Professional Documents
Culture Documents
Honglei Xu, Song Wang, Soon-Yi Wu (Eds.) - Optimization Methods, Theory and Applications-Springer (2015)
Honglei Xu, Song Wang, Soon-Yi Wu (Eds.) - Optimization Methods, Theory and Applications-Springer (2015)
Soon-Yi Wu Editors
Optimization
Methods,
Theory and
Applications
Optimization Methods, Theory and Applications
Honglei Xu • Song Wang • Soon-Yi Wu
Editors
Optimization Methods,
Theory and Applications
123
Editors
Honglei Xu Song Wang
Department of Mathematics and Statistics Department of Mathematics and Statistics
Curtin University Curtin University
Perth, WA, Australia Perth, WA, Australia
Soon-Yi Wu
Department of Mathematics
National Cheng Kung University
Tainan, Taiwan
v
vi Preface
vii
viii Contents
1.1 Introduction
The study of human motion has been of considerable interest in the field of
biomechanics. It provides detailed information to understand the human movements
that enable certain motions to be improved or made safer. The desire to understand
the mechanics behind walking has spurred the study of human locomotion (McGeer
1988) but due to its complex nature, modelling and understanding human walking
continues to be a challenging research problem in multibody systems (Hardt et al.
1999). Movement patterns can be predicted as best as possible using mathematical
models, however models can become very complicated while trying to model a
human body and its movements as closely as possible due to a body’s complexity
(Alexander 1996, 2003).
In order to replicate human walking motion realistically, a complete gait cycle,
comprising of two continuous steps, should be considered. Each step is made up
of two phases namely, single support phase and double support phase. The single
support phase occurs when one foot contacts the ground while the other leg is
swinging from rear to front, starting from the rear foot toe-off and ending when
the swinging foot lands on the ground with a heel strike. The double support phase
begins with the heel strike of the forward swing foot and ends with the toe-off of the
rear foot. As such, Bessonnet et al. (2004) described single support phase as moving
like an open tree-like kinematic chain while double support phase is kinematically
closed and overactuated.
The kinematic configuration of the model biped may change going through
one phase to the other during the collision of the foot with the ground which
results in jump conditions on the velocities (Hardt et al. 1999). This could have
been a contributing reason as to why most early research only considered single
support or assumed an instantaneous double support phase (Ren et al. 2007).
Instantaneous double support phase was being considered for walking simulations
in later research, followed by studies which considered a complete step consisting of
both the single support and double support phases (Xiang et al. 2010). In addition,
the foot segment was often neglected or assumed to be flat on the floor during
stance.
In order to achieve realistic human walking motions, both phases are necessary
and should be incorporated in the model. Modelling a biped with feet will allow
the modelling of the double support phase, from swing heel strike to stance toe off.
Hardt et al. (1999) suggested that feet bring about the addition of ankle torques and
liftoff force that is produced as the heel comes off the ground. They contemplated
that at a higher speed, the biped cannot walk effortlessly without the inclusion of
a foot. The addition of ankle actuation generated a smoother walking motion and
allowed torque inputs at hip to be distributed to the knees and ankles. Another
advantage of having a foot as a segment was the ability to distribute center of
pressure from the rear to the front of the foot during ground contact (Bullimore
and Burn 2006).
The double support phase of the gait cycle is deemed a common example of a
closed-loop problem. In a closed-loop model such as the double support phase, the
number of actuating torques is usually more than the number of degrees of freedom,
presenting a redundancy problem for inverse dynamics and control applications to
resolve (Ünver et al. 2000).
The inverse dynamics problem of biomechanics has been the most common
method used to estimate muscle forces during locomotion (Anderson and Pandy
2001) since it is computationally inexpensive and solutions can be obtained
relatively quickly on single-processor computers. However, Marshall (1985) and
Selles et al. (2001) both acknowledged that a major problem in the inverse dynamics
approach is the need for numerical differentiation of potentially noisy position data.
Even though inverse dynamics have been commonly used to estimate joint torques
during locomotion (Anderson and Pandy 2001), they produce poor results in the
presence of noisy measurements (Kuo 1998). In addition, to model a complete gait
cycle, both single and double support phases have to be considered. However, the
redundancy problem of the double support phase, leads to a state of indeterminacy,
1 Analysing Human Walking Using Dynamic Optimisation 3
unless ground reaction forces are known. The inverse dynamics method is not able
to directly solve the equations of motion for that phase as seen in Ren et al. (2007)
model.
The dynamic optimisation method integrates the equation of motion during an
optimisation process to simulate motions and solve for the optimal joint forces of
the model (Chow and Jacobson 1971; Anderson and Pandy 2001; Pandy 2001). Esti-
mating the trajectories of joint torques using this method is practical as the method
applies a forward simulation to reproduce a best observed motion (Chao and Rim
1973). A major disadvantage of dynamic optimisation is that it is more expensive
computationally (Yamaguchi and Zajac 1990), and hence has led to solutions for
walking being greatly simplified. A large amount of computation is required to
compute the trajectory of joint torques when using dynamic optimisation due to
the choice of initial guesses of torque value and the mathematical sophistication
required to understand the technique (Koh and Jennings 2003). However, as a
general rule of dynamic optimisation, an initial estimate set of torque trajectories is
required to start the optimisation. In Koh (2001), Koh used the conventional method
of inverse dynamics to find the torque trajectories as initial estimates to speed up
the convergence of the numerical process (Chao and Rim 1973). Koh had managed
to incorporate the use of both methods by applying the inverse dynamics method
to determine the initial torque estimates to be used in the dynamic optimisation
method.
Most research tends to avoid modelling double support phase as it, being a
closed loop problem, produces redundancy problem and complicates the modelling.
Studies that consider modelling both single and double support phases have chosen
to employ the method of inverse dynamics and optimsation to solve the system of
both phases, and had two different algorithms to deal with each phase. The purpose
of this study is to develop a generalised model applying optimal control theory, to
simulate normal walking motion through both the single support and double support
phases, and to gain insight into the mechanics that are involved in the overall motion.
Since normal walking can be assumed to have symmetric and cyclic characteristics,
only one step of the gait cycle needs to be modelled (Hardt et al. 1999; Anderson and
Pandy 2001; Xiang et al. 2009) with appropriate semi-periodic boundary conditions.
A combination of both inverse dynamics method and dynamic optimisation method
will be used to solve the equations of motion to model the human motion. Joint
torques and forces obtained will be studied to understand the mechanics behind
certain movements.
One major contribution is to have one model able to solve for one step cycle in
walking. The model formulated in this study can be adjusted according to cases,
hence the single support and double support phases can be solved for individually
or together. As the aim of the study is to simulate walking for one step cycle, both
phases are considered and solved.
An advantage the study brings is the application of the conventional inverse
dynamics method to obtain initial estimates of joint torques to be used in dynamic
optimisation, which is the approach taken in this study to produce the observed
motion and improve joint torques estimates. An initial joint torques “guess” from
4 M. Tan et al.
applying the inverse dynamics method can reduce computation time in computing
joint torque trajectories. In addition, dynamic optimisation solves the problem by
integrating the equation of motion forward in time, so simulation is performed
in a manner consistent with the development of motion in humans, allowing the
research to evaluate the effects of changes in ‘muscle activity’ on the outcome of
the movement.
The rest of this paper is organised as follows. In the next section, we will first set
up the geometry and the mathematical model of the walking model. In Sect. 1.3, we
will establish the objective function, its corresponding constraints and discuss the
inverse analysis used to obtain the initial joint estimates. In Sect. 1.4, we present
the analysis of numerical experiments to confirm that the mathematical model
formulated replicated human walking motions.
1.2.1 Geometry
the equation used to calculate moments of inertia of a rod, with the axis of rotation
at the CoM of the segment.
The ith segment has length `i , mass mi and moment of inertia Ii about its center
p p
of mass (CoM)(within the segment), which is a distance ri from proximal .xi ; yi /
and distance li ri from the distal .xi ; yi / end (Fig. 1.1). All segments are labelled
d d
to have a proximal and distal end. The proximal end of a segment is situated closest
to point of contact while the distal end of a segment is situated furthest from point
of contact. In this case, we have considered the point of contact to be the ball of the
stance foot. There is a global coordinate system, XOY, which has an origin fixed in
an infinite mass “ground”. Each segment’s position is known from its CoM (xi , yi )
and the angle i that the segment makes with the positive x-axis. The CoM of the
whole body, (X, Y), is given by
Pn Xn
X 1 mi x i
D Pn
iD1 where M D mi :
Y M iD1 mi yi iD1
Walk Direction
Segment 7
(Head, Arms, Trunk)
7 6
Segment 4 Segment 3
(Swing Thigh) (Stance Thigh)
4 3
Segment 5
(Swing Shank)
Segment 2
(Stance Shank)
5
Segment 6 2
(Swing Foot) Segment 1 1
X (x6d,y6d) (Stance Foot)
(x1 p,y1 p)
and
2 3
r1 0 0 0 0 0 0
6l 0 0 0 0 07
6 1 r2 7
6l 0 0 0 07
6 1 l2 r3 7
6 7
L D 6 l1 l2 l3 r4 0 0 0 7:
6 7
6 l1 l2 l3 l4 r5 0 07
6 7
4 l1 l2 l3 l4 l5 r6 05
l1 l2 l3 0 0 0 r7
The position of CoM of the whole system (MX, MY) can likewise be written in
matrix-vector form as MX D mt x and MY D mt y, where mt D .m1 ; : : : ; mn /. Hence
the relations, using mt e D M,
p p
MX D mt x D Mx1 C mt LDc e and MY D mt y D My1 C mt LDs e:
6
X 6
X
p p
xd6 D x1 C li cos i and yd6 D y1 C li sin i :
iD1 iD1
The segments are ordered so that L is lower triangular despite having multiple
branches.
To describe the topology of the body, each joint is labelled with a number, k, k D
1; : : : ; j where j D n C 1 for a body with no loops, in other words, a tree-structured
body. Some of these joints will be in contact with the ground while some are free
or constrained on a curve. This allows for externally applied forces on any joint, in
particular, from the infinite mass ground. The proximal incidence matrix is a j n
matrix Ap where
(
p 1; if segment i has proximal end at joint k;
Aki D
0; otherwise:
8 M. Tan et al.
These two matrices define the topology of the body and Ap C Ad defines the vertex-
arc, or joint-directed segment incidence matrix of a digraph (directed graph). The
joint-external contact incidence matrix B (j e) is defined as
(
1; if joint k contacts the ground at external contact i;
Bki D
0; otherwise;
where e is the number of external contacts. The possible contacts considered are at
joints numbered 1, 2, 6 and 7 which are the toes and heels of the body, where for
model simplicity, the heels are a rigid extension of the ankles.
Suppose we have a joint k, which has two proximal ends, namely i and b incident
on it and one distal end a incident on it. An external force f ek acts on the joint. The
reaction forces on segment i come from joints k proximal and .k C 1/ distal and are
p
denoted f i and f di for the proximal and distal forces respectively. It is these forces
which supply the rotational and translational motions to segment i and are given by
px dx
fkex p fi fi
f ek D ey ; f i D py ; f d
i D dy :
fk fi fi
The balance of forces at any joint k takes the form, “the sum of all reaction forces
at joint k equals the external force at joint k”. From Fig. 1.3,
px px dx ex
fi fb fa fk
py C py C C ey D 0:
fi fb fady fk
Using the incidence matrices, these equations for all joints can be combined into
matrix equations:
Ap f px C Ad f dx C B.f ex / D 0;
Ap f py C Ad f dy C B.f ey / D 0;
Af x D 0;
Af y D 0;
1 Analysing Human Walking Using Dynamic Optimisation 9
where the proximal, distal and external forces are ordered such that
2 3 2 3
f px f py
f x D 4 f ex 5 ; f y D 4 f ey 5 :
f dx f dy
When there are no external contacts at any joint, the external forces are zero.
P C Sf x D Jx !2 ;
mPu1 C Jy !
P C Sf y D gm J y !2 :
mvP 1 J x !
10 M. Tan et al.
To cater for both contact and free flight dynamics, where free flight dynamics is
p p
described when there is no external contact to the ground, during contact at .x1 ; y1 /,
.Pu1 ; vP 1 / is zero and non-zero when not in contact. Hence the translational equations
during contact are:
P C Sf x D Jx !2 ;
Jy !
P C Sf y D gm Jy !2 :
Jx !
p
In the case of non-contact, we would require the dynamic equations, xR 1 D uP 1 and
p
yR 1 D vP1 to be included.
P !
We have defined: D .1 ; : : : ; n /t , ! D , R !2 D .! 2 ; : : : ; !n2 /t ,
P D , 1
S D ŒIn ; 0; In .
Note that Jx D Dm LDc and J y D Dm LDs , where Dm D diag.m/.
For the general case, the moment equations of each segment i about the segment
CoM are given by
p px py dy
Ik !P i D i C id C fi ri sin i fidx .li ri / sin i fi ri cos i C fi .li ri / cos i ;
P C Mx f x C My f y D T;
J!
where J D diag.I1 ; I2 ; : : : ; In /,
Mx D Ds Dr j 0 j Dl Dr ,
My D Dc Dr j 0 j .Dl Dr / ,
Dr D diag.r1 ; r2 ; : : : ; rn /, Dl D diag.l1 ; l2 ; : : : ; ln /.
The external forces do not appear explicitly in these equations as they act on the
joints and transfer the forces to the segments attached to the joint.
The vector is a vector of the proximal torques appropriately ordered. These
torques can be considered as making the angle between the segments larger or
smaller. It is possible for a torque to act on a segment separated by other segments.
In this case, we assume that only one torque acts between two segments about a
common joint and that the torques are given functions of time.
A consistent notation to specify the torques needs to be established and is given
by the matrix T, which is a matrix of zeros, ones and negative ones that describes
which segments have torques acting on them. A positive torque on the distal end of a
segment contributes a negative angular acceleration to the segment, while a positive
torque on the proximal end of a segment contributes a positive angular acceleration.
1 Analysing Human Walking Using Dynamic Optimisation 11
This holds true up till Segment 6, as Segment 7 is part of a multiple branch at joint
4, together with Segment 3 and Segment 4 (Fig. 1.2). At joint 4 (hip joint), we have
a torque between Segment 3 and Segment 7, and another between Segment 4 and
Segment 7, which will be opposite of each other.
The matrix T with a torque acting between Segment 6 and the external world is
given by
1 2 3 4 5 6 7
seg1 Œ 1 0 0 0 0 0 0
seg2 Œ 1 1 0 0 0 0 0
seg3 Œ 0 1 0 0 0 1 0
TD :
seg4 Œ 0 0 1 0 0 0 1
seg5 Œ 0 0 1 1 0 0 0
seg6 Œ 0 0 0 1 1 0 0
seg7 Œ 0 0 0 0 0 1 1
Note that the sum of the each column and row has to be zero with the exception of
5 and Segment 6 in the case of one contact at Segment 1. 1 acts between Segment
1 and the external world and there is no torque between Segment 6 and the external
world.
The equations for the second order variables and forces are looked at, taking account
of variables which are zero for a time interval.
The complete equations where eight rows of the equations are included or not,
to specify the cases, are presented. Hence all dynamic variables and all forces are
included, whether zero or not. But first, the two row vectors lts and ltc are defined so
as to relate distal Segment 6 (toe, or joint 7) to proximal Segment 1 (toe or joint 1),
or proximal Segment 2 (ankle (heel) or joint 2),
1l
t
D Œl1 ; l2 ; l3 ; l4 ; l5 ; l6 ; 0; or 2l
t
D Œ0; l2 ; l3 ; l4 ; l5 ; l6 ; 0;
t
1 ls D 1 l t Ds ; or t
2 ls D 2 l t Ds ;
t
1 lc D 1 l t Dc ; or t
2 lc D 2 l t Dc ;
p
uP d6 uP 1 1 lts !P 1 ltc !2
D p C C :
vP 6d vP 1 1 lc !
t
P 1 lts !2
P D !;
p
xP 1 u1
p D ;
yP 1 v1
d
xP 6 u
D 6 ;
yP d6 v6
n 1 1 1 1 n n1 n n1 2 3
n Œ J 0 0 0 0 Mpx Mdx M py Mdy 2 3 T
!P 6 7
1 Œ lts 1 0 1 0 0t 0t 0t 0t 6 uP 1 7 6 ltc !2 7
6 7 6 t 2 7
1 Œ ltc 0 1 0 1 0t
0t 0t
0t 6 vP 7 6 l ! 7
6 17 6 s 7
.a/ 1 ŒŒ 0t 1 0 0 0 0t
0t 0t
0t 6 uP 7 6 0 7
6 67 6 7
.a/ 1 ŒŒ 0t 0 1 0 0 0t 0t 0t 0t 6 7 6 0 7
6 vP 6 7 D 6
6
7:
7
.b/ 1 ŒŒ 0t 0 0 1 0 0t
0t 0t
0t 6 px 7 6 0 7
6f 7 6 7
.b/ 1 ŒŒ 0t 0 0 0 1 0t
0t 0t
0t 6 dx 7 6 0 7
6f 7 6 7
n Œ Jy m 0 0 0 I I 0 0 6 py 7 6 J !2 7
x
4f 5 6 7
n2 Œ 0 0 0 0 0 A Ad
p
0 0 6 0 7
f dy 6 y 2 7
n Œ Jx 0 m 0 0 0 0 I I 4 J ! C mg 5
n2 Œ 0 0 0 0 0 0 0 Ap Ad 0
The above can be adjusted further to accommodate different cases such as the
single and double support phase.
The cases are broken down into:
• Case 1 (double-support phase) where Case 1A considers a scenario of two
external contacts at joint 2 and 7, and Case 1B considers a scenario of two
external contacts at joint 1 and 6
• Case 2 (free flight) considers a scenario where there is no contact with the ground
• Case 3 (single support phase with one contact at swing heel) where Case 3A
considers a scenario of one contact point at proximal Segment 2 and Case 3B
considers a scenario of one contact point at proximal Segment 6
• Case 4 (single support phase with one contact at stance toe) where Case 4A
considers a scenario of one contact point at proximal Segment 1 and Case 4B
considers a scenario of one contact point at proximal Segment 6.
These cases, showing how the matrices are modified depending whether the
weight is on the heel or toe of the foot can be found in the Appendix.
1 Analysing Human Walking Using Dynamic Optimisation 13
This study applied the conventional inverse dynamics (Winter 2009) to obtain initial
estimates of joint torques which were used in dynamic optimisation. Dynamic
optimisation approach was adopted to compute the trajectory of joint torques and to
produce the observed motion. Since we were only concerned with the single support
phase and double support phase of walking and would be considering Segment 1 to
begin from stance foot, hence our focus would be on Case 4A for single support
phase and Case 1B for double support phase.
An alternative formulation of the conventional method of inverse dynamics was
used to determine the initial estimates of joint torques, and can be found in detail
later in this section. It is used so that a more realistic set of joint torques histories
is generated. Using torque histories specific to the movement pattern of the subject
improves the convergence of the optimisation (Chao and Rim 1973). The bounds for
each control are determined using the maximum and minimum estimates obtained
from the inverse dynamics approach. Since these estimates provided approximate
torque histories that were specific to the movement pattern, it would ensure that the
optimised torque trajectories are realistic.
For the present study, 18 states (x D Œx1 ; x2 ; : : : ; x18 > ), 15 system parameters
(z D Œz1 ; z2 ; : : : ; z15 > ), and 7 controls ( D Œ1 ; 2 ; : : : ; 7 > ) namely joint torques,
were set up in MISER3.3 (Jennings et al. 2000). The 18 states consist of the angular
displacements from Segment 1 to Segment 7 (xi D i ; i D 1; : : : ; 7), angular
velocity (xi D i ; i D 8; : : : ; 14), coordinate and velocity of proximal end of
p p p p
segment one, (.x15 ; x16 ; x17 ; x18 / D .x1 ; y1 ; xP 1 ; yP 1 /). The system parameter consists
of the initial segment angular orientation (i .0/ D zi ; i D 1; : : : ; 7), initial angular
velocity at start of single support phase (!i .0/ D z7Ci ) and z15 is the step length
p
which is twice the distance of initial distance of x1 .0/ and xd6 .0/ and hence dependent
on (z1 ; : : : ; z6 ). Placing tight bounds on the system parameters allows the initial
conditions to vary by a small amount about the initial values of the data.
The variables 1 ; 2 ; 3 describe the angular displacements of the stance leg
beginning with the foot, shank, and thigh segments respectively; 4 ; 5 ; 6 describe
the angular displacements of the swing leg from thigh, shank and foot segments
respectively and 7 describes the angular displacement of the trunk segment.
p p
!1 ; : : : ; !7 are the segments’ corresponding velocities; and (x1 ; y1 ) are the coor-
dinates of the proximal end of segment one (toe of stance foot) which remains
stationary on the ground during one step of the walk cycle.
There were two parts involved in the experiment to simulate normal walking and
obtaining more precise joint torque estimates. For the first part, forward dynamics
of seven segment model during the single support phase (Case 4A) was optimised
for the joint torques and initial values of and ! such that computed .; z; t/
trajectories produced motion similar to normal walking. In doing so, the body had
to be kept upright and from falling under gravity. This was done by keeping the
y-coordinate CoM close to the initial y-coordinate CoM (at t D 0). Our initial
objective function is thus given by,
14 M. Tan et al.
Z T1
G0 .; z/ D .CoMypos CoMyinit /2 dt
0
where T1 (D 0:386 s) is the duration of the single support phase, CoMypos is the
center of mass of y-coordinate, a function of .; z; t/, and CoMyinit is the initial
center of mass of y-coordinate, a function of z, as calculated by MISER3.3.
The objective function is subject to constraints in the canonical form:
Z tk
D
Gk .u; z/ D k .x.tk ; z// C gk .t; x.t/; u.t/; z/ dt 0; k D 1; : : : ; ngc ;
0
5
X
a12 D li sin.i .T1 // C 0:1 sin.6 .T1 / 1:57/I
iD1
6
X
a22 D 0:0751 li sin.i .T1 //:
iD1
Remark. The sum of squares of two constraints equal to zero was used to
reduce the total number of constraints and to allow more leeway in reducing
the individual constraints to zero.
1 Analysing Human Walking Using Dynamic Optimisation 15
• Two all-time constraints on the swing toe and swing heel, such that the swing
foot does not penetrate the ground, are given by,
6
X
3 0; h3 ..t/; z; t/ D li sin.i .t// 0; and
iD1
5
X
4 0; h4 ..t/; z; t/ D li sin.i .t// C 0:1 sin.6 .T1 / 1:57/ 0:
iD1
• Two all-time constraints on knee, to prevent the knee from hyperextending, are
given by,
Remark. There is a possibility that segmented rigid body models moving under
the effect of gravity can spin past a natural limit during optimisation computation
in the line search. Said et al. (2006) found that hyperextension modelling done
“automatically” in the dynamics, introduced a large force to be exerted to restore
the joint if it gets close to hyperextension. However, this was not used in this
research as it creates “stiff” differential equations.
• An all-time constraint on trunk, to prevent the trunk from falling forward or
backward, is given by,
• Two constraints involving system parameters, to ensure stance foot is flat on the
ground and to position swing foot at time t D 0, is given by.
Z T1
G0 .; z/D .CoMypos CoMyinit /2 C 1;000.jankle1 j C jankle2 j C jankle3 j C jankle4 j/ dt
0
where (ankle1 ; ankle2 ; ankle3 ; ankle4 ) are the ankle constraints and only the con-
straints that are not satisfied are taken in the objective function. The final set of
controls from this optimisation was then used as the initial joint torques estimates
for the optimisation studies in the second part of the experiment.
In the second part of the experiment, the second phase of the walk cycle, double
support phase (Case1B), was incorporated with the first part for the rest of the time
interval (T D 0:486 s). Forward dynamics of seven segment model was optimised
now to simulate normal walking for both single support and double support phase
of walking. However, as the forward dynamics changed from single support phase
to double support phase, the collision of the foot with the ground resulted in jump
conditions on the model velocities (Hardt et al. 1999). The dynamics in the model
computed in MISER3.3 allowed the state to jump at particular times j , hence the
state equations have a form:
8 0
ˆ f 1 .t; x; u; z/; t 2 Œts ; 1 /; x.ts / D x .z/;
ˆ
ˆ
< f 2 .t; x; u; z/; t 2 Œ1 ; 2 /; x.1 / D h1 .x. /; z/;
1
xP .t/ D ::
ˆ
ˆ :
:̂
f p .t; x; u; z/; t 2 Œp1 ; p /; x.p1 / D hp1 .x.p1 /; z/;
where x.j / are the new states at time j . As the ankle constraints were satisfied
in the first part of the experiment, they were removed from the objective function.
1 Analysing Human Walking Using Dynamic Optimisation 17
Thus, the objective function is now similar to the original function, given by,
Z Tf
G0 .; z/ D .CoMypos CoMyinit /2 dt
0
5
X
a12 D li sin.i .T1 // C 0:1 sin.6 .T1 / 1:57/:
iD1
Remark. Foot length was defined to be from heel to toe (D 0:2368 m) and
was estimated from segment parameters, Segment 1/6 and length of heel (from
ankle), since the foot segment was assumed to take on the shape of a right-angled
triangle.
• Two terminal constraints at the end of double support phase (Tf D 0:486 s), such
that the swing foot is on the ground and does not slide, are given by,
5
X
b12 D li sin.i .Tf // C 0:1 sin.6 .Tf / 1:57/I
iD1
g4 0; 4 ..Tf /; z/ D z7 7 .Tf / D 0:
6
X
gz2 .z/ D z15 li cos.zi / D 0:
iD1
In order to solve for initial joint torques for dynamic optimisation, the conventional
method of inverse dynamics (Winter 1990) was adopted. The dynamics equation
in Case 4A and Case 1B were rearranged to solve for joint moments and reaction
forces from kinematic data of segments.
Rearranging Case 4A:
2 32 3 2 3
T M px Mdx Mpy Mdy J!P
6 0 0 7 6 px 7 6 7
0 7 6 f 7 6 J !2 J ! P
x y
6 I I 7:
6 7 6 dx 7 6 7
6 0 Ap A d
0 0 76f 7 D 6 0 7
6 7 6 py 7 6 y 2 7
4 0 0 0 I I 5 4 f 5 4 J ! C mg C Jx !P5
0 0 0 Ap Ad f dy 0
1 Analysing Human Walking Using Dynamic Optimisation 19
The coefficient matrix is now a block upper triangular and was obtained easily
at each discrete data time, ti (static inverse analysis), after solving for reaction
force vectors. As the matrix is square and invertible, a unique set of torques was
computed. This meant that for n segments, there should be n torques between
segments. However, this was not the same for Case 1B.
Even though the dynamic equations in Case 1B is a square matrix, when
rearranged, it becomes a non-invertible matrix with two rows of zeros.
Rearranging Case 1B:
2 32 3 2 3
T 0 0 M px Mdx Mpy Mdy J!P
6 0t 0t 7 6 ex 7 6 t 2 7
0 0 0t 0t 0t 7 6 f6 7 6 4 lc ! 4 ls ! P
t
6 7
6 0t 0 0 0t 0t 0t t 76 ey 7 6 t 2
0 7 6 f6 7 6 4 ls ! C 4 lc ! t
P 7
6 7
6 76 7 6 7:
6 0 0 0 I I 0 0 7 6 f px 7 D 6 J x !2 J y !P 7
6 7 6 dx 7 6 7
6 0 e5 0 Ap Ad 0 0 7 6 f 7 6 0 7
6 7 6 py 7 6 y 2 7
4 0 0 0 0 0 I I 5 4 f 5 4 J ! C mg C J ! x
P5
0 0 e5 0 0 Ap Ad f dy 0
x D At .AAt /1 y DW AC y;
Figure 1.4 shows the resulting stick diagram of a 2D human walking on level
ground, including the motion in single support phase and double support phase.
The algorithm produced 137 time points for the optimised simulation but for clarity,
only 21 were chosen at equal time intervals (t D 0:025 s) to depict the movement for
comparative purposes. A linear interpolation was done on the .x; y/ data for plotting
purposes to evenly distribute the intervals as the single support phase held a larger
percentage of the walk cycle time than the double support phase. Figure 1.5 plots
1.8
1.6
1.4
1.2
1
metres
0.8
0.6
0.4
0.2
0.2
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
meters
1.8
1.6
1.4
1.2
1 0.486
red is swing
metres
0.8
0.6
0.4
0.2
0.2
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
meters
Fig. 1.5 Start and end of walk cycle – experiment main. Dotted figure is the initial configuration
with final swing toe and initial stationary toe at the same position
1 Analysing Human Walking Using Dynamic Optimisation 21
the first and last position of the walk cycle, and showed that periodicity conditions
might have to be worked on for the last position to more accurately mirror the start
position.
As initial joint moment estimates derived from an alternate inverse dynamics
formulation were not very close to the real optimum, a large number of iterations
were involved. The computation was eventually able to come to a satisfactory
solution, satisfying constraint conditions. The final values of the canonical con-
straints are given in Table 1.1. G1 to G4 are defined as terminal time constraints
which were required to converge to 0. Their values were small enough to be
approximated to 0 and considered satisfied by MISER3.3. G5 to G13 are all time
constraints, and had to satisfy the algorithm. Constraint value of 0 occurs when
h .D 106 / ) g .h/ D 0 and are satisfied. G6 constraint had value h as
h , however it was still considered satisfied. The g value of G6 constraint not
being 0 could be explained from when swing heel struck the ground at t D 0:386 s to
t D 0:486 s, the y-position of swing heel had to be 0 or in this case approximately 0.
A closer look at the forces acting on the hip can be seen in Fig. 1.6. It was
observed that forces from the two thighs (Segment 3 and Segment 4) had to equate
to balance out the force from Segment 7. A larger force is required in the horizontal
direction from t D 0:386 s onwards as it prevented the body from continually
moving forward during double support phase.
External forces on swing ankle occurred at the start of double support phase
when swing heel struck the ground. During double support phase, the swing
ankle experienced three forces acting on it, namely distal and proximal force
from Segment 5 and Segment 6 respectively, and external force from the ground.
Figure 1.7 presents the forces acting on the ankle during double support. External
forces acting on the swing ankle were observed to be equal and opposite to the
proximal and distal forces in both x y direction. It was also observed that most of
the forces were contributed from Segment 5 rather than Segment 6 mainly because
Segment 5 was carrying the main weight of the body while Segment 6 only had the
foot.
