You are on page 1of 14

1

Elementary Properties of Positive


Concave Mappings with Applications
to Network Planning and Optimization
Renato L. G. Cavalcante, Member, IEEE, Yuxiang Shen, and Sawomir Stanczak, Senior Member, IEEE
Fraunhofer Heinrich Hertz Institute / Technical University of Berlin
Berlin, Germany
arXiv:1505.03006v5 [cs.IT] 12 Jan 2016

Email: renato.cavalcante@hhi.fraunhofer.de, yuxiangshenee@gmail.com, slawomir.stanczak@hhi.fraunhofer.de

AbstractThis study presents novel methods for computing optimization of networks based on the orthogonal frequency
fixed points of positive concave mappings and for characterizing division multiple access (OFDMA) technology [4][10]. In
the existence of fixed points. These methods are particularly im- many of these applications, the mappings are positive concave
portant in planning and optimization tasks in wireless networks. mappings, which are a strict subclass of standard interference
For example, previous studies have shown that the feasibility of
mappings [7].
a network design can be quickly evaluated by computing the
fixed point of a concave mapping that is constructed based on
For example, by using standard interference coupling mod-
many environmental and network control parameters such as els that are widely used in the literature [11], the studies in [4]
the position of base stations, channel conditions, and antenna and [6] consider a very particular case of a positive concave
tilts. To address this and more general problems, given a positive mapping for the problem of load estimation in long-term
concave mapping, we show two alternative but equivalent ways evolution (LTE) networks. The fixed point of that mapping,
to construct a matrix that is guaranteed to have spectral radius if it exists, indicates the bandwidth required by each base
strictly smaller than one if the mapping has a fixed point. This station to satisfy the data rate requirements of users. With
matrix is then used to build a new mapping that preserves
this knowledge, we can evaluate the feasibility of a network
the fixed point of the original positive concave mapping. We
design by verifying whether the required bandwidth does not
show that the standard fixed point iterations using the new
exceed the available bandwidth. However, especially in large-
mapping converges faster than the standard iterations applied
to the original concave mapping. As exemplary applications of scale planning, computation of the fixed point may require
the proposed methods, we consider the problems of power and time-consuming iterative methods. Therefore, the development
load estimation in networks based on the orthogonal frequency of fast tools to ensure the existence of a fixed point before
division multiple access (OFDMA) technology. starting a time-consuming iterative process is of high practical
relevance to network designers and to algorithms for self-
I. I NTRODUCTION organizing networks.

Problems that can be posed as that of finding fixed points In the above-mentioned load estimation problem, existence
of standard interference mappings are ubiquitous in commu- of a fixed point is fully characterized by the spectral radius
nication systems [1][9], and, in particular, in planning and of a matrix that is easily constructed from the associated
concave mapping [6]. We can also use this matrix to build
This work was partially supported by the Deutsche Forschungsgemeinschaft an affine mapping having as its fixed point a vector that gives
(DFG) under Grant STA 864/9-1. Copyright (c) 2015 IEEE. Personal use of
a lower bound of the network load. The main advantage of
this material is permitted. However, permission to use this material for any
other purposes must be obtained from the IEEE by sending a request to pubs- working with affine mappings in finite dimensional spaces is
permissions@ieee.org that computation of their fixed points reduces to solving simple
2

systems of linear equations, so we may easily obtain in this points of general positive concave mappings. We also pro-
way a certificate that the current network configuration is not pose novel low-complexity iterative methods that improve the
able to serve the demanded traffic. convergence speed of the standard fixed point algorithm. In
The first objective of this study is to show that, by using the Sect. V we revisit the problems of load and power estimation in
concept of recession or asymptotic functions in convex analysis OFDMA-based networks, and we show how the novel results
[12], [13], the technique used in [6] for the construction of and algorithms proposed here can be used in these concrete
the above-mentioned matrices (hereafter called lower bounding applications.
matrices) admits a simple extension to general positive concave
mappings. This extension has been motivated by recent results II. P RELIMINARIES
in power estimation in LTE networks [8], [9], which deal with
mappings different from that considered in [4], [6]. Concave In this study, we use the following standard definitions. By
mappings are also common in many applications in different hx, yi for arbitrary x RN and y RN , we denote the
fields [14], so the results of this study are relevant for appli- standard inner product hx, yi := xt y. Its induced norm is
p
cations outside of the wireless domain. We show alternative given by kxk := hx, xi. The set I := {1, . . . , N } is the
construction methods for lower bounding matrices that are set of indices of the components of vectors in RN . The -
very simple in many applications, including those originally norm of a vector x = [x1 , . . . , xN ] is the norm given by
considered in [6]. We also prove that the spectral radius of kxk := maxiI |xi |. We define by ek the kth standard
lower bounding matrices of general concave mappings gives basis vector of RN . Vector inequalities should be understood
a necessary condition for the existence of fixed points. For as component-wise inequalities, and we define RN + := [0, [
N

some particular concave mappings, this condition is shown to and RN ++ := ]0, [


N
(the superscript is omitted if N = 1).
be sufficient. The set of positive integers is denoted by N := {1, 2, . . .}.
The spectral radius of a matrix M RN N is given by
The second objective of this study is to develop an accelera-
(M ) := max{|1 |, . . . , |N |}, where 1 , . . . , N are the
tion method for the standard fixed-point iteration described in
eigenvalues of the matrix M . The component of the ith row
[1] when applied to concave mappings. More specifically, we
and kth column of a matrix M is denoted by [M ]i,k . For a
combine the lower bounding matrix and the original positive
vector x = [x1 , . . . , xN ] RN , the matrix diag(x) RN N
concave mapping to generate a new mapping that has the
is a diagonal matrix with [diag(x)]i,i = xi .
same fixed point of the original concave mapping. By applying
the standard fixed point iteration to this new mapping, the Concave functions and standard interference functions play
convergence speed is improved in a well-defined sense, and a crucial role in this study, so we review below basic definitions
the computational complexity is not unduly increased because and known results that are extensively used in the next sections.
only one additional matrix-vector multiplication per iteration
is required. As exemplary applications of the above results, Definition 1 (Convex set) A set C RN is said to be convex
we consider the problems of power and load estimation in if
OFDMA-based systems [8], [9].
This study is structured as follows. In Sect. II we review (x C)(y C)( ]0, 1[) x + (1 )y C.
basic results in convex analysis and in interference calculus.
The material in Sect. II can now be considered standard, but Definition 2 (Concave functions) We say that f : C R
we also show a simple proof of the fact that positive concave {} is a concave function if C RN is a convex set and
functions are standard interference functions. In Sect. II we
relate some results in [6] (used to compute lower bounds (x dom f )(y dom f )( ]0, 1[)
for load in LTE network planning) to standard results on f (x + (1 )y) f (x) + (1 )f (y),
recession functions in convex analysis. The relations are used
in Sect. IV to derive conditions for the existence of fixed where dom f := {x C | f (x) > } is the (effective)
3

