WK,((($QQXDO&RQVXPHU&RPPXQLFDWLRQV 1HWZRUNLQJ&RQIHUHQFH &&1&

Adaptive-bit Quantized Massive MIMO Systems with

MMSE-based Variational Approximate Message Passing
ற ற ș
Hong-Yunn Chen , Cheng-Fu Chou , and Leana Golubchik

ற Graduate Institute of Networking and Multimedia (GINM), National Taiwan University, Taipei, Taiwan

ș
Department of Computer Science, University of Southern California, Los Angeles, California

Abstract—Millimeter Wave (MmWave) massive multiple-input from a high-efficiency energy massive MIMO architecture
multiple-output (MIMO) has become an advantageous technology since the entire energy-consuming can be governed by the
for gigabit-per-second data transmission in 5G wireless flexible resolutions of ADC. Furthermore, the amount of data
communication. To achieve low-cost and energy-efficient generated by high-resolution ADC could be large and may have
hardware components, one-bit quantized massive MIMO systems to be transmitted over the link between the radio frequency
have been proposed for the receiver hardware architecture. The
components and Millimeter-Wave antennas. Hence, the focus
main focus of this work is leveraging the advantages of a state of
the art one-bit quantized massive MIMO system for design of an of this work is on how to design an adaptive ADC for a hybrid
adaptive-bit massive MIMO system. Hence, in this work, by beamforming architecture.
leveraging the benefits of variational approximate message Our idea is to leverage the advantages of a state of the art one-
passing (VAMP), a novel MMSE-based VAMP algorithm is bit quantized massive MIMO architecture for design of an
proposed for the adaptive-bit quantized massive MIMO system. adaptive-bit massive MIMO architecture. First, we give a brief
That is, two novel modules, i.e., an adaptive ADC bit allocation literature overview of the one-bit quantized massive MIMO
method and an MMSE-based VAMP, are proposed for mmWave architecture. Note that the signal detection in a one-bit
communications of the hybrid MIMO receiver architecture. With quantized massive MIMO architecture is different from that of
the MMSE-based VAMP, our adaptive ADC bit allocation method
a traditional MIMO architecture. The reason is that the high
is able to decrease the quantization of signals distortion by
improving the flexible resolutions of ADC. Through simulations, distortion of one-bit estimation makes it difficult to carry out
compared with existing works, our proposed adaptive ADC bit precise channel state information (CSI) when the pilot length is
allocation algorithm, together with MMSE-based VAMP, is able not prolonged enough [7]. Another method of joint channel
to achieve higher capacity, sum rate, and energy efficiency in most estimation and signal detection is adopted in [8], which utilizes
communication architectures. both the pilot patterns and data patterns to enhance the channel
Index Terms—Variational approximate message passing (VAMP), estimation, i.e., the authors design a training-based detection
Millimeter Wave (MmWave), variational Bayesian inference (VBI), approach while using a much shorter pilot sequence. In recent
massive MIMO, bilinear generalized approximated message passing times, approximate message passing (AMP) has attracted much
(BiG-AMP).
consideration because, unlike other alternatives of MP, AMP’s
I. INTRODUCTION messages are connected with nodes rather than edges on the
factor graph [9], which decreases the number of messages
Millimeter Wave (MmWave) massive multiple-input multiple-
significantly. The bilinear generalized approximate message
output (MIMO) has become an advantageous technology for
passing (BiG-AMP) [7][11] algorithm further develops AMP
gigabit-per-second data transmission in 5G wireless
from a linear model to a quantized bilinear model using Taylor
communication because of its tremendousġ potential in
expansion and Gaussian approximation to deal with the
improving spectrum utilization, overall throughput, and energy
nonlinearity. Another work is the variational Bayesian
efficiency [1]. Such large numbers of antenna elements of the
inference (VBI) [10], which calculates the posterior probability
base station result in enormous hardware expenditure and
of the goal variable in complex models by marginalizing
energy consumption. To achieve low-cost and energy-efficient
adjacent variables using variational approximation. By taking
hardware components, low-resolution quantization [2] and one-
advantages of both, Big-AMP and VBI, an innovative vertex-
bit quantization [3] have been proposed for the receiverġ
message based inference algorithm, termed variational
hardware architecture, e.g., the one-bit quantized massive
approximate message passing (VAMP) [12], is proposed for the
MIMO system [4]. In [5] and [6], authors investigate the use
one-bit quantized massive MIMO architecture; and, it could
of one-bit ADC and infinite-resolution-bit ADC in a massive
perform channel estimation, signals detection. In this paper, byġ
MIMO system. The use of infinite-resolution-bit ADC can
leveraging the benefits of VAMP, a novel inference algorithm,
achieve a higher rate as compared to one-bit ADC. We,
termed MMSE-based variational approximate message passing
however, note that such a high-resolution ADC approach is far

