Energy Efficiency in Cache Enabled Small Cell Networks With Adaptive User Clustering

TL;DR

This paper explores how adaptive user clustering based on content popularity profiles can significantly enhance energy efficiency in cache-enabled small cell networks by optimizing content placement and small cell deployment.

Contribution

It introduces a novel user clustering framework based on content popularity profiles and derives a closed-form EE expression using stochastic geometry.

Findings

Clustering improves cache hit probability substantially.

Optimized small cell density maximizes energy efficiency.

Adaptive caching outperforms unclustered approaches.

Abstract

Using a network of cache enabled small cells, traffic during peak hours can be reduced considerably through proactively fetching the content that is most probable to be requested. In this paper, we aim at exploring the impact of proactive caching on an important metric for future generation networks, namely, energy efficiency (EE). We argue that, exploiting the correlation in user content popularity profiles in addition to the spatial repartitions of users with comparable request patterns, can result in considerably improving the achievable energy efficiency of the network. In this paper, the problem of optimizing EE is decoupled into two related subproblems. The first one addresses the issue of content popularity modeling. While most existing works assume similar popularity profiles for all users in the network, we consider an alternative caching framework in which, users are clustered according to their content popularity profiles. In order to showcase the utility of the proposed clustering scheme, we use a statistical model selection criterion, namely Akaike information criterion (AIC). Using stochastic geometry, we derive a closed-form expression of the achievable EE and we find the optimal active small cell density vector that maximizes it. The second subproblem investigates the impact of exploiting the spatial repartitions of users with comparable request patterns. After considering a snapshot of the network, we formulate a combinatorial optimization problem that enables to optimize content placement such that the used transmission power is minimized. Numerical results show that the clustering scheme enable to considerably improve the cache hit probability and consequently the EE compared with an unclustered approach. Simulations also show that the small base station allocation algorithm results in improving the energy efficiency and hit probability.

Equations

$$g_{us}=\sqrt{r_{us}^{-\alpha }}{h}_{us},$$

$$\displaystyle d\big{(}N_{i},N_{c}\big{)}=$$

$$\displaystyle\int_{[0,1]^{F}}\prod_{u=1}^{U}\mathbb{P}_{u}(P_{u})\;\text{log}\big{(}\frac{\prod_{u=1}^{U}\mathbb{P}_{u}(P_{u})}{\prod_{u=1}^{U}\mathbb{P}_{u}(P_{u}|N_{i})}\big{)}\text{d}\;p_{1}...\text{d}\;p_{F},$$

$$\displaystyle d\big{(}N_{i},N_{c}\big{)}=\mathbb{E}\left\{-2\;\text{log}\;(L_{\xi_{N_{i}}}(N_{i}|P_{u},u=1...U))\right\},$$

$$\displaystyle\mathbb{E}\left\{d\big{(}N_{i},N_{c}\big{)}\right\}\approx 2k_{i}-2\text{log}\;(L_{\xi_{N_{i}}}(N_{i}|P_{u},u=1...U)).$$

$$\text{AIC}({\xi }_{N_i})=2k_i-2\text{log}(L_{{\xi }_{N_i}}(N_i|P_u,u=\mathrm{1...}U)).$$

$$\text{log}(L_{{\xi }_{N_i}}(N_i|P_u,u=\mathrm{1...}U))=$$

$$\displaystyle\sum_{u=1}^{U}\big{(}\text{log}(\frac{1}{\sqrt{2\pi}\hat{\sigma}_{\psi(u)}^{F}})-\frac{\lVert P_{u}-\hat{P}_{\psi(u)}\rVert^{2}}{2\hat{\sigma}_{\psi(u)}^{2}}+\text{log}(\frac{U_{\psi(u)}}{U})\big{)},$$

$${\hat{\sigma }}_{\psi (u)}^2=\frac{1}{(U_{\psi (u)})}\sum _{j\in U_{\psi (u)}}\Vert P_j-{\hat{P}}_{\psi (u)}{\Vert }^2.$$

$$\text{log}(L_{{\xi }_{N_i}}(N_i|P_u,u=\mathrm{1...}U))=$$

$$\sum _{k=1}^{N_i}-\frac{U_k}{2}(\text{log}(2\pi )-1+2\text{log}(\frac{U_k}{U})-F\text{log}({\hat{\sigma }}_k^2)).$$

$$\text{AIC}({\xi }_{N_i})=$$

$$2N_i(F+1)+\sum _{k=1}^{N_i}{U}_k(\text{log}(2\pi )-1+2\text{log}(\frac{U_k}{U})-F\text{log}({\hat{\sigma }}_k^2)).$$

$${\xi }_{\text{AIC}}=\underset{\xi \in \Xi }{\mathrm{argmin}}\text{AIC}(\xi ).$$

$${\hat{P}}_k=\frac{{\sum }_{u,\psi (u)=k}{P}_u}{U_k}.$$

$$\Sigma =\frac{SE}{{\rho }_{total}^c},$$

$$\displaystyle\mathbb{P}\left\{hit\right\}=\frac{1}{U}\sum_{k=1}^{N_{c}}\sum_{u=1}^{U}\big{(}\sum_{i\in\Delta_{k}}p_{iu}\big{)}\big{(}1-e^{-\lambda_{sk}\pi R^{2}}\big{)},$$

$${\rho }_{total}^c=\mathbb{E}\left\{{\rho }_I\right\}+\mathbb{E}\left\{{\rho }_T\right\}+\mathbb{E}\left\{{\rho }_f\right\},$$

$$\mathbb{E}\left\{{\rho }_I\right\}=\rho {\lambda }_s\pi R_n^2,$$

$$\displaystyle\mathbb{E}\left\{\rho_{f}\right\}=\lambda_{s}\pi R^{2}_{n}\big{(}\rho_{hd}\mathbb{P}\left\{hit\right\}+\rho_{bh}\big{(}1-\mathbb{P}\left\{hit\right\}\big{)}\big{)},$$

$$\displaystyle\mathbb{E}\left\{\rho_{T}\right\}=\frac{\lambda_{s}\pi R^{2}_{n}}{U}\sum_{k=1}^{N_{c}}\sum_{u\in\Upsilon_{k}}\big{(}\mathbb{E}\left\{\rho_{k}\right\}$$

$$\displaystyle+\sum_{j\neq k}\sum_{i\in\Delta_{s}}p_{iu}(1-e^{-\lambda_{sj}\pi R^{2}})(\mathbb{E}\left\{\rho_{j}\right\}-\mathbb{E}\left\{\rho_{k}\right\})\big{)},$$

$$\displaystyle\rho^{c}_{total}=\lambda_{s}\pi R^{2}_{n}(\rho_{hd}\mathbb{P}\left\{hit\right\}+\rho_{bh}\big{(}1-\mathbb{P}\left\{hit\right\}\big{)}+\rho)$$

$$+\frac{{\lambda }_s\pi R_n^2}{U}\sum _{k=1}^{N_c}\sum _{u\in {{\rm Y}}_k}(\frac{{\rho }_0\gamma (\frac{\alpha }{2}+1,\pi {\lambda }_{sk}{R}^2)}{({\lambda }_{sk}\pi {)}^{\frac{\alpha }{2}}}$$

$$+\sum _{j\ne k}\sum _{i\in {\Delta }_s}{p}_{iu}(1-e^{-{\lambda }_{sj}\pi R^2})(\frac{{\rho }_0\gamma (\frac{\alpha }{2}+1,\pi {\lambda }_{sj}{R}^2)}{({\lambda }_{sj}\pi {)}^{\frac{\alpha }{2}}}$$

$$-\frac{{\rho }_0\gamma (\frac{\alpha }{2}+1,\pi {\lambda }_{sk}{R}^2)}{({\lambda }_{sk}\pi {)}^{\frac{\alpha }{2}}})).$$

