Wire D 19 00453 - R1

Wireless Personal Communications
Frame Structures for Massive MIMO Communications under Channel Aging

--Manuscript Draft--
Manuscript Number: WIRE-D-19-00453R1
Full Title: Frame Structures for Massive MIMO Communications under Channel Aging
Article Type: Short Communication
Keywords: Massive MIMO, CSI estimation, Channel aging, performance analysis
Corresponding Author: Ribhu Chopra

Indian Institute of Technology Guwahati
Guwahati, Assam INDIA
Corresponding Author Secondary

Information:
Corresponding Author's Institution: Indian Institute of Technology Guwahati
Corresponding Author's Secondary

Institution:
First Author: Anubhab Chowdhury
First Author Secondary Information:
Order of Authors: Anubhab Chowdhury
Ribhu Chopra
Order of Authors Secondary Information:
Funding Information: Indian Institute of Technology Guwahati Dr. Ribhu Chopra

(xEEESUGIITG01212RBHU001)
Abstract: We study the problem of optimizing the frame structure of a massive MIMO system
under channel aging. We argue that the conventional TDD frame structure with lumped
training is suboptimal under rapidly aging channels. We, therefore discuss a
generalized frame structure allowing for the training of a fraction of the total number of
users, with the conventional lumped and interspersed training frames as its special
cases. We then derive the achievable uplink and downlink rates for this system,
incorporating the cost for switching from uplink to downlink and vice versa. The
analysis, and the subsequent numerical results clearly bring out the dependence of the
rates achievable in a massive MIMO system on the choice of the frame structure.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
Response to reviewer's comments Click here to access/download;Response to reviewer's
comments;review_response_r1.pdf
1
Response to Review Comments on

WIRE-D-19-00453, “Frame Structures for
Massive MIMO Communications under
Channel Aging”
Anubhab Chowdhury, and Ribhu Chopra
We thank the reviewers for their generous time in reviewing our paper and for their valuable feedback.
The manuscript has been modified in light of the comments by the editor and the reviewers; and a point
by point response to all the comments follows in the sequel.
Notation: The reviewers’ comments are indicated in italicized text, while our responses are in normal
text. When referring to the manuscript under review, the page, section and equation numbers correspond
to the revised version of the manuscript, and not the originally submitted version. Also, unless specified
otherwise, the notation in this document is as in the manuscript and hence is not explicitly defined here.
The references cited within the response correspond to those listed at the end of this document unless
stated otherwise.
A. Chowdhury and R. Chopra are with the Indian Institute of Technology Guwahati, Assam, India. (email: ribhu@outlook.com).
April 26, 2019 DRAFT

2
R ESPONSE TO THE C OMMENTS F ROM R EVIEWER 1
Comment 1.1: Can u discuss on the transmit and receive power for the model by considering its small
cell scenario?
Response: We thank the reviewer for this very important observation, and have now included a discussion
on downlink power control for the given system in Section III of our manuscript.
Comment 1.2: The interference which is an imp. phenomenon in such scenarios are missing....Authors
need to introduce some latest manuscripts in this regard along with analysis. some of the suggestions
are: (a) 10.1109/TCSI.2018.2866933 (b)
Response: We thank the reviewer for pointing out these important references, that have undoubtedly added
to our bibliography. We would also like to point out that we consider a single cell system. Therefore,
our main sources of interference in this case are therefore channel estimation errors, and the effects of
channel aging, already discussed in the paper in detail. We have however added explanations about the
different sources of interference after equations (4) and (10).
Comment 1.3: Please correct the references pointers....everywhere it is a question mark....
Response: We thank the reviewer for bringing this oversight to our notice, and have now corrected this
in the updated manuscript.
Comment 1.4: Equation [3] needs to be referenced.
Comment 1.5: Figure 2 is too congested and hard to read....Please either split it or increase the font
size within single plot.
Response: We agree with the reviewer’s observation. However, we require all the plots within this figure
for completeness, and cannot split it into two sub figures.
Comment 1.6: Discuss about the simulation setup in detail in a separate paragraph.... as such papers
are thoroughly based on simulations, so its important to provide proper background....
Response: We agree with the reviewer’s observation. We have now detailed all the simulation parameters

3
in a separate paragraph at the beginning of the section on numerical results.
Comment 1.7: Finally at the end compare your result in a table and show superiority of your worked in
comparison with state of the arts. It will give a proper insight for the work....
Response: We thank the reviewer for this important suggestion, and have now tabulated the gains offered
by the proposed frame structure over the conventional frame structure.
Finally, we thank the reviewer for all the constructive comments and valuable feedback, that have no
doubt improved our paper. We are very grateful for his/her time.

Manuscript Click here to
access/download;Manuscript;review_response_r1.tex
Click here to view linked References 1
Response to Review Comments on

WIRE-D-19-00453, “Frame Structures for
Massive MIMO Communications under
Channel Aging”
We thank the reviewers for their generous time in reviewing our paper and for their valuable feedback.
The manuscript has been modified in light of the comments by the editor and the reviewers; and a point
by point response to all the comments follows in the sequel.
Notation: The reviewers’ comments are indicated in italicized text, while our responses are in normal
text. When referring to the manuscript under review, the page, section and equation numbers correspond
to the revised version of the manuscript, and not the originally submitted version. Also, unless specified
otherwise, the notation in this document is as in the manuscript and hence is not explicitly defined here.
The references cited within the response correspond to those listed at the end of this document unless
stated otherwise.

2
R ESPONSE TO THE C OMMENTS F ROM R EVIEWER 1
Comment 1.1: Can u discuss on the transmit and receive power for the model by considering its small
cell scenario?
Response: We thank the reviewer for this very important observation, and have now included a discussion
on downlink power control for the given system in Section III of our manuscript.
Comment 1.2: The interference which is an imp. phenomenon in such scenarios are missing....Authors
need to introduce some latest manuscripts in this regard along with analysis. some of the suggestions
are: (a) 10.1109/TCSI.2018.2866933 (b)
Response: We thank the reviewer for pointing out these important references, that have undoubtedly added
to our bibliography. We would also like to point out that we consider a single cell system. Therefore,
our main sources of interference in this case are therefore channel estimation errors, and the effects of
channel aging, already discussed in the paper in detail. We have however added explanations about the
different sources of interference after equations (4) and (10).
Comment 1.3: Please correct the references pointers....everywhere it is a question mark....
Comment 1.4: Equation [3] needs to be referenced.
Comment 1.5: Figure 2 is too congested and hard to read....Please either split it or increase the font
size within single plot.
Response: We agree with the reviewer’s observation. However, we require all the plots within this figure
for completeness, and cannot split it into two sub figures.
Comment 1.6: Discuss about the simulation setup in detail in a separate paragraph.... as such papers
are thoroughly based on simulations, so its important to provide proper background....
Response: We agree with the reviewer’s observation. We have now detailed all the simulation parameters

3
in a separate paragraph at the beginning of the section on numerical results.
Comment 1.7: Finally at the end compare your result in a table and show superiority of your worked in
comparison with state of the arts. It will give a proper insight for the work....
Response: We thank the reviewer for this important suggestion, and have now tabulated the gains offered
by the proposed frame structure over the conventional frame structure.
Finally, we thank the reviewer for all the constructive comments and valuable feedback, that have no
doubt improved our paper. We are very grateful for his/her time.

