Hybrid Precoding Algorithm for Millimeter-

Wave Massive MIMO Systems

Mohammad Assaf Oleg G Ponomarev
Radiophysic department Radiophysic department
Tomsk State University Tomsk State University
Tomsk, Russian Federation Tomsk, Russian Federation
e-mail: [email protected] e-mail: [email protected]

Abstract— In this paper, a high-performance, low- in order to improve the performance of hybrid
complexity hybrid precoding method for millimeter-wave beamforming. In [6], the scattering characteristics of
multiple input multiple output transceiver systems is mmWave channels are exploited to construct RF/baseband
proposed. We demonstrate that, until the number of RF precoders, where the concept of orthogonal matching
chains is greater than the number of streams, the angle pursuit (OMP) is used to approximate the RF precoder as a
information of the analog phase shifters can be directly linear combination of beam steering vectors. A low-
derived from the optimal unconstrained precoder without the complexity solution to the hybrid precoding technique OMP
need for any additional information or decomposition is proposed in [7], where the concept of combining
techniques, which in turn reduces complexity and the amount
orthogonal matching and local search (OM + LS)
of feedback information. The proposed scheme works in an
iterative manner in order to minimize the residual error
techniques reduces significantly the search space in the
between the optimal matrix and the product precoding array manifold and thus leads to a lower complexity
matrices. Simulation results show that the proposed scheme procedure. A near-optimal iterative hybrid-precoding
can achieve almost the same performance as the optimal full- scheme based on subarray structure with low complexity
baseband design when the number of streams is less than the has been proposed in [8]. It was shown in [9] that the Gram
number of RF chains, while maintaining the performance of matrix of the frequency selective channel can be divided
most relevant works when the number of streams is equal to into frequency-flat and frequency-selective components,
the number of RF chains. which can be used for the design of analog and baseband
precoders, where the common analog precoder and
Keywords— Millimeter wave communication, massive combiner were produced to match the frequency-flat
MIMO, hybrid beamforming. channel subspaces. The proposed algorithm in [10]
iteratively updates the phases of the phase-shifters in the RF
I. INTRODUCTION precoder and combiner to minimize the weighted sum of
The use of millimeter (mm) wave spectrum (typically squared residuals between the optimal full-baseband design
considered 30–300 GHz) is potentially one of the key and the hybrid design. The proposed scheme in [11] takes
technologies to achieve high capacity performance in the advantage of the correlation values between the
5G/6G mobile cellular systems. It can offer multiple eigenvectors of the array response matrix and the optimal
gigabit-per-second data rates with the ability to support a precoder to choose the most relevant basis vectors, which
huge number of devices [1][2]. The small wavelength in the are used to reconstruct the representative matrix of the array.
mmWave band allows using massive multiple-input In [12], a hybrid RF and baseband precoding/combining
multiple-output (MIMO) technology, where more antennas approach for multiple stream transmission systems was
are packed in a comparably compact area, allowing for the developed by directly decomposing the predesigned
creation of high-gain antenna arrays that can compensate for unconstrained digital precoder/combiner matrix, where the
the increased propagation losses [3]. However, in massive non-convex matrix decomposition problem is turned into a
MIMO systems, the amplitudes and phases of the complex series of convex sub-problems. In [13], a hybrid precoding
transmit symbols are traditionally adjusted at the baseband method based on the singular value decomposition (SVD)
before being unconverted to the passband. The radio technique was developed, where the right singular vectors
frequency (RF) chains include the analog radio wave/digital of the channel matrix can be used to directly construct the
baseband conversion, mixer, and power amplifier. The RF precoding matrix. In [14] through the operation of the
number of RF chains in this configuration must be exactly SVD technique on the fully digital precoding matrix, the
equal to the number of antenna elements because all of the eigenvectors with the same number of the data streams are
RF chains' outputs are linked to the antenna elements. Due chosen out to obtain the angle information of the analog
to the enormous number of RF chains in this configuration, phase shifters. To obtain the remaining required
the manufacturing cost and energy consumption of such a eigenvectors, the relativity information between the left
huge MIMO system make it undesirable. Small numbers of decomposed matrix and the optimal fully digital precoding
RF chains should be employed in huge MIMO systems to matrix is exploited. In this way, the resultant hybrid
solve the aforementioned problem. This can be achieved by analog/digital precoding is achieved in a low complexity
using cost - effective variable phase shifters to apply manner. The concept of "equivalent channel" has been
analogue precoding [4]. This structure is commonly referred exploited in [15] to design a hybrid precoder and combiner
to as hybrid processing. Analog precoding is often jointly to improve the system’s spectral efficiency. A two-
implemented using phase shifters [5]. This limits the RF stage successive method is proposed in [16] to design the
precoder's components to having a constant modulus. analog precoder and combiner jointly, and the
corresponding equivalent channel is constructed. Then, the
Over the past few years, many studies have been digital precoding and combining operations are realized
conducted to optimize the precoding and combining weights
 power. Here the elements in 𝒔 are assumed to be
1
independent and have been normalized to have energy .
𝑁𝑠
The dimensions of the two matrices 𝐅𝐁𝐁 and 𝐅𝐑𝐅 are 𝑁𝑡𝑅𝐹 ×
𝑁𝑠 and 𝑁𝑡 × 𝑁𝑡𝑅𝐹 , respectively. The precoded signal vector
√𝜌𝐅𝐑𝐅 𝐅𝐁𝐁 𝒔 is then transmitted over the mmWave
narrowband propagation channel H to get the received
vector y given as:
Fig. 1. Block diagram of a hybrid transceiver for
mmWave MIMO communications. 𝒚 = √𝜌𝐇𝐅𝐑𝐅 𝐅𝐁𝐁 𝒔 + 𝒏 (1)

