You are on page 1of 13



Co-Channel Interference Modeling and Analysis in a Poisson Field of Interferers in Wireless Communications
Xueshi Yang and Athina P. Petropulu, Senior Member, IEEE
Abstract—The paper considers interference in a wireless communication network caused by users that share the same propagation medium. Under the assumption that the interfering users are spatially Poisson distributed and under a power-law propagation loss function, it has been shown in the past that the interference instantaneous amplitude at the receiver is -stable distributed. Past work has not considered the second-order statistics of the interference and has relied on the assumption that interference samples are independent. In this paper, we provide analytic expressions for the interference second-order statistics and show that depending on the properties of the users’ holding times, the interference can be correlated. We provide conditions under which the interference becomes -dependent, -mixing, or long-range dependent. Finally, we present some implications of our theoretical findings on signal detection. Index Terms—Alpha-stable distributions, interference modeling, long-range dependence, non-Gaussian, signal detection, wireless communications.

I. INTRODUCTION N WIRELESS communication networks, signal reception is often corrupted by interference from other sources, or users, that share the same propagation medium. Knowledge of the statistics of interference is important in achieving optimum signal detection and estimation. Existing models for interference can be divided into two groups: empirical models and statistical–physical models. Empirical models, e.g., the hyperbolic distribution and the -distribution [1], fit a mathematical model to the practically measured data, regardless of their physical generation mechanism. On the other hand, statistical–physical models are grounded on the physical noise generation process. Such models include the Class A noise, which was proposed by Middleton [2], and the -stable model initially proposed by Furutsu and Ishida [3] and later advanced by Giordano [4], Sousa [5], Nikias [6], and Ilow et al. [7]. A common feature in
Manuscript received March 1, 2002; revised August 8, 2002. Part of this work was presented at the IEEE—EURASIP Workshop on Nonlinear Signal and Image Processing, Baltimore, MD, June 2001, and the 11th IEEE Workshop on Statistical Signal Processing, Orchid Country Club, Singapore, August 2001. Part of this work was supported by the Office of Naval Research under Grant N00014-03-1-0123. The associate editor coordinating the review of this paper and approving it for publication was Dr. Rick S. Blum. X. Yang was with the Department of Electrical Engineering, Princeton University, Princeton, NJ 08544 USA. He is now with Seagate Research, Pittsburgh, PA 15222 USA (e-mail: A. P. Petropulu is with the Electrical and Computer Engineering Department, Drexel University, Philadelphia, PA 19104 USA (e-mail: Digital Object Identifier 10.1109/TSP.2002.806591


these interference models is the rate of decay of their density function, which is much slower than that of the Gaussian. Such noise is often referred to as impulsive noise. Impulsive noise has been observed in several indoor (cf. [8], [9]) and outdoor [2], [10] wireless communication environments. In [11], measurements of interference in mobile communications suggest that in some frequency ranges, impulsive noise dominates over thermal noise. Impulsive noise attains large values (outliers) more frequently than Gaussian noise. Such noise behavior has significant consequences in optimum receiver design [6], [12]. As optimum signal detection relies on complete knowledge of the noise instantaneous and second-order statistics [13], it is important to study the spatial and/or temporal dependence structure of the noise as well as its instantaneous statistics. In the last decade, many efforts have been devoted in this direction (see, for example, [14]–[17]). In [15], the authors consider a physical statistical noise model originating from antenna observations that are spatially dependent and show that the resulting interference can be characterized by a correlated multivariate Class A noise model. For mathematical simplicity, temporal dependence structure has been traditionally modeled by moving average (MA) [18] or Markov models [19]. However, it is not clear whether these models are consistent with the physical generation mechanism of the noise. Thus, use of these models in receiver design may lead to schemes that function poorly in practice. In this paper, we consider statistical–physical modeling for co-channel interference. In particular, we are interested in the temporal dependence structure of the interference. We adopt and extend the statistical–physical model investigated in [5]–[7]. The model considers a receiver surrounded by interfering sources. The receiver picks up the superposition of all the pulses that originate from the interferers after they have experienced power loss that is a power-law function of the distance traveled. We assume a communication network (time slot) and focus on the with basic waveform period . As assumed in [5]–[7], at interference sampled at rate any time slot, the set of interferers forms a Poisson field in space. Assuming that from slot to slot these sets of interferers correspond to independent point processes, it was shown in [5]–[7] that the sampled interference constitutes an independent identically distributed (i.i.d.) -stable process. However, the independence assumption is often violated in a practical system. To see why this would be the case, consider an interferer who starts interfering at some slot and remains active for a random number of slots. That interferer will still be active at time slots

