1

Frequency Plan Design for Multibeam Satellite

Constellations Using Linear Programming
Juan Jose Garau-Luis, Sergi Aliaga, Guillem Casadesus, Nils Pachler,
Edward Crawley, and Bruce Cameron
arXiv:2204.12494v1 [cs.CE] 26 Apr 2022

Abstract

Upcoming large satellite constellations and the advent of tighter steerable beams will offer unprece-
dented flexibility. This new flexibility will require resource management strategies to be operated in high-
dimensional and dynamic environments, as existing satellite operators are unaccustomed to operational
flexibility and automation. Frequency assignment policies have the potential to drive constellations’
performance in this new context, and are no exception to real-time and scalability requirements. The
majority of frequency assignment methods proposed in the literature fail to fulfill these two requirements,
or are unable to meet them without falling short on bandwidth and/or power efficiency. In this paper
we propose a new frequency assignment method designed to prioritize operational requirements. We
present an algorithm based on Integer Linear Programming (ILP) that is able to fully define a frequency
plan while respecting key system constraints such as handovers and interference. We are able to encode
operators’ goals such as bandwidth maximization or power reduction and produce optimal or quasi-
optimal plans according to such objectives. In our experiments, we find our method is able to allocate
at least 50% more bandwidth and reduce power consumption by 40% compared to previous operational
benchmarks. The performance advantage of our method compared to previous solutions increases with
the dimensionality of the constellation; in an experiment with a 5,000-beam MEO constellation we find
that we can allocate three times more bandwidth.

Index Terms

J. Garau-Luis, G. Casadesus, N. Pachler, E. Crawley, and B. Cameron are with the Aeronautics and Astronautics department,
Massachusetts Institute of Technology, Cambridge, MA, USA. S. Aliaga is with the Electrical Engineering and Computer Science
department, Northeastern University, Cambridge, MA, USA.
Corresponding author: Juan Jose Garau-Luis ([email protected])
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which
this version may no longer be accessible.
 2

Satellite communications, frequency assignment, multibeam constellations, resource management

I. I NTRODUCTION

Although satellite operators have been in the market for many years, their operations still
tend to be relatevely static—in time. This is especially the case for spectrum allocation policies.
When frequency plans do vary, the changes are generally manually-operated, in an effort to find
margin. This process entails dealing with a complex optimization problem that has been long
studied. However, upcoming changes in the satellite communications industry and the need for
automation pose additional challenges to spectrum optimization which previous methods might
not be able to address; new algorithmic approaches are required instead.

A. Motivation

Three trends define the new satellite communications landscape. First, new types of users
are coming into the market. Spaceborne data services are expected to substantially grow in the
coming years [1], and mobile users will constitute a significant fraction of this new demand [2].
In the past, operators could safely assume fixed terminal locations; moving forward, frequency
assignment policies must account for location changes over time, which at the technical level
involves considering time-dependant constraints. There is a need for effective mechanisms against
highly dynamic and time-dependent environments.
The remaining two trends are a consequence of technological improvements. On the one
hand, the introduction of highly flexible payloads is constituting a market disruption in terms of
spectrum management strategies. Previously, frequency plans were generally fixed, but in the next
generation of communication satellites, operators will be able to reconfigure frequency plans in
orbit, intermittently, and on a beam level. Exploiting this flexibility, as well as other mechanisms
such as frequency reuse, is key to reacting to demand changes and achieving efficient spectrum
usage.
On the other hand, the dimensionality of some of the upcoming constellations constitutes an
additional level of complexity. Examples include SpaceX’s 4,408-satellite LEO constellation with
up to 32 beams per satellite [3], [4] and SES’s O3b mPOWER MEO constellation, consisting of
11 satellites able to power thousands of beams each [5]. Any future frequency management
strategy needs to account for scalability, which challenges the adequacy of some spectrum
optimization approaches considered in the past.
 3

This new context calls for revisiting the strategies and algorithms previously utilized to make
frequency assignment decisions. This problem is not only NP-hard and highly constrained [6], but
now it also demands algorithmic solutions that are fast and scale well in order to be successfully
operated in upcoming high-dimensional and dynamic environments.

B. Literature Review

Previous studies focus on carrier assignment (i.e., which central frequency), bandwidth al-
location (i.e., how much bandwidth), or both. Reference [7] addresses the carrier assignment
problem, specifically targeting the recoloring problem by means of local search and the Simulated
Annealing algorithm, after an initial greedy combinatorial optimization approach. With the
scaling computational burden of classic optimization approaches in mind, other authors propose
Artificial Intelligence (AI) algorithms as an alternative to try to meet the new requirements of the
future scenarios. In [8], a Deep Reinforcement Learning (DRL) model is proposed to solve the
Dynamic Channel Allocation (DCA) problem, and it is shown to closely match the performance
of state-of-the-art DCA algorithms. While the online operation of such models consists of simple
forward passes of a neural network – substantially faster than other methods – the robustness
and stability of DRL models is still disputed in many real world scenarios [9].
Regarding bandwidth allocation studies, in [10], a binary search-based method is proposed as a
fairer alternative to the water-filling approach considered in [11]. A later study [12] also uses the
same fairness-centered optimization objective but allocates bandwidth using convex optimization
instead. However, the nature of future satellite constellations is not fully reflected in their analyses
since both studies test their methods on single-satellite and single-gateway use cases with no more
than 20 beams. In the AI domain, in [13] and [14], Genetic Algorithms and DRL are, respectively,
used to allocate bandwidth in single-satellite systems. Both studies consider simultaneously
optimizing other variables such as power or the roll-off factor, respectively, but the solution
quality and robustness of these AI methods are not tested for larger systems.
Both classic optimization and AI approaches are also considered for the combined carrier
assignment and bandwidth allocation problem. In [15] a neural network combined with a genetic
algorithm is used to address the problem through an interference minimization lens; it is tested
on multiple scenarios, with a maximum of 36 beams and 128 bandwidth channels. The same
problem, with an extension to also optimize for power, is addressed in [16] using a capacity-
oriented objective function and the Simulated Annealing algorithm. Tests on a single satellite
 4

with 200 beams and 16 bandwidth channels show that this method reduces both the unmet
capacity and the excess capacity in comparison to traditional approaches based on conventional
payloads. While these studies focus on optimizing bandwidth and carrier simultaneously, further
considerations such as frequency reuse mechanisms or dynamic environments are out of their
scope.
Two works do account for reuse mechanisms and prove to be substantially faster than other
approaches. First, the DRL-based method introduced in [17] uses an agent to simultaneously
make carrier assignments, bandwidth allocation, and frequency reuse selection. Despite its ability
to scale, this algorithm does not always satisfy the complete set of constraints. Second, the
constraint satisfaction algorithm presented in [18] uses a greedy assignment strategy to produce
valid frequency plans in short amounts of time. However, since beams are assigned one at a time
without an objective function, its usefulness is unclear when the operator wishes to prioritize a
specific metric (e.g., bandwidth maximization).

