Analysis of Static Cellular Cooperation between Mutually Nearest Neighboring Nodes
Luis David Alvarez Corrales, Anastasios Giovanidis, Philippe Martins, Laurent Decreusefond
11 Analysis of Static Cellular Cooperation betweenMutually Nearest Neighboring Nodes
Luis David ´Alvarez Corrales, Anastasios Giovanidis,
Member, IEEE,
Philippe Martins,
Senior Member, IEEE, and Laurent Decreusefond
Abstract —Cooperation in cellular networks is a promising scheme to improve system performance. Existing works consider that auser dynamically chooses the stations that cooperate for his/her service, but such assumption often has practical limitations. Instead,cooperation groups can be predefined and static, with nodes linked by fixed infrastructure. To analyze such a potential network, wepropose a grouping method based on node proximity. With the Mutually Nearest Neighbour Relation, we allow the formation of singlesand pairs of nodes. Given an initial topology for the stations, two new point processes are defined, one for the singles and one for thepairs. We derive structural characteristics for these processes and analyse the resulting interference fields. When the node positionsfollow a Poisson Point Process (PPP) the processes of singles and pairs are not Poisson. However, the performance of the originalmodel can be approximated by the superposition of two PPPs. This allows the derivation of exact expressions for the coverageprobability. Numerical evaluation shows coverage gains from different signal cooperation that can reach up to compared to thestandard noncooperative coverage. The analysis is general and can be applied to any type of cooperation in pairs of transmittingnodes.
Index Terms —Cooperation; Static groups; Poisson cellular network; Thinning; Interference; Poisson superposition. (cid:70)
NTRODUCTION C OOPERATION between wireless nodes, such as cellularbase stations (BSs) is receiving in recent years a lotof attention. It is considered as a way to reduce intercellinterference in future cellular networks and consequentlyimprove network capacity. It is particularly beneficial forusers located at the cell-edge, where significant
SINR gainscan be achieved in the downlink. In the wireless literature,there is a considerable amount of research on the topic,which relates to the concept of CoMP [1], [2], NetworkMIMO [3], [4], [5], or C-RAN [6], [7]. It is also expectedto play a significant role due to the coming densification ofnetworks with HetNets [8], [9]. The various strategies pro-posed differ in the number of cooperating nodes, the typeof signal cooperation, the amount of information exchange,and the way groups (clusters) are formed.Recent studies analyse such cooperative networks withStochastic Geometry as the main analytic tool [10]. Modelingthe position of wireless nodes via a Point Process givesthe possibility to include the impact of irregularity of BSlocations on the users’ performance (e.g.
SINR , throughput,delay). Furthermore, the gains from cooperation can be • Luis David ´Alvarez Corrales conducted this research while at T´el´ecomParisTech, 23 avenue d’Italie, 75013, Paris, France.E-mail: [email protected] • Anastasios Giovanidis is with the CNRS. He conducted this research whileaffiliated with T´el´ecom ParisTech, 23 avenue d’Italie, 75013, Paris, France.E-mail: [email protected] is now affiliated with University Pierre et Marie Curie, CNRS-LIP6.E-mail: [email protected] • Philippe Martins is with T´el´ecom ParisTech, 23 avenue d’Italie, 75013,Paris, France.E-mail: [email protected] • Laurent Decreusefond is with T´el´ecom ParisTech, 23 avenue d’Italie,75013, Paris, France.E-mail: [email protected] quantified in a systematic way, so there is no need to testeach different instance of the network topology by simula-tions. Closed formulas are very important for an operatorthat wants to plan and deploy an infrastructure with coop-eration functionality, because these can provide intuition onthe relative influence of various design parameters.
There are important results available for BS cooperation inwireless networks. In [11], Baccelli and Giovanidis analysethe case where BSs are modeled by a Poisson Point Process(PPP) and each user-terminal triggers the cooperation of itstwo closest BSs for its service. The authors show coverageimprovements and an increase of the coverage cell. In [12],Nigam et al consider larger size of clusters, showing thatBS cooperation is more beneficial for the worst-case user.The
SINR experienced by a typical user when served by the K strongest BSs is also investigated by Blaszczyszyn andKeeler in [13], where the authors derive tractable integralexpressions of the coverage probability for general fadingby the use of factorial moment measures. An analysis ofa similar problem with the use of Laplace Transforms (LT)is provided by Tanbourgi et al in [14]. Sakr and Hossainpropose in [15] a scheme between BSs in different tiers fordownlink CoMP. Outside the Stochastic Geometry frame-work, we find [5] and [16]. In [16], Papadogiannis et al pro-pose a dynamic clustering algorithm incorporating multi-cell cooperative processing. All the above works assumethat a user-terminal dynamically selects the set of stationsthat cooperate for its service, which changes the clusterformation for every different configuration of users. Thisis difficult to be applied in practice.Other works propose to group BSs in a static way , sothat the clusters are a-priori defined and do not change a r X i v : . [ c s . N I] N ov over time. The appropriate static clustering should resultin considerable performance benefits for the users, with acost-effective infrastructure. In favour of the static groupingapproach are Akoum and Heath [17], who randomly groupBSs around virtual centres; Park et al [18], who form clustersby using edge-coloring for a graph drawn by Delaunaytriangulation; Huang et al [19], who cluster BSs using ahexagonal lattice, and Guo et al who analyse in [20] thecoverage benefits of cooperating pairs modeled by a Gauss-Poisson point process [21]. The existing static clusteringmodels either group BSs in a random way [17], or theyrandomly generate additional cluster nodes around a clustercenter [20], [22], which is translated in the physical worldinto installing randomly new nodes in the existing infras-tructure. A more appropriate analysis should have a mapof existing BS locations as the starting point, and from thisdefine in a systematic way cooperation groups. The criterionfor grouping should be based on node proximity, in orderto limit the negative influence of first-order interference. Consider a fixed deployment of single antenna BSs on theplane. As argued above, we wish to organize these BSs (oratoms) into static cooperative groups , with possibly differentsizes. These groups must be mutually disjoint and theirunion should exhaust the whole set of BSs. Additionally,the groups must be invariable in size and elements withrespect to the random parameters of the telecommunicationnetwork (e.g. fading, shadowing, or user positions). Hence,we look for a criterion that aims at network-defined, staticclusters as opposed to the user-driven selection of otherworks.For this reason, we will propose rules that depend onlyon geometry: An atom takes part in a group, based solelyon its relative distance to the rest of the atoms. Geometryis related to the pathloss factor of the channel gain, so itencompasses important aspects that influence signal power.The specific grouping criterion (for static geometric clus-ters) that we propose in this work is the
Mutually NearestNeighbor Relation (MNNR) . The main idea is that two BSsbelong to the same group if one of the two is the nearestneighbor of the other. The MNNR is the keystone that allowsus to construct static clusters of singles and pairs. It is in-spired by a model studied by H¨aggstr ¨om and Meester [23],the
Nearest Neighbor Model (NNM) , and further analysedin [24], [25], [26], [27], [28], where each atom connects toits geometrically
Nearest Neighbor by an unidirected edge.Although we will consider here groups of size at most 2, theNNM can allow for an extension of our approach to includelarger groups. This is part of our ongoing research.
This paper provides the following contributions: • We introduce the MNNR, a grouping method for BSswhose positions are modeled by a stationary pointprocess Φ (Section 2). Most results are derived when Φ is chosen to be a Poisson Point Process (PPP). • We analyse wireless networks with two types of clus-ters (groups): Single nodes, that do not cooperate, and pairs of nodes that cooperate with each other(Section 2). • From the dependent thinning determined by theMNNR, we construct two point processes Φ (1) and Φ (2) , the processes of singles and pairs, respectively.Structural properties of both are provided: (a) theaverage proportion of atoms from Φ that belong to Φ (1) and Φ (2) , (b) the average proportion of Voronoisurface related to each one of them, (c) their re-spective Palm measures, as well as (d) propertiesconcerning repulsion/attraction (Section 2). • Our analysis is done in a general sense, withoutrestricting ourselves to specific cooperating signalschemes (Section 3). Altogether, we provide the ana-lytic tools that evaluate various strategies for trans-mitter cooperation/coordination, as those in [1], [2],[4], [29], [30]. • We provide an analysis of the interference generatedby the processes Φ (1) and Φ (2) , and derive explicitexpressions for the corresponding expected values,along with a methodology to obtain their LaplaceTransfrom (LT) (Section 4). • Based on the structural characteristics of the singlesand the cooperative pairs, we introduce an approx-imate model: the superposition of two independentPPPs. Using this, a complete analysis of the coverageprobability is provided, for two different scenarios ofuser-to-BS association (Section 5). • In Section 6 the analytic formulas are validatedthrough simulations and the gains of static nearestneighbor grouping are quantified. Section 7 presentssome pros and cons of the model. The final conclu-sions are drawn in Section 8.
Let all random elements be defined on a common probabil-ity space (Ω , F , P ) , and let E denote the expectation under P . Let Φ = { φ } be a point process, with values in R , whereevery φ represents an element of the space of simple, andlocally finite configurations of R points [31].If an atom x and a configuration φ are fixed, φ ∪ { x } denotes the simple and locally finite configuration contain-ing all the elements from φ plus x in the case where x does not belong to φ , otherwise it just denotes φ . In thesame fashion, φ \{ x } denotes the simple and locally finiteconfiguration containing all the elements from φ withoutthe point { x } in the case where x actually belongs to φ ,otherwise it represents just φ .The Euclidean distance and the Euclidean surface on R are denoted by (cid:107) · (cid:107) and S ( · ) , respectively.For x ∈ R and a A closed subset of R , we denote d ( x, A ) := inf y ∈ A (cid:107) x − y (cid:107) , the distance from x to A .Finally, for x ∈ R and r > , let B ( x, r ) := { y ∈ R | (cid:107) x − y (cid:107) < r } . HE M UTUALLY N EAREST N EIGHBOR MODEL
For two different atoms x and y in the configuration φ , wesay that x is in Nearest Neighbor Relation (NNR) with y (withrespect to φ ) if y = argmin z ∈ φ \{ x } (cid:107) x − z (cid:107) , and we write x φ → y . When the atom x is not in NNR with y , we write x φ (cid:54)→ y .Henceforth, we will only consider stationary point pro-cesses Φ = { φ } whose realisations fulfill the uniqueness ofthe nearest neighbor a.s. Note however that this conditiondoes not generally hold. For example, within the finiteconfiguration φ = { (0 , , (1 , , (0 , } , the Euclidean originis in NNR with both (1 , and (0 , . Definition 1.
Two different atoms x, y are in Mutually NearestNeighbor Relation (MNNR) if and only if (iff) x φ → y and y φ → x ,and we denote it by x φ ↔ y . In telecommunication terms, we saythat the two BSs x and y are in cooperation. Definition 2.
