies to provide outdoor coverage in a cellular system is the increased path-loss encountered at these frequencies 13 . Con-
sequently, there is a trade-off between operating broadband available at mmWave and high isotropic path-loss along with the
transmission. It is well established that the use of 𝑁 antennas improves the signal-to-noise ratio (SNR) by 10 log10 (𝑁) in the
receive path, and on the transmit side it increases the signal strength by 20 log10 (𝑁) in the desired direction 14 . For this reason,
narrow beam communications powered by multi-input multi-output (MIMO) systems are required in the mmWave communica-
tions to overcome the increased path losses thanks to suitable beamforming gains. As a result, sufficient beamforming gains are
made possible by large antenna arrays which must be packaged into small form factors. Fortunately, the short-wavelength allows
more antenna elements to be integrated into mmWave devices. Also, the large antenna array can support the multi-stream/multi-
user transmission to improve spectral efficiency. However, the feasible combination of MIMO and large available bandwidth at
mmWave can meet some of the mentioned 5G challenging requirements.
When the communication system is equipped with multiple antennas, the transmission protocol is based on precoding and
combining principles. For MIMO systems operating in conventional frequency bands, this precoding and combining is entirely
realized in the digital domain, where each RF chain per-antenna includes key components such as power amplifier, low noise
amplifier, mixer, upconverters/down converters, digital-to-analog (DAC)/analog-to digital (ADC) converters, to generate multi-
ple independently controlled beams. Furthermore, it is shown that RF components may consume up to 70% of the total transceiver
energy consumption 4 . Thus, this strict requirement of a single RF chain per-antenna substantially increases the hardware com-
plexity and power consumption, when large scale antenna arrays are employed at the transceiver 15 . To address the challenge
of this hardware complexity, i.e., the number of expensive RF chains, analog beamforming technique is one of the simplest
approaches for applying MIMO in mmWave systems 11 . The core of this technique is a network of analog phase shifters that
imposes constant modulus (CA) constraints on the beamformer and is based only on one RF chain, which is highly desirable
from a hardware complexity/power consumption point of view 12 . On the other hand, analog beamforming architecture reduces
the hardware constraints but does not exploit the benefits of the combination of mmWave and MIMO systems, because a single
beamformer only supports single-user and single-stream transmission. To strike a balance between hardware complexity and
system performance, hybrid analog/digital architecture has been proposed 16 , which split MIMO processing into the analog and
digital domains where a small number of RF chains are tied to a large antenna array. For the hybrid architecture, the number
of RF chains can be reduced to equal the number of transmitted/received data streams, which is, generally, much lower than
the number of antenna elements 17 . Recently, the investigation of hybrid precoder and combiner design has attracted extensive
attention because of its potential energy efficiency for mmWave MIMO communications 18 . Meanwhile, the hybrid beamform-
ing design is, generally, a nontrivial task, mainly due to the joint optimization of the analog and digital parts 17 . As a solution,
the sub-optimal approaches are used to found the optimal hybrid precoders that lead to maximizing spectral efficiency (SE)
or energy efficiency (EE). In 19 , the design precoding problem is based on the mutual information instead of directly using
SE, where the problem is formulated as Euclidean norm-minimization between the optimal hybrid recorders and the uncon-
strained optimal precoder which is derived from the channel’s singular value decomposition (SVD). Due to the very different
sub-channel SNRs after SVD, complicated bit allocations are usually required to match the sub-channel SNRs. The authors
in 20 proposed a geometric mean decomposition-based hybrid precoding scheme to avoid the complicated bit allocation. In 21 ,
the hybrid precoding design is considered as a matrix factorization problem and an effective alternating minimization (AltMin)
algorithms are proposed to solve the constrained matrix factorization problems for different hybrid precoding structures, i.e.,
the fully-connected and partially-connected structures. For the fully-connected structure, MO-AltMan and PE-AltMan algo-
rithms are based on manifold optimization techniques, where the PE-ALTMIN algorithm imposes unitary constraints on the
digital baseband precoder, whereas the MO-ALTMIN algorithm does not require unitary constraints on the digital precoder 15 .
While, for partially-connected, the SDR-ALTMIN algorithm is based on semidefinite relaxation to design the digital precoder.
The MO-AltMan and the SDR-ALTMIN algorithms are impractical due to their large computational complexity, especially in
B.Hadji ET AL 3

mmWave communication where we need a large antenna array system. To reduce computational complexity, low-complexity
algorithms proposed in 15 , where the hybrid precoding problem is solved using a combination of alternating minimization and
the majorization-minimization framework for the fully-connected structures, whereas for the partially-connected structures the
problem is solved using variable elimination and the minorization-maximization framework 15 . Apart from SE maximization, for
the 5G and beyond cellular communication system, the EE of the system is also a key performance indicator 22 . The design of the
hybrid precoding to maximizing EE is proposed in 22 using the SVD technique and water filling like power allocation method.
In 23 , the hybrid precoding design is proposed for the fully and partially-connected architectures to maximize the EE by deriv-
ing the precoding and combining matrices through decoupled non-convex transmitter-receiver optimization. The design in 23 is
based on the assumption of ideal circuitry and power amplifiers (PAs). In 24 , the hybrid precoding based on the EE maximiza-
tion algorithm under a practical setting of non-ideal PAs and non-negligible circuit power. The authors in 25 formulate the hybrid
precoding design problem to minimize the total power consumption of a multi-cell multi-user system, subject to the per-user
quality-of-service constraints, and take into account the hardware power consumption difference between different architectures
as well as between the base station active and sleep modes. Moreover, the EE performance was also examined w.r.t the number
of RF chains and antennas 24 . As the EE and SE are simultaneously required to enhance the wireless communication, the authors
in 17 propose a two-stage hybrid precoding beamforming design approach tackling both SE and EE maximization while taking
into account the hardware constraints and realistic circuitry power consumption. In the works discussed above, the hybrid pre-
coding design is usually based on complete Channel State Information (CSI), which is hard to acquire in mmWave systems 26 .
Hence, how to obtain the optimal precoding is the new constraints imposed for hybrid analog/digital architecture.
In the channel estimation literature, the traditional channel estimation techniques estimate the entries of the channel matrix,
but in mmWave communication, these techniques require a lot of training overhead. Because there are a lot of challenges such
as (i) the channel matrix is large due to the use of large arrays, (ii) the large mmWave communication channel bandwidth
increases noise power and makes the received SNR very low before beamforming design, and (iii) in the angular domain, the
mmWave channel has a sparse nature due to the propagation environment effect on smaller wavelength signals. On that account,
the classical techniques may still be prohibitive. To avoid classical channel estimation, Compressed Sensing (CS) tools must
be used to reduce the training overhead in mmWave communication systems, and a finite set of precoding matrices, named
codebook, known to both the receiver (Rx) and the transmitter (Tx) should be predesigned 27 . The codebook contains multiple
codewords, i.e. steering vectors. These codewords formulate the training beamforming vectors with different beamwidths. To
tackle the estimation problem of the multipath mmWave channels, the multi-resolution hierarchical beam training is typically
used to acquire the CSI. Where the angles of departure or arrival (AoDs/AoAs) of Tx and Rx are divided into several partitions,
and the best transmit/receive beams are the ones corresponding to the determined path with the maximum SNR. This results
in the selection of partitions that may contain the most dominant AoD/AoA combination of the mmWave channel. Moreover,
the multi-resolution hierarchical beam training uses an iterative process to exchange between the Tx and Rx, the indices of
the selected partitions which are further divided into smaller partitions using narrower beamwidth at each selection step until
to estimate the best AoD/AoA pair with the desired resolution. As known, the performance of hierarchical search is heavily
dependent on the codebook design 28 . Besides, the design of codebooks is surely based on the used architecture. In 29 , the hybrid
codebooks are obtained by solving the minimizing problem of Euclidean norm between the analog/digital codebooks and the
designed codebooks using sparse signal processing techniques based on the orthogonal matching pursuit (OMP). The codebook
design in 29 is based on the least square method using RF phase shifters with a large number of quantization bits to get the high-
quality wide beams. Meanwhile, the practical design of RF phases shifters with high phase shift resolution will become difficult
and complex. Furthermore, the least square solutions can include errors in codebook design that lead to wrong detections of
fixed AoAs and AoDs. Thus, the angle estimation error is probably larger than the angle resolution (the maximum estimation
error) in theory even if there is no hardware constrains and SNR is large enough 28 . To improve the estimation performance
during early training stages in 29 , a power allocation scheme has been proposed as a solution that requires an increased dynamic
range for the PAs. In 30 , the design of a hierarchical codebook for hybrid architecture is proposed by taking into account the
per-antenna power constraint. The design of the beams in 30 is based on the phase-shifted discrete Fourier Transform (PS-DFT)
that proposed in 31 and design only RF codewords because the phases of digital codewords have been absorbed into those of RF
codewords. As mentioned above, in the hybrid architecture, the digital precoder needs to be jointly designed with the analog
precoder. Therefore, the codebook design loses the hybrid aspect. On the other hand, a critical challenge to design the multi-
resolution hierarchical codebook proposed in 30 is beam widening, i.e., to design the low layer codewords which have wide beam
widths. Due to this, the authors in 30 consider using the sub-array technique to shape a wide beam, i.e., a special architecture
must be used. In 32 , the generation of beams is based on the Fourier Series Method with Kaiser Windowing (FSM-KW) to
4 B.Hadji ET AL

