U.S. patent application number 15/347027 was filed with the patent office on 2018-05-10 for beamforming in antenna systems.
The applicant listed for this patent is Facebook, Inc.. Invention is credited to Sungwoo Park, Ali YAZDAN PANAH.
Application Number | 20180131423 15/347027 |
Document ID | / |
Family ID | 62045256 |
Filed Date | 2018-05-10 |
United States Patent
Application |
20180131423 |
Kind Code |
A1 |
Park; Sungwoo ; et
al. |
May 10, 2018 |
BEAMFORMING IN ANTENNA SYSTEMS
Abstract
Apparatuses, methods, and systems for beamforming in antenna
systems are disclosed. A method includes determining an
unconstrained analog precoding matrix (F.sub.RF,UC), wherein the
unconstrained analog precoding matrix (F.sub.RF,UC) is determined
based on M dominant eigenvectors of the sum of spatial channel
covariance matrices of K users, and wherein K indicates a number of
users communicating with a base station. The method further
includes determining a constrained analog precoding matrix
(F.sub.RF) based on the unconstrained analog precoding matrix
(F.sub.RF,UC), determining a compensation matrix (F.sub.CM),
digitally multiplying K inputs with a multiple-input
multiple-output (MIMO) precoding matrix (F.sub.MU) generating M
outputs, digitally multiplying the M outputs with the compensation
matrix (F.sub.CM) generating M compensation outputs, generating M
analog frequency-up-converted signals based on the M compensation
outputs, and analog multiplying the M analog frequency-up-converted
signals with the analog precoding matrix (F.sub.RF) generating N
output signals for transmission, wherein N is greater than M.
Inventors: |
Park; Sungwoo; (Austin,
TX) ; YAZDAN PANAH; Ali; (San Francisco, CA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Facebook, Inc. |
Menlo Park |
CA |
US |
|
|
Family ID: |
62045256 |
Appl. No.: |
15/347027 |
Filed: |
November 9, 2016 |
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
H04L 25/03891 20130101;
H04B 7/0617 20130101; H04L 25/03343 20130101; H04W 88/08 20130101;
H04B 7/0456 20130101; H04B 7/0452 20130101 |
International
Class: |
H04B 7/06 20060101
H04B007/06; H04B 7/04 20060101 H04B007/04 |
Claims
1. A base station, comprising: baseband precoding circuitry,
wherein the baseband precoding circuitry receives K inputs and
digitally multiplies the K inputs with a multiple-input
multiple-output (MIMO) precoding matrix (F.sub.MU) generating M
outputs, and wherein K indicates a number of users communicating
with the base station; a compensation circuitry, wherein the
compensation circuitry digitally multiplies the M outputs of the
baseband precoding circuitry with a compensation matrix (F.sub.CM)
generating M compensation outputs; M radio frequency (RF) chains,
wherein each RF chain is configured to receive one of the M
compensation outputs, and generate an analog frequency-up-converted
signal; and analog precoding circuitry, wherein the analog
precoding circuitry receives the M analog frequency-up-converted
signals and analog multiplies the M analog frequency-up-converted
signals with a constrained analog precoding matrix (F.sub.RF)
generating N output signals for transmission, wherein N is greater
than M; wherein the constrained analog precoding matrix (F.sub.RF)
is determined based on an unconstrained analog precoding matrix
(F.sub.RF,UC), and wherein the unconstrained analog precoding
matrix (F.sub.RF,UC) is determined based on dominant eigenvectors
of the sum of spatial channel covariance matrices of the K users;
and wherein the compensation matrix (F.sub.CM) is determined based
on the constrained analog precoding matrix (E.sub.R).
2. The base station of claim 1, wherein K is less than or equal to
M.
3. The base station of claim 1, wherein multiplication of the
unconstrained analog precoding matrix (F.sub.RF,UC) with any
invertible matrix is substantially equal to the constrained analog
precoding matrix (F.sub.RF).
4. (canceled)
5. The base station of claim 1, wherein the MIMO precoding matrix
(F.sub.MU) is determined based on an effective channel matrix that
comprises one or more of the constrained analog precoding matrix
(F.sub.RF), the compensation matrix (F.sub.CM), and a raw channel
matrix.
6. The base station of claim 1, wherein the constrained analog
precoding circuitry comprises phase shifters, and wherein the
analog multiplication of the unconstrained analog precoding matrix
(F.sub.RF,UC) controls phases of the analog frequency-up-converted
signals.
7. The base station of claim 1, wherein the constrained analog
precoding circuitry comprises phase shifters, wherein
multiplication of the constrained analog precoding matrix
(F.sub.RF) and the compensation matrix (F.sub.CM) is substantially
equal to the unconstrained analog precoding matrix
(F.sub.RF,UC).
8. A method, comprising: determining an unconstrained analog
precoding matrix (F.sub.RF,UC), wherein the unconstrained analog
precoding matrix (F.sub.RF,UC) is determined based on M dominant
eigenvectors of the sum of spatial channel covariance matrices of K
users, and wherein K indicates a number of users communicating with
a base station; determining a constrained analog precoding matrix
(F.sub.RF) based on the unconstrained analog precoding matrix
(F.sub.RF,UC); determining a compensation matrix (F.sub.CM),
wherein the compensation matrix (F.sub.CM) is determined based on
the constrained analog precoding matrix (F.sub.RF); digitally
multiplying K inputs with a multiple-input multiple-output (MIMO)
precoding matrix (F.sub.MU) generating M outputs; digitally
multiplying the M outputs with the compensation matrix (F.sub.CM)
generating M compensation outputs; generating M analog
frequency-up-converted signals based on the M compensation outputs;
and analog multiplying the M analog frequency-up-converted signals
with the analog precoding matrix (F.sub.RF) generating N output
signals for transmission, wherein N is greater than M.
9. The method of claim 8, wherein K is less than or equal to M.
10. The method of claim 8, wherein multiplication of the
unconstrained analog precoding matrix (F.sub.RF,UC) with any
invertible matrix is substantially equal to the constrained analog
precoding matrix (F.sub.RF).
11. (canceled)
12. The method of claim 8, wherein the MIMO precoding matrix
(F.sub.MU) is determined based on an effective channel matrix that
comprises one or more of the constrained analog precoding matrix
(F.sub.RF), the compensation matrix (F.sub.CM), and a raw channel
matrix.
13. The method of claim 8, wherein the constrained analog precoding
circuitry comprises phase shifters, and wherein the analog
multiplication of the unconstrained analog precoding matrix
(F.sub.RF,UC) controls phases of the analog frequency-up-converted
signals.
14. The method of claim 8, wherein the constrained analog precoding
circuitry comprises phase shifters, wherein multiplication of the
constrained analog precoding matrix (F.sub.RF) and the compensation
matrix (F.sub.CM) is substantially equal to the unconstrained
analog precoding matrix (F.sub.RF,UC).
15. A system, comprising: one or more processors; and a
non-transitory computer-readable storage device including one or
more instructions for execution by the one or more processors and
when executed operable to perform operations comprising:
determining an unconstrained analog precoding matrix (F.sub.RF,UC),
wherein the unconstrained analog precoding matrix (F.sub.RF,UC) is
determined based on dominant eigenvectors of the sum of spatial
channel covariance matrices of K users, and wherein K indicates a
number of users communicating with a base station; determining a
constrained analog precoding matrix (F.sub.RF) based on the
unconstrained analog precoding matrix (F.sub.RF,UC); determining a
compensation matrix (F.sub.CM) , wherein the compensation matrix
(F.sub.CM) is determined based on the constrained analog
precoding(F.sub.RF); digitally multiplying K inputs with a
multiple-input multiple-output (MIMO) precoding matrix (F.sub.MU)
generating M outputs; digitally multiplying the M outputs with the
compensation matrix (F.sub.CM) generating M compensation outputs;
generating M analog frequency-up-converted signals based on the M
compensation outputs; and analog multiplying the M analog
frequency-up-converted signals with the analog precoding matrix
(F.sub.RF) generating N output signals for transmission, wherein N
is greater than M.
16. The system of claim 15, wherein K is less than or equal to
M.
17. The system of claim 15, wherein multiplication of the
unconstrained analog precoding matrix (F.sub.RF,UC) with any
invertible matrix is substantially equal to the constrained analog
precoding matrix (F.sub.RF).
18. (canceled)
19. The system of claim 15, wherein the MIMO precoding matrix
(F.sub.MU) is determined based on an effective channel matrix that
comprises one or more of the constrained analog precoding matrix
(F.sub.RF), the compensation matrix (F.sub.CM), and a raw channel
matrix.
20. The system of claim 15, wherein the constrained analog
precoding circuitry comprises phase shifters, and wherein the
analog multiplication of the unconstrained analog precoding matrix
(F.sub.RF,UC) controls phases of the analog frequency-up-converted
signals.
Description
FIELD OF THE DESCRIBED EMBODIMENTS
[0001] The described embodiments relate generally to wireless
communications. More particularly, the described embodiments relate
to systems, methods and apparatuses for beamforming in antenna
systems.
