You are on page 1of 14

4016 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO.

9, SEPTEMBER 2013

Cognitive Small Cell Networks:


Energy Efficiency and Trade-Offs
Matthias Wildemeersch, Student Member, IEEE, Tony Q. S. Quek, Senior Member, IEEE,
Cornelis H. Slump, Member, IEEE, and Alberto Rabbachin, Member, IEEE

Abstract—Heterogeneous networks using a mix of macrocells Conventional cellular networks based on the careful deploy-
and small cells are foreseen as one of the solutions to meet the ment of MBSs suffer from poor signal quality for indoor and
ever increasing mobile traffic demand. Nevertheless, a massive cell edge users. Furthermore, the explosive surge in mobile
deployment of small cell access points (SAPs) leads also to
a considerable increase in energy consumption. Spurred by data traffic accelerates the need for novel cellular architectures
growing environmental awareness and the high price of energy, to meet such demands [3]. The LTE-Advanced or beyond
it is crucial to design energy efficient wireless systems for both standards propose heterogeneous networks (HetNet’s) that
macrocells and small cells. In this work, we evaluate a distributed consist of a macrocell network overlayed by small cells. The
sleep-mode strategy for cognitive SAPs and we analyze the macro-tier guarantees the coverage, while the overlay network
trade-off between traffic offloading from the macrocell and the
energy consumption of the small cells. Using tools from stochastic is a means to offload the data traffic from the macrocell
geometry, we define the user discovery performance of the SAP network and to satisfy the local capacity demand. The small
and derive the uplink capacity of the small cells located in cells in this two-tier architecture can be micro-cells, pico-
the Voronoi cell of a macrocell base station, accounting for cells or femto-cells, where the distinction between the different
the uncertainties associated with random position, density, user types of small cells can be found in the size of the cells and the
activity, propagation channel, network interference generated by
uncoordinated activity, and the sensing scheme. In addition, we capability of auto-configuration and auto-optimization. Small
define a fundamental limit on the interference density that allows cells can extend the network coverage and the reduced cell size
robust detection and we elucidate the relation between energy leads to higher spatial frequency reuse and increased network
efficiency and sensing time using large deviations theory. Through capacity.
the formulation of several optimization problems, we propose a
framework that yields design guidelines for energy efficient small Although the introduction of heterogeneous networks can
cell networks. resolve the capacity demand issue [7], [8], the overall energy
Index Terms—Small cell, cognitive radio, green communica- consumption is significantly increased by the installation of
tions, stochastic geometry, energy efficiency. these additional base stations. Motivated by the high traffic
demand fluctuations over space, time, and frequency, sleep
I. I NTRODUCTION mode techniques are a promising strategy to overcome this
VER the last years mobile data traffic has risen expo- problem. For wireless sensor networks, the energy conserva-
O nentially and along with this accruing mobile data de-
mand, energy consumption has increased considerably [1]–[3].
tion by sleeping techniques for wireless devices running on
battery power has been studied in [9]. The IEEE 802.16e
Driven by growing environmental awareness and increasing and LTE standards support sleep mode strategies for mobile
electrical cost relative to the operation of mobile base stations terminals and the trade-off between energy efficiency and
(MBSs), green wireless communications has become an active response delay has been analyzed in [10]. Only some recent
field of research that tries to unite the opposing needs of work is dedicated to sleep mode techniques for small cell
growing mobile data activity and energy efficiency [4]–[6]. access points (SAPs). In [11], the authors conclude that sleep
mode operation is effective when the cell size is small and
Manuscript received August 10, 2012; revised February 13 and May 13,
2013. The editor coordinating the review of this paper and approving it for for light traffic conditions. Different sleep mode strategies
publication was H. Yanikomeroglu. for SAPs are introduced in [12], such that the wake-up
The material in this paper has been presented in part at the IEEE Wireless mechanism can be driven by the SAP, the core network or the
Communications and Networking Conference (WCNC), Shanghai, China,
April 2013, and at the IEEE International Conference on Communications user equipment (UE). For WiFi access points, the UE driven
(ICC), Budapest, Hungary, June 2013. approach has been studied in [13], but the reverse beaconing
M. Wildemeersch is with the Signals and Systems Group, University of adds complexity to the hardware and assumes knowledge
Twente, the Netherlands, and the Institute for Infocomm Research, A∗ STAR,
Singapore (e-mail: stuwma@i2r.a-star.edu.sg). about the signal structure. In [14], the overall HetNet energy
T. Q. S. Quek is with the Singapore University of Technology and Design, consumption is minimized based on sleep mode techniques
and the Institute for Infocomm Research (e-mail: qsquek@ieee.org). for network-operated femtocells. This centralized sleep and
C. H. Slump is with the Signals and Systems Group, University of
Twente, Drienerlolaan 5, 7500 AE Enschede, the Netherlands (e-mail: wake-up mechanism assumes traffic load and user localization
C.H.Slump@utwente.nl). awareness, and increases the signaling overhead. Therefore,
A. Rabbachin is with the Laboratory for Information and Decision Systems it is attractive to investigate distributed sleep mode strategies,
(LIDS), Massachusetts Institute of Technology, 77 Massachusetts Avenue,
Cambridge, MA 02139 USA (e-mail: rabalb@mit.edu). which do not involve an augmented UE complexity and require
Digital Object Identifier 10.1109/TCOMM.2013.072213.120588 neither signaling nor user localization awareness.
0090-6778/13$31.00 
c 2013 IEEE
WILDEMEERSCH et al.: COGNITIVE SMALL CELL NETWORKS: ENERGY EFFICIENCY AND TRADE-OFFS 4017

To allow a distributed decision approach, the SAP needs • We derive tractable expressions of the aggregate through-
cognitive capabilities to sense when a small cell user is put and capacity that can be offloaded by the small
active within the SAP coverage [12]. In future heterogeneous cells within an the MBS Voronoi cell. This allows us to
networks, cognitive capabilities will become essential not only formulate the trade-off between energy consumption and
for the energy efficient operation of the small cell tier, but throughput/capacity as a set of optimization problems.
also to overcome coexistence issues in multi-tier networks. • We define the interference wall, beyond which the target
Specifically, spectrum or carrier sensing by the cognitive SAPs probability of detection and false alarm can not be
enable interference avoidance and efficient spectrum allocation obtained no matter how long the sensing time. This is
[15]. Hence, cognitive small cells allow the deployment of a fundamental limit of the detection robustness, which
energy efficient and spectral efficient heterogeneous networks. confines the region of interferer densities enabling an
Many ideas can be borrowed from the field of cognitive radio energy efficient SAP design.
(CR). For cognitive radio networks (CRN), the trade-off be- • Using techniques from large deviations theory, we deter-
tween protection of the primary users (PU) and the throughput mine how fast the false alarm rate converges to a sta-
of the secondary users (SU) has been addressed in [16]– tionary value as a function of the interferer density. The
[18]. For SAP sleep mode techniques, this corresponds to the obtained rate function gives insight into how the detection
trade-off between traffic offload by means of a high detection performance affects the small cell energy consumption
probability and energy consumption, proportional to the false and traffic offload.
alarm rate. Although spectrum sensing and cognitive radio are The remainder of the paper is structured as follows. In
intensely studied, there are still many issues that need to be ad- Section II, the system model is introduced. In Section III, the
dressed [19]–[22]. For instance, the cognitive capabilities are energy consumption model and the sensing performance of
usually assessed for a deterministic (worst case) user location, the SAP are discussed. The traffic that can be offloaded from
while a random position of the user is of interest to determine the macrocell is characterized in Section IV. In Section V,
energy consumption. The capacity and energy efficiency of the trade-off between traffic offload and energy consumption
base stations in heterogeneous networks has been studied in is discussed by means of several optimization problems. The
[23], but the impact of aggregate interference in dense network limits of detection robustness and the implications on the
deployments has typically been neglected. Recent results show energy efficiency are discussed in Section VI. Numerical
that it is fundamental to consider aggregate interference to results are shown in Section VII and conclusions are drawn
accurately define the network performance [24]–[26]. in Section VIII.
In this work, we investigate how SAPs can be used to
offload the traffic from the macrocell network and how they II. S YSTEM M ODEL
can exploit their sensing capabilities to enhance the energy A. Network topology
efficiency. Specifically, these energy efficient SAPs can save We consider a cellular network model that consists of a
power by entering into sleep mode when they are not serving first tier of MBSs distributed according to a homogeneous
any active small cell users. Considering open access control, Poisson point process (PPP) Θ with density λm , overlayed
the SAPs need to sense the transmissions from a macrocell with a network of SAPs distributed according to a PPP Φ
user to an MBS, and switch on the pilot transmissions when with density λs , where usually λs > λm . Modeling the
user activity is detected within the SAP coverage. Due to the locations of SAPs and the more regular MBSs by means
simplicity of passive sensing, we assume that all SAPs perform of homogeneous PPPs is extensively discussed in literature
energy detection at the expense of being sensitive to noise and has been validated by numerical analysis and compared
and interference uncertainties. Furthermore, the transmissions with actual base station deployments [27]. Moreover, recently
from the macrocell user can be bursty, making the duty cycle theoretical evidence has been given for the accuracy of the PPP
of the sleeping mode at each SAP difficult to design to ensure model [28]. The set of macrocells is known as the Poisson-
a low probability of miss detection. Since the analysis of a Voronoi tessellation of R2 with respect to Θ. The Voronoi cell
deterministic (e.g. worst case) user location does not provide Cj corresponding to an MBS xj,m consists of those points in
guidance for the definition of the SAP energy consumption, R2 which are closer to xj,m than to any other MBS and is
it is crucial to evaluate the typical user case, that is, to defined as Cj = {y ∈ R2 | y − xj,m  ≤ y − xi,m , xi,m ∈
consider a user with random location within the SAP coverage. Θ\{xj,m }}. The SAPs that belong to Cj form the aggregate
Specifically, the main contributions of this work are listed as Aj = ∪{xi,s | xi,s ∈ Φ ∩ Cj }, and applying Campbell’s
follows: theorem the mean number of SAPs within an MBS Voronoi
• We formulate an average network energy consumption cell is given by λs /λm [29]. The mobile users are scattered
model for the cognitive SAPs that are located within the over R2 according to a PPP Ψ with density λu .1 We consider
Voronoi cell of an MBS accounting for the base station universal spectrum allocation where both tiers can access
and user densities, detection performance, the sensing the full spectrum and orthogonal frequency-division multiple
strategy, bursty macrocell user activity, and uncoordinated access (OFDMA) or time-division multiple access (TDMA) as
network interference uncertainties. multiple access technique.2 The SAPs operate in open access
• We present a unified analytical framework that models the
performance of passive sensing for a typical user within
1 Note that Θ, Φ, and Ψ are independent point processes.
2 The analysis can be extended to orthogonal spectrum allocation and affects
the SAP coverage, including the effects of propagation the interferer densities, which is important for the performance detection and
channel and aggregate network interference. the offload analysis discussed in Section III-B and IV, respectively.
4018 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 9, SEPTEMBER 2013