Manuscript Click here to
access/download;Manuscript;lett_aging_frame_r1.docx
Frame Structures for Massive MIMO

Communications under Channel Aging
Abstract
We study the problem of optimizing the frame structure of a massive MIMO system under channel
aging. We argue that the conventional TDD frame structure with lumped training is suboptimal under
rapidly aging channels. We, therefore discuss a generalized frame structure allowing for the training of a
fraction of the total number of users, with the conventional lumped and interspersed training frames as
its special cases. We then derive the achievable uplink and downlink rates for this system, incorporating
the cost for switching from uplink to downlink and vice versa. The analysis, and the subsequent numerical
results clearly bring out the dependence of the rates achievable in a massive MIMO system on the choice
of the frame structure.
Index Terms
Massive MIMO, CSI estimation, Channel aging, performance analysis
I. INTRODUCTION
The use of a large number of antennas in multiuser multiple input multiple output (MU MIMO)
systems, popularly known as massive MIMO systems, results in quasi orthogonality among the
channels for different users. This phenomenon, known as channel hardening, further results in
simplified precoder and receiver structures [1], and increased spectral and energy efficiencies
[2]–[5]. These advantages make massive MIMO a front runner technology for next generation
wireless communication systems. Since these benefits of massive MIMO are based on the quasi
orthogonality among the channels for different users, their dependence on the quality of the
channel state information (CSI) at the base station is high. However, the CSI acquired by the base
station may be inaccurate due to the presence of additive noise in channel estimation [6],

2
pilot contamination [7] and channel aging [8]. In this paper we discuss the choice of an optimal
time division duplexed (TDD) frame structure for a single cell massive MIMO system under
channel aging.
Channel aging refers to the difference in the actual channel state at the time of data
transmission, and the channel state estimate acquired at the base station during training [8]–
[11]. The phenomenon of channel aging is mainly attributed to user mobility [8] and phase noise
at the users and the BS [11]. The continuously evolving nature of the channel implies that the
accuracy of the available CSI, and consequently the system performance worsens as the time gap
between training and transmission increases. This also implies that the effects of aging become
more pronounced in large dimension systems, requiring longer training intervals, further
resulting in smaller transmission duty cycles [12]. Recent studies have shown that the inherent
advantages of massive MIMO such as precoder and combiner simplicity, and power scaling
remain unaffected by channel aging. However, aging limits the number of users being served, and
even the number of usable base station antennas in an frequency division duplexed (FDD) system
[13].
The question of optimal training schemes for aging channels has been addressed recently, and
it has been argued that the temporal correlation of the channel allows for the use of adaptive
filtering techniques, such as Kalman filtering [14]–[16] for channel tracking. It was also argued in
[16] that the conventional lumped training based frame structure may lead to unfair uplink rate
distributions among users despite power control, and interspersed frames lead to a fairer
distributions of rates among the users.
In this paper, we address the problem of the selection of an optimal frame structure for a TDD
massive MIMO system. For this purpose, we consider a generalized frame structure, formed by
stacking multiple smaller, variable length service blocks, each of which consists of training, uplink
transmission, and downlink transmission sub-blocks. In such a case, the question of the choice of
an optimal frame structure reduces to the selection of appropriate service frame durations. We
also incorporate the effects of switching from uplink to downlink transmission, and vice-versa, in
terms of an equivalent switching cost [17], thus potentially taxing the schemes that require a
large amount of switching from one mode to the other.
Our main contributions are as follows:
• We derive the achievable uplink and downlink rates of a massive MIMO system with

3
the generalized frame structure incorporating the costs of switching between uplink and
downlink. See Section III
• Via extensive numerical simulations, we characterize the behavior of the achievable downlink
and uplink rates in terms of the frame structure being employed. See Section IV.
The key finding of this work is that the achievable throughput in a massive MIMO system
depends heavily on the frame structure being used, and it is suboptimal to train all the users at
once, even in the presence of high switching costs. Also, the optimal frame duration as well as
the fraction of users to be trained per frame is a function of both the number of users and the
user mobilities. This leads to the proposed flexible frame structure yielding higher throughputs,
in comparison to the more conventional lumped or interspersed frames. In the next section, we
discuss the system model for this work.
II. SYSTEM MODEL
We consider a single cell system comprising a base station with N antennas, serving K(< N)
single antenna mobile users, as depicted in Fig. 1. We assume the channels between the BS and
the users to be reciprocal [2], and the system to operate in TDD mode. Each frame consists of
multiple service blocks, as also shown in Fig. 1. Each block further consists of three sub blocks,
viz. training, uplink transmission, and downlink transmission. During the training sub-block, L out
of the K users are scheduled for training, and transmit L orthogonal training symbols to the base
station over L channel uses. The base station uses these training symbols to update the available
channel estimates for these L users. Following this, all the K users share their uplink data during
the next λL channel uses, decoded by the BS using the available channel estimates. The BS and
the K users then switch from uplink transmission to downlink transmission, incurring a fixed cost
equivalent to c channel uses. The system then operates in the downlink mode for the next µL
channel uses. During this time, the base station transmits K appropriately precoded independent
data streams to the K users. At the end of each service block, the system again switches from
downlink to uplink transmission, incurring a cost equivalent to c channel uses. The overall average
duration of one frame, T can therefore be expressed as and that of

one service block can be expressed as TL = (1+µ+λ)L+2c,. This frame structure reduces to a
lumped training frame for L = K, and an interspersed training frame for L = 1.
The channel between the ith BS antenna and the kth user is assumed to evolve with time due

4
√
to user mobility and phase noise [11], and is modeled at the nth instant as βkhik[n], with βk as the
path loss component, and hik[n] as the fast fading component modeled as a zero mean circularly
symmetric complex Gaussian (ZMCSCG) random variable (rv) with unit variance,
Fig. 1. The system model.
denoted as CN(0,1). We assume that the relative distance between the BS antennas is much
smaller as compared to the distance between the BS and the users, and consequently, βk remains
same across all the base station antennas. Letting β, [β1,...,βK]T , the channel matrix between
√
the BS and the users at the nth instant can be expressed as H[n]diag( β), with H[n] ∈ CN×K, and
its (ik)th entry representing the ZMCSCG channel between the ith BS antenna and the kth user.
The channel evolution is modeled as a linear combination of its initial state and an
innovation component as [8], [18]
hik[n] = ρ[n]hik[0] + ¯ρ[n]zik[n] (1)
with zik[n] ∼ CN(0,1) being the innovation component, and 0 ≤ ρ[n] ≤ 1 such that ρ[τ] =
. Also, for any variable x, x¯ ,p1 − |x|2. The channel innovation processes,
zik[n] is not white, and is assumed to be stationary in the wide sense. By convention, it is assumed
that the channel evolves according to the Jakes’ model [8], [19], and the channel correlation

