The Meta Distributions of the SIR/SNR and Data Rate in Coexisting Sub-6GHz and Millimeter-Wave Cellular Networks

Using stochastic geometry tools, we develop a systematic framework to characterize the meta distributions of the downlink SIR/SNR and data rate of the typical device in a cellular network with coexisting sub-6GHz and millimeter wave (mm-wave) spectrums. Macro base-stations (MBSs) transmit on sub-6GHz channels (which we term “microwave” channels), whereas small base-stations (SBSs) communicate with devices on mm-wave channels. The SBSs are connected to MBSs via a microwave (<inline-formula> <tex-math notation="LaTeX">$\mu $ </tex-math></inline-formula>wave) wireless backhaul. The <inline-formula> <tex-math notation="LaTeX">$\mu $ </tex-math></inline-formula>wave channels are interference limited and mm-wave channels are noise limited; therefore, we have the meta distribution of SIR and SNR in <inline-formula> <tex-math notation="LaTeX">$\mu $ </tex-math></inline-formula>wave and mm-wave channels, respectively. To model the line-of-sight (LOS) nature of mm-wave channels, we use Nakagami-m fading model. To derive the meta distribution of SIR/SNR, we characterize the conditional success probability (CSP) (or equivalently reliability) and its <inline-formula> <tex-math notation="LaTeX">$b^{\mathrm {th}}$ </tex-math></inline-formula> moment for the typical device (a) when it associates to a <inline-formula> <tex-math notation="LaTeX">$\mu $ </tex-math></inline-formula>wave MBS for <italic>direct</italic> transmission, and (b) when it associates to a mm-wave SBS for <italic>dual-hop</italic> transmission (backhaul and access transmission). Performance metrics such as the mean and variance of the local delay (network jitter), mean of the CSP (coverage probability), and variance of the CSP are derived. Closed-form expressions are presented for special scenarios. The extensions of the developed framework to the <inline-formula> <tex-math notation="LaTeX">$\mu $ </tex-math></inline-formula>wave-only network or mm-wave only networks where SBSs have mm-wave backhauls are discussed. Numerical results validate the analytical results. Insights are extracted related to the reliability, coverage probability, and latency of the considered network.


I. INTRODUCTION
T HE SUB-6GHZ spectrum is running out of bandwidth to support a huge number of devices in the cellular networks.Therefore, cellular operators of the upcoming 5G networks will tap into the millimeter-wave (mm-wave) spectrum to use wider bandwidths.The mm-wave spectrum has wider bandwidths that can meet higher traffic demands and support data rates into the order of gigabits per second.The use of mm-wave spectrum is one of the key enablers of 5G and beyond networks [2] and will coexist with sub-6GHz frequencies [3], [4].However, mm-wave transmissions are highly susceptible to blockages and penetration losses; therefore the mm-wave spectrum will complement the sub-6GHz spectrum in 5G networks [5]- [8].Self-backhauling offers a simple cost-saving strategy to enable dense millimeter wave cellular networks since the optical fiber connection is always expensive [9]- [13].
In this article, we develop a framework to characterize the meta distributions of SIR/SNR as well as data rate in the coexisting sub-6GHz and mm-wave cellular network.We assume a two-tier network architecture as illustrated in Fig. 1.Tier 1 consists of macro base stations (MBSs) and tier 2 is composed of small base stations (SBSs).A MBS communicates with SBSs on backhaul links in the microwave spectrum.SBSs communicate with devices on access links in the mm-wave spectrum.This scenario supports dual-hop communications between MBSs and devices.Devices can also communicate with MBSs via direct links in the microwave spectrum, as shown in Fig. 1.Given the hybrid spectrum network, it is crucial to develop new theoretic frameworks to characterize the performance of such networks.Within this context, we consider the use of meta distributions to study the performance of such hybrid spectrum networks.
The meta distribution is first introduced by Haenggi [14] to provide a fine-grained reliability and latency analysis of wireless networks.A meta distribution 1 is defined as the distribution of the conditional success probability (CSP) of the transmission link (also termed as link reliability), conditioned on the locations of the wireless transmitters.In addition to the standard coverage (or success) probability which is equivalent to the mean of CSP, the meta distribution can capture important network performance measures such as the mean of the local transmission delay, the variance of the local transmission delay (referred to as network jitter), and the variance of the CSP which depicts the variation of the devices' performance from the mean coverage probability.Evidently, the standard coverage probability provides limited information about the performance of the typical wireless network [15]- [17].In this article, we develop a novel stochastic geometry framework based on meta distributions to estimate and analyze the communication latency and reliability of devices in a coexisting sub-6GHz and mm-wave cellular network.

A. RELATED WORK
A variety of research works studied the coverage probability of mm-wave only cellular networks [18]- [20].Di Renzo [18] proposed a general mathematical model to analyze multitier mm-wave cellular networks.Bai and Heath [19] derived the coverage and rate performance of mm-wave cellular 1.A meta distribution is a fine-grained unified performance metric that enables us to evaluate the reliability and latency of next generation wireless networks, in addition to the conventional coverage probability.The meta distribution provides answers to questions such as "What fraction of devices can achieve x% transmission success probability?"whereas the conventional success probability answers questions such as "What fraction of devices experience transmission success?" [14].
networks.Different parameters of Nakagami fading are assumed for LOS and NLOS links.Turgut and Gursoy [20] investigated heterogeneous downlink mm-wave cellular networks consisting of K tiers of randomly located BSs where each tier operates in a mm-wave frequency band.They derived coverage probability for the entire network using tools from stochastic geometry.They used Nakagami fading to model small-scale fading.Deng et al. [21] derived the success probability at the typical receiver in mm-wave device-to-device (D2D) networks.The authors considered Nakagami fading and incorporated directional beamforming.
Some recent studies analyzed the coverage or success probability of coexisting μwave and mm-wave cellular networks.A hybrid cellular network was considered by Singh et al. [10] to estimate the uplink-downlink coverage and rate distribution of self-backhauled mm-wave networks.Elshaer et al. [3] developed an analytical model to characterize decoupled uplink and downlink cell association strategies.The authors showed the superiority of this technique compared to the traditional coupled association in a network with traditional MBSs coexisting with denser mm-wave SBSs.Singh et al. [10] and Elshaer et al. [3] modeled the fading power as Rayleigh fading to enable better tractability.
Compared to traditional coverage analysis conducted in [3], [19], [20], Deng and Haenggi [22] analyzed the meta distribution of the SIR in mm-wave only single-hop D2D networks using the Poisson bipolar model and Rayleigh fading channels for analytical tractability.

