Dynamic Spectrum Management (DSM) Algorithms For Multi-User XDSL

IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 14, NO.
1, FIRST QUARTER 2012 109

Dynamic Spectrum Management (DSM)
Algorithms for Multi-User xDSL
Sean Huberman, Christopher Leung, and Tho Le-Ngoc, Fellow, IEEE
AbstractDynamic Spectrum Management (DSM) is an ef-
fective method for reducing the effect of crosstalk in Digital
Subscriber Line (DSL) systems. This paper discusses various
DSM algorithms, including Optimal Spectrum Balancing (OSB),
Iterative Spectrum Balancing (ISB), Autonomous Spectrum Bal-
ancing (ASB), Iterative Water-Filling (IWF), Selective Iterative
Water-lling (SIW), Successive Convex Approximation for Low
complExity (SCALE), the Difference of Convex functions Algo-
rithm (DCA), and Distributed Spectrum Balancing (DSB). They
are compared in terms of performance (achievable data rate)
and computational complexity.
Index TermsDigital Subscriber Line (DSL), Dynamic Spec-
trum Management (DSM), resource allocation, power allocation,
non-convex optimization.
I. INTRODUCTION
D
IGITAL Subscriber Line (DSL) continues to be the
most popular broadband access technology [1]. The most
signicant factor in the performance of DSL systems is the
interference between cables, known as crosstalk. Crosstalk has
the potential to severely limit the performance of the system
if it is not dealt with. The DSL channel is highly frequency-
selective and, in general, at high frequency the crosstalk is very
signicant. In order to make use of the higher frequency bands,
effective spectrum management techniques must be employed.
The most basic form of spectrum management is Static
Spectrum Management (SSM) [2]. SSM implements identical
spectral masks based on a worst-case scenario assumption for
all users. Clearly, this leads to inefcient spectrum manage-
ment whenever the scenario is not the worst-case and leads to
highly sub-optimal performance.
The poor performance of SSM systems led to the intro-
duction of Dynamic Spectrum Management (DSM) [3]. DSM
is a wide eld which looks to adaptively apply different
spectral masks for each user with the intent of maximizing the
throughput of the system. DSM allows for a far more efcient
use of the spectrum than SSM does. As a result, many different
DSM algorithms have been proposed. As well, it was shown
in [4] that DSM allows for more efcient power usage than
SSM techniques.
The main criteria in comparing different DSM algorithms
are their performance and their complexity. The performance
relates how well an approach succeeds at maximizing the
Manuscript received 30 September 2009; revised 26 April 2010, 27 July
2010, and 3 September 2010.
The authors are with the Department of Electrical and Computer
Engineering, McGill University, 3480 University Street, Montreal, Que-
bec, Canada, H3A 2A7 (e-mails: sean.huberman@mail.mcgill.ca; christo-
pher.leung@mail.mcgill.ca; tho.le-ngoc@mcgill.ca).
Digital Object Identier 10.1109/SURV.2011.092110.00090
achievable data rate as compared to the theoretical optimum.
The complexity of the model is related to the amount of time
required to derive the power allocation as the number of users
and frequency tones increase.
There are two main types of DSM algorithms: centralized
and distributed. Centralized systems require a central hub with
full knowledge of the network. In general, this system allows
for better performance at a cost of increasing the complex-
ity and computational time. On the other hand, distributed
systems allow for every user to self-optimize. Distributed
systems that can operate fully autonomously without the need
of explicit message passing are called autonomous systems.
In general, distributed systems reduce the complexity and
computational time but often sacrice some optimality in
terms of performance.
One of the rst distributed DSM algorithms was Iterative
Water-Filling (IWF) [5]. In IWF, each user selshly maximizes
their own data-rate. While IWF gives signicant data-rate
improvements over SSM techniques, in many situations, it
leads to sub-optimal performance.
The sub-optimality of IWF is caused by inefcient use
of the frequency spectrum. In an attempt to increase the
efciency of the frequency spectrum, many other heuristic
variations were proposed. Interested readers are referred to
[6], [7], and [8].
Two similar algorithms which improve on the performance
of IWF are Band Preference Spectrum Management (BPSM)
[9], and Iterative Power Pricing (IPP) [10]. Both apply the gen-
eral IWF algorithm but BPSM gives preferences to different
users in each frequency band, while IPP uses power pricing
for each frequency tone to allow weaker users to compete
more fairly. The BPSM and IPP algorithms are intuitively
similar, except they solve different optimization problems.
BPSM is used to maximize the achievable rate, while IPP is
used to minimize the required power. Section II-C discusses
the different optimization problems in more detail.
Another algorithm known as Selective Iterative Water-lling
(SIW) [11] was introduced. SIW selectively applies IWF to
users in the under-utilized sections of the frequency spectrum
until all users use up their total power. While SIW shows
signicant performance gains over IWF, the performance is
still sub-optimal.
A centralized algorithm called Optimal Spectrum Balancing
(OSB) [12], which maximizes the weighted sum rate across
the users, was derived to solve for the globally optimal power
allocation. OSB uses dual decomposition to solve for the
optimal transmit powers for each user separately on each
frequency tone by exhaustive search. While OSB is not
1553-877X/12/$25.00 c 2012 IEEE
110 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 14, NO. 1, FIRST QUARTER 2012
computationally tractable for many users, it serves as an upper
bound on the performance of other DSM algorithms for cases
with few users.
A similar centralized algorithm called Grouping Spectrum
Management (GSM) [13] was introduced. GSM groups users
into clusters based on similar line lengths and calculates a
reference line for each cluster. GSM then applies OSB on
the reference lines for each cluster in order to reduce the
computational complexity.
Another centralized algorithm, based on Monotomic Opti-
mization (MO), called MO-bAsed poweR controL (MARL)
[14] was introduced. MARL nds globally optimal points
by constructing a series of poly-blocks that approximate the
feasible region with increasing precision; however, MARL is
currently not feasible in the case of many users or frequency
tones.
By representing the objective function of the non-convex
optimization problem in OSB explicitly as the Difference
of two Convex functions (DC) [15], (i.e., f = g h), a
modied prismatic branch-and-bound algorithm was proposed
in [16]. The constraint set is convex since it consists of a
system of linear equations and one nonlinear constraint. The
Prismatic Branch and Bound (PBnB) algorithm [17] operates
by successively approximating the nonlinear constraint by a
piecewise linear function, and nds a globally optimal solution
by solving a sequence of Linear Programming (LP) sub-
problems. Although its computational complexity is still high,
the algorithm proposed in [16] can substantially reduce the
complexity of OSB in nding a globally optimal solution,
especially for a large number of users. It can provide an
upper-bound on the performance in many situations where
OSB is computationally intractable, and hence, can serve as a
comparison measure for other more practical DSM algorithms.
Since the previously mentioned global optimizing algo-
rithms are intractable for many users, a centralized algorithm
called Iterative Spectrum Balancing (ISB) [18] [19] was
introduced. ISB formulates the optimization problem similar
to OSB, using the weighted sum rate and dual decomposition,
but solves for the power allocations in an iterative fashion
using one-dimensional instead of N-dimensional exhaustive
searches (where N is the number of users). This allows for
ISB to be computationally tractable for many users.
A similar algorithm called Generalized Iterative Spectrum
Balancing (GISB) [20] was proposed. GISB operates similarly
to ISB and OSB, where the difference lies in terms of the
number of Power Spectral Densities (PSDs) over which the
exhaustive search is performed. For GISB, the exhaustive
search is performed over L users where 1 L N. Note that
ISB and OSB are the special cases where L = 1 and L = N,
respectively. GISB allows for a tradeoff between performance
and computational complexity. The closer the value of L is
to N, the better the performance will be at the expense of
Centralized systems are generally harder to implement in
practice. It is for this reason that many other algorithms were
introduced. One such algorithm is Autonomous Spectrum Bal-
ancing (ASB) [21]. ASB uses the same problem formulation
as OSB and ISB but operates in a distributed fashion without
the need for any explicit message passing. ASB uses the
concept of a virtual line (referred to as a reference line),
which represents the typical victim in the network. Each user
self-optimizes to protect the reference line and hence attempts
to better the overall network. Generally ASB cannot nd a
globally optimal solution; however, its performance has been
shown to be near-optimal in some situations while maintaining
a relatively low complexity.
One issue regarding the ASB algorithm is that the spectrum
update formula can be relatively time consuming. It is for this
reason that [22] proposed the ASB-2 algorithm. The ASB-
2 algorithm works exactly like the ASB algorithm but uses
a slightly different spectrum update formula. This spectrum
update formula has a signicantly lower complexity. ASB and
ASB-2 do not necessarily converge to the same solution; how-
ever, the ASB-2 algorithm converges signicantly faster than
ASB, especially when the number of users and/or frequency
tones are very large.
The ASB and ASB-2 algorithms presented in [21] and [22]
only use one reference line; however, it was shown in [23]
that using Multiple Reference Lines (MRL) can signicantly
improve performance while still maintaining a low complexity.
Guidelines for choosing reference lines to ensure consistently
strong performance were also presented in [23].
An algorithm called Semi-Blind Spectrum Balancing (2SB)
[24] builds on the concepts introduced in ASB. 2SB operates
in the same fashion as ASB but also dynamically updates
a virtual reference line parameters for each user separately,
to more accurately represent the network. The virtual lines
are updated at a Spectrum Management Center (SMC) based
on message passing to and from the users. This removes the
autonomous aspect of ASB, but generally leads to an improved
data-rate.
Several Sequential Convex Programming (SCP) algorithms
(iterative algorithms which solve a sequence of convex sub-
problems to nd a locally optimal solution) will be discussed
in this paper. One SCP algorithm, called Successive Convex
Approximation for Low-complExity (SCALE) [25] [26] can
be viewed as a hybrid centralized and distributed system since
it can be implemented in a distributed fashion with explicit
message passing. SCALE is an algorithm that applies a series
of concave lower-bounds to the maximization problem. This
enables SCALE to make use of the well researched area of
convex optimization to maximize the concave lower-bound.
Each successive iteration tightens the lower-bound towards a
locally optimal solution.
The PBnB algorithm (discussed above) makes use of the DC
property of the objective function to solve for globally optimal
solutions. Finding globally optimal solutions has a large
complexity; hence, several SCP algorithms (discussed below)
which make use of the DC property of the objective function
to solve for locally optimal solutions are more computationally
attractive.
One such SCP algorithm discussed in this paper makes
use of the Difference of Convex functions Algorithm (DCA)
presented in [27]. DCA is a centralized algorithm that begins
by re-writing the non-convex objective function in terms of the
difference of two convex functions (i.e., f = g h); however,
DCA iteratively creates an afne minorization (multivariate
rst-order approximation) of h, denoted by

h, which is used
HUBERMAN et al.: DYNAMIC SPECTRUM MANAGEMENT (DSM) ALGORITHMS FOR MULTI-USER XDSL 111
to make the objective function,

