The Approximate Sum Capacity of the Symmetric

Gaussian K-User Interference Channel

Or Ordentlich
Tel Aviv University
ordent@eng.tau.ac.il
Uri Erez
Tel Aviv University
uri@eng.tau.ac.il
Bobak Nazer
Boston University
bobak@bu.edu
AbstractWe derive a new achievable sum rate for the
symmetric Gaussian K-user interference channel. This sum rate
is shown to be within a constant gap of the outer bound on the
sum capacity of this channel for all values of interference level
outside some outage set. The result is established through the use
of lattice interference alignment. A new lattice-based extension
to the Han-Kobayshi scheme is also introduced.
I. INTRODUCTION
Handling interference efciently is a major challenge in
multi-user wireless communication. Recently, it has become
clear that this challenge can often be overcome via interfer-
ence alignment [1], [2]. However, many schemes, such as
the Cadambe-Jafar framework [2] and ergodic interference
alignment [3], require many independent channel realizations
to achieve near-perfect alignment. For static channels, it is
known that similar gains are possible but existing schemes,
such as real interference alignment [4], only apply to the very
high signal-to-noise ratio (SNR) regime.
One promising direction is the use of lattice codes, as
they can enable alignment on the signal scale. Specically,
by taking advantage of the fact that the sum of lattice code-
words is itself a lattice codeword, a receiver can treat several
users as one effective user, thereby reducing the number of
effective interferers. For instance, this approach has been
used to characterize the approximate capacity of the many-
to-one interference channel [5]. The fully connected case
presents additional challenges, many of which are captured
by the special case of symmetric channels. In this paper, we
derive a new achievable rate region for the symmetric K-user
interference channel. Our schemes rely on the compute-and-
forward framework [6], which can efciently communicate
integer combinations of codewords by exploiting the linear
mixing provided by the channel. At a high level, implementing
interference alignment with compute-and-forward boils down
to recovering just two equations per receiver. That is, if the
transmitted codewords are denoted by x
m
, then a receiver can
recover the kth codeword by rst decoding the equations
a
11
x
k
+a
12

m =k
x
m
a
21
x
k
+a
22

m =k
x
m
,
This work was supported in part by the Israel Science Foundation under
Grant No. 1557/10, a fellowship from The Yitzhak and Chaya Weinstein
Research Institute for Signal Processing at Tel Aviv University and the Feder
Family Award.
where a
11
, a
12
, a
21
, and a
22
are integer-valued coefcients.
If the vectors a
1
= [a
11
a
12
]
T
and a
2
= [a
21
a
22
]
T
are
linearly independent, then the receiver can solve for its desired
codeword x
k
. This technique performs well strating in the
moderate SNR regime (see Figure 1), and sufces to achieve
rates within a constant gap of the symmetric capacity of the
symmetric K-user interference channel.
II. APPROXIMATE SUM CAPACITY
We consider the symmetric K-user Gaussian interference
channel. This channel consists of K transmitter-receiver pairs
that wish to simultaneously communicate across a shared
channel over n time slots, where the channel gains are constant
over all n channel uses. We assume a real-valued channel
model throughout. Each transmitter has a message w
k
drawn
independently and uniformly over {1, 2, . . . , 2
nRSYM
}, and is
equipped with an encoder, E
k
: {1, 2, . . . , 2
nRSYM
} R
n
, that
maps its message into a length-n channel input x
k
= E
k
(w
k
).
The channel inputs satisfy the power constraint x
k

nSNR, where SNR > 0 is the signal-to-noise ratio. The

channel output at each receiver is a noisy linear combination
of its desired signal and the sum of the interfering terms, of
the form
y
k
= x
k
+g

m =k
x
m
+z
k
, (1)
where g > 0 parametrizes the interference strength and z
k
is an i.i.d. Gaussian vector with mean 0 and variance 1. We
dene the interference-to-noise ratio to be INR g
2
SNR and
the interference level to be
1