$${\rho }_{total}^{nc}={\lambda }_s\pi R_n^2{\rho }_{bh}+\rho {\lambda }_s\pi R_n^2+{\lambda }_s\pi R_n^2\frac{{\rho }_0\gamma (\frac{\alpha }{2}+1,\pi {\lambda }_s{R}^2)}{({\lambda }_s\pi {)}^{\frac{\alpha }{2}}}.$$

$$\text{SINR}=\frac{{\rho }_0{\Vert h_u\Vert }^2}{{\sigma }^2+{\sum }_{k=1}^{N_c}{I}_k},$$

$$\mathbb{P}\left\{\text{SINR}\ge \theta \right\}=exp(-\frac{\theta }{{\rho }_0}{\sigma }^2)\times$$

$$\displaystyle\prod_{k=1}^{N_{c}}exp\big{(}-{\pi\lambda_{sk}}\Gamma(1+\frac{2}{\alpha})\Gamma(1-\frac{2}{\alpha})(\frac{\theta}{\rho_{0}})^{\frac{2}{\alpha}}{\mathbb{E}}\left[\rho^{\frac{2}{\alpha}}_{k}\right]\big{)}.$$

$$SE={\lambda }_s\pi R_n^2\text{log}(1+\theta )\mathbb{P}\left\{\text{SINR}\ge \theta \right\}.$$

$$\Sigma =\frac{{\lambda }_s\pi R_n^2\text{log}(1+\theta )\mathbb{P}\left\{\text{SINR}\ge \theta \right\}}{{\rho }_{total}^c}.$$

$$\underset{{\Lambda }_s}{\mathrm{maximize}}$$

$${\rho }_{total}^c-{\rho }_{total}^{nc}\le 0,$$

$${\Lambda }_s^{†}1\le {\lambda }_{s_{max}}.$$

$$\nabla L({\Lambda }_s^{\ast },\varsigma ,\kappa )=\nabla \Sigma ({\Lambda }_s^{\ast })+\varsigma \nabla C({\Lambda }_s^{\ast })+\kappa \nabla H({\Lambda }_s^{\ast })=0,$$

Full text

I Introduction

The global mobile traffic is expected to increase exponentially in the coming years. This is mainly due to the wide spread of wireless devices and the emergence of video streaming as one of the main contributors in mobile data traffic. Network densification through the deployment of small base stations (SBSs) was proposed as a mean to offload an important amount of traffic from the macro base stations. However, this requires the deployment of a costly infrastructure and a considerable increase in the backhaul link capacity. Recently, information centric networks are emerging as an efficient technology to offload traffic and reduce the strains on the backhaul. In fact, a substantial part of the mobile traffic is due to several duplicate requests of the same popular contents. Consequently, proactively caching these files in the edge of the network results in enhancing user experience while reducing the needed backhaul link capacity. In addition to satisfying the increasing demand, sustainable development is also a major requirement for 5G networks. In fact, the increasing concern regarding green gazes emission made EE a major key performance indicators for future networks. Thanks to the improvement of memory devices, proactive caching offers a very practical and energy efficient alternative to network densification.

The idea of caching popular content on the edge of the network is gaining momentum as one of the most promising enablers of future generation networks [2], [3]. Based on the observation that the backhaul link is becoming a bottleneck, especially in dense networks, the idea of exploiting memory devices as a substitute to more backhaul capacity proves to be very tempting. Many existing works investigated the fundamental trade-offs in cache enabled networks. In [4], a joint routing and caching problem in small cell networks was considered, taking into account both the constrained storage and transmission bandwidth capacities of the SBSs. The authors used approximation algorithms with performance guarantees in order to derive a solution that maximizes the content requests that are satisfied by the SBSs. In [5], an information theoretic formulation of the caching problem was considered. The authors proposed coded schemes enabling a considerable improvement in peak rate compared to previously known schemes. In [6], with a limited backhaul capacity and proactive caching exploiting context awareness and social networks, the authors showed that the backhaul traffic load can be substantially reduced. A hierarchical caching system with two layers of caches was considered in [7]. The authors proposed a coded caching scheme that attains the optimal memory-rate trade-off to within a constant gap. The relation between collaboration distance and interference was studied in [8] for D2D networks. The authors showed that with enough content reuse, non-vanishing throughput per user can be attained, even with limited storage and delay. Clustering users according to their request pattern was investigated in [9] with the goal of reducing service delay. The authors showed that the clustering scheme outperforms the unclustered and random caching approach.

EE of cache enabled networks is a fundamental subject that is attracting increasing attention [10, 11]. The impact of proactive caching on EE was investigated in [13]. The key factors that impact the EE of cache enabled networks were studied. The authors showed that EE can be improved by caching at the base stations when, power efficient cache hardware and sufficient cache capacity are used. In [10], a GreenDelivery framework was proposed in cache enabled small cell networks. Using energy harvesting communications, the authors showed that this framework enables to reduce energy consumption. In [14], the energy consumption of cache enabled wireless cellular networks was investigated. The authors studied the conditions under which the area power consumption is minimized while ensuring a high coverage probability.

In this paper, we explore the impact of the content placement strategy on EE. While most previous works assume similar popularity profiles for all users, we consider an alternative caching framework in which, users are clustered according to their content popularity profiles [1, 9]. This choice is motivated by the existence of very diverse traffic patterns among users. In fact, the requested content depends on the user social network and interests that can be very different from one person to the other. Assuming a homogeneous content popularity among users can only result in loosing valuable information. Contrary to classical location based clustering approaches, we choose to use content based clustering. This can be justified by the fact that content popularity change slower than user locations. Owing to the different time scales of content popularity and user location changes, the selected cached files, which depends on the average popularity distribution per cluster, can be kept constant for long periods of time regardless of user positions.

In order to showcase the pertinence of the proposed popularity based clustering scheme, we use a statistical model selection criterion, namely, Akaike information criterion. AIC enables to measure the truthfulness of a given statistical model. It also addresses the trade-off between the fitness of a statistical model based on maximum likelihood estimation and its complexity which is given by the number of parameters to be estimated. AIC enables to adapt user clustering to any traffic pattern changes since it can detect modifications in the optimal number of clusters. We find that content popularity based clustering enables a substantial gain in term of cache hit probability and EE, even when user positions are not taken into consideration. Nevertheless, further improvement is possible by exploiting geographic information. Of course, acquiring information on user location requires a non negligible processing and signaling overhead. Consequently, this information should be leveraged whenever it is available. In the second part of the paper, we develop an optimization framework to exploit any spatial correlation in traffic patterns.

In this paper, we choose to tackle the problem of file placement once content based clustering is done. The choice is motivated by the different time scales according to which content popularity and user location evolve. In fact, while the correlation in content popularity between users from the same social group is constant for long periods of time, their location can change due to mobility. This motivates the need to adapt the cached content placement more often than the selected files in order to simplify the management of the network. In the case of low mobility, where users do not change positions too often, it makes sense to adapt the files placement based on location information. This is the case for users in confined areas (office, university campus….). To this end, we propose an optimization framework that enables to exploit any spatial correlation between users with comparable popularity profiles. We consider a setting in which user and SBS location is known based on a snapshot of their Poisson point processes (PPP). The proposed combinatorial problem aims at associating each individual SBS with the files of a given cluster. This optimization enables a more energy efficient scheme that exploit the spatial correlation in user traffic pattern in order to reduce the average consumed power. It also showcases another interesting advantage of content based clustering. In fact, the clustering that is done on the users enables also to group the files accordingly. Consequently, the complexity of the resulting optimal file placement problem is lower since the search space is reduced from the whole file catalog to groups of file of approximately equal total size. This considerably simplifies the management of the caching system compared to existing work on location based optimization where, the complexity of the formulated problems is proportional to the number of files.

I-A * Contribution and outcomes*

The main contributions of our work are presented as follows:

A clustering framework for caching: Given heterogeneous user profiles, we propose a content popularity based clustering scheme. In order to achieve an efficient user grouping, we use the Akaike Information Criterion. This allows to effectively estimate the number of clusters and the associated average popularity profiles. 2. 2.