Figure 1.8 depicts the components of the vertical and horizontal forces on the
stance toe for a one step cycle. As no ground force plate was use in this experiment,
ground reaction force cannot be accurately predicted. However, a vertical reaction
force was observed and a 9th order polynomial fit was plotted against it which
had a familiar double-peak, also known as the “M-shaped” pattern. Distribution
of the model’s weight from the stance toe to the swing ankle could be seen at
22 M. Tan et al.
500
0
500
1000 x3d
1500 x4p
x7p
2000
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time(s)
600
400
Vertical force (N)
200
0
200
400 y3d
600 y4p
y7p
800
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time(s)
Student Version of MATLAB
Fig. 1.6 Forces on hip at distal and proximal end of x and y at segments 3, 4 and 7. (xd3 ; yd3 )
p p
denotes (x; y)-coordinate of distal segment 3, (x4 ; y4 ) denotes (x; y)-coordinate of proximal segment
p p
4, (x7 ; y7 ) denotes (x,y)-coordinate of proximal segment 7
Fig. 1.7 External forces External Force on swing ankle (N) during double support
1000
from angle to ground during fe6x
double support phase. 800
fe6y
(fe6x; fe6y) denotes external
forces acting on 600
(x; y)-coordinate of proximal
External Force (N)
200
200
400
0.38 0.4 0.42 0.44 0.46 0.48 0.5
Time(s)
1 Analysing Human Walking Using Dynamic Optimisation 23
700
600
500
Force (N)
400
300 x1p
y1p
200 9th order polyfit
100
100
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time(s)
200
Force (N)
400
600
800
1000
1200
1400
0.38 0.4 0.42 0.44 0.46 0.48 0.5
Time(s)
p p p p
Fig. 1.8 Forces on stance toe (x1 ; y1 ). (x1 ; y1 ) denotes (x; y)-coordinate of proximal segment 1
(toe)
24 M. Tan et al.
50 800
Torque (Nm)
Torque (Nm)
600
0
400
50
200
100
0
150 200
0 0.1 0.2 0.3 0.4 0.4 0.45 0.5
50 0
0
200
Forces (N)
Forces (N)
50
400
100
600
150
200 800
250 1000
0 0.1 0.2 0.3 0.4 0.4 0.45 0.5
Time(s) Time(s)
1 2 3 4 5 6 7
heel-strike from end of single support phase, through to double support phase. A
high negative horizontal force kept the stance toe in position and prevented it from
moving especially during double support phase.
Figure 1.9 presents the torques and forces acting on segments to produce angular
acceleration. The torques and forces acting on each segment, including external
forces, were computed from the dynamics and plotted. It was observed that the
torques and forces on each segment for each phase were similar in pattern but were
opposite in direction to each other.
The final results of MISER3.3 and the data were plotted out and compared.
Figure 1.10 presents the .x; y/ center of mass (CoM) trajectories and velocities,
while Fig. 1.11 presents the segment angular displacements of MISER3.3 plotted
against the ones from the original data. Slight changes were observed between the
original angular displacement data and the optimised results, however the pattern
remains consistent. This was observed for the CoM trajectories and velocities as
1 Analysing Human Walking Using Dynamic Optimisation 25
CoM X CoM Y
Horizontal displacement (m)
0.4 1.18
0 1.14
0.2 1.12
0.4 1.1
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
1
0
0.5
0.5
0
0.5 1 MISER3
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
data
Time(s) Time(s)
well except for the sudden jump in velocities seen in the MISER3.3 data due to the
jump condition implemented when the heel struck the ground to account for abrupt
velocity change.
Comparisons between the results obtained by MISER3.3 and inverse analysis
were made as well. The optimised set of joint torques obtained from MISER3.3
is illustrated in Figs. 1.12 and 1.13, and plotted against the initial joint torques’
estimates. It was noted that the joint torques in MISER3.3 had to be made piecewise
constant for computation due to the jump condition occurring at t D 0:386 s.
Piecewise linear approximation was not possible as simulation could not converge
to .0:386 / D .0:386C/. Inverse analysis estimated joint torques for single
support and double support phases separately, while MISER3.3 considered the one
step walk cycle when solving for joint torques, thus requiring the jump condition.
As observed, though the joint torques required to reproduce the walking motion
in Fig. 1.4, as estimated by MISER3.3, were different to those obtained by the
conventional method of inverse dynamics, they however still followed a similar
pattern. Figure 1.14 illustrates the vertical reaction force, obtained using the results
of MISER3.3 and inverse analysis, on the stance toe during the single support phase.
A double peak pattern that was noticed in Fig. 1.13 was also observed when a 9th
order polynomial was fitted to data obtained by inverse analysis.
26 M. Tan et al.
3
Stance foot
1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Stance shank
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Stance thigh
2 MISER3
data
1.5
1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
5.5
Swing thigh
4.5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Swing shank
2
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
10
Swing foot
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
2
HAT
1.5
1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (s)
Fig. 1.11 Optimised (MISER3.3) and data segment angular displacement (rad) of a one step cycle
1.5 Conclusions
500
Stance toe
500
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Stance ankle
500
500
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Stance knee
200
MISER3
0 data
200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
50
Swing knee
50
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Swing ankle
10
10
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
200
Hip
200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
100
Hip
100
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time(s)
Fig. 1.12 Joint torque trajectories of optimised (MISER3.3) and inverse dynamics (data) for single
support phase
change in the dynamics from single support phase to double support phase. The
overall research was able to simulate normal walking motion for a full walk cycle,
based on the model developed using MISER3.3.
28 M. Tan et al.
0
Stance toe
2000
4000
0.38 0.4 0.42 0.44 0.46 0.48 0.5
Stance ankle
2000
4000
0.38 0.4 0.42 0.44 0.46 0.48 0.5
Stance knee
2000
4000
0.38 0.4 0.42 0.44 0.46 0.48 0.5
500
Swing knee
MISER3
0 data
500
0.38 0.4 0.42 0.44 0.46 0.48 0.5
Swing ankle
1
0.38 0.4 0.42 0.44 0.46 0.48 0.5
2000
Hip
2000
0.38 0.4 0.42 0.44 0.46 0.48 0.5
1000
Hip
1000
0.38 0.4 0.42 0.44 0.46 0.48 0.5
Time(s)
Fig. 1.13 Joint torque trajectories of optimised (MISER3.3) and inverse dynamics (data) for
double support phase
1 Analysing Human Walking Using Dynamic Optimisation 29
400
MISER3
300 9th order polyfit
200 data
9th order polyfit
100
100
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time(s)
n2 Œ 0 0 0 0 e1 0 0 Ap Ad f dy 0
Note that this is a square matrix of order 5n. There are n C 2 second derivatives
and 4n 2 force components. Unfortunately the order 2n 1 zero-one blocks
on the diagonal are not invertible, so the usual block inversion cannot be fol-
lowed.
30 M. Tan et al.
P D !;
p
xP 1 u
p D 1 D 0;
yP 1 v1
n 1 1 n n2 n n2 2 3 2 3
n Œ J 0 0 Mpx Mdx Mpy Mdy !P T
1 Œ 4 lts 0 0 0t 0t 0t 0t 6 f ex 7 6 l t !2 7
6 6 7 6 4 c 7
1 Œ 4 ltc 0 0 0 t
0t 0t
0t 6 f ey 7 6 l t !2 7
6 6 7 6 4 s 7
6 px 7 D 6 7:
n Œ Jy 0 0 I I 0 0 6 f 7 6 Jx !2 7
6 dx 7 6 7
n2 Œ 0 e5 0 Ap Ad 0 0 6 f 7 6 0 7
6 py 7 6 y 2 7
n Œ J x 0 0 0 0 I I 4 f 5 4 J ! C mg 5
n2 Œ 0 0 e5 0 0 A Ad
p
f dy 0
Note that this is a square matrix of order 5n 2. There are n second derivatives
and 4n 2 force components.
3 Case 2: No Contacts
px py dy
Now the forces at the joints 1 and 7, f1 , f1 , f6dx , f6 , are zero. The dynamic
equations are:
P D !;
p
xP 1 u1
p D ;
yP 1 v1
2 3
!P
n 1 n1 n2 1 n1 n2 6 uP 7 2 3
n Œ J 0 Mpx Mdx 0 Mpy Mdy 6 17
6 f px 7 T
6 7 6 Jx !2 7
n Œ Jy m I I 0 0 0 6 dx 7 6 7:
6f 7 D 6 7
n2 Œ 0 0 Ap Ad 0 0 0 6 7 6 0 7
6 vP1 7 6 y 2 7
n Œ J x 0 0 0 m I I 6 py 7 4 J ! C mg 5
4f 5
n2 Œ 0 0 0 0 0 Ap Ad 0
f dy
in total. If the position, velocity and acceleration of the distal point of Segment 6 are
needed, an extra four differential equations can be added with the four position and
velocity variables, but these can be computed from a knowledge of the variables
already in the system, that is they are dependent on the variables already in the
equations above.
2 3
n 1 1 n1 1 n2 n1 1 n2 !P 2 3
n Œ J 0 0 Mpx 0 Mdx Mpy 0 Mdy 6 uP 1 7 T
6 7
1 Œ 5 lts 1 0t 0t 0 0 0t 0 0t 6 vP 7 6 l t !2 7
6 17 6 5 c 7
1 Œ 5 ltc 0 1 0t 0 0 t
0t 0 0t 6 f px 7 6 l t !2 7
6 7 6 5 s 7
6 ex 7 D 6 7:
n Œ Jy m 0 I 0 I 0 0 0 6 f2 7 6 Jx !2 7
6 dx 7 6 7
n2 Œ 0 0 0 Ap e1 Ad 0 0 0 6f 7 6 0 7
6 py 7 6 y 2 7
n Œ Jx 0 m 0 0 0 I 0 I 6f 7 4 J ! C mg 5
6 ey 7
n2 Œ 0 0 0 0 0 0 Ap e1 Ad 4f 5 0
2
f dy
Note that this is a square matrix of order 5n 2. There are n C 2 second order
derivatives and 4n 4 force components.
The case presented here is similar to the one before except here the swing heel at
proximal Segment 6 is considered. The dynamic equations are:
2 3 2 3
P !
4 xP p 5 D 4 u1 5 ;
1
p
yP 1 v1
32 M. Tan et al.
2 3
n 1 1 n1 1 n2 n1 1 n2 !P 2 3
n Œ J 0 0 Mpx 0 Mdx Mpy 0 Mdy 6 uP 1 7 T
6 7
1 Œ 4 lts 1 0t 0t 0 0 0t 0 0t 6 vP 7 6 l t !2 7
6 17 6 4 c 7
6 f px 7 6 7
1 Œ 4 ltc 0 1 0t 0 0 0t 0 0t 6 7 D 6 4 ls !2 7 :
t t
6 ex 7 6 7
n Œ Jy m 0 I 0 I 0 0 0 6 f6 7 6 Jx !2 7
6 dx 7 6 7
n2 Œ 0 0 0 Ap e1 Ad 0 0 0 6f 7 6 0 7
6 py 7 6 y 2 7
n Œ Jx 0 m 0 0 0 I 0 I 6f 7 4 J ! C mg 5
6 ey 7
n2 Œ 0 0 0 0 0 0 Ap e1 Ad 4f 5 0
6
f dy
Note that this case and the case before are similar, the only difference falls in the
numbering of segments. This system, as before, is a square matrix of order 5n 2.
There are n C 2 second derivatives and 4n 4 force components.
P D !;
n n n2 n n2 2 3 2 3
n Œ J Mpx M dx Mpy Mdy ! P T
n Œ Jy I I 0 0 6 px 7 6
6 f 7 D 6 J !2 7 :
x 7
6 dx 7 6 7
n2 Œ 0 Ap Ad 0 0 6f 7 6 0 7
6 py 7 6 y 2 7
n Œ Jx 0 0 I I 4f 5 4 J ! C mg 5
n2 Œ 0 0 0 A Adp
f dy 0
References
Onyshko S, Winter DA (1980) A mathematical model for the dynamics of human locomotion. J
Biomech 13(4):361–368
Pandy MG (2001) Computer modeling and simulation of human movement. Annu Rev Biomed
Eng 3:245–273
Rehbock V, Teo KL, Jennings LS (1996) Optimal and suboptimal feedback controls for a class of
nonlinear systems. Comput Math Appl 31(6):71–86
Ren L, Jones RK, Howard D (2007) Predictive modelling of human walking over a complete gait
cycle. J Biomech 40(7):1567–1574
Said M, Jennings LS, Koh MT (2006) Computational models satisfying relative angle constraints
for 2-dimensional segmented bodies. Anziam J 47:541–554
Selles RW, Bussmann JBJ, Wagenaar RC, Stam HJ (2001) Comparing predictive validity of four
ballistic swing phase models of human walking. J Biomech 34(9):1171–1177
Ünver NF, Tümer ST, Özgören MK (2000) Simulation of human gait using computed torque
control. Technol Health Care 8(1):53–66
Winter DA (1990) Biomechanics and motor control of human movement, 2nd edn. Wiley, Hoboken
Winter DA (2009) Biomechanics and motor control of human movement, 4th edn. Wiley, Hoboken
Xiang Y, Arora JS, Rahmatalla S, Abdel-Malek K (2009) Optimization-based dynamic human
walking prediction: one step formulation. Int J Numer Methods Eng 79(6):667–695
Xiang Y, Arora JS, Abdel-Malek K (2010) Physics-based modeling and simulation of human
walking: a review of optimization-based and other approaches. Struct Multidiscip Optim
42(1):1–23
Yamaguchi GT, Zajac FE (1990) Restoring unassisted natural gait to paraplegics via functional
neuromuscular stimulation: a computer simulation study. IEEE Trans Biomed Eng 37(9):886–
902
Chapter 2
Rearrangement Optimization Problems Related
to a Class of Elliptic Boundary Value Problems
2.1 Introduction
After the Burton fundamental work (Burton 1987, 1989) on theory of rearrange-
ments, the rearrangement optimization problems in addressing questions such
as existence, uniqueness, symmetry and some qualitative properties of optimal
solutions have been investigated by a number of authors, see for example (Burton
1989; Kurata et al. 2004; Del Pezzo and Bonder 2009; Zivari-Rezapour 2013; Cuccu
et al. 2006a,b, 2009; Marras 2010; Marras et al. 2013; Emamizadeh and Zivari-
Rezapour 2007; Emamizadeh and Fernandes 2008; Emamizadeh and Prajapat 2009;
Chanillo et al. 2000; Chanillo and Kenig 2008; Nycander and Emamizadeh 2003;
Anedda 2011; Qiu et al. 2015) and the references therein.
This work was supported by Natural Science Foundation of China (11471235, 11171247,
11371273) and GIP of Jiangsu Province (CXZZ13_0792).
C. Qiu • Y. Huang • Y. Zhou ()
Department of Mathematics, Soochow University, Suzhou 215006, People’s Republic of China
e-mail: yuyingz@suda.edu.cn
including their similar problems involving p-Laplacian, have been studied by some
authors, see for example (Burton 1987, 1989; Cuccu et al. 2006a, 2009; Marras
2010; Marras et al. 2013; Emamizadeh and Zivari-Rezapour 2007; Emamizadeh
and Fernandes 2008), where 0 < h 2 L1 .˝/, f 2 Lq .˝/ with q > 2N=.N C
2/.
Recently, a rearrangement optimization problem related to the following quasi-
linear elliptic boundary value problem has been considered in Qiu et al. (2015):
(
p u C h.x; u/ D f .x/ in ˝;
.P/
uD0 on @˝
.Ph;f / is actually a model of the deformation problem for an elastic membrane made
out of some materials with prescribed quantities h, subject to a fixed vertical force
f . The usual goal is to identify a force function selected from R.f /, in such a way
2 Rearrangement Optimization Problems Related to a Class of Elliptic. . . 37
that the total displacement of the membrane is as small as possible. More precisely,
let I W H01 .˝/ ! R be the energy functional corresponding to the problem .Ph;f /,
which is given by
Z Z Z
1 2 2
I.u/ D jruj dx hu dx fudx; (2.1)
2 ˝ 2 ˝ ˝
and let 0 < h0 2 L1 .˝/ and f0 2 Lq .˝/ with q > 2N=.N C 2/ be two given
functions, then we will study the following minimum and maximum optimization
problems:
.Optm / Find hO 2 R.h0 /; fO 2 R.f0 / such that I.uhO ;fO / D infh2R.h0 /;f 2R.f0 / I.uh;f /
and
.OptM / Find fN 2 R.f0 / such that I.uh0 ;fN / D supf 2R.f0 / I.uh0 ;f /,
where R.h0 / and R.f0 / respectively denote the sets of all rearrangement of h0 and
f0 , uh;f (uh0 ;f ) is the unique solution of the problem .Ph;f / (.Ph0 ;f /) (the existence
and uniqueness of uh;f (uh0 ;f ) will be obtained in Propositions 2.1 of Sect. 2.3). We
will show that there exists > 0 such that for all 2 .0; /, both problems
.Optm / and .OptM / are solvable.
We note that the optimization problem considered in all the papers mentioned
above is constrained by a rearrangement set which is generated by just one fixed
function. The minimum optimization problem considered here is however con-
strained by two rearrangement sets generated by two fixed independent functions.
Moreover, Problem .Ph;f / contains .Lh / and .Pf / as special cases, the cost
functional used in our problem is more complicated than that used in the above two
problems, therefore our case needs special handling. We point out that an essential
assumption in Qiu et al. (2015) is that h.x; u/ being non-decreasing with respect to
the second variable u for almost all x 2 ˝. But in the present paper, since we assume
that 0 < and 0 < h.x/; a:e: x 2 ˝, the term h.x/u in the problem .Ph;f /
would be decreasing with respect to u for almost all x 2 ˝ and then it violates
the essential assumption given in paper Qiu et al. (2015). So the conditions and
results for the maximization problem (OptM ) here are different from those obtained
in Qiu et al. (2015). To the best of our knowledge, the results obtained in this paper
are new.
This paper is organized as follows. In Sect. 2.2, we give some preliminaries. In
Sect. 2.3, we show that the problem .Ph;f / has a unique solution. Section 2.4 is
devoted to discuss the minimization problem (Optm ) in detail. Firstly, we prove that
the minimization problem (Optm ) is solvable in the case of 0 < < , then we
obtain a representation result of the optimal solution for the minimization problem
and show that the problem (Optm ) has unique solution with some symmetric
properties if ˝ is a ball centered at the origin. In Sect. 2.5, we show that the
maximization problem (OptM ) is solvable.
38 C. Qiu et al.
2.2 Preliminaries
We denote by Lr .˝/ .1 r 1/ and H01 .˝/ the usual Sobolev spaces endowed
R 1=r
with the norms kukLr D ˝ jujr dx if 1 r < 1, kuk1 D ess supx2˝ ju.x/j and
R 2 1=2
kuk D ˝ jruj dx , respectively. Throughout the paper C will denote a positive
(possibly different) constant.
Definition 2.1. By a solution u of the problem .Ph;f / we mean that u 2 H01 .˝/
satisfying
Z
.rurv huv f v/ dx D 0; 8v 2 H01 .˝/:
˝
In this case, u 2 H01 .˝/ is a weak solution if and only if I 0 .u/v D 0, 8v 2 H01 .˝/.
The following lemmas will be used through the proofs of our main results.
Lemma 2.1 (Burton 1989, Lemma 2.1). Assume that 1 r < 1 and given f 2
Lr .˝/, then for any g 2 R.f / we have g 2 Lr .˝/ and kgkLr D kf kLr .
Lemma 2.2 (Burton 1989, Lemma 2.2). Assume that 1 r < 1 and given f 2
Lr .˝/, denote by R.f /w the weak closure of R.f / in Lr .˝/, then R.f /w is convex
and weakly compact in Lr .˝/.
Lemma 2.3 (Burton 1989, Lemma 2.9 or Cuccu et al. 2009, Lemma 2.1). Let
f ; g W ˝ 7! R be measurable functions and suppose that for each t 2 R, the level
set of g at t, i.e., fx 2 ˝ W g.x/ D tg, has zero measure. Then there exists an
increasing (decreasing) function ' such that ' ı g is a rearrangement of f where
' ı g denotes a composite function defined by
Lemma 2.4 (Burton 1989, Lemma 2.4 or Cuccu et al. 2009, Lemma 2.2). For
any 1 r < 1 define r0 D r1 r
if r > 1 and r0 D 1 if r D 1. Let f 2 Lr .˝/ and
0
g 2 Lr .˝/. Suppose that there exists an increasing (decreasing) function ' W R 7!
R such that ' ı gR 2 R.f /. Then ' ı g is the unique maximizer (minimizer) of the
linear functional ˝ hgdx, relative to h 2 R.f /w .
Lemma 2.5 (Burton 1987, Theorem 5). For any 1 r < 1 define r0 D r1 r
if
0 0
r > 1 and r D 1 Rif r D 1. Let f 2 L .˝/ and g 2 L .˝/. Suppose that the linear
r r
Lemma 2.6 (Emamizadeh and Prajapat 2009, Lemma 2.3). Suppose that f 2
Lr .˝/ and g 2 LrR.˝/, then there exists fO 2 R.f / which maximizes (minimizes) the
0
for any non-negative measurable functions f and g, where f and g are respectively
the Schwarz symmetric decreasing rearrangements of f and g, defined in the
following.
Definition 2.2 (Leoni 2009, Definition 16.5). Let f W ˝ 7! Œ0; 1/ be a measurable
function. The Schwarz symmetric decreasing rearrangement of f is the function
f W B.0; r/ 7! Œ0; 1/, defined by
˚
f .x/ D inf t 2 Œ0; 1/ W f .t/ !N jxjN ; 8x 2 B.0; r/
The following result can be deduced from Lemmas 2.3 and 3.2 and Theorem 1.1 of
Brothers and Ziemer (1988).
1;p
Lemma 2.9. Suppose that B is a ball centered at the origin. If u 2 W0 .B/ with
1 < p < 1 and u 0 then u1 .˛; 1/ is a translation of u1 .˛; 1/ for every
˛ 2 Œ0; ess supx2B u.x// and
Z Z
jrujp dx jru jp dx: (2.2)
B B
It is well known that the first eigenvalue 1 .h/ of the problem .Lh / can be
characterized by
R
jrvj2 dx
1 .h/ D inf R˝ : (2.3)
2
v2H01 .˝/;v6D0 ˝ hv dx
By Cuccu et al. (2009, Theorem 3.1), if 0 < h0 .x/ 2 L1 .˝/, then there exists
Nh 2 R.h0 / (the set of all rearrangements of h0 ) such that
R
jrvj2 dx
0< N D
WD 1 .h/ inf 1 .h/ D inf inf R˝ : (2.4)
h2R.h0 / h2R.h0 / v2H 1 .˝/;v6D0 2
0 ˝ hv dx
In this section, we will obtain the existence and uniqueness for the solution of the
problem .Ph;f /.
Proposition 2.1. Fix 0 < h.x/ 2 L1 .˝/, and f 2 Lq .˝/ with q > 2N=.N C2/ and
0 < < 1 .h/, where 1 .h/ is the first eigenvalue of the problem .Lh /. Then the
problem .Ph;f / has a unique solution uh;f 2 H01 .˝/ and I.uh;f / D infv2H 1 .˝/ I.v/.
0
Moreover, if in addition f .x/ > 0 a.e. x 2 ˝, then uh;f > 0.
Proof. First, we show that the problem .Ph;f / has a solution.
By the Hölder inequality and the Sobolev embedding inequality, we have
ˇZ ˇ
ˇ ˇ
ˇ fudxˇ kf kL kukL 0 Ckuk (2.5)
ˇ ˇ q q
˝
for all u 2 H01 .˝/ since now 1 < q0 WD q=.q 1/ < 2 where 2 WD 2N=.N 2/.
Hence we deduce from (2.1), (2.3) and (2.5) that
1
I.u/ .1 /kuk2 Ckuk ! 1
2 1 .h/
Z
I 0 .uh;f /v D ruh;f rv huh;f v f v dx D 0; 8v 2 H01 .˝/: (2.6)
˝
Next, we show that uh;f is the unique solution of the problem .Ph;f /.
Assume that wh;f 2 H01 .˝/ is another solution of the problem .Ph;f / and uh;f 6D
wh;f , then
From (2.6) and Definition 2.1 we get that for every v 2 H01 .˝/,
Z Z
ruh;f rv huh;f v dx D f vdx;
˝ ˝
Z Z
rwh;f rv hwh;f v dx D f vdx:
˝ ˝
Therefore,
Z Z
. hwh;f huh;f /vdx D .rwh;f ruh;f /rvdx; 8v 2 H01 .˝/: (2.8)
˝ ˝
Let v D wh;f uh;f . Note that 0 < < 1 .h/, then from (2.3), (2.7) and (2.8) we
obtain
Z
.jrwh;f ruh;f j2 /dx
˝
Z
D h.wh;f uh;f /2 dx
˝
Z
< 1 .h/h.wh;f uh;f /2 dx
˝
Z
.jrwh;f ruh;f j2 /dx;
˝
a contradiction. Therefore we have proved that uh;f is the unique solution of the
problem .Ph;f /.
42 C. Qiu et al.
Finally, if f .x/ > 0 then we can easily check that I.juh;f j/ I.uh;f /, which shows
that juh;f j is also a minimizer of I and thus a solution of the problem .Ph;f /. Then
uh;f D juh;f j 0 by the uniqueness of the solution. Since
we have uh;f .x/ > 0; a:e: x 2 ˝ (cf. Vázquez 1984, Theorem 5). t
u
Remark 2.1. In the case of D 1 .h/, if uh;f 2 H01 .˝/ is a solution of the problem
.Ph;f /, then for any t 2 R and v 2 H01 .˝/,
Z
r.uh;f C t'/rv h.uh;f C t'/v f v dx
˝
Z
D ruh;f rv huh;f v f v dx D 0;
˝
where ' is the eigenfunction of .Lh /. That is, uh;f Ct' is the solution of the problem
.Ph;f /. Therefore, in order to obtain the unique solution of the problem .Ph;f /, we
only consider the case of 0 < < 1 .h/ in the following.
Theorem 2.1. Suppose that 0 < h0 .x/ 2 L1 .˝/, f0 2 Lq .˝/ with q > 2N=.NC2/,
and 0 < < , where is given by (2.4). Then there exists hO 2 R.h0 /; fO 2 R.f0 /
which solves the problem .Optm /, i.e.,
AD inf I.uh;f /
h2R.h0 /;f 2R.f0 /
and then
Z Z Z
1 2
I.uh;f / D jruh;f j dx hu2h;f dx fuh;f dx
2 ˝ 2 ˝ ˝
(2.9)
1 2
.1 /kuh;f k Ckf kLq kuh;f k:
2
and
A D lim I.ui /
i!1
where ui D uhi ;fi . It follows from (2.9) that fui g is bounded in H01 .˝/, then it
has a subsequence (still denoted fui g) which weakly converges to u 2 H01 .˝/
0
and strongly converges to u in Lq .˝/ with 1 < q0 D q=.q 1/ <
2 . Since kfi kLq kf0 kLq , ffi g contain a subsequence (still denoted ffi g)
converging weakly to some fN 2 R.f0 /w , the weak closure of R.f0 / in Lq .˝/.
Then
ˇZ ˇ
ˇ ˇ
ˇ .fi fN /udxˇ ! 0 as i ! 1
ˇ ˇ
˝
0
since u 2 Lq .˝/. It follows from the Hölder inequality that
ˇZ ˇ ˇZ ˇ ˇZ ˇ
ˇ ˇ ˇ ˇ ˇ ˇ
ˇ .fi ui fN u/dxˇ ˇ fi .ui u/dxˇ C ˇ .fi fN /udxˇ
ˇ ˇ ˇ ˇ ˇ ˇ
˝ ˝ ˝
ˇZ ˇ (2.10)
ˇ ˇ
kfi kLq kui ukLq0 C ˇˇ .fi fN /udxˇˇ ! 0
˝
By (2.10) and (2.11) and the weak lower semi-continuity of the norm in the H01 .˝/,
we obtain that
Z Z Z
1 N 2 dx
A D lim I.ui / jruj2 dx hu fN udx: (2.12)
i!1 2 ˝ 2 ˝ ˝
44 C. Qiu et al.
RSimilarly we have there exists hO 2 R.h0 / which maximizes the linear functional
2
˝ lu dx, relative to l 2 R.h0 / . So that
w
Z Z
N 2 dx
hu O 2 dx:
hu
˝ ˝
By Proposition 2.1,
Z
1 O 2 dx fO v dx
I.Ou/ D inf jrvj2 hv
v2H01 .˝/ ˝ 2 2
Z Z Z (2.14)
1 O 2 dx
jruj2 dx hu fO udx:
2 ˝ 2 ˝ ˝
hO D .Ou2 / a:e: in ˝;
(2.15)
fO D '.Ou/ a:e: in ˝;
i.e.,
Z Z
2
hOu dx O u2 dx;
hO 8h 2 R.h0 /:
˝ ˝
R
So that hO is a maximizer of the linear functional L.h/ WD ˝ hOu2 dx, relative to
h 2 R.h0 /.
We claim that hO is the unique maximizer of L.h/. If not, suppose that hN is another
maximizer of L.h/. Then
Z Z
O u2 dx D
hO N u2 dx:
hO
˝ ˝
Thus
Z Z Z
1 O u2 dx
I.Ou/ D 2
jr uO j dx hO fO uO dx
2 ˝ 2 ˝ ˝
Z Z Z
1 N u2 dx
D jr uO j2 dx hO fO uO dx
2 ˝ 2 ˝ ˝
I.uhN ;fO /
I.Ou/:
So that
Z Z Z
1 N u2 dx
2
jr uO j dx hO fO uO dx D I.uhN ;fO /:
2 ˝ 2 ˝ ˝
By the uniqueness of the minimizer of the functional I, we obtain uO D uhN ;fO . Then
Z Z Z
ruhN ;fO rvdx N N O vdx D
hu fO vdx;
h;f
˝ ˝ ˝
Z Z Z
r uO rvdx O uvdx D
hO fO vdx; 8v 2 H01 .˝/:
˝ ˝ ˝
So that
Z
.hN h/O
O uvdx D 0; 8v 2 H01 .˝/;
˝
N h.x//O
.h.x/ O u.x/ D 0; a:e: x 2 ˝: (2.16)
46 C. Qiu et al.
'.Ou/ D fO ; a:e: in ˝:
and
( )!
meas x 2 ˝ W r uO D 0; 0 < uO .x/ < ess sup uO .y/ D 0: (2.18)
y2˝
Now, by using Lemma 2.9, and noting (2.17) and (2.18), we see that uO D uO .
By (2.15) in Theorem 2.2, hO D ı .Ou /2 and fO D ' ı uO are spherically symmetric
decreasing functions. It follows that hO coincides its Schwarz rearrangement, i.e.,
hO D hO D h0 , so is fO . t
u
We now consider the problem .OptM /. Our results for the problem .OptM / are the
following.
Theorem 2.4. Let 0 < h0 .x/ 2 L1 .˝/, and let f0 2 Lq .˝/ with q > 2N=.N C 2/.
Suppose that 0 < < 1 .h0 /, where 1 .h0 / is given by (2.3) and f0 .x/ 0, then
there exists a unique fN 2 R.f0 / which solves the problem .OptM /, i.e.,
2 Rearrangement Optimization Problems Related to a Class of Elliptic. . . 47
Z Z
1 .h0 / hu2h0 ;f dx jruh0 ;f j2 dx:
˝ ˝
We get
1
I.uh0 ;f / .1 /kuh0 ;f k2 Ckf kLq kuh0 ;f k:
2 1 .h0 /
The rest proof of (I), (II) and (III) is similar to the proof of Lemma 4.1 in Qiu
et al. (2015), we omit it. t
u
Similar to Lemma 4.2 in Qiu et al. (2015), we obtain
Lemma 2.11. Under the assumptions of Theorem 2.4, there exists a unique fQ 2
R.f0 /w which maximizes ˚jR.f0 /w . Moreover,
Z Z
uQ fQ dx uQ gdx; 8g 2 R.f0 /w ; (2.20)
˝ ˝
where uQ D uQf .
