1

Learning Based Uplink Interference Management in

4G LTE Cellular Systems

arXiv:1309.2543v1 [cs.NI] 10 Sep 2013

Supratim Deb, Pantelis Monogioudis

AbstractLTEs uplink (UL) efficiency critically depends on

how the interference across different cells is controlled. The
unique characteristics of LTEs modulation and UL resource
assignment poses considerable challenges in achieving this goal
because most LTE deployments have 1:1 frequency re-use, and
the uplink interference can vary considerably across successive
time slots. In this work, we propose LeAP, a measurement datadriven machine learning paradigm for power control to manage
uplink interference in LTE. The data driven approach has the
inherent advantage that the solution adapts based on network
traffic, propagation and network topology, that is increasingly
heterogeneous with multiple cell-overlays. LeAP system design
consists of the following components: (i) design of user equipment
(UE) measurement statistics that are succinct, yet expressive
enough to capture the network dynamics, and (ii) design of two
learning based algorithms that use the reported measurements to
set the power control parameters and optimize the network performance. LeAP is standards compliant and can be implemented
in centralized SON (self organized networking) server resource
(cloud). We perform extensive evaluations using radio network
plans from real LTE network operational in a major metro area
in United States. Our results show that, compared to existing
approaches, LeAP provides 4.9 gain in the 20th % tile of user
data rate, 3.25 gain in median data rate.

I. I NTRODUCTION
LTE uplink (UL) consists of a single carrier frequency
division multiple access (SC-FDMA) technique that orthogonalizes different users transmissions in the same cell, by
explicit assignments of groups of DFT-precoded orthogonal
subcarriers. This is fundamentally different from 3G, where
users interfered with each other over the carrier bandwidth and
advanced receivers, such as successive interference cancellers,
were employed to suppress same cell interference. Same cell
interference is mitigated by design in LTE, however, other
cell interference has a very different structure compared to
3G. This calls for new power control techniques for setting
user transmit power and managing uplink interference. In this
paper, we design and evaluate LeAP, a new system that uses
measurement data to set power control parameters for optimal
uplink interference management in LTE.
A. Uplink interference in 4G systems: Distinctive Properties
Uplink interference in cellular systems is managed through
careful power control which has been a topic of extensive
research for more than two decades (see Section II). However,
uplink interference in LTE networks needs to be managed over
multiple narrow bands (each corresponding to collection of a
few sub-carriers) over the entire bandwidth, thus giving rise to
unique research challenges.
To understand this better, we start by making two observations. Firstly, the uplink data rate of a user in LTE depends
The authors are with Wireless Chief Technology Office, Alcatel-Lucent
USA. e-mail: first name.last [email protected]

An illustration of three adjacent LTE cells. Depending

on MAC time-frequency assignment to users in Cell-2, 3, uplink
transmission of User-A in Cell-1 could receive high interference from
Cell-2 and low interference from Cell-3 or vice versa.
Fig. 1.

on the SINR over the resource blocks (RBs) 1 assigned to

the transmission. Secondly, LTE uses 1:1 frequency re-use
and thus an RB assigned to a user in a cell can be used by
another user in any of the neighboring cells too. The following
example illustrates how uplink interference is impacted due to
the above observations.
Example: Consider the system in Figure 1 with three
cells. Suppose the MAC of Cell-1 assigns, to an associated User-A, RBs corresponding to time-frequency tuples
(ts1 , f b1 ), (ts2 , f b2 ) over consecutive time slots ts1 , ts2 respectively. Since the MAC of Cell-2 and Cell-3 operate
independently and use the same carrier frequency due to
1:1 re-use, the MAC assignments of time-frequency tuples
(tsi , f bj ), i, j = 1, 2 at Cell-2 and Cell-3 could potentially
be the following:
Cell-2: (ts1 , f b1 ) User-C and (ts2 , f b2 ) User-B
Cell-3: (ts1 , f b1 ) User-E and (ts2 , f b2 ) User-D
As shown in the figure, since User-C and User-D are close
to the edge of Cell-1, in ts1 , User-As transmission receives
high interference from Cell-2 and low interference from Cell3, whereas, in ts2 , User-As transmission receives high interference from Cell-3 and low interference from Cell-2. In any
arbitrary time-frequency resource, interfering signal to UserAs uplink transmission at Cell-1, denoted by IntA (LT E), can
be expressed as
IntA (LT E) = PB 1 B + PC 1 C + PD1 D + PE 1 E ,
where, PU , U {B, C, D, E} denotes the received power
at Cell-1s base station due to potential uplink transmission
1 An RB is a block of 12 sub-carriers and 7 OFDM symbols and is the smallest
allocatable resource in the frequency-time domain.

of User-U in the same frequency block as User-A, and,

1 U , U {B, C, D, E} is an 0 1 indicator variable denoting
whether User-U also transmits over the same time-frequency
block as User-A. Note that, since User-B and User-C share the
same cell, only one of them can be active in a time-frequency
resource and thus 1 B +11C 1; similarly 1 D +11E 1. Since
MAC of each neighboring cell makes independent scheduling
decision on who gets scheduled in a time-frequency block, the
interference becomes highly unpredictable from transmission
to transmission; managing this unstable interference pattern
poses unique research challenges barely addressed in the
literature. Indeed, this is unlike 3G systems 2 , where the
neighboring cell interference for a similar topology with
CDMA technology (over an appropriate CDMA channel)
would simply be
IntA (CDM A) = PB + PC + PD + PE
thus leading to a more stable interference pattern across
transmissions. In general, unlike LTE, the overall CDMA
uplink interference also has an additional term for self cell
interference due to uplink users in the same cell. Of course,
for desirable user performance, the uplink interference still
has to be managed through power control algorithms that has
been the focus of much of the existing research on 3G power
control and uplink interference management.
Solution requirements: LTE networks are deployed with
self organized networking (SON) capabilities [1], [2] to maximize network performances. Todays LTE networks are also
heterogeneous (HTNs) that include high power macro cells
overlaying low power small (pico/femto) cells. Small cells
are deployed in traffic hotspots or coverage challenged areas, and thus, typically small cells outnumber macros by an
order of magnitude. This leads to a much larger, and hence
more difficult to tune and manage, cellular network where
centralized SON servers are deployed to continuously optimize
the network [1]. Thus, a good solution to uplink interference
management should satisfy the following requirements: (i) it
should be adaptive to network traffic, propagation geometry
and topology, (ii) it should scale with the size of the network
which could consist of tens of thousands of cells, and (iii)
it should be architecturally compliant in the sense that it is
implementable in a SON server and adherent to standards.
Note that, SON implementability dictates that the solution
makes use of the large amount of network measurement
data. In this paper, we design a solution that satisfy these
requirements and achieves high network performance.
B. Our Contributions
In this paper, we propose LeAP, a learning based adaptive
power control for uplink interference management in LTE
systems. We make the following contributions:
1. New framework for measurement data driven uplink
interference management: We propose a measurement data
driven framework for setting power control parameters
for optimal uplink interference management in an LTE
network. Our framework (i) models the unique interference
pattern in OFDMA based LTE systems, (ii) accounts
for varying traffic load and diverse propagation map in
2 In CDMA systems interference is summed over several simultaneous user transmissions over the entire carrier bandwidth, thus leading to more stable interference pattern.

different cells, (iii) and is implementable in a centralized

SON architecture. See Section IV-V.
2. Design of Measurement Statistics: We derive suitable
measurement statistics based on the processing of raw
data from UE measurement reports. Our measurement
statistics are succinct yet expressive enough to optimize
the uplink performance by accounting for LTEs unique
uplink interference patterns along with network state and
parameters. See Section V for details.
3. Design of learning based algorithms: Using the measurements derived in Section V, we propose two learning
based algorithms for optimal setting of cell-specific power
control parameters in LTE networks. The two algorithms
trade-off complexity and performance: one provably converges to the optimal, and the other is a fast heuristic
that can be implemented using off-the-shelf solvers. See
Section VI-VII.
4. Extensive evaluation of LeAP benefits: We evaluate
our design using a radio network plan from a real LTE
network deployed in a major US metro. We demonstrate
the substantial gains for the evaluated network: the edge
data rate (20th % tile of data rate) improves to 4.9
whereas the median gain improves to 3.25 compared to
existing approaches. The details are in Section VIII.
II. R ELATED W ORK
Cellular Power Control: Uplink power control in cellular
systems has been an active research area for around three
decades. The pioneering works in [38], [19] developed principles and iterative algorithms to achieve target SINR when
multiple users simultaneously transmit over a shared carrier.
This model is applicable to CDMA (3G) systems uplinks.
Since then, several authors have developed algorithms to
jointly optimize rate and transmit power in similar multiuser systems [32], [12], [23], [37]. In particular, [32], [23],
[37] consider utility based framework for joint optimization
of rate and transmit power. Uplink power control optimization
was made tractable in CDMA setting in [10], [35] where
log-convexity of feasible SINR region was shown. We refer
the reader to [13] for an excellent survey of the vast body
of research in power control. However, the uplink model in
LTE is fundamentally different from these systems for two
reasons. First, unlike CDMA, uplink interference in LTE is
only from neighboring cells and the interference over an
assigned frequency block could change in every transmission,
leading to a far more variable interference pattern. Secondly,
to control neighboring cell interference, the standards have
mandated cell-specific power control parameters that in turn
govern UE SINR-targets.
Fractional Power Control (FPC) in LTE: LTE power control
is FPC based which has led to some recent work [6], [18], [14],
[17]. However, unlike our work, none of previous publications
develop a framework and associated algorithms to optimize the
FPC parameters based on user path loss statistics and traffic
load. Recognizing the difficulties of setting FPC parameters in
LTE, [39] develops and evaluates closed loop power control
algorithms for dynamically adjusting SINR targets so as to
achieve a fixed or given interference target at every cell. This
work has two drawbacks: first, it is unclear how the interference targets could be dynamically set based on network and
traffic dynamics, and second, the scheme does not maximize
any network-wide SON objective.

