388

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

The Capacity of Wireless Networks

Piyush Gupta, Student Member, IEEE, and P. R. Kumar, Fellow, IEEE
AbstractWhen identical randomly located nodes, each cabits per second and using a fixed range, pable of transmitting at obtainable by each form a wireless network, the throughput

() node for a randomly chosen destination is 2 2( )

log

bits per

second under a noninterference protocol. If the nodes are optimally placed in a disk of unit area, traffic patterns are optimally assigned, and each transmissions range is optimally chosen, the bitdistance product that can be transported by the network per second is bit-meters per second. Thus even under optimal circumstances, the throughput is only bits per second for each node for a destination nonvanishingly far away. Similar results also hold under an alternate physical model where a required signal-to-interference ratio is specified for successful receptions. Fundamentally, it is the need for every node all over the domain to share whatever portion of the channel it is utilizing with nodes in its local neighborhood that is the reason for the constriction in capacity. Splitting the channel into several subchannels does not change any of the results. Some implications may be worth considering by designers. Since the throughput furnished to each user diminishes to zero as the number of users is increased, perhaps networks connecting smaller numbers of users, or featuring connections mostly with nearby neighbors, may be more likely to be find acceptance.

Index TermsAd hoc networks, capacity, multihop radio networks, throughput, wireless networks.

I. INTRODUCTION

IRELESS networks consist of a number of nodes which communicate with each other over a wireless channel. Some wireless networks have a wired backbone with only the last hop being wireless. Examples are cellular voice and data networks and mobile IP. In others, all links are wireless. One

Manuscript received December 3, 1998; revised July 1, 1999. This material is based on work supported in part by the Air Force Office of Scientific Research under Contract AF-DC-5-36128, by EPRI and the U.S. Army Research Office under Subcontract to Cornell University Contracts WO8333-04 and 35352-6086, by the U.S. Army Research Office under Contract DAAH 04-95-1-0090, the Office of Naval Research under Contract N00014-99-1-0696, and the Joint Services Electronics Program under Contract N00014-96-1-0129. Any opinions, findings, and conclusions are those of the authors and do not necessarily reflect the views of the above agencies. P. Gupta is with the Department of Electrical and Computer Engineering, and the Coordinated Science Laboratory, University of Illinois, Urbana, IL 61801 USA. P. R. Kumar is with the University of Illinois at Urbana-Champaign, Coordinated Science Laboratory, Urbana, IL 61801 USA. Communicated by V. Anantharam, Associate Editor for Communication Networks. Publisher Item Identifier S 0018-9448(00)01358-4.

example of such networks is multihop radio networks or ad hoc networks. Another possibly futuristic example, see [1], may be collections of smart homes where computers, microwave ovens, door locks, water sprinklers, and other information appliances are interconnected by a wireless network. It is to these types of all wireless networks that this paper is addressed. Such networks consist of a group of nodes which communicate with each other over a wireless channel without any centralized control; see Fig. 1. Nodes may cooperate in routing each others data packets. Lack of any centralized control and possible node mobility give rise to many issues at the network, medium access, and physical layers, which have no counterparts in the wired networks like Internet, or in cellular networks. At the network layer, the main problem is that of routing, which is exacerbated by the time-varying network topology, power constraints, and the characteristics of the wireless channel; see Ramanathan and Steenstrup [2] for an overview. The choice of medium access scheme is also difficult in ad hoc networks due to the time-varying network topology and the lack of centralized control. Use of TDMA or dynamic assignment of frequency bands is complex since there is no centralized control as in cellular networks, FDMA is inefficient in dense networks, CDMA is difficult to implement due to node mobility and the consequent need to keep track of the frequency-hopping patterns and/or spreading codes for nodes in the time-varying neighborhood, and random access appears to be the current favorite. The access problem when many nodes transmit to the same receiver has been much studied in the literature ever since the genesis of the ALOHA network, and bounds on the throughput of successful collision-free transmissions as well as transmission protocols have been devised; see Gallager [3]. Sharing channels in networks does lead to some new problems associated with hidden terminals and exposed terminals. The protocols MACA and its extension MACAW, see Karn [4] and Bhargavan et al. [5] respectively, use a series of handshake signals to resolve these problems to a certain extent. This has been standardized in the IEEE 802.11 protocol, see [6]. At the physical layer, an important issue is that of power control. The transmission power of nodes needs to be regulated so that it is high enough to reach the intended receiver while causing minimal interference at other nodes. Iterative power control algorithms have been devised, see Bambos, Chen, and Pottie [7] and Ulukus and Yates [8]. In this paper we analyze the capacity of wireless networks. We scale space and suppose that nodes are located in a region bits per second over of area 1 m2. Each node can transmit at a common wireless channel. We shall see that it is immaterial

00189448/00$10.00 2000 IEEE

GUPTA AND KUMAR: THE CAPACITY OF WIRELESS NETWORKS

389

Fig. 1.

An ad hoc wireless network.

to our results1 if the channel is broken up into several subchanbits per second, as long as nels of capacity Packets are sent from node to node in a multihop fashion until they reach their final destination. They can be buffered at intermediate nodes while awaiting transmission. Due to spatial separation, several nodes can make wireless transmissions simultaneously, provided there is no destructive interference of a transmission by others. We will describe in the sequel under what conditions a wireless transmission over a subchannel is received successfully by its intended recipient. We will consider two types of networks, Arbitrary Networks, where the node locations, destinations of sources, and traffic demands, are all arbitrary, and Random Networks, where the nodes and their destinations are randomly chosen. A. Arbitrary Networks: Arbitrarily Located Nodes and Traffic Patterns In the arbitrary setting we suppose that nodes are arbitrarily located in a disk of unit area in the plane. Each node has an arbitrarily chosen destination to which it wishes to send traffic at an arbitrary rate; thus the traffic pattern is arbitrary. Each node can choose an arbitrary range or power level for each transmission. We need to describe when a transmission is received successfully by its intended recipient. We will allow for two possible models for successful reception of a transmission over one hop, called the Protocol Model and the Physical Model, described denote the location of a node; we will also use below. Let to refer to the node itself. transmits over the 1) The Protocol Model: Suppose node th subchannel to a node Then this transmission is successif fully received by node (1) simultaneously transmitting over the for every other node same subchannel. models situations where a guard zone The quantity is specified by the protocol to prevent a neighboring node from
1We are grateful to Kimberly King for asking us to be more explicit about the prospects for routing through multiple technologies.

transmitting on the same subchannel at the same time. It also allows for imprecision in the achieved range of transmissions. Another model which is more related to physical layer considerations is be the subset of 2) The Physical Model: Let nodes simultaneously transmitting at some time instant over a be the power level chosen by node certain subchannel. Let for Then the transmission from a node , , if is successfully received by a node (2)