B. CONTRIBUTIONS
To the best of our knowledge, our work is the first to characterize the meta distributions of SIR/SNR and data rate for coexisting μwave and mm-wave networks.Different from [3], [19], [20], [22], we develop a stochastic geometry framework that considers (i) coexistence of two different network tiers with completely different channel propagation, interference, and fading models, (ii) dual-hop transmissions enabled by two different spectrums, one in each network tier, and (iii) Nakagami-m fading 2 model with shape parameter m for LOS mm-wave channels.The analysis presented in Lemma 3 to derive the b-th moment of conditional success probability (CSP) of the typical user, considering two hops with two different channel distributions along with a direct link, is a primary novelty.Such a theoretical result is not reported yet in the existing research works related to the meta-distribution.The evaluation of Lemma 3 requires Lemma 2 and Lemma 1.Another fundamental novelty is the consideration of Nakagami-m fading channels which requires a novel approach to derive the b-th moment of the CSP at each of the access link, backhaul link, and direct link.As such, the results presented in Lemma 4 and Lemma 5 2. Nakagami-m fading is a generic distribution that includes Rayleigh distribution (for non-LOS fading) as its special case when m = 1 and can well approximate the Rician fading distribution for 1 ≤ m ≤ ∞ (for LOS fading).are novel and not reported yet in the existing research works related to the meta distribution.
We assume a hybrid spectrum network architecture as illustrated in Fig. 1.Since microwave transmissions are interference limited and mm-wave transmission are noise limited, 3 we study the meta distributions of the SIR and SNR in μwave and mm-wave channels, respectively.
Our contributions and methodology include the following: (a) Different from existing works, we characterize the CSP (which is equivalent to reliability) of the typical device and its b th moment when the device either associates to (1) μwave MBS for direct transmission or (2) mm-wave SBS for dual-hop transmission (access and backhaul transmission).Using the novel moment expressions in the two scenarios, we derive a novel expression for the cumulative moment M b,T of the considered hybrid spectrum network.(b) Using the cumulative moment M b,T , we characterize the exact and approximate meta distributions of the data rate and downlink SIR/SNR of the typical device.Since the expression of M b,T relies on a binomial expansion of power b, the results for the meta distribution requiring complex values of b are obtained by applying Newton's generalized binomial theorem.(c) We characterize important network performance metrics such as coverage probability, mean local delay (which is equivalent to latency), and variance of the local delay (network jitter), using the derived cumulative moment M b,T .For metrics with negative values of b, we apply the binomial theorem for negative integers.(d) To model the LOS nature of mm-wave, we consider the versatile Nakagami-m fading channel model.
To the best of our knowledge, the meta distribution for the Nakagami-m fading channel has not been investigated yet.(e) We demonstrate the application of this framework to other specialized network scenarios where (i) SBSs are connected to MBSs via a mm-wave wireless backhaul and (ii) a network where all transmissions are conducted in μwave spectrum.Closed-form results are provided for special cases and asymptotic scenarios.We validate analytical results using Monte-Carlo simulations.Numerical results give valuable insights related to the reliability, mean local delay, variance of CSP, and standard success probability of a device.

II. SYSTEM MODEL AND ASSUMPTIONS
In this section, we describe the network deployment model (Section II-A), antenna model (Section II-B), channel model (Section II-C), device association criteria (Section II-D), 3. Given highly directional beams and high sensitivity to blockage, recent studies showed that mm-wave networks can be considered as noise limited rather than interference limited [10], [23]- [27].and SNR/SIR models for access and backhaul transmissions (Section II-E).

A. NETWORK DEPLOYMENT AND SPECTRUM ALLOCATION MODEL
We assume a two-tier cellular network architecture as shown in Fig. 1 in which the locations of the MBSs and SBSs are modeled as a two-dimensional (2D) homogeneous Poisson point process (PPP) k = {y k,1 , y k,2 , . . ., } of density λ k , where y k,i is the location of i th MBS (when k = 1) or the i th SBS (when k = 2).Let the MBS tier be tier 1 (k = 1) and the SBSs constitute tier 2 (k = 2).The locations of devices in the network are modeled as a stationary and ergodic point process of arbitrary positive intensity.We consider the typical outdoor device is located at the origin.The typical device is denoted by 0 and its tagged BS is denoted by y k,0 , i.e., tagged MBS (when k = 1) or tagged SBS (when k = 2).All BSs in the k th tier transmit with the same transmit power P k in the downlink.
We assume that a portion ηW 1 of the frequency band W 1 is reserved for the access transmission and the rest (1−η)W 1 is reserved for the backhaul transmission, where W 1 , and W 2 denote the total available μwave spectrum and mm-wave spectrum, respectively, and 0 ≤ η ≤ 1. Determining the optimal spectrum allocation ratio η will be studied in our future work.