Self Organized Networking (SON) in LTE: Study of LTE

SON algorithms have gained some attention recently mostly
for downlink related problems. [31], [28] study the problem of
downlink inter-cell interference coordination (ICIC) for LTE,
[34] optimizes downlink transmit power profiles in different
frequency carriers, and [27], [36], [24], [33] study various
forms (downlink, uplink, mobility based etc.) of traffic-load
balancing with LTE SON. To our best knowledge, ours is
the first work to develop measurement data-driven SON algorithms for uplink power control in LTE. For a very extensive
collection of material and presentations on the latest industry
developments in LTE SON, we refer the reader to [1], [2].
III. A P RIMER ON LTE U PLINK AND F RACTIONAL P OWER
C ONTROL
Terminologies used in the paper: A resource block (henceforth RB) is a block of 12 sub-carriers and 7 OFDM symbols
and is the smallest allocatable resource in the frequency-time
domain. eNodeB (eNB in short) refers to an LTE base station
and it hosts critical protocol layers like PHY, MAC, and Radio
Link Control etc. User equipment (henceforth UE) refers to
mobile terminal or user end device, and we also use UL for
uplink. Finally, reference signal received power (RSRP) is
the average received power of all downlink reference signals
across the entire bandwidth as measured by a UE. RSRP is a
measure of downlink signal strength at UE.
A. UL Transmission in LTE
LTE uplink uses SC-FDMA multiple access scheme. SCFDMA reduces mobiles peak-to-average power ratio by performing an M -point DFT precoding of an otherwise OFDMA
transmission. M depends on the number of RBs assigned
to the UE. Also, each RB assigned to an UE is mapped
to adjacent sub-carriers through suitable frequency hopping
mechanisms [16]. At the base station receiver, to mitigate frequency selective fading, the per antenna signals are combined
in a frequency domain MMSE combiner/equalizer before
the M -point Inverse DFT (IDFT) and decoding stages are
performed. The details of the above steps are not relevant for
our purpose; instead, we note two key properties that will be
useful from an interference management perspective.
P1: The equalization of the received signal is performed
separately for each RB leading to SINR being averaged
only across sub-carriers of a RB. Thus, SINR in an RB
is a direct measure of UE performance3 .
P2: Consider the adjacent subcarriers S assigned to an RB
for a UE in a Cell-1. In another neighboring Cell-2, the
same subcarriers S can be assigned to at most 1 UEs RB
because RB boundaries are aligned across subcarriers.
Put simply, at a time there is at most 1 interfering UE
per neighboring cell per RB.
B. LTE Power Control and Challenges
To mitigate uplink interference from other cells and yet
provide the UEs with flexibility to make use of good channel
conditions, LTE standards have proposed that power control
should happen at two time-scales as follows:
1. Fractional Power Control (FPC-) at slow time-scale:
At the slower time-scale (order of minutes), each cell sets
3 The final UE data rate also depends on the selected modulation and coding
for the specific scheduled HARQ process.

cell-specific parameters that the associated UEs use to set

their average transmit power and a target-SINR as a predefined function of local path loss measurements. The two
cell-specific parameters are nominal UE transmit power
P (0) and fractional path loss compensation factor < 1.
A UE-u with an average path loss PL 4 to its serving cell
transmits at a power spectral density (PSD),
Tx
P(u)
= P (0) (PL) .

(1)

Tx
where PSD P(u)
is expressed in Watt/RB. Thus, if a UE
is assigned M RBs in the uplink scheduling grant then its
Tx
max
max
total transmit power is min(M P(u)
, Ptot
) where Ptot
is
a cap on the total transmission power of a UE. We remark
that, can be interpreted as a fairness parameter that leads
to higher SINR for UEs closer to eNB.
2. At the faster time-scale, each UE is closed loop power
controlled around a mean transmit power PSD (1) to
ensure that a suitable average SINR-target is achieved.
This closed loop power control involves sending explicit
power control adjustments, via the UL grants transmitted
in Physical Downlink Control Channel (PDCCH), that can
be either absolute or relative.
Challenges: LTE standards leave unspecified how each cell
sets the value of P (0) , and average cell-specific mean
interference targets that are useful for setting UE SINR targets
(see Remark 5, Section V). Since the choice of parameters
results in a suitable mean interference level at every cell
above the noise floor, this problem of setting the cell power
control parameters is referred to as IoT control problem.
Clearly, aggressive (conservative) parameter setting in a cell
will improve (degrade) the performance in that cell but will
cause high (low) interference in neighboring cells. Since FPC scheme and its parameters are cell-specific, its configuration
lies within the scope of self-organized (SON) framework, i.e.,
the the solution should adapt the parameters based on suitable
periodic network measurements.

Remark 1 (Extensions). The main idea behind FPC- is to

have cell dependent power control parameters that change
slowly over time. Our techniques apply to any scheme where
UE transmit power is a function of cell-specific parameters
and local UE measurements (path loss, downlink SINR etc.).
IV. S YSTEM M ODEL AND C OMPUTATION A RCHITECTURE
LTE HetNet (HTN) Model: Our system model consists
of a network of heterogeneous cells all of which share the
same carrier frequency , i.e., there is 1:1 frequency re-use in
the network. Some cells are high transmit power (typically
40 W) macro-cells while others are low transmit power picocells (typically 1-5 W). In HTNs, pico cells5 are deployed
by operators in traffic hotspot locations or in locations with
poor macro-coverage. The distinction between macro and pico
is not relevant for the development of our techniques and
algorithms but it is very important from an evaluation of our
design. See Section VIII for further discussions and insights.
C denotes the set of all cells that includes macro and pico
cells. We will introduce parameters and variables as needed;
Table I lists the important ones.
4 Received Power at eNB = UE Transmit Power
PL

5 Picocells as opposed to femto cells have open subscriber group policies.

TABLE I
L IST OF PARAMETERS AND VARIABLES USED
Notation
C, c
(c C)
(0)
Pc , c
c
Uc
Jc
lc(u)
Lc
(u)
lec
Lec

c , ec
lc (b), pc (b)
c (b)

Fig. 2.

Description
Set of cells and index for a
typical cell, respectively
FPC-parameters of cell-c: nominal transmit power and
path loss compensation factor respectively.
ln(Pc(0) ), .i.e., nominal transmit power in log-scale
Set of UEs associated with cell-c
Set of cells that interfere with cell-c
Mean path loss of UE-u to eNB of its serving cell-c
Random variable for path loss of a
random UE u Ue to its serving cell-c
Mean path loss of UE-u Ue to cell-cs eNB
Random variable for path loss of a random
UE-u Ue to cell-cs eNB
c = ln Lc , ec = ln Lec
For path loss histogram bin-b of cell-c,
bin value and bin probability respectively
Expected SINR (in ln-scale) of a cell-c UE with path loss lc (b).

Underlying computation architecture.

Interfering Cell: For a typical cell indexed by c C, the

set of other cells that interfere with cell-c are denoted by Jc .
Ideally, a cell e Jc , if the uplink transmission of some UE
associated with cell-e is received at eNB of cell-c with received
power above the noise floor. Since interfering cells must be
defined based on the available measurements, we say e Jc
if some UE-u associated to cell-e reports downlink RSRP (i.e.
RSRP is above a threshold of 140 dBm) from cell-c, which is an
indication that UE-us uplink signal could interfere with eNB
of cell-c. This is reasonable assuming uplink and downlink
path-loss symmetry between UEs and eNB.
Underlying Self-Optimized Networking (SON) architecture: The underlying SON architecture is shown in Figure 2. In this architecture, network monitoring happens in
a distributed manner across the radio access network, and
the heavy duty algorithmic computations happen centrally at
the Network Management Servers. As opposed to a fully
distributed (across eNBs) computation approach, this kind of
hybrid SON architecture is the preferred by many operators for
complex SON use-cases [1], [2] for two key reasons: capability
to work across base stations from different vendors as is
typically the case, and not having to deal with convergence
issues of distributed schemes (due to asynchrony and message
latency). We note two relevant aspects of this architecture:
Main building blocks: The main building blocks are the
following: (i) monitoring components at the cells that
collect network measurements, appropriately process them
to create Key Performance Indicators (KPIs), and communicate the measurement and KPIs to the central server, and
(ii) the algorithmic computation engine (cloud servers) at
a central server that makes use of the KPIs and compute
the SON parameters that are then fed back to the network.
Time-scale of computation: The KPIs from the network
are typically communicated periodically. The period should