directly by applying the SVD technique to the matrix of Where 𝐧 ∈ ℂ𝑁𝑟 ×1 is the noise vector, representing a
equivalent channel. circularly symmetric complex Gaussian random vector with
zero-mean and covariance matrix 𝜎 2 𝐈𝑵𝒓 .
The problem with the current hybrid precoding
approaches is that they are either complex and require We use the expanded Saleh-Valenzuela geometric
several mathematical operations to improve the spectral model as in [17][18] to represent the considered narrowband
efficiency [8][9][10][16], or that they require more channel model. This model is commonly used in mmWave
information about the radio channel [6][7][13], which is massive MIMO systems. It is based on the concept that
undesirable, or that they employ matrix decomposition N cl scattering clusters, each of which contributes N ray
[11][12][15], which is also not a simple operation and does
not provide high performance in the cases where the number propagation paths, add together to form the channel matrix
of RF chains is higher than the number of data streams. In H. Therefore, the discrete-time narrowband channel H can
this research, we present an iterative hybrid-precoding be written as:
scheme focused on minimizing the spectral efficiency gap N cl N ray
ar ir,l  at ir,l 
Nt N r
 
*
between the optimal fully-digital design and low- H i ,l
(2)
complexity current schemes when the number of streams is N cl N ray i 1 l 1
less than the number of RF chains while maintaining the
same performance of the existed works when the number of where N cl and N ray are the number of scattering
streams is equal to the number of RF chains. Our hybrid
clusters and propagation rays in each cluster, respectively,
precoder/combiner is designed to take advantage of the
additional RF chains in order to minimize the distance and i ,l is the complex gain of the 𝑙 𝑡ℎ ray in the 𝑖 𝑡ℎ cluster.