B. ANTENNA MODEL
We assume that all MBSs are equipped with omnidirectional antennas with gain denoted by G o 1 dB.We consider SBSs and devices are equipped with directional antennas with sectorized gain patterns as in [18], [22], [24] to approximate the actual antenna pattern.The sectorized gain pattern is given by: where subscript a ∈ {2, D} denotes for SBSs and devices, respectively.Considering a √ N × √ N uniform planar square antenna array with N elements, 4 the antenna parameters of a uniform planar square antenna array can be given as in [22], i.e., G max a = N is the main lobe antenna gain, is the angle of the boresight direction, and θ a = √ 3 √ N is the main lobe beam width.A perfect beam alignment is assumed between a device and its serving SBS [3], [19].The antenna beams of the desired access links are assumed to be perfectly aligned, i.e., the direction of arrival (DoA) between the transmitter and receiver is known a priori at the BS and the effective gain on the intended access link can thus be denoted as G max 2 G max D .This can be done by assuming that the serving mm-wave SBS and device can adjust their 4.The antenna elements N can also be realized in practice as As such, we believe that the arrangement of antenna elements will not effect the insights related to it obtained in from model.antenna steering orientation using the estimated angles of arrivals.The analysis of the alignment errors on the desired link is beyond the scope of this work.

C. CHANNEL MODEL 1) PATH-LOSS MODEL
The signal power decay is modeled as L(r) = r α , where L(r) is the path loss for the typical receiver located at a distance r from the transmitter and α is the path loss exponent (PLE).Let L 1 (r) = r 1,D α 1 denotes the path loss of the typical device associated with the MBS tier, where α 1 is the PLE.Similarly, L 2 (r) = r 2,D α 2,l denotes the path loss of the typical device associated with the SBS tier where α 2,l = α 2,L is the PLE in the case of LOS and α 2,l = α 2,N is the PLE in the case of NLOS.It has been shown that mmwave LOS and NLOS conditions have markedly different PLEs [28].Also, we consider the near-field path loss factor ζ = ( carrier wavelength 4π ) 2 at 1 m [3], i.e., different path loss for different frequencies at the reference distance.

2) FADING MODEL
For outdoor mm-wave channels, we consider the versatile Nakagami-m fading channel model due to its analytical tractability and following the previous line of research studies [19]- [21], [29], [30].Nakagami-m fading is a general and tractable model to characterize mm-wave channels.Also, in several scenarios, Nakagami-m can approximate Rician fading for 1 ≤ m ≤ ∞ (for LOS fading) [31], [32].Rician fading is commonly used to model LOS transmissions but not tractable for meta distribution modeling since its CDF definition involves the Marcum Q-function which is not tractable.The fading parameter m l ∈ [1, 2, . . ., ∞) where l ∈ {L, N} denotes LOS and NLOS transmission links, respectively, and the mean fading power is denoted by l .The fading channel power h l follows a gamma distribution given as exp( −m l x l ), x > 0, where (.) is the gamma function, m l is the shape (or fading) parameter, and m l l is the scale parameter.That is, we consider h l ∼ (m L , 1/m L ) for the LOS links and h l ∼ (m N , 1/m N ) for the NLOS links.Rayleigh fading is a special case of Nakagami-m for m L = m N = 1.Due to the NLOS nature of μwave channels, we assume Rayleigh fading with power normalization, i.e., the channel gain g(x, y) ∼ exp(1), is independently distributed with the unit mean.

3) BLOCKAGE MODEL FOR MM-WAVE ACCESS LINKS
For mm-wave channels, LOS transmissions are vulnerable to significant penetration losses [28]; thus LOS transmissions can be blocked with a certain probability.Following [19], [29], [33], [34], we consider the actual LOS region of a device as a fixed LOS ball referred to as "equivalent LOS ball".For the sake of mathematical tractability, we consider a distance dependent blockage probability p(r) that a mmwave link of length r observes, i.e., the LOS probability p L (r) if the mm-wave desired link length is less than d and p N (r) otherwise.As such, there is a dependency in terms of the link distance r.That is, SBSs within a LOS ball of radius d are marked LOS with probability p L (r), while the SBSs outside that LOS ball are marked as NLOS with probability p N (r).Note that we will drop the notation (r) in both p L (r) and p N (r) from this point onwards and we will use only p L and p N , respectively.

D. ASSOCIATION MECHANISM
Each device associates with either a MBS or a SBS depending on the maximum biased received power in the downlink.The association criterion at the typical device can be written mathematically as follows: where , and ζ (•) denote the transmission power, biasing factor, effective antenna gain, and near-field path loss at 1 m of the intended link, respectively, in the corresponding tier (which is determined by the index in the subscript).Let L min,j (r) −1 be the minimum path loss of the typical device from a BS in the j th tier.When a device associates with a mm-wave SBS in tier-2, i.e., k = 2, the antenna gain of the intended link is 1 is defined as the omnidirectional antenna gain of MBSs and G D is the device antenna gain while operating in μwave spectrum.On the other hand, the SBS associates with a MBS offering the maximum received power in the downlink.

E. SNR/SIR MODELS FOR ACCESS AND BACKHAUL TRANSMISSIONS
The device associates to either a MBS for direct transmission or a SBS for dual-hop transmission.The first link (backhaul link) transmissions occur on the μwave spectrum between MBSs and SBSs and the second link (access link) transmissions take place in the mm-wave spectrum between SBSs and devices.Let θ 2 denotes the predefined SIR threshold for SBSs in the backhaul link and θ D denotes the predefined SIR/SNR threshold for devices.Throughout the paper, we use subscripts "1, 2", "2, D", "1, D", "D", "BH" to denote backhaul link, access link, direct link, device, and backhaul, respectively.

1) BACKHAUL TRANSMISSION
The SIR of the typical SBS associated with a MBS is modeled as: where I 1,2 = P 1 y∈ 1 \{y 1,0 } y −α 1 g(0, y) denote the backhaul interference received at a SBS from MBSs that are scheduled to transmit on the same resource block excluding the tagged MBS.