be such that, it is short enough to capture the changing

network dynamics that call for network re-configuration
and is long enough for accurate estimation the relevant
statistics. In general, most networks have periodic measurement reports with frequency 515 mins [30] but much
faster minimization of drive test (MDT) data and UE trace
information can be collected. In addition, measurement
reports can be trigerred by events like traffic load above
a certain threshold beyond normal.
Design questions: Given this architecture, to periodically
compute the FPC- parameters to maximize the network
performance in the uplink, we need to answer two questions:
Q1: What are the minimal set of network measurements
required to configure FPC- parameters?
Q2: Based on these measurements, how should we choose
, P (0) and average interference threshold for every cell ?
The algorithms should be scalable and capable of updating
the parameters as new measurement data arrives.
In Section V, we answer Q1 and propose a measurement
based optimization framework , and in Section VI-VII , we
develop learning based algorithms to answer Q2.
V. M EASUREMENT AND O PTIMIZATION F RAMEWORK
A. Network Measurement Data
Additional notations: We will derive an expression of
SINR of a typical uplink UE. Towards this goal, consider a
network snapshot with a collection of UEs U; a generic UE
is indexed by u U. Let Uc denote the set of UEs associated
(u)
with cell-c. Denote by lc path loss from u Uc to its serving
cell c. We also drop the superscript for UE and write Lc to
denote the path loss of a random UE associated with cell-c.
We denote the mean path loss from a UE u Ue to an cell
(u)
c Je by lec . We also drop the superscript for UE and
write Lec to denote the path loss from cell-e to cell-c of
a random UE belonging to cell-e. These notations are also
shown in Table I.
Remark 2 (Fast-fading and frequency selectivity). All path
loss variables must be interpreted in time-average sense
so that the effect of fast-fading is averaged out. Indeed, fastfading happens at a much faster time-scale (ms) than the
FPC- parameter computation time-scale of minutes. Also,
SC-FDMA equalization at the receiver averages out the effect
of frequency selectivity in an RB.
Expression for SINR: Consider a UE-u that transmits to
its serving cell-c over the RBs assigned to it. Consider the
transmission over any one RB that is assigned subcarriers
in the set S by the MAC scheduler. To derive the SINR
over S, we wish to quantify the interference experienced
by the received signal at serving cell-c over S. Due to our
observations on frequency hopping, the subcarriers within S
can be used in a neighboring cell by at most one UE. Note that,
depending on the load on a cell, a RB may only be utilized
only during a certain fraction of the transmission subframes.
With this motivation, we define the following binary random
variable:

if cell-e schedules a UE over the

4
1
resource block S also used at cell-c
Oec =
(2)
0
else
In the above, we assume that Oec does not depend on the
specific choice of S and has identical distribution for all S.

Denoting by v the UE that occupies S at cell-e, ignoring fastfading, the interfering signal at serving cell-c of UE-u for
transmissions over S in cell-e can be expressed as follows.
(v) 1
Tx
(v)
)
Iec = Oec P(v)
. lec
= Oec Pe(0) (le(v) )e (lec

It is instructive to note that, even without fluctuations due

to fast fading, the interference Iec is random because the
interfering UE-v transmitting over S in cell-e could be any
(v) (v)
random UE in cell-c. In other words, the quantities le , lec
can be viewed as a random sample from the joint distribution
of (Le , Lec ). Thus the total interference at cell-c over
frequency block S is a random variable Ic given by
X
Ic =
Oec Pe(0) (Le )e (Lec )1
(3)
eJ

where we drop the superscript of the loss variables to mean

that the loss is from a random UE in an interfering cell using
the RB-S. This randomness is induced by MAC scheduling of
the interfering cell. Now, the SINR of UE-u in cell-c denoted
(u)
(u)
by SINR(lc ) (as a function of UE-us path loss lc ) over
RB-S can be expressed as
SINR(lc(u) ) =
=P

Rx
P(u)

I c + N0
(0) (u)
Pc (lc )(1c )
(0)

eJc

Oec Pe (Le ) (Lec )1 + N0

(4)

Measurement variables: It follows from (4) that, for any

(u)
(u)
given lc , SINR(lc ) is a random variable that is fully
characterized by the following distributions: joint distribution
of (Le , Lec ) and Oec for all e Jc . Furthermore, as
we argue formally in Section V-B, towards computing an
average network wide performance metric over the SON
computation period, we also require c the mean uplink load
in cell-c and path loss distribution Lc . The dependence on
traffic load on the average network performance is also quite
intuitive. We thus have the following.
Design of Measurement Statistics: For a given set of
(0)
values of Pc , c , under FPC- mechanism, any expected
network wide performance metric can be fully characterized by the following statistics:
1. Joint path loss distribution (Le , Lec ) of uplink UEs
for every interfering cell e of c.
2. Pr(Oec = 1) which is the probability that cell-e
schedules a UE interfering with cell-c for transmission
over an RB.
3. Mean number of active uplink UEs (uplink load) in
every cell-c denoted by c .
4. Path loss distribution of uplink UEs of cell-c denoted
by Lc .
Histogram construction: The required distributions can be
estimated using standard histogram inference techniques [9]
using the following steps we state for completeness:
1. Collecting UE measurement samples: Since UE measurement reports6 contain reference signal strength (termed RSRP)
6 The

measurements are either available through what is called per call

measurement data or it can be seeked by eNB[3].

from multiple neighboring cells, the RSRP values can be

coverted into path losses using knowledge of cell transmit
power. Thus, for a UE associated with cell-c, if a measurement
report contains RSRP from cell-c and cell-e both, then it
provides samples for Lc , and (Le , Lec ).
2. Binning the path loss data: This is a standard step where the
range of Lc and (Le , Lec ) are divided into several disjoint
histogram bins and each data point is binned appropriately.
3. Estimating the occupancy probabilities Pr(Oec = 1):
Assuming proportional-fair MAC scheduling7 where all UEs
in a cell use the radio resources uniformly, Pr(Oec = 1) can
be estimated as the fraction of UEs in cell-e that interfere with
cell-c. This estimation can be performed by simply computing
the fraction of UEs in cell-e that report measurements from
cell-c.
We make two observations. First, the histograms are best
maintained in dB scale to so that the range of path loss
values is not too large. Second, in practice, the data samples
for constructing the histograms can come for a large enough
random subset of all the data.
B. The IoT Control Problem in LTE HTNs
We are now formulate our problem based on the measurement histograms described in the previous section.
Network wide performance metric: Given our model,
a suitable performance metric should satisfy two properties:
(i) it should account for the randomness in the SINRs in a
meaningful manner, and (ii) it should strike a balance between
aggregate cell throughputs and fairness.
We wish to propose a average performance metric where
average is over all UE path losses that can realize the measurement data. Towards this end, we first define a performance
metric for a typical UE-u in cell-c who has path loss to
(u)
(u)
its serving cell given by lc . Denote by c (lc ) as the
(u)
expected log-SINR of a typical UE-u in cell-c, i.e., c (lc ) =
(u)
(u)
E[ln SINR(lc )] where the expression of SINR(lc ) is given
by (4). We now define the UE utility function V (.) as follows:


Utility of a UE-u in cell-c as a
4
V (c (lc(u) )) =
(u)
(u)
function of c (lc ) = E[ln SINR(lc )]

(5)
Here V (.) is a concave increasing function and we have
(u)
explicitly shown the SINR depends on the path loss lc .
Choosing the utility as a function of E[ln SINRu ] has two
benefits. Firstly, ln(SINRu ) is a measure of the data rate with
SINRu . Secondly, converting the SINR into a log-scale makes
the problem tractable since it is well known that feasible power
region is log-convex [35]8 , and thus, we can use elements of
convex optimization theory.
Having defined a utility for a typical UE, we are now
in a position to define network wide performance metric
obtained by averaging over all UE path losses that can realize
the measurement data. We need additional notations for the
histogram of Lc s. Suppose the path loss histogram of Lc
is divided into k disjoint intervals with mid-point of the
intervals given by lc (1), lc (2), . . . , lc (kc ). Suppose in cellc, the empirical probability9 of items in histogram bin-b is
7 This is the most prevalent MAC scheduling in LTE.
8 In simple terms, if all powers are represented in log-scale then the feasible power

vectors of all UEs is a convex region.