(OA) mode and are accessible for all users registered with
the operator of the SAP. In order to enable a distributed
sleep/wake-up scheme, the SAPs are foreseen of cognitive
capabilities. When an SAP does not serve an active user call,
it goes into sleep mode and senses periodically the macrocell
uplink channel to detect user activity. The SAP applies passive
sensing by means of an energy detector (ED) for reasons of
low complexity and low power consumption [30]. The ED
detection performance compared with other detection schemes
has been studied exhaustively in literature [31], [32]. We
limit our analysis to the ED since this scheme provides
a lower detection bound. Once the SAP detects an active
user in the macrocell uplink band within its coverage, the
SAP switches on and starts the transmission of pilot signals.
Subsequently, the UE reports the presence of the SAP to the Fig. 1. Spatial distribution of the MBSs, SAPs, and the UEs.
MBS and the UE is handed over to the SAP. We assume that
the uplink transmission power of the UE during the sensing
UE
period is constant and that after handover to the SAP, the
UE adopts a lower and constant transmission power. It is well SAP
known that the ED has no capabilities to differentiate between
the signal of interest (SoI) and interference or noise [30].
The measurements of the SAP under realistic conditions are sensing transmission/sleep
corrupted by network interference, which can originate from
macrocell users outside the SAP coverage, macrocell users Ws T W s
within the SAP coverage that are not registered with the SAP T
operator, small cell users, cognitive devices using the same
uplink band, and spurious emissions. As a result, we consider Fig. 2. Time slotted model, representing the activity of UE and the SAP
over time.
the case of the interfering nodes spatially distributed over the
entire plane according to a homogeneous PPP Ω with density
λ.3 For a homogeneous PPP, the probability that k nodes reside is assured to be independent across different slots. By the
within a region R depends on the density λ and the area AR colouring theorem of PPPs, the active nodes that contribute to
of the region R, and can be expressed as [33] the interference form a PPP with density pI λ [33]. We will
(λAR )k −λAR further refer to ps as the sensing probability.
P[k ∈ R] = e , k = 0, 1, 2, ... (1)
k!
The spatial model consisting of a macrocell network overlayed III. C OGNITIVE SAP
with multiple small cells is illustrated in Fig. 1. In the following, we provide the average network energy
consumption model for the SAPs and we characterize the
relationship between the energy consumption and the detection
B. Activity model
performance of the cognitive SAP. The presented analysis is
We define the activity of the UEs and SAPs using a time- generic and accommodates for random locations of both the
slotted model as depicted in Fig. 2. Assuming a fixed slot UE and the interfering nodes.
duration T , the SAP senses the channel over a sensing time τs
and is in active mode over T −τs when an active mobile user is
A. Energy consumption model
detected4 . Both the UE and the interfering nodes are assumed
to have a bursty traffic mainly due to the mobility of the nodes, Three main contributions to the power consumption of
the switching between on and off states, and the switching cognitive SAPs can be identified: the power related to the
between carriers in a multi-carrier system. Thus, the activity circuit synchronization Ξc , the sensing power Ξs , and the
of an SAP, a UE and the interfering nodes in a given slot can processing power Ξt during active mode [34]. We consider the
be modeled as mutually independent Bernouilli processes with circuit synchronization is active over the entire time slot. The
success probabilities ps , pu , and pI , respectively. Moreover, SAP senses the uplink channel according to a sensing scheme
the activity of the SAP, the UE, and the interfering nodes and the corresponding energy consumption is proportional
to the sensing time. The UE signal detection is a binary
3 Note that the model allows to include the presence of groups of interferers
hypothesis test problem. In the presence of a UE signal
with different transmission power. The superposition of PPP’s with densities (hypothesis H1 ), the SAP starts the pilot transmissions when
λ1 , λ2 , ... results in a PPP with density equal to a weighted sum of the
it senses the uplink channel and correctly detects the user
densities, where the weighting factor depends  on the transmission power Pi
of the different PPPs and is given by λ = i λi (Pi /Pu )2/ν , with Pu the activity. In the absence of the UE signal (hypothesis H0 ), the
transmit power of the UE, and ν the path loss exponent. SAP starts the pilot transmissions when it incorrectly detects
4 Active mode consists of transmitting pilot signals, receiving signals, and
signal processing. Note that we neglect the time related to the handover the presence of a user. The probability that there resides at
process for simplicity. least one active UE within the coverage of a typical SAP is
WILDEMEERSCH et al.: COGNITIVE SMALL CELL NETWORKS: ENERGY EFFICIENCY AND TRADE-OFFS 4019

  
1 2 2 Rν (1 − 2jωσW2
)
ψVED |H1 (jω) = 1 −2 F1 1, ; 1 + ;
2 )N/2
(1 − 2jωσW ν ν jωPu
  2/ν    
 jω  jω
× exp −2 γ cos(π/ν) 
2/ν
2 
 1 − sign tan(π/ν) (10)
1 − 2jωσW 1 − 2jωσW
2

given by pUE = 1−exp(−pu λu πR2 ). Therefore, the expected over R2 according to a PPP, the aggregate network interference
SAP energy consumption per cell can be modeled as follows a symmetric stable distribution [35], [36]
−1
λs i ∼ S(α = 4/ν, β = 0, γ = πλC4/ν E[|hl Xl,p | ])
i 4/ν
(5)
Etot = [Ξc T + pUE {ps [Ξs τs + Pd Ξt (T − τs )]}
λm with Cx defined as
+ (1 − pUE ) {ps [Ξs τs + Pfa Ξt (T − τs )]}] (2) 
(1−x)
Γ(2−x) cos(πx/2) , x = 1,
where Pd and Pfa are the probability of user activity detection Cx  (6)
2
π, x = 1.
and false alarm, respectively.
The decision variable V determines the presence or absence
B. Non-coherent detection performance of the SoI. For the ED, V is defined as

2
Since the presence of multiple simultaneous UE transmis- 1 τs h(t)
sions in the macrocell uplink band eases the detection process, V = ν/2
s(t) + n(t) + i(t) dt. (7)
τs 0 r
we consider the more challenging scenario with a single f

