a

r
X
i
v
:
0
8
0
9
.
1
3
4
4
v
5


[
c
s
.
I
T
]


2
1

J
a
n

2
0
1
0
1
The Balanced Unicast and Multicast Capacity
Regions of Large Wireless Networks
Urs Niesen, Piyush Gupta, and Devavrat Shah
Abstract
We consider the question of determining the scaling of the n
2
-dimensional balanced unicast and the n2
n
-
dimensional balanced multicast capacity regions of a wireless network with n nodes placed uniformly at random in
a square region of area n and communicating over Gaussian fading channels. We identify this scaling of both the
balanced unicast and multicast capacity regions in terms of (n), out of 2
n
total possible, cuts. These cuts only
depend on the geometry of the locations of the source nodes and their destination nodes and the trafc demands
between them, and thus can be readily evaluated. Our results are constructive and provide optimal (in the scaling
sense) communication schemes.
I. INTRODUCTION
Characterizing the capacity region of wireless networks is a long standing open problem in information
theory. The exact capacity region is, in fact, not known for even simple networks like a three node relay
channel or a four node interference channel. In this paper, we consider the question of approximately
determining the unicast and multicast capacity regions of wireless networks by identifying their scaling
in terms of the number n of nodes in the network.
A. Related Work
In the last decade, exciting progress has been made towards approximating the capacity region of
wireless networks. We shall mention a small subset of work related to this paper.
We rst consider unicast trafc. The unicast capacity region of a wireless network with n nodes is the
set of all simultaneously achievable rates between all possible n
2
source-destination pairs. Since nding
this unicast capacity region of a wireless network exactly seems to be intractable, Gupta and Kumar
proposed a simpler but insightful question in [1]. First, instead of asking for the entire n
2
-dimensional
unicast capacity region of a wireless network with n nodes, attention was restricted to the scenario where
each node is source exactly once and chooses its destination uniformly at random from among all the other
nodes. All these n source-destination pairs communicate at the same rate, and the interest is in nding
the maximal achievable such rate. Second, instead of insisting on nding this maximal rate exactly, they
focused on its asymptotic behavior as the number of nodes n grows to innity.
This setup has indeed turned out to be more amenable to analysis. In [1], it was shown that under
random placement of nodes in a given region and under certain models of communication motivated by
current technology (called protocol channel model in the following), the per-node rate for random source-
destination pairing with uniform trafc can scale at most as O(n
1/2
) and this can be achieved (within
poly-logarithmic factor in n) by a simple scheme based on multi-hop communication. Many works since
This work was supported, in parts, by DARPA grant (ITMANET) 18870740-37362-C, by NSF grants CCR-0325673 and CNS-0519535,
and by a AFOSR grant under the complex networks program. The material in this paper was presented, in parts, at the Allerton Conference
on Communication, Control, and Computing, Monticello, IL, September 2008, and at the IEEE INFOCOM Conference, Rio de Janeiro,
Brazil, April 2009.
U. Niesen was with the Laboratory for Information and Decision Systems at the Massachusetts Institute of Technology. He is now with
the Mathematics of Networks and Communications Research Department, Bell Labs, Alcatel-Lucent. Email: [email protected]
P. Gupta is with the Mathematics of Networks and Communications Research Department, Bell Labs, Alcatel-Lucent. Email:
[email protected]
D. Shah is with the Laboratory for Information and Decision Systems at the Massachusetts Institute of Technology. Email: [email protected]
2
then have broadened the channel and communication models under which similar results can be proved
(see, for example, [2][13]). In particular, under the Gaussian fading channel model with a power-loss of
r

for signals sent over a distance of r, it was shown in [12] that in extended wireless networks (i.e., n
nodes are located in a region of area (n)) the largest uniformly achievable per-node rate under random
source-destination pairing scales essentially like
_
n
1min{3,}/2
_
.
Analyzing such random source-destination pairing with uniform trafc yields information about the n
2
-
dimensional unicast capacity region along one dimension. Hence, the results in [1] and in [12] mentioned
above provide a complete characterization of the scaling of this one-dimensional slice of the capacity
region for the protocol and Gaussian fading channel models, respectively. It is therefore natural to ask
if the scaling of the entire n
2
-dimensional unicast capacity region can be characterized. To this end, we
describe two related approaches taken in recent works.
One approach, taken by Madan, Shah, and L ev eque [14], builds upon the celebrated works of Leighton
and Rao [15] and Linial, London, and Rabinovich [16] on the approximate characterization of the unicast
capacity region of capacitated wireline networks. For such wireline networks, the scaling of the unicast
capacity region is determined (within a log(n) factor) by the minimum weighted cut of the network graph.
As shown in [14], this naturally extends to wireless networks under the protocol channel model, providing
an approximation of the unicast capacity region in this case.
Another approach, rst introduced by Gupta and Kumar [1], utilizes geometric properties of the wireless
network. Specically, the notion of the transport capacity of a network, which is the rate-distance product
summed over all source-destination pairs, was introduced in [1]. It was shown that in an extended wireless
network with n nodes and under the protocol channel model, the transport capacity can scale at most as
(n). This bound on the transport capacity provides a hyper-plane which has the capacity region and
origin on the same side. Through a repeated application of this transport capacity bound at different scales
[17], [18] obtained an implicit characterization of the unicast capacity region under the protocol channel
model.
For the Gaussian fading channel model, asymptotic upper bounds for the transport capacity were
obtained in [2], [3], and for more general distance weighted sum rates in [19].
So far, we have only considered unicast trafc. We now turn to multicast trafc. The multicast capacity
region of a wireless network with n nodes is the set of all simultaneously achievable rates between all
possible n2
n
sourcemulticast-group pairs. Instead of considering this multicast capacity region directly,
various authors have analyzed the scaling of restricted trafc patterns under a protocol channel model
assumption (see [20][24], among others). For example, in [20], Li, Tang, and Frieder obtained a scaling
characterization under a protocol channel model and random node placement for multicast trafc when
each node chooses a certain number of its destinations uniformly at random. Independently, in [21],
Shakkottai, Liu, and Srikant considered a similar setup and also obtained the precise scaling when sources
and their multicast destinations are chosen at random. Both of these results are non information-theoretic
(in that they assume a protocol channel model). Furthermore, they provide information about the scaling
of the n2
n
-dimensional multicast capacity region only along one particular dimension.
B. Our Contributions
Despite the long list of results, the question of approximately characterizing the unicast capacity region
under the Gaussian fading channel model remains far from being resolved. In fact, for Gaussian fading
channels, the only trafc pattern that is well understood is random source-destination pairing with uniform
rate. This is limiting in several aspects. First, by choosing for each source a destination at random,
most source-destination pairs will be at a distance of the diameter of the network with high probability,
i.e., at distance (

n) for an extended network. However, in many wireless networks some degree of

locality of source-destination pairs can be expected. Second, all source-destination pairs are assumed to
be communicating at uniform rate. Again, in many settings we would expect nodes to be generating
trafc at widely varying rates. Third, each node is source exactly once, and destination on average once.
3
However, in many scenarios the same source node (e.g., a server) might transmit data to many different
destination nodes, or the same destination node might request data from many different source nodes. All
these heterogeneities in the trafc demands can result in different scaling behavior of the performance of
the wireless network than what is obtained for random source-destination pairing with uniform rate.
As is pointed out in the last section, even less is known about the multicast capacity region under
Gaussian fading. In fact, the only available results are for the protocol channel model, and even there
only for special trafc patterns resulting from randomly choosing sources and their multicast groups and
assuming uniform rate. To the best of our knowledge, no information-theoretic results (i.e., assuming
Gaussian fading channels) are available even for special trafc patterns.
We address these issues by analyzing the scaling of a broad class of trafc, termed balanced trafc in
the following, in a wireless network of n randomly placed nodes under a Gaussian fading channel model.
The notion of balanced trafc is a natural generalization of symmetric trafc, in which the data to be
transmitted from a node u to a node v is equal to the amount of data to be transmitted from v to u. We
analyze the scaling of the set of achievable balanced unicast trafc (the balanced unicast capacity region)
and achievable balanced multicast trafc (the balanced multicast capacity region). The balanced unicast
capacity region provides information about n
2
n of the n
2
dimensions of the unicast capacity region;
the balanced multicast capacity region provides information about n2
n
n of the n2
n
dimensions of the
multicast capacity region.
As a rst set of results of this paper, we present an approximate characterization of the balanced
unicast and multicast capacity regions. We show that both regions can be approximated by a polytope
with less than 2n faces, each corresponding to a distinct cut (i.e., a subset of nodes) in the wireless
network. This polyhedral characterization provides a succinct approximate description of the balanced
unicast and multicast capacity regions even for large values of n. Moreover, it shows that only 2n out of
2
n
possible cuts in the wireless network are asymptotically relevant and reveals the geometric structure
of these relevant cuts.
Second, we establish the approximate equivalence of the wireless network and a wireline tree graph,
in the sense that balanced trafc can be transmitted reliably over the wireless network if and only if
approximately the same trafc can be routed over the tree graph. This equivalence is the key component
in the derivation of the approximation result for the balanced unicast and multicast capacity regions and
provides insight into the structure of large wireless networks.
Third, we propose a novel three-layer communication architecture that achieves (in the scaling sense)
the entire balanced unicast and multicast capacity regions. The top layer of this scheme treats the wireless
network as the aforementioned tree graph and routes messages between sources and their destinations
dealing with heterogeneous trafc demands. The middle layer of this scheme provides this tree abstraction
to the top layer by appropriately distributing and concentrating trafc over the wireless networkchoosing
the level of cooperation in the network. The bottom layer implements this distribution and concentration
of messages in the wireless networkdealing with interference and noise. The approximate optimality
of this three-layer architecture implies that a separation based approach, in which routing is performed
independently of the physical layer, is order-optimal. In other words, techniques such as network coding
can provide at most a small (in the scaling sense) multiplicative gain for transmission of balanced unicast
or multicast trafc in wireless networks.
C. Organization
The remainder of this paper is organized as follows. Section II introduces the network model and
notation. Section III presents our main results. We illustrate the strength of these results in Section IV
by analyzing various example scenarios with heterogeneous unicast and multicast trafc patterns. Section
V provides a high level description of the proposed communication schemes. Sections VI-VIII contain
proofs. Finally, Sections IX and X contain discussions and concluding remarks.
4
II. MODELS AND NOTATION
In this section, we discuss network and trafc models, and we introduce some notational conventions.
A. Network Model
Consider the square region
A(n) [0,

n]
2
and let V (n) A(n) be a set of |V (n)| = n nodes on A(n). Each such node represents a wireless device,
and the n nodes together form a wireless network. This setting with n nodes on a square of area n is
referred to as an extended network. Throughout this paper, we consider this extended network setting.
However, all results carry over for dense networks, where n nodes are placed on a square of unit area
(see Section IX-E for the details).
We use the same channel model as in [12]. Namely, the received signal at node v and time t is
y
v
[t]

uV (n)\{v}
h
u,v
[t]x
u
[t] + z
v
[t]
for all v V (n), t N, where the {x
u
[t]}
u,t
are the signals sent by the nodes in V (n). We impose
an average power constraint of 1 on the signal {x
u
[t]}
t
for every node u V (n). The additive noise
terms {z
v
[t]}
v,t
are independent and identically distributed (i.i.d.) circularly symmetric complex Gaussian
random variables with mean 0 and variance 1, and
h
u,v
[t] r
/2
u,v
exp(

1
u,v
[t]),
for path-loss exponent > 2, and where r
u,v
is the Euclidean distance between u and v. As a function of
u, v V (n), we assume that {
u,v
[t]}
u,v
are i.i.d.
1
with uniform distribution on [0, 2). As a function of t,
we either assume that {
u,v
[t]}
t
is stationary and ergodic, which is called fast fading in the following, or
we assume {
u,v
[t]}
t
is constant, which is called slow fading in the following. In either case, we assume
full channel state information (CSI) is available at all nodes, i.e., each node knows all {h
u,v
[t]}
u,v
at
time t. This full CSI assumption is rather strong, and so it is worth commenting on. All the converse
results presented are proved under the full CSI assumption and are hence also valid under more realistic
assumptions on the availability of CSI. Moreover, it can be shown that for achievability only 2-bit quantized
CSI is necessary for path-loss exponent (2, 3] and no CSI is necessary for > 3 to achieve the same
scaling behavior.
B. Trafc Model
A unicast trafc matrix
UC
R
nn
+
associates with each pair u, w V (n) the rate
UC
u,w
at which node
u wants to communicate to node w. We assume that messages for distinct source-destination pairs (u, w)
are independent. However, we allow the same node u to be source for multiple destinations, and the same
node w to be destination for multiple sources. In other words, we consider general unicast trafc. The
unicast capacity region
UC
(n) R
nn
+
of the wireless network is the collection of achievable unicast
trafc matrices, i.e.,
UC