Optimal active SBSs density: We derive a closed form expression of the achievable EE. We then optimize the achievable EE with respect to the density vector of active SBSs. This results in improving the achievable EE even when user positions are not taken into consideration. This is useful in practical scenarios where acquiring user locations requires substantial signaling. 3. 3.

Optimizing small base station allocation: When information about user location is available, further improvement of EE can be achieved by optimizing the SBSs allocation to the different clusters. We formulate a combinatorial content placement problem that enables to adapt the allocation of cached content based on user location.

The paper is organized as follows: We describe the system model in Section II. User clustering will be investigated in Section III. EE will then be addressed in section IV. In section V, we present the SBSs allocation algorithm. Finally, in Section VI, numerical results are presented.

II SYSTEM MODEL AND PRELIMINARIES

II-A Network Model

We consider a small cell network deployed over a disc with radius $R_{n}$. The SBSs are spatially distributed according to a homogeneous Poisson point process $\phi_{s}$ with density $\lambda_{s_{max}}$. In this paper, the available SBSs can be in idle or active modes. The density of active SBSs is given by $\lambda_{s}$ such that $\lambda_{s}\leq\lambda_{s_{max}}$. We consider an orthogonal frequency-division multiple-access (OFDMA) system where, users served by the same SBS, are scheduled on orthogonal resources. Consequently, each user will be subject to interference coming from users served by other SBSs. The users are also distributed in $\mathbb{R}^{2}$ according to an independent homogeneous PPP $\phi$ with density $\lambda$ such that, $\lambda>>\lambda_{s_{max}}$. The average number of users in the network is then given by $U=\lambda\pi R^{2}_{n}$. Each user is equipped with a single antenna and is allowed to communicate with any SBS within a radius $R$. This restriction enables to control the level of interference. We consider that $R$ is defined so that each user is covered with high probability by more than one SBS. We consider that a packet can be successfully transmitted and decoded if and only if SINR $>\theta$. This means that, if the SINR is lower than the threshold $\theta$, the link undergoes an outage and the transmission fails. A general power law pathloss model is used where, the power decay is given by $r^{-\alpha}_{us}$. $r_{us}$ represents the distance between user $u$ and its serving SBS $s$ and $\alpha>2$ denotes the pathloss exponent.The wireless channel from user $u$ to the SBS $s$ is then given by:

[TABLE]

where ${h}_{us}$ represents the small scale fading coefficient modeled as Rayleigh fading i.e., $CN\left(0,1\right)$ distributed random variable. We consider that the transmit power, used in both uplink and downlink, is defined according to channel inversion power control [21]. This is done so that the transmit power compensates the pathloss in order to keep the average signal power at the receiver (i.e., the SBS or the user terminal) equal to a certain constant value $\rho_{0}$. The transmit power used by user $u$ to communicate with SBS $s$, according to channel inversion power control, is given by: $\rho_{us}=\rho_{0}r^{\alpha}_{us}$. The channel inversion power control will ensure a limitation of the interference level since the power received at any base station from a typical user is upper bounded by $\rho_{0}R^{\alpha}$, where $R$ denotes the maximum communication radius. Controlling the level of interference, in both uplink and downlink, is a vital factor that guarantees an EE gain [13].

II-B User scheduling and caching strategy

We consider a file catalog $C$ containing $F$ files with different sizes. Each file $i$ has a size of $L_{i}$ bits. In this paper, we consider that the users have heterogeneous file popularity distributions. Each user $u$ is associated with a popularity vector $P_{u}=\left[p_{1u}...p_{Fu}\right]$, where $p_{iu}$ denotes the probability that user $u$ requests file $i$ from the catalog. We consider that these probabilities change slowly over time and that they are previously known by the network. Estimating the popularity distributions can be performed by learning from previously recorded requests [22]. In this work, we limit our analysis to the case of perfectly known popularity distributions. The study of the impact of estimation error in popularity distributions is considered in future work. Although users have heterogeneous popularity profiles, we assume that they can be grouped according to their interest into $N_{c}$ clusters. This means that the users, forming each cluster, have correlated request patterns. Meaning that the distance between their content popularity vectors is small. Each SBS is equipped with a caching capacity of $M$ bits. Each individual SBS fills its memory device with the most popular files from a given cluster. In each cluster, the most popular files are selected based on the average of the popularity vectors associated with the users forming this cluster. Therefore, appropriately clustering the users based on the similarity in their popularity distributions is of paramount importance. In this paper, not all SBSs are required to be active. We consider the density vector of active SBSs $\Lambda_{s}\in\mathbb{R}^{N_{c}\times 1}$, where each of its coefficients $\lambda_{sk},\;k=1..N_{c}$ represents the density of active SBS caching the most popular files of cluster $k$. $\Lambda_{s}$ is defined such that $\Lambda_{s}^{\dagger}\mathbf{1}=\lambda_{s}$. Each user looks for the requested file in the cache of the SBSs within a radius $R$. The user starts with the closest SBS from his own cluster. If the requested file is available in a cache within this distance, a cache hit event occurs and the user will associate with the closest SBS storing the requested file. In the event of a cache miss, the user, simply, associates with the nearest SBS from its corresponding cluster and the requested content will be retrieved from the core network through the backhaul. If a user cannot find an SBS from its own cluster within a radius of $R$, it will only communicate with SBSs from other clusters within radius $R$, in the case of a cache hit event. An example of the considered model is represented in Figure $1$ with three popularity based clusters represented, each, by a color.

III Information theoretic approach to user clustering

Proactive caching systems require an efficient characterization of content popularity in order to correctly predict the files that are most likely to be requested. While most of the existing work assume similar popularity distributions for all users, we adopt a content based clustering approach. We argue that clustering users according to their popularity distributions enables to better assess social similarities [1, 9] and, consequently, devise a more efficient caching system. In fact, supposing that all users in the network have similar popularity distributions means that the resulting statistics are just an average of content popularity over all social groups. This leads to neglecting the diversity of social behavior. Contrary to traditional location based clustering methods, content aware clustering enables to identify the main request patterns in the network which leads to a better understanding of user preferences. In this work, content popularity based clustering is considered. Users are grouped such that, the correlation between the popularity profiles of users in the same cluster is maximized. This correlation is characterized by the euclidean distance between their content popularity vectors. While content based clustering was proposed in [9] using a spectral approach, we choose to adopt an information theoretic method, namely, Akaike information criterion. AIC allows to efficiently estimate the number of clusters and to assess the information loss that results from assuming a single popularity distribution.

III-A Cluster estimation: Akaike information criterion

In the considered setting, the users have heterogeneous popularity profiles. However, the different social relations and interactions may result in some correlation in user request patterns. Consequently, content popularity based clustering is used in order to minimize the divergence among user content popularity distributions in each cluster. The number of content popularity based clusters is unknown a priori and should be estimated. Allowing the system to estimate this parameter periodically or whenever a substantial change in user interest is recorded, allows the network to cope with any modification in user request pattern.

In order to estimate the number of clusters, we use AIC [15] as a statistical model selection criterion. It allows to assess the quality of a statistical models for a given set of data. The data set to be modeled in our case is the collection of user content popularity distributions. Using AIC enables to estimate the expected Kullback-Leibler discrepancy between the data generating model and any candidate statistical model. It also addresses the trade-off between the fitness of the statistical model, based on maximum likelihood estimation, and its complexity, which is given by the number of model-characterizing parameters to be estimated.