This models a situation where a minimum signal-to-interference ratio (SIR) of is necessary for successful receptions, the ambient noise power level is , and signal power decays with disWe will suppose that , which is the usual tance as model outside a small neighborhood of the transmitter. 3) The Transport Capacity of Arbitrary Networks: Given any set of successful transmissions taking place over time and space, let us say that the network transports one bit-meter when one bit has been transported a distance of one meter toward its destination. (We do not give multiple credit for the same bit carried from one source to several different destinations as in the multicast or broadcast cases). This sum of products of bits and the distances over which they are carried is a valuable indicator of a networks transport capacity. (It should be noted square meters rather that when the area of the domain is than the normalized 1 m2, then all the transport capacity results Our main results presented below should be scaled by are the following. Recall Knuths notation: denotes that as well as . Main Result 1.: The transport capacity of an Arbitrary Netbit-meters per work under the Protocol Model is second if the nodes are optimally placed, the traffic pattern is optimally chosen, and if the range of each transmission is chosen optimally. bit-meters per Specifically, an upper bound is second for every Arbitrary Network for all spatial and temporal

390

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

scheduling strategies, while bit-meters per second (for a multiple of four) can be achieved when the nodes and traffic patterns are appropriately chosen, and the ranges and schedules of transmissions are appropriately chosen. If this transport capacity were to be equitably divided between bit-meall the nodes, then each node would obtain ters per second. If, further, each source has its destination about the same distance of 1 m away, then each node would obtain a bits per second. throughput capacity of The upper bound on transport capacity does not depend on the transmissions being omnidirectional, as implied by (1), but only on there being some dispersion in the neighborhood of the receiver; see Assumption (A.vi) in Section II. bit-meters Main Result 2: For the Physical Model, bit-meters per seconds per second is feasible, while Specifically, is not, for appropriate

1) The Protocol Model: All nodes employ a common range for all their transmissions. When node transmits to a node over the th subchannel, this transmission is successfully if received by and is no more than , i.e., i) The distance between (3) ii) For every other node the same subchannel simultaneously transmitting over

(4) 2) The Physical Model: All nodes choose a common power for all their transmissions. Let be the level subset of nodes simultaneously transmitting at some time instant over a certain subchannel. A transmission from a node , , is successfully received by a node if (5)

bit-meters per second (for a multiple of ) is feasible when the network is appropriately designed, while an upper bound is bit-meters per second. bit-meters We suspect that an upper bound of order per second may actually hold. In the special case where the ratio between the maximum and minimum powers that transmitters can employ is bounded above by , then an upper bound is in fact

3) The Throughput Capacity of Random Networks: The notion of throughput is defined in the usual manner as the time average of the number of bits per second that can be transmitted by every node to its destination. bits Definition: Feasible Throughput: A throughput of per second for each node is feasible if there is a spatial and temporal scheme for scheduling transmissions, such that by operating the network in a multihop fashion and buffering at intermediate nodes when awaiting transmission, every node can bits per second on average to its chosen destination send such that in every time innode. That is, there is a every node can send bits to its corterval responding destination node. Whether a particular throughput level is feasible may depend on the locations of the nodes. These locations are random. So is the destination for the traffic entering each node. As in PAC Learning Theory (see Valiant [9]), given the randomness involved in the problem statement, we allow for vanishingly small probabilities when defining the throughput capacity. Definition: The Throughput Capacity of Random Wireless Networks: We say that the throughput capacity of the class of bits per second if there Random Networks is of order and such that are deterministic constants is feasible is feasible Our main results are the following. Main Result 3.: In the case of both the surface of the sphere and a planar disk, the order of the throughput capacity is

bit-meters per second. It is worth noting that both bounds suggest that transport capacity improves when is larger, i.e., when the signal power decays more rapidly with distance. B. Random Networks: Randomly Located Nodes and Traffic Patterns In a random scenario, nodes are randomly located, i.e., independently and uniformly distributed, either on the surface of a three-dimensional sphere of area 1 m2, or in a disk is to of area 1 m2 in the plane. Our purpose in studying separate edge effects from other phenomena. Each node has a randomly chosen destination to which it wishes to send bits per second. The destination for each node is independently chosen as the node nearest to a randomly located point, i.e., uniformly and independently distributed. (Thus destinations are on the order of 1 m away on average.) In this random setting, we will assume that the nodes are homogeneous, i.e., all transmissions employ the same nominal range or power. As for Arbitrary Networks, we will allow for both a Protocol Model as well as a Physical Model for interference.

GUPTA AND KUMAR: THE CAPACITY OF WIRELESS NETWORKS

391

bits per second for the Protocol Model. For the upper bound we actually prove the sharp cutoff phenomenon that for some is feasible Specifically, there are deterministic constants or such that depending on and not

bits per second is feasible, and

is large. Perhaps designers should target their efforts at networks for smaller numbers of users, rather than try to develop large wireless networks. A feasible scenario is where nodes need to communicate only with nearby nodes. Then the scaled distance between sources and destinations is only meters. Thus all nodes can transmit data to nearby neighbors at a bit rate that does not decrease with Such a scenario can arise, for example, in collections of smart homes, each home having sensors and actuators communicating by wireless means. Another implication concerns the power consumption by each node for transmission. Consider Random Networks. The fraction of time that a modem is busy, whether relaying traffic or sending packets originating at the node, is only Not only that, the scaled range of each transmission is about The bounds for the Physical Model suggest that a faster rate of decay of signal power with distance, i.e., a larger , allows greater transport and throughput capacity. One more implication follows from the constructive proof of capacity. It shows that one can group the nodes into small clusters or cells, where in each cell one can designate one specific node to carry all the burden of relaying multihop packets, if so desired. Thus a division of labor is possible, were this to be found profitable. Moreover, it would further reduce the transmission power consumed by the vast majority of other nodes. This may offer some suggestive guidelines for designers of routing protocols. It should be noted that dividing the channel into subchannels does not change any of the results. Yet another issue concerns the use of relay nodes.2 Consider a Random Network with source nodes. Then the throughput that can be furnished to each of them is only under the Protocol Model. Suppose additional homogeneous nodes are deployed as pure relays in random positions, with no independent traffic needs of their own, i.e., they are not sources. Then the throughput that can be furnished to each of the sources is There is, however, a severe cost of providing this increase in throughput. The number of additional relay nodes that need to be deployed to gain an appreciable increase in capacity for the source nodes may be very large. When active nodes, to make there are equal to five times its value at , will have to be equal . The addition of nodes to serve as pure relays to at least -fold increase in this term. provides a less than One way to overcome the barrier of wireless networks is to do what is done in cellular telephonyconnect the base stations by a wired network. If, however, nondirected wireless links are used for connecting the base stations, then the capacity limitation of wireless networks remains with us, though in less obvious ways. For example, suppose a high-power base station is chosen in each cell, which communicates with other distant base stations by a wireless channel. Then the set of base stations inherits the same capacity limitation. A set of wire2We