ar ir,l  and at ir,l  are the array response vectors at the

between the unconstrained precoding matrix and the
products of the hybrid RF and baseband precoding matrices,
which leads to an approach to optimal spectral efficiency. receiver and transmitter of the 𝑙 𝑡ℎ ray in the 𝑖 𝑡ℎ cluster, with
Our proposed method only uses the optimal fully-digital
azimuth angles i ,l and i ,l , respectively. As we observe,
r r
matrix instead of the complete channel matrix, which has
larger dimensions. As a result, less feedback information is the channel matrix 𝐇 ∈ ℂ𝑁𝑟 ×𝑁𝑡 and normalized to satisfy
provided. E{‖𝐇‖2𝐹 } = 𝑁𝑡 𝑁𝑟 .
The rest of this paper is organized as follows: In Section It is worth noting that both the transmitter and receiver
II, a brief review of the system model is presented. In must have a perfect knowledge channel matrix H in order to
Section III, the proposed hybrid precoding scheme is perform precoding.
described in detail. Some simulation results are illustrated
in Section IV. Finally, Section V concludes the manuscript. At the receiver side, the data vector y is processed by the
concatenated combiner matrix 𝐖 ∗ = 𝐖𝐵𝐵 ∗ ∗
𝐖𝑅𝐹 , where
Notations: Throughout this paper, matrices and vectors are 𝑁𝑟𝑅𝐹 ×𝑁𝑠
𝐖𝐵𝐵 ∈ ℂ is the baseband combiner and 𝐖𝑅𝐹 ∈
set in boldface, with uppercase letters for matrices and 𝑅𝐹
ℂ𝑁𝑟 ×𝑁𝑟 is the RF combiner. The resultant signal at the
lowercase letters for vectors. The superscripts 𝐻 and −1 receiver is:
denote the conjugate transpose and inverse, respectively.
∗ ∗ ∗ ∗
‖ . ‖𝐹 is used to represent the Frobenius norm of a matrix. 𝑦̂ = √𝜌𝐖𝐵𝐵 𝐖𝑅𝐹 𝐇𝐅𝐑𝐅 𝐅𝐁𝐁 𝒔 + 𝐖𝐵𝐵 𝐖𝑅𝐹 𝒏 (3)
The commas and semicolons in a matrix are used as column
In a hybrid configuration, the precoding or combining is
and row separators, respectively. 𝐈𝑁 is the N × N identity actually done in two stages, where beamforming is
matrix; 𝐈𝑀×𝑁 is the M × N all-zeros matrix; performed in both the digital and analog domains. Digital
II. SYSTEM MODEL weights cannot be applied to individual antenna elements.
At the element level, the signal is adjusted by analog phase
A single-user mmWave hybrid MIMO system is shifters, which only change the phase of the signal, so we
considered as shown in Fig. 1. The transmitter is assumed should suppose that the elements of the two RF matrices
to have 𝑁𝑠 spatial streams transmitted via 𝑁𝑡 antennas using have a constant modulus.
𝑁𝑡𝑅𝐹 RF chains, while the receiver is assumed to use 𝑁𝑟
receive antennas and 𝑁𝑟𝑅𝐹 RF chains for reception. To The spectral efficiency is considered one of the most
reduce the implementation complexity in the case of multi- important performance metrics that should be maximized
stream transmission, the number of RF chains should be when choosing the precoding and combining matrices. The
smaller than the number of antennas, and hence 𝑁𝑠 ≤ spectral efficiency can be formulated as follows [19]:
𝑁𝑡𝑅𝐹 (𝑁𝑟𝑅𝐹 )<𝑁𝑡 (𝑁𝑟 ). 𝐈𝑵𝒔 +
𝜌
𝑹−1 𝐻 𝐻
𝑁𝑠 𝑛 𝐖𝐵𝐵 𝐖𝑅𝐹 𝐇𝐅𝐑𝐅 𝐅𝐁𝐁
𝑅 = 𝑙𝑜𝑔2 ( ) (4)
The 𝑁𝑠 × 1 symbol vector √𝜌𝒔 is first precoded on the ∗
× 𝐅𝐵𝐵 ∗
𝐅𝑅𝐹 𝐇 ∗ 𝐖𝐑𝐅 𝐖𝐁𝐁
baseband side by the precoder matrix 𝐅𝐁𝐁 , followed by the
RF precoder 𝐅𝐑𝐅 . Notation 𝜌 is the average received signal
 Where 𝑹𝒏 = 𝜎𝑛2 𝐖𝐵𝐵 ∗ ∗
𝐖𝑅𝐹 𝐖𝐑𝐅 𝐖𝐁𝐁 is the noise In order to show the effectiveness of using the angle
covariance matrix after combining. information of the optimal precoding matrix as a reference
for the RF matrix, suppose that 𝑁𝑠 = 𝑁𝑡𝑅𝐹 the SVD
III. PROPOSED HYBRID PRECODING METHOD decomposition of 𝐅𝑜𝑝𝑡 is:
In this section, we will review the proposed solution
approach for the hybrid mmWave MIMO system that 𝐅𝑜𝑝𝑡 = 𝐔𝑭 𝚺𝑭 𝐕 ∗ 𝑭 , (8)
presented in the previous section. where 𝐔𝑭 and 𝐕𝑭 are (𝑁𝑡 × 𝑁𝑡 ) and (𝑁𝑠 × 𝑁𝑠 ) unitary
To maximize the spectral efficiency in equation (4), the matrices, respectively, and 𝚺 is (𝑁𝑡 × 𝑁𝑠 ) diagonal matrix.
RF and baseband precoding and combining matrices should The maximum spectral efficiency depends on the optimal
∗
be designed jointly. Unfortunately, optimizing all four matrix 𝐅𝑜𝑝𝑡 through the product (𝐅𝑜𝑝𝑡 𝐅𝑜𝑝𝑡 ). Then we can
matrices at the same time is quite difficult [4]. Thus, various write:
strategies are proposed to achieve suboptimal solution while ∗
𝐅𝑜𝑝𝑡 𝐅𝑜𝑝𝑡 = 𝐔𝑭 𝚺𝑭 𝐕 ∗ 𝑭 𝐕𝑭 𝚺𝑭∗ 𝐔 ∗ 𝑭
maintaining a reasonable computational overhead. In our
paper, we will adopt the strategy proposed in [6], which = 𝐔𝑭 𝚺𝑭 𝚺𝑭∗ 𝐔 ∗ 𝑭 (9)
decouples the optimizations for precoding and combining Comparing with the spectral efficiency in equation (4),
weights. It first optimizes the precoding matrices, and then we can note that the optimal choice of the matrices 𝐅𝐑𝐅 and
uses the results to optimize the corresponding combining 𝐅𝐁𝐁 is when 𝐅𝐑𝐅 = 𝐔𝑭 (: ,1: 𝑁𝑠 ) and 𝐅𝐁𝐁 = 𝚺𝑭 (1: 𝑁𝑠 , : ) in
weights. the case where the constraint |𝐅𝐑𝐅 (𝑖, 𝑗)| = 1 ∀𝑖, 𝑗 is
The optimal unconstrained precoder 𝐅𝑜𝑝𝑡 can be defined ignored.
by the 𝑁𝑠 columns of left singular vectors of channel matrix As mentioned above, 𝐅𝑜𝑝𝑡 is a semi-unitary matrix
H that correspond to the highest singular values. Let the consists of the first 𝑁𝑠 columns of left singular vectors 𝑽 of
SVD of H be: channel matrix H, then the SVD equation can be attained
𝐇 = 𝐔𝚺𝐕 ∗ (5) by:

where U and V are (𝑁𝑟 × 𝑁𝑟 ) and (𝑁𝑡 × 𝑁𝑡 ) unitary 𝐅𝒐𝒑𝒕 = 𝐔𝑭 𝚺𝑭 𝐕 ∗ 𝑭 = 𝑽𝚺𝑭 𝐈𝑵𝒔 =
matrices, respectively, and 𝚺 is (𝑁𝑟 × 𝑁𝑡 ) diagonal matrix 𝐈𝑵𝒔
with diagonal entries arranged in decreasing order. Then = [𝐅𝒐𝒑𝒕 , 𝐅𝑼 ] [ ] [𝐈 ] (10)
𝟎(𝑵𝒕 −𝑵𝒔 )×𝑵𝒔 𝑵𝒔
𝐅𝑜𝑝𝑡 = 𝐕(: ,1: 𝑁𝑠 ). It is straightforward to see that 𝐅𝑜𝑝𝑡 is a
∗ Where 𝟎(𝑵𝒕 −𝑵𝒔)×𝑵𝒔 is a (𝑵𝒕 − 𝑵𝒔 ) × 𝑵𝒔 zero matrix
semi- unitary matrix, which means that 𝐅𝑜𝑝𝑡 𝐅𝑜𝑝𝑡 = 𝐈𝑁𝑠 .
where all the entries are zero and 𝐅𝑈 is a 𝑵𝒕 × (𝑵𝒕 − 𝑵𝒔 )
On the transmitter side, the objective of hybrid design is
matrix that contains the columns (𝑁𝑠 + 1, 𝑁𝑠 + 2, … , 𝑁𝑡 )
to optimize the two matrices 𝐅𝐁𝐁 and 𝐅𝐑𝐅 jointly that
from the matrix 𝑽.
approaches 𝐅𝑜𝑝𝑡 as in:
Algorithm 1. The proposed precoding algorithm
𝑜𝑝𝑡 𝑜𝑝𝑡 ∗
(𝐅𝑅𝐹 , 𝐅𝐵𝐵 ) = arg max tr(𝐅𝑜𝑝𝑡 𝐅𝐑𝐅 𝐅𝐁𝐁 ), Require 𝐅𝑜𝑝𝑡

|𝐅𝐑𝐅 (𝑖, 𝑗)| = 1 ∀𝑖, 𝑗 1. Operate the SVD technique on H: 𝑯 = 𝑼𝚺𝑽∗ .
S. t.
𝑁𝑡𝑅𝐹
‖𝐅𝐑𝐅 𝐅𝐁𝐁 ‖2𝐹 = 𝑁𝑠  2. 𝐅𝑅𝐹 , 𝐅𝐁𝐁 are empty matrices, 𝐿 = ⌈ ⌉
𝑁𝑠

