1

ML Codebook Design for Initial Access and

CSI Type-II Feedback in Sub-6GHz 5G NR
Ryan M. Dreifuerst, Student Member, IEEE, and Robert W. Heath Jr. Fellow, IEEE
arXiv:2303.02850v2 [eess.SP] 16 May 2023

Abstract

Beam codebooks are a recent feature to enable high dimension multiple-input multiple-output
(MIMO) in 5G new radio (NR). Codebooks comprised of customizable beamforming weights can be used
to transmit reference signals and aid the channel state information (CSI) acquisition process. Codebooks
are also used for quantizing feedback following CSI acquisition. In this paper, we characterize the role of
each codebook used during the beam management process and design a neural network to find codebooks
that improve overall system performance. Evaluating a codebook is not purely about maximizing signal
power. Instead, it is about considering the system-level dependency between the codebooks, feedback,
overhead, and spectral efficiency. The proposed neural network is built on translating codebook and
feedback knowledge into a consistent beamspace basis similar to a virtual channel model. Then, the
beamspace is fed into the neural network to generate initial access codebooks. This beamspace codebook
algorithm is designed to directly integrate with current 5G beam management standards (Releases 15-17),
without changing the feedback format, beam training timing, or requiring additional side information.
Our simulations show that the neural network codebooks improve over traditional codebooks in received
signal power, even in dispersive sub-6GHz environments. We further use our simulation framework to
evaluate type-II CSI feedback formats with regard to effective multi-user spectral efficiency. Our results
suggest that optimizing codebook performance can provide valuable spectral efficiency improvements,
but optimizing the feedback configuration can be equally important for multi-user performance in sub-
6GHz bands.

I. I NTRODUCTION

Fifth-generation (5G) cellular systems have adopted a beam-based approach for MIMO com-
munications. In such systems, broadcast control signals are beamformed to provide array gain

Ryan M. Dreifuerst and Robert W. Heath Jr. are with North Carolina State University, Raleigh, NC 27695 {rmdreifu,
rwheathjr}@ncsu.edu. This material is based upon work supported by the National Science Foundation under grant nos. NSF-
ECCS-2153698, NSF-CCF-2225555, NSF-CNS-2147955 and is supported in part by funds from federal agency and industry
partners as specified in the Resilient & Intelligent NextG Systems (RINGS) program.
 2

[1]. This helps increase coverage and meet link budgets at high carrier frequencies. While the
motivation for beam-based systems was to provide array gain in millimeter-wave (mmWave)
frequency bands, the same framework is also used at low frequencies. Unfortunately, optimization
of the beamforming framework, and subsequent feedback strategies, has not been extensively
investigated for sub-6GHz multi-user MIMO (MU-MIMO), despite the importance of low- and
mid-band MIMO in current and anticipated future standards [2].

Codebook-based beam management is a framework for beamforming and feedback that enables
transmitting beamformed reference signals and multiple forms of CSI acquisition [1]. In early
3GPP releases up to Release 12, channel state information was obtained by estimating the
channel from reference signals for every transmitter-receiver beam pair using unprecoded signals.
Unprecoded reference signals were used up through 4G LTE Release 14 because multi-antenna
beamforming was not needed to achieve higher signal-to-noise ratio (SNR) for synchronization
and channel estimation [3]. Beamformed control signals have become a necessity in mmWave
bands as a result of the increasing bandwidths (which reduce the SNR for a constant transmit
power) and shrinking antennas sizes–resulting in pre-beamformed SNR on the order of −15dB
or less [4]. By beamforming the control and reference signals, receivers are more likely to detect
initial access signals and obtain more accurate synchronization and channel measurements.

The process of obtaining CSI at the transmitter involves multiple steps and codebooks to balance
the overhead, latency, and accuracy of the CSI [5], [6]. To begin the process, the base station (BS)
transmits a synchronization signal block (SSB) with a beamformer selected from a codebook
(SSB codebook) that enables synchronization and random access. The SSB codebook is typically
small and all of the codewords in the codebook are used during each SSB period. The user
equipment (UE) will provide a small amount of feedback to conclude the initial access that
includes a beam selection from the SSB codebook. The BS will use this information to select a
subset of beamformers from a large codebook, different from the SSB codebook, for transmitting
channel state information reference signals (CSI-RS). We assign this codebook the name “CSI-
RS codebook”. The relationship between the SSB and CSI-RS codebook can be understood from
the perspective of a hierarchical beam search [7]. The CSI-RS enable hybrid beam training by
providing multiple precise beams for analog beam training–often referred to as beam refinement–
and pilots for digital channel estimation. The UE will provide an in-depth packet of feedback
following CSI-RS transmission that includes the CSI-RS codebook beamformer selection as well
 3

as information based on the channel estimation. While the first iterations of beam management
used wide and narrow discrete Fourier transform (DFT) beamformers [4], there is no requirement
that the beamformers be “wide” and “narrow” for the SSB and CSI-RS codebooks. In fact, the
codebooks can be arbitrary since the UEs do not need knowledge of the beamformers during
initial access, beam refinement, or even data transmission.

While the SSB and CSI-RS codebooks can be arbitrary, the third codebook used for quantizing
the feedback (denoted as the FB codebook) is precisely defined depending on the format (type-
I, type-II). The FB codebook is defined for specific BS and UE array geometries so that the CSI
can be represented and fed back as indices from the codebook. This codebook must be known
at the BS and UE. It is not strictly necessary that even the feedback codebook is used when
reciprocity exists, as reciprocity between the uplink and downlink channel can be exploited in
time-division-duplexing systems. Feedback from downlink training, though, is still beneficial in
these systems due to asynchronous link budgets [8] or when the number of transmitting and
receiving antennas on a UE is not equal [9].

Our work focuses on a realistic and system-level evaluation of the codebooks used during the
beam management framework in sub-6GHz bands. In particular, we unify the two stages of
codebook design with the feedback framework and MU-MIMO data transmission to achieve
good network performance. Our contributions in this manuscript are summarized as follows:

• First, we precisely describe the beam management and CSI type-II feedback approach for
5G networks. There are many misconceptions on how the codebooks, SSB, and CSI-RS
processes are used for obtaining different forms of feedback. We explain how the various
codebooks are used, what is specified in the standards versus left up to implementations,
and provide an efficient implementation for feedback quantization in CSI type-II.
• Second, we present a novel neural network architecture, Beamspace-Codex (BSC), and
feedback processing technique for SSB codebook generation. We constrain the system to
follow 5G beam management timing which presents a significant constraint for dynamic
SSB and CSI-RS codebook generation. To mitigate this constraint, the SSB codebook is
dynamically generated–which is manageable at the SSB periodicity–and the CSI-RS code-
book is determined as a precomputed DFT decomposition of the SSB codebook. With this
integration, we unify the arbitrary codebooks learned via deep learning with the feedback
 4

quantization technique to facilitate a hierarchical beam search that improves the beam
management process used in 5G NR.
• Third, we present the results of an extensive evaluation of our proposed method with
benchmark DFT codebooks and CSI-based eigen-beamforming solutions using various lev-
els of feedback and overhead. We show that our proposed solution achieves significant
performance improvements. We also show that MU-MIMO performance is dramatically
limited by the quantization resolution techniques in low-band 5G. In contrast, the frequency
selectivity and resource allocation have little impact in the current framework and scattering
environment.

There is significant work on beam training and beam alignment. Here we review a set of relevant
prior work relating to machine learning and beam training. In [10], a gradient-optimization
approach, not using deep learning, is shown to be successful in aiding the beam search process
with respect to maximizing the SNR in 60GHz personal networks. Another approach using
classical data-driven techniques is presented in [11] for vehicular networks based on historical
beam training results. While the codebooks and algorithms are relatively straightforward, the idea
of “learning” an efficient codebook from site-specific data is a cornerstone of many modern beam
alignment techniques [12]–[14]. These papers [10]–[14] all focus on mmWave channels and often
exploit the sparsity of such environments to aid the beam training process. Furthermore, all of
these papers focus exclusively on beam alignment with standard orthogonal codebooks like DFT
and phase-shifted codebooks. While DFT codebooks are sensible for certain array geometries
and limited scattering environments at mmWave bands, these codebooks do not necessarily work
well in rich scattering. Furthermore, none of these papers consider a MIMO format with multiple
UE antennas and streams of data. In this paper, we consider a full MU-MIMO OFDM system
in a rich scattering environment and use dynamic codebooks.