k,(((
 WK,((($QQXDO&RQVXPHU&RPPXQLFDWLRQV 1HWZRUNLQJ&RQIHUHQFH &&1&

(MMSE-based VAMP), is proposed for the adaptive-bit The signal ࢞ after Zero-Forcing precoding is given by
quantized massive MIMO architecture. That is, two novel ࢞ ൌ ሼ࡭ሺࣂǡ ࣘሻ ٖ ࡮ሺࣂǡ ࣘሻሽඥࢍ૚ ࢍ૛ ࢂ࢙, ሺ૝ሻ
modules, an adaptive ADC bit allocation method and MMSE- where ࡭ሺࣂǡ ࣘሻ ൌ ሾࢇሺࣂ૚ ǡ ࣘ૚ ሻǡ ࢇሺࣂ૛ ǡ ࣘ૛ ሻǡ ǥ ǡ ࢇሺࣂࡳ ǡ ࣘࡳ ሻሿ , ࡮ሺࣂǡ ࣘሻ ൌ
based VAMP, are proposed for mmWave communications of ሾ࢈ሺࣂ૚ ǡ ࣘ૚ ሻǡ ࢈ሺࣂ૛ ǡ ࣘ૛ ሻǡ ǥ ǡ ࢈ሺࣂࡳ ǡ ࣘࡳ ሻሿ are the transmit and receive
the hybrid MIMO receiver architecture. With the MMSE-based matrix, respectively. ٖ represents the Khatri-Rao product.ࡿ ൌ
VAMP, our adaptive ADC bit allocation method is able to ሾ࢙૚ ǡ ࢙૛ ǡ ǥ ǡ ࢙ࡳ ሿ ‫ א‬ԧࡳൈ૚ is the symbol vector intended for ࡳ users
decrease the quantization of signals distortion by improving the with the normalized power ࡱሾȁ࢙࢏ ȁ૛ ሿ ൌ ૚ , and ࢂ ൌ
flexible resolutions of ADC. In other words, an optimal number ሾ࢜૚ ǡ ࢜૛ ǡ ǥ ǡ ࢜ࡳ ሿ ‫ א‬ԧࡲൈࡳ is the precoding matrix consisting of ۵
of ADC bits could be allocated dynamically according to the different ࡲ-dimensional unit-norm precoding vectors ࢜࢏ ‫ א‬ԧࡲൈ૚ .
SNR of the antennas. With this, our proposed algorithms are We denote ࢁ ൌ ࡴ ෡ ሺࡴ
෡ ࡴࡴ
෡ ሻି૚ ; then the precoding vectors ࢜࢏ is
ࢁሺǣǡ࢏ሻ
able to outperform the state-of-the-art one-bit quantized ADCs, obtained as the normalized ࢏࢚ࢎ column of ࢁ , ࢜࢏ ൌ ԡࢁሺǣǡ࢏ሻԡ, the
as in [12], in the low-resolution regime of the hybrid
received signal ࢟ࢍ at the ࢍ࢚ࢎ user is described as
architecture. In addition, our proposed modules are able to
࢟ࢍ ൌ ࢎࡴࢍ ࢞ ൅ ࢔ࢍ
perform channel estimation, data detection, and noise
ൌ ඥࢍ૚ ࢍ૛ ࢎࢍ ࢜ࢍ ࢙ࢍ ൅ ඥࢍ૚ ࢍ૛ σࡳ࢏ୀ૚ǡ࢏ஷࢍ ࢎࡴ
ࡴ
ࢍ ࢜࢏ ࢙࢏ ൅ ࢔ࢍ , ሺ૞ሻ
estimation simultaneously with adaptive-bit allocation under
ADC power constraints at the receiver. Our simulation results Given the BS transmit power constraint ࡮ࡿࡳ and the UE
show that, compared with existing systems, the adaptive ADC transmit power constraint ࢁࡱࡳ , the factorsࢍ૚ andࢍ૛ are ࢍ૚ ൌ
࡮ࡿࡳ ࢁࡱࡳ
bit allocation algorithm together with MMSE-based VAMP is , ࢍ૛ ൌ , where ࢔ࢍ is the complex Gaussian noise at
ࡳ ࡹൈ࡮ࡿࡳ ାࡳ
able to achieve higher capacity, sum rate, and energy efficiency the ࢍ࢚ࢎ user with unit variance and zero mean.
in most communication scenarios.
III. ADAPTIVE ADC BIT ALLOCATION ALGORITHMS
II. SYSTEM MODEL We consider the AQNM [13] as the quantization model to
acquire a linearized quantization and each of the ADC pairs as
having b quantization bits. The rationale is that AQNM is
precise enough in low SNR [13], so, we have the quantized
signal vector after quantizing y as follows
ġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġġ࢟ࢗ ൌ छሺ‫ܡ‬ሻ ൌ ۲࢟ ൅ ࢔ࢗ  ሺ૟ሻ
where छሺǤ ሻ is the function of bit quantization utilized to the real and
imaginary parts individually, ۲ ൌ ሺ૚ െ ࣅ࢈ ሻࡵ when the number of
bits is the equivalent in ADCs. For a Gaussian input distribution,
the values of ࣅ for ࣅ ൐ ૞ can be approximated by ࣅ ൌ ࣊ξ૜ ૛
૛ି૛࢈ .
The quantization output at the ࢍ-th receiver is
࢟ࢍ ൌ छሺඥࢍ૚ ࢍ૛ ࢎࡴ
ࢍ ࢜ࢍ ࢙ࢍ ൅ ࢔ࢍ ሻ
ࡳ

ൌ ሺ૚ െ ࣅ࢈ ሻඥࢍ૚ ࢍ૛ ࢎࡴ ࡴ
ࢍ ࢜ࢍ ࢙ࢍ ൅ ሺ૚ െ ࣅ࢈ ሻඥࢍ૚ ࢍ૛ ෍ ࢎࢍ ࢜࢏ ࢙࢏
࢏ୀ૚ǡ࢏ஷࢍ
൅ሺ૚ െ ࣅ࢈ ሻ࢔ࢍ ൅ ࢔ࢗ ሺૠሻ
Therefore, the Signal-to-interference Quantization Ratio (SQR) at
Fig. 1. A hybrid beamforming architecture with adaptive ADC bit allocation.
the ࢍ-th receiver is
૛
ሺ૚ିࣅ࢈ ሻඥࢍ૚ ࢍ૛หࢎࡴ
ࢍ ࢜ࢍ ห
A. Channel Model ࡾൌ ૛ ࡳ ૛ ሺૡሻ
ࣅ࢈ ඥࢍ૚ ࢍ૛ หࢎࡴ ࡴ
ࢍ ࢜ࢍ ห ାඥࢍ૚ ࢍ૛ σ࢏స૚ǡ࢏ಯࢍหࢎࢍ ࢜ࢍ ห ା࣌
૛

We consider a millimeter wave massive MIMO system with F and we use the equality [14]
antennas at the BS, G single antenna UE. The downlink channel ॱ ቂหࢎࡴ
૛ ૛
෡ࡴ ૛ ࡲ
ቁ ሺૢሻ, ૛࡮ ࢼ ቀ૛࡮ ǡ
ࡲ
ࢍ ࢜ࢍ ห ቃ ൌ ॱ ቂฮࢎࢍ ฮ ቃ ॱ ቂหࢎࢍ ࢜ࢍ ห ቃ ൌ
of user ࢍ is ࢎࡴ ࡲ
ࢍ , where ࢎࢍ ‫ א‬ԧ is a column vector that follows Next, we define οࡾ as the difference between sum rate of perfect
ࡲି૚ ࡲି૚