An atom x ∈ φ is called single iff it is not inMNNR (does not cooperate) with any other atom in φ . That is, iffor every y ∈ φ such that x φ → y , then y φ (cid:54)→ x . (a)(b) Fig. 1: (a) The atoms x and y are mutually nearest neighbors,so, they work in pair. The atom x is the nearest neighbor of w , but w is not the closest atom to x , thus w is single. (b) APoisson realisation with its corresponding Voronoi diagram.The asterisks are the single BSs, the connected dots are thecooperating pairs.For x, y ∈ φ fixed, denote the area C ( x, y ) := B ( x, (cid:107) x − y (cid:107) ) ∪ B ( y, (cid:107) x − y (cid:107) ) . In geometric terms, the relation x φ → y holds iff the disc B ( x, (cid:107) x − y (cid:107) ) is empty of atoms from φ . Consequently, therelation x φ ↔ y holds iff, there are no atoms from φ inside C ( x, y ) . The Euclidean surface of C ( x, y ) is π (cid:107) x − y (cid:107) (2 − γ ) , where γ := − √ π is a constant number equal to thesurface, divided by π , of the intersection of two discs withunit radius and centres lying on the circumference of eachother. An illustration of the above explanations is given inFigure 1.When Φ = { φ } is a PPP, we have an expression of itsempty space function. With this in mind, along with theabove argument, it is possible to give a closed form to theprobability of two given atoms being in pair. Lemma 1.
Given a PPP Φ , with density λ > , for two differentand fixed atoms x, y ∈ R , P ( x Φ ↔ y ) = e − λπ (cid:107) x − y (cid:107) (2 − γ ) . For a stationary point process Φ , define two new pointprocesses Φ (1) and Φ (2) , that result from the dependentthinning defined above: Φ (1) := { x ∈ Φ & x is single } , Φ (2) := { x ∈ Φ & x cooperates with another element of Φ } . Both processes are stationary . This is due to the stationarityof Φ and because, by definition, they depend only on the dis-tance between the elements of Φ . From the previous Lemma,Slivniak’s Theorem and Campbell-Little-Mecke formula [10]we have the following result. Theorem 1.
Given a PPP Φ , with density λ > , for every fixedatom x ∈ R , there exists a constant δ > , independent of λ and x , such that P (cid:16) x ∈ Φ (2) (cid:17) = δ, P (cid:16) x ∈ Φ (1) (cid:17) = 1 − δ. Specifically, δ = − γ ≈ . .Proof. By definition of Φ (2) , P (cid:16) x ∈ Φ (2) (cid:17) = P (cid:16) x Φ ↔ y, for some y ∈ Φ \{ x } (cid:17) = E (cid:18) (cid:110) x Φ ↔ y, for some y ∈ Φ \{ x } (cid:111) (cid:19) ( a ) = E (cid:88) y ∈ Φ (cid:110) x Φ ↔ y (cid:111) , where ( a ) holds because for PPPs the nearest neighbor isa.s. unique. Using Campbell-Little-Mecke formula [10] for aPPP, E (cid:32) (cid:88) y ∈ Φ (cid:110) x Φ ↔ y (cid:111) (cid:33) = (cid:90) R P (cid:16) x Φ ↔ y (cid:17) λdy ( b ) = λ (cid:90) R e − λπ (cid:107) x − y (cid:107) (2 − γ ) dy = 12 − γ , where ( b ) follows from Lemma 1. Remark:
In Lemma 1 and Theorem 1 we actually makeuse of the Palm measures of the process, but avoid its nota-tion for ease of presentation, without substantial difference.
The constant δ is crucial within this work. The aboveTheorem states that, given the position of a BS (in a PPP),its probability of being in a cooperation pair is δ ≈ . ,otherwise, its probability of being single is − δ ≈ . ,irrespective of the value of the density λ > . Since we arefixing the atom location, this result should be interpretedfrom a local point of view. Nevertheless, in Section 3 wewill prove that, for a given density of the PPP λ > , theintensities of Φ (1) and Φ (2) are actually (1 − δ ) λ and δλ ,respectively. The former can be interpreted from a globalpoint of view: over any planar area in R , in average, . of atoms are singles and . belong to a cooperativepair.When Φ is a PPP, it is natural to wonder if Φ (1) and Φ (2) are also PPPs. As a matter of fact, they are not (we couldhave expected this, since they were defined by a stronglydependent thinning). Suppose that Φ (2) is actually a PPP.As shown in Theorem 1, for every atom in Φ (2) , there isa positive probability of this point not being in MNNRwith another point of Φ (2) . However, by definition, all theelements of Φ (2) are in MNNR with another element of Φ (2) ,which is a contradiction. We conclude that the process Φ (2) is not a PPP. For Φ (1) the argumentation is not as simple. Wecan show using the Kolmogorov-Smirnov test [32] that thenumber of Φ (1) atoms within a finite window is not Poissondistributed. Moreover, Monte Carlo simulations estimatethat the average proportion of single atoms from Φ (1) is farfrom the . .We can show that the percentages in Theorem 1 are notvalid just for PPPs. Take the hexagonal grid model. This iscommonly used by industry related research teams to modelthe BS positions, and then evaluate a system deploymentand performance via Monte Carlo methods. The hexagonalgrid’s centers should represent the BS locations. This is anideal scenario (the BSs are never that regular). We introduceanother point process, based on the hexagonal grid, thatactually allows for randomnes of the BS positions. Startingfrom the grid placement, let the position of each BS be ran-domly perturbed, independently of the others. For example,consider as BS location the point whose polar coordinatesaround each hexagon’s center follow two uniform randomvariables (r.v.s), one angular over [0 , π ) and the radialone over [0 , Q ] (see Figure 2). Figure 2 shows how theaverage percentage of singles and pairs for the hexagonalgrid model changes when varying the parameter Q > .Remark that these numbers are very close to the respectiveaverage percentages we found when Φ is a PPP. We can interpret the Palm probability of a stationary pointprocess as the conditional probability, given that the processhas a point inside an infinitesimal neighborhood around afixed atom [10]. Denote by P , P (1) , , and P (2) , the Palmprobabilities of the stationary point processes Φ , Φ (1) , and Φ (2) , respectively. Let A := { Φ ∈ A } , B := { Φ ∈ B } , be two events, where A := { φ | ∈ φ and is single } , B := { φ | ∈ φ and cooperates with another atom of φ } . (a) (b) Hexagonal grid model.PPP. (c)
Hexagonal grid model.PPP. (d)
Fig. 2: (a) The hexagonal grid model, without perturbation.(b) The hexagonal grid model, with the centers being per-turbed via a random experiment. (c) The average percentageof single atoms for the hexagonal grid model. (d) Theaverage percentage of atoms in cooperative pairs.We have the following result [33, pp. 35, Ex. 142].
Theorem 2.
Let Φ be a stationary point process such that P ( A ) > and P ( B ) > . Therefore, for every C ∈ Ω , P (1) , ( C ) = P ( C | A ) , P (2) , ( C ) = P ( C | B ) . When Φ is a PPP, P ( A ) = 1 − δ > and P ( B ) = δ > (Theorem 1). Then, for every C ∈ F , P (1) , ( C ) = P ( C, A )1 − δ , P (2) , ( C ) = P ( C, B ) δ . (1) Φ (2) The
Nearest Neighbor function (NN) , commonly denoted by G , is the cumulative distribution function (CDF) of thedistance from a typical atom of the process to its nearestneighboring point [34]. Denote by G (2) ( r ) the NN functionof Φ (2) , then, G (2) ( r ) = P (2) , ( d (0 , Φ (2) \ { } ) ≤ r ) , for every r > . Applying equation (1) to the above expres-sion, we have the following result. Theorem 3.
For a PPP Φ , the NN function of Φ (2) is G (2) ( r ) = 1 − e − λπr (2 − γ ) , (2) where γ is the same constant as in Lemma 1.Proof. Remark that under P A = (cid:91) y ∈ Φ \{ } { Φ ↔ y } (3) where A is the event from Theorem 2, and this union ismutually disjoint. For r > fixed, G (2) ( r ) = P (2) , (cid:16) d (0 , Φ (2) \{ } ) ≤ r (cid:17) ( a ) = P (cid:16) d (0 , Φ (2) \{ } ) ≤ r, A (cid:17) δ ( b ) = E (cid:88) y ∈ Φ \{ } { d (0 , Φ (2) \{ } ) ≤ r, Φ ↔ y } δ , ( c ) = E (cid:88) y ∈ Φ { d (0 , Φ (2) \{ } ) ≤ r, Φ ↔ y } δ , where ( a ) follows from equation (1), ( b ) after equation (3),and ( c ) from Slivkyak-Mecke’s Theorem. Observe that, ifthere is some y ∈ Φ being the mutually nearest neighbor ofthe atom , that is Φ ↔ y , then, d (0 , Φ (2) \{ } ) = d (0 , Φ \{ } ) = (cid:107) y (cid:107) a.s. Using this, Campbell-Little-Mecke formula and Lemma 1, E (cid:32) (cid:88) y ∈ Φ { d (0 , Φ (2) \{ } ) ≤ r, ↔ y } (cid:33) δ = E (cid:32) (cid:88) y ∈ Φ {(cid:107) y (cid:107)≤ r, Φ ↔ y } (cid:33) δ = E (cid:32) (cid:88) y ∈ Φ {(cid:107) y (cid:107)≤ r } { Φ ↔ y } (cid:33) δ = (cid:90) R E (cid:16) {(cid:107) y (cid:107)≤ r } { Φ ↔ y } (cid:17) λdy δ = (cid:90) R {(cid:107) y (cid:107)≤ r } E (cid:16) { Φ ↔ y } (cid:17) λdy δ = λδ (cid:90) {(cid:107) y (cid:107)≤ r } P (cid:0) Φ ↔ y (cid:1) dy = λδ (cid:90) {(cid:107) y (cid:107)≤ r } e − λπ (cid:107) y (cid:107) (2 − γ ) dy ( d ) = λ πδ (cid:90) r e − λπs (2 − γ ) sds = 1 − e − λπr (2 − γ ) , where ( d ) follows from the change of variable to polarcoordinates.The last Theorem simply states that, in the PPP case, thedistance between cooperative atoms is Rayleigh distributed,with scale parameter α := (2 λπ (2 − γ )) − / . It follows naturally to investigate the size of Voronoi cellsassociated with single atoms or pairs. A Voronoi cell ofan atom x ∈ φ is defined to be the geometric locus of allplanar points z ∈ R closer to this atom than to any otheratom of φ [35]. In a wireless network the Voronoi cell isimportant when answering the question ’which user shouldbe associated with which station?’.In a stationary framework, we examine the networkperformance at the Cartesian origin, the typical user approach .Let { (cid:121) Φ (1) } (resp. { (cid:121) Φ (2) } ) denote the event that the typical user belongs to the Voronoi cell of some atom of Φ (1) (resp. Φ (2) ). For the PPP case we have the following result. Proposition 1.
Suppose that Φ is a PPP, with density λ > .There exists a measurable function F : [0 , ∞ ) × [0 , ∞ ) × [0 , π ) × [0 , π ) −→ [0 , ∞ ) , such that P (0 (cid:121) Φ (2) )= λ (cid:90) ∞ (cid:90) ∞ (cid:90) π (cid:90) π sre − λF ( r,s,θ,ϕ ) − λπr dϕdθdrds Proof.
See Appendix A, available in the supplemental mate-rial.Since P (0 (cid:121) Φ (1) ) = 1 − P (0 (cid:121) Φ (2) ) , (4)we have also an analytic representation for P (0 (cid:121) Φ (1) ) .The function F ( r, s, θ, ϕ ) is not explicitly given, being theEuclidean surface of three overlapping discs. This is anexample of the complications that arise from the MNNR,due to numerical issues related to integration over multi-ple overlapping circles. Such complications led us to theapproximate model in Section 5. Numerical Result 1.