log(INR)
log(SNR)
.
Each receiver is equipped with a decoder, D
k
: R
n

{1, 2, . . . , 2
nRSYM
}, that produces an estimate w
k
= D
k
(y
k
) of
its desired message w
k
. A symmetric rate R
SYM
is achievable
if, for any > 0 and n large enough, there exist encoders and
decoders that can attain probability of error at most . The
symmetric capacity C
SYM
is the supremum of all achievable
symmetric rates.
1
Note that our denition of INR ignores the fact that there are K 1
interferers observed at each receiver.
2012 IEEE International Symposium on Information Theory Proceedings
978-1-4673-2579-0/12/$31.00 2012 IEEE 2072
Remark 1: Due to the symmetry of the channel, the sym-
metric capacity is equal to the sum capacity, normalized by
the number of users.
Much of the previous work on this channel focused on
nding the degrees-of-freedom (DoF) [4], [7] and generalized
DoF [8] it offers. The DoF is dened as the ratio between the
symmetric-capacity of the channel and the single-user AWGN
channel capacity, in the limit of SNR . Surprisingly,
the DoF characterization is everywhere discontinuous [7] with
respect to g. This presents an obstacle towards a clean capacity
approximation at nite SNR.
To overcome this difculty, our approximations allow for
the possibility of an outage set. Specically, in the regime
around = 1, our capacity results take the following shape:
for any constant c > 0, the capacity is approximated within
c bits over the entire range of SNR, and all channel gains g,
except for a set of Lebesgue measure (c) which vanishes
rapidly with c. This type of capacity approximation is also
used in [9] and seems to arise from the capacity region itself,
not just the lower bound. That is, it appears that the capacity
may in fact simultaneously vary rapidly with the ne scale
of the channel gains (e.g., the distance to an appropriately
scaled integer) and slowly on the coarse scale (e.g., relative
interference strength). In the high SNR limit, this behavior
shows up as a discontinuity on the rationals but, at reasonable
SNRs, our achievable scheme shows that this variation is in
fact fairly smooth. The theorem below captures our capacity
approximations in a simple form.
Theorem 1: The symmetric capacity of the symmetric
Gaussian K-user interference channel can be lower and upper
bounded as follows:
2
Noisy Interference Regime, 0 <
1
2
,
1
2
log
_
1 +
SNR
1 + INR
_

1
2
log(K 1)
C
SYM
<
1
2
log
_
1 +
SNR
1 + INR
_
+ 1
Weak Interference Regime,
1
2
<
2
3
,
1
2
log
+
(INR)
7
2
log(K) C
SYM

1
2
log
+
(INR) + 1
for all channel gains.
Moderately Weak Interference Regime,
2
3
< 1,
1
2
log
+
_
SNR

INR
_
c 8 log(K)
C
SYM

1
2
log
+
_
SNR

INR
_
+ 1
for all channel gains except for an outage set of measure
< 2
c
for any c > 0.
Strong Interference Regime, 1 < 2,
1
4
log
+
(INR)
c
2
3 C
SYM

1
4
log
+
(INR) + 1
2
All logarithms are to the base 2, log
+
(x) max (0, log(x)).
for all channel gains except for an outage set whose mea-
sure is a fraction of 2
c
of the interval 1 < |g| <