Lemma 2.12. Let fQ and uQ be as in Lemma 2.11, and let S.fQ/ D fx 2 ˝ W fQ .x/ > 0g:
Set
Then ı.
48 C. Qiu et al.
Proof. If not, we assume that > ı. Then we can choose > 1 > 2 > ı. Since
> 1 , there exists a set A S.fQ /, with positive measure, such that uQ 1 in A.
Similarly, there exists a set B ˝ n S.fQ/, with positive measure, such that uQ 2
in B. Without lose of generality we may assume that meas.A/ D meas.B/. Then
there exists a measure preserving map T W A ! B: So that we can define a particular
rearrangement of fQ as following:
8
ˆ
ˆ fQ .Tx/; x2A
<
fN .x/ D fQ .T 1 x/; x 2 B
ˆ
:̂
fQ .x/; x 2 ˝ n .A [ B/:
Thus
Z Z Z Z
uQ fQ dx uQ fN dx D uQ fQ dx uQ fN dx
˝ ˝ A[B A[B
Z Z
D uQ fQ dx uQ fN dx
A B
Z Z
1 fQ dx 2 fN dx
A B
Z
D .1 2 / fQ dx > 0:
A
R R
Therefore, ˝ uQ fQ dx > ˝ uQ fN dx, a contradiction. t
u
Proof of Theorem 2.4. Let fQ and uQ be as in Lemma 2.11. It is clear that the level sets
of uQ , restricted to S.fQ/, have measure zero. Therefore applying Lemma 2.3, there
exists a decreasing function Q such that Q ı uQ is a rearrangement of fQ relative to the
set S.fQ/. Now, define
(
;
Q t ;
.t/ D
0; t> ;
By the definition of ,
Q and , we get
By (2.22) and (2.23), in order to prove (2.21) we only need to show that
It follows from (2.25), (2.26) and (2.27) that (2.24) holds, and then (2.21) holds.
Therefore, ı uQ is a rearrangement of fQ .
Hence, applying Lemma R 2.4, we can deduce that ı uQ is the unique minimizer
of the linear functional ˝ gQudx, relative to g 2 R.f0 /w . This and (2.20) obviously
imply fQ D ı uQ 2 R.f0 /. We complete the proof by choosing fN D fQ . t
u
Acknowledgements The authors would like to thank the referees for the valuable suggestions
which have improved the early version of the manuscript.
References
Anedda C (2011) Maximization and minimization in problems involving the bi-Laplacian. Annali
di Matematica 190:145–156
Brothers JE, Ziemer WP (1988) Minimal rearrangements of Sobolev functions. J Reine Angew
Math 384:153–179
50 C. Qiu et al.
Yoshiaki Shimizu
3.1 Introduction
number of members of this class. The set of optimal solutions is known as the Pareto
front. Generally speaking, however, decision making as an engineering task aims at
obtaining a limited number of candidates for the final decision.
From this viewpoint, this study proposes a general idea for solving the many-
objective optimization (MAOP) problem in which more than several objective
functions are considered simultaneously. Effort is devoted to obtain a unique
solution known as the preferentially optimal solution or the best-compromise
solution. This approach is notably different from that of multi-objective evolutionary
algorithms (MOEA), which attempt to derive only the Pareto front (Coello 2001;
Czyzak and Jaszkiewicz 1998; Deb et al. 2000; Jaeggi et al. 2005; Robic and Filipic
2005). However, recent studies have revealed that even in MOEA, conventional
methods are not necessarily effective for dealing with MAOP problems (Hughes
2005; Sato et al. 2010).
In this context, we extend our previously proposed methods, named MOON2 and
MOON2R (Shimizu and Kawada 2002; Shimizu et al. 2004), to be able to handle
MAOP problems. Although MOON2 and MOON2R require only simple and relative
responses, handling the decision makers’ (DMs’) responses in trade-off analysis
becomes rather difficult in MAOP. To overcome this difficulty, this study proposes
an approach that is easily applicable to MAOP. Consequently, the proposed idea can
extend the applicability and practicality of existing methods to the complex decision
making environments mentioned above.
The rest of this chapter is organized as follows. In Sect. 3.2, the general
procedures of MOON2 and MOON2R are explained. Section 3.3 extends this
procedure to MAOP. In Sect. 3.4, the validity and effectiveness of the proposed
method is verified by applying it to an actual problem. A general discussion is also
presented in that section to give a definite and comprehensive outline of the direction
of future work in this area. A conclusion is given in Sect. 3.5.
each of these conventional methods has both advantages and disadvantages. For
example, since in the former method a value function is derived separately from
the search process, the DM does not need to perform repeated interactions during
the search process, whereas such interactions are required in the latter method.
On the other hand, although the latter method allows for elaborate articulation of
attainability among the conflicting objectives, such articulation is difficult to obtain
with the former method. Consequently, the derived solution may sometimes differ
substantially from the best-compromise solution provided by the DM.
MOEA methods, which differ substantially from the two methods mentioned
above, have been developed recently. However, these methods require further steps
before attaining the final solution because the DM has to find the best solution
among a potentially large number of candidates scattered along the Pareto front. In
contrast, MOON2 and MOON2R can readily derive the best-compromise solution
while being free from the requirement of repeated responses during the search,
without giving up elaborate trade-off analysis. Therefore, MOON2 and MOON2R
are expected to serve as powerful tools for enabling flexible decision making in
agile engineering under diverse customer requirements.
Because MOON2 and MOON2R belong to the prior articulation methods in MOP,
they have to identify the value function of the DM in advance. Such modeling can
be performed with a suitable artificial neural network to deal with the non-linearity
commonly seen in the value function. A back-propagation network (BPN) is used
in MOON2 , while MOON2R employs a radial-basis function network (RBFN).
To train the neural network, training data representing the preferences of the
DM should be gathered by an appropriate means. These methods use pairwise
comparison among the appropriate trial solutions, which are spread over the search
area in the objective-function space. It is natural to constrain this modeling space
to within the convex hull enclosed by the utopia and nadir solutions, which are
defined as f * D (f1 (x utop ), f2 (x utop ), : : : , fN (x utop ))T and f* D (f1 (x nad ), f2 (x nad ),
: : : , fN (x nad )) T , respectively, where x utop and x nad are the respective utopia and
nadir solutions in the decision variable space.
Then, the DM is asked to indicate the preferred solution and the spacing
between each pair of trial solutions, for example, f i D f(xi ) and f j D f(x j ), xi , x j 2X.
These responses are provided in the form of linguistic statements, which are later
transformed into scores denoted as aij (Table 3.1), similarly to the analytic hierarchy
process (AHP) (Saaty 1980). For example, when the answer is such that f i is
strongly preferable to f j , aij takes a value of 5 (Table 3.1).
aij = 1
a ij
fk 1
as utopia, nadir, a center of gravity between them, or the point where the total sum
of distances from all trial points is a minimum.
VNN f .x/ ; f R D axR > VNN f .y/ ; f R D ayR () f .x/ f .y/; 8x; y 2 X
(3.2)
Once the value function is identified, the original MOP problem is transformed
into an ordinal single-objective problem.
Because the value function is built separately from the search process, a DM
can carry out trade-off analyses whatever pace is desired without having to provide
immediate responses or wait for queries, as is often required in interactive methods.
In addition, because the required responses are simple and relative, the load on the
DM in such interaction is rather small. These are some of the notable advantages
of this approach. Moreover, the following proposition supports the validity of the
above formulation.
[Proposition] The optimal solution of Problem (p. 2) is a Pareto optimal solution
of Problem (p. 1) if the value function is chosen so as to satisfy the relation given
by Eq. (3.1).
(Proof) Let b fi , (i D 1, : : : , N) be the values of the objective functions for the
optimal solution b x of Problem (p. 2), so that b fi D fi bx . Here, let us assume for
contradiction that b f is not a Pareto optimal solution. Then there exists a certain
f 0 such that for 9j, fj 0 < b fj fj ; fj > 0 ) and fj 0 b fi , (i D 1, : : : , N,
i ¤ j).
Because the DM apparently prefers f 0 to b
f , it holds that VNN f 0 ; f R >
VNN b f ; f R . This contradicts that b f is the optimal solution of Problem (p. 2).
Hence, bf must be a Pareto optimal solution.
Once x is given, we can readily evaluate any candidate solution through VNN .
Hence, it is possible to choose the most appropriate method from among a variety of
conventional single-objective optimization methods. In addition to direct methods,
meta-heuristic methods such as genetic algorithms, simulated annealing and tabu
search are also applicable. At the same time, it is almost impossible to apply any of
the interactive methods of MOP due to the large number of interactions during the
search, which are likely to make the DM rather careless in providing responses.
When this approach is applied with an algorithm that requires the gradients of
the objective function, such as nonlinear programming, we need to obtain these
gradients by numeric differentiation. The derivative of the value function with
respect to a decision variable is calculated by using the following chain rule.
56 Y. Shimizu
Set utopia/nadir
& Searching space
No
Consistent
?
Yes
No
Need gradients
?
Yes
Incorporate Numerical differentiation
No Satisfactory
?
Yes
END
!
@VNN f .x/ ; f R @VNN f .x/ ; f R @f .x/
D (3.3)
@x @f .x/ @x
The derivative can be calculated from the analytic form of the second part in the
right-hand side of Eq. (3.3) and the following numeric differentiation. Since most
nonlinear programming software supports numeric differentiation, the algorithm can
be realized without any special concerns.
@VNN VNN . ;fi .x/C fi ; I f R /VNN . ;fi .x/; I f R /
@fi
Š fi
(3.4)
Because the aforementioned methods are natural and easy to work with for value
assessments by humans, we have applied them to various problems and confirmed
their effectiveness (Shimizu et al. 2005, 2006, 2010, 2012a; Shimizu and Tanaka
2003; Shimizu and Nomachi 2008). However, the case of MAOP is different if we
consider the limit to the abilities of humans to perform assessment. As the number
of objective functions increases, the difficulty of such value assessment through
pairwise comparison increases rapidly. For example, suppose that a customer
intends to buy a ticket for transportation in a certain situation. It seems rather easy
to choose between a pair of candidates if they are evaluated on only two objectives,
such as travel time and expense. According to the procedure outlined above, in this
case, the customer has to make a pairwise comparison between the pair of solutions
(i, j) in terms of the objectives (time i, cost i) and (time j, cost j), respectively.
However, what will happen if there are more objectives to be compared? Suppose
that the customer has to compare a pair of candidates in terms of four objectives:
time, cost, service and comfort. Undoubtedly, the difficulty of assessment will grow
substantially, and the customer may often give up on the comparison altogether,
except in special cases.
For MAOP, therefore, it is impractical to deploy the proposed idea while
maintaining the portability of the previous method. The basic idea of the proposed
procedure involves replacing the pairwise comparison on many objectives with a
comparison on a scalar objective. Assuming independence of the objective functions
of (p.1), this procedure can be realized by the following steps.
Step 1: Determine the relative importance
X among the objective functions as weights
wk ; .k D 1; : : : ; N/ ; such that wk D 1, through pairwise comparison and
k
eigenvalue calculation, as in AHP. Repeat if the pairwise comparison fails the
consistency test.
Step 2: Narrowing the focus to the kth objective function only, ask the DM to give
a preference for every pair of trial solutions f i D f xi and f j D f x j , ffk (xi ),
58 Y. Shimizu
fk (x j )g .8i; j; i > j/, and obtain the preference intensity sik ; 8i by calculating
the eigenvalues of this PCM. Repeat this process for every objective function.
X
N
Step 3: Calculate the total preference of the ith trial as Si D wk sik ; 8i.
kD1
Step 4: Finally, calculate aij , which is the PCM element corresponding to the
preference between f i and f j , as aij D Si =Sj .
Step 5: Similarly to the previous step, identify the value function of the DM from fi
and f j as the inputs and aij as the output of the neural network.
The above procedure can be easily implemented by a DM who is familiar with
AHP, and does not introduce additional complexity to the original procedures of
MOON2 and MOON2R .
( 1.
X
N t ) t
fk .x/ fk
U .f .x// D wk ; (3.5)
k
fk fk
where fk and fk denote the utopia and nadir values of the kth objective function,
respectively. Moreover, wk and t are a weight representing relative importance and
a norm parameter, respectively. Hence, U( f(x)) represents the attainability ratio for
utopia and takes a value of 1.0 for the utopia and 0.0 for the nadir.
We carried out the experiment along with the following procedures that corre-
spond to those in Sect. 3.3.
P
Step 1: Determine a set of weights wk . wk D 1/, each of which stands for
the relative importance of the corresponding objective function regarding the
preference.
Step 2: Instead of interactive pairwise comparison, for the virtual DM, obtain the
i t
f f
preference index sik of the ith trial for the kth objective as sik D f k fk ,
k k
(8k; 8i).
.
!1
XN t
Step 3: Calculate the total preference score Si as Si D wk sik , .8i/.
k
3 An Extension of the MOON2 /MOON2R Approach to Many-Objective. . . 59
l L
b
4PL3
f2 WD ı D ! min .Deflection/ (3.7)
Et3 b
r
l
f3 WD D 02 C 0 00 C 00 2 ! min .Shear stress/ (3.8)
R
6PL
f4 WD D ! min .Bending stress/ (3.9)
t2 b
60 Y. Shimizu
3.4.2.2 Constraints
hb (3.10)
Pc P (3.11)
P
0 D p (3.13)
2hl
h [m]: welding thickness; l [m]: welding length; t [m]: beam width; b [m]: beam
thickness
3.4.2.4 Parameters
First, we described the objective tree as shown in Fig. 3.4. Then, we generated six
random trials within the hyper-rectangular space enclosed by the utopia and nadir,
which are shown in Table 3.3 together with the test trials. Next, we set the weights
3 An Extension of the MOON2 /MOON2R Approach to Many-Objective. . . 61
Beam design
representing the relative importance as w D (0.4, 0.3, 0.2, 0.1), which are the same
as those given for the reference value function in Eq. (3.5). Then, the preference
intensity of every trial with respect to each objective function was derived from the
formula given in Step 2. Finally, the total preference was calculated as S D (0.293,
0.570, 0.350, 0.484, 0.439, 0.517) for t D 1. In Step 4, Si /Sj was calculated to derive
the elements of the PCM shown in Table 3.4. Based on that procedure, we built the
value function VNN (f(x); f R ) of the neural network.
Letting f R D f , we solved (p. 3) under this value function by the modified
nonlinear simplex method (Nelder and Mead 1965) so that it can accommodate the
constraints. In Table 3.5, the result is compared with that obtained by optimizing
Problem (p. 3) under the objective function in Eq. (3.5). This problem is solved by
using the commercial software package LINGO (Ver. 13.0).
62 Y. Shimizu
In a similar manner, we had S D (0.408, 0.592, 0.409, 0.534, 0.520, 0.535) for
t D 2 and obtained the results in Tables 3.6 and 3.7. Close correspondence can be
observed between the results, with a few exceptions.
3.4.4 Discussion
a b
Fig. 3.5 Post-optimal analysis in terms of MUOP. (a) Result of elite-induced MOEA. (b) Result
obtained with "-constraint method
an implicitly embedded value function, such as Eq. (3.5). This is also a basic norm
of utility theory (Fishburn 1970).
Although some results in Tables 3.5 and 3.7 seem to be somewhat far from the
reference solution, we can account for this weakness if we compare the results
in terms of the above aspects. Both results in Table 3.8 are so similar that the
DM cannot distinguish between them. Moreover, we confirmed that the best-
compromise solution could not be outperformed by any of 200 solutions obtained
with NSGA-II (Deb et al. 2000) after convergence. This numerically validates the
proposition in Sect. 3.2, which asserts that the proposed method can derive a Pareto
optimal solution.
In addition, we can use the result obtained for the post-optimal analysis combined
with a classical multi-objective analysis method, such as the " constraint method,
or recent approaches such as elite-induced evolutionary multi-objective analysis
(Shimizu et al. 2012b). As illustrated in Fig. 3.5, by producing several solutions
around the optimal result, we can move on to the next stage by choosing among
those candidates to make a final decision for actual execution. Based on the above
discussion, we again emphasize the validity of the proposed approach.
3.5 Conclusion
A MAOP method that supports flexible and adaptive decision making for complex,
diverse and competitive decision environments has been in high demand. From
this viewpoint, this study proposed a general idea for solving MAOP problems by
extending our previously proposed MUOP methods (MOON2 and MOON2R ).
64 Y. Shimizu
Although MOON2 and MOON2R require only simple and relative responses,
handling the DM’s responses in trade-off analysis becomes rather difficult in MAOP,
where more than a few objective functions are to be considered simultaneously. To
overcome this difficulty, this study proposed an approach that is easily applicable
in such cases. After presenting the general procedure, the effectiveness of the
proposed method was verified by applying it to an actual problem. The experimental
results showed that the proposed method is moderately more complex than previous
methods but maintains flexibility and adaptability. Finally, the general discussion
provided a definite and comprehensive outline of the direction of future work in
this area.
References
Coello CAC (2001) A short tutorial on evolutionary multiobjective optimization. In: Zitzler E et al
(eds) Lecture notes in computer science. Springer, Berlin, pp 21–40
Czyzak P, Jaszkiewicz AJ (1998) Pareto simulated annealing-a meta-heuristic technique for
multiple-objective combinatorial optimization. J Multi-Criteria Decis Anal 7:34–47
Deb K, Agrawal S, Pratap A, Meyarivan T (2000) A fast elitist non-dominated sorting genetic
algorithm for multi-objective optimization: NSGA-II. In: Proceedings Parallel Problem Solving
from Nature VI (PPSN-VI). Paris, France, pp 849–858
Erfani T, Utyuzhnikov SV (2012) Control of robust design in multiobjective optimization under
uncertainties. Struct Multidiscip Optim 45:247–256
Fishburn PC (1970) Utility theory for decision making. Wiley, New York
Hughes EJ (2005) Evolutionary many-objective optimization. Many once or one many? In:
Proceedings IEEE Congress on Evolutionary Computation (CEC2005). Edinburgh, UK,
pp 222–227
Jaeggi D, Parks G, Kipouros T, Clarkson J (2005) A multiobjective tabu search algorithm
for constrained optimization problems. In: Evolutionary multi-criterion optimization, third
international conference, EMO 2005, LNCS 3410. Guanajuato, Mexico, pp 490–504
Nelder JA, Mead R (1965) A simplex method for functional minimization. Comput J 7:308–313
Robic T, Filipic B (2005) DEMO: differential evolution for multi-objective optimization. In: Evo-
lutionary multi-criterion optimization, third international conference (EMO 2005), Guanajuato.
LNCS 3410, pp 520–533
Saaty TL (1980) The analytic hierarchy process. McGraw-Hill, New York
Sato H, Aguirre H, Tanaka K (2010) Many-objective evolutionary optimization by self-controlling
dominance area of solutions. Trans Jpn Soc Evol Comput 1(1):32–41
Shimizu Y (2010) An enhancement of learning optimization engineering – workbench for smart
decision making. CORONA Publishing Co., LTD. (In Japanese)
Shimizu Y, Kawada A (2002) Multi-objective optimization in terms of soft computing. Trans Soc
Instrum Control Eng 38(11):974–980
Shimizu Y, Nomachi T (2008) Integrated product design through multi-objective optimization
incorporated with meta-modeling technique. J Chem Eng Jpn 41(11):1068–1074
Shimizu Y, Tanaka Y (2003) A practical method for multi-objective scheduling through soft
computing approach. JSME Int J Ser C 46(1):54–59
Shimizu Y, Tanaka Y, Kawada A (2004) Multi-objective optimization system on the internet.
Comput Chem Eng 28(5):821–828
Shimizu Y, Miura K, Yoo J-K, Tanaka Y (2005) A progressive approach for multi-objective design
through inter-related modeling of value system and meta-model. J JSME Ser C 71(712):296–
303. (In Japanese)
3 An Extension of the MOON2 /MOON2R Approach to Many-Objective. . . 65
Shimizu Y, Yoo J-K, Tanaka Y (2006) A design support through multi-objective optimization aware
of subjectivity of value system. J JSME Ser C 72(717):1613–1620. (In Japanese)
Shimizu Y, Kato Y, Kariyahara T (2010) Prototype development for supporting multiobjective
decision making in an ill-posed environment. J Chem Eng Jpn 43(8):691–697
Shimizu Y, Waki T, Sakaguchi T (2012a) Multi-objective sequencing optimization for mixed-
model assembly line considering due-date satisfaction. J Adv Mech Des Syst Manuf 6(7):1057–
1070
Shimizu Y, Takayama M, Ohishi H (2012b) Multi-objective analysis through elite-induced
evolutionary algorithm – in the case of PSA. Trans Jpn Soc Evol Comput 3(2):22–30. (In
Japanese)
Abstract The main aim of this chapter is to investigate the existence of solutions in
connection with a class of variational-like hemivariational inequalities in reflexive
Banach spaces. Some existence theorems of solutions for the variational-like
hemivariational inequalities involving lower semicontinuous set-valued maps are
proved under different conditions. Moreover, a necessary and sufficient condition
to guarantee the existence of solutions for the variational-like hemivariational
inequalities is also given.
4.1 Introduction
Different from the fact that the variational inequality is mainly concerned with
convex energy functions, the hemivariational inequality, first introduced by Pana-
giotopulos (Panagiotopoulos 1983, 1985, 1991, 1993) in the early 1980s, is closely
G.-j. Tang
School of Science, Guangxi University for Nationalities, Nanning, Guangxi 530006, People’s
Republic of China
e-mail: guojvtang@126.com
Z.-b. Wang
Department of Mathematics, Southwest Jiaotong University, Chengdu, 610031, People’s
Republic of China
e-mail: zhongbaowang@hotmail.com
N.-j. Huang ()
Department of Mathematics, Sichuan University, Chengdu, Sichuan 610064, People’s
Republic of China
e-mail: nanjinghuang@hotmail.com
concerned with nonsmooth and nonconvex energy functions. This type of inequal-
ities and their generalization play a crucial role in describing many important
problems arising in mechanics and engineering, such as unilateral contact problems
in nonlinear elasticity, thermoviscoelastic frictional contact problems and obstacles
problems (see, for example, Carl et al. 2007; Naniewicz and Panagiotopoulos 1995;
Motreanu and Rădulescu 2003; Panagiotopoulos 1985, 1993 and the references
therein). The derivative of hemivariational inequality is based on the generalized
directional derivative introduced by Clarke (1983). In the past of almost 30 years, the
theory of hemivariational inequalities has been developed a great deal of important
results both in pure and applied mathematics as well as in other fields such as
mechanics and engineering sciences, since it allowed mathematical formulations
for some interesting problems (Carl 2001; Carl et al. 2005; Costea and Rădulescu
2009, 2010; Costea and Lupu 2010; Costea 2011; Costea et al. 2012; Costea and
Rădulescu 2012; Liu 2008; Migórski and Ochal 2004; Motreanu and Rădulescu
2000; Xiao and Huang 2009, 2008; Xiao et al. 2014; Zhang and He 2011).
On the other hand, Parida et al. (1989) introduced another new type of variational
inequality, called variational-like inequality, and showed that it can be related
to some mathematical programming problems. For more related work regarding
variational-like inequalities, we refer to Fang and Huang (2003), Bai et al. (2006),
Ansari and Yao (2001) and the references therein.
Let K be a nonempty, closed and convex subset in a real reflexive Banach space
X. Assume that A W K ⇒ X is a set-valued map, W K K ! X is a single-valued
map and W X ! R [ fC1g is a convex and lower semicontinuous functional such
that K WD K \ dom ¤ ;, where dom WD fx 2 X W .x/ < C1g is the effective
domain of . Let be a bounded open set in RN and j.x; y/ W Rk ! R be a
function. Let T W X ! Lp .I Rk / be a linear and continuous mapping, where 1 <
p < 1. We shall denote uO WD Tu and denote by jı .x; yI h/, the Clarke’s generalized
directional derivative of a locally Lipschitz mapping j.x; / at the point y 2 Rk with
respect to the direction h 2 Rk , where x 2 . We are interested in finding solutions
for the following problem:
(P) Find u 2 K such that
It is clear that a solution of problem (P) is necessarily the solution of problem (4.2)
and the converse relation is not true in general. Particularly, if A is a single-valued
4 Existence of Solutions for Variational-Like Hemivariational Inequalities. . . 69
mapping, then problems (P) coincides with (4.2) . Sometimes, a solution of problem
(P) is called as a strong solution of problem (4.2) (the similar notions can be referred
to Costea et al. 2012; Tang et al. 2014). To our best knowledge, the strong solution
of hemivariational inequalities involving set-valued maps (for example, a solution
of problem (P) other than (4.2)) was considered in few papers (Tang et al. 2014).
Moreover, we would like to mention that problems (P) and (4.2) are more general
ones because they include some problems as special cases such as:
• If j is a constant on Rk , then problems (P) and (4.2) become, respectively, as
follows:
and
and
• If A is single-valued and .v; u/ D v u, then problems both (P) and (4.2) reduce
to the problem:
Z
hA.u/; vuiC.v/.u/C jı .x; uO .x/I v.x/
O uO .x//dx 0; 8v 2 K; (4.9)
4.2 Preliminaries
Let X be a reflexive Banach space with the norm denoted by k k, X be its dual
space. For a nonempty, closed and convex subset K of X and every r > 0, we define
and
Under the hypotheses (Hj ), we can apply the Aubin-Clarke theorem (see e.g. Aubin
and Clarke 1979 or Motreanu and Rădulescu 2003) to conclude that the functional
J defined above is locally Lipschitz and
Z
J ı .wI z/ jı .x; w.x/I z.x//dx; 8w; z 2 Lp .I Rk /:
72 G.-j. Tang et al.
Consequently,
Z
J ı .OuI v/
O jı .x; uO .x/I v.x//dx;
O 8u; v 2 X: (4.11)
Definition 4.1. Let E and F be two Hausdorff topological spaces. A set-valued map
T W E ⇒ F is said to be
(i) Lower semicontinuous at x0 iff, for any open set V F such that T.x0 / \ V ¤
;, we can find a neighborhood U of x0 such that T.x/ \ V ¤ ; for all x 2 U;
(ii) Lower semicontinuous iff, it is lower semicontinuous at each x 2 E;
(iii) Lower semicontinuous iff, the restriction of T to every line segment of K is
lower semicontinuous.
We denote by G.T/ WD f.x; y/ W x 2 E and y 2 T.x/g the graph of T. It is well
known that there is an equivalent characterization for a lower semicontinuous maps
(see, for example, item (i) of Proposition 2.1 of Costea et al. 2012).
Lemma 4.2. Let E and F be two Hausdorff topological spaces. Then a set-valued
map T W E ⇒ F is lower semicontinuous iff, for any pair .x; y/ 2 G.T/ and any
net fx g 2I E converging to x, we can determine, for each 2 I, an element
y 2 T.x / such that y ! y.
The following result is a fixed point theorem for set-valued maps due to Ansari
and Yao (1999), which plays an important role in proving the existence of solutions
of problem (P) in the case of compact convex subsets in reflexive Banach spaces.
Lemma 4.3. Let K be a nonempty, closed and convex subset of a Hausdorff
topological vector space E and let S; T W K ⇒ K be two set-valued maps. Assume
that
• For each x 2 K, S.x/ be nonempty and convfS.x/g T.x/;
• K D [y2K intK S1 .y/;
• If K is not compact, there exists a nonempty, compact and convex subset C0 of K
and a nonempty and compact subset C1 of K such that, for each x 2 KnC1 , there
exists yN 2 C0 with the property that x 2 intK S1 .Ny/.
Then there exists x0 2 K such that x0 2 T.x0 /.
The next lemma is known as Mosco’s Alternative (see Mosco 1976) and plays a
crucial role in proving the existence theorems for problem (P) in the next section.
Lemma 4.4 (Mosco’s Alternative). Let K be a nonempty, compact and convex
subset of a topological space E and assume W E ! Rn [ fC1g is a proper,
convex and lower semicontinuous function such that K ¤ ;. Let ; W E E ! R
be two functions such that
• .x; y/ .x; y/ for all x; y 2 E;
• For each x 2 E, the map y 7! .x; y/ is lower semicontinuous;
• For each y 2 E, the map x 7! .x; y/ is concave.
4 Existence of Solutions for Variational-Like Hemivariational Inequalities. . . 73
Then, for each 2 R, the following alternative holds true: either there exists y0 2
K such that
In order to prove our existence results, we shall use some of the following
hypotheses, which have ever been extensively used in recent literatures (see, e.g.
Costea and Rădulescu 2009, 2012; Costea et al. 2012):
(H1A ) A W K ⇒ X is a set-valued mapping which is lower semicontinuous from K
with the strong topology into X with the weak topology, and has nonempty
values;
(H2A ) A W K ⇒ X is a set-valued mapping which is lower semicontinuous from K
with the strong topology into X with the weak topology, and has nonempty
values;
(H1 ) W K K ⇒ X is such that
(i) For all v 2 K, the map u 7! .v; u/ is continuous;
(ii) For all u; v; w 2 K and w 2 A.w/, the map v 7! hw ; .v; u/i is convex
and hw ; .v; u/i 0;
(H2 ) W K K ⇒ X is such that
(i) .u; v/ C .v; u/ D 0 for all u; v 2 K;
(ii) For all u; v; w 2 K and w 2 A.w/, the map v 7! hw ; .v; u/i is convex
and lower semicontinuous;
74 G.-j. Tang et al.
In the sequel, we shall study three cases regarding the constraint set K:
1. K a nonempty, compact and convex subset of a real reflexive Banach space X;
2. K a nonempty, bounded, closed and convex subset of a real reflexive Banach
space X;
3. K a nonempty, unbounded, closed and convex subset of a real reflexive Banach
space X.
Theorem 4.1. Let X be a real reflexive Banach space and K a nonempty, compact
and convex subset of X. Assume that (H1A ), (H1 ), (H ) and (Hj ) hold. Then problem
(P) admits at least one solution.
Proof. Arguing by contradiction, let us assume that problem (P) has no solution.
Then, for each u 2 K , there exist uN 2 A.u/ and v D v.u; uN / 2 K such that
Z
hNu ; .v; u/i C .v/ .u/ C jı .x; uO .x/I v.x/
O uO .x//dx < 0: (4.13)
Clearly, the element v for which (4.14) takes place satisfies v 2 D./, therefore
v 2 K . We consider next a set-valued map F W K ⇒ K defined by
hNu ; .w; u/i hNu ; .v1 ; u/i C .1 /hNu ; .v2 ; u/i: (4.15)
4 Existence of Solutions for Variational-Like Hemivariational Inequalities. . . 75
Since f ı . I / is positively homogeneous and subadditive (see item (i) of Lemma 4.1)
and T is linear, we get
which shows that w 2 F.u/. Therefore, F.u/ is a nonempty and convex subset of K .