2) DIRECT TRANSMISSION
The SIR of the typical device associated directly with a MBS is modeled as: where I 1,D denotes the interference received at the typical device from MBSs excluding the tagged MBS.Then I 1,D can be calculated as: I 1,D = P 1 y∈ 1 \{y 1,0 } y −α 1 g(0, y).

3) ACCESS TRANSMISSION
The SNR of the typical device associated with a mm-wave SBS is modeled as: where ζ 2 is the near-field path loss at 1 m for mmwave channels, and σ 2 2 is the variance of the additive white Gaussian noise at the device receiver.Given highly directional beams and high sensitivity to blockage, recent studies showed that mm-wave networks are typically noise limited [10], [23], [24].

III. THE META DISTRIBUTION: MATHEMATICAL PRELIMINARIES
In this section, we define the meta distribution of the SIR of the typical device and highlight exact and approximate methods to evaluate the meta distribution.

Definition 1 (Meta Distribution of the SIR and CSP):
The meta distribution FP s (x) is the complementary cumulative distribution function (CCDF) of the CSP (or reliability) P s (θ ) and given by [14]: where, conditioned on the locations of the transmitters and that the desired transmitter is active, the CSP P s (θ ) of the typical device [14] can be given as P s (θ ) = P(SIR > θ | , tx) where θ is the desired SIR.Physically, the meta distribution provides the fraction of the active links whose CSP (or reliability) is greater than the reliability threshold x.
the exact meta distribution can be given using the Gil-Pelaez theorem [35] as [14]: where (w) is imaginary part of w ∈ C and M jt (θ ) denotes the imaginary moments of P s (θ ), i.e., j = √ −1.

Using moment matching techniques and taking
, the meta distribution of the CSP can be approximated using the beta distribution as follows: where M 1 (θ ) and M 2 (θ ) are the first and the second moments, respectively; and B(a, b) is the beta function.

IV. THE META DISTRIBUTION OF THE SIR/SNR IN HYBRID SPECTRUM NETWORKS
To characterize the meta distribution of the SIR/SNR of the typical device that can associate with either a μwave MBS with probability A 1 or with a wireless backhauled mm-wave SBS with probability A 2 , the methodology of analysis is listed as follows: 1) Derive the probability of the typical device associating with μwave MBSs A 1 , LOS mm-wave SBSs A 2,L , and NLOS mm-wave SBSs A 2,N where 2) Formulate the meta distribution of the SIR/SNR of a device in the hybrid network ( Fb P s,T (x)) considering the direct link and dual-hop link with wireless backhaul transmission (Section IV-B).
Derive the b th moments of CSPs, i.e., M b,BH (θ 2 ), M b,2 (θ D ), and M b,1 (θ D ) for backhaul link, access link, and direct link transmissions, respectively (Section V). 5) Obtain the meta distributions of SIR/SNR and data rate in hybrid spectrum network using Gil-Pelaez inversion and the beta approximation (Section VI).

A. ASSOCIATION PROBABILITIES IN HYBRID SPECTRUM NETWORKS
In this subsection, we characterize the probabilities with which the typical device associates with μwave MBSs (A 1 ) or mm-wave SBSs (A 2 ).The results are presented in the following.

Lemma 1 (The Probability of Associating with mmwave SBSs):
The probability of the typical device to associate with a mm-wave SBS, using the association scheme in Eq. ( 2), can be expressed as: where ). Subsequently, the probability of a device to associate with a μwave MBS can be given as A 1 = 1 − A 2 .The conditional association probability for the typical device to associate with SBS is as follows: (10) where f (l 1 ) is given in Appendix A of our technical report [36], subsequently, Â1 (l 1 ) = 1 − Â2 (l 1 ).
Proof: Using the approach in [3], we derive Lemma 1 in Appendix A of our technical report [36].
A closed-form expression of A 1 can be derived for a case of practical interest as follows.
Corollary 1: When α 1 = 4, α 2,L = 2, and α 2,N = 4, then A 1 can be given in closed-form in Eq. ( 11) as shown at the bottom of the page, where (•) is the error function, It can be seen from Corollary 1 that when the number of antenna elements N goes to infinity, i.e., G 2 → ∞, â → ∞, then A 1 can be simplified as , which shows that association probability to MBS will be very small.Similar insights can be extracted for other parameters.
In order to derive the b th moment of CSP P s,2 (θ D ) on an access link when a device associates with a SBS (the CSP will be discussed later in Lemma 3), we have to derive the probability of a device to associate with LOS SBS A 2,L and NLOS SBS A 2,N which are defined follows.
Lemma 2 (The Probability of Associating with LOS and NLOS mm-wave SBSs): When the typical device associates with the mm-wave SBS tier, this typical device has two possibilities to connect to (a) a LOS mm-wave SBS with association probability A 2,L and (b) a NLOS mm-wave SBS with association probability A 2,N which are characterized, respectively, as follows: where Â2,L (l 2,L ) and Â2,N (l 2,N ) are the conditional probabilities with which the typical device may associate to the LOS and NLOS mm-wave SBSs, respectively, and are defined as follows: where f 2,L (l 2,L ) and f 2,N (l 2,N ) are given in [36, Eqn.(10) in Appendix B] and Proof: Using the approach in [20], we derive in [36, Lemma 2 in Appendix B].
A case of interest is when the number of antenna elements at mm-wave SBSs increases asymptotically.In such a case, the LOS and NLOS association probabilities can be given as.
Corollary 2: When the number of antenna elements at mm-wave SBSs increases, i.e., N → ∞, α 1 = 4, α 2,L = 2, and α 2,N = 4, then ā → 0. The association probabilities can be given in closed-form as follows: An interesting insight from Corollary 2 can be seen when the intensity of SBSs λ 2 → ∞ or d is large, the probability of association to LOS SBSs A 2,L becomes almost 1.On the other hand, when λ 2 → 0 or d is small, 1 F 1 [a; b; 0] = 1 thus A 2,N becomes almost 1.