9 The empirical probability of histogram bin-b is simply the number of data items
binned into the bin-b divided by the total number binned items for this histogram.

given by pc (b). Then, we define the overall system utility

as the expected utility of all UEs in all the cells where the
expectation is over empirical path loss distribution given by
the measurement data. This utility, as a function of vector of
(0)
Pc , c s (denoted by P(0) , ) is given by.
Util(P(0) , ) =

"

E[Total utility of UEs in cell-c]

kc
XX

kc
XX

E[Total utility of UE with path loss lc (b) in cell-c]

E[Num. UEs with path loss lc (b)] V (c (lc (b)))

cC b=1

kc
XX

Oec Pe(0) (Le ) (Lec )1

+ N0

(8)

eJc

cC b=1

!#
X

c = E ln

cC

random variables. We will first formulate this problem as a

mathematical program so that the problem becomes tractable.
By using motivation from [35]10 , we convert all powers, and
path losses to logarithmic scale as follows.

c pc (b) V (c (lc (b)))

(6)

cC b=1

where c (lc (b)) is the expected SINR in log-scale for any

UE with path loss lc (b) to its serving eNB of cell-c. The last
step follows because the expected number of UEs with path
loss lc (b) is product of expected number of UEs in cell-c and
the probability of a UE with path loss lc (b).
Choice of UE-utility: Though our techniques work for
a generic concave and increasing V (.), for our design and
evaluation, we use the following form of V (.) in the rest of
paper:
V () = ln ln(1 + exp()) .
(7)
One can verify that the function V (.) defined above is
concave. Roughly speaking, ln(1 + exp()) is the Shannon
data rate corresponding to log-scale (natural log) SINR of
. It is well known that utility defined by log of data rate
strikes the right balance between fairness and overall system
performance.
Remark 3 (Data rate and throughput). In this paper, we
use UE data rate or PHY data rate (in bits/sec/Hz) instead of UE throughputs (bits/sec/Hz). Under proportionalfair MAC scheduling, the dominant scheduling policy in LTE,
UE throughput Ru of UE-u in cell-c, is roughly related to
PHY data rate ru of UE-u by Ru = ru /Nc ; here Nc is
the average number of simultaneously active users in cellc. Thus, ln Ru = ln ru ln Nc . In other words, the log of
throughput and log of data rate are off by an additive constant
independent of power control parameters.
We also observe that optimizing the total log-data rate of
all UEs is aligned with the MAC objective of proportional-fair
scheduling in LTE.
Measurement based IoT Control Problem (IoTC): We
now formulate the problem of optimally configuring the parameters of the FPC-. This problem is commonly referred
to as the Interference over Thermal (IoT) control problem
(0)
because the optimal choice of Pc , c s result in a suitable
interference threshold above the noise floor in every cell. The
problem can be succinctly stated as follows:
Given the path loss histograms of Lc , (Le , Lec ),
empirical distribution Oec , and average traffic
(0)
load c at every cell-c, find optimal Pc , c for
every cell-c such that we maximize Util(P(0) , )
given by (6).
At a first glance, the above problem may look complex
due to the fact that Util(P(0) , ) is a function of several

c = ln Pc(0) , c (b) = ln lc (b) , ec (b) = ln lec (b) ,

c = ln Lc , ec = ln Lec

(9)
(10)

The notations are also shown in Table I. We can now rewrite c (lc (b)) (given by (4)) in terms of these above variables
as
c (lc (b)) = c (1 c )c (b) c
where,
"
!#
X
c = E ln
Oec e(e +e e ec ) + N0

(11)

(12)

eJc

which has the standard look of SINR expression in log-scale.

Thus, we can state the problem of maximizing Util(P(0) , )
as the problem of maximizing
XX
c pc (b)V (c (lc (b)))
c

subject to equality constraints given by (11) and (12). It

turns out that the equality constraints can be re-written as
inequalities to convert this into convex non-linear program
(NLP) with inequality constraints. This leads to the following
proposition where we have also imposed an upper bound on
the maximum transmit power per RB and maximum average
interference cap.
Proposition
V.1. Given measurement statistics of
Lc , (Lc , Lec ), Oc and average traffic load c at every
cell-c, the problem of maximizing Util(P(0) , ) given by (6),
subject to maximum transmit power constraint on RB, is
equivalent to the following convex program:
IoTC-SCP:
max

{c },{c },{c (b)}

c pc (b)V (c (b))

c,b

subject to,
c C, b [1, kc ] : c (b) c (1 c )c (b) c
(13)
"
!#
X
c C : c E ln
Oec e(e +e e ec ) + N0
eJc

c C, b [1, kc ] : c + c c (b) ln Pmax

c [0, 1], c (b) [min , ), c [ln N0 , ln Imax ]

(14)
(15)
(16)

Proof: (Outline) First, at optimality, all of the inequalities

must be met with equality from basic principles. Secondly, the
constraint set is a convex region since the constraints (14) is
convex due to convexity of LSE function11 [21].
10 [35] shows that feasible power region is log-convex in many wireless networks. This
means that, if powers are in log-domain, then the problem has convex feasible region
which makes it amenable to convex optimization theory.
11 LSE or log-sum-exponential functions are functions of the form f ({x } ) =
i i
P P
ln( i e xij aj ).

We make two important remarks.

Remark 4 (IoTC-SCP constraints). In IoTC-SCP, constraint (13) states that the expected SINR in log-scale can be
no more than the receiver UE power less the expected uplink
interference at the eNB, constraint (14) states that the expected
interference at each cell has to be at least the expected sum of
interferences from the neighboring cells, and constraint (15)
limits the maximum power allowed per RB. Finally, (16) states
the valid domain of the variables c , c s and u s, where,
min is the minimum decodable SINR. Note that the valid
domain of c is [ln N0 , ln Imax ] because maximum average
interference in cells could need a cap due to hardware design
constraints.
Remark 5 (Setting SINR target of UEs). In IoT-SCP, suppose
c , c , c represent the optimal values for cell-c and let
(0)
Pc
= exp(c ), Ic = exp(c ) be the optimal values in
linear scale. Then, the SINR target of an arbitrary UE in cell-c
(u)
with average path loss lc is set as
SINRtarget (lc(u) ) =

h
i

(u)
(0)
(u)
min Pmax (lc )1 , Pc
(lc )(1c )
Ic

.
(17)

Challenges in solving IoTC-SCP: IoTC-SCP is somewhat

different from a traditional non-linear program that are solved
using standard gradient based approaches (using Lagrangian).
The main challenge in solving IoTC-SCP comes from the
randomness in the interferers. As a consequence of this,
the expected value of r.h.s. of interference constraint (14)
is computationally-intensive even for given values of the
K
)
c s and c s. Indeed, this requires accounting for O(Bmax
number of interferer combination per cell where Bmax is the
maximum number of histogram-bins of the joint distribution
of (Le , Lec ) in any cell and K is the maximum number
of interferes of an cell. Repeating this computation for every
iteration of a gradient based algorithm is next to impossible.
By adapting elements of learning theory, we propose two
algorithms that strike a different performance-computation
trade-off. One is optimal requiring more computational resources and the other is an approximate heuristic with fast
computation times.
VI. A S TOCHASTIC L EARNING BASED A LGORITHM
We now develop a stochastic-learning based gradient algorithm.
We introduce additional notations for ease of exposition.
Denote the original optimization variables (c (b), c , c )s by
the vector z (also called primal variables). We also denote
the constraint set of the problem IoTC-SCP by the random
function h(z, O, ), i.e.
h(z, O, )

[ (b) c (1 c )c (b)]cC,b[1,kc ]
 Pc



(e +e e ec )
+ N0 c cC
= ln
eJc Oec e
[c + c c (b) ln Pmax ]cC,b[1,kc ]
(18)
Thus, the IoTC-SCP problem can be simply stated as
X
max
V (c (b)) s.t. E[h(z, O, )] 0 .
z

c,b

In the above, the expectation is with respect to the random

variables Oec s and (e , ec )s. We define the random
Lagrangian function as follows:
X
L(z, h(z, O, ), p) =
V (u ) pt h(z, O, )
(19)
u

where, p 0 denotes the so called Lagrange multiplier vector

and it has the same dimension as the number of constraints.
Since the objective function of IoTC-SCP is concave and the
constraint set is convex, one can readily show from convex
optimization theory that [8]
OPT = max min E[L(z, h(z, O, ), p)]
z

p0

(20)

= min max E[L(z, h(z, O, ), p)] ,

p0

where OPT denotes the optimal value of IoTC-SCP. Our goal

is to solve the IoTC-SCP problem by tackling the saddle point
problem (20).
Useful bounds on the optimization variables: Before we
describe our algorithm, we derive trivial but useful bounds on
feasible primal variables. The iterative scheme to be described
shortly projects the updates primal variables within these
bounds. The proof is straightforward and skipped for want
of space.
Lemma VI.1. Any feasible c (b) and c of the problem IoTCSCP satisfies
min c max , min c (b) max ,
where,
max = ln Pmax , min = min + N0 , max = ln Pmax N0 .