In our case, we aim at modeling the distribution that generates the user’s popularity vectors. We consider the true generating distribution $A(P_{1},...,P_{U})=\prod_{u=1}^{U}\mathbb{P}_{u}(P_{u})$, where $\mathbb{P}_{u}(P_{u})$ is the probability that user $u$ has a popularity vector $P_{u}$. We assume that $A(P_{1},...,P_{U})$ results from the aggregating of $N_{c}$ user clusters where, $N_{c}$ denotes the true number of clusters. Let $\xi_{N_{i}},i\in[c_{min},..,c_{max}]$ be a set of approximation models. Each approximation model $\xi_{N_{i}}$ is characterized by $N_{i}$ clusters and a popularity generating distribution $\prod_{u=1}^{U}\mathbb{P}_{u}(P_{u}|N_{i})$. The popularity generating distribution depends on the number of clusters $N_{i}$. In fact, the average and variance of content popularity vectors in each cluster depend, primarily, on $N_{i}$. The Kullback-Leibler information, which characterizes the information lost when an approximating model is used, can be written, $\forall\;\xi_{N_{i}},i\in[c_{min},..,c_{max}]$, as:

[TABLE]

After simplification, the discrepancy between the two models is given by [16]:

[TABLE]

where $\mathbb{E}\left\{.\right\}$ denotes the expectation with respect to the available data, which is the collection of all users popularity profiles, knowing $N_{i}$. $L_{\xi_{N_{i}}}(N_{i}|P_{u},u=1...U)$ denotes the likelihood of having $N_{i}$ clusters, knowing the popularity profiles of the users (the expression will be given later on in this section).

In [16], Akaike noted that $-2\;\text{log}\;(L_{\xi_{N_{i}}}(N_{i}|P_{u},u=1...U))$ is a biased estimate of the average discrepancy. After bias adjustment, the expected discrepancy can be approximated by:

[TABLE]

Here $k_{i}$ denotes the number of characterizing parameters in model $\xi_{N_{i}}$ and $\mathbb{E}\left\{.\right\}$ denotes the expectation with respect to the available data. The expected value of the discrepancy is asymptotically equal to the expected AIC of the considered statistical model which is given by:

[TABLE]

AIC allows to assess the truthfulness of any considered statistical model, and in our case, allows to estimate the number of content based clusters together with the characterizing parameters of each one. Each cluster is characterized by the average file popularity distribution and its variance within the cluster. In order to approximate the process generating the users probability vectors, we consider a set of statistical models $\varXi=\left\{\xi_{{N_{c}}_{min}}...\xi_{{N_{c}}_{max}}\right\}$ where, $\left\{{N_{c}}_{min}...{N_{c}}_{max}\right\}$ represents the range over which the search for the true number of clusters will be carried out. Each of the considered models will be typified by a number of defining parameters. In our case, each considered model $\xi_{N_{i}}$ is characterized by $N_{i}\times(F+1)$ parameters, $N_{i}\times F$ representing the average file popularity in each cluster and $N_{i}$ variance estimates. In this paper, the likelihood $L_{\xi_{N_{i}}}(N_{i}|P_{u},u=1...U)$ is computed based on a Gaussian Mixture model. This is a common assumptions for data generating models [19]. The log likelihood function $\text{log}\;(L_{\xi_{N_{i}}}(N_{i}|P_{u},u=1...U))$ is computed after clustering user with the assumption that they can be grouped into $N_{i}$ clusters. $\text{log}\;(L_{\xi_{N_{i}}}(N_{i}|P_{u},u=1...U))$ can be written as follows:

[TABLE]

where $\psi(u)$ represents the index of the cluster to which user $u$ is assigned. $\hat{P}_{\psi(u)}$ denotes the average popularity vector in cluster $\psi(u)$. $U_{\psi(u)}$ refers to the number of users in cluster $\psi(u)$. $\hat{\sigma}_{\psi(u)}^{2}$ denotes the variance of content popularity vectors in cluster $\psi(u)$ and is given by:

[TABLE]

Then the log-likelihood function can be written as:

[TABLE]

The resulting AIC for model $\xi_{N_{i}}$ is given by:

[TABLE]

The model that best describe the user popularity vectors is the one that minimizes the AIC and, consequently, the discrepancy. In order to find the best model, the user are clustered according to their content popularity vectors using the $K$-mean algorithm [17], while assuming different numbers of clusters from a search range $\left\{{N_{c}}_{min}...{N_{c}}_{max}\right\}$. The selected model $\xi_{\text{AIC}}$ verifies:

[TABLE]

The selected model which minimizes the AIC, strikes the best trade-off between fitness and complexity. This results in a truthful modeling of content popularity based clusters. The resulting model guarantees minimum discrepancy among the request patterns of the users within each cluster. We now provide the detailed description of the content based user clustering algorithm.

III-B User clustering algorithm

The proposed content popularity based clustering algorithm starts by defining a search interval $\left[{N_{c}}_{min}...{N_{c}}_{max}\right]$. The algorithm begins by assuming the existence of ${N_{c}}_{min}$ clusters. It clusters the users accordingly using the $K$-mean algorithm [17]. $K$-mean allows to assign each user to the cluster with the nearest centroid which results in minimizing the disparity between users behaviors in the same cluster. The popularity profile of the cluster is then defined as the average of the popularity vectors of all users in the cluster as:

[TABLE]

Each cluster $k$ is then associated with a vector $\hat{P}_{k}=\left[\hat{p}_{1k}...\hat{p}_{Fk}\right]$, where $\hat{p}_{fk}$ denotes the average popularity of file $f$ in cluster $k$. Once users are assigned to their respective clusters, AIC is computed. The number of clusters is incremented by adding a new centroid. The AIC is then recomputed until reaching a minimum. AIC is decreasing as a function of the number of clusters until reaching a minimum in the most accurate estimate. The AIC will then start increasing because of model complexity. Since the goal of the clustering is to reduce the divergence among users from the same cluster, a new centroids is added, at each step, in the cluster with the greatest popularity variance. The new center is selected as the user having the largest distance from the mean popularity vector of its cluster. This allows to reduce the discrepancy in user traffic pattern. The detailed clustering algorithm can be written as the following:

[TABLE]

Once content popularity based clustering is performed, the cached files of each cluster are selected based on its average popularity vector, which is given in $(11)$. For each $\{k=1...N_{c}\}$, the files in the catalog are ordered in a decreasing order of popularity according to $\hat{P}_{k}$. The set of cached files, in each cluster, $\{\Delta_{k},k=1...N_{c}\}$ is then selected as the most popular files, according to $\{\hat{P}_{k},k=1...N_{c}\}$, whose aggregate size is at maximum $M$. Apart from the maximum size constraint, we impose no restrictions on the set of cached files. Consequently, there may be some overlapping between the cached files of different clusters. Meaning that the same file can be selected in the cached sets of different clusters ($\Delta_{k}\cap\Delta_{j}\neq\emptyset,\text{for some}\;k\neq j$). Files that are selected by different clusters are very popular across user. Consequently, it makes sense to increase the number of cached copies in the network. Given the considered model in section $II$, allowing overlapping between the cached files of different clusters provides better performance. Since user preference can change over time, the algorithm in Table $I$ can be executed periodically or whenever substantial popularity profile modification is recorded. This allows the caching system to adapt the selected files accordingly. In order to investigate the performance of the proposed scheme, we assess its impact on the achievable EE of the system.

IV EE with content popularity clustering

Predicting which content is most likely to be requested and caching it in the edge can reduce the latency and backhaul load as well as increasing the overall throughput. Proactive caching also proved to be an effective technology that can improve another very important metric in future generation networks, namely, EE [13]. In what follows, we investigate the EE of cache enabled small cell networks with content popularity based user clustering. We consider the downlink of the cache enabled network. Without loss of generality, we concentrate on a reference user located at the origin of the plane. The EE of the network is given by the ratio between the average achievable spectral efficiency and the average consumed power [13].

[TABLE]

where $\Sigma$ denotes the average energy efficiency, $\rho^{c}_{total}$ denotes the average consumed power in the cache enabled small cell network and $SE$ denotes its average achievable spectral efficiency. In order to derive the expressions of $SE$ and $\rho^{c}_{total}$ and, consequently, the achievable EE, we need to start by finding the expression of the cache hit probability.