synthesize angular patterns confined in the desired angular region. Whereas the RF phase shifters are quantized with just two
bits but without the CA constraints, although the analog codebook vectors are implemented using phase shifters. The designed
codebook in 32 is dedicated to one stream data communication.
In this work, we propose a concept of a multi-resolution hierarchical codebook based on a physical design with low-complexity
architecture (i.e., 2-bit phase state and few numbers of RF chains) to avoid the codebook design errors introduced by the solution
proposed in 29 that cancel the influence of the angle estimation error. Our proposed method generates narrow beamwidth beams
that improve the channel estimation accuracy with fixed transmit power and without power allocation. This is because narrow
beams are capable of supporting high beamforming gain which aligned to the true channel directions located randomly within
the communication environment. To maintain the mathematical channel model assumptions, we formulate an adaptive CS-based
algorithm using parallel beams powered by orthogonal sequences which mean an independent selection of channel parameters.
Thus, using the parallel beams allows also to meet the parallel multi-stream mmWave communication. According to the sparsity
nature of the mmWave channel, we adopt the realistic virtual channel representation where the sparsity of mmWave channel is
more structured into a path gains matrix, which means, translate the sparsity in one matrix as criteria of CS tools. Furthermore,
this representation requires less number of grid points as computational resources.
The rest of the paper is arranged as follows. In Section 2, we introduce the system and mmWave channel models. In Section
3, we explain the formulation of the proposed adaptive CS algorithm based on the orthogonal beams training, and we describe
the way to select the beam training for a point-to-point mmWave MIMO system used to estimate mmWave channel. The multi-
resolution hierarchical codebook structure is explained in Section 4 and 5. Simulation results are conducted in Section 6 to
evaluate the design of our codebook with physical metrics, and we introduce also the mutual coherence as a new metric to
evaluate the quality of our codebook (measurement matrix ) used by the OMP algorithm. Thus, we use the SE to evaluate the
performance of the proposed multi-resolution hierarchical codebook design to estimate the mmWave channel, followed by a
hardware complexity and power consumption analysis. Finally, concluding remarks are drawn in Section 7.
We use the following notation: A is a matrix, a denotes a vector, a indicates a scalar, and  is a set. ‖A‖𝐹 stands for the
Frobenius norm of A, whereas A𝑇 ,A𝐻 ,A−1 , are its transpose, Hermitian, and inverse, respectively. [A]𝑖,∶ ([A]∶,𝑖 ) indicate the
i-th row ( i-th column) of the matrix A. A ⊗ B denotes the Kronecker product of A and B. 𝔼[.] denotes expectation.

2 SYSTEM MODEL
Let consider the time division duplexing (TDD) protocol system and downlink techniques for data transmission, we assume the
tx
point-to-point multi-stream mmWave MIMO scenario illustrated in figure 1. The Tx is equipped with 𝑁tx antennas and 𝑁RF RF

FIGURE 1 Hybrid analog/digital architecture.

rx
chains whereas, the Rx has 𝑁rx antennas and 𝑁RF RF chains. The Tx and Rx communicate via 𝑁𝑆 data streams and to ensure
the effectiveness of communications driven by the limited number of RF chains, the number of data streams is constrained to
tx rx
be bounded at the Tx and Rx by 𝑁𝑆 ≤ 𝑁RF ≤ 𝑁tx and 𝑁𝑆 ≤ 𝑁RF ≤ 𝑁rx , respectively.
In practice, the number of RF chains at the Rx is usually less than that of the Tx 19 20 33 . According to the hybrid analog/digital
tx tx
architecture, the precoder is composed of an 𝑁RF × 𝑁𝑆 baseband precoder F𝐵𝐵 followed by an 𝑁tx × 𝑁RF RF precoder F𝑅𝐹 ,
such that F = F𝑅𝐹 F𝐵𝐵 . Therefore, the sampled transmitted signal is:
r = F𝑅𝐹 F𝐵𝐵 x (1)
B.Hadji ET AL 5

[ ]
where x is the 𝑁𝑆 × 1 vector of transmitted pilot symbols, such that 𝔼 xx𝐻 = 𝑁1 I𝑁𝑆 . Since F𝑅𝐹 is implemented using analog
( ) 𝑆
−1
phase shifters, its entries are constrained to satisfy [F𝑅𝐹 ]∶,𝑖 [F𝑅𝐹 ]𝐻
∶,𝑖
= 𝑁 tx
, where (⋅)𝓁,𝓁 denotes the 𝓁 𝑡ℎ diagonal element,
𝓁,𝓁
i.e., all entries are of equal norm. The total power constraint is enforced by letting ‖ ‖2
‖F𝑅𝐹 F𝐵𝐵 ‖𝐹 = 𝑁𝑆 . As the Tx and Rx are
equipped with a uniform linear array (ULA) consisting of a large number of antenna elements, the mmWave channel with ULA
at both Tx and Rx is modeled as a weighted sum of array responses for each path 34 . Each path is composed of two spatial
frequencies which depend on the AoD at the Tx and AoA at the Rx 34 . However, we adopt a widely used Saleh-Valenzuela model
with limited 𝐿 scatterers as in 29 35 36 which is given by:
√
𝑁rx 𝑁tx ∑
𝐿
𝐻= 𝛼𝑖 arx (𝜃𝑖 )a𝐻 (𝜙 )
tx 𝑖
(2)
𝜌 𝑖=1

where 𝜌 denotes the average path-loss between the Tx and Rx, 𝛼𝑖 is the complex gain of the 𝑖𝑡ℎ path and it can define the channel
type (Rayleigh, Rician, or Nakagami). The variables 𝜙𝑖 and 𝜃𝑖 ∈ [0, 2𝜋] are the 𝑖𝑡ℎ path’s azimuth angles of departure or arrival
(AoDs/AoAs) of the Tx and Rx, respectively. The steering vector for the AoD in ULA configuration can be written as :
[ ]𝑇
1 2𝜋 2𝜋
atx (𝜙𝑖 ) = √ 1, 𝑒𝑗 𝜆 𝑑 sin(𝜙𝑖 ) , … , 𝑒𝑗(𝑁tx −1) 𝜆 𝑑 sin(𝜙𝑖 ) (3)
𝑁tx
where 𝜆 denotes the wavelength of the signal, and 𝑑 is the distance between antenna elements. The array response vectors at the
Rx, arx (𝜃𝑖 ), can be written in a similar form.
At the Rx side, to measure and to detect the signal transmitted over the precoder. The Rx uses the combiner W which is
rx rx
composed of an 𝑁rx × 𝑁𝑅𝐹 RF combiner W𝑅𝐹 and 𝑁𝑅𝐹 × 𝑁𝑆 baseband combiner W𝐵𝐵 . The processed received signal y is
given by: √
y = 𝛾W𝐻
𝐵𝐵
W𝐻 𝑅𝐹
HF𝑅𝐹 F𝐵𝐵 x + W𝐻 𝐵𝐵
W𝐻𝑅𝐹
n (4)
where 𝛾 is the average received SNR, n is the 𝑁rx ×1 noise vector with  (0, 𝜎 2 ) entries. In a similar manner
( to RF precoder,
) the
RF combining matrix is implemented using phase shifters with constant modulus constraint such that [W𝑅𝐹 ]∶,𝑖 [W𝑅𝐹 ]𝐻 ∶,𝑖
=
𝓁,𝓁
−1
𝑁rx .
As mentioned above, in the mmWave channel, there are few paths in the channel and large antenna arrays are deployed at the
Tx and Rx. Therefore, according to the mathematical model, the channel is “sparse” in the angular domain 35 .