Claim 2. For each v 2 K , the set F 1 .v/ D fu 2 K W v 2 F.u/g is open.
Let us fix v 2 K . Taking into account that
F 1 .v/ D fu 2 K W 9Nu 2 A.u/ s.t. hNu ; .v; u/iC.v/.u/CJ ı.OuI vO uO / < 0g;
we shall prove
For each u 2 A.u/ and 2 I, applying item (i) of Lemma 4.2, we can determine
u 2 A.u / such that
u * u
since A is lower semicontinuous from K with the strong topology into X with the
weak topology. This, together with (4.19), shows that
Using (4.18), (4.20), (4.21) and the lower semicontinuity of , one has
The compactness of K and the above claims ensure that all the conditions of
Lemma 4.3 are satisfied for S D T D F. Thus, we deduce that the set-valued map
F W K ⇒ K admits a fixed point u0 2 K , i.e., u0 2 F.u0 /. This can be rewritten
equivalently as
Z
J.'/ D j.x; '.x//dx;
and
.v; u/ .v; u/
D sup hu ; .u; v/i C inf hv ; .v; u/i ˛.v u/
u 2A.u/ v 2A.v/
Clearly,
It follows from conditions (H2 ) and (H˛ ) that the map defined by
˛. .v u0 //
hw ; .w ; u0 /i C J ı .uO0 I wO uO0 / C .w / .u0 /
hw ; .v; u0 /i C .1 /hw ; .u0 ; u0 /i C J ı .uO0 I vO uO0 /
C.1 /J ı .uO0 I uO0 uO0 / C .v/ C .1 /.u0 / .u0 /
D Œhw ; .v; u0 /i C J ı .uO0 I vO uO0 / C .v/ .u0 /;
which leads to
˛. .v u0 //
hw ; .v; u0 /i C J ı .uO0 I vO uO0 / C .v/ .u0 /: (4.24)
For each u0 2 A.u0 /, combining the fact that w ! u0 as # 0 with the fact that
A is semicontinuous, we deduce that, for each 2 .0; 1/, we can find w 2 A.w /
such that w * u0 as # 0. Taking the superior limit in (4.24) as # 0 and
keeping (H˛ ) in mind, we get
˛. .v u0 //
0 lim sup
#0
D hu0 ; .v; u0 /i
C J ı .uO0 I vO uO0 / C .v/ .u0 / (4.25)
Z
hu0 ; .v; u0 /i C .v/ .u0 / C jı .x; uO0 .x/I v.x/
O uO0 .x//dx: (by (4.11))
Therefore, we have
Remark 4.3. Theorem 4.2 generalizes some recent results in the following
aspects:
(i) If j is a constant on Rk , then Theorem 4.2 reduces to the corresponding
result of Theorem 3.3 presented by Costea et al. (2012);
(ii) From the relation of solutions between problems (P) and (4.2), we know that,
under the same assumptions as Theorem 4.2, problem (4.2) necessarily admits
at least one solution. In this case, if, in addition, .u; v/ D u v and ˛. / D 0,
then this conclusion reduces to Theorem 2 of Costea and Lupu (2010).
(iii) If A is single-valued, .u; v/ D u v and ˛. / D 0, then Theorem 4.2 reduces
to Theorem 2 of Motreanu and Rădulescu (2000);
(iv) If A is single-valued, .u; v/ D u v, D IK and ˛. / D 0, then Theorem 4.2
reduces to Corollary 3.3 of Costea and Rădulescu (2009) (or see Theorem 2 of
Panagiotopoulos et al. 1999).
Let us turn our attention to the case when K is an unbounded, closed and convex
subset of X. In order to establish the existence results of problem (P), we need to
introduce the following coercivity conditions:
(C1) There exists r0 > 0 such that, for each u 2 K nKr0 , we can find v 2 K with
kvk kuk such that
hu ; .v; u/i C .v/ .u/ C J ı .OuI vO uO / 0; for all u 2 A.u/I (4.26)
(C2) There exists r0 > 0 such that, for each u 2 K nKr0 , we can find v 2 K with
kvk kuk such that
Remark 4.4. (i) It is obvious that the implication (C2))(C1) holds as (4.11).
(ii) The conditions (C1) and (C2) can be regarded as generalization of some
coercivity conditions proposed recently by some authors. For example,
• If j is a constant on Rk , then condition (C1) reduces to condition (H2 )
of Theorem 3.5 presented by Costea et al. (2012);
• If .v; u/ D vu, then conditions (C1) and (C2) reduce to conditions (B) and
(C) presented in Proposition 4.1 of Tang and Huang (2013b), respectively;
if, in addition, D IK , then they become to conditions (B) and condition (C)
presented in Proposition 3.1 of Zhang and He (2011), respectively.
Theorem 4.3. Assume that all the assumptions of Theorem 4.2 hold except
the condition that K is bounded. If, in addition, the condition (C1) holds
for the functional J defined as (4.10), then problem (P) admits at least one
solution.
80 G.-j. Tang et al.
Proof. Let us fix r > r0 . Applying (4.25) of Theorem 4.2 as Kr is bounded, closed
and convex, we deduce that there exists ur 2 Kr \ D./ such that
(i) If kur k D r, then kur k > r0 . By condition (C1), we can find v0 2 K with
kv0 k < kur k such that
Let v 2 K be arbitrarily fixed. Since kv0 k < kur k D r, we know that there
exists t 2 .0; 1/ such that vt WD v0 C t.v v0 / 2 Kr \ D./. Note that T is
a linear mapping and is convex. It follows from (4.27), item (i) of (H2 ) and
item (i) of Lemma 4.1 that
Therefore, this, together with t 2 .0; 1/, shows that (4.30) also holds.
R
Since J.'/ D j.x; '.x//dx, by (Hj ) and (4.11), we conclude that
Since ku0 k > r0 , by the condition (C2), we know that there exists v0 2 K with
kv0 k < ku0 k such that
Z
8u0 2 A.u0 / W hu ; .v0 ; u0 /iC.v0 /.u0 /C j0 .x; uO0 .x/I vO0 .x/uO0 .x//dx < 0;
which contradicts with (4.33). Thus, it follows that the solution set is bounded,
completing the proof.
Using a similar technique to the one used in Panagiotopoulos et al. (1999), Costea
(2011), and Tang and Huang (2013b), we can provide a necessary and sufficient
condition for problem (P) and get the following result.
Theorem 4.5. Let T W X ! Lp .I Rk / be a linear compact operator, where 1 <
p < 1, k 1 and is a bounded open set in RN . Let K be a nonempty, closed
and convex subset of X. Assume that assumptions (H2 ), (H ) and (Hj ) hold. Then a
82 G.-j. Tang et al.
necessary and sufficient condition for problem (P) to have a solution is that there
exists a constant r > 0 with the property that at least one solution of the problem:
(Pr ) find ur 2 Kr \ D./ and such that
hur ; .w; ur /i "hur ; .v; ur /i C .1 "/hur ; .ur ; ur /i D "hur ; .v; ur /i: (4.35)
It follows from item (i) of Lemma 4.1 and the linearity of T that
Z
jı .x; uOr .x/I w.x/
O uOr .x//dx
Z Z
" jı .x; uOr .x/I v.x/
O uOr .x//dx C .1 "/ jı .x; uOr .x/I uOr .x/ uOr .x//dx
Z
D" jı .x; uOr .x/I v.x/
O uOr .x//dx: (4.36)
Applying (4.34) for v D w and assumption (H ) and combining (4.35) and (4.36),
one has
Dividing by " > 0, it follows that ur is a solution of problem (P). The proof is
complete.
Remark 4.6. For a suitable choice of maps and functionals such as A; and j, it is
easy to see that Theorem 4.5 can be reduced to Theorem 4.4 of Tang and Huang
(2013b) and Theorem 3 of Panagiotopoulos et al. (1999).
Acknowledgements This work was supported by the National Natural Science Foundation of
China (11171237), Guangxi Natural Science Foundation (2013GXNSFBA019015), Scientific
Research Foundation of Guangxi Department of Education (ZD2014045), Outstanding Young and
4 Existence of Solutions for Variational-Like Hemivariational Inequalities. . . 83
Middle-aged Backbone Teachers Training Project of Guangxi Colleges and Universities (Gui-Jiao-
Ren 2014-39) and Open Fund of Guangxi Key Laboratory of Hybrid Computation and IC Design
Analysis (HCIC201308).
References
Ansari QH, Yao JC (1999) A fixed point theorem and its applications to a system of variational
inequalities. Bull Aust Math Soc 59:433–442
Ansari QH, Yao JC (2001) Iterative schemes for solving mixed variational-like inequalities. J
Optim Theory Appl 108:527–541
Aubin JP, Clarke FH (1979) Shadow prices and duality for a class of optimal control problems.
SIAM J Control Optim 17:567–586
Bai MR, Zhou SZ, Ni GY (2006) Variational-like inequalities with relaxed -˛ pseudomonotone
mappings in Banach spaces. Appl Math Lett 19:547–554
Carl S (2001) Existence of extremal solutions of boundary hemivariational inequalities. J Differ
Eqn 171:370–396
Carl S, Le VK, Motreanu D (2005) Existence and comparison principles for general quasilinear
variational-hemivariational inequalities. J Math Anal Appl 302:65–83
Carl S, Le VK, Motreanu D (2007) Nonsmooth variational problems and their inequalities,
comparison principles and applications. Springer, New York
Clarke FH (1983) Optimization and nonsmooth analysis. Wiley, New York
Costea N, Rădulescu V (2009) Existence results for hemivariational inequalities involving relaxed
-˛ monotone mappings. Comm Appl Anal 13:293–304
Costea N, Rădulescu V (2010) Hartman-Stampacchia results for stably pseudomonotone operators
and nonlinear hemivariational inequalities. Appl Anal 89:175–188
Costea N, Lupu C (2010) On a class of variational-hemivariational inequalities involving set valued
mappings. Adv Pure Appl Math 1:233–246
Costea N (2011) Existence and uniqueness results for a class of quasi-hemivariational inequalities.
J Math Anal Appl 373:305–315
Costea N, Ion DA, Lupu C (2012) Variational-like inequality problems involving set-valued maps
and generalized monotonicity. J Optim Theory Appl 155:79–99
Costea N, Rădulescu V (2012) Inequality problems of quasi-hemivariational type involving set-
valued operators and a nonlinear term. J Glob Optim 52:743–756
Fang YP, Huang NJ (2003) Variational-like inequalities with generalized monotone mappings in
Banach spaces. J Optim Theory Appl 118:327–338
Goeleven D, Motreanu D, Panagiotopoulos D (1998) Eigenvalue problems for variational-
hemivariational inequalities at resonance. Nonlinear Anal 33:161–180
Liu ZH (2008) Existence results for quasilinear parabolic hemivariational inequalities. J Differ Eqn
244:1395–1409
Migórski S, Ochal A (2004) Boundary hemivariational inequality of parabolic type. Nonlinear
Anal 57:579–596
Mosco U (1976) Implicit variational problems and quasi-variational inequalities. In: Gossez JP,
LamiDozo EJ, Mawhin J, Waelbroek L (eds) Nonlinear operators and the calculus of variations.
Lecture notes in mathematics, vol 543. Springer, Berlin, pp 83–56
Motreanu D, Rădulescu V (2000) Existence results for inequality problems with lack of convexity.
Numer Funct Anal Optim 21:869–884
Motreanu D, Rădulescu V (2003) Variational and non-variational methods in nonlinear analysis
and boundary value problems. Kluwer Academic, Boston/Dordrecht/London
Naniewicz Z, Panagiotopoulos PD (1995) Mathematical theory of hemivariational inequalities and
applications. Marcel Dekker, New York
84 G.-j. Tang et al.
5.1 Introduction
Consider the convex feasibility problem (CFP) (Chinneck 2004), which is to find a
common point in the intersection of finitely many convex sets. CFP has extensive
applications in many areas such as approximation theory (Deutsch 1992), image
reconstruction from projections (Censor 1998; Herman 1980), optimal control (Gao
2009), and so on. A popular approach to the CFP is the so-called projection
Y. Dang ()
College of Computer Science and Technology (Software College), Henan Polytechnic University,
454000, Jiaozuo, People’s Republic of China
School of Management, University of Shanghai for Science and Technology, 200093, Shanghai,
People’s Republic of China.
e-mail: jgdyz@163.com
F. Meng
National Healthcare Group, Singapore City, Singapore
e-mail: fanwen_meng@nhg.com.sg
J. Sun
Department of Mathematics and Statistics, Curtin University, 6102, Bentley, WA, Australia
e-mail: jie.sun@curtin.edu.au
algorithm which employs orthogonal projection onto a set, see Bauschke and
Borwein (1996). An important special case of CFP is the split feasibility problem
(SFP), which deals with the case of finding a point in both the domain and the range
of a given linear operator. Namely, SFP is to find a point x satisfying
x 2 C; Ax 2 Q; (5.1)
where PC denotes the usual orthogonal projection onto C; that is, PC .x/ D
arg miny2C kx yk, for any x 2 C; 0 < < 2=.AT A/, and .AT A/ is the spectral
radius of AT A.
Another algorithm, the KM algorithm, was proposed initially for solving fixed
point problem (Crombez 2005), Byrne (2004) first applied KM iteration to the
CQ algorithm for solving the SFP. Subsequently, Zhao and Yang (2005) applied
KM iteration to a perturbed CQ algorithm, Dang and Gao (2011) combined the
KM iterative method with the modified CQ algorithm to construct a KM-CQ-Like
algorithm for solving the SFP. All these algorithms only use current iteration to find
the next iteration, so they tend to have slow convergence in practice.
The problem of finding a zero of a maximal monotone operator G in Euclidean
space RN is
02 k G.x/ C .x xk /; (5.3)
02 k G.x
kC1
/ C xkC1 xk k .xk xk1 /: (5.4)
It is well known that the proximal iteration (5.3) may be interpreted as an implicit
one-step discretization method for the evolution differential inclusion
dx
02 .t/ C G.x.t// a:e: t 0; (5.5)
dt
where a.e. stands for almost everywhere. While the inspiration for (5.4) comes from
the implicit discretization of the differential system of the second-order in time,
namely
d2 x dx
02 2
.t/ C .t/ C G.x.t// a:e: t 0; (5.6)
dt dt
where > 0 is a damping or a friction parameter. It gives rise to various numerical
methods (for monotone inclusions and fixed problems) related to the inertial
terminology (first introduced in Alvarez and Attouch 2001), all these methods,
as (5.4), achieve nice convergence properties (Alvarez 2000, 2004; Alvarez and
Attouch 2001; Mainge 2007, 2008) by incorporating second order information.
Inspired by the inertial proximal point algorithm for finding zeros of a maximal
monotone operator, in this paper, we apply the inertial technique to the algorithm
presented by Moudafi in 2010 to propose an inertial iterative algorithm to solve
the split common fixed-point problem for demicontractive mappings. Under some
suitable conditions, the asymptotical convergence is proved.
The paper is organized as follows. In Sect. 5.2, we recall some preliminaries.
In Sect. 5.3, we present an inertial iterative algorithm and show its convergence.
Section 5.4 summarizes the paper by making some concluding remarks.
5.2 Preliminaries
Throughout the rest of the paper, I denotes the identity operator, Fix.T/ denotes the
set of the fixed points of an operator T i.e., Fix.T/ WD fx j x D T.x/g:
An operator T W H ! H is called demicontractive (see for example Maruster and
Popirlan 2008) if there exists a constant ˇ 2 Œ0; 1/ such that
which is equivalent to
1ˇ
hx Tx; x zi kx Txk2 ; 8.x; z/ 2 H Fix.T/ (5.8)
2
and
1Cˇ
hx T.x/; z T.x/i kx T.x/k2 ; 8.x; z/ 2 H Fix.T/: (5.9)
2
88 Y. Dang et al.
An operator T W H ! H is called
(i) nonexpansive if kTx Tyk kx yk for all .x; y/ 2 H HI
(ii) quasi-nonexpansive if kTx zk kx zk for all .x; z/ 2 H Fix.T/I
(iii) strictly pseudocontractive if
Let us also recall that T is called demi-closed at the origin, if for any sequence
fxk g H and x 2 H, we have
In what follows, we will focus our attention on the following general two-operator
split common fixed-point problem (SCFP):
D fy 2 C j Ay 2 Qg:
Now, we give a description of the inertial algorithm and then present its asymptotic
convergence.
Algorithm 5.1
Initialization: Let x0 2 H1 be arbitrary.
Iterative step: For k 2 N, set u D I C AT .T I/A, and let
yk D xk C k .xk xk1 /
1
where ˛k 2 .0; 1/ and 2 .0; / with being the spectral radius of the operator AT A; k 2
Œ0; 1/.
that is,
1 k
'kC1 ky zk2
2
1 1
.1 /k.T I/.Ayk /k2 ˛k .1 ˇ ˛k /kU.u.yk // u.yk /k2 : (5.15)
2 2
Then we have
1 k 1
ky zk2 D kxk C k .xk xk1 / zk2
2 2
1 k 2
D kx zk2 C k hxk z; xk xk1 i C k kxk xk1 k2
2 2
k2 k
D 'k C k hxk z; xk xk1 i C kx xk1 k2 :
2
It is easy to verify that 'k D 'k1 C hxk z; xk xk1 i 12 kxk xk1 k2 . Hence
1 k k C k2 k
ky zk2 D 'k C k .'k 'k1 / C kx xk1 k2 : (5.16)
2 2
Putting (5.16) into (5.14), we get
k C k2 k
'kC1 'k C k .'k 'k1 / C kx xk1 k2
2
1 1
.1 /k.T I/.Ayk /k2 ˛k .1 ˇ ˛k /kU.u.yk // u.yk /k2
2 2
By the assumption on ˛k , we have
k C k2 k
'kC1 'k C k .'k 'k1 / C kx xk1 k2
2
1 1
.1 /k.T I/.Ayk /k2 ı 2 kU.u.yk // u.yk /k2 : (5.17)
2 2
1
Since 2 .0;
/ and k2 k , from (5.17), we derive
1
k kxk xk1 k2 ;
k2
92 Y. Dang et al.
and
C1
X
k kxk xk1 k2 < 1: (5.19)
kD1
Let ık WD k kxk xk1 k2 in Lemma 5.2. We deduce that the sequence fkxk zkg
PC1
is convergent (hence fx g is bounded) with kD1 Œkxk zk2 kxk1 zk2 C < 1.
k
1
.1 /k.T I/.Ayk /k2 'k 'kC1 C k .'k 'k1 / C k kxk xk1 k2 ;
2
and
1 2
ı kU.u.yk // u.yk /k2 'k 'kC1 C k .'k 'k1 / C k kxk xk1 k2 :
2
Hence,
C1
X 1
.1 /k.T I/.Ayk /k2 < 1
kD1
2
and
C1
X
ı 2 kU.u.yk // u.yk /k2 < 1:
kD1
Therefore,
and
U.x / D x : (5.24)
The paper developed an inertial algorithm and proved its weak convergence for
solving the split common fixed-point problem for demicontractive mappings in
Hilbert space. To some extent, the proposed algorithm and obtained results are
extensions of corresponding work in Dang and Gao (2011) and Moudafi (2010).
The inertial technique paves the way for investigating more effective and feasible
algorithm for the split common fixed-point problem. The strong convergence of the
algorithm is a possible future research topic.
Acknowledgements This work was partially supported by National Science Foundation of China
(under grant No.11171221), Basic and Frontier Research Program of Science and Technology
Department of Henan Province (under grants No.112300410277 and No.082300440150), China
Coal Industry Association Scientific and Technical Guidance to Project (under grant MTKJ-2011-
403), the NSTIP strategic technologies program in the Kingdom of Saudi Arabia – Award No.
(11-MAT1916-02), and research grant 71901 from Faculty of Science and Engineering, Curtin
University.
References
Alvarez F (2000) On the minizing property of a second order dissipative dynamical system in
Hilbert spaces. SIAM J Control Optim 39:1102–1119
Alvarez F, Attouch H (2001) An inertial proximal method for maximal monotone operators via
Discretization of a nonlinear oscillator with damping. Set-Valued Anal 9:3–11
Alvarez F (2004) Weak convergence of a relaxed and inertial hybrid projection-proximal point
algorithm for maximal monotone operators in Hilbert space. SIAM J Optim 3:773–782
Bauschke HH, Borwein JM (1996) On projection algorithms for solving convex feasibility
problems. SIAM Rev 38:367–426
Byrne C (2002) Iterative oblique projection onto convex sets and the split feasibility problem.
Inverse Probl 18:441–453
Byrne C (2004) An unified treatment of some iterative algorithm algorithms in signal processing
and image reconstruction. Inverse Probl 20:103–120
Chinneck JW (2004) The constraint consensus method for finding approximately feasible points
in nonlinear programs. INFORMS J Comput 16:255–265
Censor Y (1998) Parallel application of block iterative methods in medical imaging and radiation
therapy. Math Progr 42:307–325
94 Y. Dang et al.
S.K. Mishra
Department of Mathematics, Banaras Hindu University, Varanasi-221005, India
e-mail: bhu.skmishra@gmail.com
V. Singh
Department of Mathematics, National Institute of Technology, Chaltlang,
Izawal-796012, Mizoram, India
e-mail: vinaybhu1981@gmail.com
V. Laha ()
Department of Mathematics, Faculty of Science, Banaras Hindu University,
Varanasi-221005, India
e-mail: laha.vivek333@gmail.com
R.N. Mohapatra
Department of Mathematics, University of Central Florida, 4000 Central Florida Blvd.,
Orlando, FL 32816, USA
e-mail: ram.mohapatra@ucf.edu
6.1 Introduction
6.2 Preliminaries
where all the functions ; fQi ; gQi ; hQi W Rn ! R are continuously differentiable. The
feasible set of the MOP (6.1) is given by
Solving the MOP (6.1) is to find a local efficient solution or an efficient solution
which are defined as follows:
Definition 6.1. Let x 2 XQ be a feasible solution of the MOP (6.1). Then, x is said
to be a local efficient solution
T of the MOP (6.1), iff there exists a number ı > 0 such
that, there is no x 2 XQ B .x I ı/ satisfying
Definition 6.2. Let x 2 XQ be a feasible solution of the MOP (6.1). Then, x is said
to be an efficient solution of the MOP (6.1), iff there is no x 2 XQ satisfying
The following concept of tangent cones is well known in optimization (see, e.g.
Rockafellar 1970; Bajara et al. 1974; Clarke 1983).
Definition 6.3. Let QQ be a nonempty subset of Rn : The tangent cone to Q
Q at x 2
Q Q
clQ is the set T QI x defined by
Q ftn g # 0 W xn ! x and x x ! d ;
n
Q x / WD d 2 Rn j9fxn g
T.QI Q;
tn
The tangent cone T.QI Q x / is a nonempty closed cone and if Q Q is convex, then the
Q Q
cone T.QI x / is also convex. Let x 2 X be a feasible solution to the MOP (6.1),
and suppose that IQf ; IgQ and IhQ are the set of indices given by
IQf WD f1; 2; : : : ; mg
Q ;
IgQ WD fi 2 f1; 2; : : : ; pQ g jgQi .x / D 0g ; (6.3)
IhQ WD f1; 2; : : : ; qQ g :
Q k WD fx 2 Rn j gQ i .x/ 0; 8i D 1; 2; : : : ; pQ ;
Q
hQ i .x/ D 0; 8i D 1; 2; : : : ; qQ ; (6.4)
fQi .x/ fQi .x /; 8i D 1; 2; : : : ; m;
Q and i ¤ kg;
and
Q WD fx 2 Rn j gQ i .x/ 0; 8i D 1; 2; : : : ; pQ ;
Q
hQ i .x/ D 0; 8i D 1; 2; : : : ; qQ ; (6.5)
fQi .x/ fQi .x /; 8i D 1; 2; : : : ; mg:
Q
Q at x 2 Q
Definition 6.4. The linearizing cone to Q Q is the set L QI
Q x given by
Q x WD fd 2 Rn j r gQ i .x /T d 0; 8i 2 IgQ ;
L QI
r hQ i .x /T d D 0; 8i 2 IhQ ;
r fQi .x /T d 0; 8i 2 IQf g:
Q
\
m
Q x
L QI Q k I x ;
clcoT Q
kD1
The following constraint qualifications are sufficient conditions for the GGCQ to
hold at an efficient solution of the MOP (6.1).
Definition 6.6. Let x 2 XQ be an efficient solution of the MOP (6.1). Then,
(a) The Abadie constraint qualification (ACQ) holds at x iff
Q x
L QI Q x I
T QI
Q
\
m
Q x
L QI Q k I x I
T Q
kD1
(c) The Cottle constraint qualification (CCQ) holds at x iff for each k D
1; 2; : : : ; m;
Q the system
has a solution d 2 Rn I
100 S.K. Mishra et al.
(d) The Slater constraint qualification (SCQ) holds at x ; iff the objective functions
and the inequality constraints
fQi .i D 1; 2; : : : ; m/
Q ;
gQi .i D 1; 2; : : : ; pQ /
hQi .i D 1; 2; : : : ; qQ /
has a solution x 2 Rn :
(e) The linear constraint qualification (LCQ) holds at x iff the objective functions
fQi .i D 1; 2; : : : ; m/;
Q the inequality constraints gQi .i D 1; 2; : : : ; pQ /; and the
equality constraints hQi .i D 1; 2; : : : ; qQ /; are all affine.
(f) The linear objective constraint qualification (LOCQ) holds at x ; iff the
objective functions fQi .i D 1; 2; : : : ; m/;
Q are all affine, and the system
has a solution d 2 Rn I
(g) The Mangasarian-Fromovitz constraint qualification (MFCQ) holds at x ; iff
the gradients
r fQi .x / i 2 IQf
r hQ i .x / i 2 IhQ
r hQ i .x /T d D 0; 8i 2 IhQ ;
has a solution d 2 Rn :
(h) The linear independence constraint qualification (LICQ) holds at x ; iff the
gradients
r fQi .x / i 2 IQf ;
r gQ i .x / i 2 IgQ ;
r hQ i .x / i 2 IhQ
Q pQ qQ
X
m X X
Qi r fQi .x / C Q i r gQi .x / C Qi r hQi .x / D 0; (6.6)
iD1 iD1 iD1
and
Qi > 0; 8i D 1; 2; : : : ; m;
Q Q i 0; Q i gQi .x / D 0; 8i D 1; 2; : : : :; pQ : (6.7)
s:t: gi .x/ 0; 8i D 1; 2; : : : ; p;
hi .x/ D 0; 8i D 1; 2; : : : ; q;
Gi .x/ si 0; 8i D 1; 2; : : : ; r;
Hi .x/ 0; 8i D 1; 2; : : : ; r;
si 0; 8i D 1; 2; : : : ; r;
Hi .x/si D 0; 8i D 1; 2; : : : ; r:
In this section, we discuss the GGCQ for the MOPVC (6.8) under which the
Karush-Kuhn-Tucker (KKT) type necessary optimality conditions for a feasible
solution to be an efficient solution will be given. Suppose that the set X defined
by
X WD fx 2 Rn W gi .x/ 0; 8i D 1; 2; : : : ; p;
hi .x/ D 0; 8i D 1; 2; : : : ; q;
Hi .x/ 0; 8i D 1; 2; : : : ; r;
Gi .x/Hi .x/ 0; 8i D 1; 2; : : : ; rg
is the feasible set of the MOPVC (6.8), and x 2 X is an efficient solution. The
index sets x are defined as follows
If WD f1; 2; : : : ; mg;
Ig WD fi 2 f1; 2; : : : ; pgjgi .x / D 0g;
Ih WD f1; 2; : : : ; qg; (6.9)
IC WD fi 2 f1; 2; : : : ; rgjHi .x / > 0g;
I0 WD fi 2 f1; 2; : : : ; rgjHi .x / D 0g:
The index set IC .x / can be further divided into the following subsets
8
ˆ
<0;
ˆ if i 2 I00 ;
ri .x / D Gi .x /rHi .x /; if
i 2 I0C [ I0 ; (6.14)
ˆ
:̂H .x /rG .x /; if i 2 IC0 :
i i
Qk WD fx 2 Rn j gi .x/ 0; 8i D 1; 2; : : : ; p;
hi .x/ D 0; 8i D 1; 2; : : : ; q;
Hi .x/ 0; 8i D 1; 2; : : : ; r; (6.15)
Gi .x/Hi .x/ 0; 8i D 1; 2; : : : ; r
fi .x/ fi .x /; 8i D 1; 2; : : : ; m; i ¤ kg;
and
Q WD fx 2 Rn j gi .x/ 0; 8i D 1; 2; : : : ; p;
hi .x/ D 0; 8i D 1; 2; : : : ; q;
Hi .x/ 0; 8i D 1; 2; : : : ; r; (6.16)
Gi .x/Hi .x/ 0; 8i D 1; 2; : : : ; r;
fi .x/ fi .x /; 8i D 1; 2; : : : ; mg:
Proof. Suppose that i ; i D 1; 2; : : : ; r; is the function from (6.12). Then, using the
definitions of the index sets from (6.9)–(6.11), and in view of Definition 6.4, the
linearizing cone to Qk ; k D 1; : : : ; m at x 2 Qk is given by
6 Multiobjective Optimization Problems with Vanishing Constraints 105
Now, using the expression of the ri .x / for i 2 I0 [ IC0 from (6.14), and
rearranging the terms involved, we get the required representation (6.17) for the
linearizing cone to Qk ; k D 1; : : : ; m at x 2 Qk : t
u
Remark 6.1. The linearizing cone to Qk ; k D 1; : : : ; m at x 2 Qk is a nonempty,
closed and convex cone in Rn :
Moreover, for the scalar case it reduces to the linearized cone of the MPVC
given in Achtziger and Kanzow (2008, Lemma 4). Also, it is clear by the expres-
sions (6.15)–(6.17), that L .QI x / D \m k
kD1 L Q I x : Alternatively, the expression
for L .QI x / follows immediately from Achtziger and Kanzow (2008, Lemma 4),
by viewing it as a linearized cone to an MPVC with the constraints
gi .x/ 0; 8i D 1; 2; : : : ; p;
fi .x/ fi .x / 0; 8i D 1; 2; : : : ; m;
hi .x/ D 0; 8i D 1; 2; : : : ; q;
Hi .x/ 0; 8i D 1; 2; : : : ; r;
Gi .x/Hi .x/ 0; 8i D 1; 2; : : : ; r;
The following result gives the KKT type necessary optimality conditions for effi-
ciency, when the standard GGCQ holds at an efficient solution of the MOPVC (6.8).
