3588

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 7, JULY 2012

Joint Statistics of Radio Frequency Interference in Multiantenna Receivers
Aditya Chopra, Student Member, IEEE, and Brian L. Evans, Fellow, IEEE
Abstract—Many wireless data communications systems, such as LTE, Wi-Fi, and Wimax, have become or are rapidly becoming interference limited due to radio frequency interference (RFI) generated by both human-made and natural sources. Human-made sources of RFI include uncoordinated devices operating in the same frequency band, devices communicating in adjacent frequency bands, and computational platform subsystems radiating clock frequencies and their harmonics. Additive RFI for these wireless systems has predominantly non-Gaussian statistics, and is well modeled by the Middleton Class A distribution for centralized networks, and the symmetric alpha stable distribution for decentralized networks. Our primary contribution is the derivation of joint spatial statistical models of RFI generated from uncoordinated interfering sources randomly distributed around a multiantenna receiver. The derivation is based on statistical-physical interference generation and propagation mechanisms. Prior results on joint statistics of multiantenna interference model either spatially independent or spatially isotropic interference, and do not provide a statistical-physical derivation for certain network environments. Our proposed joint spatial statistical model captures a continuum between spatially independent and spatially isotropic statistics, and hence includes many previous results as special cases. Practical applications include co-channel interference modeling for various wireless network environments, including wireless ad hoc, cellular, local area, and femtocell networks. Index Terms—Co-channel interference, impulsive noise, poisson point processes, spatial dependence, tail probability.

W

I. INTRODUCTION

IRELESS transceivers suffer degradation in communication performance due to radio frequency interference (RFI) generated by both human-made and natural sources. Human-made sources of RFI include uncoordinated wireless devices operating in the same frequency band (co-channel interference), devices communicating in adjacent frequency bands (adjacent channel interference), and computational platform subsystems radiating clock frequencies and their harmonics [1]. Dense spatial reuse of the available radio spectrum, required to meet increasing demand in user data rates also causes severe co-channel interference and may limit communication system performance. Network planning, resource allocation [2] and

Manuscript received April 22, 2011; revised September 01, 2011 and January 06, 2012; accepted March 09, 2012. Date of publication April 03, 2012; date of current version June 12, 2012. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Benoit Champagne. This work was supported by Intel Corporation. The authors are with the Wireless Networking and Communications Group, Department of Electrical and Computer Engineering, The University of Texas at Austin, TX 78712 USA (e-mail: achopra@utexas.edu; bevans@ece.utexas. edu). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TSP.2012.2192431

user scheduling [3] are strategies typically used to avoid RFI; however, these strategies are restricted to coordinated users operating over the same or coexisting communication standards [4]. Other RFI mitigation techniques include interference cancellation [5], interference alignment [6], and receiver back-off [7]; however, these too require some level of user or base station cooperation [8]. Furthermore, some level of uncoordinated residual interference is generally present in wireless receivers regardless of the interference cancellation strategy. Residual interference includes interference from uncoordinated users (out-of-cell interference in cellular networks and co-channel interference in ad hoc networks) or users in coexisting networks (interference from hotspots in femtocell networks). Recent communication standards and research have focused on the use of multiple transmit and multiple receive antennas to increase data rate and communication reliability in wireless networks. Multiantenna wireless receivers are increasingly being used in RFI-rich network environments. Accurate statistical models of RFI observed by multiantenna receivers are needed in order to analyze communication performance of receivers [9], develop algorithms to mitigate the impact of RFI on communication performance [10], and improve network performance [11]. In this paper, we derive the joint statistics of RFI generated from a field of Poisson distributed interferers that are placed: (i) over the entire plane; or (ii) outside of a interferer-free guard zone around the receiver. Organization: Section II presents a concise survey of multidimensional statistical models of RFI. Section III discusses the system model of interference generation and propagation. The joint interference statistics are derived in Section IV. Section V discusses the impact of removing certain system model assumptions on RFI statistics. Section VI presents numerical simulations to corroborate our claims and the key takeaways are summarized in Section VII. Notation: In this paper, scalar random variables are represented using upper-case notation or Greek alphabet, random vectors are denoted using the boldface lower-case notation, and random matrices are denoted using the boldface upper-case notation. Deterministic parameters are represented using Greek alphabet, with the exception of denoting the number of redenotes the expectation of the funcceive antennas. tion with respect to the random variable denotes the probability of a random event, and denotes the vector 2-norm. II. PRIOR WORK In typical communication receiver design, interference is usually modeled as a random variable with Gaussian density distribution [12]. While the Gaussian distribution is a good model for thermal noise at the receiver [13], RFI has predominantly non-