BACKGROUND
[0002] Multiple-input multiple-output (MIMO) technology is commonly
considered a potential candidate for next generation wireless
communication, whereby a base station equipped with many antennas
simultaneously communicates with multiple users sharing time and
frequency resources. In wireless systems, transmitted signals to a
user may cause interference in other systems. Also, many antennas
require many radio frequency (RF) chains, which increase power
consumption.
[0003] It is desirable to have apparatuses, methods, and systems
for beamforming in multiple antenna systems.
SUMMARY
[0004] An embodiment includes a base station. The base station
includes a baseband precoding circuitry, wherein the baseband
precoding circuitry receives K inputs and digitally multiplies the
K inputs with a MIMO precoding matrix (F.sub.MU) generating M
outputs, and wherein K indicates a number of users communicating
with the base station. The base station further includes a
compensation circuitry, wherein the compensation circuitry
digitally multiplies the M outputs of the baseband precoding
circuitry with a compensation matrix (F.sub.CM) generating M
compensation outputs. The base station further includes M RF
chains, wherein each RF chain is configured to receive one of the M
compensation outputs, and generate an analog frequency-up-converted
signal. The base station further includes analog precoding
circuitry, wherein the analog precoding circuitry receives the M
analog frequency-up-converted signals and analog multiplies the M
analog frequency-up-converted signals with a constrained analog
precoding matrix (F.sub.RF) generating N output signals for
transmission, wherein N is greater than M, wherein the constrained
analog precoding matrix (F.sub.RF) is determined based on an
unconstrained analog precoding matrix (F.sub.RF,UC), and wherein
the unconstrained analog precoding matrix (F.sub.RF,UC) is
determined based on M dominant eigenvectors of the sum of spatial
channel covariance matrices of the K users.
[0005] Another embodiment includes a method. The method includes
determining an unconstrained analog precoding matrix (F.sub.RF,UC),
wherein the unconstrained analog precoding matrix (F.sub.RF,UC) is
determined based on M dominant eigenvectors of the sum of spatial
channel covariance matrices of K users, and wherein K indicates a
number of users communicating with a base station. The method
further includes determining a constrained analog precoding matrix
(F.sub.RF) based on the unconstrained analog precoding matrix
(F.sub.RF,UC). The method further includes determining a
compensation matrix (F.sub.CM) based on the constrained analog
precoding matrix (F.sub.RF). The method further includes digitally
multiplying K inputs with a MIMO precoding matrix (F.sub.MU)
generating M outputs. The method further includes digitally
multiplying the M outputs with the compensation matrix (F.sub.CM)
generating M compensation outputs. The method further includes
generating M analog frequency-up-converted signals based on the M
compensation outputs. The method further includes analog
multiplying the M analog frequency-up-converted signals with the
analog precoding matrix (F.sub.RF) generating N output signals for
transmission, wherein N is greater than M.
[0006] Another embodiment includes a system. The system includes
one or more processors, and includes logic encoded in one or more
non-transitory computer-readable storage media for execution by the
one or more processors. When executed, the logic is operable to
perform operations including determining an unconstrained analog
precoding matrix (F.sub.RF,UC), wherein the unconstrained analog
precoding matrix (F.sub.RF,UC) is determined based on M dominant
eigenvectors of the sum of spatial channel covariance matrices of K
users, and wherein K indicates a number of users communicating with
a base station. The logic when executed is further operable to
perform operations including determining a constrained analog
precoding matrix (F.sub.RF) based on the unconstrained analog
precoding matrix (F.sub.RF,UC). The logic when executed is further
operable to perform operations including determining a compensation
matrix (F.sub.CM) based on the constrained analog precoding matrix
(F.sub.RF). The logic when executed is further operable to perform
operations including digitally multiplying K inputs with a MIMO
precoding matrix (F.sub.MU) generating M outputs. The logic when
executed is further operable to perform operations including
digitally multiplying the M outputs with the compensation matrix
(F.sub.CM) generating M compensation outputs. The logic when
executed is further operable to perform operations including
generating M analog frequency-up-converted signals based on the M
compensation outputs. The logic when executed is further operable
to perform operations including analog multiplying the M analog
frequency-up-converted signals with the analog precoding matrix
(F.sub.RF) generating N output signals for transmission, wherein N
is greater than M.
[0007] Aspects and advantages of the described embodiments will
become apparent from the following detailed description, taken in
conjunction with the accompanying drawings, illustrating by way of
example of the principles of the described embodiments.
BRIEF DESCRIPTION OF THE DRAWINGS
[0008] FIG. 1 shows a base station environment, according to an
embodiment.
[0009] FIG. 2 shows a base station performing unconstrained hybrid
precoding, according to an embodiment.
[0010] FIG. 3 shows a base station, according to an embodiment.
[0011] FIG. 4 shows a base station performing constrained hybrid
precoding, according to an embodiment.
[0012] FIG. 5 shows a flow chart that includes acts of a method,
according to an embodiment.
[0013] FIG. 6 shows a portion of a base station including an analog
precoder with phase shifters, according to an embodiment.
DETAILED DESCRIPTION
[0014] At least some embodiments described include methods,
apparatuses, and systems for providing hybrid analogue-digital
beamforming in multi-antenna systems. For at least some
embodiments, a base station includes a baseband precoding
circuitry, wherein the baseband precoding circuitry receives K
inputs and digitally multiplies the K inputs with a MIMO precoding
matrix F.sub.MU generating M outputs, and wherein K indicates a
number of users communicating with the base station. The base
station further includes a compensation circuitry, wherein the
compensation circuitry digitally multiplies the M outputs of the
baseband precoding circuitry with a compensation matrix F.sub.CM
generating M compensation outputs. The base station further
includes M RF chains, wherein each RF chain is configured to
receive one of the M compensation outputs, and generate an analog
frequency-up-converted signal. The base station further includes
analog precoding circuitry, wherein the analog precoding circuitry
receives the M analog frequency-up-converted signals and analog
multiplies the M analog frequency-up-converted signals with a
constrained analog precoding matrix F.sub.RF generating N output
signals for transmission, wherein N is greater than M, wherein the
constrained analog precoding matrix F.sub.RF is determined based on
an unconstrained analog precoding matrix F.sub.RF,UC, and wherein
the unconstrained analog precoding matrix F.sub.RF,UC is determined
based on M dominant eigenvectors of the sum of spatial channel
covariance matrices of the K users.
[0015] At least some embodiments provide a hybrid precoding
technique for massive MIMO systems using long-term channel
statistics. As described in more detail herein, constraining the
baseband precoding matrix to a regularized zero-forcing precoder,
an unconstrained analog precoder improves the
signal-to-leakage-plus-noise ratio (SLNR) while ignoring any analog
phase shifter constraints. Subsequently, for an embodiment, a
constrained analog precoder mimics the obtained unconstrained
analog precoder under phase shifter constraints. At least some
embodiments also involve the adoption of an additional baseband
precoding matrix, referred to as a "compensation matrix." If the
number of channels correlated and the number of users is smaller
than the number of RF chains, the performance loss becomes
negligible compared to full digital precoding. Benefits of at least
some embodiments described herein stem from the use of long-term
channel statistics in such a way that the analog precoder improves
each user's desired signal, and also reduces inter-user
interference.
[0016] FIG. 1 shows a base station environment 100, according to an
embodiment. Base station environment 100 may also be referred to as
a downlink multiuser MIMO system. As shown, a base station includes
a digital baseband precoder 110, RF chains 120a and 120b, and
analog RF precoder 130. The base station transmits data across one
or more transmissions channels 140a and 140b to mobile stations
150a and 150b (also labeled MS 1 and MS K, respectively).
[0017] For an embodiment, a base station is equipped with N
antennas and M RF chains, which simultaneously communicate with K
users sharing time and frequency resources. In wireless systems, a
simultaneously transmitted signal to a user acts as an interference
to other systems. To overcome this problem, the base station
exploits so-called precoding techniques that reduce interference in
advance with prior knowledge of the wireless channel state. In
spite of its potential for increasing system throughput and
reliability, conventional hybrid precoding techniques in massive
MIMO systems require high power consumption because each antenna
requires its own RF chain, and there may be many RF chains.
[0018] FIG. 2 shows a base station 200 performing unconstrained
hybrid precoding, according to an embodiment. As shown, base
station 200 includes a baseband precoder 210, which processes an
RZF precoding matrix. Base station 200 also includes RF chains 220a
and 220b, and analog precoder 230, which processes an analog
precoding matrix 232.
[0019] For an embodiment, base station 200 performs hybrid
analog/digital precoding, which enables the use of fewer RF chains
than antennas (M<N). As described in more detail herein, for an
embodiment, the base station 200 achieves this by using long-term
channel statistics instead of instant channel state
information.
[0020] In at least one embodiments, the channels vary over time,
where, for example, the user k's N.times.1 channel vector, h.sub.k,
varies over time, where h.sub.k(1), h.sub.k(2), . . . , h.sub.k(T)
are not the same, and where N is the number of antennas. This may
be referred to an "instance channel" at a particular time. The user
k's spatial channel covariance may be defined as
R.sub.k=[h.sub.kh.sub.k*].
