1468

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

Information Theoretic Capacity of Cellular Multiple Access Channel with Shadow Fading
Dimitrios Kaltakis, Student Member, IEEE, Muhammad A. Imran, Member, IEEE, and Costas Tzaras, Member, IEEE

AbstractIn this paper, we extend the well-known model for the Gaussian Cellular Multiple Access Channel originally presented by Wyner. The rst extension to the model incorporates the distance-dependent path loss (maintaining a close relevance to path loss values in real world cellular systems) experienced by the users distributed in a planar cellular array. The density of base stations and hence the cell sizes are variable. In the context of a Hyper-receiver joint decoder, an expression for the information theoretic capacity is obtained assuming a large number of users in each cell. The model is further extended to incorporate the log-normal shadow fading variations, ensuring that the shadowing models are fairly comparable to the free space model. Using these fair models the effect of the shadow fading standard deviation on the information theoretic capacity of the cellular system is quantied. It is observed that a higher standard deviation results in lower capacity if the mean path loss is appropriately adjusted in order to model the mean loss due to the physical obstacles causing the shadow fading. The results validate that larger cell sizes and a higher standard deviation of shadowing (with appropriately adjusted mean path loss) results in lower spectral efciency. Index TermsInformation theory, land mobile radio cellular systems, multiple-access communication, Gaussian channels, multipath channels, propagation.

I. I NTRODUCTION N the past, wireless systems were designed to accommodate a large number of voice and/or low data rate users. With the emergence and continuous growth of wireless data services, the value of a wireless network is not only dened by how many users it can support, but also by its ability to support high data capacity at localized spots. Shannon developed a mathematical theory for the singlelink channel capacity in [1] providing the framework for the study of the performance limits in communications. Shannons work gave birth to the eld of information theory. In information theoretic literature different approaches have been reported for determining the maximum data rate and the means to achieve this under various assumptions and constraints. Despite the work in this eld, the rst important attempt to study the capacity of a cellular system was carried out in the previous decade by Wyner [2]. Wyners model studies
Paper approved by V. A. Aalo, the Editor for Diversity and Fading Channel Theory of the IEEE Communications Society. Manuscript received November 12, 2007; revised October 9, 2008 and May 29, 2009. D. Kaltakis and M. A. Imran are with the Centre for Communications Systems Research, University of Surrey, Guildford, Surrey, GU2 7XH UK (e-mail: D.Kaltakis@surrey.ac.uk). C. Tzaras is with Kimatica Ltd, Greece. Digital Object Identier 10.1109/TCOMM.2010.05.070591

the uplink channel and, although it considers a very crude approximation of the path loss with no path loss variability across the cell, it manages to provide an insight into the cooperation between the base stations and the benets that can be achieved through that process. Fading was incorporated in Wyners model by Somekh and Shamai in [3]. They maintained the assumption of a hyperreceiver with delay-free access to all cell-site receivers and assumed the same interference pattern as Wyner. They used a raster-scan method to transform the two-dimensional system into an equivalent linear system in order to arrange the fading coefcients and the systems path gains in the channel matrix. Their results showed that for a certain range of relatively high inter-cell interference, the fading improves the system performance when compared to the case without fading. Letzepis in [4] modied the one-dimensional version of Wyners model to account for the free-space path loss incurred between user transmitters and cell-site receivers. The optimum capacity is estimated by using the Shannon-transform [5] of the Marcenko-Pastur law [6]. To keep the problem tractable, all users are assumed to be collocated at the cell-site receivers position. The major contribution of the work is that it accounts for the interference emanating not only from the two neighboring cells but from all cells in the system, enabling the study of the effects of changing cell-density in the system. Another extension to Wyners one-dimensional model was proposed in [7] where the users are modelled as being located at the cell boundaries with their signals received only by two cell-site receivers. In [8] we nd the optimum capacity, employing a Wynerlike two dimensional model, assuming that each cell-site receiver is receiving signals from 5 tiers of users. Each tier is delineated based on the users average distance from the center of the cell of interest. Each tier is considered to have a different path gain which corresponds to a specic power law path loss. This tier-based approximation of the distance-dependent pathloss provided a good insight into how the capacity is affected if one removes the crude assumption that all users in a cell experience the same attenuation. One could obtain results even closer to reality if the users were considered to be distributed in the planar system and a different path attenuation coefcient was assigned to each one of them according to their distance from the Base Station (BS) of interest. Up till now, no attempt to incorporate shadow fading into Wyners model has been reported. There are many results on