bits per second is infeasible, both with probability approaching Since routing hot spots may form at the center one as in the case of a disk on the plane, and yet the order of throughput capacity is the same as on the surface of the sphere, it shows that the cause of the throughput constriction is not the formation of hot spots, but is the pervasive need for all nodes to share the channel locally with other nodes. Main Result 4: For the Physical Model a throughput of bits per second is feasible, while bits per second is not, for appropriate both with probSpecifically, there are ability approaching one as not depending on deterministic constants and or such that

bits per second is feasible with probability approaching one as If is the mean distance between two points independently and uniformly distributed in the domain (either surface of sphere or planar disk of unit area), then there is a deterministic not depending on or such that sequence

bit-meters per second is infeasible with probability approaching one as C. Some Possible Implications The results in this paper allow for a perfect scheduling algorithm which knows the locations of all nodes and all traffic demands, and which coordinates wireless transmissions temporally and spatially to avoid collisions which would otherwise result in lost packets. Also, the nodes are not mobile. If such perfect node location information is not available, or if nodes move, or traffic demands are not known, then the capacity can only be even smaller. There are some implications of these results which designers may want to consider. The decrease in throughput with may be regarded as unacceptable by users when the number of nodes

are grateful to Chip Elliott for raising this issue.

392

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

lessly connected base stations can provide a throughput of only for each base station.

bits per second for Random Networks on In Section V we show that bits per second and

D. A Discussion of the Tradeoffs Involved Why does the throughput capacity diminish as the number of nodes increases? For an insight into some of the tradeoffs involved, consider Random Networks. Let the mean distance to the common be traversed by a packet be , and denote by range of all transmissions. Then the mean number of hops taken Thus each node generates at least by packets is no less than bits per second of traffic for other nodes. Since the total bits number of nodes is , the total traffic is no less than per second. This has to be served by nodes each capable of bits per second. Thus one needs An upper bound Since the term on on the throughput is therefore , it might appear that to inthe right side grows linearly in crease the throughput by reducing the number of hops traversed by each packet, and thus the burden on other nodes serving as of each node. Howrelays, one should increase the range ever, the expression above is not an achievable upper bound as The reason is that we have neglected the rea function of duction in capacity due to spatial concurrency constraints, since nodes close to a receiver are required to be idle to avoid collisions which cause the loss of packets. In fact, the loss from is quadratic due to the area of the conflict inincreasing volved. Therefore, the desire to reduce the multihop burden and the desire to increase spatial concurrency and frequency reuse are in conflict. It turns out that when we consider both issues together, we find that one really needs to reduce the value of to as small a value as possible. However, there is a limit to how When the range of transmissions small one can make is too small, the wireless network loses connectivity. In a precursor result, see [10], the critical range for connectivity of networks formed by randomly located nodes on a disk in the plane has been determined. Consider the graph with random vertices uniformly and independently distributed in a disk of unit area. Join two vertices by an edge whenever they are within a disfrom each other. The critical radius for connectivity tance in the sense that the graph with is is if and connected with probability approaching one as only if For Arbitrary Networks under the Protocol model, just three constraintsthe length of routes, the consumption of valuable two-dimensional area by transmissions, and the total number of nodesare enough to force the transport capacity to be no more bit-meters per second. than The rest of this paper is organized as follows. In Section II we exhibit upper bounds on the transport capacity of the bit-meters per second and bit-meters form per second, under the Protocol and Physical Models, respectively, for Arbitrary Networks. In Section III we show that a bit-meters per second is also transport capacity of feasible for Arbitrary Networks. In Section IV we construct a scheduling and routing scheme which achieves a throughput of

bits per second are upper bounds on the throughput under the Protocol and Physical for Random Networks on Models, respectively. In Section VI we show that the above results for Random Networks also hold for a disk in the plane. II. ARBITRARY NETWORKS: AN UPPER BOUND ON TRANSPORT CAPACITY We consider the setting on a planar disk of unit area. Consider the following (nearly) minimal set of assumptions: (A.i) There are nodes arbitrarily located in a disk of unit area on the plane. (The results carry over to any domain which is the closure of its interior.) of unit area in bits over seconds. (A.ii) The network transports (A.iii) The average distance between the source and destination of a bit is Note that, together with (A.ii), this implies bit-meters per second is that a transport capacity of achieved. subchan(A.iv) Each node can transmit over any subset of bits per second, , nels with capacities . where (A.v) Transmissions are slotted into synchronized slots of length seconds. (This assumption can be eliminated, but makes the exposition easier.) (A.vi) While retaining the restriction (2) for the case of the Physical Model, we can either retain (1) in the Protocol Model or consider an alternate restriction as follows: If transmits to another node located at a disa node tance of units on a certain subchannel in a certain slot, then there can be no other receiver within a radius of around on the same subchannel in the same slot. This alternate restriction addresses situations where the transmissions are not omnidirectional, but nevertheless there is some dispersion in the neighborhood of the receiver. Theorem 2.1: i) In the Protocol Model, the transport capacity bounded as follows: bit-meters per second ii) In the Physical Model bit-meters per second iii) If the ratio between the maximum and minimum powers that transmitters can employ is strictly bounded above by , then bit-meters per second is

GUPTA AND KUMAR: THE CAPACITY OF WIRELESS NETWORKS

393

iv) When the domain is of square meters rather than 1 m2, then all the upper bounds above are scaled by Proof: Consider bit , where Let us suppose that it moves from its origin to its destination in a sequence hops, where the th hop traverses a distance of Then of from (A.iii) (6)

Summing over the subchannels and the slots gives

This can be rewritten as (9) Note now that the quadratic function is convex. Hence

nodes can transmit. Note now that in any slot at most Hence for any subchannel and any slot (10) (The th hop of bit is over subchannel in slot ) Now substituting (6) in (11) gives (12) (7) is reConsider now the Protocol Model. Suppose that over the th subchannel at the ceiving a transmission from is receiving a transmission from over the same time that same subchannel. Then from the triangle inequality and (1) Substituting (7) in (12) yields the result. Now turn to the Physical Model. The difference stems from the need to replace (8) by a different expression. Suppose is transmitting to over the th subchannel at power level at some time, and let denote the set of all simultaneous transmitters over the th subchannel at that time. Including the also in the denominator, the signal-to-intersignal power of can be written as ference requirement (2) for Combining (9) and (10) yields (11)

Summing over the subchannels and the slots, and noting that there can be no more than slots in seconds, yields

Similarly, Hence Adding the two inequalities, we obtain

Hence disks of radius times the lengths of hops centered at the receivers over the same subchannel in the same slot are essentially disjoint. (Note that this conclusion directly follows when (1) is replaced by the alternate restriction of Assumption (A.vi)). Allowing for edge effects where a node is near the periphery of the domain, and noting that a range greater than the diameter of the domain is unnecessary, we see that at least a quarter of such a disk is within the domain. Since at most bits can be carried in slot from a receiver to a transmitter over the th subchannel, we have (The th hop of bit is over subchannel in slot ) (8)

since Summing over all transmitter-receiver pairs

Summing over all slots and subchannels gives

394

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

The rest of the proof proceeds along lines similar to the Protocol instead of Model, invoking the convexity of For the consideration of the special case where , is transmitting to we start with (2). From it, it follows that if at the same time that is transmitting to , both over the same subchannel, then

Thus

domain. (This is done by noting that for a tessellation of the plane by squares of side , all squares intersecting a disk of are entirely contained within a larger concentric radius The number of such squares is greater than disk of radius Now take and ) Restricting attention to just these pairs, there are a total of simultaneous bits per second. transmissions, each of range , and each at This achieves the transport capacity indicated. For the Physical Model, a calculation of the SIR shows that Choosing it is lower-bounded at all receivers by to make this lower bound equal to yields the result.