Intuition and a stochastic iterative algorithm: The equivalent problem (20) is referred to as the saddle point problem
and one could use standard primal-dual techniques to solve
this problem [29]. In such an approach, the primal variables
and dual variables are updated alternatively in an iterative
fashion, and each update uses a gradient (of E[L(.)]) ascent
for the primal variables and gradient descent for the dual
variables. However, computing the expectation and its gradient
K
is computationally expensive (can be O(Bmax
) where Bmax
is maximum histogram bins and K is maximum number of
interferers) and this has to be done in every iteration which
is practically impossible. Instead, we perform primal-dual
iterations by substituting the random variables with random
samples of the random variables in every iteration: thus
expectation computation is replaced by a procedure to draw a
random sample from the histograms which is a computationally light procedure. The idea of replacing a random variable
by a sample in an iterative scheme is not new, the entire field
of stochastic approximation theory deals with such schemes
and conditions for this to work [11]. Our contribution is in
applying this powerful technique to the IoTC-SCP problem.
By using elements of stochastic approximation theory, we also
show that the resulting output converges to optimal.
We now describe our algorithm. Denote the value of the
primal variables in iteration-n by zn and the value of the dual
n and p
n
variables in iteration-n by pn . We also denote by z
the iteration average of the primal and dual variables respectively. The overall algorithm is summarized in Algorithm 1
and is described as follows:

n , p
n are initialized to positive
Initialization: First zn , pn , z
values within bounds specified by Lemma VI.1.
Iterative steps: Denote by an = 1/n , 0.5 < 1 as the
step size in iteration-n. The following steps are repeated for
n = 0, 1, 2, . . .:
1. Random Sampling of Interferers: This step is performed
for each interfering cell pair (c, e). For each cell-e Jc
the following steps are performed:
Toss a coin with probability of head Pr(Oec = 1).
Denote the outcome of this coin-toss by the 0 1
variable ec where ec takes value one if there is a
head.
If ec = 1, then draw a random sample from the joint
distribution of (Le , Lec ) based on the histogram of
this joint distribution. Denote the random sample, which
is a 2-tuple by (sa[1], sa[2]). Let (ec [1], ec [2]) denote
the logarithm of this random sample, i.e., ec [i] =
ln(sa[i]), i = 1, 2 .
In the function h(.) given by (18) representing the
constraints of IoTC-SCP, replace the random variables
(e , ec ) by the random sample (e c[1], ec [2]).
Denote by , as the vector of ec , ec s samples. We
write h(zn , , ), pn ) as a function of these random samples along with the primal and dual variables in iteration-n.
2. Primal Update: The primal variables are updated as
zn+1 = zn + an z L(zn , h(zn , , ), pn )

(21)

where z L(.) denotes the partial derivative of L(.) with

respect to z.
Next project updated values of e , e , e , u s
respectively within intervals [0, 1], [min , max ],
[ln N0 , ln Imax ],[min , max ]. For example, if the value
of e in iteration-(n + 1), (e )n+1 > 1, then set
(e )n+1 1; and, if any (e )n+1 < 0 set (e )n+1 0.
3. Dual update: The dual variables are updated as
+

pn+1 = [pn + an h(zn , , )]

(22)

4. Updating average iterates: The current value of the

solution is given by average iterates

Algorithm 1 I OTC-SL: Stochastic Primal Dual Algorithm for

IoTC Control.
1: Initialize the primal variables to z0 and the dual variables to p0 .
2: for iterations n = 1 to n = Nit do
3:
For each interfering cell pair (e, c) draw a random path loss

4:
5:
6:
7:
8:

sample from the joint distribution histogram of (Le , Lec ) and

the replace the path loss random variables by these samples in
the expression of h(.) in (18).
Update the primal variables according to gradient ascent based
update rule in (21).
Update the dual variables according to gradient descent based
update rule in (22).
Maintain average of the primal and dual variables over all the
iterations
end for
The average of the primal variables over all the iterations is the
output.

VII. R EGRESSION BASED C ERTAINTY E QUIVALENT

H EURISTIC
For ease of exposition, we re-write the random interference
in constraint (14) of IoTC-SCP which is
!#

"
c C : c E ln

Oec e

(e +e e ec )

+ N0

eJc

(23)

The basic idea of this heuristic consists of two high-level

methods. First is the notion of certainty equivalence where
we replace the random log-interference by log of expected
interference in (23) such that any solution with this modified
constraint is a feasible solution of IoTC-SCP. Second is
the notion of regression, where, using a known parametric
distribution as model for the path loss statistics, we perform
a fitting of the distribution parameters using the measurement
data. Now the problem can be shown to be a standard nonlinear program (NLP) that can be solved using off-the-self
NLP solvers. We next explain the different steps.
Certainty equivalent approximation: We approximate the
random interference constraint (23) as follows:
c C : c ln

E Oec e

(e +e e ec )

+ N0

!
.

eJc

(24)

n+1
z

pn+1

1
n zn + (1
1
n pn + (1

1
zn
n )
1
pn
n )

The iterative steps are repeated for Niter iterations based on

Niter , p
Niter converge within a desirable precision.
whether z
There are many techniques are testing the convergence of
stochastic iterations [5] that can be readily used for our
purpose. The step-sizes an can be set as 1/n , 0.5 < 1
or in an adaptive manner to speed up the convergence [20].
Asymptotic Optimality of Algorithm IoTC-SL We will
now state the main analytical result for Algorithm IoTCSL. The following result shows that the iterates generated
by Algorithm IoTC-SCP converge to the optimal solution for
almost every sample path of the iterates.
Theorem 1. Suppose z is the optimal solution of IoTC-SCP.
Then, the following holds:
zn z

with probability 1 .

Proof: See Appendix.

We now make a very important observation. Denote by

Fapprox the feasible solution region of the optimization problem obtained by the above modification (24) of interference
constraint (14) in IoTC-SCP. If FIoT CSCP is the feasible
region of IoTC-SCP, then it can be shown from Jensens
inequality that (24) (23), and thus,
Fapprox FIoT CSCP .
This means, if we replace constraint (23) by (24), we will
produce feasible solution to IoTC-SCP.
Denoting aec = E[Oec ], (24) can be rewritten as
!
X
c ln
aec g(Pe(0) , e ) + N0 ,
(25)
eJc

where
g(e , e ) = E[e(e +e e ec ) ] .

(26)

Model for path losses: To simplify the r.h.s. of (24), we

use the following widely used path-loss model:

D ISTRIBUTION M ODEL FOR PATH LOSS : For each interfering cell pair, (c, e), the joint distribution of the random
variables (Le , Lec ) are log-normal. In other words, the distribution of (e , ec ) = (ln Le , ln Lec ) is jointly Normal.
Indeed, log-normal distribution is a very reasonable statistical model for spatial path loss distribution [22]. For ease of
exposition, we define the following two dimensional vectors:




e
e
Zec =
, e =
,
(27)
ec
1
Zec is defined for each interfering cell pair (c, e) and e
for each cell e. Note that, under our modeling assumption,
the random variable Zec is bi-variate jointly Normal random
variable. The parameters of Zec are defined by the following
mean and correlation matrix:
t

mec = E[Zec ] , Cec = E[(Zec me )(Zec me ) ]

(28)

Note that mec is a 2 1 vector and Cec is a 2 2 symmetric

positive definite matrix. Also, mec , Cec are model parameters
that can be estimated from the measurement data using standard techniques for Gaussian parameter estimation [25] (also
outlined in Algorithm 2).
Model based simplification of IoTC-SCP: Using the
definition of Zce in (27), we can rewrite g(e , e ) in (26)
as follows.
g(e , e ) = E[exp(e + et Zec )]
Since Zec is Normal, the above expression can be simplified
using moment generating functions of multivariate Normal
random variables as [25]:


(29)
g(e , e ) = exp e + et me + 12 et Cec e
where we write g(.) to mean that it is an estimate of g(.) under
the assumed statistical model of path losses. We can now solve
the following modified version of IoTC-SCP problem where
we replace the interference constraint (14) by
!
X
Ic ln
aec g(e , e ) + N0 .
eJc

This is stated as follows:

IoTC-CE:
X

max

{e },{e },{c (b)}

c pc (b)V (c (b))

c,b

subject to,
c (b) c (1 c )c (b) c
"
#
X
1 t
t
(e +e
Cec e )
mec + e
2
c ln
aec e
+ N0
eJc

c + c c (b) ln Pmax
c [0, 1], c [ln N0 , ln Imax ], c (b) [min , )

IoTC-CE is convex deterministic NLP (and hence standard

tools of solving NLP can be used), provided the function
g(e , e )) is convex in (e , e ). The following lemma shows
shows that g(e , e )) is indeed a convex function.

Algorithm 2 S OLVE I OTC-CE: Regression based deterministic heuristic for IoTC COntrol.
1: Regression step: For each interfering cell pair (e, c), we use

the path-loss measurements to find a normal fit of the Gaussian random random variable Zec given by (27). Suppose the
histogram bins based on the measurement of joint distribution
of (Le , Lec ) are indexed by b where b = 1, 2, . . .. Suppose
bin-b represents the values (le (b), lec (b)) and has empirical
probability mass pec (b). Let (e , ec ) = (ln le (b), ln lec (b)).
Then a maximum-likelihood estimator [25] of mec , Cec is as
follows:
 P



2
e2
ec
b pec (b)e (b)
mec = P p
,
C
ec =
2
ec
e2
b ec (b)ec (b)
where,
e2 =

2 =
ec

pec (b)(e (b) mec [1])2

pec (b)(e (b) mec [1])(ec (b) mec [2])

2: Deterministic non-linear program solving: With the above es-

timates, we solve the NLP given by IoTC-CE using standard

tools for solving deterministic non-linear programs (such as, dual
based techniques, primal-dual techniques etc.) [7].

