Professional Documents
Culture Documents
Integral Methods in Science and Engineering
Integral Methods in Science and Engineering
Harris
Editors
Integral Methods
in Science and
Engineering
Mathematics Subject Classification (2010): 34-06, 35-06, 45-06, 65-06, 74-06, 76-06
www.birkhauser-science.com
To the memory of Igor Chudinovich and
Alain Largillier, former members of the
International Steering Committee of the
IMSE Consortium
Preface
vii
viii Preface
The next IMSE conference will be hosted by the Federal University of Rio
Grande do Sul and the National Institute for Space Research in Porto Alegre, RS,
Brazil, in July 2012. Further details will be posted in due course through the link
http://sites.google.com/site/imse2012.
This volume contains 37 refereed papers from among those presented in Brighton,
arranged alphabetically by the first author’s name. The editors would like to thank
the staff at Birkhäuser-Boston for their professionalism and efficient handling of the
publication process.
ix
x Contents
4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
List of Contributors
Mario Ahues
Université Jean Monnet, Université de Lyon, 23 rue du Dr. Paul Michelon,
Saint-Étienne 42023, cedex 2, France,
e-mail: mario.ahues@univ-st-etienne.fr
Lino J. Alvarez-Vázquez
Universidad de Vigo, ETSI Telecomunicación, 36310 Vigo, Spain,
e-mail: lino@dma.uvigo.es
Daniel Beck
Universidade Federal do Rio Grande do Sul, PPGMap & PROMEC, Av. Osvaldo
Aranha, 99/4, 90046-900 Porto Alegre, RS, Brazil,
e-mail: daniel.beck@passofundo.ifsul.edu.br
John R. Blake
University of Birmingham, Edgbaston, Birmingham B15 2TT, UK,
e-mail: j.r.blake@bham.ac.uk
Bardo E.J. Bodmann
Universidade Federal do Rio Grande do Sul, PPGMap & PROMEC, Av. Osvaldo
Aranha, 99/4, 90046-900 Porto Alegre, RS, Brazil,
e-mail: bardo.bodmann@ufrgs.br
Lucas N. Burigo
Universidade Federal de Rio Grande do Sul, PPGMap & PROMEC, Av. Osvaldo
Aranha, 99/4, 90046-900 Porto Alegre, RS, Brazil,
e-mail: lucas_burigo@yahoo.com
Daniela Buske
Universidade Federal de Pelotas, Instituto de Fisica e Matemática, Campus Capão
do Leão, Caixa Postal 354, 96010-900 Pelotas, RS, Brazil,
e-mail: daniela.buske@ufpel.edu.br
xix
xx List of Contributors
Dayana Q. de Camargo
Universidade Federal do Rio Grande do Sul, PPGMap & PROMEC, Av. Osvaldo
Aranha, 99/4, 90046-900 Porto Alegre, RS, Brazil,
e-mail: dayanadecamargo@gmail.com
Haroldo F. de Campos Velho
Instituto Nacional de Pesquisas Espaciais (INPE), P.O. Box 515, São José dos
Campos, SP 12245-970, Brazil,
e-mail: haroldo@lac.inpe.br
Celina Ceolin
Universidade Federal do Rio Grande do Sul, Av. Osvaldo Aranha, 99/4, 90046-900
Porto Alegre, RS, Brazil,
e-mail: celina.ce@hotmail.com
David J. Chappell
University of Nottingham, University Park, Nottingham NG7 2RD, UK,
e-mail: david.chappell@nottingham.ac.uk
Ke Chen
University of Liverpool, Peach Street, Liverpool L69 7ZL, UK,
e-mail: k.chen@liverpool.ac.uk
Subhashini Chitta
Flow Analysis, Inc., 256 93rd Street, Brooklyn, NY 11209, USA,
e-mail: subha@flowanalysis.com
Otar Chkadua
A. Razmadze Mathematical Institute, I. Javakhishvili Tbilisi State University,
2 University Street, Tbilisi 0186, Georgia,
e-mail: chkadua@rmi.ge
Christian Constanda
The University of Tulsa, 800 S. Tucker Drive, Tulsa, OK 74104, USA,
e-mail: christian-constanda@utulsa.edu
Constantin Corduneanu
University of Texas at Arlington, 411 S. Nedderman Drive, Box 19408, Arlington,
TX 76019, USA,
e-mail: concord@uta.edu
Geoff A. Curtiss
University of Birmingham, Edgbaston, Birmingham B15 2TT, UK,
e-mail: geoffc@nag.co.uk
Delfina Gómez
Universidad de Cantabria, Av. de los Castros s/n, 39005 Santander, Spain,
e-mail: gomezdel@unican.es
List of Contributors xxi
Luis Gómez
The University of Tulsa, 800 S. Tucker Drive, Tulsa, OK 74104, USA,
e-mail: lgm@Utulsa.edu
Hamza Guebbaï
Université Jean Monnet, Université de Lyon, 23 rue du Dr. Paul Michelon,
Saint-Étienne 42023, cedex 2, France,
e-mail: guebaihamza@yahoo.fr
Igor A. Guz
Centre for Micro- and Nanomechanics, University of Aberdeen, King’s College,
Aberdeen AB24 3UE, UK,
e-mail: i.guz@abdn.ac.uk
Dimiter Hadjimichef
Universidade Federal do Rio Grande do Sul, PPGMap & PROMEC, Av. Osvaldo
Aranha, 99/4, 90046-900 Porto Alegre, RS, Brazil,
e-mail: dimihadj@gmail.com
Salim M. Haidar
Grand Valley State University, 1 Campus Drive, Allendale, MI 49401, USA,
e-mail: haidars@gvsu.edu
Layal Hakim
Brunel University West London, Kingston Lane, Uxbridge, Middlesex UB8 3PH,
UK,
e-mail: layal.hakim@brunel.ac.uk
Paul J. Harris
School of Computing, Engineering and Mathematics, University of Brighton,
Lewes Road, Brighton BN2 4GJ, UK,
e-mail: p.j.harris@brighton.ac.uk
Cheryl L. Jorcyk
Boise State University, 1910 University Drive, Boise, ID 83725-1515, USA,
e-mail: cjorcyk@boisestate.edu
Jukka Kemppainen
University of Oulu, P.O.Box 4500, 90014 Oulu, Finland,
e-mail: jukemppa@paju.oulu.fi
Mikhail Kolev
University of Warmia and Mazury, Zolnierska 14, 10-561, Olsztyn, Poland,
e-mail: kolev@matman.uwm.edu.pl
Gene Kouba
Chevron Energy Technology Company, 1400 Smith Street, Houston, TX 77002,
USA,
e-mail: genekouba@chevron.com
xxii List of Contributors
Adriana Nastase
Aerodynamik des Fluges, RWTH Aachen University, Templergraben 55, 52062
Aachen, Germany,
e-mail: nastase@lafaero.rwth-aachen.de
David Natroshvili
Georgian Technical University, 77 M. Kostava, 0175 Tbilisi, Georgia,
e-mail: natrosh@hotmail.com
Sergey A. Nazarov
Institute for Problems in Mechanical Engineering, RAN V.O. Bol’shoi pr., 61,
199178 St. Petersburg, Russia,
e-mail: srgnazarov@yahoo.co.uk
Fabiana F. Paes
Instituto Nacional de Pesquisas Espaciais (INPE), P.O. Box 515, São José dos
Campos, SP 12245-970, Brazil,
e-mail: fabiana.paes@lac.inpe.br
Cláudio Pellegrini
Universidade Federal de São João del-Rei, Praça Frei Orlando 170, 36307-904 São
João del-Rei, MG, Brazil,
e-mail: pelle@ufsj.edu.br
Mikhail Perelmuter
A. Ishlinsky Institute for Problems in Mechanics, Russian Academy of Sciences,
Pr. Vernadskogo 101-1, Moscow 119526, Russia,
e-mail: perelm@ipmnet.ru
Eduardo Pereyra
The University of Tulsa, 800 S. Tucker Drive, Tulsa, OK 74104, USA,
e-mail: ep@utulsa.edu
Eugenia Pérez
Universidad de Cantabria, Avda de los Castros s/n, 39005 Santander, Spain,
e-mail: meperez@unican.es
Claudio Z. Petersen
Universidade Federal do Rio Grande do Sul, PPGMap & PROMEC, Av. Osvaldo
Aranha, 99/4, 90046-900 Porto Alegre, RS, Brazil,
e-mail: claudiopetersen@yahoo.com.br
Shirley Pomeranz
The University of Tulsa, 800 S. Tucker Drive, Tulsa, OK 74104, USA,
e-mail: pomeranz@utulsa.edu
Régis S. Quadros
Universidade Federal de Pelotas, Instituto de Fisica e Matématica, Campus Capão
do Leão, Caixa Postal 354, 96010-900 Pelotas, RS, Brazil,
e-mail: regis.quadros@ufpel.edu.br
xxiv List of Contributors
Fernando M. Ramos
Instituto Nacional de Pesquisas Espaciais (INPE), P.O. Box 515, São José dos
Campos, SP 12245-970, Brazil,
e-mail: fernando@lac.inpe.br
Paula Sanematsu
University of Tennessee Space Institute, 411 B.H. Goethert Parkway, Tullahoma,
TN 37388, USA,
e-mail: paulasan@gmail.com
Jorge San Martín
Universidad de Chile, Casilla 170/3-Correo 3, Santiago 8370459, Chile,
e-mail: jorge@dim.uchile.cl
Jean-Frano̧is Scheid
Université Henri Poincaré, BP239, F-54506 Vandoeuvre-lès-Nancy, Cedex, France,
e-mail: scheid@iecn.u-nancy.fr
Marcelo Schramm
Universidade Federal do Rio Grande do Sul, PPGMap & PROMEC, Av. Osvaldo
Aranha, 99/4, 90046-900 Porto Alegre, RS, Brazil,
e-mail: marceloschramm@hotmail.com
Cynthia F. Segatto
Universidade Federal do Rio Grande do Sul, PPGMap, Av. Bento Gonçalves 9500,
91509-900 Porto Alegre, RS, Brazil,
e-mail: csegatto@pq.cnpq.br
Ovadia Shoham
The University of Tulsa, 800 S. Tucker Drive, Tulsa, OK 74104, USA,
e-mail: ovadia-shoham@utulsa.edu
Loredana Smaranda
Universitatea din Piteşti, Str. Targu din Vale Nr. 1, 110040 Piteşti, Romania,
e-mail: smaranda@dim.uchile.cl
John Steinhoff
University of Tennessee Space Institute, 411 B.H. Goethert Parkway, Tullahoma,
TN 37388, USA,
e-mail: jsteinho@utsi.edu
Gregor Tanner
University of Nottingham, University Park, Nottingham NG7 2RD, UK,
e-mail: gregor.tanner@nottingham.ac.uk
Leslie G. Thompson
The University of Tulsa, 800 S. Tucker Drive, Tulsa, OK 74104, USA,
e-mail: lgt@utulsa.edu
List of Contributors xxv
Vladimir B. Vasilyev
Bryansk State University, Bezhitskaya 14, Bryansk 241036, Russia,
e-mail: vbv57@inbox.ru
Miguel E.Vázquez-Méndez
Universidad de Santiago, Escola Politécnica Superior, 27002 Lugo, Spain,
e-mail: miguelernesto.vazquez@usc.es
Miguel A. Vilar Rivas
Universidad de Santiago, Escola Politécnica Superior, 27002 Lugo, Spain,
e-mail: miguel.vilar@usc.es
Marco T. Vilhena
Universidade Federal do Rio Grande do Sul, Av. Osvaldo Aranha, 99/4, 90046-900
Porto Alegre, RS, Brazil,
e-mail: vilhena@pq.cnpq.br
Qian X. Wang
University of Birmingham, Edgbaston, Birmingham B15 2TT, UK,
e-mail: q.x.wang@bham.ac.uk
Jorge R. Zabadal
Universidade Federal do Rio Grande do Sul, PPGMap & PROMEC, Av. Osvaldo
Aranha, 99/4, 90046-900 Porto Alegre, RS, Brazil,
e-mail: zabadal@ufrgs.br
Gang Zhao
University of Regina, 3737 Wascana Parkway, Regina, SK, Canada S4S 0A2,
e-mail: gang.zhao@uregina.ca
Vladimir V. Zozulya
Centro de Investigación Científica de Yucatán, A.C., Calle 43, no. 130, Colonia
Chuburna de Hidalgo, C.P. 97200, Mérida, Yucatán, Mexico,
e-mail: zozulya@cicy.mx
Barbara Zubik-Kowal
Boise State University, 1910 University Drive, Boise, ID 83725-1555, USA,
e-mail: zubik@math.boisestate.edu
A Collocation Method for Cauchy Integral
Equations in L2
1 Introduction
For each nonzero real constant μ and each real function f , consider the problem of
finding a function ϕ such that
M. Ahues
Université Jean Monnet, Université de Lyon, F-42023 Saint-Étienne, France,
e-mail: mario.ahues@univ-st-etienne.fr
A. Mennouni
University of Bordj Bou-Arreridj, Algeria,
e-mail: aziz.mennouni@yahoo.fr
Let X := L2 (0, 2π ), and Xn denote the space spanned by the first 2n + 1 trigono-
metric polynomials. Define σn to be the orthogonal projection from X onto Xn .
Hence, for ψ ∈ L2 (0, 2π ),
lim σn ψ − ψ 2 = 0.
n→∞
Let be 0 < sn,1 < sn,2 < · · · < sn,2n+1 < 2π . For each i ∈ [[1, 2n + 1 ]] consider the hat
function en,i in C0 (0, 2π ), such that, for each j ∈ [[1, 2n + 1 ]],
Let Yn be the space spanned by these hat functions, which has dimension 2n + 1.
Define the interpolation projection operator πn from C0 (0, 2π ) onto Yn :
2n+1
πn h(s) := ∑ h(sn, j )en, j (s), h ∈ C0 (0, 2π ).
j=1
which is compact and skew-Hermitian from L2 (0, 2π ) into itself. Let ϕε be the so-
lution of the regularized integral equation
(μ I − Tε )ϕε = f ,
Tε ,n := πn Tε σn .
(μ I − Tε ,n )ψε ,n = f ,
we get
ψε ,n − ϕε 2 ≤ βε (Tε − Tε ,n )ϕε 2 → 0 as n → ∞.
Since Tε is skew-Hermitian,
1
(μ I − Tε )−1 ≤ ,
|μ |
3 Numerical Example
Let μ = −1 and
f (s) := −s[sin s+Si (s) cos s−sin(s)Ci (s)−Si (s−2π ) cos s+Ci (2π −s) sin s],
where Si is the sine integral function, and Ci is the cosine integral function. Then
the exact solution of (1) is ϕ (s) := s sin s. For the regularization process, take ε =
10−4 . For the numerical approximation take n = 9, n = 23 and n = 62. The results
are exhibited in Figs. 1, 2 and 3, respectively.
Fig. 1 n = 9
Fig. 2 n = 23
A Collocation Method for Cauchy Equations 5
Fig. 3 n = 62
References
[AhLaLi01] Ahues, M., Largillier, A., Limaye, B.V.: Spectral Computations for Bounded Oper-
ators, CRC Press, Boca Raton (2001).
[Mu53] Mushkelishvili, N.I.: Singular Integral Equations, Noordhoff, Groningen (1953).
[PoSt90] Porter, D., Stirling, D.: Integral Equations: A Practical Treatment, from Spectral
Theory to Applications, Cambridge University Press (1990).
[Se93] Sengupta, A.: A note on a reduction of Cauchy singular integral equation to Fred-
holm equation in L p , Applied Mathematics and Computation, 56, 97–100 (1993).
On a New Definition of the Reynolds Number
from the Interplay of Macroscopic
and Microscopic Phenomenology
1 Introduction
B.E.J. Bodmann
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: bardo.bodmann@ufrgs.br
M.T. Vilhena
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: vilhena@pq.cnpq.br
J.R. Zabadal
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: zabadal@ufrgs.br
D. Beck
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: daniel.beck@passofundo.ifsul.edu.br
measure quantities. In fact, fluids are made from atoms and molecules which obey
microscopic laws and collectively constitute a stochastic system. This system obeys
macroscopic laws provided by statistical thermodynamics and hydrodynamics. Both
realms can be described phenomenologically by macroscopic observables and mate-
rial dependent parameters (viscosity, thermal conductivity, specific heat, compress-
ibility, among others), where these parameters hide the microscopic properties. If
one traces back the parameters until its microscopic origin, in principle it should be
possible to find quantitative macroscopic-microscopic relationships beyond mere
phenomenology, where viscosity and length scales among others are macroscopic
manifestations with microscopic origin.
In this line we reason that turbulence, which is related to the Reynolds number,
may be considered an interplay of the dynamics of at least two scales, a macro-
scopic one and a microscopic one. Thus, the present work is an attempt to show
the possibilities that arise from, in our case, a simplified macroscopic-microscopic
relationship, which we derive based on a simplified model motivated by Maxwell–
Boltzmann transport. This chapter is organized as follows. In the next section we
present the microscopic approach and introduce a length scale which relates to vor-
ticity, in Sect. 3 we identify viscosity based on microscopic and thermodynamical
quantities and last (Sect. 4) we discuss our findings and give future perspectives.
fluences the remainder of the fluid. The microscopic picture suggests a non-dense
point set of interaction centers and a complementary dense set of interaction free
points, where the microscopic behavior, because of its different topology in com-
parison to a continuous approach, may give rise to a different collective behavior
(ensemble averages) on a macroscopic scale.
On the macroscopic scale we understand the velocity fieldv(x, y, z,t) = c as the
ensemble average of particle velocities c in a given volume element Δ V centered at
r = (x, y, z) at an instant t. The first difficulty arises when trying to capture a typical
macroscopic length scale, based on a microscopic property, which shall be related to
a strength with which a flow is perturbed in order to present turbulent behavior. To
this end we define the dimensionless velocity vector field G( r,t) = v and consider
Cth
the field infinitesimally displaced G →G δ R , which shall simulate the change in the
velocity field by virtue of vorticity. One may establish the relation to the original
field by an infinitesimal coordinate transformation, which reads
−1r,t)
δ R (r,t) = R G(R
G
= (1 − δ θ G) G((1
− δ θ G)r,t)
+ δ θ (−G G
=G + G (r × (∇ × G)))
where G are the generators of the transformation R represented as a vector and each
component contains a 3 × 3 transformation matrix which act on G. θ is the infinites-
imal transformation parameter, i.e. a rotation angle with respect to a given axis θ̂ .
In component form and using the convention of summing over double indices, this
reads
Γδ R i = Γi + δ θ j (−εi jmΓm + ε jmn rm ∂nΓi ) (1)
where εi jk is the complete antisymmetric Levi-Civita symbol.
These findings may be related to vorticity using a concept from differential ge-
ometry, i.e., the generating term in (1) shall arise as a closed operator sequence—
translation (Γ ), vorticity (Ω ), back translation and vorticity again, around a plaque
of infinitesimal size.
Ωi dΓj ∝ −εi jmΓm + εimn rm ∂nΓj . (2)
An expression compatible with (2) and for any volume of interest then has the form
1 ∂Γj
Ωi = (−εi jmΓm + εimn rm ∂nΓj ) d 3 r. (3)
V V ∂t
One identifies two contributions, an extrinsic one which explicitly depends on the
position and a second contribution which is position independent and thus may have
only intrinsic origin. The presence of the second term can describe vorticity without
the phenomenon of creating eddies (for instance present in shear flows), whereas
the first term creates eddies even for a macroscopic velocity field which derives as
a gradient from a scalar potential, for which the second term cancels out. A further
10 B.E.J. Bodmann et al.
comment is in order here: the intrinsic term makes sense only if microscopic degrees
of freedom exist that constitute the macroscopic field Γ , since it depends only on
the velocity field and its temporal variation in different directions.
From this quantity (3) one may derive a macroscopic length scale which shall be
used in order to define the Reynolds number. One may recognize that Ω contains the
Γ fields in a bilinear form, so that the vorticity may be generalized to a correlation
like function, henceforth called vorticity correlator
1 ∂Γj 3
ϒt (t,t ) = (t) −εi jmΓm (t ) + εimn rm ∂nΓj (t ) d r
.
V V ∂t
In the limit t → t the correlator turns vorticity. Since the vorticity may be related
to the angular frequency of an eddy, the correlator may be used to measure how
far a particle with velocity Cth propagates across an eddy with non-vanishing cor-
relations. The length scale Λ is then defined by the correlation between time and
thermal velocity via the implicit relation
1
Cth τ ϒ0 (t, 0) 2
Λ = τ Cth = √ dt ,
2 0 ϒt (t, 0)
where
1
3
ϒ0 (t, 0) = Γj (t) (−εi jmΓm (0) + εimn rm ∂nΓj (0)) d r
V V
and the thermal noise limit limτ →0 τ1 0τ ϒϒ0t dt = 2.
So far the velocity field v = c, the macroscopic length scale Λ are available
from expectation values of a microscopic ensemble. The remaining quantities like
the particle density and the viscosity may be determined only from an analysis of
the transport equation, i.e., a Navier–Stokes type equation, which may be derived
starting from the Maxwell–Boltzmann transport equation.
pμν = pδμν holds and since dissipative contributions are no longer at work, the
diagonal (equilibrium) contributions to the pressure tensor contribute only to the
homogeneous solution of the transport equation. We are interested in the dissipative
part of the equation and hence reduce the pressure tensor to the friction pressure
tensor
1 2
πμν = nm Cμ Cν − C δμν
3
with zero trace T r{πμν } = 0. The second term, after decomposition and some alge-
braic manipulations separates a term with constant temperature and a velocity field
gradient, from a term that represents heat flux, which is kinetic energy transport
n
qμ = m C2Cμ ,
2
respectively. For simplicity we ignore possible contributions of an external force
field and its resulting acceleration bμ , so that the term still to be determined is the
right hand side of (4). A convenient way to simplify the equation is to approximate
the collision term by an average friction pressure change
δ πμν πμν
≈
δt τp
which renders the original equation a transport relaxation equation [Ch95, Ba08]
∂ 1 ∂ v μ ∂ vν 1 ∂ vλ
πμν + 2nkB T + − δμν
∂t 2 ∂ rν ∂ rμ 3 ∂ rλ
4 1 ∂ q μ ∂ qν 1 ∂ qλ πμν
+ + − δμν + = 0. (5)
5 2 ∂ rν ∂ rμ 3 ∂ rλ τp
The relaxation time constant τ p for phenomenological potentials and for systems
not far from equilibrium in (5) may be related to the microscopic cross section in the
spirit of Chapman–Cowling [Ma81]. The expression below shows the mechanical
relaxation time, for an isotropic two particle interaction central potential:
∞ π −1
5 1 −u2 7
√
τp = √ e u (1 − cos (θ ))σ (θ , 2Cth u) sin(θ ) d θ du
2
16 π nCth 0 0
5 1
= √ .
16 π nCth I2
√
Here u is the relative velocity between the collision partners in multiples of 2Cth ,
θ signifies the scattering angle, and Cth a velocity scale, i.e. the thermal velocity.
A local collision operator is responsible for the space–time evolution of the dis-
tribution in consideration. The collision term depends in general on microscopic
dynamics which in many cases is not exactly known or is too complex to be eval-
uated analytically. However, for a number of applications there do exist interaction
12 B.E.J. Bodmann et al.
models [Dh07] that are sufficient to capture qualitatively as well as to a certain pre-
cision quantitatively properties of the fluid flow.
The equation above results in the Navier–Stokes type equation if the following
phenomenological identity holds:
∂ vλ 1 ∂ v μ ∂ vν 1 ∂ vλ
πμν = −ηV δμν − 2η + − δμν
∂ rλ 2 ∂ rν ∂ rμ 3 ∂ rλ
5nkB T 5n mCth
η = kB T τ p = √ = √ .
16 π Cth I2 16 π I2
Upon substitution of the found quantities into the traditional Reynolds number defi-
nition and replacing the usually employed macroscopic length by the vortex correla-
tor length Λ one arrives at an expression which is characterized by two macroscopic-
microscopic ratios, the correlation length Λ against the mean free path λ and the
macroscopic flow velocity v against the thermal velocity Cth besides a factor which
is determined from the collision integral and the total collision cross section σT .
√
ρΛ v 16 π I2 Λ v
Re = = .
η 5 σT λ Cth
√
For a collision model where the cross section σ (θ , 2Cth u) does not depend on the
scattering angle the integral can be solved analytically and is I2 = σπT .
4 Conclusion
for instance in the problem of dispersion of pollution in the atmosphere and water.
These challenges define the next steps of future activities.
References
[Ba08] Banda, M., Klar, A., Pareschi, L., Seaïd, M.: Lattice-Boltzmann type relaxation systems
and high order relaxation schemes for the incompressible Navier–Stokes equations. Math.
Comp., 77, 943–965 (2008).
[Ca07] Cao, C., Titi, E.S.: Global well-posedness of the three-dimensional viscous primitive
equations of large scale ocean and atmosphere dynamics. Annals of Mathematics, 166,
245–267 (2007).
[Ca08] Cao, C., Titi, E.S.: Regularity criteria for the three-dimensional Navier–Stokes equations.
Indiana Univ. Math. J., 57, 2643–2662 (2008).
[Ch95] Cheng, M.-C.: An efficient approach to solving the Boltzmann transport equation. Simu-
lation of Semiconductor Devices and Processes, 6, 202–205 (1995).
[Co01] Constantin, P.: Some Open Problems and Research Directions in the Mathematical Study
of Fluid Dynamics. Mathematics Unlimited and Beyond, Springer Verlag, Berlin, 353–
360 (2001).
[Co07] Constantin, P., Levant, B., Titi, E.S.: A note on the regularity of inviscid shell model of
turbulence. Physics Review E, 75, 016304-1–016304-10 (2007).
[Dh07] Dhama, A.K., McCourt, F.R.W., Dickinson, A.S.: Accuracy of recent potential energy
surfaces for the He–N2 interaction I: Virial and bulk transport coefficients. J. Chem. Phys.,
127, 054302-1–054302-13 (2007).
[Fe06] Fefferman, C.L.: Fluids and singular integrals. Contemporary Math., 411, 39–52 (2006).
[Ma81] Maitland, G.C., Rigby, M., Smith, E.B., Wakeham,W.A.: Intermolecular Forces: Their
Origin and Determination. Oxford University Press, Oxford (1981).
[Mu79] Muncaster, R.G.: On generating exact solutions of the Maxwell–Boltzmann equation.
Arch. Rational Mech. Anal., 70, 79–90 (1979).
A Self-Consistent Monte Carlo Validation
Procedure for Hadron Cancer Therapy
Simulation
1 Introduction
L.N. Burigo
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: lucas_burigo@yahoo.com.br
D. Hadjimichef
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: dimihadj@gmail.com
B.E.J. Bodmann
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: bardo.bodmann@ufrgs.br
1 ∂ φX (r, Ω , E,t)
+ Ω · ∇φX (r, Ω , E,t) + ΣtX (E)φX (r, Ω , E,t)
v ∂t
= qX (r, Ω , E,t).
Here, v is the ion velocity, r,t are the space–time coordinates, Ω indicates the direc-
tional solid angle, ΣtX is the total macroscopic cross section for processes of parti-
cles of type X ∈ {B, Be, He, Li} and qX (r, Ω , E,t) represents the heavy-ion source
(accelerator) for primary particles S and particles produced in fragmentations (the
integral term). Note that we include a pseudo Bethe–Bloch cross section into ΣtX ,
which is possible if one uses the relation of linearized energy loss for a given energy
to the macroscopic cross section.
∞
qX (r, Ω , E,t) = ∑ dE
4π
d Ω ΣXY (r, E → E, Ω → Ω )
Y =B,Be,He,Li 0
Here δXHe is the Kronecker symbol, that is, 1 for X = He and 0 otherwise; the
first term includes contributions from several nuclear reactions, including particles
produced from scattering processes;
ΣXY (r, E → E, Ω → Ω ) dE d Ω
represents the number of ions of species Y emitted in the small energy interval dE
at E and small angular range d Ω at Ω from the reaction of ion X at energy E
and direction of motion Ω , with ΣXY (r, E → E, Ω → Ω ) being the corresponding
macroscopic cross section.
The energy dependent cross sections for each individual processes are provided
by the GEANT4 data libraries [Ago03, All06] that contain also a list of relative
atomic masses or molar masses of the elements. Note that detailed data for heavy-
ion collisions are still in their consolidation phase so that there are still considerable
uncertainties with respect to cross sections. The densities of the materials used in
a simulation are specified by the user, however in the case of graphite we show
that the physical condition of the material (atomic structure) is probably not prop-
erly taken into account, whereas for the water–gelatine mixture the uncertainties are
acceptable. Note that the latter has physical properties close to biological tissues.
In the further text we show by simulation results that the accuracy of the simula-
tions are compromised because of uncertainties in the input data, of which the cross
sections form a crucial part. Differences between simulation for the same problem
occurred when comparing the results of version 8.2 with version 9.3, patch 01. With
our proposed self-consistency analysis that we will show in the next sections we pin
down uncertainties almost to input data only. Once the simulation is free of logi-
cal errors or biases (because of idealizations and simplifications) validation of input
data may be performed.
A Self-Consistent Monte Carlo Validation Procedure 19
Simulations of 106 3 He ions with initial energy E = 207.92 MeV were performed
considering two different target materials, graphite (as specified in the GEANT4
data base) and a mixture of water and gelatine (user defined) that approximates the
properties of biological tissues. The choice of the two materials was made in order to
show one example where considerable differences appear depending on the version
of the simulation platform, here GEANT4 version 8.2 and version 9.3 with patch
01, respectively. The geometry, which in more realistic cases could be of combina-
torial form was assumed to be a homogeneous cylinder aligned with the direction
of the particle beam. Individual particles were transported across the medium from
vertex to vertex and all additional particles that appeared in the specific reactions
were propagated until their energies were totally lost in the medium. From the tal-
lies that were collected along the transport of individual particles, one may project
out the energy deposition with penetration depth, the yield of β + emitters, relevant
for online monitoring of the penetration depth of the primary particles and last not
least the electromagnetic ionizing processes manifest in the Bragg curve with its
characteristic peak at the end. In Fig. 1 we show the Bragg curve which illustrates
the advantage over photon treatments due to the enhanced energy deposition at the
target position.
Fig. 1 Bragg curves of 3 He ions with initial energy E = 207.92 MeV incident on graphite (left)
and water–gelatine (right)
One clearly observes in the case of graphite a considerable difference in the two
energy deposition predictions, which by the way refer to electromagnetic processes
and not nuclear ones. We attribute this to the fact that the graphite structure depends
strongly on the physical conditions of the material production (diamond is an ob-
vious example) which shall enter in the material properties. For the gelatine-water
mixture both simulations coincide. As mentioned before, some of the produced nu-
clear fragments are β + emitters which permit to monitor the position of the Bragg
peak that should match the target location (tumor). From the comparison of the
Bragg curves (Fig. 1) and the activity of β + emitters (Fig. 2) one observes that the
20 L.N. Burigo et al.
Fig. 2 Activity curves of β + emitters in graphite (left) and in water–gelatine (right) produced by
3 He
Fig. 3 Energy deposition of fragments produced by 3 He ions in graphite (left) and in water–
gelatine (right)
peak positions are the same. However, again for graphite the two simulations yield
considerable differences in the simulated distributions.
In Fig. 3 we show the energy deposition by the fragments boron, beryllium, he-
lium and lithium, produced in particle collisions from nuclei of the target material.
A comparison to the Bragg curve shows that nuclear contributions to the energy
deposition is of the order of 101 % and thus not negligible.
Simulation fidelity may now be implemented using a self-consistency criterion.
Recalling that heavy-ion transport was implemented by a physical Monte Carlo fol-
lowing the instructions of the Boltzmann equation. In other words an ensemble of
particles together with microscopic processes mimics what we believe that happens
in the real world. In parallel to particle propagation tallies are recorded that par-
tially are used to show the depth versus energy or versus activity profiles and on the
other hand allow for particle counting at a certain position, with velocity in a certain
direction and with a specified energy. Note that the present simulation assumes a sta-
tionary case so that no explicit time dependence appears in the simulation data. The
afore mentioned self-consistency criterion makes use of the same transport equation
together with the tallies and further the Bethe–Bloch approach to invert the problem
A Self-Consistent Monte Carlo Validation Procedure 21
and solve for the nuclear cross sections by the use of a mathematical Monte Carlo.
We are aware of the fact that there exists a collection of solvers for inverse problems
which we could have employed here, but for simplicity we used a sample of cross
section values for specific energies which were accepted or rejected according to a
probability factor (in the spirit of a Boltzmann factor) that is defined in terms of the
distance between the left and right hand side of the Boltzmann equation averaged
over all energies (this procedure is incremental). Since 106 particles is not sufficient
to get good statistics for each fragment of the nuclear reaction we determine the
total nuclear cross section σt for all nuclear processes and averaged over the par-
ticle energies. For graphite we get the σt = 1.22b with GEANT4 version 9.3 and
σt = 1.11b with version 8.2. The values for water plus gelatine are σt = 1.57b and
σt = 1.59b with GEANT4 version 9.3 and with version 8.2, respectively.
5 Conclusion
present according to their respective life times, manifest in their decay curves. As a
new aspect of time dependence it will also be possible to extend the simulation with
time dependent geometries that are up to now an unsolved challenge in radiotherapy
dose calculations. These extensions are necessary because of motion of the patient
or its body, which introduce new uncertainties into the treatment and simulation
[Din03], i.e. positioning errors (different positions during treatment compared to
the diagnostic image), inter-fraction organ motion (e.g. digestive system), and intra-
fraction organ movement (e.g. respiration-induced motion). The new aspect to be
faced in future works are complex boundary and interface conditions besides the
realistic time structure of the beam. Although we believe to have made progress in
improving the GEANT4 based simulation engine, however, in order to turn it an
instrument for standard clinical use there is still much to be done.
References
[Ago03] Agostinelli, S., Allison, J., Amako, K. et al.: GEANT4—a simulation toolkit. Nucl. Instr.
Meth. Phys. Res. A, 506, 250–303 (2003).
[All06] Allison, J. et al.: Geant developments and applications. IEEE Trans. Nucl. Sci., 53, n. 1,
Part 2, 270–278 (2006).
[Cyg05] Cygler, J.E., Lochrin, C., Daskalov, G.M., Howard, M., Zohr, R., Esche, B., Eapen, L.,
Grimard, L., Caudrelier, J.M.: Clinical use of a commercial Monte Carlo treatment plan-
ning system for electron beams. Phys. Med. Biol., 50, 1029–1034 (2005).
[Deb00] Debus, J., Haberer, T., Schulze-Ertner, D., Jäkel, O., Wenz, F., Enghardt, W., Schlegel,
W., Kraft, G., Wannenmacher, M.: Fractionated carbon ion irradiation of skull base tu-
mors at GSI. First clinical results and future perspectives. Strahlenther. Onkol., 176,
211–216 (2000).
[Din03] Ding, M., Li, J., Deng, J., Fourkai, E., Ma, C.-M.: Dose correlation for thoracic motion
in radiation therapy of breast cancer. Med. Phys., 30, 2520–2529 (2003).
[Dur08] Durante, M.: Focus on heavy ions in biophysics and medical physics. New J. Phys., 10,
075002 (2008).
[Dur10] Durante, M., Loeffler, J.S.: Charged particles in radiation oncology. Nat. Rev. Clin. On-
col., 7, 37–43 (2010).
[Fra03] Fraass, B.A., Smathers, J., Deye, J.: Summary and recommendations of a National Can-
cer Institute workshop on issues limiting the clinical use of Monte Carlo calculation al-
gorithms for megavoltage external beam radiation therapy. Med. Phys., 30, 3206–3216
(2003).
[Gad90] Gademann, G., Hartmann, G.H., Kraft, G., Lorenz, W.-J., Wannenmacher M.: The med-
ical heavy ion therapy project at the Gesellschaft für Schwerionenforschung facility in
Darmstadt. Strahlenther. Onkol., 166, 34–39 (1990).
[Jer02] Jeraj, R., Keall, P.J., Siebers, J.V.: The effect of dose calculation accuracy on inverse
treatment planning. Phys. Med. Biol., 47, 391–407 (2002).
[Kea01] Keall, P.J., Siebers, J.V., Arnfield, M., Kim, J.O., Mohan, R.: Monte Carlo dose calcula-
tions for dynamic IMRT treatments. Phys. Med. Biol., 46, 929–941 (2001).
[Kra90] Kraft, G.: Radiobiological and physical basis for radiotherapy with protons and heavier
ions. Strahlenther. Onkol., 166, 10–13 (1990).
[Kra91] Kraft, G., Becher, W., Blasche, K., Böhne, D., Fischer, B., Gademann, G., Geissel, H.,
Haberer, Th., Klabunde, J., Kraft-Weyrather, W., Langenbeck, B., Mänzenberg, G., Rit-
ter, S., Rösch, W., Schardt, D., Stelzer, H., Schwab, Th.: The heavy ion therapy project
at GSI. Nucl. Tracks Radiat. Meas., 19, 911 (1991).
A Self-Consistent Monte Carlo Validation Procedure 23
[Kra94] Kraft, G.: Heavy-ion therapy at GSI. Europhys. News, 25, n. 4, 81 (1994).
[Kra95] Kraft, G., Arndt, U., Becher, W., Schardt, D., Stelzer, H., Weber, U.: Heavy ion therapy
at GSI. Th. Archinal Nucl. Instr. and Meth. in Phys. Res. A, 367, 66 (1995).
[Liu01] Liu, H.H., Verhaegen, F., Dong, L.: A method of simulating dynamic multileaf collima-
tors using Monte Carlo techniques for intensity-modulated radiation therapy. Phys. Med.
Biol., 46, 2283–2289 (2001).
[Moh01] Mohan, R., Antolak, J., Hendee, W.R.: Monte Carlo techniques should replace analyti-
cal methods for estimating dose distributions in radiotherapy treatment planning. Med.
Phys., 28, 123–126 (2001).
[Psh05] Pshenichnov, I.A., Mishustin, I.N., Greiner, W.: Neutrons from fragmentation of light
nuclei in tissue-like media: a study with GEANT4 toolkit. Phys. Med. Biol., 50, 5493–
5507 (2005).
[Psh06] Pshenichnov, I.A., Mishustin, I.N., Greiner, W.: Distributions of positron-emitting nuclei
in proton and carbon-ion therapy studied with GEANT4. Phys. Med. Biol., 51, 6099–
6112 (2006).
[Sch10] Schardt, D., Elsässer, T., Schulz-Ertner, D.: Heavy-ion tumor therapy: physical and ra-
diobiological benefits. Rev. Mod. Phys., 82, 383–425 (2010).
[Ver01] Verhaegen, F., Liu, H.H.: Incorporating dynamic collimator motion in Monte Carlo sim-
ulations: an application in modelling a dynamic wedge. Phys. Med. Biol., 46, 287–296
(2001).
A General Analytical Solution
of the Advection–Diffusion Equation
for Fickian Closure
1 Introduction
In the last few years there has been increased research interest in searching for
analytical solutions for the advection–diffusion equation (ADE). By analytical we
mean that no approximation is done along the derivation of the solution. There exists
a significant literature regarding this theme. For illustration we mention the works
of [Rou55, Smi57, ScFi75, Dem78, Van78, NiHa81, Tag85, Tir89, TiRi94, Sha96,
LiHi97, Tir03]. We note that in these works all solutions are valid for very special-
ized problems having specific wind and eddy diffusivities vertical profiles. Further,
also in the literature there is the ADMM (Advection Diffusion Multilayer Method)
approach which solves the two-dimensional ADE with variable wind profile and
eddy diffusivity coefficient [Mor06]. The main idea relies on the discretization of
the Atmospheric Boundary Layer (ABL) in a multilayer domain, assuming in each
layer that the eddy diffusivity and wind profile take averaged values. The resulting
advection–diffusion equation in each layer is then solved by the Laplace transfor-
mation technique. For more details about this methodology see the review work
done by [Mor06]. We are also aware of the recent work of [Cos06], dubbed as GI-
ADMT method (Generalized Integral Advection Diffusion Multilayer Technique),
D. Buske
Universidade Federal de Pelotas, RS, Brazil,
e-mail: daniela.buske@ufpel.edu.br
M.T. Vilhena
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: vilhena@pq.cnpq.br
C.F. Segatto
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: csegatto@pq.cnpq.br
R.S. Quadros
Universidade Federal de Pelotas, RS, Brazil,
e-mail: regis.quadros@ufpel.edu.br
In the following we derive the ADE for the simulation of pollutant releasing in the
ABL assuming Fickian closure of the turbulence. We must recall that this equation
is derived combining the continuity equation ruled by the conservation law with the
Fickian closure of turbulence. Indeed, we write the ADE in Cartesian geometry as
in [Bla97]:
∂c ∂c ∂c ∂c ∂ ∂c ∂ ∂c ∂ ∂c
+u +v +w = Kx + Ky + Kz (1)
∂t ∂x ∂y ∂z ∂x ∂x ∂y ∂y ∂z ∂z
∂c
Kz = 0 at z = 0, h, (2)
∂z
∂c
Ky = 0 at y = 0, Ly , (3)
∂y
∂c
Kx = 0 at x = 0, Lx , (4)
∂x
c(x, y, z, 0) = 0. (5)
where Ym (y) are a set of orthogonal eigenfunctions, given by Ym (z) = cos(λm y), and
λm = mπ /Ly (m = 0, 1, 2, . . .) are, respectively, the set of eigenvalues.