The above lower bounds on feasible transport capacity can be sharpened. The following bounds may be useful in the design of networks with small numbers of nodes. where Thus the same upper bound as for the Protocol Model carries over with defined as above. III. ARBITRARY NETWORKS: A CONSTRUCTIVE LOWER BOUND ON TRANSPORT CAPACITY We will now show that the order of the upper bound in the previous section is sharp for the Protocol Model, by exhibiting a scenario where it is achieved. This scenario is also feasible for the Physical Model. Theorem 3.1: There is a placement of nodes and an assignment of traffic patterns such that the network can achieve bit-meters per second under the Protocol Model, and and bit-meters per second for all Lemma 3.1: In the Protocol Model, there is a placement of nodes and an assignment of traffic patterns such that the network can achieve bit-meters per second bit-meters per second bit-meters per second for for for

bit-meters per second under the Physical Model, both whenever is a multiple of Proof: Consider the Protocol Model. Define

Recall that the domain is a disk of unit area, i.e., of radius in the plane. With the center of the disk located at the origin, place transmitters at locations

Proof: With at least two nodes, clearly bit-meters per second can be achieved by placing two nodes at diametrically opposite locations. This verifies the formula for the bound for With at least eight nodes, four transmitters can be placed at the opposite ends of perpendicular diameters, and each can totransmit toward its receiver located at a distance bit-meters ward the center of the domain. This yields per second, verifying the formula up to These bounds can be further improved slightly by tessellating the domain into hexagons, at the expense of more unwieldy expressions. IV. RANDOM NETWORKS: A CONSTRUCTIVE LOWER BOUND ON THROUGHPUT CAPACITY

and

where

is even. Also place receivers at

and

where is odd. Each transmitter can transmit to its nearest receiver, which is at a distance away, without interference from any other transmitterreceiver pair. It can be verified that there are at least transmitterreceiver pairs all located within the

Now we turn to Random Networks. Even though the setting of the problem is very different, the proof of throughput capacity is somewhat reminiscent of traditional information-theoretic arguments. We provide a constructive scheme to show that one can spatially and temporally schedule transmissions in a random graph so that when each randomly located node has a randomly chosen destination, each sourcedestination pair can indeed be bits guaranteed a virtual channel of capacity , for an per second with probability approaching as We will show how to route traffic appropriate constant

GUPTA AND KUMAR: THE CAPACITY OF WIRELESS NETWORKS

395

efficiently through the random graph so that no node is overloaded. The routing scheme will utilize a Voronoi tessellation of with some special properties. The size of each Voronoi cell is chosen carefully in relation to the number of nodes. Every cell should also be neither too thin nor too fat. The routing will be over nearly straight-line paths, which assures that it is efficient. To show that the load is balanced uniformly over the entire network, we calculate the VapnikChervonenkis dimension for certain geometrically defined random variables on the plane and the sphere, which are connected with the tessellations and routes used. We will need to ensure that the routes are independently and identically distributed. This will require us to circumvent the possible pitfall that knowledge of one route provides information on the locations of the source, destination, and intermediate relay nodes, thus possibly introducing dependencies with other routes which may depend on the locations of these nodes. We begin the constructive proof of the lower bound on the throughput capacity for Random Networks. Our treatment will be directed at the Protocol Model. Where appropriate we will comment on the arguments required for the Physical Model. A. A Spatial Tessellation of the sphere. We use a Voronoi tessellation of the surface Recall the definition of a Voronoi tessellation, see Okabe, Boots be a set of points on and Sugihara [11]. Let (or any other set for that matter). The Voronoi cell is the set of all points which are closer to than to any of the other s, i.e.,

Fig. 2. A tessellation of the surface S of the sphere.

tors. The Voronoi tessellation arising from this set of generators has the desired properties. In the sequel we will use a Voronoi tessellation (V.i) Every Voronoi cell contains a disk of area Let radius of a disk of area for which

on

Above and throughout, distances are measured on the surface of the sphere by segments of great circles connecting two points; see Stilwell [12]. The point is called the generator of Fig. 2 shows an example of a tessellation the Voronoi cell Unfortunately, the surface of the sphere does not allow of any regular tessellation where all cells look the same, except for the platonic solids; see Lyndon [13]. These latter tessellations cannot be made as fine as we need to make them. Moreover, our Voronoi tessellations will also need to be not too eccentrically shaped. We exhibit tessellations with these two special properties in the following lemma, the proof of which is constructive. , there is a Voronoi tessellation Lemma 4.1: For every with the property that every Voronoi cell contains a disk of of radius and is contained in a disk of radius . a disk of radius centered at Proof: Denote by Choose as any point in Suppose that have already been chosen such that the distance between any two s is at least . There are two cases to consider. does not intersect Suppose there is a point such that Then can be added to the collection: Define any Otherwise, we stop. This procedure has to terminate in a finite number of steps since the addition of each removes the area of a disk of radius from When we stop we will have a set of generators such that they are at least units apart, and such that all other are within a distance of from one of the generapoints on

(13) is less (Note that the area of a disk of radius on ). than (V.ii) Every Voronoi cell is contained in a disk of radius We will refer to each Voronoi cell B. Adjacency and Interference Note that all Voronoi cells are polygons since they are formed (or halfspaces in the as finite intersections of hemispheres on ). case of Definition: Adjacent Cells: Say that two cells are adjacent, if they share a common point. (Recall that every cell is a closed set). Let us choose the range of each transmission so that (14) This range allows direct communication within a cell and between adjacent cells. Lemma 4.2: Every node in a cell is within a distance from every node in its own cell or adjacent cell. see Proof: The diameter of cells is bounded by Thus the area cov(V.ii). The range of a transmission is ered by the transmission of a node includes adjacent cells. Definition: Interfering Neighbors: We say that two cells are interfering neighbors if there is a point in one cell which is of some point in the other cell. within a distance as simply a cell.