P
Lemma VII.1. The function ln( eJc g(e , e ) + N0 ) is
convex in e and e . Thus IoTC-CE is a convex NLP.
Proof: Since convex and component-wise increasing
function of a vector of convex functions is convex [8], the
result follows. The details are omitted for want of space.
Overall algorithm by putting it all together: As outlined
in Algorithm 2, we have two steps. First we use the measurement data to estimate the parameter-matrices me , Cec of the
distribution Zce . Second, we use these values to solve IoTCCE using standard NLP solving techniques.
The main benefit of the approach in this section is that, it
is computationally not so intensive and can be solved using
powerful commercial NLP solvers; the price we pay is the
sub-optimality.
VIII. E VALUATION USING P ROPAGATION M AP FROM A
R EAL LTE D EPLOYMENT
The primary goal of our evaluation is two fold. First,
to understand the gains provided by LeAP as compared to
other popular approaches. Second, to understand the effect
of important design parameters like histogram bin-size and
IoT-cap on LeAP performance. In addition, we also provide a
comparison of the two proposed LeAP algorithms.
A. Evaluation Platform
Evaluation platform: For our evaluation purpose, we
implemented the following components in Figure 2: SQL
database (DB) that saves network information, the data access
layer which interacts with the DB and creates the measurement
data discussed in Section V, and the IoT-control parameter
(P (0) , s) computation engine. All our implementations were
based on .NET using C# and are multi-threaded using the Task
Parallelization Library (TPL) facilities. We implemented Algorithm 1 within our framework using suitable data structures
to enable multi-threaded implementation. We skip these implementation details for want of space. For network measurement
information saved in the DB, we used real signal propagation

10

LTE signal map in evaluated section of a major US metro.

The heat map is based on downlink SINR as the uplink SINR depends
on specific FPC scheme.
Fig. 3.

map from an actual deployment but synthetic mobile locations

generated using a commercial network planning tool in a
manner described shortly.
Note that, in our evaluation platform, we have not yet
implemented Algorithm 2. Such implementation requires integrating a .NET compatible commercial NLP solver within
our architecture, a feature that is in progress. However, just
for the purpose of comparison, we have developed a standalone implementation of Algorithm 2 using CVX [15] a free
NLP Matlab solver. This can operate only on moderate sized
networks with restricted objective function forms and is very
slow for practical needs.
B. Evaluation Setup and Methodology
We used signals propagation maps generated by a real
operational LTE network in a major US metro area12 (see
Figure 3 for the propagation heat map). The terrain category,
cell site locations, and drive-test propagation data from the
network were fed into a commercially available Radio Network Planning (RNP) tool that is used by operators for cellular
planing [4]. The carrier bandwidth is 10 MHz in the 700 MHz
LTE band. There is no dependency of LeAP design to the
specific tool used for evaluation purpose.
For our evaluation, we selected an area of around 9 km2 in
the central business district of the city with 115 macro cells.
This part of the city has a very high density of macro cells
due to high volume of mobile data-traffic. Going forward, LTE
deployments are going to have low-power pico cells in addition
to high power macro cells. Since picos are not yet deployed
in reality, 10 pico locations were manually embedded into the
network planning tool using its built-in capabilities. Macros
have transmit power 40 W and picos have transmit power 4 W.
All including, we have 125 cells in the evaluated topology.
Though we had propagation data based on a real deployment, we do not have access to actual mobile measurements
at this stage. To generate measurement data required for our
evaluation, we generated synthetic mobile locations using the
capabilities of RNP tool. These location were used to generate
measurement data as follows:
1) Using the drive-test calibrated data, terrain information
and statistical channel models, the RNP tool was used
to generate signal propagation matrix in every pixel in
the area of interest in a major US metro as shown in
Figure 3.
2) The RNP tool was then used to drop thousands of
UEs in several locations where the density of dropped
12 The name of the city and the operator cannot be revealed for proprietary
reasons.

UEs was as per dense-urban density (450 active mobile

per sq-km). In addition, we defined mobile hotspots
around some of the picos where the mobile density was
doubled. Based on the signal propagation matrix in every
pixel and mobile drop locations, the RNP tool readily
generated all path loss data from UEs to it serving cell
and neighboring cells and also mobile to cell association
matrix.
3) The mobile path loss data was used to generate the
histograms and the other measurement KPIs proposed in
Section V. This data was then fed into LeAP database.
Once the measurement data is available, the results were
generated as follows:
1) The measurement data from LeAP database was read by
our implementation of LeAP algorithm that generated
the cell power control parameters.
2) To evaluate the gains, the parameters generated by LeAP
algorithm (or any other comparative algorithm) were
evaluated for a random UE snapshot generated by RNP
tool where each UEs data-rate based on its SINR-target
given by (17) was computed.
Comparative schemes and LeAP parameters: We compared the following schemes for our evaluation:
1) LeAP Algorithm: For most of our evaluation, we use
Algorithm 1 which has provable guarantees at a small expense
of computation time. Algorithm 2 requires expensive off-theshelf non-linear solver for large scale networks. Nevertheless,
we show the performance of Algorithm 2 using a free solver
based stand-alone implementation.
2) Fixed- FPC (FA-FPC): We compare the performance of
LeAP with the following approach popular in literature [6],
[17]: fix the same value of (typically = 0.8) for all cells
and set P (0) so that every mobiles SINR is above the decoding
threshold. SINR computation is performed using a nominal
value of interference Inominal that is typically 515 dB above
the noise floor.
There are two important design parameters for LeAP: the
histogram bin-size and the IoT-cap, that is partly dictated by
hardware designs constraints. In our evaluation, default choice
of bin-size is 1 dB and the default IoT-cap is 20 dB. In our
problem framework (see IoTC-SCP), we set
Imax = IoT-cap 2.bin-size
as a design choice to provide cushion for the path loss error
introduced by histogram binning. We also show results by
varying the bin-size and IoT-cap. Also, Pmax = 100 mW/RB.
C. Results
LeAP Performance Gains: In this section we evaluate the
performance of LeAP at various quantiles of the user data
rate CDF produced across the population of the users in our
deployment area. Figure 4 shows the LeAP performance with
FA-FPC for all users, users that are associated with macro
cells, and users that are associated with pico cells. Our main
observations are as follows:
As seen in Figure 4(a) LeAP provides a data rate improvement over the best FA-FPC scheme, of 4.9, 3.25, 2.12
for the 20th , 50th , 80th %-tile respectively. In other words,
with LeAP, at least half the users see a data rate improvement at least 3 and 80% of the users see an improvement

11

Fig. 5.
(a)

(b)

(c)

LeAP performance gain over fixed- FPC for (a) all UEs,
(b) macro cell UEs, and (c) pico cell UEs.
Fig. 4.

of at least 2. The larger gain at lower percentiles indicate

that LeAP is really beneficial to users towards the cell edge
who are most affected by inter-cell interference.
As seen in Figure 4(b),
4(c), LeAP provides higher
improvement to macro users as compared to pico users.
The median improvement of pico users data rate is 2.3
whereas it is 3.3 for macro users. This can be explained
by the fact that the macros use high power and have
larger coverage leading to more users affected by intercell interference and thus more users can reap the benefits
of LeAP.
In Figure 5, we show the gains of different cells in the
network by grouping all the 125 cells based on number of
dominant interferers. An interfering cell is dominant at least
5% of its uplink users, the users that selected this cell as
serving, that can interfere, i.e., Pr(interference) 0.05..
Figure 5 shows that the typical median data rate gain of cellgroups with 8 or fewer interferers is around 2 whereas,

Cell-wise LeAP gain vs. number of interferers.

the gains are around 4 for cell-groups with more than

10 interferers. We also show error bar for each cell group
based on the sample standard deviation of gains within each
group. We remark that the plot shows irregularity at 13 and
14 interferers due to having just 3 cells with 13-14 interferers.
These results suggest LeAP and other SON algorithms need
not be applied uniformly across the network but only in the
interference problematic regions of the network. Identifying
these regions from measurement data is an interesting problem
outside the scope of this paper.
LeAP performance with variation of histogram bin-size
and IoT-cap: In Figure 6, we show the median performance
gains of LeAP by varying histogram bin-size, with IoT-cap
either fixed at 20 dB or varying (with bin-size fixed at 2 dB)
. Our main observations are as follows:
Effect of bin-size: From Figure 6(a) we see that the
performance of LeAP and the gains deteriorate with larger
histogram bin-size which is intuitive. However, there is
marginal performance loss for 2 dB bin-size as compared
to 1 dB bin-size, but the performance deteriorates considerably for 6 dB bin-size. This suggests that the right bin-size
should not exceed 2 dB for optimal LeAP performance.
Effect of IoT-cap: The performance of LeAP with IoTcap is shown in Figure 6(b). Till an IoT-cap of around
20 dB, the median performance improves with increase in
Imax as there is greater flexibility in optimization. Past
that, the gain decreases as can be seen with an IoT-cap of
25 dB. This behavior is intriguing and can be explained as
follows. Let Iactual be the actual interference in the network for given user location. Also, Imax is the maximum
interference in IoT-SCP problem formulation. Note that,
Iactual could be higher than Imax due to approximations
introduced by histogram binning. Intuitively, with larger
Imax , the gap between Iactual and Imax is larger as there
are more interfering cells with non-negligible interference
signal. Thus, beyond a point, the binning approximations
somewhat nullify the benefits of increase in Imax .
LeAP computation times: For the 125 cell network, Algorithm 1 showed excellent gains after running for 50, 000
iterations. Using a 3.4 GHz Intel Xeon CPU with 16 GB
RAM on a 6 core machine with 64 bit OS, the running times
of our multi-threaded implementation were on an average
12 sec with a variability of around 5 sec. Since we expect
the periodicity of measurement reports to be in minutes, this
shows the scalability of our design.
Comparison of IoTC-CE and IoTC-SL: We also have developed a stand-alone implementation of IoTC-CE algorithm
using a open source NLP solver called CVX [15]. CVX is
a powerful tool for moderate sized problems that can deal
with the constraints in the IoTC-CE problem, that involve

12

Fig. 6.