[0021] The user k's spatial channel covariance may be estimated
as
R k .apprxeq. 1 T t = 1 T h k ( t ) h k * ( t ) . ##EQU00001##
[0022] This spatial channel covariance does not change at every
time slot (where a time slot is predetermined period of time), and
slowly changes in the long-term (e.g., changes less than a
threshold amount over a predetermined number of time slots), if at
all. As such, the "spatial channel covariance" is one of the
"long-term channel statistics." Other long-term channel statistics
may include mean and variance.
[0023] This spatial channel covariance matrix may contain the
information about the number of dominant channel paths between a
base station and a mobile station, and may contain the information
about the channel paths' angles relative to the antenna bore-sight.
At least some embodiments exploit this information contained in the
spatial channel covariance in that the analog precoding matrix is
based on the spatial channel covariance matrix.
[0024] More specifically, at least some embodiments, use the
spatial channel covariance, which is one of long-term channel
statistics, in the design of the analog precoding matrix, rather
than using instant channel information, h.sub.k's.
[0025] The spatial channel covariance matrix is used for the design
of the unconstrained analog precoding matrix F.sub.RF,UC, and the
unconstrained analog precoding matrix F.sub.RF,UC is used for the
design of the constrained analog precoding matrix F.sub.RF. The
process that obtains the unconstrained matrix F.sub.RF,UC is just
an interim process to obtain the constrained matrix F.sub.RF, which
is the actual analog precoding matrix. As such the spatial channel
covariance matrix is used for the design of the unconstrained
precoding matrix as well.
[0026] Hybrid analog/digital transmit precoding described herein
reduces power consumption by using a reduced set of RF chains while
maintaining a large number of physical antennas. The hybrid
precoding method essentially divides the precoding process at the
transmitter between the analog RF and digital baseband part.
[0027] At least some embodiments use long-term channel statistics,
e.g., spatial channel covariance in the analog precoder instead of
instantaneous channel knowledge with respect to analog precoding.
Firstly, as its name implies, long-term channel statistics show
much less variation in time compared to instantaneous channels,
which makes it easier to estimate. Further, long-term channel
statistics such as spatial channel covariance are typically uniform
across all subcarriers. At least some embodiments of the analog
precoder use the long-term channel statistics.
[0028] FIG. 3 shows a base station 300, according to an embodiment.
As shown, base station 300 includes a baseband precoder 310,
compensation circuitry 314, RF chains 320, and an analog precoder
330. Baseband precoder 310 receives K inputs and digitally
multiplies the K inputs with a MIMO precoding matrix F.sub.MU
generating M outputs, and where K indicates a number of users
communicating with the base station. For an embodiment, the MIMO
precoding matrix F.sub.MU is determined based on an effective
channel matrix that includes one or more of a constrained analog
precoding matrix F.sub.RF, the compensation matrix F.sub.CM, and a
raw channel matrix.
[0029] In at least some embodiments, the raw channel matrix
indicates the actual channel matrix between the physical antennas
at the base station and the single antenna at the mobile station's.
The (n, k)-th element of the actual (raw) channel matrix is the
actual channel path gain between the n-th physical antenna at the
base station and the antenna at the k-th mobile station.
[0030] In at least some embodiments, the output of the baseband
precoding in the base station is not directly connected to the
physical antennas, but connected to the physical antennas through
the analog precoding matrix. That is, the output of baseband
precoding is connected to the RF chains. As such, from the
viewpoint of the baseband precoding, the RF chains may be regarded
as effective antennas (e.g., not actual physical antennas). In this
case, the effective channel becomes the combination of the analog
precoding and the actual (raw) channel, e.g.,
H.sub.eff,UC=H'F.sub.RF,UC. More specifically, in the constrained
case, the effective channel from the viewpoint of the baseband
MU(RZF) precoding, F.sub.RZF, the effective channel is the
combination of the compensation matrix, the analog precoding
matrix, and the actual channel matrix, e.g.,
H.sub.eff,c=H'F.sub.RF,CF.sub.CM.
[0031] Compensation circuitry 314 digitally multiplies the M
outputs of baseband precoder 310 with a compensation matrix
F.sub.CM generating M compensation outputs. For an embodiment, the
compensation matrix F.sub.CM is determined based on the constrained
analog precoding matrix F.sub.RF. For an embodiment, K is less than
or equal to M. For an embodiment, analog precoder 330 includes
constrained analog precoding circuitry that includes phase
shifters.
[0032] In an embodiment, there are M RF chains, where each RF chain
is configured to receive one of the M compensation outputs, and
generate an analog frequency-up-converted signal. Analog precoder
330 receives the M analog frequency-up-converted signals and analog
multiplies the M analog frequency-up-converted signals with a
constrained analog precoding matrix F.sub.RFgenerating N output
signals for transmission, where N is greater than M.
[0033] As described in more detail herein, the constrained analog
precoding matrix F.sub.RF may be determined based on an
unconstrained analog precoding matrix F.sub.RF,UC. Also, the
unconstrained analog precoding matrix F.sub.RF,UC may be determined
based on M dominant eigenvectors of the sum of spatial channel
covariance matrices of the K users.
[0034] In at least some embodiments, the phase shifter constraint
is initially ignored in the determination of the analog precoding
matrix, and the best analog precoding matrix is found in order to
improve the spectral efficiency, based on the spatial channel
covariance matrices.
[0035] In at least some embodiments, the spectral efficiency may be
defined as {data transfer rate (bits per second)/bandwidth (Hz)}.
The "data transfer rate" is simply referred to as the rate or
throughput. In addition, if multiple users are simultaneously
receiving data from a base station, the sum of all the users' rate
may be referred to as the "sum rate." Since the bandwidth is a
fixed constant, maximizing the "spectral efficiency" is equivalent
to maximizing the "rate." The term "spectral efficiency" may be
used interchangeably with the terms "rate," "throughput," or "data
transfer rate."
[0036] After summing up the spatial channel covariance matrices of
all users, eigenvalues of the sum matrix are calculated by using an
eigenvalue decomposition. Then, the unconstrained analog precoding
matrix is constructed such that its M columns are composed of the M
eigenvectors associated with the M largest eigenvalues, which are
so-called M dominant eigenvectors.
[0037] In at least some embodiments, selecting the largest
eigenvalues depends on the number of RF chains that a base station
has. That is, the base station already knows how many eigenvalues
are to be selected. For some embodiments, the number of eigenvalues
selected is based on the number of RF chains that the base station
has. If the base station has N antennas, then there exist N
eigenvalues. If M RF chains are equipped in the base station, the
base station selects M eigenvalues among N eigenvalues. For at
least an embodiment, the selection criterion is to choose M largest
eigenvalues. In other words, the base station arranges the N
eigenvalues in a descending order, and then selects the first M
eigenvalues in that descending order.
[0038] After finding the unconstrained analog precoding matrix,
which ignores the phase shifter constraint, the actual analog
precoding matrix may be determined based on application of the
phase shifter constraint. For at least an embodiment, to reduce the
loss caused by the phase shifter constraint, the compensation
matrix is applied in order to compensate for the loss. As such, if
the phase shifter constraint is not applied, the compensation
matrix may be ignored, as it is would not be necessary.
[0039] In at least some embodiments, the constrained analog
precoding matrix combined with the compensation matrix is
equivalent to the unconstrained analog precoding matrix, and there
is no loss from the phase shifter constraint. In at least some
embodiments, the combined constrained matrix is as similar to the
unconstrained matrix as possible. The Frobenius norm of the
difference between two matrices may be used as a metric to define
similarity, where the smaller the metric, the more similar the two
matrices.
[0040] For an embodiment, multiplication of the unconstrained
analog precoding matrix F.sub.RF,UC with any invertible matrix is
substantially equal to the constrained analog precoding matrix
F.sub.RF. For an embodiment, the analog multiplication of the
unconstrained analog precoding matrix (F.sub.RF,UC) controls phases
of the analog frequency-up-converted signals.
[0041] FIG. 4 shows a base station 400 performing constrained
hybrid precoding, according to an embodiment. As shown, base
station 400 includes a baseband precoder 410 that includes a
digital precoder 412, which processes an RZF precoding matrix.
Baseband precoder 410 also includes compensation circuitry 414 that
processes a compensation matrix. Base station 400 also includes RF
chains 420a and 420b, and an analog precoder 430, which processes
an analog precoding matrix.
[0042] Baseband precoder 410 and compensation circuitry 414
constitute a digital baseband portion of base station 400, and
analog precoder 430 is an analog RF portion of base station
400.
[0043] Fixing a baseband precoder as regularized zero-forcing, at
least some embodiments obtain an unconstrained analog precoder that
improves the SLNR ignoring the fact that analog precoding is
typically realized with phase shifters (e.g., so-called phase
shifter constraints may be ignored). For an embodiment, the
obtained unconstrained analog precoder may be configured to use a
channel covariance matrix.