f = g
h, convex. Each suc-

cessive iteration more closely approximates the locally optimal
solution. For any function f, many DC decompositions exist
(e.g., g h = (g +)(h+)). The choice of decomposition
has a crucial impact on the convergence speed as well as the
performance. There are still a lot of heuristics regarding the
DCA implementation which have yet to be explored in great
detail.
A similar algorithm called Convex Approximation Dis-
tributed Spectrum Balancing (CA-DSB) was proposed in [22].
CA-DSB applies the DC property of the objective function
and follows a similar algorithm to DCA; however, CA-DSB
uses a different DC decomposition than the DCA in [27] and
therefore can operate in a distributed manner. This makes use
of parallel processing to reduce the complexity while still
maintaining strong performance through the use of message
passing.
Another algorithm called Distributed Spectrum Balancing
(DSB) was proposed in [22]. DSB also writes the objective
function as a DC function but it applies the KarushKuhn
Tucker (KKT) conditions directly. With the use of a message
passing system, the DSB algorithm solves for a locally optimal
solution in an iterative fashion. The DSB algorithm and CA-
DSB algorithms are nearly identical in terms of performance;
however, in general the runtime of DSB is slightly faster.
The Multiple Starting Point DSB (MS-DSB) algorithm,
proposed in [22], is an extension of the DSB algorithm which
makes use of multiple starting points. This allows for the
algorithm to nd solutions that are at least as good as the DSB
algorithm without signicantly increasing the complexity.
The rest of this paper is organized as follows: Section
II presents a brief overview of DSM problems in multi-
user xDSL. Section III discusses the eight algorithms: OSB,
ISB, ASB, IWF, SIW, SCALE, DCA, and DSB. Section
IV presents illustrative examples and simulation results to
evaluate and compare the performance of the eight algorithms,
while Section V discusses the complexity behaviour of the
algorithms under consideration. Section VI provides some
concluding remarks.
II. DSM IN MULTI-USER DSL NETWORK
A. xDSL Environment
xDSL is a family of technologies which transmit digital
data over twisted-pair copper telephone wires [3] [28]. xDSL
operates on the same twisted-pair copper wire as Plain Old
Telephone Service (POTS) since it uses a higher frequency
band, while POTS uses a lower frequency band (less than
4 kHz). The frequency band used for DSL depends on
the specic technology. For example, in Asymmetric Digital
Subscriber Line (ADSL) the maximum frequency used is
1.1 MHz, in ADSL2+ the maximum frequency used is 12
MHz, and in Very high bitrate DSL (VDSL) the maximum
frequency used is 30 MHz. There are dedicated frequency
bands for upstream and downstream transmission. For most
DSL technologies, a larger frequency band is allocated for
downstream transmission than for upstream transmission.
Twisted-pair copper wires are typically run underground in
binder groupings. Generally, the maximum number of twisted-
Fig. 1. DMT Block Diagram.
pair copper wires in a binder is 200. These are generally
organized into sub-bundles of 25 twisted-pair copper wires.
xDSL technology uses a scheme similar to Orthogonal
Frequency-Division Multiplexing (OFDM) known as Discrete
Multi-Tone (DMT). DMT is a multi-channel transmission
technique that divides the available spectrum into smaller sub-
channels or frequency tones [29]. DMT differs from OFDM
in that it is also capable of optimizing the bit and energy
distribution over the sub-channels (e.g., channel partitioning
or bit-loading) [3]. The basic idea of DMT transmission is
transmit the data in parallel over each frequency tone (note
that some frequency tones might transmit no data, while others
can transmit a lot of data).
Figure 1 shows the DMT transmission block diagram. The
data is put through a serial to parallel converter. This formats
the word size required for the parallel transmission. As well,
this transforms the wide-band frequency selective channel
into many parallel narrow-band frequency at tones. Each
frequency tone is modulated independently using an M-ary
modulation scheme (e.g., Quadrature Amplitude Modulation
(QAM), Phase-Shift Keying (PSK)). The Inverse Fast Fourier
Transform (IFFT) is used to convert the waveform from the
frequency domain to the time domain. A Cyclic Prex (CP) is
then added in order to avoid Inter-Symbol Interference (ISI)
and Inter-Channel Interference (ICI). After the data passes
through the DSL channel, the receiver performs the reverse
operations in order to recover the transmitted data.
A low symbol rate is used to ensure an appropriate guard-
band between frequency tones. For example, typically xDSL
systems use a frequency tone spacing of 4.3125 kHz and a
symbol rate of 4 kHz.
The main advantage of DMT is that it is well suited to
handle large attenuation at high frequencies as is typical in
xDSL systems. This is due to the fact that it converts the fre-
quency selective channel into many frequency at sub-carriers
(or frequency tones) and solves each one independently. The
number of frequency tones depends on the frequency band
allocated to the xDSL technology used (e.g., ADSL uses 256
frequency tones, whereas VDSL uses 4096 frequency tones).
One advantage of xDSL is that since it uses the same
twisted-pair copper wire as POTS, no new wiring is required.
xDSL technology makes more efcient use of the existing
phone lines; however, twisted-pair copper wire has an atten-
uation that increases with the line lengths, especially when
operating at higher frequencies as in the case of xDSL.
Due to the high attenuation at high frequencies, especially
for longer lines, xDSL is more suitable for last mile
networks or hybrid ber/copper networks. More specically,
while using optical ber instead of copper wire might yield
higher data rates over longer distances, it is far more expensive
to implement optical ber wire than to make use of the existing
copper wire. Thus, the ability to achieve similar data rates
using copper wire is particularly interesting.
High enough data rates would allow for more effective
hybrid ber/copper network architectures (referred to as FTTx
technology). One such example is Fibre To The Neighbour-
hood or Node (FTTN). FTTN uses the more expensive bre-
optic cables [30] to transmit information to a neighbourhood
(or a node closer to the customers premises) and then uses
regular twisted-pair wiring from the neighbourhood to each
and every subscribers home. This is far cheaper than a
Fibre To The Home (FTTH) [30] network architecture, where
expensive bre-optic cable is used directly to each subscribers
home.
Other such network architectures exist, such as Fibre To The
Curb or Cabinet (FTTC), Fibre To The Building or Basement
(FTTB), Fibre To The Premises (FTTP), and Fibre To The
Ofce (FTTO) [31].
Such network topologies can be found in practice in many
countries. For example, Japan deployed ADSL services and
was beginning to expand to VDSL services as of November
2005 [32]. In the United States, AT&T deploys a VDSL
service called AT&T U-verse which provides internet, TV
and phone through a FTTN network architecture to 2 million
customers across 22 states as of December 9, 2009 [33].
Similarly, Bell Canada services over 1.8 million homes with
VDSL2 through their FTTN networks in Montreal and Toronto
as of February 4, 2010 [34]. This is expected to double by the
end of 2010.
Figure 2 shows an example of an xDSL network. In this
network, optical bre is run to the Central Ofce (CO) and
to two DSL Access Multiplexers (DSLAMs). From the CO
and the DSLAMs, twisted-pair copper wire is run to various
Junction Wire Interfaces (JWIs). The JWIs run twisted-pair
copper wire to the customers premises and/or ofce building.
As shown in Figure 2, there are sections of the network where
the CO shares a binder with a DSLAM. Due to the long length
of the CO relative to the DSLAM, the DSLAM lines can cause
severe crosstalk to the CO users (this is known as the near-far
problem).
B. System Model
Consider a DSL network with a set of users (modems)
^ = 1, . . . , N and a set of tones (frequency carriers)
/ = 1, . . . , K. Using synchronous DMT modulation, there
is no ICI and transmissions can be modeled independently on
each tone k as follows:
y
k
= H
k
x
k
+z
k
.
The vector x
k
x
n
k
, n ^ contains the transmitted
signals for all users on frequency tone k, where x
n
k
is the
transmitted signal by user n on frequency tone k. Similarly,
y
k
y
n
k
, n ^ and z
k
z
n
k
, n ^ where y
n
k
is the
received signal for user n on frequency tone k. Likewise, z
n
k
is the additive noise for user n on frequency tone k which
contains thermal noise, alien crosstalk and radio frequency
interference. H
k
is an NN matrix such that [H
k
]
n,m
is the
channel gain from transmitter m to receiver n on frequency
tone k, and is dened as h
n,m
k
. The transmit PSD of user n
on frequency tone k is dened as s
n
k
c[x
n
k
[
2
/
f
, where
c denotes expected value, and
f
denotes the frequency
tone spacing. The vector containing the PSD of user n on all
frequency tones is dened as s
n
s
n
k
, k /.
When the number of users is large enough, the interfer-
ence is well approximated by a Gaussian distributed random
variable, and hence the bit loading (bits/s/Hz) of user n on
frequency tone k is dened as:
b
n
k
log
2
_
1 +
1
[h
n,n
k
[
2
s
n
k
m=n
[h
n,m
k
[
2
s
m
k
+
n
k
_
,
where is the Signal to Noise Ratio (SNR) gap which is a
function of the desired Bit Error Rate (BER), coding gain, and
noise margin [28], and
n
k
c[z
n
k
[
2
/
f
is the noise power
density of user n on frequency tone k. The achievable data
rate for user n is therefore
R
n
= f
s
k
b
n
k
.
C. Spectrum Management Problem
The spectrum management problem, based on the system
model shown in Section II-B, involves selecting the transmit
powers for each of the N users on each of the K frequency
tones.
The Static Spectrum Management (SSM) approach assumes
a worst-case scenario where all the possible users in a cable
binder are active [2] [3]. Based on this assumption, the
interference plus noise term for each user does not change
(i.e., it is static). Each user n must then adjust their transmit
power, s
n
k
, on each frequency tone k with the assumption that
the interference plus noise seen is at its maximum.
There are many different types of SSM technologies. One
example is called at Power Back-Off (PBO). Flat PBO
operates by ensuring that each user transmits the minimum
possible at PSD required to meet its Quality of Service (QoS)
requirement. Another example is the reference noise method
where each user sets its transmit PSD such that the crosstalk
felt by the virtual modem is equal to the background noise
seen by the virtual modem.
When all users are active and the situation is a worst-
case model, SSM provides reasonable spectrum management;
however, whenever the total number of active users is less than
the total number of users in the cable binder, SSM becomes
overly pessimistic and makes inefcient use of the spectrum.
The worst-case model used by SSM is entirely determined
based on the length and width of the cable used.
In practice, many bridge-taps, splices, and other cable
imperfections exist, which make the SSM worst-case model
even more inaccurate. As well, in practice, the number of
active users changes constantly and therefore in order to make
efcient use of the spectrum, it is necessary to dynamically
allocate transmit PSDs in order to maximize the achievable
bit-rate.
Fig. 2. Example of a typical xDSL Network.
Dynamic Spectrum Management (DSM) dynamically allo-
cates transmit powers for all users in response to changes in
channel conditions [3]. There are several physical constraints
imposed on the transmit powers for each user. One such
constraint is the maximum power which each user is allowed
to allocate over all of its frequency tones. The maximum power
constraint for user n is denoted by P
n
. Another constraint is
the maximum power which each user is allowed to allocate on
any particular frequency tone referred to as a spectral mask.
The spectral mask for user n on frequency tone k is denoted by
s
n,mask
k
. Therefore the power constraints can be summarized
as:
kK
s
n
k
P
n
and 0 s
n
k
s
n,mask
k
n, k. (1)
There are many different possible spectrum management
problems depending on the specic goal of the network. For
example, maximizing the total achievable data rate (i.e., Rate
Adaptive (RA)) or minimizing the total power allocated (i.e.,
Fixed Margin (FM))
1
while still ensuring some QoS require-
ments for each DSL customer. Regardless of the specic goal
of the network, the physical constraints shown in (1) are
always present.
One approach to DSM focuses on maximizing the achiev-
able data rate of one particular user (e.g., user 1), given that
the rest of the users satisfy some QoS requirement. The QoS
requirement for user n is denoted by R
n,target
. This approach
to DSM can be summarized in the following RA optimization
problem:
max
s
n
, nN
R
1
subject to: R
n
R
n,target
, n > 1 (2)
kK
s
n
k
P
n
, n
0 s
n
k
s
n,mask
k
, n, k.
1
Concepts of RA and FM are also discussed in a wireless communication
setting (e.g., [35]).
It was shown in [12] that the optimal solution to the RA
optimization problem (2) is equivalent to the optimal solution
of the RA optimization problem (3) for some w w
n
, n
^. w
n
is a weighting factor which represents the importance
of user n.
max
s
n
, nN
nN
w
n
R
n
subject to:

kK
s
n
k
P
n
, n (3)
0 s
n
k
s
n,mask
k
, n, k.
In the most general form, one can incorporate both RA and
FM into the spectrum management problem [26]. Let /
and T/ denote the index sets of the RA and FM users,
respectively. Assuming / ,= (i.e., there are some users
which are trying to maximize their rate) the joint RA and FM
spectrum management problem is shown below in (4). Note
that the rate for all of the users in / are being maximized
while the rate for all of the users in T/ are being xed at
their respective QoS requirement in order to minimize their
power consumption.
max
s
n
, nN
nRA
w
n
R
n
subject to: R
n
= R
n,target
, n T/ (4)
kK
s
n
k
P
n
, n ^
0 s
n
k
s
n,mask
k
, k /, n ^
If / = (i.e., no users are trying to maximize their rate)
the spectrum management problem reduces to the FM case.
This is shown below in (5). Note that / = ^ =
T/and the maximization problem changes to a minimization
problem. Here the total power consumed by all users on all
frequency tones is being minimized while ensuring each user
still meets their respective QoS requirements.
min
s
n
, nN
nN
kK
s
n
k
subject to: R
n
R
n,target
, n ^ (5)
kK
s
n
k
P
n
, n ^
0 s
n
k
s
n,mask
k
, k /, n ^
III. DSM ALGORITHMS UNDER CONSIDERATION
The eight DSM algorithms: OSB, ISB, ASB, IWF, SIW,
SCALE, DCA, and DSB are discussed. OSB is a centralized
algorithm which computes the globally optimal solution by
using N-dimensional exhaustive searches. ISB is a centralized
algorithms which attempts to approximate OSB by iteratively
performing 1-dimensional exhaustive searches.
ASB is an autonomous algorithm which uses the concept
of a reference line to maximize the achievable data rate.
IWF is an autonomous algorithm where every user selshly
maximizes their own transmit power, until the point is reached
where no user can benet from changing their transmit power.
SIW iteratively performs IWF, which allows for weaker users
to make more efcient use of the frequency spectrum.
SCALE is a hybrid centralized and distributed algorithm
since it is implemented in a distributed fashion by making
use of explicit message passing to gain centralized knowl-
edge. SCALE operates by successively applying convex lower
bounds to the objective function until a locally optimal point
is reached.
DCA is a centralized algorithm which makes use of a
Difference of Convex functions (DC) property in order to solve
for a locally optimal points in an iterative fashion.
DSB is a hybrid centralized distributed algorithm which
makes use of explicit message passing to gain centralized
knowledge. DSB is based on the necessary conditions for a
solution to be locally optimal.
A. OSB
As will be discussed in Section V, solving for the optimal
power allocation of all N users on each of the K frequency
tones using a brute force exhaustive search method has a
computational complexity of O(e
KN
).
OSB begins by re-writing the optimization problem shown
in (2) as the optimization problem shown in (3) [12], [18]. The
OSB algorithm then uses the concept of dual decomposition to
eliminate the exponential complexity in terms of the number of
frequency tones, K, and converts it into a linear complexity in
terms of the number of frequency tones, K. OSB converts the
constrained optimization problem in (3) into an unconstrained
optimization problem using duality theory. Therefore, the
Primal Problem, (3), is replaced by the dual problem
max
s
1
,...,s
N
L(s
1
, . . . , s
N
) (6)
where L(s
1
, . . . , s
N
)
n
w
n
R
n
n
s
n
k
.
The Lagrange multipliers (w
1
, . . . , w
N
,
1
, . . . ,
N
) are
chosen such that the KarushKuhnTucker (KKT) conditions
are satised:
n
_
P
n
k
s
n
k
_
= 0, n (7)
w
n
_
R
n
k
b
n
k
_
= 0, n. (8)
Provided that Equations (7) and (8) hold, the dual problem
(6) is equivalent to the primal optimization problem (3).
Since synchronous transmission is assumed, in Section II-B,
there is no ICI and therefore the Lagrangian can be decom-
posed over the frequency tones. This reduces the complexity
of the algorithm from exponential to linear in the number of
frequency tones, since the exhaustive search can be performed
independently on each frequency tone. Therefore, OSB re-
duces a KN-dimensional exhaustive search to K different
N-dimensional exhaustive searches.
Therefore, for each frequency tone k, the following opti-
mization problem must be solved:
max
s
1
k
,...,s
N
k
L
k
(s
1
k
, . . . , s
N
k
) (9)
where L
k
(s
1
k
, . . . , s
N
k
)
n
w
n
R
n
n
s
n
k
.
Since