SNR,
for any c > 0.
Very Strong Interference Regime, 2,
1
2
log(1 + SNR) 1 C
SYM

1
2
log(1 + SNR)
The outer bounds in Theorem 1 are based on the outer
bounds from [10] for the two-user interference channel. These
bounds for the K-user interference channel since, from the
perspective of each receiver, adding more interferers cannot
increase its own capacity.
In the noisy interference regime, the inner bound is ob-
tained simply by treating interference as noise. Conversely,
in the very strong interference regime, the inner bound can
be obtained by decoding and removing the interference, as
rst shown in [11]. Our main contributions are the inner
bounds for the strong, moderately weak and weak interference
regimes. These bounds rely on two different lattice-based
interference alignment schemes whose achievable rates are
stated in Theorems 2 and 3. Note that the rates are expressed in
terms of integer least-squares optimization problems. To obtain
our closed-form inner bounds in these regimes, we make use
of new bounds and techniques based on compute-and-forward
that are developed in [12]. Due to space limitations, proofs are
omitted and the reader is referred to [13] for the full details.
Figure 1 shows the achievable rates given by Theorems 2
and 3 for the three-user symmetric interference channel for
several values of SNR. The rates plotted in the gure were
found numerically by solving the corresponding integer least-
squares problems. For reference, we have also plotted the
upper bounds of [10] and the rate achievable by time division.
In the sequel we introduce the coding schemes used for
deriving Theorems 2 and 3.
III. PRELIMINARIES
A. Nested Lattice Codes
We employ the nested lattice framework originally proposed
in [14]. A lattice is a discrete subgroup of R
n
which is
closed under reection and real addition. Formally, for any
t
1
, t
2
, we have that t
1
, t
2
and t
1
+t
2
.
We denote the nearest neighbor quantizer associated with
the lattice by Q

() and the Voronoi region by V. The

modulo operation returns the quantization error w.r.t. the
lattice, [x] mod = x Q

(x).
A lattice is said to be nested in
1
if
1
. Our
schemes utilizes a chain of K + 1 nested lattices satisfying

K

1
. From these lattices, we construct K
codebooks, one for each user. Specically, user k is allocated
the codebook L
k
=
(k)
V, where V is the Voronoi region
of and the function (k) : {1, . . . , K} {1, . . . , K} maps
between users and lattices. The rate of each codebook L
k
is
R
k
=
1
/n log |
(k)
V|.
2073
10
1
10
0
10
1
0
0.5
1
1.5
2
2.5
3
3.5
g
S
u
m

r
a
t
e
[
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
]
(a) SNR = 20dB
10
2
10
0
10
2
0
1
2
3
4
5
6
g
S
u
m

r
a
t
e
[
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
]
(b) SNR = 35dB
10
2
10
0
10
2
0
2
4
6
8
g
S
u
m

r
a
t
e
[
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
]
(c) SNR = 50dB
10
2
10
0
10
2
0
2
4
6
8
10
g
S
u
m

r
a
t
e
[
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
]
(d) SNR = 65dB
Fig. 1. Upper and lower bounds on the sum capacity of a 3-user symmetric Gaussian interference channel with respect to the cross-gain g. The upper bound
(red line) is that of [10] and the lower bound (black line) is the maximum of the achievable rates from Theorems 2 and 3, which were computed numerically,
and the achievable rate for treating interference as noise. The lower bounds from Theorem 1 are not plotted in this gure. For reference, we have also plotted
the rate achievable via time-division (dotted blue line).
User k encodes its message onto a lattice point from its
codebook, t
k
L
k
. Each user also has a random
3
dither vector
d
k
which is generated independently and uniformly over V.
These dithers are made available to the decoder. The signal
transmitted by user k is x
k
= [t
k
d
k
] mod .
B. Effective multiple-access channel
A channel that often arises in the context of lattice interfer-
ence alignment is a K-user Gaussian multiple-access channel
(MAC) with integer-valued ratios between some of the channel
coefcients. Specically, the output of such a channel can be
written as
y =
L

=1
g

_

mK

b
m
x
m
_
+z, (2)
where K
1
, . . . , K
L
are disjoint subsets of {1, . . . , K}. We
assume that the b
m
Z are non-zero integers, which opens
up the possibility of lattice alignment.
3
It can be shown that these random dithers can be replaced with determin-
istic ones, meaning that no common randomness is required.
The channel (2) may describe the signal seen by a receiver
in an interference network, perhaps after appropriate precoding
at the transmitters. In such networks, each receiver is only
interested in the messages from some of the users while
the others act as interferers. Hence, it is benecial to align
several interfering users into one effective interferer, by taking
advantage of the fact that the sum of lattice codewords is itself
a lattice codeword.
Denition 1 (Effective users): For the MAC specied
by (2), we dene L effective users
x
eff,

mK

b
m
x
m
, = 1, . . . , L.
Denition 2 (Effective MAC): The K-user MAC (2) in-
duces the effective L-user MAC
y =
L