[0044] As described in more detail herein, analog precoder 430 of
base station 400 applies a phase shifter constraint, such that the
analog precoding matrix F.sub.RF is a constrained analog precoding
matrix F.sub.RF, composed of phase shifters such that the phase of
a signal is controlled, and where the amplitude of a signal is not
controlled.
[0045] For an embodiment, a constrained analog precoder mimics the
obtained unconstrained analog pre-coder while satisfying the phase
shifter constraints. The key idea is adopting an additional
baseband precoder, referred to as a compensation matrix whose main
role is mitigating the loss caused by using phase shifters.
Although this compensation matrix is operated in the baseband, for
an embodiment, the compensation matrix may depend on the long-term
channel statistics, thereby requiring no instantaneous channel
state information at transmitter (CSIT). Leveraging this
compensation matrix concept, the constrained analog precoder may be
improved so that the combination of the constrained analog precoder
and the compensation matrix is as similar to the unconstrained
matrix as possible.
[0046] At least some embodiments achieve long-term CSIT for the
analog precoding matrix F.sub.RF, and may be applied for wideband
systems (e.g., frequency selective). At least some embodiments may
be applied to massive MIMO cellular systems, to low-power
consumption base stations (e.g., equipped unmanned aircraft or
balloons, etc.), and to next generation wireless communication
systems (e.g., millimeter wave systems, etc.).
[0047] A benefit of at least some embodiments described herein is
that they do not dedicate an RF chain to a user (or a group of RF
chains to a user group). Dedicating RF chains to users is not
efficient in that its applicability is restricted to the case when
the number of users is equal to the number of RF chains. Instead,
at least some embodiments described herein improve the performance
in a general system environment. At least some embodiments may be
applied to the case when fewer users than RF chains are assigned,
which is beneficial in a realistic scenario where the number of
assigned users varies over time but the number of RF chains is
fixed.
[0048] Under the phase shifter constraints, the constrained analog
precoder with the compensation matrix results in almost the same
spectral efficiency as the unconstrained analog precoder. In
addition, the loss from the hybrid precoding can be low, although
the proposed technique utilizes long-term channel statistics in the
design of the analog precoder, promoting the employment of massive
MIMO systems in practical real-world cellular networks.
[0049] FIG. 5 is a flow chart that includes acts of a method,
according to an embodiment. A first step 510 includes determining
an unconstrained analog precoding matrix F.sub.RF,UC, where
unconstrained analog precoding matrix F.sub.RF,UC is unconstrained
in that a phase shifter constraint is not applied. For an
embodiment, the unconstrained analog precoding matrix F.sub.RF,UC
is determined based on M dominant eigenvectors of the sum of
spatial channel covariance matrices of K users, and where K
indicates a number of users communicating with a base station.
[0050] A second step 520 includes determining a constrained analog
precoding matrix F.sub.RF, where unconstrained analog precoding
matrix F.sub.RF,UC is unconstrained in that a phase shifter
constraint is not applied. For an embodiment, the constrained
analog precoding matrix F.sub.RF is based on the unconstrained
analog precoding matrix F.sub.RF,UC.
[0051] As indicated herein, the constrained analog precoding matrix
F.sub.RF is performed by constrained analog precoding circuitry,
which includes phase shifters.
[0052] FIG. 6 shows a portion of a base station 600 including an
analog precoder 630 with phase shifters, according to an
embodiment. As shown, base station 600 includes RF chains 620a to
620m, a power divider 632, phase shifters 634, a power combiner
636, and antennas 640a, 640b to 640n. In at least some embodiments,
base station 600 includes M RF chains and N antennas, where there
are fewer RF chains than antennas. In other words, M<N.
[0053] A third step 530 includes determining a compensation matrix
F.sub.CM. For an embodiment, the compensation matrix F.sub.CM is
determined based on the constrained analog precoding matrix
F.sub.RF.
[0054] For an embodiment, wherein multiplication of the constrained
analog precoding matrix F.sub.RF and the compensation matrix
F.sub.CM is substantially equal to the unconstrained analog
precoding matrix F.sub.RF,UC, where
F.sub.RFF.sub.CM.apprxeq.F.sub.RF,UC.
[0055] A fourth step 540 includes digitally multiplying K inputs
with a MIMO precoding matrix F.sub.MU generating M outputs. As
indicated herein, the MIMO precoding matrix F.sub.MU may be
determined based on an effective channel matrix that includes one
or more of the constrained analog precoding matrix F.sub.RF, the
compensation matrix F.sub.CM, and a raw channel matrix.
[0056] For an embodiment, the analog precoding matrix F.sub.RF,C
provides phase control of a signal, and the MIMO precoding matrix
F.sub.MU provides phase and amplitude control of a signal. For an
embodiment, where the phase shifter constraints are not applied,
the unconstrained analog precoding matrix is a conceptual precoder
that is used for the calculation of the actual analog precoder. If
the phase shifter constraints are not applied, it is sufficient to
calculate just the unconstrained analog precoding matrix
F.sub.RF,UC. In at least some embodiments, once the unconstrained
analog precoding matrix F.sub.RF,UC is obtained, the process for
the design of both the compensation matrix and the constrained
matrix is not necessary.
[0057] A fifth step 550 includes digitally multiplying the M
outputs with the compensation matrix F.sub.CM generating M
compensation outputs.
[0058] A sixth step 560 includes generating M analog
frequency-up-converted signals based on the M compensation
outputs.
[0059] A seventh step 570 includes analog multiplying the M analog
frequency-up-converted signals with the analog precoding matrix
F.sub.RF generating N output signals for transmission, wherein N is
greater than M.
[0060] The following descriptions describe further embodiments in
detail. The following notation may be used throughout. A is a
matrix, a is a vector, a is a scalar, and is a set. |a| and
.angle.a are the magnitude and phase of the complex number a.
.parallel.A.parallel..sub.F is its Frobenius norm, and A.sup.T, A*,
and A.sup.-1 are its transpose, Hermitian (conjugate transpose),
and inverse, respectively. [A].sub.m,n is the (m, n)-th element of
the matrix A. .angle.A is a matrix with the (m, n)-th element
equals e.sup.j[A]mn. I.sub.N is an N.times.N identity matrix and
0.sub.N.times.M is an N.times.M matrix whose elements are all
zeros. (m, R) is a complex Gaussian random vector with mean m and
covariance R. is used to denote expectation.
[0061] Consider a downlink system where a base station (BS)
equipped with N antennas and M (.ltoreq.N) RF chains communicates
with K (.ltoreq.M) users with a single antenna. Let
F.sub.RF.di-elect cons. .sup.N.times.M, F.sub.BB .di-elect
cons..sup.M.times.K, and s .di-elect cons..sup.K.times.1 be an
analog RF precoder, a digital baseband precoder, and a signal
vector, respectively. The transmit signal is given by
x=F.sub.RFF.sub.BBP.sub.S, (1)
where P .di-elect cons..sup.K.times.K is a diagonal matrix to
maintain the total transmit power P.sub.tx.
[0062] The received signal is given by
y=H*x+n=H*F.sub.RFF.sub.BBP.sub.S+n, (2)
where n .di-elect cons..sup.K.times.1.about.(0, .sigma..sup.2I) is
circularly symmetric complex Gaussian noise, and H* .di-elect
cons..sup.K.times.N is a downlink channel matrix for all users.
Considering spatially correlated channels, each MS has its own
spatial channel covariance matrix and the channel is modeled as
H=[h.sub.1. . . h.sub.K=[R.sub.1.sup.1/2h.sub.W,1. . .
R.sub.K.sup.1/2h.sub.W,K], (3)
where h.sub.W,K has IID complex entries of zero mean and unit
variance, and R.sub.k=[h.sub.kh.sub.k*] is a spatial channel
covariance matrix of user k. This presumes that R.sub.k's have been
obtained through covariance estimation for the hybrid
structure.
[0063] Assuming that the analog precoding is composed of phase
shifters, a constraint on F.sub.RFis imposed that all elements in
F.sub.RFhave the same amplitude. For an embodiment, there is
another constraint that F.sub.RFis designed by using R.sub.k's, not
h.sub.k's. In at least some embodiments, there are three
constraints on hybrid precoding that may be applied.
[0064] Constraint 1: The number of RF chains is less than the
number of antennas (M<N).
[0065] Constraint 2: F.sub.RFis fixed over time and/or frequency
and may depend on long-term channel statistics.
[0066] Constraint 3: F.sub.RFis composed of phase shifters (e.g.,
all the elements in F.sub.RFhave the same amplitudes).
[0067] In at least some embodiments described herein, h, h.sub.w,k,
and n are regarded as random variables, and R.sub.k's as
deterministic variables.
[0068] In at least some embodiments, hybrid precoding using
long-term channel statistics is used, where constraints 1 and 2 are
applied. For an embodiment, to determine an unconstrained analog
precoding matrix (applying constraints 1 and 2, but not constraint
3), and, presuming long-term channel statistics to design F.sub.RF,
each column of F.sub.RFis assigned to each MS as
F.sub.RF=[v.sub.1,max. . . v.sub.K,max]. (4)
where V.sub.K,max is a dominant eigenvector of R.sub.k.
