7174 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO.
10, OCTOBER 2016
Beyond One-Way Communication: Degrees of
Freedom of Multi-Way Relay MIMO
Interference Networks
Chenwei Wang, Member, IEEE
Abstract— We characterize the degrees of freedom (DoFs) of
a multiway relay wireless network consisting of four user nodes,
each with M antennas, and one N-antenna relay node. Each user
node sends one independent message to each of the other users,
and there are no direct links between any two user nodes. For
this network, we show that the symmetric DoF value per mes-
sage is given by max(min((M/3), (N/7)), min(((2M)/7), (N/6)))
normalized by space dimensions, i.e., piecewise linear depending
on M and N alternatively. While the information theoretic
DoF upper bound is established for every M and N antennas,
the achievability relying on linear signal subspace alignment
is established in the spatially normalized sense in general.
In addition, by deactivating four messages to form a two-
way relay MIMO X channel, we also present the DoF result
max(min((M/2), (N/5)), min(((2M)/5), (N/4))) in a similar
piecewise linear type. The key ingredient to these new results
is to first establish the principle of beamforming design via
linear dimension counting perspective, which is essentially the
fundamental element for establishing both the DoF converse and
the DoF achievability. Moreover, we also settle the feasibility
of linear interference alignment that extends the feasibility
framework from one-way one-hop to multiway relay interference
networks.
Index Terms— Multi-way MIMO, DoF, signal alignment.
I. I NTRODUCTION
E NLIGHTENED by interference alignment, recently the
fundamental capacity limit, particularly the the degrees
of freedom (DoF) of multiuser wireless networks, have been
widely investigated [1]. So far, the DoF characterizations
available are almost exclusively for one-way networks, i.e.,
each node either sends or demands messages, but not both.
If users are allowed to both send and demand messages to
form a class of communication networks are referred to as
multi-way networks. then channel capacity, or even the DoF
characterizations, remain widely open.
For multi-way networks, recently Lee et al. studied the DoF
of a 3-user MIMO Y channel in [4] where each user equipped
Manuscript received January 26, 2016; revised June 3, 2016; accepted
July 22, 2016. Date of publication August 4, 2016; date of current version
October 7, 2016. This paper was presented at the IEEE International Sympo-
sium on Information Theory, Honolulu, HI, USA, July 2014. The associate
editor coordinating the review of this paper and approving it for publication
was H. Suraweera.
The author is with the Mobile Network Technology Group,
DOCOMO Innovations Inc., Palo Alto, CA 94304 USA (e-mail:
cwang@docomoinnovations.com).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TWC.2016.2598330
1536-1276 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
with 2 antennas sends one independent message to each of
the other users via the help of a 3-antenna relay only, and
there are no direct links between any two users. While their
scheme can achieve the single-user DoF upper bound, the most
interesting element is their new idea of alignment for network
coding. Unlike a conventional relay which simply amplifies
or decodes and forwards its received signal, they designed a
linear beamforming scheme, so that every pair of two symbols,
i.e., the symbols of two users, each sending to and demanding
from the other, are aligned in the same dimension at the relay,
and then the relay is able to resolve the symbol summation
along each dimension. Afterwards, they applied a reciprocal
beamforming design at the relay and each user, so that finally
each user only sees two parallel interference-free point-to-
point channels after removing its self-interference. Thus, the
phrase of alignment for network coding essentially means
signal alignment at the relay and network coding at users for
decoding. The central idea in [4] can be summarised as the
principle of one-to-one vector alignment, which means that
every pair of two symbols can be perfectly aligned along the
same dimension at the relay. Naturally, we are interested in
the question whether it is a universal principle to the setting
with more than 3 users and even broader types of multi-way
relay networks. To simplify the problem as much as possible,
we consider a symmetric but general setting where there are
4 M-antenna users, each user sends one independent message
to each of the others via an N-antenna relay only, as shown
in Fig. 1, and M and N can take arbitrary positive integers.
Compared to the original 3-user MIMO Y channel, although
only one additional user is added, completely characterizing
the DoF of this network becomes quite challenging, due to the
following coupled issues. First, since M, N can take arbitrary
values, one-to-one alignment might be impossible, and we
have to resort to one-to-many alignment where one symbol
lies in the subspace jointly spanned by many symbols. While
one-to-many alignment schemes have been widely studied for
one-way one-hop networks [1], very little is known for multi-
way relay networks. Second, can we find an achievable scheme
which is information theoretically optimal? The difficulties of
DoF characterizations come from both finding tight DoF upper
bounds and coming up with clever DoF achievable schemes.
Finally, while the single-user bound and the cut-set bound are
sufficient to capture the DoF upper bound for K = 3 user Y
channel [10], it is not clear if they are sufficient for K ≥ 4
cases.
WANG: DoFs OF MULTIWAY RELAY MIMO INTERFERENCE NETWORKS
Of the vast amount of literature on beyond one-way relay
interference networks, the most closely related to this work
are references [4]–[12], [17], [21]. In particular, following
the idea in [4], the DoF of several multi-way relay networks
were studied in [5]–[7]. However, the authors only tackled
networks with special number of antennas, so that every pair
of two symbols can be aligned along the same vector at
the relay. Thus, their results, requiring one-to-one alignment
only, are established in a relatively straightforward manner.
In [10]–[12], the authors studied the DoF of two more general
multi-way relay MIMO networks based on, again, one-to-one
alignment only, and either their DoF converse or the DoF
achievability is not tight in general, thus leaving their exact
DoF characterizations still open in general. Beyond one-to-
one alignment, Wang et al. studied the DoF of the 2-pair and
3-pair two-way relay MIMO interference channel [8]. Com-
pared to [8], in this work we also need one-to-many alignment.
However, since the total number of messages increases, how
to identify the interplay between the subspaces projected from
each user at the relay becomes difficult. In [17], the authors
proposed the group alignment concept for the K -user MIMO
Y channel to study a tradeoff between the number of antennas
at the users and the relay. In [21], a more complicated system
model involving user clusters was studied. However, again,
their DoF achievability result is optimal for only a special
set of cases. Even we boil down their configurations to 4-user
MIMO Y channel, the DoF characterization is still incomplete.
Recently in [9] Yuan et al. showed an achievable DoF result of
the 4-user MIMO Y channel defined in this work, essentially
based on the idea of inter-pair signal subspace alignment
in [8]. However, their result is the DoF achievability only,
and they did not show if there exists any other more efficient
scheme to achieve more DoF, or if their DoF achievability
is optimal either by using linear beamforming schemes or in
the information theoretic sense, thus leaving this problem still
open in general.
In this paper, we show that the symmetric DoF value
per message is piecewise linear depending on M and N
alternatively. To establish this result, we provide both the
information theoretic DoF converse and the DoF achievability.
While our DoF converse is established for every M and N, the
DoF achievability relying on linear signal vector alignment is
established in the spatially-normalized sense in general, and
the key ingredient to this new result is to first establish the
principle of beamforming design for the transmission from
the users to the relay, and then designing a novel approach to
construct an efficient achievable scheme based on inter-user
signal subspace alignment for network coding. We remind the
reader that the similar observations were also illustrated in [2]
for the 3-user MIMO interference channel. However, as we
will explain in detail later, the ideas behind both the DoF
converse and the DoF achievability appear to be quite different.
Compared to the recent work [9] where only the achieved DoF
are presented, we provide both the information theoretic DoF
converse and the DoF achievability. In particular, we show
that the result in [9] is not tight when 38 ≤ M
N ≤ 2 as shown
1
in Fig. 2, by identifying a gap between their achieved DoF
and the upper bound developed in this work, and further
Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
7175
closing the gap with a new achievable scheme. For exam-
ple, consider the most interesting case (M, N) = (3, 7). It
was shown in [9] that each message can achieve 7/8 DoF.
However, we show in this paper that each message can achieve
1 DoF, which are also the information theoretically opti-
mal. Moreover, as a byproduct of this work, the DoF of
a two-way relay MIMO X channels are also studied, and
we present the DoF result in the similar piecewise linear
type.
We believe that our contribution is interesting and substan-
tial for three reasons.
• First, as mentioned earlier in this section, the open
problem studied in this paper has attracted much atten-
tion these years, and finally it is resolved in this
work.
• Second, for both the DoF converse and the achievability,
we essentially translate the original network to a one-
way one-hop channel, from the users to the relay, by
imposing additional decoding constraints at the relay.
Contrary to conventional one-hop channels where every
message is either undesired or desired but not both at
a receiver, in this work, since all communication must
pass through the relay, every message is both desired
(to its desired decoder) and undesired (to all the other
unintended decoders) at the relay. Due to the imposed
additional decoding constraints at the relay, we clearly
put forward the principle of beamforming design, which
provides a lens to study the DoF of multi-way relay
networks.
• Finally, for one-way one-hop MIMO interference net-
works, the topic feasibility of linear interference align-
ment, i.e., the feasibility of linear beamforming schemes
through interference alignment at the receivers, is one
of the most fundamental questions in information theory
and has recently drawn upon much attention. As all
existing works on the feasibility of linear alignment are
for one-way one-hop wireless networks, we settle the
feasibility of linear interference alignment for two multi-
way relay MIMO interference networks. To the best of
our knowledge, this is the first nontrivial work to establish
the feasibility of linear interference alignment for the
network beyond one-way one-hop.
II. S YSTEM M ODEL
Consider a wireless network where there are 4 user nodes,
each with M antennas, and one relay node with N antennas.
As shown in Fig. 1, each user k ∈ {1, 2, 3, 4}  K sends one
independent message Wkj , j ∈ K \ {k} to each of the other
3 users via the help of the relay node only. We denote by Hk
the N × M channel matrix from user k to the relay, and H̄k
the M × N channel matrix from the relay to user k. Moreover,
we assume that the channel coefficients are independently
drawn from continuous distributions, and stay constant during
the entire transmission once they are drawn. We also assume
that global channel knowledge is available at every node in
the network. In this work, we assume that all the 5 nodes
work in the full-duplex mode, i.e., they can hear and transmit
7176
Fig. 1. The system model.
simultaneously (If all nodes work in the half-duplex mode, the
DoF results will be scaled by a factor 1/2 due to normalization
to time.).
Since the relay hears from all the users, its received signal
vector at time t ∈ Z+ is given by:
 and the conjugate transpose of the matrix A, respectively.
Y R (t) = Hk X k (t) + Z R (t) (1)
k∈K
where X k (t) is the complex-valued M × 1 signal vector sent
from user k, the N × 1 column vector Z R (t) represents the
i.i.d. circularly symmetric complex additive white Gaussian
noise (AWGN) at the relay and Z R (t) ∼ CN (0, I N ). At the
user node, each user hears from the relay only. Thus, the
received signal vector at user k at time t is given by:
Yk (t) = H̄k X R (t) + Z k (t)
where X R (t) is the complex-valued N × 1 signal vector sent
from the relay, the M × 1 column vector Z k (t) represents the
AWGN and Z k (t) ∼ CN (0, I M ). In addition, the transmitted
signal vectors from all nodes i ∈ K ∪ {R} satisfy the average
power constraint T1 tT=1 E[X i (t)2 ] ≤ P for T channel
uses. We will study the DoF of the following two settings:
• All the 12 messages are active to form a 4-user relay
MIMO Y channel, which is a natural extension of the
3-user relay MIMO Y channel. We refer to it as the all
unicast setting.
• Among the 12 messages, let W12 = W21 = W34 =
W43 = ∅ to form a two-way relay MIMO X channel
where every user on both the left and the right sides of
Fig. 1 sends one independent message to every user on the
other side. We refer to it as the multiple unicast setting.
We denote by R(M, N) = Rkj (M, N) the symmetric rate
of the message Wkj and the corresponding symmetric DoF
metric is denoted as d(M, N). The rate and the DoF definitions
follow from their standard definitions in information theory.
In addition, the definition of the spatially normalized DoF
metric, to avoid special channel structures and to keep generic
channels for the DoF achievability, is introduced in [2], where
it is defined as d ∗ (M, N) = maxq∈Z+ d(q M,qq
N)
. Intuitively, it
means that we scale the number of antennas at each node by a
factor q so that the resulting DoF value q ·d is an integer, much
like q symbol extensions over the time/frequnecy domains.
Based on all available results so far, regarding the DoF
value, DoF normalization to spatial extensions is similar to
normalization to time/frequency.
Notations: We use a, A and A, Im , O to denote a scalar,
a column vector, a matrix, the m × m identity matrix and the
Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 10, OCTOBER 2016
Fig. 2. DoF per message for the 4-user relay MIMO Y channel (The double
dashed black curve represents the DoF achievability in [9], and the solid blue
curve represents the new DoF result in this work.)
zero matrix, respectively. Also, AT , A H stand for the transpose
In addition, we denote by A(m : n) the sub-column vector
whose entries are picked from the m t h to the n t h entries of
the vector A sequentially. Moreover, we use (x) to represent
any function so that lim x→∞ (x)/x = 0.
III. M AIN R ESULTS
We state our main DoF results and illustrate the main
insights behind the results in this section.
(2)
A. All Unicast: The 4-User Relay MIMO Y Channel
Definition 1: Define the quantity dY∗ =
max(min( M3 , 7 ), min( 7 , 6 )),
N 2M N
or equivalently
⎧
⎪
⎪ M/3, 0 < M/N ≤ 3/7,
⎨
N/7, 3/7 < M/N ≤ 1/2,
dY∗ =
⎪
⎪ 2M/7, 1/2 < M/N ≤ 7/12,
⎩
N/6, 7/12 < M/N.
Lemma 1 (DoF Converse): For the 4-user relay MIMO Y
channel defined in Section II, the information theoretic DoF
per message are upper bounded by d ≤ dY∗ .
Proof: The proof is presented in Section IV.
Lemma 2 (DoF Achievability): For the 4-user relay MIMO
Y channel defined in Section II, each message can achieve dY∗
spatially normalized DoF almost surely.
Proof: The proof is presented in Section V.
Theorem 1: For the 4-user relay MIMO Y channel defined
in Section II, each message has dY∗ spatially normalized DoF.
Proof: The proof follows directly from Lemma 1 and
Lemma 2.
We plot the DoF result implied by Theorem 1 in Fig. 2.
It can be seen that the DoF value per message is piecewise
linear depending on M and N alternatively, which implies that
there are antenna dimension redundancies at either each user
or the relay node. Recall that recently the similar observation
was also made for the 3-user M × N MIMO interference
channel by Wang et al. in [2]. However, the tools used
to obtain these results appear to be quite different. In [2],
the DoF achievability based on the new idea of subspace
alignment chains, only needs one-to-one alignment, and the
length of subspace alignment chains depends on the values of
WANG: DoFs OF MULTIWAY RELAY MIMO INTERFERENCE NETWORKS 7177