Given a PPP Φ , the average surfaceproportion of Voronoi cells associated with single atoms, and thatassociated with pairs of atoms, is independent of the parameter λ .By Monte Carlo simulations, we find that P (0 (cid:121) Φ (1) ) ≈ . , P (0 (cid:121) Φ (2) ) ≈ . . Interestingly, although the ratio of single atoms to pairsis . / . ≈ . , the ratio of the associatedVoronoi surface is . / . ≈ . , implying thatthe typical Voronoi cell of a single atom is larger than thatof an atom from a pair, as Figure 1 shows. The last remarkgives a first intuition that there is attraction between thecooperating atoms in pair and repulsion among the singles. The empty space function (ES) , commonly denoted by F , is theCDF of the distance from the typical user to the nearest atomof the point process considered [34]. The two functions NNand ES can be combined into a single expression known asthe J function . The latter is a tool introduced by van Lieshoutand Baddeley [34] to measure repulsion and/or attractionbetween the atoms of a point process. It is defined as J ( r ) = 1 − G ( r )1 − F ( r ) , (5)for every r > . In the case of the PPP, G ( r ) ≡ F ( r ) and J ( r ) = 1 , as a consequence of the fact that the reducedCampbell measure is identical to the original measure.Hence the J function quantifies the differences of anyprocess with the PPP. When J ( r ) > , this is an indicatorof repulsion between atoms, whereas J ( r ) < indicatesattraction. We use Monte Carlo simulations to plot the J function of both processes (see Figure 3). From the figureswe conclude that Φ (1) exhibits repulsion for every r ≥ , and Φ (2) attraction everywhere . However, note that the attractionin the case Φ (2) is due to the way the pairs were formed.If we consider a new process having as elements the mid-dle points between each one of the cooperating pairs, thisprocess exhibits repulsion everywhere. J−fun Singles (a)
J−fun Pairs (b)
Fig. 3: (a) The J function of the processes Φ (1) . (b) The Jfunction of the processes Φ (2) . ECEIVED S IGNALS
The analysis on this work can be applied to any typeof antennas, including directional ones. For simplicity ofpresentation we will treat here the omnidirectional case,where the emitted signal depends only on the distance of theBS from the typical user. The case of directional BSs requiresextra integration with respect to angles, which unnecessarilycomplicates the analysis, without substantial difference.In what follows, we will introduce explicit examples. Forthese, let us consider an i.i.d. family ( h r ) r> of positiveexponential variables, with parameter , also independentof the BS positions. Given p > , the couple ( h r , p ) repre-sents the random propagation effects and the power signalemitted to the typical user from a BS whose distance fromthe origin is r > . Let us also choose the path-loss functionas l ( r ) := r β , with path loss exponent β > . Consider f : R → [0 , ∞ ) and ˜ f : [0 , ∞ ) −→ R + two generic random fields. The quantity f ( x ) (and ˜ f ( r ) )represents the received signal at the typical user, whentransmitted by a single BS, whose position is x (or itsdistance from the origin is r > ). For a single BS we could,for example, consider ˜ f ( r ) = p h r r β , (6)which follows an exponential distribution, with parameter r β p . Consider g : R × R −→ R + and ˜ g : [0 , ∞ ) × [0 , ∞ ) −→ R + two generic random fields, both independent of the BSpositions. The quantity g ( x, y ) (and ˜ g ( r, z ) ) represents thereceived signal at the typical user, when transmitted by apair of BSs whose positions are x and y (or their distancefrom the origin are r > and z > ), respectively. Thereceived signal can take the following example expressions,which refer to different types of cooperation or coordination, ˜ g ( r, z ) = p h r r β + p h z z β , [NSC] on r p h r r β + (1 − on r ) p h z z β , [OFF] max (cid:110) p h r r β , p h z z β (cid:111) , [MAX] (cid:12)(cid:12)(cid:12)(cid:12)(cid:113) p h r r β e iθ r + (cid:113) p h z z β e iθ z (cid:12)(cid:12)(cid:12)(cid:12) [PH] . (7) TABLE 1: Expressions for the CCDF and the LT P ( g ( r, z ) > T ) E [ e − sg ( r,z ) ] [NSC] z β p ( r β − z β ) (cid:16) e − rβp T − e − zβp T (cid:17) r β sp + r β z β sp + z β [OFF] qe − rβp T + qe − zβp T q r β sp + r β + q z β sp + z β [MAX] e − rβp T + e − zβp T − e − (cid:18) rβp + zβp (cid:19) T r β sp + r β + z β sp + z β − r β + z β sp + r β + z β In the above, ( on r ) r> and ( θ r ) r> are two different fam-ilies of indexed identically distributed r.v.s, independentof the other random objects. They follow a Bernoulli dis-tribution, with parameter q ∈ (0 , ( q := 1 − q ), and ageneral distribution over [0 , π ) , respectively. [NSC] refersto non-coherent joint transmission, as in [12], [13], [14], [20],where each of the two BSs transmits an orthogonal signal,and the two are added at the receiver side. [OFF] refers tothe case where one of the two BSs is active and the otherinactive, according to an independent Bernoulli experiment,independent of the BS positions. [MAX] refers to the casewhere the BS with the strongest signal is actively serving auser, while the other is off. The [OFF] and [MAX] cases arerelevant to energy saving operation. In the [PH] case, twocomplex signals are combined in phase (see [11], [12]), inparticular, when cos ( θ r − θ z ) = 1 , the two signals are in thesame direction, and they add up coherently at the receiver(user side), giving the maximum cooperating signal.The above expressions in (7) are merely examples of thecooperation signals. A more general family can be proposedwith specific properties. Consider c i : [0 , ∞ ) × [0 , ∞ ) −→ R ,and d i : [0 , ∞ ) × [0 , ∞ ) −→ R + , for ≤ i ≤ n , somedeterministic and measurable functions, and suppose that P (˜ g ( r, z ) > T ) = n (cid:88) i =1 c i ( r, z ) e − d i ( r,z ) T . (8)When analysing performance related to coverage probabil-ity, the tail probability distribution functions (CCDF) for thesignals that can be written as (8) lead easier to numericallytractable formulas. However, the function defined in (8) isnot necessarily a CCDF. For this to hold, some extra condi-tions must be imposed to the functions c i ( r, z ) and d i ( r, z ) .Interestingly, the CCDF of ˜ g ( r, z ) in the [NSC] , [OFF] , and [MAX] cases fulfils equation (8) (see Table 1). Furthermore,there exist important families of r.v.s whose CCDF actuallyhas the form described in equation (8): the hypo-exponentialdistribution , the hyper-exponential distribution , the maximumover a finite number of exponential r.v.s, among others. NTERFERENCE A NALYSIS IN THE
MNNR
MODEL
The purpose of the analysis up to this point was to developthe basic tools, within a communication context, that willallow us to derive results related to cooperation. As shownin the previous Section, the cooperating BS pairs will havea different influence on the interference seen by a user inthe network, than those operating individually. The currentSection will focus on the interference field generated by Φ (1) and Φ (2) . As shown in Section 2, even when Φ is a PPP, the two processes behave differently than a PPP. We thus haveto resort to direct techniques from the theory of StochasticGeometry and point processes.If we denote by I (1) and I (2) , the interference fieldgenerated by the elements of Φ (1) and Φ (2) , then, I (1) = (cid:88) x ∈ Φ (1) f ( x ) , (9) I (2) = 12 (cid:88) x ∈ Φ (2) (cid:88) y ∈ Φ (2) \{ x } g ( x, y ) (cid:110) x Φ ↔ y (cid:111) (10)The / in front of the summation in (10) prevents us fromconsidering a pair twice. Remark that we can consider hereany type of signal (directional or not). I (1) and I (2) . The next Theorem gives an exact integral expression to theexpected value of the interference field generated by thesingles and the pairs. The proof uses the Campbell-Little-Mecke formula, Lemma 1, and Theorem 1.
Theorem 4.
For a PPP Φ , the expected value of the interferencefield generated by Φ (1) and Φ (2) is given by E (cid:104) I (1) (cid:105) = (1 − δ ) (cid:90) R E [ f ( x )] λdx, (11) E (cid:104) I (2) (cid:105) = 12 (cid:90) R (cid:90) R E [ g ( x, y )] e − λπ | x − y | (2 − γ ) λdyλdx. (12) Proof.
Let us start with I (1) . We observe that, because thenearest neighbor always exists and is unique, I (1) = (cid:88) x ∈ Φ (1) f ( x )= (cid:88) x ∈ Φ f ( x ) { x ∈ Φ (1) } = (cid:88) x ∈ Φ f ( x ) (cid:16) − { x ∈ Φ (2) } (cid:17) Thus, after applying the reduced Campbell-Little-Meckeformula and Slivnyak-Mecke’s Theorem E (cid:104) I (1) (cid:105) = E (cid:34)(cid:88) x ∈ Φ f ( x ) (cid:16) − { x ∈ Φ (2) } (cid:17)(cid:35) ( a ) = (cid:90) R E [ f ( x )] (cid:16) − P ( x ∈ Φ (2) ) (cid:17) λdx ( b ) = (1 − δ ) (cid:90) R E [ f ( x )] λdx, where ( a ) follows because f ( x ) is independent of Φ and ( b ) after Theorem 1. Then we have the desired result for I (1) .For I (2) , we make the observation that (cid:88) x ∈ Φ (2) (cid:88) y ∈ Φ (2) \{ x } g ( x, y ) { x Φ ↔ y } = (cid:88) x ∈ Φ (cid:88) y ∈ Φ \{ x } g ( x, y ) { x Φ ↔ y } , and iterating the reduced Campbell-Little-Mecke formulaand Slivnyak-Mecke’s Theorem, E (cid:104) I (2) (cid:105) = E (cid:88) x ∈ Φ (cid:88) y ∈ Φ \{ x } g ( x, y ) { x Φ ↔ y } = (cid:90) R (cid:90) R E (cid:104) g ( x, y ) { x Φ ↔ y } (cid:105) λdyλdx ( c ) = (cid:90) R (cid:90) R E [ g ( x, y )] P (cid:16) x Φ ↔ y (cid:17) λdyλdx ( d ) = (cid:90) R (cid:90) R E [ g ( x, y )] e − λπ (cid:107) x − y (cid:107) (2 − γ ) λdyλdx, where ( c ) follows because g ( x, y ) is independent of Φ and ( d ) after Lemma 1.The expected value can be finite or infinite, dependingon the choice of f ( x ) and g ( x, y ) . Observe that for [NSC]and [PH] the expected interference has the same value. Corollary 1.
For a PPP Φ , let M (1) and M (2) be the intensitymeasures of Φ (1) and Φ (2) , respectively. Then, M (1) ( dx ) = (1 − δ ) λdx,M (2) ( dx ) = δλdx. Proof.