Theorem 6.2. Let x 2 X be an efficient solution of the MOPVC (6.8). If
the standard GGCQ holds at x ; then there exist Lagrange multipliers i 2
R i 2 If ; i 2 R .i D 1; : : : ; p/ ; i 2 R .i 2 Ih / ; H
i ; i 2 R .i D 1; : : : ; r/ ; such
G
that
X
m X
p
X
q
i rfi .x /C i rgi .x / C i rhi .x /
iD1 iD1 iD1
X
r X
r
H
i rHi .x / C G
i rGi .x / D 0; (6.18)
iD1 iD1
106 S.K. Mishra et al.
and
i > 0; 8i D 1; 2; : : : ; m;
gi .x / 0; i 0; i gi .x
/ D 0; 8i D 1; 2; : : : ; p;
hi .x / D 0; 8i D 1; 2; : : : ; q; (6.19)
H
i D 0 .i 2 IC / ; H
i 0 .i 2 I00 [ I0 / ;
i free .i 2 I0C / ; i Hi .x / D 0; 8i D 1; 2; : : : ; r;
H H
i D 0 .i 2 I0 [ IC / ; i 0; .i 2 IC0 / ; i Gi .x / D 0; 8i D 1; 2; : : : ; r:
G G G
X
m X
p
X
q
X
q
i rfi .x /C i rgi .x / C iC rhi .x / i rhi .x /
iD1 iD1 iD1 iD1
X
r X
r
˛i rHi .x / C ˇi ri .x / D 0; (6.20)
iD1 iD1
and
i > 0; 8i D 1; 2; : : : ; m;
gi .x / 0; i 0; i gi .x
/ D 0; 8i D 1; 2; : : : ; p;
hi .x / 0; iC 0; iC hi .x / D 0; 8i D 1; 2; : : : ; q;
hi .x / 0; i 0; i hi .x / D 0; 8i D 1; 2; : : : ; q; (6.21)
H.x / 0; ˛i 0; ˛i Hi .x / D 0; 8i D 1; 2; : : : ; r;
i .x / 0; ˇi 0; ˇi i .x / D 0; 8i D 1; 2; : : : ; r;
iC i WD i ; 8i D 1; 2; : : : ; q;
˛i ˇi G.x / WD H
i ; 8i D 1; 2; : : : ; r; (6.22)
ˇi Hi .x / WD G
i ; 8i D 1; 2; : : : ; r;
we get the required KKT type necessary optimality conditions (6.18) and (6.19).
t
u
6 Multiobjective Optimization Problems with Vanishing Constraints 107
For the scalar case, the above KKT type necessary optimality conditions for the
MOPVC (6.8) under GGCQ reduces to the KKT conditions for the MPVC under
the standard Abadie constraint qualification given in Achtziger and Kanzow (2008,
Theorem 1). The next corollary is a direct consequence of the fact that the tangent
cones T Qk I x ; k D 1; 2; : : : ; m; contain the origin 0 2 Rn :
Corollary 6.1. Let x 2 X be an efficient solution of the MOPVC (6.8) such
that L .QI x / D f0g: Then, there exists Lagrange multipliers satisfying (6.18)
and (6.19).
Now, we give an example which verifies Corollary 6.1 with I00 ¤ :
Example 6.1. Consider the following MOPVC given by
one has
.0; 0/ D 1 rf1 .x /C 2 rf2 .x / H
1 rH1 .x / C 1 rG1 .x /
G
D 1 2; 1 1 C 1 ;
H G
and
1 0; 1 H1 .x / D 0:
H H
108 S.K. Mishra et al.
Thus 1 D 0 implies 2 D 0; and vice versa. Hence, we have 1 > 0 and 2 > 0;
and Corollary 6.1 is satisfied (Fig. 6.3).
Now, we give an example in which GGCQ does not hold for the MOPVC (6.8) with
I00 ¤
Example 6.2. Consider the following MOPVC given by
˚
L QI x D d 2 R2 jd1 D 0 ;
˚
T Q 1 I x D d 2 R2 j2d1 d12 d22 0; d2 2d1 d12 d22 0; d1 Cd22 0 ;
˚
T Q2 I x D d 2 R2 j2d1 d12 d22 0; d2 2d1 d12 d22 0; d1 0 :
Observe that
2
\
L QI x 6 clcoT Qk I x ;
kD1
and hence the standard GGCQ does not hold at x : Now, for any Lagrange
multipliers 1 0; 2 0; not both zero, and H
1 0; 1 0; one has
G
.0; 0/ D 1 rf1 .x /C 2 rf2 .x / H
1 rH1 .x / C 1 rG1 .x /
G
D 1 2 1 ; 1 ;
2H G
which does not satisfy the necessary optimality conditions (6.18) and (6.19).
Examples 6.1 and 6.2 show that GGCQ is not always violated when I00 ¤ ; but
it may not hold sometimes when I00 ¤ : Following example shows that GGCQ is
not a sufficient condition for the existence of positive Lagrange multipliers for the
MOPVC (6.8) (Fig. 6.5).
Example 6.3. Consider the following MOPVC given by
.0; 0/ D 1 rf1 .x /C 2 rf2 .x / H
1 rH1 .x / C 1 rG1 .x /
G
D 1 1 ; 1 C 1 :
H H G
1 D 1 0; 1 D 0 implies
Observe that for H G
2 0; which violets the existence
of positive Lagrange multipliers (Fig. 6.7).
6 Multiobjective Optimization Problems with Vanishing Constraints 111
It was shown in Achtziger and Kanzow (2008) that some constraint qualifications
like LICQ and MFCQ do not hold under fairly mild assumptions, whereas some
constraint qualifications like ACQ may not hold sometimes at a local minimum
of the MPVC. In this section, we investigate some more constraint qualifications
like CCQ, SCQ, LOCQ and LCQ for the MOPVC (6.8) and modify them where
necessary to use them as sufficient conditions for the GGCQ to hold at an efficient
solution of the MOPVC (6.8). The next result shows that under fairly reasonable
assumptions CCQ is not satisfied at an efficient solution of the MOPVC (6.8).
Lemma 6.2. Let x 2 X be an efficient solution of the MOPVC (6.8) with I00 [
I0C ¤ : Then, the standard CCQ is not satisfied at x :
Proof. Suppose that CCQ is satisfied at x : Then, for each k D 1; : : : ; m; the system
has a solution d 2 Rn : Using the gradient of i from (6.14) in (6.23), one has
and
1
rHi .x /T d D ri .x /d < 0; 8i 2 I0C ;
Gi .x /
Corollary 6.3. Let x be an efficient solution of the MOPVC (6.8) with I00 [ I0C ¤
: Then, the standard SCQ is not satisfied at x :
Since LICQ implies MFCQ, the following result is a direct consequence of
Corollary 6.2, and is a multiobjective analog of Achtziger and Kanzow (2008,
Lemma 2).
Corollary 6.4. Let x be an efficient solution of the MOPVC (6.8) with I00 [ I0C ¤
: Then, the standard LICQ is not satisfied at x : Moreover, if I0 ¤ ; then also the
standard LICQ is not satisfied at x :
The proof of the following result is similar to the proof of Lemma 6.2, and it
shows that LOCQ is also not satisfied at an efficient solution of the MOPVC (6.8)
when I00 [ I0C ¤ :
Lemma 6.3. Let x be an efficient solution of the MOPVC (6.8) with I00 [I0C ¤ :
Then, the standard LOCQ is not satisfied at x :
The above results show that under fairly mild assumptions most of the constraint
qualifications are violated at an efficient solution of the MOPVC (6.8), and hence
we introduce some constraint qualifications as modifications of the standard CCQ,
MFCQ, SCQ, LICQ and LOCQ for the MOPVC (6.8).
Definition 6.7. Let x be an efficient solution of the MOPVC (6.8). Then the
Cottle-Type constraint qualification for the MOPVC (6.8), denoted by CCQ-
MOPVC, holds at x iff for each k D 1; : : : ; m; the system
has a solution d 2 Rn :
It is clear that CCQ-MOPVC is different from the standard CCQ for the
MOPVC (6.8), and is a fair assumption. We now show that CCQ-MOPVC is a
sufficient condition for the GGCQ provided that the critical index set I00 D :
Lemma 6.4. Let x be an efficient solution of the MOPVC (6.8) with I00 D : If
CCQ-MOPVC is satisfied at x then the standard GGCQ is also satisfied at x :
Proof. Suppose that x is an efficient solution of the MOPVC (6.8) with I00 D :
Then, the MOPVC (6.8) is locally equivalent to the following MOP:
6 Multiobjective Optimization Problems with Vanishing Constraints 113
Now, when CCQ-MOPVC is satisfied at x with I00 D ; then the standard CCQ
for the MOP (6.25) will also hold at x ; and hence the standard GGCQ for the
MOP (6.25) will also be satisfied at x ; that is,
\
m
O x
L QI clcoT QO k I x ;
kD1
O k ; k D 1; : : : ; m; and Q
where Q O are defined as
O k WD fx 2 Rn j fi .x/ fi .x /; 8i 2 If ; i ¤ k;
Q
gi .x/ 0; 8i 2 Ig ;
hi .x/ D 0; 8i 2 Ih ; (6.26)
Hi .x/ D 0; 8i 2 I0C ;
Hi .x/ 0; 8i 2 I0 ;
Gi .x/ 0; 8i 2 IC0 g;
and
O WD fx 2 Rn j fi .x/ fi .x /; 8i 2 If ;
Q
gi .x/ 0; 8i 2 Ig ;
hi .x/ D 0; 8i 2 Ih ; (6.27)
Hi .x/ D 0; 8i 2 I0C ;
Hi .x/ 0; 8i 2 I0 ;
Gi .x/ 0; 8i 2 IC0 g:
O x is given by
Also, the linearizing cone L QI
114 S.K. Mishra et al.
O x D fd 2 Rn W rfi .x /T d 0; 8i 2 If ;
L QI
rgi .x /T d 0; 8i 2 Ig ;
rhi .x /T d D 0; 8i 2 Ih ; (6.28)
T
rHi .x / d D 0; 8i 2 I0C ;
rHi .x /T 0; 8i 2 I0 ;
rGi .x /T 0; 8i 2 IC0 g:
x / with
which in view of Lemma 6.1 is nothing but the linearizing cone L .QI
O
I00 D : Now, since Q O I x
Q , k D 1; 2; : : : ; m; it follows that T Q
k k
\
m \
m
clcoT QO k I x clcoT Qk I x ;
kD1 kD1
Observe that
2
\
L QI x 6 clcoT Qk I x ;
kD1
6 Multiobjective Optimization Problems with Vanishing Constraints 115
and hence the GGCQ-MOPVC is not satisfied at x for the given MOPVC. But, the
system given by (6.24) is solvable for x ; and hence CCQ-MOPVC holds at x for
the given MOPVC (Fig. 6.9).
Now, we give a constraint qualification for the MOPVC (6.8), which is a modified
version of the standard MFCQ, and is a multiobjective analog of VC-MFCQ
introduced in Achtziger and Kanzow (2008).
Definition 6.8. Let x 2 X be an efficient solution of the MOPVC (6.8). Then,
the Mangasarian-Fromovitz constraint qualification (MFCQ) for the MOPVC (6.8),
denoted by MFCQ-MOPVC, holds at x iff the gradients
116 S.K. Mishra et al.
rfi x i 2 If ;
rhi x .i 2 Ih / ;
rHi x .i 2 I00 [ I0C / ;
rfi x
T
d D 0; 8i 2 If ;
T
rgi x d < 0; 8i 2 Ig ;
rhi x
T
d D 0; 8i 2 Ih ; (6.29)
rHi x
T
d D 0; 8i 2 I0C [ I00 ;
rHi x
T
d > 0; 8i 2 I0 ;
rGi x
T
d < 0; 8i 2 IC0 ;
has a solution d 2 Rn :
The following result gives the relationship between the CCQ-MOPVC and the
MFCQ-MOPVC.
Lemma 6.5. Let x 2 X be an efficient solution of the MOPVC (6.8). If MFCQ-
MOPVC holds at x 2 X; then CCQ-MOPVC also holds at x :
Proof. Suppose that the MFCQ-MOPVC holds at x ; but the CCQ-MOPVC does
not hold at x : Then, there exists k 2 f1; : : : ; mg such that the system (6.24) has no
solution d 2 Rn : By Motzkin’s theorem of the alternative Mangasarian (1969), there
exist real numbers i 0 i 2 If ; i ¤ k ; i 0 i 2 Ig ; H 0 .i 2 I0 / ; G
i S i
0 .i 2 IC0 / W not all zero, and i 2 R .i 2 Ih / ;
ei 2 R .i 2 I0C I00 / ; such that
H
X
m X X
i rfi .x /C i rgi .x /C i rhi .x /
iD1 i2Ig i2Ih
i¤k
X X X
ei H rHi .x /
H
i rHi .x / C G
i rGi .x / D 0: (6.30)
S
i2I00 I0C i2I0 i2IC0
Suppose that d 2 Rn solves the systems (6.29), then from (6.30), one has
X X X
T T T
i rgi .x / d H
i rHi .x / d C G
i rGi .x / d D 0:
i2Ig i2I0 i2IC0
6 Multiobjective Optimization Problems with Vanishing Constraints 117
i D 0; 8i 2 Ig ;
H
i D 0; 8i 2 I0 ;
i D 0; 8i 2 IC0 :
G
X
m X X
i rfi .x /C i rhi .x / ei H rHi .x / D 0:
S
iD1 i2Ih i2I00 I0C
i¤k
Since
rfi x i 2 If ;
rhi x .i 2 Ih / ;
rHi x .i 2 I00 [ I0C /
i D 0; 8i 2 If ; i ¤ k;
i D 0; 8i 2 Ih ;
ei H D 0; 8i 2 I00 [ I0C ;
a contradiction to the existence of not all zero Lagrange multipliers, and hence the
result. t
u
The following constraint qualification is a modification of the standard SCQ for
the MOPVC (6.8) and serves as a sufficient condition for the CCQ-MOPVC to hold.
Definition 6.9. Let x 2 X be an efficient solution of the MOPVC (6.8). Then,
the Slater-type Constraint Qualification for the MOPVC (6.8), denoted by SCQ-
MOPVC, holds at x 2 X; iff the functions
fi i 2 If ;
gi i 2 Ig ;
Gi .i 2 IC0 / ;
Hi .i 2 I0 /
118 S.K. Mishra et al.
Hi .i 2 I00 [ I0C / ;
hi .i 2 Ih /
fi .x/ < fi x 8i 2 If ; i ¤ k;
gi .x/ < 0; 8i 2 Ig ;
hi .x/ D 0; 8i 2 Ih ; (6.31)
Hi .x/ D 0; 8i 2 I0C [ I00 ;
Hi .x/ > 0; 8i 2 I0 ;
Gi .x/ < 0; 8i 2 IC0 ;
has a solution x 2 Rn :
The following result gives the relationship between the CCQ-MOPVC and the
SCQ-MOPVC.
Lemma 6.6. Let x 2 X be an efficient solution of the MOPVC (6.8). If SCQ-
MOPVC holds at x ; then CCQ-MOPVC also holds at x :
Proof. Suppose that SCQ-MOPVC holds at x : Then, for each k D 1; : : : ; m; there
exists an xk 2 Rn such that
fi xk < fi x ; 8i 2 If ; i ¤ k;
gi xk < 0; 8i 2 Ig ;
hi xk D 0; 8i 2 Ih ; (6.32)
Hi xk D 0; 8i 2 I0C [ I00 ;
Hi xk > 0; 8i 2 I0 ;
Gi xk < 0; 8i 2 IC0 :
rfi x xk x fi xk fi x < 0; 8i 2 If ; i ¤ k;
T
rgi x xk x gi xk gi x < 0; 8i 2 Ig ;
T
rhi x xk x D hi xk hi x D 0; 8i 2 Ih ;
T
(6.33)
6 Multiobjective Optimization Problems with Vanishing Constraints 119
fi i 2 If
gi i 2 Ig ;
hi .i 2 Ih / ;
Hi .i 2 I0 / ;
Hi .i 2 IC0 / ;
fi i 2 If
rfi x
T
d 0; 8i 2 If ;
rgi x
T
d < 0; 8i 2 Ig ;
rhi x
T
d D 0; 8i 2 Ih ; (6.34)
rHi x
T
d D 0; 8i 2 I0C [ I00 ;
T
rHi x d > 0; 8i 2 I0 ;
rGi x
T
d < 0; 8i 2 IC0 ;
has a solution d 2 Rn :
120 S.K. Mishra et al.
The following result gives the relationship between the LCQ-MOPVC and the
standard GGCQ.
Lemma 6.7. Let x 2 X be an efficient of the MOPVC (6.8) such that I00 D : if
LCQ-MOPVC holds at x ; then the standard GGCQ also holds at x :
Proof. Suppose that x 2 X is an efficient solution of the MOPVC (6.8) with I00 D
: Then, the MOPVC (6.8) is locally equivalent to the MOP (6.25). Hence, for
I00 D ; LCQ-MOPVC is identical to the standard LCQ of the MOP (6.25.) Since,
LCQ of the MOP (6.25) holds at x ; it follows that GGCQ of the MOP (6.25) also
holds at x ; and proceeding as in Lemma 6.4, we get the required result. t
u
The proof of the following result is similar to the proof of Lemma 6.7.
Lemma 6.8. Let x 2 X be an efficient solution of the MOPVC (6.8) such that
I00 D . If LOCQ-MOPVC holds at x ; then GGCQ also holds at x :
Now, we give a constraint qualification of the MOPVC (6.8), which serves as a
sufficient condition for the MFCQ-MOPVC to hold, and is a multiobjective analog
of VC-LICQ introduced in Achtziger and Kanzow (2008).
Definition 6.12. Let x 2 X be an efficient solution of the MOPVC (6.8). Then,
the linear independence constraint qualification of the MOPVC (6.8), denoted by
LICQ-MOPVC, holds at x ; iff for each k D 1; 2; : : : ; m; the gradients
rfi x i 2 If ; i ¤ k ;
rgi x i 2 Ig ;
rhi x .i 2 Ih / ;
rHi x .i 2 I0 / ;
rGi x .i 2 IC0 / ;
LOCQ-MOPVC
LICQ-MOPVC SCQ-MOPVC
It was observed in Sect. 6.3 that the standard GGCQ may or may not hold at an
efficient solution of the MOPVC (6.8) when I00 ¤ : In this section, we introduce
a suitable modification of the GGCQ of the MOPVC (6.8), and use it to prove
necessary optimality conditions for efficiency in MOPVC (6.8), that are different
from the standard KKT conditions given by Theorem 6.2. We also provide various
sufficient conditions for the modified GGCQ to hold.
In order to define a modified Guignard constraint qualification, we intro-
duce a nonlinear multiobjective optimization problem (NLMOP) derived from the
MOPVC (6.8) depending on an efficient solution x 2 X as follows
k
Also, define the sets Q and Q as follows
k
Q WD fx 2 Rn j fi .x/ fi x ; 8i D 1; 2; : : : ; m; i ¤ k;
gi .x/ 0; 8i D 1; 2; : : : ; p;
hi .x/ D 0; 8i D 1; 2; : : : ; q; (6.36)
Hi .x/ D 0; Gi .x/ 0; 8i 2 I0C ;
Hi .x/ 0; Gi .x/ 0; 8i 2 I0 [ I00 [ IC0 [ IC g
and
Q WD fx 2 Rn j fi .x/ fi x ; 8i D 1; 2; : : : ; m;
gi .x/ 0; 8i D 1; 2; : : : ; p;
hi .x/ D 0; 8i D 1; 2; : : : ; q; (6.37)
Hi .x/ D 0; Gi .x/ 0; 8i 2 I0C ;
Hi .x/ 0; Gi .x/ 0; 8i 2 I0 [ I00 [ IC0 [ IC g:
k
The linearizing cone Q at x 2 X is given by
k
L.Q I x / D fd 2 Rn j rfi .x /T d 0; 8i D 1; : : : ; m; i ¤ k;
rgi .x /T d 0; 8i 2 Ig ;
rhi .x /T d D 0; 8i 2 Ih ; (6.38)
rHi .x /T d D 0; 8i 2 I0C ;
rHi .x /T d 0; 8i 2 I00 [ I0 ;
rGi .x /T d 0; 8i 2 I00 [ IC0 g:
\
m k
L QI x D L Q I x : (6.39)
kD1
The
k
following
lemma gives the relationship between the tangent cones
T Q I x ; k D 1; 2; : : : ; m; and the linearizing cone L.QI x /:
Lemma 6.10. Let x 2 be an efficient solution of the MOPVC (6.8). Then, we have
\
m k
clcoT Q I x L QI x :
kD1
6 Multiobjective Optimization Problems with Vanishing Constraints 123
\
m
clcoT Qk I x L QI x ; (6.40)
kD1
and
\
m k \
m k
clcoT Q I x L Q I x D L QI x : (6.41)
kD1 kD1
k
Also, since Q Qk , 8k D 1; : : : ; m; one has
k
T Q I x T Qk I x ; 8i D 1; : : : ; m; (6.42)
and
k
L Q I x L Qk I x ; 8i D 1; : : : ; m: (6.43)
\
m k
clcoT Q I x L QI x ; (6.44)
kD1
\
m k
L QI x clcoT Q I x :
kD1
\
m
L QI x clcoT Qk I x :
kD1
124 S.K. Mishra et al.
The following result gives the relationship between the GGCQ-NLMOP and the
GGCQ-MOPVC.
Lemma 6.11. Let x 2 X be an efficient solution of the MOPVC (6.8). If the
GGCQ-NLMOP holds at x ; then the GGCQ-MOPVC also holds at x 2 X:
Proof. Suppose that x 2 X is an efficient solution of the MOPVC (6.8) such that
GGCQ-NLMOP holds at x ; then one has
\
m k
L QI x clcoT Q I x : (6.45)
kD1
k
k
Since, Q Qk ; 8k D 1; 2; : : : ; m; it follows that T Q I x
T Qk I x ; 8k D 1; 2; : : : ; m; and hence
\
m k \
m
clcoT Q I x clcoT Qk I x : (6.46)
kD1 kD1
k
Also, we always have L Q I x L Qk I x ; 8k D 1; : : : ; m; which follows
that
\
m k \
m
L Q I x L Qk I x : (6.47)
kD1 kD1
\
m
L QI x clcoT Qk I x ;
kD1
\
m
L QI x L QI x clcoT Qk I x :
kD1
The following example says that GGCQ-MOPVC is strictly weaker than the
standard GGCQ.
6 Multiobjective Optimization Problems with Vanishing Constraints 125
It is clear that
˚
T Q1 ; x D d 2 R2 W d1 C d2 0; d1 .d1 C d2 / 0; d2 d1 0 ;
˚
T Q2 ; x D d 2 R2 W d1 C d2 0; d1 .d1 C d2 / 0; d1 C d2 0 ;
˚
L Q; x D d 2 R2 W d1 C d2 0; d2 d1 0; d1 C d2 0 ;
˚
L Q; x D d 2 R2 W d1 C d2 0; d2 d1 0; d1 C d2 0; d1 0 :
whereas
2
\
L Q; x clcoT Qk ; x ;
kD1
hence the GGCQ-MOPVC holds whereas the standard GGCQ is not satisfied at x :
When GGCQ-MOPVC holds at an efficient solution x 2 X of the MOPVC (6.8),
the KKT conditions of Theorem 6.2 may not hold, since GGCQ-MOPVC is weaker
than the standard GGCQ. Hence, in the following result, we derive KKT type
necessary optimality conditions for efficiency under GGCQ-MOPVC.
126 S.K. Mishra et al.
Theorem 6.4. Let x 2 X be an efficient solution of the MOPVC (6.8) such that
the GGCQ-MOPVC holds at x . Then, there exist Lagrange multipliers i 2 R.i D
1; : : : ; m/; i 2 R.i D 1; : : : ; p/; i 2 R.i D 1; : : : ; q/; H
i ; i 2 R.i D 1; : : : ; r/;
G
such that
X
m X
p
X
q
i rfi .x / C i rgi .x / C i rhi .x /
iD1 iD1 iD1
X
r X
r
H
i rHi .x / C G
i rGi .x / D 0; (6.48)
iD1 iD1
and
i > 0; 8i D 1; 2; : : : ; m;
gi .x / 0; i 0; i gi .x
/ D 0; 8i D 1; 2; : : : ; p;
hi .x / D 0; 8i D 1; 2; : : : ; q; (6.49)
H
i D 0 .i 2 IC / ; H
i 0 .i 2 I00 [ I0 / ;
i free .i 2 I0C / ; i Hi .x / D 0; 8i D 1; 2; : : : ; r;
H H
i Gi .x / D 0; 8i D 1; 2; : : : ; r:
G
Proof. Suppose that x is an efficient solution of the MOPVC (6.8). We will first
show that the system
rfi x
T
d 0; 8i D 1; 2; : : : ; m;
rfi x
T
d < 0; at least one i;
rgi x
T
d 0; 8i 2 Ig ;
T
rhi x d D 0; 8i 2 Ih ; (6.50)
rHi x
T
d D 0; 8i 2 I0C ;
rHi x
T
d 0; 8i 2 I00 [ I0 ;
rGi x
T
d 0; 8i 2 I00 [ IC0 ;
has no solution d 2 Rn : Suppose to the contrary that there exists d 2 Rn such that the
system (6.50) is solvable, then d 2 L QI x : Without loss of generality, we may
assume that
6 Multiobjective Optimization Problems with Vanishing Constraints 127
rfi x
T
d < 0; for some k; 1 k m;
T
rfi x d 0; 8i D 1; : : : ; m; i ¤ k:
d 2 clcoT Qk I x ; 1 k m:
Hence, there exists a sequence fdl g coT Qk I x ; 1 k m; such that
dl ! d: Now, for each dl ; l D 1; 2; : : : ; there exist number l ; lj 0; and
dlj 2 T Qk I x ; 1 k m; j D 1; 2; : : : ; l ; such that
l l
X X
lj D 1; lj dlj D dl :
jD1 jD1
rfk x
T
dlj 0; 1 k m;
rfi x
T
dlj 0; 8i D 1; : : : ; m; i ¤ k;
rgi x
T
dlj 0; 8i 2 Ig ;
T
rhi x dlj D 0; 8i 2 Ih ;
rHi x
T
dlj 0; 8i 2 I0C [ I00 [ I0 ;
ri x
T
dlj 0; 8i 2 I0C [ I00 [ I0 [ IC0 :
128 S.K. Mishra et al.
By linearity and continuity of the inner product and using the expression of the
ri .x / from (6.14), it follows that
rfk x
T
d 0; 1 k m;
rfi x
T
d 0; 8i D 1; : : : ; m; i ¤ k;
rgi x
T
d 0; 8i 2 Ig ;
T
rhi x d D 0; 8i 2 Ih ;
rHi x
T
d D 0; 8i 2 I0C ;
rHi x
T
d 0; 8i 2 I00 [ I0 ;
rGi x
T
d 0; 8i 2 IC0 ;
i 2 R .1; : : : ; m/ ; i 2 R i 2 Ig ; i 2 R .1; : : : ; q/ ;
H
i 2 R .i 2 I0C [ I00 [ I0 / ; i 2 R .i 2 I00 [ IC0 / ;
G
such that
X
m X X
q
i rfi .x / C i rgi .x / C i rhi .x /
iD1 i2Ig iD1
X X
H
i rHi .x / C G
i rGi .x / D 0;
i2I0 i2I00 [IC0
and
i > 0; 8i D 1; 2; : : : ; m;
0; 8i 2 Ig ;
i 2 R; 8i D 1; 2; : : : ; q;
i 0; 8i 2 I00 [ IC0 :
G
6 Multiobjective Optimization Problems with Vanishing Constraints 129
Setting
H
i D 0; 8i 2 IC ;
\
m
L QI x T Qk I x :
kD1
\
m k
L QI x T Q I x :
kD1
L QI x T QI x :
L QI x T QI x :
The following result is a direct consequence of Definitions 6.15, 6.17 and 6.18.
Lemma 6.13. Let x 2 X be an efficient solution of the MOPVC (6.8). If the ACQ-
MOPVC holds at x ; then the GACQ-MOPVC is also satisfied at x : Moreover, if
ACQ-NLMOP holds at x then the ACQ-MOPVC and the standard ACQ both are
satisfied at x :
We now give some more sufficient conditions which assure that the GGCQ-
MOPVC holds at an efficient solution of the MOPVC (6.8).
Theorem 6.5. Let x be an efficient solution of the MOPVC (6.8), and consider the
following conditions:
1. The standard GGCQ holds for the NLMOP (6.35) at x I
2. For each k D 1; : : : ; m; there exists a vector dO 2 Rn satisfying
rfi x dO < 0; 8i D 1; 2; : : : ; m; i ¤ k;
T
rgi x dO < 0; 8i 2 Ig ;
T
rhi x dO D 0; 8i 2 Ih ;
T
(6.53)
rHi x dO D 0; 8i 2 I0C ;
T
fi .x/ < fi x ; 8i D 1; 2; : : : ; m; i ¤ k;
gi .x/ < 0; 8i D 1; 2; : : : ; p;
hi .x/ D 0; 8i D 1; 2; : : : ; q; (6.54)
Hi .x/ D 0; Gi .x/ > 0; 8i 2 I0C ;
Hi .x/ > 0; Gi .x/ < 0; 8i 2 I0 [ I00 [ IC0 [ IC I
rfi x dO 0; 8i D 1; 2; : : : ; m;
T
rgi x dO < 0; 8i 2 Ig ;
T
rhi x dO D 0; 8i 2 Ih ;
T
(6.55)
rHi x dO D 0; 8i 2 I0C ;
T
6. The gradients
rfi x .i D 1; : : : ; m/ ;
rhi x .i D 1; : : : ; q/ ;
rHi x .i 2 I0C / ;
rfi x dO D 0; 8i D 1; 2; : : : ; m;
T
rhi x dO D 0; 8i D 1; 2; : : : ; q;
T
rHi x dO D 0; 8i 2 I0C ;
T
(6.56)
rgi x dO < 0; 8i 2 Ig ;
T
7. The gradients
rfi x .i D 1; 2; : : : m/ ;
rgi x i 2 Ig ;
rhi x .i D 1; : : : q/ ;
rHi x .i 2 I0 / ;
rGi x .i 2 I00 [ IC0 / ;
\
m k
L QI x clcoT Q I x :
kD1
k
k
Also, Q Qk ; k D 1; 2; : : : ; m; which implies that T Q I x T Qk I x ; k D
1; 2; : : : ; m; and hence
\
m k \
m
clcoT Q I x clcoT Qk I x :
kD1 kD1
\
m
L QI x clcoT Qk I x ;
kD1
GGCQ-MOPVC
GGCQ GGCQ-NLMOP
6.6 Conclusions
Acknowledgements The authors are thankful to the anonymous referees for their valuable
comments and suggestions which helped to improve this chapter in its present form. This work
was done when Vinay Singh was a Post Doctoral Fellow of National Board of Higher Mathematics
(NBHM), Department of Atomic Energy (DAE), Government of India and Vivek Laha was a
Senior Research Fellow of the Council of Scientific and Industrial Research (CSIR), New Delhi,
Ministry of Human Resources Development, Government of India at Department of Mathematics,
Banaras Hindu University.
Currently, Vivek Laha is supported by the Postdoctoral Fellowship of National Board of Higher
Mathematics, Department of Atomic Energy, Government of India (Ref. No. 2/40(47)/2014/R &
D-II/1170).
134 S.K. Mishra et al.
References
Abadie JM (1967) On the Kuhn-Tucker theorem. In: Abadie JM (eds) Nonlinear programming.