1053-587X/$31.00 © 2012 IEEE

with the signal propagation model incorporating pathloss. but does not support spatial correlation . The Middleton Class A and the Gaussian mixture distributions can also incorporate thermal noise present at the receiver without changing the nature of the distribution.. Statistical-physical models can therefore be more useful than empirical models in designing robust receivers in the presence of RFI [14]. The spherically isotropic alpha stable model is derived under both homogeneous and nonhomogeneous distribution of interferers. The symmetric alpha stable and Middleton Class A distribution functions are listed in Table I. without regard to the physical generation mechanisms behind the interference. i. Prior Work in Single-Antenna RFI Models In [17]. it was shown that interference from a homogeneous Poisson field of interferers distributed over the entire plane can be modeled using the symmetric alpha stable distribution [15]. In [14].e. MCA. The following subsections discuss key prior results in statistical modeling of RFI in singleand multiantenna receivers. whose distribution functions are listed in Table II. The two common approaches of generating multivariate extensions of uni-variate distributions assume that either: (a) RFI is independent across the receive antennas. In [18]. In [22]. B. The Middleton Class A models are also canonical. by using the Gaussian mixture distributions to model RFI statistics in network environments with clustered interferers. Prior Work in Multiantenna RFI Models Prior work on statistical modeling of RFI in multiantenna wireless systems has typically focused on using multivariate extensions of single-antenna RFI statistical models. Their results were generalized in [19].III—This model incorporates spatial dependence in multiantenna RFI. The statistical techniques used in modeling interference can be divided into two categories: (1) statistical inference methods and (2) statistical-physical derivation methods. it cannot represent independent Class-A random variables. RFI is spatially and temporally independent and identically distributed. and receiver separation is ignored. The authors assume that each receiver is surrounded by the same set of active interferers. 3) MCA.CHOPRA AND EVANS: RFI IN MULTIANTENNA RECEIVERS 3589 TABLE I KEY STATISTICAL MODELS OF RFI OBSERVED BY SINGLE-ANTENNA RECEIVER SYSTEMS Gaussian statistics [14] and is well modeled using impulsive distributions such as symmetric alpha stable [15] and Middleton Class A distributions [16].II can represent correlated or uncorrelated random variables. This result was later extended to include channel randomness in [18]. on the other hand. no knowledge of the physical environment is needed to estimate the model parameters [20]. The three multidimensional extensions. 2) MCA. lognormal shadowing and Rayleigh fading. Statistical inference approaches fit a mathematical model to interference signal measurements. The study is limited to baseband signaling and neglects any correlation in the interferer to receive antenna channel model or correlation in the interference signal generation model.I cannot capture spatial dependence or sample correlation of RFI across receive antennas. such as the Middleton Class A and symmetric alpha stable distributions. Statistical-physical models. the authors demonstrate the applicability of the spherically isotropic symmetric alpha stable model [21] to RFI generated from a Poisson distributed field of interferers. it was shown that the Middleton Class A distribution well models the statistics of sum interference from a Poisson field of interferers distributed within a circular annular region around the receiver. the author proposes three possible extensions to the univariate Class A model.I—Each receive antenna experiences additive uni-variate Class A noise with the identical parameters. A. MCA. however. unlike the symmetric alpha stable distribution. model interference based on the physical principles that govern the generation and propagation of interference-causing emissions. are as follows: 1) MCA.II—RFI is assumed to be spatially dependent and correlated across receive antennas. The impulsive nature of RFI may cause significant degradation in communication performance of wireless receivers designed under the assumption of additive Gaussian noise [10]. or (b) the multivariate RFI model is isotropic.

3590 IEEE TRANSACTIONS ON SIGNAL PROCESSING. This is contrary to their statistical-physical generation mechanism which assumes that each receive antenna has a subset of surrounding interferers that are observed exclusively by that antenna alone. Table II lists the different . These models are multivariate extensions of the Class A distribution and are not derived from physical mechanisms that govern RFI generation. the authors attempt to derive RFI statistics for a two antenna receiver based on the statistical-physical mechanisms that produce the RFI. MCA. However. 60. etc. incorporates an additive Gaussian background noise component and has found application in receiver design in the presence of interference [24]. 7. While these models have been very useful in analyzing MIMO receiver performance [23] and designing receiver algorithms [24]. NO. In [25]. fading parameters. [26]. a statistical-physical basis for these models would further enhance their appeal [25] in linking wireless network performance with environmental factors such as interferer density. VOL.III can represent uncorrelated and spatially dependent Class A random variables but it cannot represent independent or correlated random variables. and it enforces statistical dependence among receive antennas similar to MCA. Our article extends their work by using a similar interference generation and propagation framework to develop statistical-physical interference distributions for any number of receive antenna. JULY 2012 TABLE II KEY STATISTICAL MODELS OF RFI OBSERVED BY MULTIANTENNA RECEIVER SYSTEMS across antenna samples. it is incomplete as it is strictly limited to two antenna receivers [25]. The authors use a physical generation model for the received RFI at each of the two antennas that is the sum of RFI from stochastically placed interferers which include interferers observed by both antennas as well as interferers observed exclusively by a single antenna.III. Their resulting model is canonical in form.

such as one described in our system model. located in the presence of nearby interfering sources within a two-dimensional plane as shown in Fig. Sectorized antennas are geographically co-located wireless directional antennas with radiation patterns shaped as partially overlapping sectors that combine to cover the entire space around the multiantenna system [28]. our system model supports the following scenarios of interferer placement: 1) Interferer placement with guard-zones: By allowing . each receiver is under the influence of an independent set of interferers. are empty • Case II—Interferer sets and .e..CHOPRA AND EVANS: RFI IN MULTIANTENNA RECEIVERS 3591 Fig. The geometry of the individual antenna elements and the interantenna spacing is ignored. III.. It is trivial to see that the resulting RFI would also exhibit independence across the receive antennas. SYSTEM MODEL anConsider a wireless communication receiver with tennas. Illustration of interferer placement around a 3-antenna receiver at a sampling time instant.e. The Poisson point process is typically used in modeling interferer locations in large wireless networks and provides analytical tractability in mathematical derivations [30]. the origin is shifted to coincide with the receiver location. The distance of each interferer from origin (where the receiver is located) provides an ordering function. where the antenna gain for all antennas is similar. The level of correlation can be tuned by changing the intensity of the Poisson point process representing each set in the interferer field. In this scenario. the active interfering sources are clasindependent sets . The interferers are co-planar to the receiver and are transmitting in the same frequency band as the receiver.I in Table II. For ease of notation.II and MCA. Such a model can be applied to cellular and ad hoc networks with contentionbased or scheduling-based MAC protocols [3]. Interference signals common to some antennas may arise in the overlapping sections. i.III from [22]. the entire receiver is assumed to be located at a single point in space. Sectorized antennas can provide the advantages of spatial diversity while mitigating the impairments caused by multipath delay [29].e. • Case I—Interferer set is empty. effectively modeling a wireless network with an interfererfree guard zone around the receiver.. the same set of interferers. i. The common set of interferers in models correlation between the interferer fields of two antennas. or in other words.e.II. is during deployment of receivers employing sectorized antennas with full frequency reuse. . is implicitly assumed in order to define the th interferer at ..I. 1. enabling us to count the interferers in denotes the each set. i.e. denotes the set of interferers that are observed by antenna alone.. cause interference to every receive antenna. we constrain the active interferers to be located outside of a finite disk around the multiantenna wireless receiver. joint statistical models of RFI that have been proposed in prior work. and . A scenario in which an environment with interferers observed by a subset of receive antennae may arise. or MCA. We also assume a fast fading wireless channel and power-law pathloss between the interferers and the receiver.III appropriately capture the joint spatial statistics of RFI generated in such an environment. Thus. MCA.. This model is quite commonly used in multidimensional temporal [27] and spatial interference modeling [25]. shown in Table II. The resulting RFI is spatially dependent across receive antennas and neither of MCA. we can apply the term co-channel interference to describe the resulting RFI. For the th interfering source in two-dimensional coordinates of the interferer’s location. • Case III—All interferer sets are nonempty. At each sampling time instant. This models partial correlation in the interferer field observed by each of the receive antennas. and individual antennas may still see independent interference in sectors where one antenna exhibits high gain while others exhibit a null. Based on the intensity of individual interferer fields. have the same characteristic in spatial statistics but there is no statistical physical derivation linking these models to any interference generation mechanism. denotes sified into the set of interferers that are observed by all antennas. hence. We wish to highlight the fact that a spatial Poisson point process distributo have potention of interferers allows each interferer set tially infinite number of interferers. our system model yields the following three RFI generation scenarios: and . Both receivers observe sets. The joint spatial statistics of the resultant RFI from this model are derived in Section IV. Case I results in RFI that has characteristics of model MCA. i.A. MCA. i. i.e. thereby causing the resulting RFI to have spatial dependence across receive antennas. 1. two-dimensional annulus with inner and outer radii respectively. Based on values of . At each time snapshot. We also assume that the interferers are distributed within a and . we assume that are the locations of the active interferers in distributed according to a homogeneous spatial point process denoted by with the intensity of set intensity of .