Let A be a regular subset of R . For the choice f ( x ) = xA (and g ( x, y ) = xA ), the random variable I (1) (and I (2) ) counts the number of singles (pairs) within A .Applying directly the preceeding Theorem, and remarkingthat (cid:82) R e − λπ (cid:107) x − y (cid:107) (2 − γ ) λdy = δ , for every x ∈ R , we havethe desired result.The previous Corollary states that the intensities of Φ (1) and Φ (2) are (1 − δ ) λ and δλ , as stated in Section 2. Φ (1) and Φ (2) As a final discussion in this Section, we present our findingsrelated to the LT of the interference from Φ (1) and Φ (2) ,when Φ is a PPP. Fix a measurable set A ⊂ R (window).Recall that Φ( A ) denotes all the atoms of Φ inside A . Wedefine the point processes Φ (1) A = (cid:8) single atoms of Φ( A ) (cid:9) Φ (2) A = (cid:8) atoms of Φ( A ) in MNNR (cid:9) (13)where the MNNR have been considered only among the Φ atoms inside A . Consider a sequence of finite windows ( A n ) ∞ n =1 increasing to R in an appropriate sense (for exam-ple, A n = B ( n, ). We have the following result. Theorem 5.
Given a PPP Φ , then, for i = 1 , , lim n →∞ Φ ( i ) A n ( d ) = Φ ( i ) , where ( d ) = means equality in distribution.Proof. See Appendix B, available in the supplemental mate-rial.As convergence in distribution is equivalent to conver-gence of the LT [31], the previous Theorem states that, for A large enough, the LT of Φ ( i ) A approximates that one of Φ ( i ) . The benefit of this approach is that, for every finite window A , we can actually obtain an analytic representationfor the LT of Φ ( i ) A . As a sketch of the proof, fix a finitesubset A . Conditioned on the number of atoms, these arei.i.d. uniformly distributed within A . Then, using the law oftotal probability, we can express the LT as an infinite sumof terms. The probability of a PPP having a fixed number ofatoms within A is known. Thus, we only have left to findexpressions for the LT conditioned on the number of pointsinside A . For a finite number of different planar points x , . . . , x n ∈ A , define the function H i,n ( x , . . . , x n ) as theindicator function of the atom x i being in pair with anotheratom of the finite configuration { x , . . . , x n } (Definition 1).In the same fashion, define the function I i,n ( x , . . . , x n ) as the indicator function of the atom x i being single withrespect to the finite configuration { x , . . . , x n } (Definition2). Let H ( n ) ( x , . . . , x n ):= ( H (1 ,n ) ( x , . . . , x n ) , . . . , H ( n,n ) ( x , . . . , x n )) ,I ( n ) ( x , . . . , x n ):= ( I (1 ,n ) ( x , . . . , x n ) , . . . , I ( n,n ) ( x , . . . , x n )) . We have the following result.
Theorem 6 (Laplace transform) . Consider a PPP Φ , withintensity λ , a regular subset A ⊂ R , and a measurable function f : R −→ R + . Let F ( n ) ( x , . . . , x n ) := ( f ( x ) , . . . , f ( x n )) .The LT of Φ (1) A is equal to E (cid:18) e − (cid:80) x ∈ Φ(1) A f ( x ) (cid:19) = e − λ S ( A ) (cid:32) λ (cid:90) A e − f ( x ) dx + λ ∞ (cid:88) n =3 λ n n ! (cid:90) A . . . (cid:90) A e − F ( n ) ( x ,...,x n ) · H ( n ) ( x ,...,x n ) dx . . . dx n (cid:33) (14) The LT of Φ (2) A is equal to E (cid:18) e − (cid:80) x ∈ Φ(2) A f ( x ) (cid:19) = e − λ S ( A ) (cid:32) λ S ( A ) + λ (cid:90) A (cid:90) A e − ( f ( x )+ f ( y )) dydx + ∞ (cid:88) n =3 λ n n ! (cid:90) A . . . (cid:90) A e − F ( n ) ( x ,...,x n ) · I ( n ) ( x ,...,x n ) dx . . . dx n (cid:33) (15) Implementation:
The MNNR was defined in a general way(see Section 2). Then, for every natural number n , it is easyto write a program/algorithm with input ( x , . . . , x n ) andoutput H ( n ) ( x , . . . , x n ) (or I ( n ) ( x , . . . , x n ) ):1) Define a n × n matrix D = ( d i,j ) , such that d i,j = (cid:107) x i − x j (cid:107) .
2) Choose a n × vector v , such that, for each i =1 , . . . , n , v ( i ) = argmin j ∈{ ,...,n }\{ i } d i,j .
3) Define another n × vector u , such that, for each i = 1 , . . . , n • If i = v ( v ( i )) (that is, x i and x v ( i ) are inMNNR), then u ( i ) = 1 . • Else u ( i ) = 0 .4) Return u .Fixed f ( x ) , F ( n ) ( x , . . . , x n ) is also known. Thus, for everynatural number n , it is easy to set up a program thatnumerically approaches (cid:90) A . . . (cid:90) A e − F ( n ) ( x ,...,x n ) · H ( n ) ( x ,...,x n ) dx . . . dx n and (cid:90) A . . . (cid:90) A e − F ( n ) ( x ,...,x n ) · I ( n ) ( x ,...,x n ) dx . . . dx n . However, it is clear that, as n grows, computational time for H ( n ) ( x , . . . , x n ) and I ( n ) ( x , . . . , x n ) naturally increases.Since n -nested integral involving these functions needs tobe computed, complicating the problem even more. As partof the current research, the authors work on the complexityreduction of the MNNR algorightm, and investigate the rateof convergence for the expressions in equations (14) and (15)(to avoid calculating an infinite number of terms). HE S UPERPOSITION M ODEL - C
OVERAGE A NALYSIS
As a consequence of the non-Poissonian behaviour of Φ (1) and Φ (2) , a complete performance analysis of SINR relatedmetrics is analytically challenging. This is due to the factthat the expressions presented for the LT are not numericallytractable. Thus, one cannot derive simple, analytic expres-sions for the coverage probability by LT methods, as shownin [36]. Instead, we use in this work the following model toapproximate these metrics. To imitate the process of singles, we consider a PPP ˆΦ (1) ,with parameter (1 − δ ) λ . In this way, the new process ofsingles and Φ (1) share the first moment (Corollary 1).To imitate the process of pairs, we also consider a PPP ˆΦ (2) , independent of ˆΦ (1) , with intensity δ λ . We call theatoms of this process the parents . We considere the process ˆΦ (2) as independently marked. Each mark of a parent repre-sents its pairing BS, the daughter . The idea is that each couple ( parent , daughter ) imitates a cooperating pair in MNNR. Letus consider ( Z r ) r> a family of independent, real r.v.s,independent also of ˆΦ (1) and of ˆΦ (2) , where each Z r followsa Rice distribution, with parameter ( r, α ) . If Y is a randomvector representing the Cartesian coordiantes of a parent,we define its mark by Z (cid:107) Y (cid:107) .To understand the choice for the marks, suppose that aBS is placed at the polar coordinates ( r, θ ) , with r > and θ ∈ [0 , π ) fixed (see Figure 4). Assume also that this BSbelongs to a cooperating pair from the Nearest Neighbormodel, and let us denote by W the distance between thestations in pair. According to Theorem 3, W is Rayleigh dis-tributed, with scale parameter α . If Z denotes the distance Fig. 4: Two cooperating BSs, where r and Z r are theirdistances from the origin, and W is the distance betweenthem.from the typical user to the second BS, the isotropy of thePPP implies that the distribution of Z is independent of θ .Moreover, we have the following result. Proposition 2.
The r.v. Z is Rice distributed, with parameters ( r, α ) . The probability density function (PDF) of Z is given by f ( z | r ) = zα e − z r α I (cid:16) zrα (cid:17) , (16) where I ( x ) is the modified Bessel function, of the first kind, withorder zero. The angular coordinate of a PPP atom is uniformlydistributed in [0 , π ) . Moreover, the Cartesian coordinatesof a point around a center, with Rayleigh radial distance anduniform angle, are distributed as an independent Gaussianvector. Given this, Proposition 2 follows from [37, Lem. 1]. Let R and R denote the r.v.s of the distances from theclosest element of ˆΦ (1) and ˆΦ (2) to the origin, respectively.Denote also by Z the mark of the parent at R . It is knownthat the r.v.s R and R are Rayleigh distributed [36], withscale parameters ξ and ζ , where ξ := ((1 − δ )2 λπ ) − / and ζ := ( δλπ ) − / . By definition, R and Z are not mutuallyindependent, but we can derive their joint PDF. Lemma 2.
The joint PDF of the r.v. ( R , Z ) is given by f ( r, z ) = rz ( αζ ) e − r (cid:16) α + ζ (cid:17) − − z ζ I (cid:18) rzζ (cid:19) . (17) Furthermore, the r.v. Z is Rayleigh distributed, with scale pa-rameter ( α + ζ ) − / .Proof. See Appendix C, available in the supplemental mate-rial.To make a complete analysis of the coverage probability,we make use of the distribution of the random vector ( R , R , Z ) . Because R is independent of ( R , Z ) , thejoint PDF is the product of the PDF of R with the jointPDF of ( R , Z ) . It is clear from the definition of the marks of Φ (2) that, forthe superposition model, we deal only with onmidirectionalBSs. For r > , denote by L ˜ f ( s ; r ) := E (cid:104) e − s ˜ f ( r ) (cid:105) , (18a) L ˜ g ( s ; r, ρ ) := E (cid:104) e − s ˜ g ( r,Z r ) { Z r >ρ } (cid:105) , (18b) the LT of the signal generated by a single BS, and the LTof the signal generated by a cooperation pair, given that theradius of the daughter is larger than ρ ≥ . When ρ = 0 , L g ( s ; r, will be denoted just by L g ( s ; r ) . For example, ifwe take f ( r ) as in equation (6), we get L f ( s ; r ) = r β sp + r β . (19)Recall that p h r r β and p h z z β are independent, exponential r.v.s,with parameter r β p and z β p . In Table 1 we find expressionsfor E [ e − s ˜ g ( r,z ) ] in the [NSC] , [OFF] , and [MAX] cases. Byremarking that L ˜ g ( s ; r, ρ ) = E (cid:104) E (cid:104) e − s ˜ g ( r,Z r ) (cid:12)(cid:12)(cid:12) Z r (cid:105) { Z r >ρ } (cid:105) , we get analytical expressions for L ˜ g ( s ; r ) in the [NSC] , [OFF] , and [MAX] . For example, in the [NSC] we have that L ˜ g ( s ; r, ρ ) = r β sp + r β (cid:90) ∞ ρ z β sp + z β f ( z | r ) dr, where f ( z | r ) is the density function of the Rice r.v. Z r (seeequation (16)). For the more general distribution describedby equation (8), it is also possible to give analytical formulassimilar to L ˜ g ( s ; r, ρ ) . The [PH] case is more complicated(see [11, Lem. 3] for cos ( θ r − θ z ) = 1 ).We consider the interference fields generated by all theelements of ˆΦ (1) and ˆΦ (2) outside the radius ρ ≥ I (1) ( ρ ) = (cid:88) x ∈ ˆΦ (1) , (cid:107) x (cid:107) >ρ ˜ f ( (cid:107) x (cid:107) ) , (20a) ˆ I (2) ( ρ ) = (cid:88) y ∈ ˆΦ(2) (cid:107) y (cid:107) >ρ,Z (cid:107) y (cid:107) >ρ ˜ g ( (cid:107) y (cid:107) , Z (cid:107) y (cid:107) ) . (20b)When ρ = 0 , they are just denoted by ˆ I (1) and ˆ I (2) . Thetotal interference generated outside possibly different radiifor the two processes, i.e. ρ ≥ and ρ ≥ is ˆ I ( ρ , ρ ) := ˆ I (1) ( ρ ) + ˆ I (2) ( ρ ) . (21)When ρ = ρ = 0 , we write only ˆ I .The next Lemma is a well known result giving analyticalrepresentations to the LT of the PPP Interference fields [10]. Lemma 3.