UC
(n) if and only if every source-destination pair (u, w) V (n) V (n) can
reliably communicate independent messages at rate
UC
u,w
.
A multicast trafc matrix
MC
R
n2
n
+
associates with each pair u V (n), W V (n) the rate
MC
u,W
at which node u wants to multicast a message to the nodes in W. In other words, all nodes in W want
to obtain the same message from u. We assume that messages for distinct sourcemulticast-group pairs
1
It is worth pointing out that recent results [25] suggest that under certain assumptions on scattering elements, for (2, 3) and very
large values of n, the i.i.d. phase assumption does not accurately reect the physical behavior of the wireless channel. However, in follow-up
work [26] the authors show that under different assumptions on the scatterers this assumption is still justied in the (2, 3) regime even
for very large values of n. This indicates that the issue of channel modeling for large networks in the low path-loss regime is somewhat
delicate and requires further investigation.
5
(u, W) are independent. However, we allow the same node u to be source for several multicast-groups,
and the same set W of nodes to be multicast destination for multiple sources. In other words, we consider
general multicast trafc. The multicast capacity region
MC
(n) R
n2
n
+
is the collection of achievable
multicast trafc matrices, i.e,.
MC

MC
(n) if and only if every sourcemulticast-group pair (u, W) can
reliably communicate independent messages at rate
MC
u,W
.
The following example illustrates the concept of unicast and multicast trafc matrices.
Example 1. Assume n = 4, and label the nodes as {u
i
}
4
i=1
= V (n). Assume further node u
1
needs to
transmit a message m
1,2
to node u
2
at rate 1 bit per channel use, and an independent message m
1,3
to
node u
3
at rate 2 bits per channel use. Node u
2
needs to transmit a message m
2,3
to node u
3
at rate 4 bits
per channel use. All the messages m
1,2
, m
1,3
, m
2,3
are independent. This trafc pattern can be described
by a unicast trafc matrix
UC
R
44
+
with
UC
u
1
,u
2
= 1,
UC
u
1
,u
3
= 2,
UC
u
2
,u
3
= 4, and
UC
u,w
= 0 otherwise.
Note that in this example node u
1
is source for two (independent) messages, and node u
3
is destination
for two (again independent) messages. Node u
4
in this example is neither source nor destination for any
message and can be understood as a helper node.
Assume now that node u
1
needs to transmit the same message m
1,{2,3,4}
to all nodes u
2
, u
3
, u
4
at a rate
of 1 bit per channel use, and an independent message m
1,{2}
to only node 2 at rate 2 bits per channel
use. Node 2 needs to transmit a message m
2,{1,3}
to both u
1
, u
3
at rate 4 bits per channel use. All the
messages m
1,{2,3,4}
, m
1,{2}
, m
2,{1,3}
are independent. This trafc pattern can be described by a multicast
trafc matrix
MC
R
416
+
with
MC
u
1
,{u
2
,u
3
,u
4
}
= 1,
MC
u
1
,{u
2
}
= 2,
MC
u
2
,{u
1
,u
3
}
= 4, and
MC
u,W
= 0 otherwise.
Note that in this example node u
1
is source for two (independent) multicast messages, and node 2 and
3 are destinations for more than one message. The message m
1,{2,3,4}
is destined for all the nodes in the
network and can hence be understood as a broadcast message. The message m
1,{2}
is only destined for
one node and can hence be understood as a private message.
In the following, we will be interested in balanced trafc matrices that satisfy certain symmetry
properties. Consider a symmetric unicast trafc matrix
UC
satisfying
UC
u,w
=
UC
w,u
for all node pairs
u, w V (n). The notion of a balanced trafc matrix generalizes this idea of symmetric trafc.
Before we provide a precise denition of balanced trafc, we need to introduce some notation. Partition
A(n) into several square-grids. The -th square-grid divides A(n) into 4

squares, each of sidelength

2

n, denoted by {A
,i
(n)}
4

i=1
. Let V
,i
(n) V (n) be the nodes in A
,i
(n) (see Figure 1). The square
grids in levels {1, . . . , L(n)} with
2
L(n)
1
2
log(n)
_
1 log
1/2
(n)
_
,
will be of particular importance. Note that L(n) is chosen such that
4
L(n)
n = n
log
1/2
(n)
,
and hence
lim
n

A
L(n),i
(n)

= lim
n
4
L(n)
n = .
while at the same time

A
L(n),i
(n)

= 4
L(n)
n n
o(1)
,
as n . In other words, the area of the region A
L(n),i
(n) at level = L(n) grows to innity as n ,
but much slower than n.
A unicast trafc matrix
UC
is -balanced if

u/ V
,i
(n)

wV
,i
(n)

UC
u,w

uV
,i
(n)

w/ V
,i
(n)

UC
u,w
, (1)
2
All logarithms are with respect to base 2.
6
Fig. 1. Square-grids with 0 2. The grid at level = 0 is the area A(n) itself. The grid at level = 1 is indicated by the dashed
lines. The grid at level = 2 by the dashed and the dotted lines. Assume for the sake of example that the subsquares are numbered from
left to right and then from bottom to top (the precise order of numbering is immaterial). Then V0,1(n) are all the nodes V (n), V1,1(n) are
the nine nodes in the lower left corner (separated by dashed lines), and V2,1(n) are the three nodes in the lower left corner (separated by
dotted lines).
for all {1, . . . , L(n)} and i {1, . . . 4

}. In other words, for a balanced unicast trafc matrix the

amount of trafc to the nodes V
,i
(n) is not much larger than the amount of trafc from them. In particular,
all symmetric trafc matrices, i.e., satisfying
UC
u,w
=
UC
w,u
, are 1-balanced. Denote by B
UC
(n) R
nn
+
the
collection of all (n)-balanced unicast trafc matrices for some xed (n) = n
o(1)
. In the following, we
refer to trafc matrices
UC
B
UC
(n) simply as balanced trafc matrices. The balanced unicast capacity
region
BUC
(n) R
nn
+
of the wireless network is the collection of balanced unicast trafc matrices that
are achievable, i.e.,

BUC
(n)
UC
(n) B
UC
(n).
Note that (1) imposes at most n linear inequality constraints, and hence
UC
(n) and
BUC
(n) coincide
along at least n
2
n of n
2
total dimensions.
A multicast trafc matrix
MC
is -balanced if

u/ V
,i
(n)

WV (n):
WV
,i
(n)=

MC
u,W

uV
,i
(n)

WV (n):
W\V
,i
(n)=

MC
u,W
(2)
for all {1, . . . , L(n)}, i {1, . . . 4

}. Thus, for -balanced multicast trafc, the amount of trafc to the

nodes V
,i
(n) is not much larger than the amount of trafc from them. This is the natural generalization of
the notion of -balanced unicast trafc to the multicast case. Denote by B
MC
(n) R
n2
n
+
the collection
of all (n)-balanced multicast trafc matrices for some xed (n) = n
o(1)
. As before, we will refer to a
multicast trafc matrix
MC
B
MC
(n) simply as balanced multicast trafc matrix. The balanced multicast
capacity region
BMC
(n) R
n2
n
+
of the wireless network is the collection of balanced multicast trafc
matrices that are achievable, i.e.,

BMC
(n)
MC
(n) B
MC
(n).
Equation (2) imposes at most n linear inequality constraints, and hence
MC
(n) and
BMC
(n) coincide
along at least n2
n
n of n2
n
total dimensions.
C. Notational Conventions
Throughout, {K
i
}
i
, K,

K, . . . , indicate strictly positive nite constants independent of n and . To
simplify notation, we assume, when necessary, that large real numbers are integers and omit and
7
operators. For the same reason, we also suppress dependence on n within proofs whenever this dependence
is clear from the context.
III. MAIN RESULTS
In this section, we present the main results of this paper. In Section III-A, we provide an approximate
(i.e., scaling) characterization of the entire balanced unicast capacity region
BUC
(n) of the wireless
network, and in Section III-B, we provide a scaling characterization of the entire balanced multicast
capacity region
BMC
(n). In Section III-C, we discuss implications of these results on the behavior of the
unicast and multicast capacity regions for large values of n. In Section III-D, we consider computational
aspects.
A. Balanced Unicast Capacity Region
Here we present a scaling characterization of the complete balanced unicast capacity region
BUC
(n).
Dene

UC
(n)
_

UC
R
nn
+
:

uV
,i
(n)

w/ V
,i
(n)

UC
u,w
(4

n)
2min{3,}/2
{1, . . . , L(n)}, i {1, . . . , 4

},

w=u
(
UC
u,w
+
UC
w,u
) 1 u V (n)
_
,
(3)
and set

BUC
(n)

UC
(n) B
UC
(n).

BUC
(n) is the collection of all balanced unicast trafc matrices
UC
such that for various cuts S V (n)
in the network, the total trafc demand (in either one or both directions)

uS

w/ S

UC
u,w
,

uS

w/ S
(
UC
u,w
+
UC
w,u
),
across the cut S is not too big. Note that the number of cuts S we need to consider is actually quite
small. In fact, there are at most n cuts of the form S = V
,i
(n) for {1, . . . , L(n)}, and there are n
cuts of the form S = {u} for u V (n). Hence

BUC
(n) is described by at most 2n cuts.
The next theorem shows that

BUC
(n) is approximately (in the scaling sense) equal to the balanced
unicast capacity region
BUC
(n) of the wireless network.
Theorem 1. Under either fast or slow fading, for any > 2, there exist
b
1
(n) n
o(1)
,
b
2
(n) = O(log
6
(n)),
such that
b
1
(n)

BUC
(n)
BUC
(n) b
2
(n)

BUC
(n),
with probability 1 o(1) as n .
We point out that Theorem 1 holds only with probability 1 o(1) for different reasons for the fast and
slow fading cases. Under fast fading, the theorem holds only for node placements that are regular enough.
The node placement itself is random, and we show that the required regularity property is satised with
high probability as n . Under slow fading, the theorem holds under the same regularity requirements
8
on the node placement, but now it also only holds with high probability for the realization of the fading
{
u,v
}
u,v
.
Theorem 1 provides a tight scaling characterization of the entire balanced unicast capacity region

BUC
(n) of the wireless network as depicted in Figure 2. The approximation is within a factor n
o(1)
.
This factor can be further sharpened as is discussed in detail in Section IX-B.

UC
1,2

UC
2,1
b
2
(n)

BUC
(n)

BUC
(n)
b
1
(n)

BUC
(n)
Fig. 2. The set

BUC
(n) approximates the balanced unicast capacity region
BUC
(n) of the wireless network in the sense that b1(n)

BUC
(n)
(with b1(n) n
o(1)
) provides an inner bound to
BUC
(n) and b2(n)

BUC
(n) (with b2(n) = O

log
6
(n)

) provides an outer bound to

BUC
(n). The gure shows two dimensions (namely
UC
1,2
and
UC
2,1
) of the n
2
-dimensional set
BUC
(n).
We point out that for large values of path-loss exponent ( > 5) the restriction to balanced trafc can
be removed, yielding a tight scaling characterization of the entire n
2
-dimensional unicast capacity region

UC
(n). See Section IX-D for the details. For (2, 5], bounds on achievable rates for trafc that is not
balanced are discussed in Section IX-C.
B. Balanced Multicast Capacity Region
We now present an approximate characterization of the complete balanced multicast capacity region

BMC
(n).
Dene

MC
(n)
_

MC
R
n2
n
+
:

uV
,i
(n)

WV (n):
W\V
,i
(n)=

MC
u,W
(4

n)
2min{3,}/2
{1, . . . , L(n)}, i {1, . . . , 4

},

WV (n):
W\{u}=

MC
u,W
+

u=u

WV (n):
uW

MC
u,W
1 u V (n)
_
,
(4)
and set

BMC
(n)

MC
(n) B
MC
(n).
The denition of

BMC
(n) is similar to the denition of

BUC
(n) in (3).