UE within the coverage of an SAP to define the detection After sampling and considering block fading, the decision
performance. At the cognitive SAP, the received signal can be variable can be expressed as
written as N N
2
1 2 1 h
H0 : r(t) = n(t) + i(t) VED = r [k] = ν/2
s[k] + i[k] + n[k]
N N rf
k=1 k=1
h(t) (8)
H1 : r(t) = ν/2 s(t) + n(t) + i(t) (3)
rf where N = τs fs
with fs the sampling frequency equal
to the Nyquist rate6 . The probability of detection is defined
where s(t), n(t), and i(t) represent the SoI, the additive
as the probability that VED surpasses the threshold ζ in the
white Gaussian noise and the aggregate network interference,
presence of the SoI and is given by Pd = P[VED > ζ|H1 ].
respectively. The impulse response of the flat fading channel
The probability of false alarm is defined as the probability
between the UE and SAP is represented by h(t), rf is the
that VED surpasses the threshold in absence of the SoI and
distance between the UE and the SAP, and ν is the power
is given by Pfa = P[VED > ζ|H0 ]. To calculate Pd and Pfa ,
path loss exponent. To facilitate the analysis, we consider that
we propose a generic approach based on the characteristic
the typical SAP is at the origin of the Euclidean plane and the
function (CF) of the decision variable. Using the inversion
coverage of the SAP is a circular area around the origin with
theorem, the probability that the decision variable surpasses
radius R. The energy of the SoI at the SAP can be written
the threshold can be found as
as Pu rf−ν h2 , where Pu is the transmit power of the UE. We
assume there is no power control in the macrocell network, P[VED > ζ] =
and therefore, Pu is independent of the distance between the
∞ 
1 1 ψVED (−jω)ejωζ − ψVED (jω)e−jωζ
UE and the MBS.5 The noise term has a zero-mean Gaussian − Re dω
distribution n(t) ∼ N (0, σn2 ). The interfering nodes consist 2 2π 0 jω
of macrocell users, small cell users belonging to other SAPs, (9)
or (cognitive) devices which are active on the same band. where ψVED represents the CF of VED . Under H1 , ψVED is
Therefore, the aggregate network interference measured at the provided in the following theorem for a typical user, i.e. a
SAP can be written as user with random location within the small cell coverage.
∞  ∞ 
hl Xi Theorem 1: In the presence of Rayleigh block fading, the
l
i(t) = Re il (t) = Re ν/2
(4) CF of the ED decision variable for a typical user in the
l=1 l=1 rf,l presence of interference uncertainties is given by (10) where
where we model the lth interferer amplitude Xil = Xl,1 i
+ 2 F1 (.) is the Gaussian hypergeometric function.
i Proof: See Appendix A.
jXl,2 as a zero-mean complex random variable (r.v.). Note
For the calculation of Pfa in the presence of aggregate
that the r.v.’s Xil are circular symmetric, and independent and
interference, we apply the same methodology as in Appendix
identically distributed (i.i.d.) in l since the interferers transmit
A, yet in absence of the SoI. The CF of the decision variable
independently. Therefore, with the interfering nodes scattered
under H0 can be expressed as (11) at the top of the next page,
5 Since the users that benefit most from the deployment of small cells are and Pfa can be obtained applying the inversion theorem.
located far away from the MBS, our model holds for the cell edge users who
transmit at maximum power. 6 For simplicity, we assume N = τs fs
4020 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 9, SEPTEMBER 2013

1
ψVED |H0 (jω) = N/2
(1 − 2jωσW
2 )
  2/ν    
 jω  jω
× exp −2 γ cos(π/ν) 
2/ν 
 1 − sign tan(π/ν) (11)
1 − 2jωσ 2 W 1 − 2jωσ 2 W



     2/ν
R ∞
1 σn2 x ν 2π 2 2π Pi x
C̄(λ, R, ν) = 2 exp − (e − 1)rf exp − csc λ (e − 1) rf2 dx2rf drf (13)
R 0 0 Pu ν ν Pu

IV. T RAFFIC O FFLOAD where C̄ = E[ln(1 + η)] is the average channel capacity in
uplink for a typical UE7 and η is the signal-to-interference-
In order to evaluate the trade-off between the energy con- and-noise ratio (SINR). The aggregate offload capacity is a
sumption and the achievable traffic offload, we define the function of the densities of MBSs, SAPs, and mobile users,
small cell aggregate offload capacity and aggregate offload where λu affects the offload capacity by changing the number
throughput which reflect the traffic that can be accommodated of loaded small cells.8 When the SoI and all the interfering
by the SAPs in a macrocell and also account for the sensing signals are affected by Rayleigh fading, we can derive the
probability, the sensing time, and the sensing performance. average channel capacity in the next theorem as a special case
We consider a saturated resource model, intending that each of [38].
user always has a non-empty queue of data to be transmitted. Theorem 2: A typical user is uniformly distributed over
The small cell capacity assumes adaptive modulation while the coverage of the SAP, i.e. a circular area with radius R,
the small cell throughput assumes a fixed transmission rate. while the interfering nodes are spatially scattered over the two-
Note that an exact analysis of the aggregate offload capacity dimensional plane R2 according to a homogeneous PPP. The
and throughput requires an exact distribution of the area of average channel capacity in the uplink of a typical user within
the small cells which constitute a multiplicatively weighted the coverage of the associated SAP for a Rayleigh fading
Voronoi tessellation. The area distribution of a Voronoi tessel- channel is given by (13) where Pu and Pi are the transmission
lation is an open research problem for which an approximation power of the UE and of each interferer.
has been proposed in [37]. Yet, in order to present a tractable Proof: See Appendix B.
analysis, we consider a distance based association policy with Using Theorem 2, we formulate the following corollary for a
respect to the small cells. This relaxation of the original special case.
problem yields a unified framework that allows us to obtain Corollary 1: For ν = 4, the average channel capacity of a
both the detection performance and aggregate offload capacity typical user can be expressed as
and throughput.
∞   2 
1 π a (x)
C̄(λ, R, 4) = exp
2R2 b(x) 4b(x)
 0

A. Aggregate Offload Capacity a(x) a(x) + 2b(x)R2
× erfc  − erfc  dx (14)
2 b(x) 2 b(x)
The aggregate capacity that can be offloaded per macrocell

is given by the typical user uplink capacity in an SAP where a(x) = (π 2 /2)λ (ex − 1)Pi /Pu and b(x) =
multiplied with the number of loaded and active SAPs within (σn2 /Pu )(ex − 1).
the macrocell coverage. Assuming OFDMA or TDMA, the Proof: The proof of the corollary follows straightfor-
typical user uplink capacity in an SAP equals the sum-capacity wardly from [27] with some simple mathematical manipula-
of N users sharing the same resource block. The expected tions.
number of SAPs per macrocell is given by λs /λm [29]. We The expression of the average channel capacity when ν =
assume a distance based access policy with respect to the 4 can be bounded
SAPs, where access can be granted when the mobile user is √ using√ the bounds of the Q func-
tion exp(−x2 /2)/ √ π(x/ 2 + x2 /2 + 2) < Q(x) <
within the coverage R of the SAP. For this access policy, the √
exp(−x2 /2)/ π(x/ 2 + x2 /2 + 8/π).
probability that an SAP has at least one active user within
the SAP coverage can be found by the complement of the
null probability of a Poisson r.v. in (1) and is given by B. Aggregate Offload Throughput
(l)
pl = 1 − exp(−pu λu πR2 ). Let Ns = pl λs /λm be the Assuming a constant transmission rate, the throughput that
expected number of SAPs in a macrocell with active users, the small cell tier can accommodate and that accounts for the
then we define the capacity that can be offloaded from an 7 Note that we consider X
i,l complex Gaussian and the interferers quasi-
MBS that accounts for the sensing procedure as follows: static.
8 Note that we assume here that the coverage areas of different SAPs do
T − τs not overlap, which is realistic considering a reduction of the coverage range
ξc (τs ) = Ns(l) ps Pd C̄(λ, R, ν) (12) with increasing small cell density.
T
WILDEMEERSCH et al.: COGNITIVE SMALL CELL NETWORKS: ENERGY EFFICIENCY AND TRADE-OFFS 4021


   2/ν  
R2
1 σn2 ν/2 2π 2 Pi ηt 2π
Ps (ηt ) = 2 exp − ηt x exp − λ csc x dx (16)
R 0 Pu ν Pu ν

sensing strategy can be expressed as (2), respectively. If we select a detection threshold ζ  < ζ ∗
T − τs such that Pd (τs , ζ  ) > Pd (τs , ζ ∗ ) and Pfa (τs , ζ  ) > Pfa (τs , ζ ∗ ),
ξt (τs , ηt ) = Ns(l) ps Pd ρ (15) then we also have E2 (Pd (τs , ζ  ), τs ) > E2 (Pd (τs , ζ ∗ ), τs )
T
and E3 (Pfa (τs , ζ  ), τs ) > E3 (Pfa (τs , ζ ∗ ), τs ). Therefore, the
where ρ = ln(1 + ηt )Ps (ηt ) is the throughput in uplink for optimal solution is achieved under the equality constraint and
a typical UE, ηt is the target SINR, and where the success the optimization problem can be reformulated as
probability is defined as Ps (ηt ) = P[η > ηt ].
Theorem 3: The success probability for a typical user in min λs /λm [Ξc T + ps Ξs τs + pUE ps P∗d Ξt (T − τs )
the coverage of an SAP and in the presence of network τs

interference is given by (16). + (1 − pUE )ps Pfa (τs , ζ ∗ )Ξt (T − τs )]. (20)
Proof: See Appendix C.
For the interference limited case, (16) becomes For the special case of a user with fixed position and using
the Gaussian approximation for the decision variable, the
exp(−ϑR2 ) − 1 threshold value corresponding with P∗d is given by [32]
Ps (ηt ) = (17)
ϑ 
ζ ∗ = σn2 (1 + η)[1 + Q−1 (P∗d ) 2/(τs fs )] (21)
where ϑ = πλ(2π/ν) csc(2π/ν)(Pi ηt /Pu )2/ν . For the special
case of ν = 4, we formulate the following corollary. and Pfa under the constraint of Pd is given by
Corollary 2: For ν = 4, the success probability of a typical 
user can be expressed as (1 + η)[1 + Q−1 (P∗d ) 2/(τs fs )] − 1
Q  . (22)
  2 2/(τs fs )
1 π c
Ps (ηt ) = exp
2R2 d 4d Under this scenario, we formulate the following proposition.
    