IV-A *Cache hit Probability *

According to the considered system model, the cache hit probability refers to the probability of finding a requested file in the cache of a SBS within radius $R$ from a given user [18]. Our context is different from the one in [18], since the users are clustered and the SBSs cache different files depending on their associated cluster. Considering the proposed clustering model, the cache hit probability can be expressed as follows (the derivations are skipped for brevity):

[TABLE]

where $\Delta_{k}$ represents the set of the most popular files of cluster $k$ that fills the SBS caching capacity. This equation denotes the probability of finding of at least one SBS with the requested file stored in its cache within a radius $R$ from a given user. The density of SBS caching the most popular content of a cluster $\{k,k=1...N_{c}\}$ is given by $\lambda_{sk}$ and, their average number $N_{sk}$ is given by $N_{sk}=\lambda_{sk}\pi R^{2}_{n}$. The densities $\{\lambda_{sk},k=1...N_{c}\}$ are such that $\sum_{k=1}^{N_{c}}\lambda_{sk}=\lambda_{s}$. One major upside of content popularity based user clustering is content diversity. While each SBS caches the most popular files of only one cluster, users can request any of the cached files in SBSs within radius $R$, which can be fetched without additional load on the backhaul. In fact, a given user can communicate with the closest SBS caching the files of a cluster different from his own whenever the requested content is already cached. Consequently, compared with the classical approach of caching the same popular content everywhere, the users covered by several SBSs from different clusters will see an increase in their cache hit probability.

IV-B Average total consumed power

In order to gain a useful insight into the achievable EE and capture the fundamental tradeoffs, we extend the power model in [13]. The average consumed total power in the considered network with caching capabilities can be modeled as follow:

[TABLE]

where $\rho_{I}$, $\rho_{T}$ and $\rho_{f}$ denote, respectively, the power consumed by the infrastructure of active base stations, the total transmit power and the used power to fetch files from the hard disc or the core network. The expectation $\mathbb{E}\left\{.\right\}$ is taken over the users and SBSs PPPs.

The average power consumed by the infrastructure is given by:

[TABLE]

$\rho$ and $\lambda_{s}\pi R^{2}_{n}$ denote, respectively, the fix operational charge consumed by an active SBS and the average number of active SBSs. The average power used to retrieve a file either over the backhaul, when a cache miss event occurs, or from a SBS cache is given by:

[TABLE]

where $\rho_{hd}$ denotes the power needed to retrieve data from the local hard disk of a small base station when the requested content is already cached and a cache hit event occurs. $\rho_{bh}$ denotes the power needed to retrieve data from the core network through the backhaul when a cache miss event occurs. Owing to channel inversion power control, the power used for transmission depends on the distance between the communicating SBS and users. Here we consider $\Upsilon_{k}$ as the set of users associated with cluster $k,\forall k=1..N_{c}$. Each user looks for the requested file in the cache of the SBSs within a radius $R$, starting with the closest SBS from his own cluster. If the requested file is available in a cache within this distance, a cache hit event occurs and the user will associate with the closest SBS storing the requested file. In the event of a cache miss, the user associates with the nearest SBS from its corresponding cluster and the requested content is retrieved from the core network through the backhaul. If a user cannot find an SBS from its own cluster within a radius $R$, it only communicate with SBSs from other clusters within radius $R$, in the case of a cache hit event. The average total transmission power is given by:

[TABLE]

where $E\left\{\rho_{k}\right\}$ denotes the average transmit power that a typical user utilizes when communicating with SBSs associated with cluster $k,\forall k=1..N_{c}$. Finally, the expression of the average consumed total power is derived by including the expressions of $\mathbb{E}\left\{\rho_{k}\right\}$, $\forall k=1..N_{c}$.

Lemma 1**.**

The average consumed total power in the considered network with caching capabilities and content-popularity based user clustering can be modeled as follows:

[TABLE]

Proof.

See Appendix A. ∎

Following the same reasoning, the average consumed total power in the network with no proactive caching capabilities at the SBSs, is given by:

[TABLE]

$\rho^{nc}_{total}$ is taken into consideration in order to guarantee an improvement in the average EE of the network, when proactive caching is implemented.

IV-C Average Spectral Efficiency

In order to derive the achievable EE, the expression of the average spectral efficiency should be derived. The downlink SINR for a user $u$ taken at the origin is given by:

[TABLE]

where $I_{k},\;\forall k=1..N_{c}$ represents the interference coming from SBS from cluster $k$ given by $I_{k}=\sum_{i\in\phi_{sk}}\rho_{ik}\lVert h_{ui}\rVert^{2}r^{-\alpha}_{ui}$. Here $\phi_{sk}$ denotes the set of SBSs associated with cluster $k,\;k=1..N_{c}$. $\rho_{ik}$ refers to the power used in the downlink by SBS $i$ from cluster $k$. $\sigma^{2}$ represents the noise power. In order to compute the average spectral efficiency of the network, first we need to derive the achievable coverage probability which is given in the following Lemma:

Lemma 2**.**

The downlink coverage probability is given by:

[TABLE]

Proof.

See Appendix B. ∎

We can see from Lemma $2$, that increasing the SBS density enables to reduce the used transmit power. Nevertheless, we need to take into consideration the constant power consumed by the infrastructure of active SBSs which represents an important part of power consumption of the network. The average achievable spectral efficiency can be written as:

[TABLE]

Given the average achievable spectral efficiency and average consumed power, we can derive a closed form expression of the energy efficiency $\Sigma$:

[TABLE]

By substituting $(18)$ and $(21)$ into $(23)$, we obtain the average EE. We can notice from $(21)$ that the density of SBSs is a major defining parameter of $\Sigma$.

IV-D Analysis of Energy Efficiency

We can see, in $(22)$, that increasing the SBS density results in a reduction in the interference. This is mainly due to the resulting decrease in transmit power since users are closer to their serving SBSs. Nevertheless, increasing SBS density results in more power consumption due to the strain of active infrastructure. We aim at finding the optimal active SBS density vector that maximizes the achievable EE, even when user positions are not taken into consideration. We can imagine a setting in which SBS are activated and shutdown based on user density and popularity profiles. We consider a constraint in which we aim at maintaining a power budget that is lower than that used when no proactive caching is enabled. The problem can be formulated as follows:

[TABLE]

This optimization problem allows to derive the optimal density vector needed to maximize the average EE for a given user density, popularity profiles and cache size. Although the closed form expression of $\Sigma$ is complex to analyze, it can be proven that $\Sigma$ is quasi concave using an intelligent simplification by considering a composition of $\Sigma$ with an affine mapping [29].

Theorem 1**.**

The considered optimization problem is quasi concave and the optimal SBS density $\Lambda^{*}_{s}$ can be derived, with zero duality gap. If $\exists\Lambda^{*}_{s}$ such that $\nabla\Sigma(\Lambda^{*}_{s})=0$ then this vector is unique and it is the optimal solution. If this condition is not satisfied for any $\Lambda_{s}$ such that $\Lambda_{s}^{\dagger}\mathbf{1}\leq\lambda_{s_{max}}$, then the optimal solution can be found using the Karush-Kuhn-Tucker (KKT) conditions:

[TABLE]

where $L$ refers to the Lagrangian associated with problem $(24)$, $C(\Lambda_{s})=\rho^{c}_{total}-\rho^{nc}_{total}$, $H(\Lambda_{s})=\Lambda_{s}^{\dagger}\mathbf{1}$. $\varsigma\;\text{and}\;\kappa$ refers to the Lagrangian multipliers associated, respectively, with $(24a)$ and $(24b)$.

Proof.

See Appendix C. ∎

Finding the optimal density vector $\Lambda^{*}_{s}$ based on the KKT conditions in $(25)$, can be done using, for example, the sub-gradient descent method [29].

V Exploiting spatial correlation in users demand

While, in the previous sections, EE was optimized with respect to the density vector of active SBSs, further EE gain can be achieved by including spatial information whenever it is available. In this paper, the choice to decouple the two problems of cached file selection and content placement can be justified by the fact that popularity distributions change slower than user locations. Consequently, the selected cached content which depends on the average popularity distribution per cluster, is kept constant for long periods and the network can adapt its location based on user movement. Real life examples can also support this approach. While correlation in content popularity between users from the same social group stays for long periods of time, their location can change due to mobility. This motivates the need to adapt the cached content placement more often than the selected files in order to simplify the management of the network. Acquiring information on user location requires a non negligible processing and signaling overhead. Consequently, this information should be leveraged whenever it is available.