To determine the unknown coefficient cm (x, z,t) for m = 0 : M we began recasting
(1) by applying the chain rule for the diffusion terms. Substituting (7) in the resulting
L
equation and taking moments, meaning applying the operator 0 y ()Yn (y), we obtain
Ly Ly
M
∂ cm (x, z,t) ∂ cm (x, z,t)
∑ −
∂t 0
Ym (y)Yn (y)dy, −u
∂x 0
Ym (y)Yn (y)dy
m=0
Ly Ly
∂ cm (x, z,t)
− vcm (x, z,t) Ym (y)Yn (y)dy − w Ym (y)Yn (y)dy
0 ∂z 0
Ly Ly
∂ 2 cm (x, z,t) ∂ cm (x, z,t)
+ Kx Ym (y)Yn (y)dy + Kx Ym (y)Yn (y)dy
∂ x2 0 ∂x 0
Ly Ly
− λm2 cm (x, z,t) KyYm (y)Yn (y)dy + cm (x, z,t) Ky Ym (y)Yn (y)dy
0 0
Ly Ly
∂ 2 cm (x, z,t) ∂ cm (x, z,t)
+ Kz Ym (y)Yn (y)dy + Kz Ym (y)Yn (y)dy = 0. (8)
∂ z2 0 ∂z 0
The problem (10) is then solved by the GILTT method. Following the works of
[Mor09a, Mor09b] and taking advantage of the well-known solution for the station-
ary problem with advection in the x direction, we apply the Laplace transformation
technique in the t variable (t → r), obtaining the following steady-state problem:
∂ Cm (x, z, r) ∂ ∂ Cm (x, z, r)
rCm (x, z, r) + u = Kz − λm2 Ky Cm (x, z, r). (11)
∂x ∂z ∂z
A General Analytical Solution of the ADE 29
where ζl (z) are a set of orthogonal eigenfunctions, given by ζl (z) = cos(λl z), and
λl = l π /h, l = 0, 1, 2, . . . is the set of eigenvalues.
Substituting (12) into (11) and taking moments, we get the following first order
matrix differential equation:
dYm
(x, r) + G.Ym (x, r) = 0, (13)
dx
for m = 0 : M, where Ym (x, r) is the column vector whose components are {Cm,l (x, r)}
for l = 0 : L. The matrix G is defined as G = B−11 B2 . The entries of matrices B1 and
B2 are given by
h
(b1 )l, j = − u ζl (z) ζ j (z)dz
0
and
h h
(b2 )l, j = Kz ζl (z) ζ j (z)dz − λl2 Kz ζl (z)ζ j (z)dz
0 0
h
−(r + λl2 Ky ) ζl (z) ζ j (z)dz,
0
respectively.
As in the work of [Mor09b], we solve problem (13) obtaining the following so-
lution:
1 L γ +i∞
cm (x, z,t) = ∑ γ −i∞ Cm,l (x, r) ζl (z)ert dr.
2π i l=0
(14)
To overcome the drawback of evaluating the line integral appearing in the above
solution, we perform the calculation of this integral by the Gaussian quadrature
scheme, namely
K
pk pk
cm (x, z,t) = ∑ ak Cm x, z, , (15)
k=0 t t
where ak and pk are, respectively, the weights and nodes of the Gaussian quadrature
scheme [StSe66]. Regarding the issue of the adopted Laplace numerical inversion
scheme, it is important to mention that this approach is exact if the integrand is
a polynomial of degree 2M − 1 in the variable. We are aware of the existence in
the literature of methods to invert numerically the Laplace transformed functions
[VaAb04, AbVa04], but we restrict our attention in the problem considered to the
Gaussian quadrature scheme. The motivation for this choice comes not only from
the simplicity of the scheme but also the good results achieved by it.
30 D. Buske et al.
In the following we report the parameterizations adopted for the simulations re-
ported in this chapter. We need to recall that the choice of the turbulent parameter-
ization represents a fundamental aspect for pollutant dispersion modeling [Mor05].
In terms of the convective scaling parameters, the vertical eddy diffusivity can be
well formulated by the algebraic formulation proposed by [Deg02]:
where n is the non-dimensional frequency, cv,w = 0.36, cu = 0.3, and ( fm∗ )i is the
normalized frequency of the spectral peak independent of the stratification, with
( fm∗ )u = 0.67 for the longitudinal component and
z
4z
−1
8z
( fm∗ )w = 0.55 1 − exp − − 0.0003 exp
h h h
for the vertical component. Notice that these eddy diffusivities are functions of not
only turbulence but also of distance from the source [Ary95].
On the other hand, we assume that the wind profile is described by a power law,
namely [PaDu88]
n
uz z
= (17)
u1 z1
where uz and u1 are the mean wind velocity at the heights z and z1 , while n = 0.1
under unstable conditions.
In order to illustrate the aptness of the discussed formulation to simulate contam-
inant dispersion in the ABL, we evaluate the performance of the proposed solution
against experimental ground-level concentration. The experimental data set used to
evaluate the performance of the model in unstable conditions were carried out in the
northern part of Copenhagen [GrLy84]. It consists of tracer without buoyancy re-
leased from a tower at a height of 115 m, and collection of the tracer sampling units
at the ground-level positions at the maximum of three crosswind arcs. The sampling
units were positioned at two to six kilometers from the point of releasing. The site
has a roughness length of 0.6 m.
4 Numerical Results
Next we display the numerical results attained by this methodology and proceed a
statistical comparison against the GILTT two-dimensional results assuming Gaus-
sian solution in the y-direction (named here as GGILTT), as well the GIADMT
results [Cos06] and experimental data set of Copenhagen experiment. We begin, for
A General Analytical Solution of the ADE 31
Table 1 Statistical comparison between models using the Copenhagen data set
NMSE COR FA2 FB FS
GILTTG 0.33 0.80 0.87 0.28 0.09
GIADMT 0.15 0.87 0.96 0.01 −0.09
3D-GILTT 0.07 0.93 0.96 0.02 0.03
5 Conclusions
From the previous discussion, we promptly realize that the general analytical solu-
tion encountered for the time-dependent, three-dimensional ADE assuming Fickian
closure for the turbulence is an important and promising formulation to simulate
contaminant released in the ABL. In fact, besides the good results achieved, it is
relevant to reinforce the analytical character of this solution, in the sense that no
approximation is done along the solution derivation, except for the line integral ap-
proximation of the Laplace inversion of the transformed solution by the Gaussian
quadrature scheme. By general solution we mean that the solution is valid for wind
32 D. Buske et al.
Fig. 1 Scatter diagram of the observed vs predicted maximum ground-level concentrations nor-
malized by the emission rate. Data between lines correspond to a factor of two
profile and eddy diffusivity coefficient being arbitrary functions of the x and z vari-
ables. The generality also extends for the issue of atmospheric stability once the
solution is valid for any stability. Keeping in mind that the problem fulfills the hy-
pothesis of the Cauchy–Kowalewski theorem [CoHi89], we are certain to affirm that
the problem studied has a solution and its solution is unique. In addition, it is nec-
essary to underline that this sort of equation models problems in several areas of
science, and therefore this solution has a broad class of applications, for instance,
we mention the work of [CoBa07] in the area of pollutant dispersion in channels.
For the purpose of numerical computations one may make use of the Cardinal
Theorem of Interpolation Theory that allows the calculation to be made of the so-
lution for any prescribed accuracy by controlling the number of terms in the series
summation. Details of this procedure are reported elsewhere [Bod10].
Finally, we notice that because of the reduction of order of ADE reported, we are
confident to stress that the form of solution does not depend on the topology of the
problem, once the solution for the two- and three-dimensional problems are written
in terms of the same set of eigenvalues.
To extend the application of the proposed solution to more realistic problems
we shall focus our future attention to the task of extending this solution to prob-
lems requiring low wind as well time-dependent wind profile and eddy diffusivity
coefficient.
A General Analytical Solution of the ADE 33
References
[AbVa04] Abate, J., Valkó, P.P.: Multi-precision Laplace transform inversion. Int. J. for Num.
Methods in Engineering, 60, 979–993 (2004).
[Ary95] Arya, S. Pal: Modeling and parameterization of near-source diffusion in weak winds.
J. Appl. Meteor., 34, 1112–1122 (1995).
[Bla97] Blackadar, A.K.: Turbulence and Diffusion in the Atmosphere: Lectures in Environ-
mental Sciences, Springer-Verlag, 185 pp. (1997).
[Bod10] Bodmann, B., Vilhena, M.T., Ferreira, L.S., Bardaji, J.B.: An analytical solver for the
multi-group two-dimensional neutron-diffusion equation by integral transform tech-
niques. Il Nuovo Cimento C, 33, 199–206 (2010).
[Bus10] Buske, D., Vilhena, M.T., Moreira, D.M., Tirabassi, T.: An Analytical Solution for the
Transient Two-Dimensional Advection-Diffusion Equation with Non-Fickian Closure
in Cartesian Geometry by Integral Transform Technique. Integral Methods in Science
and Engineering: Computational Methods, Birkhauser, Boston, 33–40 (2010).
[Cos06] Costa, C.P., Vilhena, M.T., Moreira, D.M., Tirabassi, T.: Semi-analytical solution of the
steady three-dimensional advection-diffusion equation in the planetary boundary layer.
Atmos. Environ., 40, n. 29, 5659–5669 (2006).
[CoBa07] Cotta, R.M., Barros, F.P.J.: Integral transforms for three-dimensional steady turbulent
dispersion in rivers and channels. Applied Mathematical Modelling, 31, 2719–2732
(2007).
[CoHi89] Courant, R., Hilbert, D.: Methods of Mathematical Physics, John Wiley & Sons, New
York (1989).
[Deg02] Degrazia, G.A., Moreira, D.M., Campos, C.R.J., Carvalho, J.C., Vilhena, M.T.: Com-
parison between an integral and algebraic formulation for the eddy diffusivity using the
Copenhagen experimental dataset. Il Nuovo Cimento, 25C, 207–218 (2002).
[Dem78] Demuth, C.: A contribution to the analytical steady solution of the diffusion equation
for line sources. Atmos. Environ., 12, 1255–1258 (1978).
[GrLy84] Gryning, S.E., Lyck, E.: Atmospheric dispersion from elevated source in an urban area:
comparison between tracer experiments and model calculations. J. Appl. Meteor., 23,
651–654 (1984).
[Han89] Hanna, S.R.: Confidence limit for air quality models as estimated by bootstrap and
jacknife resampling methods. Atm. Env., 23, 1385–1395 (1989).
[LiHi97] Lin, J.S., Hildemann, L.M.: A generalised mathematical scheme to analytically solve
the atmospheric diffusion equation with dry deposition. Atm. Env., 31, 59–71 (1997).
[Mor05] Moreira, D.M., Vilhena, M.T., Tirabassi, T., Buske, D., Cotta, R.M.: Near source atmo-
spheric pollutant dispersion using the new GILTT method. Atmos. Environ., 39, n. 34,
6290–6295 (2005).
[Mor06] Moreira, D.M., Vilhena, M.T., Tirabassi, T., Costa, C., Bodmann, B.: Simulation of
pollutant dispersion in atmosphere by the Laplace transform: the ADMM approach.
Water, Air and Soil Pollution, 177, 411–439 (2006).
[Mor09a] Moreira, D.M., Vilhena, M.T., Buske, D., Tirabassi, T.: The state-of-art of the GILTT
method to simulate pollutant dispersion in the atmosphere. Atmospheric Research, 92,
1–17 (2009).
[Mor09b] Moreira, D.M., Vilhena, M.T., Buske, D.: On the GILTT Formulation for Pollutant Dis-
persion Simulation in the Atmospheric Boundary Layer. Air Pollution and Turbulence:
Modeling and Applications, CRC Press, Boca Raton – Flórida (USA), 179–202 (2009).
34 D. Buske et al.
[Mor10] Moreira, D.M., Vilhena, M.T., Tirabassi, T., Buske, D., Costa, C.P.: Comparison be-
tween analytical models to simulate pollutant dispersion in the atmosphere. Int. J. Env.
and Waste Management, 6, 327–344 (2010).
[NiHa81] Nieuwstadt, F.T.M., de Haan, B.J.: An analytical solution of one-dimensional diffu-
sion equation in a nonstationary boundary layer with an application to inversion rise
fumigation. Atmos. Environ., 15, 845–851 (1981).
[PaDu88] Panofsky, A.H., Dutton, J.A.: Atmospheric Turbulence, John Wiley & Sons, New York
(1988).
[ScFi75] Scriven, R.A., Fisher, B.A.: The long range transport of airborne material and its re-
moval by deposition and washout-II. The effect of turbulent diffusion. Atmos. Environ.,
9, 59–69 (1975).
[Rou55] Rounds, W.: Solutions of the two-dimensional diffusion equation. Trans. Am. Geophys.
Union, 36, 395–405 (1955).
[Sha96] Sharan, M., Singh, M.P., Yadav, A.K.: A mathematical model for the atmospheric dis-
persion in low winds with eddy diffusivities as linear functions of downwind distance.
Atmos. Environ., 30, n. 7, 1137–1145 (1996).
[Smi57] Smith, F.B.: The diffusion of smoke from a continuous elevated poinr source into a
turbulent atmosphere. J. Fluid Mech., 2, 49–76 (1957).
[StSe66] Stroud, A. H., Secrest, D.: Gaussian Quadrature Formulas, Prentice Hall Inc., Engle-
wood Cliffs, N.J. (1966).
[VaAb04] Valkó, P.P., Abate, J.: Comparison of sequence accelerators for the Gaver method of
numerical Laplace transform inversion. Computers and Mathematics with Application,
48, 629–636 (2004).
[Tag85] Tagliazucca, M., Nanni, T., Tirabassi, T.: An analytical dispersion model for sources in
the surface layer. Nuovo Cimento, 8C, 771–781 (1985).
[Tir89] Tirabassi, T.: Analytical air pollution and diffusion models. Water, Air and Soil Pollu-
tion, 47, 19–24 (1989).
[TiRi94] Tirabassi, T., Rizza U.: Applied dispersion modelling for ground-level concentrations
from elevated sources. Atmos. Environ., 28, 611–615 (1994).
[Tir03] Tirabassi, T.: Operational advanced air pollution modeling. PAGEOPH, 160, n. 1–2,
5–16 (2003).
[Van78] van Ulden, A.P.: Simple estimates for vertical diffusion from sources near the ground.
Atmos. Environ., 12, 2125–2129 (1978).
A Novel Method for Simulating Spectral Nuclear
Reactor Criticality by a Spatially Dependent
Volume Size Control
1 Introduction
Controlled nuclear reactions in a nuclear reactor are one of the energy resources that
may contribute to attend the increasing energy demand while minimizing impact on
the environment. Because of its efficient energy release per nuclear reaction in com-
parison to processes that involve chemical reactions for instance (which differ by
more than eight orders in magnitude) reactor control and safety is a crucial issue.
Evidently, while designing new reactor conceptions or operating existing reactors
the microscopic as well as macroscopic response of the nuclear process must be
understood in detail and described adequately in terms of mathematical models to-
gether with experimental data such as the nuclear reaction cross sections [Sek07].
The physics of the nuclear reactions taking place in a power reactor and its influence
on the neutron flux by perturbations from inside or outside the system are known
reasonably well. Nevertheless, as the present contribution will show there is still
space for progress which is manifest in a variety of recent attempts to create effi-
cient and adequate algorithms that calculate neutron fluxes, as well as other reactor
relevant quantities.
D.Q. de Camargo
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: dyanadecamargo@gmail.com
B.E.J. Bodmann
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: bardo.bodmann@ufrgs.br
M.T. Vilhena
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: vilhena@pq.cnpq.br
S.d.Q.B. Leite
Comissão Nacional de Energia Nuclear, Rio de Janeiro, RJ, Brazil,
e-mail: bogado@cnen.gov.br
Operating a nuclear reactor involves two classes of problems, the steady state
where power output is mainly constant, and power changing actions like start-up,
shut-down or changes in the power level. While in steady state condition a reactor
is operating at or near criticality, i.e. the neutron multiplication factor k is nearly
equal to unity, power changes are characterized by the reactivity ρ = (k − 1)/k.
Note that the reactivity is a convenient mathematically defined quantity, but cannot
be measured directly in practice. Nevertheless, neutron multiplication or reactivity
depends on several variables such as the temperature of the nuclear core, its chem-
ical composition and, as the following discussion will show, on energy. Usually the
multiplication factor and reactivity are used over the whole energy range. Especially
in situations of power change the neutron energy spectrum changes and so do phys-
ical quantities such as the average cross sections. Hence in order to model theses
situations a spectral description of k and consequently ρ is of need, which is the
focus of the present contribution.
In principle one may approach the problem using analytical, numerical or stochas-
tic methods [Bod10, Cam09, Lep05]. Since the energy dependence of the cross sec-
tions shows oscillations in the regions of the resonances, analytical and numerical
implementations of neutron transport equations are less adequate because they usu-
ally work with energy groups that average the cross sections over the respective
energy interval and thus would need mechanisms to correct spectral changes. From
this point of view one seems to be better off using the Monte Carlo method but
facing the task of simulating an exorbitant number of neutrons, which is in practice
impossible and has to be circumvented otherwise. In the present work we solve this
problem by a spatially dependent control volume size.
Our Monte Carlo simulation is a physical implementation using an interactive
evaluation of a deterministic model based on randomly distributed numbers accord-
ing to specified probability densities. This technique turns effective when the model
is complex and nonlinear, or when it involves an elevated number of conditions
(such as geometry with boundaries and interfaces, change of the chemical com-
position of the materials) and when the integral involves many dimensions. In our
particle transport code, the Monte Carlo technique tallies each of the particles along
its trajectories until some terminal event such as absorption or escape, amongst oth-
ers, terminates the history. In this work, the Monte Carlo implementation takes its
instructions from an integro-differential neutron transport equation in three dimen-
sions and genuine energy dependence (i.e. no energy groups). The treatment of the
cross section is one aspect that differentiates this work from others known in the
literature [Bod10, Gon10, Vil08]. Here cross sections are continuous functions of
energy, which are obtained by parametrizations and coded as program procedures.
The interaction type that a neutron will suffer and the characteristics of their dis-
placement in the environment are randomly estimated by the use of the relevant
probability distributions.
In order to render the simulation effective, the environment is divided in several
smaller volumes, where one of these volumes is chosen and plays the role of a con-
trol volume. These small volume elements are to be chosen to reconstruct a smooth
neutron distribution for the whole volume. The volume control implementation in
A Novel Method for Simulating Spectral Nuclear Reactor Criticality 37
the environment is a novel aspect in comparison to existing works and makes use of
a fixed initial number of neutrons which are inserted in a specified control volume.
Its advantage is that the neutron number may be significantly reduced while still
obtaining a good simulation. Varying the control volume is further used to simu-
late processes with different criticality. So far we show how the method works and
leave for the future a detailed examination for more realistic cases with respect to
geometries and material compositions, necessary for nuclear reactors physics.
2 Neutron Transport
Usual calculations from neutron transport express the spatial neutron population in
terms of the neutron flux, which is related to the neutron density by the velocity
modulus (Φ (r, E,t) = vn(r, E,t)). Instead, here we use the so-called neutron angular
density n(r, Ω , E,t), i.e. the neutron number per unit-volume at location r and at
instant t, moving in direction of the oriented unit-solid angle Ω with energy E. For
criticality analysis usually the neutron number (integration of the angular neutron
density over the solid angle, the energy and the volume of interest) at instant t + τNLC
and t are compared, where τNLC is the time of the neutron life cycle [CNSC03]. In
our case analyses integrate only over the solid angle and volume, but maintain the
spectral dependence.
The neutron transport equations is derived from the Boltzmann transport equation
[Vli08].
1∂
Φ (r, Ω , E,t) + Ω · ∇Φ (r, Ω , E,t) + Σt (r, E,t)Φ (r, Ω , E,t)
v ∂t
= q(r, Ω , E,t).
Here Σt (r, E,t) is the macroscopic total cross section at time t and location r; and
energy E and q(r, Ω , E,t) represents all contributions of neutrons that appear at
instant and position (t, r) and moving in the direction Ω with energy E:
∞
q(r, Ω , E,t) = dE d Ω Σ (r, E → E, Ω → Ω )Φ (r, Ω , E ,t)
0 4π
+S(r, Ω , E,t). (1)
Here S(r, Ω , E,t) represents the effect of an external neutron source. The integral
term contains contributions from the nuclear reactions, such as neutron scattering
and fission. Hence, the macroscopic cross section Σ (r, E → E, Ω → Ω ) dE d Ω
represents the number of neutrons emitted in the infinitesimal energy interval dE at
E and into the infinitesimal solid angle d Ω at Ω that have come from a reaction of
a neutron at energy E and incoming direction Ω . If importance of each individual
nuclear reaction is taken into account the cross section may be decomposed as
38 D.Q. de Camargo et al.
with νk (E) is the number of secondary neutrons released by reaction of type k (i.e.
for elastic and inelastic scattering νs = 1, for knock out (n, 2n) reactions νko = 2;
and for fission νk ∈ [2, . . . , 4]), with Σk (r, E) is the macroscopic cross section for
reaction k. pk (E → E, Ω → Ω ) dE d Ω is the probability that a neutron moving in
incoming direction Ω with energy E undergoes reaction k, with outgoing neutrons
with energy in [E, E + dE] in d Ω with director Ω .
In our Monte Carlo simulation we used the fission spectrum χk (E) in order to im-
plement the associated probabilities,
1
pk (E → E, Ω → Ω ) = χk (E). (3)
4π
Furthermore, we simplified (1) and did not consider any external source (S(r, Ω ,
E,t) ≡ 0). We also admitted no changes in the chemical composition of the nuclear
fuel or the moderator.
As already mentioned before, the present approach considers the full energy de-
pendence in the cross sections in analytical form. To this end the cross sections
obtained from nuclear data-bases were parametrized for energy intervals that per-
mitted an accurate description ( 1%) of the energy dependence. The results of
some of these parametrizations are shown in Figs. 1–5, which show the absorption
cross sections for 16 O and 235 U, the scattering cross sections for 16 O and 235 U and
the fission cross section for 235 U in comparison to the original data [Cha06]. Note
the spurious differences between the two. The preference of using parametrizations
instead of a data-base is justified by various tests that have shown us that in the
present simulation using interpolated data to get the energy corresponding to the
cross section is slower in execution time than a single function call.
The control over the chain reaction, that is, to control the number of neutrons per
energy interval in a generation relative to the number of neutrons in the previous
one, is established by the control volume. Since the number of simulated neutrons
in a volume of interest is always the same (here 106 ), adjusting the volume size con-
trols whether the multiplication factor is below or above unity. Note that simulating
criticality k = 1 is not a trivial task because one has to find the corresponding vol-
ume size, which may be obtained only by simulating super-critical and sub-critical
setups with subsequent extrapolation. Moreover, it is tedious to show that a reactor
can be critical at any (reasonable) power level.
The key to this control is the neutron life cycle. Conventionally, this operation
is done using six factors that govern the production, leakage and absorption of neu-
trons, and enable the quantitative description of the components that govern the mul-
A Novel Method for Simulating Spectral Nuclear Reactor Criticality 39
Fig. 1 Absorption cross section for 16 O in the energy range 10−11 MeV ≤ E ≤ 2 × 102 MeV,
original data (grey) and parametrization (black)
Fig. 2 Absorption cross section for 235 U in the energy range 10−11 MeV ≤ E ≤ 2 × 102 MeV,
original data (grey) and parametrization (black)
tiplication factor in the neutron cycle. Since our simulation contains the full energy
dependence, this may be extended to a spectral (i.e. energy-dependent) description
for neutron multiplication. More specifically, the relevant factors for neutron multi-
plication that are fast fission, fast leakage escape, resonance escape probability, ther-
mal non-leakage, thermal utilization and reproduction by thermal neutrons, have in
this treatment an explicit energy dependence.
The steps of the simulation are as follows:
1. The material compositions are defined, which may be time dependent. In the
present case they are assumed as constant and geometries are considered static.
40 D.Q. de Camargo et al.
Fig. 3 Scattering cross section for 16 O in the energy range 10−11 MeV ≤ E ≤ 2 × 102 MeV, orig-
inal data (grey) and parametrization (black)
Fig. 4 Scattering cross section for 235 U in the energy range 10−11 MeV ≤ E ≤ 2 × 102 MeV,
original data (grey) and parametrization (black)
2. Volume control is implemented as to scale the neutron flux in terms of the neutron
density.
3. The initial neutron energy spectrum determines the probability of neutrons that
appear with a given energy.
4. The position of the generated neutron with initial energy E is determined ran-
domly at position r within the active volume.
5. The energy value determines the cross sections and thus the scheme to score the
reaction type.
6. Once the reaction is determined the length of the trajectory, angles and the new
position are computed.
A Novel Method for Simulating Spectral Nuclear Reactor Criticality 41
Fig. 5 Fission cross section for 235 U in the energy range 10−11 MeV ≤ E ≤ 2 × 102 MeV, original
data (grey) and parametrization (black)
7. All the previous steps have been carried out without considering boundary con-
ditions, manifest in leakage, so that a decision re-directs the program to
a. either terminate the history of a neutron and restart at step 4;
b. or to determine the new interaction vertex.
8. According to the type of interaction the procedures are as follows:
a. For capture, the history of the neutron ends, and the program restarts at step
4;
b. For scattering, the new neutron energy is determined, and the program contin-
ues at step 6;
c. For fission, the number of new neutrons is determined randomly. For each
neutron the program starts at step 4.
4 Results
The Monte Carlo step so far is a quantity that may be related to a time scale by cali-
bration. A natural time scale is the neutron generation time, i.e. the time required for
neutrons from one generation to cause the fissions that produce the next generation
of neutrons. This time interval is determined by three time intervals, the time it takes
a fast neutron to slow down to thermal energy, the time the thermal neutron exists
prior to absorption in the fuel, and the time for a fissionable nucleus to emit a fast
neutron after neutron absorption. Fast neutrons may either slow down to thermal en-
ergies or leak out of the reactor, which happens in a time interval of 10−4 s–10−6 s
depending on the reactor core architecture including the moderator. Fission and fast
42 D.Q. de Camargo et al.
neutron production after neutron absorption in the nuclear fuel takes about 10−13 s
[CNSC03].
One has to consider further the neutron precursors, which we did not implement
yet into the simulation. Also here three time intervals contribute to the time scale,
the time it takes a fast neutron to slow down to thermal energy, the time the thermal
neutron bounces around before absorption, and the average time from neutron ab-
sorption to neutron emission by the precursors, which are typically organized in six
groups. The average time for decay of precursors from 235 U is tP = 12.5 s and the
average generation time is then typically of the order of magnitude τNLC ∼ 100 ms.
Since in the simulation the neutrons have a time stamp, this permits us to analyze
such a cycle. In Fig. 6 we show the cyclic behavior, which does not appear in this
form in a real reactor, but is an artifice which may be extracted from the tallies of
individual neutrons. Figure 7 shows the energy spectrum after a sequence of Monte
Carlo steps. One clearly observes more pronounced changes in the higher energy
region of the spectrum. As a novel result we show the spectral multiplication factor
change with Monte Carlo steps (Fig. 8).
5 Conclusion
In the present work we implemented a Monte Carlo simulation for neutron transport
in nuclear power reactors. Using a approach which is different from most of those
in the literature we consider three spatial dimensions and continuous energy depen-
dence, instead of the usual energy groups. To this end we supply functions with the
cross sections in parametrized and analytical form. This allows a detailed analysis
of the energy influence on the neutron population and its impact on the neutron
life cycle. As novel results we showed the changes in the neutron energy spectrum
with the steps of the simulation and further determined the spectral changes in the
neutron multiplication factor per Monte Carlo step.
A Novel Method for Simulating Spectral Nuclear Reactor Criticality 43
Fig. 7 Changes in the neutron energy spectrum after a sequence of Monte Carlo steps
The simulation uses a small sample volume in order to reproduce the neutron
population in the whole volume, which permits to simulate huge neutron numbers in
a macroscopic volume, as in the nuclear reactor core. Variation of the control volume
is further used to simulate processes with different criticality. Power changes are
usually expressed using reactivity, which in general makes use of negative feedback
(as temperature coefficients of reactivity, for instance) to control the nuclear reactor
and make it inherently self-controlling and thus safe. Results from our simulation
open new pathways the allow to harness the reactors behavior by a new degree of
freedom, i.e. its spectral behavior along the neutron life cycle.
Since there are significant changes in the neutron energy spectrum a significant
influence results on the balance of respective interactions including leakage and
especially the process where neutrons escape the resonances. Furthermore, the fact
that there is a strong influence in the spectral shape of the spectrum shows that there
∂k
is need for a spectral multiplication factor ∂eEf f . Moreover, approaches that make
use of energy groups need to associate a specific effective multiplication factor to
44 D.Q. de Camargo et al.
Fig. 8 Spectral multiplication factor per Monte Carlo step. The symbols of the sequence are de-
fined in Fig. 7
each equation representing a specific group. Until now, only one global ke f f was
attributed to the set of equations. The question which we did not answer yet but
hope to respond to in future is whether the necessary spectral ke f f ,g used per energy
group is an average or a total value.
Beyond comparison with existing approaches in the literature, we intend to sim-
ulate the transients of any given initial power to any specified final power in order
to show how for a given power set-up the stationary state is approached. Along this
line, our present work may be considered as a first step into a new direction.
References
[Bod10] Bodmann, B.E.J., de Vilhena, M.T., Ferreira, L.S., Bardaji, J.B.: An analytical solver
for the multi-group two-dimensional neutron-diffusion equation by integral transform
techniques. Il Nuovo Cimento, 33 C, n. 1, 63–70 (2010).
[Cam09] Camargo, D.Q., Bodmann, B.E.J., Garcia, R.D.M., Vilhena, M.T.M.B.: A three-
dimensional collision probability method: Criticality and neutron flux in a hexahedron
setup. Annals of Nuclear Energy, 36, 1614–1618 (2009).
[Cha06] Chadwick, M.B., et al.: ENDF/B-VII.0: Next generation evaluated nuclear data li-
brary for nuclear science and technology. Nuclear Data Sheets, 107, n. 12, 2931–3118
(2006).
[CNSC03] Canadian Nuclear Safety Commission: Science and Reactor Fundamentals – Reactor
Physics. CNSC, Ottawa, ON, Canada (2003).
[Gon10] Gonçalves, G.A., Vilhena, M.T., Bodmann, B.E.J.: Heuristic geometric “eigenvalue
universaly” in a one-dimensional neutron transport problem with anisotropic scatter-
ing. Kerntechnik, 75, 50–52 (2010).
[Lep05] Leppanen, J.: A new assembly-level Monte Carlo neutron transport code for reac-
tor physics calculations, in Mathematics and Computation, Supercomputing, Reactor
A Novel Method for Simulating Spectral Nuclear Reactor Criticality 45
Physics and Nuclear and Biological Applications, Palais des Papes, Avignon, France,
September 12–15, 2005, on CD-ROM, American Nuclear Society, LaGrange Park, IL
(2005).
[Sek07] Sekimoto, H.: Nuclear Reactor Theory, COE-INES Tokyo Institute of Technology,
ISBN978-4-903054-11-7 C3058 Part II (2007).
[Vil08] de Vilhena, M.T., Heinen, I.R., Bodmann, B.E.J.: An analytical solution for the gen-
eral perturbative diffusion equation by integral transform techniques. Annals of Nu-
clear Energy, 35, 2410–2413 (2008).
[Vli08] Van Vliet, C.M.: Equilibrium and Non-equilibrium Statistical Mechanics, World Sci-
entific, ISBN 10-981-270-478-7 (2008).
Adaptive Particle Filter for Stable Distribution
1 Introduction
Estimation theory is a central issue for several applications: filtering, signal analy-
sis, image processing, control theory, inverse problem, and data assimilation. There
is intensive research into developing a set of techniques for estimating quantities,
for example: the least squares approximation, Kalman filter, variational method,
Bayesian approach and, more recently, schemes based on artificial intelligence: neu-
ral network and fuzzy logic.
For all the methods cited, there are some constraints to be considered. For ex-
ample, in the context of inverse problem, it is important to include some a priori
information to apply the least square solution. Some authors use the expression gen-
eralized least squares to characterize the latter condition. Kalman filter is another
alternative, and it was developed for a linear dynamical system under Gaussian as-
sumptions for describing the probability density function (PDF) for the variables
involved. The extended Kalman filter is an adaptation to employ the filter to nonlin-
ear problems. But this strategy can fail under strong nonlinear regimes. Variational
methods are standard procedures in estimation problems. The goal is to calculate
a minimum of an objective function (a functional), in a similar way to the least
square approach, but the gradient of the functional is computed by using the adjoint
function associated to the problem.
For Kalman filter and the variational method, one quantity remains unknown: the
covariance matrix for representing the modeling error. Some schemes were designed
to identify this covariance matrix: the use of an estimating operator, as a secondary
2 Particle Filter
The fundamental idea underlying the Sequential Monte Carlo is to represent the
probability density function (PDF) by a set of samples with their associated weights.
This set of samples is also referred to particles [GoSaSm93, ArMaGoCl02, Do98].
In the PF, an estimation of a posteriori PDF is obtained by resampling with replace-
ment from a priori ensemble. As the PF does not require us to assume that the PDF
is linear or Gaussian, it is applicable to general nonlinear problems. In particular,
the PF can be applied to cases in which the relationship between a state and ob-
served data is nonlinear [NaUeHi07]. However, the PF often encounters a problem,
sample impoverishment, since this dependence makes convergence results harder to
obtain. One way to avoid this problem is to introduce an additional noise to make
the particles differ more from each other [GoSaSm93, ChKr04, FuCaMa08].
Two properties are relevant issues for the PF: the Bayes’ theorem, and the
Markov property.
Theorem 1. For two statistically independent events A and B, with P(B) = 0, the
Bayes rule follows:
P(B|A)P(A)
P(A|B) = .
P(B)
Adaptive Particle Filter 49
Here, the Gordon’s implementation [GoSaSm93] is considered. The PDF p(xn |Ys ) is
approximate for a function of empirical PDF (see below). According to the Bayesian
view p(wn |Ys ) contains all statistical information available about the state variable
wn , based on the information in the measurements Ys :
M
∑ q̃n
(i) (i)
p(wn |Ys ) ≈ δ (wn − wn|s ),
i=1
(1)
M
∑
(i)
q̃t =1
i=1
(i)
where δ (.) is the Dirac delta function; q̃t denotes the weight associated with parti-
(i)
cle wn|s .
(initial PDF: Gaussian, with zero mean and σ = 5, i.e., N (0, 5));
2. Compute:
(i)
rn = p(yn |wn|n−1 ) = pet (yn − h(wn ,tn ))
where
pet (z) = exp(−z2 /2)(2 π )−1
and
z = yn − h(wn ,tn ).
3. Normalize:
(i)
(i) rn
r̃n = ( j)
.
∑M
j=1 rn
The kernel of the algorithm coming from the application of Bayes’ theorem and of
the Markov property:
Equation (2) is the Bayesian estimate for the state vector wn . The expectation is
a recursive procedure (2) convergent to the true value. There is some criticism on
the Bayesian estimation (sometimes this criticism is associated to the frequentist
school in the statistics community). In general it is related to the a priori choice
for the distribution p(wn |Yn−1 ) ≡ πpr (wn ), where there is no scientific basis for the
choice. We are going to stay out of this discussion. However, it is important to
point out that starting from a distribution πpr (wn ) with finite variance and a like-
lihood function as described in step 2 (a Gaussian one) it is not possible to reach
a distribution p(wn |Yn ) ≡ πpost (wn |Yn ) with an undefined variance. There are many
distributions with no defined variance, Cauchy and Lévy distributions are two ex-
amples.
In the present paper, our development will be focused on stable distributions
[No05], the class of nondegenerate distributions.
The central point of our analysis lies in the discussion of the likelihood func-
tion p(yn |wn ) ≡ πnoise (y − A(wn )), where y is the observed value. The mathematical
model is given by
f = A(w) + δ
and the realizations are expressed by
η = A(w) + μ
where δ and μ are additive white noise with E{δ } = E{η } = 0, and they are mu-
tually independent from f and η . Therefore [Be98]:
E{η } = E{ f } = A(wo ).
In Sect. 2.1 for standard PF algorithm, in step 2, a Gaussian distribution was chosen
to represent the likelihood function. It is possible to justify such a choice due to the
central limit theorem [Pa89]. Actually, there is a more general form for this theorem.
In the distribution space, there are at least two attractors for stable distributions. If
a random variable can be described as a linear combination of a summation of M
independent random variables with finite variance, the resulting distribution in the
limit M → ∞ is a Gaussian distribution. Another attractor is the Lévy alpha-stable
distribution, resulting as a limit of sum of independent and identically distributed
random variables with no defined variance. The second form of the central limit
theorem is also called Lévy–Gnedenko’s central limit theorem.
A non-extensive form of entropy has been proposed by Tsallis [Ts88]:
!
N
k
Sq (p) = 1 − ∑ pi q
q−1 i=1
in the limit q → 1.
52 H.F. de Campos Velho and H.C. Morais Furtado
As for the extensive form of entropy, the equiprobability condition produces the
maximum for the non-extensive entropy function, and this condition leads to special
distributions [Ts99].
q > 1:
1 − q x 2 −1/(q−1)
pq (x) = αq 1 −
+
(3)
3−q σ
q = 1:
1 1 1/2 −(x/σ )2 /2
pq (x) = e
σ 2π
q < 1:
1 − q x 2 1/(q−1)
pq (x) = αq− 1 − (4)
3−q σ
where
+∞ 2
x [pq (x)]q dx
σq2 = −∞
+∞ ,
−∞ [pq (x)]q dx
1/2 Γ 1
1 q−1 q−1
αq+ = ,
σq π (3 − q) Γ 2(q−1)
3−q
1 − q 1/2 Γ 2(1−q)
5−3q
1
αq− = .
σq π (3 − q) Γ 2−q 1−q
3 Numerical Results
The new approach for a particle filter is tested to identify the state vector of the
following dynamical filter.
Prediction:
xt 25xt
xt+1 = + + 8 cos(1.2t) + μt . (5)
2 1 + xt2
Observation:
xt2
yt+1 = + νt .
20
For the system (5), 1000 particles are employed. For the initial guess: q0 = 1 (Gaus-
sian distribution), where
x0 ∼ N(0, σx = 1).
The true distribution for the random variables μt and νt are both a Tsallis distribution
with q = 2.5 (see Fig. 1), and the level of noise is 2% of the maximum value of the
noiseless data.
Table 1 shows the results obtained using different likelihood functions. The
smallest error was obtained with the Tsallis’ distribution, with q determined with
the adaptive particle filter algorithm described in Sects. 2.2 and 2.3, where we note
the convergence to the exact value q = 2.5 (see Fig. 4). Figure 2 displays true
and estimated values for the dynamics; the observation is shown by discrete points
(squares).
54 H.F. de Campos Velho and H.C. Morais Furtado
Fig. 1 Histogram of random numbers for the Tsallis distribution for q = 2.5
Table 1 Errors: Experiments with noise according to Tsallis distribution for q = 2.5
q Distribution Average error
1.0 Gaussian 6.31348
1.5 Tsallis 6.19225
2.8 Tsallis 6.97586
2.9 Tsallis 7.19318
estimated Tsallis 6.17809
Fig. 2 True state (continous line), estimated state (dot line), and observation points (squares)
Adaptive Particle Filter 55
Figure 3 shows the errors related to the estimated values obtained with the new
approach employing Tsallis’ thermostatistics (Fig. 3 right side) with q value identi-
fied (see Fig. 4), and Gaussian likelihood (Fig. 3 left side).
Fig. 3 Error: (left side): estimation using Gaussian likelihood function; (right side): new approach
for particle filter
4 Conclusion
A particle filter is a statistical procedure for dealing with nonlinear problems and
there is no assumption on the Gaussian nature of the distribution of the random vari-
ables. This approach has received more attention in different fields: control theory,
signal processing—including image filtering, inverse problem, and data assimila-
tion (a good and recent review paper was presented by Leeuwen [Le09]), this is a
very important class of inverse problem. In the book of Kaipio and Somersalo data
assimilation can be understood as belonging to the nonstationary inverse problems
[JaSo05]. Sometimes the particle filter is also called a bootstrap filter.
56 H.F. de Campos Velho and H.C. Morais Furtado
It is possible to find strategies taking into account distributions with not defined
variance, for example, Kaipio and Somersalo [JaSo05] mentioned the Cauchy dis-
tribution. However, we cannot find a general procedure to deal with this scenario.
We have shown a more general scheme to address the problem. The procedure in-
troduced here can be employed for all applications cited in the previous paragraph.
For the worked example, the procedure converges to the exact distribution.
References
[Sc06] Schön, T.B.: Estimation of Nonlinear Dynamic Systems: Theory and Appli-
cations, Dissertations no. 998 (Linköping Studies in Sciense and Tecnol-
ogy), 2006.
[Be98] Bertero, M., Boccacci, P.: Inverse Problems in Imaging, Institute of Physics
(1998).
[Pa89] Papoulis, A.: Probability and Statistics, Pearson Higher Education (1989).
[TsQu07] Tsallis, C., Queiros, S.M.D.: Nonextensive statistical mechanics and cen-
tral limit theorems I – Convolution of independent random variables and
q-product, in Complexity, Metastability and Nonextensivity (CTNEXT 07)
(Editors: S. Abe, H. Herrmann, P. Quarati, A. Rapisarda, and C. Tsallis),
American Institute of Physics (2007).