c c + 2dR2 Proposition 1: For a user with a fixed position and using
× erfc √ − erfc √ (18)
2 d 2 d the Gaussian approximation, the optimization problem as
  defined in (20) has a unique optimal solution τs∗ .
where c = π2 λ csc π2 (ηt Pi /Pu )1/2 and d = σn2 ηt /Pu .
2
Proof: See Appendix D
Proof: Similar to proof in [27]. 2) Joint optimization of sensing time and sensing probabil-
ity: Since it is energetically inefficient to sense continuously,
V. T RAFFIC O FFLOAD VERSUS E NERGY C ONSUMPTION to reduce idle listening we optimize the energy consumption
T RADE -O FF with respect to ps and τs subject to constraints on ξc and Pd .
In this section, we show how the framework developed in The optimization problem can be formulated as
Sections III and IV can be useful for the design of energy
min Etot
efficient small cell networks by means of several optimization ps ,τs
problems. s.t. ξc ≥ ξc∗ , Pd ≥ P∗d (23)

which can be solved applying a two-step approach. For every


A. Optimization of energy consumption constrained by traffic
value of τs ∈ [0, T ], the optimal sensing probability can be
offload
calculated reformulating the optimization problem as follows:
In the following, we first investigate the energy minimiza-
tion with regard to the sensing time, and then optimize the min Etot
ps
energy consumption with respect to sensing time and sensing
probability. s.t. ξc ≥ ξc∗ , Pd ≥ P∗d . (24)
1) Optimization of sensing time: In this problem, the objec- For a fix value of τs , we follow the argumentation of Section
tive of the system design is to offload the traffic whenever there V-A to conclude that the optimal sensing probability can
is an active user within the coverage of an SAP. Therefore, a be found under the equality constraint. The optimal sensing
constraint on Pd is imposed and the problem can be formulated probability is given by
as
ξc∗
min Etot p∗s = (l) ∗ T −τs
. (25)
τs Ns Pd T C̄
s.t. Pd (τs , ζ) ≥ P∗d (19)
To present a tractable analysis, we consider a user at a
where 0 ≤ τs ≤ T and P∗d is the target probability of detection. fixed distance from the SAP and apply the Gaussian ap-
For a given sensing time, a threshold ζ ∗ can be chosen as to proximation for the decision variable. The minimal energy
satisfy the constraint Pd (τs , ζ ∗ ) = P∗d . Let E2 (Pd , τs ) and consumption can be found by inserting
 (25) into (2). With
E3 (Pfa , τs ) denote the second and the third term of Etot in Pd = P∗d , Pfa = Q((ζ ∗ − σn2 )/( 2/(τs fs )σn2 )), and ζ ∗ =
4022 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 9, SEPTEMBER 2013

 
Q−1 (P∗d ) 2/(τs fs )σn2 (1 + η) + σn2 (1 + η), (23) can be With a change of variable (ζ ∗ −σn2 (1+η))/( 2/(τs fs )σn2 (1+
rewritten as η)) → u, it can be shown that there exists an optimal sensing
time that yields the optimal capacity. Specifically, the optimal
ξ∗T
min K(τs ) = min Ξs τs sensing time can be found as follows
τs τs (l)
Ns P∗d (T − τs )C̄  
e−u /2 du
2

ξ ∗ T Ξt ∂ξc (τs )
+ (pUE P∗d + (1 − pUE )Pfa (τs , ζ ∗ )) (26) = Ns ps C̄ −Q(u) + (T − τs ) √
(l)
=0
∂τs 2π dτs
Ns P∗d C̄
(l)
(30)
  
where the optimal solution can be found numerically.
where Pfa (τs , ζ ∗ ) = Q Q−1 (P∗d )(1 + η) + η τs fs /2 .
Since K(τs ) is differentiable on [0, T ], and K (τs ) is increasing VI. SAP P ERFORMANCE L IMITS
on [0, T ], K(τs ) is convex and there exists an optimal τs∗
In the former section, we illustrated how the proposed
for (26). For the limits of the sensing time interval, the first
framework can be used to design energy efficient SAPs. In
derivative of K is given by
this section, we focus on the limits of the energy detector and
lim K (τs ) = ∞ obtain an expression of the Pfa decay rate as a function of
τs →T the aggregate network interference. As such, in the context
∗ ∗
ξ Ξs ξ T Ξt
lim K (τs ) = P∗
+ P∗
(1 − pUE )Pfa (τs , ζ ∗ ) (27) of green communications, this analysis confines a region of
τs →0 d C̄ d C̄
interferer densities where the cognitive SAP can be used
where the last expression is negative if Pfa (0, ζ ∗ ) < advantageously in order to reduce the network wide energy
−Ξs /(Ξt T (1 − pUE )). Under this condition, the optimal consumption.
sensing time can be found in the range [0, T ] by applying
the bisection algorithm with tolerable sensing time error . The A. Interference wall
two-step optimization is illustrated in the following algorithm. In [32], [39], environment-dependent uncertainties are
shown to be the cause of the so-called SNR wall, below which
the detector is not robust regardless the sensing time. Noise
Algorithm 1 Joint optimization of the sensing time and uncertainty caused by the noise estimation has been discussed
sensing probability in [40]. An SAP with a UE within its coverage finds itself
Initialise τmin ← 0, τmax ← T in a high-SNR environment, and therefore, the SNR-wall due
while τmax − τmin >  do to noise estimation is not relevant in this scenario. However,
τs ← (τmin + τmax )/2 another source of uncertainty is the network interference. In
p∗s ← ξ ∗ T /(P∗d (T − τs )C̄) the following, we derive an expression of the noise uncertainty
if ∂Etot (p∗s , τs )/∂τs < 0 then due to the network interference.
τmin ← τs 1) Unbounded path-loss model: Under H0 , the decision
else variable can be written as
τmax ← τs N
end if 1
VED = (i[k] + n[k])2 . (31)
end while N
k=1
As discussed in Appendix A, the√ network interference
can be decomposed as i[k] = U G, where U is a
skewed stable distribution and G follows a normal dis-
B. Optimization of QoS under energy constraint
tribution. Therefore, the received signal under H0 condi-
The system designer can decide to limit the energy con- tioned on U can be expressed as a normal r.v. r[k]|H0 ,U ∼
sumption of a small cell. In this scenario, the optimization N (0, σn2 + 2γ ν/2 U ). With γ defined as in (5), the vari-
problem can be cast as ance of r[k]|H0 ,U can be written as σtot 2
= σn2 +
ν/2 −1 i 4/ν ν/2 ν/2
2γ U = σn (1 + 2ηi (πC4/ν E[|hl Xl,p | ]) λ U ), such
2
max ξc (τs )
τs that ηi is the interference-to-noise ratio (INR). Let G =
−1 i 4/ν ν/2
s.t. Pfa ≤ P∗fa . (28) 2ηi (πC4/ν E[|hl Xl,p | ]) U , then the total variance takes
values in the interval σtot2
∈ [σn2 , σn2 (1 + λν/2 G)] since the
Following a similar reasoning as in Section V-A, we can show
skewed stable distribution U only takes positive values. For
that the optimal solution can be found under the equality
high values of the sensing time, the central limit theorem
constraint. If we consider network interference, the detection
can be applied and Pd and Pfa can be found in terms of Q-
threshold ζ can be found numerically using (9) and (11).
functions. In order to be robust with respect to the network
In the case when we apply the Gaussian approximation for
interference, we get
the decision variable, the
 threshold satisfying the constraint is
given by ζ ∗ = σn2 (1+ 2/(τs fs )Q−1 (P∗fa )). The optimization ζ − (1 + λν/2 G)σn2
Pfa = Q 
problem can now be reformulated as 2/N (1 + λν/2 G)σn2

ζ ∗ − σn2 (1 + η) ζ − (1 + SNR)σn2
max Q  (T − τs ). (29) Pd = Q  . (32)
τs 2/(τs fs )σn2 (1 + η) 2/N (1 + SNR)σn2
WILDEMEERSCH et al.: COGNITIVE SMALL CELL NETWORKS: ENERGY EFFICIENCY AND TRADE-OFFS 4023

  2−α 1
2 + 2 (α−2)
2 (α − 3)(α − 2)(d2−ν
min − dmax )(ν − 1)
2−ν
I= (d2−ν − d2−ν
max ) ηi πσh2 . (36)
ν/2 − 1 min (dmin − dmax )(ν/2 − 1)Pi σh2
2−2ν 2−2ν