M, N only. In contrast, in this work, we have to use one-to- linear interference alignment based on only spatial beam-
many alignment, and thus there does not exist the concept of forming. The feasibility problem was first introduced by
subspace alignment chains. For example, when M N increases Gomadam et al. in [18], and then Yetis et al. in [3] draw upon
from 37 to 12 7
, the type of the achievable scheme changes classical results in algebraic geometry about the solvability of
from one-to-many alignment to one-to-one alignment. Thus, generic polynomial alignment equations, to classify a system
the most challenge is how to design an efficient one-to-many as proper if and only if the number of independent variables in
alignment scheme to achieve the DoF upper bound. Similar every set of equations is at least as large as the number of equa-
to [2], we conclude the antenna redundancies observations. tions in that set. If the involved equations are generic, Yetis
Specifically, if and only if MN = 2 , both M and N include
1 et al. appealed to the intuition that proper systems are likely
redundant antenna dimensions. If and only if M N ∈ { 7 , 12 },
3 7 to be feasible and improper systems to be infeasible. While
neither M nor N contains any redundant dimensions. Thus, the conjectured correspondence between proper/improper and
intuitively the DoF converse originates from M N = 2 (black
1 feasible/infeasible has recently attracted much attention for
circles) and N ≥ 12 , and the DoF achievability originates
M 7 one-way one-hop MIMO interference channels such as [2],
[19], [20], the feasibility of linear interference alignment has
from M N = 7 , 12 (green circles). Once we have the proofs at
3 7

these points, we can use the similar idea with additional efforts never been fully investigated if the channel is not restricted to
to solve the rest between every two adjacent transition points. one-way and one-hop.
In this paper, we also settle the issue of feasibil-
ity/infeasibility of linear interference alignment for the two
B. Multiple Unicast: The Two-Way Relay MIMO X Channel multi-way channel models defined in this paper. To the best
Definition 2: Define the quantity d X∗ = of our knowledge, this is the first time that the feasibility
max(min( M2 , 5 ), min( 5 , 4 )),
N 2M N framework is extended to interference networks beyond one-
or equivalently
⎧ way one-hop. Specifically, we also plot the DoF counting
⎪
⎪ M/2, 0 < M/N ≤ 2/5, bounds in Fig. 2 represented by the red straight line. The
⎨
∗ N/5, 2/5 < M/N ≤ 1/2,
dX = DoF counting bound, first originated from the work by
⎪
⎪ 2M/5, 1/2 < M/N ≤ 5/8, Yetis et al. in [3] for one-way one-hop MIMO interference
⎩
N/4, 5/8 < M/N. channels, identifies the system into proper and improper. We
Lemma 3 (DoF Converse): For the two-way relay MIMO follow the similar approach to find the DoF counting bounds
X channel defined in Section II, the information theoretic DoF d ≤ 2M+N 13 for the all unicast setting and d ≤ 2M+N 9 for
per message are upper bounded by d ≤ d X∗ . the multiple unicast setting in Appendix A1 and A2. Since
Proof: Since the proof is similar to that for Lemma 1, we the observations of the two settings are similar, we consider
defer the proof into Appendix B. the former setting as an example. From Fig. 2, it turns out
Lemma 4 (DoF Achievability): For the two-way relay improper is infeasible, many proper systems are infeasible, and
MIMO X channel defined in Section II, each message can if the information theoretic DoF upper bound is lower than the
achieve d X∗ spatially normalized DoF almost surely. counting bound, linear schemes are sufficient to achieve the
Proof: Since the proof is similar to that for Lemma 2, we information theoretic DoF upper bound. Interestingly, these
defer the proof into Appendix C. observations, already available for many one-way one-hop
Corollary 1: For the two-way relay MIMO X channel MIMO channels such as [2], are verified to exist in the beyond
defined in Section II, each message has d X∗ spatially normal- one-hop network as well. In the following, we directly present
ized DoF. our new results.
Proof: The proof follows directly from Lemma 3 and Theorem 2: For the 4-user relay MIMO Y channel defined
Lemma 4. in Section II, the DoF demand per message, d, are feasible
Remark: Recently, this problem was also studied by with linear interference alignment if and only if d ≤ dY∗
Xiang et al. in [13] where they only demonstrated that the DoF.
sum DoF are upper bounded by 2 min(2M, N), i.e., the DoF Proof: The proof is presented in Section VI.
per message are upper bounded by min(2M,N)4 , and this bound Corollary 2: For the two-way relay MIMO X channel
can be achieved if M ≥ 5
. As implied by Corollary 1, their defined in Section II, the DoF demand per message, d, are
N 8
DoF converse only covers the first and last regimes, and their feasible with linear interference alignment if and only if
DoF achievability is for the last regime only. In this paper, d ≤ d X∗ DoF.
our DoF converse is for every M, N, and our DoF achievability Proof: Since the channel model is a special setting of the
is for every M, N as well but in the sense of spatial extensions. 4-user relay MIMO Y channel, the proof essentially follows
from the proof of Theorem 2, and thus we omit it in this paper.