BMC
(n) is the collection of
all balanced multicast trafc matrices
MC
such that for various cuts S V (n) in the network, the total
9
trafc demand (in either one or both directions)

uS

WV (n):
W\S=

MC
u,W
,

uS

WV (n):
W\S=

MC
u,W
+

u/ S

WV (n):
WS=

MC
u,W
,
across the cut S is not too big. Note that, unlike in the denition of

BUC
(n), we count
u,W
as crossing
the cut S if u S and W \ S = , i.e., if there is at least one node w in the multicast destination group
W that lies outside S. The number of such cuts S we need to consider is at most 2n, as in the unicast
case.
The next theorem shows that

BMC
(n) is approximately (in the scaling sense) equal to the balanced
multicast capacity region
BMC
(n) of the wireless network.
Theorem 2. Under either fast or slow fading, for any > 2, there exist
b
3
(n) n
o(1)
,
b
4
(n) = O(log
6
(n)),
such that
b
3
(n)

BMC
(n)
BMC
(n) b
4
(n)

BMC
(n),
with probability 1 o(1) as n .
As with Theorem 1, Theorem 2 holds only with probability 1 o(1) for different reasons for the
fast and slow fading cases. Theorem 2 implies that the quantity

BMC
(n) determines the scaling of the
balanced multicast capacity region
BMC
(n). The approximation is up to a factor n
o(1)
as in the unicast
case, and can again be sharpened (see the discussion in Section IX-B). As in the unicast case, for > 5
the restriction of balanced trafc can be dropped resulting in a scaling characterization of the entire n2
n
-
dimensional multicast capacity region
MC
(n). The details can be found in Section IX-D. Similarly, we
can obtain bounds on achievable rates for trafc that is not balanced, as is discussed in Section IX-C.
C. Implications of Theorems 1 and 2
Theorems 1 and 2 can be applied in two ways. First, the theorems can be used to analyze the asymptotic
achievability of a sequence of trafc matrices. Consider the unicast case, and let {
UC
(n)}
n1
be a
sequence of balanced unicast trafc matrices with
UC
(n) R
nn
+
. Dene

UC
(n) sup{ :
UC
(n)
BUC
(n)},

UC
(n) sup{ :
UC
(n)

BUC
(n)},
i.e.,

UC
(n) is the largest multiplier such that the scaled trafc matrix
UC
(n) is contained in
BUC
(n)
(and similar for

UC
(n) with respect to

BUC
(n)). Then Theorem 1 provides asymptotic information about
the achievability of {
UC
(n)}
n1
in the sense that
3
lim
n
log(

UC
(n))
log(n)
= lim
n
log(

UC
(n))
log(n)
.
Theorem 2 can be used similarly to analyze sequences of balanced multicast trafc matrices. Several
applications of this approach are explored in Section IV.
3
We assume here that the limits exist, otherwise the same statement holds for limsup and liminf.
10
Second, Theorems 1 and 2 provide information about the shape of the balanced unicast and multicast
capacity regions
BUC
(n) and
BMC
(n). Consider again the unicast case. We now argue that even though
the approximation

BUC
(n) of
BUC
(n) is only up to n
o(1)
scaling, its shape is largely preserved.
To illustrate this point, consider a rectangle
R(n) [0, r
1
(n)] [0, r
2
(n)],
and let

R(n) [0, r
1
(n)] [0, r
2
(n)],
where
r
i
b
i
(n)r
i
(n)
for some b
i
(n) = n
o(1)
, be its approximation. The shape of R(n) is then determined by the ratio between
r
1
(n) and r
2
(n). For example, assume r
1
(n) = n

r
2
(n). Then
r
1
(n)
r
2
(n)
= n
o(1)
= n
o(1)
r
1
(n)
r
2
(n)
,
i.e.,
lim
n
log
_
r
1
(n)/r
2
(n)
_
log(n)
= = lim
n
log
_
r
1
(n)/ r
2
(n)
_
log(n)
,
and hence the approximation

R(n) preserves the exponent of the ratio of sidelengths of R(n). In other
words, if the two sidelengths r
1
(n) and r
2
(n) differ on exponential scale (i.e., by a factor n

for = 0)
then this shape information is preserved by the approximation

R(n).
Let us now return to the balanced unicast capacity region
BUC
(n) and its approximation

BUC
(n). We
consider several boundary points of
BUC
(n) and show that their behavior varies at scale n

for various
values of . From the discussion in the previous paragraph, this implies that a signicant part of the
shape of
BUC
(n) is preserved by its approximation

BUC
(n). First, let
UC
(n)1 for some scalar (n)
depending only on n, and where 1 is the n n matrix of all ones. If
UC

BUC
(n) then the largest
achievable value of (n) is

(n) n
min{3,}/2+o(1)
(by applying Theorem 1). Second, let
UC
such
that
UC
u

,w
=
UC
w

,u
= (n) for only one source-destination pair (u

, w

) with u

= w

and
UC
u,w
= 0
otherwise. Then

(n), the largest achievable value of (n), satises

(n) n
o(1)
. Hence the boundary
points of
BUC
(n) vary at least from n
min{3,}/2+o(1)
to n
o(1)
, and this variation on exponential scale
is preserved by

BUC
(n).
Again, a similar analysis is possible also for the multicast capacity region, showing that the approximate
balanced multicast capacity region

BMC
(n) preserves the shape of the balanced multicast capacity region

BMC
(n) on exponential scale.
D. Computational Aspects
Since we are interested in large wireless networks, computational aspects are of importance. In this
section, we show that the approximate characterizations

BUC
(n) and

BMC
(n) in Theorems 1 and 2
provide a computationally efcient approximate description of the balanced unicast and multicast capacity
regions
BUC
(n) and
BMC
(n), respectively.
Consider rst the unicast case. Note that
BUC
(n) is a n
2
-dimensional set, and hence its shape could be
rather complicated. In particular, in the special cases where the capacity region is known, its description
is often in terms of cut-set bounds. Since there are 2
n
possible subsets of n nodes, there are 2
n
possible
cut-set bounds to be considered. In other words, the description complexity of
BUC
(n) is likely to be
growing exponentially in n. On the other hand, as was pointed out in Section III-A, the description of

BUC
(n) is in terms of only 2n cuts. This implies that

BUC
(n) can be computed efciently (i.e., in
polynomial time in n). Hence even though the description complexity of
BUC
(n) is likely to be of order
11
(2
n
), the description complexity of its approximation

BUC
(n) is only of order (n)an exponential
reduction. In particular, this implies that membership
UC

BUC
(n) (and hence by Theorem 1 also the
approximate achievability of the balanced unicast trafc matrix
UC
) can be computed in polynomial time
in the network size n. More precisely, evaluating each of the (n) cuts takes at most (n
2
) operations,
yielding a (n
3
)-time algorithm for approximate testing of membership in
BUC
(n).
Consider now the multicast case.
BMC
(n) is a n2
n
-dimensional set, i.e., the number of dimensions is
exponentially large in n. Nevertheless, its approximation

BMC
(n) can (as in the unicast case) be computed
by evaluating at most 2n cuts. This yields a very compact approximate representation of the balanced
multicast capacity region
BMC
(n) (i.e., we represent a region of exponential size in n as an intersection
of only linearly many halfspacesone halfspace corresponding to each cut). Moreover, it implies that
membership
MC

BMC
(n) can be computed efciently. More precisely, evaluating each of the (n)
cuts takes at most |{(u, W) :
MC
u,W
> 0}| operations. Thus membership
MC

BMC
(n) (and hence by
Theorem 2 also the approximate achievability of the balanced multicast trafc matrix
MC
) can be tested
in at most (n) times more operations than required to just read the problem parameters. In other words,
we have a linear time (in the length of the input) algorithm for testing membership of a multicast trafc
matrix
MC
in

BMC
(n), and hence for approximate testing of membership in
BMC
(n). However, this
algorithm is not necessarily polynomial time in n, since reading just the input
MC
R
n2
n
+
itself might
take exponential time in n.
IV. EXAMPLE SCENARIOS
We next illustrate the above results by determining achievable rates in a few specic wireless network
scenarios with non-uniform trafc patterns.
Example 2. Multiple classes of source-destination pairs
There are K classes of source-destination pairs for some xed K. Each source node in class i generates
trafc at the same rate
i
(n) for a destination node that is chosen randomly within distance (n

i
/2
), for
some xed
i
[0, 1]. Each node randomly picks the class it belongs to. The resulting trafc matrix is
balanced (with (n) = n
o(1)
) with high probability, and applying Theorem 1 shows that

i
(n), the largest
achievable value of
i
(n), satises

i
(n) = n

i
(1 /2)o(1)
,
with probability 1 o(1) for all i, and where
min{3, }. (5)
Hence, for a xed number of classes K, source nodes in each class can obtain rates as a function of only
the source-destination separation in that class.
Set n
i
n

i
, and note that n
i
is on the order of the expected number of nodes that are closer to a
source than its destination. Then

i
(n) = n
o(1)
n
1 /2
i
.
Now n
1 /2
i
is precisely the per-node rate that is achievable for an extended network with n
i
nodes under
random source-destination pairing [12]. In other words, the local trafc pattern here allows us to obtain
a rate that is as good as the one achievable under random source-destination pairing for a much smaller
network.
Example 3. Trafc variation with source-destination separation
Assume each node is source for exactly one destination, chosen uniformly at random from among all
the other nodes (as in the traditional setting). However, instead of all sources generating trafc at the
12
same rate, source node u generates trafc at a rate that is a function of its separation from destination w,
i.e., the trafc matrix is given by
UC
u,w
= (r
u,w
) for some function . In particular, let us consider
(r) (n)
_
r

if r 1,
1 else,
for some xed R and some (n) depending only on n. The traditional setting corresponds to = 0,
in which case all n source-destination pairs communicate at uniform rate.
While such trafc is not balanced for small values of , the results in Section IX-C, extending Theorem 1
to trafc that is not balanced, can be used to establish the scaling of

(n), the largest achievable value

of (n), as

(n) =
_
n
1( +)/2o(1)
if 2 ,
n
o(1)
else,
with probability 1 o(1). For = 0, and noting that 2 0, this recovers the results from [12] for
random source-destination pairing with uniform rate.
Example 4. Sources with multiple destinations
All the example scenarios so far are concerned with trafc in which each node is source exactly once.
Here we consider more general trafc patterns. There are K classes of source nodes, for some xed
K. Each source node in class i has (n

i
) destination nodes for some xed
i
[0, 1] and generates
independent trafc at the same rate
i
(n) for each of them (i.e., we still consider unicast trafc). Each of
these destination nodes is chosen uniformly at random among the n1 other nodes. Every node randomly
picks the class it belongs to. Noting that the resulting trafc matrix is balanced with high probability,
Theorem 1 provides the following scaling of the rates achievable by different classes:

i
(n) = n
1
i
/2o(1)
,
with probability 1o(1) for all i. In other words, for each source node time sharing between all K classes
and then (within each class) between all its (n

i
) destination nodes is order-optimal in this scenario.
However, different sources are operating simultaneously.
Example 5. Broadcast
Consider a scenario with every node u in the network broadcasting an independent message to all other
nodes at rate (n)
u
. In other words, we have a multicast trafc matrix of the form

MC
u,W
=
_
(n)
u
if W = V (n),
0 else,
for some (n) > 0. Applying the generalization in Section IX-C of Theorem 2 yields that

(n), the
largest achievable (n), satises

(n) = n
o(1)
1

uV (n)

u
as n .
V. COMMUNICATION SCHEMES
In this section, we provide a high-level description of the communication schemes used to prove
achievability (i.e., the inner bound) in Theorems 1 and 2. In Section V-A, we present a communication
scheme for general unicast trafc, in Section V-B we show how this scheme can be adapted for general
multicast trafc. Both schemes use as a building block a communication scheme introduced in prior
work for a particular class of trafc, called uniform permutation trafc. In such uniform permutation
trafc, each node in the network is source and destination exactly once, and all these source-destination
13
pairs communicate at equal rate. For (2, 3], the order-optimal scheme for such uniform permutation
trafc (called hierarchical relaying scheme in the following) enables global cooperation in the network.
For > 3, the order-optimal scheme is multi-hop routing. We recall these two schemes for uniform
permutation trafc in Section V-C.
A. Communication Scheme for Unicast Trafc
In this section, we present a scheme to transmit general unicast trafc. This scheme has a tree structure
that makes it convenient to work with. This tree structure is crucial in proving the compact approximation
of the balanced unicast capacity region
BUC
(n) in Theorem 1.
The communication scheme consists of three layers: A top or routing layer, a middle or cooperation
layer, and a bottom or physical layer. The routing layer of this scheme treats the wireless network as a
tree graph G and routes messages between sources and their destinationsdealing with heterogeneous
trafc demands. The cooperation layer of this scheme provides this tree abstraction G to the top layer
by appropriately distributing and concentrating trafc over the wireless networkchoosing the level of
cooperation in the network. The physical layer implements this distribution and concentration of messages
in the wireless networkdealing with interference and noise.
Seen from the routing layer, the network consists of a noiseless capacitated graph G. This graph is a
tree, whose leaf nodes are the nodes V (n) in the wireless network. The internal nodes of G represent
larger clusters of nodes (i.e., subsets of V (n)) in the wireless network. More precisely, each internal node
in G represents a set V
,i
(n) for {1, . . . , L(n)} and i {1, . . . , 4