C. Paper Objectives

Given the new industry needs and the flexibility and dimensionality of new systems, an
efficient spectrum management algorithm for multibeam satellite constellations must be able
to simultaneously address the carrier assignment and the bandwidth allocation problems for
constellations with up to thousands of satellites or beams. In addition, new frequency reuse
mechanisms should be taken into account as key efficiency drivers, and an algorithm with a
small computing burden is preferred given the dynamic behavior of the users. All the studies
presented fail to address this complete set of requirements. Furthermore, none consider optimizing
for multiple satellites simultaneously, and frequency reuse mechanisms are rarely included as an
additional optimization dimension.
To try to close this research gap, we propose a frequency plan design algorithm based on
Integer Linear Programming (ILP) that optimizes for carrier assignment, bandwidth allocation,
and frequency reuse on the beam level in multibeam satellite constellations. The algorithm
produces optimal or quasi-optimal plans according to an objective function, and these plans
respect interference and handover constraints. The objective function can reflect multiple goals,
such as maximizing bandwidth, minimizing frequency reuse, and minimizing power. We show
that our algorithm can be adapted to address multiple scenarios including non-Geostationary
Orbits (NGSO), dynamic environments, and high-dimensional constellations.
 5

Fig. 1: A constellation with NS identical satellites in the same orbit and NB beams is considered.
Gateways, fixed terminals, and mobile users are connected to the network.

D. Paper Structure

The remainder of this paper is structured as follows: Section II presents the frequency plan
design problem, the assumptions considered and the constraints involved; Section III introduces
the optimization method based on Integer Linear Programming, together with the requirements
that can be encoded in the objective function; Section IV discusses the results of applying our
optimization algorithm to different scenarios; and finally Section V remarks the conclusions of
the paper.

II. P ROBLEM S TATEMENT

As depicted in Figure 1, we consider a satellite constellation with NS identical multibeam

satellites, all of them in the same orbital plane. The constellation serves multiple users and
gateways on Earth, which we assume have already been grouped into NB beams. Each beam
constantly serves its gateway or its group of users, which might be mobile —the beam then
“follows” the users. At any point in time, each beam is powered from one —and just one— of
the satellites of the constellation. If the orbit of the satellites is NGSO, then handover operations
take place and the satellite powering each beam changes over time. Our goal is to design a
frequency plan for this constellation, which consists of completely defining frequency utilization
for all beams.
In terms of frequency resources, we assume all satellites are allowed to use the exact same
part of the spectrum, which is divided into NBW equal bandwidth chunks or slots. Likewise, all
 6

Fig. 2: Frequency assignment representation in the form of a grid with NF R · NP rows and NBW
columns. In this example, NF R = 3, NP = 2, and NBW = 13. Each of the 3 beams being
powered by the satellite is assigned to a cell in the grid representing the first slot (black squares)
and to a certain number of consecutive slots (colored cells). For instance, the beam depicted in
green is assigned to reuse group 2, left polarization, and is using slots 8 and 9.

satellites have identical frequency reuse mechanisms:

there are NF R frequency reuses available, as well as NP polarizations for each reuse. Po-
larizations allow to use more spectrum in a concentrated area without incuring into additional
interference. For example, NP = 2 when using right-handed and left-handed circular polar-
izations (RHCP and LHCP, respectively). For each beam, the operator must decide how many
bandwidth slots are assigned and which reuse group and polarization should be used.
Formally, a complete frequency plan is defined as the assignment, for every beam b ∈
{1, ..., NB }, of the following elements:
• A discrete number of bandwidth slots bb , which can’t be greater than NBW .
• A positive integer fb , that indicates the first bandwidth slot used. Beam b then uses slots
fb , fb + 1, ..., fb + bb − 1, all of them part of the available spectrum.
• A positive integer rb representing a frequency reuse out of the NF R available.
• In case NP = 2, a binary variable representing the chosen polarization.
Figure 2 introduces a representation of this decision space in the form of a grid, with NF R ·NP
rows and NBW columns. Each column represents a frequency slot, whereas each row corresponds
to a combination of a frequency reuse and a polarization. As shown in the figure, rows are
sorted, first, by frequency reuse, and second, by polarization. With this representation, making a
frequency assignment for a beam turns into picking a specific cell in the grid corresponding to
the first slot (black squares in the figure) and choosing a valid number of slots (colored cells).
As previously described, based on the upcoming satellite communications landscape, we
 7

Fig. 3: Handover operation example between two satellites at time instants t1 and t2 .

assume all satellites have digital and flexible payloads. Therefore, during operations, we assume
any of the assignment parameters can be changed in real-time at negligible cost. Generally
though, the operator will choose to make changes to the frequency assignment of certain beams
whenever those beams undergo a handover. If the frequency assignment of a beam is not changed
during the handover, then the beam utilizes the same frequency assignment on the new satellite.
This is depicted in Figure 3, in which the green beam switches from satellite 1 to satellite 2 and
preserves the same frequency reuse, polarization, and bandwidth slots.
When a handover occurs, it is critical that the resources to be used on the new satellite are
not being used by any other beam already. This constitutes an important constraint when making
frequency assignment decisions for a NGSO constellation and applies to any pair of beams being
powered by the same satellite at any point in time. We define this constraint for a pair of beams
as an intra-group restriction between those beams. Then, the set RA represents all pairs (i, j)
such that beams i and j hold an intra-group restriction. We assume that this set is known in
advance and is externally updated during operation if needed.
There is another type of constraint to consider. If two beams whose footprints are close use the
same polarization, they might interfere with each other if their assigned frequency slots overlap.
This event, which we define as an inter-group restriction, is represented in Figure 4. Similarly to
the handover constraint case, the set RE encodes all pairs of beams (i, j) that hold an inter-group
restriction. This set is constructed following some interference criteria (e.g., SINR level, angular
 8

Fig. 4: Inter-group restriction example between beam 1 and beam 2. Diagram shows the moment
beam 2 is to be assigned and beam 1 has already been assigned.

distance threshold) and might be time-dependant if the constellation serves mobile users.

III. P ROPOSED O PTIMIZATION M ETHOD

In this section we encode each decision and restriction presented in the previous section
as variables and constraints of an optimization method based on Integer Linear Programming
(ILP). In the context of satellite resource allocation, methods such as ILP allow operators to
easily encode problem needs, objectives, and constraints with low granularity with respect to
the decision variables. Then, there exist multiple commercially-available tools that automatically
produce the solution of each program.
In each subsection we outline how a specific feature of the frequency assignment problem is
encoded in the ILP formulation using linear operators. We also give an overview of possible
objective functions to guide the search to select the best feasible frequency plan. We assume the
constraint sets RA and RE are an input to the model.

A. Frequency Plan Decisions

On the beam level, the frequency assignment decisions consist of choosing how many band-
width slots, which ones, and which frequency reuse and polarization to use. From the perspective
of the grid representation introduced in Figure 2, this means selecting 1) a column and 2) a row
in the grid, and then 3) a number of consecutive slots. We encode the column (i.e., the first
slot) as an integer variable fi , with domain {1, NBW }, for each beam i ∈ {1, ..., NB }. Then we
encode the row (i.e., frequency reuse and polarization) as an integer variable gi , with domain
 9

{1, ..., NF R · NP }. Finally, the number of consecutive slots is encoded as an integer variable bi ,
with the same domain as fi . We introduce these variables in the ILP formulation as follows:

fi ∈ {1, ..., NBW }, ∀ i ∈ {1, ..., NB } (1)

gi ∈ {1, ..., NF R · NP }, ∀ i ∈ {1, ..., NB } (2)

bi ∈ {1, ..., NBW }, ∀ i ∈ {1, ..., NB } (3)

Any frequency plan can then be decoded using these three variables per beam. The optimization
method returns these values to the operator.