C. Feasibility of Linear Interference Alignment

IV. DoF C ONVERSE : P ROOF OF L EMMA 1
Among the DoF studies of MIMO interference networks,
N ≤ 7 , since each user sends 3 messages
In the regime M 3
there is much interest in the DoF achievable through linear
beamforming schemes within only one channel use, i.e., and is equipped with M antennas, the DoF per message are
without symbol extensions over time/frequency/space. A cen- upper bounded by M/3. Thus, we only need to consider the
tral question in this research avenue is the feasibility of remaining regimes. Note that each user only hears from the

Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
7178
relay, thus the received signal at each user is a degraded
version of what the relay sends. Since each user is able to
decode its 3 desired messages with its 3 own messages as
side information, if a genie provides those side information to
the relay, then the relay is also able to decode the messages
desired at that user. By doing so, the DoF converse for the one-
hop multiple access channel (MAC), i.e., from the users to the
relay, is also the DoF converse of the original two-hop chan-
nel. Note that the translated one-hop MAC is fundamentally
different from the MAC that we usually deal with in one-way
communication, because here the relay can decode whatever
each user can decode provided the 3 messages originated
from that user are provided to the relay as side information.
Moreover, contrary to one-way channels where every message
at a receiver is either desired or interference but not both,
in this translated one-way MAC, since the decoder of every
user requires different side information, every message is both
desired and interference at the same relay, depending on which
decoder is of interest. In the following, for the translated one-
hop MAC with additional decoding constraints, we will first
use the linear dimension counting approach to find the linear
DoF upper bounds, and then translate them into information
theoretic DoF upper bounds. All the proofs will be built upon
the widely used genie-aided approach.
A. Linear Dimension Counting
1) M/N ≥ 7/12 ⇒ d ≤ N/6: Suppose
a genie provides to the relay the messages G =
{W12 , W13 , W14 , W32 , W42 , W34 }. Let us count the total num-
ber of dimensions occupied by the other 6 messages W21 ,
W31 , W41 , W23 , W24 , W43 . First, since user 1 can decode
W21 , W31 , W41 with its own messages W12 , W13 , W14 as side
information, which are also provided to the relay by the genie,
the relay is able to decode W21 , W31 , W41 as well. Thus, the
3 messages W21 , W31 , W41 contribute a total of 3d dimensions,
which have to be linearly independent with the interfering
dimensions occupied by the rest 3 messages W23 , W24 , W43 .
Next, note that W23 , W43 can be both decoded at user 3
where W31 , W32 , W34 are available as side information. Since
W32 , W34 are already provided by the genie at the relay,
and W31 can also be decoded first and thus available to the
relay, the relay is able to decode the messages W23 , W43
as well. Hence, W23 , W43 contribute other 2d dimensions.
Finally, consider the message W24 desired at user 4 where
W41 , W42 , W43 are available as side information. Again, W42
is already provided to the relay by the genie, and W41 , W43
can be first decoded and thus available at the relay as well.
Thus, W42 contributes another d dimensions. So far, all the
6 messages W21 , W31 , W41 , W23 , W24 , W43 contribute a total
of 3d + 2d + d = 6d, we have 6d ≤ N to produce the desired
upper bound d ≤ N/6.
Remark: Among all the non-genie messages W21 , W31 , W41 ,
W23 , W24 , W43 , it can be seen that every pair of messages
(Wi j , W j i ) with the same (i, j ), do not both appear. Intuitively,
to guarantee every message to be decoded at its own desired
receiver, we need to protect d dimensions for each (Wi j , W j i ).
Thus, all those 6 messages occupy a total of 6d dimensions.
Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 10, OCTOBER 2016
2) 3/7 ≤ M/N ≤ 1/2 ⇒ d ≤ N/7:
Suppose a genie provides to the relay the messages
G = {W12 , W13 , W14 , W42 , W43 }. Since user 1 can decode
W21 , W31 , W41 with side information W12 , W13 , W14 which
are also provided to the relay by the genie, the relay is able
to decode W21 , W31 , W41 as well. Thus, these 3 messages
contribute a total of 3d dimensions, which have to be linearly
independent with the interfering dimensions accommodating
the rest 4 messages W23 , W24 , W32 , W34 . Note that W23 , W24
both originate from user 2 and W32 , W34 both originate from
user 3, and these two users do not have a common projection
in the N-dimensional vector space at the relay because of
2M ≤ N. Thus, W23 , W24 , W32 , W34 contribute additional
4d dimensions. Therefore, these 7 messages contribute a total
of 7d dimensions in the N-dimensional vector space at the
relay, so that the symmetric DoF per message are upper
bounded by N/7.
3) 1/2 ≤ M/N ≤ 7/12 ⇒ d ≤ 2M/7: The proof
for this regime is similar to the previous case, because they
both originate from the case of M/N = 1/2, based on
the intuition of antenna dimension redundances, except for
2M ≥ N which implies that every two users have a (2M − N)-
dimensional common projection in the N-dimensional vector
space at the relay. Observing the 7 messages W21 , W31 , W41 ,
W23 , W24 , W32 , W34 that we bound their sum rate in last case,
we find that only (W23 , W32 ) form a pair. Since 2M ≥ N, the
signals carrying these two messages can occupy a common
subspace at the relay. Thus, from the DoF converse perspec-
tive, after a genie provides G  = {W12 , W13 , W14 , W42 , W43 } to
the relay, in order for the relay to decode the other 7 messages,
the genie still needs to provide additional information asso-
ciated with that common intersection projected from user 2
and user 3. To achieve this goal, the additional information
should contain the signal from, e.g., user 3, which is projected
into that common intersection at the relay. We can use a
(2M − N) × 1 column vector from user 3, denoted as X 3c
which projects into that common intersection at the relay.
With simple matrix calculation, the vector can be written as
 T
X 3c = (H3H H3 )−1 H3H [H2c H3c ]c X 3 , where Ac stands for
the arbitrary set of basis of the null space of AT . Thus, the total
number of dimensions of those 7 messages is upper bounded
by N + (2M − N) = 2M, i.e, d ≤ 2M/7.
B. Information Theoretic DoF Upper Bounds
Next, we translate the linear counting bounds above into the
information theoretic bounds.
1) M/N ≥ 7/12 ⇒ d ≤ N/6: Let a genie provide to
the relay the messages G = {W12 , W13 , W14 , W32 , W42 , W34 }.
Then for the sum rate of the 3 messages desired at user 1:
n(R21 + R31 + R41 )
≤ I (W21 , W31 , W41 ; Y1n |W12 , W13 , W14 ) + (n) (3)
≤ I (W21 , W31 , W41 ; Y Rn |W12 , W13 , W14 ) + (n) (4)
≤ I (W21 , W31 , W41 ; Y R , G|W12 , W13 , W14 ) + (n) (5)
n
= I (W21 , W31 , W41 ; G|W12 , W13 , W14 )
+ I (W21 , W31 , W41 ; Y Rn |G) + (n) (6)
= I (W21 , W31 , W41 ; Y Rn |G) + (n) (7)
WANG: DoFs OF MULTIWAY RELAY MIMO INTERFERENCE NETWORKS 7179

where (3) follows from the Fano’s inequality; (4) is obtained Next, consider the sum rate of the messages W24 , W34 desired
via the data processing inequality because (W21 , W31 , W41 ) − at user 4:
Y R − Y1 forms a Markov chain and all the messages are
independent; (5) is obtained because adding genie signals does n(R24 + R34 )
not reduce the capacity region. ≤ I (W24 , W34 ; Y4n |W41 , W42 , W43 ) + (n) (28)
Also, consider the sum rate of the 2 messages W23 , W43 ≤ I (W24 , W34 ; Y Rn |W41 , W42 , W43 ) + (n) (29)
desired at user 3:
≤ I (W24 , W34 ; Y Rn , G|W41 , W42 , W43 ) + (n) (30)
n(R23 + R43 ) = I (W24 , W34 ; G|W41 , W42 , W43 )
≤ I (W23 , W43 ; Y3n |W31 , W32 , W34 ) + (n) (8) + I (W24 , W34 ; Y Rn |W41 , G) + (n) (31)
≤ I (W23 , W43 ; Y Rn |W31 , W32 , W34 ) + (n) (9) = I (W24 , W34 ; Y Rn |W41 , G) + (n) (32)
≤ I (W23 , W43 ; Y Rn , G|W31 , W32 , W34 ) + (n) (10) ≤ I (W24 , W34 ; Y Rn |W21 , W31 , W41 , G) + (n). (33)
= I (W23 , W43 ; G|W31 , W32 , W34 )
Adding up (27), (33) we obtain:
+ I (W23 , W43 ; Y Rn |G, W31 ) + (n) (11)
= I (W23 , W43 ; Y Rn |W31 , G) + (n) (12) n(R21 + R31 + R41 + R24 + R34 )
≤ I (W23 , W43 ; Y Rn |W21 , W31 , W41 , G) + (n) (13) ≤ I (W21 , W31 , W41 , W24 , W34 ; Y Rn |G) + (n) (34)
where (13) follows from the fact that I (A; B|C) ≥ I (A; B) = h(Y Rn |G) − h(Y Rn |G, W21 , W31 , W41 , W24 , W34 ) + (n)
when A is independent of C. (35)
Next, consider the rate of W24 desired at user 4: = h(Y Rn |G) − h(Y Rn |X 1n , X 4n , W21 , W31 , W24 , W34 ) + (n)
n R24 ≤ I (W24 ; Y4n |W41 , W42 , W43 ) + (n) (14) (36)
≤ I (W24 ; Y Rn |W41 , W42 , W43 ) + (n) (15) = h(Y Rn |G) − h( X̃ 2n , X̃ 3n |W21 , W31 , W24 , W34 )
≤ I (W24 ; Y Rn , G|W41 , W42 , W43 ) + (n) (16) + n (log P) + (n) (37)
= I (W24 ; G|W41 , W42 , W43 ) = h(Y Rn |G) + n (log P) + (n)
+ I (W24 ; Y Rn |W41 , W43 , G) + (n) (17) −H (W21, W23 , W24 , W31 , W32 , W34 |W21 , W31 , W24 , W34 )
= I (W24 ; Y Rn |W41 , W43 , G) + (n) (18) (38)
≤ I (W24 ; Y Rn |W21 , W31 , W41 , W23 , W43 , G) + (n). = h(Y Rn |G) − H (W23, W32 ) + n (log P) + (n) (39)
(19) ≤ n N log P − n(R23 + R32 ) + n (log P) + (n) (40)