}. Consider two sets V

,i
(n), V
+1,j
(n)
and let , be the corresponding internal nodes in G. Then and are connected by an edge in G if
V
+1,j
(n) V
,i
(n). Similarly, for V
L(n),i
(n) and corresponding internal node in G, a leaf node u in
G is connected by an edge to if u V
L(n),i
(n) (recall that the leaf nodes of G are the nodes V (n)
in the wireless network). This construction is shown in Figure 3. In the routing layer, messages are sent
Fig. 3. Construction of the tree graph G. We consider the same nodes as in Figure 1 with L(n) = 2. The leaves of G are the nodes V (n)
of the wireless network. They are always at level = L(n) + 1 (i.e., 3 in this example). At level 0 L(n) in G, there are 4

nodes.
The tree structure is the one induced by the grid decomposition {V
,i
(n)}
,i
as shown in Figure 1. Level 0 contains the root node of G.
from each source to its destination by routing them over G. To send information along an edge of G, the
routing layer calls upon the cooperation layer.
The cooperation layer implements the tree abstraction G. This is done by ensuring that whenever a
message is located at a node in G, it is evenly distributed over the corresponding cluster in the wireless
network, i.e., every node in the cluster has access to a distinct part of equal length of the message. To
send information from a child node to its parent in G (i.e., towards the root node of G), the message
at the cluster in V (n) represented by the child node is distributed evenly among all nodes in the bigger
cluster in V (n) represented by the parent node. More precisely, let be a child node of in G, and let
V
+1,i
(n), V
,j
(n) be the corresponding subsets of V (n). Consider the cooperation layer being called by
the routing layer to send a message from to its parent over G. In the wireless network, we assume
each node in V
+1,i
(n) has access to a distinct 1/ |V
+1,i
(n)| fraction of the message to be sent. Each node
in V
+1,i
(n) splits its message part into four distinct parts of equal length. It keeps one part for itself and
14
sends the other three parts to three nodes in V
,j
(n) \ V
+1,i
(n). After each node in V
+1,i
(n) has sent
its message parts, each node in V
,j
(n) now as access to a distinct 1/ |V
,j
(n)| fraction of the message.
To send information from a parent node to a child node in G (i.e., away from the root node of G), the
message at the cluster in V (n) represented by the parent node is concentrated on the cluster in V (n)
represented by the child node. More precisely, consider the same nodes and in G corresponding to
V
+1,i
(n) and V
,j
(n) in V (n). Consider the cooperation layer being called by the routing layer to send
a message from to its child . In the wireless network, we assume each node in V
,j
(n) has access to
a distinct 1/ |V
,j
(n)| fraction of the message to be sent. Each node in V
,j
(n) sends its message part to
another node in V
+1,i
(n). After each node in V
,j
(n) has sent its message part, each node in V
+1,i
(n)
now as access to a distinct 1/ |V
+1,i
(n)| fraction of the message. To implement this distribution and
concentration of messages, the cooperation layer calls upon the physical layer.
The physical layer performs the distribution and concentration of messages. Note that the trafc induced
by the cooperation layer in the physical layer is very regular, and closely resembles a uniform permutation
trafc (in which each node in the wireless network is source and destination once and all these source-
destination pairs want to communicate at equal rate). Hence we can use either cooperative communication
(for (2, 3]) or multi-hop communication (for > 3) for the transmission of this trafc. See Section
V-C for a detailed description of these two schemes. It is this operation in the physical layer that determines
the edge capacities of the graph G as seen from the routing layer.
The operation of this three-layer architecture is illustrated in the following example.
Example 6. Consider a single source-destination pair (u, w). The corresponding operation of the three-
layer architecture is depicted in Figure 4.
u
w
Fig. 4. Example operation of the three-layer architecture under unicast trafc. The three layers depicted are (from top to bottom in the
gure) the routing layer, the cooperation layer, and the physical layer.
In the routing layer, the message is routed over the tree graph G between u and w (indicated in black
in the gure). The middle plane in the gure shows the induced behavior from using the second edge
along this path (indicated in solid black in the gure) in the cooperation layer. The bottom plane in the
gure shows (part of) the corresponding actions induced in the physical layer. Let us now consider the
specic operations of the three layers for the single message between u and w. Since G is a tree, there is
a unique path between u and w, and the routing layer sends the message over the edges along this path.
Consider now the rst such edge. Using this edge in the routing layer induces the following actions in
15
the cooperation layer. The node u, having access to the entire message, splits that message into 3 distinct
parts of equal length. It keeps one part, and sends the other two parts to the two other nodes in V
2,1
(n)
(i.e., lower left square at level = 2 in the hierarchy). In other words, after the message has traversed the
edge between u and its parent node in the routing layer, all nodes in V
2,1
(n) in the cooperation layer have
access to a distinct 1/3 fraction of the original message. The edges in the routing layer leading up the
tree (i.e., towards the root node) are implemented in the cooperation layer in a similar fashion by further
distributing the message over the wireless network. By the time the message reaches the root node of G
in the routing layer, the cooperation layer has distributed the message over the entire network and every
node in V (n) has access to a distinct 1/n fraction of the original message. Communication down the
tree in the routing layer is implemented in the cooperation layer by concentrating messages over smaller
regions in the wireless network. To physically perform this distribution and concentration of messages,
the cooperation layer calls upon the physical layer, which uses either hierarchical relaying or multi-hop
communication.
B. Communication Scheme for Multicast Trafc
Here we show that the same communication scheme presented in the last section for general unicast
trafc can also be used to transmit general multicast trafc. Again it is the tree structure of the scheme that
is critically exploited in the proof of Theorem 2 providing an approximation for the balanced multicast
capacity region
BMC
(n).
We will use the same three-layer architecture as for unicast trafc presented in Section V-A. To
accommodate multicast trafc, we only modify the operation of the top or routing layer; the lower layers
operate as before.
We now outline how the routing layer needs to be adapted for the multicast case. Consider a multicast
message that needs to be transmitted from a source node u V (n) to its set of intended destinations
W V (n). In the routing layer, we want to route this message from u to W over G. Since G is a tree,
the routing part is simple. In fact, between u and every w W there exists a unique path in G. Consider
the union of all those paths. It is easy to see that this union is a subtree of G. Indeed, it is the smallest
subtree of G that covers {u} W. Trafc is optimally routed over G from u to W by sending it along
the edges of this subtree.
The next example illustrates the operation of the routing layer under multicast trafc.
Example 7. Consider one source node u and the corresponding multicast group W {w
1
, w
2
, w
3
} as
shown in Figure 5.
u
w1
w2
w3
Fig. 5. Example operation of the routing layer in the three-layer architecture under multicast trafc.
In the routing layer, we nd the smallest subgraph G({u} W) covering {u} W (indicated by
black lines in Figure 5). Messages are sent from the source to its destinations by routing them along this
subgraph. In other words, G({u} W) is the multicast tree along which the message is sent from u to
W. The cooperation layer and physical layer operate in the same way as for unicast trafc (see Figure 4
for an example).
16
C. Communication Schemes for Uniform Permutation Trafc
Here we recall communication schemes for uniform permutation trafc on A(n), i.e., each node is
source and destination exactly once and all these n pairs communicate at uniform rate. As pointed out in
Sections V-A and V-B, these communication schemes are used as building blocks in the communication
architecture for general unicast and multicast trafc.
The structure of the optimal communication scheme depends drastically on the path-loss exponent . For
(2, 3] (small path-loss exponent), cooperative communication on a global scale is necessary to achieve
optimal performance. For > 3 (large path-loss exponent), local communication between neighboring
nodes is sufcient, and trafc is routed in a multi-hop fashion from the source to the destination. We
will refer to the order-optimal scheme for (2, 3] as hierarchical relaying scheme, and to the order
optimal scheme for > 3 as multi-hop scheme. For a uniform permutation trafc on V (n), hierarchical
relaying achieves a per-node rate of n
1/2o(1)
; multi-hop communication achieves a per-node rate of
n
1/2o(1)
. By choosing the appropriate scheme (hierarchical relaying for (2, 3], multi-hop for > 3),
we can thus achieve a per-node rate of n
1min{3,}/2o(1)
. We provide a short description of the hierarchical
relaying scheme in the following. The details can be found in [13].
Consider n nodes placed independently and uniformly at random on A(n). Divide A(n) into
n
1
1+log
1/3
(n)
squarelets of equal size. Call a squarelet dense, if it contains a number of nodes proportional to its area.
For each source-destination pair, choose such a dense squarelet as a relay, over which it will transmit
information (see Figure 6).
u
1
u
2
u
3
MAC
BC
w
1
w
2
w
3
Fig. 6. Sketch of one level of the hierarchical relaying scheme. Here {(ui, wi)}
3
i=1
are three source-destination pairs. Groups of source-
destination pairs relay their trafc over dense squarelets (shaded), which contain a number of nodes proportional to their area. We time share
between the different relay squarelets. Within each relay squarelet the scheme is used recursively to enable joint decoding and encoding at
the relay.
Consider now one such relay squarelet and the nodes that are transmitting information over it. If we
assume for the moment that the nodes within the relay squarelets could cooperate, then between the
source nodes and the relay squarelet we would have a multiple access channel (MAC), where each source
node has one transmit antenna, and the relay squarelet (acting as one node) has many receive antennas.
Between the relay squarelet and the destination nodes, we would have a broadcast channel (BC), where
each destination node has one receive antenna, and the relay squarelet (acting again as one node) has
many transmit antennas. The cooperation gain from using this kind of scheme arises from the use of
multiple antennas for this MAC and BC.
17
To actually enable this kind of cooperation at the relay squarelet, local communication within the relay
squarelets is necessary. It can be shown that this local communication problem is actually the same as
the original problem, but at a smaller scale. Indeed, we are now considering a square of size
n
1
1
1+log
1/3
(n)
with equal number of nodes (at least order wise). Hence we can use the same scheme recursively to
solve this subproblem. We terminate the recursion after log
1/3
(n) iterations, at which point we use simple
time-division multiple access (TDMA) to bootstrap the scheme.
Observe that at the nal level of the scheme, we have divided A(n) into
_
n
1
1+log
1/3
(n)
_
log
1/3
(n)
= n
1
1+log
1/3
(n)
squarelets. A sufcient condition for the scheme to succeed is that all these squarelets are dense (i.e.,
contain a number of nodes proportional to their area). However much weaker conditions are sufcient as
well, see [13].
For any permutation trafc, the per-node rate achievable with this scheme is at least n
1/2o(1)
for any
> 2 and under fast fading. Under slow fading the same per-node rate is achievable for all permutation
trafc with probability at least
1 exp
_
2
(log
2/3
(n))
_
.
Moreover, when (2, 3] and for uniform permutation trafc with a constant fraction of source-
destination pairs at distance (

n) (as is the case with high probability if the permutation trafc is

chosen at random), this is asymptotically the best uniformly achievable per-node rate.
VI. AUXILIARY LEMMAS
In this section, we provide auxiliary results, which will be used several times in the following. These
results are grouped into three parts. In Section VI-A, we describe regularity properties exhibited with high
probability by the random node placement. In Section VI-B, we provide auxiliary upper bounds on the
performance of any scheme in terms of cut-set bounds. Finally, in Section VI-C, we describe auxiliary
results on the performance of hierarchical relaying and multi-hop communication as described in Section
V-C.
A. Regularity Lemmas
Here we prove several regularity properties that are satised with high probability by a random node
placement. Formally, dene V(n) to be the collection of all node placements V (n) that satisfy the following
conditions:
r
u,v
> n
1
for all u, v V (n),

V
,i
(n)

log(n) for =
1
2
log(n),

V
,i
(n)

1 for =
1
2
log
_
n
2 log(n)
_
,

V
,i
(n)

[4
1
n, 4
+1
n] for all
_
1, . . . , L

(n)
_
,
where
L

(n)
1
2
log(n)
_
1
1
2
log
5/6
(n)
_
,
and in each case i {1, . . . , 4

}. The rst condition is that the minimum distance between node pairs is
not too small. The second condition is that all squares of area 1 contain at most log(n) nodes. The third
18
condition is that all squares of area 2 log(n) contain at least one node. The fourth condition is that all
squares up to level
1
2
log(n)
_
1
1
2
log
5/6
(n)
_
contain a number of nodes proportional to their area. Note
that, since
L(n) =
1
2
log(n)
_
1 log
1/2
(n)
_
=
1
2
log(n)
_
1
1
2
log
5/6
(n)
_
,
this holds in particular for nodes up to level L(n). The goal of this section is to prove that
P(V (n) V(n)) = 1 o(1),
as n .
The rst lemma shows that the minimum distance in a random node placement is at least n
1
with
high probability.
Lemma 3.
P
_
min
uV (n),vV (n)\{u}
r
u,v
> n
1
_
= 1 o(1),
as n .
Proof. For u, v V , let
B
u,v
{r
u,v
r}
for some r (depending only on n). Fix a node u V , then for v = u
P(B
u,v
|u)
r
2

n
(the inequality being due to boundary effects). Moreover, the events {B
u,v
}
vV \{u}
are independent
conditioned on u, and thus
P
_

vV \{u}
B
c
u,v

u
_
=

vV \{u}
P(B
c
u,v
|u)

_
1
r
2

n
_
n
.
From this,
P
_
min
uV,vV \{u}
r
u,v
r
_
= P
_

uV,vV \{u}
B
u,v
_

uV
P
_

vV \{u}
B
u,v
_
=

uV
_
1 P
_

vV \{u}
B
c
u,v
_
_
=

uV
_
1 E
_
P
_

vV \{u}
B
c
u,v

u
__
_

uV
_
1
_
1
r
2

n
_
n
_
= n
_
1
_
1
r
2

n
_
n
_
.
19
Assuming r <
_
n/, we have
n
_
1
_
1
r
2

n
_
n
_
nr
2
,
and hence
P
_
min
uV,vV \{u}
r
u,v
r
_
nr
2
,
which converges to zero for r = n
1
.
The next lemma asserts that if

L(n) is not too large then all squares {V
,i
(n)}
,i
for {1, . . . ,

L(n)}
and i {1, . . . , 4

} in the grid decomposition of V (n) contain a number of nodes that is proportional to

their area.
Lemma 4. If

L(n) satises
lim
n

L(n)
4

L(n)
n
= 0
then
P
_

L(n)

=1
4

i=1
_
|V
,i
(n)| [4
1
n, 4
+1
n]
_
_
= 1 o(1)
as n . In particular, this holds for