1
domain of f . Definition 4 (Recession or asymptotic functions) Let f :
RN R {} be upper semicontinuous, proper, and
Concave functions f : C R {} with C RN
concave. We define as its recession or asymptotic function at
for which there exists at least one vector x C satisfying
y RN the function given by (see [12, Ch. 2.5] [13, p. 152]
f (x) > are called proper concave functions. If for every
for the standard definition for convex f ):
sequence {xn }nN C converging to an arbitrary vector x
f (x + hy) f (x)
C, we have lim supn f (xn ) f (x), then we say that the (x domf ) f (y) := lim .
h h
function f is upper semicontinuous (on C).
(NOTE: The above limit is always well defined. We assume
Every concave function f : C R {} can be related
that it can take the value .)
to a convex function by f : C R {} (f takes the
value whenever f takes the value ), so the following Fact 2 If f : RN R {} is a proper, upper semicon-
results on concave functions can be directly deduced from tinuous, and concave function, then for every y domf we
standard results on convex functions found in the literature have [12, Corollary 2.5.3]
[12], [13].
f (y) = lim+ hf (h1 y), (1)
h0
Definition 3 (Superdifferentials and supergradients) Let and the above is valid for every y RN if 0 dom f .
f : C R {} be a concave function with
6= C RN . The superdifferential of f at x dom f is the Fact 3 Let f : RN R {} be proper, upper semicon-
set given by tinuous, and concave. Then [12, Proposition 6.5.1]

f (x) := u RN | (y C) hy x, ui + f (x) f (y) . f (y) = inf{hg, yi | g f (x), x dom f }.

If x / dom f , then we define f (x) := . A vector g Many estimation and optimization tasks in communication
f (x) is called a supergradient of f at x. The domain of networks can often be posed as systems coupled by standard
the superdifferential f is the set given by dom f := {x interference functions, which we define below.
RN | f (x) 6= }.
Definition 5 (Standard interference functions and mappings
In this study, if the point at which a supergradient is [1]) A function f : RN + R++ is said to be a standard
selected needs to be explicitly known, then we often use the interference function if the following properties hold:
notation g(x) f (x) to denote an arbitrary choice of the (Scalability) (x RN
1) + ) ( > 1) f (x) > f (x).
supergradient at x. (Monotonicity) (x1 RN N
2) + ) (x2 R+ ) x1 x2
As a particular case of [13, Corollary 16.15], we have the f (x1 ) f (x2 ).
following: Given N standard interference functions fi : RN + R++ ,
i = 1, . . . , N , we call the mapping T : R+ RN
N
++ : x 7
Fact 1 Let f : RN + R++ be concave. Then the superdiffer-
[f1 (x), . . . , fN (x)] a standard interference mapping.
ential f (x) is nonempty for every x RN
++ .

Fact 4 (Properties of interference mappings [1]) Let T :


The proposed acceleration methods are based on the concept
RN N
+ R++ be a standard interference mapping. Then the
of recession functions (or asymptotic functions) in convex
following holds:
analysis.

Fact 4.1 Let Fix(T ) := {x RN ++ | T (x) = x} be the set


1 In
the literature, when a concave function f is allowed to take the value of fixed points of T , then Fix(T ) is either an empty set or a
, assuming that C = RN is a common practice. If C is a proper subset
singleton.
of RN , we can define f (x) = if x RN \C to extend f from C to
RN . By doing so, the effective domain is preserved. However, for notational
Fact 4.2 Fix(T ) 6= if and only if there exists x RN such
convenience later in the text, we do not necessarily adhere to this convention,
and we allow C to be a strict subset of RN . that T (x ) x .
4

Fact 4.3 If Fix(T ) 6= , then it is the limit of the sequence conclude that
 
{xn }nN generated by xn+1 = T (xn ), where x1 RN + is ar- 1
2
f (x2 ) lim f (x1 ) f (x1 ) = f (x1 ).
bitrary. If x1 satisfies T (x1 ) x1 (resp. T (x1 ) x1 ), then
the sequence is monotonically increasing (resp. monotonically
decreasing) in each component. In particular, monotonically As every result stated in this section, Proposition 1 can
increasing sequences are produced with x1 = 0. be considered standard (see [7] and the references therein).
Nevertheless, we have decided to include a simple proof of this
The focus of this study is on (positive) concave functions,
proposition because similar statements can often be found in
which as shown below are a subclass of standard interference
the literature without proof. Furthermore, some partial proofs
functions.
available in the literature make implicit assumptions such as
the existence of the supergradients on the boundary of the
Proposition 1 Concave functions f : RN
+ R++ are
domain RN + and/or the strict concavity of the functions. We
standard interference functions:
emphasize that these assumptions are not required. As an
Proof: We need to prove that concave functions f : RN
+ example of a positive concave function (and hence a standard
R++ satisfy the scalability and monotonicity properties in interference function) not satisfying these two assumptions, we
Definition 5. have

(Scalability) Let > 1 and x RN + be arbitrary. By 1, if x = 0,
concavity of f , for every ]0, 1[, we have f (x) = f : R+ R++ : x 7
2, otherwise.
f (x+ (1 )0) f (x)+ (1 )f (0). In particular, for
= 1/, we conclude from the last inequality and positivity To characterize the existence of fixed points of affine stan-
of f that dard interference mappings, we can use the following fact:
 
1 1 1
f (x) f (x) + 1 f (0) > f (x), Fact 5 [2, Theorem A.16] For an arbitrary matrix M
P
RN N , if (M ) < 1, then k
k=1 M converges and (I
which proves the scalability property. P
M )1 = I + k=1 M k .
(Monotonicity) Let (x1 , x2 ) RN N
+ R+ satisfy x2 x1 .
As a result, x1 + (x2 x1 ) RN + for every 0. From N
Fact 6 [2, Theorem A.51] Let M RN + be a non-
the definition of concavity, we also have N
negative matrix, and let p R++ be arbitrary. A sufficient
( ]0, 1[)( 0) and necessary condition for the system x = p + M x to have
a (strictly) positive solution x RN
++ is (M ) < 1.
f ((1 )x1 + (x1 + (x2 x1 )))
(1 )f (x1 ) + f (x1 + (x2 x1 )) . We end this section with a very simple statement that is
used later to clarify an argument in Sect. V.
In particular, for an arbitrary > 1 and for = 1/, we
obtain from the positivity of f that Remark 1 Let M RN N be arbitrary and D RN N be
 
1 1 an invertible matrix. Then the eigenvalues of the matrices M
f (x2 ) 1 f (x1 ) + f (x1 + (x2 x1 ))
and DM D 1 are the same (which in particular implies that
1
> f (x1 ) f (x1 ). (M ) = (DM D1 )).