k
L
k
(s
1
k
, . . . , s
N
k
) = L(s
1
, . . . , s
N
), solving (9) in-
dependently over each frequency tone is equivalent to solving
(6). The process of writing out the Lagrangian dual problem
and decomposing the Lagrangian across all the frequency
tones is known as dual decomposition.
The full OSB algorithm is outlined in Algorithm 1. One
important note is that the choice of can drastically effect the
performance of the algorithm. If is too small, the algorithm
will take an extremely long time to converge but if is too
large, the algorithm may not converge at all. Ideally one must
tweak the value of heuristically in order to ensure proper
convergence of the algorithm. This is often a highly non-trivial
task. One method for specifying the choice of is shown in
[36].
Algorithm 1: OSB Algorithm
repeat
foreach k do
Solve by N-Dimensional exhaustive search:
(s
1
k
, . . . , s
N
k
)=arg max
{s
1
k
,...,s
N
k
}
L
k
(s
1
k
, . . . , s
N
k
) ;
end
foreach n do
w
n
= maxw
n
+ (R
n,target
k
b
n
k
), 0 ;
n
= max
n
+ (
k
s
n
k
P
n
), 0 ;
end
until s
n
k
converges k, n ;
Spectral masks can easily be incorporated into Algorithm 1
by setting the value of L
k
to a very large negative number if
s
n
k
> s
n,mask
k
for each user n.
By reducing the complexity to linear in terms of the number
of frequency tones, OSB becomes computationally tractable
for large numbers of frequency tones; however, OSB is still
computationally intractable for large numbers of users. OSB
can therefore serve as an upper-bound with which to compare
the performance of other DSM algorithms for few users.
B. ISB
In Section III-A, it is shown that OSB is computationally
intractable for large N. This motivates the choice of an
iterative algorithm which is tractable for large N. ISB operates
in a similar fashion to OSB by re-writing the optimization
problem shown in (2) as the optimization problem shown in
(3) [18] [19]. ISB also uses the concept of dual decomposi-
tion, discussed in Section III-A, to re-write the optimization
problem as:
max
s
n
k
L
k
(s
1
k
, . . . , s
N
k
), (10)
where L
k
(s
1
k
, . . . , s
N
k
)
n
w
n
R
n
n
s
n
k
.
Note that in ISB, the PSD of each user is searched for
in an iterative fashion by updating one users PSD at a time
while keeping all other users PSD values xed. This can be
seen in Equation (10) since the maximization is only over a
single users PSD, s
n
k
. It is important to note that for ISB,
the KKT conditions shown in Equations (7) and (8) must still
be satised in order to ensure that the solution to the dual
problem (10) for all k, is equivalent to the solution of the
primal problem (3).
The full ISB algorithm is outlined in Algorithm 2. As
discussed in Section III-A, one method to select is described
in [36].
Algorithm 2: ISB Algorithm
Let
R
> 0 and
P
> 0 be given ;
repeat
foreach n do
repeat
foreach k do
Fix s
m
k
m ,= n ;
Solve by 1-Dimensional exhaustive
search: s
n
k
= arg max
s
n
k
L
k
(s
1
k
, . . . , s
N
k
) ;
end
w
n
= maxw
n
+
R
(R
n,target
k
b
n
k
), 0 ;
n
= max
n
+
P
(
k
s
n
k
P
n
), 0 ;
until w
n
,
n
converge ;
end
until s
n
k
converges k, n ;
Spectral masks can be easily incorporated into Algorithm 2
by setting the value of L
k
to a very large negative number if
s
n
k
> s
n,mask
k
for each user n.
While ISB cannot guarantee to nd optimal power alloca-
tions, it has been shown to lead to optimal power allocations
in many test cases, especially with few users. Unlike OSB,
ISB has a complexity which is quadratic in the number of
users and linear in the number of frequency tones; therefore,
it is computationally tractable for many users. Even though it
cannot guarantee to nd an optimal power allocation, it can
be used as a comparative measure since its performance is
known to be close to optimal.
C. IWF
The IWF algorithm [5] consists of performing water-lling
on each user iteratively until convergence is achieved. Since
the concept of water-lling is crucial in the IWF algorithm, it
will be presented prior to discussing IWF itself.
Water-lling is the process by which one user (denoted here
by n) attempts to maximize its rate capacity regardless of the
effect on other users. More specically, water-lling attempts
to solve the following optimization problem:
max
s
n
R
n
subject to:

kK
s
n
k
P
n
,
0 s
n
k
s
n,mask
k
, k.
Since the maximization problem is concave, it can be solved
by setting the derivative of its Lagrangian to 0. This results in
the spectrum update formula with a single unknown, , shown
in Equation (11).
Dene: (s
n
k
)
m=n
[h
n,m
k
[
2
s
m
k
+
n
k
[h
n,n
k
[
2
/
.
Then s
n
k
=
s
n,mask
k
if (s
n
k
)
> s
n,mask
k
,
0 if (s
n
k
)
< 0,
(s
n
k
)
otherwise.
(11)
From a maximization point of view, the power constraint
should be active in order to achieve the highest rate. This can
be achieved by varying . The water-lling algorithm using a
bisection search to nd is presented in Algorithm 3.
Algorithm 3: Water-lling Algorithm for user n
Let > 0 be given ;
Initialize PSD: s
k
= 0, k ;
Initialize
min
= 0,
max
= 1 ;
Update PSD using (11) with
max
as ;
while

k
s
k
> P do
max
= 2
max
;
Update PSD using (11) with
max
as ;
end
while 1 do
= (
max
+
min
)/2 ;
Update PSD using (11) with ;
if

k
s
k
> P then
min
= ;
else if P
k
s
k
then
Break ;
else
max
= ;
end
end
The IWF algorithm iteratively performs water-lling one
user at a time. The users continuously perform water-lling
in turn until a Nash Equilibrium (NE) point (a point where
no user can benet from changing its power allocation) is
reached. When applying rate constraints, each user adjusts
their water-lling process to ensure that they achieve their
target rate. When applying water-lling, if a user achieves a
rate higher than its target rate, that users allowable total power
is reduced. Similarly, if a user achieves a rate lower than its
target rate, that users allowable total power is increased. Note
that the allowable total power can never exceed the total power
constraint. Therefore, the target rate set for the users must be
feasible target rates for those users in order for the algorithm
to converge. This process is summarized in Algorithm 4.
Algorithm 4: Iterative Water-lling Algorithm
Let P > 0 be given ;
Let

P
n
= P
n
n ;
Initialize PSD: s
n
k
= 0, n, k ;
repeat
foreach n do
Perform water-lling on user n (Algorithm 3)
with

P
n
as total allowable power ;
if R
n
> R
n,target
then
P
n
=

P
n
P ;
else if R
n
< R
n,target
then
P
n
=

P
n
+ P ;
if

P
n
> P
n
then
P
n
= P
n
;
end
end
until s
n
k
converges k, n ;
D. SIW
SIW [11] makes use of the fact that IWF forces users
with better channels to restrict their allowable total power.
At the same time, it acknowledges that each user is limited
by a single water-lling level on each frequency tone. SIW
attempts to remove these restrictions by allowing some users
to perform another IWF process. More specically, at the end
of every IWF process, SIW allows users who did not use up all
their transmit power to perform another IWF process without
affecting the users who already used up their transmit powers.
This allows some users to increase their water-lling level on
some frequency tones, and does so without affecting other
users which gives performance benets over IWF. Therefore,
SIW lets all users use up their allowable total power. The SIW
algorithm is summarized in Algorithm 5.
The modied IWF algorithm (Algorithm 4) used by SIW
must guarantee that at least one user is using all of its
allowable power. This is done to ensure that at each outer
iteration of SIW, at least one user is removed from the new
IWF process. One way of doing so is by maximizing one
users rate while forcing all other users to meet their target
rates.
At the end of every modied IWF process, the SIW
algorithm looks for the users that have used up all of their
Algorithm 5: Selective Iterative Water-lling Algorithm
Let

^ = ^ ;
Let

/ = / ;
Let

P
n
= P
n
n ;
while

^ ,= and

/ ,= do
Perform modied IWF with

^,

/, and power
constraints

P
n
n ;
foreach n

^ do
if

k
s
n
k
= P
n
then
^ =

^n ;
foreach k

/ do
if s
n
k
,= 0 then
/ =

/k ;
end
end
end
end
foreach n ^ do
P
n
= P
n
kK\
K
s
n
k
;
end
end
allowable power and removes them from the next IWF round.
As well, the frequency tones used by those users are also
removed. The removal of the frequency tones is crucial in
preserving the removed users data rates. This process of
performing IWF, removing users and frequency tones and re-
performing IWF, is repeated until either all the frequency tones
or all the users have been removed.
One fall-back of SIW is that it requires a messaging system
to identify the removed users and tones.
E. ASB
Autonomous algorithms are distributed algorithms which
require no explicit message passing. The idea behind ASB is
that each user tries to maximize the throughput of the system
while ensuring that their performance is above some threshold
value. This is a more seless system and hence, the overall
performance would be expected to improve since each user is
no longer acting selshly.
The ASB algorithm [21] uses the concept of a virtual
reference line which attempts to mimic a typical victim line in
the interference channel. Each user then tries to minimize the
damage done to the reference line while ensuring that its own
target data-rate is met. The reference line is determined using
statistics of the network which are known a priori and hence
the ASB can be implemented autonomously. ASB makes use
of the fact that DSL crosstalk channel gains are slowly time-
varying and hence the reference line can represent a typical
victim line.
One choice for the reference line is the longest line within
the network, since the longest line tends to have the weakest
direct transfer function and therefore is most sensitive to
crosstalk. The only knowledge a modem needs is its own direct
channel, background noise and the distance from the CO to
the Remote terminal (RT, assuming it is RT distributed). All
these values can be measured locally or programmed at the
time the RT is installed. This allows for ASB to operate fully
autonomously during run-time.
From user ns point of view, the reference lines rate is:
R
n,ref
b
n
k
,
where
b
n
k
log
2
_
1 +
1
h
n, n
k
[
2
s
k
[
h
n,n
k
[
2
s
n
k
+
k
_
.
The crosstalk channel for the reference line, denoted [
h
n, n
k
[
2
and [
h
n,n
k
[
2
, are modeled using empirical models (ANSI
models) that were developed in the standards [37], [38], and
[39]. When using this model, the only parameters required to
determine the crosstalk channel values are the length of the
coupling line, offsets (network topology) and the cable widths,
which can be programmed in by the network operator when
the modem is installed. The reference line background noise,

k
, is set to the static line noise seen by the reference line.
Finally, s
n
k
is the transmit power of user ns reference line on
frequency tone k and s
n
k
is the transmit power of user n on
tone k.
In general, the concept of a reference line could be extended
to multiple reference lines. For each M reference lines, the
optimization problem has up to M + 1 local maximums
[21]. This increases the complexity but may also increase the
performance.
The ASB optimization problem is now formulated as fol-
lows: each user n solves a different version of the following
optimization problem:
max
s
n
R
n,ref
(12)
subject to: R
n
R
n,target
k
s
n
k
P
n
0 s
n
k
s
n,mask
k
, k.
For each user n, solving (12) requires optimizing its own
PSD s
n
k
which determines its own achievable rate (R
n
) and the
reference line (R
n,ref
). After each user solves the optimization
problem, (12), the crosstalk values of the network change.
Each user must then remeasure or estimate the new crosstalk
channel values and re-optimize using (12). This process is
repeated until the PSD values of all users converge.
The approach to solving (12) is identical to those shown
in Section III-A and Section III-B where the total number of
users is two (user n and user ns reference line). The concept
of dual decomposition is then applied as in the OSB and the
ISB case giving:
L
n
k
= w
n
b
n
k
+w
n,ref
b
n
k

n
s
n
k
,
where w
n,ref
is the weight of the reference line, which is set
to 1 w
n
when applying rate constraints. If rate constraints
are not used, the weight of the reference line is set to the
weight of the user that it is attempting to emulate (which is
known a priori).
The optimal PSD, s
n,
k
, is then,
s
n,
k
arg max
s
n
k
[0, s
n,mask
k
]
L
n
k
(w
n
, w
n,ref
,
n
, s
1
k
, . . . , s
N
k
). (13)
Equation (13) can be solved for by evaluating L
n
k
/s
n
k
=
0. Equation (13) is equivalent to
0 =
n
+
w
n
fs|h
n,n
k
|
2
log(2)
_
|h
n,n
k
|
2
s
n
k
..
+
P
i=n
|h
n,i
k
|
2
s
i
k
+
n
k
_
(14)
w
n,ref
fs|h
ref
k
|
2
s
ref
k
|h
ref,n
k
|
2
log(2)
_
|h
ref,n
k
|
2
s
n
k
+
ref
k
__
|h
ref
k
|
2
s
ref
k
+|h
ref,n
k
|
2
s
n
k
+
ref
k
_
.
Equation (14) can be simplied into a cubic equation in
terms of s
n
k
and hence has three roots that can be solved
for in a closed form. It is necessary to compare the value
of L
n
k
at each of the feasible roots as well as on the boundary
conditions, s
n
k
= 0 and s
n
k
= P
n
(if no spectral mask is used,
or s
n
k
= s
n,mask
k
if one is used). The value of s
n,
k
is chosen
to be a feasible PSD corresponding to the maximum value of
L
n
k
.
In order to reduce the computational time associated with
the PSD update (solving a cubic equation many times can be
time-consuming), [22] proposed an algorithm called ASB-2,
which uses an alternate PSD update based on Equation (14).
The ASB-2 algorithm re-arranges Equation (14) by isolating
for the s
n
k
term indicated by (). The other s
n
k
terms take on
the value of the previous iteration, denoted by (s
prev
)
n
k
. This
gives the following spectrum update formula:
s
n,
k
=
_
w
n
f
s
/ log(2)
n
+P
ASB2,n
k
i=n
[h
n,i
k
[
2
s
i
k
+
n
k
[h
n,n
k
[
2
_
s
n,mask
k
0
,
where [x]
b
a
minmaxx, a, b and
P
ASB2,n
k