ࢀ
distributionऍघ൫૙ǡ ࡾࢍ ൯. The received signal ࢟ ൌ ൫࢟૚ ǡ ࢟૛ ǡ ǥ ǡ ࢟ࢍ ൯ at channel state information and imperfect channel feedback. We
the user side is now calculate the effects of imperfect channel state information by
‫ ܡ‬ൌ ࡴࡴ ࢃ‫ ܠ‬൅ ࢔. ሺ૚ሻ characterizing the οࡾ difference as follows
where ‫ א ܠ‬ԧࡳ is the vector of transmission symbol that satisfies οࡾ ൌ ࡾஶ െ ࡾ࡮
૛
ॱሼ‫ ࡴ ܠܠ‬ሽ ൌ ࡵࡳ , and ࢔̱ऍघሺ૙ǡ ࡵࡳ ሻ is the Gaussian noise. ࢃ is the ሺ૚ െ ࣅ࢈ ሻඥࢍ૚ ࢍ૛ หࢎࡴ ࢍ ࢜ࢍ ห
ൌ ॱሾ࢒࢕ࢍ૛ ሺ૚ ൅ ሻሿ
precoder. The transmit antennas are supposed to be located at ࣅ࢈ ඥࢍ૚ ࢍ૛ หࢎࡴ
૛
૛
ࢍ ࢜ࢍ ห ൅ ࣌
the position ࢖ࢌ ൌ ሾ࢞ࢌ ǡ ࢟ࢌ ሿࢀ ǡ ࢌ ൌ ૚ǡ ૛Ǥ ǥ ǡ ࡲ . At that moment, the ሺ૚ െ ࣅ࢈ ሻඥࢍ૚ ࢍ૛ หࢎࡴ
૛
ࢍ ࢜ࢍ ห
vector of the transmit array is: െॱሾ࢒࢕ࢍ૛ሺ૚ ൅ ૛ ࡳ ૛ ሻሿሺ૚૙ሻ
ࢀ ࢀ
ࢇሺࣂǡ ࣘሻ ൌ ሾࢋି࢐૛࢛࣊ ሺࣂǡࣘሻࢍ૚ ǡ ǥ ǡ ࢋି࢐૛࢛࣊ ሺࣂǡࣘሻࢍࡲ ሿࢀ , ሺ૛ሻ
 ࣅ࢈ ඥࢍ૚ ࢍ૛ หࢎࡴ ࡴ
ࢍ ࢜ࢍ ห ൅ ඥࢍ૚ ࢍ૛ σ࢏ୀ૚ǡ࢏ஷࢍหࢎࢍ ࢜ࢍ ห ൅ ࣌
૛

where ࢛ሺࣂǡ ࣘሻ ൌ ሾ‫ ࣘ ܛܗ܋ ࣂ ܖܑܛ‬ǡ ‫ࣘ ܖܑܛ ࣂ ܖܑܛ‬ሿࢀ characterizes the ሺ૚ െ ࣅ࢈ ሻඥࢍ૚ ࢍ૛ ሺࡲ െ ࡳ ൅ ૚ሻ

ൌ ࢒࢕ࢍ૛ሺ૚ ൅ ሻ െ
propagation vector, and ࣂǡ ࣘ are the elevation, and azimuth ࣅ࢈ ඥࢍ૚ ࢍ૛ ሺࡲ െ ࡳ ൅ ૚ሻ ൅ ࣌૛
individually. Equally, the vector of the receive array could be ࡲ
࢒࢕ࢍ૛ሺ૚ ൅ ሾ൭૚ െ ࣅ࢈ ሺ૚ െ ࣁ࢈ ሻඥࢍ૚ ࢍ૛ ሺࡲ െ ࡳ ൅ ૚ሻ ቆ૚ െ ૛࡮ ࢼ ൬૛࡮ ǡ ൰ቇ൱
characterized as ࡲെ૚
ࢀ ሺࣂǡࣘሻࢍ ࢀ ሺࣂǡࣘሻࢍ 
࢈ሺࣂǡ ࣘሻ ൌ ሾࢋି࢐૛࢛࣊ ૚ ǡ ǥ ǡ ࢋି࢐૛࢛࣊ ࡲ ሿࢀ , ሺ૜ሻ
 WK,((($QQXDO&RQVXPHU&RPPXQLFDWLRQV 1HWZRUNLQJ&RQIHUHQFH &&1&

ࡲ as
ൊ ࣅ࢈ ඥࢍ૚ ࢍ૛ ሺࡲ െ ࡳ ൅ ૚ሻ ቆ૚ െ ૛࡮ ࢼ ൬૛࡮ ǡ ൰ቇ ൅
ࡲെ૚ ‫ ܖܔ‬બ࢒ ൌ ‫ࡼ ܖܔ ׬‬ሺ‫܇‬ǡ શሻ ς࢒Ʈஷ࢒ሺબ࢒Ʈࢊ࣎࢒Ʈ ሻ ሺ૛૝ሻ
ࡲ ࡲ
ඥࢍ૚ ࢍ૛ ሺࡳ െ ૚ሻ ࡲି૚ ૛࡮ ࢼ ቀ૛࡮ ǡ ࡲି૚ቁ ൅ ࣌૛ ሿ ሺ૚૚ሻ This is the concept of dynamic variational approximation. In
the sum rate of user ࢍ with adaptive ADC bit allocation b is our combined channel estimation and detection problem,
શ ൌ ሼ۶ǡ ‫܆‬ǡ ઼ሽ ሺ૛૞ሻ
given as
࡮ and બ࢒ is represented as
૚ ૛ξ૜ ૛ሺ࢈ି ሻ ሼબ࢒ ሽ ൌ ሼ൛ࡼ൫ࢎࢌࢍ ȁ‫܇‬൯ൟǡ ൛ࡼ൫࢞ࢍ࢐ ȁ‫܇‬൯ൟǡ ൛ࡼ൫઼ࢍȁ‫܇‬൯ሽൟ ሺ૛૟ሻ
ൌ ૛ ૛ሺࡳష૚ሻ ሺ૚૛ሻ
ࣅ࢈ ૛ࡲష૚ ૜࣊ ‫۾‬ሺ‫܇‬ǡ ۶ǡ ‫܆‬ǡ ઼ሻ ൌ ςࢌǡ࢐ ࡼ ൫࢟ࢌ࢐ หࢠࢌ࢐ ሻ ςࢌǡ࢐ ࡼࡴ ൫ࢎࢌࢍ ห઼ሻ ൈ
The algorithm considers both the method of power constraint ςࢍǡ࢐ஸࡶ࢖࢏࢒࢕࢚ ࡼࢄ࢖ ሺ࢞ࢍ࢐ ሻ ςࢍǡ࢐வࡶ࢖࢏࢒࢕࢚ ࡼࢄࢊ ሺ࢞ࢍ࢐ ሻ ςࢍ ࡼ઼ ሺ઼ࢍ ሻ ሺ૛ૠሻ
and adaptive ADC bit allocation; thus, it can ensure that the The global bound of the logistic functionોሺǤ ሻ can be acquired
power consumption of the ADC architecture meets the when െ࢒࢕ࢍሺࢋି࢙Τ૛ ൅ ࢋ࢙Τ૛ ሻ is the Taylor series expansion at any
requirements.
࢙૛ ൌ ࢿ૛ by solving the following equations:
૚
ોሺઽሻ‫ܘܠ܍‬ሺ ൈ ሺ࢙ െ ࢿሻ െ  ሺ࢙૛ െ ࢿ૛ ሻ ൈ ૃሺઽሻሻ ൑ ોሺ‫ܛ‬ሻ ሺ૛ૡሻ
IV. MMSE-BASED VARIATIONAL APPROXIMATION ૛

