Spherical Statistics and Spatial Correlation for Multielement Antenna Systems

The well-known assumption of horizontal plane wave propagation is investigated and evidence suggests that elevation plays a crucial role in defining the spatial correlation between signals on adjacent antenna array elements. To augment previously published studies, an explicit relationship between the distribution of scatterers in three-dimensional (3D) space and the spatial correlation is formulated. A novel approach is taken for modeling of the distribution of scatterers in space. More specifically, the distribution of scatterers is modeled by the 3D von Mises-Fisher (vMF) distribution. In addition, a closed-form expression is derived for the harmonic coefficients of the vMF density. The main derivation expresses the spherical harmonic coefficients associated with an arbitrary mean direction on the sphere. Further, a closed-form expression for the spatial correlation function (SCF) is derived, based on the spherical harmonic expansion (SHE) of plane waves as well as the harmonic coefficients of the expanded vMF density. A novel approach is proposed for including the effect of directional antenna responses in the SCF. Finally, the SCF is evaluated under the existence of multiple scatterer clusters in the channel.


Introduction
Multielement antenna systems have been the centre of wireless communication research over the past decade due to the significant increase in capacity that they offer. Operators worldwide strive for capacity and coverage increase as the number of subscribers grows and demanding multimedia services are implemented. Complicated, expensive, and timeconsuming optimization techniques that aim to improve the system performance, for example, drive test analysis; antenna and/or power optimization should ideally be avoided. Multielement antenna systems have been proven to increase capacity and decrease interference through spatial multiplexing and/or beamforming. Significant efforts are continuously made to analyse and simulate such systems with a high degree of accuracy although the nature of the assumptions often being made is sometimes doubtful. For instance, some studies in the literature assume that the signals received at adjacent antenna elements are uncorrelated. This may be considered an oversimplification of the challenges encountered in a wireless propagation channel. Further, occassionally elevation is ignored, and the effect of spatial correlation on the systems performance can be misleading, especially in urban clutter types. Our conjecture that elevation can have a significant effect on the correlation function has motivated this work. Various spatial correlation models have been proposed, where elevation is ignored and the distribution of scatterers or equivalently the angle of arrival (AoA) of paths at the mobile station (MS) (or even base station (BS)) is modeled as a two-dimensional (2D) uniform, Gaussian, von Mises, or Laplacian [1][2][3][4][5] distribution. A drawback associated with distributions that are defined on the line (e.g., Gaussian) is that the summary statistics are calculated in such a way that the directionality of the variables is not taken into consideration.
The need to investigate the spatial correlation between signals on adjacent antenna elements is well known in the research community. In [6], closed-form expressions for the spatial correlation function (SCF) of antenna arrays in a three-dimensional (3D) multipath environment are derived.