B. Constraint: Limited spectrum

We now start accounting for the constraints of the problem. We first encode the constraint
that all bandwidth slots used must be within the limits imposed by the system (i.e., only the
NBW considered slots can be used). This is encoded as follows:

fi + bi − 1 ≤ NBW , ∀i ∈ {1, ..., NB } (4)

C. Auxiliary variables: relative spectrum location

The remaining constraints account for the intra-group and inter-group restrictions. Simply put,
both of these restrictions involve the need to avoid spectrum overlaps for certain pairs of beams.
For instance, if beams i and j hold a restriction, and are assigned to the same frequency reuse
and polarization (i.e., gi = gj ), we want to make sure that fi + bi ≤ fj or fj + bj ≤ fi , i.e., we
want to ensure that beam i has allocated spectrum either to the left or to the right of beam j,
but without overlapping. Before diving into the actual restrictions, these two possibilities need
to be encoded in the optimization formulation in order to preserve linearity. To that end, we use
binary variables zij and add the following constraints:

zij ∈ {0, 1}, ∀ i, j s.t. (i, j) ∈ R (5)

fj − fi ≥ 0 − M (1 − zij ), ∀ i, j s.t. (i, j) ∈ R (6)

fi − fj ≥  − M zij , ∀ i, j s.t. (i, j) ∈ R (7)

where M is a “sufficiently large” number,  is a very small positive number, and R represents
the union RA ∪ RE . Given a restriction (i, j), if zij = 1, (6) is active and enforces that fj ≥ fi
(i.e., beam j can not use lower frequencies than beam i’s). On the contrary, if zij = 0, (7) is
active and the effect is the opposite.
 10

D. Constraint: Intra-group or handover restrictions

We start by considering the intra-group restrictions, given by the set RA . This type of
restrictions are caused by handover operations and are relevant if and only if constrained beams
use the same frequency reuse and polarization. To first encode this condition linearly, we define
auxiliary binary variables yij alongside the following constraints:

yij ∈ {0, 1}, ∀ i, j s.t. (i, j) ∈ RA (8)

gi ≥ gj − M (1 − yij ), ∀ i, j s.t. (i, j) ∈ RA (9)

gi ≤ gj + M (1 − yij ), ∀ i, j s.t. (i, j) ∈ RA (10)

If yij = 1, the method enforces that gi = gj , since both (9) and (10) are active. If yij = 0, the
opposite should occur; however, this can not be achieved solely with variable yij . To enforce
strict inequality when yij = 0, we introduce binary variables pij to account for both cases gi > gj
and gj > gi . Hence, we add the following constraints:

pij ∈ {0, 1}, ∀ i, j s.t. (i, j) ∈ RA (11)

gi − gj ≥  − M (1 − pij + yij ), ∀ i, j s.t. (i, j) ∈ RA (12)

gi − gj ≤ − + M (pij + yij ), ∀ i, j s.t. (i, j) ∈ RA (13)

which are active if and only if yij = 0. In that case, and when pij = 1, (12) is active and we
have gi > gj . On the contrary, if pij = 0, (13) is active and gj > gi holds. To summarize the
effect of the auxiliary variables introduced so far, Table I shows how the different frequency
assignment cases between two beams holding an intra-group restriction are encoded by means
of variables zij , yij , and pij .
Finally we introduce the constraints that encode the intra-group restrictions:

fi + bi ≤ fj + M (2 − yij − zij ), ∀ i, j s.t. (i, j) ∈ RA (14)

fj + bj ≤ fi + M (1 − yij + zij ), ∀ i, j s.t. (i, j) ∈ RA (15)

where M is again a “sufficiently large” number. These constraints enforce a non-overlapping

frequency assignment between beams i and j holding an intra-group restriction if and only if
both beams use the same reuse group and polarization (yij = 1). Constraint (14) does so for the
case in which zij = 1 (fi ≤ fj ), whereas constraint (15) is active when zij = 0 (fi > fj ).
 11

TABLE I: Encoding of different frequency assignment cases by means of auxiliary variables

when beams i and j share an intra-group constraint.

Auxiliary variables Frequency assignment case

zij = 1 fi ≤ fj

zij = 0 fi > f j

yij = 1 gi = gj (same freq. reuse and polarization)

yij = 0 and pij = 1 gi > gj

yij = 0 and pij = 0 gi < gj

E. Constraint: Inter-group or interference restrictions

The inter-group restrictions are given by set RE and concern all pairs of beams with close
footprints, which might interfere with each other during operations. Setting a threshold for how
close two interfering beams can be is up to the operator’s policy, which might prefer to trade
additional inter-group restrictions for further interference mitigation. One way to define this set
might be to impose an angular distance threshold between beams; we leave this decision out of
the scope of this paper.
As introduced in Figure 4, this type of restrictions can negatively impact the performance of
the system when both beams are using the same polarization, regardless of their frequency reuse.
To specifically focus on polarization, we first introduce the following variables and constraints:

ki ∈ {1, ..., NF R }, ∀ i ∈ {1, ..., NB } (16)

mi ∈ {0, NP − 1}, ∀ i ∈ {1, ..., NB } (17)

gi = NP ki − mi , ∀ i ∈ {1, ..., NB } (18)

Variable ki encodes the frequency reuse assigned to beam i whereas variable mi encodes its
polarization, if any.
Similar to constraints (8) - (10), we introduce binary variable sij for each pair of beams
holding an inter-group constraint. This variable encodes whether beams i and j use the same
 12

TABLE II: Encoding of different frequency assignment cases by means of auxiliary variables
when beams i and j share an inter-group constraint.

Auxiliary variables Frequency assignment case

zij = 1 fi ≤ fj

zij = 0 fi > f j

sij = 0 mi = mj (same polarization)

sij = 1 and dij = 0 mi < mj

sij = 1 and dij = 1 mi > mj

polarization, by means of the following constraints:

sij ∈ {0, NP − 1}, ∀ i, j s.t. (i, j) ∈ RE (19)

mi ≥ mj − M sij , ∀ i, j s.t. (i, j) ∈ RE (20)

mi ≤ mj + M sij , ∀ i, j s.t. (i, j) ∈ RE (21)

If sij = 0 then both (20) and (21) are active, and mi and mj are enforced to be equal. Note that
in case NP = 1, then sij is always zero, since there is only one polarization.
Then, following the same idea behind constraints (11) - (13), we use binary variable dij to
help encoding whether mi > mj or mj > mi , i.e., enforcing mi and mj to be different in case
there is more than one polarization and sij = 1. The following constraints explain this idea:

dij ∈ {0, 1}, ∀ i, j s.t. (i, j) ∈ RE (22)

mi − mj ≤ − + M (1 + dij − sij ), ∀ i, j s.t. (i, j) ∈ RE (23)

mi − mj ≥  − M (2 − dij + sij ), ∀ i, j s.t. (i, j) ∈ RE (24)

which can only be active if sij = 1 (more than one polarization). Then, if dij = 0, (23) is active
and therefore mj > mi . On the other hand, if dij = 1, (24) is active and mi > mj . Table II
summarizes all cases that inter-group restrictions-related auxiliary variables encode.
Lastly, the actual constraints encoding the inter-group restriction interactions are the following:

fi + bi ≤ fj + M (1 + sij − zij ), ∀ i, j s.t. (i, j) ∈ RE (25)

fj + bj ≤ fi + M (sij + zij ), ∀ i, j s.t. (i, j) ∈ RE (26)

These constraints enforce a non-overlapping assignment between beams i and j holding an inter-
group restriction if and only if both beams use the same polarization (sij = 0). Constraint (25)
 13

does so for the case in which zij = 1 (fj ≥ fi ), whereas constraint (26) is active when zij = 0
(fi > fj ).