According to [6], the optimization problem in (6) is 3. 𝐅𝑜𝑝𝑡 = 𝑽(: ,1: 𝑁𝑠 ) = [𝐯1 , 𝐯2 , … , 𝐯𝑁𝑠 ].
equivalent to the minimization of ‖𝐅𝑜𝑝𝑡 − 𝐅𝐑𝐅 𝐅𝐁𝐁 ‖𝐹 . As a 4. For 𝑖 = 1: 𝐿
5. 𝐅𝑟𝑒𝑠 = 𝐅𝑜𝑝𝑡 − 𝐅𝑅𝐹 𝐅𝐁𝐁
result, the main purpose of the hybrid precoder is to make
the product 𝐅𝐑𝐅 𝐅𝐁𝐁 as close to matrix 𝐅𝑜𝑝𝑡 , as possible. The 6. 𝐅̂𝑅𝐹 = 𝐅𝑟𝑒𝑠 ./𝑎𝑏𝑠(𝐅𝑟𝑒𝑠 )
−1 ∗
main challenge when attempting to compute 𝐅𝐑𝐅 and 𝐅𝐁𝐁 7. 𝐅̂𝐵𝐵 = (𝐅̂𝑅𝐹
∗ ̂
𝐅𝑅𝐹 ) 𝐅̂𝑅𝐹 𝐅𝑟𝑒𝑠
lies in the constraint |𝐅𝐑𝐅 (𝑖, 𝑗)| = 1 ∀𝑖, 𝑗 , so it can be 8. 𝐅𝑅𝐹 = [𝐅𝑅𝐹 |𝐅𝑅𝐹 ], 𝐅𝐁𝐁 = [𝐅𝐁𝐁 |𝐅̂𝐵𝐵 ]
̂
considered that the analog phase-shifter dominates the 9. end for
performance of the hybrid precoding. 𝐅
10. 𝐅𝐁𝐁 = √𝑁𝑠 ‖𝐅 𝐁𝐁 ‖
𝐅
𝑅𝐹 𝐁𝐁 𝐹
The optimum scenario happens when the optimal
11. 𝐫𝐞𝐭𝐮𝐫𝐧 𝐅𝑅𝐹 , 𝐅𝐁𝐁
precoding matrix 𝐅𝑜𝑝𝑡 is exactly equal to the product of
𝐅𝐑𝐅 and 𝐅𝐁𝐁 , which yields the best spectral efficiency.
Unfortunately, for most of the practical channel matrices, The key observation is that 𝐅𝑜𝑝𝑡 has 𝑁𝑠 singular values
this is almost impossible. equal to 1, and the corresponding singular vectors can be
In this paper, the angle information of the optimal chosen as the column of 𝐅𝑜𝑝𝑡 . That is equivalent to choosing
precoding matrix 𝐅𝑜𝑝𝑡 will be directly used to derive the 𝐅𝐑𝐅 = 𝐅𝑜𝑝𝑡 and 𝐅𝐁𝐁 = 𝐈𝑵𝒔 , but after applying the constraint
elements of the matrix 𝐅𝐑𝐅 , unlike the OMP-based methods of constant modulus to 𝐅𝐑𝐅 , this makes it extremely difficult
in [6][7][10][11], which use the antenna array response to achieve the same performance as the optimal precoder
vectors to generate 𝐅𝐑𝐅 or in [14][15][16], which use the with low computational complexity. Therefore, it is
phase information of the principal left singular vectors of desirable to fix the RF precoder and then use the equation
the matrix 𝐅𝑜𝑝𝑡 . (7) to compute the corresponding digital precoder.
Generally, in massive MIMO systems, the number of RF
After obtaining the matrix 𝐅𝐑𝐅 , the baseband precoding chains can be bigger than 𝑁𝑠 , so the problem of choosing
matrix 𝐅𝐁𝐁 can be obtained using the least squares solution: 𝐅𝐑𝐅 = 𝐅𝑜𝑝𝑡 ./𝑎𝑏𝑠(𝐅𝑜𝑝𝑡 ) is that the dimension of 𝐅𝑜𝑝𝑡 is 𝑁𝑡 ×
∗ 𝑁𝑠 , while the dimension of 𝐅𝐑𝐅 is 𝑁𝑡 × 𝑁𝑡𝑅𝐹 , where 𝑁𝑠 ≤
𝐅𝐁𝐁 = (𝐅𝑅𝐹 𝐅𝐑𝐅 )−1 𝐅𝑅𝐹
∗
𝐅𝑜𝑝𝑡 (7)
𝑁𝑡𝑅𝐹 , so we have (𝑁𝑡𝑅𝐹 − 𝑁𝑠 ) columns should be chosen to
 (a) (b)