396

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

As the name implies, the interpretation is this: If two cells are not interfering neighbors, then in the Protocol Model a transmission from one cell cannot collide with a transmission from the other cell. C. A Bound on the Number of Interfering Neighbors of a Cell An important property of the constructed Voronoi tessellation is that the number of interfering neighbors of a cell is uniformly bounded. This will be exploited in the next section in constructing a spatial transmission schedule which allows for a high degree of spatial concurrency and thus frequency reuse. From now on s will be used to denote deterministic constants not depending on Lemma 4.3: Every cell in has no more than interfering neighbors. depends only on and grows no faster than linearly in Proof: Let be a Voronoi cell. If is an interfering neighboring Voronoi cell, there must be two points, one in and the other in , which are no more than units apart. From (V.ii), the diameter of a cell is bounded by Hence , and similarly every other interfering neighbor in the Protocol Model, must be contained within a common large disk of radius Such a disk cannot contain more than disks of radius By (V.i), there can therefore be no more This therefore is an upper than this number of cells within bound on the number of interfering neighbors of the cell The and chosen as result follows from the magnitudes of in (14). D. A Bound on the Length of an All-Cell Inclusive Transmission Schedule The bounded number of interfering neighbors for each cell allows the construction of a schedule of bounded length which to allows one opportunity for each cell in the tessellation transmit. Lemma 4.4: i) In the Protocol Model there is a schedule for transmitting slots, each cell in the tespackets such that in every gets one slot in which to transmit, and such that all sellation transmissions are successfully received within a distance from their transmitters. ii) There is a deterministic constant not depending on , , , , or such that if is chosen to satisfy

neighbors have the same color. This gives a schedule of length at most , where one can transmit one packet from each cell of the same color in a slot. For the Physical Model we will show that under the same schedule as above, the required SIR of is obtained if each that is high transmitter chooses an identical power level enough, and is large enough. Note first that any two nodes transmitting simultaneously are Hence disks of separated by a distance of at least around each transmitter are disjoint. The radius (In the case area of each such disk is at least , but it is smaller for disks on the of disks on the plane surface of the sphere). transmitting to a node at a distance Consider a node The signal power received at is at least less than Now we look at the interference power due to all the other simultaneous transmissions. Consider the annulus of all points A transmitter lying within a distance between and from within this annulus has the disk centered at itself and of raentirely contained within a larger annulus dius and of all points lying between a distance The area of this larger annulus is no more than

Each transmitter above consumes an area of at least , as noted earlier. Hence the annulus of points at a cannot contain distance between and from the receiver more than

transmitters. Furthermore, the received power at from each Noting that there can be no such transmission is at most of other simultaneous transmitter within a distance , and taking and for we see that the SIR at is lower-bounded by

then for a large enough common power level the above result i) holds even for the Physical Model. Proof: First we show the result for the Protocol Model. This follows from a well-known fact about vertex coloring of graphs of bounded degree: A graph of degree no more than can have its vertices colored by using no more than colors, with no two neighboring vertices have the same color; see Bondy and Murthy [14]. One can therefore color the cells colors such that no two interfering with no more than

Since fact smaller and greater than

, the sum in the denominator converges, and is in When is as specified than , the lower bound on the SIR converges to a value .

E. The Source-Destination Pairs Each node wishes to communicate with the node nearest to a randomly chosen location. Let be a randomly chosen location and are independently and uniformly distributed such that is i.i.d. The (i.i.d.) on , and that the sequence

GUPTA AND KUMAR: THE CAPACITY OF WIRELESS NETWORKS

397

destination node for the traffic generated at node is which is closest to chosen as the node Denote by the straight-line segment connecting and Above, and in the rest of the paper, by a straight-line segment we actually mean a segment of the great circle on the surface of the sphere; see [12]. There is one significant property enjoyed by the sequence of straight lines Lemma 4.5: The random sequence of straight-line segments is i.i.d. This has the powerful consequence of allowing us to apply the law of large numbers to the i.i.d. straight-line segments. It will be useful since the route followed by each originationdestination pair will approximate the corresponding straight-line segment, as described in the next section. F. The Routes of Packets We will choose the routes of packets to approximate these will interstraight-line segments. The straight-line segment Let denote the particsect many cells in the tessellation the cell which contains ular cell which contains , and Packets originating at will be relayed from the cell to in a sequence of hops. In each hop, the packet is the cell transferred from one cell to another in the order in which they intersect the line. (If two cells are both next cells, then either can be chosen arbitrarily). Finally, after reaching the cell containing , the packets will be sent on to their final destination, which we shall show later in Section IV-G to be no more than one hop away with high probability. Note that this is a randomized algorithm for choosing routes. It can be thought of as a load balancing scheme with some rather powerful uniformity properties, as shown in Section IV-I. G. Each Cell Contains at Least One Node To make relaying of traffic from one cell to an adjacent cell in confeasible, we need to first ensure that every cell tains at least one node. For this we use uniform convergence in the weak law of large numbers. Note that uniformity is reWe recall the following definitions; quired over all cells in see Vapnik and Chervonenkis [15] and Vapnik [16]. Let be a set of subsets. A finite set of points is said to be shattered such that by if for every subset of there is a set The VC-dimension of , denoted by VC, is defined as the supremum of the sizes of all finite sets that can For sets of finite VC-dimension, one has unibe shattered by form convergence in the weak law of large numbers. is a set of finite The VapnikChervonenkis Theorem: If , and is a sequence of i.i.d. VC-dimension VCrandom variables with common probability distribution , then for every First we will consider the case where is the set of all disks on the plane. Later we will consider the case where the disks In the planar case we can make use of results are located on from Euclidean geometry. The following result may perhaps be known already, though we have been unable to find it in the literature. Lemma 4.6: The VapnikChervonenkis dimension of the set is . of disks in Proof: It is easy to see that there is a three-point set that can be shattered by the set of disks. An example is the set of vertices of an equilateral triangle. of four points that is Suppose there is a set shattered by the set of disks. If any one of the s lies in the convex hull of the other three points, then there is no disk which too. Hence we can can contain the others without containing assume without loss of generality that the convex hull of the four points is a quadrilateral. Again, we obtain a contradiction as follows. Without loss of generality, suppose that the angles of the quadrilateral at and sum to at least 180 , i.e., 180 Suppose is a disk which contains and , but not or ; outwards in both directions see Fig. 3. Extend the diagonal at the points and till it meets the circumference of Simultaneously, let and be the points of intersection of the with the circumference of Then is diagonal a cyclic quadrilateral. However,

Fig. 3. Proof that the vertices of a quadrilateral cannot be shattered by the set of disks.