L(n) =
1
2
log(n)
_
1
1
2
log
5/6
(n)
_
,
and for

L(n) = L(n).
Proof. Let B
u
be the event that node u lies in A
,i
for xed , i. Note that

uV
1 1
Bu
= |V
,i
|
by denition, and that
P(B
u
) = 4

.
Hence, using the Chernoff bound,
P
_
|V
,i
| [4
1
n, 4
+1
n]
_
= P
_

uV
1 1
Bu
[4
1
n, 4
+1
n]
_
exp(K4

n),
for some positive constant K, and we obtain, for =

L(n),
P
_
4

L(n)

i=1
_
|V

L(n),i
| [4

L(n)1
n, 4

L(n)+1
n]
_
_
1
4

L(n)

i=1
P
_
|V

L(n),i
| [4

L(n)1
n, 4

L(n)+1
n]
_
1 4

L(n)
exp(K4

L(n)
n)
1 exp(

L(n) K4

L(n)
n), (6)
for some positive constant

K. By assumption
lim
n

L(n)
4

L(n)
n
= 0,
20
and hence
P
_
4

L(n)

i=1
_
|V

L(n),i
| [4

L(n)1
n, 4

L(n)+1
n]
_
_
1 o(1),
as n . Since the {A
,i
}
,i
are nested as a function of , we have

L(n)

=1
4

i=1
_
|V
,i
| [4
1
n, 4
+1
n]
_
=
4

L(n)

i=1
_
|V

L(n),i
| [4

L(n)1
n, 4

L(n)+1
n]
_
,
which, combined with (6), proves the rst part of the lemma.
For the second part, note that for

L(n) =
1
2
log(n)
_
1
1
2
log
5/6
(n)
_
,
we have

L(n)
4

L(n)
n
=
1
2
log(n)
_
1
1
2
log
5/6
(n)
_
2
1
2
log
1/6
(n)

log(n)
2
1
2
log
1/6
(n)
= 2
log log(n)
1
2
log
1/6
(n)
0,
and hence the lemma is valid in this case. The same holds for

L(n) = L(n) since
L(n)
1
2
log(n)
_
1
1
2
log
5/6
(n)
_
.
We are now ready to prove that a random node placement V (n) is in V(n) with high probability as
n (i.e., is fairly regular with high probability).
Lemma 5.
P(V (n) V(n)) = 1 o(1),
as n .
Proof. The rst condition,
r
u,v
> n
1
for all u, v V ,
holds with probability 1 o(1) by Lemma 3. The second and third conditions,

V
,i

log(n) for =
1
2
log(n),

V
,i

1 for =
1
2
log
_
n
2 log(n)
_
,
are shown in [12, Lemma 5.1] to hold with probability 1 o(1). The fourth condition,

V
,i

[4
1
n, 4
+1
n] for all
_
1, . . . , L

(n)
_
,
holds with probability 1 o(1) by Lemma 4. Together, this proves the result.
21
B. Converse Lemmas
Here we prove several auxiliary converse results. The rst lemma bounds the maximal achievable sum
rate for every individual node (i.e., the total trafc for which a xed node is either source or destination).
Lemma 6. Under either fast or slow fading, for any > 2, there exists b(n) = O(log(n)) such that for
all V (n) V(n),
UC

UC
(n), u V (n),

wV (n)\{u}

UC
u,w
b(n), (7)

wV (n)\{u}

UC
w,u
b(n). (8)
Proof. The argument follows the one in [12, Theorem 3.1]. Denote by C(S
1
, S
2
) the multiple-input
multiple-output (MIMO) capacity between nodes in S
1
and nodes in S
2
, for S
1
, S
2
V . Consider rst
(7). By the cut-set bound [27, Theorem 14.10.1],

w=u

UC
u,w
C({u}, {u}
c
).
C({u}, {u}
c
) is the capacity between u and the nodes in {u}
c
, i.e.,
C({u}, {u}
c
) = log
_
1 +

v=u
|h
u,v
|
2
_
log(1 + (n 1)n

)
K log(n),
with
K 2 + ,
and where for the rst inequality we have used that since V V, we have r
u,v
n
1
for all u, v V .
Similarly, for (8),

w=u

UC
w,u
C({u}
c
, {u}),
and
C({u}
c
, {u}) log
_
1 + (n 1)

v=u
|h
v,u
|
2
_
log(1 + (n 1)
2
n

)
K log(n).
The next lemma bounds the maximal achievable sum rate across the boundary out of the subsquares
V
,i
(n) for {1, . . . , L(n)}, and i {1, . . . , 4

}.
Lemma 7. Under either fast or slow fading, for any > 2, there exists b(n) = O
_
log
6
(n)
_
such that
for all V (n) V(n),
UC

UC
(n), {1, . . . , L(n)}, and i {1, . . . , 4

}, we have

uV
,i
(n)

w/ V
,i
(n)

UC
u,w
b(n)(4

n)
2min{3,}/2
.
Proof. As before, denote by C(S
1
, S
2
) the MIMO capacity between nodes in S
1
and nodes in S
2
. By the
cut-set bound [27, Theorem 14.10.1],

uV
,i

w/ V
,i

UC
u,w
C(V
,i
, V
c
,i
). (9)
22
Let
H
S
1
,S
2
[h
u,v
]
uS
1
,vS
2
be the matrix of channel gains between the nodes in S
1
and S
2
. Under fast fading
C(S
1
, S
2
) max
Q(H)0:
E(qu,u)P uS
1
E
_
log det
_
I +H

S
1
,S
2
Q(H)H
S
1
,S
2
_
_
,
and under slow fading
C(S
1
, S
2
) max
Q0:
qu,uP uS
1
log det
_
I +H

S
1
,S
2
QH
S
1
,S
2
_
.
Denote by (V
c
,i
) the nodes in V
c
,i
that are within distance one of the boundary between A
c
,i
and A
,i
.
Using the generalized Hadamard inequality yields that under either fast or slow fading
C(V
,i
, V
c
,i
) C(V
,i
, (V
c
,i
)) + C(V
,i
, V
c
,i
\ (V
c
,i
)). (10)
We start by analyzing the rst term in the sum in (10). Applying Hadamards inequality again yields
C(V
,i
, (V
c
,i
))

v(V
c
,i
)
C(V
,i
, {v}).
Since V V, we have

(V
c
,i
)

5 log(n)(4

n)
1/2
.
By the same analysis as in Lemma 6, we obtain
C(V
,i
, {v}) C({v}
c
, {v})
K
5
log(n)
for some constant K (independent of v). Therefore
C(V
,i
, (V
c
,i
)) 5 log(n)(4

n)
1/2
K
5
log(n)
K log
2
(n)(4

n)
1/2
. (11)
We now analyze the second term in the sum in (10). The arguments of [13, Lemma 12] (building
on [12, Theorem 5.2]) show that under either fast or slow fading there exists

K > 0 such that for any
V V, {0, . . . , L(n)},
C(V
,i
, V
c
,i
\ (V
c
,i
))

K log
3
(n)

uV
,i

vV
c
,i
\(V
c
,i
)
r

u,v
. (12)
Moreover, using the same arguments as in [12, Theorem 5.2] shows that there exists a constant K

> 0
such that for adjacent squares (i.e., sharing a side) A
,i
, A
,j
,

uV
,i

vV
,j
\(V
c
,i
)
r

u,v
K

log
3
(n)(4

n)
2min{3,}/2
. (13)
Consider now two diagonal squares (i.e., sharing a corner point) A
,i
, A
,j
. Using a similar argument and
suitably redening K

shows that (13) holds for diagonal squares as well.

Using this, we now compute the summation in (12). Consider rings of squares around A
,i
. The rst
such ring contains the (at most) 8 squares neighboring A
,i
. The next ring contains at most 16 squares.
In general, ring k contains at most 8k squares. Let
{A
,j
}
jI
k
23
be the squares in ring k. Then

uV
,i

vV
c
,i
\(V
c
,i
)
r

u,v
=

k1

jI
k

uV
,i

vV
,j
\(V
c
,i
)
r

u,v
. (14)
By (13),

jI
1

uV
,i

vV
,j
\(V
c
,i
)
r

u,v
8K

log
3
(n)(4

n)
2min{3,}/2
. (15)
Now note that for k > 1 and j I
k
, nodes u V
,i
and v V
,j
are at least at distance r
u,v

(k 1)(2

n). Moreover, since V V, each {V

,j
}
,j
has cardinality at most 4
+1
n. Thus

k>1

jI
k

uV
,i

vV
,j
\(V
c
,i
)
r

u,v

k>1
8k
_
4
+1
n
_
2
_
(k 1)(2

n)
_

= 128
_
4

n
_
2/2

k>1
k(k 1)

= K

_
4

n
_
2/2
, (16)
for some K

> 0, and where we have used that > 2. Substituting (15) and (16) into (14) yields

uV
,i

vV
c
,i
\(V
c
,i
)
r

u,v
8K

log
3
(n)(4

n)
2min{3,}/2
+ K

_
4

n
_
2/2
,
and hence by (12)
C(V
,i
, V
c
,i
\ (V
c
,i
))

K log
3
(n)
_
8K

log
3
(n)(4

n)
2min{3,}/2
+ K

_
4

n
_
2/2
_
. (17)
Combining (9), (10), (11), and (17) shows that

uV
,i

v/ V
,i

UC
u,v
b(n)(4

n)
2min{3,}/2
.
for every {1, . . . , L(n)}, i {1, . . . , 4

}, and under either fast or slow fading.

The following lemma bounds the maximal achievable sum rate across the boundary into the subsquares
V
,i
(n) for {1, . . . , L(n)}, and i {1, . . . , 4

}. Note that this lemma is only valid for > 5.

Lemma 8. Under either fast or slow fading, for any > 5, there exists b(n) = O
_
log
3
(n)
_
such that
for all V (n) V(n),
UC

UC
(n), {1, . . . , L(n)}, and i {1, . . . , 4

}, we have

u/ V
,i
(n)

wV
,i
(n)

UC
u,w
b(n)(4

n)
1/2
.
Proof. By the cut-set bound [27, Theorem 14.10.1],

u/ V
,i

wV
,i

UC
u,w
C(V
c
,i
, V
,i
). (18)
Denote by V
,i
the nodes in V
,i
that are within distance one of the boundary between A
c
,i
and A
,i
.
Applying the generalized Hadamard inequality as in Lemma 7, we have under either fast or slow fading
C(V
c
,i
, V
,i
) C(V
c
,i
, V
,i
) + C(V
c
,i
, V
,i
\ V
,i
)
K log
2
(n)(4

n)
1/2
+ C(V
c
,i
, V
,i
\ V
,i
),
(19)
for some positive constant K.
24
For the second term in (19), we have by slightly adapting the upper bound from Theorem 2.1 in [3]:
C(V
c
,i
, V
,i
\ V
,i
)

vV
,i
\V
,i
_

uV
c
,i
r
/2
u,v
_
2
.
Now, consider v V
,i
\ V
,i
and let d
v
be the distance of v from the closest node in V
c
,i
. Using V V
and > 5,

uV
c
,i
r
/2
u,v


K log(n)d
2/2
v
,
for some positive constant

K, and hence
C(V
c
,i
, V
,i
\ V
,i
)

vV
,i
\V
,i

K
2
log
2
(n)d
4
v
K

log
3
(n)(4

n)
1/2
,
for some positive constant K

. Combined with (19) and (18), this proves Lemma 8.

C. Achievability Lemmas
In this section, we prove auxiliary achievability results. Recall that a permutation trafc is a trafc pattern
in which each node is source and destination exactly once. Call the corresponding source-destination
pairing V (n) V (n) a permutation pairing. The lemma below analyzes the performance achievable
with either hierarchical relaying (for (2, 3]) or multi-hop communication (for > 3) applied
simultaneously to transmit permutation trafc in several disjoint regions in the network. See Section
V-C for a description of these communication schemes.
Lemma 9. Under fast fading, for any > 2, there exists b(n) n
o(1)
such that for all V (n) V(n),
{0, . . . , L(n)}, i {1, . . . , 4

}, and permutation source-destination pairing

i
on V
,i
(n), there exists

UC

UC
(n) such that
min
i{1,...,4

}
min
(u,w)
i

UC
u,w
b(n)(4

n)
1min{3,}/2
.
The same statement holds with probability 1 o(1) as n in the slow fading case.
Consider the source-destination pairing
i

i
with {
i
}
i
as in Lemma 9. This is a permuta-
tion pairing, since each
i
is a permutation pairing on V
,i
(n) and since the {V
,i
(n)}
i
are disjoint.
Lemma 9 states that every source-destination pair in can communicate at a per-node rate of at least
n
o(1)
(4

n)
1min{3,}/2
. Note that, due to the locality of the trafc pattern, this can be considerably
better than the n
1min{3,}/2o(1)
per-node rate achieved by standard hierarchical relaying or multi-hop
communication.
Proof. We shall use either hierarchical relaying (for (2, 3]) or multi-hop (for > 3) to communicate
within each square V
,i
. We operate every fourth of the V
,i
simultaneously, and show that the added
interference due to this spatial re-use results only in a constant factor loss in rate.
Consider rst (2, 3] and fast fading. The squares A
,i
at level have an area of
n

n.
In order to be able to use hierarchical relaying within each of the {A
,i
}
i
, it is sufcient to show that we
can partition each A
,i
into
n
1
1+log
1/3
(n

25
squarelets, each of which contains a number of nodes proportional to the area (see Section V-C). In other
words, we partition A into squarelets of size
n
1
1
1+log
1/3
(n

n
log
1/3
(n)
1+log
1/3
(n)
L(n)
n
1
2
log
1/3
(n)
L(n)
= 2
1
2
log
1/6
(n)
n4

1
2
log(n)
_
1
1
2
log
5/6
(n)
_
,
where we have assumed, without loss of generality, that n 2. Since V V, all these squarelets contain
a number of nodes proportional to their area, and hence this shows that all
{A
i,
}
{0,...,L(n)},i{1,...,4