c 2010 IEEE 0090-6778/10$25.00

KALTAKIS et al.: INFORMATION THEORETIC CAPACITY OF CELLULAR MULTIPLE ACCESS CHANNEL WITH SHADOW FADING

1469

while , refer to the dB power value of the respective coefcients. II. C HANNEL M ODEL A. Input-Output Equation We consider a Wyner-like [2] hexagonal cellular array model with 2 cells and uniformly distributed users in each cell. A BS in a cell will receive signals from all the users in the system. It is more practical to obtain a representation of the cellular system in terms of a rectangular array, since the representation in Figure 1 is less tractable. This can be achieved [2] by scaling and rotating the structure in Figure 1. The points of the rectangular array are indexed by (, ) {(1, 1) . . . (, )} where and are the number of the rows and columns respectively. According to the above, the received signal at the base stations antenna referring to cell (, ) is the sum of the transmitted signals, with each one being multiplied by the corresponding path gain and fading coefcients. Hence, the received signal in a cell (, ) is given by: , = +
[ =1 [ , =1

Fig. 1.

Planar array of cells.

the capacity of conventional wireless systems with shadow fading [9],[10]. In [11] the authors derive the Shannon capacity of a single-link channel under the effect of log-normal fading. The capacity is a function of the probability distribution function of the received signal-to-noise-ratio (SNR). Tse and Hanly in [12] derive the single-user delay-limited capacity with shadow fading. There have also been attempts to characterize the ergodic capacity of composite matrix channel models under the effect of large and small-scale fading (see [13] and references therein). In this paper we provide a model which can be used to evaluate the uplink capacity of a cellular system incorporating the notion of a hyper-receiver employing realistic system parameters. Using this model we were able to derive a closedform formula for the uplink capacity of a cellular system in which each transmitted signal experiences a distancedependent path loss, fast fading and shadow fading. Using this formula we study the effects that increasing cell radius and increasing the standard deviation of the shadowing component have on the capacity of the system. A. Outline This paper is organized as follows: In section II we dene the channel model. This denition includes the input-output equation and the distance dependent path loss along with the realistic system parameters that are incorporated in our model. We also dene the small-scale and large-scale fading models that are employed for the derivation of the results. In section III we present the mathematical model and derive the closed form formula for the capacity of the system. In section IV we present the denition of Rise over Thermal (RoT) and the Monte Carlo simulation setup used for validating the theoretical results. The results obtained for the systems capacity are also presented. Finally, in section V, some conclusions are drawn. B. Notation Throughout this paper the following notations are used: , refer to the linear amplitude value of the respective coefcients. Moreover, , refer to the linear power value,

] ,, ,, ,, + ] , , , , , , + , (1)

where subscript (, ) identies the cell in which a user is located and represents the user index. Subscript (, ) { (, )} represents the set of all base stations excluding the one indexed by (, ). Furthermore, , ,, , stand for the path gain coefcients, fading coefcients, complex Gaussian inputs, complex Gaussian outputs and normalized (unit power) noise respectively. A power constraint is also considered for each input, [(,, )2 ] . B. Fading Model In order to describe the fast fading coefcient, , we consider the following equation [14]: 1 = e + (0, 1) (2) + 1 + 1 where is the ratio of the Line-of Sight (LoS) and non-Lineof-Sight (NLoS) components and is the phase of the received signal. The fading coefcient is assumed to be normalized with respect to unit power. In this model, corresponds to the case where all power is concentrated in the specular path. Whereas 0 corresponds to the most severe fading i.e. Rayleigh fading with no specular component. Note that this model is valid even for a wideband system if the bandwidth is sufciently divided such that the fading can be considered to be at in each division [3], [14], [15]. Since a cellular system with distributed users is assumed, the transmitted signals cannot be synchronized with respect to the phase from the receivers point of view. This, in fact, means that the signals are received with random, independent, normally distributed phase offsets.

1470

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

250 PCS Microcell Model COST231 Extension to Hata Model =2 =3 =4

TABLE I VALUE /R ANGE OF PARAMETERS USED FOR FINDING THE CAPACITY OF

THE PRACTICAL CELLULAR SYSTEMS

200

150

100

Parameter Cell Radius Reference Distance Path Loss at 0 Path Loss Exponent UTs per cell UT Transmit Power Thermal Noise Density Carrier frequency Channel Bandwidth

Symbol 0 0 0

Value/Range (units) 0.1 10 1 103.8 2-4 usual values {2, 3.5} 20 100-200 mW -169 dBm/Hz 1.9 GHz 5 MHz

Pathloss (in dB)

50

0 0 10

10

10 10 Distance (in metres)

10

10

Fig. 2. Comparison of empirical models with simplied path loss model using appropriate parameters for simplied path loss model.