F. Objective Function

So far, we have discussed the decisions and constraints that need to be encoded to define valid
frequency plans, i.e., plans that do not violate any constraint. However, we have not addressed
how to prioritize different valid frequency plans according to the operator’s preferences and
goals. In certain occasions, operators might prefer plans that maximize bandwidth allocation or
plans that use as few frequency reuses as possible. To encode these preferences into the ILP
formulation, we propose the following objective function to be maximized:
NB
X
max (β1,i bi − |β2,i |gi − |β3,i |fi − |β4,i |Pi (fi , bi )) (27)
i=1

where β1,i , β2,i , β3,i , and β4,i are weighting parameters for beam i, with βk,i ∈ R. This function
combines four different objectives:
1) β1,i bi focuses on maximizing allocated bandwidth for beam i, or minimizing it in case
β1,i < 0. A better control over used bandwidth can lead to a more efficient power
consumption.
2) −|β2,i |gi attempts to use as few frequency reuses as possible, which is the case if |β2,i | > 0.
Setting β2,i = 0 for all beams enables a uniform use of frequency reuses and polarizations.
The coefficient β2,i uses absolute terms because the effects of maximizing and minimizing
gi are symmetric according to our formulation.
3) −|β3,i |fi seeks to have the same effect on the number of frequency slots used. If −|β3,i | <
0, lower parts of the spectrum are prioritized. If β3,i = 0, then the spectrum is used
uniformly. We specifically subtract this term given lower frequencies may require less
power consumption, and therefore are sometimes preferred by operators.
4) If available, −|β4,i |Pi (fi , bi ) directly represents the RF power consumed by beam i when
using slots fi , fi + 1, ..., fi + bi − 1. Reducing RF power when possible is reflected by this
operand when β4,i > 0. In our experiments, this metric does not correspond to a linear
combination of variables per se. In the following section we provide additional details on
its computation.
 14

The weighting parameters are used to define a priority hierarchy over these objectives. While
these parameters can be identical for all beams, operators might be interested in prioritizing
additional bandwidth or certain bands for specific beams, for example.

G. Additional design constraints

The variables and constraints presented so far encode the necessary information connected to
the main decisions and main restrictions of the problem. There are other type of considerations
that are not critical to operation or might not always be applicable but might be part of the
operator’s desiderata. We describe them next.
1) Minimum bandwidth requirement: We have defined the bandwidth variable (3) with domain
{1, ..., NBW }. However, we might be interested in specifying higher lower bounds, for contractual
reasons or since using a single bandwidth slot might not be enough to satisfy the link budget
equation [19] for certain beams. This way, we could redefine variable (3) as

bi ∈ {ci , ..., NBW }, ∀ i ∈ {1, ..., NB } (3)

where ci is the minimum number of slots that beam i requires.

2) Business constraints: While not a requirement of the problem per se, operators might prefer
to first split frequency resources between different sets of users. For instance, specific frequency
reuses might be saved for mobile users or the spectrum might be partitioned according to user
latitudes. These constraints can be easily encoded in the ILP formulation by modifying the
corresponding domains of variables (2) and (3).
3) Active and inactive beams: In this work we have assumed a feasible solution is always
possible. However, we might find ourselves in the unlikely scenario in which the operator must
turn off certain beams in order to achieve constraint satisfaction. In this case, which beams to turn
off and how to redesign the frequency assignment constitutes a different optimization problem
that needs additional variables. Specifically, variables ai are introduced to encode whether beam
i should be active or inactive/turned off in the plan. We describe the remaining changes with
respect to the presented ILP formulation in Appendix A.

IV. R ESULTS

In this section we present and discuss the results and performance of the proposed frequency
plan optimization method. Specifically, we carry out three different experiments to validate its
 15

usefulness: 1) first, we test our algorithm on three low-dimensional scenarios in which spectrum
usage maximization is required. 2) In the second experiment, we study the scalability of our
method by optimizing the frequency assignment in larger systems. To that end, we introduce an
iteration-based optimization procedure that allows us to decrease the overall optimization time
when needed. Finally, 3) in the third experiment, we consider a multibeam satellite constellation
based on a real large constellation and a realistic user distribution to assess the benefits of using
our approach to reduce on-board power consumption.

TABLE III: Dimensionality of the satellite system and frequency assignment parameters used in
each experiment. *Nch corresponds to the number of changes per iteration when the iteration-
based adaptation is used (see Section IV-B). In all cases we consider a seven-satellite constellation
(NS = 7) and two polarizations (NP = 2).

Experiment Freq. reuses NF R Bandwidth slots NBW Users NU Beams NB Changes Nch *

1 8 40 50, 60, 100 96, 118, 182 -

2 20 200 1,000 1,060 25, 50, 100

3 20 200 20,000 5,000 50

In all cases, the experiments make use of a representative example of a MEO constellation
system, the O3b mPOWER constellation filed by O3b limited [20]. In this case, it has NS = 7
satellites, each capable of using NP = 2 polarizations. User distributions as well as gateway
locations are provided by SES S.A. based on realistic configurations. In our analyses, we focus
on optimizing the downlink frequency assignment, and vary the number of beams NB , frequency
reuses NF R , and bandwidth slots NBW to represent scenarios with different spectrum availability.
The specific parameters considered in each of the three experiments are summarized in Table III.
The variable Nch corresponds to a feature of the iteration-based procedure described in Section
IV-B. We use a commercial solver, Gurobi, for our experiments [21].
Before the method begins the optimization, given a particular user distribution, we use the
beam placement algorithm described in [18] to determine the required number of beams and their
positions. We also make use of the heuristic frequency assignment algorithm described in the
same reference as a baseline benchmark throughout the remainder of the paper. In addition,
this algorithm serves as a warm-start when using the iteration-based procedure detailed in
Section IV-B. Based on this beam placement, we consider two beams will have an inter-group
 16

or interference restriction if the center of their footprints is closer than four times the half-cone
angle.