The inequality f (x2 ) > f (x1 ) (1/)f (x1 ) is valid for Proof: Assume that x is a right eigenvector associated
every > 1, so we can take the limit as goes to infinity to with an eigenvalue , and define y := Dx. As a consequence,
2 In finite dimensional spaces, all norms are equivalent. Therefore, conver- M x = x M D1 y = D 1 y DM D 1 y = y,
gence of a sequence {xn }nN RN to a point x RN does not depend
on the choice of the norm; i.e., limn kxn x ka = 0 for any norm and the result follows.
k ka .
5

III. C OMPONENT- WISE INFIMUM OF SUPERGRADIENTS OF Lemma 1.3 As in Lemma 1.2, let x RN + and k I be
POSITIVE CONCAVE FUNCTIONS arbitrary and assume that x+hek dom f for every h 0.
The main objective of this section is to propose two simple Then
techniques for computing the component-wise infimum of
(h > 0)(g(x + hek ) f (x + hek ))
supergradients of concave functions (c.f. Proposition 2 and
f (x + hek ) f (x)
Proposition 3). These techniques are motivated by the fol- gk (x + hek ) , (2)
h
lowing application. In load estimation problems in wireless
where [g1 (x + hek ), . . . , gN (x + hek )]t := g(x + hek ).
networks, the values taken by partial derivatives of functions
related to the load coupling among base stations attain their Proof:
infimum asymptotically as we move to infinity in the direction 1) We prove the result by contradiction. Assume that there
of a basis vector [4], [15]. This observation has given rise to exists a supergradient g =: [g1 , . . . , gN ]t f (x ) at
efficient techniques for the computation of lower bounds for some point x dom f such that gi < 0 for an
the load in that very particular application domain [15], and arbitrary i {1, . . . , N }. We know from the definition
extending these results to a more general class of concave of supergradients that
functions is highly desirable for other applications such as

power estimation in networks. (y RN t
+ ) f (y) f (x ) + g (y x ).

By using the concept of recession or asymptotic functions, In particular, for u : R RN : h 7 x + hei , we


we show below that the above-mentioned asymptotic result obtain
can be generalized to all positive concave functions, even if
the functions are not differentiable, in which case we use f (u(h)) f (x ) + g t (u(h) x ) = f (x ) + gi h.
supergradients instead of gradients (c.f. Proposition 3). We
Now, since gi < 0 by assumption, we obtain f (u(h))
can further show that the component-wise infimum taken by
0 for an arbitrary h f (x )/|gi |, which contradicts the
the supergradients can be easily obtained by means of simple
positivity of the range of the function f : RN + R++ .
schemes that do not require the computation of supergradients
This proves Lemma 1.1.
(c.f. Proposition 2). These infimum values are used later by
2) By Definition 3, for arbitrary x1 , x2 dom f , we
the proposed acceleration schemes to compute fixed points of
have f (x1 ) f (x0 ) + g t0 (x1 x0 ) and f (x0 )
positive concave mappings, and they can also be used to obtain
f (x1 ) + g t1 (x0 x1 ), where g 0 f (x0 ) and g 1
a certificate that the mapping does not have a fixed point. We
f (x1 ) are arbitrary supergradients. Summing these
start by formalizing some simple properties of supergradients
two inequalities yields
of concave functions.
t
(g 1 g 0 ) (x1 x0 ) 0. (3)
Lemma 1 Let f : RN
R++ be an upper semicontinuous
+
In particular, for x1 = x + hek and x0 = x, we
concave function. Then the following holds:
have 0 x1 x0 = hek 6= 0, and we can set
Lemma 1.1 All supergradients of f are non-negative vectors; g 0 = g f (x) and g 1 = g f (x + hek ). Using
i.e., these particular choices for x0 , x1 , g 0 , and g 1 in (3), we
(x dom f ) (g f (x)) g 0. obtain gk gk . Non-negativity of gk has been proved
in the first part of the lemma.
Lemma 1.2 Let x RN + and k I be arbitrary and assume
3) Use x = x in the supergradient inequality f (x )
that x + hek dom f for every h 0. Then f (x + hek ) + g(x + hek )t (x x hek ).
(h > 0) (g f (x)) (g f (x + hek ))
We can now show an efficient scheme to compute the
0 gk gk ,
element-wise infimum of supergradients.
where [g1 , . . . , gN ] := g and [g1 , . . . , gN
t
] := g .
t
6

S
Proposition 2 Let S := xRN f (x) be the set of all for any choice of x and k satisfying the assumptions of the
+
supergradients of an upper semicontinuous concave function lemma. Using (2) in Lemma 1.3 and the definition of recession
f : RN
+ R++ . For each k I, define functions, we deduce

gk := inf {gk R | g = [g1 , . . . , gN ] S} R+ , (4)


gk lim gk (x + hek )
then we have h
f (x + hek ) f (x)
lim = f (ek ).
(k I) gk = lim+ hf (h1 ek ). h h
h0
The result now follows by noticing that f (ek ) = gk by
Proof:
Fact 3. (Non-negativity of gk is immediate from Lemma 1.1.)
We have gk 0 as a direct consequence of Lemma 1.1. Now
consider the standard extension f : RN R++ {} of
f : RN
+ R++ given by IV. ACCELERATION ALGORITHMS FOR POSITIVE

f (x) if x RN CONCAVE MAPPINGS
+
f(x) =
otherwise, Having two efficient methods to compute the component-
wise infimum of supergradients of concave functions, we can
and let k I be arbitrary. By construction, f is upper now proceed with the study of general concave mappings.
semicontinuous, proper, and concave. Furthermore, dom f = To avoid unnecessary technical digressions, we do not deal
dom f, dom f = domf, and f (x) = f(x) for every with concave functions f : RN+ R++ that are not upper
x dom f. By Fact 3, we have gk = f (ek ) (see semicontinuous. To formalize this assumption, we use the
Definition 4), and the result now follows from Fact 2. following definition:
Next, we show an alternative means of computing gk in
(4). This alternative method has been used in [6] for a very Definition 6 (Positive concave mappings) We say that T :
particular concave function appearing in load estimation in RN N
+ R++ is a positive concave mapping if it is given by
LTE networks (see Sect. V-A).
T (x) := [f1 (x), . . . , fN (x)]t , (6)