The LTs of ˆ I (1) ( ρ ) and ˆ I (2) ( ρ ) , denoted by L ˆ I (1) ( s ; ρ ) and L ˆ I (2) ( s ; ρ ) , are given by L ˆ I (1) ( s ; ρ ) = e − λ π (1 − δ ) (cid:82) ∞ ρ (1 −L f ( s ; r )) rdr , (22a) L ˆ I (2) ( s ; ρ ) = e − πλδ (cid:82) ∞ ρ (1 −L g ( s ; r,ρ )) rdr . (22b)The Lemma uses the Poisson properties of ˆΦ (1) and ˆΦ (2) . The expressions given in equations (22) are the toolswhich allow us to make an entire analysis of the coverageprobability.As an example, if we replace equation (19) in equation(22a), for ρ = 0 we get the analytical representation [36] L ˆ I (1) ( s ) = e − λ (1 − δ )2 π sp )2 /ββ csc ( πβ ) , (23)where csc ( z ) is the cosecant function. In the same fash-ion, it is possible to obtain expressions for L ˆ I (1) ( s ; ρ ) and L ˆ I (2) ( s ; ρ ) . We can now make use of the PPP superposition model toevaluate the performance of the different cooperation (orcoordination) types proposed above. The beneficial signal,received at the typical user from a single BS or a pair, willbe denoted by ˜ f ( r ) and ˜ g ( r, z ) , respectively. These may notbe the same functions modeling the interference the typicaluser receives from other BSs. This is explained by the factthat the interference is the sum of the signals other BSsgenerate for their own serving users who are not locatedat the Cartesian origin.We consider two scenarios of user-to-BS association: Let us suppose that there is one BS serving the typical user,whose distance to the origin is fixed and known r > .Moreover, it serves the typical user independently of theatoms from ˆΦ (1) and ˆΦ (2) . Then the signal emitted to thetypical user is ˜ f ( r ) , and the Signal-to-Interference-plus-Noise-Ratio (SINR) at the typical user is defined by SINR := ˜ f ( r ) σ + ˆ I , (24)where σ is the additive Gaussian noise power at the re-ceiver and ˆ I is the total interference power (see equation(21)). Proposition 3.
Suppose ˜ f ( r ) as in (6) . Then, the successprobability is given by the expression P (SINR > T ) = e − Tσ rβ p L ˆ I (1) (cid:32) T r β p (cid:33) L ˆ I (2) (cid:32) T r β p (cid:33) . (25)The last proposition allows us to evaluate the SINRdirectly with the help of equations (22a) and (22b) for ρ = 0 . ˆΦ (1) or ˆΦ (2) (and hisdaughter) We consider that the typical user is connected to the BSat R (see subsection 5.2), or to the cooperating cluster(parent,daughter) at ( R , Z ) . The previous association de-pends on which one of them is closer to the typical user. If R < min { R , Z } , the single BS at R serves the typicaluser, and it emits the signal ˜ f ( R ) . In the opposite case, if R ≤ R or if Z ≤ R , the cooperating pair at ( R , Z ) serves the user, and it emits the signal ˜ g ( R , Z ) . All the BSsnot serving the typical user generate interference. Thus, SINR := ˜ f ( R ) σ +ˆ I ( R ,R ) ; R < min { R , Z } , ˜ g ( R ,Z ) σ +ˆ I ( R ,R ) ; R < min { R , Z } , ˜ g ( R ,Z ) σ +ˆ I ( Z ,R ) ; Z < min { R , R } . (26)From equation (20b), recall that once a parent generatesinterference, its respective daughter does it along with it. Forthe first term of the preceding equation, ˆ I ( R , R ) considersthat all the singles and parents lying outside R generateinterference. For the the second term we use a similarargument. For the third one, the argument is a little bit moredelicate. The r.v. ˆ I ( Z , R ) considers that all the singles lie outside the radius Z , and all of them generate interference.Nevertheless, only the parents outside R generate interfer-ence (the parent associated to R lies outside Z ). Note that,the way this user-to-BS-association is defined, for the threecases, this is the only way to assure that all the BSs notserving the typical user generate interference. Proposition 4.
Suppose f ( r ) and g ( r, z ) follow equations (6) and (8) . Then, there exist explicit functions G : [0 , ∞ ) → R + and H, K : [0 , ∞ ) × [0 , ∞ ) → R + such that P (SINR > T ) = E [ G ( R )] + E [ H ( R , Z )] + E [ K ( R , Z )] . Proof.
See Appendix D, available in the supplemental mate-rial.The expressions for G ( r ) , H ( r, z ) , and K ( r, z ) are givenby G ( r ) = ˜ G ( r ) ˆ G ( r ) ,H ( r, z ) = ˜ H ( r, z ) ˆ H ( r, z ) ,K ( r, z ) = ˜ K ( r, z ) ˆ K ( r, z ) , where ˜ G ( r ) := 1 − F R ( r ) − F Z ( r ) + F R ,Z ( r, r ) , ˜ H ( r, z ) := (1 − F R ( r )) { z>r } , ˜ K ( r, z ) := (1 − F R ( z )) { r>z } , and ˆ G ( r ) := e − Trβp σ L ˆ I (1) (cid:32) T r β p ; r (cid:33) L ˆ I (2) (cid:32) T r β p ; r (cid:33) , ˆ H ( r, z ) := n (cid:88) i =1 c i (cid:0) r, z (cid:1) e − T d i ( r,z ) σ L ˆ I (1) (cid:0) T d i ( r, z ); r (cid:1) L ˆ I (2) (cid:0) T d i ( r, z ); r (cid:1) ˆ K ( r, z ) := n (cid:88) i =1 c i (cid:0) r, z (cid:1) e − T d i ( r,z ) σ L ˆ I (1) (cid:0) T d i ( r, z ); z (cid:1) L ˆ I (2) (cid:0) T d i ( r, z ); r (cid:1) . We can find expressions for the deterministic functions L ˆ I (1) ( s ; ρ ) and L ˆ I (2) ( s ; ρ ) in equation (22), and the functions c i ( r, z ) and d i ( r, z ) are those from equation (8). Finally, thefunctions F R ( r ) , F R ( r ) , and F Z ( r ) are the CDF of the ran-dom variables R , R , Z , which are Rayleigh distributed(see Section 5.2), and F R ,Z ( r, z ) is the CDF of the randomvector ( R , Z ) , which can be obtained with equation (17). Remark: we can either calculate the expression of thecoverage probability from equation (4) via Monte Carlo sim-ulations (because we know the distribution of R , R , and ( R , Z ) ), or via numerical integration, using the formulas E [ G ( R )] = (cid:90) ∞ G ( r ) f R ( r ) dr E [ H ( R , Z )] = (cid:90) ∞ (cid:90) ∞ H ( r, z ) f R ,Z ( r, z ) dzdr E [ K ( R , Z )] = (cid:90) ∞ (cid:90) ∞ K ( r, z ) f R ,Z ( r, z ) dzdr, where f R ( r ) and f R ,Z ( r, z ) are the density functions of R and ( R , Z ) (again, see equation (17)). UMERICAL E VALUATION
We consider a density for the BSs λ = 0 . [ km ], whichcorresponds to an average closest distance of (2 √ λ ) − = 1 [km] between stations. We also consider that the power is p = 1 [ W att ].We first illustrate the validity of the expressions in Theo-rem 4. Specifically, we compare the expressions in equations(11) and (12) with simulations.
Given a fixed β > , define the random field f ( x ) = h (cid:107) x (cid:107) (cid:107) x (cid:107) β {(cid:107) x (cid:107) >R } , where R is a positive number and the fam-ily ( h r ) r> is defined in Section 3. The indicator functionserves to calculate the interference generated by the singles,outside a ball centred at and radius R . With the aid of f ( x ) , define I (1) as in equation (9). Using Theorem 4, thenumerical evaluation of the expected value of I (1) is givenin figure 5. The expression in (11) gives almost identicalresults with the simulations.Similarly, with the aid of the random field g ( x, y ) {(cid:107) x (cid:107) , (cid:107) y (cid:107) >R } , define I (2) as in equation (10).For the numerical evaluation, we consider the two cases [NSC] and [MAX] . The interference from [MAX] is alwayssmaller than that one from [NSC] , since it is received onlyfrom one of the two BSs of each pair, while the other issilent. Figure 5 shows that the numerical evaluation of theexpression in (12) gives almost identical results with thesimulations. Remark also that, for β = 4 , the two scenariosdo not numerically defer much.For the coverage probability analysis, we evaluate onlythe noiseless scenario P (SIR > T ) (with σ = 0 ). Wecompare the SIR coverage performance from the NearestNeighbor and the superposition models against the modelwithout cooperation [36]. We consider both cases ( a ) withfixed transmitter, and ( b ) where the association is done withthe (almost) closest cluster, as in (26). In this second case, forthe Nearest Neighbor model, the user-cluster association isdone differently than in the superposition model, as follows.The typical user is served by the closest BS of the originalpoint process Φ , and by its mutually nearest neighbor, if oneexists. The cooperative signals are those proposed in (7). We compare in Fig. 6 the coverage probabilities, over thethreshold T , for the Nearest Neighbor model and the super-position model, in both association cases. As we can see,the curves are very close in both cases. For the ”closest”transmission cluster the difference is more evident, becauseon the one hand the superposition model does not take intoaccount the repulsion between clusters (singles or pairs),and on the other hand the association of a cluster to theuser as done in (26) for the superposition model, sometimesmisses the actual closest daughter to the origin (which isnot necessarily the one at Z ). This never happens theway we choose the closest cluster in the Nearest Neighbormodel. Hence, the approximative model underestimates thecoverage benefits in the closest cluster association. Radius R E [ I ] β =2.5, num β =2.5, sym β =4, num β =4, sim (a)(b) Fig. 5: (a) Expected value of the interference generatedby the single atoms, outside a radius R , (b) and by thecooperative pairs, outside a radius R . −20 −15 −10 −5 0 5 10 15 2000.10.20.30.40.50.60.70.80.91 Threshold T C o ve r a g e p r ob a b ili t y NSC, NN modelNSC, superpositionOFF, NN modelOFF, superposition (a) −20 −15 −10 −5 0 5 10 15 2000.10.20.30.40.50.60.70.80.91
Threshold T C o ve r a g e p r ob a b ili t y OFF, superpositionOFF, NN modelNSC, superpositionNSC, NN model (b)
Fig. 6: Closeness of the approximation between the super-position and the Nearest Neighbor model, β = 3 . (a) Fixedtransmitter and (b) closest transmitter. In Figure 7, we compare the plots of the coverage probabilityfrom the numerical integration, against simulations, of theanalytic formula presented in Proposition 4. As we can see,they fit perfectly, both for larger values of β , like β = 4 , andfor critical ones, like β = 2 . . −20 −15 −10 −5 0 5 10 15 2000.10.20.30.40.50.60.70.80.91 Threshold T C o ve r a g e p r ob a b ili t y NSC, analyticalNSC, simulationsOFF, analyticalOFF, simulations (a) −20 −15 −10 −5 0 5 10 15 2000.10.20.30.40.50.60.70.80.91
Threshold T C o ve r a g e P r ob a b ili t y NSC, analyticalNSC, simulationsOFF, analyticalOFF, simulations (b)