3 COMPRESSED SENSING BASED CHANNEL ESTIMATION USING ORTHOGONAL

TRAINING SEQUENCE

As mentioned above, to estimate the mmWave channel, the Tx and Rx apply the training hybrid beamforming with different
beamwidths, and the best transmit/receive beams are the beams corresponding to the true path of the channel, which means each
selected transmit/receive beamwidths correspond to each channel angular parameters as clarified in 19 . Therefore, estimating the
mmWave channel is equivalent to recover channel directions with parameters (𝜃𝑖 , 𝜙𝑖 , 𝛼𝑖 ). According to the channel model in 2,
channel parameters are independent, which indicates that the beam vectors corresponding to different paths are well separated.
Due to the overlapping beams, each beam with a certain beamwidth covers the same fraction of the channel environment that is
already covered by the adjacent beam. Hence, the selected channel parameters are repeated. Therefore, the channel model loses
the mathematical assumption since the paths are not modeled independently.
In this paper, we propose using beams powered by orthogonal sequences to get orthogonal beams training to maintain the
channel model assumption because the path parameters will be selected independently thanks to the orthogonality of beams,
which means that the estimation of its best AoA/AoD pair is unique and without any redundancy. Thus, we benefit from the
maximization of the spatial diversity in the mmWave channel because its different paths are selected independently. For this
purpose, we assume that the Tx uses F precoder matrix containing 𝐾 beamforming vector (𝐾 is the number of beamforming
vector in each codebook level) to send parallel training sequences and Rx employs W combiner matrix with the same number
of vectors as precoder matrix to measure and to detect the transmitted training sequences. In this case and based on the system
models described above, the received signals are expressed as:
√
Y = 𝛾W𝐻 HFX + W𝐻 Ξ (5)
6 B.Hadji ET AL

where X denotes the training sequences matrix, such that X = [[x]𝑇1,∶ , [x]𝑇2,∶ , [x]𝑇1,∶ , … , [x]𝑇𝐾,∶ ]𝑇 , where [x]𝑖,∶ is the 𝑖𝑡ℎ transmitted
training sequence with a length of 𝐿𝑝 , and Ξ indicates the noise matrix. To detect and to obtain each transmitted training sequence
separately and independently, we use the multi-stream orthogonal training sequences given as :
⎧ 𝐿 if 𝑖 = 𝑗
⎪ 𝑝
[x]𝑖,∶ [x]𝐻
𝑗,∶
=⎨ (6)
⎪ 0 if 𝑖 ≠ 𝑗
⎩
At the Rx side, [Y]𝑖,∶ denotes the received sequences detected using the 𝑖𝑡ℎ combining vector where the model of this received
sequences can be rewritten as follows:
√ ∑
𝐾
[Y]𝑖,∶ = 𝛾[W]𝐻
∶,𝑖
H [F]∶,𝑗 [x]𝑗,∶ + [W]𝐻
∶,𝑖
Q (7)
𝑗=1

where [W]∶,𝑖 and [F]∶,𝑗 are the 𝑖𝑡ℎ combiner and 𝑗 𝑡ℎ precoder vectors respectively. Q is an 𝑁rx × 𝐿𝑝 noise matrix. After applying
the correlating [Y]𝑖,∶ with 𝑗 𝑡ℎ orthogonal sequences [x]𝐻𝑗,∶
, we can obtain:
z𝑖,𝑗 = [Y]𝑖,∶ [x]𝐻
𝑗,∶
√
= L𝑝 𝛾[W]𝐻 ∶,𝑖
H[F]∶,𝑗 + [W]𝐻
∶,𝑖
Q[x]𝐻
𝑗,∶

√
= L𝑝 𝛾[W]𝐻
∶,𝑖
H[F]∶,𝑗 + [W]𝐻 q
∶,𝑖 𝑖
(8)
where q𝑖 is a 𝑁rx × 1 noise vector. The Rx exploits the 𝐾 combining vectors [W]∶,𝑖 , 𝑖 = 1, 2, … , 𝐾 to detect the transmitted
training sequences. Thereby, the resulting vector is given as:
√
z𝑗 = L𝑝 𝛾W𝐻 H[F]∶,𝑗 + 𝑑𝑖𝑎𝑔(W𝐻 [q1 , … , q𝐾 ]) (9)
where W = [ [W]∶,1 , [W]∶,2 , … , [W]∶,𝐾 ] is the 𝑁rx × 𝐾 combining matrix. If Tx uses [F]∶,𝑗 , 𝑗 = 1, 2, … , 𝐾 to send the 𝐾
training sequences vectors simultaneously, and the Rx uses the same combining matrix W to combine the correlated received
signal.
The resultant received signal matrix by the Rx can be rewritten after the concatenating of 𝐾 vectors z𝑗 , 𝑖 = 1, 2, … , 𝐾 as:
√
Z = L𝑝 𝛾W𝐻 HF + N (10)
where F = [ [F]∶,1 , [F]∶,2 , … , [F]∶,𝐾 ] is the 𝑁tx × 𝐾 precoding matrix used by the Tx, N is an 𝐾 × 𝐾 noise matrix given by
concatenating the 𝐾 noise vectors.
Since traditional channel estimation techniques are still be prohibitive when the communication systems are equipped by a
large number of antennas, CS has been proposed to reduce the training pilot overhead. To leverage CS tools, we vectorize the
resultant matrix Z as: √
z𝑣 = L𝑝 𝛾 𝑣𝑒𝑐(W𝐻 HF) + n𝑣

(𝑎) √
= L𝑝 𝛾 (F𝑇 ⊗ W𝐻 )𝑣𝑒𝑐(H) + n𝑣 (11)
To exploit the sparse nature of the channel, we use the realistic virtual channel model, where the sparsity of the physical channel
model in 2 is translated into H𝛼 , which is a virtual channel matrix 37 , that has only 𝐿 non-zero elements 38 , and can be written as:
H = Arx H𝛼 A𝐻
tx
(12)
− 12
where Atx = [atx (𝜙̃ 1 ), atx (𝜙̃ 2 ), … , atx (𝜙̃ 𝑁tx )] , 𝜙̃ 𝑖 = + 𝑖−1
for 1, 2, … , 𝑁tx , and Arx = [arx (𝜃̃1 ), arx (𝜃̃2 ), … , arx (𝜃̃𝑁rx )],
𝑁tx
1 𝑗−1
𝜃̃𝑗 = − 2 + 𝑁 for 1, 2, … , 𝑁rx are dictionary matrices at the Tx and Rx respectively. Figure 2 shows an example of the amplitude
rx
tx
of path gains in H𝛼 with 𝑁rx = 64, 𝑁RF = 32, and 3 paths. We formulate the channel estimation problem as a sparse problem,
we substitute 12 in 11 to obtain: √
(𝑏)
z𝑣 = L𝑝 𝛾 (F𝑇 ⊗ W𝐻 )(A𝐻 tx
⊗ Arx )h𝛼 + n𝑣 (13)
⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
A𝐷
(𝑐) √
= L𝑝 𝛾 (F𝑇 A𝐻
tx
⊗ W𝐻 Arx )h𝛼 + n𝑣 (14)
B.Hadji ET AL 7

where (𝑎),(𝑏) and (𝑐) follow from the properties Kronecker product [ 39 ,13.3 and 13.26 theorems]. h𝛼 is an 𝑁tx 𝑁rx × 1 sparse
vector which contains only 𝐿 non-zero elements corresponding to path gains. From 13, it is clear that our CS model requires
less number of the grid as computational resources compared with the CS model used in 29 , where A𝐷 is considered as an
𝑁tx 𝑁rx × 𝑁 2 dictionary matrix that consists of the 𝑁 2 column vectors with choosing 𝑁 = 192 to get the best results.