FOR CHANNEL ESTIMATION AND DETECTION whereૃሺઽሻ is the invariable value calculated as
૚ ૚
ૃሺઽሻ ൌ ሺെ ൅ ࣌ሺࢿሻሻ ሺ૛ૢሻ
According to [12], every received signal is transformed into ૛ࢿ ૛

analog signal ‫ ܇‬and then quantized. Every complex valued This lower bound is an exponential approximation ofોሺ‫ܛ‬ሻ,
quantizer comprises two real-valued ADCs, which perform which is exact when ࢿ૛ ൌ ࢙૛ . Substituting (28) into (14), we
individually on the real part and on the imaginary part. obtain
࢟ࡾ ૛
‫ ܇‬ൌ ‫ܖ܏ܛ‬ሺ‫ܡ‬ሻ ൌ ‫ܖ܏ܛ‬ሺ࢟ࡾ ሻ ൅ ܑǤ ‫ܖ܏ܛ‬ሺ࢟ࡵ ሻ ‫ࡼ ܖܔ‬൫࢟ࡾࢌ࢐ ȁࢠࡾࢌ࢐ ൯ ൌ ࢌ࢐
ࢠࡾࢌ࢐ െ ࣅ൫ࢿࡾࢌ࢐ ൯൫ࢠࡾࢌ࢐ ൯  ሺ૜૙ሻ
૛
So, the conditional probability function is calculated as ࢟ࡵࢌ࢐ ૛
ࡶ ‫ࡼ ܖܔ‬൫࢟ࡵࢌ࢐ ȁࢠࡵࢌ࢐ ൯ ൌ ࢠࡵࢌ࢐ െ ࣅ൫ࢿࡵࢌ࢐ ൯൫ࢠࡵࢌ࢐ ൯  ሺ૜૚ሻ
ࡼሺࢅȁࡴǡ ࢄሻ ൌ ςࡲࢌୀ૚ ς࢐ୀ૚ ࡼሺ࢟ࢌ࢐ ȁࢠࢌ࢐ ሻ ሺ૚૜ሻ ૛

where each ࡼሺ࢟ࢌ࢐ȁࢠࢌ࢐ሻ is factored as whereሺࢿࡾࢌ࢐ ሻ૛ ൌ ࡱሾሺࢠࡾࢌ࢐ ሻ૛ ሿand ሺࢿࡵࢌ࢐ ሻ૛ ൌ ࡱሾሺࢠࡵࢌ࢐ ሻ૛ ሿ are invariable
ࡼሺ࢟ࢌ࢐ ȁࢠࢌ࢐ ሻ ൌ ࡼ൫࢟ࡾࢌ࢐ ȁࢠࡾࢌ࢐ ൯ࡼ൫࢟ࡵࢌ࢐ ȁࢠࡵࢌ࢐ ൯ ሺ૚૝ሻ values. Therefore, we can adopt (13), (30), and (31) to acquire
The prior probability is written as the exponential approximation of ‫۾‬ሺ‫܇‬ȁ۶ǡ ‫܆‬ሻ , therefore ࡼሺ‫܆‬ȁ‫܇‬ሻ
࢖࢏࢒࢕࢚
ࡼࢄ ሺࢄሻ ൌ ςࡳࢍୀ૚ ς࢐ୀ૚
ࡶ ࡶ
ࡼࢄ࢖ ሺ࢞ࢍ࢐ ሻ ςࡳࢍୀ૚ ς࢐வࡶ࢖࢏࢒࢕࢚ ࡼࢄࢊ ሺ࢞ࢍ࢐ ሻ ሺ૚૞ሻ and ࡼࢄ ሺࢄሻ are expressed by using complex Gaussian
distributions. In case of ࡼࢄ ሺࢄሻ , we can express ࡼࢄࢊ ሺ࢞ࢍ࢐ሻ and
Entries in the channel matrix ۶ are assumed to be complex
ࡼࢄ࢖ ሺ࢞ࢍ࢐ ሻ as
Gaussian variables, calculated as
ࡼࡴ ሺ۶ȁ઼ࢍ ሻ ൌ ςࡲࢌୀ૚ ςࡳࢍୀ૚ ࡼࡴ ൫ࢎࢌࢍ ห઼ࢍ ሻ ሺ૚૟ሻ ࡼࢄࢊ ൫࢞ࢍ࢐ ൯ ൌ ऍघ൫࢞ࢍ࢐ Ǣ ૙ǡ ૚൯ ሺ૜૛ሻ
ࡼࡴ ൫ࢎࢌࢍ ห઼ࢍሻ ൌ ऍघ൫ࢎࢌࢍ Ǣ ૙ǡ ઼ି૚ ࡼࢄ࢖ ൫࢞ࢍ࢐ ൯ ൌ ऍघ൫࢞ࢍ࢐ Ǣ࢞૙ࢍ࢐ ǡ ૙൯ ሺ૜૜ሻ
ࢍ ൯ ሺ૚ૠሻ
The combined effect of transmit power, channel fading, and As forࡼሺ‫܆‬ȁ‫܇‬ሻ, it should be noted that ࡼ൫࢞ࢍ࢐ȁ‫܇‬൯ is calculated from
adaptive ADC quantization resolution is characterized by the (24), which is not the entire goal posterior. The signals detector
parameter ઼, whose is ruled by the Gamma distribution swaps the log-likelihood ratio (LLR) information with a
ࡼ઼ ൫઼ࢍ ൯ ൌ ડሺ઼ࢍ Ǣ ࢇ૙ ǡ ࢈૙ ሻ ሺ૚ૡሻ transceiver. For variable ࢞ࢍ࢐, the ࢒ି࢚ࢎ bit ̓s LLR information is
Assume the posterior is calculated as calculated as
σ ෥ ࢍ࢐ǡ࣓
ऍघ൫࢞ࢍ࢐ Ǣ࣐ ෥ ࢍ࢐൯
෫ࢍ࢐ǡ࢒ ൌ ‫ ܖܔ‬ሼ࢞ࢍ࢐‫א࢐ࢍ࢞׷‬ऋǡࢉࢍ࢐࢒స૚ሽ
ࡸࡸࡾ ሺ૜૝ሻ
‫۾‬ሺ۶ǡ ‫܆‬ǡ ઼ȁ‫܇‬ሻ ൌ ෑ ‫۾‬൫઼ࢍ ȁ‫܇‬൯ ෑ ‫۾‬൫ࢎࢌࢍ ȁ‫܇‬൯ ෑ ‫۾‬൫࢞ࢍǤ࢐ ȁ‫܇‬൯ σ ෥ ࢍ࢐ǡ࣓
ሼ࢞ࢍ࢐ ‫א ࢐ࢍ࢞׷‬ऋǡࢉࢍ࢐࢒స૙ ሽ ऍघ൫࢞ࢍ࢐ Ǣ࣐ ෥ ࢍ࢐൯

ࢍ ࢌǡࢍ ࢍǡ࢐ Therefore, we formulate the output message adopting ෫ࢍ࢐ǡ࢒.