EURASIP Journal on Wireless Communications and Networking
However, in [6] the distribution of the angle of arrival of multipaths is assumed to be uniform over a given range of angles, and the azimuth is independently distributed from elevation. In [7], a thorough review of different 2D spatial correlation models is provided. By including the spatial information, the authors derive a closed-form expression for the ergodic capacity. In [8], the authors provide a capacity analysis of a dense multielement antenna system having obtained the spatial correlation values between the antenna elements. In addition, the idea of 3D propagation has also been considered in a number of other research works [9][10][11]. Toeltsch et al. provided a statistical analysis of 3D wideband measurement data after the application of a multipath parameter extraction algorithm, namely Unitary ESPRIT. The multipath components estimated were then clustered in the angular and delay domains, allowing a detailed statistical analysis of the propagation properties. In [10], the authors derived a cross-polarized multielement antenna channel model that takes into consideration the azimuth as well as the elevation domains. In [11], a spherical outdoor to indoor power spectrum model is developed for the MS, where the necessity of including the third dimension, that is, elevation becomes evident.
In this paper, the spatial correlation model derived in [12] is reviewed under a three-dimensional scattering distribution function, namely the von Mises-Fisher (vMF) distribution. It is shown that the correlation is not insensitive to elevation and that it is dependent on three key parameters, that is, angular spread (concentration of scatterers within a cluster), mean AoA in azimuth and elevation, and position of the elements. The results of this paper suggest that the correlation decreases significantly as elevation increases while this is dependent on the three parameters stated previously. To better understand this effect, the vMF distribution is expanded into spherical harmonics, and the harmonic coefficients are obtained in a closed-form expression. This closed-form expression is later used in the formulation of the spatial correlation function (SCF).
The arguments for proposing the vMF model are outlined in [13,14], where it is shown that this distribution possesses excellent characteristics in describing the AoA of paths at the MS. In [14], a mixture of distributions is used in order to describe the scattering distribution function. This is in agreement with the natural phenomenon reported in [15,16], that the channel clusters the incoming paths. For the same reasons, in this paper, the concept of mixture modeling is adopted for the characterization of the spatial correlation function. Consequently, the spatial correlation is scaled by the probability that each cluster contributes to the channel. The estimated priors of the vMF mixture model may define the magnitude of this relationship in terms of an assumed or otherwise estimated power value emanating from each scatterer cluster. The estimation of the parameters of the vMF mixture model is achieved using a soft-Expectation Maximization (soft-EM) algorithm [13]. The organization of this paper is as follows. In Section 2, the basics on spherical harmonics are provided, and a closed-form expression for the harmonic coefficients of the vMF density is further derived. In Section 3, a closed-form expression for the spatial correlation based on the unimodal vMF assumption is provided. In Section 3, the spatial correlation function is evaluated for various scenarios (of a constrained spherical array topology), where it is shown that elevation plays a major role in defining correlation. Later in Section 4, one element is elevated with respect to the other, and the correlation is re-examined. In this section, a detailed analysis on the effect of directional array responses on the SCF is provided. In Section 5, the SCF is extended so to capture the effect of multiple clusters of scatterers in the channel. Finally, the conclusions are presented in Section 6.

Spherical Phase Modes and the vMF Density
The vMF is a generic 3D model for directional data distributed unimodally on the sphere. It can therefore be considered as a suitable candidate for applications where the variables are angular in nature [17]. These variables are signified by ϑ and ϕ denoting the angles of the random vector Ω widely used in the forthcoming analysis. The distribution is closely related, and it is in fact analogous to the multivariate Gaussian distribution. If the distribution is defined on the normal sphere S 2 , its density function is significantly simplified. More specifically, the density function takes the following form: This equation is known as the general form of the vMF distribution [17] and is dependent on the modal direction µ (known as the axis of symmetry) and the dispersion parameter κ. Note that µ is defined in terms of (ϑ o , ϕ o ). The interested reader is referred to [14] for further mathematical details of the distribution and graphical illustrations for its interpretation within a scattering context. The spherical phase modes of the vMF are analyzed in the sequel.

Spherical Phase
Modes. The spherical phase modes or spherical harmonics are the angular portion of an orthonormal set of solutions (on the surface of the sphere) to Laplaces equation. The spherical harmonics are defined as follows [18,19]: where ϑ and ϕ are the elevation and azimuthal angles in the spherical co-ordinate system, n is the degree of the spherical harmonic, and m is the order. P m n are the associated Legendre functions. The orthonormality principle of spherical harmonics may be stated as [18] where δ n1,n2 , δ m1,m2 denote the Kronecker delta defined by δ n,m = 1, n = m, δ n,m = 0, n / = m. Knowing that the spherical harmonics form a complete orthonormal basis over S 2 allows EURASIP Journal on Wireless Communications and Networking 3 the expansion of any square-integrable function (i.e., vMF) as a linear combination of these The spherical harmonic coefficients f n,m can be obtained (in a closed-form expression) using the spherical Fourier transform, otherwise the use of numerical integration techniques might be necessary. The spherical harmonic coefficients can be found by where † denotes complex conjugation and dS = sin ϑdϑdϕ is the element of the solid angle. Note that since we introduce dS inside the double integral, there is no need to include the sin ϑ term of the vMF distribution in (1). It is important to distinguish between the pdf of (1) and the volumetric pdf [20]. In the next sections, the volumetric pdf is used, that is ignoring the last sin ϑ term in (1).