1053-587X/03$17.00 © 2003 IEEE

if (1) for . . or session life. It should be noted that in this paper. In this paper. i. Codifference -stable distributions are known for their lack of moments of .e. A particular class of heavy-tail distributions with infinite variance is the -stable distribution. and for sign sign where is the scale parameter. we give conditions for the interference to be -dependent.and second-order characteristic function of the sampled interference and show that under certain assumptions. From (6). is replaced by . Section III details the interference model. is the skewness parameter. In particular. although it reduces to the Gaussian . where A quantity that is closely related to the codifference is [20] is slowly varying function. at most. e. This paper is organized as follows. It reduces to processes. Numerical simulations are presented in Section VI. infinite variance if it is regularly varying with index . if the session life is heavy-tail distributed.YANG AND PETROPULU: CO-CHANNEL INTERFERENCE MODELING AND ANALYSIS 65 . the role of covariance is played by the covariation or the codifference [20]. we assume that the interferer’s holding time. In the latter case. is a random variable with some known distribution. such a univariate -stable . If .. Section VIII contains some concluding remarks. whereas the dependence structure of the interference is studied in Section V. one time slot.g. it can for some positive constant and be seen that a long-memory process is characterized by an autocorrelation that decays hyperbolically as the lag increases. for second-order statistics do not exist. Long-Range Dependence is called a (wide-sense) staA second-order process tionary with long memory. B. with certain probability. C. We obtain the first. II. is -stable in if and A vector of only if there exists a finite measure on the unit sphere and a vector in such that the characteristic function is given by [20] sign (2) (4) and is referred to as generalized codifference [21]. such slowly varying functions are constant functions or ratios of two polynomials with identical degree). we only consider modeling of co-channel interference. This is in contrast with the exponential decay corresponding to short memory processes. If . In short. Section VII discusses some implications of long-range dependence in signal detection. consists of two points In the case when and . . -stable distributions are more conveniently characterized by their characteristic functions. which does not include thermal noise presented in any practical communication systems. the order greater than or equal to . Independence can only be valid if the interferer remains active for. Finally. .e. and will be one of the . Section II provides the relevant mathematical background. In particular. the distribution becomes a univariate -stable . the codifference for the case of jointly (5) is defined for any stationary heavy-tailed random process. In such a case. The parameter is the characteristic expo- nent. for all positive . MATHEMATICAL BACKGROUND A. -mixing. or simply . . . random The codifference of two jointly and equals variables (3) is the scale parameter of the variable . The -stable distribution is often classified as non-Gaussian.. auto-regressive moving average (ARMA) processes. the variance of is infinite (if and moments of order greater than or equal to are infinite). and the spectral measure is concentrated on them. The statistics of the interference are derived in Section IV. the mean In those cases. Correspondingly. random variable will be denoted by the distribution is symmetric about and is termed symmetric -stable. if its is finite and satisfies [22] autocorrelation function (6) . Heavy-Tail Distributions and -Stable Distributions A random variable is regularly-varying with index as Here. Due to lack of closed-form expression for their probability density functions. the interference constitutes a long-range dependent process in a generalized sense (which is defined in Section II-C). i. . (typically. the interference becomes jointly -stable.. A direct consequence of this is interferers at are that location-wise. or long-range dependence. its characteristic function equals distribution. the interferers at time slots and not independent of each other. The difference between -stable and Gaussian case when densities is that the tails of the former are heavier than those of the latter. and is the location parameter. The variable is said to be heavy-tailed with . or long-range dependent.

random variables. [23]: Let is a long-memory process in a generalized sense if say that . satisfies (7) where is some real positive constant and III. NO. the power attenuation increases logarithmically with the distance between the transmitter and the receiver (cf. propagates through the transmission medium and the i. cf. As the sampling time tends to infinity. where is a random variable with known distribution. Let us define the term emerging interferers at time interval to describe the interfering sources whose contribution arrives for the first time at the receiver in the beginning of time interval . IV. as a result. Proposition 1: Assume that the mean of the interferers’ session life is finite. we will assume that the users. i. The modeling is performed first for a a disk is taken. where the asterisk denotes convolution. to larger than 5. the signal transmitted from the th interfering user.e.e. [5] and [7]. We will assume that a user. and the density of the emerging interferers at each time slot is . and correlations of represents interference. the limit for three-dimensional (3-D) scenario is a straightforward extension of the two-dimensional (2-D) one. The power loss function can be expressed in terms of signal ampli. reasonable to assume that the interferers that emerged at two different time slots are independent or. Let be the projection of onto at time slot . 51. For the sake of simplicity. At time . We will focus for a particular .. as defined in (4). where a receiver receives the signal of interest in the presence of other interfering signals. A nonhomogeneous Poisson process can be mapped to a homogeneous one through transformations. the result of the where and with the basis functions .e. let us consider the instantaneous interference at time . which is an extension to the model described in [5]–[7]. once started transmission. Due to filtering only. receiver filters. the total signal at the receiver is (9) is the signal of interest. [2]). we drop the subscript . more precisely. STATISTICAL ANALYSIS OF THE RESULTED INTERFERENCE A. VOL... A correlation demodulator decomposes the received signal into a -dimensional vector. The finite . the contribution of the th insponse . be a stationary process. which are contributions of active interferers.e. the propagation channel and the receiver can be represented by a time-invariant filter with impulse re. Thus. tude loss function (8) where is the path loss exponent.66 IEEE TRANSACTIONS ON SIGNAL PROCESSING. and the users are distributed within the disk according to a 2-D Poisson point process. we assume that distortion and attenuation can be separated.i. . JANUARY 2003 The following generalization of the concept of long memory process can be useful for processes which lack autocorrelation. respectively. (10) It holds that (11) (12) and are. The interferers that emerged at any time interval are located according to a Poisson point process in the space (Poisson field) with density .. thus denoting by and . correspond two independent Poisson point processes. terfering source at the receiver is of the form 1 may be a function of time and the locations of the unit area/volume. which forms a nonhomogeneous Poisson point process. . At time . i. In wireless communications. Instantaneous Statistics Now. Consider a wireless communication scenario without power control. The is independent of for ].d. the number and locations of the active interferers at any given time interval need to be specified. The receiver is placed at the origin of the coordinate system. and at the final step. Let us first consider only the filtering effect. we only consider the homogeneous case. in dense urban environments and hard partitioned office buildings [11]. we present the interference model. We Definition 1 [21]. For simplicity. and. For notational conon the statistics of and venience. For short time intervals. at time .  is a constant. for hallways within buildings. continuously emits pulses for a duration of time slots. INTERFERENCE MODEL In the sequel. and denotes the set of where interferers at time . the active interferers become Poisson distributed in the space with asymptotic density . In order to calculate the instantaneous statistics. i. Note that the transmitting power of user has been implicitly incorporated into . In this paper.. is the sum of a random number of i. respectively. which is denoted by . of course. The combination of these two groups would make the interferers at slots and dependent (location-wise). in addition to interferers that emerge at . and is a function of the antenna height and the signal propagation environment. s are assumed to be spatially independent [e. The signal is expanded into a series of orthonormal basis functions . The receiver consists of a signal demodulator followed by the detector. there could be interferers that emerged at some slot and still stay active at . It may vary from slightly less than 2. One issue that the model of [5]–[7] does not take into account is that. and the receiver are all on the same plane and concentrated in of radius . 1. It is understood that the latter group would exist if the holding times of the users were longer than one time slot .1 It is. which are the potential interferers. gets attenuated and distorted.g.