Proposition 3 Let x RN + and k I be arbitrary. In addi-


where all functions f1 : RN N
+ R++ , . . . , fN : R+ R++
N
tion, assume that f : R+ R++ is an upper semicontinuous are concave and upper semicontinuous.
concave function and that x+hek dom f for every h 0.
By Proposition 1, we know that positive concave mappings
Define by [g1 (x + hek ), . . . , gN (x + hek )] := g(x + hek )
are standard interference mappings. The remaining of this
f (x + hek ) an arbitrary supergradient at x + hek . Then
section has the objective of investigating the following prob-
lim gk (x + hek ) = gk 0, (5) lems associated with a positive concave mapping T (which, as
h
shown in Sect. V, are problems that need to be addressed in
where gk is defined in (4). many network planning and optimization tasks):

Proof: Let k I be arbitrary. It follows from Lemma 1.2 P1) Verify whether T has a fixed point by using computa-
that, irrespective of the criterion we use to select a super- tionally efficient algorithms.
gradient g(x + hek ) f (x + hek ), its kth component P2) Improve the convergence speed of the standard iteration
gk (x + hek ) should be monotonically non-increasing as h in Fact 4.3 to obtain the fixed point of T (if it exists).
increases (and lower bounded by 0). As a result, the limit
limh gk (x + hek ) exists. By definition, gk is the infimum A. Conditions for the existence of fixed points of positive
of the kth component of all supergradients, hence we have that concave mappings
To address problem P1), we use the concept of lower
gk lim gk (x + hek )
h bounding matrices, which we define as follows:
7

Definition 7 The lower bounding matrix of a positive concave Lemma 2 Let M be the lower bounding matrix of a positive
mapping T : RN N t
+ R++ : x 7 [f1 (x), . . . , fN (x)] is the concave mapping T : RN N
+ R++ in accordance with
N
non-negative matrix M RN + with its ith row and kth Definition 7. Then
column given by
(y RN N
+ )(x R+ ) x y T (x) T (y) + M (x y).

[M ]i,k := inf {gk R | [g1 , . . . , gN ] Si } R+ , (7)


S (10)
where Si := xRN fi (x).
+
Proof: We prove the inequality for an arbitrary component
Note that Proposition 2 and Proposition 3 show two simple of the mapping T ; i.e., for the function fi , where i I is
techniques to compute each component of lower bounding arbitrary. Let y RN + and x y be arbitrary vectors, and
matrices. construct the sequence {xn := (1/n)1 + x}nN RN ++ . By
Fact 1, we have xn dom fi for every n N. From the
Example 1 (Construction of lower bounding matrices with
definition of supergradients, we know that, for every n N,
the results in Propositions 2 and 3) Let the functions f1 :
RN N fi (y) + g tn (xn y) fi (xn ), (11)
+ R++ , . . . , fN : R+ R++ be concave and upper
semicontinuous. Using Proposition 2, we can compute the where g n fi (xn ) is an arbitrary supergradient. By Defini-
lower bounding matrix M of the mapping T : RN N
+ R++ : tion 7, the ith row of M , denoted by mi 0 as a column
t
x 7 [f1 (x), . . . , fN (x)] by vector, is the component-wise infimum of all supergradients of
the function fi , hence 0 mi g n for every n N. Using
M=
this last relation together with xn y in (11), we deduce:
limh0+ hf1 (h1 e1 ) limh0+ hf1 (h1 eN )
(n N)
... ..
. ... .

limh0+ hfN (h1 e1 ) limh0+ hfN (h1 eN ) fi (y) + mti (xn y) fi (y) + g tn (xn y) fi (xn ).
(8) By construction, limn xn = x. As a result, we con-
Equivalently, by fixing x RN clude from the continuity of affine functions and upper semi-
++ arbitrarily, we can also
compute the lower bounding matrix M with the results in continuity of fi that
Proposition 3 and Fact 1 as follows: fi (y) + mti (x y) = lim sup(fi (y) + mti (xn y))
n
M= lim sup fi (xn ) fi (x).
n
limh g11 (x + he1 ) 1
limh gN (x + heN )
.. .. ..
. ,
. . The next proposition addresses problem P1) stated in the
N N
limh g1 (x + he1 ) limh gN (x + heN ) beginning of this section:
(9)
Proposition 4 Let M be the lower bounding matrix of a
where we denote by gki (x) the kth element of a supergradient
positive concave mapping T : RN N
+ R++ . A necessary
of fi at x RN i i t
++ (i.e., [g1 (x), . . . , gN (x)] fi (x)). condition for Fix(T ) 6= is (M ) < 1.
Proposition 2 and Proposition 3 also show that the lower Proof: Use y = 0 in (10) to verify that the affine mapping
bounding matrix is non-negative. The name lower bounding T L : RN N
+ R++ : x 7 T (0) + M x satisfies T (x)
matrix stems from the fact that this matrix is constructed TL (x) > 0 for every x RN + . Being an affine mapping,
with component-wise lower bounds of supergradients. Lower TL is a positive concave mapping, hence it is also a standard
bounding matrices can also be used to construct affine map- interference mapping by Proposition 1. Now let x RN ++ be
pings that serve as lower bounds of their corresponding posi- the fixed point of the mapping T . By Lemma 2, we obtain:
tive concave mappings, in the following sense:
TL (x ) T (x ) = x , (12)
8

which implies the existence of the (unique) fixed point of the Proposition 5 Let T : RN N
+ R++ be a positive concave
mapping TL by Fact 4.2. In other words, there exists a unique mapping with lower bounding matrix M satisfying (M ) < 1.
positive vector xb RN b = T (0) + M x
++ satisfying x b , and we In addition, assume that
know by Fact 6 that there exists a positive vector satisfying this
(y RN N
++ )(x R+ ) T (x) y + M x. (13)
last equality if and only if (M ) < 1 (recall that T (0) > 0
N
and that M RN + by construction). Then the mapping T has a fixed point.
An immediate consequence of Fact 5 and Proposition 4 is
the following useful result: Proof: Let y RN ++ be a vector satisfying T (x)

y + M x for every x RN + . By Fact. 6, we know that x :=
Corollary 1 Let T : RN N
+ R++ be a positive concave (I M )1 y is a strictly positive vector. Therefore,
N
mapping with Fix(T ) 6= , and denote by M RN + its
T (x ) y + M x = (I M )x + M x = x ,
existing lower bounding matrix given by Definition 7. Then
(I M )1 exists, and it is a non-negative matrix. and the above implies that Fix(T ) 6= by Fact. 4.2.