Fig. 2. Spectral efficiencies with varying SNRs when 𝑁𝑡 = 64 and 𝑁𝑟 = 16. (a) Data stream 𝑁𝑠 = 1 and 2. (b) Data stream 𝑁𝑠 = 4.

(a)
(b)
Fig. 3. Spectral efficiencies with varying SNRs when 𝑁𝑡 = 256 and 𝑁𝑟 = 64. (a) Data stream 𝑁𝑠 = 1 and 4. (b) Data stream 𝑁𝑠 = 6.

achieve maximum spectral efficiency. This is the reason between the matrix 𝐅𝑜𝑝𝑡 and the product 𝐅𝑅𝐹 𝐅𝐁𝐁 . The
why many researchers use antenna array response vectors or pseudo-code for the proposed precoder is given in
left 𝐅𝑜𝑝𝑡 ′s singular vectors. In our proposed algorithm, the Algorithm 1. The precoding matrices are divided into L
remaining (𝑁𝑡𝑅𝐹 − 𝑁𝑠 ) columns will be calculated in an parts, each part consists of 𝑁𝑠 columns for the RF matrix
iterative manner in order to minimize the residual error
𝐅𝑅𝐹 and 𝑁𝑠 rows for the baseband matrix 𝐅𝐁𝐁 . The 𝑖 𝑡ℎ part IV. NUMERICAL RESULTS
at the 𝑖 𝑡ℎ iteration is computed based on 𝐅𝑜𝑝𝑡 after
To show the spectral efficiency improvement of our
removing the contribution of the selected vector until the 𝑖 𝑡ℎ proposed hybrid precoding scheme, our results are
iteration. At the end of the loop, the product of the precoding compared with the OMP algorithm in [6], SVD-Based Low-
matrices will be the sum of these parts’ products Complexity algorithm in [13], general hybrid precoding
L
FRF FBB   FRF are the 𝑖 𝑡ℎ part scheme in [14], Hybrid Precoding by Equivalent Channel
i i
i i
FBB , where FRF and FBB
i1
(HP-EC) in [15] and Joint Hybrid Precoding/Combining in
of RF and baseband matrices, respectively. [16].

In summary, the first 𝑁𝑠 column of the RF precoder is The radio channel is generated based on the following
computed in steps 6 and 7 by setting 𝐅̂𝑅𝐹 = 𝐅𝑜𝑝𝑡 ./ settings: the number of clusters 𝑁𝑐𝑙 = 8, the number of rays
𝑎𝑏𝑠(𝐅𝑜𝑝𝑡 ), where at the first iteration 𝐅𝑟𝑒𝑠 = 𝐅𝑜𝑝𝑡 , and then 𝑁𝑟𝑎𝑦 = 10 , and all the rays are spread with Laplacian
the first 𝑁𝑠 rows of the 𝐅𝐁𝐁 is calculated based on equation distributed azimuth and elevation angles of arrival and
(7) in step 7. In the next iteration, the residual error between departure. The spread angles in the azimuth and elevation
the matrix 𝐅𝑜𝑝𝑡 and the product 𝐅𝑅𝐹 𝐅𝐁𝐁 is calculated in step directions are set as 20𝑜 and 60𝑜 , respectively. Moreover, it
is assumed that every cluster has constant transmit power.
5 as a matrix 𝐅𝑟𝑒𝑠 , then the next 𝑁𝑠 columns of 𝐅𝑅𝐹 and 𝑁𝑠 Our simulation takes into account two antenna array
rows of 𝐅𝐁𝐁 are computed to minimize the residual error. structures: the first is 64×16 and the second is 256×64,
It is worth noting that the hybrid RF and baseband allowing us to demonstrate the effects of increasing the
combiners on the receiver side can be obtained in the same antenna array size on system performance. Each simulation
way as the hybrid precoder design in Algorithm 1. point is averaged across 5000 different channel realizations.
 (a) (b)
Fig. 4. Spectral efficiency with different values of 𝑁𝑠 under the condition that (𝑁𝑡 = 64 and 𝑁𝑟 = 16). (a) SNR = -20 dB, (b) SNR = 0 dB.