A. Experiment 1: Maximizing Bandwidth Allocation

In this experiment, our goal is to verify the behavior of the optimization method by max-
imizing bandwidth allocation in three scenarios with an increasing number of users and, in
turn, an increasing number of beams, as described in Table III. We encode bandwidth allocation
maximization in the proposed objective function (27) by adjusting the weighting parameters as
follows:
β1,i = 1, β2,i = β3,i = β4,i = 0, ∀i ∈ {1, ..., NB }

To quantify the performance of our method and compare it against others, we use the normalized
total assigned bandwidth for both the heuristic (BWHEU ) and optimized (BWILP ) frequency
plans. This metric is defined as follows:
NB
1 X
BW = bi (28)
Ctot i=1
where Ctot is the total system capacity, defined as the total number of frequency slots available
in the constellation:
Ctot = NS · NBW · NF R · NP (29)

As mentioned, the rationale behind improving spectrum utilization is that it generally results in
lower power consumption, without explicitly resorting to the introduction of power considerations
in the formulation by means of parameters β4,i . We explore this case in Section IV-C. While we
do not restrict the maximum bandwidth that can be assigned per beam, in real operations this
might be necessary due to spectral efficiency considerations [19]. If needed, a bandwidth upper
bound can be also encoded in our formulation.
Figure 5 shows the frequency assignment at a specific instant for one of the satellites when
using the heuristic algorithm and when using our method. Each colored region represents the
frequency slots assigned to a beam. As observed, the colored region of the plot is substantially
larger when using our method. Since the plot represents the frequency assignment for a specific
time instant corresponding to a specific routing of beams to satellites, unassigned slots might be
reserved for beams that are soon to undergo or have recently undergone handover operations in
the simulation.
 17

(a) Baseline freq. plan [18] (b) Optimized freq. plan (ILP)

Fig. 5: Comparison for one specific satellite in the constellation between the frequency
assignment produced by the heuristic algorithm serving as baseline (left) and our method (right)
when bandwidth maximization is prioritized. The scenario is defined with NS = 7, NP = 2,
NF R = 8, NBW = 40, and NB = 182. The horizontal axis represents the frequency slots
fi ∈ {1, ..., NBW }, whereas the vertical axis represents the frequency reuse and polarizations
gi ∈ {1, ..., NF R · NP }. Gold coloring is used for gateway beams, while a different hue is used
depending on the central frequency assigned to each user beam.

To quantify the increase in the assigned bandwidth, Table IV presents the normalized total
assigned bandwidth for the heuristic algorithm and our method in the three test scenarios. The
results show that our optimization method achieves up to a 73% bandwidth increase with respect
to the heuristic baseline solution in the scenario with 100 users, while satisfying all frequency
and handover constraints.
An increase in the number of beams might not be proportional to an increase in the total
number of frequency slots that we are able to allocate. This is clearly reflected by the values
of BWILP presented in Table IV, since moving from 96 beams to 118 beams entails a total
assigned bandwidth increase, but the effect of moving from 118 beams to 182 beams is the
opposite. In the latter case, when increasing the number of beams, the number of intra- and
inter-group constraints also increases, and therefore frequency assigments can not exploit more
frequency slots in order to avoid frequency overlaps. Still, our method makes the best assignment
possible regardless of the number of constraints.
 18

TABLE IV: Results for Experiment 1, focused on maximizing bandwidth allocation.

Scenario Users NU Beams NB BWHEU BWILP BW Increase

1.1 50 96 0.18 0.27 51%
1.2 60 118 0.25 0.41 61%
1.3 100 182 0.21 0.37 73%

B. Experiment 2: Iteration-based extension and computing time tradeoffs

In this experiment we test the scalability of our method and develop an iteration-based
procedure to reduce the computing time in high-dimensional scenarios. As the dimensionality of
both the constellation and user base increase, making frequency decisions in real time becomes
more challenging, since the number of constraints scales quadratically with the number of beams.
If computing time is a constraint, there is a point in which every algorithm—even state-of-the-art
ones—must make tradeoffs in order to meet time requirements. We explore the tradeoffs of our
method in this section.
A method to speed up the optimization when dimensionality is large is presented in Appendix
B. In this modified approach, we transform the prior formulation to an iteration-based approach
where the search space is restricted to interesting regions. For each beam, we select a few
variable assignments for each decision variable and let the solver choose the best combinations
across beams. In each iteration, the optimization algorithm changes the frequency assignment of
Nch beams, while the rest are kept fixed. Although the algorithm can make frequency decisions
without any prior input, we found that using a complete suboptimal frequency plan as a warm-
start substantially decreases runtime. There is no requirement of validity for the warm-start
frequency plan, since our method can amend violated constraints or deactivate beams. In our
experiments, we use the frequency plan provided by the heuristic algorithm presented in [18]
as a warm-start. In general, we achieve a significant reduction in the number of variables and
constraints that are considered in each call to the optimizer.
To test the iterative optimization method, we use a scenario with higher dimensionality, with
NF R = 20 frequency reuses, NBW = 200 bandwidth slots, and NU = 1, 000 users. The users and
their corresponding gateways are connected using NB = 1, 060 beams. We therefore increase
both the search space —more resources— and the constraint set —more beams— compared to
the first experiments. We still optimize the frequency plan to maximize the allocated bandwidth.
 19

We run the algorithm using three different configurations: Nch = 25, 50, and 100 changes at
each iteration. The beams assigned at every iteration are chosen randomly and we do not impose
any restriction on how many times a single beam can undergo assignment. Given the nature of
the procedure, we might require specific beams to be reassigned in different iterations before
converging to a stable solution. In our experiments we run the algorithm until getting this stable
solution; we consider it is reached when the total allocated bandwidth does not increase during
50 consecutive iterations. Since we introduce stochastic elements, we run the algorithm 10 times
for each of the selected Nch and report statistics on the 10 runs.
Table V presents the results of using the iterative approach to reach a stable solution in less
time when using Nch = 25, 50, and 100. We compare the total assigned bandwidth increase with
respect to the warm-start and the computing time (wall-clock time). The value of Nch is directly
related to the tradeoff between computing time and number of iterations. In all cases, we are
able to find a global frequency assignment that more than triples the total assigned bandwidth
without violating any constraints of the system. When using Nch = 25 we do so in less than one
third the amount of time required for Nch = 100, although, since we are assigning fewer beams
at a time, we observe that the improvement per iteration is lower, as graphically represented in
Figure 6. The average total number of assignments, i.e. Nch · Nit , is approximately 1.1 × 105 ,
1.6×105 , and 2.0×105 for Nch = 25, 50, and 100, respectively. Note that this is about two orders
of magnitude higher than the number of beams (NB = 1, 060 in these experiments). Table VI
provides additional insights on the time and iterations it takes to increase the bandwidth by 100%,
200%, and 300%. Generally, we can observe our method achieves 80% of the improvement in
less than 20% of the convergence time.
With these experiments, we assess that the iteration-based procedure enables the scalability
of our optimization method, which is an important requirement for future high-dimensional
constellation operations. Although limiting the number of beams that the optimization algorithms
can modify significantly reduces the dimensionality of the solution space, the performance
improvement comes at the expense of converging to suboptimal solutions and, therefore, needing
to assign around one hundred times more beams in total. Nonetheless, robustness and fast
reaction times are generally preferred over optimality when operating in highly dynamic and
time-dependent environments, where the feasibility of the frequency assignment might be only
temporary. Appropriately setting hyperparameters such as Nch is key in those cases [22].
 20

TABLE V: Results for Experiment 2, reported over 10 different runs and focused on maximizing
bandwidth allocation using the iterative optimization method. All simulations run on a server
with 20 cores of an Intel 8160 processor and 192 GB of RAM. SEM = Standard Error of the
Mean.

BWILP BW Increase (%) Iterations Nit Comp. time (min)

Changes Nch BWHEU
Mean±SEM Mean±SEM Mean±SEM Mean±SEM
25 0.61±0.00 333±1 4578±130 10.4±0.2
50 0.14 0.63±0.00 342±1 3188±138 18.9±0.8
100 0.63±0.00 344±1 1967±101 46.3±1.8

TABLE VI: Results for Experiment 2, number of iterations and the computing time required to
reach an increase in allocated bandwidth of 100%, 200%, and 300% for three different algorithm
configurations: Nch = 25, 50, and 100. Number of iterations and computing time required to
converge are also included. All simulations run on a server with 20 cores of an Intel 8160
processor and 192 GB of RAM. SEM = Standard Error of the Mean.