In another line of related work, algorithms have been proposed that design beamformers or
codebooks [7], [13], [15]–[20]. Hierarchical codebooks and the associated design process were
proposed in [7], [15], [16] which separated the beam training into multiple stages of wide-beam
and narrow-beam codebooks. A similar strategy was adopted by 5G using the SSB and CSI-RS
codebooks and plays an important role in systems designed for 5G deployment. Deep learning
was applied to the task in [17], which uses deep reinforcement learning to design broadcast beam
 5

patterns. While designed for small-scale MIMO, the results suggest that arbitrary learned beam
patterns can be an effective way to initiate communication in broadband networks. A supervised
and an unsupervised learning approach were presented in [18], which showed promising results
when used to design beamforming vectors from channel measurements. These results were based
on a simplified channel model, though, and perfect CSI was assumed at the BS. Recently, a
deep, federated, reinforcement learning strategy was proposed [13] that jointly trained a beam
management model over a network using user location information. The focus was on deciding
which sectors would be active, thereby narrowing the “codebook,” although the paper ignores all
beamforming besides sector-level association and large-scale pathloss. Another paper considered
the task of learning probing beams [19], which is one of the only papers to consider both the
SSB and CSI-RS codebooks. The results show significant gains can be achieved over a basic
hierarchical search, although perfect CSI is assumed to be available for training and the channels
are assumed to be narrowband. The results from [19] provide motivation for researching realistic
deep learning codebook methods. In our previous work we designed a neural network for SSB
codebook generation in narrowband channels as well but trained using an angular representation
of the channel [20]. The results, however, did not extend well to wideband systems and did not
consider the impact that the SSB codebook has on system-level performance. In this investigation,
we now model wideband channels and fully incorporate the entire beam management and data
transmission into our evaluation.

An important part of the 5G beam management framework is the connection between beam
training and feedback. An enhanced CSI feedback strategy was proposed for full-dimension
MIMO [3] in 4G Release 13 which is based on a linear combination of DFT codewords. This
strategy is very similar to the type-II feedback format introduced in 5G Release 15. One key
difference is that the initial idea assumed the users must know the beamforming vector used
during pilot transmission, which was not adopted into the standards. There has also been growing
interest in using machine learning to design or modify the feedback [21]–[23]. All that work
assumes that both the network and the users can share models for encoding and decoding the
feedback, which is not supported or easily introduced into network operation. In this paper, we do
not use deep learning to change the feedback methods. Instead, we characterize the relationship
and limitations the feedback and beam management strategy have on MU-MIMO performance.

Notation: A is a matrix, a and {a[i]}N T

i=1 are column vectors and a, A denote scalars. A , A,
 6

Beam Management Steps

1)
2)
SSB transmission
SSB Report received
UE
a) CSI-RS transmission
b) SSB codebook update
3) CSI-RS Report received

4)
a) RZF Precoder calculated
Data transmission with RZF
UE
UE
BSC
Algorithm
𝑡, 𝑘
Beamformer Selection

SSB CSI-RS RZF

Codebook Codebook Precoding
CSI-RS Report
SSB Report PMI, RSRP, etc.
SSBRI, RSRP

Fig. 1. Beamforming codebooks are used to transmit reference signals for synchronization from the SSB codebook. Based on the
SSB report, additional reference signals are beamformed using the CSI-RS codebook for channel estimation by the user equipment
(UEs). Using the CSI-RS report, the base station can design a precoder (i.e. RZF precoder) that attempts to maximize the spectral
efficiency. The proposed BSC algorithm updates the SSB codebook, which in turn affects the CSI-RS codebook and feedback.

A∗ , and A† represent the transpose, conjugate, conjugate transpose, and psuedo-inverse of A.

A[k, `] denotes the entry of A in the k th row and the `th column. The same meaning is also
associated with Ak,` . Similarly, A[:, k] refers to the k th column of A. Unspecified norm equations
p
are kak2 = a∗ a for vectors and the Frobenius norm kAkF = Tr(AA∗ ) for matrices. We make
use of the disjoint operator A \ α to describe the matrix (or vector) A excluding the subset
√
(or element) α. We define j = −1. The operator E[·] is used for the expectation of a random
variable. Due to the notational complexity of MU-MIMO with OFDM, we will always use u to
refer to a specific UE, t as a specific time, k as a specific frequency resource, and nt /nr to refer
to a specific transmit or receive antenna.

The remainder of our paper is organized as follows. First, we introduce the system model
beginning from an arbitrary MU-MIMO OFDM channel and continuing through the SSB and
CSI-RS beam management processes. We finish this section with a description of the channel
estimation, feedback, and data transmission stages. Next, we introduce the beamspace observation
and neural network architecture that form the basis of our work. In the final sections, we present
the simulation setup and evaluate the various codebooks and feedback formats to address the
 7

lack of prior work when it comes to codebooks and CSI formats in sub-6GHz 5G NR.

II. S YSTEM MODEL

We begin the investigation with an overview of the channel definition and assumptions before
introducing the beam management steps including SSB transmission, CSI-RS transmission,
feedback, and downlink data transmission, as shown in Figure 1. Afterward, we outline the
problem formulation that influences the neural architecture. Throughout this paper, we will limit
the system configuration to a single cell with multiple UEs each equipped with multiple antennas
and all arrays are fully digital and single-polarization. While the inclusion of multiple cells is
more realistic, the impact of multi-cell interference on beam management and feedback can be
mitigated with limited coordination between nearby cells. The use of single-polarized arrays is
also a simplification, although during beam training the second polarization operates as a simple
extension with little impact on the overall system operation. We will investigate multiple BS
deployments with dual-polarized arrays in future work.

A. MU-MIMO channel model

We model the system so that it is representative of real-world conditions to evaluate performance

in a realistic beam training scenario for sub-6GHz massive MIMO. We do not explicitly specify
TDD or FDD because this work does not rely on channel reciprocity. In addition, the channel
will always be handled in the frequency domain as the multi-user downlink OFDM channel,
however, we do not assume any specific channel model at this point. The OFDM MU-MIMO
channel is defined for U users, over T time slots, and K subcarriers between each of the
NR receive antennas and NT transmit antennas as H ∈ CU ×T ×K×NR ×NT . Note that we can
generate this channel model using a clustered channel model as typically done by QuaDRiGa
[24], 3GPP, and other raytracing or statistical models with the array response vectors aN (θ) =
√1 [1, ejπ cos θ , ej2π cos θ , ...,
N
ej(N −1)π cos θ ]T . Defining the time and frequency responses for every
antenna pair is critical to accurately model the beam training process with specific resource
elements allocated for pilots (reference signals) and data. With this setup in mind, a generic
received signal model for signal s transmitted with precoder/beamformer F and received with a
combiner W over a band of subcarriers K is
U −1
(u) 1 (u)
X (u)
yt,k =√ Wt,k Hu,t,k Fu,t,k su,t,k + Wt,k Nt,k . (1)
KNT u=0
 8

The noise N is modeled as IID, complex Gaussian random values to account for thermal noise
and the noise figure of the receiver. Note that because users do not share information, the
combiners and received signals have superscript (u), while the beamformers and channels have
subscript u to imply that an aggregation is possible i.e. the BS has knowledge of the precoders
for all users. During the next subsections, we will specify how W(u) , Fu , and su are used during
the SSB, CSI-RS, and data transmission stages. We will assume throughout the paper that all
beamformers are normalized according to a per-symbol power constraint, i.e. kFk2 = NT . In
the SSB and CSI-RS stages, F is selected from one or more codebooks that play critical roles
in beam management.

B. Codebooks

Communication protocols employ codebooks in many different ways for beam training, feedback
representations, information quantization, and more. Generally, codebooks are understood as
an ordered set of elements called codewords. Throughout this work on beam management,
codewords will correspond to a vector of complex values and we will treat codebooks as
matrices to simplify notation in later steps. There are three codebooks that we will reference and
directly impact beam management: a codebook of beamforming vectors for SSB transmission
(SSB codebook), a codebook for beamforming CSI-RS transmissions (CSI-RS codebook), and
a codebook for feedback quantization (FB codebook). Each of these codebooks is designed to
serve different purposes, although there are many inter-relationships that make designing and
evaluating codebooks challenging.