=1
g

x
eff,
+z, (3)
with the vector of effective channel coefcients g =
[g
1
g
L
]
T
R
L
. The effective channel is further char-
acterized by the effective users weights b
2
eff,
=

mK

b
2
m
2074
for = 1, . . . , L, and the effective (diagonal) weight matrix
B = diag(b
2
eff,1
, . . . , b
2
eff,L
).
Denition 3 (Effective lattice points): Let t
m
be the lattice
point transmitted by user m. We dene the effective lattice
point corresponding to effective user as
t
eff,
=
_

mK

b
m
t
m
_
mod .
Let
eff
() = min
mK

(m) be the index of the densest lattice

contributing to t
eff,
. Since all lattices are nested, it follows that
t
eff,

eff ()
.
When lattice interference alignment schemes are designed
properly, the message intended for the receiver is mapped into
a unique effective user, while multiple interfering users are
folded into a smaller number of effective users. In this case, it
sufces for the receiver to decode only the L effective lattice
points corresponding to the effective users, rather than the K
lattice points transmitted by all users. In our considerations,
the effective lattice points are recovered by rst decoding
L lattice equations of the type v =
_

L
=1
a

t
eff,
_
mod
with linearly independent coefcient vectors, and then solving
for t
eff,1
, . . . , t
eff,L
. We use the compute-and-forward frame-
work [6] for decoding the lattice equations.
Denition 4: The computation rate for decoding the lattice
equation v =
_

L
=1
a

t
eff,
_
mod with coefcient vector
a = [a
1
a
L
]
T
Z
L
over the effective MAC (3) is dened
as
R
comp
(g, a, B) =
1
2
log
_
SNR

2
eff
(g, a, B)
_
where

2
eff
(g, a, B) =
_
_
_
_
SNR
1
B
1
+gg
T
_
1/2
a
_
_
_
2
.
The following denition identies the L linearly indepen-
dent coefcient vectors which yield the highest computation
rates.
Denition 5: We say that an ordered set of linearly indepen-
dent integer coefcient vectors {a
1
, . . . , a
L
} with correspond-
ing computation rates R
comp,
R
comp
(g, a

, B) is optimal if
R
comp,1
R
comp,L
and for any = 1, . . . , L and any set
of integer coefcient vectors { a
1
, . . . , a

} of rank
min
i=1,...,
R
comp
(g, a
i
, B) R
comp,
. (4)
IV. ACHIEVABLE SCHEMES
In this section, we use the compute-and-forward framework
in order obtain two new achievable symmetric rate-regions.
We develop two coding schemes. The rst sufces to approx-
imate the symmetric capacity for the very strong, strong and
noisy interference regimes. The second scheme is a lattice-
based variant of the well-known Han-Kobayashi scheme [15],
which approximately achieves the symmetric capacity in the
moderately weak and weak interference regimes.
First scheme - A single-layer lattice code: This scheme uses a
pair of nested lattices
1
to construct the codebook L =

1
of rate R
SYM
. All users encode their messages using
this codebook. Since all interferers arrive at the kth transmitter
with the same gain, they will be aligned into one effective
lattice point. Thus, the K-user MAC seen by the kth receiver
becomes an effective two-user MAC of the form dened in
Section III-B
y
k
= x
eff,k1
+gx
eff,k2
+ z
k
, (5)
where x
eff,k1
= x
k
, x
eff,k2
=