(a) (b)
Fig. 5. Spectral efficiency with different values of 𝑁𝑠 under the condition that (𝑁𝑡 = 256 and 𝑁𝑟 = 64). (a) SNR = -20 dB, (b) SNR = 0 dB.

Figs. 2 and 3 illustrate the spectral efficiencies versus proposed method can always achieve the best in the
the varying SNRs under the conditions of (𝑁𝑡 = 64 , 𝑁𝑟 = spectral-efficiency performance, and when 𝑁𝑠 = 𝑁𝑡𝑅𝐹 , the
16, 𝑁𝑡𝑅𝐹 = 𝑁𝑟𝑅𝐹 = 4) , and (𝑁𝑡 = 256, 𝑁𝑟 = 64, 𝑁𝑡𝑅𝐹 = proposed method preserves the performance of the current
𝑁𝑟𝑅𝐹 = 8), respectively, for different values of 𝑁𝑠 . approaches.
From the simulation results in Fig. 2, when 𝑁𝑠 = 1, the It is important to note that when 𝑁𝑠 = 𝑁𝑡𝑅𝐹 , the number of
proposed scheme performs the same as the optimal iterations L = 1, and the proposed approach is less complex
precoding scheme, when 𝑁𝑠 = 2 , the gap between our than all of the compared works since the angle information
proposed scheme and the optimal scheme increases, but is of the analog phase shifters can be directly derived from the
still less than the gap with the current schemes, and when optimal unconstrained precoder without the need for any
𝑁𝑠 = 𝑁𝑡𝑅𝐹 = 4 , the gap increases and the achieved additional information or decomposition techniques. As a
performance is similar to the current schemes. As shown in result, when 𝑁𝑠 = 𝑁𝑡𝑅𝐹 , the performance is the same, but
Fig 3, increasing the size of the antenna array yields the the complexity is reduced.
same results. So we can say that the proposed hybrid-
precoding scheme can clearly outperform the current V. CONCLUSION
precoding technique in the case 𝑁𝑠 < 𝑁𝑡𝑅𝐹 . However, when In this paper, we considered single-user precoding and
𝑁𝑠 increases, the spectral efficiency gap between our combing in mmWave hybrid systems. In order to get the
proposed scheme and the optimal scheme increases until it product of the RF and BB precoding matrices closer to the
performs similarly to the current methods in the case optimal matrix, which would improve the spectral
𝑁𝑠 = 𝑁𝑡𝑅𝐹 . efficiency performance, the RF and BB precoding matrices
vectors were iteratively constructed using the angle
Firstly, the spectral efficiency does not increase with the
information derived directly from the optimal precoder. As
increasing of 𝑁𝑆 in a proportional manner for all the studied a result, the proposed method performs better in real-time
cases, for example, when (𝑁𝑡 = 64 , 𝑁𝑟 = 16, 𝑁𝑡𝑅𝐹 = and is more realistic since it does not require any additional
𝑁𝑟𝑅𝐹 = 4) and 𝑆𝑁𝑅 = −20 𝑑𝐵, we can note that the SE information other than the fully digital matrix without using
reaches the maximum point at 𝑁𝑆 = 3 and then decreases any decomposition techniques. The obtained results show
gradually. Secondly, we can note that when 𝑁𝑠 < 𝑁𝑡𝑅𝐹 , the that the spectral efficiency gap between the optimal fully-
digital design and low-complexity current schemes is [9] W. M. Chan, T. Kim, H. Ghauch and M. Bengtsson, "Subspace
minimized when 𝑁𝑠 < 𝑁𝑡𝑅𝐹 , while providing the same estimation and hybrid precoding for wideband millimeter-wave
MIMO systems," 2016 50th Asilomar Conference on Signals,
performance when 𝑁𝑠 = 𝑁𝑡𝑅𝐹 Systems and Computers, Pacific Grove, CA, USA, 2016, pp. 286-
290
REFERENCES [10] C. -E. Chen, "An Iterative Hybrid Transceiver Design Algorithm for