Wiley, New York, pp 21–36
Achtziger W, Kanzow C (2008) Mathematical programs with vanishing constraints: optimality
conditions and constraint qualifications. Math Progr 114:69–99
Aghezzaf B, Hachimi M (2004) Second order duality in multiobjective programming involving
generalized type I functions. Numer Funct Anal Optim 25(7–8):725–736
Aghezzaf B, Hachimi M (2001) Sufficiency and duality in multiobjective programming involving
generalized .F; /convexity. J Math Anal Appl 258:617–628
Bajara BS, Goode JJ, Nashed MZ (1974) On the cones of tangents with applications to
mathematical programming. J Optim Theory Appl 13:389–426
Bigi G, Pappalardo M (1999) Regularity conditions in vector optimization. J Optim Theory Appl
102(1):83–96
Bao TQ, Mordukhovich BS (2007) Existence of minimizers and necessary conditions in set-valued
optimization with equilibrium constraint. Appl Math 52:453–472
Bao TQ, Gupta P, Mordukhovich BS (2007) Necessary conditions in multiobjective optimization
with equilibrium constraints. J Optim Theory Appl 135:179–203
Bao TQ, Gupta P, Mordukhovich BS (2008) Suboptimality conditions for mathematical programs
with equilibrium constraints. Taiwan. J Math 12(9):2569–2592
Chinchuluun A, Pardalos PM (2007) A survey of recent developments in multiobjective optimiza-
tion. Ann Oper Res 154:29–50
Clarke FH (1983) Optimization and nonsmooth analysis. Wiley-Interscience, New York
Eschenauer H, Koski J, Osyczka A (eds) (1990) Multicriteria design optimization: procedures and
applications. Springer, Berlin
Facchinei F, Pang J-S (2003) Finite-dimensional variational inequalities and complementarity
problems. Springer, New York
Gould FJ, Tolle JW (1971) A necessary and sufficient qualification for constrained optimization.
SIAM J Appl Math 20:164–172
Guignard M (1969) Generalized Kuhn-Tucker conditions for mathematical programming problems
in a Banach space. SIAM J Contr 7:232–241
Hoheisel T, Kanzow C (2007) Würzburg: first and second order optimality conditions for
mathematical programs with vanishing constraints. Appl Math 52(6):495–514
Hoheisel T, Kanzow C (2008) Stationary conditions for mathematical programs with vanishing
constraints using weak constraint qualifications. J Math Anal Appl 337:292–310
Hoheisel T, Kanzow C (2009) On the Abadie and Guignard constraint qualifications for mathemat-
ical programmes with vanishing constraints. Optimization 58(4):431–448
Hoheisel T, Kanzow C, Outrata JV (2010) Exact penalty results for mathematical programs with
vanishing constraints. Nonlinear Anal 72:2514–2526
Izmailov AF, Solodov MV (2009) Mathematical programs with vanishing constraints: optimality
conditions, sensitivity, and a relaxation method. J Optim Theory Appl 142:501–532
Koski J (1993) Multicriteria optimization in structural design: state of the art. In: Proceedings of
the 19th design automation conferences, Albuquerque. ASME, pp 621–629
Li XF (2000) Constraint qualifications in nonsmooth multiobjective optimization. J Optim Theory
Appl 106(2):373–398
Liang Z-A, Huang H-X, Pardalos PM (2003) Efficiency conditions and duality for a class of
multiobjective fractional programming problems. J Global Optim 27:447–471
Lin P, Zhou P, Wu CW (2011) Multiobjective topology optimization of end plates of proton
exchange membrane fuel cell stacks. J Power Sources 196(3):1222–1228
Luo Z-Q, Pang J-S, Ralph D (1996) Mathematical programs with equilibrium constraints.
Cambridge University Press, Cambridge
Maeda T (1994) Constraint qualifications in multiobjective optimization problems: differentiable
case. J Optim Theory Appl 80(3):483–500
6 Multiobjective Optimization Problems with Vanishing Constraints 135
Maeda T (2004) Second order conditions for efficiency in nonsmooth multiobjective optimization
problems. J Optim Theory Appl 122(3):521–538
Mangasarian OL (1969) Nonlinear programming. McGrawHill, New York
Min S, Nishiwaki S, Kikuchi N (2000) Unified topology design of static and vibrating structures
using multiobjective optimization. Comput Struct 75:93–116
Mishra SK, Wang, S-Y, Lai KK (2005) Optimality and duality for multiple-objective optimization
under generalized type I univexity. J Math Anal Appl 303:315–326
Mordukhovich BS (2004) Equilibrium problems with equilibrium constraints via multiobjective
optimization. Optim Methods Softw 19:479–492
Mordukhovich BS (2006) Variational analysis and generalized differentiation, II: applications.
Grundlehren series (fundamental principles of mathematical sciences), vol 331. Springer,
Berlin
Mordukhovich BS (2009) Multiobjective optimization problems with equilibrium constraints.
Math Progr Ser B 117:331–354
Outrarata JV, Kočvara M, Zowe J (1998) Nonsmooth approach to optimization problems with
equilibrium constraints. Kluwer Academic, Dordrecht
Peterson DW (1973) A review of constraint qualifications in finite-dimensional spaces. SIAM Rev
15:639–654
Preda V, Chitescu I (1999) On constraint qualifications in multiobjective optimization problems:
semidifferentiable case. J Optim Theory Appl 100(2):417–433
Rockafellar RT (1970) Convex analysis. Princeton University Press, Princeton
Stadler W (1984) Multicriteria optimization in mechanics. Appl Mech Rev 37:277–286
Chapter 7
A New Hybrid Optimization Algorithm for
the Estimation of Archie Parameters
Jianjun Liu, Honglei Xu, Guoning Wu, and Kok Lay Teo
Abstract Archie formula, which contains three fundamental parameters (a, m, n),
is the basic equation to compute the water saturation in a clean or shaly formation.
These parameters are known as Archie parameters. To identify accurately the water
saturation for a given reservoir condition, it depends critically on the accurate
estimates of the values of Archie parameters (a, m, n). These parameters are
interdependent and hence it is difficult to identify them accurately. So we present
a new hybrid global optimization technique, where a gradient-based method with
BFGS update is combined with an intelligent algorithm called Artificial Bee Colony.
This new hybrid global optimization technique has both the fast convergence of
gradient descent algorithm and the global convergence of swarm algorithm. It is
used to identify Archie parameters in carbonate reservoirs. The results obtained
are highly satisfactory. To further test the effectiveness of the new hybrid global
optimization method, it is applied to ten non-convex benchmark problems. The
outcomes are encouraging.
7.1 Introduction
J. Liu () • G. Wu
College of Science, China University of Petroleum, Beijing 102249, China
e-mail: liujj@cup.edu.cn
H. Xu
Department of Mathematics and Statistics, Curtin University, Perth, WA 6845, Australia
K.L. Teo
School of Mathematics and Statistics, Curtin University, Perth, WA, 6845, Australia
Archie equation, which relates the resistivity of the formation to the porosity, the
water saturation and the formation water resistivity, is expressed as:
7 A New Hybrid Optimization Algorithm for the Estimation of Archie Parameters 139
1n
F Rw
Sw D (7.1)
Rt
CAPE, the error function, expressed as the difference between the computed and
the measured water saturations given by
X
M X
N 1n !2
aRwij
f .m; n; a/ D Swij (7.3)
jD1 iD1
ˆj m Rt ij
The methods to solve the global optimization problems (GOPs) can be classified into
two main classes: deterministic methods, and stochastic methods. The first class,
which makes use of the some deterministic information to solve GOPs, includes
the tunneling method (Levy and Gomez 1985), and filled function methods (Liang
et al. 2007; Liu and Xu 2004). The second class, which relies on probabilistic
techniques, includes Ant Colony Optimization algorithm (Toksari 2006), Genetic
Algorithm (Goldberg 1989; Michaelewicz 1996), Simulated Annealing algorithm
(Kirpatrick et al. 1983), and Particle Swarm Optimization (Eberhart and Kennedy
1995; Kennedy and Eberhart 1995; Shi and Eberhart 1998). Especially, Artificial
Bee Colony (ABC) holds better performance than most of the stochastic algorithms
mentioned above (Karaboga and Basturk 2007, 2008). The deterministic algorithms
which are gradient-based converge rapidly. However, they will get stuck in local
minima of a multimodal function. On the other hand, stochastic optimization algo-
rithms, which tend to perform global optimization, are computationally expensive
doing random searches. Therefore, an approach that combines thestrengths of
7 A New Hybrid Optimization Algorithm for the Estimation of Archie Parameters 141
5
6
4
4
3
2 2
1
0
3
2 3
1 1 2
n a
n=2
40
35
30 30
25
20
20
10 15
10
0
3 5
2
1
1 2
3
a m
a=1
20 14
12
15
10
10
8
5 6
4
0
3 2
2 3
2
m 1 1
n
142 J. Liu et al.
fiti
pi D XSN (7.5)
fiti
jD1
where fiti is the fitness value of the solution i and SN is the number of food sources.
When the nectar of the food source is abandoned by employed bees, the scout
bees replace it with a new one. In the ABC algorithm, if the quality of a solution
cannot be improved after a predetermined number of cycles called “limit”, the scout
bee replaces the abandoned solution with a new one chosen randomly. In such a
condition, the new solution is constructed according to the following equation
whereyjmin and yjmax are, respectively, the lower and upper bounds on the value of the
jth parameter.
In this paper, the search mechanism, proposed in (Karaboga and Basturk 2007),
is chosen to escape from a local solution in the new hybrid method. The search
144 J. Liu et al.
strategy is determined by the parameters, SN, the number of the food sources which
is equal to the number of employed or onlooker bees; limit, a predetermined number
of cycles; and the maximum cycle number, MCN. The ABC algorithm can be
implemented as follows:
Initiation. Set SN, limit, MCN, and cycle D 1. Generate an initial population of the
potential solutions Yi ; i D 1; 2; : : : ; SN, based on x*, and Evaluate f (Yi ).
repeat
1. Produce new solutions Vi for the employed bees by using Eq. (7.4)
2. Assume that the greedy selection process is adopted by the employed bees
3. Calculate the probability values pi for the solutions Yi by using Eq. (7.5)
4. With the probability values pi , produce the new solutions Vi for the onlookers
from the solutions Yi
5. Assume that the greedy selection process is adopted by the onlookers
6. Determine the abandoned solution for the scout, if exists, and replace it with
a new randomly produced solution Yi by using Eq. (7.6)
7. Memorize the best solution achieved so far
8. cycle D cycle C 1
until cycle D MCN
In the proposed hybrid algorithm, BFGS algorithm is executed from initial point
x0 to find a local minimum of f (x) with high-speed descent. Then the ABC algorithm
is used to escape from the local solution to find a better solution, y*, which will be
taken as the new starting point for the BFGS algorithm in the next cycle. A better
minimum x* found by the BFGS algorithm is taken as the best solution (food source)
in the ABC algorithm and is memorized. If the best solution y* found by the ABC
algorithm has a better value than the former memorized solution x*, then y* is used
as the starting point for the BFGS algorithm. This guarantees that a local search
operates in the neighborhood of the best solution found by the proposed algorithm
in all previous iterations.
Incidentally, the proposed algorithm can be generalized by using more powerful
local search techniques to refine the best solution found in the ABC. Derivative-free
techniques like Nelder–Mead Simplex method or Hooke-Jeeves (Fei Kang et al.
2011) method can be used for the local search when the objective function is not
continuously differentiable.
In this section, the focus is on the estimation of the Archie parameters for carbonate
reservoirs. The carbonate reservoirs are of great importance because they contain
almost 60 % of world’s oil reserves. The accuracy of Eq. (7.1) depends on the
accuracy of the estimates of the input parameters Rw , Rt , and F.
7 A New Hybrid Optimization Algorithm for the Estimation of Archie Parameters 145
Table 7.2 Error analysis on the determined Archie parameters by three techniques
Absolute error
Method Ea Emin Emax Erms S
3D method (Hamada et al. 2013) 0.102 0.002 0.51 0.14 0.10
CAPE (a, m, n) (Hamada et al. 2013) 0.095 0.001 0.33 0.12 0.08
BFGSABC 0.035 0.004 0.026 0.08 0.07
Note: Ea the average absolute relative error, Emin /Emax the minimum/maximum absolute error, S
the standard deviation. Erms the root mean square error
Hereinafter, 29 carbonate core samples taken from (Hamada et al. 2013) are
selected as the simulations of certain wells, that is N D 29. For each core sample, the
electrical resistivity Rw and Rt at different water saturation percentages are measured
at room temperature, that is M D 30. Set x D .m; n; a/, then BFGSABC algorithm
can be applied to solve the model indicated by Eq. (7.3).
Based on 30 independent core samples measurements, the data obtained by the
proposed hybrid method are compared with the data computed by 3D method and
CAPE.
Typical values of the Archie parameters obtained using the CAPE method,
the 3D method (Hamada et al. 2002, 2013) and the proposed hybrid method are
as shown in Table 7.1. Table 7.2 demonstrates the average error, the root mean
square error and standard deviation of the water saturation computed by three
techniques.
The average error, the root mean square error and standard deviation are shown
in Fig. 7.2. Figure 7.3 displays the linear regression estimation for three techniques.
Typical results of the measured water saturation and the estimated water saturation
profiles for different Archie parameters obtained by using CAPE, 3D and the
proposed hybrid method (BFGSABC) are illustrated in Fig. 7.4a. Figure 7.4b
depicts water saturation relative error profiles calculated by the three options against
selected interval for core samples.
From Figs. 7.2, 7.3, and 7.4, the measured water saturation and the estimated
water saturation profiles calculated by different methods are clearly demonstrated.
These profiles support the accuracy analysis in regards of the performance of
different techniques in order to get the most accurate Archie parameters. Note that
water saturation computed by the proposed technique has a better matching with the
measured water saturation than other two methods.
146 J. Liu et al.
0.035
Fig. 7.2 The average error, RMS error and standard deviation between the three techniques
0.9
BFGSABC
0.8 3D Method
CAPE (a,m,n) Method
0.7
0.6
computed Sw
0.5
0.4
0.3
0.2
0.1
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
measured Sw
a b
1 1
3 3
5 5
7 7
9 9
11 11
No. of core samples
19 19
21 21
23 23
25 25
27 27
29 29
0.2 0.4 0.6 0.8 1 1.2 0 50 100
Ws Relative Error
Fig. 7.4 (a) Comparison between the measured water saturation with the calculated water
saturation for three techniques, and (b) Relative error between three techniques
many local minima. The following are some of such multimodal functions: Rastrigin
function(f3), Ackley function(f4), Griewank function(f5), Schaffer function(f6) and
Levy function (f10 ).The ten test functions, their dimensions and modalities are listed
in details in the table of the Appendix.
In the experiments of BFGSABC on test problems, the number of maximum
generations are 50, 100 and 500 for the dimensions of 2, 10 and 50, and the
population sizes are 20, 20 and 200, respectively. Because BFGS is the deter-
ministic algorithm, only ABC and BFGSABC performed for 50 independent runs
on 10 functions. The calculation is done within the Matlab 7.70 environment.
The computer was functioned with double cores 2.5 GHz CPU PC running in
windows 7.
148 J. Liu et al.
Table 7.3 The global minima of test functions in 2D found by BFGS, ABC and BFGSABC
Best value Mean
Fun. Dim. BFGS ABC BFGSABC ABC BFGSABC
f1 2 1.256E-16 1.12e-12 5.59E-20 1.05e-09 2.29E-20
f2 4.30E-14 3.32E-04 1.94E-18 0.1437 1.32E-18
f3 9.949 8.84E-11 0.0 0.0017 0.0
f4 17.612 5.98E-07 8.88E-16 5.43E-05 3.23E-15
f5 97.449 1.05E-05 0.0 0.0101 0.0
f6 0.494 9.91E-05 0.0 0.0127 0.0
f7 0.0 0.9886 1.0 0.1835 1.0
f8 84.000 3.0000 2.99 3.5450 2.99
f9 0.397 0.3979 0.39 0.3983 0.39
f10 1.844E-15 1.35E-17 4.822E-20 2.47E-12 7.22E-20
Table 7.4 The global minima found by BFGS, ABC and BFGSABC for 10 dimensions
Best value Mean
Fun. Dim. BFGS ABC BFGSABC ABC BFGSABC
f1 10 1.12E-16 3.5834E-04 4.55E-17 0.0883 8.51E-17
f2 1010.67 1.0431 5.91E-11 21.9949 6.91E-11
f3 97.50 1.0878 1.13E-9 5.1409 5.12E-9
f4 19.20 0.2565 3.60E-5 1.9434 6.89E-5
f5 0.147 0.0436 2.65E-14 0.2768 4.46E-14
f10 7.85E-11 2.7850E-05 4.79E-14 0.0011 9.62E-14
Table 7.5 The global minima found by BFGS, ABC and BFGSABC for 50 dimensions
Best value Mean
Fun. Dim. BFGS ABC BFGSABC ABC BFGSABC
f1 50 7.94E-16 4.13E-06 3.58E-16 4.0423E-05 5.45E-16
f2 6.47 E C 6 52.3644 7.6 E-3 225.8868 8.6 E-3
f3 4.64 E C 2 28.6034 5.96E-1 40.8226 9.96E-1
f4 1.93 E C 2 1.8572 3.233E-8 3.3657 6.73E-7
f5 1.60E-2 0.0230 2.90E-13 0.4623 2.90E-12
f10 8.46E-7 5.92E-04 1.39E-11 0.0670 1.39E-11
All the ten functions considered are in 2 dimensions and some of these functions
are in higher dimensions. The results of experiments are listed in following tables.
Bold fonts in Tables 7.3, 7.4, and 7.5 indicate that the BFGS algorithm fails to solve
the problems because of being trapped into a local minimum.
After comparison of the data in Table 7.3, for 10 test functions of 2 dimensions, it
can be show that the global minimums of f3 f8 cannot be found by BFGS. Although
ABC finds the approx global minimums, their accuracy of minimal function values
and the mean value for all of 10 functions is lower than BFGSABC.
7 A New Hybrid Optimization Algorithm for the Estimation of Archie Parameters 149
a b 105
105
initial point
10-5 10-5
10-10 10-10
10-15 10-15
0 2 4 6 8 10 0 5 10 15
Iteration Iteration
c 102
d 100
initial point
Logarithm of function value
initial point
after BFGS search Logarithm of function value after BFGS search
After ABC search 10-5 After ABC search
0
10
10-10
10-2
10-15
10-4 10-20
1 2 3 4 5 6 7 1 2 3 4 5 6 7
Iteration Iteration
Fig. 7.5 Typical convergence history for the new algorithm for multimodal functions in 2
dimensions. (a) f3 Rastrigin function; (b) f4 Ackley function; (c) f5 Griewank function; (d) f6
Schaffer function
From Tables 7.4 and 7.5, it is observed that the hybrid method has better
performance than both BFGS and ABC in terms of the best values found for higher
dimensions.
Monotonic convergence, which is a very desirable property, is observed for
the proposed hybrid method. See, for example, the typical convergence histories
for the algorithm on the test functions f3 , f4 , f5 and f6 in 2 dimensions, which
are displayed in Fig. 7.5. Since the ABC method is mainly used for bypassing
the previously converged local minimum and discovering the descent point, the
decrease in function value after executing each ABC search might be small.
In addition, we study two multimodal functions f5 and f10 in 1,000 dimensions.
The maximum numbers of generations are 2,000 and the population size is 200.
These two functions have many local minima, which are regularly distributed.
Table 7.6 shows the best and mean values, CPU time and numbers of function
evaluation.
150 J. Liu et al.
Table 7.6 The optimal information on two 1,000 dimensions test functions by the proposed
method
Function Best value Mean CPU time (s) Number of function evaluation
f5 1.33E-8 7.90E-15 1,755 6.3E C 7
f10 1.01E-5 5.22E-2 2,713 6.6E C 7
From Tables 7.4, 7.5, and 7.6, it can be observed that the proposed hybrid method
can find the best “global” minima when compared with BFGS and ABC methods
available for the ten test functions. Furthermore, the success rate of finding the
“global” minima is 100 % for the new proposed hybrid method.
It can be concluded that the hybrid method proposed in this paper has better
performance in solving global optimization problems, especially in the rate of
convergence speed, reliability and the quality of the solution obtained.
7.6 Conclusions
For an accurate estimation of the Archie parameters, a new hybrid global search
method, which combines the well-known quasi-Newton algorithm (BFGS) and the
populated global search algorithm (ABC), is proposed. The ABC technique plays
the role of escaping from a local minimum to a better descent point from which
the local search can restart to find a better minimum. The hybrid method inherits
both the convergent rate and accuracy of the BFGS and the capability of escaping
from local minima of the ABC. Numerical results on ten benchmark problems have
shown that global minimum, especially for multimodal continuous functions, can
be sought using this hybrid descent method with very nice monotonic convergence
history. The results obtained from the simulation experiments on carbonate core
samples show that the proposed method has better performance than other methods
in terms of the accurate estimates of the Archie parameters. From the experiment
results, we observe that the water saturation computed by the proposed method
matches well with the measured water saturation when compared with the other
two methods.
Acknowledgements The authors gratefully acknowledge the financial support from the National
Natural Science Foundation of China (Grant No. 11371371, No. 11171079) and the Foundations
of China University of Petroleum (No. KYJJ2012-06-03, KYJJ2012-12).
Appendix: Expressions and Properties of Ten Test Problems
Optimal Optimal
Fun Expression Range solution value Modalities
Xn
f1 fSph D x2i 100 xi 100 .0; ; 0/ 0 Uni
iD1
n1 h
X i
2
f2 fRos D 100 xi1 x2i C .xi 1/2 30 xi 30 .0; ; 0/ 0 Uni
iD1
n
X
f3 fRas D 10n C x2i 10 cos .2xi / 5:12 xi 5:12 .0; ; 0/ 0 Multi
iD1 v
u n
u X n
X
u
1u1
5tn x2i 1n cos.2xi /
iD1 iD1
f4 fAck D 20 C e 20e e 32:768 xi 32:768 .0; ; 0/ 0 Multi
n
X n
Y
1 xi
f5 fGri D 4000 x2i cos p i
C1 600 xi 600 .0; ; 0/ 0 Multi
iD1 iD1
q
sin2 x21 C x22 0:5
f6 fSch D 0:5 C 2 100 xi 100 (0, 0) 0 Multi
1 C 0:001 x21 C x22
2
.x2 /2
f7 fEas D cos x1 cos x2 e.x1 / 100 xi 100 .; / 1 Uni
(continued)
7 A New Hybrid Optimization Algorithm for the Estimation of Archie Parameters
151
152
(continued)
Optimal Optimal
Fun Expression
h i Range solution value Modalities
fGP D 1 C .x1 C x2 C 1/2 19 14x1 C 3x21 14x2 C 6x1 x2 C 3x22
f8 h i 2 xi 2 .0; 1/ 3 Uni
30 C .2x1 3x2 /2 18 32x1 C 12x21 C 48x2 36x1 x2 C 3x22
2
f9 fBra D a x2 –bx21 C cx1 r C s .1 t/ cos .x1 / C s; with 5 x1 100 x2 15 .; 12:275/ ; 0.397887 Uni
5:1 5 1 (, 2.275),
a D 1; b D 4 2 ; c D ; r D 6; s D 10; t D 8
(9.42478, 2.475)
n1 h
X i
fLve D sin2 .y1 / C .yi 1/2 1 C 10sin2 .yi C 1/
f10 iD1
5 xi 10 (1, 1, : : : , 1) 0 Uni
2 2 xi 1
C.yn 1/ 1 C 10sin .2yn / ; where yi D 1 C 4
:
J. Liu et al.
7 A New Hybrid Optimization Algorithm for the Estimation of Archie Parameters 153
References
Andrei N (2008) An unconstrained optimization test functions collection. Adv Model Optim
10(1):147–161
Ara TS, Talabani S, Atlas B, Vaziri HH, Islam MR (2001) In-depth investigation of the validity of
the Archie Equation in carbonate rocks, SPE 67204, pp 1–10
Archie GE (1942) The electrical resistivity log as an aid in determining some reservoir character-
istics. Trans AIME 146:54–62
Chen DS et al (1995) Novel approaches to the determination of Archie parameters I: simplex
method. SPE Adv Technol Ser 3(1):39–43
Eberhart RC, Kennedy J (1995) A new optimizer using particle swarm theory. In: Proceedings of
the 6th symposium on micro machine and human science, Nagoya, Japan, pp 39–43
Enikanselu PA, Olaitan OO (2013) Determination of Archie parameters and the effect on water
saturation over “honey” field, Niger-delta. Can J Comput Math Nat Sci Eng Med 4(4):306–314
Fei Kang, Junjie Li, Zhenyue Ma, Haojin Li (2011) Artificial bee colony algorithm with local
search (HJ)for numerical optimization. J Softw 6(3):490–497
Garcia-Palomares UM, Gonzalez-Castan FJ, Burguillo-Rial JC (2006) A Combined Global &
Local Search (CGLS) approach to global optimization. J Glob Optim 34:409–426
Godarzi AA et al (2012) The simultaneous determination of Archie’s parameters by application
of modified genetic algorithm and HDP methods: a comparison with current methods via two
case studies. Pet Sci Technol 30(1):54–63
Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. Mas-
sachusetts: Addison-Wesley
Hamada GM, Al-Awad MNJ, Alsughayer AA (2002) Water saturation computation from labora-
tory, 3D Regression. Oil Gas Sci Technol Rev. IFP 57(6):637–651
Hamada GM, Almajed AA, Okasha TM et al (2013) Uncertainty analysis of Archie’s parameters
determination techniques in carbonate reservoirs. J Pet Explor Prod Technol 3(1):1–10
Karaboga D, Basturk B (2007) A powerful and efficient algorithm for numerical function
optimization: artificial bee colony (ABC) algorithm. J Glob Optim 39(3):459–471
Karaboga D, Basturk B (2008) On the performance of artificial bee colony (ABC) algorithm. Appl
Soft Comput 8(1):687–697
Kennedy J, Eberhart RC (1995) Particle swarm optimization. In: Process of IEEE international
conference on neural networks, Piscataway, pp 1942–1948
Kirpatrick S, Gelatt CD, Vecchi MP (1983) Optimization by simulated annealing. Science, New
Series 220(4598):671–680
Levy AV, Montalvo A (1985) The tunneling algorithm for the global minimization of functions.
SIAM J Sci Stat Comput 6(1):15–29
Liang YM, Zhang LS, Li MM, Han BS (2007) A filled function method for global optimization. J
Comput Appl Math 205:16–31
Liu X, Xu W (2004) A new filled function applied to global optimization. Comput Oper Res
31:61–80
Mabrouk WM, Soliman KS, Anas SS (2013) New method to calculate the formation water
resistivity (Rw). J Pet Sci Eng 10:49–52
Makar KH, Kamel MH (2012) An approach for velocity determination from merging Archie and
Raymer–Hunt–Gardner transform in reservoir of clean nature. J Pet Sci Eng 86–87:297–301
Maute RE, Lyle WD, Sprunt E (1992) Improved data-analysis method determines Archie
parameters from core data. J Pet Technol 44(1):103–107
Michael R, Collett TS et al (2013) Large-scale depositional characteristics of the Ulleung Basin and
its impact on electrical resistivity and Archie-parameters for gas hydrate saturation estimates.
Mar Pet Geol 47:222–235
Michaelewicz Z (1996) Genetic algorithms C Data structures D Evolution programs. Springer,
Berlin
154 J. Liu et al.
Momin Jamil, Xin-She Yang (2013) A literature survey of benchmark functions for global
optimization problems. Int J Math Model Numer Optim 4(2):150–194
Noel MM (2012) A new gradient based particle swarm optimization algorithm for accurate
computation of global minimum. Appl Soft Comput 12:353–359
Shi Y, Eberhart RC (1998) A modified particle swarm optimizer. In: Proceedings of the IEEE
Congress on Evolutionary Computation (CEC 1998), Piscataway, pp 69–73
Toksari MD (2006) Ant colony optimization for finding the global minimum. Appl Math Comput
176(1):308–316
Yiu KFC, Liu Y, Teo KL (2004) A hybrid descent method for global optimization. J Glob Optim
28:229–238
Chapter 8
Optimization of Multivariate Inverse Mixing
Problems with Application to Neural Metabolite
Analysis
8.1 Introduction
8.2 Methodology
AT AX D AT C (8.1)
AA C C D E (8.2)
ek;j
Then the entries of E% are given by e%k;j D 100 % ck;j
for metabolite j and
measurement k.
8 Optimization of Multivariate Inverse Mixing Problems with Application to. . . 159
Then we wish to minimize f .y/ D kE% .y/k, the L2 -norm of E% , subject to equality
and inequality constraints on A. Our optimization problem becomes:
hi D 0 (8.4)
gj 0 (8.5)
and this is subject to only equality constraints, which is an easier class of problem
to solve. Note that as approaches 0, the minimum in Eq. 8.6 should equal
the minimum in Eq. 8.3. Solving the approximated problem is done by taking a
sequence of steps using one of two methods. The first, and default, method linearizes
the Lagrangian of the approximated problem and attempts to find a solution that
satisfies the Karush-Kuhn-Tucker (KKT) conditions (Byrd et al. 2000). If this
160 A. Tamura-Sato et al.
Dy L .y ;
/ D 0 where
i 0 8i;
j gj .y / D0 (8.7)
and
zT r 2 L .y ;
/z > 0 8z 2 T 0 ; z ¤ 0
In this section we apply our methodology to the analysis of brain metabolite data.
Numerous studies have measured the concentrations of major brain metabolites
using 1H MRS, to evaluate brain biochemical changes and maturation in adults and
neonates (Kreis et al. 1993). However, the interpretation in terms of the health or
specific status of the brain cells is unclear since the data are obtained from a region
with a heterogeneous mixture of cells.
8 Optimization of Multivariate Inverse Mixing Problems with Application to. . . 161
Fig. 8.1 Age division for each subject visit. Box and whisker plots represent the minimum, second
quartile, median, third quartile, and maximum value of all subject data
Nine healthy newborns were studied, two boys and seven girls. Pregnant mothers
were recruited from the maternity ward at the Queen’s Medical Center in Honolulu
and through physician referrals. Each parent or legal guardian signed an informed
consent form approved by our Institutional Review Board, and completed detailed
interviews regarding their medical and drug use histories. Mothers were 18 years or
older at the time of giving birth, and had minimal or no drug use or any other prenatal
complications during pregnancy or perinatal problems during delivery. All babies
were at or near full term for gestational age (36 weeks or more), and were evaluated
thoroughly to ensure they were healthy. Each neonate was scanned three times:
within 1 week of birth, and at approximately 1 and 2 months thereafter. Although
many more neonates were studied, only nine infants with complete datasets for all
five metabolites and good quality data for both brain regions studied are presented.
The age distribution of these nine infants at each visit is shown in Fig. 8.1.