[0069] Once F.sub.RFis decided, the baseband precoder adopts
conventional MU-MIMO techniques such as zero-forcing (ZF) or
regularized zero-forcing (RZF) with respect to the combined
effective channel H*.sub.eff=H*F.sub.RF. The rationale behind this
technique is to improve the long-term average power of the desired
signal in the analog part. The main drawback of this approach is,
however, not considering the interference in the analog part, which
results in performance degradation unless the channel is ideally
orthogonal. Moreover, this technique cannot be directly applied
when K<M.
[0070] For an embodiment, a focus on the RZF case may be
F.sub.BB=[f.sub.BB,1. . .
f.sub.BB,K]=(F*.sub.RFHH*F.sub.RF+.beta.I.sub.M).sup.-1F*.sub.RFH,
(5)
where .beta. is a regularization parameter and is set as
.beta. = K .sigma. 2 P tx = K .rho. , where .rho. = P tx .sigma. 2
##EQU00002##
denotes the transmit SNR. Considering an equal power strategy that
makes each user's power equal after precoding including both
F.sub.RFand F.sub.BB, the k-th diagonal element of P in (1) is
p k = P tx K F RF f bb , k = P tx Kh k * W 2 h k , where W = F RF (
F RF * HH * F RF + K .rho. I M ) - 1 F RF * . ( 6 )
##EQU00003##
[0071] The instantaneous SLNR can be written as
SLNR k = h k * F RF f bb , k 2 i .noteq. k h i * F RF f bb , k 2 +
F RF f bb , k 2 ( .sigma. 2 p k 2 ) = h k * Wh k h k * Wh k i
.noteq. k h k * Wh i h i * Wh k + K .rho. h k * W 2 h k = h k * Wh
k h k * Wh k h k * W ( HH * - h k h k * + K .rho. I N ) Wh k . ( 7
) ##EQU00004##
[0072] The goal is to find F.sub.RFto improve the SLNR in (7).
Instead of assigning each column of F.sub.RFto each user, a
subspace spanned by orthonormal bases {V.sub.1, ..., V.sub.M} may
be found where v.sub.m .di-elect cons..sup.N and |V.sub.m|=1,
.A-inverted.m=1, . . . , M. Therefore, there is no constraint such
as K=M, so this approach can be applied for the case of K<M as
well. In at least some embodiments, allocating smaller users than M
is better than allocating M users when M RF chains are given.
[0073] Let each column of F.sub.RF a linear combination of the
bases, then F.sub.RF can be represented as
F.sub.RF=VA (8)
where A .di-elect cons..sup.M.times.M is an invertible matrix and
V=[V.sub.1. . . V.sub.M] .di-elect cons..sup.N.times.M, where
.sup.N.times.M is a set of N.times.M semi-unitary matrices as
.sup.N.times.M={X|X*X=I.sub.M, X .di-elect cons..sup.N.times.M}.
(9)
[0074] In the following proposition, SLNR in (7) has a maximum
value when A is unitary, and thus F.sub.RFis semi-unitary.
[0075] Proposition 1: If V and P.sub.tx is given, SLNR in (7) is
improved when A is unitary.
[0076] Proof: Let {tilde over (H)}*=H*V and =h*.sub.kV. Then, SLNR
in (7) can be written as
SLNR k = AWA * AWA * AWA * ( - + K .rho. I M ) AWA * = ( - + K
.rho. I M ) = ( + K .rho. I M ) - , where = ( + K .rho. ( AA * ) -
1 ) - 1 . Let .delta. A be defined as ( 10 ) .delta. A = = ( - + K
.rho. I M ) , ( 11 ) ##EQU00005##
where =. Then, the SLNR in (10) can be rewritten as
SLNR k = .delta. A 1 - .delta. A = 1 1 .delta. A - 1 ( 12 )
##EQU00006##
[0077] Note that SLNR is improved when (S.sub.A has a predetermined
increased value. The that improves .delta..sub.A has the same
direction as the generalized eigenvector of ({tilde over
(H)}+k/.rho.I.sub.M. Since
+ K .rho. I M ##EQU00007##
is invertible, a solution of has a form as
.varies. + K .rho. I M ) - 1 .varies. ( + K .rho. I M ) - 1 , ( 13
) ##EQU00008##
which implies that (AA*).sup.-1=I.sub.M.
[0078] When A is unitary, .delta..sub.A has the maximal value of
.delta..sub.A=({tilde over (H)}+k/.rho.I.sub.M).sup.-1and SLNR in
(10) is expressed as
SLNR k = h k * V ( V * HH * V + K .rho. I M ) - 1 V * h k 1 - h k *
V ( V * HH * V + K .rho. I M ) - 1 V * h k = h k * V ( V * ( i
.noteq. k K h i h i * ) V + K .rho. I M ) - 1 V * h k = h w , k * R
k 1 / 2 V ( i .noteq. k K V * R i 1 2 h w , i h w , i * R i 1 2 V +
K .rho. I M ) - 1 V * R k 1 2 h w , k ( 14 ) ##EQU00009##
where the second equality comes from the matrix inversion
lemma.
[0079] Since SLNR is independent on A as long as A is unitary,
constructing V in (8) may be performed to improve SLNR. Note that
SLNR in (14) is a random variable due to h.sub.W,k. The random
variable SLNR, however, converges to a deterministic value as the
number of antennas becomes large.
[0080] Let h.sub.k=R.sub.k.sup.1/2h.sub.W,k= {square root over
(N)}R.sub.k.sup.1/2g.sub.k, where g.sub.k has IID complex entries
with zero mean and variance of 1/N. Then, as N goes to infinity,
the SLNR in (14) converges to
SLNR k = Ng k H R k 1 2 V ( N i .noteq. k K V * R i 1 2 g i g i * R
i 1 2 V + K .rho. I M ) - 1 V * R k 1 2 g k a . s . Tr ( R k 1 2 V
( N i .noteq. k K V * R i 1 2 g i g i * R i 1 2 V + K .rho. I M ) -
1 V * R k 1 2 ) a . s . Tr ( R k 1 2 V ( N i = k K V * R i 1 2 g i
g i * R i 1 2 V + K .rho. I M ) - 1 V * R k 1 2 ) = Tr ( V * R k V
( i = 1 K V * R i 1 2 g i g i * R i 1 2 V + K .rho. I M ) - 1 ) (
15 ) ##EQU00010##
where the first convergence comes from the trace lemma, and the
second convergence comes from the rank-1 perturbation lemma.
[0081] The random variable SLNR is converged to a deterministic
SLNR value, as N goes to infinity, as
SLNR k a . s . .gamma. k , ( 16 ) ##EQU00011##
where .gamma..sub.1, . . . , .gamma..sub.k are the unique
nonnegative solution of
.gamma. k = Tr ( V * R k V ( j = 1 K V * R j V 1 + .gamma. j + K
.rho. I M ) - 1 ) . ( 17 ) ##EQU00012##
[0082] The solution of .gamma..sub.1, . . . , .gamma..sub.k can be
obtained in fixed-point equations as
.gamma..sub.k=log.sub.t.fwdarw..infin..gamma..sub.k.sup.(t)
where
.gamma. k ( t ) = Tr ( V * R k V ( j = 1 K V * R j V 1 + .gamma. j
( t - 1 ) + K .rho. I M ) - 1 ) . ( 18 ) ##EQU00013##
[0083] Let consider the problem that improves the asymptotic SLNR
averaged over all users as
s . t . .gamma. k = Tr ( V * R k V ( j = 1 K V * R j V 1 + .gamma.
j + K .rho. I M ) - 1 ) , .A-inverted. k . ( 19 ) ##EQU00014##
[0084] Since this is difficult to solve directly due to K fixed
point equations in (19), to relax the problem, presuming that all
users have the same SLNR as
.gamma. 1 = = .gamma. k = .gamma. = 1 K k = 1 K .gamma. k .
##EQU00015##
[0085] Then, the problem becomes
max V .di-elect cons. N .times. M .gamma. s . t . .gamma. = Tr ( V
* R tot V ( KV * R tot V 1 + .gamma. + K .rho. I M ) - 1 ) , ( 20 )
##EQU00016##
where
R tot = 1 K k = 1 K R k . ##EQU00017##
Let V*R.sub.totV be decomposed as UAU*by eigenvalue decomposition
and have eigenvalues of v.sub.1, . . . , v.sub.M in descending
order. Then, .gamma. is rewritten as
.gamma. = 1 K Tr ( U .LAMBDA. U * ( U .LAMBDA. U * 1 + .gamma. + 1
.rho. I M ) - 1 ) = 1 K m = 1 M ~ 1 1 1 + .gamma. + 1 .rho. v m (
21 ) ##EQU00018##
where {tilde over (M)}=min(M, rank(R.sub.tot)). Then, a solution to
(20) is given in the following proposition.