The sizes of the codebooks are based on the supported usage for each one. The SSB codebook
BSSB ∈ CNP ×Lmax has Lmax codewords which is between 1 and 64 depending on the carrier
frequency [1]. The CSI-RS codebook BCSI-RS ∈ CNP ×N is at least the same size, where N is
typically large but only a small subset of beamformers is selected to be used in one period.
The FB codebook BDFT ∈ CNP ×NX OH NY OV is an oversampled DFT codebook with horizontal and
vertical oversampling factors OH and OV . The number of logical ports, NP , determines the digital
precoding dimension and is equal to NX NY in a fully digital array. The first two codebooks
are transparent to the UE, meaning that they are configurable and any arbitrary beamforming
codebooks can be employed. The FB codebook must stay constant and is fully determined by the
BS logical array size and oversampling factors, which are sent to the UE during initial access so
 9

that both the BS and the UE can use the same codebook for feedback representation. Previously,
4G used a feedback codebook based upon predefined matrices called CSI type-I precoding matrix
indicator (PMI). 5G augmented the feedback to include type-II PMI which corresponds to a set
of directional array responses with amplitude and phase combining factors.

C. SSB transmission and reception

Initial access is used in mobile networks to obtain limited channel information, achieve synchro-
nization, and set up random access procedures. In 5G NR, a cell may initiate the initial access
period at regular intervals of {5, 10, 20, 40, 80, 160}ms [6], to control the periodicity a UE must
be active (not in power-saving mode) and provide feedback to the network. During this period,
the cell will transmit SSBs that contain primary and secondary synchronization signals (PSS,
SSS) as well as demodulation reference signals (DMRS) [6]. These SSBs are beamformed using
a specific beamforming vector selected from BSSB and associated with a beam index. Each SSB
is transmitted sequentially in time, using 20 contiguous resource blocks. This is a much smaller
bandwidth than the downlink data transmission and is a key limitation of relying on SSB feedback
to obtain CSI in frequency-selective channels. Furthermore, the resource allocation of SSBs is es-
sentially fixed to assist new users with joining a network. Depending on the cell carrier frequency
and subcarrier spacing, a cell may transmit up to Lmax = {1, 4, 8, 64} beams in a period, and all
cells must transmit at least one beam. During transmission, the UE will measure the Reference
Signal Received Power (RSRP) and report the measurement using the random access channel
slot corresponding to the index of the beam with the highest RSRP. This way the base station
knows how strong the signal was and to which of the SSB codebook’s codewords it corresponds.

The SSB is transmitted over a band of KSSB frequency resources at times Ti,SSB for the ith SSB
using beamformer fi ∈ BSSB . While transmitting the reference signal, the rest of the band is still
available for data transmission, so the system power is shared over all K resource blocks. In
total, all Lmax SSBs are transmitted from the codebook BSSB . During SSB reception, the generic
received signal model from (1) is modified to account for the single data stream

(u) 1 (u)∗ (u)∗

ySSBi ,t,k = √ w Hu,t,k fi st,k + wt,k Nt,k . (2)
KNT t,k

In mmWave bands, it is common that w is also determined through beam training. In sub-6GHz
bands, the receiver instead uses antenna selection or simply the first antenna for power saving
 10

[25]. With the received signal model (2) in mind, we can now define an important quantity
provided as feedback during initial access, the RSRP.

D. Reference signal receive power

RSRP is one of the primary metrics that the receiver will measure during initial access for
determining the channel quality. The base station will use the RSRP feedback to determine the
strength of the signal, the initial code rate to be used, and what CSI-RS beamformers should be
employed at the next iteration. The RSRP measures the received power during a given reference
signal (SSB or CSI-RS). In the SSB case, the demodulation reference signal (DMRS), known
at the transmitter and receiver, is a pilot signal used for RSRP measurement and decoding
the master information block provided in the SSB. We will assume antenna selection for the
nr ∈ {0, ..NR − 1} antenna during SSB reception, but because it is fully digital the UE can
determine the antenna selection after receiving the full SSB and selecting the strongest receive
branch. The RSRP is then

(u) 1
ySSBi ,t,k = √ Hu,t,k fi st,k + Nt,k (3)
KNT
X X 2
(u) (u)
RSRPi = max ySSBi ,t,k,nr . (4)
nr
k∈KDMRS t∈Ti,DMRS

Note that (4) is for a specific UE (u) and SSB (i), so the total SSB information aggregated at
the base station is comprised of the strongest RSRP and corresponding SSB index (SSBRI) for
each of the UEs, denoted as (p, m), and defined as
n oU −1
(u)
p=
max RSRPi (5)
i u=0
 U −1
(u)
m = argmax RSRPi . (6)
i u=0

It is worth noting that the actual SSBRI is not transmitted by the users; the random access slot
that the UE responds during will correspond to a specific SSBRI. This is an efficient way to
reduce the amount of feedback, allow the base station to perform receive combining for the UE
feedback, and still share the necessary information.
 11

𝑥𝑃−1
𝑦𝐿−1 Resource Mapper

𝑦𝐿−2 𝑥𝑃−2
Resource Mapper
Precoding Beamforming
Matrix Matrix
𝑦1 𝐿×𝑃 𝑃×𝑃
𝑥1
Resource Mapper
𝑦0
𝑥0
Resource Mapper

DMRS CSI-RS
𝐿×𝑃 𝑃×𝑃 𝑁𝑥 × 𝑁𝑦
𝐿 layers Precoder Beamformer 𝑃 Logical Ports Physical Antenna Array

Fig. 2. The data processing flow from data streams (layers) to antenna outputs in 5G. First, the streams (and DMRS) are
multiplexed according to the precoding matrix, which essentially assigns the layers to the ports. Then, the result, along with
CSI-RS, is beamformed to the logical ports. The logical processes are then mapped onto OFDM resource grids and potentially
beamformed again onto the corresponding physical antennas.

E. CSI-RS transmission and reception

Beam refinement is the second step of the hierarchical beam management process that is used to
obtain better beam alignment and enable wideband channel estimation. Beam refinement is done
through the transmission of CSI-RS which is an encompassing term used for the reference pilots,
the beamformed reference signal, and the logical ports that the pilots are sent on. In contrast
to SSB signals, CSI-RS are extremely flexible with many combinations of time and frequency
resources allocatable. There are some limitations on how often CSI-RS can be transmitted,
but the default periodicity a UE expects when joining the network is 80ms and the symbols
must span a larger bandwidth than the SSB–at least 24 resource blocks and up to the entire
bandwidth should contain CSI-RS symbols for a measurement report. Furthermore, CSI-RS are
also transmitted both periodically and aperiodically to assist beam tracking with highly mobile
users. All told, the beam refinement stage requires more resources than initial access, which
is why a hierarchical beam search is used to limit the number of CSI-RS processes needed to
achieve a strong connection.

Once a BS receives the SSB feedback for a new user, it can attempt to gather more precise
CSI through channel state information reference signals. The CSI-RS (and associated CSI-
RS beamformers FCSI-RS ) are used for refined channel estimation, although the estimation is
performed on the beamformed channel Hu FCSI-RS , rather than the physical channel. This enables
 12

larger hybrid arrays and improves the SNR for more accurate estimation. Readers are encouraged
to review [5], [6], [9] for more details of CSI-RS. The BS can use the SSB feedback to determine
and transmit a set of CSI-RS signals with beamformers that are intended to further increase the
RSRP compared to the SSB.

The design of CSI-RS codebooks has been studied to some degree, but DFT beams are common
[3], [13], [26] (and references therein). Often, it is assumed that the set of beamformers is
selected from the same codebook as the codebook used for feedback, which is specified as an
oversampled DFT codebook. There is no specification or need for the SSB or CSI-RS codebooks
to be DFT codebooks, although using DFT beamformers during CSI-RS has advantages related
to fast and efficient processing. For example, the CSI-RS beamformers can be efficiently selected
as the largest DFT components of the SSB beamformers. We will assume the CSI-RS codebook,
BCSI-RS , is comprised of NX NY OH OV oversampled DFT vectors throughout this paper, although
additional investigations into joint SSB–CSI-RS codebook algorithms are planned in future work.
The selection of the CSI-RS beams from SSB decomposition is further refined in Section III.