Closed-Form Expression for the Harmonic Coefficients of the vMF Density.
It is known that a function, f (γ) = e κ cos γ , can be expanded as follows [21]: By (6), it follows that (1) can be expressed as where cos γ = sin ϑ o sin ϑ cos(ϕ − ϕ o ) + cos ϑ o cos ϑ is the displacement vector between the mean direction vector and a vector pointing in a scatterer's direction. Making use of (5) and the Addition Theorem for spherical harmonics in (A.2), the coefficients can be written in a closed-form expression as follows: The derivation of (8) for an arbitrary direction on the sphere is detailed in Appendix A.1. In Appendix A.2, it is shown that by conditioning the mean elevation angle, ϑ o = 0 • (North Pole), a different spherical harmonic coefficient formulae is obtained. The latter being a special case of the main derivation.

Wavefield Decomposition and Spatial Correlation Using Constrained Spherical Antenna Arrays
Prior to proceeding into the relationship between the spherical harmonic expansion (SHE) of plane waves and the proposed vMF correlation function, it is vital to explain how the characteristics of the antenna array are integrated into this model. In the initial setup, each antenna comprising the array is assumed to be an isotropic radiator (three dimensions). In two dimensions, this corresponds to a circular aperture being equally sensitive in all azimuthal directions, thus revealing an omnidirectional pattern. Whilst recognizing that this assumption hinders the realistic performance of a multielement antenna system, it enables us to study the broader picture of the spatial correlation effect on the system's performance (e.g., capacity). Including realistic antenna patterns greatly complicates further analysis, making analytical solutions very difficult if not impossible. This simplification has helped us to decompose the rather complicated problem posed by the inclusion of mathematically untractable antenna patterns. We proceed in the following directions. In Section 3.1, a definition of spatial correlation is provided. The vMF SCF is derived in Section 3.2. The rest of the parts comprising this section are organised as follows. We begin with analyzing the effect of an elevated cluster of scatterers in a constrained spherical antenna topology (e.g., two elements are placed in the median plane θ = 90 • ) in Section 3.3. Our primary interest is to later compare the spatial correlation estimation of a 3D incoming wavefield (i.e., vMF distributed) to a 2D and 3D element arrangements. Finally, it is shown in Section 3.4 that under the condition of isotropy the vMF SCF reduces to the sinc function.

Spatial
Correlation of Narrowband Sources. In [12], an expression is derived that characterizes the SCF for an arbitrary distribution of scatterers on the sphere. It was shown that if the harmonic coefficients of the scattering distribution function are obtained, a closed-form expression for the spatial correlation can be derived. Note that it is important to obtain those coefficients in an analytical form which is often a difficult task. In case the coefficients are not obtainable in a compact closed-form, then other numerical integration techniques need to be identified, that is, Monte Carlo integration. In [12], the authors conclude that the correlation is relatively insensitive to restriction in elevation having assumed a field uniform in azimuth. One of the arguments of this paper is concerned with the scattering distribution function to be used during this analysis. It is reasonable that the correlation will be relatively insensitive to elevation if the azimuth is uniformly distributed. In order to identify the effect of elevation on the correlation function, more realistic distribution functions need to be used. In the sequel, it will be shown that using the vMF distribution we can see exactly how the correlation scales as the elevation changes. We observed that correlation decreases as elevation increases and that the spatial correlation is dependent on elevation.
First consider the signals s 1 and s 2 impinging on an array with elements m 1 and m 2 . Each element pattern is assumed to be isotropic, and therefore each signal is assumed to be received by an isotropic sensor. Under these conditions the normalized spatial correlation between the complex envelopes of the two received signals is defined as in [12] and may be expressed as where k = ω/c is the wave number and y is a unit vector in the direction of wave propagation. Note that P( y) is the distribution function of scatterers on the unit sphere. Since P( y) is defined on S 2 , the scattering environment can be specified given the spherical harmonic coefficients in (4), so that P( y) = f (ϑ, ϕ). In the following section, we attempt to integrate the spatial correlation definition presented above with the wavefield decomposition using a spherical antenna array.