across and . as well as platform noise [10] in small form-factor devices such as laptops and mobile phones. combined with the three cases of correlation within interferer distribution.d. NO.i. we study the impact of removing many of these assumptions on the statistics of RFI. JOINT CHARACTERISTIC FUNCTION MULTIANTENNA RFI OF propagation through the wireless medium. Such a scenario can model interference in an ad hoc network without any contention-based medium access control protocol [3]. SPATIALLY INDEPENDENT INTERFERENCE IS TRIVIALLY DERIVED FROM PRIOR WORK ON SINGLE-DIMENSIONAL STATISTICAL MODELS OF INTERFERENCE.3592 IEEE TRANSACTIONS ON SIGNAL PROCESSING. . We can express the interference signal at the th antenna as (1) where represent the sum interference signal from interfering sources in (visible only to antenna ). 60. and is the . and (3) we can write the resultant interference signal at the th receive antenna as At each receive antenna . denotes the complex fast-fading channel between the interferer and receiver . is independent of the antenna under observation . denotes the distance between the receiver (located at origin) and the interferer. VOL. and that the amplitude channel phase is uniformly distributed on . The majority of Section IV is dedicated to deriving interference statistics for isotropic interference (Case II) in network scenarios not studied in prior work. In Sections V-A to -D. the baseband sum interference signal at any sampling time instant can be expressed as the sum total of the interference signal observed from common interferers and the interference signal from interferers visible only to antenna . In Section V-C. across all interfering sources . yields a total of twelve possible scenarios of interferer distribution. The signal from the common set of interferers is expressed as (3) Note that the difference between (2) and (3) is that the interferer and the distance between interferer emission signal and the receive antenna (placed at origin). JULY 2012 TABLE III PRIOR RESULTS ON STATISTICAL-PHYSICAL INTERFERENCE DISTRIBUTION MODELS IN MULTIANTENNA RECEIVERS. The spatially independent interference statistics (Case I) are a trivial extension to the single antenna interference statistics. (SYMMETRIC ALPHA STABLE: SAS) 2) Interferer placement without guard-zones: By setting . IV. 7. we will derive the joint characteristic function of the in-phase and quadrature-phase components of . consequently. and The possible parameter values .d.B. where (5) (6) In order to study the spatial statistics of interference across multiple receive antennas. we allow interferers to be located arbitrarily close to the wireless receiver. we assume that the channel follows the Rayleigh distribution. respectively. denotes complex baseband emissions from interfering source is the emission signal envelope and is the phase where of the emission. denoted by . Combining (1). (2). Interference statistics for Case III require knowledge of interference statistics for Cases I and II. and interfering sources in (visible to all antennas). Table III lists the scenarios where prior statistical models of interference exist in literature. The channel between the interferer and th receiver. power pathloss coefficient indicates the reduction in interfering signal amplitude during (4) The complex baseband interference at each receive antenna can then be decomposed into its in-phase and quadrature components . and are assumed to be i. The sum from interferers in can be written as interference signal [16] (2) is the sum of interfering signals emitted by each interferer located at . is also assumed i. This is because we ignore interantenna spacing and assume all antennas are located at the origin. The joint spatial statistics of the resultant RFI from this model are derived in Section IV. For the fast fading channel model. we will discuss the impact of spatial correlation of the channel model on interference statistics.i.

(12) (13) where and .i. Substituting (4) in (7) and separating independent terms in the expectation. We can rewrite (10) and (11) in their polar forms as A. since we assumed that the interferers are ordered according to how close they are located to the origin.CHOPRA AND EVANS: RFI IN MULTIANTENNA RECEIVERS 3593 interference. and . This allows us to remove the interferer index and treat the contribution to RFI from each interferer as an independent random variable. We will now evaluate each component term in (9) for interferer environments with and without guard zones. Conditioned on . denoted by the term in (7). we get (8). by virtue of a property of Poisson point processes. Consequently. we evaluate the contribution of the interferers in set to the joint spatial statistics of RFI. and are all i. the joint characteristic function can be written as (7) . shown at the bottom of the page. To simplify notation. Each component term in (9) is the characteristic function of the interference contribution by one of each interferer sets . as described in Section III. and their emissions and channel realizations are also independent as well. However. where Note that the expectation is evaluated over the random variables . Thus. For the sake of brevity and readability. (8) is decomposed into the product form (9) where [see (10)–(11) at the bottom of the page]. Using (1). and . the points in a Poisson point process are distributed independently and uniformly across the region in consideration. Joint Spatial Statistics of RFI in Presence of Guard Zones (RFI Contribution From ): In 1) Evaluation of this section. (5) and (6). we refrain from listing all of these random variables in the subscript of the expectation operator.d. We can separate the independent terms in the expectation by noting that the interferers in the different sets are distributed in space as independent homogeneous processes. (8) (10) (11) . the number of interferers is a Poisson random variable with parameter . (12) can be expressed as (14) (15) (16) Once conditioned on a fixed number of total points. respectively. respectively. the points are uniformly distributed within the region of the point process when conditioned on the total number of points . the interferers within are distributed according to a homogeneous spatial Poisson point process inside an annulus with finite inner and outer radii and . In the constrained interferer placement system model. and can be replaced in (16) by . is assumed to be increasing with the index .