The CSI-RS process involves transmitting a set of pilots s on the NP ports along with PCSI
beamformers FCSI-RS = {fi }Pi=0
CSI
⊆ BCSI-RS chosen by the serving base station to improve the
UE channel estimation SNR for the effective channel. The beamformers are used during non-
overlapping time-frequency resources (Ti,CSI , Ki,CSI ), with the aggregate set of beamformers
corresponding to what we define as FCSI-RS . Each CSI-RS allocation will have at least PCSI
OFDM resource elements allocated for channel estimation in one resource block and only one
CSI-RS beam is active for a given resource to allow for a similar, interference-free beam selection
as the SSB process. This is done by setting all other CSI-RS resources and ports to be zero-
power CSI-RS except the active one for the specified resources. In a dual-polarized system, the
two polarizations are transmitted together to estimate a co-polarization factor that would be fed
back during CSI reporting. The whole CSI-RS process is allocated over NRB resource blocks,
which is a configurable parameter to use wider bandwidths and noise averaging during channel
estimation. The beam reception and selection process is identical to the RSRP and SSB selection
in the previous subsection. For more information on the CSI-RS beam and port allocation please
see [3] and references therein. At this stage, the entire UE antenna array is active and the UE
 ∗
(u) \
can perform maximum ratio combining wi,t,k = Hu,t,k fi based on the estimated beamformed
 13

channel and appropriate normalization. The estimated SNR at the receiver is

2
\
1 X X 1 Hu,t,k fc
SNR(u) = max . (7)
c∈PCSI K
k∈Kc,CSI t∈Tc,CSI
NT E[kNt,k k2 ]

Note that the noise power E[kNt,k k2 ] can be estimated using zero-power CSI-RS resources. The
SNR can be reported during feedback, although one of the most important aspects of the SNR
is how the channel estimation (and subsequent feedback) is impacted. For the next part of this
section, we will focus on a single user (u) with SNR corresponding to the channel estimation
SNR for the beamformed channel (Hu,t,k FCSI-RS ).

F. Channel estimation

Channel estimation accuracy depends on the number of pilot symbols and the received SNR of
the pilots. The algorithm for UE channel estimation is not defined in the specification and is not
critical for our system-level analysis. For simplicity, we will assume channel estimation via a
simple least squares algorithm so that only knowledge of the pilots is required. Now, we motivate
the inclusion of channel estimation in beam management by considering the relationship between
the channel estimation error, the codebook-dependent-SNR, and the number of pilot resources
Npilots . The resulting mean squared error (MSE) of the channel estimate is [27, section 3.7]

1
MSE = Npilots
. (8)
NT
SNR

Therefore, we can see that a primary feature of the beam management process is to enable more
accurate channel estimation by beamforming the reference pilot signals. It might seem that the
overhead of using multiple beams is wasted given that the number of pilots could alternatively
be increased according to (8). The issue with such logic is that: 1) it assumes that the signal
is always detectable over the noise floor, which is not necessarily the case for non-beamformed
signals and 2) the beamforming can provide a gain on the order of 20dB or more, yet that would
require 100 times more pilot symbols, far larger than the number of beams used.

Calculating the estimated effective channel, HF,

d which may not be frequency selective, is the
first step in determining the precoder matrix information (PMI), which is one of the quantized CSI
elements provided during feedback. The PMI can correspond to different aspects of the estimated
channel because it is left to operator implementation as part of the PMI selection procedure.
 14

For example, the user could provide PMI according to its desired beamformers (calculated
via singular value decomposition of HF).
d The base station could then use this information to
beamform to a given user and null steer [28] the beams for other users. This puts the majority of
the computational burden, however, on the UE. Instead, we follow models similar to [29] where
the user provides PMI corresponding to the effective channel estimate, rather than the SVD
of the effective channel estimate. This way the BS handles most of the computation, channel
reconstruction, and precoder determination.

G. CSI feedback and reconstruction

We will focus on CSI type-II feedback (Release 16), as our feedback baseline, which is intended
for MU-MIMO with up to 2 layers per user [25], and is also compliant with Release 17. The UE
quantizes the channel estimate according to the FB codebook for a specified LCSI beams per rank
up to the rank indicator, which is the maximum supported MIMO rank, R. A straightforward
method for selecting the PMI would be to exhaustively try every possible combination of LCSI
beams from the codebook and select the best one (according to some metric). Assuming the
feedback codebook is an oversampled DFT codebook, we can make use of efficient algorithms
for quantizing the beamformed channel estimate into the PMI. Although not presented here,
codebooks can also be used for SU-MIMO feedback; we use this form of feedback when
comparing against the SU-MIMO results in Section IV.

First, we define the oversampled DFT codebook BDFT beginning with the (non-oversampled)
DFT codebook for a uniform linear array UN defined in the same way as e.g. [30]. Then we
can define a diagonal oversampling matrix DO,N for an oversampling O and N elements as
    
j2π j2π(N − 1)
DO,N = diag 1, exp − , ..., exp − . (9)
ON ON

Then, the oversampled DFT codebook for a one-dimensional array can be concisely written as

N ×ON
BO,N = [UN , DO,N UN , D2O,N UN ..., DO−1
O,N UN ] ∈ C . (10)

Finally, the UPA oversampling codebook can be defined as the kronecker product of the NX and
NY dimensions

BDFT = BOH ,NX ⊗ BOV ,NY ∈ CNX NY ×OH NX OV NY . (11)

 15

With the FB codebook defined (11), the PMI selection procedure can begin. The UE first
calculates the inner product for all beams from the oversampled codebook BDFT and selects the
first beam (and by association the oversampling basis) as

d DFT ∈ CNR ×NX OH NY OV

C = HFB (12)

Q0 = argmax |C| (13)

where the argmax is determined from the absolute value of all elements of C and the resulting Q0
is an index corresponding to oversampling powers Q0,X and Q0,Y . Because (non-oversampled)
DFT beams are orthogonal, the PMI can be calculated iteratively by selecting the first beam from
the entire oversampled codebook BDFT , reducing the codebook to just the orthogonal subset, and
selecting the beams from the subset. The orthogonal subset can be understood as the kronecker
product of the oversampling-shifted DFT codebooks

Q Q
Bortho,Q0 = DOH0,X 0,Y
,NX UNX ⊗ DOH ,NX UNY (14)

where Q0,X and Q0,Y are the oversampling powers in the X and Y directions of the codebook
index Q0 . With the orthogonal subset, the next step is to select the LCSI beam indices Q ∈
R ×LCSI
ZN
+ and combining coefficients A ∈ CNR ×LCSI

d nr Bortho,Q0 ∈ CNX NY
cortho = HF (15)

Qnr = (LCSI )- argmax |cortho | ∀nr (16)

Anr = cortho [Qnr ] ∀nr (17)

Anr
Anr = ∀nr . (18)
Anr [argmax |Anr |]

First, the orthogonal beams are “evaluated” in (15), then we make use of the k- argmax operation
defined as the k indices for the k largest values in (16) to select the LCSI beam indices. Finally,
the combining coefficients are determined as the complex values of the beam products (17)
and normalized according to the largest coefficient (18). The oversampling basis Q0,X , Q0,Y is
reported back in the PMI so that the overall beam selections can be reconstructed at the BS.

The CSI report contains the CSI-RS indicator (CRI) which is the strong CSI-RS beamforming
codeword, the rank of the PMI, the RSRP or SNR, and the PMI feedback beam indices Q, as
 16

well as 4 bit amplitude and 8-PSK phase values. The amplitude and phase values come from A
which are the corresponding complex values for each [nr , `] feedback beam. There is an additional
overhead reduction step in the specification where instead of indexing each element of Q from
the entire codebook, the set of indices is translated to one of the NLXCSI
NY


combinations and one
index corresponding to the oversampling factors of the strongest index (Q0,X , Q0,Y ) is used. In
this paper, we will ignore the amplitude and phase quantization (assuming perfect amplitude and
phase knowledge) to focus on the effects of the codebooks, SNR, and feedback. The CSI report
can also contain both wideband and narrowband components for a degree of frequency selective
precoding. The number of narrowband components is small, at most BWP = 8 in FR1, to mitigate
the growing feedback and complexity. All of the previous and subsequent steps are defined for
BWP = 1, but extending to frequency selective feedback and precoding is straightforward. We
investigate how the amount of feedback (LCSI , BWP) impacts performance in Section IV.