Spherical Harmonic Expansion of Plane Waves and the
Proposed vMF SCF. From the Jacobi-Anger expansion in (10) and the Addition Theorem of spherical harmonics in (A.2), the spherical harmonic expansion of plane waves can be written as in (11) [22] where j n denotes the spherical Bessel function of order n This equation expresses the expansion of plane waves into spherical co-ordinates and shows that the spherical harmonics only exhibit a spatial dependence related to the directions of the distribution of scatterers (i.e., (ϑ, ϕ)). This result is due to a spherical aperture that employs a unity weighting function, and therefore is equally sensitive in all directions (isotropic). Confining the elements in the 2D plane results in the expansion of a plane wave on a circular aperture of omnidirectional elements. Combining (9) and (11) results in [22] ρ( In (12), f n,m is obtained from (8) to finally produce the closed-form expression for the vMF spatial fading correlation function The vMF correlation function proposed in (13) is expressed as a sum of terms, where each term either characterizes the scattering environment or the antenna locations. As a result, the effects of arbitrary antenna configurations and the complex scattering environment on the SCF can easily be investigated [7]. The correlation function in (13) has been evaluated for element position m 1 = 45 • and m 2 = 90 • . In this array architecture, the elements are positioned in the median plane. Consequently, the limiting case of a spherical array reveals a circular topology of elements. A circular aperture cannot discriminate 3D incoming signals with the same level of accuracy as a spherical aperture. This becomes apparent by realizing that the elevation of the incoming source is considered only within the argument of the nth order Bessel function. The azimuthal dependence, however, is expressed in the exponential term. Thus, circular harmonics can be rotated in azimuth but not in elevation. Also if the directions of sources are not significantly attenuated in elevation, then a circular topology cannot unambiguously resolve the angles of arrival of the impinging signals.
Herein, a comparison is performed between the spatial correlation estimates of the two array topologies under the assumption of a 3D incoming wavefield. The choice of angular position of elements was selected at random. Note that the angular position of elements in (13) produces a new two-dimensional (2D) vector given by m res ≡ m 2 − m 1 = 157.5(π/180). Note that the vector m resultant depicted in Figure 4 will be abbreviated as m res for convenience. In spatial correlation terms, the quantity m 1 − m 2 should give the same value although this time its corresponding resultant vector m res = 337.5(π/180). Subsequently, the correlation function is dependent on the angular position of elements. The assumptions above can be generalized to any antenna array topology, that is, linear or rectangular, which is a distinctive feature of the proposed spatial channel model. To carry out the multiple simulation scenarios presented in the following section, the number of terms (N) in the sum of (13) was set equal to thirteen (N = 23). The decision on the appropriate number of terms to be used in the sum of (13) was based on the extensive analysis provided in [23]. The authors showed that the multipath field can be safely truncated by (N + 1) 2 terms, where N = πer/λ , · is the ceiling operator, e ≈ 2.7183, and r indicates the region of space to which the field is coupled or equivalently the radius of the circular or spherical region in consideration. This region of space is where multiple antenna elements may be located. In the analysis provided in this section, we consider that all elements lie in a circular ring with a maximum interelement spacing of m 2 − m 1 /λ = 2. Note that m 2 − m 1 /λ denotes the interelement spacing distance, where m 2 − m 1 = R, and this should not be confused with the radius r of the circular array. With the use of simple trigonometrical identities, it can be shown that if m 2 − m 1 /λ = 2 and the angle between the two elements is 45 • , the radius r 2.6. As a result, the maximum number of terms required is N = 23. However, due to the restriction in the multipath field, as defined by the concentration parameter κ, the total number of terms required may be further reduced in accordance with N = ωer/λ , where ω defines the spread of the distribution [23].

Evaluation of the Closed-Form SCF.
The aim of this section is threefold: firstly, to identify the effect of elevation on the vMF SCF proposed in (13) (Section 3.3.1); secondly, to examine the effect of the azimuth of arrival of propagating waves on the SCF (Section 3.3.2); finally, to investigate the influence of the concentration parameter κ on the vMF correlation function (Section 3.3.3).