(a)

Fig. 7. Median data rate comparison for IoTC-CE and IoTC-SL.

Histogram bin width is 0.5 dB.

log-sum-exponential functions. However, CVX cannot operate

for 125 cells in our evaluation network due to associated
variable limits in the solver, and so we choose a sub-region
consisting of 35 cells including 4 pico cells. Furthermore,
since CVX only takes objective functions of certain forms,
we modified our objective by choosing utility as V () =
ln(SINRmin + ln(exp())), clearly an approximation.
In Figure 7, we compare the median data rate of the two
LeAP algorithms and the best FPC scheme in our evaluation.
IoTC-CE still provides 90% improvement for all cells and
around 2 improvements for macro cells. For pico cells, IoTCCE outperforms IoTC-SL. Also, the performance of median
data rate of IoTC-CE is 36% less compared to IoTC-SL. This
could be partially due to that the approximated V (.) worsens
at lower SINR.
IX. C ONCLUDING R EMARKS
In this work we have proposed LeAP, a measurement data
driven approach for optimizing uplink performance in an LTE
network. Using propagation data from a real operational LTE
network, we show the huge gains to be had using our approach.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]

[7]
[8]
[9]
[10]

(b)

Effect of (a) histogram bin size on median LeAP performance, and (b) effect of IoT-cap on median LeAP performance.

http://www.sonlte.com/technology/.
http://lteworld.org/category/tags/son.
3GPP Technical Specification 36.331. Available from www.3gpp.org.
Atoll 3.2, Radio Network Planning Tool. http://www.forsk.com/atoll/.
A LEXOPOULOS , C. In Winter Simulation Conference.
A MIRIJOO , M., J ORGUSESKI , L., L ITJENS , R., AND NASCIMENTO ,
R. Effectiveness of cell outage compensation in lte networks. In
Consumer Communications and Networking Conference (CCNC), 2011
IEEE (2011), pp. 642-647.
B ERTSEKAS , D. Nonlinear Programming. Athena Scientific, 1999.
B ERTSEKAS , D., N EDIC , A., AND O ZDAGLAR , A. Convex Analysis
and Optimization. Athena Scientific, 2003.
B ISHOP, C. M. Pattern Recognition and Machine Learning (Information
Science and Statistics). Springer-Verlag New York, Inc., Secaucus, NJ,
USA, 2006.
B OCHE , H., AND S TANCZAK , S. Strict convexity of the feasible log-sir
region. IEEE Transactions on Communications 56, 9 (2008), 1511-1518.

[11] B ORKAR , V. Stochastic Approximation: A Dynamical Systems Viewpoint. Cambridge University Press, 2008.
[12] C HIANG , M., AND B ELL , J. Balancing supply and demand of bandwidth
in wireless cellular networks: Utility maximization over powers and
rates. In IEEE INFOCOM (2004).
[13] C HIANG , M., H ANDE , P., L AN , T., AND TAN , C. W. Power Control
in Wireless Cellular Networks Power Control in Wireless Cellular
Networks: Foundation and Trends in Networking. now, 2008.
[14] C OUPECHOUX , M., AND K ELIF, J.-M. How to set the fractional power
control compensation factor in lte ? In Sarnoff Symposium, 2011 34th
IEEE (2011).
[15] CVX R ESEARCH , I. CVX: Matlab software for disciplined convex
programming, version 2.0 beta. http://cvxr.com/cvx, Sep 2012.
[16] DAHLMAN , E., PARKVALL , S., S KOLD , J., AND B EMING , P. 3G
Evolution: HSPA and LTE for Mobile Boradband. Academic Press,
2008.
[17] ET. AL ., C. U. C. Performance of Uplink Fractional Power Control in
UTRAN LTE. In VTC Spring (2008), pp. 2517-2521.
[18] ET. AL ., M. B. Interference Based Power Control Performance in LTE
Uplink. In Wireless Communication Systems. 2008. ISWCS 08. IEEE
International Symposium on (2008), pp. 698-702.
[19] F OSCHINI , G. J., AND M ILJANIC , Z. A simple distributed autonomous
power control algorithm and its convergence. Vehicular Tech., IEEE
Trans. on 42, 4 (Nov 1993), 641-646.
[20] G EORGE , A. P., AND P OWELL , W. B. Adaptive stepsizes for recursive
estimation with applications in approximate dynamic programming.
Mach. Learn. 65, 1 (Oct 2006), 167198.
[21] G HAOUI , L. E. Hyper-Textbook: Optimization Models and Applications.
[22] G OLDSMITH , A. Wireless Communications. Cambridge University
Press, New York, NY, USA, 2005.
[23] H ANDE , P., R ANGAN , S., C HIANG , M., AND W U , X. Distributed
uplink power control for optimal SOR assignment in cellular data
networks. IEEE/ACM Trans. Netw. 16, 6 (2008), 1420-1433.
[24] H U , H., Z HANG , J., Z HENG , X., YANG , Y., AND W U , P. Selfconfiguration and self-optimization for LTE networks. Communications
Magazine, IEEE 48, 2 (February), 94-100.
[25] K AY, S. M. 1 ed. Apr.
[26] K HALIL , H. Nonlinear Systems. Prentice Hall, 2002.
[27] L OBINGER , A., S TEFANSKI , S., JANSEN , T., AND BALAN , I. Load
balancing in downlink lte self-optimizing networks. In Vehicular
Technology Conference (VTC 2010-Spring), 2010 IEEE 71st (May),
pp. 1-5.
[28] M ADAN , R., B ORRAN , J., S AMPATH , A., B HUSHAN , N., K HAN DEKAR , A., AND J I , T. Cell association and interference coordination in
heterogeneous lte-a cellular networks. IEEE Journal on Selected Areas
in Communications 28 (2010), 14791489.
[29] N EDIC , A., AND O ZDAGLAR , A. Subgradient methods for saddle-point
problems. Journal of Optimization Theory and Applications (2009),
205-228.
[30] P ENTTINEN , J. T. J. The LTE/SAE Deployment Handbook. John Wiley
& Sons, January 2012.
[31] R AHMAN , M., YANIKOMEROGLU , H., AND W ONG , W. Enhancing
cell-edge performance: a downlink dynamic interference avoidance
scheme with inter-cell coordination. IEEE Transactions on Wireless
Communications 9, 4 (2010), 14141425.
[32] S ARAYDAR , C., M ANDAYAM , N. B., AND G OODMAN , D. J. Pricing
and power control in a multicell wireless data network. IEEE JOURNAL
ON SELECTED AREAS IN COMMUNICATIONS 19 (2001), 1883
1892.

13

[33] S ON , H., L EE , S., K IM , S., AND S HIN , Y. Soft load balancing

over heterogeneous wireless networks. Vehicular Technology, IEEE
Transactions on 57, 4 (July), 2632-2638.
[34] S TOLYAR , A., AND V ISWANATHAN , H. Self-organizing dynamic fractional frequency reuse in ofdma systems. In INFOCOM (2008), IEEE,
pp. 691699.
[35] S UNG , C. W. Log-convexity property of the feasible sir region in powercontrolled cellular systems. IEEE Communications Letters 6, 6 (2002).
[36] T URKKA , J., N IHTILA , T., AND V IERING , I. Performance of LTE son
uplink load balancing in non-regular network. In Personal Indoor and
Mobile Radio Communications (PIMRC), 2011 IEEE 22nd International
Symposium on (Sept.).
[37] X IAO , M., S HROFF , N. B., AND C HONG , E. K. P. A utility-based
power-control scheme in wireless cellular systems. IEEE/ACM TRANS.
ON NETWORKING 11, 2 (2003), 210221.
[38] YATES , R. D. A framework for uplink power control in cellular radio
systems. IEEE Journal on Selected Areas in Communications 13 (1996),
1341-1347.
[39] Z HANG , H., P RASAD , N., R ANGARAJAN , S., M EKHAIL , S., S AID ,
S., AND A RNOTT, R. Standards-compliant lte and lte-a uplink power
control. In Communications (ICC), 2012 IEEE International Conference
on (2012).