In the case of low mobility, where users do not change positions too often, it makes sense to adapt the files placement based on location information. Adapting the cached files placement can be done periodically or whenever the backhaul load allows it.

Practically, we may observe a spatial correlation in user file demand. This can be explained by the fact that people from the same social group (living or working in the same place) are most likely to have similar preferences. We aim at finding an effective allocation of the SBSs to the different clusters in order to minimize transmit power and, consequently, to improve the achievable EE. In fact, decreasing the distance between a given user and the SBS storing its requested file results in lower transmit power. Consequently, by effectively allocating the SBSs to the different clusters, we are able to lower the level of interference in the network, which results in increasing the EE [13]. The problem of cache placement can be tackled by adopting a hybrid approach where, a clustering based on both the location and content popularity is performed. This approach is more complex and do not necessarily produce better results since content popularity stays constant for long periods of time. In addition, the clustering that is done on the users enables also to group the files accordingly. Consequently, the complexity of the resulting optimal file placement problem is lower since the search space is reduced from the whole file catalog to groups of file of approximately equal total size. This simplifies the management of the caching system compared to existing work on location based optimization where, the complexity of the formulated problems is proportional to the number of files.

We consider a setting in which the location of all users and SBSs are known. This is implemented by considering a snapshot of the users and SBSs PPPs. We develop an integer optimization problem where we aim at minimizing the used power over the possible SBS-cluster affectation. Thanks to channel inversion power control, minimizing the used power is equivalent to reducing the distance between the users and the SBS caching the files they are most likely to request. We define $\omega_{u,s}$, the weight of the link between user $u$ and the SBS $s$ as follows:

[TABLE]

where $r_{us}$ representing the distance between user $u$ and the SBS $s$. $\omega_{\infty}$ is an arbitrarily large value. $\omega_{\infty}$ assures that no user can communicate with a SBS at a distance larger than $R$. We sort the links pathloss coefficients in decreasing order and denote by $(s)_{u}$ the SBS with the $s$-th greatest pathloss coefficient to user $u$. Based on the considered system model, less power is used when a user is served from a SBS within its neighborhood. Consequently, maximizing $\omega_{u,s}$ is equivalent to minimizing the transmit power and the distance between the user and its serving SBS. The average number of SBS associated with each cluster $k$ is given by $N_{sk}=\lambda_{sk}\pi R^{2}_{n}$. Since $\lambda_{sk},\;k=1..N_{c}$ are computed in $(24)$ so that EE is maximized, it does not take into consideration their spatial repartition in the network. The transmit power increases when users from the same cluster are not located within a reduced area. In order to deal with this problem, we relax the constraint on the number of SBS per cluster and we replace $N_{sk}$ by $N^{\prime}_{sk}$ where $N^{\prime}_{sk}>N_{sk}$. We consider the adjacency matrix $Y$, where $y_{s,k},\forall s=1..N_{s},k=1..N_{c}$ is given by:

[TABLE]

The problem of optimal SBS allocation to their respective clusters can be formulated as:

[TABLE]

Here, $(28a)$ captures the fact that each SBS stores the most popular files of one unique cluster. $(28b)$ indicates that the number of SBSs allocated to each cluster should respect the density vector $\Lambda^{*}_{s}$ which maximizes EE. The objective function in $(28)$ guarantees that each SBS caches the files that are most likely to be requested by nearby users. In fact, $(y_{(s)_{u},k}\prod\limits_{i=1}^{s-1}(1-y_{(i)_{u},k}))$ is an indicator function that refers to the case where the most popular files of cluster $k$ are cached in SBS $(s)_{u}$ and not in SBSs $(i)_{u},i=1,...,s-1.$ Consequently, the objective function value is equal to the expected pathloss between users and their serving SBSs. We show that the considered optimization problem is NP-hard. We then prove that it can be formulated as the maximization of a submodular function over matroid constraints and we provide an algorithm that enables to derive a $(1-\frac{1}{e})$ approximation of the optimal solution of problem $(28)$. This formulation looks somehow similar to the considered problem in [20] where, the authors aim at optimizing the allocation of each individual file to a set of femto access points in order to minimize the expected downloading time. Nevertheless, problem $(28)$ consider a different objective function where, the aim is to minimize the transmit power. While [20] aims at optimizing the assignment of each individual file to the different femto access points, the objective in $(28)$ is, actually, to assign predefined batches of files from each cluster to the SBSs. Consequently, the problem formulation in the present paper enables to considerably reduce the complexity of deriving a solution. In fact, the running time depends on the number of popularity based clusters rather than the number of files. This is an important impact of the present formulation in $(28)$ since the number of files is typically very large. The considered setting enables to solve problem $(28)$ using sophisticated algorithms that can be computationally prohibitive otherwise.

V-A Computational Intractability

We start by showing the computational intractability of problem $(28)$.

Theorem 2**.**

The considered optimization problem in $(28)$ is NP-hard.

Proof.

In order to show that $(28)$ is NP-hard, we consider a special case of our setting where $N^{\prime}_{sk}=N\;\forall k=1..N_{c}$ and $N_{s}=N_{c}$. This special case means that the number of SBS associated with each cluster is the same, which is the case when $\sum_{u=1}^{U}\sum_{f\in\Delta_{k}}p_{fu}=C,\;\forall k=1..N_{c}$. In this case, the resulting optimization problem can be written as follow:

[TABLE]

In order to show NP-hardness, we use a reduction from the following NP-hard problem:

Weighted $K$-Set Packing Problem: $K$-Set packing is an NP-hard combinatorial problem. It is one of the 21 problems of Karp [26]. The $K$-Set packing problem aims to find a maximum number of pairwise disjoint sets, with at most $K$ elements, in a family $S$ of subsets of a universal set $V$. The weighted version of the $K$-Set packing problem is obtained by assigning a real weight to each subset and maximizing the total weight.

We consider a collection of SBS sets $\{v_{i},i=1,...,n\}$, associated each with a weight $\omega_{v_{i}}=\sum_{u=1}^{U}\underset{s\in v_{i}}{\text{max}}\;\omega_{u,s}$. Problem $(28)$ can then be formulated as a Weighted K-Set Packing Problem:

[TABLE]

Solving $(30)$ results in at most $N_{c}$ sets of SBSs. Since the resulting sets are disjoint, each of them will be associated with a given cluster. The number of resulting sets could not exceed $N_{c}$ since $N_{s}=N_{c}$. We can see that solving the weighted $K$-Set Packing Problem, for $K=N$ and where the weight of each subset is given by $C\omega_{v_{i}}$, is equivalent to solving the special case of the SBS allocation problem in $(29)$. Knowing that the Weighted $K$-Set Packing Problem is NP-hard, we can then conclude that $(28)$ is also NP-hard. ∎

V-B Optimizing small base station allocation

In order to solve the considered optimization problem in $(28)$, we start by showing that it is equivalent to the maximization of a sub-modular set function over matroid constraints. The definitions of matroids and sub-modular set functions can be found in [24]. This structure allows the use the randomized algorithm proposed in [27] which achieves, at least, $(1-\frac{1}{e})$ of the optimal value. Taking into consideration the problem constraints we have the following:

Lemma 3**.**

The Considered Optimization problem in $(28)$ is equivalent to a maximization of a sub-modular set function over matroid constraints.

Proof.

See Appendix D. ∎

In order to solve the considered problem, we use the randomized algorithm proposed in [27]. This algorithm provides a $(1-\frac{1}{e})$-approximation of the optimal solution for sub-modular set function maximization with matroid constraints. This algorithm consists of two steps. In the first one, a fractional solution of the relaxed problem is obtained using a continuous greedy process. In the second part of the algorithm, the derived fractional solution is rounded using a variant of the pipage rounding technique [28]. In a typical setting, randomly rounding a fractional solution of an optimization problem does not preserve the feasibility of the solution, in particular when equality constraints are considered. Nevertheless, the pipage rounding technique in [28] enables to round a fractional solution so that the problem constraints are not violated. In our case the running time of the algorithm is $O((N_{s}N_{c})^{8})$ [27], where $N_{s}=\lambda_{s}\pi R^{2}_{n}$. This is quite convenient since the running time of the algorithm does not depend on the number of files in the catalog which can be very large. This an interesting result of content based clustering since it reduces the search space from the whole catalog to bins of files of approximately equal size.