We define the sample complexity N ∗ as the sensing time B. False alarm decay
needed to obtain the target Pd and Pfa . Eliminating ζ from To express the relationship between energy consumption
the equations in (32) and solving to N , the sample complexity and interfering node density, we determine how fast Pfa
can be written as converges to its target value. As such, we provide a tool to
2(Q−1 (Pfa )(1 + Gλ2 ) − Q−1 (Pd )(1 + SNR))2 evaluate how increasing the sensing time affects the energy
N∗ = . efficiency. A direct method to obtain the PDF of VED as
(SNR − Gλν/2 )2
(33) a function of the sensing time is cumbersome9. Instead, we
It follows that as λ → (SNR/G)2/ν , we have N ∗ → ∞. will use tools from large deviations theory to determine how
In other words, a target Pfa and Pd cannot be attained fast the target Pfa can be reached. According to the Cramer
within a finite sensing time for some interfering node density theorem, we have for interference-limited networks

approaching (SNR/G)2/ν . We call this limit the interference Pfa (ζ) = P[1/N i2 [k] > ζ] ≤ e−N I(ζ) (38)
wall λwall,s and note that λwall,s is a function of the SNR, INR, N
the power path loss exponent, and the fading. However, to which decays exponentially with the sensing time and the
calculate G, a percentile of the distribution has to be selected decay rate is determined by the rate function I(x). In order
that corresponds to a worst case, due to the heavy tails of the to have finitemoments, we model the aggregate interference
stable distribution. power IΩ = i2 [k] according to a truncated stable distribu-
2) Bounded path-loss model: The heuristic approximation tion. The CF of the truncated stable distribution is given by
based on the stable distribution is computationally intensive [42]
and sensitive to the selection of the percentile. In order to    

find a solution with higher accuracy, we consider that the ψIΩ (jω) = exp γ  Γ(−α ) (g − jω)α − g α (39)
interferers are located in the annulus A, defined by the radii
where α is chosen equal to the characteristic exponent of
dmin and dmax . Under such conditions, it can be shown that
the stable distribution in the unbounded path loss model. The
i[k] in (31) follows a truncated stable distribution [41]
parameters of the truncated stable distribution can be found
i[k] ∼ St (γ  , α = 4/ν, g) (34) using the method of the cumulants. From (39), the cumulants
of the truncated stable distribution can be expressed as
1 dn 
where α corresponds to the characteristic exponent of the sta- 
κI (n) = n n ln ψIΩ (jω)
ble distribution, γ  corresponds to the dispersion, and g reflects j dω ω=0

the decaying of the tail of the truncated stable distribution. = (−1)n γ  Γ(−α )g α −n Πn−1 
i=0 (α − i). (40)
The coefficients γ  and g can be determined by the method
of the cumulants, by imposing the equality of the second and Building on Cambell’s theorem [41], the cumulants of the
fourth cumulant of the truncated stable distribution with the aggregate interference can be expressed as
respective cumulants of the network interference. Applying 2πλ
the approach of Section VI-A1, we further approximate the κ(n) = Pin (d2−nν − d2−nν
min )μh2 (n). (41)
2 − nν max
interference by a Gaussian r.v., such that the received signal
Using (40) and (41), the parameters γ  and g can be written
follows a normal distribution r[k]|H0 ∼ N (0, σn2 + σi2 ), where
as a function of the first two cumulants as follows
σi2 represents the second order moment of the truncated stable
−κ(1)
distribution γ =   
 ) α −1

∂ 2 Mi (t, λ) Γ(−α )α κ(1)(1−α


κ(2)
σi2 = E[i2 [k]] = |t=0 = σn2 Iλ (35)
∂t2 κ(1)(1 − α )
g = . (42)
κ(2)
where Mi (t, λ) is the moment generating function (MGF) of
the truncated stable distribution. After some manipulations, Since i2 [k] are assumed to be i.i.d., the Cramer theorem can be
I can be expressed as (36). Note that the variance of the applied and we can express the rate function as the Legendre-
network interference is linear in the interferer density λ and Fenchel transform of the logarithmic MGF
the parameter I = f (dmin , dmax , ν, ηi , σh2 ). For ν = 4,   

I(x) = sup θx − γ  Γ(−α )[(g − θ)α − g α ] . (43)
the parameter I further simplifies to 2(d−2 −2
min − dmax )Pi πσh .
2
θ
Eliminating ζ from the expressions of Pd and Pfa , the sample For θ < g, let the first derivative be equal to zero and solving
complexity can now be expressed as to θ for ν = 4, we have
  2
2(Q−1 (Pfa )(1 + Iλ) − Q−1 (Pd )(1 + SNR))2 −γ Γ(−α )
N∗ = (37) θ=g− . (44)
(SNR − Iλ)2 2x
and the interference wall is given by λwall,t = SNR/I. 9 Note that we expressed the CF of VED in Section III-B
4024 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 9, SEPTEMBER 2013

1
1.4 Pu /Pi = 5dB
Pu /Pi = 0dB
0.9 Pu /Pi = −5dB
1.2

Typical Uplink Capacity [nats/s/Hz/SAP]


0.8
sparse
0.7
1

0.6
CCDF

0.8
0.5
dense

0.4 0.6

0.3
0.4
0.2
λ = 0.015
λ = 0.01
λ = 0.005
0.1 λ = 0.0005 0.2
λ = 0.0001
λ=0
0
−30 −25 −20 −15 −10 −5 0 5 10 15 20 0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
SINR [dB] λ [m−2 ]

Fig. 3. The success probability in the presence of sparse and dense interferers. Fig. 4. The average channel capacity corrected by the detection probability
for different values of the power ratio between the SoI and the interferers.

Substituting θ in (43), the expression of the rate function is


1
given by
√ (γ  Γ(−α ))2 0.98

I(x) = gx + gγ  Γ(−α) + . (45)


4x 0.96

VII. N UMERICAL R ESULS 0.94


Etot /Etot,max

In this section, we present some numerical results that 0.92

provide insight into the trade-off between energy consumption


and traffic offload of the SAPs within a MBS Voronoi cell. 0.9

0.88

A. Traffic offload
0.86
τs = 100, ζ = cst
Figure 3 depicts the success probability for a typical user τs = 15, ζ = cst
τs = 100, Pd = cst
within the coverage of an SAP. With the interferers distributed 0.84
τs = 15, Pd = cst

according to a PPP, this figure illustrates that the success 0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018
λ [m−2 ]
probability decreases drastically with increasing interfering
node density. Figure 4 shows Pd C̄ (the uplink capacity of
a typical user in the SAP coverage corrected for the detection Fig. 5. Total energy consumption for different values of the sensing time
and with fix or variable detection threshold value.
probability) as a function of the interferer density. The figure
illustrates the combined effect of the increasing detection
performance and the decreasing average channel capacity with Figure 5 shows the total energy consumption considering
λ. a constant target Pd and a constant threshold, respectively.
We observe that for all scenarios the energy consumption
B. SAP energy efficiency grows with increasing interfering node density. This can be
Since this work focuses on the trade-off between the energy ascribed to the fact that the energy consumption is linear
consumption and the aggregate offload throughput and capac- in Pd and Pfa . As more energy is provided to the ED
ity, it is meaningful to analyze how the energy consumption with increasing interferer density, Pd and Pfa increase for
depends on the interferer density. In the following numerical the constant threshold scenario. In this numerical example,
examples, we consider that SNR = 3 dB defined for the UE pUE = 0.1 and therefore, the increase of energy consumption
at the edge of the SAP coverage R = 20, while the INR = is dominated by Pfa . If information is available about the
10 dB defined at a distance of 1 meter (far-field assumption). interference environment, the threshold can be altered such
We consider the densities of MBSs, SAPs, and users to be that Pd = P∗d . Raising the threshold with the interferer density,
λm = 10−6 m−2 , λs = 10−5 m−2 , and λu = 10−4 m−2 . moderates the increase of Pfa and the energy consumption,
From (2), we observe that the power consumption consists which is reflected in the figure. This means that the knowledge
of three components Ξc , Ξs , and Ξt , which we assume to be of the interference environment allows to improve the energy
constant and given by Ξc = 6W , Ξs = 4W , and Ξt = 5W efficiency of small cell networks.
[43]. Unless otherwise specified, we set the probabilities Figure 6 illustrates the effect of the SAP coverage range
pI = 0.1, ps = 1, and pUE = 0.1. Furthermore, we let the and the interferer node density on the energy consumption.
frame length T = 400/fs and the maximum duty cycle (DC) The energy consumption increases with almost 50 % when R
is 25 %. varies from 15 to 50 m. The sensing time is fixed to N = 15
WILDEMEERSCH et al.: COGNITIVE SMALL CELL NETWORKS: ENERGY EFFICIENCY AND TRADE-OFFS 4025