}
are simultaneously regular enough for hierarchical relaying to be successful under fast fading. This achieves
a per-node rate of

UC
u,w
n
o(1)
(4

n)
1/2
(20)
for any (u, w)
i
(see Section V-C, or [13, Theorem 1]).
We now show that (20) holds with high probability also under slow fading. For V V hierarchical
relaying is successful under slow fading for all permutation trafc on V with probability at least
1 exp
_
2
K log
2/3
(n)
_
for some constant K (see again Section V-C). Hence, hierarchical relaying is successful for all permutation
trafc on V
,i
with probability at least
1 exp
_
2
K log
2/3
(4

n)
_
1 exp
_
2
K log
2/3
(4
L(n)
n)
_
= 1 exp
_
2
K log
1/3
(n)
_
.
And hence hierarchical relaying is successful under slow fading for all {1, . . . , L(n)} and all
permutation trafc on every {V
,i
}
4

i=1
with probability at least
1 L(n)4
L(n)
exp
_
2
K log
1/3
(n)
_
1 n
2
exp
_
2
K log
1/3
(n)
_
1 o(1)
as n .
We now argue that the additional interference from spatial re-use results only in a constant loss in
rate. This follows from the same arguments as in the proof of [13, Theorem 1] (with the appropriate
modications for slow fading as described there). Intuitively, this is the case since the interference from a
square at distance r is attenuated by a factor r

, which, since > 2, is summable. Hence the combined

interference has power on the order of the receiver noise, resulting in only a constant factor loss in rate.
For > 3, the argument is similarinstead of hierarchical relaying we now use multi-hop communi-
cation. For V V and under either fast or slow fading, this achieves a per-node rate of

UC
u,w
n
o(1)
(4

n)
1/2
(21)
for any (u, w)
i
. Combining (20) and (21) yields the lemma.
26
VII. PROOF OF THEOREM 1
The proof of Theorem 1 relies on the construction of a capacitated (noiseless, wireline) graph G and
linking its performance under routing to the performance of the wireless network. This graph G = (V
G
, E
G
)
is constructed as follows. G is a full tree (i.e., all its leaf nodes are on the same level). G has n leaves,
each of them representing an element of V (n). To simplify notation, we assume that V (n) V
G
, so that
the leaves of G are exactly the elements of V (n) V
G
. Whenever the distinction is relevant, we use
u, v for nodes in V (n) V
G
and , for nodes in V
G
\ V (n) in the following. The internal nodes of G
correspond to V
,i
(n) for all {0, . . . , L(n)}, i {1, . . . , 4

}, with hierarchy induced by the one on

A(n). In particular, let and be internal nodes in V
G
and let V
,i
(n) and V
+1,j
(n) be the corresponding
subsets of V (n). Then is a child node of if V
+1,j
(n) V
,i
(n).
In the following, we will assume V V, which holds with probability 1o(1) as n by Lemma 5.
With this assumption, nodes in V
G
at level < L(n) have 4 children each, nodes in V
G
at level = L(n)
have between 4
L(n)1
n and 4
L(n)+1
n children, and nodes in V
G
at level = L(n) +1 are the leaves of
the tree (see Figure 7 below and Figure 3 in Section V-A).
= L(n) + 1
= L(n)
.
.
.
= 1
= 0
Fig. 7. Communication graph G constructed in the proof of Theorem 1. Nodes on levels {0, . . . , L(n) 1} have four children each,
nodes on level = L(n) have

n
log
1/2
(n)

children each. The total number of leaf nodes is n, one representing each node in the wireless
network V (n). An internal node in G at level {0, . . . , L(n)} represents the collection of nodes in V
,i
(n) for some i.
For V
G
, denote by L() the leaf nodes of the subtree of G rooted at . Note that, by construction
of the graph G, L() = V
,i
(n) for some and i. To understand the relation between V
G
and V (n), we
dene the representative R : V
G
2
V (n)
of as follows. For a leaf node u V (n) V
G
of G, let
R(u) {u}.
For V
G
at level L(n), choose R() L() V (n) such that
|R()| = 4
L(n)1
n.
This is possible since V (n) V(n) by assumption. Finally, for V
G
at level < L(n), and with
children {
i
}
4
j=1
, let
R()
4
_
j=1
R(
j
).
We now dene an edge capacity c
,
for each edge (, ) E
G
. If is a leaf of G and its parent,
set
c
,
= c
,
1. (22)
If is an internal node at level in G and its parent, then set
c
,
= c
,
(4

n)
2min{3,}/2
. (23)
Having chosen edge capacities on G, we can now dene the set
UC
G
(n) R
nn
+
of feasible unicast
trafc matrices between leaf nodes of G. In other words,
UC

UC
G
(n) if messages at the leaf nodes of
27
G can be routed to their destinations (which are also leaf nodes) over G at rates
UC
while respecting the
capacity constraints on the edges of G. Dene

BUC
G
(n)
UC
G
(n) B
UC
(n).
We rst prove the achievability part of Theorem 1. The next lemma shows that if trafc can be routed
over the tree G then approximately the same trafc can be transmitted reliably over the wireless network.
Lemma 10. Under fast fading, for any > 2, there exists b(n) n
o(1)
such that for any V (n) V(n),
b(n)
UC
G
(n)
UC
(n).
The same statement holds for slow fading with probability 1 o(1) as n .
Proof. Assume
UC

UC
G
, i.e., trafc can be routed between the leaf nodes of G at a rate
UC
, we need
to show that n
o(1)

UC

UC
(i.e., almost the same ow can be reliably transmitted over the wireless
network). We use the three-layer communication architecture introduced in Section V-A to establish this
result.
Recall the three layers of this architecture: the routing, cooperation, and physical layers. The layers of
this communication scheme operate as follows. In the routing layer, we treat the wireless network as the
graph G and route the messages between nodes over the edges of G. The cooperation layer provides this
tree abstraction to the routing layer by distributing and concentrating messages over subsets of the wireless
networks. The physical layer implements this distribution and concentration of messages by dealing with
interference and noise.
Consider rst the routing layer, and assume that the tree abstraction G can be implemented in the
wireless network with only a n
o(1)
factor loss. Since
UC

UC
G
by assumption, we then know that the
routing layer will be able to reliably transmit messages at rates n
o(1)

UC
over the wireless network. We
now show that the tree abstraction can indeed be implemented with a factor n
o(1)
loss in the wireless
network.
This tree abstraction is provided to the routing layer by the cooperation layer. We will show that the
operation of the cooperation layer satises the following invariance property: If a message is located at a
node G in the routing layer, then the same message is evenly distributed over all nodes in R() in
the wireless network. In other words, all nodes u R() V contain a distinct part of length 1/ |R()|
of the message.
Consider rst a leaf node u V V
G
in G, and assume the routing layer calls upon the cooperation
layer to send a message to its parent V
G
in G. Note rst that u is also an element of V , and it has
access to the entire message to be sent over G. Since for leaf nodes R(u) = {u}, this shows that the
invariance property is satised at u. The message is split at u into |R()| parts of equal length, and one
part is sent to each node in R() over the wireless network. In other words, we distribute the message
over the wireless network by a factor of |R()|. Hence the invariance property is also satised at .
Consider now an internal node V
G
, and assume the routing layer calls upon the cooperation layer to
send a message to its parent node V
G
. Note that since all trafc in G originates at the leaf nodes of G
(which are the actual nodes in the wireless network), a message at had to traverse all levels below in
the tree G. We assume that the invariance property holds up to level in the tree, and show that it is then
also satised at level . By the induction hypothesis, each node u R() has access to a distinct part of
length 1/ |R()|. Each such node u splits its message part into four distinct parts of equal length. Node
u keeps one part for itself, and sends the other three parts to nodes in R(). Since |R()| = 4 |R()|,
this can be performed such that each node in R() obtains exactly one message part. In other words, we
distribute the message by a factor four over the wireless network, and the invariance property is satised
at V
G
.
Operation along edges down the tree (i.e., towards the leaf nodes) is similar, but instead of distributing
messages, we now concentrate them over the wireless network. To route a message from a node V
G
with internal children {
j
}
4
j=1
to one of them (say
1
) in the routing layer, the cooperation layer sends
28
the message parts from each {R(
j
)}
4
j=2
to a corresponding node in R(
1
) and combines them there. In
other words, we concentrate the message by a factor four over the wireless network.
To route a message to a leaf node u V V
G
from its parent in G in the routing layer, the cooperation
layer sends the corresponding message parts at each node R() to u over the wireless network. Thus, again
we concentrate the message over the network, but this time by a factor of |R()|. Both these operations
along edges down the tree preserve the invariance property. This shows that the invariance property is
preserved by all operations induced by the routing layer in the cooperation layer.
Finally, to actually implement this distribution and concentration of messages, the cooperation layer
calls upon the physical layer. Note that at the routing layer, all edges of the tree can be routed over simul-
taneously. Therefore, the cooperation layer can potentially call the physical layer to perform distribution
and concentration of messages over all sets {R()}
V
G
simultaneously. The function of the physical
layer is to schedule all these operations and to deal with the resulting interference as well as with channel
noise.
This scheduling is done as follows. First, the physical layer time shares between communication up the
tree and communication down the tree (i.e., between distribution and concentration of messages). This
results in a loss of a factor 1/2 in rate. The physical layer further time shares between all the L(n) + 1
internal levels of the tree, resulting in a further
1
L(n)+1
factor loss in rate. Hence, the total rate loss by
this time sharing is
1
2(L(n) + 1)
. (24)
Consider now the operations within some level 1, . . . , L(n) in the tree (i.e., for edge (, ) on this
level, neither nor is a leaf node). We show that the rate at which the physical layer implements the
edge (, ) is equal to n
o(1)
c
,
, i.e., only a small factor less than the capacity of the edge (, ) in
the tree G. Note rst that the distribution or concentration of trafc induced by the cooperation layer to
implement one edge e at level (i.e., between node levels and 1) is restricted to V
1,i
for some
i = i(e). We can thus partition the edges at level into {E
j
G
}
4
j=1
such that the four sets
_
eE
j
G
V
1,i(e)
of nodes are disjoint. Time sharing between these four sets yields an additional loss of a factor 1/4 in
rate. Fix one such value of j, and consider the operations induced by the cooperation layer in the set
corresponding to j. We consider communication up the tree (i.e., distribution of messages), the analysis
for communication down the tree is similar. For a particular edge (, ) E
j
G
with the parent of ,
each node u R() has split its message part into four parts, three of which need to be sent to the nodes
in R() \ R(). Moreover, this assignment of destination nodes in R() \ R() to u is performed such
that no node in R() \ R() is destination more than once. In other word, each node in R() is source
exactly three times and each node in R() \R() is destination exactly once. This can be written as three
source-destination pairings {
k
i(,)
}
3
k=1
, on V
1,i(,)
. Moreover, each such
k
i(,)
can be understood as a
subset of a permutation source-destination pairing. We time share between the three values of k (yielding
an additional loss of a factor 1/3 in rate). Now, for each value of k, Lemma 9 shows that by using
either hierarchical relaying (for (2, 3]) or multi-hop communication for ( > 3), we can communicate
according to {
k
i(e)
}
eE
j
G
at a per-node rate of
n
o(1)
(4
1
n)
1min{3,}/2
under fast fading, and with probability
4
1 o(1) also under slow fading. Since R() contains 4
1
n
nodes, and accounting for the loss (24) for time sharing between the levels in G and the additional loss
4
Note that Lemma 9 actually shows that all permutation trafc for every value of can be transmitted with high probability under slow
fading. In other words, with high probability all levels of G can be implemented successfully under slow fading.
29
of factors 1/4 and 1/3 for time sharing between j and k, the physical layer implements an edge capacity
for e at level of
1
2(L(n) + 1)

1
4

1
3
4
1
n n
o(1)
(4
1
n)
1min{3,}/2
= n
o(1)
(4

n)
2min{3,}/2
= n
o(1)
c
e
.
Consider now the operations within level = L(n) + 1 in the tree (i.e., for edge (u, ) on this level,
u is a leaf node). We show that the rate at which the physical layer implements the edge (u, ) is equal
to n
o(1)
c
u,
. We again consider only communication up the tree (i.e., distribution of messages in the
cooperation layer), communication down the tree is performed in a similar manner. The trafc induced
by the cooperation layer at level L(n) + 1 is within the sets V
L(n),i
for i = {1, . . . , 4
L(n)
}. Consider now
communication within one V
L(n),i
, and assume without loss of generality that in the routing layer every
node u V
L(n),i
needs to send trafc along the edge (u, ). In the physical layer, we need to distribute a
1/ |R()| fraction of this trafc from each node u V
L(n),i
to every node in R() V
L(n),i
. This can be
expressed as

V
L(n),i

source-destination pairings, and we time share between them. Accounting for the
fact that only 1/ |R()| of trafc needs to be sent according to each pairing and since V V, this results
in a time sharing loss of at most a factor
|R()|

V
L(n),i

1
16
.
Now, using Lemma 9, all these source-destination pairings in all subsquares {V
L(n),i
} can be implemented
simultaneously at a per node rate of
n
o(1)
(4
L(n)
n)
1min{3,}/2
n
o(1)
(n
log
1/2
(n)
)
1/2
n
o(1)
.
Accounting for the loss (24) for time sharing between the levels in G, the additional factor 1/16 loss for
time sharing within each V
L(n),i
, the physical layer implements an edge capacity for e at level = L(n)+1
of
1
2(L(n) + 1)