For the capacity analysis that will be presented later in this paper, the expected value of the product of two fading coefcients will be needed. Since the received signals have random phase offsets, the mean value of the product of two independent fading realisations is zero for Rayleigh ( 0) as well as for Rician ( = 0) fading distributions. On the other hand, the expectation of the product of a fading coefcient with the complex conjugate of itself is equal to its power, and thus equal to one (since the fading coefcients are assumed to be normalized with respect to unit power). C. Path Loss Model We use a modied (by adding 1 at the denominator to avoid possible divisions with zero during the analysis) power law path loss model which can be expressed as:
2 =( 0 0 0

1+

(3)

where 0 is dened as the power received at the reference distance (0 ) when the transmitted power is one unit and + 0 is the actual distance from the BS. Note that 0 is also a function of the carrier frequency . It should be noted that the modied path loss model is not completely arbitrarily selected, but has a strong one-to-one correspondence to a practical system; a system into which any transmitter can be placed at a distance from the receive antenna which is constrained to the range [0 , ]. This also implies that ranges in [0, ], which creates an exclusion zone with a diameter of 20 around the BS of interest. The modied path loss model attempts to capture two important phenomena in the physical system. The parameter 0 captures the actual attenuation at the reference distance and the path loss exponent captures the rate at which the attenuation increases with the distance. In order to provide a one-to-one correspondence between the modied path loss model and the empirical models, the values of these parameters need to be determined with the objective of providing a best-t of the

empirical data obtained in the eld. We have selected two well-known empirical models. For a micro-cellular system the wideband PCS model [16] is selected, while for a macrocellular system we selected the PCS extension to the HATA model [16]. In order to compare the modied path loss model to the empirical models we plot the results obtained by the two empirical models and the results for the modied model with varying in Figure 2. Based on the limitations of the two models (in terms of the parameters that can be used) we employ the following parameters to approximate the path loss. For both models we use carrier frequency =1.9GHz, transmit antenna height = 1.5 and power loss at the reference distance 10 log10 0 = 38dB. We use the minimum allowed receiver antenna height for the macrocellular system model (30m) and the maximum allowed for the microcellular system (13.3m). It can be observed that the microcellular model suggests a smaller value of = 2, and the macrocellular model suggests a much larger value of = 3.5. Even for the microcellular model, the path loss exponent after the Fresnel zone clearance is larger, i.e. 3.5. The macrocellular model also suggests a further loss of around 20 dB. Hence, for practical results, we use a value of 10 log10 0 = 38 dB and a path loss exponent of 2 for small cells and 3.5 for large cells. In Table I we summarize the value/range of parameters that will be used in this paper for nding the capacity of the Multiple Access Channel with shadow fading. D. Shadow Fading Model To incorporate shadowing into the system, we need to nd a model which can accurately represent the variation in received amplitude in the system due to the shadow fading [15], [16]. Goldsmith in [15] proposes the log-normal shadowing model which accurately models the above-mentioned variation. In the log-normal shadowing model, the pdf of the random 2 ) is power gain (ratio of receive to transmit power, given by: (10 log10 )2 (4) ( ) = exp 2 2 2

10 where = ln 10 , (average dB path gain due to the distance-dependent path loss [15]) is the mean of = 10 log10 in dB, and is the standard deviation of

KALTAKIS et al.: INFORMATION THEORETIC CAPACITY OF CELLULAR MULTIPLE ACCESS CHANNEL WITH SHADOW FADING

1471

Users with approximately the same path loss

r dr

2
r dr 2
^

received signal experiences an independent fading coefcient, . The statistical properties of these coefcients determine the capacity. To evaluate [] , the following denition is used. Denition: The expectation of the product of a complex fading coefcient and the complex conjugate of another complex fading coefcient is dened as: [ 2 ] ) ( This expectation is required, since through the knowledge of it, every element of [ [] becomes known. ) ( ] - If = then: ( ) = 2 . This is the case with the diagonal elements of [], giving all diagonal entries in a closed form for the expectation of the covariance matrix. [ ] - If = then one needs to nd the value of ( ) = 2 that will appear in some off-diagonal elements of []. As a result, according to the fading model presented in section II.B in this paper, all the off-diagonal entries of the expected covariance matrix are zero and all the diagonal entries are the same (with [ ] = 1) and are given by: [ 2] [[] [] ] = (8)
1 2 lim , , 2 =1
2

2 K rdr
Fig. 3.

users

Reference area containing users with the same path loss.