YANG AND PETROPULU: CO-CHANNEL INTERFERENCE MODELING AND ANALYSIS 67 Proof: See Appendix A. we will set the time origin for the system confrom now on. Hence. it is posing the condition that distributed (cf. at some constant level be uniformly distributed between [ 1/2. However. (8) may serve as a good approximation [5]. let us also assume that the th interferer (21) Here. provides the joint characteristic function of . (22) Proof: See Appendix B. which is the mean of Recognizing that the session life. It holds that (14) (15) and represent the set of interferers that are acwhere and . it is conventional to use a truncated form (27) B. In many cases.. respectively. Joint Statistics We next study the joint statistics of interference samples obtained at different time slots and . i. the interference does become jointly -stable. [5. shown that the resulting interference is (13)–(15)]). The latter can be verified by checking the characteristic function of . We assume . 1/2]. Let As before. tends to zero as tends to infinity and 3) If approaches the mean of session life . 2) In the special case (24) we have Hence for and . i. Hence. does not vary with . we obtain the first-order characteristic 1) Setting function of the interference process. and . Then. . this is consistent with the findings in [5] and [7]. is the characteristic function of .. at any time . by imis symmetrically distributed. Now. Indeed.e. we can conclude the following. for some special (16) for general distributed . they are becoming independently -stable distributed (see also Section V-B). the joint characteristic function may be simplified as (26) Equation (26) implies that when the distance between two samples becomes asymptotically large. Let us denote the interference at and by respectively. Remarks on Proposition 2 . where. we get a result consistent with Theorem 1. in general. we can use related results in [5] and [7]. 1) Such a case would arise if the th interferer remained throughout its session life. the set sidered to be of active interferers form a Poisson field in the space with density . As approaches zero. cases of We next consider two such cases. Special Cases—Jointly -Stable Interference It is interesting to note that although the sampled interference is marginally -stable. Theorem 1: The instantaneous interference observed at any distributed with chargiven time interval at the receiver is and scale parameter acteristic exponent (13) where is the common characteristic function of . C. (25) Equation (25) implies that the interference samples obtained at different time slots are independent and jointly -stable distributed. 4) The loss function in (8) is a far field approximation. it is not jointly -stable. (23) . such truncation leads for some constant to an intractable analytical solution. D. Proposition 2: Let the session life have finite mean. the we denote by and joint characteristic function of equals (16) where (17) (18) (19) (20) . and its survival distribution function. For convenience. Therefore. the power of the signal becomes infinite. However. and is the second-order characteristic function of . for typical values of . The following proposition tive at time and .e.