5
coefficient, ρ[n] is modeled as ρ[n] = J0(2πfdTsn), where fd is the Doppler frequency, Ts is the
sampling period, and J0(·) is the Bessel function of the first kind and zeroth order [20, Eq.
(9.1.18)].
We assume that the channel estimates at the BS are acquired and updated using a Kalman
filtering based technique [14]–[16]. We denote the set of channel indices being used for training
in the mth frame by Im, the training instant for the kth user as tk,m ∈ Im, the estimate of the channel
vector between the BS and the kth user obtained at tk,m as hˆk,m, and the mean squared channel
estimation error for hˆk,m as ψk,m. It can then be shown that the channel hk[n] can be
expressed in terms of hˆk,m as [16]
h , (2)
with ψk,m[n] , |ρ[n − tk,m]|ψk,m + |ρ¯[n − tk,m]|2, and h˜k,m[n] being the zero mean channel
estimation error uncorrelated with hˆk,m. We do not list the Kalman filter update equations for
the sake of brevity. However, it is easy to show that ψk,m is upper bounded as
with Ep,k being the pilot power for the kth user. We next use this bound to obtain achievable
uplink and downlink rates for the proposed system.
III. ACHIEVABLE UPLINK AND DOWNLINK RATES
In this section, we derive the achievable uplink and downlink rates for the given system
assuming maximal ratio combining (MRC) in the uplink, and matched filter precoding (MFP) in the
downlink. During the uplink sub-frame, all the K users simultaneously transmit their signals
to the base station, such that the received signal at the nth instant is given as [13],
y , (3)
where sk[n] is the symbol being transmitted by the kth user at the nth instant, and Es,u,k is the
uplink signal power of the kth user. It has been argued in the literature [2], that channel hardening
results in the elimination of fast fading effects, thus making the effective downlink channel
coefficient proportional to the slow fading coefficient βk. Therefore, recent works [2], [13]

6
advocate the use of channel inversion based power control in massive MIMO, such that βkEs,u,k =
Es,u, and βkEp,k = Ep,
Using the most recent channel estimate for the kth user’s channel, hˆk,m, for combining, the
combined signal at the base station can be expanded as
vk[n] = (1 − ψk,m[n])1/2pβkEs,u,kE[hˆHk,mhˆk,m]su,k[n]
+ (1 − ψk,m[n])1/2pβkEs,u,k(E[hˆHk,mhˆk,m] − hˆHk,mhˆk,m)su,k[n]
q
+ ψk,m[n]βkEs,u,khˆHk,mh˜k,m[n]su,k[n]
,mhl
In the above, the first term corresponds to the desired signal, the second to the interference due
to the difference between the true uplink channel, and its mean value considered at the BS, the
third to the self interference due to the estimation error at the BS, the fourth to the inter user
interference, and the last to the additive noise. It has been argued in [21] that in the worst case,
all the interference terms can be treated as AWGN. Treating interference as noise, assuming
channel inversion based uplink pilot and uplink/downlink data power control [2], [13] the uplink
SINR for the kth user can be obtained via the use and forget bounds detailed in [2] as
. (5)
Since (1−ψk,m[n]) can be lower bounded as , and defining ,

we can lower bound the per user achievable uplink rate as
Ru ≤ X log2(1 + |ρ[n − tk,m]|2γu[0]). (6)

n∈Tu
where Tu is the set of time indices during which uplink data is transmitted, and is defined as Tu =
{mTL + L + 1,...,mTL + (λ + 1)L}, m ∈ Z. This can be maximized by maximizing
γu[0] in terms of the pilot and signal powers, with a constraint on the per user average received
power [12].

7
Similarly, defining the set of time indices for downlink transmission, Td, as Td = {mTL +
(λ+1)L+c+1,...,(m+1)TL −c} m ∈ Z. The downlink signal transmitted by the BS at the
nth (n ∈ Td) instant is
x (7)
with Ed,s,k being the transmit power allocated to the kth user, and λ being the power normalization
factor at the BS. The power normalization factor λ used to restrict the total transmitted power
by the BS, such that [2],
2 Et
λ= , (8)
E[Tr{Hˆ Hdiag(Ed)Hˆ }]
where, Et is the total power radiated by the BS, and the vector Ed is defined as Ed , [Ed,s,1,...,Ed,s,K].
Using channel hardening arguments [22], it is easy to show that λ reduces to,
. (9)
Defining, ,
the signal received at the kth user at the nth instant, rk[n], can be expanded in the form
In the above, the first term corresponds to the desired signal, the second to the interference
due to the difference between the true effective downlink channel and its mean, the third to the
interference due to the estimation error of the true channel vector at the BS, the fourth to the
inter user interference, and the last to the AWGN. Again employing use and forget analysis
from [2], and upper bounding the channel estimation error as , the effective SINR
for the kth user at the nth instant can be reduced to γd,k[n] = |ρ[n − tk,m]|2γd,k[0] where

8
. (11)
Further, the achievable downlink rate of the kth user is obtained as
Rd,k = X log2(1 + γd,k[n]). (12)

n∈Td
Now, based on the frame structure, it can be observed that the channels available for the
downlink will, in general be aged more as compared to the channels available for uplink. This is
in order to facilitate communication for the battery powered users, as opposed to the centrally
powered base station. Also, by virtue of the frame structure, the number of ‘less aged’ channel
slots available for uplink transmission will be an increasing function of L, and therefore the
achievable uplink rate will be maximized for L = K. However, no such observation can be made
about the achievable down-link rates. Therefore, in the next section, we numerically evaluate the
achievable downlink rates, and optimize these in terms of L.
IV. NUMERICAL RESULTS
For the purpose of these experiments, we consider a single cell massive MIMO system
containing a 1000 (N) antenna base station, transmitting at a carrier frequency (fc) of 2 GHz. We
assume that the BS serves a maximum of K = 200 users, uniformly distributed over a circular cell
with radius 5 km. The reference distance (d0) in this case is assumed to be 1 km, and the path loss
coefficient (βk) for the kth user, located at a distance dk from the B.S. is
calculated as
, (13)
with η being the propagation constant. In this setup we assume η = 4. Also, similar to other past
works on channel agingin massive MIMO systems [9], [11], [13], we consider a narrowband
channel, with a signal bandwidth of 1 MHz. We assume that the base station samples at the
Nyquist rate of the complex baseband signal, i.e., at 1 MHz.
Fig. 2 plots the achievable downlink rates against the fraction for different user velocities
and for different values of λ = µ = a with no switching cost, i.e. for c = 0. It can be observed that
at low user mobilities the achievable uplink rates are independent of the chosen value of L;