in dB. In our case, is a function of the path power loss for a distance from the Base Station. The conditional pdf of the total power gain , given the distance dependent power loss , is given by: 2 (10 log ( )) 10 ( ) = exp 2 2 2 (5) where ( ) is a function of which needs to be determined. III. C APACITY OF THE S YSTEM A. General Capacity Formula Following the classical Information Theoretic approach [17], the per-cell uplink capacity is given by: ( ) ( ) 1 log2 det[I + HH ] = lim (6) 2 where the matrix H is an 2 2 channel attenuation matrix for users per cell and a planar array of cells (or a linear array of 2 cells) and . Note that H can be modelled as the overall system matrix which is constituted of the amplitude/envelope attenuation (path-loss) factor and the fading coefcients, i.e. H = G where contains the shadow fading coefcients ( ), G contains the fast fading coefcients ( ) and denotes the Hadamard operator (element-wise multiplication of matrices). We make a basic simplication here by assuming that there are a sufciently large number of users in each cell. When 1, the expectation in (6) can be moved inside the logdet and then we have [3]: ) ( ( [ ]) 1 (7) log det I + HH = lim 2 2 This gives an upper bound for the capacity which is tight when is large (refer to Appendix A). So, the problem reduces to that of characterizing the mean value of I + HH for the given statistical properties of fading and path gain coefcients. Path attenuation coefcients are a function of the User Terminals (UTs) positions. As all these coefcients are dependent on the distances between the transmitter and the receivers whose positions are xed, is a deterministic matrix for a given snapshot of the system. Each

(9)

where , is the attenuation factor between the base station of interest and the user in the system. All the diagonal entries are the same and each entry corresponds to a specic BS in the system. This simplies (7) to the following form: )]) ( [ ( 1 2 log2 det I + diag 2 [ ] (10) = lim 2 But: [ 2] = [ ] = ( )d , (11) with ( ) denoting the probability distribution function of the power attenuation factor given by (5). B. pdf of the Distance Dependent Path Loss It is known that if the pdf of a random variable is given by (), then it is possible to calculate the pdf of some variable , which depends on via = (). Assuming is a monotonic function the resulting density function is: ( ) = 1 (1 ( )) ( ) 1 ( ) , (12)

where = 1 ( ) is the inverse function of = () and is the rst derivative of that function. Let us now consider the planar cellular system described in users per unit area. The probability density section II with function of the user position needs to be dened as in (12) and the pdf of the distance dependent path loss in the system needs to be found. To do this, consider that is the radial distance from the BS of interest. Let us also consider that all users in the disc formed by the two circles with radii + 2 and 2 (where is very small) experience approximately

1472

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

14

x 10

Pdf of path loss

12

10

1) Capacity of a System without shadowing : Using equation (10) one can derive the per-cell capacity of a system where each users signal experiences multipath fading and distancedependent path loss as: ( [ 2 ]) (19) no-shadowing = log2 1 + 2 with: [ 2] =
0
0 (1+)/2

8 p ()

2 ( )d

(20)

Using (18) one can calculate (20) and thus provide a closed form solution for the per-cell capacity of the system (equation A).
60 50 40 30 path gain value (dB) 20 10 0

0 70

C. Capacity with Shadow Fading Let us consider the log-normal shadow fading model that was presented in section II.D of this paper. We nd the capacity as follows. The conditional pdf of the two random variables, the total power gain and its mean is given by (5). Since in (15) represents the distance-dependent attenuation, equation (15) needs modication in order to represent the distance-dependent power loss, at position , as: = () = 0 (1 + ) (21)

Fig. 4. pdf of the distance-dependent attenuation for the planar system. Coverage 3 Km, path loss factor 3.5, 0 = 0 dB.

the same path loss. The number of users with the same path loss can be found as (see Figure 3): [ ( ( )2 )2 ] + = = 2 2 2 (13) Assuming a maximum radial distance , the probability of a user being in this area (thus experiencing this path loss) is 2 . Thus, the pdf of the user position is given by: 2 () = 2 , [0, ] 2 (14)

Let us consider the path loss model described in section II.C. The distance dependent path attenuation for each user in the system is given by: 0 , (15) = () = (1 + )/2 where , 0 are as dened above. Now we can use (14), (15) in conjunction with (12) to nd the probability density function of the distance dependent free space attenuation. From (15) we have:
2/ = 1 ( ) = 0 1 1/

Now, following the same steps as in (16), (17), (18) equation (12) will yield the pdf of the distance-dependent power loss: ( ) +2 +1 2 2/ 1/ ( ) = (22) 0 0 2 Since, in (5), the value of the mean distance-dependent power loss is in dB, it is considered useful to transform eq. (22) so as to represent the pdf of the distance-dependent power loss also in dB. To do that, equation (12) is once more used for the following transformation: = 10 log10 , (23)

with ( ) given by (22). Thus, the pdf of the random variable which is the distance dependent power loss in dB is: ( (+2) (+1) ) 2 2/ 1/ 10 10 10 10 10 10 0 0 2 (24) The capacity can be calculated using (10) and by noticing that: ( ) = [ 2] E = E [ ] = 10 log10 0 = ( )
10 log10
0 (1+)

(16)

and () =

0 2(1 + ) 2 +1

(17)

Thus using (14), (16), (17) equation (12) yields: ) ( +4 +2 4 2/ 1/ ( ) = 2 0 0

(18)

Equation (18) is the probability density function of the random variable , the distance-dependent attenuation in a planar system with the users uniformly distributed and the attenuation 0 taking values between (1+)/2 and 0 (See Figure 4).