w
n,ref
fs|h
ref
k
|
2
s
ref
k
|h
ref,n
k
|
2
/ log(2)
_
|h
ref,n
k
|
2
(sprev)
n
k
+
ref
k
__
|h
ref
k
|
2
s
ref
k
+|h
ref,n
k
|
2
(sprev)
n
k
+
ref
k
_
.
After solving for the optimal PSD value, user n updates
n
to enforce the total power constraint and w
n
to enforce the
target rate constraint. The algorithm presented uses a simple
bisection search to nd the Lagrange multiplier values. Users
iterate this process until all the users PSDs converge. The
complete algorithm is shown in Algorithm 6. The value of
the upper bound on
n
is denoted by
n
max
. It must be found
prior to beginning the
n
bisection. One method for computing
n
max
is by initializing it to 1 and increasing it (e.g., doubling)
until the power allocated is less than or equal to the total power
constraint. This will ensure a feasible solution for
n
exists.
For more information about ASB, interested readers are
referred to [40].
Thus far, the discussion regarding the ASB and ASB-2
algorithms only involved one reference line. As mentioned,
extending ASB to Multiple Reference Lines (MRL) provided
Algorithm 6: ASB Algorithm
Let
R
> 0,
P
> 0 be given ;
Initialize PSDs: s
n
k
= P
n
/K, n, k ;
repeat
foreach n do
Initialize w
n
min
= 0, w
n
max
= 1 ;
while [
k
b
n
k
R
n,target
[ >
R
do
w
n
= (w
n
max
+w
n
min
)/2 ;
Initialize
n
min
= 0,
n
max
=
n
max
;
while
[
k
s
n
k
P
n
[ >
P
and
k
s
n
k
P
n
do
n
= (
n
max
+
n
min
)/2 ;
s
n
k
= arg max
s
n
k
[0, s
n,mask
k
]
L
n
k
k ;
if

k
s
n
k
> P
n
then
n
min
=
n
;
end
otherwise
n
max
=
n
;
end
end
if

k
b
n
k
> R
n,target
then
w
n
max
= w
n
;
end
otherwise
w
n
min
= w
n
;
end
end
end
until s
n
k
converges k, n ;
a signicant increase in complexity (for every M reference
lines, the PSD update requires solving a polynomial of order
2M + 1 [22]); however, the ASB-2 algorithm can be easily
extended to MRL without a signicant increase in complexity.
As such, [23] extended the ASB algorithm using MRL
(ASB-MRL). It was shown that using MRL can provide
signicant performance increases, and can drastically reduce
the number of time-consuming interference measurements re-
quired to be taken. This provides a very signicant complexity
reduction for practical systems.
One issue regarding any ASB-based algorithm is the choice
of reference line(s). [23] presented three conditions in order to
ensure strong performance for ASB using MRL for a general
network topology. The three conditions include:
1) Together, the reference lines must have non-zero PSD
over every tone k / and must represent the general
characteristics of the DSL network.
2) Each reference lines total transmit power must represent
what an actual line with similar characteristics might
require.
3) Reference lines require other reference lines in order to
better reect the interference they perceive.
These conditions ensure that the reference lines are repre-
sentative of the actual network and therefore each user can
self-optimize with local information, as well as the reference
lines that act as approximations to global information in order
to gain performance improvements. The resulting algorithm
is still fully distributed; however, the use of MRL as side
information provides meaningful information to each user.
The reference lines are used to create a virtual network
representative of the original network, but consisting of M
reference lines and the local user. Therefore, when each
user self-optimizes, they solve an M + 1-user, K-frequency
tone optimization problem (i.e., solving for the optimal PSD
of the virtual network). Since the reference line parameters
and transfer functions are known locally, solving this virtual
network requires no additional time-consuming interference
measurements. As such, algorithms like DSB and SCALE,
which have been shown to give strong performance but require
many interference measurements (see Section IV-B) can be
used to optimize the virtual network, and therefore update
each users PSD. By selecting the reference line parameters
according to the three conditions outlined above, the M + 1-
user virtual network seems to approximate the true network
well. As such, approximate global knowledge can be obtained
locally.
The ASB-MRL algorithm referred to for the remainder
of this paper will be referring to the ASB-DSB algorithm
presented in [23]. The ASB-DSB algorithm is named after
the fact that every outer iteration is an ASB iteration, while
ever inner iteration is a DSB iteration. More specically,
suppose each user, n, has M reference lines. Each user self-
optimizes simultaneously (the ASB outer iteration); however,
when each user self-optimizes with its reference lines, the
DSB algorithm is run locally to solve the virtual network,
using only the characteristics of the reference lines and locally
available information (the DSB inner iteration).
F. SCALE
Successive Convex Approximation for Low complExity
(SCALE) was rst introduced in [25] and [26]. It solves
problem (3) by approximating it with a concave lower bound,
maximizing the approximation, and repeating the process
with another approximation. SCALE introduces a method of
distributing the required processing over multiple users.
SCALE uses the approximation log(z) + log(1 +z)
with z
0
/(1 +z
0
) and log(1 +z
0
) z
0
log(z
0
)/(1 +
z
0
). The inequality is tight when z = z
0
. By applying the
inequality to the total rate capacity equation and replacing the
variables s
n
k
with s
n
k
where s
n
k
= exp( s
n
k
), a concave equation
is obtained yielding a concave maximization problem:
max
s
k,n
_
w
n
n
k
log
2
_
[h
n,n
k
[
2
exp ( s
n
k
)/
m=n
[h
n,m
k
[
2
exp ( s
m
k
) +
n
k
_
+
n
k
_
,
which can be re-written as:
max
s
k,n
_
w
n
n
k
_
log
2
_
|h
n,n
k
|
2
_
+ s
n
k
(15)
log
2
_
m=n
[h
n,m
k
[
2
e
s
m
k
+
n
k
_
_
+
n
k
_
.
Any convex optimization tool can be used to solve Problem
(15), but SCALE proposes a gradient approach that can also
be distributed among each users by using the functions
Lagrangian:
L(s, ) =
n
_
k
e
s
n
k
P
n
_
+
n
w
n
n
k
_
log
2
_
[h
n,n
k
[
2
_
+ s
n
k
log
2
_

m=n
[h
n,m
k
[
2
e
s
m
k
+
n
k
_
_
+
n
k
.
By taking the Lagrangians gradient and setting it equal to
zero, the spectrum update formula shown in Equation (16) is
obtained:
s
n
k
=
w
n
n
k
n
+
m=n
|h
m,n
k
|
2
w
m
m
k
P
q=m
|h
m,q
k
|
2
s
q
k
+
m
k
. (16)
The denominator of the update equation, Equation (16),
contains information about other users. This additional infor-
mation allows the distributed algorithm to converge to a locally
optimal point. In order to gather this information, a message
passing system is required. Every user measures their total
interference and noise on every tone and transmits it back
to the Spectrum Management Center (SMC). Since the SMC
has partial channel knowledge, the message passing system
and update formula can be simplied as follows:
^
n
k
=
w
n
n
k
m=n
[h
n,m
k
[
2
s
m
k
+
n
k
/
n
k
=
m=n
[h
m,n
k
[
2
^
m
k
s
n
k
=
w
n
n
k
n
+/
n
k
. (17)
At every iteration, every user n calculates ^
n
k
on every
tone and sends it to the SMC. The SMC produces the /
n
k
values and distributes them to each user n. The full SCALE
algorithm is summarized in Algorithm 7.
G. DCA
DCA [27] and [41] solves problem (3) by iteratively writ-
ing the objective function as the difference of two convex
functions, f = g h, applying an afne minorization onto h
(multivariate rst order approximation), and applying standard
convex optimization techniques.
DCA is based on the local optimality conditions of
min
s
f(s) : s (, where ( represents the constraints set.
The general DCA approach makes use of the concept of sub-
differentials, which generalizes the derivative at points where
the objective function is non-differentiable; however, since for
DSM purposes the objective function is differentiable, this
paper uses derivatives instead of sub-differentials.
Starting from an initial point, s
(0)
(, two sequences
s
(k)
and q
(k)
are constructed such that, q
(k)
h(s
(k)
)
and s
(k+1)
g
(q
(k)
), where g
is the conjugate of g.
Algorithm 7: SCALE Algorithm
At each user ns modem :
Initialize PSD: s
n
k
= 0, k ;
Initialize
n
k
= 0, k ;
repeat
Receive /
n
k
from SMC ;
Update spectrum using (17) ;
At every m iterations, update
n
k
, k ;
Generate ^
n
k
and send to SMC ;
indenitely
At the SMC :
repeat
Receive ^
n
k
from every user ;
Generate /
n
k
and send to SMC ;
indenitely
The rst step of applying DCA to the DSM problem is to
re-write the RA optimization problem (3) in the form of a
minimization problem,
min
s
n
, nN
nN
w
n
R
n
subject to:

kK
s
n
k
P
n
, n (18)
0 s
n
k
s
n,mask
k
, n, k.
The next step is to write the objective function in terms of
the difference of two convex functions,
nN
w
n
R
n
=
nN
w
n
_
f
s
kK
b
n
k
_
=
n,k
w
n
f
s
log
2
_
1 +
1
|h
n,n
k
|
2
s
n
k
P
m=n
|h
n,m
k
|
2
s
m
k
+
n
k
_
=
n,k
w
n
f
s
log
2
_
P
m=n
|h
n,m
k
|
2
s
m
k
+
_
|h
n,n
k
|
2
/
_
s
n
k
+
n
k
P
m=n
|h
n,m
k
|
2
s
m
k
+
n
k
_
.
For simplicity, dene:
H
n,m
k

_
[h
n,n
k
[
2
/ if m = n
[h
n,m
k
[
2
if m ,= n
, s s
n
k
: n ^, k /
A
n
k
(s)

m=n
H
n,m
k
s
m
k
+
n
k
, B
n
k
(s) A
n
k
+H
n,n
k
s
n
k
.
The objective function of the DCA optimization problem
(18) can then be re-written as:
f(s) =
nN
w
n
_
f
s
kK
log
2
_
B
n
k
(s)
A
n
k
(s)
_
_
g(s) h(s),
where g(s) f
s
n, k
w
n
log
2
_
B
n
k
(s)
_
and h(s)
f
s
n, k
w
n
log
2
_
A
n
k
(s)
_
.
Using the fact that a summation of convex functions forms
a convex function, and the fact that the negative log-sum is
convex, and that w
n
and f
s
are positive, it follows that g(s)
and h(s) are also both convex functions. Therefore, f(s) =
g(s)h(s) is a DC function. There are many different possible
methods of applying DCA to solve this problem, one method
is discussed in the following section.
Consider the derivatives of g(s) and h(s).
g(s)
s
n
k
= f
s
n
w
n
ln(2)
H
n, n
k
A
n
k
(s)
(19)
h(s)
s
n
k
= f
s
n= n
w
n
ln(2)
H
n, n
k
B
n
k
(s)
(20)
Note that the denominators of Equations (19) and (20) re-
quire evaluating N-dimensional and (N-1)-dimensional sum-
mations respectively.
Since the sub-problem for the DC decomposition (f(s) =
g(s)h(s)) is highly nonlinear, one method to solve it involves
creating a new DC decomposition to simplify the sub-problem
calculation. One DC decomposition (as shown in [27] and
[42]) denes a new objective function as:
f(s)
1
2
[[s[[
2
_
1
2
[[s[[
2
f(s)
_
,
where > 0 to enforce convexity on g(s)
1
2
[[s[[
2
. It
is shown in [43] that if > |
2
g(s)|
, then

h(s)
1
2
[[s[[
2
f(s) is convex over the region of interest. Applying
DCA on

f(s) = g(s)

h(s) involves applying an afne
minorization (multi-variant rst order approximation) to

h(s),
which transforms the original general nonlinear optimization
problem into a quadratic programming optimization problem.
The benet of this DC decomposition (shown in [27] and
[42]) is that it reduces the problem from a general nonlinear
optimization problem into a simple quadratic optimization
problem, which is a very well-researched eld containing
many algorithms to nd solutions.
The algorithm for applying this method of DCA to

f(s) is
shown in Algorithm 8. On the r-th iteration, DCA approxi-
mates the function

h by its afne minorization (i.e., taking
q
r

h(s
r
)) which forces f to be convex. DCA then
minimizes the convex approximation using standard convex
optimization techniques (i.e., determining a point s
r+1
(q
r
)). It is shown in [43] that it is sufcient to nd a
point s
r+1
arg min
s
g(s) s, q
r
) : s R
p
.
Algorithm 8: DCA Algorithm
Initialize s
(0)
to be a best guess ;
Initialize step count to r = 0 ;
repeat
Calculate q
r
=
h(s
(r)
) = s
(r)
f(s
(r)
) ;
Calculate
s
r+1
= arg min
s
1
2
[[s[[
2
s, q
(r)
_
: s ( ;
r = r + 1 ;
until s
r
converges ;
H. DSB
The Distributed Spectrum Balancing (DSB) algorithm [22]
begins by writing out the objective function as a Difference of
Convex functions (DC) function and is based on applying the
KKT conditions of the DC problem directly. The optimization
problem for each user n is dened as:
max
s
n
1
,...,s
n
K
n,k
w
n
f
s
log
2
_
m
H
n,m
k
s
m
k
+
n
k
_
n,k
w
n
f
s
log
2
_