SCF for Varying Mean Elevation of Arrival.
To generate the coefficients f n,m , the concentration parameter κ and the MAoA of the cluster must be specified a priori (or estimated [13,14]). To proceed, the initial κ value was set to twenty (κ = 20) (Please note that in the proposed unimodal vMF model the distance between the scatterer cluster and the antenna array is not explicitly considered. If the MS is moving towards this cluster of scatterers, its associated κ value as seen at the MS will decrease. Consequently, a decreasing concentration parameter κ can capture this effect. By setting κ to a large value, that is, κ = 20, scenarios where the scatterers are located in the far field of the multielement antenna system can be modeled. A decreasing κ value can therefore model the contribution of local scattering to the spatial correlation experienced between the receiving antennas. For this reason, a smaller value of κ is later introduced in this section, and its effect on the correlation function is investigated. Because of symmetry, there is no need to evaluate (13) for ϑ o greater than 90 • . It is evident that the correlation decreases as elevation increases (as measured from 90 • ) and vice versa. Further, consider that a cluster is located at mean azimuth ϕ o = 30 • with concentration parameter κ = 20. The evaluation of (13) for this mean azimuth angle results in Figure 1(b). Note again that as elevation increases (from ϑ o = 90 • → 0 • ) correlation decreases, although the difference between the lowest and maximum elevation points (ϑ o = 90 • and ϑ o = 0 • , resp.) is now smaller. The correlation experienced between elements m 1 and m 2 at a mean azimuth cluster angle ϕ o = 157.5 • is identical to the one presented in Figure 1(a) (due to symmetry). The underlying reason is illustrated by geometry in Figure 4. In Figure 1(c), the scaling of correlation with elevation at ϕ o = 300 • is further provided. The spatial correlation difference between minimum and maximum ϑ o 's has now been increased, since the vector ϕ o = 300 • is closer to the resultant m res = 337.5 • than the vector ϕ o = 30 • is to m res . Intuitively, this can be explained by the fact that the angular difference between the first pair of vectors is only 37.5 • compared to the second pair whose angular difference equals 52.5 • . The wider angular range implies lower correlation values since as the distance between any given pair of vectors increases as their associated correlation decreases. Additionally, an interpretation of the observable decrease in the SCF during the ascent of the scatterer cluster to the north region of the sphere is that as the spatial cluster is elevated, a larger deviation is projected in the azimuthal plane. Consequently, maximum deviation is realized when the mean direction vector aligns with the vector directed along the z-axis. In view of this, we speculate that the decrease in spatial correlation due to the spatial cluster's disposition from the azimuthal plane is justified.
The assumption of elements being positioned strictly on the azimuthal plane needs to be retained.

SCF for Different Angular Spread.
To extend the results presented above, the concentration parameter κ is now varied in order to investigate its effect on the SCF. As expected, decreasing the value of κ decreases the spatial correlation experienced between adjacent antenna elements. In Figure 3, the estimated spatial correlation values for κ = 10, κ = 5, and κ = 1, respectively, are shown. The results can be compared with the correlation values depicted in Figure 1(a), where the concentration parameter κ = 20. A decrease in the spatial correlation is evident.
The correlation function is clearly not insensitive to elevation. This was also confirmed in [24], where the authors made use of the more complex Fisher-Bingham 5-parameter model. The Fisher-Bingham 5-parameter (FB5) model does not have a well-known series expansion, andas a result, its spherical harmonic coefficients cannot be obtained in a closed-form expression. Although the FB5 model is more generic, its complexity is a limiting factor. In [25], Shafi et al. also showed that ignoring elevation can significantly underestimate capacity. This can be thought of as a direct consequence of the fact that at higher elevation angles correlation decreases, and therefore capacity increases. Interestingly, the only angles at which the elevation effect vanishes are at ϕ o = 67.5 • and ϕ o = 247.5 • , respectively (see Figure 1(d)). The reason behind this observation is that at ϕ o = 67.5 • and ϕ o = 247.5 • , the cosine of the angle between the m res and both of the vectors is 0. The vectors are at right angles to each other. Figure 4 provides a geometrical explanation for this effect. The spatial correlation is minimum at both locations even for different elevation angles. Therefore, according to this study, the correlation becomes insensitive to elevation when (i) the concentration parameter κ = 0, in which case the distribution of scatterers over the sphere becomes completely isotropic; (ii) the mean azimuthal angle of a cluster is on the plane orthogonal to m res .
The above results are directly comparable to the results obtained in [12]. Here, we will not present the results of the uniform (although limited) azimuth/elevation scattering distribution function presented in [12]. It was shown that if the scatterer distribution is uniform in azimuth, the correlation becomes insensitive to restriction in elevation. As an addition to the significant work of [12], in this work, we show exactly how elevation affects correlation if a more realistic distribution of scatterers (i.e., vMF) is to be used. In [13,14], it was shown that the vMF provides a very accurate fit for the AoA of multipaths at the MS and hence the distribution of scatterers in space. This result has therefore motivated the investigation of spatial correlation in antenna arrays under the 3D vMF model. In Section 3.4, it is shown that the proposed vMF model can also reduce to the wellknown sinc function when one of its parameters, namely κ, is set to zero. This implies that the scatterers are isotropically distributed on the surface of the sphere.