>0

( )d d = (25)

10 log10 0
0 (1+)

10 log10

( ) E [ ] d

KALTAKIS et al.: INFORMATION THEORETIC CAPACITY OF CELLULAR MULTIPLE ACCESS CHANNEL WITH SHADOW FADING

1473

no-shadowing

)) ( ( 2 log2 1 + 2 2 0 ln(1 + ) + 1 1 = 2 1+ ( ) = 23 1 2 2 ( 1)(1+) +( 2)(1+) log2 1 + 2 2 0 (1+) > 2 ( 2)( 1)(1+)23

(A)

Mean value of path gain without shadowing Mean value of path gain with shadowing Variation due to shadowing

10

pdf of the shadowing coefficient for different confidence intervals

Area under the curve: 0.5

10

10

Area under the curve: 0.00135

10
p ( )
p

negative offset path gain

10

10

10

Area under the curve: 3.3x105

10

10

10

12

no shift shift by 3 standard deviations (99.85% confidence) shift by 4 standars deviations (99.995% confidence) path loss value 10
5

10

total path gain

10
p

10

10

distance

Fig. 5. Simple explanation for the negative offset that needs to be found so as to be able to compare the capacity with and without shadowing

Fig. 6. pdf of the shadowing coefcients for different values of the standard deviation multiplier and respective constant path loss value. The areas under the curve values denote the probability there will be points with value higher than that of the path loss in the absence of shadowing.

But: E [ ] = 2 (10 log ( )) 10 = = exp 2 2 2 >0 2 ( ln ( )) = exp d = 2 2 2 >0 2 ) ( , (26) + = exp 2 2 where we have used the substitution = . Using equations (24), (26) it is derived from (25) that: 2 E [ ] = 2 0 10
1/

of the standard deviation of the shadowing component, since the higher the standard deviation, the further away the points of the path gain lie from their mean value (See Figure 5). As the normal distribution has an innite tail, one possible way to quantify this negative offset is to use the condence intervals of the log-normal distribution so as to claim that a certain percentage (e.g. 99.99%) of the shadowing coefcient points lie below the mean value of the no-shadowing path gain. In Figure 6 we plot the pdf of the shadowing coefcient (5) that is obtained by setting its mean to a specic value. It is obvious that if no negative offset is introduced then there is a high probability (50%) that the shadowing coefcient value will be higher than the path loss value in the absence of shadowing, which is not possible. Thus we dene: ( ) , (28)

10 log10 0
0 (1+)

10 log10
(+1) 10

10

10

0 10

2/

(+2) 10

(27)

) ( exp 2 + 2
2

( ) is the distance-dependent mean of the shadowing coefcient. It depends on , the distance dependent power loss when no shadowing is present. We need to nd a relationship between these in order to be able to derive the capacity of the system. One observation that can be made here is that by incorporating shadowing in a system, the mean value of the path gain, when compared to the situation without shadowing, should be decreased. This means that a negative offset in the mean of the random shadowing coefcient for our model needs to be found. The distribution of this random variable is given by (5). This negative offset shall be a function

where is the standard deviation multiplier determining the condence interval of the normal distribution. A value of 3 means that we are operating at 99.85% condence, while a value of 4 means a 99.995% condence. In Figure 6 we have calculated the probability, that for two different shifts of the mean value, there will be shadowing coefcients with greater values than the path loss coefcients in the absence of shadowing. Using (27), (28) equation (10) yields the per-cell capacity of the system with shadow fading, which can be expressed by equation (B). IV. R ESULTS A. Denition of Rise over Thermal (RoT) In the practical engineering design of cellular systems, the main gure of merit that determines the capacity (maximum reliable transmission rate with vanishingly small error rate) of

1474

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

2 2 = log2 1 + 2

10 log10 0
0 (1+)