9
however, the achievable rate becomes highly sensitive to the choice of L. It is also observed that
longer frame durations are preferable for scenarios with low user mobilities, and vice versa, which
is in accordance with intuition. It is also clear that the optimal value of L depends on the overall
frame length, as well as the user mobility, and needs to be determined numerically.
In Figs. 3 and 4, we plot the achievable uplink rates against the number of users (K) for different
user mobilities for λ = µ = 2. In these figures we also compare the achievable rates for lumped,
and interspersed training schemes against the schemes corresponding to the optimal value of the
parameter L for different switching costs. It is observed that for a large number of users at high
user mobilities, it is preferable to use interspersed training as compared to lumped training when
no switching cost is involved, which is as per the expectations. However, lumped training is
preferable when the switching cost is high.
In Fig. 5 we plot the achievable per user uplink rates for different cases against the number of
users at a user velocity of 150 km/h. As discussed earlier, the uplink rates are maximized

10
Fraction of users trained per block (L/K)
Fig. 2. Achievable per user rates at different user mobilities for different frame durations and different values of L.
for L = K. However, the choice of a value of L, optimized for downlink transmission does not
affect the achievable uplink rates significantly for either c = 0, or c = 1.
In Table I, we tabulate the percentage gains achieved by using the proposed frame structure
over the conventional lumped frame structure for 200 users and different switching costs for
different user velocities. It is observed that the performance of a massive MIMO system can be
increased by as much as 38% by using the proposed frame structure instead of the more
conventional lumped structure. It is also observed that the impact of an increased switching cost
on these gains is minimal.
TABLE I
GAINS OVER THE CONVENTIONAL LUMPED TRAINING STRUCTURE
Switching Cost 0 0.25 0.5 0.75 1 1.5 2 5

Percentage 37.93% 37.77% 37.6% 37.58% 37.46% 37.16% 36.68% 37.56%
Gain
Number of Users (K)
Fig. 3. Achievable downlink rates for different numbers of users under different mobilities with no switching cost(c=0).

11
V. CONCLUSIONS
In this paper, we considered a generalized frame structure for TDD massive MIMO systems. It
was argued that the more traditional interspersed and lumped frame structures are special cases
of the proposed frame. We derived bounds on the achievable uplink and downlink rates
considering MRC and MFP respectively. Finally, using numerical results, we illustrated that, in
general, the achievable downlink rates depend on the number of users being trained per service
block, and the length of a service block. It was also shown that the proposed block structure can
be used to fine tune these parameters to optimize the frame rates for given channel and mobility
conditions, resulting in a gain of more than 35% under high user mobilities.
REFERENCES
[1] T. Narasimhan and A. Chockalingam, “Channel hardening-exploiting message passing (CHEMP) receiver in large-scale
MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 8, pp. 847–860, Oct. 2014.
[2] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals of Massive MIMO. Cambridge University Press,
Cambridge, UK, 2016.
[3] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL of cellular networks: How many antennas do we
need?,” IEEE J. Sel. Areas Commun., vol. 31, pp. 160–171, Feb. 2013.
Number of Users (K)
Fig. 4. Achievable downlink rates for different numbers of users under different mobilities with c=1.

12
[4] S. Blandino, G. Mangraviti, C. Desset, A. Bourdoux, P. Wambacq, and S. Pollin, “Multi-user hybrid MIMO at 60 GHz using
16-antenna transmitters,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 66, pp. 848–858, Feb. 2019.
[5] M. Rahman and J.-D. Park, “The smallest form factor UWB antenna with quintuple rejection bands for iot applications
utilizing RSRR and RCSRR,” Sensors, vol. 18, p. 911, Feb. 2018.
[6] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach to channel estimation in large-scale multiple-antenna
systems,” IEEE J. Sel. Areas Commun., vol. 31, pp. 264–273, Feb. 2013.
[7] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot contamination and precoding in multi-cell TDD systems,”
IEEE Trans. Wireless Commun., vol. 10, pp. 2640–2651, Aug. 2011.
[8] K. T. Truong and R. W. Heath, “Effects of channel aging in massive MIMO systems,” Journal of Communications and
Networks, vol. 15, pp. 338–351, Aug. 2013.
[9] A. K. Papazafeiropoulos and T. Ratnarajah, “Deterministic equivalent performance analysis of time-varying massive MIMO
systems,” IEEE Trans. Wireless Commun., vol. 14, pp. 5795–5809, Oct. 2015.
[10] G. Amarasuriya and H. V. Poor, “Impact of channel aging in multi-way relay networks with massive MIMO,” in Proc. IEEE
Intl. Conf. Commun. (ICC 2015), (London, UK), pp. 1951–1957, June 2015.
[11] A. K. Papazafeiropoulos, “Impact of general channel aging conditions on the downlink performance of massive MIMO,”
IEEE Trans. Veh. Technol., vol. 66, pp. 1428–1444, Feb. 2016.
[12] R. Chopra, C. R. Murthy, and H. A. Suraweera, “On the throughput of large MIMO beamforming systems with channel
aging,” IEEE Signal Process. Lett., vol. 23, pp. 1523–1527, Nov. 2016.
[13] R. Chopra, C. Murthy, H. Suraweera, and E. Larsson, “Performance analysis of FDD massive MIMO systems under

13
Number of users (K)
Fig. 5. Achievable uplink rates for different numbers of users under different configurations for a user velocity of 150 km/h.
channel aging,” IEEE Trans. Wireless Commun., vol. 17, pp. 1094–1108, Feb. 2018.
[14] S. Kashyap, C. Mollen, E. Bj´ ornson, and E. G. Larsson, “Performance analysis of TDD massive MIMO with Kalman¨ channel
predication,” in Proc. Intl. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2017), (New Orleans, LA),
pp. 3554–3558, March 2017.
[15] V. Arya and K. Appaiah, “Kalman filter based tracking for channel aging in massive MIMO systems,” in 2018 International
Conference on Signal Processing and Communications (SPCOM) (SPCOM 2018), (Indian Institute of Science, Bangalore,
India), July 2018.
[16] R. Chopra, “Uplink training for massive MIMO systems under channel aging,” in 2018 International Conference on Signal
Processing and Communications (SPCOM) (SPCOM 2018), (Indian Institute of Science, Bangalore, India), July 2018.
[17] L. Chen, S. Iellamo, and M. Coupechoux, “Opportunistic spectrum access with channel switching cost for cognitive radio
networks,” in 2011 IEEE International Conference on Communications (ICC), pp. 1–5, June 2011.
[18] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors, and F. Tufvesson, “Scaling up MIMO:
Opportunities and challenges with very large arrays,” IEEE Signal Process. Mag., vol. 30, pp. 40–60, Jan. 2013.
[19] W. C. Jakes and D. C. Cox, eds., Microwave Mobile Communications.
[20] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables.
New York: Dover, 9th ed., 1964.

14
3
[21] B. Hassibi and B. M. Hochwald, “How much training is needed in multiple-antenna wireless links?,” IEEE Trans. Inf. Theory,
vol. 49, pp. 951–963, Apr. 2003.
[22] R. Couillet and M. Debbah, Random matrix methods for wireless communications. Cambridge University Press, Cambridge,
UK, 1st ed., 2011.