ࡸࡸࡾ
Using the Bayesian method, we derive the minimum mean where ࡼ࢘࢕࢐ [.] represent the projection operation, and the
square error (MMSE) estimator; the Bayesian mean square distributions of the projection is represented as
error is given in ሺ૚ૢሻ̱ሺ૛૚ሻ. ෥ ࢍ࢐ ǡ ࣓
ऍघ൫࢞ࢍ࢐ Ǣ࣐ࢍ࢐ ǡ ࣓ࢍ࢐ ൯ ൌ ࡼ࢘࢕࢐ሾऍघ൫࢞ࢍ࢐ Ǣ࣐ ෥ ࢍ࢐ ൯] ሺ૜૞ሻ
෡ࢌࢍ ሻ૛ ‫۾‬൫‫܇‬ȁࢎࢌࢍ ൯ࢊࢅࢊࢎࢌࢍ
ࡹࡿࡱࢎࢌࢍ ൌ  ‫׭‬ሺࢎࢌࢍ െ ࢎ ሺ૚ૢሻ For Binary Phase Shift Keying (BPSK), the new message can
෡ࢌࢍ ൌ ॱ൫ࢎࢌࢍ ȁ‫܇‬൯ ൌ ‫ ࢍࢌࢎ ׬‬Ǥ ‫۾‬൫ࢎࢌࢍ ȁ‫܇‬൯
ࢎ ሺ૛૙ሻ be formulated by solving the following equations:
‫۾‬൫‫܇‬ȁࢎࢌࢍ ൯‫۾‬൫ࢎࢌࢍ ൯ ෫ࢍ࢐ǡ࢒ ൌ ࡸࡸࡾࢍ࢐ǡ࢒ ǡ
‫۾‬൫ࢎࢌࢍ ȁ‫܇‬൯ ൌ ሺ૛૚ሻ ࡸࡸࡾ
‫۾ ׬‬൫‫܇‬ȁࢎࢌࢍ൯‫۾‬൫ࢎࢌࢍ ൯ࢊ൫ࢎࢌࢍ ൯ ૚ ࡾȀࡵ ࡾȀࡵ
Consider the problem of approximating a set of hidden   ൌ  ሺ࣐ࢍ࢐ ሻ૛ ൅ ࣓ࢍ࢐ ሺ૜૟ሻ
૛

variablesશ ൌ ሼ࣎૚ ǡ ࣎૛ ǡ ǥ ሽ from signal ‫܇‬. It can be approximated by where ࡸࡸࡾࢍ࢐ǡ࢒ is obtained by substituting the output
the variational approximation બሺશሻ , which reduces the message घሺ࢞ࡾȀࡵ ࡾȀࡵ ࡾȀࡵ
ࢍ࢐ Ǣ ࣐ࢍ࢐ ǡ ࣓ࢍ࢐ ሻ into (34). Then, the entire LLR
Kullback-Leibler divergence ࡷࡸሺબሺશሻȁȁࡼሺશȁ‫܇‬ሻሻ when the information of this symbol could be maintained. For
posterior ࡼሺશȁ‫܇‬ሻ is not obtainable. Because ࡼሺ‫܇‬ሻ is the Quadrature amplitude modulation (QAM), the output message
invariable value, the minimum value of Kullback-Leibler cannot comprise the entire LLR information of this symbol
divergence can be formulated as the maximum value of ࡸሺબሻ, because the three variables ሺ࣐ࡾࢍ࢐ ǡ ࣐ࡵࢍ࢐ ǡ ࣓ࢍ࢐ ሻ cannot satisfy all LLR
which is characterized as equations in (36). Therefore, ࣐ࢍ࢐ , ࣓ࢍ࢐ are calculated
ࡼሺ‫܇‬ǡશሻ
ࡸሺબሻ ൌ ‫ ׬‬બ ሺશሻ‫ܖܔ‬
બሺશሻ
ࢊદ ሺ૛૛ሻ alternatively based on expectation propagation. Then, ࡼ࢘࢕࢐(.)
Under the independence assumption for the hidden variables, can be described using ࡲ૚ ൫࣐෥ ࢍ࢐ ǡ ࣓
෥ ࢍ࢐ ൯ and ࡲ૛ ൫࣐
෥ ࢍ࢐ ǡ ࣓
෥ ࢍ࢐ ൯ as
બሺશሻ ൌ ς࢒ બ࢒ ሺ࣎࢒ ሻ ൌ ς࢒ બ࢒ ሺ૛૜ሻ ෥ ࢍ࢐ ൯ ൌ  σ࢞ࢍ࢐‫א‬ऋ ࢞ࢍ࢐ ࡼࡸࡸࡾ
෥ ࢍ࢐ ǡ ࣓
࣐ࢍ࢐ ൌ  ࡲ૚ ൫࣐ ෫ ሺ࢞ࢍ࢐ ሻ ሺ૜ૠሻ
૛ ૛
we can characterize ࡸሺબሻ with regards to બ࢒ and obtain the ࣓ࢍ࢐ ൌ  ࡲ૛ ൫࣐ ෥ ࢍ࢐ ൯ ൌ  σ࢞ࢍ࢐‫א‬ऋห࢞ࢍ࢐ ห ࡼࡸࡸࡾ
෥ ࢍ࢐ ǡ ࣓ ෫ ൫࢞ࢍ࢐ ൯ െ ห࣐ࢍ࢐ ห ሺ૜ૡሻ
maximum value, and then acquire બሺદሻ with each બ࢒ calculated where ࡼࡸࡸࡾ
෫ ሺ࢞ࢍ࢐ ሻ is the discrete probability distribution of ࢞ࢍ࢐
 WK,((($QQXDO&RQVXPHU&RPPXQLFDWLRQV 1HWZRUNLQJ&RQIHUHQFH &&1&