[Le09] Leeuwen, P.J.V.: Particle filtering in geophysical systems. Monthly Weather
Review, 137, 4089 (2009).
On the Analytical Solution of the Multi-Group
Neutron Diffusion Kinetic Equation
in One-Dimensional Cartesian Geometry
by an Integral Transform Technique
1 Introduction
C. Ceolin
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: celina.ce@hotmail.com
M.T. Vilhena
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: vilhena@pq.cnpq.br
B.E.J. Bodmann
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: bardo.bodmann@ufrgs.br
to be more specific we present an analytical solution for the neutron diffusion ki-
netic equation in Cartesian geometry. For simplicity, and without losing generality,
we discuss the one-dimensional, multi-group diffusion kinetic equation with six de-
layed neutron precursor concentration groups. Bearing in mind that the equation for
the delayed neutron precursor concentration is a first-order linear differential equa-
tion in the time variable, one unique procedure (i.e. the GITT approach) for the set
of equations is made possible introducing a spurious diffusion term with (small)
diffusion constant ε . Applying the GITT technique to the modified diffusion kinetic
equation, we come out with a matrix differential equation which has a well known
solution in the limit ε → 0. Moreover, a comparison of arbitrary but small values
of ε to the zero limit solution permits us to evaluate the influence of diffusion of
the delayed neutron precursor concentration on the time evolution behavior of the
neutron flux. Note that the presence of this diffusion term may seem an arbitrary
complication of the equation system, but as the forthcoming discussion will show,
the presence of a spatial second-order derivative term opens the possibility to also
express the precursors in terms of the same set of eigenfunctions used for the neu-
tron diffusion problem. Thus the neutron flux as well as the precursor concentrations
are written as an expansion of orthogonal eigenfunctions of a self-adjoint compact
operator related to the diffusion term.
Upon application of the GITT technique to the extended diffusion kinetic equa-
tion results in a matrix differential equation that may be solved in general even for
large matrix order, due to the non-degeneracy of eigenvalues. This is especially an
advantage by virtue of the stiff character of this sort of problems, generated by the
considerable differences in time scale of the prompt and delayed neutrons. Unlike
other methods, the methodology adopted here is robust in that it allows us to work
out the one-dimensional diffusion kinetic equation, in a straightforward manner, for
problems that consider a significant number of energy groups (up to 200), and fur-
ther may be extended to multi-dimensional and multi-layered problems.
1 ∂ ∂2
φ1 (x,t) = D1 2 φ1 (x,t) − Σa1 φ1 (x,t)
v1 ∂ t ∂x
6
+(1 − β ) ν1 Σ f 1 φ1 (x,t) + ν2 Σ f 2 φ2 (x,t) + ∑ λiCi (x,t), (1)
i=1
1 ∂ ∂2
φ2 (x,t) = D2 2 φ2 (x,t) − Σa2 φ2 (x,t) + Σs12 φ1 (x,t), (2)
v2 ∂ t ∂x
Analytical Solution of the Neutron Diffusion Kinetic Equation 61
∂
Ci (x,t) = −λiCi (x,t) + βi ν1 Σ f 1 φ1 (x,t) + ν2 Σ f 2 φ2 (x,t) , (3)
∂t
where i = 1, . . . , 6, for t > 0 and 0 < x < L. Here φ1 (x,t) and φ2 (x,t) denote the fast
and the thermal neutron flux, Ci (x,t) is the delayed neutron precursor concentration
of the ith group, v is the neutron velocity, D is the diffusion coefficient, Σa is the
absorption cross section, Σs is the scattering cross section, Σ f is the fission cross
section, ν is the average number of neutrons emitted by fission, β is the delayed
neutron fraction and λ is the delayed neutron decay constant.
The system of equations (1)–(3) is subject to the zero flux boundary conditions
φ (0,t) = φ (L,t) = 0,
∂ ∂2
Ci (x,t) = ε 2 Ci (x,t) − λiCi (x,t) + βi ν1 Σ f 1 φ1 (x,t)
∂t ∂x
+ν2 Σ f 2 φ2 (x,t) ,
where i = 1, . . . , 6, and χn (x) = sin(γn x) are the eigenfunctions that are solutions
of an auxiliary Sturm–Liouville problem constructed from the adjoint terms of the
original problem. Replacing this ansatz in the system of equations (1)–(3), taking
62 C. Ceolin et al.
X (t) + A X(t) = 0,
Recalling that in the present problem the eigenvalues of A are nondegenerate allows
us to rewrite the exponential term
where Y is the matrix of the eigenvectors of the matrix A and Y−1 its inverse. Dg
is the diagonal matrix with the eigenvalues of A. There are a variety of ways at
hand from the literature to compute the solution in (7), as shown in a nineteen fold
way by [Mol78]. In a more recent publication [Seg08] one finds the application
of the combined Laplace transformation technique and matrix decomposition. This
method has the advantage of being general, in the sense that it can be applied to
solve problems with repeated eigenvalues as shown in [Seg99, Seg08]. Concluding
our derivation, we observe that the solution of the problem (1)–(3) is well defined
by (4), (5) and (6) where the expansion coefficients of the solution are evaluated by
the formula (8) in the limit ε → 0.
3 Numerical Results
To show the aptness of the proposed method to handle the diffusion kinetic equation,
in the following we solve a problem considering a slab with L = 160 cm and the
nuclear parameters given in Tables 1 and 2.
We present the influence of the spurious diffusion term on the results encountered
for the kinetic equation in the limit where ε goes to zero. In Tables 3 and 4 the
results obtained by this methodology for the fast and the thermal neutron flux are
shown for a selection of numerical values for 10−3 ≥ ε ≥ 10−7 . From a simple
Analytical Solution of the Neutron Diffusion Kinetic Equation 63
inspection of the displayed results in Table 3 stability of the solution for the fast
neutron flux becomes apparent; especially for ε = 10−6 and ε = 10−7 the solution
may be considered exact for applications. Also for the thermal neutrons (see Table 4)
one observes a comparable precision already for ε = 10−5 . Therefore, one may
conclude that the problem extended by a spurious term solves the diffusion kinetic
equation exactly.
In the following we show the numerical stability of the approach, i.e. the variation
of the numerical values when increasing the number of terms in the solution series.
It is noteworthy that for a neutron flux precision of the order of 1% only a small
number of terms in the expansion is necessary. In Tables 5 and 6, we show the
numerical variation of the results obtained for the fast and the thermal neutron flux
with increasing number of terms in the solution series up to N = 150. From Table 5,
one observes that an accuracy of eight significant digits for the fast neutron flux is
attained for a summation of 40 terms. For the thermal neutron flux (see Table 6), we
obtain the same precision with only 20 terms. It is evident that the results as shown
in Tables 5 and 6 are not a proof of convergence. However, a proof may be set up by
following the reasoning of [Bod10], upon which we will elaborate for the present
case in a future work.
For further illustration we show the plots for the fast and the thermal neutron
fluxes in Figs. 1 and 2, respectively. Further, we show in Fig. 3 also the delayed
neutron precursor concentration for the group i = 1.
Analytical Solution of the Neutron Diffusion Kinetic Equation 65
4 Conclusions
procedure analogous to the one discussed here. Indeed, the natural interest of our
future work shall focus on this topic.
References
[Abd06] Ait-Abderrahim, H., Stanculescu, A.: IAEA Coordinate Research Project on Analytical
and Experimental Benchmark Analyses of Accelerator Driven Systems. Proceedings of
ANS Topical Meeting on Reactor Physics, Paper C074 (2006).
[Bod10] Bodmann, B.E.J., de Vilhena, M.T., Ferreira, L.S., Bardaji, J.B.: An analytical solver
for the multi-group two-dimensional neutron-diffusion equation by integral transform
techniques. Il Nuovo Cimento, 33 C, n. 1, 63–70 (2010).
[Cor08] Corno, S.E., Dulla, S., Picca, P., Ravetto, P.: Analytical approach to the neutron kinetics
of the non-homogeneous reactor. Progress in Nuclear Energy, 50, 847–865 (2008).
[Cot93] Cotta, R.M.: Integral Transforms in Computational Heat and Fluid Flow, CRC Press,
Florida (1993).
[Cot98] Cotta, R.M. (Org.): The Integral Transform Method in Thermal and Fluids Science and
Engineering, Begell House, Inc., New York (1998).
[Cou89] Courant, R., Hilbert, D.: Methods of Mathematical Physics – II, John Wiley & Sons
(1989).
[Dul07] Dulla, S., Ravetto, P., Picca, P., Tomatis, D.: Analytical benchmarks for the kinetics
of accelerator-driven systems. Joint International Topical Meeting on Mathematics &
Computation and Supercomputing in Nuclear Applications, (2007).
[Mai07] Maiorino, J.R.: The Participation of IPEN in the AIEA Coordinate Research Projects on
Accelerator Driven Systems (ADS). International Nuclear Atlantic Conference – INAC,
(2007).
[Mol78] Moler, C., Van Loan, C.: Nineteen dubious ways of computing the exponential of a
matrix. SIAM Review, 801–836 (1978).
[Oli07] Oliveira, F.L., Maiorino, J.R., Santos, R.S.: The analytical benchmark solution of spa-
tial diffusion kinetics in source driven systems for homogeneous media. International
Nuclear Atlantic Conference – INAC, (2007).
[Seg99] Segatto, C.F., Vilhena, M.T., Gomes, M.G.: The one dimensional LTSn solution in a slab
with high degree of quadrature. Annals of Nuclear Energy, 26, 925–934 (1999).
[Seg08] Segatto, C.F., Vilhena, M.T., Marona, D.V.: The LTSn solution of the transport equation
for one-dimensional Cartesian geometry with c = 1. Kerntechnik, 73, 57–60 (2008).
Estimating the Validity of Statistical Energy
Analysis Using Dynamical Energy Analysis:
A Preliminary Study
1 Introduction
Dynamical energy analysis was recently introduced as a new approach toward de-
termining the distribution of mechanical and acoustic wave energy in complex built
up structures [Ta09]. The technique interpolates between standard statistical energy
analysis and full ray tracing, containing both of these methods as limiting cases.
Statistical energy analysis (SEA) is a highly efficient method for analyzing energy
distributions in large scale industrial problems, but remains an expert tool since ex-
plicit bounds on its validity are generally difficult to establish. In this work we aim
to address this problem through a study of the spectral properties of the transfer
operator arising in dynamical energy analysis. Through this study one can estimate
escape and correlation decay rates for SEA subsystems and use these quantities to
indicate the validity of an SEA approach.
D.J. Chappell
University of Nottingham, UK,
e-mail: david.chappell@nottingham.ac.uk
G. Tanner
University of Nottingham, UK,
e-mail: gregor.tanner@nottingham.ac.uk
ki = ω /ci + iμi /2
|G(r, r0 ; ω )|2
ε (r, r0 ; ω ) = , (2)
mi c2i
where L j is the length of the ray trajectory between r0 and r including possible re-
flections on boundaries. The amplitudes A j may be written as a product of three
terms as in [Ta09] due to damping, mode conversion and reflection/transmission co-
efficients, and geometrical factors. The phase index ν j contains contributions from
the reflection/transmission coefficients at interfaces between subsystems and from
caustics along the ray path.
Analogous representations to (3) have been considered in detail in quantum me-
chanics [Gu90] and are also valid for general wave equations in elasticity such as the
biharmonic equation and the Navier–Cauchy equation, see [Ta07] for an overview.
In the latter case G becomes matrix valued. Note that the summation in (3) is typ-
ically over infinitely many terms, where the number of contributing rays increases
(in general) exponentially with the length of the trajectories included. This gives rise
to convergence issues, especially in the case of low or no damping [Ta07].
Validity of Statistical Energy Analysis 71
The wave energy density (2) can now be expressed as a double sum over classical
trajectories and hence ε (r, r0 ; ω ) may be written as
with C = π 2 /(ρi c2i (2π )(d+1) ). The dominant contributions to the double sum arise
from terms in which the phases cancel exactly; one thus splits the calculation into a
diagonal part
ρ (r, r0 ; ω ) = ∑ |A j |2 (5)
j:r0 →r
A brief outline of the derivation of the DEA flow equations is now given; for details,
see [Ta09] and [Ch10]. The time dependence of a density of ray trajectories can
be described in terms of a linear phase space operator L τ (X,Y ) = δ (X − ϕ τ (Y )),
known as a Perron–Frobenius operator in dynamical systems theory, thus
ρ (X, τ ) = L τ (X,Y )ρ0 (Y )dY. (6)
Here X = (r, p) denotes the phase space coordinate with position r and momentum
p. The phase space flow ϕ τ (Y ) gives the position of the particle after time τ starting
at Y = (r , p ) when τ = 0. Furthermore, ρ0 denotes the initial ray density at time
τ = 0 and the domain of integration is over the whole of phase space.
72 D.J. Chappell and G. Tanner
with initial density ρ0 (Y, ω ) = δ (r −r0 )δ (k02 −H(Y )), where H = |p|2 is the Hamil-
ton function for the wave operator Ĥ. It can be shown that (7) is equivalent to the
diagonal approximation (5). A weight function w is included to incorporate damp-
ing and reflection/transmission coefficients. It is assumed that w is multiplicative,
(w(ϕ τ1 (X), τ2 )w(X, τ1 ) = w(X, τ1 + τ2 )), which holds for standard absorbtion mech-
anism and reflection processes.
In order to solve the stationary flow problem (7) a boundary mapping technique
is employed. For the time being let us consider a problem with a single (sub-)system
Ω = Ω1 with boundary Γ . The boundary mapping procedure involves first mapping
the ray density emanating continuously from the source onto the boundary Γ . The
(0)
resulting boundary layer density ρΓ is equivalent to a source density on the bound-
ary producing the same ray field in the interior as the original source field after one
reflection. Secondly, densities on the boundary are mapped back onto the bound-
ary by a boundary transfer operator B(X s ,Y s ; ω ) = w(Y s )δ (X s − φ ω (Y s )), where
X s = (s, ps ) represents the coordinates on the boundary (s parameterizes Γ and ps
denotes the momentum component tangential to Γ at s), likewise Y s = (s , ps ), and
φ ω is the invertible boundary map. Note that convexity is assumed to ensure φ ω is
well defined; non-convex regions could be handled by introducing a cut-off function
in the shadow zone as in [Le02] or by subdividing the regions further.
The stationary density on the boundary induced by the initial boundary distribu-
(0)
tion ρΓ (X s , ω ) can then be obtained using
∞
∑ Bn (ω )ρΓ (ω ) = (I − B(ω ))−1 ρΓ (ω ),
(0) (0)
ρΓ (ω ) = (8)
n=0
3.2 Subsystems
elements can then be treated as losses in one subsystem and source terms in another.
Typical subsystem interfaces are surfaces of reflection/ transmission due to sudden
changes in material parameters or local boundary conditions. We describe the full
dynamics in terms of subsystem boundary transfer operators Bi j ; flow between Ω j
and Ωi is only possible if Ωi ∩ Ω j = 0/ and one obtains
where φiωj is the boundary map in subsystem j mapped onto the boundary of the
adjacent subsystem i and Xis are the boundary coordinates of Ωi . The weight wi j
contains, among other factors, reflection and transmission coefficients characteriz-
ing the coupling at the interface between Ω j and Ωi .
equations [Ly95]. The resulting full NΩ -dimensional B matrix yields a set of SEA
equations using (8) after mapping the boundary densities back into the interior.
An SEA approximation is justified if the ray dynamics within each subsystem
are sufficiently chaotic such that a trajectory entering subsystem j forgets every-
thing about its past before exiting Ω j ; SEA can thus be viewed as a Markov ap-
proximation of the deterministic dynamics. Thus correlations within the dynamics
must decay rapidly on the time scale it takes for a typical ray to leave Ω j . This is
discussed further in relation to the spectral properties of the transfer operator in the
next section. Conditions typically cited for SEA to be a good approximation include
that the subsystem boundaries are sufficiently irregular, the subsystems are dynam-
ically well separated and absorption and dissipation is small. In this case SEA is
an extremely efficient method compared to standard ray tracing. However, for sub-
systems with regular features, such as rectangular cavities or corridor-like elements,
incoming rays are directly channeled into outgoing rays, thus rendering the Markov
approximation invalid and introducing memory effects. Likewise, strong damping
may lead to a significant decay of the signal before reaching the exit channel intro-
ducing geometric (system dependent) effects.
The spectrum of the transfer operator contains information about the asymptotic
behavior of the ray dynamics and in particular the escape and correlation decay
rates. As we have discussed above, for SEA to be a good approximation requires
correlations within the ray dynamics to decay rapidly on the time scale it takes for a
typical ray to leave a particular subsystem. Hence we expect to see high correlation
decay rates and low escape rates in cases where SEA works well. We will put this
expectation to the test in what follows.
We conduct our analysis for each subsystem i = 1, . . . , NΩ and consider the di-
agonal components of the transfer operator corresponding to self-interactions of
subsystems, that is, Bii . Once a basis function representation has been applied Bii
will be represented by a square matrix whose size depends on the order at which we
truncate the basis expansion. The size also depends on the form of the approxima-
tion, for example whether the basis approximation is applied globally as in [Ta09]
or if a subsystem boundary is split into a number of approximation regions due to
non-smooth boundary conditions at corners or subsystem interfaces as in [Ch10]. As
mentioned above, for SEA the matrix representation of Bii corresponds to a single
number.
The first quantity of interest is the escape rate for a subsystem i for some
i = 1, . . . , NΩ . This may be estimated from the maximum eigenvalue λ0 of the DEA
matrix representation Bii of Bii , which gives the probability that a ray trajectory
will remain in subsystem i (see Chap. 22 of [Cv06]). Therefore if subsystem i is
closed and undamped, that is, there is no transfer of energy between it and any other
subsystem and the damping coefficient μi = 0, then λ0 = 1. The asymptotic escape
Validity of Statistical Energy Analysis 75
rate γ is given as
γ = − ln λ0 . (10)
In this work we assume μi = 0 and therefore obtain an estimate of the minimum
escape rate for the subsystem i. Incorporating damping would simply correspond to
an increase in the escape rate proportional to the level of damping.
The second quantity of interest is the correlation decay rate. This may be es-
timated from the DEA matrix representation Bii of Bii when the subsystem i is
hypothetically considered to be closed and we take the damping coefficient μi = 0.
In particular, for ergodic ray dynamics the correlation decay rate may be estimated
from the gap between the largest two eigenvalues of Bii ([Cv06], Chap. 22). Since
we are considering closed subsystems with zero damping, then the maximum eigen-
value λ0 = 1. The correlation decay rate ν may be estimated from the second largest
eigenvalue λ1 as
ν = − ln λ1 . (11)
Note that the quantities computed as described above will be estimates since we
are only able to work with a basis approximation of the transfer operator its spectrum
is dependent on the underlying function space, see for example [Ba06]. In the next
section we study numerically the relative sizes of γ and ν for some examples where
we already know how well SEA works.
4 Numerical Examples
A variety of two-cavity systems are considered as in Ref. [Ta09] and are shown in
Fig. 1. Configuration A features irregular shaped well separated pentagonal subsys-
tems and SEA has been shown to work well [Ta09, Ch10]. In configuration B the
size of the interface between the subsystems is increased reducing their dynamical
separation and therefore the applicability of SEA. Configuration C includes a rect-
angular left-hand subsystem channeling rays out of the subsystem and introducing
long-range correlations in the dynamics. In addition, the source is further from the
intersection of the two subsystems. It has been demonstrated that SEA does not
work well for this configuration, particularly for large damping levels [Ta09, Ch10].
Figure 2 shows the approximations to the escape and correlation decay rates for
the two subsystems in configuration A. The quantities for subsystem 1 (the left-
hand region) are referred to as “Escape 1” and “Decay 1”, and analogous names are
used for subsystem 2 (the right-hand region). The order of the basis expansion N is
plotted on the horizontal axis. Given that SEA works well for this configuration one
would expect relatively low escape rates and large correlation decay rates. In fact
we see that only for subsystem 2 is ν > γ , for subsystem 1 this is only true for the
lowest level of basis approximation employed (N = 4). The reason the results are
not more conclusive is possibly due to the ray dynamics in the irregular polygons
not being sufficiently chaotic, leading to algebraic, rather than exponential decay of
correlations (see for example [Ca99]), that is, ν = 0. From the figure it is evident
that the decay rate is decreasing as we compute to a greater degree of accuracy using
larger N, and may be converging to zero. It is expected that more conclusive results
would be obtained if noise was incorporated in the model. In fact one could view
the error in the lower order approximations as a form of noise and in this case it is
clear that ν > γ for both subsystems.
Figure 3 shows the approximations to the escape and correlation decay rates for
configuration B and is laid out in the same way as Fig. 2. It is clear for both subsys-
tems here that γ > ν and thus the escape rate is large enough so that correlations in
the ray dynamics will not have decayed sufficiently before a typical ray exits either
subsystem. Therefore the Markov approximation underlying an SEA treatment will
Validity of Statistical Energy Analysis 77
not be valid, and as we have seen in previous studies SEA will not be a good model
in this case [Ta09, Ch10]. Note that the escape rates shown in Figs. 2 and 3 are lower
for subsystem 1 than for subsystem 2, which is expected since in both cases the in-
terface between the subsystems forms a smaller part of the total boundary length for
subsystem 2 than for subsystem 1.
Figure 4 shows the approximations to the escape and correlation decay rates for
configuration C. For subsystem 2 the results are similar to those for configuration
A and hence a similar explanation applies. For subsystem 1 the correlation decay
rate is zero because 1 is a repeated eigenvalue and hence ν = − ln λ1 = − ln 1 = 0.
The reason why 1 is a repeated eigenvalue here is the regular geometry, and in
particular the absolute values of the momenta on each pair of opposite sides being
time invariant for each admissible trajectory.
5 Conclusions
The escape and correlation decay rates of the transfer operator arising in dynami-
cal energy analysis have been studied numerically, with the aim of using them to
indicate the validity of an SEA approach. While the results clearly indicated when
SEA was not expected to be valid, they were less conclusive in the case where SEA
78 D.J. Chappell and G. Tanner
works well. These studies suggest that exponential decay of correlations is perhaps
too strong an indicator in the models employed here and weaker measures of the
decay of correlations in the ray dynamics may prove more suitable.
References
[Ba06] Bai, Z-Q.: On the discrete Frobenius–Perron operator of the Bernoulli map. J. Phys. A:
Math. Theor., 39, 4945–4953 (2006).
[Ca99] Casati, G., Prosen, T.: Mixing property of triangular billiards. Phys. Rev. Lett., 83, n. 23,
4729–4732 (1999).
[Ch10] Chappell, D.J., Giani, S., Tanner, G.: Dynamical energy analysis for built-up acoustic
systems at high frequencies. J. Acoust. Soc. Amer., submitted.
[Cv06] Cvitanović, P., Artuso, R., Mainieri, R., Tanner, G., Vattay, G.: Chaos: Classical and
Quantum, Niels Bohr Institut, Copenhagen, ChaosBook.org (2009).
[Gu90] Gutzwiller, M.C.: Chaos in Classical and Quantum Mechanics, Springer, New York
(1990).
[Le02] Le Bot, A.: Energy transfer for high frequencies in built-up structures. J. Sound. Vib., 250,
247–275 (2002).
[Ly95] Lyon, R.H., DeJong, R.G.: Theory and Application of Statistical Energy Analysis, 2nd
edn., Butterworth-Heinemann, Boston, MA (1995).
[Ta07] Tanner, G., Søndergaard, N.: Wave chaos in acoustics and elasticity. J. Phys. A: Math.
Theor., 40, R443–R509 (2007).
[Ta09] Tanner, G.: Dynamical energy analysis—Determining wave energy distributions in vibro-
acoustical structures in the high-frequency regime. J. Sound Vib., 320, 1023–1038 (2009).
Efficient Iterative Methods for Fast Solution
of Integral Operators Related Problems
K. Chen
1 Introduction
∂φ
The Helmholtz equation, with a Neumann boundary condition ∂n = g,
K. Chen
University of Liverpool, UK,
e-mail: k.chen@liverpool.ac.uk
∇2 φ + k2 φ = 0, p∈Ω (1)
where nq is the unit normal exterior to the boundary ∂ Ω away from Ω with
eik|p−q|
Gk (p, q) = ,
4π |p − q|
Au = f (3)
where u = φ ; for collocation method Ai, j = L(pi )ψ j and for Galerkin method Ai, j =
(ψi , L(pi )ψ j ).
Assume that A is not symmetric (if it is symmetric, simpler variants can be used).
Normal equation approach. The idea is to consider instead of (3)
AAT y = f , u = AT y,
since AAT is symmetric positive definite. Given an initial guess u(0) with residual
r(0) = f − AAT y(0) = f − Au(0) , we obtain α = (r(k) )T r(k) /(p(k) )T AAT p(k) from
solving
min rAAT = f − AAT yAAT
y
Iterative Methods for Integral Problems 81
in the form y = y(k+1) = y(k) + α p(k) with p(k) a new search direction. Here at k = 0,
we take p(0) = r(0) and for k ≥ 1, with r(k) = f − AAT y(k) = f − Au(k) , we use the
conjugate gradient direction p(k) = r(k) + β p(k−1) with (p(k) )T AAT p(k−1) = 0 i.e.
Equivalently let p(k) denote AT p(k) for the purpose of eliminating the intermedi-
ate variable y. Then u = u(k) + α p(k) , p(k) = AT r(k) + β p(k−1) with (p(k) )p(k−1) = 0
and
α = (r(k) )T r(k) /(p(k) )T AAT p(k) = (r(k) )T r(k) /(p(k) )T p(k) .
where em ∈ Rn is the mth unit vector. Further the above minimization is reduced to
the simple least squares’ problem
where y ∈ Rm and ê1 ∈ Rm is the first unit vector. After finding y, we obtain the next
iterate u(m) = u(0) + Qm y.
82 K. Chen
Multigrid Method
L u = f , = 1, 2, . . . , L
such that ∂ Ω 1 is the finest grid where we desire to solve and the other coarser grids
are here to speed up the computation. See [TrOoSc2001, Ch05].
The idea of a MGM relies on the residual correction principle. If ũ1 is a known
approximation of u1 with a nonzero residual r1 = f1 − L1 ũ1 . Then solving the resid-
ual equation
L2 v2 = R1 r1
or for a nonlinear case
L2 w2 = L2 R1 ũ1 + R1 r1 , v2 = w2 − R1 ũ1
MGM(uk , fk , k):
(1) Pre-smoothing over ∂ Ω k : Lk uk = fk
(2) Computation of the residual ∂ Ω k : rk = fk − Lk uk
(3) Set k = k + 1
(4) Restriction to coarse grid ∂ Ω k : uk = Rk−1 uk−1 and rk = Rk−1 rk−1
(5) If k = L (coarsest grid), solve Lk vk = rk ‘accurately’; otherwise
call the MGM step again: MGM(uk , fk , k)
(6) Set k = k − 1
(7) Interpolation of the correction vk = Pk vk+1
(8) Update the fine grid solution uk = uk + vk
(9) Return to continue
The most expensive part of the above two methods CGM and MGM is in the
matrix-vector products, which may be speeded up.
Two excellent ideas for speeding up matrix-vector products shown below lead to
fundamentally new and fast methods.
Iterative Methods for Integral Problems 83
The FMM makes to the top 10 algorithms [Ci00]. For boundary integral equations,
the decay properties of the integral kernel can be utilized analytically to design
hierarchical expansions that may be arranged to give the FMM approximation for
computing each row of matrix L1 multiplying a vector quickly. Similar algorithms to
the FMM can be derived using a function-free and H-matrix approach, see [BrGr97,
Fo09, AmPr99, Ba08].
We remark that while the FMM offers a fast solution per iteration for an iterative
method, the overall number of iterations required is dependent on the conditioning
of an underlying problem. To speed up such iterations effective preconditioning is
necessary. However, the computation of a preconditioner is usually restricted by the
inaccessibility of matrix elements from far field.
Wavelets Methods
A rich class of problems involving an integral operator arise from high resolution
image processing [Vo02, ChSh05]. Among others, one example is the image deblur-
ring problem (as shown in Fig. 1) of reconstructing image u from
given a noisy and blurred image z in Ω = [0, 1]2 ⊂ R2 , where η is unknown but
assumed to be a Gaussian white noise with 0 mean and estimable standard deviation
σ [Vo02].
Although more recent models exist, the most well-known model for the above
problem is due to Rudin–Osher–Fatemi [RuOsFa92]
* +
min E(u) ≡ α |∇u|dxdy + |Ku − z|2 dxdy , (4)
u Ω Ω
where
the first term is the total-variation (TV) semi-norm regularizer with |∇u| =
ux + u2y and the integral operator K is assumed to have a spatially-invariant kernel
2
i.e.
(Ku)(x, y) = k(x − x , y − y )u(x , y )dx dy .
Ω
The Euler–Lagrange partial differential equation of (4) is the following:
∇u
− α∇ · + K ∗ Ku = K ∗ z (5)
|∇u|β
where
|∇u|β = u2x + u2y + β
For either method, alternating minimization leads to simple solutions: For (6) solv-
ing for u is a simple denoising problem and solving for v can be through FFT be-
cause the regularizer now adds a constant diagonal to K ∗ K. For (7) solving for ω
can be done analytically while the solution of u involves a simpler semi-norm to
enable fast solvers.
Although elegant, both methods involve many (up to 20) iterations between two
sub-problems and hence are non-optimal. Besides, practically, only a nearby prob-
lem is solved as γ cannot be taken too large.
Below we consider how to solve (4) directly after discretizing it. Given z ∈ Rn×n ,
the above model [RuOsFa92] rewritten as
1
min E(u), E(u) = α ux + uy + (Ku − z) dxdy,
2 2 2
u Ω 2
n−1 n−1
E(u) = α ∑∑ (ui, j − ui, j+1 )2 + (ui, j − ui+1, j )2
i=1 j=1
1 n n n n 2
+ ∑ ∑ ∑ ∑
2 i=1 j=1 =1 m=1
Ki, j;,m u,m − zi, j ,
with α = α /h and h = 1/(n − 1). Here we shall assume that K = (Ki, j;,m ) is a
block circulant matrix with circulant blocks (BCCB). This is the case if we adopt
the periodic boundary condition [NgBo03, Vo02, ChSh05].
Now solve (8) by the coordinate descent method on the finest level 1:
86 K. Chen
⎧ (0)
⎪
⎪ Given u(0) = (ui, j ) = (zi, j ) with l = 0,
⎪
⎨ (l)
Solve ui, j = argminui, j ∈R E loc (ui, j ) for i, j = 1, 2, . . . , n
(9)
⎪ Set u(l+1) = (u(l) ) and repeat the above step with l = l + 1
⎪
⎪
⎩ i, j
until a prescribed stopping step on l,
where
1 (l) (l)
E loc
(ui, j ) = Ku − z + α
2
(ui, j − ui+1, j )2 + (ui, j − ui, j+1 )2
2
(l) (l) (l)
+ (ui, j − ui−1, j )2 + (ui−1, j − ui−1, j+1 )2
(l) (l) (l)
+ (ui, j − ui, j−1 )2 + (ui, j−1 − ui+1, j−1 )2 . (10)
min J(0
u + Pk ci, j )
ci, j ∈R
2 k2 −1
T (ci, j ) = ∑ (ci, j − hk1 −1, )2 + v2k1 −1, + ∑ (ci, j − vm,2 )2 + h2m,2
=1 m=k1
2 −1 k2
+ ∑ (ci, j − hk2 , )2 + v2k2 , + ∑ (ci, j − vm,1 −1 )2 + v2m,1 −1
=1 m=k1
√ 2
+ 2 (ci, j − vk2 ,2 )2 + hk2 ,2 ,
Iterative Methods for Integral Problems 87
1
vm, = u0m,+1 − u0k, , hm, = u0m+1, − u0m, ,
vk , + hk2 ,2 vk , − hk2 ,2 (12)
vk2 ,2 = 2 2 , hk2 ,2 = 2 2 .
2 2
Clearly since each iteration would take O(n2 ) per block on any level, the overall
algorithm will have O(n4 ) at least and is hence not optimal.
A new breakthrough on the issue was made in [ChCh10], based on the reorga-
nizing the solution of the above coarse level subproblems. We first observe that the
first order condition of (11) takes the form
In this short paper, we discussed two problems solved by integral methods. Both
problems may be studied further.
Firstly, for the Helmholtz equation, an optimal way of combining the FMM and
wavelets with suitable preconditioning is still to be found out. The more recent
interests of many researchers have turned to high frequency modeling (where the
wavenumber k is large) and inverse problems where the boundary information (ei-
ther boundary conditions or the boundary itself) is missing while some measurement
of the solution is known.
Secondly, for the image deblurring problem, it remains to develop effective iter-
ative methods for cases where a more sophisticated choice of regularizers (e.g. the
mean curvature [BrCh10]) is used or a spatially variant blur kernel is used. A more
practical problem is the blind deblurring, where the kernel k is not known and must
be estimated together with u. See [ChWo98, MoKa08].
Iterative Methods for Integral Problems 89
Acknowledgements Most of the results reviewed in this paper were achieved through joint work
with collaborators including Paul J. Harris (University of Brighton), Siamiak Amini (University of
Salford), Tony F. Chan (University of California at LA) and Raymond H. Chan (Chinese University
of Hong Kong).
References
[AmPr99] Amini, S., Profit, A.J.: Analysis of a diagonal form of the fast multipole algorithm
for scattering theory. BIT, 39, 585–602 (1999).
[AmHaWi92] Amini, S., Harris, P.J., Wilton, D.T.: Coupled Boundary and Finite Element Meth-
ods for the Solution of the Dynamic Fluid-Structure Interaction Problem, Lecture
Note in Engineering, 77, C.A. Brebbia and S.A. Orszag, eds., Springer, London
(1992).
[Ba08] Banjai, L., Hackbusch, W.: Hierarchical matrix techniques for low and high fre-
quency Helmholtz equation. IMA J. Numer. Anal., 28, 46–79 (2008).
[BrGr97] Beatson, R., Greengard, L.: A short course on fast multipole methods, in
Wavelets, Multilevel Methods and Elliptic PDEs, Oxford University Press, 1–37
(1997). See also http://math.nyu.edu/faculty/greengar/shortcourse_fmm.pdf.
[BrCh10] Brito, C., Chen, K.: Multigrid algorithm for high order denoising. SIAM J. Imag-
ing Sci., 3, 363–389 (2010).
[ChCh10] Chan, R.H., Chen, K.: A multilevel algorithm for simultaneously denoising and
deblurring images. SIAM J. Sci. Comput., 32, 1043–1063 (2010).
[ChCh02] Chan, T.F., Chen, K.: On two variants of an algebraic wavelet preconditioner.
SIAM J. Sci. Comput., 24, 260–283 (2002).
[ChSh05] Chan, T.F., Shen, J.H.: Image Processing and Analysis – Variational, PDE,
Wavelet, and Stochastic Methods, SIAM Publications (2005).
[ChWo98] Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE Trans. Image
Process, 7, 370–375 (1998).
[Ch05] Chen, K.: Matrix Preconditioning Techniques and Applications, Cambridge Uni-
versity Press, Cambridge (2005).
[Ci00] Cipra, B.A.: The Best of the 20th Century: Editors Name Top 10 Algorithms.
SIAM News, 33 (2000).
[Da97] Dahmen, W.: Wavelet and multiscale methods for operator equations. Acta Nu-
merica, 1997, 55–228 (1997).
[Fo09] Fong, W., Darve, E.: The black-box fast multipole method. J. Comp. Phys., 228,
8712–8725 (2009).
[HaCh05] Hawkins, S., Chen, K.: An implicit wavelet approximate inverse preconditioner.
SIAM J. Sci. Comput., 27, 667–686 (2005).
[HuNgWe08] Huang, Y., Ng, M., Wen, Y.: A fast total variation minimization method for image
restoration. SIAM J. Multiscale Modeling and Simulation, 7, 774–795 (2008).
[MaSc07] Mazya, V., Schmidt, G.: Approximate Approximations, American Mathematical
Society, Providence, RI (2007).
[MoKa08] Money, J.H., Kang, S.H.: Total variation minimizing blind deconvolution with
shock filter. Image Vision Comput., 26, 302–314 (2008).
[NgBo03] Ng, M., Bose, N.K.: Mathematical analysis of super-resolution methodology.
IEEE Signal Proc. Magazine, 20, 62–74 (2003).
[RuOsFa92] Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal
algorithms. Physica D, 60, 259–268 (1992).
[SaSc96] Saad, Y., Schultz, M.H.: GMRES: A generalized minimal residual algorithm for
solving nonsymmetric linear systems. SIAM J. Sci. Stat. Comput., 7, 856–869
(1986).
90 K. Chen
[TrOoSc2001] Trottenberg, U., Oosterlee, C., Schuller, A.: Multigrid, Academic Press (2001).
[Vo02] Vogel, C.R.: Computational Methods for Inverse Problems, SIAM Publications,
Philadelphia, PA (2002).
[WaYaYihZ08] Wang, Y., Yang, J., Yin, W., Zhang, Y.: A new alternating minimization algo-
rithm for total variation image reconstruction. SIAM J. Imaging Sci., 1, 248–272
(2008).
Analysis of Some Localized Boundary–Domain
Integral Equations for Transmission Problems
with Variable Coefficients
1 Introduction
We consider the basic and mixed transmission problems for scalar second order
elliptic partial differential equations with variable coefficients and use the localized
parametrices to reduce the problems to direct segregated boundary–domain integral
equations.
The treatment, by variational methods, of the transmission problems considered
in this paper have been investigated in the research literature, and the corresponding
uniqueness and existence results are well known (see, e.g., [HW08, LiMa72]).
For the special cases, when the fundamental solution is available, the Dirichlet-
and Neumann-type boundary value problems have also been investigated by the
classical potential method (see [Mir70, HW08] and the references therein).
Our goal here is to show that the problems can be equivalently reduced to some
localized boundary–domain integral equations (LBDIEs) and the corresponding lo-
calized boundary–domain integral operators (LBDIOs) are invertible which, beside
a pure mathematical interest, may also have some applications in numerical analysis
for construction of efficient numerical algorithms (see, e.g., [Mik02, MN05, SSA00,
ZZA98, ZZA99] and the references therein).
In our case, the localized parametrix is represented as the product of a Levi func-
tion of the differential operator under consideration and an appropriately chosen
O. Chkadua
A.Razmadze Mathematical Institute, Tbilisi, Georgia,
e-mail: chkadua@rmi.acnet.ge
S.E. Mikhailov
Brunel University West London, UK,
e-mail: sergey.mikhailov@brunel.ac.uk
D. Natroshvili
Georgian Technical University, Tbilisi, Georgia,
e-mail: natrosh@hotmail.com
Further, for sufficiently smooth functions (from the space H 2 (Ωq ), say) we intro-
duce the co-normal derivative operator on ∂ Ωq , q = 1, 2, in the usual trace sense:
3
Tq± (x, ∂x ) u(x) := ∑ aq (x) nk (x) γq± [∂xk u(x)],
(q)
(1)
k=1
where x ∈ ∂ Ωq and the symbols γq+ and γq− denote the trace operators on ∂ Ωq from
the domain Ωq+ := Ωq and its complement Ωq− := R3 \ Ω q , respectively.
We set
The classical co-normal derivative operators given by (1) can be continuously ex-
tended to functions from the spaces H 1, 0 (Ωq± ; Aq ) by the (generalized) canonical
1
co-normal derivative operators Tq± : H 1, 0 (Ωq± ; Aq ) → H − 2 (∂ Ωq ) (cf., for example,
[Co88, Lemma 3.2], [McL00, Lemma 4.3]) defined as
± 4 5
Tq u , w := ± (± ±
q w) Aq u + Eq (u, q w) dx (2)
∂ Ωq
Ωq
1
for all w ∈ H 2 (∂ Ωq ). Here ± ±
q are continuous linear extension operators, q :
1
H 2 (∂ Ωq ) → H 1 (Ωq± ) which are the right inverse of the trace operators γq± , while
and the central dot denotes the scalar product in R3 . The symbol g1 , g2 ∂ Ωq in (2)
−2 1 1
duality brackets between the spaces H (∂ Ωq ) and H (∂ Ωq ), coincid-
denotes the 2
Find functions u1 ∈ H 1,0 (Ω1 ; A1 ) and u2 ∈ H 1,0 (Ω2 ; A2 ) satisfying the differential
equations
Aq (x, ∂x ) uq = fq in Ωq , q = 1, 2, (3)
the transmission conditions on the interface
γ1 u1 − γ2 u2 = ϕ0i on Si , (4)
T1 u1 + T2 u2 = ψ0i on Si , (5)
γ2 u2 = ϕ0e on Se , (6)
We will refer these boundary transmission problems as (TD) and (TM) problems,
respectively.
For the data in the above formulated problems we assume
1 1
ϕ0i ∈ H 2 (Si ), ψ0i ∈ H − 2 (Si ),
1 1
ϕ0e ∈ H 2 (Se ), ψ0e ∈ H − 2 (Se ),
1 1
ϕ0e ∈ H 2 (SeD ), ψ0e ∈ H − 2 (SeN ),
(M) (M)
fq ∈ H 0 (Ωq ), q = 1, 2.
Equations (3) are understood in the distributional sense, the Dirichlet-type bound-
ary and transmission conditions are understood in the usual trace sense, while the
Neumann-type conditions for the co-normal derivatives are understood in the sense
of the canonical co-normal derivatives defined by (2).
We recall that the normal vectors n(1) and n(2) in the definitions of the co-normal
derivatives T1 u and T2 u on Si have opposite directions.