[1] P. Sagazio, S. Callender, W. Shin, O. Orhan, S. Pellerano, and C. Millimeter Wave MIMO Systems," in IEEE Wireless
Hull, “Architecture and circuit choices for 5G millimeter-wave Communications Letters, vol. 4, no. 3, pp. 285-288, June 2015,
beamforming transceivers,” IEEE Commun. Mag., vol. 56, no. 12, [11] X. Liu et al., "Hybrid Precoding for Massive mmWave MIMO
pp. 186–192, Dec. 2018. Systems," in IEEE Access, vol. 7, pp. 33577-33586, 2019.
[2] Z. Xiao et al., “A survey on millimeter-wave beamforming enabled [12] W. Ni, X. Dong and W. -S. Lu, "Near-Optimal Hybrid Processing
UAV communications and networking,” IEEE Commun. Surveys for Massive MIMO Systems via Matrix Decomposition," in IEEE
Tuts., vol. 24, no. 1, pp. 557–610, 5th Quart., 2022. Transactions on Signal Processing, vol. 65, no. 15, pp. 3922-3933, 1
[3] S. A. Busari, K. M. S. Huq, S. Mumtaz, L. Dai, and J. Rodriguez, Aug.1, 2017
“Millimeter-wave massive MIMO communication for future [13] D. Zhang, P. Pan, R. You and H. Wang, "SVD-Based Low-
wireless systems: A survey,” IEEE Commun. Surveys Tuts., vol. 20, Complexity Hybrid Precoding for Millimeter-Wave MIMO
no. 2, pp. 836–869, 2nd Quart., 2017. Systems," in IEEE Communications Letters, vol. 22, no. 10, pp.
[4] A. F. Molisch et al., “Hybrid beamforming for massive MIMO: A 2176-2179, Oct. 2018,
survey,” IEEE Commun. Mag., vol. 55, no. 9, pp. 134–141, Sep. [14] S. Wang, L. Li, R. Ruby, and P. Li, “A general hybrid precoding
2017 scheme for millimeter wave massive MIMO systems,” Wireless
[5] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile Netw., vol. 26, pp. 1331–1345, Mar. 2020.
broadband systems,” IEEE Commun. Mag., vol. 49, no. 6, pp. 101– [15] S. Wang, M. He, J. Wang, R. Ran, H. Ji and V. C. M. Leung, "A
107, 2011. Family of Hybrid Precoding Schemes for Millimeter-Wave Massive
[6] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi and R. W. Heath, MIMO Systems," in IEEE Systems Journal, vol. 16, no. 3, pp. 4881-
"Spatially Sparse Precoding in Millimeter Wave MIMO Systems," 4891, Sept. 2022
in IEEE Transactions on Wireless Communications, vol. 13, no. 3, [16] S. Wang et al., "A Joint Hybrid Precoding/Combining Scheme
pp. 1499-1513, March 2014 Based on Equivalent Channel for Massive MIMO Systems,"
[7] Rusu, C., Me´ndez-Rial, R., Gonza´lez-Prelcic, N., & Heath, R. W. in IEEE Journal on Selected Areas in Communications, vol. 40, no.
Jr. (2015). Low complexity hybrid sparse precoding and combining 10, pp. 2882-2893, Oct. 2022,
in millimeter wave MIMO systems. In Proceedings of the IEEE [17] V. Raghavan and A. M. Sayeed, “Sublinear capacity scaling laws for
international conference on communication, London, UK (pp. 1340– sparse MIMO channels,” IEEE Trans. Inf. Theory, vol. 57, no. 1, pp.
1345). 345–364, 2011.
[8] L. Dai, X. Gao, J. Quan, S. Han and C. -L. I, "Near-optimal hybrid [18] A. M. Sayeed, “Deconstructing multiantenna fading channels,”
analog and digital precoding for downlink mmWave massive MIMO IEEE Trans. Signal Process., vol. 50, no. 10, pp. 2563–2579, 2002
systems," 2015 IEEE International Conference on Communications
[19] A. Goldsmith, S. Jafar, N. Jindal, and S. Vishwanath, “Capacity
(ICC), London, UK, 2015, pp. 1334-1339.
limits of MIMO channels,” IEEE J. Sel. Areas Commun., vol. 21,
no. 5, pp. 684–702, 2003.