Proposition 4 is interesting in its own right because it


enables us to certify that a given positive concave mapping
has no fixed point. We only need to show that the spectral B. Acceleration techniques for positive concave mappings
radius of its lower bounding matrix has spectral radius greater
We now turn our attention to problem P2). To address this
than or equal to one. This result is highly relevant in network
problem, we use the concept of accelerated mappings, which
optimization and planning problems. As already mentioned
we define as follows:
in the introduction, in these applications, the feasibility of a
network design follows from the existence of the fixed point
Definition 8 (Accelerated mappings) Let T : RN N
+ R++ be
of a mapping that is constructed based on antenna tilts, power
a positive concave mapping and M be its lower bounding
allocations, the position of base stations, etc. Optimization of
matrix. If (M ) < 1, the accelerated mapping TA : RN
+
the network performance (e.g., in terms of energy efficiency,
RN
++ of T is the mapping given by:
capacity, coverage, etc.) over the joint set of all control
parameters is typically an NP-hard problem. As a result, many TA (x) := (I M )1 (T (x) M x). (14)
optimization algorithms proposed in the literature are greedy
heuristics that need a fast feasibility check of multiple network To see that the codomain of TA in the above definition is
configurations at each iteration [4]. Proposition 4 opens up indeed RN++ , note that, by Lemma 2, we have that T (x)
the door to the development of efficient and fast methods M x T (0) > 0 for x RN + . Now use Fact 5 to conclude
1
for excluding many infeasible network configurations from that (I M ) (T (x) M x) is a (strictly) positive vector
consideration, which can significantly accelerate the overall for every x RN+.
optimization process. The next lemma shows an alternative way to compute
We emphasize that the converse of Proposition 4 does not accelerated mappings. The main advantage of this alternative
hold in general. There are mappings for which the lower expression is computational. We have to perform only one
bounding matrix has spectral radius strictly less than one, and matrix-vector multiplication.
yet mappings do not have a fixed point (see the application
in Sect. V-B). Therefore, to characterize the existence of a Lemma 3 Let TA : RN N
+ R++ be the accelerated mapping
fixed point based on the spectral radius of the lower bounding of the concave mapping T : RN N
+ R++ , where we assume
matrix, we need additional assumptions on the mapping. The that the lower bounding matrix M of T satisfies (M ) < 1.
next proposition shows a particularly useful assumption that is Then TA in (14) can be equivalently expressed as
satisfied in load estimation problems (see Sect. V-A and [6]
for a particular application of this proposition). TA (x) = (I M )1 (T (x) x) + x. (15)
9

Proof: Recalling that the matrix I M is invertible as a are estimates of the power allocation or of the load at the base
direct consequence of Fact 5, we deduce: stations [4][9]. Therefore, even if the mapping has a fixed
point, the network design is invalid if the power or load of any
TA (x) = (I M )1 (T (x) x) + x
base station exceeds its physical limit. If the iterative algorithm
(I M )TA (x) = T (x) x + (I M )x produces a monotonically increasing sequence, we obtain a
(I M )TA (x) = T (x) M x certificate that the design is invalid as soon as any element of
TA (x) = (I M )1 (T (x) M x). the vector sequence exceeds its limit. It is particularly in these
cases that the standard iteration with the accelerated mapping
TA converges faster than the standard iteration with the original
Positive concave mappings and their corresponding ac-
mapping T , in the following sense:
celerated mappings have many common characteristics. In
particular, they are both standard interference mappings, and Definition 9 (Faster convergence) Let {xn }nN RN and
they have the same fixed point, as shown below. {y n }nN RN be two sequences converging to the same
vector u RN . We say that the sequence {xn }nN RN
Lemma 4 Assume that T : RN N
+ R++ is a positive concave converges faster than {yn }nN RN if kxn u k ky n
mapping with lower bounding matrix M satisfying (M ) <
u k for every n N.
1. Then the accelerated mapping TA : RN N
+ R++ of T
is a standard interference mapping. Furthermore, Fix(T ) = With the above definition, we can now formally state the
Fix(TA ). improvement obtained by using TA instead of T with the
standard iteration in Fact 4.3.
Proof: By (M ) < 1, the matrix inverse (I M )1
exists, and it is a non-negative matrix (Fact 5). As a result, Proposition 6 Assume that T : RN N
+ R++ is a positive
each component of the mapping T (x) := (I M )1 T (x) concave mapping with lower bounding matrix M RN N
(x RN + ) is a positive sum of concave functions, hence satisfying (M ) < 1. Let TA : RN N
+ R++ be the accel-
the resulting function is also concave. Observing that linear erated mapping of T . Consider the following two sequences:
functions are both concave and convex, we verify that each {xn+1 := T (xn )}nN and {xn+1 := TA (xn )}nN . Assume
component of TA (x) = T (x) (I M )1 x + x > 0 is that both sequences start from the same vector u RN +;
a positive sum of concave functions, hence TA is a positive i.e., u = x1 = x1 . If {xn }nN is monotonically increasing
concave mapping. Proposition 1 now shows that TA is a (resp. monotonically decreasing) in each component, then the
standard interference mapping. Consequently, TA has a unique following holds:
fixed point, if it exists (Fact 4.1). If x Fix(T ), then
TA (x ) = (I M )1 (T (x ) M x ) = (I M )1 (x Proposition 6.1 {xn }nN is monotonically increasing (resp.
M x ) = (I M )1 (I M )x = x . The converse is monotonically decreasing) in each component.
also immediate. Note that T (x) = (I M )TA (x) + M x,
hence T (x ) = x if TA (x ) = x , and we conclude that Proposition 6.2 xn xn (resp. xn xn ) for every n N.
Fix(T ) = Fix(TA ).
Proposition 6.3 If the mapping T has a fixed point (which,
The practical implication of Lemma 4 is that, to compute the
in particular, it is automatically guaranteed if {xn }nN is
fixed point of a positive concave mapping T , we can instead
monotonically decreasing in each component), then {xn }nN
compute the fixed point of its accelerated version TA by using
converges faster than {xn }nN to x Fix(T ), in the sense
the standard iteration xn+1 = TA (xn ) shown in Fact 4. In
of Definition 9.
many applications, having a monotone sequence {xn }nN is
desirable, and a sequence of this type can be constructed with Proof: We prove the proposition only for monotonically
the standard fixed point iteration by starting the iterations from increasing sequences (in each component). The proof for
x1 = 0 (see Fact 4.3). For example, in network planning and monotonically decreasing sequences can be obtained in a
optimization tasks, the fixed points of the concave mappings similar fashion by reversing all inequalities.
10