m=n
H
n,m
k
s
m
k
+
n
k
_
subject to:

k
s
n
k
P
n
, n (21)
0 s
n
k
s
n,mask
k
, n, k,
where H
n,m
k
is dened as in Section III-G, as follows:
H
n,m
k
= [h
n,m
k
[
2
if m ,= n and H
n,m
k
= [h
n,m
k
[
2
/ if m = n.
By incorporating the constraints into the Lagrangian func-
tion, taking its derivative and setting it to zero, the following
equation can be derived in order to satisfy the KKT conditions
(for all n and k):
0 =
n
+
w
n
f
s
[h
n,n
k
[
2
/ log(2)
[h
n,n
k
[
2
s
n
k
..
i=n
[h
n,i
k
[
2
s
i
k
+
n
k
+
m=n
w
m
f
s
[h
m,n
k
[
2
/ log(2)
p=m
[h
m,p
k
[
2
s
p
k
+[h
m,m
k
[
2
s
m
k
+
m
k
m=n
w
m
f
s
[h
m,n
k
[
2
/ log(2)
p=m
[h
m,p
k
[
2
s
p
k
+
m
k
. (22)
The spectrum update formula for DSB can be found by
isolating for the s
n
k
term indicated by () in Equation (22).
This gives the following spectrum update formula:
s
n
k
=
_
w
n
f
s
/ log(2)
n
+P
DSB,n
k
m=n
[h
n,m
k
[
2
s
m
k
+
n
k
[h
n,n
k
[
2
_
s
n,mask
k
0
,
(23)
where
P
DSB,n
k

m=n
w
m
f
s
[h
m,n
k
[
2
log(2)
_
1
int
n
k
1
rec
n
k
_
,
int
n
k

m=n
[h
n,m
k
[
2
s
m
k
+
n
k
,
rec
n
k
int
n
k
+[h
n,n
k
[
2
s
n
k
.
The DSB algorithm makes use of some message pass-
ing in order to evaluate the spectrum update formula. Each
user n measures the values for int
n
k
and sends the message
_
1
int
n
k
1
rec
n
k
_
to the SMC. The SMC then computes P
DSB,n
k
for
all n and sends each user their corresponding P
DSB,n
k
. Each
user can then update their spectrum using Equation (23). This
process is summarized in Algorithm 9. The DSB algorithm
loops until the PSD of each user has converged.
The DSB algorithm has been shown to be quite powerful
but since it is a local optimizing algorithm it can be quite
sensitive to the choice of initial point. It is for this reason
that the DSB algorithm was extended in [22] to the Multiple
Starting point DSB (MS-DSB). For each frequency tone, the
Algorithm 9: DSB Algorithm
At each user ns modem :
Initialize PSD: s
n
k
= s
n,mask
k
/2, k ;
repeat
Receive message P
DSB,n
k
from SMC ;
Update spectrum using (23) ;
Generate message
_
1
int
n
k
1
rec
n
k
_
and send to SMC ;
indenitely
At the SMC :
repeat
Receive message
_
1
int
n
k
1
rec
n
k
_
from every user ;
Generate P
DSB,n
k
for all users and send to SMC ;
until s
n
k
has converged N ;
Fig. 3. Network Topology.
MS-DSB algorithm tries multiple initial points, runs a few
iterations with each starting point and selects the starting point
corresponding to the best so far (based on the few iterations).
This increases the probability (on a per-tone basis) that the
locally optimal solution found by the algorithm is in fact the
globally optimal one.
The authors in [22] proposed testing N +2 different initial
transmit spectrum for each frequency tone. Specically, the
vector whose components are all zero, s
n
k
= s
n,mask
k
/2 and
e
n
s
n,mask
k
, where e
n
is the vector of all zeros and a one
in the n-th location. Intuitively, these correspond to no users
transmitting, every user transmitting evenly, and each user
transmitting while no other user does (for all N users). After
selecting the initial PSD for each frequency tone, the standard
DSB algorithm is applied (with the new initial PSD).
IV. ILLUSTRATIVE EXAMPLES
It has been shown that in symmetric or frequency-at
environments, many DSM algorithms perform near-optimally.
Therefore the non-symmetric frequency selective near-far de-
ployment, which is more common in practice, is of interest.
This section will discuss the performance (achievable data
rate) for various DSM algorithms in a typical near-far de-
ployment.
A. 2-User Scenario
One way to visualize the performance of different DSM
algorithms is to construct (or plot) and compare the rate
regions for each algorithm. The rate region is a plot where
each axis represents the rate achieved for a different user in
the system. Therefore, the rate region of an algorithm shows
all the possible achievable data rate combinations of the users
that can be achieved. In general, for N users this would be
0 2 4 6 8 10 12 14 16 18
140
120
100
80
60
40
20
0
Frequency (MHz)
T
r
a
n
s
f
e
r

F
u
n
c
t
i
o
n

(
d
B
)

H11
H21
H12
H22
Fig. 4. Two-User Test Case.
an N-dimensional plot; therefore, the two-user case allows
for easier visualization. In the two-user case, the rate region
is a 2-dimensional plot of the achievable rate of user 1 vs.
the achievable rate of user 2. This plot is represented by a
curve. The further away the curve is from the origin, the
higher the achievable rate of the algorithm and therefore the
better the performance. Points operating on the boundary of
the rate region are considered to be optimal points. The goal
of DSM algorithms is to generate points which are closest to
the optimal values, while operating at a low computational
complexity. The concepts extend naturally to the multi-user
case (as will be seen in Section IV-B).
This section uses a non-symmetric frequency-selective net-
work topology. The test case considers an upstream transmis-
sion only. The rst line (user 1) runs 1500 ft from user 1 to the
Central Ofce (CO), and the second line (user 2) runs 3000 ft
from user 2 to the CO, as shown in Figure 3. For this upstream
scenario, the transmitting modems (TX) are located at the
users premises and the receivers (RX) are located at the CO.
Each DSL modem transmits over multiple frequency tones.
Note that in a general setting both upstream and downstream
transmission occurs. For completeness, more general scenarios
are presented in Section IV-B.
The test case represents a realistic upstream VDSL scenario
[37]. The test case assumes that 26-gauge (0.4 mm) lines are
used. The target symbol error probability is set at 10
7
. The
coding gain and noise margin are set to 3 dB and 6 dB,
respectively. The frequency tone spacing is
f
= 4.3125 kHz
and the DMT symbol rate is f
s
= 4 kHz. The FDD band plan
998 [44] is used, which consists of two separate upstream
transmission bands: 3.75 5.2 MHz and 8.5 12 MHz. The
optional 30 138 kHz band is not used. A maximum transmit
power of 11.5 dBm is applied to each modem for each DSM
algorithm.
The direct and the Far-End crossTalk (FEXT) transfer
functions for the two-user test case are shown in Figure 4.
User 1 represents the 1500 ft line and user 2 represents the
3000 ft line. Therefore, H11 and H22 are user 1 and user
2s direct transfer functions, respectively. As well, H21 is the
0 10 20 30 40 50 60
0
2
4
6
8
10
12
User 1 Rate (Mbps)
U
s
e
r

2

R
a
t
e

(
M
b
p
s
)