Finally, adding up (7), (13) and (19), we have:
n(R21 + R31 + R41 + R23 + R43 + R24 )
≤ I (W21 , W31 , W41 , W23 , W43 , W24 ; Y Rn |G) + (n)
≤ h(Y Rn |G) + (n) ≤ n N log P + (n)
where (20) is obtained due to the mutual information chain
rule, and (21) is obtained since the relay has a total of N
antennas. Then dividing n log P on both sides, and letting first
n → ∞ and then P → ∞, we obtain the desired DoF upper
bound
6n R ≤ n N log P + (n) ⇒ d ≤ N/6.
2) 3/7 ≤ M/N ≤ 1/2 ⇒ d ≤ N/7: Let a genie provide
the messages to the relay G = {W12 , W13 , W14 , W42 , W43 }.
Consider the sum rate of the 3 messages desired at user 1:
n(R21 + R31 + R41 )
≤ I (W21 , W31 , W41 ; Y1n |W12 , W13 , W14 ) + (n)
≤ I (W21 , W31 , W41 ; Y Rn |W12 , W13 , W14 ) + (n)
≤ I (W21 , W31 , W41 ; Y Rn , G|W12 , W13 , W14 ) + (n)
= I (W21 , W31 , W41 ; G|W12 , W13 , W14 )
+ I (W21 , W31 , W41 ; Y Rn |G) + (n)
where (36) is obtained because X 1 is an encoding function of
the messages W12 , W13 , W14 , and X 4 is an encoding function
of the messages W41 , W42 , W43 ; (37) follows from the fact that
(20) by subtracting the contributions of X 1 and X 4 from Y R , i.e.,
(21) Y R − H1 X 1 − H4 X 4 = H2 X 2 + H3 X 3 + Z R , the relay only
sees N linear combinations of the signals X 2 , X 3 sent from
user 2 and user 3 subject to the noise. Since they both have
M antennas only and 2M ≤ N, the relay is able to recover
the signals X 2 and X 3 subject to the noise distortion; (38) is
obtained because X̃ 2 and X̃ 3 are the noisy versions of X 2 ,
X 3 , and we can recover all their own messages, respectively,
subject to the noise distortion.
(22) Finally, by rearranging (40) we have the following rate
inequality:
n(R21 + R31 + R41 + R24 + R34 + R23 + R32 )
≤ n N log P + n (log P) + (n). (41)
Dividing n log P on both sides, and letting first n → ∞ and
(23)
then P → ∞, we obtain the bound
(24)
(25) 7n R ≤ n N log P + n (log P) + (n) ⇒ d ≤ N/7. (42)
3) 1/2 ≤ M/N ≤ 7/12 ⇒ d ≤ 2M/7: Let a genie
(26) provide the information G = {G  , X 3cn } to the relay where

= I (W21 , W31 , W41 ; Y Rn |G) + (n). (27) 

G = {W12 , W13 , W14 , W42 , W43 }. Then we consider the sum

Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
7180 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 10, OCTOBER 2016

rate of the 3 messages desired at user 1: V. D O F ACHIEVABILITY: P ROOF OF L EMMA 2

n(R21 + R31 + R41 ) With the intuition of antenna dimension redundancies earlier
≤ I (W21 , W31 , W41 ; Y1n |W12 , W13 , W14 ) + (n) (43) mentioned in Section III, we only need to first demonstrate
the achievability at MN = 7 and N = 12 , and then extend
3 M 7
≤ I (W21 , W31 , W41 ; Y Rn |W12 , W13 , W14 ) + (n) (44) the DoF achievability to other regimes via spatial extensions.
≤ I (W21 , W31 , W41 ; Y Rn , G|W12 , W13 , W14 ) + (n) (45) Basically, the DoF achievability at MN = 12 , also reported in
7

= I (W21 , W31 , W41 ; G  |W12 , W13 , W14 ) [9], is somewhat a straightforward extension of [6] because
+ I (W21 , W31 , W41 ; X 3cn
, Y Rn |G  ) + (n) (46) only one-to-one alignment is sufficient. In contrast, the DoF
N = 7 is quite challenging. In this section,
achievability at M 3
n
= I (W21 , W31 , W41 ; X 3c , Y Rn |G  ) + (n). (47) we are primarily interested in this case.
Next, consider the sum rate of the messages W24 , W34 desired
at user 4:
Principle of the Beamforming Design in the First Phase
n(R24 + R34 ) As mentioned in Section IV, the received signal at each
≤ I (W24 , W34 ; Y4n |W41 , W42 , W43 ) + (n) (48) user only depends on the transmitted signal from the relay.
≤ I (W24 , W34 ; Y Rn |W41 , W42 , W43 ) + (n) (49) If the side information available at one user is also available
≤ I (W24 , W34 ; Y Rn , G|W41 , W42 , W43 ) + (n) (50) at the relay, then the relay is able to decode whatever that
user can decode. From the DoF achievability perspective, it
= I (W24 , W34 ; G  |W41 , W42 , W43 ) implies that it does not matter whether the two signals carrying
+ I (W24 , W34 ; X 3cn
, Y Rn |W41 , G  ) + (n) (51) Wi j , W j i respectively align or not at the relay, because given
= I (W24 , W34 ; X 3c , Y Rn |W41 , G  ) + (n)
n
(52) a linear combination of the signals carrying Wi j and W j i ,
≤ I (W24 , W34 ; X 3c
n
, Y Rn |W21 , W31 , W41 , G  ) + (n). (53) once one message is available as the side information, the
subspace occupied by other message is uncovered. Thus, the
Adding up (47), (53) we obtain: principle to design an achievable scheme is that at the relay,
for every (Wi j , W j i ) pair, the signals carrying all the other
n(R21 + R31 + R41 + R24 + R34 ) messages occupy only N −d dimensions, and the d dimensions
≤ I (W21 , W31 , W41 , W24 , W34 ; X 3c
n
, Y Rn |G  ) + (n) (54) accommodating the each message of Wi j , W j i are linearly
n
= h(X 3c , Y Rn |G  ) independent with those N − d dimensions, so that the relay
− h(X 3cn
, Y Rn |G  , W21 , W31 , W41 , W24 , W34 ) + (n) (55) is able to resolve d linearly independent combinations of the
signals associated with Wi j , W j i only, subject to the noise
= n
h(X 3c , Y Rn |G  ) distortion.
− h(X 3cn
, Y Rn |X 1n , X 4n , W21 , W31 , W24 , W34 ) + (n) (56)
= n
h(X 3c , Y Rn |G  ) − h( X̃ 2n , X̃ 3n |W21 , W31 , W24 , W34 )
A. (M, N) = (3,7) ⇒ d = 1
+ n (log P) + (n) (57)
= h(X 3c
n
, Y Rn |G  ) − H (W23, W32 ) + n (log P) + (n) (58) 1) From the Users to the Relay (1st Phase): User k encodes
the symbol u kj using a 3 × 1 beamforming vector Vkj , j ∈
≤ n(2M − N) log P + n N log P − n(R23 + R32 )
K \ {k}, and the transmitted signal from user k is given by
+ n (log P) + (n) (59)

where (57) is obtained because by removing the contribu- Xk = Vkj u kj . (62)
tions of X 1 and X 4 from Y R , the relay only sees N linear j ∈K \{k}

combinations of the signals X 2 , X 3 sent from user 2 and

user 3 subject to the noise. Note that X 3c , the transmitted
signal from user 3, is projected into that common subspace.
Thus, from (X 3c , Y R − H1 X 1 − H4 X 4 ), the relay is able
to reconstruct the signals X 2 and X 3 subject to the noise
distortion.
Finally, by rearranging (59) we have the following rate
inequality:
n(R21 + R31 + R41 + R24 + R34 + R23 + R32 )
≤ 2n M log P + n o(log P) + (n).
Then dividing n log P on both sides, and letting first n → ∞
and then P → ∞, we obtain
7n R ≤ 2n M log P + n o(log P) + (n) ⇒ d ≤ 2M/7.
(61)
Consider the received signal at the relay node. Note that
there are a total of 12 symbols, comprising 6 pairs of two
symbols (u kj , u j k ). Our goal is that for every pair of sym-
bols, the 10 symbols associated with the other 5 pairs span
7 − 1 = 6 dimensions only, and the signal vector carrying
each symbol in that pair is linearly independent with those
6 dimensions. By doing so, the relay can precode the 6 linear
equations via reciprocal beamforming and broadcast to the
users, so that each user finally sees 3 linear combinations,
(60) each of which is associated with one of its own symbols and
one of its desired symbols. To achieve this goal, for every pair
of symbols, the other 10 symbols have to be accommodated
into a 6-dimensional subspace at the relay. That is, we need
to align 4 symbols into the subspace spanned by the other
symbols. Since there are 6 pairs, the total number of alignment
equations is given by 4 × 6 = 24, and each equation can be

Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
WANG: DoFs OF MULTIWAY RELAY MIMO INTERFERENCE NETWORKS
Fig. 3. The inter-user signal subspace alignment at the relay node (Each
circle represents one symbol carrying one DoF. In the rectangle, the lower
7 symbols occupy 7 dimensions, one for each, where each dimension Fi
can be represented by a linearly independent combinations of the received
signals from the 7 antennas. The other upper 5 symbols are aligned into the
subspace jointly spanned by three dimensions where the arrows represent the
dimensions that we pick for each symbol.)
written as: for every (i, j ) where i, j ∈ K and i = j ,
 
Hi αi j  Vi j  + H j α j i V j i
j  ∈K \{i, j } i  ∈K \{ j,i}
 
+ Hk αkk  Vkk  = 0
k∈K \{i, j } k  ∈K \{k}
where on the left-hand side, the first two terms represent
the two vectors carrying the transmitted symbols of user i
other than u i j , and of user j other than u j i , respectively,
and the third term captures the other six vectors carrying the
six symbols sent from the other two users. Also, each vector
in (63) is also scaled by a variable α(··) that needs design.
Thus, we need to specify a total of 10 × 24 = 240 scalar
variables α(··) as well as 12 3 × 1 beamforming vectors V(··).
While this approach is straightforward, it is extremely hard to
find the solution due to three reasons. First, the number of the
variables associated with alignment equations is too large to
handle with. Second, the 24 alignment equations have strong
dependencies among themselves, because only four channel
matrices Hk , k ∈ K are involved. It is not straightforward
how to reduce the redundancies of the equations. Finally, the
equations in (63) are not linear but bilinear, but solvability of
general bilinear equations remains widely open even in the
mathematic field.
Recognizing the challenges emerging out from the standard
approach, let us reconsider the signal vectors at the relay
carrying each symbol. In order for the reader to follow easily,
let us look into Fig. 3, and see how to align the symbols as
many as possible.
First, consider the signal vectors sent from user 3 and
user 4. Since each user is equipped with 3 antennas and
the relay is equipped with 7 antennas, the subspaces pro-
jected from these two users do not have overlap at the
relay, i.e., alignment is impossible. Thus, the 3 brown sym-
bols u 31 , u 32 , u 34 from user 3 and the 3 grey symbols
u 41 , u 42 , u 43 from user 4 occupy a total of 6 dimensions at
the relay. For simplicity, we use the 7 × 1 column vectors
F2 , F3 , F4 , F5 , F6 , F7 to present their directions. Hence, these
Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
7181
6 vectors can be written as:
H3 V31 = F2 , H3 V32 = F3 , H3 V34 = F4 ,
H4 V41 = F5 , H4 V42 = F6 , H4 V43 = F7 . (64)
So far, a total of 6 symbols are accommodated in the
7-dimensional signal space at the relay.
Second, consider the signal carrying the green symbol u 21 .
Since u 21 cannot be paired with any one of the 6 symbols
already dwelling at the relay, it has to occupy the remaining
dimension, i.e., alignment is unnecessary. For simplicity, we
denote it by the 7 × 1 column vector F1 :
H2 V21 = F1 . (65)
Next, we look at the other two green symbols u 23 , u 24 .
Because the 7-dimensional signal space at the relay is already
occupied by 7 symbols, we have to use signal alignment.
Let us consider the two paired symbols u 23 and u 23 first.
According to the principle of the beamforming design, we
have to guarantee that after zero-forcing the other 10 symbols,
we can obtain a clean observation associated with u 23 , u 32
only. Thus, u 23 has to be projected into a subspace jointly
spanned by symbols including at least u 32 , i.e., the vector F3 .
However, since one-to-one alignment is impossible in this
(63) channel, we need to find some other basis together with F3
to accommodate u 23 . Here, we choose the directions F4 , F7
which carry the two paired symbols u 34 , u 43 , so that u 23 is
aligned in the subspace joint spanned by F2 , F4 , F7 . Without
loss of generality, we obtain the alignment equation H2 V23 =
−(F3 + α4(1) F4 + α7(1) F7 ) where α4(1) , α7(1) are two non-zero
scalar variables that we can design, and the minus sign is
used for simple calculation later. For an intuitive presentation,
we connect u 23 to u 32 , u 34 , u 43 with red dashed lines in
Fig. 3. While the construction of H2 V23 seems to be somewhat
artificial, we provide the intuitions in the following why we
construct the alignment equation in such a manner.
• Similar to the pair (u 23 , u 32 ), we also need to guarantee
that for each of u 21 , u 31 , u 41 , u 42 along the dimensions
F1 , F2 , F5 , F6 , a clean observation associated with itself
and its paired symbol can be obtained, after zero-forcing
the other 10 symbols. Thus, we cannot connect u 23 to
any of u 21 , u 31 , u 41 , u 42 with red dashed lines.
• Among the remaining two dimensions F4 , F7 carrying
u 34 , u 43 , if we connect u 23 to only one of u 34 , u 43 with
a red dashed line, say, u 34 , then for u 34 , u 43 , after zero
forcing the other 10 symbols, only u 43 remains there and
we cannot see u 34 , thus violating our principle. Hence,
we have to connect u 23 to both u 34 and u 43 in Fig. 3.
With the similar analysis, for the symbol u 24 , we need to
connect u 24 to its paired u 42 along F6 and u 34 , u 43 along
(2) (2)
F4 , F7 , and thus we have H2 V24 = −(F6 + α4 F4 + α7 F7 )
(2) (2)
where α4 , α7 are another two non-zero scalar variables that
we can design. Moreover, for the last three blue symbols
u 12 , u 13 , u 14 sent from user 1, we can automatically write
(3)
down the alignment equations as H1 V12 = −(F1 + α4 F4 +
(3) (4) (4)
α7 F7 ), H1 V13 = −(F2 + α4 F4 + α7 F7 ) and H1 V14 =
−(F5 + α4 F4 + α7 F7 ), where α4 , α7(i) are the non-zero
(5) (5) (i)
scalar variables for the i t h alignment equation. In Fig. 3,
7182 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 10, OCTOBER 2016

we use the black dashed lines to represent the corresponding So far, we settle the challenges that we encounter by using
alignment equations. the standard approach as we mentioned at the beginning
Finally, consider the symbols (u 34 , u 43 ) again. Clearly, all of this section. Next, we rewrite the 5 alignment equations
the other 5 pairs connect to both u 34 and u 43 in Fig. 3. Recall from (67) to (71) into a compact form:
that after zero-forcing the other 10 symbols, what remains ⎡ ⎤
H1 O O H2 O O O O H3 O O H4
should be a clean linear combination of only (u 34 , u 43 ). ⎢ O H1 O O O O H3 O H3 O O H4 ⎥
That is, although the 5 alignment equations above push the ⎢ ⎥
⎢ O O H1 O O O O O H3 H4 O H4 ⎥
interfering symbols into the 2-dimensional subspace with basis ⎢ ⎥
⎣ O O O O H2 O O H3 H3 O O H4 ⎦
F4 , F7 , we have to keep one-dimension clean for (u 34 , u 43 ).
O O O O O H2 O O H3 O H4 H4
Therefore, in the 2-dimensional subspace with basis F4 , F7 ,   
all the other interfering symbols must align along the same H35×36
direction to occupy only one dimension, which implies α4(i) = ⎡ ⎤
V12
(i)
α4 , α7 = α7 , i = 1, 2, 3, 4, 5. Note that this is a significant ⎢ V13 ⎥
⎢ ⎥
condition that requires 5 signal vectors to align, when we × ⎢ . ⎥ = 0. (72)
⎣ .. ⎦
consider (u 34 , u 43 ). In this work, we choose α4 = α7 = 1,
and we can rewrite the 5 alignment equations as V43