30 gain of path

25
j,index of receive virtual angle

20

15

10

10 20 30 40 50 60
i,index of transmit virtual angle

tx
FIGURE 2 Amplitude of path gains in the virtual channel representation of H𝛼 with 𝑁rx = 64, 𝑁RF = 32, and 3 paths

The final aim is to design hybrid precoders/combiners at the Tx and Rx based on the physical design codebook with low-
complexity to estimate the mmWave channel for the point-to-point MIMO system. For our goal, we adopt the way of the beam
training algorithm that is used in 29 . During the channel estimation algorithm, we assume the availability of a feedback channel
between the Tx and Rx. In general, the training process is divided into 𝑆 stages, then the best training precoding (F𝑠 ) and the
measurements (W𝑠 ) are determined by selecting the ones with maximum SNR. The algorithm starts initially by dividing the
received signal vector into several of partitions, which equivalently divides the AoAs/AoDs range [0, 2𝜋] into several angular
intervals, and design the 𝐾 training precoding and measurement of each stage to sense those partitions. For instance, at the
first stage, the Tx employs two beamforming vectors, F1,1 and F1,2 to transmit the two orthogonal pilot signals generated by
Hadamard Matrix code at two successive time slots according to system protocol (TDD). At the reception side, the Rx employs
two measurement vectors W1,1 and W1,2 which are multiplied with the conjugate of the orthogonal pilot signals transmitted by
the Tx to detect separately the pilot signals carried on each beamforming vector. As mentioned about the selection criteria, the Rx
compares the SNR of the received pilot signals to determine which one has the maximum SNR. This translates into selecting the
partitions which are highly likely to contain the most dominant AoD/AoA combination of the mmWave channel. The best training
precoding (F) and the measurements (W) are used in the later stages with orthogonal pilot signals and their conjugate, where
their corresponding range of AoAs/AoDs are further divided into smaller partitions until the required resolution is achieved.
The high performance of estimating the mmWave channel is the result of a judicious codebook design. Thus, as mentioned
above, the approximation of analog and digital precoding is accomplished by a sparse coding algorithm (OMP). However, the
design problem is equivalent to a sparse decomposition problem, we know that the sparse decomposition problem is trivial if
the dictionary is well designed. For this reason, the design of the codebook is a critical point in estimating the mmWave channel
which is necessary for communications.

4 STRUCTURE OF MULTI-RESOLUTION HIERARCHICAL CODEBOOK DESIGN

In this section, we present the design of the multi-resolution hierarchical beamforming codebook according to architectural con-
straints, particularly the number of RF phase shifters with a small number of quantization bits. In 29 , multi-resolution hierarchical
beamforming codewords are proposed but assumes that RF phase shifters with a large number of quantization bits are available,
which is not the case in practice. Then, the design of the codebook is chosen by the projection of its vectors on dictionary matrix
vectors in a predefined sub-range of AoD, which is equal to the product of G matrix and normalization constant. Each column
of G matrix contains 1′ 𝑠 in this sub-range and zeros elsewhere sub-range. In this perspective, proposed in 29 , the design of the
8 B.Hadji ET AL

𝐾 𝑡ℎ codebook vector in level 𝑠, denoted F(𝑠,𝑘) can be written in the following way:
A𝐻 F
𝐵𝑆 (𝑠,𝑘)
= 𝐶𝑆 G(𝑠,𝑘) (15)
where A𝐵𝑆 is an 𝑁tx × 𝑁 dictionary matrix with 𝑁 > 𝑁tx , such that 𝑁 and 𝑁tx are the number of grid points and the number
of antennas at the Tx, respectively. 𝐶𝑆 is a normalization constant. This strategy constructs an inconsistent system, while the
solution depends mainly on the least square method to obtain the codebook, which is given by F(𝑠,𝑘) = 𝐶𝑆 (A𝐵𝑆 A𝐻
𝐵𝑆
)−1 A𝐵𝑆 G(𝑠,𝑘) .
Therefore, this solution is found by the pseudo-inverse because the dictionary matrix is not square, which means that an error
is introduced into the design of the codebook. The design of our codebook is based on a proposed multi-resolution hierarchical
FSM-KW method to transmit a parallel multiple data streams as illustrated in figure 3.

FIGURE 3 First three levels of the proposed multi-resolution codebook with 𝐾 = 2.

The beams designed by our codebook are generated into a strict beam range and have high directivity without any approximate
solution. Our codebook consists of 𝑆 levels with indices from 𝑠 = 1, 2, … 𝑆 to estimate the most dominant AoD/AoA with
beamwidth vector equal to 2𝜋∕𝐾 𝑆 , such as 𝐾 is the design parameter. Therefore, each level contains 𝐾 𝑠 beamforming vectors. In
each codebook level 𝑠, the beamwidth of each vector used in the corresponding training stage of the adaptive mmWave channel
estimation algorithm is defined by:
𝜑𝑤(𝑠) = 2𝜋∕𝐾 𝑠 (16)
and scanning phase of each vector named 𝜑𝑐(𝑠) is equal to:
𝜑𝑐(𝑠) = 2𝜋∕2𝐾+𝑖 (17)
such that, 𝑖 : index of incrementation, it is initialized to zero at the first level. Finally, 𝜑𝑤(𝑠) and 𝜑𝑐(𝑠) of each vector in each
codebook level 𝑠 are used to compute its codewords F𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒 (𝑚) according to the FSM-KW method as:
−𝛽𝜑𝑐(𝑠) sin(𝜑𝑤(𝑠) 𝛽)
F𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒 (𝑚) = 𝑤(𝑚)𝑒 𝜋𝛽 (18)
𝑚 = 0, 1, ⋯ , 𝑁tx − 1
𝑁
where 𝑤(𝑚) are the samples of the Kaiser window, 𝛽 = 𝑚 − 2tx (see [ 40 , Chapter 19, pp. 638-642] for more details about the
Kaiser window parameters ). Each beam pattern is confined in the desired angular region to avoid the overlap of adjacent beam
patterns to obtain a codebook with small mutual coherence.

5 DESIGN OF HYBRID ANALOG/DIGITAL CODEBOOK

Due to the hardware architecture constraints, the design of codebooks with low-complexity is essential in estimating of mmWave
channel, to reduce the system’s constraints (RF phase shifters with a large number of quantization bits) and to obtain a feasible
architecture (few RF chains number).
In this section, we explain how to generate the beams based on hybrid analog/digital architecture to meet the multi-stream
communication. We focus on the hybrid codebook design at the Tx , a similar manner can be applied to construct the Rx hybrid
B.Hadji ET AL 9

codebook. Therefore, the design of the hybrid analog and digital codewords matrices is given by:

{ } ‖ ‖
F𝑜𝑝𝑡 , F𝑜𝑝𝑡 = argmin ‖F𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒 − F𝑅𝐹 F𝐵𝐵 ‖
𝑅𝐹 𝐵𝐵 ‖ ‖𝐹
[ ]
s.t. F𝑅𝐹 ∶,𝑖 ∈ {[𝐴𝑐𝑎𝑛 ]∶,𝓁 |1 ≤ 𝓁 ≤ 𝑁𝑐𝑎𝑛 }, 𝑖 = 1, 2, ⋯ , 𝑁𝑅𝐹 (19)

‖F𝑅𝐹 F𝐵𝐵 ‖2 = 𝑁𝑆
‖ ‖𝐹

where F𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒 is the global training beams to approach the digital and analog beams, calculated as in the equation (18), and
𝐴𝑐𝑎𝑛 is an 𝑁tx × 𝑁𝑐𝑎𝑛 matrix which defines the finite set of possible analog beam vectors. The 𝐴𝑐𝑎𝑛 matrix is known as an analog
codebook where its columns will be designed to satisfy the analog beam constraints.
In 32 , 𝐴𝑐𝑎𝑛 is based on the development of the IEEE802.15.3c codebook design to generate the analog codebook where the RF
phase shifters with just four-phase values (0, ±𝜋∕2, 𝜋), i.e., with just two quantization bits and without amplitude adjustment.
The 2-bit resolution codebook vectors in 32 have unitary modulus but do not take into account any constraint on its entries,
although the analog codebook vectors are implemented using phase shifters.
In this paper, we chose another way to generate 𝐴𝑐𝑎𝑛 matrix based on the development of the analog codebook proposed in 41
to satisfy analog codebook constraints with the low resolution based on the phase state. We define 𝐴𝑐𝑎𝑛 ∈ 𝑁tx × 𝑁tx with all
its entries having constant modulus amplitude √ 1 as:
𝑁tx