Proposition 3: A necessary condition for to be -mixing is as was defined in (20). (20). 51. it implies [24. tends to zero as so that the condition is automatically satisfied. particularly in spread spectrum networks. it is straightforward to show that if the max. . A more realistic model would be to assume that the symbols are either 1 or 1 with equal probability and independent from slot to slot. In that case. For which is the -algebra generated by the indicated random variis -mixing if there exists a ables. 2) Bernoulli Distributed In wireless communications. Note that for most distributions. In this section. define . In applications of particular interest is the class of -mixing for which (37) Under this condition. For example. For convenience. Such a property is useful when asymptotic convergence problem is considered. are defined as (17). . and (33) Again. and . Bernoulli distributed for different . and long-range dependent. The following provides the necessary condition for the interference to be -mixing. JANUARY 2003 emerges at . by (36). the contribution from interferer is (31) is i.i. If -mixing [24. and its session life is . the imum possible value of the session life of users is interference becomes -dependent. we will tribution give conditions under which the resulted interference is -dependent. a necessary condition of (37) is given by (38) (35) From (32). -Dependent is A stationary random sequence said to be -dependent if there is a non-negative integer . DEPENDENCE STRUCTURE OF CO-CHANNEL INTERFERENCE Equation (16) implies that the dependence structure of the co-channel interference is determined by the session life dis. then is also the past and the present. samples of the -mixing noise (or their nonlinear transformations) comply with the standard central limit theorem (CLT). -Mixing (30) denotes the characteristic function of . with such that for each non-negative sequence . (32) implies that the interference at jointly -stable. 182].68 IEEE TRANSACTIONS ON SIGNAL PROCESSING. (32) where . Then. Intuitively. If together imply . p. it can be shown (see Appendix C) that (29) where . variables. (37) is imposed on the -mixing noise considered. although it is not difficult to verify that the results .d. VOL. 1] (34) -mixing. Then. It lies in the fact that if . Now. . where is Proof: Assuming that Lem. NO. (19). such that the sequences and are statistically independent for all satisfying . we will focus on the case of Bernoulli distributed [see (32)] for which the interference is jointly -stable distributed. -mixing. where taking 1 or 1 with equal probability of 1/2. the left side of (35) is V. satisfying the following. provided presented also hold for generally distributed has finite second-order statistics for all that such that (21) is well defined. respectively. The contribution from this interferer can be written as (28) is the indicator funcwhere is integer number. 1. it can be shown that (see Appendix D) A more general dependence structure of importance is -mixing. and for each . B. where Equation (29) implies that the interference is jointly -stable (see Appendix C). are defined as before. we we also have say that the process is symmetrically -mixing. A. and so is the interference. most of the interference is due to co-channel users who are transmitting data sequences during their session life. Then. the distant future is virtually independent of is -mixing. and are (36) and the result follows. and tion. in [25]. in a -mixing process. -mixing defines for a stationary sequence of random . say . For the interference integers model on hand. the interferers at time instances separated by more than are independent. to apply the asymptotic relative efficiency as a criterion for signal detection.

Ch.e. A random variable has a Zipf distribution [28] if (a) (b) Fig.i. A more rigorous statistic method to test distributed is the whether the experimented data is indeed QQ-plot. In this case. with tail index Bernoulli random variables taking possible values 1 and 1 with equal probability 1/2. The path loss was power-law with . multiple access terminals utilize the same frequency channel. and scale parameter tributed. 1. in spread spectrum packet radio networks. The motivation behind considering heavy-tail distributed session life is that such a distribution can well characterize many current and future communication systems. they can be modeled by Pareto distributions [23]. [27]. we set . Note that eter . The wireless communication network link was subjected to interferers that are spatially Poisson distributed over a plane . we find that . We employ the method of sample fractiles [6. and is the tail index. for simplicity. conventional tools such as covariance that measures the dependence structure are not applicable. 1(b).. [26]). [27]. e. i. which compares the quantiles of experimented data to distributed ( ) data. Should the exthose of the ideally distributed. and are i. can be characterized by our interference model. Interference presented in a communication link in a Poisson field of interferers: (a) = 1. we study the asymptotic dependence structure of the interference when the session life is heavy-tail distributed. Then. as described in Section IV-D. (41) . We synthesized ideally same parameters ( and ) and plotted the quantiles plot versus (39) is an integer denoting the location parameter. jointly -stable distributed with long-range dependence in the generalized sense when users’ holding time is heavy-tail distributed. In this . we here assume that the session life is Zipf distributed (a discrete version of the Pareto distribution). The session life distributions in future wireless systems are expected to be similar to the ones in current wireline networks. distributed with scale paramstantaneous interference is . In particular. For example. Assuming the data is -stable dis. Bernoulli distributed. respectively. For a discrete-time communication system. 1(a)]. which paper. which are very close to their theoretical value and . Instead. 5. with density Once the interferers become active. we simulate a wireless network link with Poisson distributed interferers and Bernoulli distributed s.g. as shown in trace with Fig. (40) where time lag between time intervals.i.e. it exhibits strong impulsiveness [impulsiveness becomes weaker as decreases. Assuming that multiuser detection and power control are not implemented. as we will see. their resource-request holding times (session life) are distributed with much fatter tail than that of voice only network users (cf. the interference from other users.d. As increasingly more wireless users are equipped with internet-enabled cell phones. where is the Riemann Zeta implies that function. and .YANG AND PETROPULU: CO-CHANNEL INTERFERENCE MODELING AND ANALYSIS 69 which can be satisfied by most distributions. as defined in (33) Proof: See Appendix D. (38) is not met.3] to estimate the various parameters for the first trace [see Fig. The signals received at the receiver consist of superposition of the signals from all the users in the network. it forms another interesting class of dependence structures. where is the mean of the session life. we use the codifference [see (4)] to explore the dependence structure.d. as evidenced by Fig. according to (33). VI. Our goal is to show that the simulated interference is consistent with our theoretical findings. where denotes the scale parameter. the QQ-plot would perimental data be indeed distributed noise with the be linear.. We will denote the tail index of the session life by to avoid confusion with the defined in (17). then the resulted interference is longrange dependent in the generalized sense. and the inequal probabilities. the exponential family. taking values 1 with . was [28] distributed with taken to be i. Both traces have been normalized with respect to their empirical standard deviation]. As expected. Long-Range Dependent In this section. which in our simulations are Zipf and . One segment of the simulated interference process is shown in Fig. i. they stay in the active state for random time durations. NUMERICAL SIMULATION RESULTS In this section. However. or otherwise referred to as self-interference. for which extensive statistical analysis of high-definition network traffic measurement has shown that the holding times of data network users are heavy-tailed distributed [23]. Since the interference is marginally heavy-tail distributed. tail index of the session life distribution. 1(a) [for comparison purposes. (b) = 1:82. Proposition 4: If the session life of the interferers are Zipf .. we also simulate another corresponding to . distributed. 1(b)]. C.