Flat PBO
IWF
ASB
SIW
OSB \ ISB \ DCA \ SCALE \ DSB
Fig. 5. Two-User Rate Regions.
FEXT transfer function from user 1 to user 2, and H12 is
the FEXT transfer function from user 2 to user 1. Figure 4
shows the transfer function over all 4096 frequency tones;
specically, the bolded parts of the curves are the upstream
bands used for DSM.
Simulations were run for: OSB, ISB, ASB, IWF, SIW,
SCALE, DCA, DSB and Flat PBO (a SSM technique (dis-
cussed in Section II-C)). The results are summarized in Figure
5. It is important to note that because of the simplicity of the
two-user channel, without the presence of alien crosstalk, the
rate regions are more optimistic than usual. The test case used
has few locally optimal points and therefore SCALE and DCA
are able to converge to globally optimal points. In general,
this is not the case, as with more interfering users the number
of locally optimal points grows signicantly, which leads to
reduced performance in locally optimizing algorithms (e.g.,
DCA, DSB, and SCALE). As such, the algorithms must be
compared in a multi-user setting to distinguish between the
stronger algorithms, which is shown in Section IV-B.
Based on the performance of the SSM algorithm (Flat
PBO) it is evident that the performance benets of DSM are
signicant. In more complicated channels, where the number
of users increases, the performance of Flat PBO (and all SSM
techniques) is even worse as compared to the DSM techniques.
B. 25-User Scenarios
In order to better compare the performance of the DSM
algorithms, a multi-user setting is required. A 25-user scenario
is very practical since while binders might contain up to 200
lines, they generally contain sub-bundles of 25 users each.
DSM can be applied to each sub-bundle of 25 users separately
while treating the interference of the other bundles (alien
crosstalk) as background noise.
As well, due to the large complexity of some of the
centralized algorithms, they are not practical for a 25-user
scenario. The focus then becomes on distributed (e.g., IWF,
ASB, ASB-MRL, SIW) and semi-centralized algorithms (e.g.,
SCALE, DSB).
Fig. 6. 25-User Test Case.
TABLE I
USER DISTRIBUTIONS FOR THE 25-USER SCENARIOS.
25 User Scenario - A 25 User Scenario - B
# of CO users 12 9
# of RT #1 users 13 8
# of RT #2 users 0 8
The network topology used for the 25-user scenarios is
shown in Figure 6. The lengths L
1
, . . . , L
N
are random
variables uniformly distributed over the interval [0, 1000] feet.
PSD masks were applied using VDSL Prole 17a band plan
[45], which consists of three downstream bands (0.276 3.75
MHz, 5.2 8.5 MHz, and 12 17.664 MHz) and two upstream
bands (3.75 5.2 MHz and 8.5 12 MHz).
Two 25-user test cases were conducted. Each of these test
cases consisted of both upstream and downstream transmis-
sion. The user distributions for the two test cases, 25 User
Scenario - A and 25 User Scenario - B are summarized in
Table I.
One hundred network realizations were simulated and each
of IWF, ASB-2, ASB-MRL, SCALE, DSB and ISB solved
the RA optimization problem (3) with w
n
= 1 n. This
avoids the use of target rates, which speeds up simulation
time. Each algorithm was run until the sum of the L-2 norms
(or Euclidean norms) of all the users PSDs converged to
within 1E-10 of the previous iteration (or a maximum of
1000 iterations was reached). The ASB-2 algorithm was only
run to an accuracy of 1E-8 since it was observed that in a
large number of realizations, convergence was not achieved
and instead the algorithm toggled between several points.
Due to the fact that the SIW algorithm requires target rates
to operate, it was omitted from these comparisons. As well,
the limitations of SIW was shown in the two-user test case
(Section IV-A). DCA was omitted since the algorithm cannot
be applied with PSD masks, which is an important physical
constraint in practical DSL systems.
Since the choice of reference line for an arbitrary network
topology is still an open question, the ASB (respectively ASB-
2) results vary depending on the choice of reference line. For
downstream transmission, the reference line of each user was
selected as the longest possible CO line (i.e., a 2000-ft CO
line). For the upstream transmission, the reference line was
chosen as the longest possible RT user which is most affected
by crosstalk (i.e., a 2000-ft RT #1 line for Scenario A and a
1500-ft RT #2 line for Scenario B).
The ASB-MRL algorithm was simulated using two pairs
of reference lines for each RT and CO subgroup. More
specically, for each RT and CO subgroup, four reference
lines were selected (2 1000-ft line and 2 2000-ft line).
The simulation results are based on the ASB-DSB approach
outlined in [23]; as well, the reference PSDs were generated
by performing single-user water-lling assuming no crosstalk.
The PSDs of the stronger reference lines (most signicant
interferers) were then scaled back to reect a more accurate
representation of the PSD of each user.
For the downstream 25 User Scenario - A, the RT reference
lines reduced their PSDs by 10 dB. For the upstream 25 User
Scenario - A, the CO reference lines reduced their PSDs by
20 dB and the RT reference lines reduced their PSDs by 10
dB.
For the downstream 25 User Scenario - B, the CO and
RT #1 reference lines reduced their PSDs by 10 dB and the
RT #2 reference lines reduced their PSDs by 20 dB. For the
upstream 25 User Scenario - B, the RT #1 and RT #2 reference
lines reduced their PSDs by 10 dB and the CO reference lines
reduced their PSDs by 20 dB.
The PSDs were scaled by either 10 dB, 20 dB or not at
all, since these quantities heuristically seemed to work fairly
well. It is important to note that while some PSD scaling is
necessary, the relative scaling of the reference lines is very
signicant (e.g., strong reference lines should have higher
PSD scaling than weak reference lines). The sensitivity to the
choice of scaling for an arbitrary network still needs to be
investigated.
The 25-user simulations are intended to distinguish between
the various high-performance (in terms of both achievable
rate and runtime) distributed algorithms in a realistic scenario.
The ISB algorithm was included as a performance measure of
how well the distributed algorithms (with local knowledge of
the network) performed compared to a near-optimal algorithm
(with global knowledge of the network).
1) Scenario - A: A summary of the results for 25 User
Scenario - A is shown in Table II. The sum rate (objective
function value) is the sum of the rates of all 25 users (since
w
n
= 1 n). The average number of iterations refers to
the outer iterations of the algorithm, and is also equal to
the number of interference measurements required to take
when using simultaneous spectrum updates for each user. It is
important to note that the time per iteration of IWF, ASB-2,
ASB-MRL, DSB, and SCALE are all very small and therefore
the main constraint in terms of practical runtime is the
number of iterations, since taking interference measurements
can be much more time-consuming than the runtime of each
iteration (e.g., one set of channel measurements over the 4096
frequency tones is on the order of minutes using a typical
VDSL2 DSLAM)
2
. The total power used refers to the total
power consumed across the 25 users.
Based on the results in Table II, it is clear that in both
the downstream and upstream scenarios, DSB, SCALE and
ASB-MRL achieve the best overall performance. For the
downstream scenario, SCALE provides a 0.4% performance
increase over DSB, while still requiring fewer iterations (num-
ber of time-consuming interference measurements required),
on average. In the upstream scenario, both algorithms provide
the same performance but DSB requires fewer time-consuming
2
Based on channel measurements taken in the Broadband Communications
Research Lab at McGill University
TABLE II
SUMMARY OF 25 User Scenario - A.
Downstream
Avg. Avg. Avg. Avg.
Sum Rate Runtime Number of Total Power Used
(Mbps) (s) Iterations (dBm)
IWF 783.87 0.17 11.9 25.48
ASB-2 810.76 0.58 22.0 24.50
ASB-MRL 858.15 0.85 13.4 22.96
DSB 858.81 7.67 380.0 22.62
SCALE 862.36 4.99 317.7 22.10
ISB 847.62 hours 47.9 22.89
Upstream
Avg. Avg. Avg. Avg.
IWF 195.32 0.05 6.8 25.48
ASB-2 247.54 0.04 2.9 12.09
ASB-MRL 264.39 0.20 7.2 21.96
DSB 267.75 1.93 251.6 20.48
SCALE 267.75 1.97 293.7 20.36
ISB 243.09 hours 59.4 20.36
interference measurements while SCALE requires less total
power consumption, on average.
Table II also shows that ASB-MRL achieved strong per-
formance while maintaining a low number of iterations. In
particular, the average performance of ASB-MRL was within
1% of the average performance of DSB and SCALE in
the downstream scenario but it required an average of 13.4
iterations, as opposed to more than 300 iterations required
for DSB and SCALE. Similarly, in the upstream scenario, the
average performance of ASB-MRL was within 1.3% of the
average performance of DSB and SCALE but it only required
an average of 7.2 iterations, compared to more than 200
iterations required for DSB and SCALE. The reduction of 200-
300 time-consuming interference measurements demonstrates
that ASB-MRL can provide a signicant practical complexity
reduction over semi-centralized algorithms like SCALE and
DSB, while operating fully autonomously and maintaining
strong overall performance.
As well, the performance of the centralized ISB algorithm
for both the downstream and upstream scenarios was inferior
to that of SCALE and DSB, on average. For the downstream
scenario, in 67% of realizations the sum rate of ISB was
superior to that of SCALE and DSB. The largest sum rate
increase in ISB over SCALE and DSB was 8.4 Mbps (1%
performance increase); however, the largest sum rate decrease
by ISB over DSB and SCALE was 76.6 Mbps (8.6% per-
formance decrease). Even though the sum rate achieved by
ISB was superior to that of DSB and SCALE in 67% of
the realizations, the sum rate in the remaining 33% of the
realizations was considerably lower than that of DSB and
SCALE. This resulted in the average sum rate of ISB being
smaller than that of DSB and SCALE. For the upstream
scenario, the performance of ISB was inferior to that of
SCALE and DSB for all network realizations, resulting in an
average sum rate of approximately 24 Mbps (9.2%) lower.
It is worthwhile to note that the performance of ISB was
often stronger than DSB and SCALE in the downstream
scenario; however, at times the ISB algorithm yielded a
signicantly worse sum rate. This could be due to the fact that
when the ISB algorithm was implemented, the PSD of each
user was updated in turn from user 1 to user 25, regardless
of the network realization. The order in which the users
TABLE III
INITIAL BEHAVIOR FOR 25 User Scenario - A.
Downstream
Average Sum Rate (Mbps) for Iteration #
1 2 3 Final
ASB-MRL 856.61 858.30 858.10 858.15
DSB 765.81 795.93 811.62 858.81
SCALE 557.20 670.92 712.54 862.36
Upstream
1 2 3 Final
ASB-MRL 263.11 264.35 264.40 264.39
DSB 205.65 221.86 232.18 267.75
SCALE 148.64 187.72 204.24 267.75
sequentially update their PSD can have a crucial impact on the
nal achievable rate. In order to apply ISB more appropriately,
the algorithm would need to be re-run using various different
PSD update orders and selecting the maximum value.
For the sake of time, the ISB algorithm was only run for
the one sequence of PSD updates for each network realization.
These results suggest that for the realizations where this
sequence of PSD updates was suitable, the sum rate achieved
by ISB was superior to DSB and SCALE (on average by
4 Mbps or 0.4%); however, for the realizations where this
sequence of PSD updates was not suitable, the sum rate
achieved by ISB was signicantly inferior to DSB and SCALE
(on average 40 Mbps or 5.1%).
Table II shows that the average performance of IWF is infe-
rior to DSB and SCALE in terms of the total sum rate and the
total power consumed; however, it requires signicantly fewer
iterations which means it requires signicantly fewer costly
interference measurements, on average. The performance of
IWF in the downstream scenario is signicantly better than in
the upstream scenario (relative to the other DSM algorithms).
More specically, IWF performs at 90.9% of the rate of
SCALE in the downstream scenario, while it can only perform
at 72.9% of the rate of DSB and SCALE in the upstream
scenario.
Table II shows that when running ASB-2 to an accuracy
of 1E-8, signicant performance improvements can be seen.
In particular, ASB-2 achieves 94.0% of the SCALE rate in
the downstream scenario and 92.5% of the SCALE rate in
the upstream scenario. This performance is achieved at a
signicant complexity reduction in terms of the number of
iterations. The main issue regarding the ASB-2 algorithm is
its limiting behaviour.
Beyond what is shown in Table II, when ASB-2 was run
to an accuracy of 1E-9, its downstream sum rate dropped by
167 Mbps (20.6%) and its upstream sum rate dropped by 102
Mbps (41.4%). Similar results were obtained when ASB-2 was
run to an accuracy of 1E-10 (i.e., a 22.7% decrease in the
downstream scenario and a 49.9% decrease in the upstream
scenario, as compared to an accuracy of 1E-8). Regardless
of the accuracy on ASB-2, the maximum sum rate was often
achieved on the rst or rst few iterations; afterwards, the
performance of ASB-2 rapidly decreased as the number of
iterations increased.
Similarly, the average rst iteration of ASB-2 for the down-
stream scenario yielded a sum rate of 844.01 Mbps (104.1%
of the converged 1E-8 result). For the upstream scenario,
the average number of iterations before convergence when
TABLE IV
SUMMARY OF 25 User Scenario - B.
Downstream
Avg. Avg. Avg. Avg.
IWF 811.90 0.11 6.1 25.48
ASB-2 868.23 0.35 15.7 23.29
ASB-MRL 932.51 3.42 38.6 21.06
DSB 940.72 8.11 392.5 22.50
SCALE 942.09 5.11 287.6 20.91
ISB 932.33 hours 60.93 22.82
Upstream
Avg. Avg. Avg. Avg.
IWF 217.40 0.05 6.7 25.48
ASB-2 221.91 0.28 26.9 18.03
ASB-MRL 298.45 0.27 7.2 22.54
DSB 304.52 1.51 184.9 20.66
SCALE 304.42 1.76 255.6 20.59
ISB 303.78 hours 52.46 20.80
running an accuracy of 1E-8 was 2.9. As such, the average
third iteration of ASB-2 yielded a similar sum rate of 247.98
Mbps (100.1% of the converged 1E-8 result) as the average
converged sum rate.
Due to the limiting behavior of ASB-2, care must be taken
to not over-run the algorithm, since running too many
iterations could drastically reduce the performance. As such,
it is important to be aware of the potential benets of ASB-2
but even more so, the potential dangers of using it.
The ideal algorithm would achieve strong performance in
very few iterations, while at the same time ensuring perfor-
mance increases (or does not signicantly decrease) as the
number of iterations increases. DSB, SCALE and ASB-MRL
(with appropriately chosen parameters) all exhibit this type
of behavior. A summary of the sum rate for the rst three
iterations and the nal sum rate for DSB, SCALE and ASB-
MRL is shown in Table III.
Table III shows that while SCALE achieves the largest
nal sum rate, it approaches its nal point slowly. ASB-MRL
approaches its nal sum rate extremely quickly, on average
converging to within 99% of the nal sum rate in only a
couple of iterations (both upstream and downstream). For the
downstream scenario, DSB and SCALE on average achieved
94.5% and 82.7% of their nal sum rate after three iterations,
respectively. For the upstream scenario, DSB and SCALE on
average achieved 86.7% and 76.3% of their nal sum rate after
three iterations, respectively.
It is interesting to note that while Table II shows SCALE
requires fewer iterations on average than DSB in the down-
stream scenario (317 compared to 380), Table III shows that
DSBs initial steps are much closer to the nal sum rate than
that of SCALE.
Overall, for the 25 User Scenario - A, ASB-MRL achieved
the best balance between maximizing the sum rate, while
requiring fewer iterations to converge, for both the upstream
and downstream scenarios.
2) Scenario - B: A summary of the results for 25 User
Scenario - B is shown in Table IV.
Based on the results of Table IV, it is clear that DSB,
SCALE and ASB-MRL are the highest performing non-
centralized algorithms. For both the upstream and downstream
scenarios, DSB and SCALE provide roughly the same perfor-
mance but SCALE required fewer iterations in the downstream