10 log10

10

10

2 ( (+2) (+1) ) 2/ 1 / d 0 10 10 0 10 10 exp 2 + 2

(B)

a UT, is the Signal-to-Interference-Noise-Ratio (SINR) at the BS receiver, given as SINR = + 0 (29)

where is the received power at the BS of interest, 0 is the thermal Additive White Gaussian Noise (AWGN) at the receiving BS and is the inter-cell and intra-cell interference received from other UTs of the system. However, in the information-theoretic analysis of hyper-receiver cellular systems, the main gure of merit that determines the per-cell capacity (at any BT) is: , (30) RoT = 0 assuming that all UTs in the system transmit at their maximum allowable power constraint, . The factor denotes the relative attenuation experienced by the transmitted signal of each UT until it reaches the receiver. The numerator term is the total received signal power (desired signal power for the base station in consideration and also the power of the signals intended for the other base stations in the system). Splitting the numerator into desired , , and (conventionally termed) undesired signals, , we can express RoT as: + , (31) RoT = 0 Which shows that the information theoretic approach of using a joint decoder has the potential of converting the conventionally harmful interference into a factor that increases the gure of merit RoT by moving the interference term from the denominator to the numerator. Note that the numerator of (30) is a direct function of the transmit power of the UT. Therefore, we can dene the ratio 0 and can also use this as the gure of merit. With this denition incorporated, the RoT is given by: , (32) RoT =

whole cell. In this direction, the per-cell capacity can be calculated as a function of , which is a xed system parameter, common to all the UTs of a cell. It shall be noted that the problem of nding the (per-cell) capacity of a cellular system can be greatly simplied by focusing on the single BS receiver and its RoT. Due to the symmetry of the problem (ignoring the edge effects) all BS receivers are identical and system capacity is simply the percell capacity times the number of cells. From (10) it follows that: 2 [ ] 1 2 1 + 2 [ log2 ] = lim 2 =1 ( ) 2 = log2 1 + 2 [ ] = log2 (1 + RoT) , where RoT
2

(33)

2 2 , = 2 [ ],

(34)

=1

as discussed before. This result suggests that the capacity formulations of the cellular system all fall on the same graph and the different system parameters dene the range of operation on this graph by controlling the RoT in the system. B. System Simulation For the Monte Carlo simulations used to compare the theoretical results to simulations, the users are uniformly distributed on the coverage area and their distances from each Base Station are calculated. The distance-dependent path attenuation for each user is then calculated according to (15) and the respective values are placed into the 2 2 matrix s . Fading coefcients are simulated as random normally distributed complex numbers with both real and imaginary parts having zero mean and a standard deviation of 1/ 2. In order to generate the shadowing coefcients for the Monte Carlo capacity simulation, we generate log-normally distributed random variables for each users position, with the respective mean given by (28). The value for each is taken from s . These variables are then placed in the 2 2 matrix s . Then, the capacity formula given by (6) is evaluated for many realizations of the Hs = s Gs matrix. C. Capacity Results In this section, we present the capacity results obtained by plotting (A) and (B) versus various system parameters. Following the denition of RoT above, in Figure 7, we illustrate where our system operates on the capacity (bps/Hz/cell) v.s. RoT curve. The capacity is plotted for different cases per path loss exponent value; no shadowing, mild shadowing conditions

The main reasons that the SINR does not constitute an appropriate gure of merit for information-theoretic analysis are: Inter-cell and intra-cell interference is not harmful and thus the term cannot be used in the denominator when a joint decoder is considered. Since there is no harmful interference, there is no need for power control and thus the UTs constantly transmit with the maximum available power . In this context, the transmit power remains xed for all the UTs, whereas the received power at each BS differs for each UT. In addition, since the objective function is the percell capacity, the power variable affecting the value of this function should have a constant value throughout the

KALTAKIS et al.: INFORMATION THEORETIC CAPACITY OF CELLULAR MULTIPLE ACCESS CHANNEL WITH SHADOW FADING

1475

RoT curve 25
Capacity region with no shadowing =2

20 Capacity in bps/Hz per cell

Capacity region with 4dB shadowing =2 Capacity region with 8dB shadowing =2

15

10
Capacity region with 4dB shadowing =3.5

5
Capacity region with 8dB shadowing =3.5

0 20

10

10

20 30 RoT in dB

40

50

60

70

Fig. 7. Per cell capacity of the system with and without shadowing versus Rise over Thermal in dB. =80 users per-cell, reference distance power loss -38 dB. Cell radius 1000m. User power ranging from 100 mW to 200 mW. The RoT curve is also plotted.
Capacity for different parameters No Shadowing Shadowing with standard deviation 4 dB, 2 Shadowing with standard deviation 8 dB, 2 Shadowing with standard deviation 4 dB, 3.5 Shadowing with standard deviation 8 dB, 3.5 Simulation results