[0086] Proposition 2: The V that improves the SLNR in (20) is the
matrix whose columns are composed of M eigenvectors associated with
the M largest eigenvalues of
R tot = 1 K k = 1 K R k . ##EQU00019##
[0087] Proof: Let .lamda..sub.1, . . . , .lamda..sub.N be the
eigenvalues of R.sub.tot in descending order and V.sub.A=[V
V.sub.0] be a unitary matrix such that V.sub.0*V.sub.0=I.sub.N-M
and V*V.sub.0=0.sub.M.times.(N-M). Since V.sub.A is a unitary
matrix, V.sub.A*R.sub.totV.sub.A has the same eigenvalues as
R.sub.tot and can be represented as
V A * R tot V A = [ V * R tot V V * R tot V 0 V 0 * R tot V V 0 * R
tot V 0 ] . ( 22 ) ##EQU00020##
[0088] Let the eigenvalues of V.sub.A*R.sub.totV.sub.A be denoted
as v.sub.1.gtoreq.. . . .gtoreq.v.sub.M. Then, the eigenvalues of
the leading principal submatrix, V*R.sub.totV have the interlacing
property such as
.lamda..sub.N-M+i.ltoreq.v.sub.i.ltoreq..lamda..sub.i, for i=1, . .
. M. (23)
[0089] Since R.sub.tot is Hermitian, .lamda..sub.i for i=1, . . . ,
rank(R.sub.tot) have positive real values, and .lamda..sub.i for
i>rank(R.sub.tot) have zero values. Consequently, v.sub.i for
i>rank(R.sub.tot) become zeros, and
.lamda..sub.i.sup.-1.ltoreq.v.sub.i.sup.-1.ltoreq..lamda..sub.N-M+i.sup.-
-1, for i=1, . . . , {tilde over (M)}. (2)
[0090] From (24), the constraint in (20) becomes
.gamma. = 1 K m = 1 M ~ 1 1 1 + .gamma. + 1 .rho. v m .ltoreq. 1 K
m = 1 M ~ 1 1 1 + .gamma. + 1 .rho..lamda. m , ( 25 )
##EQU00021##
where the equality holds if V is composed of M dominant
eigenvectors of R.sub.tot. Since the solution of the fixed point
equation with respect to .gamma. has the maximum value if the
equality holds, the proof is completed.
[0091] The above proposition indicates that, in at least some
embodiments, the analog precoding F.sub.RFto the problem in (20)
uses the M dominant eigenvectors of
R tot = 1 K k = 1 K R k , ##EQU00022##
e.g., the sum of the spatial covariance matrices of K users.
[0092] Although the derived solution is based on the relaxed
problem assuming that large antenna arrays are equipped and the
SLNR per user is approximated to the average value over users, this
approximated solution significantly outperforms conventional
techniques and has spectral efficiency close to that of the fully
digital precoding in spatially correlated channels. In addition, it
can be proved that the proposed solution has exactly the same
spectral efficiency as that of the fully digital precoding if
R.sub.tot is rank-deficient and its rank is less than or equal to
M.
[0093] Even when the number of antennas is not so large and thus
the SLNR does not converge to a certain value, it can be proved
that the proposed analog precoding is beneficial in the sense that
at least some implementations improve the lower bound of the
expectation of the SLNR averaged over K users. The expectation of
the average SLNR over K users can be expressed as
[ 1 K k = 1 K SLNR k ] = Tr ( 1 K k = 1 K [ V * R k 1 2 h w , k h w
, k * R k 1 2 V ( i .noteq. k K V * R i 1 2 h w , i h w , i * R i 1
2 V + K .rho. I M ) - 1 ] ) = Tr ( 1 K k = 1 K [ V * R k 1 2 h w ,
k h w , k * R k 1 2 V ] [ ( i .noteq. k K V * R i 1 2 h w , i h w ,
i * R i 1 2 V + K .rho. I M ) - 1 ] ) .gtoreq. Tr ( 1 K k = 1 K V *
R k V ( i .noteq. k K V * R i V + K .rho. I M ) - 1 ) .gtoreq. Tr (
1 K k = 1 K V * R k V ( i = k K V * R i V + K .rho. I M ) - 1 ) . (
26 ) ##EQU00023##
where the second equality comes from the fact that h.sub.w,k's are
independent, and the first inequality comes from the fact that
[A.sup.-1]-([A]).sup.-1 is a positive semidefinite for a positive
semidefinite matrix A, and the second inequality comes from the
fact that A.sup.-1-(A+B).sup.-1 is a positive semidefinite for a
positive semidefinite matrix A and B.
[0094] With the same notation used in (20) and (21), the lower
bound in (26) can be represented as
[ 1 K k = 1 K SLNR k ] .gtoreq. Tr ( 1 K k = 1 K V * R tot V ( V *
R tot V + K .rho. I M ) - 1 ) = m = 1 M ~ 1 K + .rho. v m . ( 27 )
##EQU00024##
[0095] This lower bound expression in (27) has a similar form to
(21), and it can be easily proved that the V that improves the
expected average SLNR is the same as the solution in Proposition
2.
[0096] For an embodiment, RF chains construct a subspace for all
users as a whole in the analog precoding. For this reason, there is
no limitation on assigning the exactly same number of the users to
the number of the RF chains, providing a wide range of
applicability of the proposed method.
[0097] In at least some embodiments, hybrid precoding under a phase
shifter using long-term channel statistics may be used, where
constraints 1, 2, and 3 are applied, where F.sub.RFis composed of
phase shifters. Specifically, at least some embodiments provide a
technique to mimic F.sub.RFunder the phase shifter constraint. The
unconstrained F.sub.RFmay be derived as F.sub.RF,UC and its
constrained version as F.sub.RF,C.
[0098] For an embodiment, the following algorithm may be used.
TABLE-US-00001 Algorithm 1 Find F.sub.RF,C Input: F.sub.RF,UC
Initialization: F.sub.(0) = .angle.(F.sub.RF,UC), n = 0 repeat n
.rarw. n + 1 F.sub.(n) =
.angle.(F.sub.RF,UCF.sub.RF,UC*F.sub.(n-1)) until
||F.sub.RF,UCF.sub.RF,UC*F.sub.(n-1) - F.sub.(n)||.sub.F converges
Output: F.sub.RF,UC = F.sub.(n)
[0099] The previous way to make F.sub.RF,C as similar to
F.sub.RF,UC as possible is presented. A simple way to find the most
similar F.sub.RF,C is solving
min F RF , C , [ F RF , C ] i , j = 1 N F RF , UC - F RF , C F 2 .
( 28 ) ##EQU00025##
[0100] F.sub.RF,C that decreases the Frobenius norm of the
difference between F.sub.RF,C and F.sub.RF,UC is known as a
reasonable approximation of F.sub.RF,UC. The solution of (28) is
given by
[ F RF , C ( opt ) ] i , j = 1 N e j.angle. ( [ F RF , UC ] i , j )
, ##EQU00026##
where .angle.(.alpha.) denotes the phase of a complex number
.alpha.. The weakness of this approach is that F.sub.RF,C loses the
orthogonality that F.sub.RF,UC retains. Recall that F.sub.RF,UC may
be semi-unitary according to Proposition 1.
[0101] In at least some embodiments, to overcome this weakness, a
compensation matrix in the baseband part is applied to restore the
orthogonality lost in the analog part as shown in FIG. 2. The
compensation matrix F.sub.CM is designed by
F.sub.CM=(F*.sub.RF,CF.sub.RF,C).sup.-1/2, (29)
which makes F.sub.RF,CF.sub.CM semi-unitary. By applying the
compensation matrix, additional room is made for further
improvement in designing F.sub.RF,C. Denote A as an arbitrary
invertible matrix, which is decomposed by SVD as
U.sub.AD.sub.AV.sub.A*. Suppose that F.sub.RF,UCA is used instead
of F.sub.RF,UC in the unconstrained case. The unconstrained analog
precoder combined with the compensation matrix becomes
F.sub.RF,UCAF.sub.CM=F.sub.RF,UCA(A*F*.sub.RF,CF.sub.RF,C).sup.-1/2=F.su-
b.RF,UCU.sub.A, (30)
which satisfies the criterion as in the unconstrained analog
precoding. Therefore, the unconstrained analog precoding
F.sub.RF,UC can be replaced by F.sub.RF,UCA for any invertible
matrix A without any performance loss. Using this property, a
modified problem may be used instead of (28) as
min F RF , C , [ F RF , C ] i , j = 1 , A F RF , UC A - F RF , C F
2 . ( 31 ) ##EQU00027##
[0102] Thanks to the increased degrees of freedom of the design,
the constrained analog precoding F.sub.RF,C can be made closer to
constrained analog precoding. The solution to (31) can be obtained
by an alternating technique. In at least some embodiments, the
algorithm firstly finds the optimal A assuming that F.sub.RF,C is
fixed. Given a fixed F.sub.RF,C, the optimal A is given by
A ( opt ) = arg min A F RF , UC A - F RF , C F 2 = F RF , UC * F RF
, C . ( 32 ) ##EQU00028##
[0103] Then, assuming that A is fixed, the optimal F.sub.RF,C is
given by
F RF , C ( opt ) = min F RF , C , [ F RF , C ] i , j = 1 F RF , UC
A - F RF , C F 2 = .angle. ( F RF , UC A ) . ( 33 )
##EQU00029##
where .angle.(X) is a matrix whose (i, j)-th element is
e.sup.j.angle.([X].sup.i,j) Using (32) and (33), the solution can
be obtained from an iterative algorithm described in Algorithm
1.