As we mentioned in the introduction, all the transmission problems formu-
lated above have been investigated in the literature using the variational methods
and the corresponding uniqueness and existence results are well known (see, e.g.,
[LiMa72]). Our goal here is to show that the problems can be equivalently reduced
to some LBDIEs and to investigate the Fredholm and invertibility properties of the
corresponding LBDIOs.
Localized BDIEs for Transmission Problems 95
based on the localized parametrices Pq . Here S ∈ {Si , Se , ∂ Ω2 }. Note that for the
layer potentials we will drop the subindex S when S = ∂ Ωq , i.e.,
96 O. Chkadua et al.
(q) (q)
V (q) := V∂ Ωq , W (q) := W∂ Ωq .
For the pseudodifferential operator (13), we employ also the notation LS(q) :=
LS+(q) . Note that the kernel functions of the operators (11) and (12) are at most
weakly singular if the localizing function χ ∈ X 2 and the surface S is C1,β smooth
with β > 0. The mapping properties of the operators (9)–(13) are studied in
[CMN09-L].
Further on we assume that the following relation holds on the interface:
Finally, we present some auxiliary propositions which play a crucial role in our
analysis and which can be proved by extending the arguments similar to those ap-
plied in the proof of Lemmas 6.3 and 6.4 in [CMN09-L] from the case χ ∈ X1+ 3 to
Lemma 1. Let χ ∈ X1∗ 3 , and let condition (14) hold. Further, let G ∈ H 0 (Ω ), g ∈
q q 1
1 1 1
H 2 (Si ), g2 ∈ H 2 (Si ), ge ∈ H − 2 (Se ) and
−
Then Gq = 0 in Ωq , q = 1, 2, g1 = 0, g2 = 0 on Si , and ge = 0 on Se .
Lemma 2. Let χ ∈ X1∗ 3 , and let condition (14) hold. Further, let G ∈ H 0 (Ω ), g ∈
q q 1
1 1
H − 2 (Si ), g2 ∈ H 2 (Si ), geD ∈ H0 − 21 (SeD ), geN ∈ H
0 12 (SeN ), and
The Second Green’s identity holds for the operator Aq (x, ∂x ) and u, v ∈ H 1, 0 (Ωq ; Aq ),
see, e.g., [Co88, Lemma 3.2], [McL00, Lemma 4.3],
[v Aq u − u Aq v] dx = [(γq v)Tq u − (γq u)Tq v] dS, q = 1, 2.
Ωq ∂ Ωq
By the standard limiting procedure near the singular point of the parametrix (see,
e.g., [Mir70]), we obtain the following parametrix-based third Green’s identity for
arbitrary u = uq ∈ H 1, 0 (Ωq ; Aq ),
Recall that for layer potentials we drop the subindex S when S = ∂ Ωq . Bearing in
mind the properties of the localized potentials given in Appendix B, the trace and
co-normal derivative of (15) are given by
1
γq uq +γq Rq uq −V (q) Tq uq +W (q) γq uq = γq Pq Aq uq on ∂ Ωq , (16)
2
1 (q)
Tq uq +Tq Rq uq −W Tq uq +L (q) γq uq = Tq Pq Aq uq on ∂ Ωq . (17)
2
With the help of these relations one can construct various types of LBDIE systems
for the above formulated transmission BVPs.
where
1 1 1
H(T D) := H 1, 0 (Ω1 ; A1 ) × H 1, 0 (Ω2 ; A2 ) × H − 2 (Si ) × H 2 (Si ) × H − 2 (Se ), (21)
(T D)
u2 + R2 u2 +VS(2) ψi +WS(2) ϕi −VS(2)
e
ψe = F2 in Ω2 , (23)
i i
where
(T D) 1 1
F1 = P1 f1 + VS(1) ψ0i − WS(1) ϕ0i ,
2 i 2 i
(T D) 1 1
F2 = P2 f2 + VS(2) ψ0i + WS(2) ϕ0i −WS(2) ϕ0e .
2 i 2 i e
K U (T D) = F (T D) ,
(T D)
(27)
then the five-vector U (T D) ∈ H(T D) given by (20), where ψi , ϕi and ψe are defined
by (19), solves LBDIEs system (22)–(26).
(ii) Vice versa, if a five-vector U (T D) ∈ H(T D) solves LBDIEs system (22)–(26)
and condition (14) holds, then (u1 , u2 ) ∈ H 1, 0 (Ω1 ; A1 ) × H 1, 0 (Ω2 ; A2 ) solves Prob-
lem (TD) and relations (19) hold.
A1 u1 − f1 = 0 in Ω1 , A2 u2 − f2 = 0 in Ω2 .
Due to this equivalence theorem we conclude that the LBDIE system (22)–(26)
with the special right hand side which belongs to the space F(T D) is uniquely solv-
able in the space H(T D) defined by (21). In particular, the corresponding homoge-
neous LBDIEs system possesses only the trivial solution. By the way, one can eas-
ily check that the right hand side expressions in LBDIEs system (22)–(26) vanish if
fq = 0 in Ωq , q = 1, 2, ϕ0i = 0 and ψ0i = 0 on Si , and ϕ0e = 0 on Se .
Our next aim is to establish the invertibility of the matrix operator generated
by the left hand side expressions in the LBDIEs system (22)–(26) both in already
introduced and in wider function spaces.
Let us introduce the notations
1 1 1
X(T D) := H 1 (Ω1 ) × H 1 (Ω2 ) × H − 2 (Si ) × H 2 (Si ) × H − 2 (Se ) ,
1 1 1
Y(T D) := H 1 (Ω1 ) × H 1 (Ω2 ) × H 2 (Si ) × H − 2 (Si ) × H 2 (Se ) .
K (T D)
: H(T D) → F(T D) (28)
: X(T D) → Y(T D) . (29)
Proof. We can easily see that the upper triangular matrix operator
⎡ ⎤
I 0 −rΩ VS(1) rΩ WS(1) 0
1 1
⎢ ⎥
i i
⎢0 I rΩ VS(2) rΩ WS(2) −rΩ VS(2) ⎥
⎢ e ⎥
⎢
2 2 2
i i
⎥
:= ⎢ rSi γ2VS(2) ⎥
(T D)
K0 ⎢0 0 −VS(1) − VS(2) WS(1) − WS(2) ⎥ (30)
⎢ ⎥
i i i i e
⎢0 0 0 L (1) + L (2) −rSi T2VSe ⎥
(2)
⎣ Si Si ⎦
0 0 0 0 −VS(2)
e
(T D)
K0 : X(T D) → Y(T D) , (31)
by standard arguments it can be shown that the operators in the main diagonal in
(30) are Fredholm of index zero in the corresponding function spaces. Therefore the
operator (29) is also Fredholm with index zero.
It remains to show that the null space of the operator (29) is trivial. We proceed
as follows. Let U ∈ X(T D) be a solution to the homogeneous system K (T D)U = 0.
Then the first two equations of the system imply that U ∈ H(T D) due Theorems 6
and 7, and by the equivalence Theorem 1 we conclude U = 0. Thus the kernel of the
operator (29) is trivial and consequently (29) is invertible.
To prove invertibility of operator (28), we remark that for any F (T D) ∈ F(T D) a
unique solution U ∈ X(T D) of (27) is delivered by the inverse of the operator (29).
On the other hand, since F (T D) ∈ F(T D) , the first two lines of the matrix operator
K (T D) imply that in fact U ∈ H(T D) and the mapping F(T D) → H(T D) delivered by
the inverse of the operator (29) is continuous, i.e., this operator is inverse to operator
(28).
Let us consider the mixed-type transmission problem (3), (4), (5), (7), (8), with the
right hand sides
1 1
ϕ0i ∈ H 2 (Si ), ψ0i ∈ H − 2 (Si ),
1 1
(32)
ϕ0e ∈ H 2 (SeD ), ψ0e ∈ H − 2 (SeN ),
(M) (M)
fq ∈ H 0 (Ωq ), q = 1, 2.
1 1
Let us denote by Φ0e ∈ H 2 (Se ) and Ψ0e ∈ H − 2 (Se ) some fixed extensions of the
(M) (M)
boundary functions ϕ0e and ψ0e from SeD and SeN , respectively, onto the whole
(M) (M)
surface Se , preserving the space. Then rSeD Φ0e = ϕ0e , rSeN Ψ0e = ψ0e .
Similar to (19) for the Problem (TD), let us introduce the following combinations
of the unknown boundary functions:
1 1 1 1
ψi = (T1 u1 − T2 u2 ) ∈ H − 2 (Si ), ϕi = (γ1 u1 + γ2 u2 ) ∈ H 2 (Si ),
2 2 (33)
0 − 21 (SeD ),
ψe = T2 u2 − Ψ0e ∈ H ϕe = γ2 u2 − Φ0e ∈ H0 12 (SeN ).
0 − 21 (SeD ) × H
×H 0 12 (SeN ),
Let us employ the third Green identities (15) in Ω1 and Ω2 , the difference of
their traces (16) and the sum of their co-normal derivatives (17) on Si ; and also the
trace (16) on SeD and the co-normal derivative (17) on SeN . Then after substituting
the transmission conditions (4)–(5) and the mixed boundary conditions (7)–(8) we
arrive at the following system of direct segregated LBDIEs for the components of
the vector U (T M) ,
(T M)
u1 + R1 u1 −VS(1) ψi +WS(1) ϕi = F1 in Ω1 , (35)
i i
(T M)
u2 +R2 u2 +VS(2) ψi +WS(2) ϕi −VS(2)
e
ψe +WS(2)
e
ϕe = F2 in Ω2 , (36)
i i
where
(T M) 1 1
F1 = P1 f1 + VS(1) ψ0i − WS(1) ϕ0i , (41)
2 i 2 i
(T M) 1 1
F2 = P2 f2 + VS(2) ψ0i + WS(2) ϕ0i +VS(2)Ψ0e −WS(2) Φ0e . (42)
2 i 2 i e e
As in the case of the problem (TD), we have here the following equivalence theorem.
1
Theorem 3. Let χ ∈ X1∗ 3 and conditions (32) hold. Further, let Φ ∈ H 2 (S ) and
0e e
− 1 (M) (M)
Ψ0e ∈ H 2 (Se ) be some fixed extensions of the boundary functions ϕ0e and ψ0e
from SeD and SeN , respectively, onto the whole surface Se .
(i) If a pair (u1 , u2 ) ∈ H 1, 0 (Ω1 ; A1 )×H 1, 0 (Ω2 ; A2 ) solves the transmission mixed
problem (TM), then the six-vector U (T M) ∈ H(T M) given by (34), where ψi , ϕi , ψe
and ϕe are defined by (33), solves the LBDIEs system (35)–(42).
(ii) Vice versa, if a six-vector U (T M) ∈ H(T M) solves the LBDIEs system (35)–
(42) and condition (14) holds, then the pair (u1 , u2 ) solves the Problem (TM) and
the relations (33) hold.
Proof. The claim (i) immediately follows from the deduction of (35)–(42).
Now, let a six-vector U (T M) solve the LBDIEs system (35)–(42). Subtracting
from (37) the trace γ1 of (35) and adding the trace γ2 of (36), we prove (4). Simi-
Localized BDIEs for Transmission Problems 103
larly, subtracting from (38) the co-normal derivative T1 of (35) and the co-normal
derivative T2 of (36), we prove (5). Subtracting from (39) the trace γ2 of (36), we
prove (7). Similarly, subtracting from (40) the co-normal derivative T2 of (36), we
prove (8). That is, the transmission conditions on Si and the mixed boundary condi-
tions on Se are fulfilled.
It remains to show that the (3) and the relations (33) hold true. Due to the embed-
ding U (T M) ∈ H(T M) , the third Green identities (15) hold. Comparing these identities
with the first two equations of the LBDIEs system, (35) and (36), and taking into
account transmission conditions (4)–(5) and mixed boundary conditions (7)–(8), al-
ready satisfied, we arrive at the relations
1 1
VS(1) [T1 u1 − T2 u2 ] − ψi +WS(1) ϕi − [γ1 u1 + γ2 u2 ]
i 2 i 2
= P1 ( f1 − A1 u1 ) in Ω1 ,
1 1
VS(2) [T1 u1 − T2 u2 ] − ψi −WS(2) ϕi − [γ1 u1 + γ2 u2 ]
i 2 i 2
+VS(2)
e
(−T2 u2 + ψe + Ψ0e ) +WS(2)
e
(γ2 u2 − ϕe − Φ0e )
= P2 (A2 u2 − f2 ) in Ω2 .
K (T M)
U (T M) = F (T M) , (44)
K (T M)
: H(T M) → F(T M) , (45)
104 O. Chkadua et al.
where
1 1
0 − 21 (SeD )×H
X(T M) := H 1 (Ω1 )×H 1 (Ω2 )×H − 2 (Si )×H 2 (Si )×H 0 12 (SeN ) ,
1 1 1 1
Y(T M) := H 1 (Ω1 )×H 1 (Ω2 )×H 2 (Si )×H − 2 (Si )×H 2 (SeD )×H − 2 (SeN ) .
As follows from the mapping properties of the potentials (see Theorems 6 and 7),
operators (45) and (46) are bounded.
Let us show that operator (46) is Fredholm with zero index and thus (46) and
consequently (45) are invertible.
Consider the upper triangular operator
⎡ ⎤
I 0 −rΩ VS(1) rΩ WS(1) 0 0
1 1
⎢ ⎥
i i
⎢0 I rΩ VS(2) rΩ WS(2) −rΩ VS(2) rΩ WS(2) ⎥
⎢ 2 2 2 e 2 e ⎥
⎢ i i
⎥
⎢0 0 −VS(1)−VS(2) WS(1)−WS(2) rSi γ2VS(2) −rSi γ2WS(2) ⎥
(T M) ⎢ e ⎥
K0 := ⎢ i i e i i
⎥.
⎢0 0 0 LS(1)+LS(2) −rSi T2VS(2) rSi T2WS(2) ⎥
⎢ ⎥
⎢ i e
i e
⎥
⎢0 0 0 0 −rSeD VS(2) rSeD WS(2) ⎥
⎣ e e ⎦
00 0 0 0 rSeN LSe
(2)
is a compact operator.
On the other hand, as has been mentioned above, in the proof of Theorem 2, the
third and forth operators in the main diagonal
1 1 1 1
−[VS(1) + VS(2) ] : H − 2 (Si ) → H 2 (Si ) , LS(1) + LS(2) : H 2 (Si ) → H − 2 (Si )
i i i i
Therefore, (47) and consequently (46) is a Fredholm operator with zero index.
It remains to show that the null space of operator (46) is trivial. Let U ∈ X(T M)
be a solution to the homogeneous equation K (T M)U = 0. Then due to the first
two lines of the matrix equation and mapping properties (52), (53) and (54) we see
that U ∈ H(T M) and by the equivalence Theorem 3 we conclude U = 0 due to the
uniqueness theorem for the problem (TM) in the space H(T M) . Thus the operator
(46) is invertible.
To prove the invertibility of operator (45), we note that for any F (T M) ∈ F(T M) a
unique solution U ∈ X(T M) of (44) is delivered by the inverse to the operator (46).
On the other hand, since F (T M) ∈ F(T M) , the first two lines of the matrix operator
K (T M) imply that in fact U ∈ H(T M) and the mapping F(T M) → H(T M) delivered by
the inverse to the operator (46) is continuous, i.e., this operator gives the inverse to
operator (45) as well.
Now we can summarize the results obtained above in the following theorem.
Theorem 4. Let χ ∈ X1∗
3 , and let condition (14) hold. Then the operators (45) and
1
σχ (ω ) := χ8s (ω ) > 0 for a.e. ω ∈ R, (48)
ω
where χ8s (ω ) denotes the sine-transform of χ̆ :
∞
χ8s (ω ) := χ̆ (ρ ) sin(ρ ω ) d ρ .
0
Here we collect some assertions describing the properties of the localized potentials
following from [CMN09-1, CMN09-L].
Theorem 5. The following operators are continuous:
0 s (Ωq ) → H s+2 (Ωq ),
Pq : H χ ∈ X 1,
s ∈ R, (49)
1 1
Pq : H s (Ωq ) → H s+2 (Ωq ), − < s < k − , χ ∈ X k , k = 1, 2, 3. (50)
2 2
Continuity of (49) is given by [CMN09-L, Theorem 5.4] while (50) can be proved
using [CMN09-L, Lemma 5.9] and [CMN09-1, Theorem 3.8].
are compact.
Acknowledgements This research was supported by the EPSRC grant EP/H020497/1: “Math-
ematical analysis of Localized Boundary–Domain Integral Equations for Variable-Coefficient
Boundary Value Problems” and partly by the Georgian Technical University grant in the case of
the third author.
References
[BCN09] Buchukuri, T., Chkadua, O., Natroshvili, D.: Mixed boundary value problems of
thermopiezoelectricity for solids with interior cracks. Integral Equations and Oper-
ator Theory, 64, n. 4, 495–537 (2009).
[CMN09-1] Chkadua, O., Mikhailov, S., Natroshvili, D.: Analysis of direct boundary–domain
integral equations for a mixed BVP with variable coefficient. Part I. Equivalence and
invertibility. J. Integral Equations Appl., 21, 499–542 (2009).
[CMN09-L] Chkadua, O., Mikhailov, S., Natroshvili, D.: Analysis of some localized boundary–
domain integral equations. J. Integral Equations Appl., 21, 407–447 (2009).
[CMN11] Chkadua, O., Mikhailov, S.E., Natroshvili, D.: Analysis of segregated boundary-
domain integral equations for variable-coefficient problems with cracks. Numerical
Methods for PDEs, 27, n. 1, 121–140 (2011).
[Co88] Costabel, M.: Boundary integral operators on Lipschitz domains: elementary results.
SIAM J. Math. Anal., 19, 613–626 (1988).
[DaLi90] Dautray, R., Lions, J.L.: Mathematical Analysis and Numerical Methods for Science
and Technology, vol. 4, Integral Equations and Numerical Methods, Springer-Verlag,
Berlin (1990).
[HW08] Hsiao, G.C., Wendland, W.L.: Boundary Integral Equations, Applied Mathematical
Sciences, Springer-Verlag, Berlin–Heidelberg (2008).
[LiMa72] Lions, J.-L., Magenes, E.: Non-Homogeneous Boundary Value Problems and Appli-
cations, vol. 1, Springer-Verlag, New York–Heidelberg (1972).
108 O. Chkadua et al.
[McL00] McLean, W.: Strongly Elliptic Systems and Boundary Integral Equations, Cambridge
University Press, Cambridge, UK (2000).
[Mik02] Mikhailov, S.E.: Localized boundary–domain integral formulation for problems with
variable coefficients. Int. J. Engineering Analysis with Boundary Elements, 26, 681–
690 (2002).
[MN05] Mikhailov, S.E., Nakhova, I.S.: Mesh-based numerical implementation of the local-
ized boundary–domain integral equation method to a variable-coefficient Neumann
problem. J. Engineering Math., 51, 251–259 (2005).
[Mir70] Miranda, C.: Partial Differential Equations of Elliptic Type, Second revised edition,
Springer-Verlag, New York–Berlin (1970).
[SSA00] Sladek, J., Sladek, V., Atluri, S.N.: Local boundary integral equation (LBIE) method
for solving problems of elasticity with nonhomogeneous material properties. Com-
put. Mech., 24, n. 6, 456–462 (2000).
[Mik06] Mikhailov, S.E.: Analysis of united boundary-domain integro-differential and inte-
gral equations for a mixed BVP with variable coefficient. Math. Methods in Applied
Sciences, 29, 715–739 (2006).
[ZZA98] Zhu, T., Zhang, J.-D., Atluri, S.N.: A local boundary integral equation (LBIE)
method in computational mechanics, and a meshless discretization approach. Com-
put. Mech., 21, n. 3, 223–235 (1998).
[ZZA99] Zhu, T., Zhang, J.-D., Atluri, S.N.: A meshless numerical method based on the local
boundary integral equation (LBIE) to solve linear and non-linear boundary value
problems. Eng. Anal. Bound. Elem., 23, n. 5–6, 375–389 (1999).
Analysis of Segregated Boundary–Domain
Integral Equations for Mixed
Variable-Coefficient BVPs in Exterior Domains
1 Introduction
The direct segregated boundary–domain integral equations for the mixed boundary-
value problem for a scalar second order elliptic partial differential equation with
variable coefficient in an exterior domain in R3 are analyzed in this paper. In
the literature the boundary-value problems considered here have been investigated
using variational methods in weighted Sobolev spaces, particularly in [Han71,
NP73, GN78, Mäu83, Gir87, DL90, Néd01]. For some cases of the PDE with con-
stant coefficients, when the fundamental solution is available, the Dirichlet and
Neumann type boundary-value problems in exterior domains were also investi-
gated by the classical potential (indirect boundary integral equation) method, see
[NP73, GN78, Gir87, DL90, CC00, Néd01] and the references therein.
Our goal here is to show that the mixed problems with variable coefficients
can be reduced to some systems of boundary–domain integral equations (BDIEs)
and investigate equivalence of the reduction and invertibility of the corresponding
boundary–domain integral operators in the weighted Sobolev spaces. To do this,
we extend to the exterior domains and weighted spaces the methods developed in
[CMN09a] for the interior domains and standard Sobolev (Bessel potential) spaces.
O. Chkadua
A. Razmadze Mathematical Institute, Tbilisi, Georgia,
e-mail: chkadua@rmi.acnet.ge
S.E. Mikhailov
Brunel University West London, UK,
e-mail: sergey.mikhailov@brunel.ac.uk
D. Natroshvili
Georgian Technical University, Tbilisi, Georgia,
e-mail: natrosh@hotmail.com
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 109
DOI 10.1007/978-0-8176-8238-5_11, © Springer Science+Business Media, LLC 2011
110 O. Chkadua et al.
H 1 (Ω ) := {g ∈ L2 (ρ −1 ; Ω ) : ∇g ∈ L2 (Ω )},
g2H 1 (Ω ) := ρ −1 g2L2 (Ω ) + ∇g2L2 (Ω ) .
Using the corresponding property for the space H 1 (Ω ), it is easy to prove that
D(Ω ) is dense in H 1 (Ω ), cf. [Han71, Theorem I.1], [Gir87, Theorem 2.2]. If Ω
is unbounded, then the semi-norm |g|H 1 (Ω ) := ∇gL2 (Ω ) is equivalent to the norm
gH 1 (Ω ) in H 1 (Ω ), see e.g. [DL90, Ch. XI, Part B, §1]. If Ω is bounded, then
H 1 (Ω ) = H 1 (Ω ). If Ω is a bounded subdomain of an unbounded domain Ω and
g ∈ H 1 (Ω ), then g ∈ H 1 (Ω ).
Let us define H 91 (Ω ) as a completion of D(Ω ) in H 1 (R3 ), H 9−1 (Ω ) :=
1 ∗ −1 9 1 ∗
[H (Ω )] , H (Ω ) := [H (Ω )] , L2 (ρ ; Ω ) := {g : ρ g ∈ L2 (Ω )}. Evidently
L2 (ρ ; Ω ) ⊂ H −1 (Ω ). Any distribution g ∈ H 9−1 (Ω ) has a representation g =
Analysis of Segregated BDIEs in Exterior Domains 111
where
E(u, ϕ )(x) := a(x)∇u(x) · ∇ϕ (x).
Thus by density of D(Ω ) in H 91 (Ω ), we see that A : H 1 (Ω ) → H −1 (Ω ) is con-
tinuous.
From the trace theorem (see, e.g., [LM72]) for u ∈ H 1 (Ω ) it follows that if u ∈
H 1 (Ω ± ), then γ ± u ∈ H 2 (∂ Ω ), where γ ± = γ∂±Ω are the trace operators on ∂ Ω
1
from Ω . We will use γ for γ ± if γ + = γ − . We will use also notations u± for the
±
H 1,0 (Ω ; A) := {g ∈ H 1 (Ω ) : Ag ∈ L2 (ρ ; Ω )}
cf. [GN78].
For u ∈ H 1 (Ω ) (as well as for u ∈ H 1 (Ω )) the co–normal derivative operators
a∂n u on ∂ Ω may not exist in the classical (trace) sense. However if u ∈ H 1,0 (Ω ; A),
one can correctly define the (generalized) canonical co–normal derivative T + u ∈
1
H − 2 (∂ Ω ) similar to, for example, [Cos88, Lemma 3.2], [McL00, Lemma 4.3]) as
+ 4 5 1
T u, w := (γ−1
+
w)Au + E(u, γ−1
+
w) dx ∀ w ∈ H 2 (∂ Ω ),
∂Ω
Ω
1
where γ−1
+
: H 2 (∂ Ω ) → H 1 (Ω ) is a bounded right inverse of the trace operator γ + :
1
H 1 (Ω ) → H 2 (∂ Ω ). The symbol g1 , g2 ∂ Ω denotes the duality brackets between
1 1
the spaces H − 2 (∂ Ω ) and H 2 (∂ Ω ), coinciding with ∂ Ω g1 (x)g2 (x)dS if g1 , g2 ∈
1
L2 (∂ Ω ). The operator T + : H 1,0 (Ω ; A) → H − 2 (∂ Ω ) is continuous and gives the
continuous extension on H 1,0 (Ω ; A) of the classical co-normal derivative operator
a∂n , where ∂n = n · ∇ and n = n+ is the normal vector on ∂ Ω directed outward with
respect to Ω .
Similar to the proofs available in [Cos88, Lemma 3.4], [McL00, Lemma 4.3] for
H 1,0 (Ω ; A) (see also [Mik08, Mik11] for the more general spaces H 1,t (Ω ; A)), one
can prove that for u ∈ H 1,0 (Ω ; A) the first Green identity holds in the form
112 O. Chkadua et al.
+ 4 5
T u , γ +v = v Au + E(u, v) dx ∀ v ∈ H 1 (Ω ). (3)
∂Ω
Ω
Then for arbitrary functions u, v ∈ H 1,0 (Ω ; A) we have the second Green identity,
4 5
v Au − u Av dx = T + u , γ + v ∂ Ω − T + v , γ + u ∂ Ω . (4)
Ω
A u = f in Ω , (5)
r∂ Ω γ + u = ϕ0 on ∂ ΩD , (6)
D
r∂ Ω T + u = ψ0 on ∂ ΩN , (7)
N
where 1 1
ϕ0 ∈ H 2 (∂ ΩD ), ψ0 ∈ H − 2 (∂ ΩN ), f ∈ L2 (ρ ; Ω ). (8)
Here ∂ Ω = ∂ ΩD ∪ ∂ ΩN , where ∂ ΩD and ∂ ΩN are nonintersecting simply con-
nected submanifolds of ∂ Ω with an infinitely smooth boundary curve := ∂ ΩD ∩
∂ ΩN ∈ C∞ .
The first Green identity (3) immediately implies the following uniqueness theo-
rem.
Theorem 1. The homogeneous version of BVP (5)–(7), i.e. with ϕ0 = 0, ψ0 = 0,
f = 0, has only the trivial solution, while the non-homogeneous problem (5)–(7)
with ϕ0 , ψ0 and f satisfying (8) has at most one solution in H 1,0 (Ω ; A).
Remark 1. We note that the existence of a solution in H 1 (Ω ; A), and thus in
H 1,0 (Ω ; A), can be proved using the variational setting and the Lax–Milgram theo-
rem, cf. [GN78, Mäu83, Gir87], where this was done for the Dirichlet and Neumann
problems for the Poisson equation.
where
3
xi − yi ∂ a(x)
R(x, y) = ∑ , x, y ∈ R3 . (10)
i=1 4 π a(y) |x − y| 3 ∂x
i
u + Ru −V (T + u) +W (γ + u) = PAu in Ω (11)
are the parametrix-based volume Newton-type and remainder potentials defined for
y ∈ R3 , while
V g(y) := − P(x, y) g(x) dSx , W g(y) := − [Tx P(x, y)]g(x) dSx (13)
∂Ω ∂Ω
are surface single layer and double layer potentials, defined for y ∈ R3 \∂ Ω . The
Newton-type and the remainder potential operator given by (12) for Ω = R3 will
be denoted as P and R, respectively. Recall that in the definition of W we assumed
Tx = a(x) n(x) · ∇x , where n = n+ is normal vector on ∂ Ω directed outward with
respect to Ω .
From (9), (10), (12), and (13) one can obtain representations of the parametrix-
based potential operators in terms of their counterparts for a = 1, i.e. associated with
the Laplace operator Δ ,
1 1 3
Pg =
a
PΔ g , R g = −
a ∑ ∂ j PΔ g ∂ j a) , (14)
j=1
1 1
V g = VΔ g, W g = WΔ (ag). (15)
a a
Theorem 2. The following operators are continuous:
114 O. Chkadua et al.
PΔ φ = Δ −1 φ . (23)
Δ v = 0 in Ω , γ + v = w on ∂ Ω (24)
VΔ φ = Qγ VΔ φ .
Analysis of Segregated BDIEs in Exterior Domains 115
1 1
Taking into account the continuity of the operator γ VΔ : H − 2 (∂ Ω ) → H 2 (∂ Ω ) and
the density of C∞ (∂ Ω ) in H − 2 (∂ Ω ) we arrive at the continuity of VΔ : H − 2 (∂ Ω ) →
1 1
H 1 (Ω ; Δ ). The first relation in (15) implies continuity of (21) under conditions (1).
Continuity of (22) is proved by a similar argument.
Let us prove continuity of (19). To this end, let us consider the second relation in
(14) for a density φ ∈ D(R3 ) and apply the Gauss divergence theorem
3
1 1
R φ (y) =
4π a(y) ∑ ∂y j |x − y| φ (x)∂ j a(x)dx
j=1 Ω
3
1 1
=− ∑ Ω j |x − y| φ (x)∂ j a(x)dx
4π a(y) j=1
∂x
3
1 1
=−
4π a(y) ∑ |x − y|
(γφ (x))∂n a(x)dSx
j=1 ∂ Ω
3
1 1
+
4π a(y) ∑ |x − y|
∂ j (φ (x)∂ j a(x))dx,
j=1 Ω
that is,
3
R φ (y) = −V [(γφ )∂n a](y) − ∑ P[∂ j (φ ∂ j a)](y). (25)
j=1
Due to the density of D(R3 ) in H 1 (Ω ), the continuity of the operators (18) and (21)
and conditions (1), relation (25) are valid also for φ ∈ H 1 (R3 ), thus implying (19).
For φ ∈ D(R3 ) the representation similar to (25) when Ω = R3 takes the form
3
Rφ (y) = − ∑ P[∂ j (φ ∂ j a)](y). (26)
j=1
where y ∈ S.
They can be also presented in terms of their counterparts for a = 1, i.e. associated
with the Laplace operator Δ , see [CMN09a],
1 1
V g = VΔ g, W g = WΔ (ag), (31)
a
a
∂ 1
W g = WΔ g + a VΔ g, (32)
∂n a
∂ 1
L ± g = LΔ (ag) + a WΔ± (ag) (33)
∂n a
where, as usual, the subscript Δ means that the corresponding surface potentials are
constructed by means of the harmonic fundamental solution PΔ (x, y) = −(4 π |x −
y|)−1 . It is taken into account that a and its derivatives are continuous in R3 and
γ ±V g1 (y) = V g1 (y),
1
γ ±W g2 (y) = ∓ g2 (y) + W g2 (y),
2
1
T ±V g1 (y) = ± g1 (y) + W g1 (y),
2
where y ∈ ∂ Ω .
Taking trace and co-normal derivative of the third Green identity (11) on ∂ Ω , we
obtain,
1 +
γ u + γ + Ru − V T + u + W γ + u = γ + PAu on ∂ Ω , (35)
2
1 +
T u + T + Ru − W ∂Ω
T + u + L∂+Ω γ + u = T + PAu on ∂ Ω . (36)
2
Analysis of Segregated BDIEs in Exterior Domains 117
u(y) + Ru −V Ψ +W Φ = P f in Ω , (37)
and prove for the weighted spaces the analog of [CMN09a, Lemma 4.1].
1 1
Lemma 1. Let u ∈ H 1 (Ω ), f ∈ L2 (ρ ; Ω ), Ψ ∈ H − 2 (∂ Ω ), Φ ∈ H 2 (∂ Ω ) satisfy
(37). Then u belongs to H 1,0 (Ω ; A) and is a solution of the equation
Au = f in Ω (38)
and
u = P f − Ru +V Ψ −W Φ ,
we conclude by Theorem 2 that u ∈ H 1,0 (Ω ; A). Thus we can write the third Green
identity (11) for the function u.
Subtracting (37) from the identity (11), we obtain
−V Ψ ∗ +W Φ ∗ = P[Au − f ] in Ω , (40)
Applying the Laplace operator Δ to the last equation and taking into consideration
that both the functions in the left-hand side are harmonic surface potentials, while
the right-hand side function is the classical Newtonian volume potential, we arrive
at (38). Substituting (38) back into (40) leads to (39).
V Ψ ∗ = 0 in Ω ,
then Ψ ∗ = 0.
1
(ii) Let Φ ∗ ∈ H 2 (∂ Ω ). If
W Φ ∗ (y) = 0 in Ω ,
(iii) Let ∂ Ω = S1 ∪S2 , where S1 and S2 are nonempty nonintersecting simply con-
nected submanifolds of ∂ Ω with infinitely smooth boundaries. Let Ψ ∗ ∈ H 0 − 21 (S1 ),
Φ∗ ∈ H 0 12 (S2 ). If
V Ψ ∗ (y) −W Φ ∗ (y) = 0 in Ω ,
then Ψ ∗ = 0 and Φ ∗ = 0 on ∂ Ω .
Proof. The proofs of items (i) and (iii) coincide with the proofs of their counterparts
for an interior domain in [CMN09a, Lemma 4.2].
To prove item (ii), we first remark that the Gauss lemma implies that ΦΔ = −C
satisfies the equation WΔ ΦΔ = 0 in the exterior domain Ω for any C = const. Let us
check that there is no other solutions of the equation in Ω . By the usual argument,
T +WΔ ΦΔ = T −WΔ ΦΔ = 0 on ∂ Ω , which implies WΔ = const in the interior domain
Ω − due to the uniqueness up to a constant of the solution of the Neumann problem
in H 1 (Ω − ). Then the jump property of WΔ gives ΦΔ = const. Applying the second
relation of (15) finalizes the proof of item (ii).
γ + u = Φ0 + ϕ , 0 12 (∂ ΩN );
ϕ ∈H T + u = Ψ0 + ψ , 0 − 21 (∂ ΩD ),
ψ ∈H
and we will regard the new unknown functions ϕ and ψ as formally segregated of u.
Thus we will look for the triplet
0 − 21 (∂ ΩD ) × H
U = (u, ψ , ϕ ) ∈ H := H 1,0 (Ω ; A) × H 0 12 (∂ ΩN )
0 − 2 (∂ ΩD ) × H
⊂ X := H 1 (Ω ) × H
1
0 2 (∂ ΩN ).
1
First, using (11) in Ω , the restriction of (35) on ∂ ΩD , and the restriction of (36)
on ∂ ΩN , we arrive at the BDIE system (M11) of three equations for the triplet of
unknowns, (u, ψ , ϕ ),
Analysis of Segregated BDIEs in Exterior Domains 119
u + Ru −V ψ +W ϕ = F0 in Ω , (41)
2 3
r∂ Ω γ + Ru − V ψ + W ϕ = r∂ Ω γ + F0 − ϕ0 on ∂ ΩD ,
D D
6 7
r∂ Ω T + Ru − W ψ + L∂+Ω ϕ = r∂ Ω T + F0 − ψ0 on ∂ ΩN ,
N N
where
F0 := P f +V Ψ0 −W Φ0 in Ω .
We denote the matrix operator of the left-hand side of the systems (M11) as
⎡ ⎤
I +R −V W
⎢ ⎥
⎢ ⎥
M 11 := ⎢ r∂ ΩD γ + R −r∂ ΩD V r∂ ΩD W ⎥ .
⎣ ⎦
r∂ Ω T + R −r∂ Ω W r∂ Ω L +
N N N
The notation (M11) and the corresponding superscripts mean that system in-
cludes the integral operators of the first kind on both the Dirichlet and Neumann
parts of the boundary. The other BDIE systems below are also denoted, respectively.
Here we use (11) in Ω and (35) on the whole of ∂ Ω to arrive at the BDIE system
(M12) of two equations for the triplet (u, ψ , ϕ ),
u + Ru −V ψ +W ϕ = F0 in Ω ,
1
ϕ + γ + Ru − V ψ + W ϕ = γ + F0 − Φ0 on ∂ Ω .
2
The left-hand side matrix operator of the system is
⎡ ⎤
I + R −V W
⎢ ⎥
M 12 := ⎣ 1 ⎦.
γ R −V
+ I +W
2
To arrive at the BDIE system (M21) of two equations for the triplet (u, ψ , ϕ ), we
use (11) in Ω and (36) on the whole of ∂ Ω ,
u + Ru −V ψ +W ϕ = F0 in Ω , (42)
120 O. Chkadua et al.
1
ψ + T + Ru − W ψ + L + ϕ = T + F0 − Ψ0 on ∂ Ω . (43)
2
The left-hand side matrix operator of the system is
⎡ ⎤
I +R −V W
M 21 := ⎣ 1 ⎦.
T +R I −W L+
2
Finally, using (11) in Ω , the restriction of (36) on ∂ ΩD , and the restriction of (35) on
∂ ΩN , we arrive for the triplet (u, ψ , ϕ ) at the BDIE system (M22) of three equations
of “almost” the second kind (up to the spaces),
u + Ru −V ψ +W ϕ = F0 in Ω ,
1 6 7 6 7
ψ + r∂ Ω T + Ru − W ψ + L + ϕ = r∂ Ω T + F0 − Ψ0 on ∂ ΩD ,
2 D D
1 6 7 6 7
ϕ + r∂ Ω γ + Ru − V ψ + W ϕ = r∂ Ω γ + F0 − Φ0 on ∂ ΩN .
2 N N
The matrix operator of the left-hand side of the system (M22) takes the form
⎡ ⎤
I +R −V W
⎢ 1 ⎥
⎢ ⎥
⎢r +R − W
L + ⎥
M := ⎢ ∂ ΩD
22 T r∂ ΩD
I r∂ ΩD ⎥.
⎢ 2 ⎥
⎣ 1 ⎦
r∂ Ω γ + R −r∂ Ω V r∂ Ω I +W
N N N 2
Remark 2. Note that the second relation (14) means that if a = const outside a
bounded subdomain Ω ⊂ Ω , then the operator R acts only on the restriction rΩ u.
This implies that all the BDIE systems reduce in this case to the BDIEs over Ω and
∂ Ω , and are supplemented with the integral representations for u in Ω \Ω̄ given by
the first equations of the systems.
Denoting the right-hand sides of the systems (M11), (M12), (M21) and (M22) as
F αβ , the systems can be rewritten as
M αβ U = F αβ ,
6 BDIE Analysis
Proof. Item (i) immediately follows from the deduction of the BDIE systems (M11),
(M12), (M21), (M22).
Using the similarity of Lemma 1 and items (i, iii) of Lemma 2 to their coun-
terparts Lemma 4.1 and Lemma 4.2 (i, iii) in [CMN09a] for the bounded domain
Ω , the proof of item (ii) of the theorem follows word-for-word the corresponding
proofs of Theorems 5.2, 5.5 and 5.12 in [CMN09a].
The situation with regards to uniqueness and equivalence for system (M21) dif-
fers from the one for other systems and from its counterpart BDIE system (T T ) in
[CMN09a], particularly because item (ii) of Lemma 2 is different from its analog,
[CMN09a, Lemma 4.2(ii)]. Thus system (M21) will be further analyzed elsewhere.
To prove the invertibility of the counterparts of the operators M αβ for bounded
domains in [CMN09a], we essentially used there the compactness of the operator
R : H 1 (Ω ) → H 1 (Ω ) based on the Rellich compactness theorem. However, the
latter theorem does not hold for unbounded domains with compact boundaries, and
so to cope with this, we will split the operator R into two parts, one of which can
be made arbitrarily small while the other one is compact.
122 O. Chkadua et al.
Lemma 3. Let ρ |∇a(x)| → 0 as |x| → ∞. Then for any ε > 0 the operator R can
be represented as R = Rc + Rs , where Rc : H 1 (Ω ) → H 1 (Ω ) is compact, while
Rs H 1 (Ω )→H 1 (Ω ) < ε .
where
3 3
Q := ∑ ∂ j [(1 − μ )g∂ j a]H9−1 (Ω ) ≤ ∑ (1 − μ )g∂ j aL2 (Ω )
j=1 j=1
Thus
as claimed.
Let us prove the claim about the operator Rc . Since the support of μ belongs
to B2η , for any fixed η the operator Rc : : H 1 (Ω ) → H 1 (Ω ) can be represented
as Rc g = RΩ2η [μ rΩ g], where Ω2η = Ω B2η and the operator RΩ2η is given by
2η
the second relation (12) with Ω replaced by Ω2η . The operator RΩ2η : L2 (Ω2η ) →
H 1 (Ω ) is continuous by (20) since L2 (Ω2η ) = L2 (ρ −1 ; Ω2η ) for the bounded do-
main Ω2η . On the other hand, the restriction operator rΩ : H 1 (Ω ) → H 1 (Ω2η ) =
2η
H 1 (Ω2η ) is continuous while the imbedding of H 1 (Ω2η ) to L2 (Ω2η ) is compact,
which implies that the operator Rc : H 1 (Ω ) → H 1 (Ω ) is compact.