VI Numerical Results AND Discussion

In this section, we investigate the impact of the different system parameters on the Cache hit probability and EE. We then investigate the impact of the SBS allocation algorithm on the performances of the network. We consider a circular region with an area of $A=10Km^{2}$. We simulate two PPP processes, one for the users and another for the SBSs over this area. The respective densities of these process are $\lambda$ and $\lambda_{s_{max}}$ with $\lambda>>\lambda_{s_{max}}$ . The considered SBS density values are defined based on the typical communication range of a SBS. We consider a catalog constituted of $F=2000$ files with different randomly generated sizes $L_{i},\;i=1...F$ in the range $\left[10\;\text{MB}...100\;\text{MB}\right]$ [13]. We also consider the normalized cache size $\eta=\frac{M}{\sum_{i=1}^{F}L_{i}}$. We characterize each cluster by a given popularity based file ordering. For each user $u$, $P_{u}$ is generated according to a Zipf distribution with parameter $1$ [3], after randomly selecting a cluster file ordering. This results in a random allocation of the users to the different clusters. In order to run the clustering algorithm, we only need an interval over which the search of the number of clusters is carried on. In our simulations we take $\left[{N_{c}}_{min}...{N_{c}}_{max}\right]=\left[5...30\right]$. We consider a pathloss exponent $\alpha=2.5$. We consider the following power values [13]:

[TABLE]

Figure 2 shows the evolution of the cache hit probability as a function of the communication radius $R$ for different SBS densities. Figure 2 also shows a comparison between the content popularity based clustering approach and the classical method of supposing the same content popularity among all users. As an example, for a communication radius of $0.9\;Km$ and a SBS density of $\lambda_{s}=1$, we notice a substantial gain with a cache hit probability of $0.736$ for the clustering scheme compared with a probability equal to $0.41$ when caching the most popular files in all SBSs. We notice that the hit probability for the scheme without clustering saturates at a low value. This is due to the fact that the same set of files is cached in all the SBSs, which is clearly a suboptimal approach, especially when users are covered by multiple SBSs. The increase in hit probability for the clustering method is mainly due to the diversity of files cached in the SBS. Increasing the SBS density results in reducing the average distance from mobile users, which, consequently, results in improving the cache hit probability.

Figure 3 shows the evolution of the achievable EE as a function of the normalized cache size for different SBS densities. Figure 3 also shows a comparison between the content popularity based clustering approach and the classical method of supposing the same content popularity among all users. We can see that the proposed clustering method outperforms the classical approach of caching the same most popular files in all SBSs. For a normalized cache size of $0.4$ and an SBS density of $1.6\;\text{SBS}/Km^{2}$, we notice an increase of $12.5\%$ in the achievable EE. This gain is mainly due to the fact that the proposed method scores a higher hit probability than the unclustered approach. Consequently, the average energy needed to fetch the requested content is lower when user clustering is used. Even though restricting users to communicate with the closest SBS from their cluster, in the case of a cache miss event, can lead to an increase in the average transmit power, the observed gain in the energy used to fetch the desired content compensates for that. The gain in EE increases as a function of the SBS density. For a normalized cache size of $0.4$ and an SBS density of $1.9\;\text{SBS}/Km^{2}$, we notice an increase of $14.2\%$ in the achievable EE. This increase in EE gain can be explained by the fact that the average transmit power is a decreasing function of SBS density.

Figure 4 shows the performance of AIC model selection. We consider three settings in which, the true numbers of clusters are $10$, $15$ and $20$, respectively. The figure represents the computed AIC per point over the estimation range. The lowest AIC value represents the model that strikes the best trade-off between fitness and complexity. Note that the negative values of the AIC are due to a negative bias which characterize the AIC with a small sample number.

Figure 5 shows the impact of the SBS allocation algorithm on EE. We can see that, for different values of the SBS density, optimizing the SBS allocation results in a considerable gain in the EE. For a SBS density of $\lambda_{s}=1.9$ and a normalized cache size of $0.4$, optimizing the allocation of the SBSs results in an EE gain of $42.2\%$. As the SBSs density increases, the allocation algorithm results in greater improvement in EE. Optimizing the cluster-SBS association results in less average transmit power which reduces the interference and improves the achievable EE.

VII Conclusion

In this paper, we studied a cache enabled small cell network. A content popularity based clustering approach was considered in order to exploit the correlation between the request patterns of users. We develop an algorithm that enables to estimate the number of content popularity based clusters and to efficiently assign users to their respective groups. By considering the distribution of SBSs to be a Poisson point process, we investigated the impact of proactive caching with user clustering on the achievable EE . We proposed an optimization framework that enables to derive the optimal active SBS density vector in order to maximize the EE. In order to exploit any spatial correlation in user request patterns, we proposed a SBS allocation algorithm that aims at minimizing the transmit power by bringing the cached files closer to the users that are most likely to request them. Finally, we performed numerical analysis which show that the proposed clustering framework considerably outperforms the scheme in which the most popular files are cached in all SBSs. It also shows that optimized SBS allocation results in an improvement in the achievable hit probability and EE. For future work, the optimal cache placement strategy with heterogeneous popularity profiles and mobility patterns will be investigated.

VIII Appendix

**Appendix A proof of Lemma 1:

** We derive the expression of the average consumed power in the network with cache enabled SBSs. The average total power $\rho^{c}_{total}$ is given by:

[TABLE]

where $\mathbb{E}\left\{\rho_{I}\right\}=\rho\lambda_{s}\pi R^{2}_{n}$ and \mathbb{E}\left\{\rho_{f}\right\}=\lambda_{s}\pi R^{2}_{n}\big{(}\rho_{hd}\mathbb{P}\left\{hit\right\}+\rho_{bh}\big{(}1-\mathbb{P}\left\{hit\right\}\big{)}\big{)}. Taking into account the considered system model, the average transmit power used by a given user from cluster $k$, $\mathbb{E}\left\{\rho^{[k]}_{T}\right\}$ can be written as follows:

[TABLE]

After averaging over all users in the network, the average consumed transmit power is given by:

[TABLE]

We need then to compute the average power used by the users to communicate with the nearest SBS from any given cluster $k\;,k=1..N_{c}$.

According to the PPP assumption for the location of the SBSs, the distance from a user to its nearest SBS from cluster $k$, denoted by $r_{k}$, has the following pdf [21]:

[TABLE]

The transmit power used by the user in this case is given by $\rho_{k}=\rho_{0}r^{\alpha}_{k}$. Then:

[TABLE]

Following the same calculus for ${\mathbb{E}}\left[\rho_{k}\right],\;k=1..N_{c}$, we obtain the final expression of the average consumed power in the network:

[TABLE]

**Appendix B proof of Lemma 2:

**We derive the achievable coverage probability when using channel inversion power control:

[TABLE]

We use the fact that $\left\|h_{u}\right\|^{2}$ is exponentially distributed and $\mathcal{L}_{I_{k}}(s)$ is the Laplace transform of $I_{k}$ at $s$. To prove Lemma 3, we need to compute the Laplace transform of $I_{k},\forall k$. The interfering base stations constitute multiple PPP processes $\phi_{sk},k=1...N_{c}$, each associated with a given cluster. The Laplace transform of $I_{k}$ for a given $k$ is obtained as:

[TABLE]

${\mathbb{E}}\left[\rho^{\frac{2}{\alpha}}_{k}\right]$ depends on the density of the small cells caching files from cluster $k$. ${\mathbb{E}}\left[\rho^{\frac{2}{\alpha}}_{k}\right]$ can be deduced from Appendix A as; ${\mathbb{E}}\left[\rho^{\frac{2}{\alpha}}_{k}\right]=\frac{\rho_{0}^{\frac{2}{\alpha}}\gamma(2,\pi\lambda_{sk}R^{2})}{\lambda_{sk}\pi}$. Based on Slivnyak’s Theorem for Poisson Point Processes [21], the obtained Expression is valid for any user within the network.