for any value of and modulo is also uniformly distributed over . all terms in (25) with reduce to zero after taking expectation with respect to .3594 IEEE TRANSACTIONS ON SIGNAL PROCESSING. Using the series expansion of the zeroth-order Bessel function (28) we can rewrite (27) as shown in (29)–(30) at the bottom of the next page. 7. NO. the th moment of is . JULY 2012 in (16) we replaced lows the distribution by the random variable . the log-characteristic function is (21) (26) Note that the expectation is now with respect to the remaining random variables in (26). shown at in the bottom of the page. and (25) reduces to This distribution arises when we consider the annular disc with inner and outer radius of and . Next we use the identity (24) for . Combining this notion with (16) we get (17)–(20). the logcan be expressed as characteristic function (25) (35) (17) (18) (19) (20) . that fol- Since and are uniform random variables within . To evaluate the expectation in (26). we have (22) (23) where . and points distributed uniwithin the anformly in this region. Expanding the expectation in (34) with respect to the random variable . Under the assumption that the fading channel is Rayleigh distributed. The number of interferers nular region is a Poisson random variable with mean . at the bottom of the next page. Thus. we get (31)–(34). By taking the logarithm of (20). 60. Applying to (30). and denotes the where Bessel function of order . Combining (24) and (23). we rewrite it as (27) Note that we used the assumption that the fast-fading channel between interferer and receive antennas are spatially independent across the antenna index . VOL.

(35) can be written as (36)–(37). in which case and we are left with . Similar to an approach used in [19]. The multiplicative term prevents us from simplifying (38) into an exponential. we are able to determine the appropriate values for for any with MSE .CHOPRA AND EVANS: RFI IN MULTIANTENNA RECEIVERS 3595 Using Taylor series expansion of . has conAssuming that the interferer emission amplitude stant value . In the case where cannot less than be ignored. (29) (30) (31) (32) (33) (34) (36) (37) . we get . valid for . we approxthe term as an power series for and choose paimate and to minimize the mean squared error (MSE) rameters (39) (38) In Section V-A. shown at the bottom of the page. Note that in reasonable to assume that for many wireless network scenarios Using the nonlinear unconstrained optimization functionality provided by MATLAB. we discuss the impact of applying any general interferer emission amplitude distribution on the interference statistics. we can again use the power series approximation to find parameters such that .

The next section incorporates receiver thermal noise which gives rise to the parameter seen in MCA. At each antenna . JULY 2012 Using the aforementioned approximation. MCA.II or more generally. Such events cause a high degree of impulsiveness in interference subsequently leading to heavy-tails in first-order interference statistics. we derive joint spatial statistics of sum interference with the parameter . we set in (35) to get (48) Combining and . is very high. 1) Evaluation of (RFI Contribution From ): The contribution to the joint spatial statistics of RFI from the interferers in is expressed as the term in the joint characteristic function of RFI given by (9).III. and (44).III with a containing equal values in the diagonal. we get into a temporary variable (46) where is the Dirac-delta function. resulting in zero RFI. Since is also proportional to . an active interferer may arise quite close to the receiver causing a sudden large burst of interference. Thus. B. It indicates the probability that there are no interferers in the annular region around (49) . we derive contribution of interferer sets to the spatial joint statistics of RFI. we can incorporate it into our model resulting in the following distribution: (47) . asymptotically increasing the number of interferers changes the nature of the distribution as well. Thus. where It is interesting to note that the parameter is linearly dependent on the corresponding interferer density . (43).3596 IEEE TRANSACTIONS ON SIGNAL PROCESSING. Combining (9). the Middleton Class A distribution is less impulsive in nature and resembles the Gaussian distribution. the number of interferers If becomes very large.II and MCA. NO. 60. however. In order to derive closed form expressions for (10). VOL. Assuming that antenna observes independent thermal noise with variance . we can see that the interference statistics become less impulsive as the radius of the guard zone increases. Joint Spatial Statistics of RFI in Absence of Guard-Zones In this section. we can write (10) as (43) which is the characteristic function of the isotropic multivariate Middleton Class A distribution shown in Table II as MCA. In practical receivers. grows large as well and it is well known that for large values of . or in other words. 7. the log-characteristic exponent in (38) can be expressed as (40) (41) (42) Equation (40) is the log-characteristic function of a Middleton Class A where is the overlap index indicating the impulsiveness of the interference. thermal background noise is always present and is well modeled by the Gaussian distribution. Thus we can write (11) as (44) where and are the parameters of the Class A distribution. 2) Evaluation of (RFI Contribution From ): In this section. and is the mean intensity of the interference [14]. the interference from the exclusive set of interferers is identical to RFI seen by a single antenna receiver surrounded by interferers distributed according to a Poisson point process. In the absence of guard zones. we get the joint characteristic function of RFI as (45) The corresponding probability density function can be written as the receiver. The statistics of such RFI have been derived in [19] and shown to be well modeled by the univariate Middleton Class A distribution.