1 1
λ = 0.015 λ = 0.01
λ = 0.01 λ = 0.005
λ = 0.005 0.9 λ = 0.001
0.95

0.8

0.9
0.7
Etot /Etot,max

Etot /Etot,max
0.85
0.6

0.5
0.8

0.4
0.75

0.3

0.7
0.2

0.65 0.1
15 20 25 30 35 40 45 50 0.05 0.1 0.15 0.2 0.25
coverage [m] sensing time (DC)

Fig. 6. Total energy consumption for different values of the interferer density. Fig. 7. Etot for different values of the interferer density and with pUE =
0.1. The optimization is performed subject to P∗d = 0.9 and ξc∗ = 0.5
nats/s/Hz/macrocell.

samples and Pd = P∗d = 0.9 for all values of the coverage. To


satisfy the constraint on Pd , the detection threshold decreases
for a larger coverage, and as a consequence Pfa increases. 1
λ = 0.015
λ = 0.01
This is reflected in the total energy consumption. From the λ = 0.005
0.9
numerical results presented in Fig. 6, we can conclude that
the energy consumption is affected considerably by changing 0.8

the SAP coverage due to the strong relation between the user
detection performance and the coverage range. 0.7
ξc /Etot

In Section V-A2, we defined an optimization problem that


0.6
minimizes the energy consumption subject to constraints on
both the traffic offload and the detection probability. Figure 7 0.5

depicts the objective function (23) of the energy consumption


optimization problem for different interferer densities. We 0.4

ascertain the convexity of the energy consumption of all SAPs


0.3
belonging to a macrocell Voronoi cell subject to constraints
on Pd and ξc . The initial decrease of the energy consumption 0.05 0.1 0.15
sensing time (DC)
0.2 0.25

is due to the decrease of Pfa with the sensing time until


Pfa reaches a stationary value. Any further increase of the
Fig. 8. Energy efficiency for different values of λ as a function of the sensing
energy consumption is related to the sensing energy accruing duty cycle.
with the sensing time. Furthermore, this example reveals that
for increasing interferer density, the optimal sensing time
decreases, which can be ascribed to a higher stationary Pfa
with increasing λ. The impact of λ on the optimal sensing Defining the load of a typical base station (SAP or MBS) as
time is however negligible. This result highlights that the the number of active users within the coverage, we notice that
proposed framework can be used to find the optimal sensing the small cell load is related to the density of the active users
time and sensing probability. Moreover, we observe that the and the small cell coverage. Owing to the major coverage
energy consumption can be improved considerably and the differences between the tiers, the load of a typical SAP
energy consumption gain by joint optimization of τs and ps and MBS differ considerably such that the aggregate offload
increases with larger interferer density. capacity depends on λm , λs , and λu . Figure 9 illustrates how
To elucidate the trade-off between the energy consumption the energy consumption and the traffic offload of the SAPs
and the traffic offload, we define the energy efficiency as ς = within a macrocell vary as a function of λu and how they
ξc /Etot . Figure 8 shows the effect of the sensing time and relate. The numerical results illustrate the impact of the quality
the interferer density on the energy efficiency. We observe that of the cognitive capabilities on both the energy consumption
there exists an optimal value of the sensing time balancing the and traffic offload. In this scenario, we assume λu to vary as
two opposing requirements. With respect to the optimization 10−6 < λu < 10−3 and that SIR = 0dB. For sensing DC =
of the energy consumption subject to a constraint on Pd , this 0.05, the energy efficiency ς is clearly superior than for higher
figure shows that with increasing interferer density the optimal values of the duty cycle. In case of perfect sensing and low
sensing time shifts to smaller values. This effect is related to values of the load, the energy consumption is lower than the
the linear decrease of the traffic offload with increasing sensing curves with realistic detection performance due to Pfa = 0,
time, as can be verified in (12) or (15). and this effect diminishes with increasing load.
4026 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 9, SEPTEMBER 2013

1 0.45
DC = 0.05 λ = 0.001
DC = 0.15 λ = 0.005
DC = 0.25 0.4 λ = 0.01
DC = 0.25 − perfect sensing
0.95
0.35
σi2 = 10 dB

0.3
0.9
Etot /Etot,max

0.25

I(x)
0.85 0.2

0.15
0.8 σi2 = 15 dB

0.1

0.05
0.75

0
−1 0 1 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
10 10 10
ξc [bits/s/Hz/macrocell] x

Fig. 9. Energy consumption for different values of the sensing time and for Fig. 11. The rate function is given for different values of the interference
the ideal case of perfect sensing as a function of the load. power and density.

800
simulation
consumption and reflects the achievable energy efficiency.
stable distribution
truncated distribution
700

VIII. C ONCLUSIONS
600
In this paper, we proposed an analytical framework that
sample complexity

500 allows to analyze the trade-off between the energy consump-


tion and the traffic offload of cognitive SAPs located in an
400
MBS Voronoi cell. In particular, we defined the detection
300 performance of an energy detector for a typical user, as
well as the aggregate capacity and throughput that can be
200
offloaded from an MBS. The model accounts for channel
100
fading, aggregate network interference, bursty activity, net-
work topology, and load. The proposed model allows to
0
10
−4 −3
10
−2
10
quantify the effect of critical system parameters such as the
λ [m−2 ] interferer density and the SAP coverage on the detection
performance, the aggregate offload capacity and throughput,
Fig. 10. The interference wall using the approximation with the 87 percentile and the total energy consumption. Numerical results reveal
of the stable distribution, using the truncated stable distribution and using that the knowledge of the interference environment can lead
numerical simulations for P∗fa = 0.1 and P∗d = 0.9, SNR = 3 dB and INR =
20 dB. to a substantial reduction of the SAP energy consumption.
We defined several optimization problems and showed that
the proposed framework can be used both for the design of
C. Performance limits the optimal sensing time and sensing probability, as for the
evaluation of the energy efficiency with respect to network
Figure 10 shows the sample complexity to satisfy a target topology and load. Further, fundamental limits of the detection
Pfa and Pd for increasing interferer density. This figure robustness are defined by introducing the interference wall and
illustrates the fundamental limit of the interfering node density the speed of convergence of Pfa . In conclusion, the framework
under which the SAP can robustly detect the macrocell user can be used for the energy efficient design and operation of
presence. Beyond the interference wall, the noise uncer- cognitive SAPs in heterogeneous networks. Possible future
tainty becomes too big to distinguish between SoI or noise. directions to extend this work are to consider the combined
The curves are drawn using the 87 percentile of the stable energy consumption of multiple tiers, include constraints on
distribution, the truncated stable distribution and numerical the user quality of service, and compare the effectiveness of
simulations. For dmin = 1 and dmax = 100, the approaches the distributed sleep mode scheme with strategies such as
with the stable and the truncated stable distribution are in random sleeping or centralized sleep mode schemes exploiting
good agreement and correspond well with the numerical user location awareness.
simulation. Figure 11 shows the Pfa decay rate as expressed
in (45) as a function of the threshold x for different values
of the interference power and the interferer density. It can be IX. ACKNOWLEDGMENT
observed that with increasing density and interference power This work was partly supported by the SRG ISTD 2012037,
the rate function decreases. Thus, the rate function can reveal CAS Fellowship for Young International Scientists Grant
more insight into the effect of λ and Pi on the SAP power 2011Y2GA02, SUTD-MIT International Design Centre under
WILDEMEERSCH et al.: COGNITIVE SMALL CELL NETWORKS: ENERGY EFFICIENCY AND TRADE-OFFS 4027

Grant IDSF1200106OH, the A∗ STAR Research Attachment A PPENDIX B


Programme, and the European Commission Marie Curie In- P ROOF OF T HEOREM 2
ternational Outgoing Fellowship under Grant 2010-272923.
Proof: The average capacity of a typical user can be
written as
A PPENDIX A
P ROOF OF T HEOREM 1 C̄ = Erf ,φ,h [ln(1 + η)]
√ √
Proof:
√ Let S = s[k]/ N , I = i[k]/ N and W = where the expectation is taken over the distance rf between
n[k]/ N with σn2 /N = σW 2
, then the discrete decision the UE and the SAP, over the spatial PPP φ of the interferers
statistic in (8) can be rewritten as and over the fading distribution h. For a positive r.v. X,

N 
2 E[X] = 0 1 − FX (x)dx with FX (x) the cumulative distri-
h
VED|H1 = ν/2
S[k] + I[k] +W [k] . bution function. Following the approach of [27], the average
k=1 r throughput can be expressed as
f  
K  

R

σn2 x ν
Conditioning on K, VED follows a non-central chi-square C̄ = exp − (e − 1)rf
Pu
distribution and therefore, the CF of VED is given by 0 0
 x 
  e −1 ν 2rf
1 jωN K 2 × LI r dx 2 drf (47)
ψVED |H1 ,K (jω) = exp . Pu f R
(1 − 2jωσW 2 )N/2 1 − 2jωσW2

(46) where LI (s) is the Laplace transform of the network inter-


Since the network interference follows a symmetric stable ference I. Applying the probability generating functional of
distribution, the decomposition property can be √ applied and the PPP and assuming that the interfering signal is affected
the interference can be represented as I = U G, with by Rayleigh fading, we can write
U ∼ S(2/ν, 1, cos( πν )) and G ∼ N (0, 2γ ν/2 /N ). If we  
assume that the SoI has a normal distribution, then K 2 in ex − 1 ν
LI rf =
(46) stands for the power of a normally distributed r.v. with P
variance h2 Pu /(rfν N ) + 2γ ν/2 U/N . Let V = 2γ ν/2 U , then  u
∞ 
1
we can write exp −2πλ 1− udu
0 1 + PPui (ex − 1)(rf /u)ν
ψVED |H1 ,h,rf ,V (jω) =
  which by a change of variables can further be simplified to
1 jωN (h2 Pu /(rfν N ) + V /N )
exp .  
(1 − 2jωσW 2 )N/2 1 − 2jωσW 2
ex − 1 ν
LI rf =
P
The exponential can be expanded as the product of two u    2/ν
exponentials. Further deconditioning on h and V , we get 2π 2 2π Pi
exp − csc λ (ex − 1)2/ν rf2 . (48)
ν ν Pu
1
ψVED |H1 ,rf (jω) = N/2−1
(1 − 2jωσW 2 )
Inserting (48) in (47), the proof is concluded.
 