1) Recall that, by Lemma 4, TA is a standard interference that T : RN N


+ R++ is a positive concave mapping with
mapping, so, in light of Fact 4.3, we only need to prove a fixed point denoted by x RN ++ . Since this fixed point is
that TA (u) u if T (u) u. strictly positive, for an arbitrary vector x RN + there always
By assumption, T (u) u 0 and (M ) < 1. In exists > 1 satisfying x x . From Definition 5, we verify
particular, by using Fact 5 and non-negativity of M , that T (x) T (x ) < T (x ) = x . These inequalities
the last inequality implies that (I M )1 is a non- imply that (see also Fact. 4.3) i) T n (0) T n (x) T n (x )
negative matrix. Consequently, from (15), we deduce: for every n N, ii) the sequence {T n (x )}nN is mono-
tonically decreasing, and iii) {T n(0)}nN is monotonically
TA (u) = (I M )1 (T (u) u) + u u.
increasing. In other words, each term of the monotone se-
2) We show the result by using induction. Assume that quences {T n (0)}nN and {T n(x )}nN are, respectively,
xn xn for a given n N. From the definition of the (element-wise) lower and upper bounds for each term of
mapping TA in (14), we deduce: the sequence {T n (x)}nN . All the above arguments are also
valid if we exchange T by its corresponding accelerated
TA (xn ) = xn+1 = T (xn ) + M (xn+1 xn ). (16)
mapping TA , and we note that lower and upper bounding
We have already proved that {xn } is monotonically sequences {TAn (0)}nN and {TAn(x )}nN for the sequence
increasing with the assumptions of the proposition {TAn (x)}nN converge faster to the fixed point x when com-
(hence xn+1 xn 0), M is a non-negative ma- pared to the lower and upper bounding sequences {T n(0)}nN
trix, and T is a mapping satisfying the monotonicity and {T n (x )}nN for the sequence {T n(x)}nN . In other
property of standard interference functions. Using these words, the sequence produced by xn+1 = TA (xn ) with
observations in (16), we verify that: x1 = u RN + arbitrary has sharper element-wise bounds
than the sequence produced by xn+1 = T (xn ) for the same
xn+1 = TA (xn ) T (xn ) T (xn ) = xn+1 .
starting point x1 = u.
The above arguments are valid, in particular, for n = 1,
because x1 = x1 = u by assumption. Remark 3 The price we pay to use the accelerated iteration
3) First recall that both {xn }nN and {xn }nN converge xn+1 = TA (xn ) instead of using xn+1 = T (xn ) is the need
to the uniquely existing fixed point x Fix(T ) for a matrix-vector multiplication, if TA is evaluated by using
(Lemma 4 and Fact 4.3). The desired result kxn (15) (assuming that the lower bounding matrix is not the zero
x k kxn x k , valid for every n N, follows matrix). Furthermore, a matrix inversion is required (or, for
directly from Proposition 6.2. increased numerical stability, a matrix decomposition), but this
operation needs to be done only once. One situation where the
As an immediate consequence of Proposition 6 and Fact 4.3, proposed scheme is particularly useful is when the evaluation
we have the following. of the mapping T is time consuming when compared to the
matrix-vector multiplication. In this situation, for all practical
Corollary 2 Assume that T : RN N purposes, the time to compute xn or xn is roughly equivalent
+ R++ is a positive

concave mapping with x Fix(T ) 6= , and denote by for a given n N sufficiently small. However, for every n N,
T A : RN N xn is guaranteed to be a better approximation of the fixed
+ R++ its corresponding accelerated mapping.
Then the sequence {TAn (0)}nN converges faster to x than point of the mapping T than xn . This situation is common in
the sequence {T n (0)}nN , in the sense of Definition 9 (we network planning.
assume that both sequences start from the vector 0).
V. A PPLICATIONS IN N ETWORK P LANNING AND

Remark 2 Following Yates arguments [1] to prove the con- O PTIMIZATION


vergence of the iteration in Fact. 4.3, we can also argue We now apply the general results in the previous sections
that the proposed accelerated scheme is expected to be fast to two concrete estimation problems in LTE networks. First,
when the initial point is arbitrary. More precisely, assume we consider the load estimation task discussed in [4][7],
11