Line Figure Click here to access/download;Line Figure;model.eps
Line Figure Click here to access/download;Line Figure;fig2.eps
3
v=10 km/h
v=100 km/h
2.5 v=200 km/h
Achivable per user rate (bps)
1.5
Solid: Optimal
0.5
Dashed: Purely Interspersed
Dotted: Purely Lumped
0
50 100 150 200
Number of Users (K)
1.2
1
0.8
0.6
0.4
a=1
a=2
0.2 a=3 Solid: v=10km/h
a=4 Dashed: v=150 km/h
a=5
0
0 0.2 0.4 0.6 0.8 1
3
v=10 km/h
v=100 km/h
2.5 v=200 km/h
Achivable Rate Per User
1.5
0.5 Solid: Optimal

0
50 100 150 200
Number of Users (K)
3
Lumped Training
Achivable Per user uplink rate (bps)
Interspersed Training c=0

2.5 Optimal for downlink c=0
Optimal for downlink c=1
1.5
0.5
0 50 100 150 200
Number of users (K)
Author's Picture & Biography Click here to access/download;Author's Picture &
Biography;bio.pdf
1
Author Bio
Ribhu Chopra
Ribhu Chopra received the B.E. degree in Electron-

ics and Communication Engineering from Panjab
University, Chandigarh, India in 2009, and the M.
Tech. and Ph. D. Degrees in Electronics and Com-
munication Engineering from the Indian Institute
of Technology Roorkee, India in 2011 and 2016
respectively. He worked as a project associate at
Department of Electrical Communication Engineer-
ing, Indian Institute of Science, Bangalore from Aug.
2015, till May 2016. From May 2016 to March 2017
he worked as an institute research associate at the
Department of Electrical Communication Engineering, Indian Institute of Sci-
ence, Bangalore, India. In April 2017, he joined the department of Electronics
and Electrical Engineering, Indian Institute of Technology Guwahati, Assam,
India. His research interests include statistical and adaptive signal processing,
massive MIMO communications, and cognitive communications.
Attachment to Manuscript Click here to access/download;Attachment to
Manuscript;lett_aging_frame_r1.pdf
Frame Structures for Massive MIMO

Communications under Channel Aging
Abstract
We study the problem of optimizing the frame structure of a massive MIMO system under channel
aging. We argue that the conventional TDD frame structure with lumped training is suboptimal under
rapidly aging channels. We, therefore discuss a generalized frame structure allowing for the training of a
fraction of the total number of users, with the conventional lumped and interspersed training frames as its
special cases. We then derive the achievable uplink and downlink rates for this system, incorporating the
cost for switching from uplink to downlink and vice versa. The analysis, and the subsequent numerical
results clearly bring out the dependence of the rates achievable in a massive MIMO system on the
choice of the frame structure.
Index Terms
Massive MIMO, CSI estimation, Channel aging, performance analysis
I. I NTRODUCTION
The use of a large number of antennas in multiuser multiple input multiple output (MU
MIMO) systems, popularly known as massive MIMO systems, results in quasi orthogonality
among the channels for different users. This phenomenon, known as channel hardening, further
results in simplified precoder and receiver structures [1], and increased spectral and energy
efficiencies [2]–[5]. These advantages make massive MIMO a front runner technology for next
generation wireless communication systems. Since these benefits of massive MIMO are based on
the quasi orthogonality among the channels for different users, their dependence on the quality of
the channel state information (CSI) at the base station is high. However, the CSI acquired by the
base station may be inaccurate due to the presence of additive noise in channel estimation [6],

2
pilot contamination [7] and channel aging [8]. In this paper we discuss the choice of an optimal
time division duplexed (TDD) frame structure for a single cell massive MIMO system under
channel aging.
Channel aging refers to the difference in the actual channel state at the time of data transmis-
sion, and the channel state estimate acquired at the base station during training [8]–[11]. The
phenomenon of channel aging is mainly attributed to user mobility [8] and phase noise at the
users and the BS [11]. The continuously evolving nature of the channel implies that the accuracy
of the available CSI, and consequently the system performance worsens as the time gap between
training and transmission increases. This also implies that the effects of aging become more
pronounced in large dimension systems, requiring longer training intervals, further resulting in
smaller transmission duty cycles [12]. Recent studies have shown that the inherent advantages of
massive MIMO such as precoder and combiner simplicity, and power scaling remain unaffected
by channel aging. However, aging limits the number of users being served, and even the number
of usable base station antennas in an frequency division duplexed (FDD) system [13].
The question of optimal training schemes for aging channels has been addressed recently, and
it has been argued that the temporal correlation of the channel allows for the use of adaptive
filtering techniques, such as Kalman filtering [14]–[16] for channel tracking. It was also argued
in [16] that the conventional lumped training based frame structure may lead to unfair uplink
rate distributions among users despite power control, and interspersed frames lead to a fairer
distributions of rates among the users.
In this paper, we address the problem of the selection of an optimal frame structure for
a TDD massive MIMO system. For this purpose, we consider a generalized frame structure,
formed by stacking multiple smaller, variable length service blocks, each of which consists of
training, uplink transmission, and downlink transmission sub-blocks. In such a case, the question
of the choice of an optimal frame structure reduces to the selection of appropriate service frame
durations. We also incorporate the effects of switching from uplink to downlink transmission,
and vice-versa, in terms of an equivalent switching cost [17], thus potentially taxing the schemes
that require a large amount of switching from one mode to the other.
Our main contributions are as follows:
• We derive the achievable uplink and downlink rates of a massive MIMO system with
the generalized frame structure incorporating the costs of switching between uplink and
downlink. See Section III

3
• Via extensive numerical simulations, we characterize the behavior of the achievable down-
link and uplink rates in terms of the frame structure being employed. See Section IV.
The key finding of this work is that the achievable throughput in a massive MIMO system
depends heavily on the frame structure being used, and it is suboptimal to train all the users at
once, even in the presence of high switching costs. Also, the optimal frame duration as well as
the fraction of users to be trained per frame is a function of both the number of users and the
user mobilities. This leads to the proposed flexible frame structure yielding higher throughputs,
in comparison to the more conventional lumped or interspersed frames. In the next section, we
discuss the system model for this work.
II. S YSTEM M ODEL
We consider a single cell system comprising a base station with N antennas, serving K(< N )
single antenna mobile users, as depicted in Fig. 1. We assume the channels between the BS and
the users to be reciprocal [2], and the system to operate in TDD mode. Each frame consists of
multiple service blocks, as also shown in Fig. 1. Each block further consists of three sub blocks,
viz. training, uplink transmission, and downlink transmission. During the training sub-block, L
out of the K users are scheduled for training, and transmit L orthogonal training symbols to
the base station over L channel uses. The base station uses these training symbols to update the
available channel estimates for these L users. Following this, all the K users share their uplink
data during the next λL channel uses, decoded by the BS using the available channel estimates.
The BS and the K users then switch from uplink transmission to downlink transmission, incurring
a fixed cost equivalent to c channel uses. The system then operates in the downlink mode for
the next µL channel uses. During this time, the base station transmits K appropriately precoded
independent data streams to the K users. At the end of each service block, the system again
switches from downlink to uplink transmission, incurring a cost equivalent to c channel uses. The
overall average duration of one frame, T can therefore be expressed as T = (1+µ+λ)K+2 K