M 11 : X → Y11 , (45)
M 11 : H → F11 . (46)
where ⎡ ⎤
I −V W
⎢ ⎥
⎢ −r∂ Ω V r∂ Ω W ⎥
M011 := ⎢ 0 ⎥,
⎣ D D ⎦
0 0 r∂ ΩN
Lˆ
and Lˆ is defined in (34). Evidently operator (47) is continuous. The diagonal op-
erators of the triangular matrix operator M011 are continuously invertible (cf. the
proof of [CMN09a, Theorem 5.3]), implying that the operator (47) is continuously
invertible as well.
Let us now represent R = Rs + Rc by Lemma 3 so that the operator Rs is suffi-
ciently small for the operator
⎡ ⎤
Rs 0 0
⎢ ⎥
⎢ ⎥
Ms11 := ⎢ r∂ ΩD γ + Rs 0 0 ⎥
⎣ ⎦
r∂ Ω T + Rs 0 0
N
to satisfy the inequality Ms11 X→Y11 < (M011 )−1 Y11 →X , where (M011 )−1 is the
operator inverse to M011 . Then the operator M011 + Ms11 : X → Y11 is continuously
invertible, while the operator
M 12 : X → Y12 ,
M 12 : H → F12 .
where
124 O. Chkadua et al.
⎡ ⎤
I −V W
⎢ ⎥
M012 := ⎣ 1 ⎦.
0 −V I
2
Evidently operator (48) is continuous. Any solution U = (u, ψ , ϕ ) ∈ X of the
1
equation M012 U = F , where F = (F1 , F2 ) ∈ H 1 (Ω ) × H 2 (∂ Ω ) will solve
also the following extended system of three equations,
u +W ϕ − V ψ = F1 in Ω ,
1
ϕ − V ψ = F2 on ∂ Ω ,
2
−r∂ Ω V ψ = r∂ Ω F2 on ∂ ΩD ,
D D
I : H 1 (Ω ) → H 1 (Ω ),
1 1 1
I : H 2 (∂ Ω ) → H 2 (∂ Ω ),
2
−r∂ Ω V :H 0 − 21 (∂ ΩD ) → H 12 (∂ ΩD ),
D
are continuously invertible, implying that the triangular matrix operator of the sys-
1
tem mapping H 1 (Ω )×H − 2 (∂ Ω )× H 0 12 (∂ ΩN ) to H 1 (Ω )×H 12 (∂ Ω )×H 12 (∂ ΩD )
is also invertible. Taking into account that if ψ ∈ H 0 − 21 (∂ ΩD ) solves the third equa-
tion of the system, then ϕ = 2(F2 + V ψ ) ∈ H 0 2 (∂ ΩN ), we arrive at the invertibility
1
of the operator (48). The rest of the proof coincides word-for-word with the one for
Theorem 5.
such that
F1 = P f∗ +V Ψ∗ −W Φ∗ in Ω + , (50)
F2 = rS1 T + F1 − rS1 Ψ∗ on S1 , (51)
F3 = rS2 γ + F1 − rS2 Φ∗ on S2 . (52)
Analysis of Segregated BDIEs in Exterior Domains 125
M 22 : H → F22 (53)
M 22 U = F (54)
ψ = T + u − (C∂ ΩD ,∂ ΩN F )2 ,
ϕ = γ + u − (C∂ ΩD ,∂ ΩN F )3 ,
and is evidently continuous. Thus the operator (M 22 )−1 is the right inverse of the
operator (53) but due to the injectivity of the latter implied by the equivalence The-
orem 4, the operator (M 22 )−1 is the two-side inverse of it.
u = F0 +VΔ ψ −WΔ ϕ in Ω .
126 O. Chkadua et al.
where UˆΔ := (ψ , ϕ ) ∈ H0 − 2 (∂ ΩD ) × H
1
0 2 (∂ ΩN ), 1
⎡ ⎤
r 1
I − W r L +
∂ Ω Δ ∂ Ω Δ
MˆΔ22 := ⎣ D D ⎦,
2
−r∂ Ω VΔ r∂ Ω 12 I + WΔ
N N
! (58)
r T F0 − r∂ Ω Ψ0
+
1 1
FˆΔ22 := ∂ ΩD + D ∈ H − 2 (∂ ΩD ) × H 2 (∂ ΩN ).
r∂ Ω F0 − r∂ Ω Φ0
N N
0 − 21 (∂ ΩD )× H
Moreover, the operator MˆΔ22 : H 0 12 (∂ ΩN ) → H − 21 (∂ ΩD )×H 12 (∂ ΩN )
is bounded and injective. Similar to [CMN09a, Theorem 5.18], one can prove the
following corollary from Theorem 7.
0 − 21 (∂ ΩD ) × H
MˆΔ22 : H 0 12 (∂ ΩN ) → H − 21 (∂ ΩD ) × H 12 (∂ ΩN )
is invertible.
where ⎡ ⎤
I −V W
⎢ 1 ⎥
⎢ ⎥
M022 := ⎢ 0 r∂ Ω 2I −W Δ r∂ Ω Lˆ ⎥.
⎣ D D ⎦
1
0 −r∂ Ω V r∂ Ω 2 I + W
N N
where diag(1, 1/a) and diag(1, a) are diagonal 2 × 2 matrices. The operator MˆΔ22
given by (58) is invertible by Theorem 8. Since 0 < a0 < a(x) < a1 < ∞, this implies
the invertibility of the operator
0 − 21 (∂ ΩD ) × H
Mˆ022 : H 0 12 (∂ ΩN ) → H − 21 (∂ ΩD ) × H 12 (∂ ΩN )
and thus of operator (59). The rest of the proof coincides word-for-word with the
one for Theorem 5.
7 Concluding Remarks
Acknowledgements This work was supported by the grant EP/H020497/1 “Mathematical analy-
sis of localized boundary–domain integral equations for BVPs with variable coefficients” from the
EPSRC, UK.
References
[CC00] Chudinovich, I., Constanda, C.: Variational and Potential Methods in the Theory of
Bending of Plates with Transverse Shear Deformation, Chapman & Hall/CRC, Boca
Raton – London – New York – Washington, DC (2000).
[CMN09a] Chkadua, O., Mikhailov, S.E., Natroshvili, D.: Analysis of direct boundary–domain
integral equations for a mixed BVP with variable coefficient, I: Equivalence and in-
vertibility. Journal of Integral Equations and Applications, 21, 499–543 (2009).
[CMN09b] Chkadua, O., Mikhailov, S.E., Natroshvili, D.: Analysis of some localized boundary–
domain integral equations. Journal of Integral Equations and Applications, 21, 405–
445 (2009).
128 O. Chkadua et al.
[Cos88] Costabel, M.: Boundary integral operators on Lipschitz domains: elementary results.
SIAM J. Math. Anal., 19, 613–626 (1988).
[DL90] Dautray, R., Lions, J.: Mathematical Analysis and Numerical Methods for Science
and Technology, vol. 4, Integral Equations and Numerical Methods, Springer, Berlin
– Heidelberg – New York (1990).
[Gir87] Giroire, J.: Étude de Quelques Problèmes aux Limites Extérieurs et Résolution par
Équations Intégrales. Thése de Doctorat d’État, Université Pierre-et-Marie-Curie
(Paris-VI) (1987).
[GN78] Giroire, J., Nedelec, J.: Numerical solution of an exterior Neumann problem using a
double layer potential. Mathematics of Computation, 32, 973–990 (1978).
[Han71] Hanouzet, B.: Espaces de Sobolev avec poids application au probleme de Dirichlet
dans un demi espace. Rend. del Sem. Mat. della Univ. di Padova, XLVI, 227–272
(1971).
[LM72] Lions, J.-L., Magenes, E.: Non-Homogeneous Boundary Value Problems and Appli-
cations, vol. 1, Springer, Berlin – Heidelberg – New York (1972).
[Mäu83] Mäulen, J.: Lösungen der Poissongleichung und harmonishe Vektorfelder in un-
beshränkten Gebieten. Math. Meth. in the Appl. Sci., 5, 233–255 (1983).
[McL00] McLean, W.: Strongly Elliptic Systems and Boundary Integral Equations, Cambridge
University Press, Cambridge, UK (2000).
[Mik02] Mikhailov, S.E.: Localized boundary–domain integral formulations for problems with
variable coefficients. Engineering Analysis with Boundary Elements, 26, 681–690
(2002).
[Mik05] Mikhailov, S.E.: Analysis of extended boundary–domain integral and integro-
differential equations of some variable-coefficient BVP, in: Advances in Boundary
Integral Methods — Proc. of the 5th UK Conference on Boundary Integral Methods
(Editor: K. Chen), University of Liverpool Publ., Liverpool, UK, 106–125 (2005).
[Mik06] Mikhailov, S.E.: Analysis of united boundary–domain integro-differential and inte-
gral equations for a mixed BVP with variable coefficient. Math. Methods in Applied
Sciences, 29, 715–739 (2006).
[Mik08] Mikhailov, S.E.: About traces, extensions and co-normal derivative operators on Lip-
schitz domains, in: Integral Methods in Science and Engineering: Techniques and
Applications (Editors: C. Constanda, S. Potapenko), Birkhäuser, Boston, 149–160
(2008).
[Mik11] Mikhailov, S.E.: Traces, extensions and co-normal derivatives for elliptic systems on
Lipschitz domains. J. Math. Analysis and Appl., 378, 324–342 (2011).
[Mir70] Miranda, C.: Partial Differential Equations of Elliptic Type, Second edition, Springer,
Berlin – Heidelberg – New York (1970).
[Néd01] Nédélec, J.-C.: Acoustic and Electromagnetic Equations, Springer-Verlag, New York
(2001).
[NP73] Nedelec, J., Planchard, J.: Une méthode variationelle d’éléments finis pour la résolu-
tion numérique d’un probléme extérieur dans R3 . RAIRO, 7, n. R3, 105–129 (1973).
Thermoelastic Plates with Arc-Shaped Cracks
1 Prerequisites
Bending theories of elastic plates describe the behavior of thin structures under ex-
ternal mechanical and thermal influences. Among the most accurate ones, Mindlin-
type models provide information not only on the bending and twisting moments gen-
erated in the body, but also on the transverse shear forces acting across the thickness.
In what follows we use a combination of variational and boundary integral equation
techniques to study the mathematical properties and solution of the theory proposed
in [ScTa93], when the plate is weakened by an arc-shaped crack. The corresponding
results in the absence of the temperature factor can be found in [ChCo06, ChCo00],
and [ChCo05].
Consider a homogeneous and isotropic elastic material that occupies a region
S̄ × [−h0 /2, h0 /2] in R3 , where S is a domain in R2 and h0 = const is such that
0 < h0 diam S. A generic point in the plate is written in terms of Cartesian coor-
dinates as x = (x, x3 ), x = (x1 , x2 ) ∈ S̄. The displacement vector at x at time t ≥ 0 is
denoted by v(x ,t) = (v1 (x ,t), v2 (x ,t), v3 (x ,t))T , where the superscript T indicates
matrix transposition, and the temperature is denoted by θ (x ,t).
In the plate model investigated in [Co90] it is assumed that
If thermal effects are also taken into account, then we must consider the “averaged
temperature moment”, defined by (see [ScTa93])
h0 /2
1 h20
u4 (x,t) = x3 θ (x, x3 ,t) dx3 , h2 = .
h2 h0 12
−h0 /2
C. Constanda
The University of Tulsa, OK, USA,
e-mail: christian-constanda@utulsa.edu
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 129
DOI 10.1007/978-0-8176-8238-5_12, © Springer Science+Business Media, LLC 2011
130 I. Chudinovich and C. Constanda
The right-hand side in (1) has the form Q = (qT , q4 )T , where q = (q1 , q2 , q3 )T is a
combination of the forces and moments acting on the plate and its faces and q4 is a
combination of the averaged heat-source density, temperature, and heat flux on the
faces.
Without loss of generality [ChCoVe04], we assume that the initial conditions are
homogeneous; in other words,
Let ∂ S be a simple, closed C2 -curve that divides R2 into interior and exterior
domains S+ and S− , let ∂ S0 be an open connected arc of ∂ S which models a crack,
and let ∂ S1 = ∂ S \ ∂ S̄0 . Modifying the original meaning of the symbol S, we now
set S = R2 \ ∂ S̄0 and assume this infinite domain to be the middle-plane section of
the plate. For ν = 0, 1, we write
We restrict our attention to the initial-boundary value problem (TCD) that con-
sists in finding the solution of system (1) satisfying initial conditions (2) and the
edge-of-the-crack Dirichlet boundary conditions
where the superscripts + and − denote the limiting values of the corresponding
functions as (x,t) tends to Γ from inside Σ + or Σ − , respectively. The vector-valued
functions F ± = (( f ± )T , f4± )T , f ± = ( f1± , f2± , f3± )T , are prescribed.
As a first step in the solution of the variational, or weak, formulation of (TCD),
we consider the boundary value problem (TCD p ) to which the former is reduced by
the application of the Laplace transformation with respect to t.
We denote the Laplace transform of a function by a superposed hat and the (com-
plex) transformation parameter by p.
Let Cl (S̄), l = 0, 1, 2, . . . , be the space of functions U defined in both S̄+ and S̄−
and satisfying
where the superscripts ± indicate the limiting values of the corresponding functions
as x tends to ∂ S from inside S± . We remark that U and its derivatives may have
jump discontinuities across arc ∂ S0 .
The classical problem (TCD p ) depending on the complex parameter p consists
in finding Û ∈ C2 (S) ∩C(S̄), where C(S̄) = C0 (S̄), such that
Hm,p (R2 ) = Hm,p (R2 ) × Hm (R2 ), with the norm of the elements V̂ = (v̂T , v̂4 )T
defined by |V̂ |m,p = v̂m,p + v̂4 m ;
132 I. Chudinovich and C. Constanda
Hm (S± ) and Hm,p (S± ) consist of the restrictions to S± of all v̂4 ∈ Hm (R2 ) and
v̂ ∈ Hm,p (R2 ), respectively, with norms
Hm,p (S± ) = Hm,p (S± ) × Hm (S± ), with the norm of Û = (ûT , û4 )T given by
|Û|m,p;S± = ûm,p;S± + û4 m;S± .
If p = 0, then we write
Hm (R2 ) = Hm,0 (R2 ) = [Hm (R2 )]3 , Hm (R2 ) = Hm (R2 ) × Hm (R2 ) = [Hm (R2 )]4 ,
Hm (S± ) = Hm,0 (S± ) = [Hm (S± )]3 , Hm (S± ) = Hm (S± ) × Hm (S± ) = [Hm (S± )]4 .
The norms on [Hm (R2 )]n and [Hm (S± )]n are denoted by the same symbols · m and
· m;S± , respectively, for all n = 1, 2, . . . , as this does not cause any ambiguity.
We now introduce spaces of functions defined on ∂ S.
H1/2 (∂ S) and H1/2,p (∂ S) are the spaces of the traces (see [LiMa72]) on ∂ S of all
û4 ∈ H1 (S± ) and û ∈ H1,p (S± ), with norms
H1/2,p (∂ S) = H1/2,p (∂ S) × H1/2 (∂ S), with the norm of F̂ = ( fˆT , fˆ4 )T given by
|F̂|1/2,p;∂ S = fˆ1/2,p;∂ S + fˆ4 1/2;∂ S .
For simplicity, we denote by the same symbols γ ± the continuous (uniformly
with respect to p ∈ C) trace operators from H1 (S± ) to H1/2 (∂ S), from H1,p (S± ) to
H1/2,p (∂ S), and from H1,p (S± ) to H1/2,p (∂ S). If p = 0, then we write H1/2 (∂ S) =
H1/2,0 (∂ S) = [H1/2 (∂ S)]3 and H1/2 (∂ S) = H1/2 (∂ S) × H1/2 (∂ S) = [H1/2 (∂ S)]4 .
The norms on [H1/2 (∂ S)]n are denoted by · 1/2;∂ S for all n = 1, 2, . . . .
If Û(x,t) = (ûT (x,t), û4 (x,t))T is defined for x ∈ S, then its restrictions to S± are
denoted by Û± = (ûT± , û±,4 )T , and we write û4 = {û+,4 , û−,4 }, û = {û+ , û− }, and
Û = {Û+ , Û− }. Let πν , ν = 0, 1, be the operators of restriction from ∂ S to ∂ Sν , and
let γν± = πν γ ± be the trace operators on ∂ Sν .
H1 (S), H1,p (S), and H1,p (S) are the spaces of all û4 = {û+,4 , û−,4 }, û = {û+ , û− },
and Û = {Û+ , Û− } such that û±,4 ∈ H1 (S± ), û± ∈ H1,p (S± ), Û± ∈ H1,p (S± ), and
γ1+ û+,4 = γ1− û−,4 , γ1+ û+ = γ1− û− , γ1+Û+ = γ1−Û− . The norms on H1 (S), H1,p (S), and
H1,p (S) are û4 1;S = û+,4 1;S+ +û−,4 1,p;S− , û1,p;S = û+ 1,p;S+ +û− 1,p;S− ,
and |Û|1,p;S = |Û+ |1,p;S+ + |Û− |1,p;S− .
H̊1 (S), H̊1,p (S), and H˚1,p (S) denote the subspaces of H1 (R2 ), H1,p (R2 ), and
H1,p (R2 ) of all û4 , û, and Û such that γ0+ û+,4 = γ0− û−,4 = 0, γ0+ û+ = γ0− û− = 0,
and γ0+Û+ = γ0−Û− = 0.
H̊−1 (S̄) and H−1 (S), H̊−1,p (S̄) and H−1,p (S), H˚−1,p (S̄) and H−1,p (S) are the
duals of H1 (S) and H̊1 (S), H1,p (S) and H̊1,p (S), H1,p (S) and H˚1,p (S), respectively,
defined by the duality generated by the inner product (· , ·)0 in the corresponding
Thermoelastic Plates with Arc-Shaped Cracks 133
spaces [L2 (S)]n . The norms on these spaces are denoted by · −1;S̄ and · −1;S ,
·−1,p;S̄ and ·−1,p;S , |·|−1,p;S̄ and |·|−1,p;S . On H−1,p (S) it is also convenient
to use the equivalent norm
where
1/2 1/2
ϒp (Û, Ŵ ) = a(û, ŵ) + (∇û4 , ∇ŵ4 )0 + p2 (B0 û, B0 ŵ)0
+ κ −1 p(û4 , ŵ4 )0 − h2± ϖ (û4 , div ŵ)0 + η p(div û, ŵ4 )0 , (6)
a(û, ŵ) = 2 E(û, ŵ) dx,
S
B0 = diag{ρ h2 , ρ h2 , ρ },
2E(û, ŵ) = h2 E0 (û, ŵ) + h2 μ (∂2 û1 + ∂1 û2 )(∂2 ŵ¯ 1 + ∂1 ŵ¯ 2 )
+ μ [(û1 + ∂1 û3 )(ŵ¯ 1 + ∂1 ŵ¯ 3 ) + (û2 + ∂2 û3 )(ŵ¯ 2 + ∂2 ŵ¯ 3 )],
134 I. Chudinovich and C. Constanda
E0 (û, ŵ) = (λ + 2μ )[(∂1 û1 )(∂1 ŵ¯ 1 ) + (∂2 û2 )(∂2 ŵ¯ 2 )]
+ λ [(∂1 û1 )(∂2 ŵ¯ 2 ) + (∂2 û2 )(∂1 ŵ¯ 1 )].
L L L L L
H1/2,l, κ (∂ S0 ), H1/2,k,κ (∂ S0 ), and H1/2,l,k,κ (∂ S0 ) = H1/2,k,κ (∂ S0 )×H1/2,l,κ (∂ S0 )
are the spaces of elements fˆ4 (x, p), fˆ(x, p), and F̂(x, p) that
(i) define holomorphic mappings fˆ4 (· , p) : Cκ → H1/2 (∂ S0 ), fˆ(· , p) : Cκ →
[H1/2 (∂ S0 )]3 , and F̂(· , p) : Cκ → [H1/2 (∂ S0 )]4 ;
(ii) have norms defined by
∞
fˆ4 21/2,l,κ ;∂ S0 = sup (1 + |p|2 )l fˆ4 (x, p)21/2;∂ S0 d τ < ∞,
σ >κ
−∞
∞
fˆ21/2,k,κ ;S = sup (1 + |p|2 )k fˆ(x, p)21/2,p;∂ S0 d τ < ∞,
σ >κ
−∞
L L ˚L L L
H̊1/2,l, κ (∂ S0 ), H̊1/2,k,κ (∂ S0 ), and H1/2,l,k,κ (∂ S0 ) = H̊1/2,k,κ (∂ S0 )× H̊1/2,l,κ (∂ S0 )
consist of elements δ fˆ4 (x, p), δ fˆ(x, p), and δ F̂(x, p) that
(i) define holomorphic mappings
∞
δ fˆ21/2,k,κ ;∂ S = sup (1 + |p|2 )k δ fˆ(x, p)21/2,p;∂ S d τ < ∞,
σ >κ
−∞
L −1 L −1 L −1 −1
˚L
Let H1,l,k, κ (Σ ), H−1,l,k,κ (Σ ), H1/2,l,k,κ (Γ0 ), and H1/2,l,k,κ (Γ0 ), κ > 0, l, k ∈ R,
L
be the spaces of the inverse Laplace transforms of the elements of H1,l,k, κ (S),
H L (S), H L (∂ S0 ), and H˚ L (∂ S0 ), endowed with the norms
−1,l,k,κ 1/2,l,k,κ 1/2,l,k,κ
136 I. Chudinovich and C. Constanda
We continue to use the symbols γν± , ν = 1, 2, for the trace operators from Σ ± to
Γν , and the symbols πν for the operators of restriction from Γ to Γν . We denote by
C0∞ (Σ ) the subspace of C0∞ (R̄+
3 ) consisting of all W such that W (x,t) = 0, (x,t) ∈ Γ0 .
L −1 (Σ ) a weak solution of (TCD) if
We call U = (uT , u4 )T ∈ H1,0,0, κ
(i) γ0 u = 0, where γ0 is the trace operator on S × {t = 0};
(ii) γ0+U = F + and γ0−U = F − ;
∞
(iii) for all W = (wT , w4 )T ∈ C0∞ (Σ ), ϒ (U,W ) = (Q,W )0 dt, where
0
∞
1/2 1/2
ϒ (U,W ) = {a(u, w) + (∇u4 , ∇w4 )0 − (B0 ∂t u, B0 ∂t w)0
− κ −1 (u4 , ∂t w4 )0 − h2 ϖ (u4 , div w)0 − η (div u, ∂t w4 )0 } dt. (7)
0
Theorem 3. Let U(x,t) = L −1Û(x, p) be the inverse Laplace transform of the weak
L −1
solution Û(x, p) of problem (TCD p ). If κ > 0, l ∈ R, and Q ∈ H−1,l+1,l, κ (Σ ) and
± L −1 − L −1
F ∈H (Γ0 ) are such that δ F = F − F ∈ H
1/2,l+1,l+1,κ
+ ˚ (Γ0 ), then
1/2,l+1,l+1,κ
−1
L (Σ ) and
U ∈ H1,l,l, κ
C0∞ (S̄) is the space of pairs of functions Û = {Û+ , Û− }, where Û+ ∈ C∞ (S̄+ ),
Û− ∈ C0∞ (S̄− ), and their limiting values and those of all their derivatives from inside
S+ and S− coincide on ∂ S1 .
Let Û be a classical solution of (TCN p ). Choosing an arbitrary Ŵ ∈ C0∞ (S̄), mul-
tiplying the equation in (9) by Ŵ in [L2 (S)]4 , and integrating the result over S (or,
what is the same, over R2 ), we arrive at
Theorem 4. For any κ > 0, p ∈ Cκ , Q̂ ∈ H˚−1,p (S), and Ĝ± ∈ H−1/2,p (∂ S0 ) such
that δ Ĝ = Ĝ+ − Ĝ− ∈ H˚−1/2,p (∂ S0 ), problem (TCN p ) has a unique weak solution
Û ∈ H1,p (S) and
L L
H−1/2,l, κ (∂ S0 ), H−1/2,k,κ (∂ S0 ), and
L L L
H−1/2,l,k, κ (∂ S0 ) = H−1/2,k,κ (∂ S0 ) × H−1/2,l,κ (∂ S0 )
are the spaces of all ĝ4 (x, p), ĝ(x, p), and Ĝ(x, p) that
(i) define holomorphic mappings
∞
ĝ4 2−1/2,l,κ ;∂ S0 = sup (1 + |p|2 )l ĝ4 (x, p)2−1/2;∂ S0 d τ < ∞,
σ >κ
−∞
∞
ĝ2−1/2,k,κ ;S = sup (1 + |p|2 )k ĝ(x, p)2−1/2,p;∂ S0 d τ < ∞,
σ >κ
−∞
|Ĝ|−1/2,l,k,κ ;∂ S0 = ĝ−1/2,k,κ ;∂ S0 + ĝ4 −1/2,l,κ ;∂ S0 .
L L
H̊−1/2,l, κ (∂ S0 ), H̊−1/2,k,κ (∂ S0 ), and
L L L
H˚−1/2,l,k, κ (∂ S0 ) = H̊−1/2,k,κ (∂ S0 ) × H̊−1/2,l,κ (∂ S0 )
are the spaces of all δ ĝ4 (x, p), δ ĝ(x, p), and δ Ĝ(x, p) that
(i) define holomorphic mappings δ ĝ4 : Cκ → H̊−1/2 (∂ S0 ), δ ĝ(x, p) : Cκ →
[H̊−1/2 (∂ S0 )]3 , and δ Ĝ(x, p) : Cκ → [H̊−1/2 (∂ S0 )]4 ;
(ii) have norms
∞
δ ĝ4 2−1/2,l,κ ;∂ S = sup (1 + |p|2 )l δ ĝ4 (x, p)2−1/2;∂ S d τ < ∞,
σ >κ
−∞
∞
δ ĝ2−1/2,k,κ ;∂ S = sup (1 + |p|2 )k δ ĝ(x, p)2−1/2,p;∂ S d τ < ∞,
σ >κ
−∞
|δ Ĝ|−1/2,l,k,κ ;∂ S = δ ĝ−1/2,k,κ ;∂ S + δ ĝ4 −1/2,l,κ ;∂ S .
L
Theorem 5. If κ > 0, l ∈ R, and Q̂ ∈ H˚−1,l+1,l, ± L
κ (S) and Ĝ ∈ H−1/2,l+1,l,κ (∂ S0 )
are such that δ Ĝ = Ĝ+ − Ĝ− ∈ H˚ L (∂ S0 ), then the weak solution Û of
−1/2,l+1,l,κ
L (S) and
problem (TCN p ) belongs to H1,l,l, κ
2
|Û|1,l,l,κ ;S ≤ c |Q̂|−1,l+1,l,κ ;S̄ + |δ Ĝ|−1/2,l+1,l+1,κ ;∂ S
3
+ |Ĝ− |−1/2,l+1,l+1,κ ;∂ S0 .
C0∞ (Σ̄ ) is the space of all U = {U+ ,U− } such that U+ ∈ C0∞ (Σ̄ + ), U− ∈ C0∞ (Σ̄ − ),
and γ1+ ∂ α U+ = γ1− ∂ α U− for all multi-indices α . We say that a vector distribution
L −1 (Σ ) is a weak solution of (TCN) if
U = (uT , u4 )T ∈ H1,0,0, κ
140 I. Chudinovich and C. Constanda
∞
ϒ (U,W ) = (Q,W )0 dt + L(W ),
0
References
[Co90] Constanda, C.: A Mathematical Analysis of Bending of Plates with Transverse Shear
Deformation, Longman, Harlow (1990).
[ScTa93] Schiavone, P., Tait, R.J.: Thermal effects in Mindlin-type plates. Quart. J. Mech.
Appl. Math., 46, 27–39 (1993).
[ChCo06] Chudinovich, I., Constanda, C.: Potential representations of solutions for dynamic
bending of elastic plates weakened by cracks. Math. Mech. Solids, 11, 494–512
(2006).
[ChCo00] Chudinovich, I., Constanda, C.: Variational and Potential Methods in the Theory
of Bending of Plates with Transverse Shear Deformation, Chapman & Hall/CRC,
Boca Raton, FL (2000).
[ChCo05] Chudinovich, I., Constanda, C.: Variational and Potential Methods for a Class of
Linear Hyperbolic Evolutionary Processes, Springer, London (2005).
[ChCoVe04] Chudinovich, I., Constanda, C., Colín Venegas, J.: The Cauchy problem in the the-
ory of thermoelastic plates with transverse shear deformation. J. Integral Equations
Appl., 16, 321–342 (2004).
[LiMa72] Lions, J.-L., Magenes, E.: Non-homogeneous Boundary Value Problems and Appli-
cations, vol. 1, Springer, Berlin (1972).
Almost Periodicity in Semilinear Systems
C. Corduneanu
In our paper [Co11], the linear system, associated to (S), has been investigated:
where A stands for a constant n by n matrix, with complex entries, while f (t) ∈
APr (R, C n ). The following result has been proven in [Co11].
Theorem 1. Assume the following conditions are satisfied by the system (1):
Then, for each f ∈ APr (R, C n ), 1 ≤ r ≤ 2, there exists a unique solution of (1),
x(t) ∈ APr (R, C n ). Moreover, there exists M > 0, depending only on A and r such
that
|x(t)|r ≤ M| f |r . (4)
The proof is provided in our paper [Co11], except for the inequality (4), which
follows from the estimate (4.20) in [Co11]. Namely, it has been proven in [Co11]
that
|xk |r ≤ m| fk |r , k ≥ 1, (5)
where
∞ ∞
f∼ ∑ fk exp(iλkt), x∼ ∑ xk exp(iλkt). (6)
k=1 k=1
C. Corduneanu
The University of Texas at Arlington, TX, USA, and the Romanian Academy,
e-mail: concord@uta.edu
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 141
DOI 10.1007/978-0-8176-8238-5_13, © Springer Science+Business Media, LLC 2011
142 C. Corduneanu
which implies, taking into account the norm in APr (R, C n ), the inequality (5), with
M = m1/r .
We will consider now the system (S) in APr (R, C n ), and use Theorem 1 and the
contraction mapping principle to obtain the existence and uniqueness of solution.
The following hypotheses will be imposed on the “nonlinear” part f x:
(c) f : APr (R, C n ) → APr (R, C n ) satisfies the Lipschitz-type condition
Remark 1. Since for any function x ∈ AP2 (R, C n ) ⊃ APr (R, C n ), 1 ≤ r < 2, it follows
that any solutions of (S) will satisfy the system in Carathéodory’s sense, i.e., almost
everywhere. See our book [Co09], where AP2 (R, C n ) is shown to consist of locally
square integrable functions and, therefore ẋ ∈ AP2 (R, C ) enjoys the same property.
Proof of Theorem 2. One considers the operator T : APr (R, C ) into itself, by letting
T x = y, with x and y related by the equation
ẏ = Ay + ( f x)(t), (9)
with y ∈ APr (R, C ) the unique solution of (9) in APr (R, C ), according to Theorem 1.
If T x = y and Tu = z, then v = y − z satisfies
which is linear in v and of the form (1). Applying (4) to (10), and taking into
account (8), one obtains the inequality
λ < M −1 , (12)
3 An Integro-Differential System
If one assumes that f ∈ APr (R, C n ), then it appears natural to look for a solution
x ∈ APr (R, C ) to (13). In [Co11] we have established the inequalities
in case 1 < r ≤ 2.
Let us rewrite (13) in the form
with k ∗ x given by (16). We now substitute into (13) the Fourier series, which leads
to the following identity:
∞
∞
iλ j x j exp(iλ j t) ∼ ∑ k(s) exp(−iλ j s)ds x j exp(iλ j t) + ∑ f j exp(iλ j t).
j=1 R j=1
and finally
iλ j − k(s) exp(−iλ j s)ds x j = f j , j ≥ 1. (17)
R
each x j is uniquely determined, for any f ∈ APr (R, C n ), i.e., for any sequence
{λ j } ∈ R.
The condition (d) can be rewritten in the form
4 5
(d) det iω I − 0
k(iω ) ≥ m > 0, ω ∈ R, (18)
where 0
k(iω ) is the Fourier transform of the kernel k in (13) or (13) :
0
k(iω ) = k(s) exp{−iω s}ds, ω ∈ R.
R
|x j | ≤ m−1 | f j |, j ≥ 1. (19)
At this point in the discussion, we are proceed as in Theorem 1 and conclude that
f ∈ APr (R, C n ) implies x ∈ APr (R, C n ), where f and x are like in (13) . Moreover,
an estimate of the form (4) is valid, with M = m−1/r .
Summarizing the above discussion on system (13) , we can state the following
result.
(d) | det[iω I − 0
k(iω )]| ≥ m > 0, ω ∈ R.
(e) |k| ∈ L1 (R, R), f ∈ APr (R, C n ).
Remark 3. It is obvious that condition (d) above is the equivalent of condition (a) in
Theorem 1, serving the same purpose. It is more complex, and it would be interest-
ing to see whether or not it can be replaced by its weaker counterpart
| det(iω I − 0
k(iω ))| > 0, ω ∈ R. (20)
Actually, (20) must be valid only when ω runs over the set {λk ; k ≥ 1} of Fourier
exponents of f . When f is fixed in APr (R, C n ), we deal with much less than required
in (20). If we want the result to be valid for any f ∈ APr (R, C n ), then (20) must be
verified in full, due to the arbitrariness of the Fourier exponents in case of almost
periodic functions.
Let us consider this problem. Since |0 k(iω )| → 0 as |ω | → ∞, on behalf of as-
sumption (e) in Theorem 3, and due to the continuity of the value of a determinant
with respect to its entries, we can write
| det(iω I − 0
k(iω ))| > m1 , |ω | > M1 . (21)
Indeed, the dominating term is |ω |n , and (21) follows immediately. But in the closed
interval |ω | ≤ M1 , det(iω I − 0
k(iω )) is a continuous function of ω , and therefore,
having | det(iω I − 0k(iω ))| > 0, there is a positive minimum of this quantity, say
m2 > 0. These estimates lead to the conclusion
inf | det(iω I − 0
k(iω ))| = m > 0, ω ∈ R,
with m = min{m1 , m2 }. Hence, condition (d) in Theorem 3 can be replaced with its
weaker form (20).
with k ∗ x defined by (16), and f x as in Theorem 1. Let us formulate now the condi-
tion guaranteeing the existence and uniqueness of solution in APr (R, C n ).
|x(t)|r ≤ M| f |r , (23)
146 C. Corduneanu
for f and x related by (13) . If we denote by T the operator from APr (R, C n ) into
itself, defined by ẏ(t) = (k ∗ y)(t) + ( f x)(t), t ∈ R, then for v = y − z = T x − Tu one
obtains v̇ = (k ∗ v) + ( f x)(t) − ( f u)(t), which leads to the inequality
after using the Lipschitz condition (8). For λ satisfying (12), one obtains the result
of Theorem 4, which ends the proof.
Remark 4. In dealing with the system (S), we have considered two particular choices
for the linear operator A: the matrix case and the convolution type operator.
A problem arising in this context is to find other linear operators acting on
APr (R, C n ), such that the linear equation ẋ(t) = (Ax)(t) + f (t) has unique solu-
tion in APr (R, C n ), for each f in the same space. Then, the associated semilinear
equation will enjoy the same property.
Remark 5. It is certain that the equations we have investigated possess other solu-
tions, besides the almost periodic one. It is interesting to build up such solutions and
compare them with the one already emphasized by the above results.
det(I − 0
k(iω )) = 0, ω ∈ R.
References
[Co11] Corduneanu, C.: A scale of almost periodic functions spaces. Differential and Integral
Equations, 24 (2011), 1–27.
[Co09] Corduneanu, C.: Almost Periodic Oscillations and Waves, Springer, New York (2009).
[Co73] Corduneanu, C.: Integral Equations and Stability of Feedback Systems, Academic Press,
New York (1973).
Bubble Behavior Near a Two Fluid Interface
1 Introduction
The influence of rigid and free boundaries on the dynamics of bubbles has been re-
searched extensively, both experimentally and theoretically. Experiments by [Be66]
showed that the presence of a solid boundary caused the formation of a liquid jet
through the bubble, forming a toroidal bubble. This behavior has been observed in
many other experiments since, including [Br02, Pi98, To86, La75]. Similar behav-
ior is also observed when a bubble collapses near a free surface. In such conditions
bubble jetting may be directed away from the surface, with a counter-jet forming out
of the free surface. Experiments using spark-generated bubbles by [Bl81] under free
fall conditions and [Ch80] in standard gravity showed this counter-jet to be greatly
influenced by the standoff distance. Bubbles formed very close to the surface gener-
ate severe vertical surface spikes, and those at greater distances create much smaller
and smoother deformations to the surface.
Numerical investigations into both of these phenomena have been undertaken us-
ing boundary integral methods to great effect. Simply connected simulations were
first carried out by [Le76], and [Ta85, Bl86] into cavitation collapse near rigid
boundaries. These results showed jet direction in the absence of gravity to be di-
G.A. Curtiss
University of Birmingham, UK,
e-mail: geoffc@nag.co.uk
D.M. Leppinen
University of Birmingham, UK,
e-mail: d.m.leppinen@bham.ac.uk
Q.X. Wang
University of Birmingham, UK,
e-mail: q.x.wang@bham.ac.uk
J.R. Blake
University of Birmingham, UK,
e-mail: j.r.blake@bham.ac.uk
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 147
DOI 10.1007/978-0-8176-8238-5_14, © Springer Science+Business Media, LLC 2011
148 G.A. Curtiss et al.
rected toward the wall, and that the expansion phase of the bubbles is of importance
for the later collapse phase, in spite of the high level of sphericity observed about
the maximum bubble volume. Simulations of behavior near free surfaces were also
carried out by [Ta85, Bl87] with some success. More accurate implementations have
since modeled pre-toroidal behavior near free surfaces with a high degree of accu-
racy [Wa96, Pe03, Ro01].
This paper seeks to investigate the behavior of high pressure cavitation bubbles
in the vicinity of a two fluid interface, that is, the interface between two immis-
cible liquids of different densities. Where appropriate this is continued into the
toroidal phase using the vortex ring method of [Lu91]. Previously [Kl04a] have
used a boundary integral approach to examine the dynamics of a single bubble near
a density interface where the authors decoupled the normal velocity of the fluid–
fluid interface from the discretized matrix equations using geometrical arguments
derived from the equal density case. This does not decrease the time required to cal-
culate the Green’s function block matrices, yet it does significantly decrease the size
of the matrix to be inverted. Their results showed various types of behavior in the
presence of gravity, including jetting toward and away from the interface, and bubble
pinching near the null impulse line, in agreement with the model presented herein.
A further advancement has been presented to account for a linearly elastic fluid in
the uncavitated layer, by means of a displacement parameter on the fluid–fluid in-
terface [Kl04b]. The results obtained are in agreement with experiments involving
PAA samples by [Br01a, Br01b].
In the current work we extend the results of [Kl04a] by implementing a mathe-
matical model to examine the evolution of multiple bubbles on either side of the in-
terface without making any unnecessary assumptions about the fluid flow. In Sect. 2
we introduce the mathematical model, with details of the numerical implementation
given in Sect. 3. The model validation is briefly discussed in Sect. 4 and results are
presented in Sect. 5. The overall conclusions are made in Sect. 6.
Fig. 1 (a) A sketch of bubbles interacting with a density interface with fluid of density ρ1 in fluid 1
underlying fluid of density ρ2 in fluid 2. (b) The axisymmetric geometry employed in the numerical
implementation
Bubble Behavior Near a Two Fluid Interface 149
2 Mathematical Formulation
The problem we are considering is sketched in Fig. 1(a) with a layer of fluid of
density ρ1 underlying a layer of fluid of density ρ2 with ρ1 > ρ2 . In each layer the
fluid is assumed to be inviscid, incompressible and irrotational so that the velocity
field can be written as the gradient of a potential with
ui = ∇φi (1)
for i = 1, 2 where
∇ · ui = ∇2 φi = 0 (2)
and
∇ × ui = 0. (3)
There is at least one bubble in either the upper or lower layer, with an arbitrary
number of additional bubbles. Our objective is to examine the interaction of the
bubbles with the density interface and with each other.
With the above assumptions Bernoulli’s equation is valid everywhere in both
exterior fluids which gives
∂ φi |ui |2 p p∞
+ + + g(z − zo ) = (4)
∂t 2 ρi ρi
where t is time, p is pressure, g is the acceleration of gravity, z is the vertical coordi-
nate of a fluid particle with reference to the initial vertical coordinate of the interface
zo and p∞ is the far field pressure at zo .
The exterior fluid layers are taken as isothermal. Furthermore it is assumed that
no mass or heat transfer will occur across the gas/liquid boundaries. Thus the bubble
gases obey an adiabatic law. The pressure inside each bubble is given by
γ
V0
pb = pv + p0 , (5)
V
where pv represents the condensible gas vapor pressure, V, V0 are the bubble volume
and initial reference volume, respectively, p0 is the initial reference pressure for the
bubble at reference volume and γ is the ratio of specific heats of the incondensible
gas (taken as 1.4 in this paper). The pressure of the surrounding liquid at the bubble
surface is given by the Young–Laplace condition as
p = pb − σb,i ∇ · nb (6)
where σb,i is the surface tension of a bubble in fluid i and nb is the outward pointing
unit normal to the bubble. The pressure jump across the density interface is also
calculated using the Young–Laplace condition as
p1 = p2 − σI ∇ · n1 (7)
150 G.A. Curtiss et al.
where σI is the surface tension of the density interface and n1 is the unit normal
pointing from fluid 1 into fluid 2. In this paper p∞ is taken as constant, and the early
bubble behavior observed is the result of an initially high internal gas pressure,
p0 p∞ − pv . This is valid for spark and laser generated bubbles, where the initial
pressure may reach many thousands of atmospheres.