1
16
n
o(1)
= n
o(1)
= n
o(1)
c
e
,
under either fast or slow fading.
Together, this shows that the physical and cooperation layers provide the tree abstraction G to the
routing layer with edge capacities of only a factor n
o(1)
loss. Hence, if messages can be routed at rates

UC
between the leaf nodes of G, then messages can be reliably transmitted over the wireless network at
rates n
o(1)

UC
. Hence

UC

UC
G
n
o(1)

UC

UC
.
Noting that the n
o(1)
factor is uniform in
UC
, this shows that
n
o(1)

UC
G

UC
.
We have seen that the unicast capacity region
UC
G
(n) of the graph G under routing is (appropriately
scaled) an inner bound to the unicast capacity region
UC
(n) of the wireless network. Taking the intersec-
tion with the set of balanced trafc matrices B
UC
(n) yields that the same holds for
BUC
G
(n) and
BUC
(n).
The next lemma shows that ((n) + 1)
BUC
G
(n) (with (n) = n
o(1)
as in the denition of B
UC
(n) in (1))
is an outer bound to the approximate unicast capacity region

BUC
(n) of the wireless network as dened
in (3) Combining Lemmas 5, 10, and 11 below, yields that with high probability
n
o(1)

BUC
(n) n
o(1)

BUC
G
(n)
BUC
(n),
proving the achievability part of Theorem 1.
30
Lemma 11. For any > 2 and any V (n) V(n),

BUC
(n) ((n) + 1)
BUC
G
(n),
where (n) = n
o(1)
is the factor in the denition of B
UC
(n) in (1).
Proof. We rst relate the total trafc across an edge e in the graph G to the total trafc across a cut V
,i
for some and i.
Consider an edge e = (, ) E
G
, and assume rst that e connects nodes at level and 1 in the
tree with L(n). We slight abuse of notation, set
c
e
c
,
.
Note rst that by (23) we have
c
e
= (4

n)
2min{3,}/2
. (25)
Moreover, since G is a tree, removing the edge e from E
G
separates the tree into two connected
components, say S
1
, S
2
V
G
. Consider now the leaf nodes in S
1
. By the construction of the tree structure
of G, these leaf nodes are either equal to V
,i
or V
c
,i
for some i {1, . . . , 4

}. Assume without loss of

generality that they are equal to V
,i
. Then V
c
,i
are the leaf nodes in S
2
. Now since trafc is only assumed
to be between leaf nodes of G, the total trafc demand between S
1
and S
2
is equal to

uV
,i

w/ V
c
,i
(
UC
u,w
+
UC
w,u
). (26)
By the tree structure of G, all this trafc has to be routed over edge e.
Consider now an edge e connecting a node at level L(n) +1 and L(n), i.e., a leaf node u to its parent
. Then, by (22),
c
e
= 1. (27)
The total trafc crossing the edge e is equal to

w=u
(
UC
u,w
+
UC
w,u
). (28)
We now show that

BUC
((n) + 1)
BUC
G
. (29)
Assume
UC

BUC
, then

uV
,i

w/ V
,i

UC
u,w
(4

n)
2min{3,}
for all {1, . . . , L(n)}, i {1, . . . , 4

}, and

w=u
(
UC
u,w
+
UC
w,u
) 1
for all u V . Since
UC
is balanced, this implies that
1
(n) + 1

uV
,i

w/ V
,i
(
UC
u,w
+
UC
w,u
) (4

n)
2min{3,}
for L(n). By (25), (26), (27), (28), we obtain that the trafc demand across each edge e of the graph
G is less than (n) + 1 times its capacity c
e
. Therefore, using that G is a tree,
1
(n)+1

UC
can be routed
over G, i.e.,
UC
((n) + 1)
BUC
G
. This proves (29).
31
We now turn to the converse part of Theorem 1. The next lemma shows that

UC
(n) (appropriately
scaled) is an outer bound to the unicast capacity region
UC
(n) of the wireless network. Taking the
intersection with the collection of balanced trafc matrices B
UC
(n) and combining with Lemma 5, this
shows that with high probability

BUC
(n) O(log
6
(n))

BUC
(n),
proving the converse part of Theorem 1.
Lemma 12. Under either fast or slow fading, for any > 2, there exists b(n) = O(log
6
(n)) such that
for any V (n) V(n),

UC
(n) b(n)

UC
(n).
Proof. Assume
UC

UC
. By Lemma 7, we have for any {1, . . . , L(n)} and i {1, . . . , 4

},

uV
,i

w/ V
,i

UC
u,w
K log
6
(n)(4

n)
2min{3,}/2
(30)
for some constant K not depending on
UC
.
Consider now u V . Lemma 6 shows that

w=u

UC
u,w


K log(n),

w=u

UC
w,u


K log(n),
with constant

K not depending on
UC
, and therefore,

w=u
(
UC
u,w
+
UC
w,u
) 2

K log(n). (31)
Combining (30) and (31) proves that there exists b(n) = O(log
6
(n)) such that
UC

UC
implies

UC
b(n)

UC
, proving the lemma.
VIII. PROOF OF THEOREM 2
Consider again the tree graph G = (V
G
, E
G
) with leaf nodes V (n) V
G
constructed in Section VII. As
before, we consider trafc between leaf nodes of G. In particular, any multicast trafc matrix
MC
R
n2
n
+
for the wireless network is also a multicast trafc matrix for the graph G. Denote by
MC
G
(n) R
n2
n
+
the set of feasible (under routing) multicast trafc matrices between leaf nodes of G, and set

BMC
G
(n)
MC
(n) B
MC
(n).
The next lemma shows that if multicast trafc can be routed over G then approximately the same
multicast trafc can be transmitted reliably over the wireless network. Taking the intersection with B
MC
(n)
implies that the same result holds also for balanced trafc.
Lemma 13. Under fast fading, for any > 2, there exists b(n) n
o(1)
such that for all V (n) V(n),
b(n)
MC
G
(n)
MC
(n).
The same statement holds under slow fading with probability 1 o(1) as n .
Proof. The proof follows using the same construction as in Lemma 10.
We now show that, since G is a tree graph,

BMC
(n) is an inner bound (up to a factor (n) +1) to the
the multicast capacity region
BMC
G
(n). The fact that G is a tree is critical for this result to hold.
32
Lemma 14. For any > 2,

BMC
(n) ((n) + 1)
BMC
G
(n).
where (n) = n
o(1)
is the factor in the denition of B
MC
(n) in (2).
Proof. Assume
MC
B
MC
\
BMC
G
. Since G is a tree, there is only one way to route multicast trafc
from u to W, namely along the subtree G({u} W) induced by {u} W (i.e., the smallest subtree of
G that covers {u} W). Hence for any edge e E
G
, the trafc d

MC(e) that needs to be routed over e

is equal to
d

MC(e) =

uV,WV :
eE
G({u}W)

MC
u,W
.
Now, since
MC
B
MC
\
BMC
G
, there exists e E
G
such that
d

MC(e) > c
e
. (32)
Let be the level of this edge e in G. We have
c
e
=
_
_
4

n
_
2min{3,}/2
if L(n),
1 else.
(33)
Assume rst that L(n) and let i be such that the removal of the edge e in G disconnects the leave
nodes in V
,i
from the ones in V
c
,i
. Then we have
d

MC(e) =

uV
,i

WV :
W\V
,i
=

MC
u,W
+

u/ V
,i
(n)

WV :
WV
,i
=

MC
u,W
. (34)
Assume then that = L(n) + 1, and assume e separates the leaf node u from {u}
c
in G. Then
d

MC(e) =

WV :
W\{u}=

MC
u,W
+

u=u

WV :
uW

MC
u,W
.
If = L(n) + 1, then (32), (33), and (34) imply that
MC
/

BMC
and therefore
MC
/
1
(n)+1

BMC
. If
L(n) then, since
MC
is (n)-balanced, we have

uV
,i

WV :
W\V
,i
=

MC
u,W
+

u/ V
,i
(n)

WV :
WV
,i
=

MC
u,W
((n) + 1)

uV
,i

WV :
W\V
,i
=

MC
u,W
. (35)
Combining (32), (33), (34), and (35) shows that
MC
/
1
(n)+1

BMC
for L(n) as well.
Hence, we have shown that
MC
B
MC
\
BMC
G
implies
MC
/
1
(n)+1

BMC
, proving the lemma.
Combining Lemmas 13, and 14, and 5 shows that, with probability 1 o(1) as n ,
n
o(1)

BMC
(n) n
o(1)

BMC
G
(n)
BMC
(n),
proving the inner bound in Theorem 2.
We now turn to the proof of the outer bound to
MC
(n). The next lemma combined with Lemma 5,
and taking the intersection with B
MC
(n), proves the outer bound in Theorem 2.
Lemma 15. Under fast fading, for any > 2, there exists b(n) = O(log
6
(n)) such that for all V (n)
V(n),

MC
(n) b(n)

MC
(n).
The same statement holds under slow fading with probability 1 o(1) as n .
33
Proof. We say that a unicast trafc matrix
UC
is compatible with a multicast trafc matrix
MC
if there
exists a mapping f : V (n) 2
V (n)
V (n) such that f(u, W) W {u}, for all (u, W), and

UC
u,w
=

WV (n):
f(u,W)=w

MC
u,W
for all (u, w). In words,
MC
is compatible with
UC
if we can create the unicast trafc matrix
UC
from

MC
by simply discarding the trafc for the pair (u, W) at all the nodes W \ {f(u, W)}.
Note that if
MC

BMC
and if
UC
is compatible with
MC
then
UC

UC
. Indeed, we can reliably
transmit at rate
UC
by using the communication scheme for
MC
and discarding all the unwanted messages
delivered by this scheme. Now consider a cut V
,i
with L(n) in the wireless network, and choose a
mapping f : V (n) 2
V (n)
V (n) such that

uV
,i

w/ V
,i

UC
u,w
=

uV
,i

WV :
W\V
,i
=

MC
u,W
.
Since
UC

UC
, we can apply Lemma 7 to obtain

uV
,i

WV :
W\V
,i
=

MC
u,W
=

uV
,i

w/ V
,i

UC
u,w
b(n)
_
4

n
_
2min{3,}/2
,
with b(n) = O(log
6
(n)). Repeating the same argument for cuts of the form {u} and {u}
c
and using
Lemma 6, shows that
MC
b(n)

MC
. Noting that the b(n) term is uniform in
MC
yields that

MC
b(n)

MC
,
concluding the proof of the lemma.
IX. DISCUSSION
We discuss several aspects and extensions of the three-layer architecture introduced in Section V-C and
used in the achievability parts of Theorems 1 and 2. In Section IX-A, we comment on the various tree
structures used in the three-layer architecture. In Section IX-B we show that for certain values of the
bounds in the theorems can be signicantly sharpened. In Section IX-C, we discuss bounds for trafc
that is not balanced. In Section IX-D, we show that for large values of path-loss exponent ( > 5) these
bounds are tight. Hence in the large path-loss regime the requirement of balanced trafc is not necessary,
and we obtain a scaling characterization of the entire unicast and multicast capacity regions. In Section
IX-E, we point out how the results discussed so far can be used to obtain the scaling of the unicast and
multicast capacity regions of dense networks (where n nodes are randomly placed on a square of unit
area).
A. Tree Structures
There are two distinct tree structures that are used in the construction of the three-layer communication
scheme proposed in this paperone explicit and one implicit. These two tree structures appear in different
layers of the communication scheme and serve different purposes.
The rst (explicit) tree structure is given by the tree G utilized in the routing layer and implemented
in the cooperation layer. The main purpose of this tree structure is to perform localized load balancing.
In fact, the distribution and concentration of trafc is used to avoid unnecessary bottlenecks. Note that
the tree G is used by the scheme for any value of .
The second (implicit) tree structure occurs in the physical layer. This tree structure appears only for
(2, 3]. In this regime, the physical layer uses the hierarchical relaying scheme. It is the hierarchical structure
of this scheme that can equivalently be understood as a tree. The purpose of this second tree structure is
to enable distributed multiple-antenna communication, i.e., to perform cooperative communication.
34
B. Second-Order Asymptotics
The scaling results in Theorems 1 and 2 are up to a factor n
o(1)
and hence preserve information at
scale n

for constant (see also the discussion in Section III-C). Here we examine in more detail the
behavior of this n
o(1)
factor and show that in certain situations it can be signicantly sharpened.
Note rst that the outer bound in Theorems 1 and 2 hold up to a factor O(log
6
(n)), i.e., poly-logarithmic
in n. However, the inner bound holds only up to the aforementioned n
o(1)
factor. A closer look at the
proofs of the two theorems reveals that the precise inner bound is of order

1
(n)n
O
_
log
1/3
(n)
_
,
where (n) is the factor in the denition of B
UC
(n) and B
MC
(n) (see (1) and (2)). With a more careful
analysis (see [13] for the details), this can be sharpened to essentially