L
c,
and that of one service block can be expressed as TL = (1 + µ + λ)L + 2c,. This frame structure
reduces to a lumped training frame for L = K, and an interspersed training frame for L = 1.
The channel between the ith BS antenna and the kth user is assumed to evolve with time due
√
to user mobility and phase noise [11], and is modeled at the nth instant as β k hik [n], with βk
as the path loss component, and hik [n] as the fast fading component modeled as a zero mean
circularly symmetric complex Gaussian (ZMCSCG) random variable (rv) with unit variance,

4
Fig. 1. The system model.
denoted as CN (0, 1). We assume that the relative distance between the BS antennas is much
smaller as compared to the distance between the BS and the users, and consequently, βk remains
same across all the base station antennas. Letting β , [β1 , . . . , βK ]T , the channel matrix between
√
the BS and the users at the nth instant can be expressed as H[n]diag( β), with H[n] ∈ CN ×K ,
and its (ik)th entry representing the ZMCSCG channel between the ith BS antenna and the
kth user. The channel evolution is modeled as a linear combination of its initial state and an
innovation component as [8], [18]
hik [n] = ρ[n]hik [0] + ρ̄[n]zik [n] (1)
with zik [n] ∼ CN (0, 1) being the innovation component, and 0 ≤ ρ[n] ≤ 1 such that ρ[τ ] =
p
E[hik [m]h∗ik [m−τ ]]. Also, for any variable x, x̄ , 1 − |x|2 . The channel innovation processes,
zik [n] is not white, and is assumed to be stationary in the wide sense. By convention, it is assumed
that the channel evolves according to the Jakes’ model [8], [19], and the channel correlation
coefficient, ρ[n] is modeled as ρ[n] = J0 (2πfd Ts n), where fd is the Doppler frequency, Ts is
the sampling period, and J0 (·) is the Bessel function of the first kind and zeroth order [20, Eq.
(9.1.18)].
We assume that the channel estimates at the BS are acquired and updated using a Kalman
filtering based technique [14]–[16]. We denote the set of channel indices being used for training
in the mth frame by Im , the training instant for the kth user as tk,m ∈ Im , the estimate of the

5
channel vector between the BS and the kth user obtained at tk,m as ĥk,m , and the mean squared
channel estimation error for ĥk,m as ψk,m . It can then be shown that the channel hk [n] can be
expressed in terms of ĥk,m as [16]
(1/2)
hk [n] = (1 − ψk,m [n])(1/2) ĥk,m + ψk,m [n]h̃k,m [n], (2)
with ψk,m [n] , |ρ[n − tk,m ]|ψk,m + |ρ̄[n − tk,m ]|2 , and h̃k,m [n] being the zero mean channel
estimation error uncorrelated with ĥk,m . We do not list the Kalman filter update equations for
N0
the sake of brevity. However, it is easy to show that ψk,m is upper bounded as ψk,m ≤ βk Ep,k +N0
with Ep,k being the pilot power for the kth user. We next use this bound to obtain achievable
uplink and downlink rates for the proposed system.
III. ACHIEVABLE U PLINK AND D OWNLINK R ATES
In this section, we derive the achievable uplink and downlink rates for the given system
assuming maximal ratio combining (MRC) in the uplink, and matched filter precoding (MFP) in
the downlink. During the uplink sub-frame, all the K users simultaneously transmit their signals
to the base station, such that the received signal at the nth instant is given as [13],
K
X p
p
y[n] = βk Es,u,k hk [n]su,k + N0 w[n], (3)
k=1
where sk [n] is the symbol being transmitted by the kth user at the nth instant, and Es,u,k is
the uplink signal power of the kth user. It has been argued in the literature [2], that channel
hardening results in the elimination of fast fading effects, thus making the effective downlink
channel coefficient proportional to the slow fading coefficient βk . Therefore, recent works [2],
[13] advocate the use of channel inversion based power control in massive MIMO, such that
βk Es,u,k = Es,u , and βk Ep,k = Ep ,
Using the most recent channel estimate for the kth user’s channel, ĥk,m , for combining, the
combined signal at the base station can be expanded as
p
vk [n] = (1 − ψk,m [n])1/2 βk Es,u,k E[ĥH
k,m ĥk,m ]su,k [n]
p
+ (1 − ψk,m [n])1/2 βk Es,u,k (E[ĥH H
k,m ĥk,m ] − ĥk,m ĥk,m )su,k [n]
q
+ ψk,m [n]βk Es,u,k ĥH k,m h̃k,m [n]su,k [n]

6
K
X p
p
+ βl Es,u,l ĥH
k,m hl [n]su,l [n] + N0 ĥH
k,m w[n]. (4)
l=1
l6=k
In the above, the first term corresponds to the desired signal, the second to the interference due
to the difference between the true uplink channel, and its mean value considered at the BS, the
third to the self interference due to the estimation error at the BS, the fourth to the inter user
interference, and the last to the additive noise. It has been argued in [21] that in the worst case,
all the interference terms can be treated as AWGN. Treating interference as noise, assuming
channel inversion based uplink pilot and uplink/downlink data power control [2], [13] the uplink
SINR for the kth user can be obtained via the use and forget bounds detailed in [2] as
N (1 − ψk,m [n])
γu,k [n] = . (5)
K + ENs,u0
Since (1−ψk,m [n]) can be lower bounded as |ρ[n−tk,m ]|2 Ep E+N

p
0
, and defining γu [0] = N Ep Es
(Ep +N0 )(K+N0 )
,
we can lower bound the per user achievable uplink rate as
X
Ru ≤ log2 (1 + |ρ[n − tk,m ]|2 γu [0]). (6)
n∈Tu
where Tu is the set of time indices during which uplink data is transmitted, and is defined as
Tu = {mTL + L + 1, . . . , mTL + (λ + 1)L}, m ∈ Z. This can be maximized by maximizing
γu [0] in terms of the pilot and signal powers, with a constraint on the per user average received
power [12].
Similarly, defining the set of time indices for downlink transmission, Td , as Td = {mTL +
(λ + 1)L + c + 1, . . . , (m + 1)TL − c} m ∈ Z. The downlink signal transmitted by the BS at the
nth (n ∈ Td ) instant is
K
X
ĥ∗k,m [n] Ed,s,k sk [n]
p
x[n] = λ (7)
k=1
with Ed,s,k being the transmit power allocated to the kth user, and λ being the power normalization
factor at the BS. The power normalization factor λ used to restrict the total transmitted power
by the BS, such that [2],
Et
λ2 = , (8)
E[Tr{ĤH diag(E d )Ĥ}]
where, Et is the total power radiated by the BS, and the vector E d is defined as E d , [E d,s,1 , . . . , E d,s,K ].