we assumed a Rayleigh distributed fast fading channel between interfering sources and the multiantenna receiver. Combining (9). If is very high. If is finite. the characteristic function form is of the (51) where . the Rayleigh distribution of the fading channel between interferers and the receiver was used to simplify the integral of a Bessel function in (30). Since we are interested in the RFI amplitude tail statistics. In this scenario. our model degenerates into spatially independent RFI. we arrive at the characteristic function of the Y as (58) Equation (58) is the characteristic function of random variable that is a mixture of independent and spherically isotropic symmetric alpha stable vectors. While the Rayleigh distribution is a reasonably accurate and frequently used model of fading channel amplitude. the resulting interference is still distributed as a symmetric alpha stable random variable. The characteristic function expression in (49) offers little insight when is finite. Interference Statistics in General Fading Channel Models In developing our system model in Section III. With constrained interferer locations. respectively. we get with our knowledge of the distri- (53) A. but converges rapidly as increases. (55) can be written as (50) consequently. By setting to 0. For unconstrained interferer location distribution. or in other words. however. The evaluation of for interferers in is identical for all . albeit with a very high intensity. and This is the characteristic function of an isotropic symmetric alpha stable random vector. IMPACT OF SYSTEM MODEL ASSUMPTIONS ON INTERFERENCE STATISTICS (52) Taking expectation over bution of . it has been shown in [18] that a series summation of the form (4) with finite terms converges very rapidly to the alpha stable distribution. This is an interesting property in contrast to interference with guard zones. Starting with (13) and Using a derivation similar to the steps from (16) to (25). Removing the Rayleigh distribution assumption would prevent this step in the proof. the integral is reduced to When .CHOPRA AND EVANS: RFI IN MULTIANTENNA RECEIVERS 3597 When . 2) Evaluation of (RFI Contribution From ): can be considered as the contribution to the joint characteristic function of RFI. where interference statistics under a large value of are less impulsive in nature and start resembling the Gaussian distribution. and (56) (57) where . and (51). the number of interferers becomes very large. while setting to 0 for all causes isotropic RFI at the receiver. made by interfering sources visible only to one receive antenna. (57). the characteristic function expression is not very useful. we get that Equation (57) is the characteristic function of a symmetric alpha stable random variable. we do not require any assumptions on the channel amplitude distribution to derive RFI statistics. The dispersion parameters and depend linearly on the intensities and . and from Fourier (54) Thus we have evaluated (13) as (55) . V. the parameter is linearly dependent on the corresponding interferer density . other distributions have also been widely used to characterize wireless channel amplitudes [31].

we used (61) when stepping from neighborhood of (63) to (64). assuming that the interferer emissions are uniformly distributed in phase . we assume that the wireless fading channel between the interfering source and receiver is spatially independent and identically distributed across the multiple antennas in the receiver. we assumed that the emission amplitude was constant in (38) in order to simplify the interference statistics into the form of a Middleton Class A distribution. Interference Statistics in Random Interferer Emissions In deriving interferer statistics in the presence of guard zones. yielding joint statistics of the form given in (46). therefore is independent of . even if the interferer emission envelopes are randomly distributed. Interference Statistics in Spatially Correlated Wireless Channel Models To derive the joint statistics of RFI in Sections IV–A and -B. by using the approximation (66) and (68) to get . However. We use the approximation [16] (59) where is a correction factor expressed as . If there is phase correlation between receive antennas and . (66) As discussed in Section V-A. [33]. 60. the rest of the joint RFI statistics derivation continues from (34) as shown in Section IV-A. we can combine (69) (62) (63) (64) (65) Since we are interested in evaluating (26) primarily in the . (40) would be replaced by (70) In Section IV-B we used the assumption that the fast-fading channel phase is uncorrelated across receive antennas. in order to accurately model tail probabilities. We start from (25). we used the assumption that the fading channel is i. Note that when we derive RFI statistics in the absence of guard zones we make no assumptions regarding the interferer emission amplitude. In this region. we can use the approximation arrive at . In this paper. VOL. to show are equal to zero.d. we only study the impact of channel correlation on the joint interference statistics in the presence of guard zones. across the receive antennas. Without making this assumption. then the terms containing are not equal to zero for . [19] that as . 7. the slowest decaying term in is of the order . It has been shown [14]. and we can write (59) as (61) Using (61). Since (65) has the same form as (34). we can simplify (26) as which has the same form as (40). which shows that the joint log-characteristic function of RFI can be expressed as Again. the behavior of the characteristic function in the neighborhood of governs the tail probability of the random envelope [14].3598 IEEE TRANSACTIONS ON SIGNAL PROCESSING. especially in the tail probability region.i. In stepping from (64) to (65). B. C. JULY 2012 analysis. to (67) (60) (68) is the confluent hypergeometric function of where the first kind. This assumption may not be true if two antennas are close to each other. resulting in the that all terms with simplified product of terms expression in (26). Thus. Spatial correlations in the wireless channel are routinely modeled when studying the performance of multiantenna receivers [32]. NO. we are concerned with the region around . the interference statistics can be approximated using the Middleton Class A form.

for example the coefficient of the is indicative of correlation between and . many indoor wireless channel models account for variable pathloss exponents [34]. we assumed that the pathloss exponent is fixed across emissions from all possible interferers. the elements of matrix. contains nonzero terms with variables . We again use the approximation result provided in (59) to get (74)–(75). These terms indicates sample level correlation in term the sum RFI. and . Note that in addition to the terms. Interference Statistics in Randomly Distributed Pathloss Exponents In our system model. and considering only these terms (70) reduces to The probability distribution corresponding to the characteristic function in (45) can be written as (77) where the matrix is a . Assuming that the pathloss exponent is independent of the interferers. we can write (4) as (72) Equation (71) can be expressed as shown in (73) at the bottom of the page. this coefficient of the term is equal to . For all integers are given as (78) (79) (80) (81) (71) By applying the following identity: (82) (83) D. Analogous to the cross terms in multidimensional Gaussian distributions. the Poisson field of interferers gets randomly thinned (73) (74) (75) . However.CHOPRA AND EVANS: RFI IN MULTIANTENNA RECEIVERS 3599 only the terms with and are nonzero. If we consider the pathloss exponent to be a random variable with support and a discrete probability density . shown at the bottom of the page. Thus we can now update our model to include correlation (76) (84) where each term in the outer summation is the sum interference from a Poisson field of interferers with pathloss exponent of .