H. MU-MIMO data transmission

After the CSI-RS period, the base station must determine the precoder that serves a group of
users, often with the goal of maximizing the sum rate or proportional fair rate. Until this stage,
all of the downlink transmissions have been multi-cast in the sense that the same pilots and
data are intended for all users. Therefore, there is no interference assuming a well-designed
OFDM cyclic prefix and subcarrier spacing is used with accurate synchronization. In contrast, the
downlink data transmission carries different data for each user, resulting in interference between
data streams. The remainder of this section outlines the channel reconstruction process, precoder
selection, and achievable spectral efficiency in the downlink channel.

First, because the feedback is type-II, the BS can reconstruct the PMI (corresponding to the
beamformed channel estimate here) with the corresponding complex scaling A as
CSI −1
LX
HF[n
d r] = A[nr , `]BDFT [:, Qnr ,` ] ∀ nr . (19)
d
`=0

Then, assuming the beamformers were well-chosen such that FCSI-RS is right-invertible, channel
reconstruction can be performed successfully. We reintroduce the UE (u) notation here to allow
for aggregating all user channels in subsequent steps by

[ †
H
bu = H
[ u FFCSI-RS . (20)
b
 17

In the case of DFT beams, this process simply becomes an IDFT operation [3], which is
significantly more efficient than the matrix inverse of an NP × NP matrix. The efficiency at this
step is critical to quickly using the feedback to serve users and is one reason we will tailor our
algorithm to integrate with a DFT codebook for CSI-RS. Aggregating the results for each user,
b ∈ CU ×NR ×NP
the base station obtains the downlink channel estimate H
b
 
H
b = Hb 0, H
b 1 , ..., H
b U −1 . (21)
b b b b

The base station can design the precoders for the set of estimated channels using any form of
precoding. To reasonably evaluate the system we will assume a regularized zero forcing precoder
is used with regularization determined to minimize the signal-to-leakage noise ratio [27, section
9.9]. The RZF precoder is built up for each user as
PU −1 b ∗
b i + U NT E[N2 ])−1 H
b u∗
r
NT ( i=0 Hb H b b
i t,k,u
Fu = ∗
. (22)
U PU −1 b 2
b i + U NT E[N ])−1 Hb u∗
( i=0 Hb H b b
i t,k,u

The noise factor, E[N2t,k,u ] can be estimated based on the RSRP/SNR reported and the bandwidth
of the downlink signal. The UE will also perform a combining strategy based on the previously
determined rank. In this work, we will assume an LMMSE receiver is used to maximize
the signal-to-interference noise ratio (SINR) for any precoder, using the embedded DMRS to
determine the combined channel and precoder. The resulting SINR at resource element (t, k) for
a given user u and layer r with equal power allocation per user is obtained as
−1
 UX −1
It,k,u = Hu,t,k Fi F∗i H∗u,t,k + U NT E[N2t,k,u ] (23)
i=0

1 F∗u [:, r]H∗u,t,k It,k,u Hu,t,k Fu [:, r]

SINRt,k,u,r = . (24)
K 1 − F∗u [:, r]H∗u,t,k It,k,u Hu,t,k Fu [:, r]

While the SINR expression (24) appears complicated, it is a simplified ratio between the signal
power for the data stream (u, r) versus the interference power of all other streams. The SINR at
this step is not calculated in a physical system, but it is necessary for evaluating the achievable
spectral efficiency.

Finally, the most critical metric in the wireless network is the sum SE or sum rate when applied
to a specific bandwidth. A fairness constraint can also be applied, however, we have not specified
 18

ML/AI BS UE

5~20 ms
typically

Fig. 3. A timing diagram of the process for codebook learning and evaluation. The AI/ML engine provides new codebooks at each
SSB interval for the base station. At each CSI-RS period, the BS transmits pilots using the CSI-RS codebook, which is determined
based on the SSB codebook and feedback. The CSI-RS feedback is used to determine the precoder used for data transmission.

any form of scheduling so we are most interested in maximizing the sum spectral efficiency or
rate. Assuming Gaussian signaling and treating interference as Gaussian noise, the achievable
spectral efficiency, SEu,t , is
K−1
XX R−1
SEu,t = log2 (1 + SINRu,t,k,r ). (25)
k=0 r=0

In the final post-processing, we consider the effective sum SE, which accounts for the overhead
due to beam training by removing the corresponding time/frequency resources due to training and
feedback of the beam management system (TBM , KBM ) from the spectral efficiency calculation
U
X −1 R−1
X X X
Eff-SSE = log2 (1 + SINRu,t,k,r ). (26)
u=0 t∈T / BM r=0
/ BM k∈K

With this goal and the beam management framework in mind, we can now define the beamspace
processing and machine learning codebook design algorithm.
 19

Neural architecture

RSRP
SSBRI
Convert the SSB codebook Convert the predicted
to the beamspace beamspace to beamformers

Fig. 4. A depiction of the BSC pre-processing, neural architecture, and post-processing. The feedback is converted to an angular
representation (beamspace) and the strongest directions are identified based on the feedback reported. The neural network jointly
processes the SSB feedback and beamspace representation to produce a new beamspace codebook that is then converted back to
complex beamforming coefficients. This unifies the various codebooks with a consistent input dimensionality even with varying
numbers of users or even antenna sizes.

III. B EAMSPACE -C ODEX

In this section, we present the Beamspace-Codex which is a neural network architecture and
processing setup that is used to generate the SSB codebooks, with considerations for joint SSB-
CSI-RS codebook design left to future work. We start by introducing the beamspace observation
that provides a consistent basis for a learning algorithm using dynamic codebooks. We then
present the supervised learning formulation that enables convergent and dynamic codebooks. Fi-
nally, we define the neural network architecture that enables learning the underlying relationship
between the beamspace observation and the SSB codebook. The full methodology–processing,
architecture, training, and evaluation–is critical to integrating wireless domain knowledge and
physical structure into the learning algorithm.

One of the primary steps in designing an appropriate neural network is formulating the input-
output relationship to be conducive and consistent for learning. There are two problems with
neural networks within the context of 5G beam management. First, the SSBRI feedback corre-
sponds to a beamformer from a codebook that was used previously, which does not provide a
consistent representation or meaning with dynamic codebooks. Second, the number of users is
not constant within a cell, so the size of the feedback and data dimension changes over time.
Although there are specific ways to overcome dynamic data sizes, it often requires sacrificing
 20

hardware optimizations and impedes inference and training times.

We propose transforming the codebook into the beamspace domain. The beamspace is an angular
representation of a beamformer corresponding to the array factor evaluated over a range of
directions [31]. This ensures that regardless of the specific beamforming codeword, or even
physical antenna dimension, the corresponding input from a set of previous beamformers still
represents a two-dimensional grid of projections. The beamspace conversion for Nx0 azimuth
directions and Ny0 elevation direction is calculated as

1
θNm = [0, 1, ..., Nm − 1]T (27)
π
UNs ,Nm , [aNs (θ0 ), ...aNs (θNm −1 )] ∈ CNs ×Nm (28)
(N ,NY )
FSSBX [i] = [FSSB [i, 0 : NX ]T , FSSB [i, NX : 2NX ]T , ..., FSSB [i, (NX − 1)NY : NX NY ]T ] (29)
(N ,NY )
OBSC [i] = UNx0 ,NX FSSBX [i]UNY ,Ny0 ∀i ∈ {1, 2, ...Lmax }. (30)

The conversion starts by generating an angular matrix from (28) for the azimuth directions
UNx0 ,NX and elevation directions UNY ,Ny0 as a series of array responses. Then, the codebook
must be reshaped from a vector of size NT to the planar dimensions NX × NY in (29) before
the beamspace conversion in (30) that produces the beamspace observation, OBSC . In addition
to the angular representation, the input is also concatenated with the feedback corresponding
to the number of users reporting each beam and the sum RSRP after min-max normalization.
Note that the beamspace conversion is a reversible operation, so we also train the network to
predict the beamspace of the desired output, rather than direct beamforming coefficients. The
predicted beamformers are obtained in post-processing, and the computational complexity can be
controlled by changing the observation sizes Nx0 and Ny0 . With the pre/post-processing defined,
we can now define the supervised learning environment and desired outputs used for training.