[0104] Once F.sub.RF,C is decided, the compensation matrix F.sub.CM
is obtained from F.sub.RF,C and (29). The overall baseband
precoding in the constrained case is
F.sub.BB,C=F.sub.CMF.sub.RZF, (34)
where F.sub.RZF is an RZF precoder with respect to the effective
channel H.sub.eff,c=H*F.sub.RF,CF.sub.CM as
F.sub.RZF=(H*.sub.eff,cH.sub.eff,c+.beta.I.sub.M).sup.-1H*.sub.eff,c.
(35)
[0105] Algorithm 2 summarizes the overall process for the hybrid
precoding design under Constraint 1, 2, and 3.
[0106] For an embodiment, the following algorithm may be used.
TABLE-US-00002 Algorithm 2 Hybrid precoding design for multiuser
massive MIMO Step 1: Find an unconstrained analog precoding matrix
F.sub.RF,UC F RF , UC = M dominant eigenvectors of k = 1 K R k
##EQU00030## Step 2: Find a constrained analog precoding matrix
F.sub.RF,C using Algorithm 1 Step 3: Construct a baseband
compensation matrix, F.sub.CM as F.sub.CM = (F.sub.RF,C*F.sub.RF,C
).sup.-1/2 Step 4: Construct a baseband RZF precoding matrix,
F.sub.RZF, as F.sub.RZF = (F.sub.CM*F.sub.RF*HH*F.sub.CM +
.beta.I.sub.M).sup.-1F.sub.CM*F.sub.RF*H Step 5: Construct an
overall baseband precoding matrix F.sub.BB,C and an overall hybrid
precoding matrix F.sub.HB as F.sub.BB,C = F.sub.CMF.sub.RZF,
F.sub.HC = F.sub.RF,CF.sub.BB,C = F.sub.RF,CF.sub.CMF.sub.RZF
[0107] For an embodiment, as a measure of the loss, the ratio of
the asymptotic SLNR may be averaged over K users of the hybrid
precoding to that of the full digital precoding. Similarly to the
hybrid precoding case, the asymptotic SLNR of user k in the fully
digital precoding case can be represented as
SNLR k ( FD ) a . s . .gamma. k ( FD ) , ( 36 ) ##EQU00031##
where .gamma..sub.1.sup.(FD), . . . , .gamma..sub.k.sup.(FD) are
the unique nonnegative solution of
.gamma. k ( FD ) = Tr ( R k ( j = 1 K R j 1 + .gamma. j ( FD ) + K
.rho. I N ) - 1 ) . ( 37 ) ##EQU00032##
[0108] Let .gamma..sub.k.sup.(HB) denote the asymptotic SLNR of the
hybrid precoding in (16). Then, the performance metric may be
defined as
.eta. = 1 K k = 1 K .gamma. k ( HB ) 1 K k = 1 K .gamma. k ( FD ) ,
( 38 ) ##EQU00033##
and .eta. satisfies 0.ltoreq..eta..ltoreq.1. Note that 10
log.sub.10 .eta. indicates the average SLNR loss in dB caused by
the hybrid precoding compared to the fully digital precoding.
Therefore, if R.sub.1, . . . , R.sub.k are given, the SLNR loss can
be calculated by using (16), (36), and (38). The SLNR loss,
however, does not have a closed form due to the fixed point
equations.
[0109] In the following propositions, some special cases are
introduced where the SLNR loss metric has a closed form. For a
general case, an approximation of the SLNR loss metric in
Proposition 5 may be derived. Notations of
.kappa. = K N and .mu. = M N ##EQU00034##
may be used that denote the relative number of users and RF chains
compared to the number of antennas in the SLNR loss analysis. This
presumes that .kappa. and .mu. have constant values without
converging to zero as N goes to infinity. Note that
0.ltoreq..kappa..ltoreq..mu..ltoreq.1.
[0110] Proposition 3: For uncorrelated channels, e.g.,
R.sub.k=I.sub.n, .A-inverted.k, the SLNR loss metric .eta. is a
function of .kappa., .mu., and .rho. as
.eta. = ( ( .mu. - .kappa. ) .rho. - .kappa. ) + ( ( .mu. - .kappa.
) .rho. - .kappa. ) 2 + 4 .mu. .kappa. .rho. ( ( 1 - .kappa. )
.rho. - .kappa. ) + ( ( 1 - .kappa. ) .rho. - .kappa. ) 2 + 4
.kappa. .rho. , ( 39 ) ##EQU00035##
and if .rho..fwdarw..infin.and .mu.>.kappa., then .eta. can be
approximated to
.eta. = .mu. - .kappa. 1 - .kappa. . ( 40 ) ##EQU00036##
[0111] Proof: When R.sub.k=I.sub.N, .A-inverted.k,
.gamma..sub.k.sup.(FD) in (37) is given by
.gamma. k ( FD ) = Tr ( ( ( j = 1 K 1 1 + .gamma. j ( FD ) + K
.rho. ) I N ) - 1 ) = N j = 1 K 1 1 + .gamma. j ( FD ) + K .rho. ,
.A-inverted. k , ( 41 ) ##EQU00037##
which implies
.gamma. 1 ( FD ) = = .gamma. K ( FD ) = .gamma. ( FD ) = N K 1 +
.gamma. j ( FD ) + K .rho. = 1 .kappa. 1 + .gamma. j ( FD ) +
.kappa. .rho. , ##EQU00038##
and the positive solution of .gamma..sup.(FD) to this equation
becomes the numerator in (39). In a similar way, it can be proved
that .gamma..sub.1.sup.(HB)=. . .
=.gamma..sub.K.sup.(HB)=.gamma..sup.(HB) and .gamma..sup.(HB) is
given by the denominator in (39), using the fact that R.sub.tot is
an identity matrix. If .rho..fwdarw..infin. and .mu.>.kappa.,
then .eta. converges as (40).
[0112] At high SNR region (.rho..fwdarw..infin.) in the
uncorrelated channels, the SLNR loss caused by the hybrid precoding
in (40) is negligible if .mu..fwdarw.1, e.g., M.apprxeq.N.
Furthermore, as .kappa. approaches to .mu., e.g.,
K M .fwdarw. 1 , ##EQU00039##
the SLNR loss becomes disastrous.
[0113] In the next proposition, the SLNR loss decreases as the
channels become more spatially correlated. In this correlated case,
the covariance matrix R.sub.k is likely to be ill-conditioned,
e.g., the eigenvalues are not evenly distributed, and a few
dominant eigenvalues account for most of the sum of all the
eigenvalues. The following proposition shows an extreme case where
there is no SLNR loss from the hybrid precoding in the correlated
channels.
[0114] Proposition 4: For correlated channels, if .SIGMA.k=1.sup.K
R.sub.k is rank-deficient and its rank is lower than or equal to M,
then the SLNR loss metric .eta. is equal to one, e.g., the hybrid
precoding has the same asymptotic SLNR as that of the fully digital
precoding.
[0115] Proof: Let the rank of R.sub.tot be {tilde over
(M)}.ltoreq.M and V.sub.{tilde over (M)} be the eigenvector
associated with its nonzero eigenvalues. Since the rank of
R.sub.tot=.SIGMA..sub.k=1.sup.K R.sub.k is {tilde over
(M)}.ltoreq.N, the rank of each user's covariance matrix R.sub.k
becomes at most {tilde over (M)} and thus can be represented as
R.sub.k=V.sub.{tilde over (M)}Q.sub.kV*.sub.{tilde over (M)} where
Q.sub.k .di-elect cons.C.sup.{tilde over (M)}.times.{tilde over
(M)}. Note that this is not an eigenvalue decomposition, so Q.sub.k
is generally not a diagonal matrix. In the proposed hybrid
precoding technique, the analog precoding without the phase shifter
constraint is given by F.sub.RF=[V.sub.{tilde over (M)}0] which
means that {tilde over (M)} RF chains are used among M ones. From
(17), the deterministic SLNR of user kin the hybrid precoding case
is the unique nonnegative solution of
.gamma. k ( HB ) = Tr ( F RF * R k F RF ( j = 1 K F RF * R j F RF 1
+ .gamma. j ( HB ) + K .rho. I M ) - 1 ) = Tr ( [ Q k 0 0 0 M - M ~
] ( j = 1 K 1 1 + .gamma. j ( HB ) [ Q j 0 0 0 M - M ~ ] + K .rho.
I M ) - 1 ) = Tr ( Q k ( j = 1 K Q j 1 + .gamma. j ( HB ) + K .rho.