1 jω
× ψV .
1 − jω(Pu /(2rfν ) + σW2 ) 1 − 2jωσW
2
A PPENDIX C
Using the scaling property of a stable random variable, the CF P ROOF OF T HEOREM 3
of the decision variable can be written as
1 Proof: The success probability for a typical user in the
ψVED |H1 ,rf (jω) = N/2−1 coverage of the SAP is given by
(1 − 2jωσW
2 )

1  2 
× h Pu rf−ν
1 − jω(Pu /(2rfν ) + σW
2 ) Ps (ηt ) = Erf ,φ,h P > ηt .
  2/ν σn2 + I
 jω 
× exp −2 γ cos(π/ν) 
2/ν 
 For a typical user uniformly distributed over the coverage of
1 − 2jωσ 2 W
    the SAP and considering Rayleigh fading, we can write

 ν    ν 
× 1 − sign tan(π/ν) . R
rf ηt σn2 2rf
1 − 2jωσW
2
Ps (ηt ) = exp − Eφ exp −
rf ηt I
drf
0 Pu Pu R2
We take the expectation with respect to rf bearing in mind
that the UE is within the range [0, R] and that the probability where Eφ [exp (−rfν ηt I/Pu )] = LI (rfν ηt /Pu ). Evaluating the
density function (PDF) of rf is given by fX (rf ) = 2rf /R2 . Laplace transform of the network interference (48) at rfν ηt /Pu ,
Solving the expectation, the proof is concluded. the proof is concluded.
4028 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 9, SEPTEMBER 2013

A PPENDIX D [19] E. Axell, G. Leus, E. G. Larsson, and H. V. Poor, “Spectrum sensing


P ROOF OF P ROPOSITION 1 for cognitive radio,” IEEE Signal Process. Mag., vol. 29, no. 3, pp.
101–116, May 2012.
Proof: Let G(τs ) be the objective function in (20). If [20] Q. Zhao and B. M. Sadler, “A survey of dynamic spectrum access,”
G(τs ) is differentiable over [0, T ], then G(τs ) is convex iff IEEE Signal Process. Mag., vol. 24, no. 3, pp. 79–89, May 2007.
[21] K. C. Chen and R. Prasad, Cognitive Radio Networks. Wiley, 2009.
G  (τs ) is increasing over the considered interval. We write the [22] T. V. Nguyen, H. Shin, T. Q. S. Quek, and M. Z. Win, “Sensing and
derivative of G(τs ) with respect to τs as probing cardinalities for active cognitive radios,” IEEE Trans. Signal
Process., vol. 60, no. 4, pp. 1833–1848, Apr. 2012.
G  (τs ) = λs /λm [ps Ξs − pu ps P∗d Ξt − (1 − pu )ps Pfa (ζ ∗ , τs )Ξt [23] L. Saker, S. E. Elayoubi, L. Rong, and T. Chahed, “Capacity and energy
efficiency of picocell deployment in LTE-A networks,” in Proc. 2011
+ (1 − pu )ps Ξt (T − τs )Pfa (ζ ∗ , τs )]. (49) IEEE Int. Veh. Technol. Conf., pp. 1–5.
[24] S. M. Cheng, S. Y. Lien, F. S. Chu, and K. C. Chen, “On exploiting
For ζ ∗ > σn2 , Pfa is positive and decreasing with τs , while cognitive radio to mitigate interference in macro/femto heterogeneous
networks,” IEEE Wireless Commun. Mag., vol. 18, no. 3, pp. 40–47,
Pfa is negative and increasing with τs . Therefore, the third Jun. 2011.
and fourth term in (49) are increasing over [0, T ]. Since the [25] C. H. M. de Lima, M. Bennis, and M. Latva-aho, “Coordination mech-
first two terms in the expression of G  (τs ) are constant, this anisms for self-organizing femtocells in two-tier coexistence scenarios,”
IEEE Trans. Wireless Commun., vol. 11, no. 6, pp. 2212–2223, Jun.
concludes the proof. 2012.
[26] Y. S. Soh, T. Q. S. Quek, M. Kountouris, and G. Caire, “Cognitive
hybrid division duplex for two-tier femtocell networks,” IEEE Trans.
R EFERENCES Wireless Commun., 2013, accepted for publication.
[1] T. Q. S. Quek, G. de la Roche, I. Guvenc, and M. Kountouris, Small [27] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to
Cell Networks: Deployment, PHY Techniques, and Resource Allocation. coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59,
Cambridge University Press, 2013. no. 11, pp. 3122–3134, Nov. 2011.
[2] D. Lopez-Perez, I. Guvenc, G. de la Roche, M. Kountouris, T. Q. [28] B. Blaszczyszyn, M. K. Karray, and H.-P. Keeler, “Using Poisson
S. Quek, and J. Zhang, “Enhanced intercell interference coordination processes to model lattice cellular networks,” in Proc. 2013 IEEE Joint
challenges in heterogeneous networks,” IEEE Wireless Commun. Mag., Conf. of the IEEE Computer and Commun. Societies, pp. 797–805.
vol. 18, no. 3, pp. 22-30, Jun. 2011. [29] S. G. Foss and S. A. Zuyev, “On a Voronoi aggregative process related
[3] J. G. Andrews, H. Claussen, M. Dohler, S. Rangan, and M. C. Reed, to a bivariate Poisson process,” Advances in Applied Probability, pp.
“Femtocells: past, present, and future,” IEEE J. Sel. Areas Commun., 965–981, 1996.
vol. 30, no. 3, pp. 497–508, Apr. 2012. [30] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms for
[4] H. Bogucka and A. Conti, “Degrees of freedom for energy savings cognitive radio applications,” IEEE Commun. Surveys Tuts., vol. 11,
in practical adaptive wireless systems,” IEEE Commun. Mag., vol. 49, no. 1, pp. 116–130, First Quarter 2009.
no. 6, pp. 38–45, Jun. 2011. [31] D. Bhargavi and C. R. Murthy, “Performance comparison of energy,
[5] Y. S. Soh, T. Q. S. Quek, M. Kountouris, and H. Shin, “Energy efficient matched-filter and cyclostationarity-based spectrum sensing,” in Proc.
heterogeneous cellular networks,” IEEE J. Sel. Areas Commun., vol. 31, 2010 IEEE Workshop on Signal Proc. Advances in Wireless Commun.,
no. 5, pp. 840–850, May 2013. pp. 1–5.
[6] Y. Chen, S. Zhang, S. Xu, and G. Li, “Fundamental trade-offs on green [32] R. Tandra and A. Sahai, “SNR walls for signal detection,” IEEE J. Sel.
wireless networks,” IEEE Commun. Mag., vol. 49, no. 6, pp. 30–37, Topics Signal Process., vol. 2, no. 1, pp. 4–17, Feb. 2008.
Jun. 2011. [33] J. F. C. Kingman, Poisson Processes. Oxford University Press, 1993.
[7] W. C. Cheung, T. Q. S. Quek, and M. Kountouris, “Throughput opti- [34] I. Ashraf, L. Ho, and H. Claussen, “Improving energy efficiency of
mization, spectrum allocation, and access control in two-tier femtocell femtocell base stations via user activity detection,” in Proc. 2010 IEEE
networks,” IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 561–574, Wireless Commun. and Networking Conf.
Apr. 2012. [35] M. Z. Win, P. C. Pinto, and L. A. Shepp, “A mathematical theory of
[8] A.-H. Tsai, L.-C. Wang, J.-H. Huang, and R.-B. Hwang, “High-capacity network interference and its applications,” Proc. IEEE, vol. 97, no. 2,
OFDMA femtocells by directional antennas and location awareness,” pp. 205–230, Feb. 2009.
IEEE Systems J., vol. 6, no. 2, pp. 329–340, Jun. 2012. [36] P. C. Pinto and M. Z. Win, “Communication in a Poisson field of
[9] S. Wu, C. Chen, and M. Chen, “Collaborative wakeup in clustered ad interferers—part I: interference distribution and error probability,” IEEE
hoc networks,” IEEE J. Sel. Areas Commun., vol. 29, no. 8, pp. 1585– Trans. Wireless Commun., vol. 9, no. 7, pp. 2176–2186, Jul. 2010.
1594, Sep. 2011. [37] J. S. Ferenc and Z. Néda, “On the size distribution of Poisson Voronoi
[10] A. Azad, S. Alouf, E. Altman, V. Borkar, and G. Paschos, “Optimal cells,” Physica A: Statistical Mechanics and its Applications, vol. 385,
control of sleep periods for wireless terminals,” IEEE J. Sel. Areas no. 2, pp. 518–526, Nov. 2007.
Commun., vol. 29, no. 8, pp. 1605–1617, Sep. 2011. [38] T. D. Novlan, H. S. Dhillon, and J. G. Andrews, “Analytical modeling
[11] W. Guo and T. O’Farrell, “Green cellular network: deployment solutions, of uplink cellular networks,” IEEE Trans. Wireless Commun., 2013,
sensitivity and tradeoffs,” in Proc. 2011 Wireless Advanced, pp. 42–47. accepted for publication.
[12] I. Ashraf, F. Boccardi, and L. Ho, “Sleep mode techniques for small [39] A. Sahai, S. M. Mishra, and R. Tandra, Cognitive Radios: System Design
cell deployments,” IEEE Commun. Mag., vol. 49, no. 8, pp. 72–79, Perspective. Springer, 2009, ch. Spectrum Sensing: Fundamental Limits.
Aug. 2011. [40] A. Mariani, A. Giorgetti, and M. Chiani, “Effects of noise power
[13] I. Haratcherev, M. Fiorito, and C. Balageas, “Low-power sleep mode estimation on energy detection for cognitive radio applications,” IEEE
and out-of-band wake-up for indoor access points,” in Proc. 2009 IEEE Trans. Commun., vol. 59, no. 12, pp. 3410–3420, Dec. 2011.
Global Telecomm. Conf., pp. 1–6. [41] A. Rabbachin, T. Q. S. Quek, H. Shin, and M. Z. Win, “Cognitive
[14] L. Saker, S. E. Elayoubi, R. Combes, and T. Chahed, “Optimal control network interference,” IEEE J. Sel. Areas Commun., vol. 29, no. 2, pp.
of wake up mechanisms of femtocells in heterogeneous networks,” IEEE 480–493, Feb. 2011.
J. Sel. Areas Commun., vol. 30, no. 3, pp. 664–672, Apr. 2012. [42] A. Rabbachin, A. Conti, and M. Z. Win, “Intentional network interfer-
[15] H. El Sawy and E. Hossain, “Two-tier hetnets with cognitive femtocells: ence for denial of wireless eavesdropping,” in Proc. 2011 IEEE Global
downlink performance modeling and analysis in a multi-channel envi- Telecomm. Conf., pp. 4268–4273.
ronment,” IEEE Trans. Mobile Comput., 2013, accepted for publication. [43] G. Auer, V. Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson,
[16] Y. C. Liang, Y. Zeng, E. C. Y. Peh, and A. T. Hoang, “Sensing- M. Imran, D. Sabella, M. Gonzalez, O. Blume, and A. Fehske, “How
throughput tradeoff for cognitive radio networks,” IEEE Trans. Wireless much energy is needed to run a wireless network?” IEEE Wireless
Commun., vol. 7, no. 4, pp. 1326–1337, Apr. 2008. Commun. Mag., vol. 18, no. 5, pp. 40–49, Oct. 2011.
[17] F. Moghimi, A. Nasri, and R. Schober, “Adaptive Lp -norm spectrum
sensing for cognitive radio networks,” IEEE Trans. Commun., vol. 59,
no. 7, pp. 1934–1945, Jul. 2011.
[18] S. M. Cheng, W. C. Ao, and K. C. Chen, “Efficiency of a cognitive radio
link with opportunistic interference mitigation,” IEEE Trans. Wireless
Commun., vol. 10, no. 6, pp. 1715–1720, Jun. 2011.
WILDEMEERSCH et al.: COGNITIVE SMALL CELL NETWORKS: ENERGY EFFICIENCY AND TRADE-OFFS 4029