[ ]
(𝑛−1)(𝓁−1)−𝑁tx ∕2
𝑓 𝑙𝑜𝑜𝑟 Φ
1
[𝐴𝑐𝑎𝑛 ]𝑛,𝓁 = √ 𝑗 4
(20)
𝑁tx
𝑛 = 1, 2, ⋯ , 𝑁tx ; 𝓁 = 1, 2, ⋯ , 𝑁tx

where Φ = 2𝑛 , such that 𝑛 is a positive integer number 41 . In our analog codebook, we choose 𝑛 equal to two bits. To solve the
minimization problem in (19), we use the OMP algorithm summarized in Algorithm 1.
In hybrid analog/digital architecture, the digital precoder F𝐵𝐵 needs to be jointly designed with the analog precoder F𝑅𝐹 .
In the first, Algorithm 1 starts by finding the vector [𝐴𝑐𝑎𝑛 ]∶,𝓁 along which the matrix F𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒 has the maximum projection to
define the best F𝑅𝐹 vectors. In step (5), the selected column vector [𝐴𝑐𝑎𝑛 ]∶,𝓁 appends to the RF precoder F𝑅𝐹 . After the dominant
vector is found, and the least-squares solution to F𝐵𝐵 is calculated in step (7), the contribution of the selected vector is removed
in step (8) and the algorithm proceeds to find the column along which the “residual precoding matrix ” F𝑟𝑒𝑠 has the largest
rx
projection. The step (10) ensures that the transmit power constraint is exactly satisfied, and the process continues until all 𝑁RF
beamforming vectors have been selected.

Algorithm 1 Hybrid analog-digital codewords approximation via OMP

Require: F𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒 , 𝐴𝑐𝑎𝑛
1: F𝑅𝐹 = [ ]
2: F𝑟𝑒𝑠 = F𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒
3: for 𝑖 ⩽ N𝑡𝑥
𝑅𝐹
4: Ψ = 𝐴𝐻 F
𝑐𝑎𝑛 𝑟𝑒𝑠
5: 𝑘 = argmax𝓁=1,2,⋯,𝑁𝑐𝑎𝑛 (ΨΨ𝐻 )𝓁,𝓁
6: F𝑅𝐹 = [F𝑅𝐹 |𝐴𝑘𝑐𝑎𝑛 ]
7: F𝐵𝐵 = (F𝐻 F )−1 F𝐻
𝑅𝐹 𝑅𝐹
F
𝑅𝐹 𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒
F𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒 −F𝑅𝐹 F𝐵𝐵
8: F𝑟𝑒𝑠 = F
‖ 𝑀𝑢𝑙𝑠𝑡𝑎𝑔𝑒 −F𝑅𝐹 F𝐵𝐵 ‖𝐹
9: end for √
F
10: F𝐵𝐵 = 𝑁𝑠 F 𝐵𝐵
‖ 𝑅𝐹 F𝐵𝐵 ‖
11: return F𝑅𝐹 , F𝐵𝐵
10 B.Hadji ET AL

6 SIMULATION RESULTS
In this section, we first present some numerical results to evaluate the design of the proposed multi-resolution hierarchical
codebook described in previous sections and compared with the hybrid codebook proposed in 29 and 32 . Second, we use SE as a
performance to evaluate the results of channel estimation with our codebook and the codebook proposed in 29 . Then, we analyze
the hardware complexity and power consumption.
Hybrid design in [29] - Stage 1 Hybrid design in [32] - Stage 1 Proposed hybrid design - Stage 1
90 2 90 2 90 2.3194
120 60 120 60 120 60

150 1 30 150 1 30 150 1.1597 30

180 0 180 0 180 0

210 330 210 330 210 330

240 300 240 300 240 300

270 270 270

(a) Codebooks for stage 1.

Hybrid design in [29] - Stage 2 Hybrid design in [32] - Stage 2 Proposed hybrid design - Stage 2
90 2.3702 90 2.0654 90 4
120 60 120 60 120 60

150 1.1851 30 150 1.0327 30 150 2 30

180 0 180 0 180 0

210 330 210 330 210 330

240 300 240 300 240 300

270 270 270

(b) Codebooks for stage 2.

Hybrid design in [29] - Stage 3 Hybrid design in [32] - Stage 3 Proposed hybrid design - Stage 3
90 4 90 4 90 4
120 60 120 60 120 60

150 2 30 150 2 30 150 2 30

180 0 180 0 180 0

210 330 210 330 210 330

240 300 240 300 240 300

270 270 270

(c) Codebooks for stage 3.

FIGURE 4 Comparisons of the adjacent beams patterns for the first three stages of codebook designed with the method proposed
in 29 , the method proposed in 32 and our method based on multi-resolution hierarchical FSM-KW design.

6.1 Performance of codebook design

In this subsection, we compare beam patterns design generated by our proposed method with beam patterns designed by the
solutions proposed in 29 and 32 . For our codebook, we assume a ULA antenna configuration, 𝜆∕2-spaced. The Tx is equipped with
tx
𝑁tx = 64 and 𝑁𝑅𝐹 = 16 RF chains, where the analog codebook is generated as expressed in (20) based on the phase state with
𝑛 = 2-bits. The same architecture parameters are used to generate the beams proposed in 32 , and the RF phase shifters with two
B.Hadji ET AL 11

quantization bits while the hybrid codebooks in 29 are generated by the same number of ULA antennas, but with 𝑁𝑅𝐹 tx
= 32 RF
chains and 7-bits RF phase shifters. Figure 4 shows the beam pattern comparison of our proposed codebook and the codebooks
proposed in 32 , 29 . From this figure, we find that our proposed method generates flatter beam patterns with limited overlap between
adjacent beams into a strict beam range. Different physical metrics are needed to analyze the beam design. Figure 5 shows the
percentage of gain, the overlap between the adjacent beams, and the normalized peak-valley (P-V) value. As we know, the high
beamforming gain is still necessary to improve the mmWave communication, as each increasing of beamforming gain means
an increase in the SE due to the direct relationship. The P-V is a metric commonly adopted in optics to characterize the surface
flatness of laser beams. It is calculated as the difference between the “highest” and “lowest” values on the sector normalized
to the mean value and expressed in percentage terms 32 . As shown, our proposed hybrid analog-digital beam patterns have high
physical metrics with low-complexity compared to the method proposed in 29 .

Gain in %
100
Hybrid design in [29]
Hybrid design in [32]
75 Proposed hybrid design
%

50

25

0
Level 1 Level 2 Level 3
Levels

(a) Gain.

Overlap between adjacent beams in %

100
Hybrid design in [29]
Hybrid design in [32]
75 Proposed hybrid design
%

50

25

0
Level 1 Level 2 Level 3
Levels

(b) Overlap between adjacent beams.

peak-valley value (P-V) %
100
Hybrid design in [29]
Hybrid design in [32]
75 Proposed hybrid design
%

50

25

0
Level 1 Level 2 Level 3
Levels

(c) Normalized peak-valley value (P-V).

FIGURE 5 Physical metrics of beam patterns in terms of (a) percentage of gain,(b) overlap between adjacent beams and (c)
normalized peak-valley (P-V) value.
As mentioned above, the mmWave channel has sparse nature. In fact, the design of hybrid precoders/combiners is based on
the OMP algorithm. The performance and complexity of this algorithm are based on codebook quality. For this aim, we used the
codebook orthogonality constant as a new metric to evaluate the codebook design through the mutual coherence, where denoted
12 B.Hadji ET AL

as 𝜇(𝐷) and it can be given by:

𝜇(𝐷) = max|⟨𝑔𝑖 , 𝑔𝑗 ⟩|
(21)
𝑖≠𝑗
where 𝐷 ∈ ℂ𝑚×𝑁 is codebook with columns 𝑔1 , … 𝑔𝑁 (unitary vectors). Figure 6 shows the mutual coherence results of our
proposed codebook and the codebook designed in 29 which are used to estimate the mmWave channel. The analysis of sparse
algorithms usually involves a quality of the measurement matrix, and its smaller mutual coherence contributes to the good
performance of these algorithms. Our proposed codebook displays a better mutual coherence results to guarantee the success of
any sparse algorithms used to design the analog/digital precoder and combiner.