B. FORMULATION OF THE META DISTRIBUTION, CSP AND ITS B TH MOMENT IN THE HYBRID NETWORK
When a device associates with a mm-wave SBS, the overall CSP depends on the CSPs of the SIR and SNR on both the backhaul link and the access link, respectively.On the other hand, when a device associates to MBS the CSP depends on the SIR of the direct link.It is thus necessary to formulate the relationship between the meta distribution, CSP, and its b th moment in the considered hybrid network as follows.
Lemma 3 (Meta Distribution of the Typical Device in the Hybrid Network): The combined meta distribution of the SIR/SNR in the hybrid spectrum network can be characterized as follows: where M jt,T (θ ) can be characterized by deriving the b th moment of the P s,T (θ ). 5 where Step (a) follows from breaking the correlation between 1 and 2 for tractability and the results show that the correlation is weak and also follows from the fact that the b th moment of the SIR or SNR of a device associated to tier i can be defined as where Âi is the conditional association probability to tier i and M b|i = P b s,i 5.The b th moment of a random variable X is the expected value of random variable to the power b, i.e., E[X b ].
is the conditional b th moment of the SIR or SNR in tier i.In our case, we have Â2 (l 2 ) which is the conditional association probability to mm-wave SBS where l 2 ∈ {L, N} since a device can associate to either LOS or NLOS mm-wave SBS.The step (b) follows from the fact that the CSP of the dual-hop transmission depends on the CSP of access and backhaul link; therefore, we have a product of the access and backhaul CSPs, i.e., P s,BH (θ 2 )P s,2 (θ D ) that are independent random variables.There is no correlation since μwave backhaul does not interfere with mm-wave transmissions.The step (c) follows from the fact if X and Y are inde- Finally, the step (d) follows from the definition of Â2 (l 2 ) in Lemma 2 and the step (e) follows by applying the definition of moments.
In the next section, we derive the CSP of access, backhaul, and direct links along with their respective b th moments, as needed in Lemma 4 to characterize the overall moment as well as the meta distribution.

V. CHARACTERIZATION OF THE CSPS AND MOMENTS
In this section, we derive the CSPs P s,BH (θ 2 ), P s,2 (θ D ), P s,1 (θ D ) and the b th moments M b,BH (θ 2 ), M b,2 (θ D ), and M b,1 (θ D ) for backhaul link, access link, and direct link, respectively.

A. CSP AND THE B TH MOMENT -ACCESS LINK
We condition on having a device at the origin which becomes the typical device.The CSP of the typical device at the origin associating with the mm-wave SBS-tier (when k = 2) can be described as follows: The CSP of the SNR of a device on the access link with LOS can be defined by substituting SNR 2,D defined in Eq. ( 5) into Definition 1 as follows: where (a) follows from the definition of and the fact that the channel gain h L (0, y) is a normalized gamma random variable and γ (., .) is the lower incomplete gamma function and (s) = γ (s, x) + (s, x), where (., .) is the upper incomplete gamma function.Similarly, CSP of the SNR on the access link for NLOS case can be given as follows: where As such, the b th moment of the CSP on the access link for the typical device when it is served by the mm-wave SBS tier is given by the following.

B. CSP AND MOMENT -BACKHAUL LINK
For the backhaul link, we condition on having a SBS at the origin which becomes the typical SBS.Using the expression of SIR 1,2 in Eq. (3) the CSP of the backhaul link P s,BH (θ 2 ) can be given as: where (a) follows from the Rayleigh fading channel gain g(0, y) ∼ exp( 1) and (b) is found by taking the expectation with respect to g(0, y).Following [14], the b th moment of the CSP on the backhaul link is given as: where (a) follows from the probability generating functional (PGFL) of the relative distance process generated by a PPP, i.e., G R Lemma 1] and 2 F 1 (., .; .; .)represents Gauss' hypergeometric function where R is the relative distance process defined in [37, Definition 2].

C. CSP AND MOMENT -DIRECT LINK
Using the expression of SIR 1,D in Eq. ( 4), we calculate the CSP of the direct link P s,1 (θ D ) as follows: where (a) follows from the channel gain g(0, y) ∼ exp(1) and is independently exponentially distributed with unit mean and (b) is obtained by taking the expectation with respect to g(0, y).While taking the association probabilities into consideration, the b th moment of the CSP P s,1 (θ D ) of the typical device when it is served by a μwave MBS is characterized in the following lemma.

Lemma 5 [The b th moment of the CSP (P s,1 (θ D )) when a device associates with a MBS]:
The b th moment of the CSP experienced by a device, when the device associates with a MBS, can be characterized in Eq. ( 22) as shown at the bottom of the page. Proof: +1 dv is independent of l 1 , thus where N → ∞ or α 1 = 4, α 2,L = 2, and α 2,N = 4, then we can get a closed-form for the three integral over l 1 using Corollary 1 and Corollary 2.

D. COMBINED B TH MOMENT OF THE CSP IN HYBRID NETWORKS
After substituting the values of M b,BH (θ 2 ), M b,2 (θ D ), and M b,1 (θ D ) in Eq. ( 20), Eq. ( 18), and Eq. ( 22), respectively into the total meta distribution for the entire network in Eq. ( 14), we get the b th moment of the CSP at the typical device as shown in Eq. ( 23) as shown at the bottom of the page.
In the next section, we use the combined b th moment in (23) to compute the meta distributions of SIR/SNR and data rate using Gil-Pelaez inversion and the beta approximation.