Iterations Nit Comp. time (min)

Changes Nch
Mean ± SEM Mean ± SEM
BW Increase 100% 200% 300% Convergence 100% 200% 300% Convergence
25 195±1 534±3 1673±21 4578±130 2.3±0.7 3.6±1.1 5.9±1.9 10.4±3.3
50 101±0 281±1 831±7 3188±138 2.9±0.9 4.8±1.5 8.1±2.6 18.9±6.0
100 61±0 165±1 469±5 1967±101 6.0±1.9 10.1±3.2 18.0±5.7 46.3±14.6

C. Experiment 3: Minimizing Power Consumption in a Real-World Scenario

Finally, in this third experiment, we demonstrate the benefits of using our method to optimize
the frequency assignment in a realistic scenario with a large-scale constellation and user base.
The user data is provided by SES S.A. and includes throughput demand for 20,000 users
located around the globe which, after running our beam placement algorithm, are grouped
into approximately 5,000 beams (see Appendix C). In this experiment we prioritize directly
reducing power consumption by leveraging the fourth term of the objective function equation
(27), −|β1,i |Pi (fi , bi ). For the purposes of this section, we understand Pi as a linear function that
returns the consumed power for beam i given its fi and bi assignments. Its full implementation
details into the formulation are described in Appendix D.
Figure 7 shows the warm-start (heuristic baseline algorithm) and the optimized frequency
 21

350

300

Total Assigned Bandwidth Increase (%)

250

200

150

100

50 Nch = 25
Nch = 50
0 Nch = 100
0 1000 2000 3000 4000 5000
Iteration

Fig. 6: Evolution of the percentual increase of the total allocated bandwidth at each iteration for
different values of Nch .

(a) Baseline frequency plan used as warm-start (computed using algorithm from [18])

(b) Optimized frequency plan (our method)

Fig. 7: Comparison between the warm-start and the optimized frequency assignment in a scenario
with NS = 7, NP = 2, NF R = 20, NBW = 200, and NB = 5000.
 22

TABLE VII: Results for Experiment 3, minimizing power consumption in a real-world scenario.

Changes Nch Power Decrease BWHEU BWILP BW Increase Iterations Nit

50 39.8% 0.175 0.614 251% 2464

plan (our method) for the seven satellites of the constellation at one moment of the simulation.
For this experiment, we run the iterative implementation of the optimization algorithm with
Nch = 50 until a stable solution is reached, as it can provide a solution under real-time operational
constraints. The results presented in Table VII show that using our method reduces the power
consumption by up to a 40% after fewer than 2,500 iterations. This is achieved by increasing the
total allocated bandwidth by more than 250%, which can be observed in Figure 7: the optimized
plan has significantly larger colored regions, i.e., assigned frequency slots, as well as several
unused combinations of a frequency reuse and polarization.
The results of this experiment demonstrate that the framework with the iteration-based pro-
cedure is able to substantially improve the frequency assignment in a high-dimensional, real-
world-based system. The algorithm is able to navigate the search space of this highly-constrained
problem and change the frequency allocation of beams to decrease the overall power consump-
tion. The number of changes per iteration Nch and the total number of iterations can be modified
to adjust the system’s operation to meet any possible time requirements. Although we used
a specific representative constellation for our experiments, our framework can be adapted to
constellations with different flexibilities: number of satellites and planes, number of frequency
reuses and slots, additional operational constraints, etc.

V. C ONCLUSION

In this paper, we have addressed the need to develop a frequency assignment optimization algo-
rithm that accounts for the flexibility and dimensionality of the upcoming communication satellite
operations. Our optimization framework leverages the use of frequency reuse mechanisms to ef-
ficiently handle carrier and bandwidth assignment for multisatellite constellations with thousands
of beams while respecting handover and frequency constraints, a set of requirements that has
previously not been addressed altogether. Our method is based on Integer Linear Programming
and produces optimal or quasi-optimal frequency plans according to a flexible objective function
that allows encoding multiple goals such as maximizing bandwidth, minimizing the number of
 23

active frequency reuses, and minimizing RF power consumption. We have also presented an
iteration-based implementation of the framework that not only enables its operation in high-
dimensional use cases but also introduces degrees of flexibility to configure it depending on the
scenario and computing time constraints.
We have carried out three different experiments to validate the performance, scalability, and
potential real-world operability of our method, respectively. The results of the first experiment
prove the ability of the framework to efficiently solve the frequency plan design problem,
achieving a 73% improvement on the total allocated bandwidth in simple scenarios. The iterative
optimization method analyzed in the second experiment is able to optimize the frequency
assignment in higher-dimensional scenarios, achieving over a 300% improvement on spectrum
usage in a time-efficient manner. Finally, the results from the third experiment demonstrate the
benefits of using our framework in a real-world scenario with approximately 20,000 users and
5,000 beams. In this case, the framework increases the allocated bandwidth but prioritizes making
an efficient use of frequency reuses and polarizations in order to achieve a 40% reduction in
on-board power consumption.

A PPENDIX A
O PTIMIZATION WITH ACTIVATION VARIABLES

It is possible that, given the constraints of the problem, not all conditions can be met and
hence feasibility can not be guaranteed. To minimize the impact in those situations, some beams
can be turned off and their assigned resources reallocated. We introduce binary variables ai to
encode whether beam i should be active or not:

ai ∈ {0, 1}, ∀ i ∈ {1, . . . , NB } (30)

If beam i is active, then ai = 1. Now these variables need to be taken into consideration when
specifying the rest of the constraints of the program. Below we describe the changes to the
constraints presented in the original framework.

A. Limited spectrum

If the maximum number of bandwidth slots available are not enough for beam i to satisfy the
link budget equations [19], the beam should be turned off. To encode this scenario we modify
 24

constraint (4) and add the activation variables:

fi + bi − 1 ≤ NBW + M (1 − ai ), ∀i ∈ {1, ..., NB } (4)

Here, variable M represents a “sufficiently large” number, such that this constraint has no effect
if beam i is not active (ai = 0).

B. Intra-group restrictions

Other reasons why we might need to turn off certain beams are related to intra-group and inter-
group restrictions. Regarding the former, constraints (14) and (15) can be extended to account
for beam activation:

fi + bi − M (1 − ai ) ≤ fj + M (3 − aj − yij − zij ), ∀ i, j s.t. (i, j) ∈ RA (14)

fj + bj − M (1 − aj ) ≤ fi + M (2 − ai − yij + zij ), ∀ i, j s.t. (i, j) ∈ RA (15)

Now these constraints prevent frequency overlap between beams i and j holding an intra-group
restriction if and only if both beams are active (ai = 1 and aj = 1).