180 This is a contradiction since the sum of the opposite angles of a cyclic quadrilateral is exactly 180 . Now we address the problem of determining the VC-dimension of disks on the surface of a sphere. It is sufficient for us to restrict attention to disks strictly smaller than hemispheres. To convert results from the plane to , we use a mapping called the inversion map which maps the punctured surface of the sphere onto the plane. Since the radius of the sphere is

whenever VC-

398

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

Fig. 4. The inversion mapping maps disks on S into disks on R

immaterial for the remainder of this discussion, we consider a Let us refer sphere of radius , centered at the point Also let us refer to the plane to it temporarily as as Then the mapping

and excludes and However, we have seen the impossibility of this happening on the plane in Lemma 4.6. Since each cell in the tessellation contains a disk of area (from V.i), we can appeal to uniform convergence in the law of large numbers. such that Lemma 4.8: There is a sequence Every cell contains a node)

where is the Euclidean norm, has several useful properties (see [11]). (i.e., except for the (i) It maps the punctured surface In fact, each point on is origin) onto the plane mapped to the point obtained by extending the ray from the origin to until it hits the plane (ii) (iii) It maps disks on not containing the origin into disks See Fig. 4. on the plane For our purposes, the last property is most important. It is used in the following lemma. Lemma 4.7: The VC-dimension of the set of disks on strictly smaller than hemispheres is . Proof: The proof parallels the contradiction argument of Lemma 4.6. Suppose that there is a set of four points which is shattered by such disks. They all have and to be contained in a disk smaller than a hemisphere. Let be opposite vertices of the quadrilateral formed. Since the set is shattered, there are two disks, each of radius less than that and necessary to form a hemisphere, one of which contains but excludes and , while the other contains and but excludes and Since each disk is strictly less than a hemisphere, there is a point in the complement of their union. Rotate the sphere so that this point is at the top. Without loss of generality we can scale the sphere so that its radius is , and then translate it so that its top is at the origin. Applying the inversion map shows that there is a disk on the which contains and and excludes plane and , and another disk on which contains and

Proof: Let denote the class of disks of area Note that the VC-dimension of is also . Hence Number of nodes in

whenever

This is satisfied when

Since each cell we have

in

contains a disk of area

Number of nodes in for every The result follows. Hence every cell in contains at least one node to relay ). Morethe traffic (with probability exceeding over, every such node has enough range to communicate with all nodes in any adjacent cell (see Lemma 4.2). Hence packets can be relayed from one cell intersecting a line to the next cell intersecting the line. Hence the routing scheme given above can indeed work as planned with probability exceeding

GUPTA AND KUMAR: THE CAPACITY OF WIRELESS NETWORKS

399

Fig. 5.

Transforming great circles intersecting disks into points lying in equatorial bands.

From now on we will use the phrase with high probability, abbreviated as whp to stand for with probability approaching as The multihop relaying scheme can therefore function as planned whp. H. The Mean Number of Routes Served by Each Cell and , where Recall that the straight line connects and are independently and uniformly distributed on By is our assumption (V.ii) on the tessellation , each cell (Note that contained in a disk of radius no more than is less than This allows the area of a disk of radius on intersects a given cell us to bound the probability that a line in Lemma 4.9: For every line and cell

Let denote the great circle containing the line , i.e., the extension of the line so that it wraps around the sphere. The same proof technique shows the following. Lemma 4.10: For every great circle (Great circle intersects and cell

There being a total of lines , one connecting each with , the mean number of lines or great circles passing through a cell is bounded as follows: Number of lines in Number of great circles in intersecting a cell intersecting a cell

intersects I. The Actual Traffic Served by Each Cell Proof: As noted above, from property (V.ii) of the tesis contained in a disk of radius sellation, every cell If lies at a distance from the disk, then the by the disk is no more than angle subtended at The area of the sector so formed is no more than If does joining and cannot not lie in this sector, then the line at intersect the disk containing the cell Hence, for a point containing the a distance from the disk of radius and intercell , the probability that the line connecting sects the disk is no more than Since is uniformly distributed on , the probability density that it is at a distance from the disk is bounded above by Integrating, we obtain Above, since routes follow lines, we have bounded the mean number of routes passing through each cell. However, what we need to bound is the actual random number of routes served by every cell. To do this we make use of the critical property that the seis i.i.d. Hence, so are the straight lines quence This allows us to exploit uniform convergence in the law of large numbers. is contained in a disk of radius Recall that each cell We will bound the number of great circles intersecting This is clearly an upper bound on the such disks of radius passing through cells. number of lines We transform the problem of counting intersections of disks of radius with great circles into a shattering problem let denote the (unique) as follows. For every point on great circle containing all points equidistant from it. This is akin to associating an equator with a pole. Given a great circle , the inverse of this map is not well defined since every equator has two poles. However, we arbitrarily choose one of these two poles and designate it as the inverse Consider a disk of radius centered at a point on Let denote the set of all points ; it is a band of width which are within a distance from around the great circle See Fig. 5.

intersects

400

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

Let denote the set of all disks on following lemma and corollary.

It is easy to see the if and .

rate can be accommodated by all cells if it is less than the rate available, i.e., if

Lemma 4.11: The great circle intersects the disk is contained in the band only if the point

denote the set of all great circles Corollary 4.1: Let The VC-dimension of which intersect is the same as the VC-dimension of Let denote the set of all disks strictly smaller than hemispheres. To appeal to uniform convergence in the law of large numbers we only have to show that the VC-dimension of is bounded. Note that for each is the intersection of two disks, each strictly larger band than a hemisphere. It is trivial that the VC-dimension of a class of sets is the same as the VC-dimension of the class formed by the complements of the sets. It is also known (see Vidyasagar [17]) that if is a set of sets, and consists of sets which are each obtained by intersecting two sets in , then VCVC-

Moreover, within a cell, the traffic to be handled by the entire cell can be handled by any one node in the cell, since each node bits per second whenever necessary. In can transmit at rate fact, one can even designate one node in each cell as a relay node. This node can handle all the traffic needing to be relayed. The other nodes can simply serve as sources or sinks. We have proved the following theorem, noting the linear in Lemma 4.3, and the choice of in growth of in Lemma 4.4 for the Physical Model. Theorem 4.1: in the Protocol Model, there (i) For Random Networks on not depending on , , or , is a deterministic constant such that

Hence we obtain the following lemma. is Lemma 4.12: The VC-dimension of no more than ten times the VC-dimension of In Lemma 4.7, we have already shown that the VC-dimension is . Hence uniform convergence in the weak law of large of numbers holds, and we obtain the following. Lemma 4.13: There is a (Number of lines such that intersecting ) bits per second is feasible whp. It should be noted that these throughput levels have been attained without subdividing the wireless channel into subchannels of smaller capacity. V. RANDOM NETWORKS: AN UPPER BOUND ON THROUGHPUT CAPACITY Now we turn to the proof of the upper bound on the capacity for Random Networks. First we will show that that when the range is too small not every source will be able to communicate with its desired destination. A. Asymptotic Probability of an Isolated Node From [10] we know that a necessary condition for connectivity whp for the problem of nodes strewn on a disk of unit , where The area in the plane is setting here requires a slightly different treatment. The area is not A saving grace in comof a disk of radius on parison to a disk on the plane is that there is no need to consider the tedious issue of edge effects. Another subtle issue is that we may not need connectivity of the entire graph. Strictly speaking, we only need that every source be able to communicate with its chosen destination. What we will show below is that disconnectedness manifests itself by the presence of isolated nodes. These nodes will then be unable bits per second is feasible whp. in the Physical Model, there ii) For Random Networks on are deterministic constants and not depending on , , , , or , such that