m =k
x
m
are the effective
users, b
2
eff,k1
= 1, b
2
eff,k2
= K 1 are the effective users
weights and g = [1 g]
T
is the vector of channel gains.
The next theorem gives an achievable rate region for the K-
user interference channel when each receiver jointly decodes
both the effective user x
eff,k1
which carries the desired infor-
mation, and the effective user x
eff,k2
which carries the sum
of interfering codewords. The theorem relies on decoding two
independent linear combinations of the effective lattice points.
This is in contrast to the successive decoding technique used
in [11], where rst the interference is decoded and removed,
and only then the desired lattice point is decoded.
Theorem 2: Let R
comp,1
R
comp,2
be the optimal com-
putation rates for the effective MAC (5) induced by the
symmetric K-user interference channel (1). Any symmetric
rate R
SYM
< R
comp,2
is achievable for the symmetric K-user
interference channel (1).
For the two-user case, it is known that in the weak and
moderately weak interference regimes each receiver should
decode only part of the message transmitted by the other
user [10]. A natural extension of this Han-Kobayashi [15]
approach to the K-user case is for each receiver to decode
linear combinations that only include parts of the interfering
messages. This is enabled by using a superposition of two
lattice codewords at each transmitter. See Figure 2 for an
illustration.
Second scheme - Lattice Han-Kobayashi: This scheme em-
ploys a chain of nested lattices
2

1
to construct
two codebooks L
1
and L
2
with rates R
1
and R
2
, respectively.
Each user k splits its message w
k
into two messages, a public
message w
k1
that is mapped into codebook L
1
and a private
message w
k2
that is mapped into codebook L
2
. It is convenient
to treat each user k as two virtual users with codewords x
k1
and x
k2
that carry messages w
k1
and w
k2
, respectively. User
k transmits a superposition of its virtual users codewords,
x
k
=
_
1
2
x
k1
+x
k2
,
for [0, 1). The signal seen by the kth receiver is
y
k
=
_
1
2
x
k1
+x
k2
+g
_
1
2

m =k
x
m1
+g

m =k
x
m2
+z
k
, (6)
2075
Transmit
Equations Decoded by Receivers
x11 x12
a11x11 +a12x12 +a13

m =1
xm1 a21x11 +a22x12 +a23

m =1
xm1 a31x11 +a32x12 +a33

m =1
xm1
x21 x22
a11x21 +a12x22 +a13

m =2
xm1 a21x21 +a22x22 +a23

m =2
xm1 a31x21 +a32x22 +a33

m =2
xm1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xK1 xK2
a11xK1 +a12xK2 +a13

m =K
xm1 a21xK1 +a22xK2 +a23

m =K
xm1 a31xK1 +a32xK2 +a33

m =K
xm1
Fig. 2. Illustration of the lattice Han-Kobayashi scheme. Each transmitter sends a public (blue) and a private (red) lattice codeword. Each receiver decodes
three equations of the public codewords as well as its desired private codeword while treating the other private codewords as noise. From these equations,
the receivers can infer their desired public and private messages.
which induces the effective four-user MAC
y
k
=
_
1
2
x
eff,k1
+x
eff,k2
+g
_
1
2
x
eff,k3
+gx
eff,k4
+z
k
, (7)
with effective users x
eff,k1
= x
k1
, x
eff,k2
= x
k2
, x
eff,k3
=

m =k
x
m1
and x
eff,k4
=

m =k
x
m2
. The effective users
weights are b
2
eff,k1
= 1, b
2
eff,k2
= 1, b
2
eff,k3
= K 1 and
b
2
eff,k4
= K 1.
The receiver aims to decode the effective codewords x
eff,k1
,
x
eff,k2
and x
eff,k3
while treating the fourth effective code-
word x
eff,k4
as noise. This is done by using the compute-
and-forward framework to decode three independent linear
combinations of x
eff,k1
, x
eff,k2
and x
eff,k3
in which x
eff,k4
does not participate. It can be shown [13] that for this aim
the effective channel (7) is equivalent to the effective three-
user MAC
y
k
=
_
_
1
2
x
eff,k1
+x
eff,k2
+g
_
1
2
x
eff,k3
_
+z
k
,
where = 1/
_
1 + (K 1)SNRg
2