VI. COMPUTING THE META DISTRIBUTIONS AND SPECIAL CASES
In this section, we compute the meta distribution of SIR/SNR using Gil-Pelaez inversion and beta approximation by applying the derived result of M b,T .Special cases where b = 1 provides the standard coverage probability and b = −1 provides the mean local delay are discussed.Further, we show how to evaluate the data rate meta distribution from the derived framework.

A. COMPUTING THE META DISTRIBUTION OF SIR/SNR
Technically, substituting b = jt in (23), we should obtain the imaginary moments M jt,T .However, since the expression of M jt,T relies on a binomial expansion of power b, the results cannot be obtained directly through substitution.Therefore, we apply Newton's generalized binomial theorem given as follows.
Definition 2: The binomial expansion for an imaginary power r is given as follows (x + y) r = r k=0 r k x r−k y k using the Isaac Newton's generalized binomial theorem to allow real exponents other than non-negative integers, i.e., imaginary exponent r, as , where (.) k is the Pochhammer symbol, which stands here for a falling factorial.
Applying Definition 2 in step (e) of Appendix C, we then obtain the expression for M jt,T as shown in Eq. ( 24) as shown at the bottom of the page.
The imaginary moments can be substituted in the Gil-Pelaez inversion theorem as in Definition 1 to obtain FP s ,T .Furthermore, we follow [14], [16], [38] to approximate the meta distribution by a beta distribution by matching the first and second moments, which are easily obtained from the general result in Eq. ( 23) by substituting b = 1 and b = 2 to get M 1,T and M 2,T , respectively.Taking , the meta distribution using beta approximation can be given as follows:

B. MEAN AND VARIANCE OF THE LOCAL DELAY
The mean local delay is the mean number of transmission attempts, i.e., re-transmissions, needed to successfully transmit a packet to the target receiver.The mean local delay M −1,T which is the −1 st moment of the CSP of the typical device should be calculated by substituting b = −1 in Eq. ( 23).However, since the expression of M b,T relies on a binomial expansion of power b, the results cannot be obtained directly through substitution.Therefore, we apply binomial theorem for the negative integers as follows: the binomial theorem for a negative integer power n can be given [39] as Applying Eq. ( 26) in step (e) of Appendix C, we then obtain the expression for M −1,T as in Eq. ( 27) at the bottom of the next page.
Remark: In order to better characterize the fluctuation of the local delay, the variance of the local delay also referred to as network jitter can be given by NJ = M −2,T − M 2 −1,T .

C. THE META DISTRIBUTION OF THE DATA RATE IN HYBRID SPECTRUM NETWORKS
Let T denote the data rate (in bits/sec) of the typical device on a specific transmission link which is a random variable and is defined as R = W log 2 (1 + SIR) using Shannon capacity.Using the meta distribution of the SIR, the meta distribution of the data rate can be derived to present the fraction of active devices in each realization of the point process that have a data rate R greater than T with probability at least x, i.e., devices data rate reliability threshold.That is, first deriving the CSP of the data rate as follows: where P s (2 ) denote the CSP of the device data rate over single link.Finally, deriving the b th moment of the CSP of the data rate and applying Gil-Pelaez inversion we can obtain the meta distribution of the data rate.
Corollary 3: Similar to the meta distribution of the SIR/SNR derived in Lemma 3 and conditioned on the location of the point process, we derive the meta distribution of the data rate in hybrid 5G cellular networks, using the moment Q b of the conditional data rate as follows: where P s,1 (2 tx) denote the CSP of the device data rate on the direct, backhaul, and access link, respectively.
In the following section, we discuss the application of this framework to two scenarios (i) μwave only network and (ii) mm-wave backhauls and microwave access links.

VII. EXTENSIONS OF THE MODEL TO OTHER NETWORK ARCHITECTURES
The framework discussed above can be flexibly applied to different network architectures.In this section we discuss how to extend the framework to two other network architectures: 1) both tiers operating in the sub-6GHz (microwave) spectrum as in traditional cellular networks; and 2) the two tiers operating in two millimeter-wave spectrums that are orthogonal to each other.Due to space limitation, we provide only general directions of how to extend the earlier framework to these two other network architectures.

A. THE META DISTRIBUTION OF THE SIR IN MICROWAVE-ONLY CELLULAR NETWORKS
We characterize the meta distribution of the downlink SIR attained at the typical device in a μwave-only cellular network, i.e., the access and backhaul links of SBSs operate in the μwave frequency.A device associates with either a serving MBS for direct transmissions (when k = 1) or a SBS for dual-hop transmissions (when k = 2), depending on the biased received signal power criterion.MBSs and SBSs are assumed to operate on orthogonal spectrums; thus, there is no inter-tier interference.On the other hand, each SBS associates with a MBS based on the maximum received power at the SBS.The association criterion for the typical device can be described as follows [40]: where .denotes the Euclidean distance.The typical device associates with a serving node (given by Eq. ( 30)), which is termed the tagged SBS.For the sake of clarity, we define Pjk = [40], the conditional association probability for the typical device connecting to the k th tier (conditional over the desired link distance r D,k ) is as follows: where n denotes the index of the tier associating with the typical device.We calculate the CSP P s,2 (θ D ) (when k = 2) of the access link operating in the μwave band as follows: where (a) follows from the channel gain g(0, y 2,0 ) ∼ exp (1) and is independently exponentially distributed with unit mean and (b) is obtained by taking the expectation with respect to g(0, y 2,i ).Lemma 6: Using Eq. ( 21) and Eq. ( 32), we calculate a general expression for the b th moment of the CSP on direct link M b,k' (when k = 2) and the b th moment of the CSP at access link (when k = 1) as: Proof: See Appendix C. Note that Lemma 5 is novel and different from [41] where we derive the b th moment of CSP for orthogonal spectrum two tier network while the work in [41] is done for shared spectrum tiers.Similarly, the moment of the CSP of the typical device with offloading biases is as follows: where M b,BH (θ 2 ), M b,2 (θ D ), and M b,1 (θ D ) are defined in Eq. ( 20), Eq. ( 33) (when k = 2), and Eq. ( 33) (when k = 1), respectively.The step (a) follows from the similar approach as taken in Lemma 4.