0.08
hybrid codebook with hierarchical FSM-KW
0.07 hybrid codebook proposed in [29]
Normalized coherence constant

0.06

0.05

0.04

0.03

0.02

0.01

0
0 5 10 15 20 25 30
Codeworks

FIGURE 6 Mutual coherence comparison of our proposed codebook and the codebook proposed in 29 .

6.2 Downlink channel estimation comparison

In this subsection, we provide the simulation results to analyze the effect of hardware complexity of the hybrid architecture,
namely the number of RF chains at each end and the number of quantization bits on the performance of the mmWave channel
estimation by using our multi-resolution hierarchical codebook design and codebook design proposed in 29 and compared with
the optimal unconstrained precoding based on the SVD decomposition of the channel matrix (perfect case). Spectral efficiency
results are used as metrics to evaluate this performance. In hybrid analog/digital, when Gaussian symbols are transmitted over
the mmWave channel, the achieved SE is given by 19 :
𝛾 −1 𝐻 𝐻
𝑅 = log2 (|I𝑁𝑠 + R W𝐵𝐵 W𝑅𝐹 HF𝑅𝐹 F𝐵𝐵 F𝐻 F𝐻 H𝐻 W𝑅𝐹 W𝐵𝐵 |) (22)
𝑁𝑠 𝑛 𝐵𝐵 𝑅𝐹

where R𝑛 = 𝜎𝑛2 W𝐻 𝐵𝐵
W𝐻𝑅𝐹
W𝑅𝐹 W𝐵𝐵 is the noise covariance matrix at the Rx. As described in Section 2, the point-to-point
mmWave MIMO system and the channel model are used in this simulation. The Tx and Rx are equipped with 𝑁tx = 64 and
𝑁rx = 32 antennas, where the separation distance between the Tx and Rx is equal to 50 meters, and the path-loss exponent of n
=3 for shadowed urban cellular radio environment. The downlink channel has 𝐿 = 3 paths and Rayleigh distributed, the azimuth
AoAs and AoDs are uniformly distributed in [0, 2𝜋]. The mmWave MIMO system operates at 28 GHz with the transmission
bandwidth 𝐵𝑤 = 100 MHz and the noise power is 𝜎 = −173 + 10 log10 (𝐵𝑤 ).
For comparison purposes, we evaluate the average SE by varying the number of RF chains and the number of bit phase
shifters. Figure 7 shows the average SE comparison in mmWave MIMO system between our codebook used to estimate the
mmWave channel, codebook design proposed in 29 , and unconstrained precoders and combiners. For our proposed codebook, the
tx rx
Tx and Rx are equipped with 𝑁𝑅𝐹 = 10 and 𝑁𝑅𝐹 = 4 RF chains respectively, with 2-bit resolution analog phase shifters and 3
codebook stages. As shown in figure 7 our proposed codebook design has a better performance compared to the result obtained
by the method proposed in 29 , which is achieved with 𝑁𝑅𝐹
tx
= 24 and 𝑁𝑅𝐹 rx
= 12 RF chains at the Tx and Rx respectively, with
5 quantization bits and 6 codebook stages. Then, as we increase the number of RF chains and the quantization bits number, the
tx rx
average SE is augmented. In the proposed codebook, we assume that the Tx and Rx are equipped with 𝑁𝑅𝐹 = 16 and 𝑁𝑅𝐹 =8
29
RF chains with only 2-bit resolution and 3 codebook stages. Whereas, for the proposed method in , we choose the mmWave
MIMO system with 7-bits resolution analog phase shifters and 6 codebook stages, where the Tx and Rx are equipped with
tx rx
𝑁𝑅𝐹 = 32 RF chains and 𝑁𝑅𝐹 = 16 RF chains respectively. Besides, the higher complexity hardware using by the proposed
29
method in , our proposed method achieves high SE and better results with less number of codebook stages.
From figure 7, we found that our codebook designed with low complexity, (i.e., 2 quantization bits and few numbers of RF
chains) has a good performance and outperforms the results obtained by 29 because the precoders and combiners generated by
B.Hadji ET AL 13

15
Perfect CSI
HP (q=7,N RFrx =16,NRFtx =32)[29]
Proposed HP (q=2,N RFrx =8,NRFtx =16)

Spectral Efficiency (bps/ Hz)

HP (q=5, NRFrx =12,NRFtx =24)[29]
Proposed HP (q=2, NRFrx =4,NRFtx =10)
10

0
-40 -35 -30 -25 -20 -15 -10 -5 0
SNR (dB)
FIGURE 7 Average achievable SE performance comparison of mmWave channel estimation when there are L = 3 paths using
our codebook, codebook design proposed in 29 , and optimal unconstrained precoding.

our proposed codebook have suitable gains to increase the SE and to compensate the low complexity. Also, the orthogonal
training beamforming vectors allow selecting the channel parameters independently, which maximizes the spatial diversity. Our
proposed method is designed to meet parallel multi-stream data. Consequently, the efficiency of the system increases with the
number of the data stream. The average achievable SE is plotted versus received SNR over 1000 channel realizations.

6.3 Hardware complexity and power consumption analysis

In hybrid architecture, the hardware complexity is associated with the RF chain and bit quantization numbers. Therefore, their
numbers should be reduced to a minimum due to cost and power consideration. Table 1 shows the hardware complexity com-
parison at the 𝑇𝑥 side based on step 4 and 7 from algorithm 1 using our proposed codebook and the codebook proposed in 29 .
tx
The computational complexity is a function of the number of antennas 𝑁tx , the number of RF chains 𝑁RF , the number of data
𝑞
𝑁𝑆 , and 𝐿 = 2 with 𝑞 defined as the bit quantization numbers. The computational complexity considered here is expressed in
terms of the total number of flops 1 . In real arithmetic, a multiplication followed by an addition needs 2 flops. With complex-
valued quantities, a multiplication followed by an addition needs 8 flops. Thus, the complexity of complex matrix multiplication
is nearly 4 times its real counterpart. For a complex 𝑚 × 𝑛 matrix A, we summarize the total flops needed for the matrix opera-
tions as shown the following: (1) Multiplication of 𝑚 × 𝑛 and 𝑛 × 𝑝 complex matrices = 8𝑚𝑛𝑝 − 2𝑚𝑝. (2) Pseudo-inversion of
an 𝑚 × 𝑛 (𝑚 ≤ 𝑛) complex matrix = 34 𝑚3 + 7𝑚2 𝑛 − 𝑚2 − 2𝑚𝑛.

TABLE 1 Computational complexity at TX side for hybrid precoding using our proposed codebook and the codebook proposed
in 29 .

Hybrid beamforming method Complexity (flops)

2 4
Hybrid precoding in 29 tx
𝑁RF tx
(2𝑁𝑠 (4𝑁tx − 1)(𝑁RF + L) + 𝑁tx ( 3 𝑁tx + 7𝑁RFtx tx
− 1) − 𝑁tx (𝑁tx + 2𝑁RF ))
tx tx 2 4 tx tx
Proposed hybrid precoding 𝑁RF (2𝑁𝑠 (4𝑁tx − 1)(𝑁RF + 𝑁tx ) + 𝑁tx ( 3 𝑁tx + 7𝑁RF − 1) − 𝑁tx (𝑁tx + 2𝑁RF ))

Figure 8 illustrates the complexity of our hybrid beamforming method compared with 29 . From this figure, we can observe
that the complexity of algorithm 1 using our proposed codebook is significantly lower compared to 29 . It is clear that the method
in 29 performs an exhaustive search over the codebook of RF beamformer that consequently results in higher complexity due to

1A flop stands for floating point operation. Operations such as addition, multiplication, subtraction, division and compare are considered as one flop.
14 B.Hadji ET AL

the quantization bits number. Besides, the RF chains number increases the number of iterations which can indeed be prohibitive
in terms of complexity in practice. Power consumption circuits can poses difficulties in implementing the hybrid architecture,
9.5

log10 (complexity)
8.5

HP (q=7,N RFtx =32)[29]