I M ~ ) - 1 ) ( 42 ) ##EQU00040##
[0116] Let V.sub.A=[V.sub.{tilde over (M)}V.sub.N-{tilde over (M)}]
be a unitary matrix where V.sub.N-{tilde over (M)} is the null
space of V.sub.{tilde over (M)} such that V*.sub.{tilde over (M)}*
V.sub.N-{tilde over (M)}=0.sub.{tilde over (M)}.times.(N-M) and
V*.sub.N-{tilde over (M)}V.sub.N-{tilde over (M)}=I.sub.N-{tilde
over (M)}. In the fully digital precoding case, the fixed point
equation of the deterministic SLNR of user kin (37) can be
reformulated as
( 43 ) ##EQU00041## .gamma. k ( FD ) = Tr ( V M ~ Q k V M ~ * ( j =
1 K V M ~ Q j V M ~ * 1 + .gamma. j ( FD ) + K .rho. I N ) - 1 ) =
Tr ( V A [ Q k 0 0 0 M - M ~ ] V A * ( j = 1 K 1 1 + .gamma. j ( HB
) V A [ Q j 0 0 0 M - M ~ ] V A * + K .rho. I N ) - 1 ) = Tr ( [ Q
k 0 0 0 M - M ~ ] ( j = 1 K 1 1 + .gamma. j ( HB ) [ Q j 0 0 0 M -
M ~ ] + K .rho. I N ) - 1 ) = Tr ( Q k ( j = 1 K Q j 1 + .gamma. j
( FD ) + K .rho. I M ~ ) - 1 ) ##EQU00041.2##
[0117] Since (42) is identical to (43), and the solution of these
fixed point equations have a unique solution, the proof is
completed.
[0118] Consider a general correlated channel case where the rank of
R.sub.tot is not strictly less than M. Although the (N-M) smallest
eigenvalues are not exactly zeros, it is possible for those
eigenvalues to become much smaller than the other dominant
eigenvalues in the highly correlated channels. It is intuitive that
the smaller those non-dominant eigenvalues are, the smaller the
loss from the hybrid precoding. A question still remains about how
much the exact loss will be according to the portions of the small
eigenvalues. For quantitative analysis, Let .lamda..sub.1, . . . ,
.lamda..sub.N be the nonnegative eigenvalues of R.sub.tot in
descending order and define a metric, .tau., as the ratio of the
sum of M largest eigenvalues to the sum of all eigenvalues,
e.g.,
.tau. = i = 1 M .lamda. i i = 1 N .lamda. i . ##EQU00042##
[0119] This metric .tau. ranging from
.mu. ( = M N ) ##EQU00043##
to 1 can be regarded as the metric that indicates how concentrated
the eigenvalues are. The goal of the quantitative analysis here is
to express the SLNR loss .eta. as a function of the concentration
metric .sup.-c and other system parameters such as
.kappa. ( = K N ) and .mu. ( = M N ) , ##EQU00044##
which can provide a useful insight to the relation between both
metrics.
[0120] The closed form expressions on the SLNR metric .eta. in
Proposition 3 and 4 are the special cases when .tau.=.mu. and
.tau.=1, respectively. The SLNR loss metric .eta., however, does
not have a closed form expression if .mu.<.tau.<1. Instead of
pursing exact expressions, two approximations may be used to get an
insight to the impact of .tau. and other parameters on .eta..
First, all users' deterministic SLNR's are the same as the average
value used previously. Second, all the M largest eigenvalues have
an identical value that is their average values as
.lamda. L _ = 1 M i = 1 M .lamda. i = .tau. M Tr ( R tot ) , ( 44 )
##EQU00045##
and the N-M remaining eigenvalues have the same value as
.lamda. _ S = 1 N - M i = M + 1 N .lamda. i = 1 - .tau. N - M Tr (
R tot ) . ( 45 ) ##EQU00046##
[0121] From (21) and the above assumptions, the deterministic SLNR
of the hybrid precoding and the fully digital precoding are the
nonnegative unique solution of
.gamma. ( HB ) = 1 K m = 1 M 1 1 1 + .gamma. ( HB ) + 1 .rho.
.lamda. m = 1 K M 1 1 + .gamma. ( HB ) + 1 .rho. .lamda. L _ , and
( 46 ) .gamma. ( FD ) = 1 K m = 1 N 1 1 1 + .gamma. ( FD ) + 1
.rho. .lamda. m = 1 K M 1 1 + .gamma. ( FD ) + 1 .rho. .lamda. L _
+ 1 K N - M 1 1 + .gamma. ( FD ) + 1 .rho. .lamda. _ S ( 47 )
##EQU00047##
[0122] respectively. In the following proposition, the approximate
SLNR loss metric .eta. using the above two assumptions is derived
in a closed form.
[0123] Proposition 5: For the spatially correlated channels where
Tr(R.sub.k)=N for all k as the uncorrelated channel case in (40),
the SLNR loss metric .eta. approximates to
.eta. .apprxeq. ( B - A - AB ) + ( B - A - AB ) 2 + 4 AB 2 - 6 ( 1
D + 1 .omega. H + .omega. H G ) - 2 , where .omega. = - 1 2 + 1 2 3
i and ( 48 ) A = .kappa. .rho. .tau. , B = .mu. .rho. .tau. , C = 1
- .mu. ( .rho. ( 1 - .tau. ) ) , D = B + C + 1 - .kappa. .kappa. ,
E = BC ( .rho. + .kappa. .kappa. ) - B - C G = D 2 - 3 E , H = ( G
+ ( ( 2 D 3 - 9 DE - 27 BC ) 2 - 4 G 3 ) 1 2 2 ) 1 3 . ( 49 )
##EQU00048##
[0124] Proof: From (21) and the above assumptions, the
deterministic SLNR of the hybrid precoding is the nonnegative
unique solution of
.gamma. ( HB ) = 1 K M 1 1 + .gamma. ( HB ) + 1 .rho. .lamda. L _ =
.mu. / K 1 1 + .gamma. ( HB ) + .mu. .rho. .tau. , ( 50 )
##EQU00049##
and the solution is given by
.gamma. ( HB ) = ( ( 1 .kappa. - 1 .mu. ) .rho. .tau. - 1 ) + ( ( 1
.kappa. - 1 .mu. ) .rho. .tau. - 1 ) 2 + 4 .rho. .tau. .kappa. 2 ,
= ( ( 1 A - 1 B ) - 1 ) + ( ( 1 A - 1 B ) - 1 ) 2 + 4 A 2 , ( 51 )
##EQU00050##
where
A = .kappa. .rho. .tau. and B = .mu. .rho. .tau. . ##EQU00051##
In the fully digital precoding case, the deterministic SLNR is the
solution of
.gamma. ( FD ) = 1 K M 1 1 + .gamma. ( FD ) + 1 .rho. .lamda. L _ =
1 K N - M 1 1 + .gamma. ( FD ) + 1 .rho. .lamda. _ S , = 1 .kappa.
( M 1 .mu. ( 1 + .gamma. ( FD ) ) + 1 .rho. .tau. = N - M 1 ( 1 -
.mu. ) ( 1 + .gamma. ( FD ) ) + 1 .rho. ( 1 - .tau. ) ) . ( 52 )
##EQU00052##
[0125] Let
C = 1 - .mu. .rho. ( 1 - .tau. ) , ##EQU00053##
then the equation (52) can be simplified as
( .gamma. ( FD ) ) 3 + ( B + C + 1 - .kappa. .kappa. ) ( .gamma. (
FD ) ) 2 + ( BC ( .rho. - .kappa. .kappa. ) - B - C ) .gamma. ( FD
) - BC = 0. Let D = B + c + 1 - .kappa. .kappa. and E = BC ( .rho.
+ .kappa. .kappa. ) - B - C . ( 53 ) ##EQU00054##
[0126] The nonnegative solution of (53) is given by
.gamma. ( FD ) = - 3 ( 1 D + 1 .omega. H + .omega. H G ) - 1 where
.omega. = - 1 2 + 1 2 3 i , G = D 2 - 3 E , and H = ( G + ( ( 2 D 3
- 9 DE - 27 BC ) 2 - 4 G 3 ) 1 2 2 ) 1 3 . ( 54 ) ##EQU00055##
[0127] From (51) and (54), the approximate SLNR loss metric becomes
(48).
[0128] The approximate SLNR loss metric in (48) is a decreasing
function with respect to .tau.. Since the range of .tau. is
.mu..ltoreq..tau..ltoreq.1, the metric has a minimum value of (39)
when .tau.=.mu. (uncorrelated channels), and a maximum value of one
when .tau.=1 (correlated channels with rank(R.sub.tot)=M). The
approximate SLNR loss metric also depends on three other
factors:
.mu. ( = M N ) , .kappa. ( = K N ) , and .rho. ( = P tx .sigma. 2 )
. ##EQU00056##
[0129] Although specific embodiments have been described and
illustrated, the embodiments are not to be limited to the specific
forms or arrangements of parts so described and illustrated. The
described embodiments are to only be limited by the claims.
* * * * *