Supervised learning provides many advantages over unsupervised or reinforcement learning

when there is a desired output compared to just a black-box metric. The task of determining
beamforming vectors that maximize the power to a set of users has an optimal solution if CSI is
available via singular value decomposition (SVD) beamforming. To that end, we assume a dataset
of channels has been built up or simulated to allow for offline training, although our site-specific
results in Section IV suggest that transferring a learned model between sites is still more effective
 21

than DFT codebooks. Then, the solution to maximize the power received for a user is the right
singular vectors (RSV) of the SVD of the channel [4]. In the MU-MIMO OFDM case, this means
taking the RSVs corresponding to the Lmax strongest singular values of all users. Additionally,
the timing of the beams is important and each beam has a specified set of resources on which
to be transmitted. We must sacrifice the time relationship of the data, however, to build a true
supervised setting. In particular, we do not use information from the past SSB feedback beyond
the previous step. This limitation removes the capability of using historical data as inputs but
significantly improves convergence compared to a reinforcement learning setting. The process
is restricted to a two-step relationship where we include the previous codebook and feedback to
help design the next codebook. We now have a supervised framework and pre/post-processing
step that converts the previous codebook and RSV codebook to the beamspace domain. The final
component of the Beamspace-Codex is the neural architecture that facilitates efficient learning.

Determining the best neural network architecture is often challenging even with hyperparameter
tuning and an intuitive understanding of the problem. Using the beamspace conversion, the
problem can roughly be seen as a translation task from the initial beamspace to the RSV
beamspace. It is not necessarily obvious how the limited feedback information could be used
to recover the RSV codebook. There is some underlying information available, however, by
the selection of one beam over Lmax − 1 other options. Additionally, the environment and user
densities follow a pattern that is also possible to exploit. At the same time, the inputs and
outputs are similar to an image of the angular directions of transmitted power. Therefore, we
considered traditional architectures (fully-connected networks), image-based architectures (con-
volutional and diffusion [32] networks), and translation-based architectures (transformers [33]
and autoencoders). Although the problem has many similarities with both vision and translation
problems, state-of-the-art architectures were not necessarily best under reasonable computational
limits i.e. 24GB of VRAM and 5ms inference time. Throughout our testing, the fully connected
architecture showed the strongest results with an average of 1dB higher average RSRP than other
architectures, while also being the most computationally efficient one.

We propose a multi-layer fully connected network with significant regularization through dropout
comprised of 6 densely connected layers with parameters and sizes determined through Bayesian
hyperparameter tuning and defined in Table I. The model is trained using an Adam optimizer
[34] with early stopping and cosine learning rate decay. Training is performed with a 550, 000
 22

TABLE I
B EAMSPACE -C ODEX NETWORK PARAMETERS

Layer Primary Parameter Activation Func. Output Dimension

Flatten (2Lmax (Nx0 + 2)(Ny0 + 2))
Fully Conn. 144 Neurons ReLU (144)
Dropout 0.2 Rate (144)
Fully Conn. 1808 Neurons ReLU (1808)
Dropout 0.4 Rate (1808)
Fully Conn. 272 Neurons ReLU (272)
Dropout 0.4 Rate (272)
Fully Conn. 240 Neurons ReLU (240)
Fully Conn. 80 Neurons ReLU (80)
Fully Conn. (2Nx0 Ny0 Lmax ) Neurons (Nx0 , Ny0 , 2Lmax )
Reshape (Nx0 , Ny0 , 2Lmax )

sample training dataset following the setup in Section V with validation and test sets each
from an unseen portion of the data corresponding to 160, 000 and 80, 000 samples respectively.
Gradients are determined by the cosine distance (dcos (a, b) = 1 − a∗ b/(kak kbk)) between the
RSV beamspace and the neural network-predicted beamspace. The choice of the loss function is
critical for performance as the same network architectures trained on mean squared error results
in significantly worse (> 5dB) performance due to focusing on the pixel-wise magnitude, rather
than the relative magnitude or direction of the RSV beamspace.

IV. DATA GENERATION SETUP

We require an accurate channel simulator coupled with a flexible framework that retains the
timing of the 5G beam training process to fairly evaluate the performance of ML-designed
codebooks. We integrate the channel generation from QuaDRiGa [24] with a post-processing
suite that enables an initial access scenario, beamspace generation, codebook determination,
and evaluation of the network performance after accounting for beam management overhead.
QuaDRiGa generates channels through a stochastic process with additional features for spatial
consistency, correlations, and spherical wave modeling. In our simulations, we simulate 200
users scattered over a base station sector site. The BS is equipped with an NX = 4, NY = 8
planar array using a 3GPP 3D antenna model with half-wavelength spacing. Each user has
NR = 4 antennas all tuned for a 3.5GHz carrier frequency. 10% of users are given a vehicular
 23

mobility pattern such that they travel along a roadway through the center of the cell with speeds
normally distributed with mean 25m/s and variance 5m/s. The remaining users are uniformly
scattered within 450m and travel in any direction with speeds uniformly drawn from an interval
of [0, 3]m/s. Channel distributions follow a 3GPP urban macrocell environment and channels are
sampled over a bandwidth of 100MHz every 1ms for 2s. A large number of users (≈ 200) are
simulated to build up a database that is spatially and environmentally consistent. In the final
section of the results, we address how the model generalizes to a new setting with different
environmental parameters and different roadway locations.

Once the channels are generated, post-processing is necessary to produce a realistic setting and
build up the algorithm’s datasets. The process starts by randomly selecting a number of active
users (uniformly between [4, 12]) and gathering the channels and time slots from the channel
set. Then randomly selecting Lmax DFT beams for the ‘prior’ codebook and calculating the RSV
codebook. The codebooks are converted to beamspace representations and stored in a dataset
along with the SSB feedback from a subset of the active users where each active user is assigned
an 80% probability of its feedback being included. This represents an initial access channel
where not all users are previously known and introduces an additional regularization term into
our algorithm to prevent over-focusing on known users. It is assumed that UEs use antenna
selection for the RSRP reception during the SSB process, digital combining for CSI-RS, and
LMMSE combining during data transmission as defined in the system model.

V. S IMULATION RESULTS

It is critical to extensively evaluate a neural network model to fairly represent and compare it
against traditional methods. In our previous work [35], we have considered a learning threshold
as a minimal condition for determining if meaningful relationships have been learned by the
network. In realistic wireless channels, such settings are much harder to characterize, so we
instead benchmark the performance against industry-standard DFT techniques and RSV code-
books assuming CSI were known perfectly. This helps to identify the upper-performance limits
(with respect to maximum RSRP) as well as the minimal performance needed to justify moving
beyond traditional codebooks. We then use our robust simulation framework to better understand
how the various codebooks and type-II CSI feedback parameters affect network performance in
realistic FR1 channels. In the first result, the RSRP performance of the SSB process with various
 24

codebooks is plotted. Then the subsequent SNR achieved during CSI-RS is evaluated, where the
CSI-RS beams are determined by decomposing the SSB beam into a subset of DFT beams with
an equal proportion to the number of users selecting the beam during SSB.

A. RSRP and SNR

The RSRP is a basic metric that can be used to directly demonstrate the ability of the neural net-
work to output a valuable SSB codebook. Following the SSB process, we also highlight the per-
formance that the dynamic codebook provides when selecting CSI-RS beams as a decomposition
of the SSB codebook. While these two metrics–RSRP and SNR–do not ultimately characterize the
network performance as a result of the codebook algorithm, they do provide meaningful insight
into the performance of various codebook methods. Figure 5a shows the empirical cumulative
distribution function (CDF) of the RSRP reported using various codebook/beamforming methods
compared to our proposed BSC algorithm. Without any beamforming, more than 50% of users
would be below the minimum RSRP for signal detection, which is often around −120dBm.
The SSB codebooks are decomposed into CSI-RS codebooks in Figure 5b, where we see the
RSV codebook is not consistently the best SSB codebook when directly decomposed into CSI-
RS codebooks, although the decomposed BSC is not universally better than decomposed DFT
codebooks either. We can understand this as a result of the rich scattering environment that causes
the RSV to not effectively decompose into a small number of incoherent DFT components. While
RSV beamformers maximize the received power for a user, this maximization is based on the sig-
nal coherently combining and is not well-represented by a narrow DFT beam. From these results,
we can see that the BSC algorithm is advantageous compared to traditional methods, but we also
see that restricting the CSI-RS codebook to oversampled DFT beams is a significant limitation.