Fig. 7: Validity of the analysis for the superposition modelfor the fixed single transmitter. (a) β = 2 . (b) β = 4 . The possible coverage gains, compared to the non-cooperative network, in the case of an association with afixed transmitter, are shown in Fig. 8(a). As a first remark,for the fixed association, the [NSC] case for the NearestNeighbor model and the non-cooperative model are practi-cally the same. The coverage probability in the [MAX] caseis close to the coverage probability in the [NSC] case. Thissuggests that the strongest signal in each cooperating pairinfluences interference the most. For the [OFF] case there isa benefit compared to the non-cooperative case, in thelargest part of the domain in T .The gains are also evaluated in the case of associationwith the closest cluster. For the SINR, let us call [MAX/OFF] the case where the closest cluster emits a signal to the typicaluser according to [MAX] , i.e. only the max signal is sent,while the pairs generate interference, according to [OFF] .The idea is that when all network pairs choose [MAX] cooperation for their own users, this choice of one-station-out-of-two is random for the typical user point of view. This [MAX/OFF] case shows a absolute gain from the noncooperative case, which is around for the [NSC] (see Fig. −20 −15 −10 −5 0 5 10 15 2000.10.20.30.40.50.60.70.80.91 Threshold T C o ve r a g e p r ob a b ili t y Non−cooperative modelNSC cooperationOFF cooperationMAX cooperation (a) −20 −15 −10 −5 0 5 10 15 2000.10.20.30.40.50.60.70.80.91
Threshold T C o ve r a g e p r ob a b ili t y Non−cooperative modelNSC cooperationOF2/OF1 cooperation (b)
Fig. 8: Coverage gains from Nearest Neighbor cooperationcompared to no cooperation, β = 3 . (a) Fixed transmitterand (b) closest transmitter.8(b)). This gain is almost equal with the dynamic clusteringin [11]. ODEL B ENEFITS AND F UTURE R ESEARCH
The static grouping model presented in this paper has thefollowing network benefits. • By definition, the MNNR reduces the generated in-terference. • The percentage of stations that are in cooperativepairs and the percentage of those left single (both forthe PPP and the hexagonal grid model) show that theMNNR is a reasonable grouping strategy. • Our approach can be applied to many cooperationvariations, ranging from simple coordination of theBSs in group, to fully cooperative transmission usingknowledge of the channel states.The mathematical innovations are the following. • The derivation of structural properties for Φ (1) and Φ (2) by classical Stochastic Geometry tools comesnaturally. • Two repulsive point processes can be constructed ina natural way. • Based on simple geometrical concepts, the MNNRcan be easily implemented in any programming lan-guage. This simplifies the numerical evaluation ofthe system. • The superposition approach makes possible a com-plete analysis of the coverage probability, which canbe extended to other performance metric, such as thethroughput. • The superposition is equal in distribution to a Gauss-Poisson process. However, the mark related to eachpair is chosen in an original way that allows thederivation of simple formulas, that cannot be ob-tained by directly using a Gauss-Poisson.As stated in Theorem 5, the resulting edge effects ofthe MNNR are negligible for a PPP. This makes possible afinite window analysis for the interference generated by thesingles and pairs. Part of the current research of the authorsis the convergence rate for this result, as well as bounds andthe respective convergence rate for Theorem 6.The static cluster methodology proposed is based onthe Euclidean distance between BSs, and fixes groups ofsingles and pairs over time. This approach doest not allowfor flexibility in the way the groups are created. We couldimagine introducing different methods, for the formation ofthe desirable clusters, that take into account the availabilityof each BS as well as the Euclidean disatnce among them.Analysis and applications of this type of cooperation can beconsidered as future extensions of the MNNR.
ONCLUSIONS
The MNNR is a reasonable methodology to define single BSsand cooperative pairs. In spite of the analytical difficulties(due to overlapping discs with different radii), it is possibleto use an approximate model and provide a complete anal-ysis of SINR-related metrics. Moreover, it could be possibleto generalise the idea and include clusters of size greaterthan two. The analysis should remain the same and similarresults could be derived via Monte Carlo simulations.The coverage benefits of the MNNR, with respect to thenon-cooperative case, can reach a of absolute gain,although around of stations is single and do not co-operate. Similar gains can be achieved by some dynamicclustering methodologies. For the MNNR, this is impressive,considering that only of the BSs cooperate. Differentkinds of cooperative signals reported different coveragebenefits. Thus, cooperation benefits fundamentally dependon the choice of the grouping method, the allowed max-imum cluster size, as well as the appropriate cooperationsignals. This works provides an important step towardsresolving this complex problem. A CKNOWLEDGMENT
The authors would like to thank Aur´elien Vasseur for thehelp on the proof of Theorem 5, and for the suggestedrelated literature. R EFERENCES [1] R. Irmer, H. Droste, P. Marsch, S. Brueck, H. P. Mayer, L. Thiele,and V. Jungnickel. Coordinated Multipoint: Concepts, Perfor-mance, and Field Trials Results.
IEEE Communications Magazine ,49:102–111, 2011. [2] V. Jungnickel, K. Manolakis, W. Zirwas, B. Panzner, V. Braun,M. Lossow, M. Sternad, Apelfr¨ojd R., and T. Svensson. The Roleof Small Cells, Coordinated Multipoint, and Massive MIMO in 5g.
IEEE Communications Magazine , 52(44-51), 2014.[3] V. Jungnickel, S. Jaeckel, L. Thiele, L. Jiang, U. Kr ¨uger, A. Brylka,and C. Helmont. Capacity Measurements in a Cooperative MIMONetwork.
IEEE Transactions on Vehicular Communications , 58:2392 –2405, Jun 2008.[4] D. Gesbert, S. Hanly, H. Huang, S. Shamai-Shitz, O. Simeone, andW. Yu. Multi-Cell MIMO Cooperative Networks: A New Look atInterference.
IEEE JSAC , 28(9), 2010.[5] A Giovanidis, J. Krolikowski, and S. Brueck. A 0-1 program toform minimum cost clusters in the downlink of cooperating basestations. In
WCNC , pages 940–945, 2012.[6] A. Checko, H. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M. S.Berger, and L. Dittmann. Cloud RAN for Mobile Networks–ATechnology Overview.
IEEE Communications Surveys and Tutorials ,17:405–426, 2015.[7] M. Lyazidi, N. Aitsaadi, and R. Langar. Dynamic resource alloca-tion for Cloud-RAN in LTE with real-time BBU/RRH assignment.
IEEE International Conference on Communications (ICC) , 2016.[8] H.S. Dhillon, R.K. Ganti, F. Baccelli, and J.G. Andrews. Modelingand analysis of K-tier downlink heterogeneous cellular networks.
IEEE JSAC, vol. 30, no. 3, pp. 550-560, Apr. , 2012.[9] A. Checko, H. Holm, and H. Christiansen. Optimizing small cellsdeployment by the use of C-RANs.
European Wireless 2014 , 2014.[10] F. Baccelli and B. Błaszczyszyn.
Stochastic Geometry and WirelessNetworks, Volume I — Theory , volume 3, No 3–4 of
Foundations andTrends in Networking . NoW Publishers, 2009.[11] F. Baccelli and A. Giovanidis. A Stochastic Geometry Frameworkfor Analyzing Pairwise-Cooperative Cellular Networks.
IEEETrans. on Wireless Communications , 14(2):794–808, 2015.[12] G. Nigam, P. Minero, and M. Haenggi. Coordinated MultipointJoint Transmission in Heterogeneous Networks.
IEEE Trans. onCommunications , 62(11):4134–4146, 2014.[13] B. Blaszczyszyn and H. P. Keeler. Studying the SINR Process ofthe Typical User in Poisson Networks Using its Factorial MomentMeasures.
IEEE Trans. on Information Theory , 61(12):6774–6794,2015.[14] R. Tanbourgi, S. Singh, J. G. Andrews, and F. K. Jondral. ATractable Model for Noncoherent Joint-Transmission Base StationCooperation.
IEEE Trans. on Wireless Communications , 13(9):4959–4973, 2014.[15] A. H. Sakr and E. Hossain. Location-aware cross-tier coordinatedmultipoint transmission in two-tier cellular networks.
IEEE Trans.on Wireless Communications , 2014.[16] A. Papadogiannis, D. Gesbert, and E. Hardouin. A DynamicClustering Approach in Wireless Networks with Multi-Cell Co-operative Proocessing.
ICC , 2008.[17] S. Akoum and R. W. Heath. Interference Coordination: RandomClustering and Adaptive Limited Feedback.
IEEE Trans. on SignalProcessing , 61(7):1822–1834, 2013.[18] J. Park, N. Lee, and R. W. Heath. Cooperative Base StationColoring for Pair-wise Multi-Cell Coordination.
IEEE Trans. onWireless Communications , 2016.[19] H. Huang and J. Andrews. A Stochastic Geometry Aproachto Coverage in Cellular Networks With Multi-Cell Cooperation.
Global Telecommunications Conference , 2011.[20] Anjin Guo, Yi Zhong, M. Haenggi, and Wenyi Zhang. Thegauss-poisson process for wireless networks and the benefits ofcooperation. In
IEEE Trans. on Wireless Communications , 2016.[21] M. Haenggi.
Stochastic Geometry for Wireless Networks . CambridgeUniversity Press New York, NY, USA, 1st edition, 2012.[22] M. Afshang, H.S. Dhillon, and P.H.J. Chong. Fundamentals ofCluster-Centric Content Placement in Cache-Enabled Device-to-Device Networks.
IEEE Trans. on Communications , 2016.[23] O. H¨aggstr¨om and R. Meester. Nearest neighbour and hard spheremodels in continuum percolation.
Wiley Random Struct. Alg. ,9(3):295–315, 1996.[24] D. J. Daley and G. Last. Descending chains, the lilypond model,and mutual-nearest-neighbour matching.
Advances in AppliedProbability , 37(3):604–628, 2005.[25] I. Kozakova, R. Meester, and S. Nanda. The size of components incontinuum nearest-neighbor graphs.
The Annals of Probability , 34,no.2:528–538, 2006. [26] D.J. Daley, H. Stoyan, and D. Stoyan. The volume fraction of aPoisson germ model with maximally non-overlapping sphericalgrains. Adv. Appl. Prob. (SGSA) , 31:610–624, 1999.[27] A. Giovanidis, L. David ´Alvarez Corrales, and L. Decreusefond.Analyzing interference from static cellular cooperation using theNearest Neighbour Model. In
WiOpt , pages 576–583, 2015.[28] L. David ´Alvarez Corrales, A. Giovanidis, and P. Martins. Cov-erage Gains from the Static Cooperation of Mutually NearestNeighbours. In
Globecom , 2016.[29] P. Kerret and D. Gesbert. CSI Sharing Strategies for TransmitterCooperation in Wireless Networks.
IEEE Wireless Communications ,20:43–49, 2013.[30] P. Kerret.