෫ࢍ࢐ǡ࢒
given by ࡸࡸࡾ Vertex messages such as ષ࢚ࢌࢍ ൫ࢎࢌࢍ ൯ ൌ ऍघ൫ࢎࢌࢍ Ǣ ࢚ࣖࢌࢍ ǡ ࢚ࣚࢌࢍ ൯ are
defined, with
‫ܖܔ‬ષ࢚ା૚ ࢚
ࢌࢍ՜ࢌ࢐ ൫ࢎࢌࢍ ൯ ൌ ඵ ‫ ࡼ ܖܔ‬൫࢟ࢌ࢐ หࢠࢌ࢐ ሻ ෑ ષࢌࢍƮ՜ࢌ࢐ ൫ࢎࢌࢍƮ ൯ ࢊࢎࢌࢍƮ ൈ ࢚ ૛
࢚ࣚା૚ ࢚ ࢚ ࢚
ࢌࢍ ൌ ሺ઼ࢍ ൅  σ࢐Ʈ ࢽࢌ࢐Ʈ՜ሺࢌࢍǡࢍ࢐Ʈሻ ሺห࣐ࢍ࢐Ʈ՜ࢌ࢐Ʈ ห ൅ ࣓ࢍ࢐Ʈ՜ࢌ࢐Ʈ ሻሻ
ି૚
ሺ૞૚ሻ
ࢍƮஷࢍ ‫࢚ כ‬
ςࢍƮƮ ષ࢚ࢍƮƮ࢐՜ࢌ࢐ ൫࢞ࢍƮƮ࢐ ൯ ࢊ࢞ࢍƮƮ࢐  ሺ૜ૢሻ ࢚ࣖା૚
ࢌࢍ ൌ ࢚ࣚା૚ ࢚
ࢌࢍ ሺσ࢐Ʈ ൫࣐ࢍ࢐Ʈ՜ࢌ࢐Ʈ ൯ ࣀࢌ࢐Ʈ՜ሺࢌࢍǡࢍ࢐Ʈሻ ሻ ሺ૞૛ሻ

‫ܖܔ‬ષ࢚ା૚ ࢚ V. RESULTS
ࢌ࢐՜ࢍ࢐ ൫࢞ࢍ࢐ ൯ ൌ ඵ ‫ ࡼ ܖܔ‬൫࢟ࢌ࢐ หࢠࢌ࢐ ሻ ෑ ષࢌࢍƮ՜ࢌ࢐ ൫ࢎࢌࢍƮ ൯ ࢊࢎࢌࢍƮ ൈ
ࢍƮ In this section, we conduct a performance study to compare our
ςࢍƮƮஷࢍ ષ࢚ࢍƮƮ࢐՜ࢌ࢐ ൫࢞ࢍƮƮ࢐ ൯ ࢊ࢞ࢍƮƮ࢐  ሺ૝૙ሻ proposed adaptive bit allocation system with other one-bit
෪ ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ ൫ࢎࢌࢍ ࢞ࢍ࢐ ൯ ൌ
‫ܖܔ‬ષ quantized systems.
‫ࡼ ܖܔ ׬‬൫࢟ࢌ࢐ หࢠࢌ࢐ ሻ ςࢍƮஷࢍ ષ࢚ࢌࢍƮ՜ࢌ࢐ ൫ࢎࢌࢍƮ ൯ ࢊࢎࢌࢍƮ ൈ Fig. 3 depicts the results of theġ sum rate of the adaptive bit
ષ࢚ࢍƮ࢐՜ࢌ࢐ ൫࢞ࢍƮ࢐ ൯ࢊ࢞ࢍƮ࢐  ሺ૝૚ሻ allocation and one-bit quantized systems under different
transmit powers with 8 usersġ and 256 antennas. In Fig. 3, the
adaptive bit allocation scheme is able to achieve higher sum rate
than the one-bit quantized systems. In particular, the adaptive
bit allocation outperforms the one-bit quantized systems in the
low SNR even under Gaussian noise. This demonstrates that the
adaptive bit allocation is reliable to Gaussian noise. Fig. 4
depicts a comparison of the sum rate of the adaptive bit
allocation and one-bit quantized systems over a different
number of BS antennas with 8 users and 20 dBm transmit
power. It is not surprising that our adaptive bit allocation
system still obtains the best sum under different number of BS
antennas.
Next, we compare the performance of four inference detectors
Fig. 2. Factor graph representation of the adaptive bit quantized including our proposed MMSE-based VAMP, VAMP [12],
BiG-AMP [15], and VBI [16]. The BiG-AMP algorithm is
massive MIMO systems, with F = 3; G = 2, and J = 2.
realized based on the GAMPmatlab [17]. The linear detector in
Fig. 2. presents the factor graph of the joint distribution in (27). [18] is used as the benchmark detector. Since the linear detector
Messages about the random variables are propagated among the applies least-squares (LS) channel estimation and maximal
edges of the factor graph. Let ષ࢚ା૚ ࢚ା૚ ratio combining (MRC) processing for channel estimation and
ࢌ࢐՜ࢌࢍ ൫ࢎࢌࢍ ൯ and ષࢌ࢐՜ࢍ࢐ ൫࢞ࢍ࢐ ൯ denote
detection, it is denoted as LS+MRC in this work.
the messages from the channel transition nodeࡼሺ࢟ࢌ࢐ ȁࢠࢌ࢐ ሻ to the
To validate the correctness of our program and perform a fair
cloning node of variables ࢎࢌࢍ and ࢞ࢍ࢐ , individually. Based on
comparison of the proposed algorithm with the benchmark
variational approximation (24), the two messages are calculated
algorithms, we choose the same simulation parameters as used
as (39) and (40). To eradicate the identical computation, their
in previous works. That is, in this experiment, we estimate the
common portion is derived from (41), which is defined as
adaptive bit allocation MIMO system with ۴ = 200 antennas, ۵
෩ ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ ൫ࢎࢌࢍ ࢞ࢍ࢐ ൯ to denote the message from the channel
ષ
= 50 users, block length of ۸ = 500 and pilot length of ۸‫ = ܜܗܔܑܘ‬50.
transition node ࡼሺ࢟ࢌ࢐ ȁࢠࢌ࢐ ሻ to the cloning node of ࢎࢌࢍ and ࢞ࢍ࢐ The results show that MMSE-based VAMP could perform the
simultaneously. We define a circularly symmetric complex channel estimation and detection; it is able to achieve better
Gaussian message characterized as performance than VAMP and linear detector, but it is not as
ષ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ ൫ࢎࢌࢍ ࢞ࢍ࢐ ൯
ࣀ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ
good as BiG-AMP in high SNR. On the other hand, in low SNR,
૚
ऍघ ൬ࢎࢌࢍ ࢞ࢍ࢐ Ǣ ǡ ൰ ሺ૝૛ሻ MMSE-based VAMP provides better reliability in comparison
ࢽ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ ࢽ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ

whose mean and variance are averaged over the distribution of with BiG-AMP. Fig. 5 presents the results of the BER versus
෩ ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ ൫ࢎࢌࢍ ࢞ࢍ࢐ ൯ǣ
ષ SNR performance with parameter settings; we see that MMSE-
ࢽ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ ൌ  ࡾ
૚
ࡵ ሺ૝૜ሻ based VAMP could perform well in the scenarios with small ۴,
෩ ሾ൫ࢎࢌࢍ ࢞ࢍ࢐ ൯ ሿାࢂࢇ࢘ષ
ࢂࢇ࢘ષ ෩ ሾ൫ࢎࢌࢍ ࢞ࢍ࢐ ൯ ሿሻ i.e., low SNR and small pilot percentage ( ۸‫ ܜܗܔܑܘ‬Ȁ۸), while BiG-
ࣀ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ
ࢽ࢚ࢌ࢐՜ሺࢌࢍǡࢍ࢐ሻ
ൌ ࡱષ෩ ሾࢎࢌࢍ ࢞ࢍ࢐ ሿ ሺ૝૝ሻ AMP could not. That is, in Fig. 5, BiG-AMP performs better
Corresponding messages in the inverse direction are than the other four methods when SNR= 3dB, but it cannot
ષ࢚ା૚ ࢚ ࢚ା૚ comprise to converge with SNR lower than 0dB when
ࢌࢍ՜ࢌ࢐ ൫ࢎࢌࢍ ൯ ൌ ࡼࡴ ൫ࢎࢌࢍ ห઼ࢍ ሻ ς࢐Ʈஷ࢐ ષࢌ࢐Ʈ՜ࢌࢍ ൫ࢎࢌࢍ ൯ ሺ૝૞ሻ
෩ ࢚ା૚ ࢚ା૚
ષࢍ࢐՜ࢌ࢐ ൫࢞ࢍ࢐ ൯ ൌ ࡼࢄ ሺ࢞ࢍ࢐ ሻ ςࢌƮஷࢌ ષࢌƮ࢐՜ࢍ࢐ ൫࢞ࢍ࢐ ൯ ሺ૝૟ሻ parameters are ۴ = 200, ۵ = 10, ۸ = 1000 and ۸‫ = ܜܗܔܑܘ‬10. The
Specifically, ࡼ࢘࢕࢐[.] is applied to the message ષ ෩ ࢚ା૚ SNR threshold of convergence reduces to -2dB with longer
ࢍ࢐՜ࢌ࢐ ൫࢞ࢍ࢐ ൯
࢚ା૚ ෩ ࢚ା૚
ષࢍ࢐՜ࢌ࢐ ൫࢞ࢍ࢐ ൯ ൌ ࡼ࢘࢕࢐ሾષࢍ࢐՜ࢌ࢐ ൫࢞ࢍ࢐ ൯ሿ ሺ૝ૠሻ pilot length ۸‫ = ܜܗܔܑܘ‬20 (dashed line) in Fig. 5; or, with shorter
Through Gaussian approximation, the messages given above block length ۸ = 700 (solid line) in Fig. 6. Fig.6 also
are denoted as demonstrates that the SNR diminishes significantly with
ષ࢚ࢌࢍ՜ࢌ࢐ ൫ࢎࢌࢍ ൯ ൌ ऍघ൫ࢎࢌࢍ Ǣ ࢚ࣖࢌࢍ՜ࢌ࢐ ǡ ࢚ࣚࢌࢍ՜ࢌ࢐ ൯ ሺ૝ૡሻ number of antennas, F, e.g., from 0dB for ۴ = 150 to -2 dB for
෩ ࢚ࢍ࢐՜ࢌ࢐ ൫࢞ࢍ࢐ ൯ ൌ ऍघ൫࢞ࢍ࢐ Ǣ ࣐
ષ ෥ ࢚ࢍ࢐՜ࢌ࢐ ǡ ࣓
෥ ࢚ࢍ࢐՜ࢌ࢐ ൯ ሺ૝ૢሻ ۴ = 200, with ۵ = 10, ۸ = 700, ۸‫ = ܜܗܔܑܘ‬10. Again, we observe in
ષࢍ࢐՜ࢌ࢐ ൫࢞ࢍ࢐ ൯ ൌ ऍघ൫࢞ࢍ࢐ Ǣ ࣐ࢍ࢐՜ࢌ࢐ ǡ ࣓࢚ࢍ࢐՜ࢌ࢐ ൯
࢚ ࢚
ሺ૞૙ሻ Fig. 6 that VAMP is able to provide almost the same reliability
 WK,((($QQXDO&RQVXPHU&RPPXQLFDWLRQV 1HWZRUNLQJ&RQIHUHQFH &&1&

as the proposed MMSE-based VAMP, but gets little worse

performance than MMSE-based VAMP under all SNR. Note
that the LS+MRC detector does not perform well with ۸‫ = ܜܗܔܑܘ‬۵
in Figs. 5 and Figs. 6. This is because its detection performance
is lower than ૚૙ି૚ in Fig. 5 with ۸‫ = ܜܗܔܑܘ‬2۵ = 20, which is worse
than the other four inference algorithms.
In Fig. 7, we study the impact of the role of number of antennas,
F, on the performance of signal detection. The parameter
settings are SNR= 0dB, ۵ = 10, ۸ = 1000, and ۸‫ = ܜܗܔܑܘ‬10, 20, 30.
We see that BERs of all five algorithms can be improved in the
cases of greater SNR and bigger ۴. The LS+MRC detector is
impressionable to pilot length, and it needs more than ۸‫ = ܜܗܔܑܘ‬30
to reach its optimization. When the pilot length is sufficient, the
performance of MMSE-based VAMP is better than that of
VAMP, VBI, or LS+MRC; but, it is inferior to BiG-AMP. We
note that the gap between BiG-AMP and other inference Fig. 3. Sum rate performance: sum rate versus the number of antennas.
schemes is insignificant in low SNR cases but becomes more
significant when SNR become larger. In Fig. 7, BiG-AMP
needs ۴ = 210 antennas to achieve the detection with ۸‫ = ܜܗܔܑܘ‬10.
This threshold on number of antennas decreases as pilot length
increases, e.g., ۴ = 130 for ۸‫ = ܜܗܔܑܘ‬20. Additionally, ۸‫ = ܜܗܔܑܘ‬30 is
required to perform the detection with F = 100 antennas. Fig. 8
depicts the ۸‫ ܜܗܔܑܘ‬effect on the performance. ۴ = 100, ۵ = 10, ۸ =
1000 and SNR= 0dB, 5dB are the parameter settings. The
results show that the increase in ۸‫ ܜܗܔܑܘ‬helps with improving the
performance of the five algorithms. Therefore, additional pilot
symbols are unnecessary once the convergence is guaranteed.
Particularly, LS+MRC needs about ۸‫ = ܜܗܔܑܘ‬20 ‫ ׽‬50 to prevent
performance deficiency. BiG-AMP needs ۸‫ = ܜܗܔܑܘ‬30 to prevent
divergence with SNR= 0dB, and only needs ۸‫ = ܜܗܔܑܘ‬20 when
SNR is raised to 5dB. On the contrary, MMSE-based VAMP
and VAMP still do well even when ۸‫ = ܜܗܔܑܘ‬۵ = 10. Fig. 4. Sum rate performance: sum rate versus the total downlink