25

Capacity in bps/Hz per cell

20

(4 dB standard deviation) and more severe shadow fading (8 dB standard deviation). The capacity region is obtained by varying the UT transmit power. The standard deviation multiplier that was selected for plotting the results was the one that provides 99.995% condence. As the RoT curve illustrates the optimum capacity of the Multiple-access Channel under the notion of a hyper-receiver, Figure 7 provides a very useful insight into the operating region of systems under real-world assumptions. In Figures 8 and 9 we plot the per-cell capacity of the system for different path loss exponents v.s. the cell radius and the per-user transmit power respectively. The capacity difference for the same path loss factor is almost constant with respect to the cell radius, 4.7bps/Hz capacity loss when we compare the no-shadowing and 4dB shadowing cases and 3.47 bps/Hz capacity loss when the severity of the shadowing increases from 4dB to 8dB for both path loss exponent values. On the other hand, when we compare capacities with different path loss exponents, it is observed that the capacity loss increases slightly as the cell radius increases (capacity loss ranging from 4.48 to 5.18 bps/Hz when we compare the two 4dB shadowing cases). Another interesting insight provided by Figure 8 is that the capacity for path loss exponent 3.5 and shadowing standard deviation 4dB is always less than the capacity for path loss exponent 2 and standard deviation 8 dB. V. C ONCLUSIONS In this paper, we proposed a model which can be used to evaluate the uplink capacity of a cellular system under the notion of a hyper-receiver incorporating realistic system parameters. We derived a formula for the uplink capacity of a cellular system in which each transmitted signal experiences a distance-dependent path loss, fast fading and shadow fading. The results illustrate the uplink capacity range, in bps/Hz/cell, with which a system employing full receiver co-operation is expected to operate. By incorporating some realistic assumptions about the system parameters we endeavoured to make the results as close to reality as possible. The increasing standard deviation of the shadowing component, as well as the increasing cell radius, have, as expected, a negative impact on the capacity. Another interesting insight is that increasing the path loss exponent value has a larger impact on the capacity than increasing the shadowing components standard deviation. Due to the fact that we did not assume distancedependent standard deviation of the shadowing component, the capacity gap between different shadowing conditions is constant with respect to the cell radius, but it decreases as the severity of the fading increases. A PPENDIX A L AW OF L ARGE N UMBERS IN THE C APACITY F ORMULA Consider the 2 2 matrix H with its elements to be random variables following some distribution. Consider also the following multiplication: = HH (35)

15

10

500

1000

1500 2000 Cell radius in meters

2500

3000

Fig. 8. Per cell capacity of the system with and without shadowing. =80 users per-cell, user power 200 mW, reference distance power loss -38 dB. Cell radius ranging from 100m to 3Km.
Capacity v.s. user power 26 24 22 Capacity in bps/Hz per cell 20 18 16 14 12 10 8 6 4 100 110 120 130 No shadowing (=2) Shadowing with standard deviation 4 dB (=2) Shadowing with standard deviation 8 dB (=2) Shadowing with standard deviation 4 dB (=3.5) Shadowing with standard deviation 8 dB (=3.5) Simulation Results 140 150 160 User power in mW 170 180 190 200

Fig. 9. Per cell capacity of the system with and without shadowing. =80 users per-cell, reference distance power loss -38 dB. Cell radius 100m. User power ranging from 100 mW to 200 mW

where H is the 2 2 Hermitian transpose matrix of H . Each element of the matrix is the result of the multiplication of a row of matrix H (which is a 2 vector)

1476

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

with a column of matrix H (again a 2 vector). Hence, each element of matrix is the 2 sum of random variables multiplied with the conjugate transpose of other random variables with all of them following the same distribution: =
2

[ ]

(36)