7.5
Proposed HP (q=2,N RFtx =16)
HP (q=5,N RFtx =24)[29]
7
Proposed HP (q=2,N RFtx =10)

6.5
100 150 200 250 300
Number of Antennas Ntx

FIGURE 8 Computational complexity comparison between our proposed codebook and codebook design proposed in 29 , with
𝑁𝑆 = 3
where the beamforming gain depends on the communication level and the possibility of reconfiguring the analog beamforming
network. Fully and partially-connected are the most common types of hybrid architectures. Although partially-connected hybrid
architectures reduce implementation complexity, the fully-connected hybrid architecture provides full beamforming gain and
higher beam steering flexibility. For this reason, the hybrid architecture used in our method and the proposed method in 29
are based on fully-connected hybrid implementation. Additionally, hybrid precoding/combination designs have a significant
impact on power consumption due to computational complexity. Hence, the power consumed for the computation of the hybrid
precoding/combination matrices of fully-connected hybrid architecture can be modeled as 42 :
𝐶
𝜇𝐻𝐹 𝐶 = 𝑁𝑅𝐹 tx tx
𝑃𝑅𝐹 + 2𝑃𝐿𝑂 + 𝑁𝑅𝐹 rx rx HFC
𝑃𝑅𝐹 + 𝑃𝐿𝑜𝑠𝑠 + 𝑁rx (𝑁𝑅𝐹rx
+ 1)𝑃𝐿𝑁𝐴 + HFC (23)
𝐷𝑁𝑆
where 𝐶HFC is the computational complexity of the algorithm 1 expressed in the number of the required flops and 𝐷 is the
tx rx
transceiver’s computational efficiency in flops/W. 𝑃𝑅𝐹 = 𝑃𝐷𝐴𝐶 + 𝑃𝑚𝑖𝑥 + 𝑃𝐿𝑃 𝐹 and 𝑃𝑅𝐹 = 𝑃𝐴𝐷𝐶 + 𝑃𝑚𝑖𝑥 + 𝑃𝐿𝑃 𝐹 is the power
consumed in a transmitter’s and a receiver’s RF chain, respectively, 𝑃𝐷𝐴𝐶 , 𝑃𝑚𝑖𝑥 , 𝑃𝐿𝑃 𝐹 , 𝑃𝐴𝐷𝐶 , 𝑃𝐿𝑂 and 𝑃𝐿𝑁𝐴 is the power
consumed by the Digital-to-Analog Converter (DAC), the mixer, Low-pass filter, the Analog-to-Digital Converter (ADC), local
oscillator and the low noise amplifier (LNA), respectively. The losses of RF hardware are introduced in the analysis of such
hybrid systems due to the presence of the power splitters, the phase shifters, and the power combiners. On the transmitter side,
these losses were quantified based on the S-parameter representation of the RF hardware component and it was shown to be
HFC tx
equal to 𝑃𝐿𝑜𝑠𝑠 = 𝐿𝑆 𝐿𝑃 𝑆 𝑁tx 𝐿𝐶 𝑁𝑅𝐹 , where 𝐿𝑆 , 𝐿𝑃 𝑆 , and 𝐿𝐶 are the static power losses due to the splitters, phase shifters, and
42
combiners . The power consumption of the different electronic components used in the simulation is summarized in table 2.
TABLE 2 Simulation parameters

Component Notation Value

Local oscillator 42 𝑃𝐿𝑂 5 mW
Mixer 42 𝑃𝑚𝑖𝑥 19 mW
Low-pass filter 42 𝑃𝐿𝑃 𝐹 14 mW
ADC,DAC 42 𝑃𝐴𝐷𝐶,𝐷𝐴𝐶 200 mW
Low-Noise Amplifier 43 𝑃𝐿𝑁𝐴 20 mW
Two-way power divider 44 𝐿𝑆 0.6 dB
Two-way power combiner 44 𝐿𝐶 0.6 dB + 6dB
Phase-shifter 45 𝐿𝑃 𝑆 2 dB
transceiver’s computational efficiency 42 𝐷 5 Gflops/W

Figure 9 compares the power consumption results of our proposed codebook design and the codebook proposed in 29 by varying
the number of antennas 𝑁tx for a fixed antennas number at Rx. Due to the low-complexity used in the design of our proposed
B.Hadji ET AL 15

codebook, it is easy to see the proposed method has less power consumption than and the method proposed in 29 . Therefore, our
proposed method can implement fully-connected hybrid architecture with low complexity, less power consumption, and low
cost.
49
HP (q=7,N RFrx =16,N RFtx =32)[29]
Proposed HP (NRFrx =8,N RFtx =16)
48
HP (q=5,N =12,N =24)[29]
RFtx RFtx
47 Proposed HP (N =4,N =10)

power consumption (dB)

RFrx RFtx

46

45

44

43

42

41
100 150 200 250 300
Number of Antennas Ntx

FIGURE 9 Power consumption varying versus the number of antennas 𝑁tx

7 CONCLUSIONS

In this paper, we developed a multi-resolution hierarchical codebook to meet the communication of multiple data streams in
the multipath environment. The training beamforming vectors generated by our method are based on the physical design and
low-complexity hybrid analog/digital architecture with different beamwidths. Then, we exploited these training beamforming
vectors to estimate the mmWave channel. Simulation results compared with the codebook design proposed in 29 used in mmWave
channel estimation has verified that our proposed method has two advantages: (i) can achieve the better performance of channel
estimation, and (ii) can reduce the communication cost.

References

1. Rappaport TS, Sun S, Mayzus R, et al. Millimeter wave mobile communications for 5G cellular: It will work!. IEEE access
2013; 1(14): 335–349.

2. Uwaechia AN, Mahyuddin NM. A Comprehensive Survey on Millimeter Wave Communications for Fifth-Generation
Wireless Networks: Feasibility and Challenges. IEEE Access 2020; 8: 62367–62414.

3. Yassein MB, Aljawarneh S, Al-Sadi A. Challenges and features of IoT communications in 5G networks. 2017 International
Conference on Electrical and Computing Technologies and Applications (ICECTA) 2017: 1–5.

4. Wang B, Dai L, Wang Z, Ge N, Zhou S. Spectrum and energy-efficient beamspace MIMO-NOMA for millimeter-wave
communications using lens antenna array. IEEE Journal on Selected Areas in Communications 2017; 35(10): 2370–2382.

5. Wei Z, Yuan J, Ng DWK, Elkashlan M, Ding Z. A survey of downlink non-orthogonal multiple access for 5G wireless
communication networks. arXiv preprint arXiv:1609.01856 2016.

6. Karjalainen J, Nekovee M, Benn H, Kim W, Park J, Sungsoo H. Challenges and opportunities of mm-wave communication
in 5G networks. 2014 9𝑡ℎ international conference on cognitive radio oriented wireless networks and communications
(CROWNCOM) 2014: 372–376.

7. Gupta A, Jha RK. A survey of 5G network: Architecture and emerging technologies. IEEE access 2015; 3: 1206–1232.

8. Fischione C, Koutsonikolas D, Rangan S, et al. Guest editorial millimeter-wave networking. IEEE Journal on Selected
Areas in Communications 2019; 37(12): 2649–2652.
16 B.Hadji ET AL

9. Noh S, Zoltowski MD, Love DJ. Multi-resolution codebook and adaptive beamforming sequence design for millimeter wave
beam alignment. IEEE Transactions on Wireless Communications 2017; 16(9): 5689–5701.

10. Rappaport TS, Xing Y, MacCartney GR, Molisch AF, Mellios E, Zhang J. Overview of millimeter wave communica-
tions for fifth-generation (5G) wireless networks With a focus on propagation models. IEEE Transactions on Antennas and
Propagation 2017; 65(12): 6213–6230.

11. Heath RW, Gonzalez-Prelcic N, Rangan S, Roh W, Sayeed AM. An overview of signal processing techniques for millimeter
wave MIMO systems. IEEE journal of selected topics in signal processing 2016; 10(3): 436–453.

12. Tsinos CG, Chatzinotas S, Ottersten B. Hybrid analog-digital transceiver designs for multi-user MIMO mmwave cognitive
radio systems. IEEE Transactions on Cognitive Communications and Networking 2019; 6(1): 310–324.