(a) Log–log plot of the codifference of the interference in 40 Monte Carlo simulations. In the second case. respectively. 4(a) and (b).4. 3(a). despite the fact that Cauchy noise only consti). We consider the problem of binary signal detection in non-i. when the session life is heavy-tail distributed. indicating that the interference is long-range dependent in the generalized sense. and the mean slope of the fitted lines is found to be 0. 1. In the first case.d. If . the test statistic is vation (43) is the dispersion of the interference ( ). In Fig.i. A least squares line was fitted to each simulation run. that of the interference (length of 5000 points) in Fig. (b) Mean (solid line) and standard deviation (dotted lines) of the codifference. NO. which should lead to a long-range dependent interference. We estimated the codifference of the inaccording to [29]: terference (42) is the data length. The figure clearly demonstrates that the instantaneous interference distribution can be modeled well by the As shown in Section V-C. Diamond-marked lines represent the long-range dependent (LRD) case. we set the session life to be Zipf distributed with .d. JANUARY 2003 Fig. IMPLICATIONS OF IMPULSIVE AND LRD INTERFERENCE The dependence structure of the interference should be taken into account in signal detection and estimation. the interference exhibits long-range dependence in the generalized sense. we also plotted the mean (solid line) and the standard deviation (dotted line) of the estimated codifference. In particular. We observe that although the mean BER where . 2.d. Let us assume that the transmitter is sending binary signals or with equal probability and the transmission is corrupted by the noise . which is the Cauchy distribution.i. The linear trend is clearly seen in the graph.i. Note that according to Proposition 4.i. For deciding between the two hypotheses. This point is highlighted in the following example. For 40 Monte Carlo simulations. case.3829. star-marked lines are for the i. noise modeled as -stable for a wide range of .d. . which should result in an i. Other parameters are selected such that the dispersion of the interference are the same for the two situations. It has been shown in [30] that the Cauchy receiver performs robustly in the presence of i. let us adopt the Cauchy receiver developed in [30]. we decide that has been transmitted at time . we overlap the estimated codifferand ences in a log–log scale. Forty Monte Carlo simulations were run for the parameters described above. QQ-plot of simulated interference and ideally S S distributed random variable with the same parameter. 3(b). In Fig. where is the mean the Zipf distribution. interference.d. note that We performed 40 Monte Carlo simulations with data length of 5000 each. VOL. Fig.i. Given the obsertutes a special case of -stable noise ( . the theoretical value is 0. 51. and is some small multiplicative where constant. both having identical marginal distributions. otherwise. 2. VII. 3. we overlook the noise dependence structure and treat it as if it were white.70 IEEE TRANSACTIONS ON SIGNAL PROCESSING. The consequence of ignoring the dependence is studied via the bit-error-rate (BER). The estimated value is in good agreement with the theory. noise. = 1. we ferers in first scenario by . corresponding to various values of . Each run was based on 5000 points . noise to scenarios where noise is highly correlated can yield unexpected degradation of receiver performance. the corresponding mean BER along with the standard deviation (normalized by the mean BER) are shown in Fig. but for detection. we set the session life to be equal to 1. Denoting the density of the interand in second scenario by . applying signal detection algorithms optimized for i. we decide in favor of Two different noise processes are simulated based on the proposed model.