**Appendix C proof of Theorem 1:

**We start by showing that the objective function is quasi-concave. Given the expression of $\Sigma$ as a function of the density vector $\Lambda_{s}$, it is difficult to prove its quasi-concavity by using its gradient or Hessian matrix. However, using the fact that a composition with an affine function preserves quasi-concavity [29], this proof can be considerably simplified. To prove the quasi-concavity of $\Sigma$ as a function of the density vector $\Lambda_{s}$, we consider an affine function $f(t)$ given by:

[TABLE]

where $\Lambda^{0}_{s}\in\mathbb{R}^{N_{c}\times 1}$ such that $\sum_{k=1}^{N_{c}}\lambda^{0}_{sk}\leq\lambda_{s_{max}}$, $Z\in\mathbb{R}^{N_{c}\times 1}$ and $t\in\mathbb{R}$. Since a composition with an affine function preserves quasi-concavity, it is sufficient to prove the quasi-concavity of $\Sigma(tZ+\Lambda^{0}_{s})$ with respect to $t$ in order to show the quasi-concavity of $\Sigma$ with respect to $\Lambda_{s}$.

The objective $\Sigma(tZ+\Lambda^{0}_{s})$ can be written as a product of two nonnegative functions:

$U(tZ+\Lambda^{0}_{s})=\sum_{k=1}^{N_{c}}(tz_{k}+\lambda^{0}_{sk})\pi R^{2}_{n}\;\text{log}(1+\theta)\;\mathbb{P}\left\{\text{SINR}\geq\theta\right\}$ and $V(tZ+\Lambda^{0}_{s})=\frac{1}{\rho^{c}_{total}}$.

We start by computing the derivatives of $U(tZ+\Lambda^{0}_{s})$ and $V(tZ+\Lambda^{0}_{s})$ with respect to $t$:

[TABLE]

Then $\frac{\rm dU(tZ+\Lambda^{0}_{s})}{\rm dt}>0$. We do the same to $V(tZ+\Lambda^{0}_{s})=\frac{1}{\rho^{c}_{total}}$. We have:

[TABLE]

where $\chi$ is given by:

[TABLE]

Here $\chi_{1}$ and $\chi_{2}$ are respectively given by:

[TABLE]

[TABLE]

Consequently $\frac{\rm dV(tZ+\Lambda^{0}_{s})}{\rm dt}<0$. In what follows we distinguish two cases depending on the existence of a point $t^{*}$ such that $\frac{\rm d\Sigma(tZ+\Lambda^{0}_{s})}{\rm dt}_{t=t^{*}}=0$. If $\exists t^{*}$ such that $\frac{\rm d\Sigma(tZ+\Lambda^{0}_{s})}{\rm dt}_{t=t^{*}}=0$ then:

[TABLE]

We compute the derivative of $L(t)=\frac{-U^{\prime}V}{V^{\prime}U}$ with respect to $t$. The expression of the derivative is omitted here for brevity. We find that $L^{\prime}(t)>0$. Since $L(t)$ is a strictly increasing function then, according the Theorem of intermediate value, if $\exists t^{*}$ such that $L(t^{*})=1$ then this point is unique. Finally, depending on the existence of $t^{*}$, we have two cases:

• •

If $\exists\;t^{*}$ such that $L(t^{*})=1$ then this point is unique and $\Sigma(f(t))$ is increasing for $t<t^{*}$ and decreasing for $t>t^{*}$.

• •

If, on the other hand, $t^{*}$ does not exists, then $\Sigma(f(t))$ is a strictly monotone function.

This proves that $\Sigma(f(t))$ is a quasi-concave function of $t$. Since composition with an affine function preserves quasi-concavity, we can deduce that $\Sigma(\Lambda_{s})$ is a quasi-concave function of $\Lambda_{s}$ and that, if $\exists\Lambda^{*}_{s}$ such that $\nabla\Sigma(\Lambda^{*}_{s})=0$ then this vector is unique .

In the second step of the proof, we need to show that the constraint $C(\Lambda_{s})=\rho^{c}_{total}-\rho^{nc}_{total}$ is also quasi concave. This is done in a similar way as in the first step by considering $C(f(t))$. After computing the derivative of $C(f(t))$ with respect to $t$, we find that:$\frac{\rm dC(f(t))}{\rm dt}<0$. Then the first constraint is quasi concave. Using the same method, it is trivial to show that the second constraint is also quasi-concave. In order to finish the proof, we need to show that the optimal solution can be found with zero duality gap. This will be done using results on quasi-concave programming from [23]. Since $\frac{\rm d(\rho^{c}_{total}-\rho^{nc}_{total})}{\rm d\lambda_{sk}}\neq 0,\forall k=1...N_{c}$ then, according to the Necessity Theorem in [23], any solution of the optimization problem $(23)$ satisfies the KKT conditions. We, now, distinguish between two case:

• •

If $\exists\;\Lambda^{*}_{s}$ such that $\nabla\Sigma(\Lambda_{s})_{\Lambda_{s}=\Lambda^{*}_{s}}=0$ and $\Lambda^{*}_{s}$ satisfies the constraints then, $\Lambda^{*}_{s}$ is unique and it is a global optimum of $(23)$. The uniqueness of $\Lambda^{*}_{s}$, if it exists, was shown in the first step of the proof.

• •

If $\nabla\Sigma(\Lambda_{s})\geq 0$, $\forall\Lambda_{s}\;\text{such that}\;\Lambda_{s}^{\dagger}\mathbf{1}\leq\lambda_{s_{max}}$, the sufficiency Theorem in [23] is verified. Consequently, by combining the necessity and sufficiency results, the optimal SBS density vector can be derived using KKT.

**Appendix D proof of Lemma 3:

**First we need to prove that the objective function $\Omega$ is sub-modular. We consider two SBSs allocations $X\;\text{and}\;Y$ such that $X\subseteq Y$ and we need to prove that the marginal value of adding a new allocated SBS $l$ to cluster $i$ in $X$ and $Y$ verifies:

[TABLE]

Monotonicity is trivial since any new SBS allocation cannot decrease the value of the objective function. In order to show submodularity of the function, we compare the marginal values of adding $y_{li}$ to $X$ and $Y$. Here we consider $\Pi_{i}(X\cup\left\{y_{li}\right\})$ referring to the users that change their serving SBS from cluster $i$. $\mu(u,i)$ refers to the index of the SBS from cluster $i$ serving user $u$. A user changes its serving SBS when the new allocated one is closer which induces less transmit power. Consequently, the marginal values of adding $y_{li}$ to $X$ and $Y$ are as follows:

[TABLE]

Since $X\subseteq Y$ we can deduce that $\Pi_{i}(Y\cup\left\{y_{li}\right\})\subseteq\Pi_{i}(X\cup\left\{y_{li}\right\})$. Since a user changes its serving SBS only when a closer allocated one is available then $\omega^{(Y\cup\left\{y_{li}\right\})}_{u\mu(u,i)}-\omega^{(Y)}_{u\mu(u,i)}>0$ which proves that \Omega\big{(}X\cup\left\{y_{li}\right\}\big{)}-\Omega\big{(}X\big{)}\geq\Omega\big{(}Y\cup\left\{y_{li}\right\}\big{)}-\Omega\big{(}Y\big{)}. Consequently, $\Omega$ is a sub-modular set function. It is simple to verify that the constraints $\sum_{s=1}^{N_{s}}y_{sk}\leq N_{sk},\forall k=1..N_{c}$ are equivalent to a matroid constraints [20]. Then the considered optimization problem is equivalent to maximizing a sub-modular function subject to matroid constraints.