scenario, while DSB required fewer iterations in the upstream
scenario. For both scenarios, SCALE consumed less total
power than DSB.
Table IV also shows that the average performance of ASB-
MRL was within 1.0% of the average performance of DSB
and SCALE achieved in the downstream scenario; as well,
the average performance of ASB-MRL was within 1.9% of
the average performance of DSB and SCALE achieved in the
upstream scenario. For both scenarios, ASB-MRL required
substantially fewer iterations to achieve convergence than DSB
or SCALE (i.e., 38.6 compared to 392.5 and 287.6 in the
downstream scenario and 7.2 compared to 184.9 and 255.6 in
the upstream scenario).
As shown in Table IV, the performance of the centralized
algorithm ISB for both the downstream and upstream scenarios
was inferior to that of SCALE and DSB, on average. For
the downstream scenario, 49% of the network realizations
simulated resulted in a higher sum rate for ISB than for both
DSB and SCALE. The largest sum rate increase by ISB over
DSB and SCALE was 7 Mbps (0.8% performance increase);
however, the largest sum rate decrease by ISB over DSB
and SCALE was 80 Mbps (8.7% performance decrease). In
general, it was observed that in the 49% of realizations where
ISB was superior, the performance gain was minor (on average
0.2%); whereas, in the 51% of realizations where ISB was
inferior, the performance drops were more severe (on average
2.1%). This resulted in an average sum rate for ISB below
that of DSB and SCALE. In the upstream scenario, ISB was
inferior to DSB and SCALE for each network realization, but
since the achieved values were fairly close for each algorithm,
the average sum rates achieved by all three are nearly identical.
As discussed in Section IV-B1, the order in which each
users PSD was updated can greatly affect the results for a
given realization and could explain why for some realizations
ISB achieved strong performance, but for other realizations
ISB achieved poor performance. For the downstream scenario,
in the 49% of realizations where ISB achieved better perfor-
mance than DSB and SCALE, the average sum rate of ISB was
only approximately 2 Mbps larger (less than 1% performance
improvement).
Table IV shows that the performance of IWF is much
stronger in the downstream scenario (86.3% of the DSB
sum rate) than the upstream scenario (71.4% of the DSB
sum rate). Similar to the Section IV-B1 results, IWF requires
fewer iterations to achieve convergence than the other DSM
algorithms for both the upstream and downstream scenarios.
Beyond what is shown in Table IV, the limiting performance
of ASB-2 was very poor. More specically, when ASB-2 was
run to an accuracy of 1E-10, the sum rate achieved was 549.74
Mbps (63.3% of the 1E-8 results) in the downstream scenario
and 101.68 Mbps (45.8% of the 1E-8 results) in the upstream
scenario.
A summary of the sum rate for the rst three iterations and
the nal sum rate for DSB, SCALE and ASB-MRL is shown
in Table V.
Table V shows that ASB-MRL approached its nal sum rate
extremely quickly, on average converging to within 99% of the
TABLE V
INITIAL BEHAVIOR FOR 25 User Scenario - B.
Downstream
1 2 3 Final
ASB-MRL 932.00 933.30 933.01 932.51
DSB 795.85 833.97 854.25 940.72
SCALE 617.39 742.63 790.92 942.09
Upstream
1 2 3 Final
ASB-MRL 297.02 298.54 298.51 298.45
DSB 232.50 253.77 266.73 304.52
SCALE 147.78 194.56 211.76 304.42
nal sum rate in only a couple of iterations (both upstream and
downstream). For the downstream scenario, DSB and SCALE
achieved 90.8% and 84.0% of their nal sum rate after three
iterations, respectively. For the upstream scenario, DSB and
SCALE achieved 87.5% and 69.6% of their nal sum rate
after three iterations, respectively.
Overall, for the 25 User Scenario - B, ASB-MRL achieved
the best balance between maximizing the sum rate, while
ensuring fewer iterations to converge, for both the upstream
and downstream scenarios.
V. COMPLEXITY
Another signicant factor in choosing a DSM algorithm
is the run-time (or implementation time). The run-time of
different algorithms depends on many factors including the
processor(s) used, the coding used, the programming lan-
guage used, etc. Therefore, the performance metric used in
this section to compare the DSM algorithms will be their
Computational complexity represents the number of op-
erations required for the execution of an algorithm. More
specically, it shows the behavior of different algorithms
as the input sizes increase. In order to represent this, O()
notation is used. It describes the limiting behavior of a function
when the arguments tend towards innity. This is described
mathematically as follows:
f(x) = O(g(x)) [f(x)[ c[g(x)[ x as x ,
where c R is some constant.
Therefore, for an arbitrary algorithm dependent on some
variable X, a O(X) algorithm will converge faster than a
O(X
2
) algorithm for sufciently large X. The convergence
time of an algorithm is said to be exponential if it is O(e
X
),
and is said to be polynomial if it is O(X
a
) for a > 1. It is
important to note that in algorithmic terms, exponential time
is extremely slow.
For DSM purposes, there are two main inputs (variables)
of interest: the number of users, N, and the number of
frequency tones, K. Hence, all algorithms will be of the form
O(f(N, K)) for some function f(N, K). The total power
and target data rate constraints, shown in Equation (2), cause
coupling over the frequency tones, which requires that each
users PSD be searched for jointly over all frequency tones.
Solving an X-dimensional exhaustive search has a compu-
tational complexity of O(e
X
). Therefore, solving the DSM op-
timal resource allocation through an exhaustive search in both
the number of users, N, and the number of frequency tones,
K, would lead to a computational complexity of O(e
NK
),
which is computationally intractable.
In Section I, the concept of centralized versus distributed
(and autonomous) systems was introduced. One key difference
is in how the total interference plus noise is found. For
autonomous algorithms, the total interference plus noise is
measured locally by each user. For centralized algorithms,
since the DSL channel is generally time-invariant, channel
measurements can be taken a priori and stored at the CO (these
measurements would need to be updated periodically). Hence,
full channel knowledge is assumed (i.e., [h
n,m
k
[
2
is known
n, m ^, k /). Depending on the implementation,
distributed algorithms can either operate fully autonomously
or in a semi-centralized manner by making use of explicit
message passing.
Operating in a centralized manner may require the time-
consuming computation of the interference plus noise for each
user on each frequency tone in the system. More specically,
to compute the total interference plus noise for some user n
on a given frequency tone k, one must cycle through all the
users m ,= n and multiply the channel transfer function by
user ms transmit PSD for the specic frequency tone k and
sum them. Hence, in order to compute the total interference
plus noise for all N users on some frequency tone k, the total
complexity will be:
N (N 1) = N
2
N.
Therefore, the computational complexity of calculating the
total interference plus noise for all N users on some frequency
tones k will be:
C
Interference
= O(N
2
).
In addition, computing the bit-rates of all N users on
some frequency tone k is dominated by computing the total
interference plus noise and therefore, is also O(N
2
). Note
that in order to compute the bit-rates of all N users on all
K frequency tones, the computational complexity will be
O(KN
2
).
A. IWF Complexity
Iterative Water-Filling (IWF) is a distributed algorithm
briey described in Section III-C (Algorithm 4). IWF consists
of two loops where the outer loop cycles through the inner
loop for all users. The inner loop performs water-lling for
the n-th user over all the K frequency tones. The water-lling
process involves a bisection search for that users Lagrange
multiplier, ; Assuming that the number of bisection iterations
required for the PSD to converge is at most v
1
and assuming
that the number of iterations required for the IWF algorithm
to converge is v
2
, the total complexity of the IWF algorithm
will be v
1
v
2
N K, which implies:
C
IWF
= O(KN).
It is important to note that for IWF, whenver a user is
required to perform water-lling, that user must re-measure
its total interference plus noise. Therefore, for N users and v
iterations, a total of vN measurements must be performed. If
the channel conditions are poor, the total number of iterations,
v, may be large. Therefore, if the values of N and v are large,
taking vN measurements may be a bottleneck in terms of run-
time.
For such situations, IWF can be implemented in a central-
ized manner. As explained previously in Section V, since the
DSL channel is very slow time-varying, channel measurements
can be taken a priori and stored at the CO. The CO can then
compute the total interference plus noise seen by user n in a
centralized manner, which requires O(KN) operations. The
CO runs a virtual IWF game to generate the new PSD values.
More specically, on user ns turn in the virtual game, the CO
computes the total interference plus noise seen by user n and
then updates user ns virtual PSD. The CO plays this virtual
game until all of the users PSDs converge, at which point the
CO informs each user of their new PSD value.
Implementing IWF in a centralized manner, using the same
denitions as for the distributed manner, would require a total
complexity of N (KN +v
1
v
2
K), which implies:
C
IWF
(centralized) = O(KN
2
).
Note that C
IWF
(centralized) has a higher complexity than
C
IWF
since it performs more computations during every
iteration. In practice, the run-time between the two methods
can vary greatly depending on the number of users in the sys-
tem, as well as the channel conditions, which both affect the
number of iterations and interference measurements required.
B. SIW Complexity
The Selective Iterative Water-lling (SIW) algorithm (de-
scribed in Section III-D, Algorithm 5) performs two loops
where the outer loop cycles through the inner loop for all
users where the inner loop performs IWF for all users. At
every iteration of the outer loop, users (and their corresponding
frequency tones) that have already achieved their maximum
power are removed. It has been shown that each sequential
inner loop removes at least one user and frequency tone. This
process is repeated until either all the users have been removed
from the inner loop (maximum of N iterations), or until all
frequency tones have been removed (possibly in less than
N iterations). Therefore, the maximum number of additional
outer loop iterations is N, which leads to a total complexity
of up to N times that of IWF, which implies:
C
SIW
= O(KN
2
).
C. OSB Complexity
The Optimal Spectrum Balancing (OSB) Algorithm is de-
scribed in Section III-A (Algorithm 1). Assuming that the
number of discrete bit-loadings is X, then the search space
for all users PSDs will contain X
N
elements. Therefore, to
solve for the optimal PSD for each user, an exhaustive search
must be performed over the X
N
combinations. For each bit-
loading combination, the total rate must be calculated, which
adds a multiplicative factor of N
2
.
The OSB algorithm performs the above exhaustive search
independently on each of the K frequency tones, and re-
peats itself until each users Lagrange multipliers converge.
Assuming that the number of iterations until all the Lagrange
multipliers converge is v
1
, the total complexity of the OSB
algorithm is v
1
K N
2
X
N
, which implies:
C
OSB
= O(KN
2
X
N
).
D. ISB Complexity
The Iterative Spectrum Balancing (ISB) algorithm is de-
scribed in Section III-B (Algorithm 2). Assuming that the
number of discrete bit-loadings is X, then the search space
for each users PSD will contain X elements. For ISB, there
are two loops where the outer loop cycles through the inner
loop for all users. The inner loop performs N different 1-
dimensional exhaustive searches over each of the X elements
for each user. For each bit-loading combination, the total rate
must be calculated, which adds a multiplicative factor of N
2
.
Therefore, with respect to the number of users, N, each inner
loop consists of N
2
+ N calculations. Assuming that the
number of iterations required for the users PSDs to converge
is v
1
, and that the number of iterations for w
n
and
n
to
converge is v
2
, the total complexity of the ISB algorithm is
v
1
v
2
K (N
2
+N) N X, which implies:
C
ISB
= O(KN
3
X).
E. ASB Complexity
The Autonomous Spectrum Balancing (ASB) algorithm is
described in Section III-E (Algorithm 6). ASB has three
nested iterations; the outermost cycle iterates through the
users. Within each cycle, each user runs an outer loop which
updates w
n
until the target data rate is achieved, and an
inner loop which updates
n
until the power constraint is
satised. Bisection search is used in both loops and therefore,
the number of iterations required for both the outer and inner
loops to converge, while achieving an accuracy of
R
and
P
respectively is log
2
(1/
R
) log
2
(1/
P
).
For ASB, the complexity of each iteration is dominated
by nding the roots of the cubic polynomial (i.e., solving
equation (14)). Finding the roots of a cubic equation requires
44 operations in total [46]. This equation must be solved for
each of the K frequency tones, hence the total complexity
of nding the roots is 44K. Therefore, assuming that the
algorithm converges in a nite number of iterations, v, the
total complexity of the ASB algorithm is vNlog
2
(1/
R
)
log
2
(1/
P
) 44K which implies:
C
ASB
= O(KN).
Note that the overall complexity of ASB-2 (described in
Section III-E) is the same as ASB even though the complexity
of each PSD update is reduced (i.e., solving a cubic equation
is replaced by a simple water-lling like update). While the
overall computational complexities are the same, the ASB-2
algorithm has a signicantly lower runtime.
The overall complexity of ASB-MRL is roughly the same as
that of ASB-2. The difference is that the P
ASB2,n
k
becomes a
summation over the number of reference lines. Therefore since
the number of reference lines is a constant, the complexity of
ASB-MRL is also O(KN). Note that even though ASB-MRL
uses multiple reference lines, since it is based on the ASB-2
algorithm, its PSD update formula still provides computational
advantages over the original ASBs.
F. SCALE Complexity
The Successive Convex Approximation for Low complExity
(SCALE) algorithm is described in Section III-F (Algorithm
7). SCALE is an iterative algorithm, each iteration begins
by user n updating its PSD on all K frequency tones. Until
the Lagrange multipliers for each user converge, the process
of measuring their total interference plus noise, sending the
measurements to the Spectrum Management Center (SMC),
waiting for a reply and nally updating values accordingly,
is repeated. Assuming v is the number of operations required
to update a given users PSD, the total complexity from a
distributed point of view is vK.
The SMC receives a message constructed from the mea-
sured total interference plus noise from all users. The SMC
then sends each user the sum of the messages received from
all the other users. Therefore, generating the message for each
user requires KN additions. This process must be repeated
for all N users, leading to a complexity of KN
2
at the SMC.
Therefore, the total complexity of the SCALE algorithm is
KN
2
+vK, which implies:
C
SCALE
= O(KN
2
).
G. DCA Complexity
The Difference of Convex functions Algorithm (DCA) is
described in Section III-G (Algorithm 8). The complexity
will be dominated by calculating the gradient of the objec-
tive function and applying convex optimization techniques.
Since there are K N variables (e.g., s
n
k
for some k and
some n), calculating the gradient requires taking K N
derivatives. For each of the K N derivatives, as shown in
Equations (19) and (20), a nested summation over all N and
N1 elements must be calculated, respectively. This requires
K N N
2
computations the rst time it is computed.
While the denominators of Equations (19) and (20) are not
the same, by noticing that each is the same for all K N
derivatives, additional computations can be avoided by re-