the discrepancy beas smaller tween the BER standard deviations further increases. of the BER statistics on the noise LRD index [ This will be subject of future investigation. Key assumptions were that individual interferers have certain random session life. The dependence structure of the interference must be taken into account to attain optimum signal detection. Mean and standard deviation of BER of Cauchy receiver in the = 1:2) and i. the active interferers are spatially Poisson distributed in the space. such as the interferers sending BPSK signals. Fig. APPENDIX A PROOF OF PROPOSITION 1 denote the number of active interferers at Proof: Let the th time interval. and there exist long periods where the sequence stays in low levels [22].d.d. be the survival function of the session life.d. We showed that although the interference process is marginally .d. -stable noise for 40 Monte Carlo simulations.i. i. It would be of interest to study analytically the dependence .e. whereas star-marked lines denote the i. for signal detection in LRD interference based on finite data length. -stable noise for 40 Monte presence of LRD ( Carlo simulations. case.i. It is interesting to see that in this case. The standard deviation of the BER in the LRD case is significantly larger than in the i. the asymptotically is . one often observes cycles and trends. see (40)]. alternatively. There exist long periods where the maximal level tends to stay high. taking the time origin to be . except in some special cases. This observation implies that the highly correlated interference can degrade the performance of the Cauchy receiver. Thus.. . or LRD. corresponding counting quantity .d. implies stronger dependence.i.i. Signal detection in the presence of impulsive and LRD noise is still an open problem and is currently under investigation. although the process is stationary. Letting tend to with parameter or. in a short segment of the series.d. case. Some preliminary results shown in this paper indicate that the performance of traditional detectors may deteriorate significantly under LRD noise. the probability of error oscillates in a wider range than in an i. where At any given time interval. by noting that a combined Poisson point process is still Poisson. case. . we investigated the statistics of the interference resulted from a Poisson field of interferers. We obtained the instantaneous and second order distributions of the interference. Therefore.2. It holds that (44) is the time when the sources emerged. 4. and where instability of the receiver may lead to erroneous performance.i. VIII. vis-à-vis the i. -mixing.YANG AND PETROPULU: CO-CHANNEL INTERFERENCE MODELING AND ANALYSIS 71 (a) (a) (b) Fig. and s correwhere is the indicator spond to their session life (in multiples of ).i. whose distribution is a priori known.d. where finite data length is available. 5(a) and (b) illustrate the case when =1. there is a large discrepancy between the standard deviation of the BER. function. (b) Fig. A heuristic explanation for this phenomenon lies in the observation that LRD time series exhibit strong low-frequency component. CONCLUSIONS In this paper. Let (45) Since the number of new interferers at each time interval is Poisson distributed is a Poisson random variable. Mean and standard deviation (normalized by mean) of BER of Cauchy receiver in the presence of LRD ( = 1:5) and i. case. This is of particular importance in practice. noise environment. whereas star-marked lines denote the i. and the signal propagation attenuation follows a power law function. Diamond-marked lines represent the LRD case.i. s are chosen such that the mean BERs are similar to the previous example. 5. are close to each other in the two cases. We provided conditions under which the interference becomes -dependent. or constant amplitude signals. it is in general not jointly -stable distributed. Diamond-marked lines represent the LRD case.

The same independence . than session life ) 2) active at but not ( (denoted by ). . From (50). We then have for (46) represents the set of interferers which are active at time . which has a radius of . and . Thus. Let us use . Forinterferers starting at time . NO. i. the summations of can be replaced by the double summation the form . we get and sources started in at in at (48) in (49) and can be easily deduced where the definitions of by (48). the sum is first taken for all sources restricted in a disk centered at the receiver . respectively.72 IEEE TRANSACTIONS ON SIGNAL PROCESSING.they can be classified into three groups according to their active/inactive states in and : 1) inactive at both and .. this set can be further partitioned as (47) contains the interferers that are active where the set and emerged at time slot . . . VOL. Therefore. and . . 1. According to their emerging time. Thus. JANUARY 2003 APPENDIX B PROOF OF PROPOSITION 2 Proof: The joint characteristic function of the total interference at time and is A. due to the Poisson assumption. the conclusion applies to interferers in joint characteristic function becomes (50) To compute (50). . and into cases for We divide the calculation . ) (denoted 3) active at both and (session life by . 51. are inBy assumption. We should note here that at will not contribute the interferers that emerged after time to the interference received at . respectively. their session life are shorter (let denote this set of interferers). and to denote the number of elements in sets set . the interferers in dependent of each other for different . at (51) . their locations are independent and uniform distributed on the disk . Assuming there are total sources that started emission at time .e.

from (56). we get (58) Proceeding in the manner of [5]. the Since the survival function of the session life is probability that an interferer that emerged at time will remain active until is (52) The probability that an interferer that emerged at will survive until but will die out at is (53) while the probability that an interferer will be inactive at is (54) If there are interferers beginning their emission at . we get (57) The logarithm of (50) equals (55) . where Combining (55) and (51).and second-order . the second term in (58) can be simplified as (59) where (56) (60) where Denoting and . characteristic function of . the probability that of them are active until time and of them are active at but not is Then.YANG AND PETROPULU: CO-CHANNEL INTERFERENCE MODELING AND ANALYSIS 73 where and are the first.