VI. CONCLUSION transmission power.ġ

This paper proposes a novel adaptive ADC bit allocation for

mmWave communication. That is, with an MMSE-based
VAMP inference module, the adaptive ADC bit allocation
algorithm makes the quantization error of signals insignificant
which leads to the higher energy efficiency than a one-bit
quantized ADC system. Note that our proposed modules can be
easily equipped with state-of-the-art beamforming methods and
are applicable to the 5G mmWave system. Moreover, our
proposed modules are able to do channel estimation, data
detection and noise estimation simultaneously with adaptive-bit
allocation under ADC power constraints at the receiver. This is
because our proposed modules not only leverage VAMP
advantages of tremendous performance, low computational
Fig. 5. BER performance vs. SNR in low SNR with parameter settings: Jpilot
complexity and overhead, but also compensate for its
vulnerability in convergence under different schemes. = (solid) 10, (dashed) 20, F = 200; G = 10; J = 1000.
 WK,((($QQXDO&RQVXPHU&RPPXQLFDWLRQV 1HWZRUNLQJ&RQIHUHQFH &&1&

wave communications for future mobile networks,” IEEE J. Sel. Areas

Commun., vol. 35, no. 9, pp. 1909–1935, Sept 2017.
[2] S. Han, I. Chih-Lin, Z. Xu, and C. Rowell, “Large-scale antenna systems
with hybrid analog and digital beamforming for millimeter wave 5G,”
IEEE Commun. Mag., vol. 53, no. 1, pp. 186–194, Jan. 2015.
[3] J. Mo and R. W. Heath, “Capacity analysis of one-bit quantized MIMO
systems with transmitter channel state information,” IEEE Trans. Signal
Process., vol. 63, no. 20, pp. 5498–5512, Jul. 2015.
[4] Y. Li, C. Tao, G. S. Granados, A. Mezghani, A. L. Swindlehurst, L. Liu,
“Channel Estimation and Performance Analysis of One-Bit Massive
MIMO Systems,” IEEE Trans. Signal Process., vol. 65, no. 15, pp. 4075 -
4089, Aug. 2017.
[5] N. Liang and W. Zhang, “Mixed-ADC massive MIMO,” IEEE J. Sel. Areas
Commun., vol. 34, no. 4, pp. 983–997, Mar. 2016.
[6] T. C. Zhang, C.-K.Wen, S. Jin, and T. Jiang, “Mixed-ADC massive MIMO
detectors: Performance analysis and design optimization,” IEEE Trans.
Wireless Commun., vol. 15, no. 11, pp. 7738–7752, Nov. 2016.
[7] C. K. Wen, C. J. Wang, S. Jin, and K. K. Wong, “Bayes-Optimal Joint
Fig. 6. BER performance vs. SNR in low SNR with parameter settings: Channel-and-Data Estimation for Massive MIMO With Low-Precision
ADCs,” IEEE Trans. Signal Process., vol. 64, no. 10, pp. 2541–2556, 2015.
[8] V. Buchoux, O. Capp´e, ´ E. Moulines, and A. Gorokhov, “On the
G = 10; J = 700; Jpilot = 10. F = (solid) 200, (dashed) 150.
performance of semi-blind subspace-based channel estimation,” IEEE
Trans. Signal Process., vol. 48, no. 6, pp. 1750–1759, 2000.
[9] M. Bayati and A. Montanari, “The dynamics of message passing on dense
graphs, with applications to compressed sensing,” IEEE Trans. Inform.
Theory, vol. 57, no. 2, pp. 764–785, 2011.
[10] D. G. Tzikas, A. C. Likas, and N. P. Galatsanos, “The variational
approximation for bayesian inference,” IEEE Trans. Signal Process., vol.
25, no. 6, pp. 131–146, 2008.
[11] J. T. Parker, P. Schniter, and V. Cevher, “Bilinear Generalized
Approximate Message Passing—Part II: Applications,” IEEE Trans.
Signal Process., vol. 62, no. 22, pp. 5854–5867, 2014.
[12] Z. Zhang, X. Cai, C. Li, C. Zhong, and H. Dai “One-bit Quantized Massive
MIMO Detection Based on Variational Approximate Message Passing,”
IEEE Trans. Signal Process., vol. 66, no. 9, pp. 2358-2373, 2018.
[13] O. Orhan, E. Erkip, and S. Rangan, “Low power analog to digital
conversion in millimeter wave systems: Impact of resolution and
bandwidth on performance,” in Proc. IEEE Inf. Theory Appl. Workshop,
Feb. 2015, pp. 191–198.
[14] N. Jindal, “MIMO broadcast channels with finite-rate feedback,” IEEE
Trans. Inf. Theory, vol. 52, no. 11, pp. 5045–5060, Nov 2006.
Fig. 7. BER performance vs. antenna number F, with Jpilot = 10(solid), [15] E. De Carvalho and D. Slock, “Semi-blind maximum-likelihood
multichannel estimation with gaussian prior for the symbols using soft
20(dashed), 30(dot-dash), SNR= 0dB, G = 10. J = 1000. decisions,” in IEEE Veh. Technol. Conf. (IEEE VTC), vol. 2, 1998, pp.
1563–1567.
[16] D. G. Tzikas, A. C. Likas, and N. P. Galatsanos, “The Variational
Approximation for Bayesian Inference,” IEEE Trans. Signal Process., vol.
25, no. 6, pp. 131–146, 2008.
[17] S. Rangan, J. T. Parker, P. Schniter, and etc., “Gampmatlab,” https://
sourceforge.net/projects/gampmatlab/.
[18] C. Risi, D. Persson, and E. G. Larsson, ȾMassive MIMO with 1-bit
ADC,ȿ ArXiv e-prints, arXiv:1404.7736, Apr. 2014.
https://arxiv.org/pdf/1404.7736

Fig. 8. BER performance vs. pilot length Jpilot with F = 100; G = 10; J =

1000, SNR = 0dB and 5dB.

REFERENCES
[1] M. Xiao, S. Mumtaz, Y. Huang, L. Dai, Y. Li, M. Matthaiou, G.
K.Karagiannidis, E. Bjornson, K. Yang, C. L. I, and A. Ghosh, “Millimeter