B. CSI-RS reporting

In this subsection, we provide an in-depth evaluation of the CSI-RS process without BSC code-
books. We evaluate performance based on the effective downlink SE after overhead due to beam
training, uplink feedback, and an additional signaling overhead factor of 10%. Uplink feedback
is assumed to be the only resource allocated for uplink transmission, and we assume a worst-case
scenario where users transmit with MCS0 from Table 5.1.3.1-1 [25] with users allocated at the
resource block level to account for additional overhead. User selection during the data phase is
performed based only on the knowledge at the BS by first estimating the sum spectral efficiency
 25

1.0 DFT DFT (CSI-RS)

BSC BSC (CSI-RS)
RSV 24 RSV (CSI-RS)
0.8 Non-Beamformed

Average SNR Reported [dB]

22
0.6
Empirical CDF

Gain of BF
20
0.4
18
0.2 Gain of BSC
2.8 dB 16
0.0
160 140 120 100 80 60 0 2 4 6 8 10 12 14 16 18
RSRP [dBm] Sample Number

(a) SSB RSRP (b) CSI-RS SNR

Fig. 5. Two plots of the codebook performance measuring (a) RSRP received during SSB and (b) SNR during CSI-RS reception.
During the SSB stage, each of the codebooks is used directly and then in (b) the SSB codebooks are decomposed to form the
active CSI-RS codebook subset. Non-beamformed transmission presents a significant loss in performance shown by the “gain of
BF” arrow, while the learned BSC codebook shows gains of almost 3dB over traditional DFT beams. While RSV beamforming
is RSRP-maximizing during the SSB, the DFT decomposition of an RSV codebook is not necessarily SNR-maximizing due to
multipath propagation.

achieved with the imperfect CSI for all users with {1, 2} data layers. The best estimated SE
combination is selected and used. Such a format is naturally suboptimal, even with perfect CSI,
due to a lack of rank and power adaption. User selection is not the focus of this investigation,
so we only use an exhaustive search over the reconstructed channel knowledge available at the
BS to provide a basic but reasonable scheduling algorithm. The SE is also compared with rank-
1 non-PMI beamforming where RZF is not employed, the channel is not estimated, and the only
feedback is obtained from the CRI, with user selection and precoder determination simply being
the CSI-RS beamforming codewords with the strongest reported user RSRP. This is often an envi-
sioned beam management strategy for sparse mmWave channels with short, low-latency feedback,
but is not assumed to be effective in sub-6GHz channels due to the rich scattering environment.

The size of the CSI-RS codebook is related to the channel estimation SNR because larger
CSI-RS codebooks, made of orthogonal DFT beamformers, will result in equivalent or higher
SNR compared to smaller codebooks with the same channels and feedback. We characterize the
performance of larger or smaller CSI-RS codebooks in Figure 6a. Although the estimation SNR,
and by extension performance before quantization, will always improve with a larger codebook,
there is also a modest amount of overhead with larger PCSI so that there is little spectral efficiency
 26

gain beyond PCSI ≥ 8.

In the next comparison, the feedback quantization is evaluated by modifying the parameter LCSI
in type-II formats. We can see from Fig. 6b that higher resolution feedback is beneficial, with
the effective data rate exceeding 2 times higher when the resolution is improved from LCSI = 1
to LCSI = 32. From there on, the other parameters CSI-RS resource block allocation (NRB)
and the number of frequency selective resources (BWP), are less influential compared to the
quantization resolution. Most of the modest performance gains achieved with multiple frequency
selective precoders and larger downlink CSI-RS resource usage are offset by the increasing
overhead resulting in negligible or even reductive performance impacts. Joining the information
from Figures 6a-7a, a good balance of feedback and performance can be achieved with a setting
of LCSI = 32, BWP = 1, NRB = 24, PCSI = 8. It is important to note that the current 5G Release
16, even with enhanced type-II feedback, only supports LCSI <= 6, which means most situations
will only yield about 1.5 times higher effective spectral efficiency relative to SU-MIMO, and
essentially equivalent performance to the non-PMI beamforming. Perhaps even more surprising
is that using a single CSI-RS beam (PCSI = 1) has less degradation than using LCSI = 8 feedback
beams per rank, relative to the best performance. This suggests that feedback limitations are
more likely to restrict MU-MIMO performance in sub-6GHz systems compared to codebook
optimization or accurate channel estimation.

C. Site generalization

Up to now, we have used a test set of data drawn from the same statistical environment as
the training set. While a network operator will likely gather data for a specific site, here we
examine the impact of testing on a new distribution of data. We simulate a new environment with
completely different user distributions and mobility patterns. Now 50% of users are vehicular with
roadways defined in new locations and a new stochastic realization of the channel environment.
In Figure 8 the agnostic (orange) data is obtained without updating the model, whereas the fine-
tuned data (blue) shows the results after we allow for 60s of retraining on data from the new
environment. Our time limit is arbitrary and could ultimately be orders of magnitude shorter than
the timescale a network operator might use to update the model, but corresponds to approximately
1% of the training time used in our setup for a balance of fine-tuning and generalization. It can
be seen in Figure 8 the performance in the new environment (agnostic) is significantly worse
 27

1.0 1.0

0.8 0.8
Pcsi=1 Lcsi=1
Empirical CDF

Empirical CDF
Pcsi=2
0.6 Pcsi=4 0.6 Lcsi=4
Pcsi=8 Lcsi=8
Pcsi=12 Lcsi=16
0.4 SU-Pcsi=1 0.4
SU-Pcsi=2 Lcsi=24
0.2 SU-Pcsi=4
0.2 Lcsi=32
SU-Pcsi=8 SU-MIMO
SU-Pcsi=12
Non-PMI Non-PMI
0.0 0.0
0 10 20 30 0 10 20 30
Effective Sum SE [bps/Hz] Effective Sum SE [bps/Hz]

(a) Codebook size PCSI (b) Quantization paths LCSI

Fig. 6. Evaluation of the effects of active CSI-RS codebook size (a) and quantization using LCSI feedback components (b).
We can use the size of the codebook as a means of looking at how codebook optimization improves performance because the
overhead with larger PCSI is very small compared to the overhead of type-II feedback. SU-MIMO results using type-I PMI are
shown by dashed lines while MU-MIMO results are shown by solid lines and Non-PMI refers to rank 1 beamforming using the
CRI. SU-MIMO does not include multipath feedback so only a single SU-MIMO result is shown in (b). Other parameters are
set to LCSI = 32 for (a), PCSI = 16 for (b), BWP = 1, NRB = 24. We can see that the loss from using LCSI = 4 (3GPP Release
16 maximum support) compared to LCSI = 32 is more detrimental than restricting the system to a single CSI-RS beamformer.

1.0 1.0

0.8 0.8
Empirical CDF

Empirical CDF

0.6 BWP=1 0.6 N_RB=24

BWP=2 N_RB=48
0.4 BWP=4 0.4 N_RB=96
SU-BWP=1 SU-N_RB=24
0.2 SU-BWP=2 0.2 SU-N_RB=48
SU-BWP=4 SU-N_RB=96
Non-PMI Non-PMI
0.0 0.0
0 10 20 30 0 10 20 30
Effective Sum SE [bps/Hz] Effective Sum SE [bps/Hz]

(a) BWP (b) NRB

Fig. 7. A comparison of the effective spectral efficiency with different numbers of frequency selective precoders BWP and
varying the number of resource blocks (NRB) per BWP. It is assumed that LCSI = 32, PCSI = 16, NRB = 24 × BWP in (a)
and BWP = 1 in (b). While increasing the number of resource blocks, therefore increasing the number of pilots, also increases
the channel estimation performance, the rising feedback overhead causes the performance to decrease in (b). In comparison,
the number of BWP increases the overhead linearly, yet the performance is essentially equivalent in (a). Therefore we see that
frequency selective feedback and beamforming is not necessarily advantageous when overhead is accounted for.

relative to the RSV. The performance loss is noticeably improved during fine-tuning, especially
for the lowest 10% of users.
 28

BSC(Agnostic)
0.200 BSC (fine-tuned)

0.175

0.150

0.125

Probability
0.100

0.075

0.050

0.025

0.000
20 15 10 5 0 5 10
RSRP Offset from RSV [dB]

Fig. 8. Histogram representing the performance delta between BSC and RSV codebooks in the agnostic setting and after fine-
tuning. It can be seen that the worst-case differences are resolved and in some cases (when more users are present than SSB beams)
the RSV codebook can be outperformed based on the higher proportion of the blue columns to the right of the orange columns.