Isotropic Distribution of Scatterers.
If the distribution of scatterers is isotropic on the sphere, then the plane waves will be incident on the elements from all directions. Isotropic distribution implies that κ = 0, and therefore only the order n = 0 is required for the calculation of the spherical harmonic coefficients. Consequently, the vMF density reduces to f (ϑ, ϕ) = sin ϑ/4π(Appendix A.2) and (5) becomes f 0,0 = √ 1/4πP 0 (cos ϑ) = 1/ √ 4π. Upon substitution of f 0,0 and Y 0 0 in (12), the vMF SCF function reduces to the classic result obtained in [26]. Note that the spherical harmonic of degree n = 0 is given by Y 0 0 = 1/2 √ π. The same result was obtained in [12] under a uniform (limited) azimuth-elevation distribution on the sphere.

Wavefield Decomposition and Spatial Correlation Using Directional Antenna Responses
In the previous section, the so-called constrained spherical array topology was examined under the condition of omnidirectional antenna responses. In this section, focus is mainly shifted towards antenna responses that are directional in nature, and the effect of an elevated element on SCF is only briefly discussed. To achieve this, we introduce a weighting function G(ϑ, ϕ, ϑ o , ϕ o ) in order to include the effect of a realistic antenna pattern while computing the spatial correlation. This function describes the degree of sensitivity of the antenna pattern towards some specific direction of interest. Unfortunately, its introduction hinders the emergence of an analytical solution as it will soon become evident. Therefore, here, a numerical solution is sought due to the complexity of the problem. This complexity escalates by adding other impacting factors of antenna pattern properties, that is, mutual coupling and antenna mounting platform effects. The effect of the antenna mounting platform on the spatial correlation function will be investigated in future work. ϑ, ϕ = 45 • , 180 • . Interestingly, the mean direction vector that produces minimum correlation is at the mean elevation angles ϑ o1 , ϑ o2 = 90 • , 0 • . A similar interpretation may be given as the one provided earlier for the constrained spherical topology. the key point in here is that as the cluster mean direction vector approaches the resultant vector, correlation increases.