We proceed by non-dimensionalizing our equations using the following scales:
the density of the lower layer, ρ1 , for density; the difference in the far field pres-
sure and the condensible gas vapor pressure, Δ p = p∞ − pv , for pressure; the maxi-
mum radius a bubble would reach in an infinite fluid of density ρ1 , Rmax for length;
Rmax (ρ1 /Δ p)1/2 for time; (Δ p/ρ1 )1/2 /Rmax for potential; Rmax Δ p for surface ten-
sion. Thus along a bubble interface in fluid 1 the Bernoulli equation becomes
γ
∂ φ1 |u1 |2 V0
+ = 1−ε − δ (z − z0 ) + σb,1 ∇ · nb (8)
∂t 2 V
where ρ = ρ1 /ρ2 . Along the density interface we use (4) and (7) to obtain
∂ φ1 |u1 |2 ∂ φ2 |u2 |2
+ −ρ + = δ z (1 + ρ ) − σI ∇ · n1 . (10)
∂t 2 ∂t 2
3 Numerical Implementation
The general problem considered in Fig. 1(a) is fully three-dimensional. For compu-
tational convenience we consider the axisymmetric problem sketched in Fig. 1(b)
where (b, c) refer to points on a bubble surface in fluid 1, (d, e) refer to points on a
bubble surface in fluid 2 and (p, q) refer to points on the interface between fluids 1
and 2. The substantial derivatives on all surfaces are given by
D ∂
= + u · ∇, (11)
Dt ∂t
where u = ui on the bubble surface in fluid i and u = (u1 + u2 )/2 on the two fluid
interface. Particles on the bubble surfaces and the density interface are treated as
material points and are advected according to
db dd dp u1 + u2
= u1 , = u2 , = . (12)
dt dt dt 2
For convenience we define
Bubble Behavior Near a Two Fluid Interface 151
as the density-weighted potential along the density interface which satisfies the evo-
lution equation
DF(p)
= (1 − ρ ) (u1 · u2 ) − 2(1 + ρ )δ z + 2σI ∇ · n1 . (14)
Dt
Along the bubble interfaces in fluid 1 we have
γ
Dφ1 (b) |u1 |2 V0
= +1−ε − δ (z − z0 ) + σb,1 ∇ · nb (15)
Dt 2 V
where G(x0 , x) is the free space Green’s function given by G(x0 , x) = 1/|x − x0 |,
∂ G(x0 , x)/∂ n is the normal derivative of the Green’s function with respect to the
outward normal of the fluid domain and x ∈ ∂ D. In our implementation all of the
surfaces are represented by quintic splines and are smooth so that
Equation (17) is appropriate in both fluid 1 and fluid 2. In fluid 1 ∂ D is the union
of the surfaces of all bubbles located in fluid 1 and of the interface separating fluid
1 and 2. Similarly, in fluid 2 ∂ D is the union of the surfaces of all bubbles located
in fluid 2 and of the interface separating fluid 1 and 2. The central portion of the
density interface is represented using nodal points. The truncated interface is then
extended to infinity using a least squares type approximation [Cu09] with all inte-
grals evaluated analytically along the infinite extension.
152 G.A. Curtiss et al.
The procedure for discretizing and solving the boundary integral equation given
in (17) is well known. Details of the exact numerical implementation, including
the non-trivial incorporation of the vortex ring method of [Lu91] for when a bub-
ble becomes toroidal is given in the thesis by Curtiss [Cu09]. Briefly, all quantities
of interest along the central portion of the density interface and along the bubble
surfaces are represented using quintic splines based upon an arc length formula-
tion. The logarithmic singularities associated with the boundary integral method
are removed analytically and all of the resultant integrals are approximated using
high-order (typically 20) Gaussian quadrature. In summary, when (17) is discretized
along each of the bubble surfaces in fluid 1 and 2, and along the density interface
between the fluids, we obtain the following system of equations:
⎡ ρ ⎤⎡ ∂ φ1 (b) ⎤
Gbc 0 −Gbq 1+ρ DGbq ∂ n1
⎢ 0 Gde Gdq −1 ⎥⎢ ∂ φ2 (d) ⎥
⎢ 1+ρ DGdq ⎥⎢ ∂ n2 ⎥
⎢ ρ −1 ⎥⎢ ⎥
⎣ pc
G G pe 0 ρ +1 DG pq − 2 π I pq ⎦ ⎣ ∂ φ2 (p) ⎦
ρ ∂ n 2
G pc −G pe −2G pq DG pq + 2π 1− 1+ρ I pq φ1 (p) + φ2 (p)
⎡ −1 ⎤
2π Ibc + DGbc 0 2+2ρ DGbq ⎡ ⎤
⎢ π −1 ⎥ φ1 (b)
⎢ 0 2 I + DG de 2+2ρ DG dq ⎥
⎥ ⎣ φ2 (d) ⎦
de
=⎢ −1 (20)
⎣ DG pc DG pe 1+ρ DG pq ⎦
F(p)
DG pc −DG pe 2π 1+1 ρ I pq
where G is used to indicate the discretized Green’s function and DG the discretized
normal derivative of the Green’s function. The block matrix structure is used to em-
phasize the bubble/bubble, bubble/interface and interface/interface contributions to
the overall result with reference to the notation introduced in Fig. 1(b). The RHS
of (17) is known at any time step and the resultant set of equations can be in-
verted to determine the normal velocities ∂ φ1 (b)/∂ n1 , ∂ φ2 (d)/∂ n2 , ∂ φ2 (p)/∂ n2
and the sum φ1 (p) + φ2 (p) and hence to determine the velocity along all inter-
faces.
4 Validation
The interaction between single/multiple bubbles and a density interface can be ex-
amined by time-stepping (12) and (14)–(16), with the interfacial velocities deter-
mined by inverting (20) and using splining techniques to calculate tangential ve-
locities. The current numerical implementation has been validated by comparison
with the case of ρ = 1 with σi = 0 (i.e. the case of a Rayleigh–Plesset bubble in
an infinite ambient liquid) and with previous experimental results for the interaction
of a bubble with a free surface [Bl81] (i.e. the case of ρ = 0) and a rigid bound-
ary [Br02] (i.e. the case of ρ = ∞). Full validation details are available [Cu09].
Arguably the best validation of the numerical implementation is with reference to
Bubble Behavior Near a Two Fluid Interface 153
Fig. 2 Comparisons between experiments near a water/white spirit interface with ρ = 0.76, δ =
0.0147 by [Ch80], used with permission, and the numerical code for h = 0.87 (top) and h = 2.2
(base). The final frames in both cases are where the simulation ends due to either non-axial jet
impact (top) or jet disconnection (base), and the resultant time of the final frames for h = 2.2 are
displaced. Other parameters are estimated as ε = 100, σb = 0.001
154 G.A. Curtiss et al.
the experiments of Chahine and Bovis [Ch80] of the interaction of a bubble with a
water/white spirit interface with a non-trivial value of ρ = 0.76 as seen in Fig. 2.
In this case a spark-generated centimeter-sized bubble is formed in the water layer.
Figure 2 compares two of these experiments with BIM simulations acting under
gravity with a buoyancy parameter of δ = 0.0147. With the shallow standoff of
h = 0.87, the simulation is halted after one oscillation as a non-axial jet impact
occurs. In the deeper standoff case of h = 2.2, a buoyancy driven jet forms, lead-
ing to an axial jet impact and is hence simulated further using the toroidal bubble
model. In both cases excellent agreement is seen between the experiment and the
simulation.
Fig. 3 The evolution of the bubble and fluid–fluid interface with h = 0.5 for density ratios ρ =
0, 0.1, 0.2, 0.3, 0.4 with the smaller density ratios showing greater surface deformation and counter-
jetting in the bubble
5 Results
The numerical model above has been used to examine the interaction of a bubble (or
bubbles) with a density interface. The direct output from (12) is the time evolution
of all of the free surface shapes. In addition, information concerning the pressure
field and the velocity field in fluids 1 and 2 can be obtained using (8), (9) and (17)
as detailed in [Cu09]. Below we will give some examples of the results that can be
obtained using the current numerical method. Figure 3 shows the time evolution of
Bubble Behavior Near a Two Fluid Interface 155
a single bubble interacting with a density interface. Figure 4 shows the interaction
of two bubbles below a density interface. Figure 5 shows the interaction of two
bubbles on opposite sides of a density interface. Figure 6 shows the interaction of
two bubbles below and one bubble above a density interface. In all cases the symbol
h refers to the standoff distance of the bubble which is distance of the centroid of
the initially spherical bubble from the fluid–fluid interface divided by the maximum
radius the bubble would obtain in an infinite quiescent ambient liquid given its initial
internal pressure.
Figure 3 shows the evolution of the density interface and bubble shape for the
case of one bubble with h = 0.5 for ρ = 0, 0.1, 0.2, 0.3 and 0.4. The case of ρ = 0
corresponds to the interaction of a bubble with a free surface where an intense up-
wards jet is formed on the density interface with a corresponding counter-jet on the
bubble surface. As the density ratio increases away from the free surface case, the
intensity of the free surface jet and the counter-jet decreases. The bubbles become
toroidal when the counter-jet pierces the lower face of the bubble. The simulations
proceed into the toroidal regime until they become numerically unstable. While it is
not shown here, simulations for ρ > 1 approach the known limit for the interaction
of a bubble with a rigid boundary in the limit as ρ → ∞.
Fig. 4 Collapse and re-expansion of two bubbles on the same side of a density interface with
ρ = 0.5 and h1 = 1.5, h2 = 4.5
156 G.A. Curtiss et al.
Figure 4 follows the evolution of two bubbles on the same side of a density
interface with ρ = 0.5 for the case of h1 = 1.5 and h2 = 4.5. We consider the
case of ε = 100 with γ = 1.4 so that initially the bubbles expand spherically to
their maximum radius and then they collapse. As they collapse the bubbles strongly
interact with each other. While the density interface remains predominantly hori-
zontal throughout jetting occurs in both bubbles during the collapse and then re-
expansion phase and ultimately both bubbles become toroidal. The relative strength
of the bubble–bubble interaction and the bubble–interface interactions depends on
the standoff distances and the density ratio.
Fig. 5 Two bubble interactions about an interface with ρ = 1.3. The initial standoff distances are
h1 = 0.75 in the lower layer and h2 = 1.5 in the top layer
An interesting interaction with one bubble on each side of the density interface
is shown for ρ = 1.3 in Fig. 5 where the standoff distance of the bubble in the lower
layer is h1 = 0.75 and the standoff distance of the bubble in the upper layer is h2 =
1.5. The deformation of both the upper and lower bubble is significantly different
at this value of ρ from what would be the case if either of the bubbles was absent.
The bubble–bubble interaction and the bubble–interface interaction depends on the
value of ρ and increases as the standoff distances decrease. For the case considered
the bubble in the upper layer is strongly deformed by the interaction. The upper
bubble rapidly becomes toroidal with the bubble jet impacting upon the density
Bubble Behavior Near a Two Fluid Interface 157
interface and deforming the lower bubble. This suggests a possible mechanism for
the enhancement of mixing across a density interface.
Fig. 6 Three bubble interaction with ε = 100, γ = 1.4, ρ = 1.3 and h1 = 1.5 in the upper layer and
h2 = 1.5 and h3 = 4.5 in the lower layer
The case of ρ = 1.3 with one bubble above (h1 = 1.5 in the upper layer) and two
bubbles below the density interface (h2 = 1.5, h3 = 4.5 in the lower layer) is shown
in Fig. 6. As expected for this value of ρ it is bubble in the upper layer which is most
strongly deformed. The downward liquid jet in the upper bubble deflects the density
interface downward and deforms the middle bubble. The lower bubble in the dense
layer interacts with the bubble above it to form a strong upward axial jet ultimately
becoming toroidal.
6 Conclusions
References
[Be66] Benjamin, T.B., Ellis, A.T.: The collapse of cavitation bubbles and the pressures thereby
produced against solid boundaries. Phil. Trans. R. Soc. Lond. A, 260, 221–240 (1966).
[Bl81] Blake, J.R., Gibson, D.C.: Growth and collapse of a vapour cavity near a free surface. J.
Fluid Mech., 111, 123–140 (1981).
[Bl86] Blake, J.R., Taib, B.B., Doherty, G.: Transient cavities near boundaries. Part 1. Rigid
boundary. J. Fluid Mech., 170, 479–497 (1986).
[Bl87] Blake, J.R., Taib, B.B., Doherty, G.: Transient cavities near boundaries. Part 2. Free
surface. J. Fluid Mech., 181, 197–212 (1987).
[Br02] Brujan, E.A., Keen, G.S., Vogel, A., Blake, J.R.: The final stage of the collapse of a
cavitation bubble close to a rigid boundary. Phys. Fluids, 14, 85–92 (2002).
[Br01a] Brujan, E.A., Nahen, K., Schmidt, P., Vogel, A.: Dynamics of laser-induced cavitation
bubbles near an elastic boundary. J. of Fluid Mech., 433, 251–281 (2001).
[Br01b] Brujan, E.A., Nahen, K., Schmidt, P., Vogel, A.: Dynamics of laser-induced cavitation
bubbles near elastic boundaries: Influence of the elastic modulus. J. Fluid Mech., 433,
283–314 (2001).
[Ch80] Chahine, G.L., Bovis, A.: Oscillation and collapse of a cavitation bubble in the vicinity
of a two-liquid interface. Springer Series in Electrophysics 4 - Cavitation and inhomo-
geneities in underwater acoustics, 23–29 (1980).
[Cu09] Curtiss, G.A.: Non-linear, non-spherical bubble dynamics near a two fluid interface, PhD
Thesis, The University of Birmingham (2009).
[Kl04a] Klaseboer, E., Khoo, B.C.: Boundary integral equations as applied to an oscillating bub-
ble near a fluid–fluid interface. Comp. Mech., 33, 129–138 (2004).
[Kl04b] Klaseboer, E., Khoo, B.C.: An oscillating bubble near an elastic material. J. Appl. Phys.,
96, 5808–5818 (2004).
[La75] Lauterborn, W., Bolle, H.: Experimental investigations of cavitation-bubble collapse in
the neighbourhood of a solid boundary. J. Fluid Mech., 72, 391–399 (1975).
[Le76] Lenoir, M.: Calcul numérique de l’implosion d’une bulle de cavitation au voisinage
d’une paroi ou d’une surface libre. J. Mécanique, 15, 725–751 (1976).
[Lu91] Lundgren, T.S., Mansour, N.N.: Vortex ring bubbles. J. Fluid Mech., 224, 177–196
(1991).
[Pe03] Pearson, A.: Hydrodynamics of jet impact in a collapsing bubble. PhD Thesis, The Uni-
versity of Birmingham (2003).
[Pi98] Phillip, A., Lauterborn, W.: Cavitation erosion by single laser-produced bubbles. J. Fluid
Mech., 361, 75–116 (1998).
[Ro01] Robinson, P.B., Blake, J.R., Kodama, T., Shima, A., Tomita, Y.: Interaction of cavitation
bubbles with a free surface. J. Appl. Phys., 89, 8225–8237 (2001).
[To86] Tomita, Y., Shima, A.: Mechanisms of impulsive pressure generation and damage pit
formation by bubble collapse. J. Fluid Mech., 169, 535–564 (1986).
[Ta85] Taib, B.B.: Boundary integral method applied to cavitation bubble dynamics. Phd Thesis,
The University of Wollongong (1985).
[Vo80] Vogel, A., Schweigner, P., Frieser, A., Asiyo, M.N., Birngruber, R.: Intraocular Nd:YAG
Laser surgery: Light tissue interaction, damage range, and reduction of collateral effects.
J. Quant. Elec., 26, 2240–2260 (1990).
[Wa96] Wang, Q.X., Yeo, K.S., Khoo, B.C., Lam, K.Y.: Nonlinear interaction between gas bub-
ble and free surface. Computers and Fluids, 25, 607–628 (1996).
Spectral Stiff Problems in Domains
with a Strongly Oscillating Boundary
D. Gómez
Universidad de Cantabria, Santander, Spain,
e-mail: gomezdel@unican.es
S.A. Nazarov
Institute for Problems in Mechanical Engineering, St. Petersburg, Russia,
e-mail: srgnazarov@yahoo.co.uk
E. Pérez
Universidad de Cantabria, Santander, Spain,
e-mail: meperez@unican.es
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 159
DOI 10.1007/978-0-8176-8238-5_15, © Springer Science+Business Media, LLC 2011
160 D. Gómez et al.
⎧
⎪
⎪ −AΔxU ε = λ ε U ε in Ω ,
⎪
⎪
⎪
⎪ −aε −t Δx uε = λ ε ε −1 uε in ωε ,
⎪
⎨
U ε = uε on Γ , (3)
⎪
⎪
⎪
⎪ A∂ν U ε = aε −t ∂ν uε on Γ ,
⎪
⎪
⎪ −t
⎩
aε ∂n uε = 0 on Γε .
In (3), A and a are two positive constants while ∂ν and ∂n denote the derivatives
along the outward normal vectors ν and n to the curves Γ and Γε , respectively, and t
is a parameter such that 0 ≤ t < 1. We study the asymptotic behavior, as ε → 0, of the
eigenvalues λ ε of (3), in the range of the low frequencies, and of their corresponding
eigenfunctions {U ε , uε }.
Stiff Problems in Domains with a Strongly Oscillating Boundary 161
− aε −t Δx uε = λ ε ε −t−m uε in ωε (4)
The parameters t and t + m reflect the relative stiffness of the band and the dead
weight of the band, respectively, in mechanical problems. These stiff problems have
been introduced in [Go06a] and are of interest, for instance, in the study of rein-
forcement problems for solid media and in vibrations for a two-phase system in
fluid mechanics.
A characterization of the limiting problems for the eigenpairs of (3)1 , (4), (3)3 ,
(3)4 , (3)5 , for the different values of t and m has been obtained in [Go06a] by means
of asymptotic expansions. In [Go06a], we also provide sharp bounds for conver-
gence rates of the eigenpairs {λ ε , {U ε , uε }} in the case where t = 1 and m = 0
by using the so-called inverse-direct reduction method. A different approach for
the eigenpairs is provided in [Go06b] for the case where t > 1 and m = 0 where,
in addition to the convergence, a complete asymptotic expansion for the eigen-
pairs has been obtained, and a connection of this problem with Wentzell problems
with small parameters has been shown. Asymptotics for the middle and high fre-
quencies have been considered in [GoNa11]. We refer to [Go06a] and [Go06b]
for a review of previous works in the literature on the subject which in fact are
scarce.
In contrast, this is the first time in the literature that the boundary homogeniza-
tion of a heavy stiff band surrounding a fixed domain is considered. For the homog-
enization of a soft oscillating band and a stationary problem, see [BuKo87]. For
homogenization of spectral problems in domains with oscillating boundary, let us
mention [Sa80, Na90, OlSh92] and [LoSa79]. See [Na07] in connection with the
elasticity operator and [Na08] for further references.
In this paper, we deal with the case where 0 ≤ t < 1 and m = 1 − t, namely with
problem (3) where the strip ωε is defined by (1) and (2). In Sect. 2 we consider
the shape of ωε in (1) and in Sect. 3 that in (2). We obtain the limiting problems
associated for each case and prove the convergence of the eigenvalues with conser-
vation of the multiplicity toward those of the limiting problems using the Rayleigh
quotient and contradictory arguments (see, for instance, Chapter III of [At84] and
[GoLo99]). Here, since we are dealing with variable domains we need extension
operators (10) for the proof. We show that, for 0 < t < 1 and m = 1 − t, the stiffer
band does not influence the limit behavior which is the same when t = 0, that is, the
band is only heavy and the density takes the value O(ε −1 ) in the band.
162 D. Gómez et al.
The variational formulation of (3) for (1) and (2) reads: Find λ ε and {U ε , uε } ∈
H 1 (Ωε ), {U ε , uε } = 0, satisfying
a
A ∇xU ε · ∇x G dx + t ∇x uε · ∇x g dx
Ω ε ωε
ε ε 1 ε
=λ U G dx + u g dx ∀{G, g} ∈ H 1 (Ωε ). (5)
Ω ε ωε
that is, the sequence of eigenvalues repeated according to their multiplicities. Let
On the other hand, on account of the continuity of the function h and of the
curvature κ, for a certain sufficiently small d > 0, there exist constants c, C1 , C2 and
C3 independent of ε such that
denotes the Jacobian of the transformation from (x1 , x2 ) to (ν , τ ). Here and in the
sequel, c,C,Ci denote different constants independent of ε . In (9), h(τ ) is hε (τ ) in
the case of (2).
For each F ∈ C∞ (Ω ), let us define the function {F, f˜} ∈ H 1 (Ωε ) where
Let us also introduce the following three inequalities which will be useful
throughout this paper:
T
Z(0) − 1 Z(t) dt ≤ T 1/2 Z L2 (0,T ) ∀Z ∈ H 1 (0, T ), T > 0, (12)
T 0
ε h(τ ) ε h(τ )
|u(ν , τ )−u(0, τ )|2 d ν ≤ ε 2 h(τ )2 |∂ν u(ν , τ )|2 d ν ∀u ∈ H 1 (ωε ) (13)
0 0
and
U −CU 2L2 (Ω ) + U −CU 2L2 (Γ ) ≤ C∇xU2L2 (Ω ) ∀U ∈ H 1 (Ω ) (14)
h
The two first inequalities are obtained from the Newton–Leibnitz formula (cf.
[Go06a] for instance). In order to prove the Sobolev type inequality (14), we con-
164 D. Gómez et al.
Taking into account that the eigenfunctions corresponding to λ̃0 = 0 are the con-
stants and that, for all U ∈ H 1 (Ω ), U − CU is orthogonal to the constants with the
scalar product defined by (24), the minimax principle ensures that
Ω |∇x G| Ω |∇xU|
2 dx 2 dx
λ̃1 = min
≤
,
Ω |G| dx + Γ h|G| d τ |U −CU |2 dx + Γ h|U −CU |2 d τ
2 2
G∈Ẽ0⊥ Ω
G=0
where Ẽ0⊥ denotes the space of the functions of H 1 (Ω ) which are orthogonal to the
constants.
We obtain the following bound for the eigenvalues of (3).
Lemma 1 Let λkε be the eigenvalues of (3) with 0 ≤ t < 1 and ωε be defined by (1).
For each fixed k ∈ N, and ε sufficiently small, we have
where the minimum is taken over all the subspaces Ek ⊂ H 1 (Ωε ) with dim Ek =
k + 1.
Let {λi }∞
i=0 be the eigenvalues of the Dirichlet problem in Ω and {Vi }i=0 the
∞
This inequality provides the right-hand side of (16). As regards the left-hand side
of (16), it is enough to prove it for k = 1, namely λ1ε > C > 0. Let {U1ε , uε1 } be an
eigenfunction of (5) corresponding to λ1ε and satisfying (6). Then,
Stiff Problems in Domains with a Strongly Oscillating Boundary 165
1
A∇xU1ε 2L2 (Ω ) + ∇x uε1 2L2 (ωε ) = λ1ε (18)
εt
and the proof is completed by showing that ∇xU1ε L2 (Ω ) > C > 0 for sufficiently
small ε .
Introducing in (6) the curvilinear coordinates in the integral in ωε and using
inequalities (9) and (13) and the fact that U1ε = uε1 on Γ , we have
1 ≤ C(U1ε 2L2 (Ω ) + U1ε 2L2 (Γ ) + uε1 2L2 (ωε ) + ε ∇x uε1 2L2 (ωε ) ). (19)
h
In order to estimate the two first terms on the right-hand side of (19), we use (14)
and the definition of CU1ε . Then,
U1ε 2L2 (Ω ) + U1ε 2L2 (Γ ) = U1ε −CU1ε 2L2 (Ω ) + U1ε −CU1ε 2L2 (Γ )
h h
ε 2
+ (CU1ε )2 [ Γ h d τ ] ≤ C∇xU1 L2 (Ω ) + (CU1ε ) [ h dτ ] .
dx + 2 dx +
Ω Ω Γ
Besides, taking into account the orthogonality condition of {U1ε , uε1 } to the constants
for the scalar product (7) (cf. (6)), (3)3 and inequalities (9) and (12) yields
1 ε 1
CU1ε =
hu1 d τ − u dx ≤ Cε 1/2 uε1 H 1 (ωε )
ε
Ω dx + Γ h d τ Γ ε ωε 1
and consequently
U1ε 2L2 (Ω ) + U1ε 2L2 (Γ ) ≤ C(∇xU1ε 2L2 (Ω ) + ε uε1 2H 1 (ωε ) ). (20)
h
Finally, from (19), (20), (18), the normalization condition (6) for {U1ε , uε1 } and
the boundedness of λ1ε , we deduce
1 ≤ C(∇xU1ε 2L2 (Ω ) + uε1 2L2 (ωε ) + ε ∇x uε1 2L2 (ωε ) ) ≤ C(∇xU1ε 2L2 (Ω ) + ε )
and so, for sufficiently small ε , ∇xU1ε 2L2 (Ω ) > 1/(2C). Therefore, the lemma is
proved.
Estimate (16) indicates the order of magnitude of the so-called low frequencies;
that is, for fixed k, λkε = O(1), and its asymptotic behavior as ε → 0 and that of
the corresponding eigenfunctions {Ukε , uεk } has been predicted by means of matched
asymptotic expansions in [Go06a] where the limiting spectral problem
1
−AΔxV = μ V in Ω ,
(21)
A∂ν V = μ hV on Γ ,
Problem (22) has a nonnegative discrete spectrum (see [Go06a] for details). Let
k→∞
0 = μ0 < μ1 ≤ μ2 ≤ · · · ≤ μk ≤ · · · −−−−−−→ ∞
be the eigenvalues of (22) with the usual convention of repeated eigenvalues. Let
{Vk }∞
k=0 be the corresponding eigenfunctions which are subject to the orthonormal-
ization condition
Vk Vl dx + hVk Vl d τ = δk,l (23)
Ω Γ
and form a basis in H 1 (Ω ). The eigenfunctions associated with μ0 = 0 are the con-
stants.
Let us denote by (·, ·)0 the scalar product defined by the left-hand side of (23) in
H 1 (Ω ), that is,
(V,W )0 = V W dx + hV W d τ ∀V,W ∈ H 1 (Ω ). (24)
Ω Γ
The main convergence result for the low frequencies of (3) with 0 ≤ t < 1 is
stated in the following theorem.
Theorem 1. Let λkε be the eigenvalues of (3) with 0 ≤ t < 1 and ωε defined by (1).
For each fixed k ∈ N ∪ {0}, the sequence λkε converges toward the eigenvalue μk
of (22) as ε → 0. In addition, for each sequence {Ukε , uεk } of eigenfunctions of (5),
{Ukε , uεk } satisfying the normalization condition (6), we can extract a subsequence
(still denoted by ε ) such that Ukε → Vk∗ weakly in H 1 (Ω ), as ε → 0, where Vk∗ is an
eigenfunction of (22) corresponding to μk and the set {Vk∗ }∞ k=0 forms an orthonor-
mal basis in H 1 (Ω ) for the scalar product defined by (24).
Proof. Taking into account (16), the orthonormalization condition (6) and the fact
that (λkε , {Ukε , uεk }) is an eigenelement of (5), we can extract a subsequence (still
denoted by ε ) such that for each k = 1, 2, 3 . . .
ε →0 ε →0
λkε −−−−−−→ λk∗ and Ukε −−−−−−→Vk∗ weakly in H 1 (Ω ),
where λk∗ > 0 and Vk∗ is certain function in H 1 (Ω ). Moreover, the following esti-
mates hold:
uεk 2L2 (ωε ) ≤ ε and ∇x uεk 2L2 (ωε ) ≤ Ck ε t . (25)
and, using (25), (9), (13) and the fact that uεk converges to Vk∗ strongly in L2 (Γ ),
we deduce that the three first terms on the right-hand side converge toward zero as
ε → 0. Then, we take limits in (6), as ε → 0, and we get
Vk∗Vl∗ dx + hVk∗Vl∗ d τ = δk,l .
Ω Γ
In order to identify (λk∗ ,Vk∗ ), for any fixed W ∈ C∞ (Ω ), we consider the vari-
ational formulation (5) for λ ε = λkε , U ε = Ukε , uε = uεk and the test function
{G, g} = {W, w̃} where w̃ is defined by (10), and introduce the curvilinear coor-
dinates in ωε :
ε h(τ )
a
A ∇xUkε · ∇xW dx + t ∂τ uεk ∂τ w̃K −1 d ν d τ
Ω ε Γ 0
ε h(τ )
ε ε 1 ε
= λk Uk W dx + uk w̃K d ν d τ . (26)
Ω ε Γ 0
Taking into account estimates (9), (11), (13) and (25) and that 0 ≤ t < 1, we pass to
the limit in (26) and obtain
A ∇xVk∗ · ∇xW dx = λk∗ Vk∗W dx + hVk∗W d τ . (27)
Ω Ω Γ
and taking into account the uniqueness of problem (30) we identify Φ as the eigen-
function V1 corresponding to μ1 . Also, taking {G, g} = {Φ ε , φ ε } in (28), we prove
the convergence, as ε → 0,
a 1
A |∇x Φ ε |2 dx + |∇x φ ε |2 dx + |Φ ε |2 dx + |φ ε |2 dx → μ1 + 1. (31)
Ω εt ωε Ω ε ωε
for all {G, g} ∈ H 1 (Ωε ), and using (28), (29), (9), (11), (13) and the convergence of
Φ ε toward V1 weakly in H 1 (Ω ) (cf. (31)), we have
lim Rε {Ψ ε , ψ ε } = lim Rε {Φ ε , φ ε } = μ1 + 1.
ε →0 ε →0
Stiff Problems in Domains with a Strongly Oscillating Boundary 169
But the minimax principle allows us to write λ1ε +1 ≤ Rε {Ψ ε , ψ ε } and, taking limits
as ε → 0, λ1∗ + 1 ≤ μ1 + 1. Thus, the result λk∗ = μk holds for k = 1.
Let us assume that λi∗ = μi holds for 0 ≤ i ≤ k. Since the eigenfunctions {Vi∗ }ki=1
corresponding to {λi∗ }ki=1 are orthonormal with the scalar product (24), it is clear
that μk+1 ≤ λk+1∗ . In order to prove λ ∗ ≤ μ ε ε
k+1 , we consider {Φ , φ } ∈ H (Ω ε )
1
k+1
the solution of (28) with μ1 ,V1 , ṽ1 replaced by μk+1 ,Vk+1 , ṽk+1 , respectively, where
Vk+1 is an eigenfunction of (22), with norm 1, corresponding to μk+1 and such that
it is orthogonal to {Vi∗ }ki=0 with the scalar product (24), and ṽk+1 is defined by (10).
We use the same argument as in (28) to prove that Φ ε converges toward Vk+1 weakly
in H 1 (Ω ) as ε → 0, and Rε {Φ ε , φ ε } converges to μk+1 + 1. Now, we define
k
{Ψ ε , ψ ε } = {Φ ε , φ ε } − ∑ ({Φ ε , φ ε }, {Uiε , uεi })ε {Uiε , uεi },
i=0
which is orthogonal to {{Uiε , uεi }}ki=0 with the scalar product (7) and satisfies
ε →0
({Ψ ε , ψ ε } − {Φ ε , φ ε }, {Ψ ε , ψ ε } − {Φ ε , φ ε })ε −−−−−−→ 0 and
1 ε →0
∇x (Ψ ε − Φ ε )2L2 (Ω ) + ∇x (ψ ε − φ ε )2L2 (ωε ) −−−−−−→ 0.
εt
Then
lim Rε {Ψ ε , ψ ε } = lim Rε {Φ ε , φ ε } = μk+1 + 1.
ε →0 ε →0
ε + 1 ≤ Rε {Ψ ε , ψ ε }, taking limits as ε
Since the minimax principle gives λk+1 → 0,
∗ ∗ =μ
we obtain λk+1 + 1 ≤ μk+1 + 1 and so λk+1 ∗
k+1 . Therefore, the result λk = μk
holds for k = 0, 1, 2 . . .
We have proved the result for the eigenvalues for a certain subsequence ε , but
taking into account that for any sequence it is possible to extract a subsequence
satisfying the same result, the statement holds for the sequence ε . In addition, the
fact that the set {Vk∗ }∞ k=0 forms a basis in H (Ω ) is obtained by contradiction. In-
1
∞
deed, let us assume that {Vk }k=0 is not a basis in H 1 (Ω ). Let V ∗ an eigenfunction
∗
In this section we study the asymptotic behavior of the eigenvalues of (3) and their
corresponding eigenfunctions when 0 ≤ t < 1 and the band ωε is a thin domain with
strongly oscillating boundary. The aim is to show the convergence, as ε → 0, of the
eigenpairs of (5) when ωε is defined by (2) toward those of the resulting problem
(34) (cf. Theorem 2).
170 D. Gómez et al.
As is well known, the following convergence holds (cf. Chapter V of [Sa80] for
instance):
hε → h̃ weakly in L2 (Γ ) as ε → 0. (32)
We previously proved that the bound for the eigenvalues of (3) given by Lemma 1
also holds in this case.
Lemma 2 Let λkε be the eigenvalues of (3) with 0 ≤ t < 1 and ωε defined by (2).
For each fixed k ∈ N, and ε sufficiently small, we have
Proof. We observe that the proof of the right-hand side of (16) in Lemma 1 is inde-
pendent of the domain ωε . To prove the left-hand side of (33) we use the technique
in Lemma 1 with minor modifications. Considering problem (15) with h the constant
h̃, from (14) we have
that is to say, C0U ≡ CU when h(τ ) = h̃. Thus, if {U1ε , uε1 } is an eigenfunction of (5)
corresponding to λ1ε satisfying (6), we have
where
ε
2
Ω U1 dx + h̃ Γ U1ε d τ 1
r1ε = and r2ε = |uε1 |2 dx − h̃ |U1ε |2 d τ .
|Ω | + h̃|Γ | ε ωε Γ
On account of the normalization of {U1ε , uε1 }, the variational formulation (5) and
the boundedness of λ1ε , we can extract a subsequence (still denoted by ε ) such that
U1ε converges toward some function V1∗ weakly in H 1 (Ω ) as ε → 0. Then, from the
orthogonality condition of {U1ε , uε1 } to the constants given by (6), (3)3 , (9), (13),
(32) and the strong convergence of U1ε to V1∗ in L2 (Γ ) yields
Stiff Problems in Domains with a Strongly Oscillating Boundary 171
ε h (τ )
1 1 ε
ε
r1 = uε1 (ν , τ )(K(ν , τ ) − 1) d ν d τ
|Ω | + h̃|Γ | ε Γ 0
ε hε (τ ) 2
1 ε →0
+ (uε1 (ν , τ ) − uε1 (0, τ )) d ν d τ + (hε − h̃)U1ε d τ −−−−−−→ 0.
ε Γ 0 Γ
1 = U1ε − C0U1ε 2L2 (Ω ) + h̃U1ε − C0U1ε 2L2 (Γ ) + oε (1) ≤ C∇xU1ε 2L2 (Ω ) + oε (1),
where λk∗ is some constant and Vk∗ is some function of H 1 (Ω ). In order to identify
(λk∗ ,Vk∗ ), we take limits in the variational formulation (5) with the test functions
{G, g} = {W, w̃} where W ∈ C∞ (Ω ) and w̃ is defined by (10). Arguments similar
to those in the proof of Theorem 1 (cf. (25)–(27)) as well as the convergence (32)
lead us to the limit problem: Find μ and V ∈ H 1 (Ω ), V = 0, satisfying the integral
identity
A ∇V · ∇W dx = μ VW dx + h̃ VW d τ ∀W ∈ H 1 (Ω ). (34)
Ω Ω Γ
(cf. (21) to compare). We note that the limiting problem depends only on the average
of h or equivalently on the area of the unit cell.
Problem (34) has a nonnegative discrete spectrum that we denote by {μk }∞ k=0 .
Let {Vk }∞k=0 be the corresponding eigenfunctions which are subject to the orthonor-
malization condition
Vk Vl dx + h̃ Vk Vl d τ = δk,l (35)
Ω Γ
Theorem 2. Let λkε be the eigenvalues of (3) with 0 ≤ t < 1 and ωε defined by (2).
For each fixed k ∈ N ∪ {0}, the sequence λkε converges toward the eigenvalue μk
of (34) as ε → 0. In addition, for each sequence {Ukε , uεk } of eigenfunctions of (5),
{Ukε , uεk } satisfying the normalization condition (6), we can extract a subsequence
(still denoted by ε ) such that Ukε → Vk∗ weakly in H 1 (Ω ), as ε → 0, where Vk∗ is an
eigenfunction of (34) corresponding to μk and the set {Vk∗ }∞ k=0 forms an orthonor-
mal basis in H 1 (Ω ) for the scalar product defined by the left-hand side of (35).
Proof. We rewrite the proof of Theorem 1 with minor modifications (let us to refer
those outlined in the proof of Lemma 2).
Acknowledgements This work has been partially supported by the Spanish grant MTM2009-
12628. The work of S.A. Nazarov has also been partially supported by the RFBR: 09-01-00759.
References
[At84] Attouch, H.: Variational Convergence for Functions and Operators, Pitmann (1984).
[BuKo87] Buttazzo, G., Kohn, R.V.: Reinforcement by a thin layer oscillating thickness. Appl.
Math. Optim., 16, 247–261 (1987).
[Go06a] Gómez, D., Lobo, M., Nazarov, S.A., Pérez, E.: Spectral stiff ptoblems in domains
surrounded by thin bands: asymptotic and uniform estimates for eigenvalues. J. Math.
Pures Appl., 85, 598–632 (2006).
[Go06b] Gómez, D., Lobo, M., Nazarov, S.A., Pérez, E.: Asymptotics for the spectrum of the
Wentzell problem with a small parameter and other related stiff problems. J. Math.
Pures Appl., 86, 369–402 (2006).
[GoLo99] Gómez, D., Lobo, M., Pérez, E.: On the eigenfunctions associated with the high fre-
quencies in systems with a concentrated mass. J. Math. Pures Appl., 78, 841–865
(1999).
[GoNa11] Gómez, D., Nazarov, S.A., Pérez, E.: Spectral stiff problems in domains surrounded by
thin stiff and heavy bands: local effects for eigenfunctions. Netw. Heterog. Media, 6,
1–35 (2011).
[LoSa79] Lobo, M., Sanchez-Palencia, E.: Sur certaines propriétés spectrales des perturbations
du domaine dans les problèmes aux limites. Comm. Partial Differential Equations, 4,
n. 10, 1085–1098 (1979).
[Na90] Nazarov, S.A.: Binomial asymptotic behavior of solutions of spectral problems with
singular perturbations. Mat. Sb., 181, n. 3, 291–320 (1990) (English transl: Math.
USSR-Sb., 69, n. 2, 307–340 (1991)).
[Na07] Nazarov, S.A.: Eigenoscillations of an elastic body with a rough surface. Prikl. Mekh.
Tekhn. Fiz., 48, n. 6, 103–114 (2007) (English transl: J. Appl. Mech. Tech. Phys., 48,
n. 6, 861–870 (2007)).
[Na08] Nazarov, S.A.: Asymptotic behavior of solutions and the modeling of problems in the
theory of elasticity in a domain with a rapidly oscillating boundary. Izv. Ross. Akad.
Nauk Ser. Mat., 72, n. 3, 103–158 (2008) (English transl: Izv. Math., 72, n. 3, 509–564
(2008)).
[OlSh92] Oleinik, O.A., Shamaev, A.S., Yosifian, G.A.: Mathematical Problems in Elasticity and
Homogenization, North-Holland (1992).
[Sa80] Sanchez-Palencia, E.: Nonhomogeneous Media and Vibration Theory, Springer-Verlag
(1980).
Spectra and Pseudospectra
of a Convection–Diffusion Operator
1 Introduction
Au := −Δ u − (y, x) · ∇u + (x2 + y2 )u
H. Guebbaï
Université Jean Monnet, Université de Lyon, Saint-Étienne, France,
e-mail: guebaihamza@yahoo.fr
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 173
DOI 10.1007/978-0-8176-8238-5_16, © Springer Science+Business Media, LLC 2011
174 H. Guebbaï and A. Largillier
and verifies
1 4 k
ϕ ( f , g) = ∑ i Q( f + ik g).
4 k=1
By the Cauchy–Schwarz inequality,
16 7
(−yu, −xu) · ∇u ≤ ∇uL2 (Ω ) + (x2 + y2 )|u|2 .
2
Ω 2 Ω
Our goal is to determine the spectrum of A. We remark that D(A) includes a Dirichlet
boundary condition. Consider first the eigenvalue problem: Find λ ∈ C and u ∈
D(A) \ {0} such that
3 The Spectrum of Aη
We observe that there exists η0 for which M vanishes and changes of sign (see
Fig. 1).
176 H. Guebbaï and A. Largillier
Fig. 1 M as a function of η
1 1
ReAη u, u = [Aη u, u + Aη u, u] = [Aη u, u + u, Aη u].