VI. C ONCLUSION

In this paper, we presented a novel framework for learning dynamic codebooks for sub-6GHz 5G
NR. We set up a system with multiple UEs each equipped with a uniform linear array to receive
from a base station with a fully digital planar array. We simulate realistic sub-6GHz channels
using QuaDRiGa and build a post-processing framework for evaluating the system performance
of SSB and CSI-RS codebooks and feedback. Using this framework we address two questions
related to codebook performance in 5G: 1) Can machine learning design more effective SSB
codebooks? 2) How should feedback be configured for MU-MIMO in sub-6GHz environments?

We first proposed a codebook transformation based on beamspace projections that allowed for
efficiently and consistently representing SSB codebooks in a format conducive to learning.
This beamspace representation can circumvent challenges with dynamic user numbers, changing
codebooks, and even deploying across different array sizes. Using the beamspace as a translation
tool, we designed a neural architecture, Beamspace-Codex, to output new codebooks from SSB
feedback. The proposed method directly integrates with current 5G standards and improves SSB
RSRP by 2-3dB on average. We also show that only 1% of the training compute budget is
needed for retraining in new environments to significantly improve the worst-case performance.
 29

Secondly, we found that the PMI quantization resolution, determined by the number of feedback
DFT beams, is a limiting factor in the performance of MU-MIMO in current standards.

R EFERENCES
[1] R. M. Dreifuerst and R. W. Heath, “Massive MIMO in 5G: How beamforming, codebooks, and feedback enable larger
arrays,” in submission to IEEE Commun. Mag, 2023. [Online]. Available: https://arxiv.org/abs/2301.13390

[2] K. Lee, J. Kim, E. W. Jin, and K. S. Kim, “Extreme massive mimo for upper-mid band 6g communications,” in 2022 13th
International Conference on Information and Communication Technology Convergence (ICTC), 2022, pp. 997–999.

[3] G. Morozov, A. Davydov, and V. Sergeev, “Enhanced CSI feedback for FD-MIMO with beamformed CSI-RS in LTE-A
Pro systems,” in Proc. of the IEEE Veh. Technol. Conf., Jul. 2016, pp. 1–5.

[4] V. Raghavan et al., “Beamforming Tradeoffs for Initial UE Discovery in Millimeter-Wave MIMO Systems,” IEEE J. Sel.
Top. Signal Process., vol. 10, no. 3, pp. 543–559, Apr. 2016.

[5] Y. Heng et al., “Six key challenges for beam management in 5.5G and 6G systems,” IEEE Commun. Mag., vol. 59, no. 7,
pp. 74–79, Jul. 2021.

[6] M. Giordani et al., “A tutorial on beam management for 3GPP NR at mmwave frequencies,” IEEE Commun. Surv. Tutorials,
vol. 21, no. 1, pp. 173–196, Jan. 2019.

[7] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel estimation and hybrid precoding for millimeter wave
cellular systems,” IEEE J. Sel. Top. Signal Process., vol. 8, no. 5, pp. 831–846, Oct. 2014.

[8] E. Becirovic, E. Bjornson, and E. G. Larsson, “Combining reciprocity and CSI feedback in MIMO systems,” IEEE Trans.
Wirel. Commun., May 2022.

[9] Samsung, “Massive MIMO for new radio,” Dec. 2020. [Online]. Available: https://www.samsung.com/global/business/
networks/insights/white-papers/1208-massive-mimo-for-new-radio/

[10] B. Li et al., “On the efficient beam-forming training for 60GHz wireless personal area networks,” IEEE Trans. Wireless
Commun., vol. 12, no. 2, pp. 504–515, 2013.

[11] V. Va et al., “Inverse multipath fingerprinting for millimeter wave V2I beam alignment,” IEEE Trans. Veh. Technol., vol. 67,
no. 5, pp. 4042–4058, May 2018.

[12] Y. Wang, N. J. Myers, N. González-Prelcic, and R. W. Heath, “Site-specific online compressive beam codebook learning
in mmWave vehicular communication,” IEEE Trans. Wirel. Commun., pp. 1–14, 2021.

[13] Q. Xue et al., “Beam management in ultra-dense mmWave network via federated reinforcement learning: An intelligent
and secure approach,” IEEE Trans. Cogn. Commun. Netw., 2022.

[14] J. Yang, W. Zhu, and M. Tao, “Deep learning for hierarchical beam alignment in mmWave communication systems,” in
Proc. of the IEEE Global Communications Conference, 2022.

[15] J. Wang et al., “Beam codebook based beamforming protocol for multi-gbps millimeter-wave WPAN systems,” IEEE
Journal on Selected Areas in Communications, vol. 27, no. 8, pp. 1390–1399, 2009.

[16] Z. Xiao, T. He, P. Xia, and X.-G. Xia, “Hierarchical codebook design for beamforming training in millimeter-wave
communication,” IEEE Trans. Wireless Commun., vol. 15, no. 5, pp. 3380–3392, 2016.
 30

[17] R. Shafin et al., “Self-tuning sectorization: Deep reinforcement learning meets broadcast beam optimization,” IEEE Trans.
Wirel. Commun., vol. 19, no. 6, pp. 4038–4053, Jun. 2020.

[18] W. Xia et al., “A deep learning framework for optimization of MISO downlink beamforming,” IEEE Trans. Commun.,
vol. 68, no. 3, pp. 1866–1880, Mar. 2020.

[19] Y. Heng, J. Mo, and J. G. Andrews, “Learning site-specific probing beams for fast mmwave beam alignment,” IEEE Trans.
Wirel. Commun., vol. 21, no. 8, pp. 5785–5800, 2022.

[20] R. M. Dreifuerst, R. W. Heath, and A. Yazdan, “Massive MIMO beam management in sub-6 GHz 5G NR,” in Proc. of
the IEEE Veh. Technol. Conf., 2022, pp. 1–5.

[21] C. K. Wen, W. T. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wirel. Commun. Lett., vol. 7,
no. 5, pp. 748–751, 2018.

[22] J. Kim, H. Lee, and S.-H. Park, “Learning robust beamforming for MISO downlink systems,” IEEE Communications
Letters, vol. 25, no. 6, pp. 1916–1920, 2021.

[23] M. Chen et al., “Deep learning-based implicit CSI feedback in massive MIMO,” IEEE Trans. Commun., vol. 70, pp. 935–
950, Dec. 2022.

[24] S. Jaeckel, L. Raschkowski, K. Börner, and L. Thiele, “QuaDRiGa: A 3-D multi-cell channel model with time evolution
for enabling virtual field trials,” IEEE Trans. Antennas Propag., pp. 3242–3256, 2014.

[25] 3GPP, “TS 38.214 V16.7.1 NR; Physical layer procedures for data.”

[26] A. Singh and S. Joshi, “A survey on hybrid beamforming in mmWave massive MIMO system,” J. Sci. Res., vol. 65, no. 01,
pp. 201–213, 2021.

[27] R. W. Heath Jr. and A. Lozano, Foundations of MIMO Communication. Cambridge University Press, 2018.

[28] L. Godara, “Application of antenna arrays to mobile communications. ii. beam-forming and direction-of-arrival consider-
ations,” Proceedings of the IEEE, vol. 85, no. 8, pp. 1195–1245, 1997.

[29] M. R. Castellanos et al., “Channel-Reconstruction-Based Hybrid Precoding for Millimeter-Wave Multi-User MIMO Sys-
tems,” IEEE J. Sel. Areas Commun., vol. 12, no. 2, pp. 383–398, May 2018.

[30] R. W. Heath et al., “An overview of signal processing techniques for millimeter wave MIMO systems,” IEEE J. Sel. Top.
Signal Process., vol. 10, no. 3, pp. 436–453, Apr. 2016.

[31] A. Sayeed, “Deconstructing multiantenna fading channels,” IEEE Transactions on Signal Processing, vol. 50, no. 10, pp.
2563–2579, 2002.

[32] J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” 2020. [Online]. Available: https:
//arxiv.org/abs/2006.11239

[33] Z. Liu et al., “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proc. of the IEEE Int. Conf.
Comput. Vis., Mar. 2021, pp. 9992–10 002.

[34] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” International Conference on Learning Representations,
Dec. 2014.

[35] R. M. Dreifuerst and R. W. Heath, “SignalNet: A low resolution sinusoid decomposition and estimation network,” IEEE
Trans. Signal Process., vol. 70, pp. 4454–4467, Aug. 2022.