Matthias Wildemeersch (S’09) received the M.S. Cornelis H. Slump received the M.Sc. degree
degree in Elecromechanical Engineering from the in Electrical Engineering from Delft University of
University of Ghent (Belgium) in 2003. He is cur- Technology, the Netherlands in 1979. In 1984 he
rently working towards the PhD degree in Electri- obtained his Ph.D. in physics from the University
cal Engineering at the University of Twente (the of Groningen, the Netherlands. From 1983 to 1989
Netherlands). From 2008 to 2011 he was at the he was employed at Philips Medical Systems in Best
Joint Research Centre of the European Commis- as head of a predevelopment group on image quality
sion, where he worked on the impact assessment and cardiovascular image processing. In 1989 he
of interference related to GNSS. Since 2012, he is joined the Network Theory group from the Univer-
with the Institute for Infocomm Research (Singa- sity of Twente, Enschede, the Netherlands. His main
pore) where his work broadly covers the analysis research interest is in detection and estimation, in-
of heterogeneous networks. His research interests span various aspects of terference reduction and pattern analysis as a part of digital signal processing.
wireless communications and signal processing and focus on the application In June 1999 he was appointed as a full professor and heads the Signals and
of probability theory and stochastic geometry to green communications, Systems group.
cognitive radio, and interference mitigation. He received the IEEE SPAWC
Alberto Rabbachin (S’03-M’07) received the Mas-
2013 Best Student Paper Award and is an awardee of the A*STAR Research
ter Degree in telecommunications engineering from
Attachment Programme (2012-2013).
the University of Bologna (Italy) and the Ph.D.
degree in electrical engineering from the University
Tony Q. S. Quek (S’98-M’08-SM’12) received the
of Oulu (Finland), in 2001 and 2008, respectively.
B.E. and M.E. degrees in Electrical and Electronics
In 2011, he joined the Laboratory for information
Engineering from Tokyo Institute of Technology,
and decision systems (LIDS) at the Massachusetts
Tokyo, Japan, in 1998 and 2000, respectively. At
Institute of Technology (MIT), where he is currently
Massachusetts Institute of Technology (MIT), Cam-
a Marie Curie Fellow. He was a postdoctoral re-
bridge, MA, he earned the Ph.D. in Electrical Engi-
searcher with the Institute for the Protection and
neering and Computer Science in Feb. 2008.
Security of the Citizen of the European Commission
Currently, he is an Assistant Professor with the
Joint Research Center from 2008 to 2011. His research interests involve
Information Systems Technology and Design Pillar
communication theory and stochastic geometry applied to real-problems in
at Singapore University of Technology and Design
wireless networks including network secrecy, cognitive radio, ultrawide-band
(SUTD). He is also a Scientist with the Institute
transceiver design, network synchronization, ranging techniques, and interfer-
for Infocomm Research. His main research interests are the application
ence exploitation. Dr. Rabbachin is an Editor of the IEEE C OMMUNICATION
of mathematical, optimization, and statistical theories to communication,
L ETTERS and served as symposium and workshop chair in various IEEE
networking, signal processing, information theoretic, and resource allocation
conferences. He received the IEEE Communications Society William R.
problems. Specific current research topics include cooperative networks,
Bennett Prize in the Field of Communications Networking (2012), the IEEE
heterogeneous networks, green communications, smart grid, wireless security,
Globecom best paper award (2010), the European Commission JRC best
compressed sensing, and cognitive radio.
young scientist award (2010), and the Nokia Fellowship (2005, 2006). He
Dr. Quek has been actively involved in organizing and chairing sessions,
is a recipient of the International Outgoing Marie Curie Fellowship (2011-
and has served as a member of the Technical Program Committee (TPC) in
2014).
a number of international conferences. He served as the Cognitive Radio &
Cooperative Communications Track Chair for the IEEE VTC in spring 2011,
the Wireless Communications Symposium Chair for the IEEE Globecom
in 2011, and the Communication Theory Symposium Chair for the IEEE
WCSP in 2013. He is currently an Editor for the IEEE T RANSACTIONS ON
C OMMUNICATIONS and the IEEE W IRELESS C OMMUNICATIONS L ETTERS .
He was Guest Editor for the IEEE Communications Magazine (Special
Issue on Heterogeneous and Small Cell Networks) in 2013. He received the
Singapore Government Scholarship in 1993, Tokyu Foundation Fellowship
in 1998, and the A∗ STAR National Science Scholarship in 2002. He was
honored with the 2008 Philip Yeo Prize for Outstanding Achievement in
Research, the IEEE Globecom 2010 Best Paper Award, the 2011 JSPS Invited
Fellow for Research in Japan, the CAS Fellowship for Young International
Scientists in 2011, the 2012 IEEE William R. Bennett Prize, and the IEEE
SPAWC 2013 Best Student Paper Award.

You might also like