1
(n)n
O
_
log
1/2
(n)
_
.
The exponent log
1/2
(n) in the inner bound has two causes. The rst is the use of hierarchical relaying
(for (2, 3]). The second is the operation of the physical layer at level L(n) + 1 of the tree (i.e., to
implement communication between the leaf nodes of G and their parents). Indeed at that level, we are
operating on a square of area
4
L(n)
n = n
log
1/2
(n)
,
and the loss is essentially inversely proportional to that area. Now, the reason why L(n) can not be
chosen to be larger (to make this loss smaller), is because hierarchical relaying requires a certain amount
of regularity in the node placement, which can only be guaranteed for large enough areas.
This suggests that for the > 3 regime, where multi-hop communication is used at the physical layer
instead of hierarchical relaying, we might be able to signicantly improve the inner bound. To this end, we
have to choose more levels in the tree G, such that at the last level before the tree nodes, we are operating
on a square that has an area of order log(n). Changing the three-layer architecture in this manner, and
choosing (n) appropriately, for > 3 the inner bound can be improved to (log
2
(n)) in n. Combined
with the poly-logarithmic outer bound, this yields a O(log
8
(n)) approximation of the balanced unicast
and multicast capacity regions for > 3.
C. Non-Balanced Trafc
Theorems 1 and 2 describe the scaling of the balanced unicast and multicast capacity regions
BUC
(n)
and
BMC
(n), respectively. As we have argued, the balanced unicast region
BUC
(n) coincides with the
unicast capacity region
UC
(n) along at least n
2
n out of n
2
total dimensions, and the balanced multicast
region
BMC
(n) coincides with the multicast capacity region
UC
(n) along at least n2
n
n out of n2
n
total dimensions. However, the proofs of these results provide also bounds for trafc that is not balanced,
i.e., for the remaining 2n dimensions.
Dene the following two regions:

UC
1
(n)
_

UC
R
nn
+
:

uV
,i
(n)

w/ V
,i
(n)
(
UC
u,w
+
UC
w,u
) (4

n)
2min{3,}/2
{1, . . . , L(n)}, i {1, . . . , 4

},

w=u
(
UC
u,w
+
UC
w,u
) 1 u V (n)
_
,
35
and

MC
1
(n)
_

MC
R
n2
n
+
:

uV
,i
(n)

WV (n):
W\V
,i
(n)=

MC
u,W
+

u/ V
,i
(n)

WV (n):
WV
,i
(n)=

MC
u,W
(4

n)
2min{3,}/2
{1, . . . , L(n)}, i {1, . . . , 4

},

WV (n):
W\{u}=

MC
u,W
+

u=u

WV (n):
uW

MC
u,W
1 u V (n)
_
.

UC
(n) and

UC
1
(n) differ in that for {1, . . . , L(n)},

UC
(n) only bounds trafc ow out of V
,i
(n),
whereas

UC
1
(n) bounds trafc in both directions across V
,i
(n) (and similar for

MC
(n) and

MC
1
(n)).
The analysis in Sections VII and VIII shows that
n
o(1)

UC
1
(n)
UC
(n) O(log
6
(n))

UC
(n),
n
o(1)

MC
1
(n)
MC
(n) O(log
6
(n))

MC
(n),
with probability 1 o(1) as n . In other words, we obtain an inner and an outer bound on the
capacity regions
UC
(n) and
MC
(n). These bounds coincide in the scaling sense for balanced trafc, for
which we recover Theorems 1 and 2.
D. Large Path-Loss Exponent Regime
The discussion in Section IX-C reveals that in order to obtain scaling information for trafc that is
not balanced, a stronger version of the converse results in Lemma 7 is needed. In particular, Lemma 7
bounds the sum-rate

uV
,i
(n)

w/ V
,i
(n)

UC
u,w
for
UC

UC
(n). The required stronger version of the lemma would also need to bound sum rates in
the other direction, i.e.,

u/ V
,i
(n)

wV
,i
(n)

UC
u,w
.
For large path-loss exponents > 5, such a stronger version of Lemma 7 holds (see Lemma 8). With
this, we obtain that for > 5,
n
o(1)

UC
1
(n)
UC
(n) O(log
6
(n))

UC
1
(n),
n
o(1)

MC
1
(n)
MC
(n) O(log
6
(n))

MC
1
(n),
with probability 1o(1) as n . In other words, in the high path-loss exponent regime > 5,

UC
1
(n)
and

MC
1
(n) characterize the scaling of the entire unicast and multicast capacity regions, respectively.
E. Dense Networks
So far, we have only discussed extended networks, i.e., n nodes are located on a square of area n. We
now briey sketch how these results can be recast for dense networks, in which n nodes are located on
a square of unit area.
Note rst that by rescaling power by a factor n
/2
, a dense network with any path-loss exponent can
essentially be transformed into an equivalent extended network with path-loss exponent . In particular,
any scheme for extended networks with path-loss exponent yields a scheme with same performance
for dense networks with any path-loss exponent (see also [12, Section V.A]). To optimize the resulting
scheme for the dense network, we start with the scheme for extended networks corresponding to close to
36
2. Hence an inner bound for the unicast and multicast capacity regions for dense networks with path-loss
exponent can be obtained from the ones for extended networks by taking a limit as 2. Moreover,
an application of Lemma 6 yields a matching (in the scaling sense) outer bound.
The resulting approximate balanced capacity regions

BUC
(n) and

BMC
(n) have particularly simple
shapes in this limit. In fact, the only constraints in (3) and (4) that can be tight are at level = log(n).
Moreover, as in Section IX-D, it can be shown that the restriction of balanced trafc is not necessary for
dense networks. This results in the following approximate capacity regions for dense networks:

UC
(n)
_

UC
R
nn
+
:

w=u
(
UC
u,w
+
UC
w,u
) 1, u V (n)
_
for unicast, and

MC
(n)
_

MC
R
n2
n
+
:

WV (n):
W\{u}=

MC
u,W
+

u=u

WV (n):
uW

MC
u,W
1, u V (n)
_
for multicast. We obtain that for dense networks, for any > 2,
n
o(1)

UC
(n)
UC
(n) O(log
6
(n))

UC
(n),
n
o(1)

MC
(n)
MC
(n) O(log
6
(n))

MC
(n),
with probability 1 o(1) as n .
X. CONCLUSIONS
In this paper, we have obtained an explicit information-theoretic characterization of the scaling of the
n
2
-dimensional balanced unicast and n2
n
-dimensional balanced multicast capacity regions of a wireless
network with n randomly placed nodes and assuming a Gaussian fading channel model. These regions
span at least n
2
n and n2
n
n dimensions of R
nn
+
and R
n2
n
+
, respectively, and hence determine the
scaling of the unicast capacity region along at least n
2
n out of n
2
dimensions and the scaling of the
multicast capacity region along at least n2
n
n out of n2
n
dimensions. The characterization is in terms
of 2n weighted cuts, which are based on the geometry of the locations of the source nodes and their
destination nodes and on the trafc demands between them, and thus can be readily evaluated.
This characterization is obtained by establishing that the unicast and multicast capacity regions of a
capacitated (wireline, noiseless) tree graph under routing have essentially the same scaling as that of the
original network. The leaf nodes of this tree graph correspond to the nodes in the wireless network, and
internal nodes of the tree graph correspond to hierarchically growing sets of nodes.
This equivalence suggests a three-layer communication architecture for achieving the entire balanced
unicast and multicast capacity regions (in the scaling sense). The top or routing layer establishes paths
from each of the source nodes to its destination (for unicast) or set of destinations (for multicast) over
the tree graph. The middle or cooperation layer provides this tree abstraction to the routing layer by
distributing the trafc among the corresponding set of nodes as a message travels up the tree graph, and
by concentrating the trafc on to the corresponding set of nodes as a message travels down the tree. The
bottom or physical layer implements this distribution and concentration of trafc over the wireless network.
This implementation depends on the path-loss exponent: For low path loss, (2, 3], hierarchical relaying
is used, while for high path loss ( > 3), multi-hop communication is used.
This scheme also establishes that a separation based approach, in which the routing layer works
essentially independently of the physical layer, can achieve nearly the entire balanced unicast and multicast
capacity regions in the scaling sense. Thus, for balanced trafc, such techniques as network coding can
provide at most a small increase in the scaling.
37
XI. ACKNOWLEDGMENTS
The authors would like to thank David Tse, Greg Wornell, and Lizhong Zheng for helpful discussions,
and the anonymous reviewers for their help in improving the presentation of this paper.
REFERENCES
[1] P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE Transactions on Information Theory, 46(2):388404, March 2000.
[2] L.-L. Xie and P. R. Kumar. A network information theory for wireless communication: Scaling laws and optimal operation. IEEE
Transactions on Information Theory, 50(5):748767, May 2004.
[3] A. Jovi ci c, P. Viswanath, and S. R. Kulkarni. Upper bounds to transport capacity of wireless networks. IEEE Transactions on
Information Theory, 50(11):2555 2565, November 2004.
[4] F. Xue, L.-L. Xie, and P. R. Kumar. The transport capacity of wireless networks over fading channels. IEEE Transactions on Information
Theory, 51(3):834847, March 2005.
[5] O. L ev eque and

I. E. Telatar. Information-theoretic upper bounds on the capacity of large extended ad hoc wireless networks. IEEE
Transactions on Information Theory, 51(3):858865, March 2005.
[6] L.-L. Xie and P. R. Kumar. On the path-loss attenuation regime for positive cost and linear scaling of transport capacity in wireless
networks. IEEE Transactions on Information Theory, 52(6):23132328, June 2006.
[7] M. Franceschetti, O. Dousse, D. N. C. Tse, and P. Thiran. Closing the gap in the capacity of wireless networks via percolation theory.
IEEE Transactions on Information Theory, 53(3):10091018, March 2007.
[8] P. Gupta and P. R. Kumar. Towards an information theory of large networks: An achievable rate region. IEEE Transactions on
Information Theory, 49(8):18771894, August 2003.
[9] L.-L Xie and P. R. Kumar. An achievable rate for the multiple-level relay channel. IEEE Transactions on Information Theory,
51(4):13481358, April 2005.
[10] G. Kramer, M. Gastpar, and P. Gupta. Cooperative strategies and capacity theorems for relay networks. IEEE Transactions on
Information Theory, 51(9):30373063, September 2005.
[11] S. Aeron and V. Saligrama. Wireless ad hoc networks: Strategies and scaling laws for the xed SNR regime. IEEE Transactions on
Information Theory, 53(6):20442059, June 2007.
[12] A.

Ozg ur, O. L ev eque, and D. N. C. Tse. Hierarchical cooperation achieves optimal capacity scaling in ad hoc networks. IEEE
Transactions on Information Theory, 53(10):35493572, October 2007.
[13] U. Niesen, P. Gupta, and D. Shah. On capacity scaling in arbitrary wireless networks. IEEE Transactions on Information Theory,
55(9):39593982, September 2009.
[14] R. Madan, D. Shah, and O. L ev eque. Product multicommodity ow in wireless networks. IEEE Transactions on Information Theory,
54(4):14601476, April 2008.
[15] F. T. Leighton and S. Rao. Multicommodity max-ow min-cut theorems and their use in designing approximation algorithms. Journal
of the ACM, 46(6):787832, November 1999.
[16] N. Linial, E. London, and Y. Rabinovich. The geometry of graphs and some of its algorithmic applications. Combinatorica, 15(2):215
245, June 1995.
[17] S. Subramanian, S. Shakkottai, and P. Gupta. On optimal geographic routing in wireless networks with holes and non-uniform trafc.
In Proc. IEEE Infocom, pages 13281336, May 2007.
[18] S. Subramanian, S. Shakkottai, and P. Gupta. Optimal geographic routing for wireless networks with near-arbitrary holes and trafc.
In Proc. IEEE Infocom, pages 10191027, April 2008.
[19] S. H. A. Ahmad, A. Jovi ci c, and P. Viswanath. On outer bounds to the capacity region of wireless networks. IEEE Transactions on
Information Theory, 52(6):27702776, June 2006.
[20] X-Y. Li, S-J. Tang, and O. Frieder. Multicast capacity for large scale wireless ad hoc networks. In ACM MobiCom, pages 263277,
September 2007.
[21] S. Shakkottai, X. Liu, and R. Srikant. The multicast capacity of large multihop wireless networks. In ACM MobiHoc, pages 247255,
September 2007.
[22] A. Keshavarz-Haddad, V. Ribeiro, and R. Riedi. Broadcast capacity in multihop wireless networks. In Proc. ACM MobiCom, pages
239250, September 2006.
[23] P. Chaporkar and S. Sarkar. Wireless multicast: Theory and approaches. IEEE Transactions on Information Theory, 51(6):19541972,
June 2005.
[24] P. Jacquet and G. Rodolakis. Multicast scaling properties in massively sense ad hoc networks. In Proc. International Conference on
Parallel and Distributed Systems, volume 2, pages 9399, July 2005.
[25] M. Franceschetti, M. D. Migliore, and P. Minero. The capacity of wireless networks: Information-theoretic and physical limits. IEEE
Transactions on Information Theory, 55(8):34133424, August 2009.
[26] M. Franceschetti, M. D. Migliore, and P. Minero. The degrees of freedom of wireless networks: Information theoretic and physical
limits. In Allerton Conference on Communication, Control, and Computing, pages 15781581, September 2008.
[27] T. M. Cover and J. A. Thomas. Elements of Information Theory. Wiley, 1991.