C. Inter-group restrictions

Similarly, the constraints encoding the inter-group restrictions (25) and (26) are also extended:

fi + bi − M (1 − ai ) ≤ fj + M (2 − aj + sij − zij ), ∀ i, j s.t. (i, j) ∈ RE (25)

fj + bj − M (1 − aj ) ≤ fi + M (1 − ai + sij + zij ), ∀ i, j s.t. (i, j) ∈ RE (26)

D. Objective function

Finally, to prevent the framework from unnecessarily turning off beams to improve the ob-
jective function (27), we include an additional term in the function that accounts for beam
activation:
NB
X
max (β1,i bi − |β2,i |gi − |β3,i |fi − |β4,i |Pi (fi , bi ) + |β5,i |ai ) (27)
i=1

By increasing the weight of parameter β5,i ∈ R the framework will prioritize having beam i
active to maximizing any of the other elements of the function.
 25

A PPENDIX B
I TERATION - BASED OPTIMIZATION FORMULATION

There are two main factors that influence the computing time of an ILP formulation: the
number of decisions to make and the dimensionality of the search space. An efficient formula-
tion is able to reduce both of those factors without compromising the quality of the solution.
The following lines describe the improved iteration-based ILP formulation used in Sections
IV-B and IV-C. Note that, contrary to the initial formulation, the speed-ups proposed do not
guarantee convergence or optimality over the final solution, but substantially reduce the amount
of computing time needed to obtain a good enough result.

A. From full-optimization to an iteration-based approach

The first element to notice is that optimizing a large number of beams at the same time entails
a high-dimensional combinatorial problem. Without accounting for interference restrictions, each
2
beam has an order of NF R NP NBW possibilities: for each beam, we can decide the frequency
reuse, polarization, initial frequency, and total bandwidth to use. Then, the number of frequency
2
plan combinations is in the order of (NF R NP NBW )NB . While this number is still tractable for
low dimensional scenarios, optimizing for a large number of beams is infeasible from a time
perspective. A way to reduce the dimensionality of the problem is to tune only certain beams at
a time, while keeping the rest fixed, and iterate over the optimization procedure while selecting
different sets of beams so that all beams are optimized. The number of beams that are allowed
to change at every iteration will be denoted by Nch .

B. Reducing the search space by ranking

2
Next, we note that, although we can decide from NF R NP NBW possibilities for each beam,
only a small subset of those options is interesting from an optimization point of view; we
explain this below. Specifically, considering the activation variables introduced in Appendix A,
each beam i contributes to the objective function in the form of:

β1,i bi − |β2,i |gi − |β3,i |fi − |β4,i |Pi (fi , bi ) + |β5,i |ai (27)

This allows us to rank the different options based on their contribution to the total objective
value. The most interesting options will be the ones with the highests ranks. Let us denote xf,g,b,i
as a binary decision variable that represents choosing the option for beam i which has initial
 26

frequency f , frequency reuse g, and bandwidth b. We can rewrite the contribution of each beam
as:
X X
(β1,i b − |β2,i |g − |β3,i |f − |β4,i |Pi (f, b) + |β5,i |)xf,g,b,i = lf,g,b,i xf,g,b,i (31)
f,g,b f,g,b

Where lf,g,b,i is the contribution of the option denoted by xf,g,b,i . Note that now the last factor
|β5,i | does not include the activation variable ai . This is addressed in the following section. Now,
instead of considering all the feasible options, we will consider only a subset of alternatives
for each beam, denoted by the symbol Vi . By considering a strict subset, we are limiting the
options of the algorithm and reducing the search space of solutions, with the expectation that
the optimal or close to optimal solution lies within the selected subset. By including a ranking
system, we can define this subset as the X most interesting options. It is important to highlight
that, while solutions with higher bandwidth will tend to be preferred, those are also the most
difficult to allocate in the frequency plan. Therefore, instead of a single ranking system for all
options, we pre-compute a ranking for each possible bandwidth, and include the options in the
top of each ranking in the set. In our experiments we select the top 10 candidates per bandwidth
assignment. Also, note that the interference constraints (inter-group restrictions) with beams that
are not being changed can be computed a priori, and options that are not feasible (i.e., solutions
that have interference constraints with beams that are not being changed), do not need to be
included in the ranking, thus reducing the number of active constraints in the problem. The new
problem formulation can be stated as:
NB X
X
max lf,g,b,i xf,g,b,i (32)
i=1 f,g,b∈Vi

Additionally, we need to impose that at most one option can be active at a time:
X
xf,g,b,i = 1 ∀i (33)
f,g,b∈Vi

C. Dealing with activation variables

Now we explain how the activation variables can come into play in this new formulation. As
we use a warm-start in our iteration-based experiments, we first need to distinguish between
cases where the beam in the warm-start plan is active or not:
1) According to the warm-start, the original beam i is active, and it has initial frequency
forig , frequency reuse gorig , and bandwidth borig . Then, we include an additional option
 27

xorig,i that corresponds to the equivalent option in the original plan. Then, the previous
constraint can be expressed as:
X
xf,g,b,i + xorig,i = 1 (33)
f,g,b∈Vi

2) According to the warm-start, the original beam i is not active. Then, we include the
additional variable ai that indicates if the beam is active (ai = 1) or not (ai = 0). The
previous constraint can be expressed as:
X
xf,g,b,i − ai = 0 (33)
f,g,b∈Vi

With this formulation, a feasible option is guaranteed to exist, although not all beams might be
active in the final solution. If a feasible solution with all beams active does exist, this formulation
will provide it.

D. Including intra- and inter-group restrictions

The final aspect to address in this new approach is how to include intra- and inter-group
constraints. As mentioned, interference with fixed beams is passively included by disregarding
the infeasible options. The only restrictions to be included are the ones between beams that
are allowed to change in the same iteration. Since the decision variables are now just binary
activation variables with implicit frequency reuse, initial frequency, and bandwidth, we can easily
precompute if two options from two different beams collide or not. If they do, we only need to
add a restriction of the type:
xfi ,gi ,bi ,i + xfj ,gj ,bj ,j ≤ 1 (34)

Which ensures that at most one of the two options will be active, guaranteeing constraint
satisfaction in the final solution. Note that, since we need to add a constraint for each pair
of colliding options, the number of constraints may increase quadratically with the number
of options, which imposes computational constraints in the number of options that can be
considered.
Now, the complexity of the problem is solely determined by the number of beams that are
allowed to change at each point, and the number of options considered for each beam, which
are engineering decisions that can be assessed independently of the problem characteristics. As
a final remark, note that if we set Nch = NB and consider all valid options, the formulation is
equivalent to the original formulation presented in this work.
 28

A PPENDIX C
B EAM PLACEMENT DISTRIBUTION

Figure 8 shows the beam placement distribution that results after using the beam placement
algorithm from [18] on a set of approximately 18,000 users. 5,000 beams are used to serve these
users.

Fig. 8: Beam placement for scenario with approximately 18,000 users and 5,000 beams.