In the case where a guard zone exists around the receiver. resulting in another Poisson field with intensity . the clustered Poisson point process is a good interferer location distribution model for heterogeneous and two-tier networks [19]. Thus. i. we have shown that interference statistics exhibit spatial isotropy. NO. antenna spacing may be ignored if the radius of the guard zone is much larger than the interantenna distances at the receiver. 2 shows a scatter plot of the amplitude of simulated random RFI observed at two antennas over a span of 40 time samples using different values of and . impulsive events at the two receive antennas occur with different strengths but at the same location in time. i. we numerically simulate the random variable describes the observed RFI across a given in (5) and (6). the statistics of resulting interference are unchanged [19]. it is often used to compare probability density distributions and can be computed efficiently. Low KL divergence between two density distributions implies high similarity between the density functions. Poisson Point Process Distribution of Interferer Locations: The Poisson point process distribution of interferer locations has been shown to well model cellular networks. giving rise to a possible different sets of interferers for a antenna receiver. We choose such that . In the case when . we assumed interferers that were either observed by all receive antennas or by only a single receive antenna. it has been shown that if an antenna is placed a small distance away from the center of the guard zone. in order to normalize our data as the variance of interference observed at each antenna (proportional to ) remains the same regardless of the value taken by or . each antenna would observe interference from different interferers causing interference to be independent across antennas. which move from spatially isotropic to spatially independent. i. E. In the scenario where from there are no guard zones around the receiver. To validate our amplitude distribution model. SIMULATION RESULTS To study the accuracy of our joint amplitude distribution as model. It intuitively follows that increasing interantenna spacing reduces the spatial dependence in the resulting interference statistics. simulated multiantenna receiver in a Poisson field of interferers. for example. the density of total interferers observed by each antenna is the same. the total interference is the sum of interference independent fields of interferers. indicating that impulsive events (samples with a large amplitude) occur in a spatially dependent manner. 7. VI. antennas and the interferers excluOur receiver uses sive to each antenna are distributed with equal densities. which exhibits the Gaussian mixture distribution. the resulting interference is the sum of independent multidimensional Middleton Class A random variables.e. and (2) the tail probabilities of the numerically simulated and our proposed amplitude distribution. ad-hoc networks and uncoordinated radiative sources. It TABLE IV PARAMETER VALUES USED IN SIMULATIONS may therefore be possible to apply our proposed continuum between isotropy and independence to approximate the statistics of interference in receivers with significantly separate antennas. indicating its dependence on . which can be written as (85) (86) is the same as derived in (58) with indexing In (85). Modeling joint statistics of RFI between separate antennas is beyond the scope of this paper and an avenue of future work. The horizontal axis denotes the RFI amplitude at antenna 1 and the vertical axis deincreases. This is simply because the distance between any interferer and an antenna is larger than the guard zone radius.. Other point processes can also be used to model a variety of applications.. we can employ a single random variable to denote the distance from a interferer in the set to each receive antennas. In the presence of guard zones. 60. In single antenna receivers.e. Interferers Visible to a Subset of Receive Antennae: In this paper. . the RFI notes the RFI amplitude at antenna 2. JULY 2012 by probability . the characteristic function of interference would represent the sum of many alpha stable random variables. Neglecting Interantenna Spacing at the Receiver: Using this assumption. If the interantenna spacing is asymptotically large. As samples move towards the top left of the scatter plot area. Fig. Although KL divergence is not strictly a distance metric. This assumption can be extended to include interferers that are visible by only a subset of receiver antennas. Fig. and there is no interantenna spacing. as each set of interferers would result in RFI statistics that are statistically isotropic across the subset of receiver antennas observing the interferer set. we use two metrics: (1) the Kullback-Liebler (KL) divergence between numerically simulated interference amplitude distribution and our proposed amplitude distribution. Discussions on Remaining System Model Assumptions The key assumptions that remain in our system model are as follows. 3 shows the KL divergence between the numerically simulated distribution of interference in the presence of . Table IV lists the values for the rest of the simulation parameters. consequently antenna geometry will have minimal impact on the pathloss attenuation term in (1). VOL. The characteristic function of interference from set can be expressed by modifying (43) to write it as (87) (88) It may require some extra mathematical rigor to prove the same results when the pathloss exponent is a continuous random variable. Our statistical derivations are also amenable to incorporating such scenarios..3600 IEEE TRANSACTIONS ON SIGNAL PROCESSING.e.

and the parameters values used in simulation. Using (78). The tail probability of the multidimensional Middleton Class A distribution can be evaluated as a mixture of Gaussian tail probabilities.CHOPRA AND EVANS: RFI IN MULTIANTENNA RECEIVERS 3601 Fig. our proposed distribution. . Otherwise. Given a threshold . indicating that the approximation used in (59) is close to its true value. Fig. 6 shows that the KL divergence with Rayleigh fading channel model is the lowest. a Rician model is commonly used. the two receive antennas has a predicted value of . Fig. We simulate correlation between the in-phase components of the channel between an interferer and the two receive antennas. and it is very large owing to the fact that the Gaussian distribution cannot accurately model impulsiveness in simulated RFI. The Pearson product moment correlation coefficient (PMCC) [35] between these two random variables is chosen as 0. while the Gaussian distribution is clearly unable to capture the large tail probabilities of impulsive interference. We also show the KL divergence between the numerically simulated distribution and the isotropic and independent-only multivariate distributions given in Table II as MCA. respectively.III and MCA. Next. the tail probability of interference is related to the outage performance of receivers.I. Finally. we define tail probability as . KL divergence is also calculated between the simulated RFI distribution and models MCA. The tail probabilities of our proposed distributions match closely to simulated interference. We use three common stochastic fading channel models [31]. The KL divergence increases slightly upon changing the channel model. 3. indicating that our proposed distribution is able to well capture partial spatial dependence in interference. 4 and 5 shows a comparison of the tail probabilities of the numerically simulated distribution. The tail probability is the complementary cumulative distribution function of a random variable and in performance analysis of communication systems. where a lower KL divergence means posed model versus denotes the KL divergence between distributions and a better fit. we compare the tail probabilities of the numerically simulated distribution and our proposed distribution models. and the Gaussian distribution for interference with and without guard zones. 2.III (Isotropic Class A) from Table II. When the dominant propagation path is line-of-sight. the PMCC between the in-phase components of .3. We also employ KL divergence to study the impact of different channel models on the resulting spatial distribution of RFI. and the approximate tail probability of the symmetric alpha stable distribution is given in [15]. respectively. which is to be expected.I (Independent Class A) and MCA. The KL divergence between the numerically simulated RFI amplitude distribution and our proposed models is lowest across all values of . we simulate correlated fading channels to test whether our model correctly accounts for spatial correlation. The system parameter values are chosen from Table IV. The KL divergence between simulated RFI and multivariate Gaussian distribution of equal variance is also shown in Fig. 3. Figs. Estimated KL divergence between simulated RFI distribution and pro. a Rayleigh or Nakagami model is commonly used. Table V defines the three distributions. Scatter plot of RFI amplitude observed at 2 receive antennas. guard zones and our proposed multivariate Class A model.