using the previously computed summations (i.e., A
n
k
(s) and
B
n
k
(s), respectively). For each additional user the number of
computations required will be K N N. Therefore the
complexity of taking the derivative will be:
C
DCA
(Derivative) = O(KN
3
KN
2
) = O(KN
3
)
at each iteration.
Each iteration requires solving a convex quadratic program-
ming optimization problem (discussed in Section III-G). It is
shown in [47] that a convex quadratic program can be solved
in O(n
3
L) arithmetic operations, where n is the number of
variables (which is equal to N since the program can be solved
independently on each frequency tone) and L is the encoding
length (which is a constant). Therefore, solving the convex
quadratic program on each of the K frequency tones has a
complexity of O(KN
3
). Note that this algorithm solves for
PSDs without using the spectral mask constraint. If spectral
masks are used, this complexity would increase.
Therefore, assuming that the number of iterations required
until convergence is achieved is v, the total complexity is v
[(KN
3
KN
2
) +KN
3
], which implies:
C
DCA
= O(KN
3
).
TABLE VI
SUMMARY OF DSM ALGORITHM COMPLEXITIES.
DSM Complexity Message Passing
Algorithms Overall At SMC At User Per Update
IWF O(KN) O(K)
ASB O(KN) O(K)
SIW O(KN
2
) O(K) KN
OSB O(Ke
N
)
ISB O(KN
2
)
DCA O(KN
3
)
SCALE O(KN
2
) O(K) 2KN
DSB O(KN
2
) O(K) 2KN
H. DSB Complexity
The Distributed Spectrum Balancing (DSB) algorithm is
described in Section III-H (Algorithm 9). DSB is an iterative
algorithm, each iteration begins by user n measuring its
channel and sending the message
_
1
int
n
k
1
rec
n
k
_
to the SMC.
The SMC gathers the messages from all users, computes
P
DSB,n
k
on all K frequency tones for all N users and sends
P
DSB,n
k
to each user, n. Each user then updates its PSD
on all K frequency tones. Until the Lagrange multipliers
for each user converge, the process of measuring their total
interference plus noise, sending the measurements to the
Spectrum Management Center (SMC), waiting for a reply
and nally updating values accordingly, is repeated. Since the
number of operations required to update a given users PSD
is a constant, the complexity from a distributed point of view
is O(K).
Calculating the message that the SMC sends each user, n, is
a sum over N 1 users of the product between the crosstalk
channel and the message received from user n. Therefore,
generating the message for each user requires K (N 1)
additions. This process must be repeated for all N users,
leading to a complexity of KN
2
KN at the SMC.
Therefore, assuming that v is the number of iterations
required for convergence, the total complexity of the DSB
algorithm is v [KN
2
KN +O(K)], which implies:
C
DSB
= O(KN
2
).
I. Summary of Complexities
The complexities of the DSM algorithms are summarized
in Table VI.
VI. CONCLUDING REMARKS
This paper presented an overview of algorithms that adap-
tively allocate resources in a multiuser DSL system. The
concept of DSM and its advantages in comparison to SSM was
discussed. The following DSM algorithms were discussed in
detail: OSB, ISB, ASB, IWF, SIW, SCALE, DCA and DSB.
Through the use of two illustrative examples (Section IV),
it was shown that semi-centralized based algorithms with only
local knowledge of the network (i.e., SCALE and DSB) were
able to achieve comparable performance (achievable data rate)
to a known full-centralized technique with global knowledge
of the network (e.g., ISB). For some network realizations,
ISB outperformed DSB and SCALE by a maximum of less
than 1%; however, the performance of both DSB and SCALE
was superior to ISB on average, by as much as 10% for
some realizations. It was also shown that a fully distributed
algorithm, ASB-MRL with the appropriate initialization pa-
rameters, achieved comparable performance to SCALE and
DSB.
Through the use of computational complexity analysis
(Section V), it was shown that most of the DSM algorithms
would be computationally tractable for large numbers of users
and frequency tones. By comparing the computational com-
plexities derived in Section V with the runtime and number
of iterations (interference measurements required) results of
Section IV-B, it is clear that IWF, ASB and ASB-MRL have
signicant computational advantages over DSB, SCALE and
ISB.
As well, it was shown that SCALE often produced
marginally better results in terms of performance over DSB;
nevertheless, DSB approached its solution at a much faster
rate. It was also shown that ASB-MRL approached its nal
rate signicantly faster than DSB or SCALE.
Based on the above observations, it is clear that the com-
plexity benets of IWF, ASB and ASB-MRL type algorithms
are very apparent over algorithms like DSB and SCALE, since
they require signicantly fewer interference measurements.
Interference measurements can be time-consuming, and due to
the relatively fast time per iteration of all the non-centralized
algorithms, they appear to be the limiting factor in terms of
the practical implementation of the DSM algorithms.
ACKNOWLEDGMENTS
The authors would like to thank Dr. Raphael Cendrillon for
his discussions and advice on the implementation of ASB,
Dr. Paschalis Tsiaakis for his discussions on DSB, and Prof.
Hoai-An Le-Thi, Mr. Anh Son Ta, Mr. Duc Quynh Tran and
Mr. Thuan Quang Nguyen of Universite Paul Verlaine - Metz
for various interesting discussions on DCA. The authors would
also like to thank the reviewers for their helpful comments.
Finally, the authors would like to acknowledge the nancial
support of NSERC, Bell Canada, and FQRNT.
REFERENCES
[1] Broadband Forum, 3 million new north-american broadband sub-
scribers added in rst quarter 2009 despite economic downturn, News
Release., Jun. 16, 2009.
[2] ANSI Standard T1.417-2001, Spectrum management for loop trans-
mission systems, 2001.
[3] T. Starr, M. Sorbara, J. M. Ciof, and P. J. Silverman, DSL Advances.
Prentice-Hall, 2003.
[4] J. M. Ciof, S. Jagannathan, W. Lee, H. Zou, A. Chowdhery, W. Rhee,
G. Ginis, and P. Silverman, Greener copper with dynamic spectrum
management, in Proc. IEEE Global Telecommun. Conf., pp. 15, Dec.
2008.
[5] W. Yu, G. Ginis, and J. Ciof, Distributed multiuser power control for
digital subscriber lines, IEEE J. Sel. Areas Commun., vol. 20, pp. 1105
1115, Jun. 2002.
[6] Y. Liu and Z. Su, Distributed dynamic spectrum management for digital
subscriber lines, IEICE Trans. Commun., vol. E90B, no. 3, pp. 491
498, Mar. 2007.
[7] H. Bagheri, M. R. Pakravan, and B. H. Khalaj, Iterative multi-user
power allocation: Performance evaluation & modication, in Proc.
IEEE Region 10 TENCON, vol. 3, pp. 216219, Nov. 2004.
[8] D. Statovci, T. Nordstrom, and R. Nilsson, The normalized-rate itera-
tive algorithm: A practical dynamic spectrum management method for
DSL, EURASIP J. Appl. Signal Process., vol. 2006, Article ID 95175,
17 pages, DOI: 10.1155/ASP/2006/95175, 2006.
[9] W. Lee, Y. Kim, M. H. Brady, and J. M. Ciof, Band-preference
dynamic spectrum management in a DSL environment, in Proc. IEEE
Global Telecommun. Conf., pp. 15, Nov. 2006.
[10] Y. Noam and A. Leshem, Iterative power pricing for distributed
spectrum coordination in DSL, IEEE Trans. Commun., vol. 57, pp. 948
953, Apr. 2009.
[11] Y. Xu, S. Panigrahi, and T. Le-Ngoc, Selective iterative water-lling
for digital subscriber lines (DSL), EURASIP J. Appl. Signal Process.,
vol. 2007, Article ID 59068, 11 pages, DOI: 10.1155/2007/59068, 2007.
[12] R. Cendrillon, W. Yu, M. Moonen, J. Verlinden, and T. Bostoen,
Optimal multiuser spectrum management for digital subscriber lines,
IEEE Trans. Commun., vol. 54, pp. 922933, May 2006.
[13] S. Wei, L. Youming, and Y. Miaoliang, Low-complexity grouping
spectrum management in multi-user DSL networks, in Proc. IEEE Int.
Conf. Commun., pp. 381385, Jun. 2009.
[14] L. P. Qian and Y. Jun, Monotonic optimization for non-concave power
control in multiuser multicarrier network systems, in Proc. IEEE
INFOCOM, pp. 172180, Apr. 2009.
[15] T. Hoang, Convex Analysis and Global Optimization. Kluwer Academic
Publishers, 1998.
[16] Y. Xu, T. Le-Ngoc, and S. Panigrahi, Global concave minimization for
optimal spectrum balancing in multi-user DSL networks, IEEE Trans.
Signal Process., vol. 56, pp. 28752885, Jul. 2008.
[17] R. Horst, T. Q. Phong, N. V. Thoai, and J. de Vries, On solving a D.C.
programming problem by a sequence of linear programs, Journal of
Global Optimization, vol. 1, pp. 183204, 1991.
[18] R. Cendrillon and M. Moonen, Iterative spectrum balancing for digital
subscriber lines, in Proc. IEEE Int. Conf. Commun., May 2005.
[19] R. Lui and W. Yu, Low-complexity near-optimal spectrum balancing
for digital subscriber lines, in Proc. IEEE Int. Conf. Commun., May
2005.
[20] A. Forouzan and L. Garth, Generalized iterative spectrum balancing
and grouped vectoring for maximal throughput of digital subscriber
lines, in Proc. IEEE Global Telecommun. Conf., vol. 4, pp. 23592363,
Nov. 2005.
[21] R. Cendrillon, J. W. Huang, M. Chiang, and M. Moonen, Autonomous
spectrum balancing for digital subscriber lines, IEEE Trans. Signal
Process., vol. 55, pp. 42414257, Aug. 2007.
[22] P. Tsiaakis, M. Diehl, and M. Moonen, Distributed spectrum man-
agement algorithms for multiuser DSL networks, IEEE Trans. Signal
Process., vol. 56, pp. 48254842, Oct. 2008.
[23] C. Leung, S. Huberman, and T. Le-Ngoc, Autonomous spectrum
balancing using multiple reference lines for digital subscriber lines,
in Proc. IEEE Global Telecommun. Conf., Dec. 2010.
[24] R. Moraes, B. Dortschyy, A. Klautau, and J. R. i Riuy, Semi-blind
power allocation for digital subscriber lines, in Proc. IEEE Int. Conf.
Commun., pp. 14201425, May 2008.
[25] J. Papandriopoulos and J. S. Evans, Low-complexity distributed alo-
girthms for spectrum balancing in multi-user DSL networks, in Proc.
IEEE Int. Conf. Commun., Jun. 2006.
[26] J. Papandriopoulos and J. S. Evans, Scale: A low-complexity distributed
protocol for spectrum balancing in multiuser DSL networks, IEEE
Trans. Inf.. Theory, vol. 55, pp. 37113724, Aug. 2009.
[27] D. T. Pham and H. A. Le Thi, Convex analysis approach to d.c.
programming: Theory, algorithms and applications, Acta Mathematica
Vietnamica, vol. 22, pp. 289355, 1997.
[28] T. Starr, J. M. Ciof, and P. Silverman, Understanding Digital Sub-
scriber Line Technology. Upper Saddle River, NJ: Prentice-Hall, 1999.
[29] J. M. Ciof, S. Jagannathan, and W. Lee, Digital subscriber
line (DSL). http:// www.scholarpedia.org/ article/ Digital subscriber
line (DSL), Aug. 2008.
[30] R. van der Berg, Developments in bre technologies and investment
(working party on communication infrastructures and services policy),
Organisation for Economic Co-operation and Development, Apr. 3,
2008.
[31] ZTE, FTTx solution white paper, http:// www.telecomasia.net/ pdf/
ZTE/ ZTE 091709.pdf , 2009.
[32] S. Yoshii, Spectrum management of DSL systems in Japan, in TEN-
CON 2005 2005 IEEE Region 10, pp. 16, Nov. 2005.
[33] AT&T u-verse TV marks 2 million customer milestone, http:// www.
digitaltvnews.net/ content/ ?p=11786, Dec. 9, 2009.
[34] Bell high speed bre rollouts enable new internet and TV services,
http:// www.bce.ca/ en/ news/ releases/ bev/ 2010/ 02/ 04/ 75344.html, Feb.
4, 2010.
[35] S. Sadr, A. Anpalagan, and K. Raahemifar, Radio resource allocation
algorithms for the downlink of multiuser OFDM communication sys-
tems, IEEE Commun. Surveys Tuts., vol. 11, no. 3, pp. 92106, 2009.
[36] P. Tsiaakis, J. Vangorp, M. Moonen, J. Verlinden, and K. Van Acker,
An efcient search algorithm for the Lagrange multipliers of optimal
spectrum balancing in multi-user xDSL systems, in Acoustics, Speech
and Signal Processing, 2006. ICASSP 2006 Proceedings. 2006 IEEE
International Conference on, vol. 4, pp. IVIV, May 2006.
[37] ETSI, Transmission and multiplexing (TM); access transmission sys-
tems on metallic access cables; very high speed digital subscriber line
(VDSL); part 1: Functional requirements, ETSI Std. TS 101 270-1,
vol. Rev. V. 1.3.1, 2003.
[38] Very high speed digital subscriber line (VDSL) metalic interface, ANSI
Std., vol. T1.424, 2004.
[39] Very high speed digital subscriber line transceivers 2, ITU Draft Std.,
vol. G.993.2, 2006.
[40] M. Chiang, Advances in Applied Mathematics and Global Optimization,
ch. Nonconvex Optimization for Communication Systems. Springer,
2009.
[41] H. A. Le Thi and D. T. Pham, The DC programming and DCA revisted
with DC models of real world non-convex optimization problems,
Annals of Operations Research, vol. 133, pp. 2346, 2005.
[42] R. Blanquero and E. Carrizosa, On covering methods for D.C. opti-
mization, Journal of Global Optimization, vol. 18, pp. 265274, 2000.
[43] S. Huberman, C. Leung, and T. Le-Ngoc, DSM algorithms in multi-
user xDSL, Tech. Rep., McGill University, Aug. 2009. http://www.ece.
mcgill.ca/
shuber6/.
[44] K. McCammon, G. VDSL: VDSL band plan for North America, ITU
contribution D, vol. 715, 2000.
[45] Very high speed digital subscriber line transceivers 2, ITU Draft Std.,
vol. G.993.2 Amendment 1, 2007.
[46] R. Nickalls, A new approach to solving the cubic: Cardans solution
revealed, Mathematical Gazette, vol. 77, pp. 354359, 1993.
[47] D. Goldfarb and S. Liu, An O(n
3
L) primal interior point algorithm
for convex quadratic programming, Math. Program., vol. 49, no. 3,
pp. 325340, 1991.
Sean Huberman received his B.Sc. in engineering
(with rst-degree honours) from Queens University,
Kingston, Canada, in applied mathematics and engi-
neering control and communications, in May 2008.
He began his M.Eng. in electrical engineering at
McGill University, Montreal, Canada, in September
2008. In January 2010, he transferred into the PhD
program in electrical engineering at McGill Univer-
sity.
His research interests include techniques of math-
ematical optimization, dynamic resource allocation,
channel modeling and channel measurements.
Mr. Huberman was the recipient of the Hydro Quebec Engineering Doctoral
Award in 2009. In 2010, he was the recipient of the Vadasz Doctoral
Fellowship in Engineering. He was also the recipient of a three-year doctoral
National Science and Engineering Research Council (NSERC) award and a
three-year McGill Engineering Doctoral Award (MEDA).
Christopher Leung received his B.Eng. (with
Distinction and Deans Honour List) in electri-
cal engineering in 2009 from McGill University,
Montr eal, Qu ebec, Canada. He is currently nishing
his M.Eng., also in electrical engineering at McGill
University.
Mr. Leung has received the Alexander Graham
Bell Canada Graduate Scholarship from the Na-
tional Science and Engineering Research Council
(NSERC) and the Bourses de matrise en recherche
from the Fonds qu ebecois de la recherche sur la
nature et les technologies (FQRNT) for pursuing his M.Eng.
Tho Le-Ngoc Tho Le-Ngoc obtained his B.Eng.
(with Distinction) in Electrical Engineering in 1976,
his M.Eng. in 1978 from McGill University, Mon-
treal, and his Ph.D. in Digital Communications in
1983 from the University of Ottawa, Canada. During
1977-1982, he was with Spar Aerospace Limited and
involved in the development and design of satellite
communications systems. During 1982-1985, he was
an Engineering Manager of the Radio Group in the
Department of Development Engineering of SRT-
elecom Inc., where he developed the new point-to-
multipoint DA-TDMA/TDM Subscriber Radio System SR500. During 1985-
2000, he was a Professor at the Department of Electrical and Computer
Engineering of Concordia University. Since 2000, he has been with the
Department of Electrical and Computer Engineering of McGill University.
His research interest is in the area of broadband digital communications.
He is the recipient of the 2004 Canadian Award in Telecommunications
Research, and recipient of the IEEE Canada Fessenden Award 2005. He holds
a Canada Research Chair (Tier I) on Broadband Access Communications, and
a Bell Canada/NSERC Industrial Research Chair on Performance & Resource
Management in Broadband xDSL Access Networks.

Dynamic Spectrum Management (DSM) Algorithms For Multi-User XDSL

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Dynamic Spectrum Management (DSM) Algorithms For Multi-User XDSL

Hochgeladen von

Copyright:

Verfügbare Formate

IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 14, NO.

1, FIRST QUARTER 2012 109

h, convex. Each suc-

Das könnte Ihnen auch gefallen