=1

From the above one can say that for a large number of for every xed ,the law of large numbers applies to each element of [ HH ]and thus reaches its expected value, 2 . i.e. ACKNOWLEDGMENT The work reported in this paper has formed part of the Fundamental Limits to Wireless Network Capacity Elective Research Programme of the Virtual Centre of Excellence in Mobile & Personal Communications, Mobile VCE, www.mobilevce.com. This research has been funded by the following Industrial Companies who are Members of Mobile VCE - BBC, BT, Huawei, Nokia, Nokia Siemens Networks, Nortel, Vodafone. Fully detailed technical reports on this research are available to staff from these Industrial Members of Mobile VCE. The authors would like to thank Prof. G. Caire and Prof. D. Tse for the fruitful discussions. R EFERENCES
[1] C. E. Shannon, A mathematical theory of communication, Bell Systems Technical J., vol. 27, pp. 379423, 1948. [2] A. Wyner, Shannon-theoretic approach to a Gaussian cellular multipleaccess channel, IEEE Trans. Inf. Theory, vol. 40, no. 6, pp. 17131727, Nov. 1994. [3] O. Somekh and S. Shamai, Shannon-theoretic approach to a Gaussian cellular multiple-access channel with fading, IEEE Trans. Inf. Theory, vol. 46, no. 4, pp. 14011425, July 2000. [4] N. Letzepis and A. Grant, Information capacity of multiple spot beam satellite channels, in Proc. 6th Australian Communications Theory Workshop, 2005, Feb. 2005, pp. 168174. [5] A. M. Tulino and S. Verdu, Random matrix theory and wireless communications, Commun. Inf. Theory, vol. 1, no. 1, pp. 1182, 2004. [6] V. Marcenko and L. Pastur, Distributions of eigenvalues of some sets of random matrices, Math. USSR-Sb, vol. 1, pp. 507536, 1967. [7] O. Somekh, B. M. Zaidel, and S. Shamai, Sum rate characterization of joint multiple cell site processing, to appear in IEEE Trans. Inf. Theory. [8] D. Kaltakis, M. A. Imran, and C. Tzaras, Information theoretic capacity of cellular networks with multiple-tier interference, submitted to Vehicular Technology Conference, 2008. [9] K. Gilhousen, I. Jacobs, R. Padovani, A. Viterbi, J. Weaver, and I. Wheatley, On the capacity of a cellular CDMA system, IEEE Trans. Veh. Technol., vol. 40, no. 2, pp. 303312, May 1991. [10] A. Naguib, A. Paulraj, and T. Kailath, Capacity improvement with basestation antenna arrays in cellular CDMA, IEEE Trans. Veh. Technol., vol. 43, no. 3, pp. 691698, Aug. 1994. [11] A. J. Goldsmith and P. P. Varaiya, Capacity of fading channels with channel side information, IEEE Trans. Inf. Theory, vol. 43, no. 6, pp. 19861992, Nov. 1997. [12] S. V. Hanly and D. N. C. Tse, Multiaccess fading channelspart II: delay-limited capacities, IEEE Trans. Inf. Theory, vol. 44, no. 7, pp. 28162831, Nov. 1998.

[13] C. B. Chae, D. Mazzarese, and J. R. W. Heath, Ergodic capacity of spatial multiplexing MIMO channels with log-normal shadowing and rayleigh fading, in IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, 2007 (PIMRC 07), Sep. 2007. [14] D. Tse and P. Viswanath, Fundamentals of Wireless Communications. Cambridge University Press, 2005. [15] A. Goldsmith, Wireless Communications. Cambridge University Press, 2005. [16] T. S. Rappaport, Ed., Wireless Communications: Principles and Practice. Upper Saddle River, NJ: Prentice-Hall, Inc., 1995. [17] I. E. Telatar, Capacity of multi-antenna Gaussian channels, Eur. Trans. Telecommun., vol. 10, no. 6, pp. 585595, Nov. 1999. Dimitrios Kaltakis received the B.Sc. degree in electrical and computer engineering from the National Technical University of Athens (NTUA), Greece, in 2005 and the M.Sc. and Ph.D. degrees in Mobile Communications from the University of Surrey, U.K. Currently, he is a Member of the Centre for Communications Systems Research, University of Surrey. His main research interests are in the area of wireless systems and focused on fundamental performance limits of communication systems. He has signicant contributions to the research project on fundamental capacity limits for Mobile Virtual Centre of Excellence (MVCE), UK. During his research, he has published several articles in international journals and conferences. Muhammad Ali Imran is currently a Lecturer in the Centre for Communication Systems Research (CCSR) of University of Surrey. Dr Imran received his B.Sc. degree in Electrical Engineering from University of Engineering and Technology Lahore, Pakistan and the M.Sc. and Ph.D. degrees from Imperial College London, UK. He has done his post doctoral research at University of Surrey for two years. His main research focus is on the physical layer, fundamental performance limits of communication systems, resource allocation and signal processing at transmitter and receiver end for MIMO and cooperative systems. He has signicant contributions to the research project on fundamental capacity limits for Mobile Virtual Centre of Excellence (MVCE), UK. He is coordinator of self-organising networks theme in India-UK Advanced Technology Centre project funded by Engineering and Physical Sciences Research Council (EPSRC), UK. He is also contributing to European Commission funded ROCKET project and preparing for the future European project EARTH for Green Wireless Communications. He has an extensive experience of simulation and modelling tools for communication systems. During his research, he has published several articles in international journals and conferences. Costas Tzaras received the B.Sc. degree in physics from the University of Thessaloniki, Greece and the M.Sc. and Ph.D. degrees in Mobile and satellite communications from the University of Surrey, U.K. In 1997, he worked as a Research Fellow at the University of Surrey and later he was employed by Vodafone R&D in the Radio Modelling Group to act as the Principal Investigator in radio modelling issues across the Vodafone Group. In 2002, he returned to the University of Surrey as a Lecturer in Mobile Communications and he continued to lead the Antennas and Propagation Team at Vodafone Global. Since 2007, he is working for Kimatica Ltd which is a software company that specializes in the eld of radio modelling. His main interests lie in the eld of wireless technology with emphasis on radio channel modelling, network planning and optimization and capacity investigations on wireless systems.