A PPENDIX A
T WO S IMPLE U SEFUL L EMMAS
We will need the following inequalities in our proof. The
proofs of the inequalities are there in a longer version of the
paper.
Lemma A.1. The following holds for a0 = 1 and real ai >
0, i = 1, 2, . . . , k:
k
X

ln2 (

eai ) ln2 (k + 1) + (1 + 2 ln(k + 1))

i=0

k
X

a2i .

i=0

Lemma A.2. For real positive x1 , x2 , the following holds:

ln2 (x1 + x2 ) ln2 (e + x1 + x2 ) + ln2 x2 .

In the above, zi,min and zi,max are the lower and the upper
bounds of the corresponding primal variables. Similarly, define
the functions +
p (q) parametrized by p = [p1 , p2 , . . . pn ] as
 +

+
+
+
p (q) = p1 (q1 ) p2 (q2 ) . . . pn (qn )
where,
+
pj (qj ) =

We will use results from stochastic approximation theory

to prove our results. Broadly speaking, stochastic approximation theory states that, under suitable conditions, discrete
iterations with random variables converge to an equilibrium
point of an equivalent deterministic ODE (ordinary differential
equation) [11]. In other words, to show that a discretestochastic update rule converges to a desired point, we need
to show the following steps [11]: (i) define an equivalent
deterministic ODE, (ii) argue that the ODE has an equilibrium
point identical to to the underlying discrete-stochastic update
rule, and (iii) argue that conditions for limiting behavior of
discrete-stochastic update rule to be similar to that of the
ODE is satisfied (these conditions are provided by stochastic
approximation theory).
Defining a limiting ODE: Let m be the dimension of z or
equivalently the number of primal variables. First, define the
following operator z (y) parametrized by primal variable z:
h
it
(2)
(m)
z (y) = (1)
(y
)

(y
)
.
.
.

(y
)
.
.
.
,
m
1
2
z1
z2
zm

qj I{qj >0}
qj

, if pj = 0
, if pj > 0

We now define the following equivalent ODE for the

iterative update rule in Algorithm IoTC-SDP.

 

z(t) (E[Lz (z(t), h(z(t), O, ), p(t))])
z (t)
=
+
p(t)

p(t) (E[h(z(t), O, )])

(30)
Asymptotic equivalence of ODE: The fundamental result
in stochastic approximation theory [11] implies that the ODE
defined in (30) has similar asymptotic behavior to the updates
defined in Algorithm IoTC-SL provided certain conditions
hold. The following lemma establishes this equivalence by
verifying these conditions in our setup.
Lemma B.1. The stochastic primal-dual update rule of Algorithm IoTC-SL given by (21) and (22) converges to the
invariant set of ODE given by (30).
Proof: We introduce some notations to start with. Rewrite
the iterative update steps (21) and (22) in Algorithm IoTC-SL
in a compact form as


 

zn+1
zn
=P
+ an Mn
pn+1
pn
where,
Mn =

A PPENDIX B
P ROOF OF T HEOREM 1



z L(zn , h(zn , , ), pn )
h(zn , , )

(31)

and P(.) denotes the projection of the primal variables into

the compact set defined by bounds in Lemma VI.1 and dual
variables into the positive axis.
According to Chapter 5.4 [11], the candidate equivalent
ODE is given by





z (t)
(E[Lz (z(t), h(z(t), O, ), p(t))])
= (z(t),p(t))
p(t)

E[h(z(t), O, )]

where
(x) (y) = lim
0

P(x + y) P(y)
.

One can verify that this candidate equivalent ODE is precisely

the one defined by (30). Also, with a step-size of an = 1/n,
for iterative update steps (21) and (22) in Algorithm IoTC-SL
to converge to an invariant set of the candidate ODE, we need
to verify two conditions for our case (Chapter 5.4 [11]):
C1: The functions, z(t) (E[Lz (z, h(z, O, ), p)]) and
13
+
in z and p.
p (E[h(z, O, )) are Lipschitz
C2: The random variable Mn satisfies

where

yi I{yi >0}
yi I{yi <0}
(i)
(y
)
=
i
zi

yi

, if zi = zi,min
, if zi = zi,max
, else

E[kMn k ] K(1 + kzn k + kpn k )

13 A map x F (x) is called Lipschitz if kF (x ) F (x )k Kkx x k
1
2
1
2
for scalar constant K (0, ).

14

for some constant K > 0.

The conditions are not difficult to verify for our problem.
We outline the steps in the following.
Verifying C1: We need to show the Lipschitz continuity of
the r.h.s. of (30) in z and p. First note that the ODE (30)
is defined such that z(t), t 0 lies within the compact set
defined in Lemma VI.1 (provided the initial conditions are
also within that set) from which it follows that all terms (in
the r.h.s. of (30)) containing the primal variables are bounded.
Also, the r.h.s of (30) is a linear function of the dual variables
p. One can combine these two facts and argue easily l.h.s.
of (30) is Lipschitz in z and p. We skip the details.
Verifying C2: Note from (18) that, the randomness in
h(zn , , ) is only due to randomness in Zc,n where
!
X
(e,n +e,n e ec )
+ N0 .
Zc,n = ln
e e

Lemma B.2. The ODE given by (30) is globally asymptotically stable and the system converges to saddle point solution
of L(z, p) which also corresponds to the optimai value of z.
Proof: Consider the Lyapunov function
2

where, z and p are the primal and dual optimal solution of

the problem IoTC-SCP. Now note the following.
(t) = 2(z(t) z )t z(t)
+ 2(p(t) p )t p(t)

W
t
= 2(z(t) z ) z(t) (E[Lz (z(t), h(z(t), O, ), p(t))])
+ 2(p(t) p )t +
p(t) (E[h(z(t), O, )])

cC

for suitable positive constants K1 and K2 . To see this, first

observe the following.
2
E[Zc,n
]
"

(zi zi )(i)
zi (yi ) (zi zi )yi ,

(pi pi )+
pi (qi ) (pi pi )qi .

It thus follows from (33) that

(t) 2(z(t) z )t E[z L(z(t), h(z(t), O, ), p(t))]
W
2(p(t) p )t E[p L(z(t), h(z(t), O, ), p(t))] (34)
For notational convenience, let G(.) denote the function

e+

e e(e,n +e,n e ec ) + N0

G(z, p) = E[L(z, h(z, O, ), p)] .

+ ln2 N0

eJc

where we have applied Lemma A.2. Define A > 1 as the upper

bound on the path loss in linear-scale in the entire network.
Now notice that the function f (x) = ln2 (x + e) is concave for
x > 0. By applying Jensens inequality and definition of A,

Note that G(z, p) is concave in z and convex in p. Thus, from

the proerty of convex and concave functions[8]
(z z )t z G(z, p) G(z, p) G(z , p)
and
(p p )t p G(z, p) G(z, p) G(z, p ) ,

2
E[Zc,n
]

h
i
E e e(e,n +e,n e ec ) + e + N0

!
+ ln2 N0

eJc

!
ln2

since zi,min zi zi,max . Also,

!#

E ln2

ln2

(33)

Now note from the definition of z (.) that,

eJc

We can use the boundedness of the primal variables z to show

that
X
2
E[Mn2 ] K1 +
K2 E[Zc,n
],
(32)

W (t) = kz(t) z k + kp(t) p k ,

ee,n + e + N0

+ ln2 N0

eJc

which applied to (34) implies

(t)
W
G(z(t), p(t)) G(z , p(t)) (G(z(t), p(t)) G(z(t), p ))
= G(z(t), p ) G(z , p(t))
(35)
Since (z , p ) is a saddle point solution of

Applying Lemma A.1 to the previous step, we have for suitable

constants K3 and K4 (i.e., K3 , K4 are independent of e,n s),
X
X
2
2
E[Zc,n
] K3 + K4
ln2 (ee,n ) = K3 + K4
e,n
eJc

eJc

The above combined with (32) implies that,

2

E[Mn2 ] K5 + K6 k n k

for system dependent constants K5 , K6 (i.e., K5 , K6 could

depend on the network topology, but they are independent of
the primal and dual variables). Condition C2 is thus verified.
The lemma is proved since we have verified C1 and C2.
Global Stability of the ODE: Now that we have argued that
Algorithm IoTC-SL behaves like the ODE (30) asymptotically,
we prove that ODE (30) converges to the optimal point of
problem IoTC-SCP.

max min G(z, p) = min max G(z, p) ,

z

p0

p0

we have, G(z, p ) G(z , p ) G(z , p) for any z and

p 0, from which it follows that
(t) 0 ,
W
with equality iff z(t) = z and p(t) = p . Thus, it follows
from LaSalles invariance principle [26] that the ODE given
by (30) converges to (z , p ).
Putting it all together: Proof of Theorem 1. Lemma B.1
shows that Algorithm IoTC-SL converges to the invaraint
set of ODE (30) and Lemma B.2 shows that the ODE
converges to the optimal point of problem IoTC-SCP. Thus
Algorithm IoTC-SL converges to the optimal point of problem
IoTC-SCP. Hence the proof.