Spatial Correlation with Directional Antenna Patterns.
Directional antenna patterns are commonly used in practice, and therefore the extension of this work to capture their influence on spatial correlation is noteworthy. Base stations usually employ directional patterns as opposed to mobile terminals that should not in most cases favor any direction. The influence of antenna directivity on the performance of multipath channels is multidimensional. From a propagation viewpoint, the antenna directivity not only affects the power received at a terminal, but it also has a significant impact on the angle spread of the channel. It is an important realization that a directive antenna has the potential to change our perception on the physical distribution of scatterers. Of course, this is dependent on the antenna effective aperture, which is in turn related to the gain of the antenna. Also, nonisotropic antenna patterns yield weighted AoA components, and as a result, they alter the power angular spectrum received at the MS/BS. At this moment, the weighting function G(ϑ, ϕ, ϑ o , ϕ o ) is introduced, which essentially describes the sensitivity of an infinitesimal point located on the array at ϑ, ϕ. Note that the weighting function is also associated with the look-direction ϑ o , ϕ o of the antenna pattern. This means that the pattern may be rotated at any azimuth-elevation pair of angles. Following the approach presented in [27], the weighting function may be expressed as where γ was defined earlier in (7). Applying the Addition Theorem for spherical harmonics (A.2) to (14) transforms the above equation to The application of this weighting factor to the inverse spherical harmonics transform of a plane wave (5) results in an expression for the harmonic coefficients (decomposed wavefield) due to a directional antenna being present. This expression can be written as After substitution of the weighting function in (15) and the equation corresponding to the plane wave expansion in (11) to (16), it can be seen that the spherical harmonic coefficients cannot be obtained in an analytical form since the integral contains the product of three spherical harmonics. Consequently, the equation that defines the expansion of plane waves in spherical co-ordinates under a directional antenna cannot be derived in a closed-form solution. In this respect, (17) can only be evaluated numerically, and may be rewritten as follows: where G( y, µ) = G(ϑ, ϕ, ϑ o , ϕ o ). This rearrangement has a meaningful interpretation since the product P( y) × G( y, µ) defines the overall scattering response through the multiplication of the azimuth-elevation scattering response P( y) with the complex antenna pattern G( y, µ). To exemplify this, let us consider the case where the elements exhibit dipolelike characteristics. Under this assumption, the weighting function in (15) may be rewritten as whose associated expansion coefficients are given by However, this is of little significance at the moment since the SCF will be evaluated numerically. Note that this is a firstorder dipole response whose elevation beamwidth is 90 • . Increasing the order of the array reduces the corresponding beamwidth. Second-and third-order dipole responses have elevation beamwidths of 65 • and 54 • , respectively, which can be met in practical scenarios. The elevation beamwidth of dipoles varies depending on the application. Typically, antenna designs with high-directivity index have reduced elevation beamwidths. Of course, elevation beamwidths of azimuthally omnidirectional dipole patterns can be as small as 7 • , but this should not compromise the generality of the approach followed in here. Other practical designs include the second-and third-order dipole, cardioid, and hypercardioid directional responses. In general, a Jth order array response may be expressed as [28] This is the canonic form of the antenna response, and it actually factors (21) into its first-order terms. The normalised directional array response may be rewritten in the following form as a function of γ: G γ = α 0 + α 1 cos γ + α 2 cos 2 γ + · · · + α j cos j γ.
Solving (20) results in expressions for the s in terms of the αs. For a third-order array, for example, (20) and (21) become  respectively. In Table 2, some of these patterns are listed and subsequently plotted in Figure 5. In Figure 6, the correlation estimates of (17) are displayed for various interelement spacings and concentration parameters. The look direction for all antenna responses is set at ϑ o , ϕ o = 90 • , 157.5 • , and therefore the cluster mean direction vector coincides with the look direction of each antenna. This configuration produces the highest correlation estimates. It becomes clear that the isotropic element pattern assumption serves as a lower bound to the spatial correlation function. Also, as the concentration parameter increases, the weighted (including antenna effect) SCF approaches the unweighted SCF. For low concentration parameter values, that is, κ = 5, the dipole response produces the highest spatial correlation estimates due to possessing the highest directivity index in comparison to the other responses. The choice of antenna response clearly depends on different factors, for example, front-to-back ratio, beamwidth, and directivity, and therefore an antenna response is application specific. The approach followed is generic enough to allow any antenna response to be included in the model. In the next section, we mathematically describe the scenario of multiple scatterer clusters and its inclusion on the spatial correlation function.