MRI studies were performed on a Siemens Trio 3.0 T scanner while the infants
slept (typically after nursing) and were unsedated. All babies had a sagittal
3D-magnetization prepared rapid acquisition by gradient echo (MP-RAGE)
sequence. Also, a T2-weighted 3D-SPACE sequence was acquired to ensure no
lesions were present. Based on anatomical landmarks in the MP-RAGE scan,
spectra were acquired in the right basal ganglia (BGR, 6.0 cm3 ) and frontal
white matter, right side (FWR, 5.0 cm3 ); see Fig. 8.2. A short echo-time Point
RESolved Spectroscopy (PRESS) acquisition sequence (relaxation time/echo
time = 3,000/30 ms, 2.5 min acquisition) was used (Chang et al. 1996), and
metabolite concentrations for five major metabolites were determined as described
162 A. Tamura-Sato et al.
Fig. 8.2 Neonate spectroscopic voxel locations shown on the MPRAGE images in all three
orientations (top left: coronal; bottom left: sagittal; top and bottom right: axial)
previously (Kreis et al. 1993, 2002). For each subject and each time point, a
complete MRS data set was available, which included the concentrations of:
• N-acetyl compounds (NAA): This metabolite is exclusively found in the nervous
system (peripheral and central) and is detected in both grey and white matter. It
is thought to be a marker of neuronal and axonal viability and density. Decreased
concentration of NAA is a sign of neuronal loss or degradation (Cheong et al.
2006; Urenjak et al. 1992, 1993).
• Total creatine (tCR): The role of tCR is to supply energy to all cells in the body,
including brain cells (Cheong et al. 2006).
• Choline-containing compounds (CHO): A marker of nerve signaling, myelin, and
cellular membrane turnover which can also reflect cellular proliferation. Reduced
CHO may be related with delayed myelination or apoptosis (Cheong et al. 2006).
8 Optimization of Multivariate Inverse Mixing Problems with Application to. . . 163
Table 8.1 and Fig. 8.3 show the mean values and standard deviation for the
metabolite concentrations in the basal ganglia, right side (BGR) and data for the
frontal white matter, right side (FWR) is also provided in Table 8.1. We note that
with the exception of [MI] in visits II and III, the metabolite concentrations are
higher in the BGR compared to the FWR. We can compare the measurements from
subjects in Visit I to those obtained in Kreis et al. (2002) for full term neonates
(38 weeks < GA < 43 weeks). In Kreis et al. (2002), all metabolite concentrations
in the ROI placed in the centrum semiovale for developing white matter (based on 11
subjects) were higher compared to our values in the frontal white matter (based on
nine subjects). Indeed, in their paper they obtained [NAA]:3:5˙0:5, [tCR]:4:8˙0:6,
[CHO]: 2:3 ˙ 0:1, [mI]:5:9 ˙ 0:7 and [GLX]:6:3 ˙ 1:1.
Table 8.1 Mean and standard deviation of metabolic concentrations in the two regions of interest
by visit
Visit I
Region [NAA] [tCR] [CHO] [MI] [GLX]
BGR 4:23 ˙ 0:36 5:17 ˙ 0:58 2:19 ˙ 0:13 4:80 ˙ 0:39 7:80 ˙ 1:10
FWR 2:85 ˙ 0:33 3:10 ˙ 0:24 1:83 ˙ 0:21 4:54 ˙ 0:49 5:90 ˙ 0:85
Visit II
Region [NAA] [tCR] [CHO] [MI] [GLX]
BGR 4:85 ˙ 0:19 5:59 ˙ 0:28 1:90 ˙ 0:18 4:00 ˙ 0:64 7:96 ˙ 1:15
FWR 3:71 ˙ 0:52 3:40 ˙ 0:52 1:62 ˙ 0:21 4:06 ˙ 0:71 5:95 ˙ 0:78
Visit III
Region [NAA] [tCR] [CHO] [MI] [GLX]
BGR 5:06 ˙ 0:51 5:62 ˙ 0:55 1:73 ˙ 0:22 3:13 ˙ 0:60 8:01 ˙ 1:03
FWR 4:70 ˙ 0:49 3:76 ˙ 0:58 1:53 ˙ 0:17 3:28 ˙ 0:70 7:00 ˙ 0:77
164 A. Tamura-Sato et al.
Fig. 8.3 Metabolite Concentrations in the BGR Region. Nine subjects were each sceanned three
times. Visit I was scanned between 39.6 and 41.1 weeks (postmenstrual age). Visit II was scanned
between 43.9 and 46.7 weeks. Visit III was scanned between 48.0 and 52.7 weeks
Graphs for each of the five metabolite concentrations over time for the two ROIs are
displayed in Fig. 8.4. The average rate of growth for each metabolite on the graph
is shown in Table 8.2. There is a greater increase over time in [NAA], [tCR], and
[GLX] levels in the FWR compared to the BGR, and a greater decrease in [CHO]
and [MI] levels in the BGR compared to the FWR. We note the smallest difference in
growth between the two regions is for [CHO], and the greatest difference in growth
rate is for [GLX].
A two-factor (age and region) repeated measures ANOVA analysis was con-
ducted on the metabolite data (Table 8.3). All metabolites except CHO had a
significant regional dependence, and all metabolites except GLX had significant age
dependence. Only NAA showed a significant interaction effect between region and
age.
8 Optimization of Multivariate Inverse Mixing Problems with Application to. . . 165
Fig. 8.4 Graphs for each of the five metabolite concentrations over time. Each of the five vertical
planes displays graphs of a specific metabolite concentration vs time for the two ROI. Since a
traditional line of best fit does not take into account the dynamics of the data within a subject
over time, we use a new fitting procedure. We calculate the slope between sequential data points
(subjects with three measurements in the same ROI have two pairs of sequential points). We set
an initial point using the average age and average metabolite concentration of Visit I scans. To
create the fit, we use a variation of Euler’s method: taking a small step along the x-axis (age), and
using the average of the calculated slopes at that age to find the change in the y-value (metabolite
concentration), then repeating the process to create a piecewise linear function that fits the data
Table 8.2 Metabolite concentration growth rates. Average rate of change for metabolite concen-
tration vs age in each region [mM/week]
NAA tCR CHO MI GLX
BGR 0:0701 0:0378 0:0403 0:1731 0:0116
FWR 0:1862 0:0715 0:0263 0:1389 0:1486
Since NAA and MI are theorized to be markers for neurons and glia (Brand
et al. 1993 and Guimaraes et al. 1995), this supports an approach with these two
populations of cells. We will call them type I and type II, respectively. Cells of type
I are characterized by a negligible concentration of NAA, a common assumption for
glia, and cells of type II are characterized by negligible MI concentration, a common
assumption for neurons.
We have a total of m D 27 measurements for the BGR and m D 27 measurements
for the FWR. In our model, each measurement is treated independently even though
it might represent a repeat measurement in a given subject. Based on our analysis
of metabolite concentrations in Sect. 8.4 and Table 8.3, we will consider region-
specific composition matrices. Thus, each region will have its own composition
matrix A. The methodology for setting up the model for each region is essentially
the same, however.
The matrix C is a 5 m matrix as follows. Each column of C corresponds
to one measurement, and therefore contains five values: one for each metabolite
concentration. We order the concentrations as follows: [NAA], [tCR], [CHO], [MI],
[GLX]; therefore, 11 reflects the NAA concentration per unit of density in type I
cells. Because of the exclusivity of the NAA and MI markers, we set 11 D 0 and
4
2 D 0, and can therefore remove them as variables.
Since n D 2, the composition matrix A is 2 5. Since negative values would
j
be unrealistic, we add the non-negative constraints that i 0 and xki 0 for
each i; j; k. We can compute explicitly the elements of the residual matrix E D
.ei;j . //1i5;1jm defined by Eq. (8.2)
5
" 5
! 5
!#
X cj;k X X
e . /D
i;j k
1
i
1 i
2
l l
1 2 C k
2 i
1
l l
1 2 C i
2 ci;j ;
kD1
det.AT A/ lD1 lD1
(8.8)
As in Eq. (8.3), our optimization problem becomes:
X
.e% .y//2
i;j
min f .y/; f .y/ D (8.9)
y
i;j
4
X
h1 .y/ D .yi /2 1 D 0 (8.10)
iD1
8
X
h2 .y/ D .yi /2 1 D 0 (8.11)
iD5
gi .y/ D yi 0; i D 1; 8 (8.12)
8 Optimization of Multivariate Inverse Mixing Problems with Application to. . . 167
5 5
!
1 X X
G1;j .y/ D c k;j a 1k a 2k a1m a2m 0 (8.13)
det .AT A/ kD1 mD1
5 5
!
1 X X
G2;j .y/ D c k;j a 2k a 1k a1m a2m 0 (8.14)
det .AT A/ kD1 mD1
where Eqs. (8.10)–(8.11) are the L2 -normalization for each column of A, Eq. (8.12)
is the non-negativity constraint on elements of A, and Eqs. (8.13)–(8.14) are the
non-negativity constraints on cell density.
We solve the above optimization problem numerically using MATLAB 2010b.
The results of our numerical calculations can be found in Table 8.4.
The Lagrangian multipliers were calculated numerically for all metabolite
spectra matrices in Table 8.4 as well as the Hessian of the Lagrangian using Matlab’s
fmincon solver. The Hessians were found to be strictly positive definite, therefore
satisfying the sufficient condition for local minima. We then checked for global
optima with MATLAB’s GlobalSearch. Across all regions and all starting points
chosen, the GlobalSearch solver reported the same minima and these minima were
consistent with the local minima found by the fmincon solver.
Inspection of optimized regional A values in Table 8.4 (see also Fig. 8.5) shows
that the metabolite concentration per unit of cellular density is similar for [CHO] in
both type I and type II cells for both regions examined. [tCR] is higher in the FWR
Fig. 8.5 Metabolite concentration per unit cell density for each region. The left graph shows the
concentrations per unit type I density, and the right graph shows the concentrations per unit type
II density. Note the NAA concentrations for type I cells and the MI concentrations for type II are
zero across both regions due to constraints
168 A. Tamura-Sato et al.
than the BGR in type I cells, but in type II cells [tCR] levels are higher in BGR than
FWR. The neuronal marker [NAA] has a higher value per unit of cellular density in
the frontal region compared to the deep grey matter, while the glial marker [MI] is
higher in the BGR compared to the FWR, suggesting a greater dependence on each
marker for their respective cell type. [GLX] has a substantially higher concentration
per unit density in the FWR compared to BGR for type I cells, whereas in type II
cells BGR has a slightly greater [GLX] value than FWR.
The matrix of relative cellular density X can now be computed from the equality
X D A C where A represents the Moore-Penrose pseudo-inverse regional matrices
(see Table 8.5). From this Table it is clear that for the cells of type I, almost all the
information is embedded in the MI concentration. For the cells of type II, however,
the interplay between the various metabolites is more pronounced but dominated by
the concentrations of NAA and GLX.
The results show a general decrease in density of type I cells over time, and an
increase in density of type II cells. The frontal region shows a greater increase in
density for type II cells than the basal ganglia, see Fig. 8.6 and Table 8.6. Repeated
Table 8.5 Pseudo-inverse regional matrices, where A represents the Moore-Penrose pseudo-
inverse matrix of matrix A
ABGR AFWR
-0.09012 0.03959 0.22237 0.98842 0.00283 -0.30849 0.12036 0.22813 1.00653 0.18156
0.48271 0.48135 0.05099 -0.18454 0.73127 0.74378 0.31288 -0.0053 -0.41747 0.61835
Fig. 8.6 Graphs of density levels for type I and type II vs age
Table 8.8 Mean and standard deviation of control residual values, in percent
Residual values region [NAA] [tCR] [CHO] [MI] [GLX]
BGR 0:11 ˙ 9:6 0:37 ˙ 7:8 1:1 ˙ 11 0:19 ˙ 1:6 1:0 ˙ 7:0
FWR 0:033 ˙ 7:7 0:096 ˙ 7:7 1:5 ˙ 13 0:37 ˙ 3:2 0:75 ˙ 5:4
measures 2-factor ANOVA on the density levels shows both populations have a
significant dependence on region and age, but only type II cells have a significant
interaction effect on region and age, see Table 8.7.
8.5.2 Residuals
Finally, to validate our model with the specific A and X matrices obtained, we
can recover C D AX and compare it with the measured data C. Recall that our
optimization process seeks to minimize the sum of the squares of the percentage
differences between our calculated C values and the metabolite data from the MRS
experiments C. Table 8.8 shows the mean and standard deviation of the residuals.
The measured data C are subject to error due to noise in the MRS experiments, so
we expect some error in our results. The standard deviations of residual values are
approximately 10 % or less, which is consistent with typical errors associated with
in vivo MRS metabolite levels.
When the methodology is applied to all of the data together, rather than separately
for each region, the residuals are larger. Thus, the region-specific A matrices give a
more accurate representation.
Conversely, type II cells are designed to have no MI, and were found to have similar
contributions of NAA, tCR, and GLX. Therefore, type II cells most likely represent
the neuronal population.
Importantly, over the age range evaluated, a simple 2-source model is able to
explain most of the variance in the metabolite data, with residuals that approach
typical errors associated with in vivo MRS measurements (approximately 10 %).
Furthermore, a 3-source model did not result in substantial improvements in fitting
accuracy, suggesting that a glial and neuronal compartment are sufficient for
representing the measured 1H MRS data within experimental errors.
Assuming the type I compartment represents glial cells and the type II com-
partment represents the neuronal compartment, our findings are in agreement with
those of prior studies. Specifically, both regions examined showed a pronounced
increase in the neuronal compartment with age, probably representing neuronal
maturation during the first few months of life. Of note, the neuronal compartment
of the frontal white matter increased at over three times the rate compared to that
of the basal ganglia. This suggests that the basal ganglia are more mature at birth
compared to white matter, in agreement with the literature (Kostović and Jovanov-
Milošević 2006). In parallel, there is a substantial decrease in the glial (type I)
compartment over the first few months of life, perhaps due to replacement of the
glial cell compartment with maturing neurons or osmotic effects.
Table 8.9 Coefficients of first and second principal components in BGR and FWR
BGR FWR
Metabolite Comp. 1 Comp. 2 Metabolite Comp. 1 Comp. 2
NAA 0.18166 0.33287 NAA 0.62662 0.03408
Cr 0.2051 0.16135 Cr 0.30277 0.28121
CHO 0.01854 0.15223 CHO 0.02391 0.14161
MI 0.27759 0.84329 MI 0.37771 0.84766
GLX 0.92062 0.35897 GLX 0.61028 0.42566
8 Optimization of Multivariate Inverse Mixing Problems with Application to. . . 171
1 1
MI MI
0.5 0.5
GLX GLX
Cr
Component 2
Component 2
CHO CHO
0 0 NAA
Cr
NAA
-0.5 -0.5
-1 -1
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Component 1 Component 1
BGR FWR
Fig. 8.7 Reduction of system to two PCA Components for BGR and FWR. Contribution of five
measured metabolites to each component shown as vectors
Table 8.10 Calculated values for A without non-negativity constraints and zero constraints
BGR FWR
Metabolite Type I Type II Metabolite Type I Type II
NAA 0:02263 0:46374 NAA 0:16187 0:61143
Cr 0:15187 0:48950 Cr 0:17478 0:36445
CHO 0:23401 0:09748 CHO 0:22052 0:09245
MI 0:94423 0:02221 MI 0:90005 0:01410
GLX 0:17348 0:73167 GLX 0:29074 0:69612
sources (matrix A in the model), since negative composition values are unrealistic.
For instance, the potential glial principal component of the BGR region, having a
high loading of MI (Table 8.9), would also have a substantial negative contribution
.0:33/ of NAA, which is physically impossible in a mixing situation. Conversely,
the calculated A matrices from our methodology produce results that are non-
negative, and the residual values using the model results are smaller than the PCA
results.
The principal components in PCA are calculated to optimally explain variance in
the data. Our calculated A matrices are chosen to minimize residual values, but still
explain most of the variance in the data. In the BGR, type I and type II compositions
(the first and second columns of A) explain 30:1 % and 37:6 % of the variance,
respectively. In the FWR, type I and type II compositions explain 20:4 % and 60:4 %
of the variance, respectively.
If the non-negative constraints and zero constraints on NAA and MI are removed
from our analysis, we obtain the solutions in Table 8.10. Comparing the first and
second PCA components in Table 8.9 to the Type II and Type I results, respectively,
for the model with reduced constraints, the FWR results are similar, while the BGR
results are less comparable.
172 A. Tamura-Sato et al.
8.6 Conclusion
Acknowledgements We are grateful to the research participants in this study. We also thank all
of our clinical and technical research staff who helped with the data collection (S. Buchthal, A.
Hernandez, E. Cunningham, H. Johansen, J. Skranes, R. Yamakawa).
Funding: This work was supported by the National Institute on Drug Abuse (K24-DA016170;
K02-DA016991; 1R01 DA021146), the National Institute on Minority Health and Health Dis-
parities (8G12-MD007601-27), the National Institute of Neurological Diseases and Stroke (U54-
NS056883) and the Office of National Drug Control Policy. M. Chyba and A. Tamura-Sato were
partially supported by the National Science Foundation (NSF) Division of Graduate Education,
award #0841223, and the NSF Division of Mathematical Sciences, award #1109937.
8 Optimization of Multivariate Inverse Mixing Problems with Application to. . . 173
References
Ancora G, Soffritti S, Lodi R, Tonon C, Grandi S, Locatelli C, Nardi L, Bisacchi N, Testa C, Tani
G, Ambrosetto P, Faldella G (2010) A combined a-EEG and MR spectroscopy study in term
newborns with hypoxic-ischemic encephalopathy. Brain Dev 32:835–842
Befroy DE, Shulman GI (2011) Magnetic resonance spectroscopy studies of human metabolism.
Diabetes 60(5):1361–1369
Blüml S, Wisnowski JL, Nelson Jr. MD, Paquette L, Gilles FH, Kinney HC, Panigrahy A (2013)
Metabolic Maturation of the Human Brain from Birth Through Adolescece: Insights from In
Vivo Magnetic Resonance Spectroscopy. Cerebral Cortex 23(12):2944–2955
Brand A, Richter-Landsberg C, Leibfritz D (1993) Multinuclear NMR studies on the energy
metabolism of glial and neuronal cells. Dev Neurosci 15:289–298
Byrd RH, Hribar M, Nocedal J (1999) An interior point algorithm for large-scale nonlinear
programming. SIAM J Optim 9(4):877–900
Byrd RH, Gilbert JC, Nocedal J (2000) A trust region method based on interior point techniques
for nonlinear programming. Math Program 89(1):149–185
Cantrell, CA (2008) Technical note: review of methods for linear least-squares fitting of data and
application to atmospheric chemistry problems. Atmos Chem Phys 8:5477–5487
Chang L, Ernst T, Poland RE, Jenden DJ (1996) In vivo proton magnetic resonance spectroscopy
of the normal aging human brain. Life Sci 58(22):2049–2056
Cheong JLY, Cady EB, Penrice J, Wyatt JS, Cox IJ, Robertson NJ (2006) Proton MR spectroscopy
in neonates with perinatal cerebral hypoxic-ischemic injury: metabolite peak-area ratios,
relaxation times, and absolute concentrations. Am J Neuroradiol 27:1546–1554
Christopherson N, Hooper RP (1992) Multivariate analysis of stream water chemical data: the
use of principal components analysis for the end-member mixing problem. Water Resour Res
28(1):99–107
Christopherson N, Neal C, Hooper RP, Vogt RD, Andersen S (1990) Modelling streamwater
chemistry as a mixture of soilwater end-members: a step toward second-generation acidification
models. J Hydrol 116:307–320
Constrained Nonlinear Optimization Algorithms. The MathWorks Inc. Web. Accessed 6
Aug. 2014. http://www.mathworks.com/help/optim/ug/constrained-nonlinear-optimization-
algorithms.html
Guimaraes AR, Schwartz P, Prakash MR, Carr CA, Berger UV, Jenkins BG, Coyle JT, González
RG (1995) Quantitative in vivo 1H nuclear magnetic resonance spectroscopic imaging of
neuronal loss in Rat brain. Neuroscience 69(4):1095–1101
Haseler LJ, Arcinue E, Danielsen ER, Bluml S, Ross BD (1997) Evidence from proton magnetic
resonance spectroscopy for a metabolic cascade of neuronal damage in shaken baby syndrome.
Pediatrics 99(1):4–14
How GlobalSearch and MultiStart Work. The MathWorks Inc. Web. Accessed 6 Aug. 2014. http://
www.mathworks.com/help/gads/how-globalsearch-and-multistart-work.html#bsc9eec
Hunter MA, Takane Y (2002) Constrained principal component analysis: various applications. J
Educ Behav Stat 27(2):105–145
Kirov I, Flaysher L, Fleysher R, Patil V, Liu S, Gonen O (2008) The age dependence of regional
proton metabolites T2 relaxation times in the human brain at 3 Tesla. Magn Reson Med
60(4):790–795
Kostović I, Jovanov-Milošević N (2006) The development of cerebral connections during the first
20–45 weeks’ gestation. Assess Brain Funct Perinat Period 11(6):415–422
Kreis R, Ernst T, Ross BD (1993) Development of the human brain: in vivo quantification of
metabolite and water content with proton magnetic resonance spectroscopy. Magn Reson Med
30(4):424–437
Kreis R, Hofmann L, Kuhlmann B, Boesch C, Bossi E, Hüppi PS (2002) Brain metabolite
composition during early human brain development as measured by quantitative in vivo 1H
magnetic resonance spectroscopy. Magn Reson Med 48(6):949–958
174 A. Tamura-Sato et al.
Manganas L, Zhang X, Li Y, Hazel RD, Smith SD, Wagshul ME, Henn F, Benveniste H, Djurić PM,
Enikolopov G, Maletić-Svatić M (2007) Magnetic resonance spectroscopy identifies neural
progenitor cells in the live human brain. Science 318(5852):980–985
Mangia S, Giove F, DiNuzzo M (2012) Metabolic pathways and activity-dependent modulation of
glutamate concentration in the human brain. Neurochem Res 37(11):2544–2561
Pouwels P, Brockmann K, Kruse B, Wilken B, Wick M, Hanefeld F, Frahm J (1999) Regional age
dependence of human brain metabolites from infancy to adulthood as detected by quantitative
localized proton MRS. Pediatr Res 46:474–474
Snieder R, Trampert J (1999) Inverse problems in geophysics. In: Wirgin A (ed) Wavefield inver-
sion. International centre for mechanical sciences, vol 398. Springer, New York, pp 119–190
Takane Y, Hunter MA (2001) Constrained principal component analysis: a comprehensive theory.
Appl Algebra Eng Commun Comput 12(5):391–419
Urenjak J, Williams SR, Gadian DG, Noble M (1992) Specific expression of N-acetylaspartate
in neurons, oligodendrocyte-type-2 astrocyte progenitors, and immature oligodendrocytes in
vitro. J Neurochem 59(1):55–61
Urenjak J, Williams SR, Gadian DG, Noble M (1993) Proton nuclear magnetic resonance
spectroscopy unambiguously identifies different neural cell types. J Neurosci 13(3):981–989
van Doormaal P, Meiners L, ter Horst H, van der Veere C, Sijens P (2012) The prognostic value
of multivoxel magnetic resonance spectroscopy determined metabolite levels in white and grey
matter brain tissue for adverse outcome in term newborns following perinatal asphyxia. Eur J
Radiol 22(4):772–778
Waltz RA, Morales JL, Nocedal J, Orban D (2006) An interior algorithm for nonlinear optimization
that combines line search and trust region steps. Math Program 107(3):391–408
Wang ZJ, Vigneron DB, Miller SP, Mukherjee P, Charlton NN, Lu Y, Barkovich AJ (2008) Brain
metabolite levels assessed by lactate-edited MR spectroscopy in premature neonateswith and
without pentobarbital sedation. Am J Neuroradiology 29:798–801
Zou H, Hastie T, Tibshirani R (2006) Sparse principal component analysis. J Comput Graph Stat
15(2):265–286
Chapter 9
Exact Regularization, and Its Connections to
Normal Cone Identity and Weak Sharp Minima
in Nonlinear Programming
S. Deng
9.1 Introduction
Ax D b;
S. Deng ()
Department of Mathematical Sciences, Northern Illinois University, DeKalb, IL, USA
e-mail: deng@math.niu.edu.
y D Ax0 C w;
The role of problem (9.3) is to analyze problem (9.2) through certain regulariza-
tion process (see Definition 9.1 for more details).
The technique of regularization is a common approach used to solve an ill-
posed nonlinear optimization problem with non-unique solutions by constructing
a related problem whose solution is well behaved and deviated only slightly from a
9 Exact Regularization, and Its Connections to Normal Cone Identity and. . . 177
solution of the original problem. Deviations from solutions of the original problem
are generally accepted as a trade-off for obtaining solutions with other desirable
properties. However, it would be more desirable if solutions of the regularized
problem were also solutions of the original problem. In a recent paper by Friedlander
and Tseng (2007), the authors presented a systematic study for exact regularization
of a convex program. The term exact regularization was coined in Friedlander and
Tseng (2007); according to Friedlander and Tseng (2007) the regularization is exact
if the solutions of perturbed problems are also solutions of the original problem
for all values of penalty parameters below some positive threshold value. In Deng
(2012), we demonstrate that the main results of Friedlander and Tseng (2007) can
be extended to non-convex programs thereby the application domain of this exact
regularization technique has been significantly expanded. In this paper, for convex
programs, we examine inner-connections among exact regularization, normal cone
identity, and the existence of a weak sharp minimum for certain associated nonlinear
programs. Specifically, we show that strongly exact regularization is equivalent to
normal cone identity, and weak sharp minima implies normal cone identity. Along
the way, we illustrate by examples, how to obtain both new results and reproduce
many existing results from a fresh perspective.
The notation used in this note is standard. See e.g. Rockafellar and Wets (1998).
where p denotes the optimal value of (P). Let SQ be the set of optimal solutions of
(Q), and suppose that SQ 6D ;. Let the Lagrangian function of (Q) be
Problem (Q) may not have a Lagrange multiplier even for the convex case as
illustrated by the following example.
Example 9.3. Let g.x/ D x2 , C D R, and f .x/ D x. Then S D argminx2C g D f0g.
For L.x; y/ D x C yx2 , there are no saddle points for L over R RC . In fact, for
y 0,
(
1 if y D 0;
inf L.x; y/ D 1
x2C 4y if y > 0:
If (Q) is a convex program, then the existence of Lagrange multiplies for (Q) is
equivalent to the existence of a saddle point for L, and the set of Lagrange multiplies
is the same for any solutions of (Q). Hence Theorem 9.1 generalizes one of main
results in Friedlander and Tseng (2007).
The following theorem generalizes the main results (Theorem 2.1) of Friedlander
and Tseng (2007) on exact regularization to the non-convex case.
Theorem 9.2 (Theorem 6 of Deng (2012)). Consider problems (P), (Q), and
(P.ı/). Then the following statements are true.
(a) For any ı > 0, S \ Sı SQ .
(b) If there exists a saddle point .Nx; yN / of L for (Q) with xN 2 SQ , then S \ Sı D SQ
for all ı 2 .0; 1=Ny. Here we use the convention 1=Ny D C1 when yN D 0.
(c) If there exists ıN > 0 such that S \ SıN 6D ;, then .Nx; 1=ı/ N is a saddle point of L
for (Q) with any xN 2 S \ Sı D SQ for all ı 2 .0; ı. N
(d) If there exists ıN > 0 such that S \ SıN 6D ;, then Sı S for all ı 2 .0; ı/.
N
In this section, we assume that C is a closed convex set, and f ; g are finite convex
functions. We begin with two definitions.
Definition 9.2. We say that strongly exact regularization holds for (P) if for any
f with SQ 6D ;, (P) is exactly regularized with respect to f in the sense of
Definition 9.1.
Definition 9.3. We say that the normal cone identity holds for (P) if, for each x 2 S,
where @g.x/ is the subdifferential of g at x, and NS .x/; NC .x/ are normal cones of S
and C at x respectively.
The equivalence of strongly exact regularization and normal cone identity
follows.
Theorem 9.4. For (P), strongly exact regularization holds if and only if the normal
cone identity holds for (P).
Proof. ())Let xN 2 S be given. For any v 2 NS .Nx/, let f .x/ D 1=2jjx .v C xN /jj2 :
Then xN is the unique minimizer for f over S, and rf .Nx/ D v. Since (P) is exactly
regularized with respect to f , by Theorem 9.2, there is some yN 0 such that .Nx; yN / is
a saddle point of L. So
Hence,
that is, v 2 NC .Nx/ C RC @g.Nx/, which implies that NS .Nx/ NC .Nx/ C RC @g.Nx/. On
the other hand, as S D fx jg.x/ p g \ C, one always has
Thus, the inclusion NS .Nx/ NC .Nx/ C RC @g.Nx/ always holds. Therefore, the normal
cone identity holds.
(() Let f be a finite convex function. Suppose that SQ 6D ;. Then for xN 2 S,
0 2 @f .Nx/ C NS .Nx/:
This shows that .Nx; yN / is a saddle point of L. By Theorem 9.2, exact regularization
holds for (P) with respect to f . Since f is any finite convex function, strongly exact
regularization holds for (P). This completes the proof.
For (P), the normal cone identity property is closely related to the notion
of weak sharp minima, which has found a number of important applications in
mathematical programming. See Burke and Ferris (1993) and Burke and Deng
(2002) and references therein. Recall Burke and Ferris (1993) and Burke and Deng
(2002) that S is said to be a set of weak sharp minima for g over the set C with
modulus ˛ > 0 if
˛dist.x; S/ g.x/ p 8x 2 C;
where ıC is the indicator function of the set C. By Part 2 of Theorem 2.3 in Burke
and Deng (2002), for any x 2 S,
This shows that NS .x/ D RC @g.x/ C NC .x/; that is, the normal cone identity holds.
This completes the proof.
We conclude the paper with an example which shows that the normal cone
identity holding for (P) does not imply that S is a set of weak sharp minima for
g over C in general.
Example 9.5. For (P), let g W R3 ! R be given by g.x/ D x3 , and C D
Œ0; 13 \ .\1 3 2
kD2 C /; where C D fx 2 R j x1 .k 1/x2 k x3 1=kg: Then the
k k
optimal value p D 0 and S D fx 2 C j x3 D 0g. It was shown in the Appendix of
Friedlander and Tseng (2007) that S is not a set of weak sharp minima for g over C,
and that for any given z 2 Rn and f .x/ D 1=2jjx zjj2, exact regularization holds for
182 S. Deng
(P) with respect to f . The proof of necessary part of Theorem 9.4 shows that this
implies the normal cone identity holds; that is, NS .x/ D NC .x/ C RC @g.x/ for each
x 2 S.
Acknowledgements We wish to thank the referees for their useful comments, which helped us
improve the presentation of the paper.
References
Burke JV, Ferris M (1993) Weak sharp minima in mathematical programming. SIAM J Control
Optim 31:1340–1359
Burke JV, Deng S (2002) Weak sharp minima revisited, part I: basic theory. Control Cybern
31:439–469
Deng S (2012) A saddle point characterization of exact regularization of non-convex programs.