it remains active until . Note that (68) where . This far. Nevertheless. 51. Since interferers emerged between and may die out before .74 IEEE TRANSACTIONS ON SIGNAL PROCESSING. . we need to consider them separately from the interferers that emerged at . JANUARY 2003 we finally get ilar reasoning as before. Reasoning as before. for we have (64) represents the interferers started at but died out where and to enumerate and before . We will use . For any interferer starting at time with probability sion before with probability . be grouped as either active or inactive at . respectively. for as defined in (60). Hence. we get For started in in at (67) (65) where as defined before in (60). contributing interference only at time . whereas it ends emis. these interferers can . Next. (66) (63) Equations (61)–(63) define the contributions from interferers that emerged before time . we may proceed to calculate (65) as in at (61) Letting . These interferers emerged after time . we are done with the computation for interferers that emerged at or before time . Interferers that emerged at time need to be treated differently since they always contribute to the interference at while not necessarily at . then (62) and where C. 1. NO. Following sim- . B. we are continuing with the calculation of the second product term in (49). VOL.

following similar steps. Proof: We need to calculate gives us . we prove (29). To Next. we can show that for (69) Plugging (61)–(67) and (69) into (49). The log-characteristic function of the random equals variable (77) Now. we get APPENDIX D PROOF OF (32)—BERNOULLI CASE The characteristic function of a Bernoulli random variable. which can be (75) (70) Since (76) we have (71) plugging (71) into (16). Hence. Plugging (39). Th. we have is independent of and (81) Since the upper bound (80) and lower bound (81) converge as . we conclude that (82) . see this. is Note that (80) (73) In (73). and are jointly -stable. Using the (72) which indeed is symmetrically -stable. and the interference process is LRD in the generalized sense. Finally. . 2. (79) (78) . we obtain (29). we show that and are symmetrically -stable. APPENDIX E PROOF OF PROPOSITION 4 Proof: The codifference of the interference separated by can be calculated as The above integral exists as long as integral identity [31]. where is as given as in (33).YANG AND PETROPULU: CO-CHANNEL INTERFERENCE MODELING AND ANALYSIS 75 D. APPENDIX C PROOF OF (29)—CONSTANT CASE Here. we have used the fact that for . Equation (28) (74) where calculated as . Hence which takes 1 or 1 with equal probability 1/2. (16)–(21) follow. and is defined as in (60). for .5]. (32) is readily verified. noting that we only need to show that the linear combination is symmetrically -stable for any real number [20. Therefore.1. is positive. Note that for and are positive.