Mixtures of vMF Distributions in a Spatial Correlation Context
During the last decade, directional channel measurements have been extensively used in various measurement campaigns worldwide. This is due to the fact that directional channel measurements can reveal the clustered nature of the propagation channel. Important findings related to these measurements have been reported in [9,15,[29][30][31][32], all of which establish that channels with multiple scatterer clusters often occur in nature. In [29], the authors identify the existence of far scatterer clusters using measurements from a campaign in Frankfurt, Germany. In [30], Kuchar et al. not only mention the existence of far scatterer clusters that result in increased delay dispersion but also report that 65% of the incoming signal energy is incident with an elevation larger than 10 • . The latter experiment was carried out in Paris, France. The work in [31,32] also confirmed the existence of multiple scatterer clusters. Standardized models, such as the one developed by the third-generation partnership project (3GPP), namely the spatial channel model (SCM), also accounts for multiple scatterer clusters. It is therefore important to include the multiple scatterer cluster effect in future channel models and to also recognize that far scatterer clusters are an important propagation scenario in urban environments.
The definition of a cluster is subject to different analysis constraints. Usually the clusterization of multipath components takes place in spatial, power, and temporal domains. This paper aims at understanding of fading effects by exploiting the spatial domain. In [13,14], clustering is performed in the spatial domain. It was shown that the vMF mixture model provides a good fit to the angular estimates of the multidimensional parameter extraction algorithm, namely RiMAX [33]. The model-based approach followed uses the soft-EM for mixtures of vMF distributions. Mathematically, each distribution in the mixture of Q VMFs can be described The overall density of the mixture model comprising of Q vMF distributions can be described as where the parameter vector V = {χ 1 , . . . , χ Q ; v 1 , . . . , v Q }, and χ Q defines the prior probability that vector Ω was generated by the qth component. After taking into consideration (24), the SCF in (13) can be modified accordingly, where the priors Q q=1 χ q = 1. The resultant equation is the multimodal vMF SCF function, the mathematical details of which are omitted in here. A similar approach was followed in [34], where the authors considered a mixture of 2D von Mises distributions. To understand the implications of the multimodal density on the SCF, let us assume that in a radio propagation channel exist 3 clusters with different MAoAs, and concentration parameters κ. In Table 1, the parameters are shown. We have assumed equal contribution from each cluster. However, the results can be generalized to any set of parameters, any number of clusters, and any contribution from each cluster. To sample a point from the mixture density in (24), the qth vMF was randomly chosen with probability χ q and subsequently sample a point from f q (Ω | v q ) [35]. The evaluation of the multimodal SCF for this set of parameters produces Figure 7. It is evident that maximum correlation is obtained at ϑ o = 90 • . To further explore the concept of mixture modeling, we consider that the 2nd cluster has a higher assigned prior (χ 2 = 0.6) and that the rest of the parameters remain the same (as in Table 1). In this case, the contribution of the rest of the spatial clusters in the mixture is χ 1 = 0.2 and χ 3 = 0.2, respectively. Evidently, the priors of each cluster in the mixture can have a dramatic effect on the correlation function as shown in Figure 8. The mixture  correlation function does not only depend on the prior contribution of each component, but on the mean elevation AoA, mean azimuth AoA, and concentration parameter of each cluster. Note that the priors can be a function of the total power in a cluster. It is of common interest to direct focus on the power-dominating clusters. Therefore, the priors χ q in the mixture density proposed can be changed to normalized power-weighted values (after the soft-EM estimation). This can be achieved by estimating the total power in each cluster, and changing the Q q=1 = χ q term to Q q=1 = p q , where p q is the total power in each cluster.
It is important to observe that in the proposed vMF mixture model, the distance from the scatterer cluster to the antenna array system is considered, although indirectly, as opposed to the unimodal case presented earlier. This is due to the fact that the contribution of each cluster in the vMF mixture is scaled in accordance with the power emanated from it. It is therefore valid to expect that scatterer clusters in the vicinity of the MS will have a higher associated power, contributing more to the spatial correlation between the receiving antennas. Equivalently, scatterer clusters in the far field of the multielement antenna system are expected to have a higher value of concentration and a lower value of power. Evidently, the proposed model can fit to a wider range of propagation scenarios.

Conclusions
In this paper, the assumption of traditional horizontal wave propagation and its effect on the SCF have been reviewed. A novel approach is proposed (vMF) for modeling the distribution of scatterers in space. The spherical harmonic coefficients of the vMF distribution were derived and various simulations on the modified SCF function indicate   In this case, the priors are set such that the 2nd cluster in the mixture has a higher value, that is, ρ 2 = 0.6, and the remaining two spatial clusters have ρ 1 = 0.2 and ρ 3 = 0.2, respectively. that the correlation decreases as elevation increases when the elements are positioned on the median plane. The proposed method directly links the spatial correlation with the elevation of the incoming cluster of multipaths. It was also observed that if an element is elevated while the other is kept on the horizontal plane, signal correlations change significantly. Since most of the practical array designs include two-dimensional element configurations, the first case applies. The proposed model was further enhanced by the inclusion of an arbitrary weighting function to describe the effect of the antenna pattern on the spatial correlation. Finally, the effect of multiple clusters of scatterers that is frequently reported in the literature was also included in the SCF. The proposed model can be used to study the capacity in multielement antenna systems. Upon substitution of (A.8) into (4) and making use of the Y 0 n (ϑ, 0) relationship stated earlier, the vMF pdf when µ = [0 0 1] is obtained in (A.5). In the above derivation, we have used the orthogonality principle of Legendre polynomials which states that It is important to note that the spherical harmonic coefficients derived are independent of the antenna positions.