Depending on the region within which interferers are distributed. CONCLUSION In this paper. Tail probability versus threshold for interference in the absence of . where a lower KL divergence means a better fit. Tail probability versus threshold for interference in the presence of guard zones. The tail probabilities are generated for isotropic and a mixture of isotropic and indepeninterference . The density function corresponding to each . 7. TABLE V COMMONLY USED FAST FADING CHANNEL MODELS [31] USED IN FIG. Some of these distributions find use in designing interference mitigation algorithms or analyzing communication performance of receivers in the presence of interference. The tail probability is compared between the numerically guard zones simulated interference (“Sim”) and the tail of the multivariate symmetric alpha stable distribution (“Expr”). JULY 2012 Fig. and other parameter values are Using our proposed framework. and physical mechanisms describing the generation and propagation of interference through the wireless medium. 60. 7. Our framework incorporates random distribution of interferer locations in two-dimensional space around the receiver with an optional interferer-free guard zone. we derive the joint statistics of interference observed across a multiantenna receiver.3602 IEEE TRANSACTIONS ON SIGNAL PROCESSING. 6 Fig. Fig. Remaining parameter values are dent interference given in Table IV. the interference statistics can be modeled using the Middleton Class A or the symmetric alpha stable distribution. NO. VOL. The tail probabilities are generated for isotropic and a mixture of isotropic and indepeninterference . Estimated KL divergence between simulated RFI distribution and proposed model versus . 5. Fig. 4. VII. with the resulting amplitude distribution modeling both spatially isotropic and spatially independent observations of RFI as special cases. and remaining paramchannel model is provided in Table V. denotes KL divergence between distrieters are given in Table IV. butions and . 7 shows that the empirically estimated value of the PMCC matches our prediction quite accurately. such as fast fading and pathloss attenuation. Fig. The tail probability is compared between the numerically simulated interference (“Sim”) and the tail of our proposed multivariate Middleton Class A distribution (“Expr”). Our framework also incorporates partial statistical dependence of RFI across the receive antennas and captures a continuum between spatially independent and spatially isotropic interference. Estimated PMCC versus given in Table IV. Remaining parameter values are dent interference given in Table IV. using different fast fading channel models. 6. . we propose a statistical-physical framework for modeling RFI observed by a multiantenna receiver surrounded by interference causing emitters. By .