Note that if a cell contains , it needs to forward the packet This final destination is at most to its final destination Else, if a cell does not contain , then the one hop away traffic is relayed to the next cell. Hence the traffic handled by a cell is proportional to the number of lines passing through it. carries traffic of rate bits per second, Since each line we have obtained the following bound. Lemma 4.14: There is a such that )

(Traffic needing to be carried by cell

J. Lower Bound on Throughput Capacity of Random Networks From Lemma 4.4 we know that there exists a schedule for slots, each cell transmitting packets such that in every gets one slot to transmit, and such that each in the tessellation of the transmitter. transmission is received within a range Thus the rate at which each cell gets to transmit is bits per second. On the other hand, the rate at which each cell needs to transmit whp. With high probability, this is less than

GUPTA AND KUMAR: THE CAPACITY OF WIRELESS NETWORKS

401

to communicate with any other node. Hence the absence of isolated nodes is indeed a necessary condition for feasibility of any throughput. We recall two results from [10]. Lemma 5.1: (i) For any

(ii) For any given

, there exists for all

, such that

If , then Lemma 5.2: If and for all sufficiently large

, then, for any fixed

Fig. 6. Computing surface area.

A(r), the area of a disk of radius r on a sphere of unit

on From

Given the nodes, denote by the graph which results from connecting nodes separated by a distance less than by an edge. Let denote the has at least one orderprobability that a graph component, i.e., a set of nodes which form a connected set, but which are not connected with any other node. Also, let denote the probability that is disconnected. The main necessary condition for the absence of a single isolated node, and consequently also for connectivity, is the following. Lemma 5.3: If where

Next we compute the area of a disk of radius Note that the radius of the sphere itself is as indicated in Fig. 6, we get

(16) Hence (17) Now is isolated in

then

and Also Proof: Consider first the case where for a fixed Consider , the probability that has at least one order- component. Then and isolated in

(18)

(19) where the first term on the right-hand side above takes into account the case where the distance between and is between and Substituting (18) and (19) in (15) and using (17), we get

is the only isolated node in is an isolated node in and are isolated nodes in

is isolated in and are isolated in (15)

402

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

Using Lemmas 5.1 and 5.2, for and we have

, and any fixed

for all Now, replace by where Then, for any , for all Also, the probability of an isolated node is monotone decreasing in . Hence

Fig. 7.

cannot receive at the same time as

on the same subchannel.

for

Taking limits

Since this holds for all

and

and since

mean length of the path of packets is at least since there of a point on the sphere is always a node within a distance whp. (This was shown in Lemma 4.8). Thus the mean number Since each source of hops taken by a packet is at least bits per second, there are sources, and each bit generates nodes, it needs to be relayed on the average by at least follows that the total number of bits per second served by the To ensure that entire network needs to be at least all the required traffic is carried, we therefore need

the results follow. Corollary 5.1: The asymptotic probability that graph has an isolated node and is disconnected is strictly and positive if B. Upper Bound on Throughput Capacity of Random Networks The key to the upper bound, as in the case of Arbitrary Networks, is to note that each transmission consumes valuable area. Lemma 5.4: The number of simultaneous transmissions on any particular subchannel is no more than

Thus

From the previous section we know that is necessary to guarantee connectivity whp. Hence we obtain the following upper bound. under the ProTheorem 5.1: For Random Networks on , not tocol Model, there is a deterministic constant depending on , , or , such that is feasible

in the Protocol Model. in Fig. 7 transmits successfully to Proof: Suppose node on the th subchannel. Then no other node within node of can be simultaneously receiving a sepa distance arate transmission on the same subchannel due to the requirements (3) and (4) and the triangle inequality. centered at each receiver on Hence disks of radius the th subchannel are disjoint. Since the area of each such , it follows that the network can support no disk is simultaneous transmissions on the th more than subchannel. Noting that each transmission over the th subchannel is of bits per second, by adding all the transmissions taking place subchannels, we see that they at the same time over all the cannot total more than

Note that just as in Theorem 4.1 the number of subchannels is irrelevant. For the Physical Model, the upper bound is as follows. under the PhysTheorem 5.2: For Random Networks on , not ical Model, there is a deterministic sequence depending on , , , or , such that Prob is feasible

bits per second in the Protocol Model. Now let denote the mean length of a line connecting two Then the independently and uniformly distributed points on

where is the mean distance between two points independently and uniformly distributed on the unit area surface of the sphere. bitProof: In Section II we have shown that meters per second is an upper bound on the transport capacity for an Arbitrary Network under the Protocol Model. We will now show that any upper bound on the transport capacity for Arbitrary Networks under the Protocol Model is also an upper bound on the transport capacity for Random Networks under the Physical Model. This will prove the assertion since there are nodes, each having its destination at least meters away on average.

GUPTA AND KUMAR: THE CAPACITY OF WIRELESS NETWORKS

403

Consider any set of successful simultaneous transmissions is sucunder the Physical Model for Random Networks. If over a the th subchannel, at the cessfully transmitting to is also successfully transmitting to over same time that the same subchannel, then from (5)

and so

where Hence any set of simultaneous transmissions feasible for Random Networks under the Physical Model is also feasible in the Protocol Model for Arbitrary Networks. Thus the upper bound on the transport capacity for the latter also holds for the former. VI. THROUGHPUT CAPACITY OF RANDOM NETWORKS ON PLANAR DISK The reader may wonder if the capacity is much different when the network is located on a disk in the two-dimensional plane, rather than on the surface of a sphere. The key issue is whether hot spots created at the center of the domain by several origindestination pairs routing their traffic through the center will make it a bottleneck. The answer is no. The order of the capacity is unchanged for the Protocol Model, and the earlier orders for the lower and upper bounds for the Physical Model continue to hold. Clearly, the arguments for the earlier upper bounds still hold, in view of the same necessary condition on the radius for connectivity (see [10]) in Random Networks under the Protocol Model, and the same reduction of Random Networks under the Physical Model to Arbitrary Networks under the Protocol Model. The critical issue is to show that the earlier lower bounds can still be achieved. We show this by using the same tesselLet be the disk of unit area on lation-based scheme as on the plane on which the nodes are randomly located. Note that just as on , the probability that a randomly chosen line on intersects a disk of radius is no more than This applies even to disks of radius in the center of Thus no unduly hot spots are expected to occur at the center of the domain The key result to show however is that with high probability no hot spots are created anywhere. That is, we need to show the analog of Lemma 4.13 that the number of lines inwhp. Lemma 4.11 tersecting every cell is less than and Corollary 4.1 are not applicable any more since we are not However, we can circumvent this problem as follows. on by using an inverWe map into a large sphere of radius Consider a straight line on Let denote sion map which is the image of the line, and let dethe curve on connecting the two end note the corresponding geodesic on deviates from points. When is large enough, every such by no more than a distance That is, the distortion between the images of straight lines on the disk and the geodesics is very small.