Pac J Optim 8(1):27–32
Friedlander M, Tseng P (2007) Exact regularization of convex programs. SIAM J Optim
18:1326–1350
Rockafellar RT (1993) Lagrange multipliers and optimality. SIAM Rev 35:183–238
Rockafellar RT, Wets R (1998) Variational analysis. Springer, Berlin/Heidelberg
Chapter 10
The Worst-Case DFT Filter Bank Design with
Subchannel Variations
Lin Jiang, Changzhi Wu, Xiangyu Wang, and Kok Lay Teo
10.1 Introduction
Filter banks play an important role in a wide range of signal processing applications
such as echo cancellation (Kellermann 1988), microphone arrays (de Haan et al.
2003), speech enhancement and equalization (Vaidyanathan 1993), as well as image
and speech processing. Owing to their wide range of applications, they have been
extensively studied in the past two decades (Dam et al. 2005; de Haan et al. 2001,
2003; Harteneck et al. 1999; Kellermann 1988; Kha et al. 2009; Mansour 2007;
Nguyen 1994; Sturm 1999; Vaidyanathan 1993; Wilbur et al. 2004; Wu and Teo
2010, 2011; Wu et al. 2008, 2013; Yiu et al. 2004; Zhang et al. 2008). In multirate
digital signal processing, an analysis filter is used to divide the signal to be processed
into subbands. They are then decimated according to the new bandwidth of the
subbands. The decimation process causes aliasing of the subband signals. It is
L. Jiang ()
School of Mathematics, Anhui Normal University, Wuhu, 241000, China
C. Wu • X. Wang
Australasian Joint Research Centre for Building Information Modelling, School of Built
Environment, Curtin University, Perth, WA 6845, Australia
K.L. Teo
School of Mathematics and Statistics, Curtin University, Perth, WA, 6845, Australia
possible to cancel this aliasing through the design of a synthesis filter bank in such
a way that the whole multirate chain yields no distortion; the total transfer function
is reduced to a simple delay. This is often referred to as the perfect reconstruction
(PR) property (Harteneck et al. 1999).
However, any filtering operation in the subbands will cause phase and amplitude
changes, thereby altering this property. Thus, aliasing may be caused in the
reconstructed output of the subband adaptive filter. To overcome this problem,
optimization methods are often used in the design of filter banks, where both
the aliasing effect and the distortion levels in the filter bank are optimized. In
de Haan et al. (2003), the design of a uniform discrete Fourier transform (DFT)
filter bank is solved by a two-step optimization problem. In the first step, the
analysis filter bank is designed in such way that the aliasing terms in each subband
are minimized individually, contributing to minimal aliasing at the output without
aliasing cancellation. In the second step, the synthesis filter bank is designed to
match the analysis filter bank where the analysis-synthesis response is optimized
while all aliasing terms in the output signal are individually suppressed, rather than
aiming at aliasing cancellation. Since the analysis and synthesis filter banks are
designed separately, this method does not produce a good result. To improve this
method, a bi-iterative method is used in Dam et al. (2005), where the design of
this filter bank is formulated as a constrained fourth order polynomial optimization
problem with continuous constraints. This optimization problem is hard to solve.
Thus, a bi-iterative computational scheme is incorporated to solve the formulated
optimization problem. More specifically, the analysis filter bank is fixed when
solving the synthesis filter bank, while the synthesis filter bank is fixed when
solving the analysis filter bank. In this way, only quadratic optimization problem
is required to be solved during the design process of this filter bank. Since the
original optimization problem is a fourth order polynomial optimization problem,
this bi-iterative scheme does not provide a global optimal solution. To overcome
this problem, a global optimization method based on the filled function method is
introduced to solve the corresponding optimization problem in Wu et al. (2008).
By using the global method, better results are obtained. In Yiu et al. (2004), the
design problem has been formulated as a multicriteria optimization problem. Then, a
nonlinear programming methods can be used to solve such an optimization problem.
In Wilbur et al. (2004), the design of a generalized DFT filter bank is formulated as a
cone program with a combination of linear, second-order, and semi-definite cones. It
is solved by using an existing convex optimization software package (Sturm 1999).
Although the DFT filter bank has been extensively studied, the actual filtering
operation in each frequency band has only been taken into limited consideration.
In this paper, we propose a new formulation which includes a filtering operation in
each subband in addition to the optimization and control of each individual criterion.
The aliasing effects for the filter bank are minimized subject to the constraints on
the distortion level for all the frequencies. The formulation is such that it includes a
term measuring the deviation from the nominal value in each subband. Comparing
with the earlier formulations, our formulation is more ideal for use in adaptive filters
or speech enhancement applications since the filter operation with distortion in each
10 The Worst-Case DFT Filter Bank Design with Subchannel Variations 185
subband has been taken into consideration. In this problem formulation, the study
is to provide simultaneous optimization on both the analysis and synthesis filter
banks, while maintaining robustness against deviation from unity in each subband.
The advantage of this problem formulation is that it provides less overall distortion
when the subchannels are subject to distortions. This filter bank design problem is
formulated as a constrained minimax optimization problem.
The second contribution of this paper is that we proved that the formu-
lated constrained minimax optimization problem was equivalent to a semi-infinite
optimization problem where the continuous constraints are only with respect to
frequency. Although a minimax optimization problem is actually a semi-infinite
optimization problem, we cannot use available methods for solving semi-infinite
optimization to solve it directly. This is due to the fact that the semi-infinite
optimization problem is known to suffer from the curse of dimensionality (Lopez
and Still 2007). In this paper, we will show that this minimax optimization problem
is equivalent to a standard semi-infinite optimization problem with non-smooth
constraints. Then, an iterative computational scheme is developed to solve this
optimization problem. Some simulation examples are presented to illustrate the
method proposed. Simulation results show that, for a fixed distortion level, the
aliasing between different subbands is significantly reduced, in some cases up to
28 dB, when compared with those obtained by using the bi-iterative optimization
method developed in Dam et al. (2005) which is not taken into the consideration of
the subband variations.
Uniformly modulated filter banks are employed where the filter banks are formed
by modulated versions of the analysis and synthesis prototype filters. Denote h D
Œh.0/; ; h.La 1/T as the prototype filter of length La for the analysis filter
bank with the corresponding transfer function H.z/ D hT a .z/, where a .z/ D
Œ1; z1 ; ; z.La 1/ . Similarly, denote g D Œg.0/; ; g.Ls 1/T as the prototype
filter of length Ls for the synthesis filter bank with the transfer function G.z/ D
gT s .z/, where s .z/ D Œ1; z1 ; ; z.Ls 1/ . For a system with M subbands, the
subband filters Hm .z/ and Gm .z/, 0 m M 1, are obtained from the prototype
filters H.z/ and G.z/, respectively, as follows:
Hm .z/ D H.zWM
m
/ and Gm .z/ D G.zWM
m
/ (10.1)
1X
D1
Xm .z/ D H.z1=D WM WD /X.z1=D WDd /
m d
(10.2)
D dD0
186 L. Jiang et al.
X (z) X0 ( z) Y0 ( z )
H0 ( z ) ↓D ↑D G0 ( z )
X1 ( z ) Y1 ( z )
H1 ( z ) ↓D ↑D G1 ( z ) +
Y (z)
X M −1 ( z ) YM −1 ( z ) G ( z)
HM−1 ( z) ↓D ↑D M−1 +
In the synthesis filter bank, the subband signals Ym .z/ are interpolated by a factor D
and then added to form the output signal Y.z/ given by
1X X
D1 M1
Y.z/ D X.zWDd / m .zD /H.zWM WD /G.zWM
m d m
/: (10.4)
D dD0 mD0
P
M1
The term m .zD /H.zWM WD /G.zWM
m d m
/ can be viewed as the transfer function
mD0
which contributes to the aliasing terms in the output signal for 1 d D 1 and to
the desired output signal for d D 0. Normally, the analysis and synthesis filter banks
are designed for the case m .zD / D 1 as filtering operation is unknown. However,
this performs poorly in real situations when m .zD / can take on arbitrary values.
Thus, the optimization should be performed subject to an allowable variation for
m .zD / D 1 C ım ; where ım is a random variation. Thus the design of this filter bank
should include this variation constraint. Let ı D Œı0 ; ı1 ; ; ıM1 T : We suppose
that the random vector ı is restricted to the following box constrained set
˚
U D ı D Œı0 ; ı1 ; ; ıM1 T 2 RM W jıi j "i ;
The objective is to optimize the analysis and synthesis prototype filters with respect
to both the aliasing power and the distortion for the filter bank. The aliasing power
for all the aliasing terms in (10.4) for a frequency ! 2 Œ; is given by
10 The Worst-Case DFT Filter Bank Design with Subchannel Variations 187
1 XX
D1 M1
m 2
A.!/ D j m .zD /H.ej! WM WD /G.ej! WM
m d
/j : (10.5)
D dD1 mD0
This can be rewritten in terms of the analysis and synthesis prototype FIR filter
coefficients as follows:
1 XX
D1 M1
A.!/ D j m .zD /hT ˆm;d .ej! /gj2 (10.6)
D dD1 mD0
where
Thus, the total aliasing power for all the frequencies ! 2 Œ; is defined as
Z
1
.h; g; ı/ D A.!/d!
2
Z X M1
D1 X
1
D j m .zD /hT ˆm;d .ej! /gj2 d!
2D dD1 mD0
Z X
M1 X
D1
1
D j .1 C ım / hT ˆm;d .ej! /gj2 d!; (10.8)
2D mD0 dD1
1 X
M1
T.z/ D m .zD /H.zWM
m
/G.zWM
m
/ D hT ‰.z; ı/g; (10.9)
D mD0
where
1 X
M1
‰.z; ı/ D .1 C ım /a .zWM
m
/sT .zWM
m
/: (10.10)
D mD0
Since the analysis and synthesis prototype filters are to be designed subject to an
allowable small distortion from a desired response, the worst case scenario can be
formulated using the mini-max constraint on the total response of the filter bank
below,
where Re fzg and Im fzg denote the real part and the imaginary part of z; respectively.
Now the optimal design for the worst-case scenario of this filter bank can be posed
as the following optimization problem:
Problem 10.1.
As a result of the presence of the max operator, the cost function (10.16) is non-
ı2U
smooth. Thus, Problem 10.1 cannot be solved by the gradient-based optimization
methods. By introducing a new variable ; Problem 10.1 can be reformulated as the
following semi-infinite optimization problem:
Problem 10.2.
min
h;g;
.h; g; ı/ ; 8ı 2 U : (10.17)
10 The Worst-Case DFT Filter Bank Design with Subchannel Variations 189
Clearly, D has 2M distinct extreme points. Let these 2M distinct points be denoted as
ıN 1 ; ; ıN 2M : For the set U , we have the following theorems:
Theorem 10.1. U is a convex set and ıN 1 ; ; ıN 2M are extreme points of U :
Furthermore, U D co ıN 1 ; ; ıN 2M :
Proof. See Appendix 2.
Theorem 10.2. For each given h and g; the maximum of .h; g; ı/ in U is attained
at one of the 2M extreme points ıN 1 ; ; ıN 2M :
Proof. See Appendix 2.
By Theorem 10.2, it is clear that the constraint
.h; g; ı/ ; 8ı 2 U ;
190 L. Jiang et al.
.h; g; ıN i / ; i D 1; ; 2M : (10.19)
and
˚
i;m .!; h; g/ D Im hT a .ej! WM
m
/sT .ej! WM
m
/g :
1 X
M1
D . r;m .!; h; g/ C "m j r;m .!; h; g/j/ ; (10.20)
D mD0
˚
min hT Re .ej! ; ı/ g
ı2U
1 X
M1
D . r;m .!; h; g/ "m j r;m .!; h; g/j/ ; (10.21)
D mD0
˚
max hT Im .ej! ; ı/ g
ı2U
1 X
M1
D . i;m .!; h; g/ C "m j i;m .!; h; g/j/ ; (10.22)
D mD0
and
˚
min hT Im .ej! ; ı/ g
ı2U
1 X
M1
D . i;m .!; h; g/ "m j i;m .!; h; g/j/ : (10.23)
D mD0
min (10.24)
h;g;
subject to
.h; g; ıN i / ; i D 1; ; 2M ; (10.25)
1 X
M1
. r;m .!; h; g/ C "m j r;m .!; h; g/j/
D mD0
p
cos .d !/ = 2; 8! 2 Œ; ; (10.26)
1 X
M1
. r;m .!; h; g/ "m j r;m .!; h; g/j/
D mD0
p
C cos .d !/ = 2; 8! 2 Œ; ; (10.27)
1 X
M1
. i;m .!; h; g/ C "m j i;m .!; h; g/j/
D mD0
p
C sin .d !/ = 2; 8! 2 Œ; ; (10.28)
1 X
M1
. i;m .!; h; g/ "m j i;m .!; h; g/j/
D mD0
p
sin .d !/ = 2; 8! 2 Œ; : (10.29)
1 X
M1
p 1 X
M1
2 "m j r;m .!; h; g/j ;
D mD0
192 L. Jiang et al.
and
!
1 X
M1
r;m .!; h; g/ cos .d !/
D mD0
p 1 X
M1
= 2 "m j r;m .!; h; g/j ;
D mD0
respectively, we have
ˇ M1 ˇ
ˇ1 X ˇ
ˇ ˇ
ˇ r;m .!; h; g/ cos .d !/ˇ
ˇD ˇ
mD0
p 1 X
M1
= 2 "m j r;m .!; h; g/j :
D mD0
Thus, a necessary condition for the solvability of Problem 10.1 is that there exits h
and g; such that
1 X p
M1
"m j r;m .!; h; g/j = 2; 8! 2 Œ; : (10.30)
D mD0
Taking "Q D min f"0 ; "1 ; ; "M1 g and ! D 0 in (10.12) and (10.13), we obtain
ˇM1 ˇ ˇM1 ˇ
p "Q ˇˇ X ˇ
ˇ "Q ˇˇ X ˇ
ˇ
"Q 1 = 2 ˇ r;m .0; h; g/ˇ ˇ r;m .0; h; g/ˇ
D ˇmD0 ˇ Dˇ
mD0
ˇ
"Q X 1 X p
M1 M1
j r;m .0; h; g/j "m j r;m .0; h; g/j = 2: (10.31)
D mD0 D mD0
p
If "Q 2 1; then "Q by (10.31). This illustrates the relationship between
the disturbance and the tolerance in (10.11). Actually, from our computational
experience, should be much larger than max f"0 ; "1 ; ; "M1 g :
Let
ˇ M1 ˇ
ˇ1 X ˇ
ˇ ˇ
F1 .!/ D ˇ r;m .!; h; g/ cos .d !/ˇ (10.32)
ˇD ˇ
mD0
1 X
M1
C "m j r;m .!; h; g/j ;
D mD0
10 The Worst-Case DFT Filter Bank Design with Subchannel Variations 193
ˇ M1 ˇ
ˇ1 X ˇ
ˇ ˇ
F2 .!/ D ˇ i;m .!; h; g/ C sin .d !/ˇ (10.33)
ˇD ˇ
mD0
1 X
M1
C "m j i;m .!; h; g/j :
D mD0
Then, according
p to Remark 10.1,pthe constraints (10.26)–(10.29) are equivalent to
F1 .!/ = 2 and F2 .!/ = 2:
Note from Theorem 10.4 that the continuous inequality constraints in Problem
10.3 are only with respect to w: However, some of the constraint functions are
non-smooth since they appear in the form as absolute value functions. Thus,
gradient-based optimization methods cannot be applied directly. In order to remove
this obstacle, we introduce the following smoothing approximation (Teo and Goh
1988):
8
< y; if y ;
' .y/ D 2 C y2 = .2/ ; if jyj < ; (10.34)
:
y; if y ;
where > 0 and y 2 R. Now we replace j r;m .!; h; g/j and j i;m .!; h; g/j
in the continuous inequality constraints (10.26)–(10.29) and obtain the following
optimization problem
min (10.35)
h;g;
subject to
.h; g; ıN i / ; i D 1; ; 2M ; (10.36)
1 X
M1
1 X
M1
r;m .!; h; g/ "m ' . r;m .!; h; g// (10.38)
D mD0
p
C cos .d !/ = 2; 8! 2 Œ; ;
194 L. Jiang et al.
1 X
M1
1 X
M1
i;m .!; h; g/ "m ' . i;m .!; h; g// (10.40)
D mD0
p
sin .d !/ = 2; 8! 2 Œ; :
Algorithm 10.1
• Step 1: Initialize 1 > 0 and k D 1: Set ıi D 0; i D 0; 1; ; M 1; and use the bi-iterative
optimization method in Dam et al. (2005) to design the filter bank. Let the solution obtained be
denoted by h
0 ; g0 .
• Step 2: Solve Problem .Pk / with the initial condition h
k1 ; gk1 . Let the cost and solution
obtained be denoted by k and hk ; gk ; respectively.
• Step 3: Set kC1 D k =L; where L > 1 is a pre-specified number.
• Step 4: If k1 k κ; where κ > 0 is a prescribed small number, stop. Otherwise, set
k D k C 1, go to Step 2.
In this section, we will use our developed algorithm to design the analysis and syn-
thesis prototype filters with subchannel variations. In the following discussion, D
102 : Furthermore, the continuous constraints interval Œ; is discretized into
512 equally spaced frequency points for the optimization Problem .P /. Consider
the case with M D 4; D D 2; " D Œ"0 ; "1 ; "2 ; "4 T D Œ0:001; 0:002; 0:004; 0:004T :
Let La D Ls D 4M C 1; d D 4M.
First, we consider the case without subchannel variations, i.e., ıi D 0; i D
0; 1; ; M 1: For such a filter bank design, the bi-iterative algorithm developed
in Dam et al. (2005) is introduced to design the initial analysis and synthesis
prototype filters. The cost obtained is 167:0414 dB. Let the prototype filters h and
g obtained be collected together and denote as x.0/ : We substitute x.0/ into (10.16)
and find that the largest value of .h; g; ı/ is 90:7572 dB which is obtained at
the extreme point Œ0:001; 0:002; 0:004; 0:004T of U . The maximum violation of
the constraints (10.12)–(10.15) is 0:003973: From these results, we can see that the
aliasing .h; g; ı/ will have a large increase if the subchannels with variations. Thus,
it is necessary to consider the robust optimal filter bank design.
Now, we consider the case of the subchannels with variations. We use the
prototype filters x.0/ obtained by setting ıi D 0 as the initial condition and use
Algorithm 10.1 to optimize Problem .P / with 1 D 103 ; L D 10; κ D
106 . After three iterations, the optimal solution is obtained. The optimal cost is
106:4557 dB which is obtained at the extreme point Œ0:001; 0:002; 0:004; 0:004T
of U . Clearly, there is not only about 16 dB improvement by using the developed
robust optimization method when compared with the result obtained by the bi-
iterative algorithm developed in Dam et al. (2005), but also maintaining the
constraints (10.12)–(10.15) when the subchannels with variations. However, it
should be noted that the problem considered in Dam et al. (2005) is an optimization
problem of a uniform FIR filter bank with group delay specification but without the
consideration of the variations in subchannels. The coefficients of the prototype
analysis and synthesis filters are presented in Table 10.1 and the corresponding
frequency responses are plotted in Figs. 10.2 and 10.3, respectively. The frequency
response of T ej! with ı D Œ0; 0; 0; 0T in (10.9) is plotted in Fig. 10.4 and F1 .w/
and F2 .w/ are plotted in Fig. 10.5. If we take all the "i to be the same, i.e., "i D
0:005: Then the analysis and synthesis prototype filters designed by our method are
plotted in Fig. 10.6 and F1 .!/ and F2 .!/ are depicted in Figs. 10.7 and 10.8. The
cost obtained is 109:7622 dB. Thus, there is about 19 dB improvement by using
our design method. Furthermore, the filter bank design by our method satisfies the
constraints (10.12)–(10.15) for any ı in U since the constraints (10.26)–(10.29)
are satisfied. From Fig. 10.7, we can see that this is not the case if the bi-iterative
algorithm developed in Dam et al. (2005) is used.
Now let M D 8; D D 4; " D Œ0:002; 0:001; 0:002; 0:001; 0:002; 0:001; 0:002;
0:001T . La D Ls D 4M; d D 4M. We use the bi-iterative algorithm in Dam
et al. (2005) to design the initial analysis and synthesis prototype filters with
196 L. Jiang et al.
10
0
Analysis prototype filter
−10
Frequency response in dB
−20
−30
−40
−50
−60
−70
−80
−90
0 0.2 0.4 0.6 0.8 1
The normalized frequency
Fig. 10.2 The frequency response of the analysis prototype filter with " D Œ0:001; 0:002;
0:004; 0:004T
10 The Worst-Case DFT Filter Bank Design with Subchannel Variations 197
10
Frequency response in dB
−20
−30
−40
−50
−60
−70
−80
0 0.2 0.4 0.6 0.8 1
Fig. 10.3 The frequency response of the synthesis prototype filter with " D Œ0:001; 0:002;
0:004; 0:004T
−3
x 10
3.5
2.5
Frequency response in dB
1.5
0.5
0
−1 −0.5 0 0.5 1
Fig. 10.4 The frequency response of T ejw with ı D Œ0; 0; 0; 0T and " D Œ0:001; 0:002;
0:004; 0:004T
198 L. Jiang et al.
−3
x 10
8
7 F (w)
1
F (w)
2
6
0
−1 −0.5 0 0.5 1
Fig. 10.5 The figures of F1 .!/ and F2 .!/ with " D Œ0:001; 0:002; 0:004; 0:004T
20
−20
−40
−60
−80
−100
0 0.2 0.4 0.6 0.8 1
The normalized frequency
Fig. 10.6 The frequency responses of the analysis and synthesis prototype filters with " D
Œ0:005; 0:005; 0:005; 0:005T
10 The Worst-Case DFT Filter Bank Design with Subchannel Variations 199
0.014
F (w)
1
F (w)
2
0.012
0.01
0.008
0.006
0.004
0.002
0
−1 −0.5 0 0.5 1
Fig. 10.7 The figures of F1 .!/ and F2 .!/ by the method in Dam et al. (2005)
−3
x 10
8
F1(w)
1
F2(w)
0
−1 −0.5 0 0.5 1
Fig. 10.8 The figures of F1 .w/ and F2 .w/ with " D Œ0:005; 0:005; 0:005; 0:005T
of U : Now we use Algorithm 10.1 to solve Problem 10.3 with the obtained
prototype filters as initial guess. After two iterations, the optimal cost 98:5855 dB
is obtained. The analysis and synthesis prototype filters are plotted in Fig. 10.9. The
200 L. Jiang et al.
20
Frequency response in dB
−20
−40
−60
−80
−100
0 0.2 0.4 0.6 0.8 1
The normalized frequency
Fig. 10.9 The frequency response of the synthesis prototype filter with M D 8, D D 4, " D
Œ0:002; 0:001; 0:002; 0:001; 0:002; 0:001; 0:002; 0:001 T
−3
x 10
8
F1(w)
7 F (w)
2
6
Frequency response in dB
0
−1 −0.5 0 0.5 1
The normalized frequency
Fig. 10.10 The figures of F1 .!/ and F2 .!/ with " D Œ0:002; 0:001; 0:002; 0:001; 0:002; 0:001;
0:002; 0:001T
corresponding F1 .!/ and F2 .!/ are plotted in Fig. 10.10. From the results obtained,
we can see that the aliasing can achieve about 28 dB improvement by using our
robust design method.
10 The Worst-Case DFT Filter Bank Design with Subchannel Variations 201
10.6 Conclusions
We have proposed a new formulation of the DFT filter bank problem in which the
subchannel are with variations. Comparing with the earlier formulations, our formu-
lation is more realistic since the filter operation with distortion in each subband has
been taken into consideration. It is in the form of a minimax optimization problem
with continuous inequality constraints. Although this minimax optimization can be
reformulated as a semi-infinite optimization problem by introducing an additional
variable, it is still cannot be solved directly by any existing method for semi-infinite
optimization problems. This is because the continuous constraints are not only with
respect to frequency, but also with respect to variations in subchannels. However, by
exploiting its properties, we proved that such a semi-infinite optimization problem is
equivalent to a semi-infinite optimization problem where the continuous constraints
are only with respect to frequency. Then, an approximate computation scheme is
developed to solve the transformed semi-infinite optimization problem. Simulation
results showed that the new method achieved a very high aliasing suppression while
maintaining the distortion under variations in the different filter bands to be a small
level.
Appendix 1
1 XX X D1
X
M1 M1 D1
.h; g; ı/ D .1 C ım / .1 C ın / hT ˆm;n;d;l .g/ h; (10.41)
D mD0 nD0 dD1 lD1
where ˆm;n;d;l .g/ is a La La matrix. The .i; j/-th element of ˆm;n;d;l .g/ is given by
s 1 L
LX Xs 1
2
Œˆm;n;d;l .g/i;j D cos .m n/ .i C t 2/
tD0 sD0
M
2
C .d .i 1/ l .j 1// ı .i C t j s/ g .t/ g .s/ ;
D
1; if t D 0;
ı .t/ D
0; if t ¤ 0:
202 L. Jiang et al.
Appendix 2
ˇ
jık j D ˇ N ˇ
i ıi;k ˇ
ˇN ˇ
i ıi;k D i "k D "k :
ˇ iD1 ˇ iD1 iD1
we have
Since jı1 j "1 ; there exists a 1 such that ı1 D 1 "1 .1 1 / "1 and 0 1 1:
Thus,
.h; g; ı/ D ı T Qı C qT ıC q;
10 The Worst-Case DFT Filter Bank Design with Subchannel Variations 203
T
where Q D Q1=2 Q1=2 is a semi-positive definite matrix, q and q are correspond-
ing vector and constant. For any ; 0 1, ı 1 and ı 2 ; we have
.h; g; ı 1 / C .1 / .h; g; ı 2 /
.h; g; ı 1 C .1 / ı 2 /
2 T
D ı T1 Qı 1 C .1 / ı T2 Qı 2 ı 1 Qı 1
.1 /2 ı T2 Qı 2 2 .1 / ı T1 Qı 2
D .1 / ı T1 Qı 1 C ı T2 Qı 2 2ı T1 Qı 2
h T
D .1 / Q1=2 ı 1 Q1=2 ı 1
T T
i
C Q1=2 ı 2 Q1=2 ı 2 2 Q1=2 ı 1 Q1=2 ı 2
0:
˚ with respect to ı. Suppose that ı D
Thus, .h; g; ı/ is a convex function
arg max .h; g; ı/ and N D max .h; g; ıN 1 /; ; .h; g; ıN 2M / : From Theo-
ı2U
P
2M
rem 10.1, we know that there exists i 0; 0 i 2M ; with i D 1;
iD1
P
2M
such that ı D N Since .h; g; ı/ is a convex function with respect to ı, we
iıi :
iD1
have
2
X
M
.h; g; ı / D .h; g; N
i ıi /
iD1
2M
X 2
X
M
i .h; g; ıN i / i N D N:
iD1 iD1
1 X
M1
D .1 C ım / r;m .!; h; g/
D mD0
1 X 1 X
M1 M1
D r;m .!; h; g/ C ım r;m .!; h; g/ :
D mD0 D mD0
204 L. Jiang et al.
Hence,
˚
max hT Re ‰.ej! ; ı/ g
ı2U
1 X 1 X
M1 M1
D r;m .!; h; g/ C max ım r;m .!; h; g/ :
D mD0 ı2U D mD0
1 X
M1
jım j j r;m .!; h; g/j
D mD0
1 X
M1
"m j r;m .!; h; g/j :
D mD0
Hence,
1 X 1 X
M1 M1
max ım r;m .w; h; g/ "m j r;m .!; h; g/j : (10.42)
ı2U D mD0 D mD0
h iT
On the other hand, taking ıQ D ıQ0 ; ıQ1 ; ; ıQM1 ; where ıQm D "m r;m .!; h; g/ =
j r;m .!; h; g/j ; yields
1 XQ
M1
ım r;m .!; h; g/
D mD0
1 X
M1
D "m . r;m .!; h; g//2 = j r;m .!; h; g/j
D mD0
1 X
M1
D "m j r;m .!; h; g/j :
D mD0
1 X 1 X
M1 M1
max ım r;m .!; h; g/ D "m j r;m .!; h; g/j : (10.43)
ı2U D mD0 D mD0
References
Dam HH, Nordholm S, Cantoni A (2005) Uniform FIR filterbank optimization with group delay
specification. IEEE Trans Signal Process 53(11):4249–4260
de Haan JM, Grbic N, Claesson I, Nordholm S (2001) Design of oversampled uniform DFT filter
banks with delay specification using quadratic optimization. In: Proceedings of ICASSP’2001,
Salt Lake City, pp 3633–3636
de Haan JM, Grbic N, Claesson I, Nordholm S (2003) Filter bank design for subband adaptive
microphone arrays. IEEE Trans Speech Audio Process 11(1):14–23
Harteneck M, Weiss S, Stewart RW (1999) Design of near perfect reconstruction oversampled filter
banks for subband adaptive filters. IEEE Trans Circuits Syst 46:1081–1085
Kellermann W (1988) Analysis and design of multirate systems for cancellation of acoustic echoes.
In: Proceedings of ICASSP’88, New York, pp 2570–2573
Kha HH, Tuan HD, Nguyen TQ (2009) Efficient design of cosine-modulated filter banks via convex
optimization. IEEE Trans Signal Process 57(3):966–976
Lopez M, Still G (2007) Semi-infinite programming. Eur J Oper Res 180:491–518
Mansour MF (2007) On the optimization of oversampled DFT filter banks. IEEE Signal Process
Lett 14(6):389–392
Nguyen TQ (1994) Near-perfect-reconstruction pseudo-QMF banks. IEEE Trans Signal Process
42(1):65–75
Sturm JF (1999) Using SeDuMi 1.02, a Matlab toolbox for optimization over symmetric cones.
Optim Methods Softw 11–12:625–653
Teo KL, Goh CJ (1988) On constrained optimization problems with nonsmooth cost functionals.
Appl Math Optim 18:181–190
Vaidyanathan PP (1993) Multirate systems and filter banks. Prentice-Hall, Englewood Cliffs
Wilbur MR, Davidson TN, Reilly JP (2004) Efficient design of oversampled NPR GDFT
filterbanks. IEEE Trans Signal Process 52(7):1947–1963
Wu CZ, Teo KL (2010) A dual parametrization approach to Nyquist filter design. Signal Process
90:3128–3133
Wu CZ, Teo KL (2011) Design of discrete Fourier transform modulated filter bank with sharp
transition band. IET Signal Process 5:433–440
Wu CZ, Teo KL, Rehbock V, Dam HH (2008) Global optimum design of uniform FIR filter bank
with magnitude constraints. IEEE Trans Signal Process 56(11):5478–5486
Wu CZ, Gao D, Teo KL (2013) A direct optimization method for low group delay FIR filter design.
Signal Process 93:1764–1772
Yiu KFC, Grbic N, Nordholm S, Teo KL (2004) Multicriteria design of oversampled uniform DFT
filter banks. IEEE Signal Process Lett 11(6):541–544
Zhang ZJ, Shui PL, Su T (2008) Efficient design of high-complexity cosine modulated filter banks
using 2Mth band conditions. IEEE Trans Signal Process 56(11):5414–5426