2 2
But
∂u
−Δ u(x, y)u(x, y)dxdy = − (x, y) u(x, y) d σ
Ωη ∂ Ωη ∂ν
=0 over ∂ Ω η
+ ∇u(x, y).∇u(x, y)dxdy = |∇u(x, y)|2 dxdy,
Ωη Ωη
0= y∂x (uu)(x, y) dxdy = y ∂x u(x, y)u(x, y) + u(x, y)∂x u(x, y) dxdy
Ωη Ωη
and
0= x∂y (uu)(x, y) dxdy = x ∂y u(x, y)u(x, y) + u(x, y)∂y u(x, y) dxdy.
Ωη Ωη
Thus
Spectra and Pseudospectra of a Convection–Diffusion Operator 177
Aη u, u = |∇u|2 + yu∂x u + xu∂y u + (x2 + y2 )|u|2 .
Ωη
Similarly,
u, Aη u = |∇u|2 − yu∂x u − xu∂y u + (x2 + y2 )|u|2 ,
Ωη
where
−2
ReAη u, u = ∇u2L2 (Ωη ) + (x2 + y2 )|u(x, y)|2 dxdy ≥ (CPF + η 2 )u2L2 (Ωη ) .
Ωη
−2 −2
Hence Aη is injective for all λ < CPF + η 2 and thus sp p (Aη ) is included in [CPF +
η , +∞). Let H = H0 (Ωη ), g ∈ L (Ωη ) and λ ∈ (−∞, 5η /4). The sesquilinear
2 1 2 2
form defined on H by
ϕλ (u, v) := ∇u.∇v + 5(x2 + y2 )/4 − λ uv dxdy
Ωη
verifies
|ϕλ (u, v)| ≤ ∇uL2 (Ωη ) ∇vL2 (Ωη ) +CuL2 (Ωη ) vL2 (Ωη ) ,
where 6 7
C := sup 5(x2 + y2 )/4 : (x, y) ∈ Ωη + |λ |,
and 6 7
|ϕ (u, u)| ≥ min 1, (5η 2 /4 − λ ) u2H .
is continuous, it follows from the Lax–Milgram theorem that the equation ϕλ (u, v) =
L(v) for all v ∈ H, has a unique solution u in H. Let g ∈ L2 (Ωη ). Consider the prob-
lem
(P) Find u ∈ L2 (Ωη ) : Aη u − λ u = g on Ωη , u = 0 on ∂ Ωη .
xy xy xy
Multiply the equation by e 2 , set g0(x, y) := g(x, y)e 2 , u0(x, y) := u(x, y)e 2 and let
g0 ∈ L2 (Ωη ). Then (P) is equivalent to
4 Pseudospectra
Extending f by 0 on Ω
A f − λ f L2 (Ω )
≤ ε.
f L2 (Ω )
It follows that λ belongs in sp ε (A) and
;
sp ε (Aη ) ⊂ sp ε (A).
0<η <1
Conversely, let ε ∈ [0, ε ) and λ ∈ sp ε (A). Then there is f ∈ D(A) such that
A f − λ f L2 (Ω )
≤ ε < ε.
f L2 (Ω )
Since Cc∞ (Ω ) is dense in D(A) respect to the graph norm. For all f ∈ D(A), there is
a sequence ( fn )n∈N in Cc∞ (Ω ) such that
A fn − λ fn L2 (Ω ) A f − λ f L2 (Ω )
lim = .
n→+∞ fn L2 (Ω ) f L2 (Ω )
Spectra and Pseudospectra of a Convection–Diffusion Operator 179
)
Since sp ε (A) = sp ε (A), it follows that
0<ε <ε
;
sp ε (A) = sp ε (Aη ).
0<η <1
5 Spectra
In this section we will establish a new result which relates the spectra of A to that of
Aη . Let S ⊂ C. Recall that the ε -neighborhood of S is
Nε (S) := {s + z : s ∈ S, |z| ≤ ε } .
Theorem 4. For all ε > 0
; ;
Nε sp (Aη ) = sp ε (Aη ).
0<η <1 0<η <1
)
Proof. Let λ ∈ sp ε (Aη ). There is η1 ∈ (0, 1) such that
0<η <1
Conversely, let ;
λ ∈ Nε sp (Aη ) .
0<η <1
)
Then λ = s + z, where s ∈ sp (Aη ) and |z| ≤ ε . There is a sequence
0<η <1
;
(sn )n∈N ⊂ sp (Aη )
0<η <1
180 H. Guebbaï and A. Largillier
which converges to s. So
and then ;
λ∈ spε (Aη ).
0<η <1
Theorem 5. We have ;
sp (A) = sp (Aη ).
0<η <1
)
Proof. Let λ ∈ sp (Aη ). There is η1 ≤ η0 such that λ ∈ sp (Aη1 ). Then there
0<η <1
is f ∈ D(Aη1 ), where Aη1 f − λ f =0. Extending f by 0 on Ω , we obtain λ ∈ sp (A).
) c )
Conversely, let λ ∈ [ sp (Aη ) where d(λ , sp (Aη )) > ε . Then
0<η <1 0<η <1
;
λ ∈ Nε sp (Aη ) .
0<η <1
By Theorem 4,
;
λ∈
/ sp ε (Aη ),
0<η <1
References
[Da00] Davies, E.B.: Pseudospectra of differential operators. J. Operator Theory, 43, 243–262
(2000).
[Da95] Davies, E.B.: Spectral Theory and Differential Operators, Cambridge University Press
(1995).
[Ka80] Kato, T.: Perturbation Theory of Linear Operators, Second edition, Springer-Verlag,
Berlin, Heidelberg and New York (1980).
[Bo99] Boulton, L.: Non-self-adjoint harmonic oscillator, compact semigroups and pseu-
dospectra. Preprint, King’s College London (1999).
[ReTr94] Reddy, S.C., Trefethen, L.N.: Pseudospectra of the convection–diffusion operator.
SIAM J. Appl. Math., 54, 1634–1649 (1994).
[RoSi96] Roch, S., Silbermann, B.: C*-algebra techniques in numerical analysis. J. Operator
Theory, 35, 241–280 (1996).
[Tr92] Trefethen, L.N.: Pseudospectra of Matrices, Longman Sci. Tech. Publ., Harlow, UK
(1992).
[Tr97] Trefethen, L.N.: Pseudospectra of linear operators. SIAM Review, 39, 383–406 (1997).
A Necessary and Sufficient Condition
for the Existence of Absolute Minimizers
for Energy Functionals with Scale Invariance
S.M. Haidar
1 Introduction
r(R)
u(x) = x
R
in W 1,1 (Ω , R 3 ), where R := |x|, we constructed strongly elliptic strain-energy den-
sity functions of the form
S.M. Haidar
Grand Valley State University, Allendale, MI, USA,
e-mail: haidars@gvsu.edu
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 181
DOI 10.1007/978-0-8176-8238-5_17, © Springer Science+Business Media, LLC 2011
182 S.M. Haidar
where F := u(x), γ ∈ (0, 1), and a and b are positive real numbers. In the absence
of external forces the associated total energy functional was given by
1
J(u(x)) = W (x, u(x))dx = f (R, r, r )dR := J(r(R)), (1)
Ω 0
We then showed, in [Ha09], that for certain choices of a and b, the equilibrium
equations associated with the total energy functional admitted nontrivial weak so-
lutions of the form r(R) = λ Rγ with J(r = λ Rγ ) = 0. That is, these solutions also
served as the absolute minimizers for the corresponding energy functional. See Re-
mark 2.
The question now is how regular can such solutions be? Can they be in W 1,s (0, 1)
for s ≥ 1/(1 − γ ) or even in a smaller space? Note that if we were to restrict s to be
greater than 1/(1 − γ ), we would loose the actual minimizer r(R) = λ Rγ . In effect,
this is then a question that involves two essential parts: On the one hand, it is a ques-
tion about the possible existence of a different (more regular) equilibrium solution
and, on the other, it must be that this new solution also serves as the new absolute
minimizer for the corresponding total energy functional. Let us keep in mind that the
Euler–Lagrange equations, while being generally necessary, need not be sufficient
for the existence of minimizers for the corresponding total energy functional, see
[BaMi85]. Here, as we generalize to arbitrary scale-invariant energies, we are led in
a natural way to consider formulating a new optimality condition for the equilibrium
solutions to be absolute minimizers of the corresponding total energy functional.
One can argue that it may be possible to still show that equilibrium solutions are
optimal, and therefore avoid the need for a new sufficiency criterion for a global
minimizer, by using the well-known Hamilton–Jacobi theory. However, this is not
always guaranteed to work as it involves using similarity transformations and solv-
ing a very complicated nonlinear partial differential equation for the underlying
value function associated with the minimization problem in question. For a full
treatment on the subject see [Ru66, He66].
Partly motivated by [HeMi88, BoFoLeMa02], we prefer to pursue another line of
reasoning, via the Noether identity and field theory, along which we achieve at least
the same goal. This approach manifests itself in the fact that we will be dealing di-
rectly with a nonlinear differential equation (Noether identity) the solution to which
serves as the absolute minimizer for the total energy functional, without having to
obtain the underlying value function as required by the alternative Hamilton–Jacobi
theory. It also has an advantage in the following sense: the Noether identity, which
is, in essence, a first integral of the Euler–Lagrange equations, is inherently a con-
servation law, which would correspond naturally with scale-invariant constitutive
properties such as (2) above. Therefore, the main outcome in this paper is our pre-
Absolute Minimizers for Energy Functionals 183
sentation of a new result that the Noether identity would provide for a necessary
and sufficient condition for the existence of global minimizers for functionals with
scale- invariant energies. With this criterion, one will be able to obtain the new mini-
mizer over the smaller space as well as gain more information regarding the optimal
regularity of weak equilibria in general. The following preliminary, artificial exam-
ple illustrates the subtleties involved in the formulation of our main result, which
we present later on in Theorem 3.
over the set A := {u ∈ W 1,1 (0, 1) : u(0) = 0, u(1) = 1}. Within the context of field
theory of the calculus of variations,the value function associated with this optimiza-
tion problem is S(x0 = 1, y0 = 1) := inf{I(y) : y ∈ A }, which is given by the fol-
lowing Hamilton–Jacobi problem:
1
Sx (s, y) + f ∗ (x, y, Sy (x, y)) = 0
(4)
S(0, 0) = 0,
where
f ∗ (x, y, z) := sup[pz − f (x, y, p)]
p
denotes the Fenchel dual (known in the older literature as Legendre transform).
Moreover, f ∗ (x, y, z) is obtained as the real-valued continuous function (x, y) →
p̂(x, y) that specifies the slope at (x, y) of the extremal passing through the point
(x, y). Hence, any admissible trajectory given by y = p̂(x, y) is a minimizer for the
functional I. We note further that, since the slope field p̂(x, y) exists at every point in
the domain [0, 1] × [0, 1] =: M, the collection of resulting extremals covers the entire
domain to provide what is known as Mayer field. For a comprehensive treatment of
these topics, see [GeFo63, Ru66, He66].
In our example, f ∗ is obtained as the p̂(x, y) given by
1 1
p̂(x, y) = 1 (y2 − x)−2/5 [Sy (x, y)] 5 . (5)
6 5
where α = 5/(66/5 ). We would like to point out that the Lagrangian of the func-
tional I satisfies property (2) with γ = 0.5. Let us seek similarity solutions of (6)
1
of the form S(x, y) = (h ◦ ξ )(x, y), where ξ (x, y) := yx− 2 and h is a scalar-valued
function of ξ . Substituting for S into (6) yields
1
− (ξ + α (ξ 2 − −1)−2/5 (h)1/5 ) = 0. (7)
2
Later on in Theorem 2 we show that κ = 0, but for now let us accept this fact and
rewrite (11) as
(y )5 (y2 − x)2 [−5y x + 3y] = 0. (12)
For y2 = x and y = 0, (12) implies that −5y x + 3y = 0, or
3
y = yx−1 ,
5
which is the same relation as (8) for p̂(x, y).
2 Optimality Criterion
To begin with let us consider the following: at each point (x0 , y0 ) ∈ [0, 1] × [0, 1] =:
M, put A0 := {u ∈ W 1,1 (0, 1) : u(0) = 0, u(x0 ) = y0 } and
x0
I(y) := f (x, y, y )dx. (13)
0
The key point to address in this section is the fact that the admissible solution to the
Noether identity is indeed the minimizer for the associated energy functional (13).
For convenience, putting (R, r, r ) = (x, y, y ) and taking ε = 1/x in property (2)
yields
f (x, y, y ) = x−1 f (1, X, P) := x−1 e(P, x), (14)
where P(x, y) := y x1−γ and X(x, y) := yxγ . In terms of P, X, and e(., .), the Noether
identity (10) takes the form
We also make the following two standing assumptions on the integrand f (x, y, .) the
roles of which become relevant in the proofs of the theorems and the remarks below:
(H0) f (x, y, .) has at least superlinear growth at ∞.
(H1) f,33 (x, y, ·) ≥ 0 with f,33 (x, y, z) > 0 provided yx−γ = 1, z = 0.
Proof. The main issue here is to prove that κ = 0 since, by Theorem 1, identity (15)
is a first integral of the associated Euler–Lagrange equation which is necessarily
satisfied by the minimizer ŷ.
In effect, from field theory as illustrated in the preliminary example above, it is
known that ŷ is the admissible solution of the equation
y = p̂(x, y),
where (x, y) → p̂(x, y) is the slope field function. The corresponding Hamilton–
Jacobi equation is
S,x + f ∗ (x, y, S,y ) = 0, (16)
where
f ∗ (x, y, z) = sup[tz − f (x, y,t)].
t
Due to the superlinear growth of the function f (x, y, ·) at ∞, by (H0), this sup is
attained at p̂ given implicitly by
V = e,P . (17)
Since e,P (·, X) is strictly monotone (increasing), it is clear that such a p̂ exists and
is unique (by the implicit function theorem). Equation (16) may now be written as
where P̂ = p̂x1−γ . On the other hand, ŷ satisfies the Noether identity (15), namely,
Hence, 0 = κ .
Theorem 3. (Sufficiency) Let y(·) ∈ A0 , be a solution to the Noether identity (15)
with κ = 0. Let f (x, y, y ) satisfy property (14), and hypotheses (H0) and (H1). Then
y(·) is the global minimizer for the functional I(y).
Proof. First we use (14) to derive an equivalent form for identity (10). Then we
construct the minimizer in question to complete the proof.
Step 1: An equivalent form for the Noether identity.
In terms of P and X, identity (10) takes the form
Z → m(X, Z).
It now follows that for each x for which X(x) = 1, the Noether identity (20) is
equivalent to
P(x) = m(X(x), γ X(x)). (21)
−1
dx X(x) = x (P(x) − γ X(x)),
d
Since identity (21) takes the form
d
X(x) = x−1 [m(X(x), γ X(x)) − γ X(x)] (22)
dx
for all x ∈ (0, x0 ) for which X(x) = 1. If X̂(x) is a solution of (22) then so is X̂(λ x)
for any λ > 0. Then, for any λ > 0, consider
y(·, λ ) satisfies Noether identity (15) or equivalently (10). Such a family of trajecto-
ries provides a Mayer field which simply covers the region M (see the discussion in
Sect. 1.1). Thus, to each point (x0 , y0 ) ∈ M, there is an associated slope value given
by
p̂(x0 , y0 ) = y (x0 , λ0 ),
where λ0 = λ (x0 , y0 ) is the unique value of λ for which the curve x → y(x, λ ) passes
through (x0 , y0 ) [observe that X̂(·) is strictly monotone increasing].
Step 2: We are now in position to show that y(x, λ ) yields the global minimizer
in question.
Observe that the slope field function p̂(·, ·) is smooth. Also the Hilbert integral
(x0 ,y0 )
H := [ f (x, y, p̂) − p̂ f,3 (x, y, p̂)]dx + f,3 (x, y, p̂)dy (24)
(0,0)
is equal to I(y; x0 , y0 ) because of the fact that along an arc of an extremal of the
field, y = p̂(x, y). With M simply connected, it now follows that the Hilbert inte-
grand L(x, y, y ) := f (x, y, p̂(x, y)) + (y − p̂(x, y)) f,3 (x, y, p̂(x, y)) of (24) is an exact
differential on M. In fact, the function
V :M→R
defined by
x0
V (x0 , y0 ) = L(x, y(x, λ0 ), y (x, λ0 ))dx
0
x0
= f (x, y(x, λ0 ), y (x, λ0 ))dx
0
x0
= x−1 e(X̂(λ0 x), m(X̂(λ0 x), γ X̂(λ0 x)))dx (25)
0
188 S.M. Haidar
where
w(x, y, p, q) = f (x, y, q) − f (x, y, p) − (q − p) f p (x, y, p)
is the Weierstrass excess function.
By (A), the second term on the right side of (26) constitutes a nonnegative
bounded sequence of real numbers, and it tends to zero as n → ∞. By (25),
αn
V (αn , y(αn )) = x−1 e(X̂(λn x), m(X̂(λn x), γ X̂(λn x)))dx
0
λn αn
= t −1 e(X̂(t), m(X̂(t), γ X̂(t)))dt,
0
where λn = λ (αn , y(αn )) is the unique value for λ for which the curve x → y(x, λ )
passes through (αn , y(αn )), (per Step 1). Thus y(αn ; λn ) = y(αn ).
Hypothesis (H1) or (3) implies that w(x, y, p̂, y ) ≥ 0 [in fact, w(x, y, p̂, y ) > 0 for all
unequal values of p̂ and y as long as (x, y, p̂) and (x, y, y ) lie in the domain of f with
X(x) = 1]. Hence,
I(y(·)) − I(y(·; λ0 )) ≥ 0.
It now follows that the minimization problem possesses y(·) as its solution as
claimed, and this completes the proof.
3 Concluding Remarks
Remark 2. The homogeneity property (2) and (H1) were used in [BaMi85] and
[HeMi88] to study the regularity of minimizers for one-dimensional variational
problems in the calculus of variations. See [Ha00] for a physical interpretation of
this scale-invariance property.
Remark 3. Let us recall from Sect. 1 that the energy functional (1) admitted r(R) =
λ Rγ as its absolute minimizer with J(r = λ Rγ ) = 0. Note that these functions corre-
spond to a point along the line P = γ X in the PX−plane. For s ≥ 1−1 γ , such functions
are no longer admissible in A0 ∩W 1,s (0, 1). In our hypotheses (H1) or (3), we ruled
out the existence of such solutions and, instead, established, in Sect. 1.1, a result for
existence of more regular equilibria. For instance, could there be admissible solu-
tions to the Noether identity (15) with κ = 0 of the form r = λ Rβ γ , where β γ < 1
for β > 1. Clearly r(R) = λ Rγ presents a stronger singularity at the center of the
unit ball in the deformation gradient than do r(R) = λ Rβ γ . Following further the
regularity question, it is natural to inquire about the physical significance of such
minimizers. Could such weak singularities at the center of the ball be connected to
elastic disarrangements responsible for the onset of fracture somewhere inside the
material body? This is physically plausible keeping in mind that the structure of
energy, which allows for discontinuities in the deformation gradients, has been con-
nected to material defects in crystals just like the Gibbs phenomenon in fluids. For
details, see [Er83, Mo87, MuPa08]. Before, we go on, we must also mention that our
mathematical approach here to model fracture in nonlinear mechanics is not the only
one available. For example, relaxation methods using special functions of bounded
variations have also been studied for this purpose, see [AF00, MuPa08]. An inter-
esting feature about our approach, however, is the fact that most existence results
in the literature do not show the actual form of the minimizers in any specific case,
for example when the domain Ω is a ball, as we did in [Ha09]. The questions raised
above remained wide open, and it would be particularly interesting to see at least
examples in nonlinear elasticity where such minimizers do exist in A0 ∩W 1,s (0, 1)
for s ≥ 1/(1 − γ ) and are associated with material instabilities.
Acknowledgements This work was partially supported by a grant from The Center for Scholarly
and Creative Excellence at Grand Valley State University.
190 S.M. Haidar
References
[AF00] Ambriosio, L., Fusco, N., Pallara, D.: Functions of Bounded Variations and Free
Discontinuity Problems, Clarendron Press, Oxford University Press, New York
(2000).
[Ba82] Ball, J.M.: Discontinuous equilibrium solutions and cavitation in nonlinear elas-
ticity. Phil. Trans. Roy. Soc. Lond. A, 306, 557–611 (1982).
[BoFoLeMa02] Bouchitte, G., Fonseca, I., Leoni, G., Mascarenhas, L.: A global method for re-
laxation in W (1,p) and in SBV. Arch. Ration. Mech. Anal., 165, no. 3, 187–242
(2002).
[BaMi85] Ball, J.M., Mizel, L.V.: One-dimensional variational problems whose solutions
do not satisfy the Euler–Lagrange equations. Arch. Rat. Mech. Anal., 90, 325–
388 (1985).
[Er83] Ericksen, J.L.: Ill-posed problems in thermoelasticity theory, in: Systems of
Nonlinear Partial Differential Equations (Editor: J.M. Ball), Reidel, Dordrecht
(1983).
[GeFo63] Gelfand, I.M., Fomin, S.V.: Calculus of Variations, Prentice-Hall (1963).
[Ha00] Haidar, S.M.: Existence and regularity of weak solutions to the displacement
boundary value problem of nonlinear elastostatics, in: Integral Methods in Sci-
ence and Engineering, CRC Press, 161–166 (2000).
[Ha07] Haidar, S.M.: Convexity conditions in uniqueness and regularity of equilibria in
nonlinear elasticity, in: Integral Methods in Science and Engineering: Techniques
and Applications (Editors: C. Constanda, S. Potapenko), Springer–Birkhäuser,
Boston, MA, 109–119 (2007).
[Ha09] Haidar, S.M.: On J. Ball’s fundamental existence theory and regularity of weak
equilibria in nonlinear radial hyperelasticity, in: Integral Methods in Science
and Engineering: Techniques and Applications, vol. 1 (Editors: C. Constanda,
M. Pérez), Springer–Birkhäuser, Boston, MA, 161–171 (2009).
[He66] Hestenes, M.R.: Calculus of Variations and Optimal Control Theory, Wiley
(1966).
[HeMi88] Heinricher, A., Mizel, V.: The Lavrentiev phenomenom for invariant variational
problems. Arch. Rat. Mech. Anal., 102, no. 1, 57–93 (1988).
[Lo77] Logan, J.D.: Invariant Variational Principles, Academic Press (1977).
[MuPa08] Muller, S., Palombaro, M.: Existence of minimizers for a polyconvex energy in a
crystal with dislocations. Cal. Var., 31, 473–482 (2008).
[Mo87] Modica, L.: Gradient theory of phase transitions with boundary contact energy.
Ann. Inst. H. Poincaré, Anal. Non Linéaire, 4, 487–512 (1987).
[Ol86] Olver, P.J.: Applications of Lie-Groups to Differential Equations, Springer-
Verlag, New York (1986).
[Ru66] Rund, H.: The Hamilton–Jacobi Theory in the Calculus of Variations, D. Van
Nostrand, London (1966).
Nonlinear Abel-Type Integral Equation
in Modeling Creep Crack Propagation
1 Introduction
A nonlinear Abel-type equation is obtained in this paper to model creep crack time-
dependent propagation in the infinite viscoelastic plane. A finite time when the in-
tegral equation solution becomes unbounded is obtained analytically as well as the
equation parameters when solution blows up for all times. A modification to the
Nyström method is introduced to numerically solve the equation and some compu-
tational results are presented.
The strength of materials under creep conditions depends not only on the instant
value of the load but also on the load duration and generally on the temporal load
history. This effect is essential for concrete and some plastics under room temper-
ature, and for structural metals under elevated temperatures. The temporal strength
condition under constant uniaxial stress, starting at time t = 0, takes the form
|σ |
|σ | < σ ∗ (t) i.e., < 1, (1)
σ ∗ (t)
where the dependence of the temporal strength on time has a form shown in Fig. 1.
A popular approximation of the temporal strength σ ∗ (t) is the power-type func-
tion, given by σ ∗ (t) = σ 0t −1/b that we will call (similar to fatigue) the Basquin
diagram, where σ 0 and b are constant material parameters, see e.g. [Ra69, PeMa71].
When the load is not constant but varies with time, different generalizations of
relation (1), often called accumulation rules, are possible. A novel accumulation rule
for life time under variable loading, given in [MiNa09, MiNa11] has the following
form,
L. Hakim
Brunel University West London, UK,
e-mail: layal.hakim@brunel.ac.uk
S.E. Mikhailov
Brunel University West London, UK,
e-mail: sergey.mikhailov@brunel.ac.uk
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 191
DOI 10.1007/978-0-8176-8238-5_18, © Springer Science+Business Media, LLC 2011
192 L. Hakim and S.E. Mikhailov
Fig. 1 Durability diagram: material strength σ ∗ (t) under constant uniaxial stress as function of
life time
t β !1/β
Λ (σ ;t) :=
d σ (τ ) d τ < 1, (2)
dt 0
σ ∗ (t − τ )
where σ ∗ (t) is the durability under constant loading and β = 0 is another material
constant.
Particularly for the Basquin diagram σ ∗ (t) = σ 0t −1/b , temporal strength condi-
tion (2) takes the form
t β !1/β
Λ (σ ;t) =
d σ (τ ) (t − τ )β /b d τ
dt 0
σ0
t β !1/β
β β
= σ (τ ) (t − τ ) b −1 d τ < 1. (3)
β
bσ0 0
Note that due to the problem symmetry, σ22 is a principal stress component at x2 = 0.
Assuming that the durability of the viscoelastic material is described by the dura-
bility condition given in (3) with σ = σ22 (τ , x1 ), we substitute there stress distribu-
tion (4). Replacing the inequality with the equality at the tip of the propagating
Nonlinear Abel-Type Integral Equation 193
crack, x1 = a(t), we arrive at the following nonlinear Volterra integral equation and
initial condition for a(τ ),
t β
(t − τ ) b −1
d τ = ca−β (t), t > 0, (5)
0 [a2 (t) − a2 (τ )]β /2
a(0) = a0 , (6)
where b, β > 0, c = βb ( σq0 )−β and the initial crack half-length a0 are known con-
stants.
Normalizing the variables t → t/t∞ and a → a/a0 , where t∞ = (σ0 /q)b is the
life time of the infinite plane without a crack under the same load q, simplifies the
problem to
t β
(t − τ ) b −1 b −β
dτ = a (t), (7)
0 [a2 (t) − a2 (τ )]
β
2 β
a(0) = 1. (8)
The function to be sought is a(t), which denotes the normalized length of the crack,
and t denotes the normalized time.
Equation (7) is a nonlinear Volterra integral equation and more specifically, since
the integrand has a weak singularity as τ approaches t, this is an integral equation of
the Abel type. The unknown function participates nonlinearly both in the integrand
and in the out-of-integral term, and the integrand depends not only on a(τ ) but also
on a(t).
For β = b, (7) was obtained and solved analytically in [MiNa03a, MiNa03b]. For
this case the equation is reduced to a simpler Abel-type integral equation,
194 L. Hakim and S.E. Mikhailov
t
1
ab (t) b
d τ = 1. (9)
0 [a2 (t) − a2 (τ )] 2
After the change of dependent and independent variables (cf. also [GoVe91]) the
equation becomes linear, and the solution can be written as
a 2−b 2 sin ( π (b−2) )
0 b b b a20
t(a) = 1 − 2
2 F1 1 − , 1 − , 2 − ; 2 (10)
a π (b − 2) 2 2 2 a
as τ → t. Thus for the integral in (7) to exist, we obtain the same condition 0 < b < 2
as for (9).
To find the normalized breaking time t as a(t) tends to infinity, let us multiply
both sides of (7) by aβ (t) to arrive at the equation
t β
(t − τ ) b −1 aβ (t) b
dτ = (11)
0 [a2 (t) − a2 (τ )] 2
β
β
aβ (t)
lim β
= 1.
a(t)→∞ [a2 (t) − a2 (τ )] 2
Assuming that the limit can be transferred under the integral, the limit of (11) re-
duces to t
β b
(t − τ ) b −1 d τ = .
0 β
Evaluating the integral yields
β
t b = 1,
which gives t = 1. Therefore, for any β > 0 and b ∈ (0, 2), the normalized breaking
time is t = 1 if the assumption on the limit transfer is correct.
Nonlinear Abel-Type Integral Equation 195
3 Numerical Algorithms
No exact or numerical solution of (7) seems to be available in the literature for cases
when β = b. This paper aims to solve numerically (7) in order to find the normalized
crack length, a, as the normalized time, t, increases, for different values of β .
There is a wide range of numerical methods for solving Fredholm and Volterra
equations. One of the methods is known as the Nyström method and is explained
in detail in, for example, [At97]. However, the standard Nyström method is not
well suited for integral equations with singularities. Therefore, to use the Nyström
method for finding approximate solutions to (7), we modified it as described in the
following.
Let us introduce a mesh of n node points, containing equidistant time steps, ti =
ih, where h is the step size, i = 0, 1, 2, . . . , n. Then, we apply collocations for (7) in
the node points starting from i = 1,
ti β
(ti − τ ) 2 −1 b −β
dτ = a (ti ), i = 1, 2, . . . , n. (12)
0 [a2 (t ) − a2 (τ )] 2
β
β
i
ti−1 β ti β
(ti − τ ) b −1 (ti − τ ) b −1 b −β
dτ + dτ = a (ti )
0 [a2 (ti ) − a2 (τ )]
β
2 ti−1 [a2 (ti ) − a2 (τ )]
β
2 β
i = 1, 2, . . . , n, (13)
a0 = 1, (15)
β
β
i−1
(ti − t j ) b −1 h 2 ti β β
∑ β
wi−1, j +
a2i − a2i−1
(ti − τ ) b −1− 2 d τ
j=0 [a2i − a2j ] 2 ti−1
b −β
= a , i = 1, 2, . . . , n, (16)
β i
196 L. Hakim and S.E. Mikhailov
where wi−1, j represent the quadrature weights. To make (16) valid for i = 1 we set
w0, j = 0.
Evaluating the integral and simplifying the result, we can rewrite (16) as
β β
i−1
(ti − t j ) b −1 hb b −β
∑ β
wi−1, j + β
= a , i = 1, . . . , n.
β i
(17)
j=0 [a2i − a2j ] 2 ( βb − β
2 )[ai − ai−1 ]
2 2 2
Since the summation term in (17) vanishes for i = 1, we can find from (15) and
(17) an explicit expression for a1 in terms of h, β and b,
=
−1 2 >− 21
β β β β β
a1 = 1 − hb − . (18)
b b 2
For a fixed i, we will solve nonlinear equation (19) for ai using the Newton
method. Thus, each ai will be approximated using the following iterative relation
F(ai,k , i)
ai,k+1 = ai,k − , i = 2, . . . , n, (20)
F (ai,k , i)
where F (ai , i) denotes the derivative of F(ai , i) with respect to ai . Here, i refers to
the collocation point in the mesh, while k is the iteration step in the Newton method.
Differentiating and simplifying the result yields the following expression for the
derivative of F(ai , i),
Nonlinear Abel-Type Integral Equation 197
β β
hβ ai (ti − t0 ) b −1 i−2
β ai (ti − t j ) b −1
F (ai , i) = − β
−h ∑ β
2[a2i − a20 ] 2 +1 j=1 [a2i − a2j ] 2 +1
β
h b β ai (1 + 2( βb − β2 )−1 ) −β −1
− β
+ bai , i = 2, . . . , n.
2[a2i − a2i−1 ] 2
To avoid the numerical problems related with the infinite growth of a(t) as t → 1
while computing the iterations in the Newton method solving (16), we introduced a
new variable,
1
Ai = 2 . (21)
ai
β /2
Thus, by substituting (21) in (16) and dividing it by Ai , we arrive at the equation
β β β
h(ti − t0 ) b −1 h 2 (1 + 2( βb − β2 )−1 ) i−2
h(ti − t j ) b −1
G(Ai , i) = + β
+∑ β
2[1 − AA0i ]β /2 2[1 − AAi−1
i
]2 j=1 [1 − AAij ] 2
b
− = 0, i = 2, 3, . . . , n.
β
Calculating the derivative of G with respect to Ai gives
β β
β h(ti − t0 ) b −1 β h b (1 + 2( βb − β2 )−1 )
G (Ai , i) = β
+ β
4A0 [1 − AA0i ] 2 +1 4Ai−1 [1 − AAi−1
i
] 2 +1
β
i−2
β h(ti − t j ) b −1
+∑ β
,
j=1 2A j [1 − AAij ] 2 +1
G(Ai,k , i)
Ai,k+1 = Ai,k − where i = 2, 3, . . . , n
G (Ai,k , i)
√
and then obtain ai = 1/ Ai .
This algorithm modification can be associated with solving numerically the equa-
tion in the form of (11) instead of (7).
198 L. Hakim and S.E. Mikhailov
4 Numerical Results
Fig. 3 Creep crack length vs. normalized time for different n, β = b = 1.5
One can expect that the convergence at other values of the parameter β , where
no analytical solution is available, will have the same character as for β = b and,
particularly, for the number of steps, n = 500, the error in A will not exceed 1%.
Figures 6, 7 present numerical results for the cases β = 2b, β = b/2, and β = 1
with n = 500 on the graphs of ln[a(t)/a0 ] verses t, and the graphs of A(t)/A0
verses t.
Graphs representing A(t) are particularly useful for analyzing the breaking time;
this is when the curve crosses the time axes, since the crack length a = 1/A2 tends to
infinity when A tends to zero. From the graph in Fig. 7, we can see that when β = 2b
the normalized breaking time is t = 1. For other values of β , it is not completely
clear from the graphs but we come to the same conclusions under the assumption
made in Sect. 2.
5 Conclusion
References
[At97] Atkinson, K.E.: The Numerical Solution of Integral Equations of the Second Kind,
Cambridge University Press, New York–Melbourne (1997).
[GoVe91] Gorenflo, R., Vessella, S.: Abel Integral Equations Analysis and Aplications, Springer
Verlag, Berlin–New York (1991).
[MiNa03a] Mikhailov, S.E., Namestnikova, I.V.: Local and non-local approaches to fatigue
crack initiation and propagation, in: IUTAM Symposium on Asymptotics, Singulari-
ties and Homogenisation in Problems of Mechanics (Editor: A.B. Movchan), Kluwer,
The Netherlands, 285–294 (2003).
[MiNa03b] Mikhailov, S.E., Namestnikova, I.V.: Local and non-local approaches to creep crack
initiation and propagation, in: Proceedings of the 9th International Conference on the
Mechanical Behaviour of Materials, Geneva, Switzerland (2003).
[MiNa09] Mikhailov, S.E., Namestnikova, I.V.: About accumulation rules for life-time predic-
tion under variable loading, in: Proceedings of the 12th International Conference on
Fracture, Ottawa, Canada. Natural Resources Canada & National Research Council
of Canada (2009).
[MiNa11] Mikhailov, S.E., Namestnikova, I.V.: History-sensitive accumulation rules for life-
time prediction under variable loading. Archive of Applied Mechanics (2011), DOI:
10.1007/s00419-011-0511-6.
[PeMa71] Penny, R., Marriott, D.: Design for Creep, McGraw-Hill, London (1971).
[Ra69] Rabotnov, Y.N.: Creep Problems in Structural Members, North-Holland Publ.,
Amsterdam–London (1969).
[Ra80] Rabotnov, Yu.N.: Elements of Hereditary Solid Mechanics, Mir Publishers, Moscow
(1980).
[Sa61] Savin, G.N.: Stress Concentration around Holes, Pergamon Press, Oxford (1961).
Some Thoughts on Methods for Evaluating
a Class of Highly Oscillatory Integrals
P.J. Harris
1 Introduction
When using the boundary integral method for solving problems in acoustics it is
necessary to evaluate integrals of the form
1
I= f (x)eikg(x) dx 1-D
0
1 1 (1)
ikg(x,y)
I= f (x, y)e dx dy 2-D
0 0
P.J. Harris
University of Brighton, UK,
e-mail: p.j.harris@brighton.ac.uk
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 203
DOI 10.1007/978-0-8176-8238-5_19, © Springer Science+Business Media, LLC 2011
204 P.J. Harris
2 Quadrature Methods
Consider the one-dimensional integral given in (1). One commonly used method for
deriving a quadrature rule is to pick a set of points {x1 , x2 , . . . , xn }, interpolate the
integrand at these points by some polynomial (or piecewise polynomial) and then
obtain an approximation to the integral by integrating the interpolating polynomial
exactly. However, for large values of k in integrals of the form (1) the interpolating
polynomial (or piecewise polynomial) is going to need a large number of interpola-
tion points if it is going to accurately approximate the original integrand. An alter-
native strategy is to use simple piecewise polynomials to independently interpolate
the functions f and g appearing in (1). This will require relatively few interpola-
tion points to accurately approximate the integrand. Figure 1 shows the results of
interpolating the real part of the integrand of (1) with
√ √
f (x) = x4 + 1, g(x) = x2 + 4x + 5 (2)
and k = 200 using 10 interpolation points. Here the curve labeled Integrand is the
actual integrand, the curve labeled Global is the obtained by interpolating the whole
integrand and the curve labeled Separate the the curve obtained by interpolating
f and g individually. Whilst the curves labeled Integrand and Separate are indis-
tinguishable, the curve where global interpolation has been used gives a very poor
approximation to the original integrand.
Fig. 1 The results of interpolating the real part of (1), with f and g given by (2), both globally and
separately using 10 interpolation points
0
f (x)eikg(x) dx ≈ ∑ fi (x)eikgi (x) dx (3)
i=1 xi
where
xi+1 − x x − xi
fi (x) = f (xi ) + f (xi+1 ),
xi+1 − xi xi+1 − xi
xi+1 − x x − xi
gi (x) = g(xi ) + g(xi+1 ).
xi+1 − xi xi+1 − xi
The integral on the right-hand side of (3) can be evaluated exactly to give a quadra-
ture rule in terms of the values of the functions f and g at the quadrature points. We
shall refer to this rule as the Parts Rule Method 0 (PRM0) rule.
We now follow the analysis presented in [HaCh09], based on the work of
[IsNo06], to obtain an improved rule. Assume that the function g (x) is strictly non-
zero for x ∈ [0, 1] and apply the parts rule to the original integral (1) to obtain
1
!1 1
f (x)eikg(x) g (x) f (x) − f (x)g (x) ikg(x)
f (x)eikg(x) dx = − e dx (4)
0 ikg (x) 0
ik[g (x)] 2
0
where the integral on the right-hand side can be evaluated using the above rule. This
defines the Parts Rule Method 1 (PRM1) rule.
Applying the parts rule a second time yields
1
f (x) g (x) f (x) − f (x)g (x) ikg(x) 1
f (x)e ikg(x)
dx = − e
0 ikg (x) i2 k2 (g (x))3 0
1
f (x)(g (x))2 − f (x)g (x)g (x) − 3 f (x)g (x)g (x) + 3 f (x)(g (x))2
+
0 (ik)2 (g (x))4
× e ikg(x)
dx (5)
where, again, the final integral can use evaluated using the rule (3) above. This is
the Parts Rule Method 2 (PRM2) rule.
Here it is worth noting that in (4) the final integral which has to be evaluated
approximately contains a factor of 1/k and in (5) there is a corresponding factor of
1/k2 . Hence, the numerical size of the integral that needs approximating diminishes
rapidly as the parameter k increases.
For the two-dimensional case, we subdivide the domain of the integral into trian-
gular sub-domains. The simplest way of doing this is to divide the original square
domain into an n × n grid of smaller squares and then divide each sub-square into
two triangles. In each triangle we can apply the change of variables
where f1 , f2 , f3 , g1 , g2 , g3 are the values of f and g at the three vertices of the triangle
and
J j = |(x2 − x1 )(y3 − y1 ) − (x3 − x1 )(y2 − y1 )|.
The integral on the right-hand side of (6) can be evaluated exactly, although the
exact form depends on the relative values of g1 , g2 and g3 (see [HaCh09] for further
details). This is the two-dimensional version of the PRM0 rule.
As in the one-dimensional case, we can apply the parts rule to the integral over
T j to obtain
f eikg f eikg
Ij = −
(ik)2 gx gy x=1,y=1 (ik)2 gx gy x=0,y=1
f eikg f eikg
− +
2
(ik) gx gy x=1,y=0 (ik) gx gy x=0,y=0
2
1
= > !y=1
1 gxy gx f + gxx gy f − gy gx fx
− 2
e ikg
dx
0 (ik) g2x g2y
y=0
1 1
1 fy gy − f gyy
dy dx
0 0 ik g2y
where I j has been used to denote the integral over the jth triangle. The is the two-
dimensional PRM1 rule.
Applying the parts rule a second time yields
= >
fy gy − gyy f
f ikg
Ij = − e
(ik)2 g2x g2y (ik)3 g3y gx
x=1,y=1
= >
fy gy − gyy f
f ikg
− − e
2 2
(ik) gx gy 2 3 3
(ik) gy gx
x=1,y=0
= >
fy gy − gyy f
f
− − eikg
(ik)2 g2x g2y (ik)3 g3y gx
x=0,y=1
= >
f fy gy − gyy f
+ − eikg |x=0,y=0
(ik)2 g2x g2y (ik)3 g3y gx
1
= > !y=1
gx gy fx − gxx gy f − gxy gx f w1
− − e ikg
dx
0 (ik)2 g2x g2y (ik)3 g2x g4y
y=0
Evaluating Oscillatory Integrals 207
1 1
= >
w2
+ eikg dx dy, (7)
0 0 (ik)2 g4y
where
w1 = −2gxy gx gy fy + gx g3y fxy + 3gyy gxy gx f − gyy gx gy fx − gyyx gx gy f
−gxx g2y fy + gxx gyy gy f ,
w2 = −3gyy gy fy + g2y fyy + 3g2yy f − gyyy gy f
where the remaining integrals can be evaluated using the appropriate one-dimen-
sional or two-dimensional rule from above. This is the two-dimensional PRM2 rule.