B. EXTENSIONS TO MILLIMETER-WAVE BACKHAULS NETWORKS
The proposed framework can be extended to a scenario where the backhaul and access transmissions are conducted on orthogonal mm-wave spectrums.Note that Eq. ( 3) will be changed similar to Eq. ( 5).Then, only the first term, M b,BH (θ 2 ) in the main Eq. ( 14) of our model that characterizes the moment of the CSP in the backhaul will be re-defined as The framework can also be extended to a scenario where the backhaul transmissions are conducted on the mm-wave spectrum and the access links of SBSs operate on μ-wave.In this case, we will need to use the results in Section VII-A while redefining the term (35).

VIII. NUMERICAL RESULTS AND DISCUSSIONS
We present the simulation parameters in Section VIII-A.Then, we validate our numerical results using Monte-Carlo simulations in Section VIII-B.In Section VIII-B, we use the developed analytical models to obtain insights related to the meta distribution of the SIR/SNR of the typical device, mean and variance of the success probability, transmission delay, and the reliability of the typical device in the downlink direction.

A. SIMULATION PARAMETERS
Unless otherwise stated, we use the following simulation parameters throughout our numerical results.The transmission powers of MBSs and SBSs in the downlink are P 1 = 50 Watts and P 2 = 5 Watts, respectively.The size of the simulated network is 90km × 90km.We assume that the density of MBSs is λ 1 = 2 MBSs/km

B. NUMERICAL RESULTS AND DISCUSSIONS 1) COVERAGE AND VARIANCE AS A FUNCTION OF SIR/SNR THRESHOLD IN HYBRID SPECTRUM NETWORKS
Fig. 2 illustrates the standard success probability M 1,T and its variance M 2,T − M 2  1,T as a function of target SIR/SNR threshold θ of devices in a hybrid spectrum network.As we can see in Fig. 2 that the simulation results match the analytical results, however the slight gap is due to the Alzer's inequality considered in Appendix C.This gap will be zero when Nakagami fading turns into Rayleigh fading as shown in the next figure.By examining Fig. 2, a numerical evaluation shows that the variance is maximized at θ = −3 dB where the success is M 1,T = 0.49.For moderate values of θ , there is a trade-off between maximizing coverage or reducing variance because the variance first increases and then decreases while the coverage probability is monotonically decreasing.For higher values of θ , lower coverage probabilities have lower variance so its a low-reliability regime where more devices' performances are spread around low coverage probability.As such, the low values of θ provides a higher reliability regime.Fig. 3 illustrates the standard success probability M 1,T and the variance as a function of θ with Rayleigh fading (i.e., m L = m N = 1).As we can see in Fig. 3 that the simulation results closely match the analytical results.The reason is that the approximation of the incomplete gamma function (also referred to as Alzer's inequality) becomes exact when m L becomes equal to unity.Subsequently, this figure explains the reason for the gap between the simulation and the analytical curves in Fig. 2.

2) COVERAGE AND VARIANCE AS A FUNCTION OF THE NUMBER OF ANTENNA ARRAY ELEMENTS IN HYBRID SPECTRUM NETWORKS
Fig. 4 depicts the coverage probability and variance as a function of θ considering the number of antenna array elements as N = 10, 20, and 30 to show the effect of higher directional antenna gains.The general trends for the coverage probability and its variance are found to be the same as in previous figures.The main observation is that although the coverage enhancement is not significant with increasing antenna elements, the reduction in the variance is noticeable which supports higher directional antenna gains and the importance of analyzing the higher moments of the CSP.

3) COVERAGE AND VARIANCE AS A FUNCTION OF B 2 IN μWAVE-ONLY NETWORKS
In Fig. 5, we study the effect of offloading devices from the MBS tier to the SBSs tier in terms of the coverage probability (which is the mean reliability) and the variance of the CSP (or reliability).By offloading devices from the MBS tier to the SBSs tier when B 2 = 30, the coverage probability M 1,T suffers due to the dual-hop transmission effect in wireless backhauled SBSs; however the variance of the results reduces which is a novel and positive insight.Another observation is that the variance of the CSP in μwave-only network is high compared to the hybrid network.This can be shown by comparing points V 1 = (1, 0.1) in Fig. 4 and V 2 = (4, 0.19) in Fig. 5, for the case of B 1 = B 2 = 1.We noticed that the variance has decreased from 0.19 to 0.1 when the SBS antenna array size is increased to N = 20.This implies that the hybrid spectrum network outperforms the μwave-only network due to the directional antenna gains.

4) MEAN LOCAL DELAY (μWAVE VS MM-WAVE SBSS)
Fig. 6 depicts the mean local delay experienced by the typical device as a function of the SBSs density λ 2 in a hybrid spectrum network.The mean local delay is the mean number of transmission attempts to successfully transmit a packet.The mean local delay increases by increasing λ 2 .After the SBS density reaches λ 2 = 20 SBSs/km 2 , the mean local delay stays constant at value 1.11.This result can be intuitively explained as follows.When the mm-wave SBS density is low, the typical device has a higher probability to connect to a MBS, i.e., the mean local delay of the network results from only one hop communication (from the MBS to the device).However, when the λ 2 increases, the typical device has a higher probability to connect to a mm-wave SBS, i.e., the network local delay results from two hops communication (from the MBS to the SBS then from the SBS to the device).Furthermore, the beamforming high directional gain steerable antennas will push more devices to associate with SBSs thus a higher network delay is observed.Fig. 7 shows that, all else being equal, the mean local delay of the hybrid spectrum network is lower than that of the μwave-only network.Fig. 7 depicts the mean local delay for a μwave-only network as a function of λ 2 .When λ 2 increases the mean local delay of the total network increases again due to the increase in interference which is not the case in the hybrid spectrum network.The network mean local delay in the case of α 1 = α 2 = 3 is higher than that in the case of α 1 = α 2 = 4 due to higher path loss degradation for higher PLEs.8, the fraction of devices achieving a required rate increases as the number of antennas elements increases.In other words, increasing the number of antenna elements of SBSs has a positive effect on the achievable rate and its meta distribution.This insight helps 5G cellular network operators to find the most efficient operating antenna configuration to achieve certain reliability for certain 5G applications.