[23] P.” U. F. [13] A. pp. “Statistical-Physical Models of Man-Made and Natural Radio Noise Part II: First Order Probability Models of the Envelope and Phase. 1–5. and K. no. [4] J. IEEE Workshop on Signal Process. 220–229. Sang-Hyun. Poor. pp. Mogensen. Veh. Biglieri. Conf. no. Senst and G.. Prof.. Wireless Commun. A. 1992. [27] X.-S. 29.. [20] S.S. Jan. pp. 190–208. Aug. Quek. 66–68. Acoust. 1998. no. Gulati. K. Dept. He was a Research Assistant with the Embedded Signal Processing Laboratory. vol. 2011. 2nd ed. K. S. C..” in Proc. 12. Aug.” in Proc. “A study of platform EMI from LCD panels—Impact on wireless... 1995.11b indoor system design. 4. Vari.” in Proc. Pers. and Ph. ser. 2044–2056. no.” Amer. Digital Communications Over Fading Channels. K. 2006. pp. 7. Signal Process. Veh.” IEEE Trans. Aghasadeghi. Luo. Apr.” U. [19] K. 2005. Chinn. May 1999. pp. Inf. . 45. 5. 1–5. vol. 1–6. J. 2. Lin. K. P. vol.” IEEE Trans. IEEE Int. Jafar. no. Feb. “Efficient estimation of class A noise parameters via the EM algorithm. Signal Process. and Ph.. pp. vol. real-time multichannel testbeds for underwater communication systems and powerline communication systems. no. pp. pp.” IEEE Trans. B. 1405–1408. [33] J.S. 2009. 50.. Electromagn.D. vol. Juttner. R. E. L.” IEEE Trans. 587–592. 2005. Technol. degrees in electrical engineering from The University of Texas at Austin in 2008 and 2011. Conf. [28] M. Apr. 60–72. Statist. [2] D. [26] K.. 4. Jun. Ilow and D. Gao and C. vol. Feb. “Dynamic frequency planning versus frequency reuse schemes in OFDMA networks. degree in electrical engineering from the Indian Institute of Technology. and P. Portland. and C.” in Proc. McDonald and R. “Co-channel interference modeling and analysis in a Poisson field of interferers in wireless communications. Commun. A. a Silicon Valley-based startup developing mobile backhaul solutions. The University of Texas at Austin. Evans. 38. “Cognitive network interference. [25] K. A. Zeidler. Conf. pp.S. Jun. Feb. [35] J. OR. Middleton. Technol. vol. 3. Evans. 46. M. Ascheid. Nassar. Q. [29] T. Fall Conf.” IEEE Trans. M. 3701–3705. respectively. S. Zhang. His current research interests include interference modeling and mitigation for wireless systems. A.. Deutsche Bundesbank’s Ann. 1743–1754. Technol... and M. 42. Gulati. Tinsley.. He has graduated 20 Ph. vol. K. statistical modeling of radio frequency interference. His research interests included real-time multichannel multicarrier testbeds for ADSL communication systems. McLaughlin. 480–493. 6. 2009. “A statistical and physical mechanisms-based interference and noise model for array observations. Taqqu. Wiley Series in Telecommunications and Signal Processing. “Performance analysis of compact antenna arrays with maximal ratio combining in correlated Nakagami fading channels. [11] A. vol. L. Goldsmith. [7] M. 897–906. [8] C.. “From antenna spacings to theoretical capacities—Guidelines for simulating MIMO systems. Taipei. W. Theory. “Performance of a spread spectrum packet radio network link in a Poisson field of interferers. J. REFERENCES [1] J. 58. Symp. Hatzinakos. “Mitigating near-field interference in laptop embedded wireless transceivers. 6. Evans. Sousa. Blum. 1998.. Jan. pp. 2. and the Texas Exes Teaching Award from the UT Austin alumni association in 2011. 4. [22] P. B. Commun.” IEEE Trans. Speech Signal Process.” IEEE Trans.. Zabin and H.” IEEE Trans. S. Signal Process. Andrews. 2010. Patent 5 742 911.. S. vol.” in Proc.. Office of Telecommun. Gulati. IEEE Int. New York: Wiley. U.Tech. Evans. “Thirteen ways to look at the correlation coefficient. Sreerama.” IEEE Trans. “Signal detection in multivariate Class-A interference. 2005. Tinsley.. 37. pp. 1991. Delaney. Chen and R. and S. Cadambe and S. Chan-Byoung. Goldsmith. Dong. double major in electrical engineering and computer science from the Rose-Hulman Institute of Technology in 1987. Lee. [10] M. Mar. from 1993 to 1996. 1249–1252. 2001. 22.D. Lopez-Perez.. vol. 2007.. Pedersen.” in Proc. Tepedelenlioglu. 1503–1510. 626–631. 8. “Statistics of co-channel interference in a field of Poisson and Poisson-Poisson clustered interferers. Evans received the National Science Foundation CAREER award in 1997. Signal Process. May 2008. Nolan. He is currently a systems engineer at Fastback Networks. K. Commun. no. [9] A. vol. pp. “Procedures for determining the properties of the firstorder canonical models of class A and class B electromagnetic interference. IEEE Veh. 6. June 2009.: Cambridge Univ. “Multivariate stable densities and distribution functions: General and elliptical case. Middleton. Lee and Z. 1601–1611. Schumacher. 2009. our proposed models can better inform such analysis. Chopra. T. Inf. “Analytic alpha-stable noise modeling in a Poisson field of interferers or scatterers. statistical signal processing for high-speed MIMO communication system design. pp. Blum. Paulraj. and F. Jan. 267–277. Pers. He was a Postdoctoral Researcher in electronic design automation at the University of California. [16] D. [21] J. no. 1994.” in Proc. 43. This leads to the design of robust receivers that are better suited to operate in the presence of interference in different network environments.CHOPRA AND EVANS: RFI IN MULTIANTENNA RECEIVERS 3603 providing a link between network models and interference distribution. Sel. V. Nov.D. 51. pp. Wireless Commun. Indoor Mobile Radio Commun. Hasan and J. “Performance bounds of MIMO receivers in the presence of radio frequency interference.. K. no. no. [5] J. May 2008. Inf. vol. [32] L. Evans (S’87–M’93–SM’97–F’09) received the B. “Channel models for IEEE 802. Commun. MIMO Wireless Commun. Jan. 2005.” IEEE Wireless Commun. 19–29. respectively. A. Berkeley. “Non-Gaussian noise models in signal processing for telecommunications: New methods and results for class A and class B noise models. The University of Texas at Austin (UT Austin). Gulati.” in Proc. and J. July 2000. students. Heath. Shi. Acoust.S. Middleton. [34] A. Compat. pp. Wireless Communications. R. Commun. [24] Y. M. A. vol. A. “Wireless communications and coexistence for smart environments. “Interference alignment and spatial degrees of freedom for the K user interference channel. 48. Commerce. 2.. Speech. K. vol.. 6. 1.” IEEE Trans. IEEE Int. Aug. R. [31] M. [18] J. no. 48. vol. [3] A. Press. “The guard zone in wireless ad hoc networks. B. and R. 1999. 2007. “Space-time coding over fading channels with impulsive noise. Wireless Commun. and H.. Alouini. Rep. V. and the M. 534–538. Electromagn. G. Chopra. Win. Stable Non-Gaussian Random Processes: Stochastic Models With Infinite Variance. He has published more than 200 refereed conference and journal papers. Simon and M. Calderbank. IEEE Int.” IEEE Trans. vol. Monti. Mazzenga. 2002. pp. N. 2007. in 1988 and 1993. 2003. pp.” in Proc. S. pp. Dec. 1129–1149. 971–975. Rees. 2000. pp. Rodgers and W. “Linear network coordinated beamforming for cell-boundary users.” IEEE Trans. and X. Andrews. vol.. Dumbrill and I. L. Conf. Conf. IEEE Int. “A sectorized beamspace adaptive diversity combiner for multipath environments.K. pp. Dec.. Aditya Chopra (S’06) received the B. degrees in electrical engineering from the Georgia Institute of Technology. [14] D..” IEEE Trans. Theory. Slattery. “Optimal output back-off in OFDM systems with nonlinear power amplifiers. Nicewander. 1988. New York: Chapman and Hall. 2008. “MIMO receiver design in the presence of radio frequency interference. Tinsley. Adv. J. Sujeeth. 19–24. Delhi. [30] D. H. Yang and A. G. pp. Peha. A. Commun. K. root causes and mitigation methods. Heath. and joined the faculty at the UT Austin in 1996. 3.S. Theory. vol. L. R. 2004. 1. Poor. Tinsley. “Efficient algorithms for sequence detection in non-Gaussian noise with intersymbol interference. and mitigation of synchronous noise in test and measurement platforms. Taiwan. 1979. R. vol..” in Proc. S.” IEEE Trans. [12] E. pp. “Interference cancellation for cellular systems: A contemporary overview. W. 1. Shin. R. no.” IEEE Trans. Conf. Cambridge. [15] G.. and 9 M. 48. S. and electronic design automation tools for multicore embedded systems. Signal Process.. 5. IEEE Global Telecommun. Samorodnitsky and M. vol. IEEE Int. G. Petropulu. IEEE Int. Mag. Compat. B. Rabbachin.S. [6] V. and X. Z. He is the Engineering Foundation Professor in the Department of Electrical and Computer Engineering (ECE).” in Proc. R. Sep. Andrews. in 2006 and the M. no.” IEEE Trans. Constantinides. Conf. Apr. Bettner. A. and K. A. Atlanta. May 2000. Apr. J. “Sectorized Cellular Radio Base Station Antenna. Borrelli. Tech. [17] E. Areas Commun. Brian L. Symp. 1976. pp.