Transmission coop´erative dans les r´eseaux sans-fil avec feed-back distribu´e . R´eseaux et t´el´ecommunications, T´el´ecom ParisTech,2013.[31] Olav Kallemberg.
Random Measures . Academic Pr, 1984.[32] D.J. Sheskin.
Handbook of parametric and nonparametric statisticalprocedures . Chapman & Hall/CRC, 4th Edition, 2007.[33] F. Baccelli and P. Br´emaud.
Elements of Queueing Theory . Springer,second edition edition, 2003.[34] A.J. Baddeley.
Spatial Point Processes and their Applications . Lec-ture Notes in Mathematics: Stochastic Geometry, Springer Verlag,Berlin Heidelberg, 2007.[35] M. de Berg, O. Cheong, M. van Kreveld, and M. Overmars.
Com-putational Geometry: Algorithms and Applications . Springer-Verlag,3rd rev. ed. 2008.[36] J. G. Andrews, F. Baccelli, and R. K. Ganti. A Tractable Approachto Coverage and Rate in Cellular Networks.
IEEE Trans. onCommunications , 59(11):3122–3134, 2011.[37] M. Afshang, H.S. Dhillon, and P.H.J. Chong. Modeling andPerformance Analysis of Clustered Device-to-Device Networks.
IEEE Transactions on Wireless Communications , 15:4957 – 4972, April2016.[38] Sheldon M. Ross.
A first course in probability . Pearson Pretince Hall,8th edition, 2009.[39] Eric W. Weisstein. Modified Bessel Function ofthe First Kind.
MathWorld–A Wolfram Web Resource.http://mathworld.wolfram.com/ModifiedBesselFunctionoftheFirstKind.html . Luis David ´Alvarez Corrales is a Ph.D. stu-dent at the Network and Computer Science De-partment, Telecom ParisTech. He received theMaster’s degree in Probability and StochasticModels from the University Pierre et Marie Curie(UPMC), laboratory LPMA. He has further com-pleted the Master’s degree in Mathematical Fi-nance and the Bachelor’s degree in Mathematicsat the National Autonomous University of Mexico(UNAM). In Mexico, he has worked in CRM anal-ysis, energy regulation, and insurance regulation(Solvency II). His research lies in the area of applied probability, stochas-tic modeling, and stochastic geometry applications for the performanceevaluation of telecommunication networks.
Anastasios Giovanidis is a permanent re-searcher of the French National Center forScientific Research (CNRS). After spendingthree years with Telecom ParisTech, laboratoryCNRS-LTCI, he is now affiliated with the Uni-versity Pierre et Marie Curie (UPMC), laboratoryCNRS-LIP6. He has been a postdoctoral fellow,first with the Zuse Institute Berlin, Germany andlater with INRIA, Paris, France. He received theDr.-Ing. degree in Mobile Communications fromthe Technical University of Berlin, Germany andthe Diploma in Electrical and Computer Engineering from the NationalTechnical University of Athens, Greece. His research lies in the areaof performance evaluation and optimization of telecommunication net-works, with emphasis in queuing and stochastic geometry applications.
Philippe Martins is Professor in the networkingand computer science department, at T´el´ecomParisTech (Paris, France). He is an IEEE se-nior member. His main research interests lie inprotocol design and performance evaluation ofmobile networks. He worked on stochastic ge-ometry models for dimensioning and planning.He is also investigating on coverage modellingand energy harvesting using models based onsimplicial homology.
Laurent Decreusefond is a former student ofEcole Normale Sup´erieure de Cachan. He ob-tained his Ph.D. degree in Mathematics in 1994from Telecom ParisTech and his Habilitation in2001. He is currently a Professor in the Networkand Computer Science Department, at TelecomParisTech. His main fields of interest are theMalliavin calculus, the stochastic analysis of longrange dependent processes, random geome-try and topology and their applications. With P.Moyal, he co-authored a book about the stochas-tic modelling of telecommunication A PPENDIX
A (S
UPPLEMENTAL M ATERIAL ) Proof of Proposition 1
Denote by ( R, Θ) the polar coordinates of the closest Φ atom from thetypical user. The r.v. R is Rayleigh distributed, with scale parameter ( λ π ) − / . Because of the isotropy of a stationary PPP, R and Θ areindependent r. v. and Θ is uniformly distributed over [0 , π ) . Therefore,its density function is given by f ( R, Θ) ( r, θ ) = λre − λπr { r> } { θ ∈ [0 , π ) } Then, P (0 (cid:121) Φ (2) ) = E (cid:16) P (0 (cid:121) Φ (2) | R, Θ) (cid:17) = (cid:90) ∞ (cid:90) π P (0 (cid:121) Φ (2) | R = r, Θ = θ ) λre − λπr dθdr Fix a realisation φ . Denote by x and y two different atoms from φ ,whose polar coordinates are ( r, θ ) and ( s, ϕ ) , respectively. If ρ denotesthe Euclidean distance between x and x , then, ρ = r + s − rscos ( θ − ϕ ) If we suppose that x is the nearest neighbor atom from φ to the origin,then, the atoms x and y are in MNNR iff D ( x, y ) := ( B ( x, ρ ) ∪ B ( y, ρ )) \ B (0 , r ) (27)is empty of atoms from φ \{ x, y } . Denote by F ( r, s, θ, ϕ ) the Euclideansurface of D ( x, y ) , the empty space function of a PPP implies that P (0 (cid:121) Φ (2) | R = r, Θ = θ )= (cid:90) ∞ (cid:90) π e − λF ( r,s,θ,ϕ ) sdsdϕ, and therefore, P (0 (cid:121) Φ (2) )= λ (cid:90) ∞ (cid:90) π (cid:90) ∞ (cid:90) π e − λF ( r,s,θ,ϕ ) − λπr rsdsdϕdθdr In some cases, it is actually possible to find explicit values for F ( r, s, θ, ϕ ) . For example, the case ρ ≥ r implies that B (0 , r ) ⊂ B ( x, ρ ) . Thus, S ( D ( x, y )) = S ( B ( y, ρ ) \ B ( x, ρ )) + S ( B ( x, ρ )) − S ( B (0 , r )) Then, we have that S ( B ( y, ρ ) \ B ( x, ρ )) = πρ (1 − γ ) , and S ( B ( x, ρ )) − S ( B (0 , r )) = πρ − πr . Unfortunately, in other cases is arduous to obtain F ( r, s, θ, ϕ ) . A PPENDIX
B (S
UPPLEMENTAL M ATERIAL ) Proof of Theorem 5
For a natural number n , denote B n := B (0 , n ) and fix Φ n := Φ (1) B n , asdone in equation (13). We will prove that, for every compact subset E of R ,(i) lim n →∞ P (Φ n ( E ) = 0) = P (Φ (1) ( E ) = 0) (ii) lim sup n →∞ P (Φ n ( E ) ≤ ≥ P (Φ (1) ( E ) ≤ (iii) lim t (cid:37)∞ lim sup n →∞ P (Φ n ( E ) > t ) = 0 The previous being equivalent to convergence in distribution of thesequence of point processes (Φ n ) to Φ (1) [31]. Fix a compact E ⊂ R .Let us start to prove ( i ) . Being Φ n a thinning of the PPP Φ , P (Φ n ( E ) = 0) = e − λ S ( E ) + P (Φ n ( E ) = 0 , Φ( E ) > e − λ S ( E ) + P (Φ n ( E ) = 0 , Φ (1) ( E ) = 0 , Φ( E ) > P (Φ n ( E ) = 0 , Φ (1) ( E ) > , Φ( E ) > Given that the compact subset E is fixed, take a natural number n such that n > sup y ∈ E (cid:107) y (cid:107) and such that E ⊂ B n , for every n > n . Therefore, for every atombelonging to Φ (1) , but not to Φ n , the distance to its nearest neighbormust exceed sup y ∈ E (cid:107) y (cid:107) . Thus, there exits a constant C > suchthat, for every n > n , P (Φ n ( E ) = 0 , Φ (1) ( E ) > , Φ( E ) > ≤ e − λπC n In the same fashion, P (Φ (1) ( E ) = 0) = e − λ S ( E ) + P (Φ n ( E ) = 0 , Φ (1) ( E ) = 0 , Φ( E ) > P (Φ (1) ( E ) = 0 , Φ n ( E ) > , Φ( E ) > and there must exists a natural number n , and a constant C > suchthat, for every n > n , P (Φ (1) ( E ) = 0 , Φ n ( E ) > , Φ( E ) > ≤ e − λπC n Take N = max { n , n } , then, for every n > N , | P (Φ n ( E ) = 0) − P (Φ (1) ( E ) = 0) | ≤ e − λπC n + e − λπC n We conclude that lim n →∞ P (Φ n ( E ) = 0) = P (Φ (1) ( E ) = 0) To prove ( ii ) , remark that P (Φ n ( E ) = 1) = P (Φ n ( E ) = 1 , Φ (1) ( E ) (cid:54) = 1)+ P (Φ n ( E ) = 1 , Φ (1) ( E ) = 1) P (Φ (1) ( E ) = 1) = P (Φ (1) ( E ) = 1 , Φ n ( E ) (cid:54) = 1)+ P (Φ (1) ( E ) = 1 , Φ n ( E ) = 1) hence | P (Φ n ( E ) = 1) − P (Φ (1) ( E ) = 1) |≤| P (Φ n ( E ) = 1 , Φ (1) ( E ) (cid:54) = 1) − P (Φ (1) ( E ) = 1 , Φ n ( E ) (cid:54) = 1) | In the same way as we did before, we can prove that lim n →∞ | P (Φ n ( E ) = 1 , Φ (1) ( E ) (cid:54) = 1) − P (Φ (1) ( E ) = 1 , Φ n ( E ) (cid:54) = 1) | = 0 , and this leads to lim n →∞ P (Φ n ( E ) = 1) = P (Φ (1) ( E ) = 1) Finally, we prove ( iii ) . Being Φ n a thinning of the PPP Φ , P (Φ n ( E ) > t ) ≤ P (Φ( E ) > t ) ≤ E Φ( E ) t = λ S ( E ) t which goes to zero, as t (cid:37) ∞ .Take a sequence ( A n ) of compact sets. To conclude that (Φ (1) A n ) converges in distribution to Φ (1)) , then it must fulfil that, for everynatural number m , there exits another natural number N such that, forevery n > N , then, B m ⊂ A n . We can prove that (Φ (2) A n ) converges indistribution to Φ (2) . A PPENDIX
C (S
UPPLEMENTAL M ATERIAL ) Proof of Proposition 2.