H2 V23 = −(F3 + F4 + F7 ), H2 V24 = −(F6 + F4 + F7 ), Since Hk , k ∈ K are generic, it is not difficult to verify the
35 × 36 matrix H in (72) has full rank, by picking a special
H1 V12 = −(F1 + F4 + F7 ), H1 V13 = −(F2 + F4 + F7 ), set of Hk , k ∈ K and compute det(HH H ) = 0 (This approach
H1 V14 = −(F5 + F4 + F7 ). (66) has been widely used in information theory to study linear
As shown above, the signal alignment happens among the independencies among vectors of a matrix, such as [2].). For
subspaces projected from different user nodes. Such an opera- example, we pick the following set of Hk , k ∈ K :
⎡ ⎤ ⎡ ⎤
tion, based on one-to-many alignment, is referred to as inter- 1 0 0 0 0 0
user signal subspace alignment. Also, note that all the 4 users ⎢0 1 0⎥ ⎢0 0 0⎥
⎢ ⎥ ⎢ ⎥
involve in each alignment equation, which has not been shown ⎢0 0 1⎥ ⎢0 0 0⎥
⎢ ⎥ ⎢ ⎥
in prior works, e.g., [8], [9], [17], [21]. H1 = ⎢ ⎥ ⎢
⎢ 0 0 0 ⎥, H2 = ⎢ 1 0 0 ⎥,
⎥
So far, we have established all the alignment equations that ⎢0 0 0⎥ ⎢0 1 0⎥
⎢ ⎥ ⎢ ⎥
we need. Before we prove the linear independencies among ⎣0 0 0⎦ ⎣0 0 1⎦
the seven column vectors Fi , i = 1, 2, · · · , 7, let us first 0 0 0 0 0 0
examine the scheme above does not violate the principle of ⎡ ⎤ ⎡ ⎤
0 1 0 0 1 0
beamforming design introduced earlier. Among the 6 pairs ⎢0 0 1⎥ ⎢0 1 1⎥
(u 12 , u 21 ), (u 13 , u 31 ), (u 14 , u 41 ), (u 23 , u 32 ), (u 24 , u 42 ) and ⎢ ⎥ ⎢ ⎥
⎢0 0 0⎥ ⎢1 0 0⎥
(u 34 , u 43 ), we only examine the first pair as an example, and ⎢ ⎥ ⎢ ⎥
H3 = ⎢ ⎥ ⎢
⎢ 0 1 0 ⎥, H4 = ⎢ 0 1 0 ⎥,
⎥ (73)
the reader can verify the others with the similar approach. ⎢0 0 1⎥ ⎢0 0 1⎥
For (u 12 , u 21 ) at the relay, the vectors carrying the other ⎢ ⎥ ⎢ ⎥
⎣0 0 0⎦ ⎣1 0 1⎦
10 symbols u 13 , u 14 , u 31 , u 41 , u 23 , u 24 , u 32 , u 34 , u 42 and 1 0 0 0 1 1
u 43 are given by −(F2 + F4 + F7 ), −(F4 + F5 + F7 ), F2 , F5 ,
−(F3 + F4 + F7 ), −(F4 + F6 + F7 ), F3 , F4 , F6 and F7 , as shown and the reader can verify H has full rank by computing
in (66) and (64). It can be seen that all these 10 symbols lie in a det(HH H ) = 0. As a consequence, from (72), we obtain the
6-dimensional subspace spanned by the 6 linearly independent beamforming vectors of each user as follows:
basis F2 , F3 , F4 , F5 , F6 and F7 . In addition, the two symbols V = (I36 − H H (HH H )−1 H) det(HH H )Q 36×1 , (74)
u 12 (desired at user 2) and u 21 (desired at user 1) arrive at the
V12 = V (1 : 3), V13 = V (4 : 6), V14 = V (7 : 9),
relay along the vectors −(F1 + F4 + F7 ) and F1 respectively.
Thus, each of them does not lie in that 6-dimensional subspace V21 = V (10 : 12), V23 = V (13 : 15), V24 = V (16 : 18),
spanned by F2 , F3 , F4 , F5 , F6 , F7 . V31 = V (19 : 21), V32 = V (22 : 24), V34 = V (25 : 27),
What remains to be shown is to specify the 7 linearly V41 = V (28 : 30), V42 = V (31 : 33), V43 = V (34 : 36),
independent column vectors Fi , i = 1, 2, · · · , 7 given the
(75)
matrices Hk , k ∈ K . After observing the 12 vectors on the
right-hand side of equations from (64) to (66), we can rewrite where Q 36×1 is a randomly picked 36 × 1 vector. Finally,
all those equations by eliminating all Fi to obtain the following it suffices to only prove that the 7 column vectors Fi , i =
5 alignment equations: 1, · · · , 7 are linearly independent, or equivalently to show that
H1 V12 + H2 V21 + H3 V34 + H4 V43 = 0, (67) G  [F1 , F2 , F3 , F4 , F5 , F6 , F7 ]
H1 V13 + H3 V31 + H3 V34 + H4 V43 = 0, (68) = [H2 V21 , H3 V31 , H3 V32 , H3 V34 , H4 V41 , H4 V42 , H4 V43 ]
H1 V14 + H3 V34 + H4 V41 + H4 V43 = 0, (69)
has full rank. Since each entry of G is a polynomial of
H2 V23 + H3 V32 + H3 V34 + H4 V43 = 0, (70) only the entries of Hk once Q 36×1 is drawn, it suffices
H2 V24 + H3 V34 + H4 V42 + H4 V43 = 0. (71) to show det(G) = 0 via picking a special set of Hk ,

Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
WANG: DoFs OF MULTIWAY RELAY MIMO INTERFERENCE NETWORKS
in (73), and the resulting beamforming vectors of each user are
given by:
⎡ ⎤
1 0 0
[V12 V13 V14 ] = α ⎣ 0 1 0 ⎦,
0 0 −1
⎡ ⎤
1 0 0
[V21 V23 V24 ] = α ⎣ −1 −1 0 ⎦,
0 0 1
⎡ ⎤
0 0 1
[V31 V32 V34 ] = α ⎣ 1 1 0 ⎦,
−1 0 1
⎡ ⎤
1 0 0 number of antennas at each node by 3, to obtain a (M  , N  ) =
[V41 V42 V43] = α ⎣ 1 1 −1 ⎦, (76)
−1 −1 0 of antennas at the relay from 3N to 7M. For this reduced
where α is a non-zero scalar only depending on the vector
Q 36×1 to satisfy the power constraint. Then it is easy to verify
det(G) = 0, i.e., the matrix G has full rank.
So far, we finish the design of the beamforming vectors in
the first phase.
2) From the Relay to the Users (2nd Phase): The relay
is able to resolve a clean linear combination of every pair
of symbols (u i j , u j i ), by projecting the received signal into
the nullspace of the 6-dimensional subspace spanned by the
other 10 symbols. By doing so, the relay obtains 6 linear
equations associated with every pair of symbols individually
si j = βi j u i j + β j i u j i + z i j , i, j ∈ K , i < j , where the
linear combination coefficients βi j , βij depend on the channel
matrices Hk , k ∈ K only and z i j is the AWGN with bounded
variance, which does not depend on the power P, and thus it
does not affect the DoF analysis.
In this phase, each user follows from a reciprocal approach
to design its receiving beamforming matrix. Specifically, by
replacing Hk in the first phase with H̄kT , and by using the
same design as in the first phase, the transmitting beamforming
vector for each symbol in the first phase, is the receiving
beamforming vector for its paired desired symbol in the second
phase. Recall that in the first phase, at the relay side, for
each pair, the other 10 symbols only project a 6-dimensional
subspace at the relay. Thus, in the second phase, the relay
sends each linear equation si j to both user i and user j
with the beamforming vector which is orthogonal to the
6-dimensional subspace projected back from the subspaces of
all users accommodating the other 10 symbols. By doing so,
each user j finally only sees 3 linear equations si j , i = j .
Note that each linear equation is associated with u i j , u j i only
subject to the noise. After subtracting the signal carrying its
own symbol u j i from its received signal (ηsi j + z j ), user j
sees an equivalent interference-free point-to-point channel
(ηi j si j + z j ) − ηi j β j i u j i = ηi j βi j u i j + z j + ηi j z i j , where
ηi j is a constant scalar to represent the amplifying coefficient
for the stream u i j at the relay, z j is the AWGN with bounded
variance at user j , and both ηi j and z j , z i j do not depend on
the power P. Thus, user j is able to decode u i j , so that each
message can achieve one DoF.
Remark: Recall the idea of alignment for network coding
for the 3-user relay MIMO Y channel [4], it turns out that
Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
7183
they had β = β  = 1 via one-to-one alignment. In contrast,
β and β  in this work depend on the channel matrices Hk , k ∈
K only, via one-to-many alignment.
B. General Cases in the Regime M/N ≤ 1/2
For general M/N = 3/7 cases, i.e., (M, N) = (3β, 7β)
and β ∈ Z+ , β > 1, we can still use the same achievable
scheme that we show in last section, as long as each symbol
is a β-dimensional symbol, i.e., each message carries β DoF.
For the other cases in the regime M/N ≤ 1/2, similar
to [2], we establish the DoF achievability in the sense of
spatial extensions. Specifically, if M/N < 3/7, we scale the
(3M, 3N) network. Since 3N > 7M, we reduce the number
network, we again can use the same approach that we show in
last section to achieve M DoF per message, so that the DoF per
message normalized to spatial extensions are given by M/3,
as shown in Theorem 1. On the other hand, if M/N > 3/7,
we scale the number of antennas at each node by 7, to obtain
a (M  , N  ) = (7M, 7N) network. Since 7M > 3N, we reduce
the number of antennas at each user from 7M to 3N. For
this reduced network, we again use the same approach that
we show in last section to achieve N DoF per message, so
that the DoF per message normalized to spatial extensions are
given by N/7, implied by Theorem 1.
VI. F EASIBILITY OF L INEAR A LIGNMENT:
P ROOF OF T HEOREM 2
For the 4-user relay MIMO Y channel defined in Section II,
Theorem 2 implies that the DoF demand per user, d, is feasible
with linear interference alignment if and only if d ≤ dY∗ .
Since the upper bound follows directly from Lemma 1, we
only need to provide the achievability to show d ≤ dY∗
DoF per message are achievable using linear schemes without
symbol extensions in time/frequency/space. As implied by
Theorem 2, the feasible DoF value is again presented by four
pieces. We will consider each regime individually.
A. M/N ≤ 1/2
The idea behind the proof for the regime M/N ≤ 1/2 is
based on reducing the number of antennas at both each user
and the relay to obtain a reduced network, to which we directly
apply the achievable scheme designed for M/N = 3/7.
Let us consider the regime M/N ≤ 3/7 first. In this regime,
our goal is to show that each message can achieve M/3 DoF.
To see this, each message is encoded to M/3 independent
symbols. At each user, we reduce the number of antennas from
M to M  = 3 M 3 . Also, at the relay node, we reduce the
number of antennas N to N  = 7 M 3 . This can be done since
M ≥ M  and N ≥ N  . By doing so, we form a new 4-user
relay MIMO Y channel where each user has M  antennas, the

relay has N  antennas and M  /N  = 3/7. Thus, M3 = M 3
DoF can be achieved, by the use of he achievable scheme that
we present in Section V by replacing each one-dimensional

symbol with an M3 -dimensional symbol.
7184 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 10, OCTOBER 2016