among other studies. Briefly, the objective is to determine the where


bandwidth required to satisfy the data rate demand of all users f i : RM M
+ R++ R++
in the network, by assuming that the transmit power of all base P dj
(, p)
7 jNi .
stations is given. Ki,j (, p)
The second application we consider is the reverse of load Note that, for each fixed p RM++ and i M, the function
estimation. For a given load allocation at the base stations, M
hp,i : R+ R++ : 7 fi (, p) is concave, hence the
the objective is to estimate the power allocation inducing that solution of (17) with fixed p can be obtained by computing
load. This reverse problem has been motivated by the study in the fixed point of the positive concave mapping given by [5],
[8], which has proved that using all available bandwidth is ad- [7]
vantageous from various perspectives, and, in particular, from
the perspective of transmit energy savings and interference Tp () := [hp,1 (), . . . , hp,M ()]t . (18)
reduction. That study also proves that there exists a standard Therefore, all the theory developed in the previous sections ap-
interference mapping having as its fixed point the solution of plies to this problem, and, in particular, the novel acceleration
the power estimation problem, and the study in [9] has shown schemes for the computation of fixed points. Before proceeding
that the interference mapping can take the form of a positive with numerical examples of the acceleration schemes, we
concave mapping. revisit known results related to this problem, and we show how
the application-agnostic approaches developed in Sect. III and
A. Load estimation
in Sect. IV can be used to reach these known results in a more
We consider an LTE network with M base stations and N
convenient way.
users represented by elements of the sets M = {1, . . . , M }
In particular, the authors of [6] construct a matrix by com-
and N = {1, . . . , N }, respectively. The set of users connected
puting the values that the partial derivatives of the functions
to base station i M is denoted by Ni , and the data rate
hp,1 , . . . , hp,M attain when a given component of the argument
requirement of user j N is given by dj > 0. The propagation
of these functions goes to infinity. It has been shown in [16]
loss between user j N and base station i M is denoted by
that the system of nonlinear equations in (17) has a solution if
gi,j > 0. Each base station i M has K resource units that
and only if the spectral radius of this matrix proposed in [6]
can be assigned to users, and the transmit power per resource
is strictly less than one. Using the terminology and results in
unit for each base station i M is pi > 0. The reliable
Sect. III and in Sect. IV, we note that the matrix suggested
downlink data rate for each resource unit connecting base
in [6] is a particular case of a lower bounding matrix in
station i M to user j N is approximated by the following
Definition 7 constructed with the technique in Proposition 3.
well-established interference-coupling model [4][7], [10]:
The fact that the spectral radius of this lower bounding matrix
! gives sufficient and necessary conditions to characterize the
pi gi,j
i,j (, p) = B log2 1+ P 2
, existence of a solution of the nonlinear system is a direct con-
kM\{i} k pk gk,j +
sequence of the application-agnostic results in Proposition 4
where 2 is the noise power per resource unit, p = and Proposition 5.
[p1 , . . . , pM ]t is the downlink power vector per resource unit, To be more precise, we can use (9) to construct the lower
and = [1 , . . . , M ]t is the load vector. Here, the load i is bounding matrix M p of the mapping Tp as follows:
fraction of the number of resource units in the time-frequency
grid that users in the set Ni require from base station i. For Mp =

fixed power allocation p RM ++ , the load is the solution to limh g11 ( + he1 ) limh gM
1
( + heM )
.. .. ..
the following system of nonlinear equations [4][7]: ,
. . .
1 = f1 (, p) limh g1M ( + he1 ) limh gMM
( + heM )
..
. (17)
where RM i
++ is arbitrary and gk () is the kth component
M = fM (, p), of a supergradient of the function hp,i at an arbitrary point
12

= [1 , . . . , M ]t . By noticing that the function hp,i is


100
differentiable in the interior of its domain, gki () is simply T
101 Tp A
the partial derivative hp,i () for every RM ++ . As a
k
result, we can verify that the lower bounding matrix of the 102

Normalized mean error


mapping Tp is given by M p = diag(p)1 M diag(p), where 103

104

0, if i = k
[M ]i,k = P ln(2)d j g k,j (19) 105

jNi otherwise.
KBgi,j 106

(We can also obtain (19) by constructing the lower bounding 107

matrix with the approach in (8).) 108


By Remark 1, (M p ) < 1 is equivalent to (M ) < 1, and 109
0 2 4 6 8 10
we note that M does not depend on the power allocation p, Iteration
a fact originally stated in [16]. Therefore, to verify whether
Fig. 1. NME of the load estimate as a function of the number of iterations.
the mapping Tp has a fixed point by using the results in Confidence intervals (95%) have been computed, but they are not visible in
Proposition 4 and Proposition 5, we can compute (M ) the figure.
instead of (M p ). In other words, knowledge of (M )
is sufficient to determine whether the system of nonlinear
equations in (17) has a solution, as already stated in [16] for where Fix(Tp ). We approximate the expectation operator
this particular application. by averaging the results of 100 runs of the simulation, and
Having the lower bounding matrix in closed form, we in each simulation the positions of the users (and hence the
can now proceed to the numerical evaluations of the novel propagation loss) are the random variables. All iterations start
acceleration schemes. In the simulations we show here, we from the zero vector, and networks where the corresponding
compare the accuracy of the load estimates generated by the concave mapping does not have a fixed point are discarded.
standard iteration n+1 = Tp ( n ) with its accelerated version Therefore, the expectation in (20) is conditioned to the fact
n+1 = Tp A ( n ). Table I lists the main parameters of the that spectral radius of the lower bounding matrix is strictly
network. smaller than one.
TABLE I. N ETWORK PARAMETERS OF THE SIMULATION Fig. 1 shows results obtained by using the iterative scheme
Parameter Value
in Fact 4.3 with the original mapping Tp and with its proposed
Carrier frequency 900 MHz
Number of resource units (K) 25 accelerated version Tp A . We verify that the mapping Tp A
Transmit power per resource unit (pi , i M) 1.6W requires fewer iterations than Tp to obtain a given numerical
System bandwidth (K B) 5 MHz
precision, which is an expected result by considering Proposi-
Noise power spectral density -145.1 dBm/Hz
Propagation model Okumura-Hata, urban tion 6.
Antenna height of base stations 30m
Antenna height of the users 1.5m
Number of users (N ) 200
Number of base stations (M) 25
Data rate of each user (dj , j N ) 768 kbps
Dimension of the field 2500m2500m B. Power estimation
User distribution Uniformly distributed at random
Base station distribution Uniformly distributed
We now turn our attention to the problem of power esti-
The figure of merit used in the comparisons is the expected mation in LTE networks. The objective is to solve (17) for
normalized mean error (NME), which we define by p = [p1 , . . . , pM ]t with the load RM ++ being the fixed
parameter. It is shown in [9] that the solution of this nonlinear
eNME () := E[k k/k k], (20) system is the fixed point of the positive concave mapping given
13