Denote by A and B the Cartesian coordinates of the nearest parentto the typical user and his daughter, respectively, and also denoteby ( R , Θ) and ( Z , Ψ) their respective polar coordinates. Define C := A − B and denote its polar coordinates by ( W, Ω) (see Figure9). The random variables R and W are Rayleigh distributed, withscale parameters ζ and α , respectively. Moreover, the random angles Θ , Ψ and Ω are considered uniformly distributed over [0 , π ) , to preservethe isotropy in the PPP case. Also, the random variables R , Θ , W , and Ω are independent between them, as in the PPP case.Our first goal is to find the joint distribution of the random vector ( R , Z ) and, as a consequence, find also the distribution of Z . Thecartesian coordinates of a point around a center, wich has rayleighradial distance from the origin and uniform angle, are distibuted asan independent Gaussian vector [38, pp. 276, Ex. 7b]. Hence, thereexist independent random variables A x , A y , C x , C y , where A x , A y are Normal distributed, with parameter (0 , ζ ) , and C x , C y are alsoNormal distributed, with parameters (0 , α ) , and such that ( A x , A y ) d = ( R cos Θ , R sin Θ) , ( C x , C y ) d = ( W cos Ω , W sin Ω) . By definition, (a)
Fig. 9 A x d = R cos Θ , C x d = R cos Θ − Z cos Ψ ,A y d = R sin Θ , C y d = R sin Θ − Z sin Ψ . The absolute value of the Jacobian of the above transformation is R Z .Denote by f A x ,A y ,C x ,C y the joint PDF of ( A x , A y , C x , C y ) , if f R, Θ ,Z, Ψ denotes the joint PDF of ( R , Θ , Z , Ψ) , then, by the change of variableTheorem [38, pp. 274], f R, Θ ,Z, Ψ ( r, θ, z, ψ )= f A x ,A y ,C x ,C y ( rcosθ, rsinθ, rcosθ − zcosψ, rsinθ − zsinψ ) rz ( a ) = rz (2 παζ ) e − (cid:18) r cos θ ζ + r sin θ ζ + ( rcosθ − zcosψ )22 α + ( rsinθ − zsinψ )22 α (cid:19) ( b ) = rz (2 παζ ) e − (cid:18) r (cid:18) α + ζ (cid:19) + z α − rzcos ( θ − ψ ) α (cid:19) , where ( a ) comes from the formula of the distribution of independentGaussian random variables, and ( b ) follows from the trigonometricidentities cos θ + sin θ = 1 and cosθcosψ + sinθsinψ = cos ( θ − ψ ) . Toobtain the joint PDF of ( R , Z ) , denoted by f R ,Z , we integrate theprevious expression over [0 , π ) × [0 , π ) , with respect to the variables θ and ψ , f R ,Z ( r, z ) ( c ) = rz ( αζ ) e − (cid:18) r (cid:18) α + ζ (cid:19) + z α (cid:19) π (cid:90) π e rzcoswα dw ( d ) = rz ( αζ ) e − (cid:18) r (cid:18) α + ζ (cid:19) + z α (cid:19) I (cid:16) rzα (cid:17) , where ( c ) comes from the change of variable w = θ − ψ and ( d ) followsbecause the integral representation I ( x ) = π (cid:82) π e xcosw dw [39]. Letus denote by f Z the PDF of the random variable Z and by η = (cid:16) α + ζ (cid:17) . To obtain f Z , we integrate over [0 , ∞ ) with respect to thevariable r the preceding equation f Z ( z ) ( e ) = (cid:90) ∞ rz ( αζ ) e − (cid:18) r η + z α (cid:19) ∞ (cid:88) n =0 (1 / n ( n !) (cid:16) rzα (cid:17) n dr = z ( αζ ) e − z α ∞ (cid:88) n =0 (1 / n ( n !) (cid:18) z α (cid:19) n (cid:90) ∞ r n e − r η rdr ( f ) = z ( αζ ) η e − z α ∞ (cid:88) n =0 (cid:16) z α η (cid:17) n n != z ( αζ ) η e − z α e z α η ( g ) = zα + ζ e − z α ζ , where ( e ) comes from the series representation I ( x ) = (cid:80) ∞ n =0 (1 / n ( n !) x n [39], while ( f ) follows after the formula (cid:90) ∞ r n e − r η rdr = 2 n η n +1 n ! , and ( g ) after soma algebraic manipulations and from the definition of η . A PPENDIX
D (S
UPPLEMENTAL M ATERIAL ) Proof of Proposition 4
We split the proof in three parts.
A first expression
The events { R < min { R , Z }} , { R < min { R , Z }} , and { Z < min { R , R }} are mutually independent, then, from equation (26), P (SINR > T )= P (cid:32) ˜ f ( R ) σ + ˆ I ( R , R ) > T, R < min { R , Z } (cid:33) + P (cid:32) ˜ g ( R , Z ) σ + ˆ I ( R , R ) > T, R < min { R , Z } (cid:33) + P (cid:32) ˜ g ( R , Z ) σ + ˆ I ( Z , R ) , Z < min { R , R } (cid:33) = E (cid:34) (cid:26) ˜ f ( R σ I ( R ,R >T (cid:27) { R < min { R ,Z }} (cid:35) + E (cid:34) (cid:26) ˜ g ( R ,Z σ I ( R ,R >T (cid:27) { R < min { R ,Z }} (cid:35) + E (cid:34) (cid:26) ˜ g ( R ,Z σ I ( Z ,R >T (cid:27) { Z < min { R ,R }} (cid:35) (28)For the first term we have that E (cid:34) (cid:110) ˜ f ( R σ I ( R ,R >T (cid:111) { R < min { R ,Z }} (cid:35) = E (cid:34) E (cid:34) (cid:110) ˜ f ( R σ I ( R ,R >T (cid:111) { R < min { R ,Z }} (cid:12)(cid:12)(cid:12) R , R , Z (cid:35)(cid:35) ( a ) = E (cid:34) { R < min { R ,Z }} E (cid:34) (cid:110) ˜ f ( R σ I ( R ,R >T (cid:111)(cid:12)(cid:12)(cid:12) R , R , Z (cid:35)(cid:35) ( b ) = E (cid:34) { R < min { R ,Z }} E (cid:34) (cid:110) ˜ f ( R σ I ( R ,R >T (cid:111)(cid:12)(cid:12)(cid:12) R (cid:35)(cid:35) = E (cid:34) { R < min { R ,Z }} P (cid:32) ˜ f ( R ) σ + ˆ I ( R , R ) > T (cid:12)(cid:12)(cid:12) R (cid:33)(cid:35) , where ( a ) comes from the properties of the conditional expectation and ( b ) follows because the event (cid:110) ˜ f ( R ) σ +ˆ I ( R ,R ) > T (cid:111) is independent of R and Z . After a similar analysis for the two terms with the cooperativesignal E (cid:34) (cid:110) ˜ g ( R ,Z σ I ( R ,R >T (cid:111) { R < min { R ,Z }} (cid:35) = E (cid:34) { R < min { R ,Z }} P (cid:32) ˜ g ( R , Z ) σ + ˆ I ( R , R ) > T (cid:12)(cid:12)(cid:12) R , Z (cid:33)(cid:35) , E (cid:34) (cid:110) ˜ g ( R ,Z σ I ( Z ,R >T (cid:111) { Z < min { R ,R }} (cid:35) = E (cid:34) { Z < min { R ,R }} P (cid:32) ˜ g ( R , Z ) σ + ˆ I ( Z , R ) > T (cid:12)(cid:12)(cid:12) R , Z (cid:33)(cid:35) Some functions
Denote by ˆ G ( r ) := P (cid:32) ˜ f ( R ) σ + ˆ I ( R , R ) > T (cid:12)(cid:12)(cid:12) R = r (cid:33) ˆ H ( r, z ) := P (cid:32) ˜ g ( R , Z ) σ + ˆ I ( R , R ) > T (cid:12)(cid:12)(cid:12) R = r, Z = z (cid:33) ˆ K ( r, z ) := P (cid:32) ˜ g ( R , Z ) σ + ˆ I ( Z , R ) > T (cid:12)(cid:12)(cid:12) R = r, Z = z (cid:33) . For a given r > , because R is independent from ˆ I ( R , R ) , ˆ G ( r ) = P (cid:16) ˜ f ( r ) > T ( σ + ˆ I ( r, r )) (cid:17) Consider ˜ f ( r ) as in (6), then it follows an exponential distribution withparameter r β p . Since ˆ I ( r, r ) is independent of ˜ f ( r ) , ˆ G ( r ) = E (cid:104) P (cid:16) ˜ f ( r ) > T (cid:16) σ + ˆ I ( r, r ) (cid:17) (cid:12)(cid:12)(cid:12) ˆ I ( r, r ) (cid:17)(cid:105) = e − Trβp σ L ˆ I (1) (cid:32) T r β p ; r (cid:33) L ˆ I (2) (cid:32) T r β p ; r (cid:33) , where the deterministic functions L ˆ I (1) ( s ; ρ ) and L ˆ I (2) ( s ; ρ ) are givenby (22).In the same fashion, for r > and z > , because ( R , Z ) isindependent of ˆ I ( R , R ) , ˆ H ( r, z ) = P (cid:16) ˜ g ( r, z ) > T (cid:16) σ + ˆ I ( r, r ) (cid:17)(cid:17) Using the general expression in (8) for ˜ g ( r, z ) , ˆ H ( r, z ) = n (cid:88) i =1 c i (cid:0) r, z (cid:1) e − Td i ( r,z ) σ L ˆ I (1) (cid:0) T d i ( r, z ); r (cid:1) L ˆ I (2) (cid:0) T d i ( r, z ); r (cid:1) We do the same to find and expression for ˆ K ( r, z ) . Final expression
To complete the analysis, we need to find the coverage probabilityexpressed in equation (28), thus, we need expressions for E (cid:104) ˆ G ( R ) { R < min { R ,Z }} (cid:105) , E (cid:104) ˆ H ( R , Z ) { R < min { R ,Z }} (cid:105) , E (cid:104) ˆ K ( R , Z ) { Z < min { R ,R }} (cid:105) . Let us begin by the first one, E (cid:104) ˆ G ( R ) { R < min R ,Z } (cid:105) ( a ) = E (cid:104) E (cid:104) ˆ G ( R ) { R < min R ,Z } | R (cid:105)(cid:105) = E (cid:104) ˆ G ( R ) E (cid:2) { R < min R ,Z } | R (cid:3)(cid:105) = E (cid:104) ˆ G ( R ) P (min { R , Z } > R | R ) (cid:105) , where ( a ) follows by properties of the conditional expectation. Define G ( r ) = ˆ G ( r ) P (min { R , Z } > R | R = r ) , we only have left to find an explicit expression for P (min { R , Z } > R | R = r ) . Because R is independent of ( R , Z ) , P (min { R , Z } > R | R = r ) = P (min { R , Z } > r ) , and then P (min { R , Z } > r ) = 1 − F R ( r ) − F Z ( r ) + F R ,Z ( r, r ) , where F R , F Z , and F R ,Z are the CDF of R , Z , and ( R , Z ) thatcan be explicitly obtained from equation (17).In the same fashion, E (cid:104) ˆ H ( R , Z ) { R < min R ,Z } (cid:105) = E (cid:104) ˆ H ( R , Z ) P (min { R , Z } > R | R , Z ) (cid:105) , E (cid:104) ˆ K ( R , Z ) { Z < min R ,R } (cid:105) = E (cid:104) ˆ K ( R , Z ) P (min { R , R } > Z | R , Z ) (cid:105) Define H ( r, z ) := ˆ H ( r, z ) P (min { R , Z } > R | R = r, Z = z ) ,K ( r, z ) := ˆ K ( r, z ) P (min { R , R } > Z | R = r, Z = z ) To obtain explicit formulas for H ( r, z ) and K ( r, z ) , we proceed asbefore to find out that P (min { R , Z } > R | R = r, Z = z ) = (1 − F R ( r )) { z>r } , P (min { R , R } > Z | R = r, Z = z ) = (1 − F R ( z )) { r>z } , where F R is the CDF of R1