13. Stirling-Gallacher RA, Rahman MS. Multi-user MIMO strategies for a millimeter wave communication system using hybrid
beam-forming. 2015 IEEE International Conference on Communications (ICC) 2015: 2437–2443.

14. Banerjee A, Vaesen K, Visweswaran A, et al. Millimeter-Wave Transceivers for Wireless Communication, Radar, and
Sensing. 2019 IEEE Custom Integrated Circuits Conference (CICC) 2019: 1–11.

15. Arora A, Tsinos CG, Rao BSMR, Chatzinotas S, Ottersten B. Hybrid Transceivers Design for Large-Scale Antenna Arrays
Using Majorization-Minimization Algorithms. IEEE Transactions on Signal Processing 2019; 68: 701–714.

16. Zhang X, Molisch AF, Kung SY. Variable-phase-shift-based RF-baseband codesign for MIMO antenna selection. IEEE
Transactions on Signal Processing 2005; 53(11): 4091–4103.

17. Ardah K, Fodor G, Silva YC, Freitas WC, Almeida dAL. Hybrid analog-digital beamforming design for SE and EE
maximization in massive MIMO networks. IEEE Transactions on Vehicular Technology 2019; 69(1): 377–389.

18. Wang Z, Li M, Liu Q, Swindlehurst AL. Hybrid precoder and combiner design with low-resolution phase shifters in
mmWave MIMO systems. IEEE Journal of Selected Topics in Signal Processing 2018; 12(2): 256–269.

19. El Ayach O, Rajagopal S, Abu-Surra S, Pi Z, Heath RW. Spatially sparse precoding in millimeter wave MIMO systems.
IEEE transactions on wireless communications 2014; 13(3): 1499–1513.

20. Xie T, Dai L, Gao X, Shakir MZ, Wang Z. GMD-based hybrid precoding for millimeter-wave massive MIMO systems.
arXiv preprint arXiv:1605.04686 2016.

21. Yu X, Shen JC, Zhang J, Letaief KB. Alternating minimization algorithms for hybrid precoding in millimeter wave MIMO
systems. IEEE Journal of Selected Topics in Signal Processing 2016; 10(3): 485–500.

22. Zhu F, He S, Li R, Huang Y, You X. Energy-efficient hybrid precoding for broadband millimeter wave communication
systems. 2017 9th International Conference on Wireless Communications and Signal Processing (WCSP) 2017: 1–5.

23. He S, Qi C, Wu Y, Huang Y. Energy-efficient transceiver design for hybrid sub-array architecture MIMO systems. IEEE
Access 2016; 4: 9895–9905.

24. Zhang Y, Cui Q, Ni W, Zhang P. Energy-efficient transmission of hybrid array with non-ideal power amplifiers and circuitry.
IEEE Transactions on Wireless Communications 2018; 17(6): 3945–3958.

25. Fang C, Makki B, Li J, Svensson T. Coordinated hybrid precoding for energy-efficient millimeter wave systems. 2018 IEEE
19𝑡ℎ International Workshop on Signal Processing Advances in Wireless Communications (SPAWC) 2018: 1–5.

26. Alkhateeb A, Heath RW, Leus G. Achievable rates of multi-user millimeter wave systems with hybrid precoding. 2015
IEEE International Conference on Communication Workshop (ICCW) 2015: 1232–1237.

27. Su X, Zeng J, Li J, et al. Limited feedback precoding for massive MIMO. International Journal of Antennas and Propagation
2013; 2013.
B.Hadji ET AL 17

28. Yang J, Wei Z, Li N, Sang L, Li P. Enhanced multi-resolution hierarchical codebook design for adaptive compressed sensing
based millimeter wave channel estimation. 2016 IEEE/CIC International Conference on Communications in China (ICCC)
2016: 1–5.

29. Alkhateeb A, El Ayach O, Leus G, Heath RW. Channel estimation and hybrid precoding for millimeter wave cellular systems.
IEEE Journal of Selected Topics in Signal Processing 2014; 8(5): 831–846.

30. Xiao Z, Xia P, Xia XG. Codebook design for millimeter-wave channel estimation with hybrid precoding structure. IEEE
Transactions on Wireless Communications 2016; 16(1): 141–153.

31. Noh S, Zoltowski MD, Love DJ. Multi-resolution codebook based beamforming sequence design in millimeter-wave
systems. 2015 IEEE Global Communications Conference (GLOBECOM) 2015: 1–6.

32. De Donno D, Beltrán JP, Giustiniano D, Widmer J. Hybrid analog-digital beam training for mmWave systems with low-
resolution RF phase shifters. 2016 IEEE International Conference on Communications Workshops (ICC) 2016: 700–705.

33. Seleem H, Sulyman AI, Alsanie A. Hybrid precoding-beamforming design with Hadamard RF codebook for mmWave
large-scale MIMO systems. IEEE Access 2017; 5: 6813–6823.

34. Manoj A, Kannu AP. Multi-user millimeter wave channel estimation using generalized block OMP algorithm. 2017 IEEE
18th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC) 2017: 1–5.

35. Mendez-Rial R, Gonzalez-Prelcic N, Heath RW. Adaptive hybrid precoding and combining in mmWave multiuser MIMO
systems based on compressed covariance estimation. 2015 IEEE 6th International Workshop on Computational Advances
in Multi-Sensor Adaptive Processing (CAMSAP) 2015: 213–216.

36. Liang L, Xu W, Dong X. Low-complexity hybrid precoding in massive multiuser MIMO systems. IEEE Wireless
Communications Letters 2014; 3(6): 653–656.

37. Sayeed AM. Deconstructing multiantenna fading channels. IEEE Transactions on Signal processing 2002; 50(10): 2563–
2579.

38. Seo J, Sung Y, Lee G, Kim D. Pilot beam sequence design for channel estimation in millimeter-wave MIMO systems: A
POMDP framework. 2015 IEEE 16th International Workshop on Signal Processing Advances in Wireless Communications
(SPAWC) 2015: 236–240.

39. Laub AJ. Matrix analysis for scientists and engineers. 91. Siam . 2005.

40. Orfinidas SJ. (2013).Electromagnetic waves and antennas. [Online]. Available:. https://www.ece.rutgers.edu/~orfanidi/
ewa/; .

41. Abdo AMA, Zhao X, Zhang R, Zhang Y. Codebook metrics evaluation for millimeter wave communications by antenna
array response and signal to noise ratio. In: . 17. EDP Sciences. ; 2018: 03011.

42. Tsinos CG, Maleki S, Chatzinotas S, Ottersten B. On the energy-efficiency of hybrid analog–digital transceivers for single-
and multi-carrier large antenna array systems. IEEE Journal on Selected Areas in Communications 2017; 35(9): 1980–1995.

43. Méndez-Rial R, Rusu C, Alkhateeb A, González-Prelcic N, Heath RW. Channel estimation and hybrid combining for
mmWave: Phase shifters or switches?. 2015 Information Theory and Applications Workshop (ITA) 2015: 90–97.

44. Ribeiro LN, Schwarz S, Rupp M, Almeida dAL. Energy efficiency of mmWave massive MIMO precoding with low-
resolution DACs. IEEE Journal of Selected Topics in Signal Processing 2018; 12(2): 298–312.

45. Ha VN, Nguyen DH, Frigon JF. System Energy-Efficient Hybrid Beamforming for mmWave Multi-user Systems. IEEE
Transactions on Green Communications and Networking 2020.
18 B.Hadji ET AL

AUTHOR BIOGRAPHY

Baghdad Hadji . received the State Engineering degree (Ingéniorat d’état) in Biomedical Electronics Engi-
neering in 2010 from the University of Tlemcen, Algeria, and the Magister degree in electronics (advanced
techniques in signal processing) in 2016 from the Military Polytechnic School (EMP), Algiers. He is cur-
rently pursuing the Ph.D. degree in Telecommunication, at the University of Science and Technology Houari
Boumediene (USTHB) Algiers. His research interests are focused on signal processing, 5G Communication
Systems, advanced multi-access techniques, compressive sensing, channel estimation

How to cite this article: Channel Estimation Based on Low-Complexity Hierarchical Codebook Design for Millimeter-Wave
MIMO Systems