Next, consider the regime 3/7 < M/N ≤ 1/2. Our goal where user 1 encodes its 3m 1 symbols for W12 , W13
is to show that each message can achieve N/7 DoF. To see and W14 sequentially, user 2 encodes its 3m 1 symbols for
this, each message is encoded to N/7 symbols. At the relay, W21 , W23 and W24 sequentially, user 3 encodes its 3m 1
we reduce the number of antennas from N to N  = 7 N7 . symbols for W31 , W32 and W34 sequentially, and user 4
Also, at each user node, we reduce the number of antennas M encodes its 3m 1 symbols for W41 , W42 and W43 sequentially.
to M  = 3 N7 . Again, this can be done due to the fact that Therefore, the linear signal alignment solution exists almost
M ≥ M  and N ≥ N  . Hence, we form a new 4-user relay surely.
MIMO Y channel where each user has M  antennas, the relay Next, consider the regime 1/2 ≤ M/N ≤ 7/12. In this

has N  antennas and M  /N  = 3/7. Thus, N7 = N7 DoF regime, our goal is to show that each message can achieve
2M
can be achieved, by the use of the achievable scheme that 7 DoF. To see this, each message is encoded to 2M 7
we present in Section V by replacing each one-dimensional symbols. At the relay node, we reduce the number of antennas

symbol with an N7 -dimensional symbol. from N to N  = 6 2M 7 . Now, consider the number of
common intersection projected from every two users, which
is given by 2M − N  ≥ 2M − N ≥ 2M 7 ≥
2M
7 , implying
B. M/N ≥ 1/2 2M
that we have freedom to choose 7 dimensions to send
2M
The idea behind the proof for the regime M/N ≥ 1/2 is 7 independent symbols per message. Again, we need a
based on reducing the number of antennas at the relay node proof to show that is true through constructing special specific
only to obtain a reduced network, to which we directly apply channels. With the same construction of the specific channels
the achievable scheme designed for M/N = 7/12. as we present for M/N ≥ 7/12 by letting m 1 = 2M 7 ,
First, we consider the regime M/N ≥ 7/12. In this regime, it can be seen that the alignment solution exists almost
our goal is to show that each message can achieve N/6 DoF. surely.
To see this, each message is encoded to N/6 symbols. At
the relay node, we reduce the number of antennas from N to VII. C ONCLUSION
N  = 6 N6 . Since M/N ≥ 7/12, we must have M/N  ≥ 7/12
as well. Now, consider the number of common intersection In this paper, we characterize the DoF of two kinds of
projected from every two users, which is given by 2M − N  ≥ multi-way relay MIMO interference networks: the 4-user relay
2M − N ≥ N6 ≥ N6 , implying that we can randomly pick MIMO Y channel and the two-way relay MIMO X channel.
N To establish the DoF converse, we begin with the linear
6 dimensions in that common intersection, along which the
two signals carrying the corresponding paired N6 symbols dimension counting approach to produce the linear DoF upper
per user are aligned. Note that the analysis above is carried out bounds, and then translate them into the information theoretic
via linear dimension counting. We still need a proof to show upper bounds. The information theoretic DoF upper bounds
that is true through constructing special specific channels. For are facilitated by translating the multi-way two-hop channel
example, assuming m 1 = N6 , we choose the N  × M reduced to a one-hop multiple access channel with additional decoding
channel matrices, still denoted by Hk , k ∈ K for brevity, as constraints at the relay. To establish the DoF achievability,
Hk = [Hk Rk ], k ∈ K where Rk are randomly generated we propose linear beamforming schemes, in the sense of
N  × (M − 3m 1 ) matrices and Hk , k ∈ K are given by spatially-normalized extension in general, to show that the
information theoretically optimal DoF can be achieved almost
⎡ ⎤ ⎡ ⎤
Im 1 O O Im 1 O O surely.
⎢ O Im O ⎥ ⎢ O O O ⎥ Several interesting observations follow as a byproduct of our
⎢ 1 ⎥ ⎢ ⎥
⎢ O O Im 1 ⎥ ⎥ ⎢ O O O ⎥ analysis. First, we precisely identify settings with redundant

H1 = ⎢ ⎢ 
, H2 = ⎢ ⎢ ⎥,
⎢ O O O ⎥⎥ ⎢ O Im 1 O ⎥ ⎥
dimensions at each user node, the relay node, both or neither. It
⎣ O O O ⎦ ⎣ O O Im 1 ⎦ implies that the similar observations, available for the one-way
O O O O O O one-hop MIMO interference channel [2], exist in the multi-
⎡ ⎤ ⎡ ⎤ way relay MIMO interference networks as well. Second, our
O O O O O O
⎢ Im O O ⎥ ⎢ O O O ⎥ results in this paper also shed lights on the feasibility of linear
⎢ 1 ⎥ ⎢ ⎥
⎢ O O O ⎥⎥ ⎢ O ⎥ interference alignment. Finally, while we find that the DoF
H3 = ⎢ , H 
= ⎢ Im 1 O ⎥,
⎢ O Im O ⎥ 4 ⎢ O O O ⎥ counting bound also serves as the upper bound for linear beam-
⎢ 1 ⎥ ⎢ ⎥ forming schemes, does the DoF decomposition bound exist in
⎣ O O O ⎦ ⎣ O Im 1 O ⎦
O O Im 1 O O Im 1 multi-way relay communication networks? The DoF decompo-
sition is already well known for several one-way interference
(77)
channels such as [2], [14] and the X channels [15]. While the
where each O stands for the m 1 × m 1 zero matrix. As a relationship between the DoF counting bound and the DoF
consequence, using the achievable scheme for M/N  = 7/12, decomposition bound motivates many recent interesting works
the beamforming matrix for each message at each user can be for the one-way one-hop channels such as [2], [14]–[16], it
automatically determined as is not clear if we have the same observation or fundamental
  principles behind the results for multi-way communication net-
I3m 1 works. This research avenue would be of interest in the future
Vk = , k∈K (78)
O(M−3m 1 )×3m 1 work.

Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
WANG: DoFs OF MULTIWAY RELAY MIMO INTERFERENCE NETWORKS 7185

A PPENDIX Xiang et al. in [13], we only study the case (M, N) = (2, 5)
A. Development of the DoF Counting Bound where Theorem 1 tells d = 1 DoF. Again, we only need to
carefully design the achievability in the first phase, so that
1) All Unicast Setting: 4-User Relay MIMO Y Channel:
a clean linear combination of each pair of symbols can be
Let us first count the number of variables. At the user
resolved at the relay. In the first phase, each user k encodes
side, each user encodes the d independent symbols carrying
each symbol u k· using a 2 × 1 beamforming vector Vk· , and
each message with an M × d beamforming matrix. Thus,
the transmitted signal vector of user k can be written as
the beamforming matrix corresponding to each message con-
tributes (M − d)d independent variables. Since there are a 
Vk3 u k3 + Vk4 u k4 , k = 1, 2,
total of 12 messages, the number of variables contributed Xk = (79)
Vk1 u k1 + Vk2 u k2 , k = 3, 4.
by the beamforming matrices of all the 12 messages is
given by 12(M − d)d. At the relay node, to protect a In this case, at least we need to ensure that for every user,
d-dimensional clean subspace for each pair of messages, the except for its 2 desired symbols and its 2 own symbols,
relay employs an N × d receiving beamforming matrix. Since the signal subspace occupied by the other 4 symbols has
the 12 messages comprise a total of 6 pairs, the number of 5 − 2d = 3 dimensions only. For example, consider user 1
variables contributed by the receiving beamforming matrices who desires the symbols u 31 and u 41 , and its own transmitted
at the relay node is 6(N − d)d. Thus, the total number of symbols are u 13 and u 14 . Except for these 4 symbols, the other
variables is given by 12(M −d)d +6(N −d)d. Next, we count 4 symbols u 23 , u 24 u 32 and u 42 are interfering symbols. In the
the number of nulling equations. For each pair of messages, 5-dimensional space at the relay, to protect a 2-dimensional
when the relay employs an N × d receiving beamforming subspace for the 2 desired symbols, those 4 vectors pro-
matrix to protect a d-dimensional clean subspace for that pair, jected from users 2, 3 and 4 can only span 3 dimensions.
the relay essentially zero forces all the signals carrying the Thus, we align the vector carrying u 42 into the subspace
other 10 messages, each encoded to d independent symbols. spanned by the vectors carrying u 23 , u 24 and u 32 at the relay.
Since the total number of nulling equations is 10d 2 for every That is,
pair of messages, the number of nulling equations is 60d 2 .
Following the intuition explained in [3], we define a system as H2 V23 + H2 V24 + H3 V32 + H4 V42 = 0. (80)
proper if the number of variables is no fewer than the number
of nulling equations, and improper otherwise. Thus, we obtain Next, consider user 2 who desires the symbols u 32 and u 42 ,
the DoF counting bound 12(M −d)d +6(N −d)d ≥ 60d 2 ⇒ and its own transmitted symbols are u 23 and u 24 . Thus,
d ≤ 2M+N13 .
the remaining 4 symbols u 13 , u 14 u 31 and u 41 constitute
2) Multiple Unicast Setting: Two-Way Relay MIMO X interfering symbols. have to span no more than 3 dimensions.
Channel: Following the similar analysis as in last section, In particular, we align the vector carrying u 41 into the subspace
we directly count the number of variables to be 8(M − d)d + spanned by the vectors at the relay. That is,
4(N − d)d where we replace “12” and “6” in last section
with “8” and “4” here, and the number of equations associated H1 V13 + H1 V14 + H3 V31 + H4 V41 = 0. (81)
with nulling interference to be 24d 2 . Thus, the DoF counting
bound can be written as 8(M − d)d + 4(N − d)d ≥ 24d 2 ⇒ With the similar analysis, for user 3, we let
d ≤ 2M+N 9 .
H1 V14 + H2 V24 + H4 V41 + H4 V42 = 0. (82)
B. DoF Converse: Proof of Lemma 3 Finally, for user 4, we need another alignment equation. How-
As implied by Lemma 3, the DoF result can be represented ever, after carefully observing the alignment equations (80),
by 4 linear pieces depending on M and N alternatively. For (81) and (82), we find that (80) + (81) − (82) produces:
the regime M/N ≤ 2/5, since each user equipped with
M antennas sends 2 messages, the DoF per message are upper H1 V13 + H2 V23 + H3 V31 + H3 V32 = 0 (83)
bounded by M/2. For the remaining regimes, we follow the
same analysis that we show for the all unicast setting in which implies that the vector carrying u 13 is automatically
Section IV by deactivating the 4 messages W12 = W21 = aligned into the subspace spanned by the vectors carrying
W34 = W43 = ∅ and thus R12 = R21 = R34 = R43 = 0. u 23 u 31 and u 32 at the relay. That is, the equation (83) is
Then the DoF converse proofs for MN ≥ 8 , 5 < N ≤ 2 and
5 2 M 1 redundant.
2 < N ≤ 8 are essentially the same as those for N ≥ 12 ,
1 M 5 M 7 Collecting the 3 equations (80), (81) and (82), we rewrite
7 < N ≤ 2 and 2 < N ≤ 12 in Section IV, respectively.
3 M 1 1 M 7 them into a compact matrix form:
⎡ ⎤ ⎡V ⎤
O O H2 H2 O H3 O H4 13
⎢ ⎥
O ⎦ ⎣ ... ⎦ = 0.
C. DoF Achievability: Proof of Lemma 4 ⎣ H1 H1 O O H3 O H4
With the antenna dimension redundances intuition illus- O H1 O H2 O O H4 H4
trated in Section III, Section V and [2], it suffices to present    V42
the achievability at M/N = 2/5 and M/N = 5/8. Since H15×16
the DoF achievability at M/N = 5/8 was already shown by (84)

Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.
7186 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 10, OCTOBER 2016

Similarly to prove H has full rank, we pick a special set of [7] H. Chung, N. Lee, B. Shim, and T. Oh, “On the beamforming design for
matrices Hk , k ∈ K : MIMO multipair two-way relay channels,” IEEE Trans. Veh. Technol.,
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ vol. 61, no. 7, pp. 3301–3306, Sep. 2012.
1 0 0 0 0 1 [8] C. Wang and S. A. Jafar, “Degrees of freedom of the two-
⎢0 1⎥ ⎢0 0⎥ ⎢1 1⎥ way relay MIMO interference channel,” e-print UC-escholarship,
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ UCI CPCC Rep. 9qc3343h, Jan. 2013.
H1 = ⎢ ⎢ 0 0 ⎥, H2 = ⎢ 1 0 ⎥, H3 = ⎢ 0 0 ⎥,
⎥ ⎢ ⎥ ⎢ ⎥ [9] B. Yuan, X. Liao, F. Gao, and X. Luo, “Achievable degrees of freedom
⎣0 0⎦ ⎣0 1⎦ ⎣1 0⎦ of the four-user MIMO Y channel,” IEEE Commun. Lett., vol. 18, no. 1,
pp. 6–9, Jan. 2014.
0 0 0 0 0 1
⎡ ⎤ [10] A. Chaaban, K. Ochs, and A. Sezgin, “The degrees of freedom of the
0 1 MIMO Y-channel,” in Proc. IEEE ISIT, Jul. 2013, pp. 1581–1585.
⎢0 0⎥ [11] Y. Tian and A. Yener, “Degrees of freedom for the MIMO multi-way
⎢ ⎥ relay channel,” in Proc. IEEE ISIT, Jul. 2013, pp. 1576–1580.
H4 = ⎢ ⎢ 0 1 ⎥,
⎥ (85) [12] A. A. Zewail, M. Nafie, Y. Mohasseb, and H. El Gamal, “Achievable
⎣0 0⎦ degrees of freedom region of MIMO relay networks using Detour
schemes,” in Proc. IEEE ICC , Jun. 2014, pp. 1860–1865.
1 0 [13] Z. Xiang, M. Tao, J. Mo, and X. Wang, “Degrees of freedom for MIMO
and then it is easy to verify that H has full rank by computing two-way X relay channel,” IEEE Trans. Signal Process., vol. 61, no. 7,
pp. 1711–1720, Apr. 2013.
det(HH H ) = 0. Also, we can obtain the beamforming vector [14] C. Wang, H. Sun, and S. A. Jafar, “Genie chains and the degrees of
of each symbol via solving the equation (84): freedom of the K -user MIMO interference channel,” in Proc. IEEE ISIT,
Jul. 2012, pp. 2476–2480.
V = (I16 − H H (HH H )−1 ) det(HH H )Q 16×1 , [15] H. Sun, C. Geng, T. Gou, and S. A. Jafar. (Jul. 2012). “Degrees of
freedom of MIMO X networks: Spatial scale invariance, one-sided
V13 = V (1 : 2), V14 = V (3 : 4), V23 = V (5 : 6), decomposability and linear feasibility.” [Online]. Available: http://arxiv.
org/abs/1207.6137
V24 = V (7 : 8), V31 = V (9 : 10), V32 = V (11 : 12), [16] O. Gonzalez, C. Beltran, and I. Santamaria. (Feb. 2012). “A feasibility
V41 = V (13 : 14), V42 = V (15 : 16). (86) test for linear interference alignment in MIMO channels with constant
coefficients.” [Online]. Available: http://arxiv.org/abs/1202.0186
where Q 16×1 is a randomly picked 16×1 vector. Note that the [17] H. Mu and J. K. Tugnait, “Achievable degrees of freedom for K -user
MIMO Y channels using signal group based alignment,” IEEE Trans.
vector V in (86), lying in the null space of H, is a polynomial Wireless Commun., vol. 13, no. 8, pp. 4520–4533, Aug. 2014.
of all the entries of Hk , k ∈ K only once Q 16×1 is drawn. [18] K. Gomadam, V. R. Cadambe, and S. A. Jafar, “A distributed numerical
Thus, if we still pick the matrices in (85), then the resulting approach to interference alignment and applications to wireless interfer-
ence networks,” IEEE Trans. Inf. Theory, vol. 57, no. 6, pp. 3309–3322,
beamforming vector of each symbol can be simplified as: Jun. 2011.
  [19] G. Bresler, D. Cartwright, and D. Tse. (Apr. 2011). “Settling the
0 −1
[V13 V14 ] = [V23 V24 ] = α , feasibility of interference alignment for the MIMO interference channel:
−1 0 The symmetric square case.” [Online]. Available: https://arxiv.org/abs/
    1104.0888
0 1 −1 1
[V31 V32 ] = α , [V41 V42 ] = α . (87) [20] M. Razaviyayn, G. Lyubeznik, and Z.-Q. Luo, “On the degrees of free-
1 −1 0 1 dom achievable through interference alignment in a MIMO interference
channel,” IEEE Trans. Signal Process., vol. 60, no. 2, pp. 812–821,
where α is a non-zero scalar depending on Q 16×1 only to meet Feb. 2012.
the power constraint. What remains to be shown is to verify [21] R. Wang, X. Yuan, and M. Tao, “Degrees of freedom of MIMO multiway
for each pair of symbols, the other 6 symbols only occupy relay channel with clustered pairwise exchange,” IEEE J. Sel. Areas
Commun., vol. 33, no. 2, pp. 337–351, Feb. 2015.
4 dimensions, which are also linearly independent with the
vector carrying each symbol in that pair. Since it is not difficult
to verify, we omit the verification due to the space limitation.

R EFERENCES
[1] S. A. Jafar, “Interference alignment—A new look at signal dimensions in
a communication network,” Found. Trends Commun. Inf. Theory, vol. 7,
no. 1, pp. 1–134, 2010.
[2] C. Wang, T. Gou, and S. A. Jafar, “Subspace alignment chains and the
degrees of freedom of the three-user MIMO interference channel,” IEEE
Trans. Inf. Theory, vol. 60, no. 5, pp. 2432–2479, May 2014.
[3] C. M. Yetis, T. Gou, S. A. Jafar, and A. H. Kayran, “On feasibility of
interference alignment in MIMO interference networks,” IEEE Trans.
Signal Process., vol. 58, no. 9, pp. 4771–4782, Sep. 2010.
[4] N. Lee, J.-B. Lim, and J. Chun, “Degrees of freedom of the MIMO
Y channel: Signal space alignment for network coding,” IEEE Trans.
Inf. Theory, vol. 56, no. 7, pp. 3332–3342, Jul. 2010.
[5] Y. Wang, L. Yang, Y. Ai, H. Li, and W. Zhang, “Joint signal space
alignment and precoding in two-way relay multi-user networks,” in
Proc. 4th Int. Conf. Intell. Netw. Collaborative Syst. (INCoS), Bucharest,
Romania, Sep. 2012, pp. 385–389.
[6] K. Lee, N. Lee, and I. Lee, “Achievable degrees of freedom on
K-user Y channels,” IEEE Trans. Wireless Commun., vol. 11, no. 3,
pp. 1210–1219, Mar. 2012.
Chenwei Wang received the B.Eng. degree in infor-
mation engineering and the M.S. degree in commu-
nications and information systems from the Beijing
University of Posts and Telecommunications, China,
in 2005 and 2008, respectively, and the Ph.D. degree
in electrical and computer engineering from the
University of California, Irvine, CA, USA, in 2012.
Since 2013, he has been a Research Engineer with
the Mobile Network Technology Group, DOCOMO
Innovations Inc., Palo Alto, CA, USA. His current
research interests include the 5G key technologies,
network information theory, wireless communications, and networks. Before
joining DOCOMO Innovations Inc., his industry experience includes intern
positions with the Research, Technology and Platforms Group, Nokia Siemens
Networks, Beijing, China, in 2008, and the Network Innovation Lab, DoCoMo
USA Labs, Palo Alto, in 2010. He was a recipient of the 2014 IEEE Signal
Processing Society Young Author Best Paper Award, and the Honor of the
IEEE C OMMUNICATIONS L ETTERS Exemplary Reviewer in 2011 and 2014.

Authorized licensed use limited to: Aristotle University of Thessaloniki. Downloaded on March 14,2024 at 12:30:14 UTC from IEEE Xplore. Restrictions apply.