2
, and x
eff,k1
, x
eff,k2
,
x
eff,k3
, b
2
eff,k1
, b
2
eff,k2
, b
2
eff,k3
are as before.
As in the two-user interference channel, the value of
should be optimized such as to maximize the achievable
rates for decoding the effective users x
eff,k1
and x
eff,k2
from
the effective three-user MAC. We follow the approach taken
in [10] and choose
2
= 1/(g
2
SNR). With this choice, we
can obtain the following theorem (see [13]).
Theorem 3: Assume g
2
SNR > 1 and consider the effective
MAC
y
k
=

g
2
SNR 1
K g
2
SNR
x
eff,k1
+
_
1
K g
2
SNR
x
eff,k2
+g

g
2
SNR 1
K g
2
SNR
x
eff,k3
+z
k
, (8)
with effective users weights b
2
eff,k1
= 1, b
2
eff,k2
= 1, and
b
2
eff,k3
= K1. Let {a
HK
1
, a
HK
2
, a
HK
3
} and R
HK
comp,1
R
HK
comp,2

R
HK
comp,3
be the optimal coefcient vectors and computation
rates for this effective MAC. Any symmetric rate
R
SYM
< R
HK
comp,2
+R
HK
comp,3
is achievable for the symmetric K-user interference chan-
nel (1).
REFERENCES
[1] M. A. Maddah-Ali, A. S. Motahari, and A. K. Khandani, Communi-
cation over MIMO X channels: Interference alignment, decomposition,
and performance analysis, IT Trans., Aug. 2008.
[2] V. R. Cadambe and S. A. Jafar, Interference alignment and the degrees
of freedom for the K-user interference channel, IT Trans., Aug. 2008.
[3] B. Nazer, M. Gastpar, S. A. Jafar, and S. Vishwanath, Ergodic inter-
ference alignment, IT Trans., Submitted Aug. 2011, available online
http://arxiv.org/abs/0901.4379.
[4] A. S. Motahari, S. O. Gharan, M.-A. Maddah-Ali, and A. K. Khan-
dani, Real interference alignment: Exploiting the potential of single
antenna systems, IT Trans., Submitted Nov. 2009, available online
http://arxiv.org/abs/0908.2282.
[5] G. Bresler, A. Parekh, and D. Tse, The approximate capacity of
the many-to-one and one-to-many Gaussian interference channels, IT
Trans., Sept. 2010.
[6] B. Nazer and M. Gastpar, Compute-and-forward: Harnessing interfer-
ence through structured codes, IT Trans., Oct. 2011.
[7] R. Etkin and E. Ordentlich, The degrees-of-freedom of the K-user
Gaussian interference channel is discontinuous at rational channel
coefcients, IT Trans., Nov. 2009.
[8] S. A. Jafar and S. Vishwanath, Generalized degrees of freedom of the
symmetric Gaussian K-user interference channel, IT Trans., July 2010.
[9] U. Niesen and M. A. Maddah-Ali, Interference alignment: From
degrees-of-freedom to constant-gap capacity approximations, IT Trans.,
Submitted Dec. 2011, available online: http://arxiv.org/abs/1112.4879.
[10] R. H. Etkin, D. N. C. Tse, and H. Wang, Gaussian interference channel
capacity to within one bit, IT Trans., Dec. 2008.
[11] S. Sridharan, A. Jafarian, S. Vishwanath, and S. A. Jafar, Capacity
of symmetric K-user Gaussian very strong interference channels, in
GLOBECOM, Dec. 2008.
[12] O. Ordentlich, U. Erez, and B. Nazer, The compute-and-forward
transform, in ISIT, July 2012.
[13] , The approximate sum capacity of the symmetric Gaussian K-
user interference channel, IT Trans., Submitted May 2012, available
online.
[14] U. Erez and R. Zamir, Achieving
1
2
log (1 + SNR) on the AWGN
channel with lattice encoding and decoding, IT Trans., Oct. 2004.
[15] T. S. Han and K. Kobayashi, A new achievable rate region for the
interference channel, IT Trans., Jan. 1981.
2076