Consider now a cell of the tessellation of It is This disk is mapped into contained in a disk of radius Let be a disk in with another disk larger than that of the same center as , but with a radius It follows that a straight on intersects the disk only if on intersects the disk the corresponding geodesic (The reason is that the enlargement of the radius of accounts for the distortion involved in replacing the images of straight line by geodesics). We have already shown in Section IV-I that the uniform law of large numbers holds for the probability of randomly chosen geodesics intersecting disks. Mapping back on the plane shows that the uniform upper bound on into the number of straight lines passing through the disks of radius applies with high probability. Thus the same results for the capacity continue to hold. Theorem 6.1: For Random Networks on a planar disk of unit area, the results of Theorems 4.1, 5.1, and 5.2 continue to hold, except that in Theorem 5.2, is the mean distance between two points independently and uniformly distributed in the planar disk of unit area. VII. CONCLUDING REMARKS We have shown that under a Protocol Model of noninterference, the capacity of wireless networks with randomly located nodes each capable of transmitting at bits per second and employing a common range, and each with randomly chosen and therefore likely far away destination, is This is true whether the nodes are located on the surface of a three-dimensional sphere or on a planar disk. Even when the nodes are optimally placed in a disk of unit area, and the range of each transmission is optimally selected, a wireless network cannot bits per second provide a throughput of more than to each node for a distance of the order of 1 m away. In fact, summing over all the bits transported, a wireless network on a disk of unit area in the plane cannot transport a total of more bit-meters per second, irrespective of how the than load is distributed. Under a Physical Model of noninterference, the lower bounds are the same as those above for the Protocol for Model, while the upper bounds on throughput are for Arbitrary Networks. Random Networks and Splitting the channel into several subchannels does not change any of these results. These results have some implications that designers may want to consider. Perhaps efforts should be targeted at designing networks with small numbers of nodes. On the positive side, the results show that modulo further medium access or adaptive routing restrictions, communication with nearby neighbors at constant bit rates can be provided in a dense clusters of nodes, since the sourcedestination distances This shows that scethen shrink in scaled length as narios envisaged in collections of smart homes, or networks with mostly close-range transactions and sparse long-range demands, are feasible. We have not considered in this paper the additional burden in coordinating access to the wireless channel, and the additional

404

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 46, NO. 2, MARCH 2000

burden caused by mobility and link failures and the consequent need to route traffic in a distributed and adaptive way. These can only further throttle capacity. It would be useful to quantify these additional burdens. Another issue to be studied is delay. This will arise when the traffic is bursty or when nodes are mobile. These two sources of delay are markedly different. Finally, spatial directivity in the antennas or beamforming will be advantageous in increasing the spatial concurrency of transmissions, since wireless networks can then behave like wired ones. Ephremides [18] has analyzed the medium access problem for a single channel and shown that when only ternary feedback from the channel can be used to schedule transmissions, the throughput of collision-free successful transmissions is the same as in the usual omnidirectional case. When node locations and demands are known and do not have to be figured out purely from ternary feedback, transmissions can be advantageously scheduled so that collisions are avoided, and the throughput can consequently be increased. However, this is a challenging proposition since transmissions from nodes will have to be carefully orchestrated. Such schemes may pose some technological challenges though for low-cost networks. Finally, there is the challange of a more information-theoretic formulation. REFERENCES
[1] What you will want next and the really smart house, Newsweek, May 30, 1999. [2] S. Ramanathan and M. Steenstrup, A survey of routing techniques for mobile communication networks, Mobile Networks and Appl., vol. 1, no. 2, pp. 89104, 1996.

[3] R. Gallager, A perspective on multiaccess channels, IEEE Trans. Inform. Theory (Special Issue on Random Access Communications), vol. IT-31, Mar. 1985. [4] P. Karn, MACA: A new channel access method for packet radio, in Proc. 9th Computer Networking Conf., Sept. 1990, pp. 134140. [5] V. Bharghavan, A. Demers, S. Shenkar, and L. Zhang, MACAW: A media access protocol for wireless LANs, in Proc. SIGCOMM94 Conf. on Communications Architectures, Protocols and Applications, Aug. 1994, pp. 212225. [6] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE Standard 802.111997, IEEE Computer Society LAN MAN Standards Committee, Ed., 1997. [7] N. Bambos, S. Chen, and G. Pottie, Radio link admission algorithms for wirelessd networks with power control and active link quality protection, in Proc. IEEE INFOCOM, Boston, MA, 1996. [8] S. Ulukus and R. Yates, Stochastic power control for cellular radio systems, Preprint, 1996. [9] L. Valiant, A theory of the learnable, Commun. ACM, vol. 27, pp. 11341142, Nov. 1984. [10] P. Gupta and P. R. Kumar, Critical power for asymptotic connectivity in wireless networks, in Stochastic Analysis, Control, Optimization and Applications: A Volume in Honor of W. H. Fleming, W. M. McEneany, G. Yin, and Q. Zhang, Eds. Boston, MA: Birkhauser, 1998, pp. 547566. [11] A. Okabe, B. Boots, and K. Sugihara, Spatial Tessellations Concepts and Applications of Voronoi Diagrams. New York: Wiley, 1992. [12] J. Stilwell, Geometry of Surfaces. New York: Springer-Verlag, 1992. [13] R. C. Lynwood, Groups and Geometry. Cambridge, U.K.: Cambridge Univ. Press, 1985. [14] J. A. Bondy and U. Murthy, Graph Theory with Applications. New York: Elsevier, 1976. [15] V. N. Vapnik and A. Chervonenkis, On the uniform convergence of relative frequencies of events to their probabilities, Theory Probab. its Applic., vol. 16, no. 2, pp. 264280, 1971. [16] V. N. Vapnik, Estimation of Dependences Based on Empirical Data. New York: Springer-Verlag, 1982. [17] M. Vidyasagar, A Theory of Learning and Generalization. London, U.K.: Springer-Verlag, 1997. [18] A. Ephremides, Some wireless networking problems with a theoretical conscience, in Codes, Curves, and Signals: Common Threads in Communications, A. Vardy, Ed., Aug. 1998.