6) THE META DISTRIBUTION IN A MICROWAVE-ONLY NETWORK
In Fig. 9, we validate our analysis by depicting the exact (Gil-Pelaez) meta distribution in a μwave-only network defined in Eq. ( 13), and the beta approximation for the meta distribution defined in Eq. ( 8).Our simulation result provides a good match for a wide range of θ values and this validates the accuracy of our analytical model.The small gap between the analytical model and simulation results when θ = 1 is due to breaking the correlation between 1 and 2 for tractability in Lemma 3. Fig. 9 also serves as an illustration of the meta distribution of the SIR of the typical device in a μwave-only network.We note that about 23% of devices (when θ = 10), 72% of devices (when θ = 1), and 98% of devices (when θ = 0.1) have reliability, i.e., success probability, at least 0.3.

IX. CONCLUSION
This paper characterizes the meta distributions of the SIR/SNR and data rate of the typical device in a hybrid spectrum network.The meta distribution is evaluated first by formulating and then characterizing the moments of the CSP of the typical device in the hybrid network.Important performance metrics such as the mean local delay, coverage probability, network jitter, and variance of the CSP (or reliability) are studied.Numerical results demonstrate the significance of evaluating the meta distribution which requires a systematic evaluation of the generalized moment of order b that helps in evaluating network metric such as coverage probability when b = 1, mean local delay when b = −1, network jitter using b = −2 and b = −1, etc.

APPENDIX A PROOF OF LEMMA 4
The b th moment of the CSP of the typical device served by the mm-wave SBS is derived as:

APPENDIX B PROOF OF LEMMA 5
The b th moment of the CSP of the typical device when associated to μwave MBS is derived as: where (a) follows from taking expectation over l 1 = r α 1 and considering the conditional association probability for the typical device connecting to the MBSs tier given in Lemma (1) and substituting the value of P s,1 (θ D ) from Eq. (21).In step (b) we apply PGFL of the PPP [43,Ch. 4].
Step (c) follows from averaging over l 1 .In step (d), we use the change of variable v = l 1 y α 1 , dy =

APPENDIX C PROOF OF LEMMA 6
While taking the association biases effect in consideration, the b th moment of the CSP P s,k (θ D ) of the typical device when it is served by the k th tier is given as follows: where (a) follows from considering the conditional association probability for the typical device connecting to the k th tier given in Eq. (31).In step (b), we apply PGFL of the PPP [43,Ch. 4].
Step (c) follows from averaging over r k,D , step (d) is by using variable substitution q = πλ k r 2 and v = r/y.In step (e), we perform variable substitution v = u( Pjk Bjk ) −1/α j and step (f) follows from the fact that
M b,Dual−Hop is the b th moment of the SIR/SNR when a device associates to mm-wave SBS for dual-hop transmission and M b,Single−Hop is the b th moment of the SIR when a device associates to MBS for direct transmission.After reformulation, we define M b,BH (θ 2 ) as the unconditional b th moment of the backhaul SIR, M b,2 (θ D ) as the unconditional b th moment of the SNR at access link when a device associates to mm-wave SBS, and M b,1 (θ D ) as the unconditional b th moment of the SIR at direct link when a device associates to μwave BS.Note that P s,1 (θ D ) = P(SIR 1,D > θ D | 1 , tx) denotes the CSP of device over the direct link, P s,BH (θ 2 ) = P(SIR 1,2 > θ 2 | 1 , tx) denotes the CSP at backhaul link, and P s,2 (θ D ) = P(SNR 2,D > θ D | 2 , tx) denotes the CSP for the access link transmission.Proof:

2 and the density of SBSs is λ 2 =
70 SBSs/km 2 .The offloading biases for the MBSs and the SBSs are B 1 = B 2 = 1, respectively.The PLE for MBSs is set to α 1 = 4 and for mm-wave SBSs, α 2,L = 2 in the case of LOS and α 2,N = 4 in the case of NLOS.The network downlink bandwidth is 100 MHz for μwave MBSs and 1 GHz for mm-wave SBSs with channel frequency 28 GHz.The LOS (NLOS) states are modeled by large (small) values of m, i.e., m L = 2 and m N = 1 [20].SBSs number of antenna elements is N = 10.The receiver noise is calculated as [10] σ 2 2 = −174 dBm/Hz + 10 log 10 (W 2 ) + 10 dB, where W 2 = 1 GHz is bandwidth allocated to the mm-wave SBSs.The antenna gains of MBSs are G o 1 = 0 dB and devices directional antenna gain is G max D = 10 dB.

5 )
Fig.8depicts the meta distribution of the data rate in hybrid spectrum networks as a function of reliability x for different number of antenna elements N = 10, 20, 40, and 50 with rate threshold T = 1 Gbps.As shown in Fig.8, the fraction of devices achieving a required rate increases as the number of antennas elements increases.In other words, increasing the number of antenna elements of SBSs has a positive effect on the achievable rate and its meta distribution.This insight helps 5G cellular network operators to find the most efficient operating antenna configuration to achieve certain reliability for certain 5G applications.