7
Using channel hardening arguments [22], it is easy to show that λ reduces to,
Et
λ2 = PK . (9)
N l=1 Es,d,l
Et
Defining, Fs,d,k , PK
Es,d,l
,
l=1
the signal received at the kth user at the nth instant, rk [n], can be expanded in the form
r
Fs,d,k
rk [n] = (1 − ψk,m [n])βk E[ĥTk,m ĥ∗k,m ]sk [n]
N
r
Fs,d,k T ∗
+ (1 − ψk,m [n])βk (ĥk,m ĥk,m − E[ĥTk,m ĥ∗k,m ])sk [n]
N
r
βk Fs,d,k
+ ψk,m [n]h̃Tk,m [n]ĥ∗k,m sk [n]
N
K
r
X βk Fs,d,l T p
+ hk [n]ĥ∗l,m sl [n] + N0 wk [n]. (10)
l=1
N
l6=k
In the above, the first term corresponds to the desired signal, the second to the interference
due to the difference between the true effective downlink channel and its mean, the third to
the interference due to the estimation error of the true channel vector at the BS, the fourth to
the inter user interference, and the last to the AWGN. Again employing use and forget analysis
N0
from [2], and upper bounding the channel estimation error as ψk,m = Ep +N0
, the effective SINR
for the kth user at the nth instant can be reduced to γd,k [n] = |ρ[n − tk,m ]|2 γd,k [0] where
N βk Fs,d,k Ep E+N
p
0
γd,k [0] = PK . (11)
βk l=1 Fs,d,l + N0
Further, the achievable downlink rate of the kth user is obtained as
X
Rd,k = log2 (1 + γd,k [n]). (12)
n∈Td
Now, based on the frame structure, it can be observed that the channels available for the
downlink will, in general be aged more as compared to the channels available for uplink. This
is in order to facilitate communication for the battery powered users, as opposed to the centrally
powered base station. Also, by virtue of the frame structure, the number of ‘less aged’ channel
slots available for uplink transmission will be an increasing function of L, and therefore the
achievable uplink rate will be maximized for L = K. However, no such observation can be made

8
about the achievable down-link rates. Therefore, in the next section, we numerically evaluate the
achievable downlink rates, and optimize these in terms of L.
IV. N UMERICAL R ESULTS
For the purpose of these experiments, we consider a single cell massive MIMO system
containing a 1000 (N) antenna base station, transmitting at a carrier frequency (fc ) of 2 GHz.
We assume that the BS serves a maximum of K = 200 users, uniformly distributed over a
circular cell with radius 5 km. The reference distance (d0 ) in this case is assumed to be 1 km,
and the path loss coefficient (βk ) for the kth user, located at a distance dk from the B.S. is
calculated as −η
dk
βk = , (13)
d0
with η being the propagation constant. In this setup we assume η = 4. Also, similar to other past
works on channel agingin massive MIMO systems [9], [11], [13], we consider a narrowband
channel, with a signal bandwidth of 1 MHz. We assume that the base station samples at the
Nyquist rate of the complex baseband signal, i.e., at 1 MHz.
L
Fig. 2 plots the achievable downlink rates against the fraction K
for different user velocities
and for different values of λ = µ = a with no switching cost, i.e. for c = 0. It can be observed
that at low user mobilities the achievable uplink rates are independent of the chosen value of
L; however, the achievable rate becomes highly sensitive to the choice of L. It is also observed
that longer frame durations are preferable for scenarios with low user mobilities, and vice versa,
which is in accordance with intuition. It is also clear that the optimal value of L depends on the
overall frame length, as well as the user mobility, and needs to be determined numerically.
In Figs. 3 and 4, we plot the achievable uplink rates against the number of users (K) for
different user mobilities for λ = µ = 2. In these figures we also compare the achievable rates
for lumped, and interspersed training schemes against the schemes corresponding to the optimal
value of the parameter L for different switching costs. It is observed that for a large number of
users at high user mobilities, it is preferable to use interspersed training as compared to lumped
training when no switching cost is involved, which is as per the expectations. However, lumped
training is preferable when the switching cost is high.
In Fig. 5 we plot the achievable per user uplink rates for different cases against the number
of users at a user velocity of 150 km/h. As discussed earlier, the uplink rates are maximized

9
1.2

0.8
0.6
0.4
a=1
a=2
0.2 a=3 Solid: v=10km/h
a=4 Dashed: v=150 km/h
a=5
0
0 0.2 0.4 0.6 0.8 1
Fig. 2. Achievable per user rates at different user mobilities for different frame durations and different values of L.
for L = K. However, the choice of a value of L, optimized for downlink transmission does not
affect the achievable uplink rates significantly for either c = 0, or c = 1.
In Table I, we tabulate the percentage gains achieved by using the proposed frame structure
over the conventional lumped frame structure for 200 users and different switching costs for
different user velocities. It is observed that the performance of a massive MIMO system can
be increased by as much as 38% by using the proposed frame structure instead of the more
conventional lumped structure. It is also observed that the impact of an increased switching cost
on these gains is minimal.
TABLE I
G AINS OVER THE CONVENTIONAL LUMPED TRAINING STRUCTURE
Switching Cost 0 0.25 0.5 0.75 1 1.5 2 5

Percentage Gain 37.93% 37.77% 37.6% 37.58% 37.46% 37.16% 36.68% 37.56%

10
3
v=10 km/h
v=100 km/h
2.5 v=200 km/h

2
1.5
Solid: Optimal
0.5
0
50 100 150 200
Number of Users (K)
Fig. 3. Achievable downlink rates for different numbers of users under different mobilities with no switching cost(c=0).
V. C ONCLUSIONS
In this paper, we considered a generalized frame structure for TDD massive MIMO systems.
It was argued that the more traditional interspersed and lumped frame structures are special
cases of the proposed frame. We derived bounds on the achievable uplink and downlink rates
considering MRC and MFP respectively. Finally, using numerical results, we illustrated that, in
general, the achievable downlink rates depend on the number of users being trained per service
block, and the length of a service block. It was also shown that the proposed block structure can
be used to fine tune these parameters to optimize the frame rates for given channel and mobility
conditions, resulting in a gain of more than 35% under high user mobilities.
R EFERENCES
[1] T. Narasimhan and A. Chockalingam, “Channel hardening-exploiting message passing (CHEMP) receiver in large-scale
MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 8, pp. 847–860, Oct. 2014.
[2] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals of Massive MIMO. Cambridge University Press,
Cambridge, UK, 2016.
[3] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL of cellular networks: How many antennas do we
need?,” IEEE J. Sel. Areas Commun., vol. 31, pp. 160–171, Feb. 2013.