by T (p) := [h,1 (p), . . . , h,M (p)]t , where


p P 100
i dj T

jNi , if pi 6= 0 T A

i K i,j (, p)
P dj ln 2 P  101
h,i (p) := p g + 2
,

jNi
KBgi,j i kM\{i} k k k,j

Normalized mean error





otherwise. 102

By using (8) to construct the lower bounding matrix M


of the mapping T , we deduce: 103

104
M =

limx0+ xh,1 (x1 e1 ) limx0+ xh,1 (x1 eM ) 105
.. .. .. 0 50 100 150 200
. Iteration
. .
limx0+ xh,M (x1 e1 ) 1
limx0+ xh,M (x eM ) Fig. 2. NME of the power estimate as a function of the number of iterations.
1 Confidence intervals (95%) have been computed, but they are not visible in
= diag() M diag(),
the figure.
where M is the same matrix defined in (19). (The same result
can be obtained by using Proposition 3 to construct the lower provides us with clear advantages over the standard iterative
bounding matrix, but here applying Proposition 2 is easier than approach, in accordance to the analysis in Sect. IV-B.
applying Proposition 3.)
From Proposition 4 and the definition of M , we conclude
VI. C ONCLUSIONS
that a necessary condition for existence of the fixed point of
T is (M ) < 1, which is the same requirement for the We have shown that the results in [6] for the construction
existence of the fixed point of Tp in (18). However, there of lower bounding matrices in a very particular application
is a fundamental difference between these two mappings. As domain can be generalized to a large class of positive concave
proved in [16], (M ) < 1 (note: this spectral radius does not mappings where even differentiability is not required. More
depend on ) is both a sufficient and necessary condition for specifically, we proved that positive concave mappings with
the existence of the fixed point of Tp . In contrast, the study nonempty fixed point set can be associated with a non-negative
in [8] has shown that the existence of the fixed point of T lower bounding matrix having spectral radius strictly smaller
also depends on . Therefore, T is an example of a mapping than one. By imposing additional assumptions on the mapping,
proving that the converse of Proposition 4 does not hold in having spectral radius strictly smaller than one also implies
general. the existence of the fixed point of the concave mapping.
We now turn the attention to the acceleration schemes in this We also demonstrated that the lower bounding matrix can
particular application. We use the same network considered be constructed with two simple and equivalent methods, and
in the load estimation task. The desired load is obtained this matrix can be combined with its generating concave
by solving (17) with the power fixed to the value shown in mapping to build a new mapping that preserves the fixed
Table I. Then we solve the reverse problem; we compute point. The standard fixed point iterations applied to this new
the power shown in Table I by using the standard iteration mapping typically requires fewer evaluations of the original
pn+1 = T (pn ) and its accelerated version pn+1 = T A (pn ). mapping to obtain an estimate of the fixed point for any given
Both algorithms start from the zero vector. The normalized precision. The additional computational complexity of this
mean error is again used as the figure of merit (which in novel approach is very modest. In the tasks of load and power
this application is defined by replacing the load vector by estimation in LTE networks, where we are mostly interested
the power vector in (20)). We can see in Fig. 2 that in in the precision of the estimates after a limited number of
this application the proposed acceleration scheme once again iterations, numerical examples show that the improvement in
14

convergence speed obtained with the proposed method can be [15] I. Siomina and D. Yuan, Load balancing in heterogeneous LTE:
substantial. Range optimization via cell offset and load-coupling characterization,
in Communications (ICC), 2012 IEEE International Conference on.
IEEE, 2012, pp. 13571361.
R EFERENCES
[16] C. Ho, D. Yuan, and S. Sun, Data offloading in load coupled networks:
[1] R. D. Yates, A framework for uplink power control in cellular radio A utility maximization framework, IEEE Trans. Wireless Commun.,
systems, IEEE J. Select. Areas Commun., vol. 13, no. 7, pp. pp. 1341 vol. 13, no. 4, pp. 19211931, April 2014.
1348, Sept. 1995.
[2] S. Stanczak, M. Wiczanowski, and H. Boche, Fundamentals of Resource
Allocation in Wireless Networks, 2nd ed., ser. Foundations in Signal
Processing, Communications and Networking, W. Utschick, H. Boche,
and R. Mathar, Eds. Berlin Heidelberg: Springer, 2009.
[3] M. Schubert and H. Boche, Interference Calculus - A General Frame-
work for Interference Management and Network Utility Optimization.
Berlin: Springer, 2011.
[4] K. Majewski and M. Koonert, Conservative cell load approximation
for radio networks with Shannon channels and its application to
LTE network planning, in Telecommunications (AICT), 2010 Sixth
Advanced International Conference on, May 2010, pp. 219 225.
[5] A. Fehske, H. Klessig, J. Voigt, and G. Fettweis, Concurrent load-
aware adjustment of user association and antenna tilts in self-organizing
radio networks, IEEE Trans. Veh. Technol., no. 5, June 2013.
[6] I. Siomina and D. Yuan, Analysis of cell load coupling for LTE network
planning and optimization, IEEE Trans. Wireless Commun., no. 6, pp.
22872297, June 2012.
[7] R. L. G. Cavalcante, S. Stanczak, M. Schubert, A. Eisenblater, and
U. Turke, Toward energy-efficient 5G wireless communication tech-
nologies, IEEE Signal Processing Mag., vol. 31, no. 6, pp. 2434, Nov.
2014.
[8] C. K. Ho, D. Yuan, L. Lei, and S. Sun, Power and load coupling
in cellular networks for energy optimization, IEEE Trans. Wireless
Commun., vol. 14, no. 1, pp. 500519, Jan. 2015.
[9] R. L. G. Cavalcante, E. Pollakis, and S. Stanczak, Power estimation
in LTE systems with the general framework of standard interference
mappings, in IEEE Global Conference on Signal and Information
Processing (GlobalSIP 14), Dec. 2014.
[10] K. Majewski, U. Turke, X. Huang, and B. Bonk, Analytical cell load
assessment in OFDM radio networks, in Proc. IEEE PIMRC07, 2007,
pp. 15.
[11] P. Mogensen, W. Na, I. Z. Kovacs, F. Frederiksen, A. Pokhariyal, K. I.
Pedersen, T. Kolding, K. Hugl, and M. Kuusela, LTE capacity com-
pared to the Shannon bound, in the 65th IEEE Vehicular Technology
Conference, VTC Spring 2007, Apr. 2007.
[12] A. Auslender and M. Teboulle, Asymptotic Cones and Functions in
Optimization and Variational Inequalities. New York: Springer, 2003.
[13] H. H. Bauschke and P. L. Combettes, Convex Analysis and Monotone
Operator Theory in Hilbert Spaces. Springer, 2011.
[14] J. Kennan, Uniqueness of positive fixed points for increasing concave
functions on Rn: an elementary result, Review of Economic Dynamics,
vol. 4, no. 4, pp. 893899, 2001.