[22] J. Topics in NonGaussian Signal Processing. Kogon and D. he was a visiting researcher at the Laboratoire des Signaux et Systemes. New York: Chapman & Hall. Areas Commun. Signal Processing. Aug. [14] P. J. China. pp. vol. “Performance of optimum and suboptimum receivers in the presence of impulsive noise modeled as an alpha-stable process. 1349–1363. “Locally optimum detection in moving average non-Gaussian noise. and J./Apr. vol. “Signal detection in multivariate Class A interference. REFERENCES [1] E. Tsinghua University. Billingsley. degree in electrical and computer engineering in 1991. 43. A. L. 1942–1946. Feb. Mar.” in Proc. 1992. 1996. Petropulu. H. Nov. We also thank the anonymous reviewers for the careful reading of earlier versions of this paper.-C. Petropulu. Theory.. 1883–1892.. “Analytic alpha-stable noise modeling in a Poisson field of interferers or scatters. 43. pp. 1972. June 1998. Baltimore.-C. pp. Thomas. Eds. ICASSP. [24] P.” IEEE Trans. [31] I.” IEEE Trans.Sc. She is a member of the IEEE Conference Board and the Technical Committee on Signal Processing Theory and Methods..D. Willinger. degree in electrical and computer engineering in 1988 and the Ph. Arnold. Tsihrintzis and C. Pareto Distributions. vol. Wiley Encyclopedia of Telecommunications. 1987. JANUARY 2003 ACKNOWLEDGMENT The authors would like to thank Dr. CNRS-Université Paris Sud. [19] . Jan. “Signal modeling with self-similar -stable processes: The fractional levy stable motion model. Ryzhik. New York: Wiley. 32. Petropulu. [30] G. 71–86. degree in electrical engineering from Drexel University. Blum. Networking. both from Northeastern University. R. [7] J. S. B. Halverson and G. From January 2002 to August 2002. vol. In 1992. Manolakis. She is the co-author of the textbook Higher-Order Spectra Analysis: A Nonlinear Signal Processing Framework. vol. His research interests are in the areas of non-Gaussian signal processing for communications. PA. 1994. no.” in Proc. Commun. [4] A. Workshop Modeling. S./Mar. P. [5] E. vol. Beran.. Petropulu. vol. A. vol. 1989.. 904–914. 106–127.” J. “Estimating long-range dependence in impulsive traffic flows. Blackard. as a Research Staff Member. Parsons. S. Convergence of Probability Measures. Petropulu (SM’01) received the Diploma in electrical engineering from the National Technical University of Athens. NO. and signal processing for data-storage channels. vol. 41–55. vol. vol. pp. Theory. pp. “Discrete-time detection in -mixing noise. Taqqu. [2] D. Rappaport. Hatzinakos. vol. in 2001. “Performance of a spread spectrum packet radio network link in a Poisson field of interferers. 40.” in Trends in Wireless Indoor Networks. pp. M. Pesquet. he joined Seagate Research. L. “The extended alternating fractal renewal process for modeling traffic in high-speed communication networks. 1993. Gradshteyn.” IEEE Trans.. [29] X. 3. 48. Yang and A. [23] X. C. 2000. 1968. degree from the Electronic Engineering Department. New York: Wiley. Kunz et al. Yang. C. “Interference modeling in radio communication networks. M. vol. From September 1999 to April 2000. [26] T. “Power-law shot noise and relationship to long-memory processes. J. he was a Postdoctoral Research Associate with the Electrical Engineering Department.” IEEE Trans. L. [13] H. Wegman. Ecole Supérieure d’Electrícité. In September 2002.” IEEE Trans. Schwartz. Inform. New York: Wiley. [20] G. Signal Processing. S.” Radio Sci. 1994. 46. S. Petropulu received the 1995 Presidential Faculty Fellow Award. vol. K.” IEEE Trans. 1961. CNRS-Université Paris Sud. New York: Academic. Pesquet. Furutsu and T. France. Inform. pp. PA. Inform. AZ. 1995. P.” IEEE Trans. Middleton. Her research interests span the area of statistical signal processing. 41. Sousa. 1988. PA. Boston. Apr. Technol. Signal Processing. 11. New York: Springer-Verlag. Paris. Nikias. Thomas. Wise.” IEEE Trans. [25] D. France. vol. where she is now a Professor. S. in 1998 and the Ph.. “Locally optimum Bayes detection in ergotic Markov noise. July 2000. G. “Statistical–physical models of electromagnetic interference. she was an Associate Professor at LSS. 1983. “On the theory of amplitude distribution of impulsive random noise. July 2000. New York: Chapman & Hall. Drexel University. MD: International. [11] J.” IEEE Trans. [3] K. Commun. Lowen at Harvard Medical School for the interesting discussions on Proposition 4. Aug. S. T. D. 365–373. fractional order statistics and ultrasound imaging. pp. Samorodnitsky and M.. Aug. M. [8] K.” IEEE Trans. Krizman. V. NJ: Prentice-Hall. communications. NJ. Jeffrey. Salt Lake City. 1993). 1994. Giordano and F. vol. CT: JAI. “WAP traffic: Description and comparison to WWW traffic. pp. EMC-19. Applied Phys. 2001. Eds. Blankenship. Signal Processing./Apr. Yang and A. B. Poor and J. and A. Xueshi Yang received the B. Inform. 189–198. Anal. 1997. B. 976–1000. From 1999 to 2000. A. SUPELEC. 44. Taqqu. 1993. Bostian./Mar. 1995. (Englewood Cliffs. she joined the Department of Electrical and Computer Engineering. communication system modeling and analysis. [28] B. Electromagn. New York: Wiley. 51. Sherman. and D. McDonald and R. [12] S. higher order statistics. Poor and J. Athina P. pp. J. Boston. pp. Nov. 11. “A statistical and physical mechanismsbased interference and noise model for array observations. [18] A. Theory. Shao. Yang. Signal Processing With Alpha-Stable Distributions and Applications. “Signal detection in dependent non-Gaussian noise. [27] W. 1995. Rappaport. and X. Theory. M. 47th IEEE Veh. I. “Measurements and models of radio frequency impulsive noise for indoor wireless communications.” in Proc. Select. 49. Princeton University. Greece in 1986 and the M.. 3rd ACM Int. . [6] C. Signal Detection in Non-Gaussian Noise. Maras. VOL. Proakis. 2044–2056. Feb. 1743–1754. “Modeling of atmosphere noise. 7. no.Sc. Ilow and D. Commun. 36.. P. Ishida. Nikias and M. S. pp. Table of Integrals. 1980. W. pp. Thomas. July 1995. New York: Springer.D.” IEEE Trans.. and T. 38. 1977. She has served as an associate editor for the IEEE TRANSACTIONS ON SIGNAL PROCESSING and the IEEE SIGNAL PROCESSING LETTERS.” IEEE Trans. L. Stable Non-Gaussian Random Processes: Stochastic Models With Infinite Variance. Wilson. pp..” IEEE J. P. “Threshold detection in correlated non-Gaussian noise fields. 5. July 2001.” IEEE Trans.” in Advances in Statistical Signal Processing. Haber. 7. 907–912. Philadelphia. Delaney. “Self-similarity through high-variability: Statistical analysis of Ethernet LAN traffic at the source level. Series and Products. Greenwich. [21] A. 1. vol. [17] X. [15] K. Princeton. and J. Phoenix. 1601–1611. May 1997. Signal Processing.” IEEE/ACM Trans. 2002. Eds. D. Dr. Compat. F. and C. Feb. The Mobile Radio Propagation Channel.76 IEEE TRANSACTIONS ON SIGNAL PROCESSING. Athens. A. 1996. pp. Middleton. Statistics for Long-Memory Processes. 48. vol. UT. Sept. Simulation Wireless Mobile Syst. 1994. MA. Pittsburgh. IT-26. Paris. 991–1001. Ed. Kassam. MA. “Measurements and simulation of radio frequency impulsive noise in hospitals and clinics. M. [9] T. 1006–1010. [10] S. [16] D.. Philadelphia. R. V. Conf. pp. V. Beijing. S.