11
3
v=10 km/h
v=100 km/h
2.5 v=200 km/h
Achivable Rate Per User

2
1.5
0.5 Solid: Optimal

0
50 100 150 200
Number of Users (K)
Fig. 4. Achievable downlink rates for different numbers of users under different mobilities with c=1.
[4] S. Blandino, G. Mangraviti, C. Desset, A. Bourdoux, P. Wambacq, and S. Pollin, “Multi-user hybrid MIMO at 60 GHz
using 16-antenna transmitters,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 66, pp. 848–858, Feb.
2019.
[5] M. Rahman and J.-D. Park, “The smallest form factor UWB antenna with quintuple rejection bands for iot applications
utilizing RSRR and RCSRR,” Sensors, vol. 18, p. 911, Feb. 2018.
[6] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach to channel estimation in large-scale multiple-antenna
systems,” IEEE J. Sel. Areas Commun., vol. 31, pp. 264–273, Feb. 2013.
[7] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot contamination and precoding in multi-cell TDD systems,”
IEEE Trans. Wireless Commun., vol. 10, pp. 2640–2651, Aug. 2011.
[8] K. T. Truong and R. W. Heath, “Effects of channel aging in massive MIMO systems,” Journal of Communications and
Networks, vol. 15, pp. 338–351, Aug. 2013.
[9] A. K. Papazafeiropoulos and T. Ratnarajah, “Deterministic equivalent performance analysis of time-varying massive MIMO
systems,” IEEE Trans. Wireless Commun., vol. 14, pp. 5795–5809, Oct. 2015.
[10] G. Amarasuriya and H. V. Poor, “Impact of channel aging in multi-way relay networks with massive MIMO,” in Proc.
IEEE Intl. Conf. Commun. (ICC 2015), (London, UK), pp. 1951–1957, June 2015.
[11] A. K. Papazafeiropoulos, “Impact of general channel aging conditions on the downlink performance of massive MIMO,”
IEEE Trans. Veh. Technol., vol. 66, pp. 1428–1444, Feb. 2016.
[12] R. Chopra, C. R. Murthy, and H. A. Suraweera, “On the throughput of large MIMO beamforming systems with channel
aging,” IEEE Signal Process. Lett., vol. 23, pp. 1523–1527, Nov. 2016.
[13] R. Chopra, C. Murthy, H. Suraweera, and E. Larsson, “Performance analysis of FDD massive MIMO systems under

12
3
Lumped Training
Achivable Per user uplink rate (bps)

2.5 Optimal for downlink c=0
Optimal for downlink c=1
1.5
0.5
0 50 100 150 200
Number of users (K)
Fig. 5. Achievable uplink rates for different numbers of users under different configurations for a user velocity of 150 km/h.
channel aging,” IEEE Trans. Wireless Commun., vol. 17, pp. 1094–1108, Feb. 2018.
[14] S. Kashyap, C. Mollén, E. Björnson, and E. G. Larsson, “Performance analysis of TDD massive MIMO with Kalman
channel predication,” in Proc. Intl. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2017), (New Orleans, LA),
pp. 3554–3558, March 2017.
[15] V. Arya and K. Appaiah, “Kalman filter based tracking for channel aging in massive MIMO systems,” in 2018 International
Conference on Signal Processing and Communications (SPCOM) (SPCOM 2018), (Indian Institute of Science, Bangalore,
India), July 2018.
[16] R. Chopra, “Uplink training for massive MIMO systems under channel aging,” in 2018 International Conference on Signal
Processing and Communications (SPCOM) (SPCOM 2018), (Indian Institute of Science, Bangalore, India), July 2018.
[17] L. Chen, S. Iellamo, and M. Coupechoux, “Opportunistic spectrum access with channel switching cost for cognitive radio
networks,” in 2011 IEEE International Conference on Communications (ICC), pp. 1–5, June 2011.
[18] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors, and F. Tufvesson, “Scaling up MIMO:
Opportunities and challenges with very large arrays,” IEEE Signal Process. Mag., vol. 30, pp. 40–60, Jan. 2013.
[19] W. C. Jakes and D. C. Cox, eds., Microwave Mobile Communications.
[20] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables.
New York: Dover, 9th ed., 1964.
[21] B. Hassibi and B. M. Hochwald, “How much training is needed in multiple-antenna wireless links?,” IEEE Trans. Inf.
Theory, vol. 49, pp. 951–963, Apr. 2003.
[22] R. Couillet and M. Debbah, Random matrix methods for wireless communications. Cambridge University Press, Cambridge,
UK, 1st ed., 2011.

Wire D 19 00453 - R1

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wire D 19 00453 - R1

Uploaded by

Copyright:

Available Formats

Wireless Personal Communications

Frame Structures for Massive MIMO Communications under Channel Aging

Manuscript Number: WIRE-D-19-00453R1

Article Type: Short Communication

Keywords: Massive MIMO, CSI estimation, Channel aging, performance analysis

Corresponding Author: Ribhu Chopra

Corresponding Author Secondary

Corresponding Author's Institution: Indian Institute of Technology Guwahati

Corresponding Author's Secondary

First Author: Anubhab Chowdhury

First Author Secondary Information:

Order of Authors: Anubhab Chowdhury

Order of Authors Secondary Information:

Funding Information: Indian Institute of Technology Guwahati Dr. Ribhu Chopra

Response to Review Comments on

April 26, 2019 DRAFT

R ESPONSE TO THE C OMMENTS F ROM R EVIEWER 1

Comment 1.3: Please correct the references pointers....everywhere it is a question mark....

size within single plot.

are thoroughly based on simulations, so its important to provide proper background....

April 26, 2019 DRAFT

in a separate paragraph at the beginning of the section on numerical results.

April 26, 2019 DRAFT

Response to Review Comments on

April 30, 2019 DRAFT

R ESPONSE TO THE C OMMENTS F ROM R EVIEWER 1

Comment 1.3: Please correct the references pointers....everywhere it is a question mark....

size within single plot.

are thoroughly based on simulations, so its important to provide proper background....

April 30, 2019 DRAFT

in a separate paragraph at the beginning of the section on numerical results.

April 30, 2019 DRAFT

Frame Structures for Massive MIMO

Massive MIMO, CSI estimation, Channel aging, performance analysis

April 26, 2019 DRAFT

April 26, 2019 DRAFT

II. SYSTEM MODEL

duration of one frame, T can therefore be expressed as and that of

April 26, 2019 DRAFT

Fig. 1. The system model.

hik[n] = ρ[n]hik[0] + ¯ρ[n]zik[n] (1)

April 26, 2019 DRAFT

expressed in terms of hˆk,m as [16]

III. ACHIEVABLE UPLINK AND DOWNLINK RATES

April 26, 2019 DRAFT

combined signal at the base station can be expanded as

Since (1−ψk,m[n]) can be lower bounded as , and defining ,

Ru ≤ X log2(1 + |ρ[n − tk,m]|2γu[0]). (6)

{mTL + L + 1,...,mTL + (λ + 1)L}, m ∈ Z. This can be maximized by maximizing

April 26, 2019 DRAFT

(λ+1)L+c+1,...,(m+1)TL −c} m ∈ Z. The downlink signal transmitted by the BS at the

nth (n ∈ Td) instant is

April 26, 2019 DRAFT

Rd,k = X log2(1 + γd,k[n]). (12)

achievable downlink rates, and optimize these in terms of L.

IV. NUMERICAL RESULTS

April 26, 2019 DRAFT

April 26, 2019 DRAFT

Fraction of users trained per block (L/K)

Switching Cost 0 0.25 0.5 0.75 1 1.5 2 5

April 26, 2019 DRAFT

Number of Users (K)

April 26, 2019 DRAFT