A PPENDIX D
P OWER COMPUTING ALGORITHM

One of the elements in the objective function equation (27) corresponds to a metric for power
consumption. As power equations are not linear [19], we can not directly use them in our
framework. As a solution, we precompute, for each of the NB beams, the consumed power for
each possible frequency assignment.
To that end, we use the satellite communications models described in [13]. For simplicity, we
describe the procedure to compute the necessary power Pb for one beam b given its data rate
demand Db and a certain allocated bandwidth BWb . We assume the satellites use the MODCOD
 29

schemes defined in the standards DVB-S2 and DVB-S2X [23]. Given a certain roll-off factor
αb , we can compute the lower bound of the required spectral efficiency as
Db (1 + αb )
Γreq = (35)
BWb
We select the MODCOD whose spectral efficiency is the lowest such that Γ ≥ Γreq . If no such
MODCOD exists, we set a power value of Pb = M , where M is a sufficiently large number.
Otherwise, we can get the appropriate value for Eb /N from the MODCOD scheme. Since in
our work we have considered interference mitigation mechanisms by means of the inter-group
constraints, we assume interference is negligible. We then compute the necessary C/N0 as
 
C  Eb  Db
= · (36)
N0 b N b BWb


With C/N0 in dB, we can then compute the power as:


C 
Pb = + OBO − GTx − GRx
N0 b
+ FSPL + 10 log10 (kTsys ) (37)

where OBO is the power-amplifier output back-off, GTx and GRx are the transmitting and
receiving antenna gains, respectively, k. is the Boltzmann constant, and Tsys is the system
temperature (assumed to be 290K). FSPL and Latm account for the free-space path losses,
respectively. We assume FSPL are significantly larger than atmospheric losses, and losses at the
transmitting and receiving antennas, so we neglect all the latter.
We compute a power value Pb for each possible assignment of BWb , and repeat the process
for all beams in the constellation.

ACKNOWLEDGMENT

This work was supported by SES S.A.. The authors would like to thank SES S.A. for their input
to this paper and their financial support. The project that produced these results also received
the support of a fellowship from “la Caixa” Foundation (ID 100010434). The fellowship code
is LCF/BQ/AA19/11720036.

R EFERENCES

[1] Northern Sky Research, “VSAT and Broadband Satellite Markets,” Tech. Rep., 2019.
[2] ——, “IFC Market project to reach $36 billion in cumulative revenue by 2028,” Tech. Rep., 2019.
 30

[3] I. del Portillo, B. G. Cameron, and E. F. Crawley, “A technical comparison of three low earth orbit satellite constellation
systems to provide global broadband,” Acta Astronautica, 2019.
[4] N. Pachler, I. del Portillo, E. F. Crawley, and B. G. Cameron, “An Updated Comparison of Four Low Earth Orbit Satellite
Constellation Systems to Provide Global Broadband,” in 2021 IEEE International Conference on Communications
Workshops (ICC Workshops). IEEE, jun 2021, pp. 1–7. [Online]. Available: https://ieeexplore.ieee.org/document/9473799/
[5] SES S.A., “O3b mPOWER Press Factsheet,” Tech. Rep., Aug. 2021. [Online]. Available: https://www.ses.com/sites/
default/files/2022-02/SES O3b mPOWER Factsheet EN.pdf
[6] T. Mizuike and Y. Ito, “Optimization of frequency assignment,” IEEE Transactions on Communications, vol. 37, no. 10,
pp. 1031–1041, 1989. [Online]. Available: http://ieeexplore.ieee.org/document/41158/
[7] J.-T. Camino, S. Mourgues, C. Artigues, and L. Houssin, “A greedy approach combined with graph coloring for non-uniform
beam layouts under antenna constraints in multibeam satellite systems,” in 2014 7th Advanced Satellite Multimedia Systems
Conference and the 13th Signal Processing for Space Communications Workshop (ASMS/SPSC). IEEE, sep 2014, pp.
374–381.
[8] X. Hu, S. Liu, R. Chen, W. Wang, and C. Wang, “A Deep Reinforcement Learning-Based Framework for Dynamic
Resource Allocation in Multibeam Satellite Systems,” IEEE Communications Letters, vol. 22, no. 8, pp. 1612–1615, 2018.
[9] G. Dulac-Arnold, N. Levine, D. J. Mankowitz, J. Li, C. Paduraru, S. Gowal, and T. Hester, “Challenges of real-world
reinforcement learning: definitions, benchmarks and analysis,” Machine Learning, apr 2021. [Online]. Available:
https://link.springer.com/10.1007/s10994-021-05961-4
[10] U. Park, “A Dynamic Bandwidth Allocation Scheme for a Multi-spot-beam Satellite System,” ETRI Journal, vol. 34, no. 4,
pp. 613–616, aug 2012.
[11] J. P. Choi and V. W. S. Chan, “Optimum power and beam allocation based on traffic demands and channel conditions over
satellite downlinks,” IEEE Transactions on Wireless Communications, vol. 4, no. 6, pp. 2983–2993, nov 2005.
[12] Heng Wang, Aijun Liu, Xiaofei Pan, and Luliang Jia, “Optimal bandwidth allocation for multi-spot-beam satellite
communication systems,” in Proceedings 2013 International Conference on Mechatronic Sciences, Electric Engineering
and Computer (MEC). IEEE, dec 2013, pp. 2794–2798.
[13] A. Paris, I. del Portillo, B. G. Cameron, and E. F. Crawley, “A Genetic Algorithm for Joint Power and Bandwidth Allocation
in Multibeam Satellite Systems,” in 2019 IEEE Aerospace Conference. IEEE, 2019.
[14] P. V. R. Ferreira, R. Paffenroth, A. M. Wyglinski, T. M. Hackett, S. G. Bilen, R. C. Reinhart, and D. J. Mortensen,
“Multiobjective Reinforcement Learning for Cognitive Satellite Communications Using Deep Neural Network Ensembles,”
IEEE Journal on Selected Areas in Communications, vol. 36, no. 5, pp. 1030–1041, 2018.
[15] S. Salcedo-Sanz and C. Bousoño-Calzón, “A Hybrid Neural-Genetic Algorithm for the Frequency Assignment Problem in
Satellite Communications,” Applied Intelligence, vol. 22, no. 3, pp. 207–217, may 2005.
[16] G. Cocco, T. De Cola, M. Angelone, Z. Katona, and S. Erl, “Radio Resource Management Optimization of Flexible
Satellite Payloads for DVB-S2 Systems,” IEEE Transactions on Broadcasting, vol. 64, no. 2, pp. 266–280, 2018.
[17] J. J. G. Luis, E. Crawley, and B. Cameron, “Applicability and Challenges of Deep Reinforcement Learning for Satellite
Frequency Plan Design,” in 2021 IEEE Aerospace Conference. IEEE, mar 2021, pp. 1–11.
[18] N. Pachler de la Osa, M. Guerster, I. Portillo Barrios, E. Crawley, and B. Cameron, “Static beam placement and frequency
plan algorithms for LEO constellations,” International Journal of Satellite Communications and Networking, p. sat.1345,
aug 2020.
[19] G. Maral and M. Bousquet, Satellite communications systems: systems, techniques and technology. John Wiley & Sons,
2011.
 31

[20] Federal Communications Commission, “O3b Limited FCC International Bureau Filings [IBFS].” [Online]. Available:
https://fcc.report/IBFS/Company/O3b-Limited
[21] Gurobi Optimization, LLC, “Gurobi Optimizer Reference Manual,” 2022. [Online]. Available: https://www.gurobi.com
[22] S. Aliaga, J. J. Garau-Luis, E. Crawley, and B. Cameron, “Dynamic resource management algorithm reconfiguration for
multibeam satellite constellations,” in 2021 IEEE Aerospace Conference. IEEE, mar 2022.
[23] ETSI EN 302 307-2, “Digital Video Broadcasting (DVB); Second generation framing structure, channel coding and
modulation systems for Broadcasting, Interactive Services, News Gathering and other broadband satellite applications;
Part 2: DVB-S2 Extensions (DVB-S2X),” Tech. Rep., 2015.