Again, we note that the final integral in (7) contains a factor of 1/k2 and so will
rapidly diminish as k increases.
Consider the one-dimensional integral given in (1). Make the change of variables
u = g(x) to give
1 g(1)
ikg(x) f (g−1 (u)) iku
f (x)e dx = −1
e du (8)
0 g(0) g (g (u))
where, again, we require g (x) to be non-zero in the range of integration. The integral
on the right-hand side of (8) is now simply a standard Fourier type integral and can
be efficiently evaluated using existing methods. However, this method require us to
be able to evaluate g−1 (u) for any value of u. In examples where it is possible to
obtain a simple expression for g−1 this is not a problem, but for other functions g it
may be necessary to use a numerical method. If the rule is used in conjunction with
a boundary integral method then it should be possible to devise an efficient scheme
for evaluating g−1 as the general form of g will be the same for all the integrals that
need to be evaluated. We shall refer to this rule as the Change Of Variables (COV)
rule.
For the two-dimensional case, the change of variables can be applied to either
x or y, and the choice of which to use may be problem or context dependent. For
brevity, here we shall only apply the change of variables to y. Making the change of
variables u = g(x, y) (as proposed in [KiDoGrSm09]) leads to
1 1 1 g(x,1)
f (x, g−1 (x, u)) iku
f (x, y)eikg(x,y) dy dx = −1
e du dx (9)
0 0 0 g(x,0) gy (x, g (x, u))
where the notation gy is used to denote the derivative with respect to y and the
inverse g−1 is the inverse function with respect to y. This is now a standard Fourier
type integral with respect to u and a straightforward integral with respect to x.
208 P.J. Harris
3 Numerical Examples
Figure 2 shows the results of applying the different quadrature rules to the integral
1< √
x2 +4x+5
x4 + 1eik dx (10)
0
for different values of the frequency parameter k using 10 subdivisions. The corre-
sponding results using 20 subdivisions are shown in Fig. 3.
Fig. 2 The log of the relative error (expressed as a percentage) when the various quadrature rules
are used to evaluate (10) using 10 subintervals
Fig. 3 The log of the relative error (expressed as a percentage) when the various quadrature rules
are used to evaluate (10) using 20 subintervals
Evaluating Oscillatory Integrals 209
These results show that PRM2 rule performs better than the PRM1 rule, and that
both perform better than the PRM0 rule. As noted above, due to the 1/k and 1/k2
factors in the integrals to be evaluated numerically in the PRM1 and PRM2 rules,
respectively, this is not entirely surprising. Overall, the COV rule gives the same
level of accuracy as the PRM1 rule and is much simpler to implement than either
the PRM1 or PRM2 rules. As expected, increasing the number of subintervals has
improved the accuracy of the rules.
In the two-dimensional case
1 1 √
1 2 2
< eik x +y +2x+2y dx dy (11)
0 0 2 2
x + y + 2x + 2y
was used due to its similarity to the integral which can arise when using the bound-
ary integral method to solve acoustic problems in three dimensions. The results of
applying each of the quadrature rules is shown in Fig. 4. This shows that for rela-
Fig. 4 The relative error (expressed as a percentage) when the various quadrature rules are used to
evaluate (11)
tively low frequency, the COV rule is more accurate than the others, but for higher
frequencies there is little to choose between the PRM1, PRM2 and COV methods
in terms of accuracy.
One of the main applications of the quadrature rules discussed above is in the so-
lution of integral equations with highly oscillatory kernel functions. Consider, for
example, the integral equation
210 P.J. Harris
1 √
− λ f (x) + eik x+y+1
f (y) dy = g(x) (12)
0
where g(x) is chosen to make f (x) = (x + 1)3 be the exact solution. Clearly as the
parameter k increases the kernel function will oscillate more rapidly (see Fig. 5,
which shows the real part of this kernel function for different values of k). The in-
Fig. 5 The kernel function appearing in (12) for different values of the parameter k
tegral equation (12) can be solved using the usual piecewise constant collocation
method, the details of which are not given here, where the PRM0 and COV methods
have been used to evaluate the required integrals. The errors in the computed solu-
tion using 100 elements are shown in Fig. 6, where both the PRM0 and the COV
rules have 11 quadrature points per element. The results using the two rules are in-
distinguishable, indicating that the choice of rule is not significant and that the error
is being dominated by the discretization error in the piecewise constant collocation
method and not the error in the numerical integration. For comparison purposes, a
compound Gauss rule was also used to carry out the integration, and it was found
that 2000 quadrature points per element were needed to obtain approximately the
same overall level of accuracy.
5 Conclusions
The methods introduced in this paper have shown that it is possible to develop
quadrature rules which can be used to quickly evaluate integrals of the form (1).
This paper compares two types of rule for evaluating such integrals and show that
they yield results which have broadly similar levels of accuracy.
The rules were then utilized in the solution of a simple integral equation with an
oscillatory kernel function and it was found that by using the high-frequency rules
were much more efficient at evaluating the necessary integrals where compared to a
conventional compound Gauss rule.
Evaluating Oscillatory Integrals 211
Fig. 6 The log of the relative error (expressed as a percentage) in the computed solution of (12)
using 100 piecewise constant elements
References
[HaCh09] Harris, P.J., Chen, K.: An efficient method for evaluating the integral of a class of
highly oscillatory functions. Journal of Comp. and App. Maths., 230, n. 2, 433–
442 (2009).
[HuVa05] Huybrechs, D., Vandewalle, S.: The efficient evaluation of highly oscillatory in-
tegrals in BEM by analytical continuation, in: Adv. Boundary Integral Methods
(Editor: K. Chen), Liverpool University Press, 20–30 (2005).
[IsNo06] Iserles, A., Norsett, S.P.: Quadrature methods for multivariate highly oscillatory
integrals using derivatives. Math. Comp., 75, n. 255, 1233–1258 (2006).
[KiDoGrSm09] Kim, T., Dominguez, V., Graham, I.G., Smyshlyeav, V.P.: Recent progress on
hybrid numerical-asymptotic methods for high-frequency scattering problems, in:
Proceeding of the 7th UK Conference on Boundary Integral Methods, University
of Nottingham, UK (2009).
[Le96] Levin, D.: Fast integration of rapidly oscillating functions. J. Comput. Appl.
Math., 67, n. 1, 95–101 (1996).
[Ol06] Olver, S.: On the quadrature of multivariate highly oscillatory integrals over non-
polytope domain. Numer. Math., 103, 645–665 (2006).
[XiWa08] Xiang, S., Wang, H.: On the Levin iterative method for oscillatory integrals. Jour-
nal of Computational and Applied Mathematics, 217, 38–45 (2008).
Numerical Experiments for Mammary
Adenocarcinoma Cell Progression
1 Introduction
An estimated 192, 370 women in the United States were diagnosed with breast can-
cer during 2009 [BCF]. Of these, over 40, 000 will succumb to the disease as a result
of widespread metastasis to secondary organs such as bone, lungs, and liver. These
statistics suggest the need for improved early detection techniques and treatment
options.
The most common types of breast cancer originate either in the milk ducts or
in the milk producing glands of the breast. Breast cancer arising from ducts is
called invasive ductal carcinoma (IDC), while cancer arising from glandular tis-
sue is called invasive lobular carcinoma (ILC). IDC comprises 70–80% of all breast
cancer, while ILC makes up only 8–14% of breast cancer cases [ArBaClEl04]. More
rare forms of breast cancer include inflammatory breast cancer and Paget’s disease
of the breast.
Ductal breast cancer progresses from noninvasive precancerous lesions to inva-
sive carcinoma. Inevitably, invasive adenocarcinoma will enter into the systemic
circulation and potentially develop secondary tumors as metastases. These highly
aggressive, metastatic cells are the source of many complications associated with
cancer and are the ultimate cause of death. In addition to the cancer cells themselves,
the tumor microenvironment plays an important role in successful tumor progres-
sion. Cells in this environment include endothelial cells that make up the blood ves-
C.L. Jorcyk
Boise State University, ID, USA,
e-mail: cjorcyk@boisestate.edu
M. Kolev
University of Warmia and Mazury, Olsztyn, Poland,
e-mail: kolev@matman.uwm.edu.pl
B. Zubik-Kowal
Boise State University, ID, USA,
e-mail: zubik@math.boisestate.edu
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 213
DOI 10.1007/978-0-8176-8238-5_20, © Springer Science+Business Media, LLC 2011
214 C.L. Jorcyk et al.
in media supplemented with charcoal-stripped fetal bovine serum (CS-FBS), the cell
lines were unable to proliferate and could not be rescued by supplemental estrogen
[H03]. Interestingly, the M28N2 cells are one of only a few mouse mammary cell
lines to express PgR, but 6-methyl-17-hydroxyprogesterone acetate (MPA) did not
have a proliferative effect on this cell line [H03].
The four cell lines can be distinguished by their growth characteristics in vitro.
Using a proliferation assay, 2 × 103 cells/well (passage 28–34) in a 24-well dish
were cultured in Dulbecco’s Modified Eagle Media (DMEM) containing 10% fetal
bovine serum (FBS). The cells were harvested by trypsinization and counted on
days 3, 6, and 9 using a hemacytometer. The in vitro growth rates were consistent
with increasing tumorigenicity in vivo [H03], and each cell line had the following
doubling time: M28N2 → 37 h, M27H4 → 25 h, M6→ 18 h, and M6C → 15 h [H03].
Subcutaneous injection of the cell lines into athymic, nude mice resulted in a
growth pattern that is consistent with the doubling times obtained in the in vitro
proliferation assay [H03]. Injection of the cell lines into nude mice yielded tumors in
100% of mice for the M6 (n = 8) and M6C (n = 7) cell lines, while only one of three
nude mice inoculated with the M27H4 and M28N2 cells displayed tumor growth.
This series of mammary adenocarcinoma cell lines, representing progressive stages
of tumorigenesis, presents a useful system for studying the molecular and genetic
alterations that accompany the development of invasive carcinoma.
3 Mathematical Model
to tumor growth, the inner state of the interacting individuals (cells, particles etc.)
usually denotes the biological activity of the individuals or the ability to express
their main functions. This is the reason for using the term “active particles” for such
microscopic interacting entities [LiSaBe07].
Here, we follow the idea of [Ko05, Ku97] and consider the interactions between
six main populations, each denoted by the subscript i. The subscript i = 1 refers to
the population of cancer cells (CCs). The other populations described in our model
are the helper T cells (Th cells, i = 2), the cytotoxic T lymphocytes (CTLs, i = 3), the
antigen presenting cells (APCs, i = 4), the antigen-loaded APCs (Ag-APC, i = 5),
and the cells of the host environment (HE, i = 6). As in [Ko05], the distribution
density of the i-th population with activation state u ∈ [0, 1] at time t ≥ 0 is denoted
by fi (t, u), fi : [0, ∞) × [0, 1] → R+ , i = 1, . . . , 6. Moreover, we denote by
1
ni (t) = fi (t, u)du, ni : [0, ∞) → R+ , i = 1, . . . , 6, (1)
0
the CCs. Therefore, we suppose that the activation state of CCs decreases, i.e. they
become more dangerous. The rate of generation of new tumor cells is assumed to
be proportional to the concentration of the existent CCs.
Specific Th1 cells and CTLs participate in the destruction of tumor cells. We
suppose that the rate of killing of CCs is proportional to the activation states of Th
cells and CTLs and obtain the equation describing the temporary evolution of the
population of cancer cells:
1
= 1
>
∂ f1 (1) (1)
(t, u) = p16 f1 (t, v)dv + t16 2u f1 (t, v)dv − u f1 (t, u)
2
∂t 0 u
1 1
− d12 f1 (t, u) v f2 (t, v)dv − d13 f1 (t, u) v f3 (t, v)dv+S1 (t), (2)
0 0
where S1 (t) is a source term describing the possible influx of CCs. Further, we
assume that the dynamics of the populations denoted by i = 2, 3 depend on the
following factors: the constant generation of T cells (Th cells and CTLs) by HE, the
production of T cells as well as the increase in the activation states of Th cells and
CTLs as a result of the interactions between Th cells and antigen-loaded APCs, the
destruction of T cells due to their interactions with CCs, the natural death of T cells,
and the possible influx of T cells.
It is observed that the HE constantly produces T cells and APCs [LyWhFa00].
The activity of newly generated T cells is low and it increases during their devel-
opment and selection [Ku97]. The interactions between Th cells and antigen-loaded
APCs lead to the production of new helper T cells and CTLs. We assume that the
rate of this generation is proportional to the states of activity of Th cells and that
the probability of production of highly active T cells is lower than the probability
of creation of less active T cells. The interactions between Th cells and antigen-
loaded APCs result in an increase in the activation states of Th cells. We suppose
that this change of the activity of Th cells depends on the concentration of antigen-
loaded APCs. The interactions between Th cells and CTLs lead to an increase in
the states of activity of CTLs through cytokines secreted by Th cells. We assume
that the increase of the activity of CTLs depends on the concentration of Th cells.
The interactions between T cells and CCs can result in apoptosis of Th cells and
CTLs [LyWhFa00]. Thus, the equations describing the temporary evolution of the
populations of Th cells and CTLs have the following forms:
1
∂ f2 (2) (2)
(t, u) = p16 (1 − u) + p25 (1 − u)n5 (t) v f2 (t, v)dv
∂t 0
= >
u
(2) 2
+ t25 n5 (t) 2 (u−v) f2 (t, v)dv−(1−u) f2 (t, u)
0
1
− d21 f2 (t, u) f1 (t, v)dv − d26 f2 (t, u) + S2 (t, u) (3)
0
and
218 C.L. Jorcyk et al.
1
∂ f3 (3) (3)
(t, u) = p16 (1 − u) + p25 (1 − u)n5 (t) v f2 (t, v)dv
∂t 0
= >
u 1
(3)
+ t23 2 (u−v) f3 (t, v)dv−(1−u)2 f3 (t, u) f2 (t, v)dv
0 0
1
− d31 f3 (t, u) f1 (t, v)dv − d36 f3 (t, u) + S3 (t, u). (4)
0
The functions Si (t, u), i = 2, 3 describe the possible inlet of Th cells and CTLs,
respectively.
For the temporary evolution of the fourth population of APCs, we suppose that
they are constantly generated by HE and additionally created (with a rate propor-
tional to the state of activity of Th cells) due to the interactions between Th cells
and antigen-loaded APCs. Some APCs are loaded with tumor antigens due to the
interactions between APCs and CCs [Ku97]. We suppose that the concentration of
new antigen-loaded APCs is proportional to the activation states of CCs. Let us note
that the corresponding term is a loss term for the fourth population of APCs and it
is a gain term for the fifth population of [Ag-APCs]. We also take into account the
process of natural death of APCs and consider the following equation describing the
dynamics of the population i = 4 of APCs:
1 1
d (4) (4) (5)
n4 (t) = p16 + p25 n5 (t) v f2 (t, v)dv − b14 n4 (t) v f1 (t, v)dv
dt 0 0
− d46 n4 (t). (5)
We also consider the possible source term S5 (t) of antigen-loaded APCs, the
natural death of [Ag-APCs] as well as their elimination by cancer cells. The equation
describing the time evolution of the population i = 5 of antigen-loaded APCs has
the following form:
1 1
d (5)
n5 (t) = b14 n4 (t) v f1 (t, v)dv − d51 n5 (t) f1 (t, v)dv
dt 0 0
− d56 n5 (t) + S5 (t). (6)
Recently, it has been demonstrated in [DrKoMaZu10] that the (2)–(6) have the
potential to describe the growth of breast tumor cells in humans. In this paper, we ap-
ply (2)–(6) to the C3(1)/Tag transgenic mouse model of mammary cancer described
in Sect. 2.
The entire model (2)–(6) for the interacting populations is a system of partial
integro-differential equations. Note that (2)–(6) is not complete and has to be sup-
plemented by initial conditions. We apply the experimental data of [H03] for the
initial conditions and for estimating the parameters of the model.
Mammary Adenocarcinoma 219
experiments lead to the source function S1 (t) = a1t + b1 and the parameter values,
which are listed in Table 1. The solutions n1 (t) computed for the mammary can-
cer cell lines from [H03] with the corresponding parameter values from Table 1 are
presented in Fig. 1. The numerical data are presented by the solid lines and com-
pared with the in vitro experimental data from [H03] indicated by ◦, , , ! for the
M28N2 , M27H4 , M6, and M6C cells, respectively.
We apply the in vivo data [H03] to the entire system (2)–(6) and by minimizing
the error between its numerical solutions and the laboratory data we estimate the
parameter values, which are listed in Table 2. The results of the numerical experi-
ments with the parameter values from Table 2 are presented and compared with the
in vivo data [H03] in Fig. 2.
Table 2 Parameter values for tumor growth in nude mice
Symbol M6C M6 M28N2 M27H4
1.7 · 10−1 1.3 · 10−1 5.9 · 10−3 9.7 · 10−5
(1)
p16
1.1 · 10−1 1.0 · 10−1 1.3 · 10−1 1.1 · 10−1
(2)
p16
6.6 · 10−2 8.4 · 10−2 1.1 · 10−1 1.4 · 10−1
(3)
p16
8.7 · 10−2 9.4 · 10−2 1.0 · 10−1 1.0 · 10−1
(4)
p16
1.0 · 10−1 1.0 · 10−1 1.2 · 10−1 1.1 · 10−1
(2)
p25
8.9 · 10−2 1.0 · 10−1 1.1 · 10−1 1.0 · 10−1
(3)
p25
1.0 · 10−1 9.9 · 10−2 1.1 · 10−1 1.1 · 10−1
(4)
p25
1.1 · 10−1 1.0 · 10−1 9.8 · 10−2 8.3 · 10−2
(1)
t16
1.1 · 10−1 9.0 · 10−2 9.4 · 10−2 1.5 · 10−1
(2)
t25
1.1 · 10−1 1.0 · 10−1 1.0 · 10−1 1.0 · 10−1
(3)
t25
9.9 · 10−2 1.0 · 10−1 1.0 · 10−1 1.0 · 10−1
(5)
b14
d12 8.7 · 10−2 9.7 · 10−2 9.9 · 10−2 9.5 · 10−2
d13 5.5 · 10−2 8.6 · 10−2 9.7 · 10−2 9.9 · 10−2
d21 1.3 · 10−1 1.1 · 10−1 1.1 · 10−1 1.1 · 10−1
d26 7.2 · 10−2 9.6 · 10−2 1.0 · 10−1 7.9 · 10−2
d31 9.1 · 10−2 8.6 · 10−2 8.9 · 10−2 8.9 · 10−2
d36 1.4 · 10−1 9.3 · 10−2 1.1 · 10−1 1.2 · 10−1
d46 1.0 · 10−1 1.0 · 10−1 1.1 · 10−1 1.0 · 10−1
d51 1.1 · 10−1 1.0 · 10−1 1.3 · 10−1 9.4 · 10−2
d56 8.4 · 10−2 9.4 · 10−2 8.3 · 10−2 1.1 · 10−1
Si 0 0 0 0
The parameter values from Table 2 and the comparison (Fig. 2) between the pre-
dicted and in vivo data allow us to make the following observations. The parameter
(1)
p16 , which characterizes the generation of new tumor cells, is larger for the more
advanced tumorigenic cell lines M6 and M6C than for the weakly tumorigenic cell
(1)
lines M27H4 and M28N2 . Additionally, p16 is larger for M6C than for M6 with the
difference being about 4 · 10−2 . This numerical observation is in agreement with
the fact that M6C and M6 were derived from tissues with different stages of tumor
progression and, contrary to M6, the M6C cell line was derived from a lung metasta-
Mammary Adenocarcinoma 221
sis consequently displaying more aggressive tumor invasion after the subcutaneous
injection than the in vivo growth rate displayed by the M6 cell line.
The M27H4 and M28N2 lines are less tumorigenic than M6 as they were devel-
oped from tissue representative of earlier stages of carcinogenesis. The differences
(1)
between the values of the parameter p16 for M6 and for the weakly tumorigenic
M27H4 and M28N2 cell lines is about 10−1 and the curves for M27H4 and M28N2
in Fig. 2 display slower growth rates than the curve for M6. The M28N2 cells were
(1)
derived from histologically normal mammary tissue and, interestingly, p16 is larger
for M28N2 than for M27H4 . Additionally, only one of three mice inoculated with the
M27H4 cells displayed tumor growth, which grew more slowly than one of the tu-
(1)
mors that was formed by M28N2 . However, the difference between the values of p16
−3
for M28N2 and M27H4 is only about 6 · 10 , which is in agreement with the fact
that the difference between the in vitro growth rates of the two cell lines was not sta-
tistically significant and the tumors formed by M28N2 and M27H4 were carcinomas
without metastases.
The parameter d12 , which characterizes the destruction of cancer cells due to the
activity of Th cells, has the smallest value for the most aggressive M6C cell line, and
interestingly, the largest value is obtained not for M27H4 (which displays the slowest
in vivo growth rate) but for M28N2 (displaying the second slowest in vivo growth
rate). This numerical observation is in agreement with the fact that the M28N2 cells
are derived from normal mammary tissue and are susceptible to transformations
(2)
and genetic and molecular alterations. This is also in agreement with p16 , which
characterizes the generation of new Th cells by HE, having the largest value for
M28N2 and approximately the same value for the other cell lines. The parameter d13
portraying the destruction of cancer cells due to the response of CTLs is smallest
in value for the most aggressive M6C cell line, larger for the second aggressive M6
222 C.L. Jorcyk et al.
cell line, again larger for M28N2 , and holds the largest value for the weakest M27H4
cell line; we conclude that d13 decreases with increasing tumorigenicity. The same
(3)
feature is observed for the parameter p16 , which characterizes the production of
new CTLs by HE; its value is smallest for the most aggressive M6C cell line, larger
for the second aggressive M6 cell line, then yet larger for M28N2 , and the largest for
(4)
the weakest M27H4 cell line. This is similar for the parameter p16 characterizing
the constant production of new APCs by HE.
We have analyzed cell culture and animal and mathematical models for the investi-
gation of mammary carcinoma progression. The development and characterization
of a progressive series of C3(1)/Tag mammary carcinoma cell lines have been de-
scribed by means of integro-differential equations of Boltzmann type. The numeri-
cal approximations to the solutions of the proposed mathematical model have shown
a good agreement with the laboratory data of the tumor growth rates. As the devel-
opment and utilization of experimental and computational methods can improve the
understanding of the carcinogenic processes, our future work will address expanded
mathematical models and new mammary adenocarcinoma cell lines derived from
mouse models.
References
[ArBeAnLa03] Arlotti, L., Bellomo, N., De Angelis, E., Lachowicz, M.: Generalized Kinetic
Models in Applied Sciences, Series on Advances in Mathematics for Applied
Sciences, 64, World Scientific Publishing Co., Inc., River Edge, NJ, xii+199
(2003).
[ArBaClEl04] Arpino, G., Bardou, V.J., Clark, G.M., Elledge, R.M.: Infiltrating lobular car-
cinoma of the breast: tumor characteristics and clinical outcome. Breast Can-
cer Research, 6, 149–156 (2004).
[BeFo94] Bellomo, N., Forni, G.: Dynamics of tumor interaction with the host immune
system. Math. Comput. Modelling, 20, 107–122 (1994).
[BeBe06] Bellomo, N., Bellouquid, A.: On the mathematical kinetic theory of active
particles with discrete states. The derivation of macroscopic equations. Math.
Comput. Modelling, 44, 397–404 (2006).
[BeSl06] Bellomo, N., Sleeman, B.: Preface in: Multiscale Cancer Modelling. Comput.
Math. Meth. Med., (special issue) 20, 67–70 (2006).
[BeMa06] Bellomo, N., Maini, P.: Preface in: Cancer Modelling (II). Math. Models Meth-
ods Appl. Sci., (special issue) 16, n. 7b, iii–vii (2006).
[Be07] Bellomo, N.: Modelling Complex Living Systems, Birkhäuser, Boston (2007).
[BeLiMa08] Bellomo, N., Li, N.K., Maini, P.K.: On the foundations of cancer modelling:
selected topics, speculations, and perspectives. Math. Models Methods Appl.
Sci., 18, 593–646 (2008).
Mammary Adenocarcinoma 223
[BeDe08] Bellomo, N., Delitala, M.: From the mathematical kinetic, and stochastic
game theory to modelling mutations, onset, progression and immune com-
petition of cancer cells. Physics of Life Reviews, 5, 183–206 (2008).
[BeDe06] Bellouquid, A., Delitala, M.: Modelling Complex Biological Systems – A Ki-
netic Theory Approach, Birkhäuser, Boston (2006).
[BCF] Breast Cancer Facts & Figures 2009–2010. American Cancer Society, Inc.,
Atlanta.
[Ca02] Calvo, A., Xiao, N., Kang, J., Best, C.J.M., Leiva, I., Emmert-Buck, M.R.,
Jorcyk, C., Green, J.E.: Alterations in gene expression profiles during prostate
cancer progression: functional correlations to tumorigenicity and down-regul-
ation of selenoprotein-P in mouse and human tumors. Cancer Research, 62,
5325–5335 (2002).
[CoChLe94] Cox, L.A., Chen, G., Lee, E.Y.: Tumor suppressor genes and their roles in
breast cancer. Breast Cancer Research Treat, 32, 19–38 (1994).
[AnLo08] De Angelis, E., Lodz, B.: On the kinetic theory for active particles: A model
for tumor-immune system competition. Math. Comput. Modelling, 47, 196–
209 (2008).
[LiSaBe07] De Lillo, S., Salvatori, M.C., Bellomo, N.: Mathematical tools of the kinetic
theory of active particles with some reasoning on the modelling progression
and heterogeneity. Math. Comput. Modelling, 45, 564–578 (2007).
[DrKoMaZu10] Drucis, K., Kolev, M., Majda, W., Zubik-Kowal, B.: Nonlinear modeling with
mammographic evidence of carcinoma, Nonlinear Anal. Real World Appl., 11,
4326–4334 (2010).
[DyBuWhHaH89] Dyson, N., Buchkovich, K., Whyte, P., Harlow, E.: The cellular 107K protein
that binds to adenovirus E1A also associates with the large T antigens of SV40
and JC virus. Cell, 58, 249–255 (1989).
[Ge00] Green, J.E., Shibata, M., Yoshidome, K., Liu, M., Jorcyk, C.L., Anver, M.R.,
Wigginton, J.M., Wiltrout, R., Shibata, E., Kaczmarczyk, S., Wang, W., Liu,
Z., Calvo, A., Couldrey, C.: The C3(1)/SV40 T-antigen transgenic mouse
model of mammary carncer: ductal epithelial cell targeting with multistage
progression to carcinoma. Oncogene, 19, 1020–1027 (2000).
[H03] Holzer, R.G., MacDougall, C., Cortright, G., Atwood, K., Green, J.E., Jorcyk,
C.L.: Development and characterization of a progressive series of mammary
adenocarcinoma cell lines derived from the C3(1)/SV40 large T-antigen trans-
genic mouse model. Breast Cancer Research Treat, 77, 65–76 (2003).
[Ja09] Jackiewicz, Z., Jorcyk, C.L., Kolev, M., Zubik-Kowal, B.: Correlation be-
tween animal and mathematical models for prostate cancer progression. Com-
put. Math. Methods Med., 10, 241–252 (2009).
[JaSe92] Jäger, E., Segel, L.: On the distribution of dominance in a population of inter-
acting anonymous organisms, SIAM J. Appl. Math., 52, 1442–1468 (1992).
[Jo98] Jorcyk, C.L., Liu, M.L., Shibata, M.A., Maroulakou, I.G., Komschlies, K.L.,
McPhaul, M.J., Reseau, J.H., Green, J.E.: Development and characterization
of a mouse prostate adenocarcinoma cell line: ductal formation determined by
extracellular matrix. Prostate, 34, 10–22 (1998).
[Ko05] Kolev, M.: A mathematical model of cellular immune response to leukemia.
Math. Comput. Modelling, 41, 1071–1081 (2005).
[Ku97] Kuby, J.: Immunology, 3rd edition, W.H. Freeman, New York (1997).
[LyWhFa00] Lydyard, P.M., Whelan, A., Fanger, M.W.: Instant Notes in Immunology,
BIOS Scientific Publishers Ltd., Oxford (2000).
[Ma94] Maroulakou, I.G., Anver, M., Garrett, L., Green, J.E.: Prostate and mammary
adenocarcinoma in transgenic mice carrying a rat C3(1) simian virus 40 large
tumor antigen fusion protein. Proc. Natl. Acad. Sci. USA, 91, 11236–11240
(1994).
224 C.L. Jorcyk et al.
[M9i2] Mietz, J.A., Unger, T., Huibregtse, J.M., Howly, P.M.: The transcriptional
transactivation function of wild-type p53 is inhibited by SV40 large T anti-
gen and by HPV-16 E6 oncoprotein. EMBO J., 11, 5013–5020 (1992).
[Os91] Osborne, R.J., Merlo, G.R., Mitsudomi, T., Venesio, T., Liscia, D.S., Cappa,
A.P., Chiba, I., Takahashi, T., Nau, M.M., Callahan, R., et al.: Mutations in the
p53 gene in primary human breast cancers. Cancer Research, 51, 6194–6198
(1991).
[Pa83] Parker, M.G.R., White, H., Hurst, M., Needham, M., Tilly, R.: Prostatic
steroid-binding protein: isolation and characterization of C3 genes. J. Biol.
Chem., 258, 12–15 (1983).
[ScVa10] Schmid, M.C., Varner, J.A.: Myeloid cells in the tumor microenvironment:
modulation of tumor angiogenesis and tumor inflammation. Journal of On-
cology, 1–10 (2010).
[SoshGrJo02] Soares, C.R., Shibata, M.A., Green, J.E., Jorcyk, C.L.: Development of PIN
and prostate adenocarcinoma cell lines: a model system for multistage tumor
progression. Neoplasia, 4, 112–20 (2002).
[Yo98] Yoshidome, K., Shibata, M., Maroulakou, I.G., Liu, M., Jorcyk, C.L., Gold,
L.G., Welch, V.N., Green, J.E.: Genetic alterations in the development of
mammary and prostate cancer in the C3(1)/Tag transgenic mouse model. Int.
J. Oncol., 12, 449–453 (1998).
Limiting Cases of Subdiffusion
J. Kemppainen
1 Introduction
We study the limiting cases with respect to the order of the time derivative of a time
fractional diffusion equation (TFDE) in a subdiffusive case, i.e. the order of the time
derivative is α ∈ (0, 1). The problem reads as follows:
⎧ α
⎨ ∂t u(x,t) − Δx u(x,t) = 0, (x,t) ∈ QT = Ω × (0, T ),
u(x,t) = g(x,t), (x,t) ∈ ΣT = Γ × (0, T ), (TFDE)
⎩
u(x, 0) = 0, x ∈ Ω,
J. Kemppainen
University of Oulu, Finland,
e-mail: jukemppa@paju.oulu.fi
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 225
DOI 10.1007/978-0-8176-8238-5_21, © Springer Science+Business Media, LLC 2011
226 J. Kemppainen
In our analysis we use the boundary integral approach to study the limits. Here we
recall some known results of the solvability of (TFDE) and describe the technique
to obtain the limits.
First of all, we have to clarify the function spaces which we are using. We assume
that the boundary datum g is continuous on ΣT . Moreover, for the unique solvability
of (TFDE) it is required that the compatibility condition
g(·, 0) = 0 on Γ
holds. It can then be proved that the (TFDE) admits a unique solution u ∈ C (QT )
[Ke10].
We recall the main steps for the existence and uniqueness proof. It is known that
a fundamental solution of (TFDE) is
1
20 ( 1 |x|2t −α |(α ,α )
π −n/2t α −1 |x|−n H12 ), x ∈ Rn ,t > 0,
Gα (x,t) = 4 (n/2,1),(1,1)
0, x ∈ Rn ,t < 0,
where x ∈ Ω , t ∈ (0, T ) and n(y) is the unit outward normal at y ∈ Γ . Clearly the
double-layer potential is an infinitely differentiable solution of (∂tα − Δ )u = 0 in QT
with u(·, 0) = 0. Therefore we need to choose ϕ such that the Dirichlet boundary
condition is satisfied.
Determining the boundary values of Dα ϕ leads to the corresponding boundary
integral equation (− 12 I +Wα )ϕ = g, where
t
(Wα ϕ )(x,t) = ∂n(y) Gα (x − y,t − τ )ϕ (y, τ )dσ (y)dτ , (x,t) ∈ ΣT .
0 Γ
Then Fredholm theory applies and we can prove existence and uniqueness of the so-
lution of the boundary integral equation. Therefore solution of the original problem
Limiting Cases of Subdiffusion 227
(ii) If z ≤ 1, then
|∂n(y) Gα (x − y,t)|
⎧
⎪ −α −1 |x − y|λ +1 | log(|x − y|2 t −α )| if n = 2,
⎨t
≤ C t −α −1 |x − y|λ if n = 3,
⎪
⎩ −α −1
t |x − y|λ +3−n if n ≥ 4.
for all 0 < t ≤ T in the case n = 3. The other values of n may be treated similarly.
Iterating the estimate (2) we obtain the estimate
t
(Wαk ϕ )(·,t)L∞ (Γ ) ≤ Ck I k−1 (t − τ )kαγ −1 ϕ (·, τ )L∞ (Γ ) dτ
0
228 J. Kemppainen
For the last equality we have used the properties of the Beta function.
Now there exists an integer k0 such that k0 αγ − 1 ≥ 0. Hence we have an estimate
t
(Wα 0 ϕ )(·,t)L∞ (Γ ) ≤ M ϕ (·, τ )L∞ (Γ ) dτ
k
(4)
0
for all 0 ≤ t ≤ T and M = Ck0 I k0 −1 max{1, (2T )k0 γ −1 }, which is uniformly bounded
on α ∈ (α0 , 1).
From (4) we can derive that
(MT )l
Wα 0 ϕ L∞ (ΣT ) ≤ ϕ L∞ (ΣT )
k l
(5)
l!
k l
for all l ∈ N. The estimate (5) implies that for some l the operator Wα 0 is a contrac-
tion. Since the same is true for some power of the operator 2Wα , we may conclude
that − 12 I + Wα is invertible and the inverse may be represented as the Neumann
series 1 −1 ∞
− I +Wα = −2 ∑ (2Wα )k .
2 k=0
The rest of this section is dedicated to the calculation of the limit limα →1− Wα ϕ .
We show that there exists an integrable majorant which is independent of α ∈
(α0 , 1). As before, we have to treat the cases z ≥ 1 and z ≤ 1 separately. More-
over, for small values of z we have to consider different values of n’s separately.
Since the technique is the same for all n, we restrict ourselves to the case n = 3.
From Lemma 1 we can obtain an estimate for large values of z:
∂n(y) Gα (x − y,t − τ )ϕ (y, τ )dσ (y)dτ
Γ ×(0,t)∩{|z|≥1}
t
(t − τ )α0 (γ − 2 )−1 |x − y|λ +1−2γ dσ (y)dτ
n
≤Cϕ L∞ (ΣT )
0 Γ
n+λ
for any γ > 0. The upper bound is integrable if we choose n
2 <γ < 2 . Similarly,
if z ≤ 1, we can obtain an estimate
Limiting Cases of Subdiffusion 229
∂n(y) Gα (x − y,t − τ )ϕ (y, τ )dσ (y)dτ
Γ ×(0,t)∩{|z|≤1}
t
≤Cϕ L∞ (ΣT ) (t − τ )α0 (γ −1)−1 |x − y|λ −2γ dσ (y)dτ ,
0 Γ
which is finite for any 1 < γ < 1 + λ2 . The previous two estimates together with
Lebesgue’s theorem of dominated convergence imply that the limit can be taken
inside the integral sign.
The next step is to calculate the limit limα →1− ∂n(y) Gα (x − y,t). Since
x − y, n(y) α −1
∂n(y) Gα (x − y,t) = −π −n/2 t Hα (z),
|x − y|n+2
where ·, · denotes the inner product in Rn , z = 14 |x − y|2t −α and we have denoted
1 Γ ( n2 + s)Γ (1 + s) Γ (1 + s) −s
Hα (z) = n+2 z ds,
2π i C Γ (α + α s) Γ (s)
where C is an infinite contour on the complex plane circulating the negative real
axis counterclockwise [Ke09]. Using the Stirling relation for the Gamma function,
we can conclude that there exists an integrable majorant for the integrand in Hα that
is independent of α . Therefore we can take the limit under the integral sign in Hα .
Using the properties of the Fox H-function and the Gamma function, we see that
1 x − y, n(y) 6 |x − y|2 7
lim ∂n(y) Gα (x − y,t − τ ) = exp − ,
α →1− 2n+1 π n/2 (t − τ )n+1 4(t − τ )
which is nothing but the normal derivative of the fundamental solution of the heat
equation. Therefore we have proved the following Theorem, which is our main re-
sult.
Theorem 1. The solution ϕα of the equation (− 12 I + Wα )ϕα = g converges in
C (ΣT ) as α → 1− to the solution of the equation (− 12 I +W1 )ϕ = g, where
t 6 |x − y|2 7
1 x − y, n(y)
W1 ϕ (x,t) = exp − ϕ (y, τ )dσ (y)dτ .
2(4π )n/2 0 Γ (t − τ )n+1 4(t − τ )
is uniformly bounded for α ∈ (0, 1). The same argument as in the previous section
implies that limα →0+ Wα exists in L (C 1 ([0, T ]; C (Γ )), C ([0, T ]; C (Γ ))).
For the limit we first move (t − τ )α −1 inside the integral which defines Hα and
observe that
t t
(t − τ )α s+α −1
lim ϕ (y, τ )dτ = ∂τ ϕ (y, τ )dτ = ϕ (y,t)
α →0+ 0 Γ (α + α s) 0
if Re s > 0. Due to the analyticity of the Fox H-function we may take the contour C
in the definition of Hα to be C = {s ∈ C : Re s = s0 > 0}. Then, using the properties
of the Fox H-function and the Gamma function, we have the following result.
Remark 2. Note that W0 is the usual boundary integral operator corresponding to the
Helmholz operator Δ + I (see [McL00] and [PrBrMa90]).
References
[EiKo04] Eidelman, S.D., Kochubei, A.N.: Cauchy problem for fractional diffusion equations.
J. Differential Equations, 199, 211–255 (2004).
[Ke09] Kemppainen, J.: Properties of the single layer potential for the time fractional diffu-
sion equation. Journal of Integral Equations and Applications (to appear).
[Ke10] Kemppainen, J.: Existence and uniqueness of the solution for a time-fractional dif-
fusion equation (submitted for publication).
[KiSa04] Kilbas, A.A., Saigo, M.: H-transforms: Theory and Applications, CRC Press, LLC
(2004).
[Kr99] Kress, R.: Linear Integral Equations, Second edition, Springer-Verlag Inc., New
York (1999).
[McL00] McLean, W.: Strongly Elliptic Systems and Boundary Integral Equations, Cam-
bridge University Press, USA (2000).
[Po99] Podlubny, I.: Fractional Differential Equations, Academic Press, San Diego, Cali-
fornia (1999).
[PrBrMa90] Prudnikov, A.P., Brychkov, Y.A., Marichev, O.I.: Integrals and Series, vol. 3. More
Special Functions, Overseas Publishers Association, Amsterdam (1990).
A New Hybrid Method to Predict
the Distribution of Vibro-Acoustic Energy
in Complex Built-Up Structures
1 Introduction
D.N. Maksimov
University of Nottingham, UK,
e-mail: dmitrii.maksimov@nottingham.ac.uk
G. Tanner
University of Nottingham, UK,
e-mail: gregor.tanner@nottingham.ac.uk
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 233
DOI 10.1007/978-0-8176-8238-5_22, © Springer Science+Business Media, LLC 2011
234 D.N. Maksimov and G. Tanner
same order as the characteristic scale are labeled “deterministic” and are treated with
FEM. The other subsystems are labeled “stochastic”—their response is modeled by
an SEA-like approach. A simple example of a structure split into stochastic and
deterministic components is shown in Fig. 1. We will also use this example for
numerical validations of our method.
We will work with the 2D Helmholtz equation with Neumann boundary condi-
tions on all boundaries, though the concepts are valid for general wave equations
such as vectorial elasticity as well as for 3D problems. The global solution is given
by a scalar wave field which we decompose in the form
ψ = ψ (d) + ψ (r) .
The direct field ψ (d) satisfies Neumann boundary conditions on the boundaries of
deterministic subsystems and absorbing boundary conditions on the boundaries of
the stochastic subsystems, also referred to as random boundaries in what follows.
The stochastic reverberant field ψ (r) arises from scattering from the random bound-
aries and subsequent interference. The total field ψ is the solution of the Helmholtz
equation over the whole system satisfying Neumann boundary conditions on all
boundaries, that is,
2 ψ + (ω 2 − iηω )ψ = f ;
here, η is a damping coefficient and f is the force leading to an excitation of the
system. Throughout this paper all quantities are given in dimensionless units.
Discretizing the problem by FEM, we obtain a set of linear equations
(K + M + R) q = g, (1)
where q is the vector of nodal degrees of freedom, g is the vector of nodal forces,
and K, M, and R are stiffness, mass, and damping matrices, respectively. Our aim is
to find the mean wave energies in the stochastic subsystems, E p , p = 1, . . . , P and
in the deterministic subsystems, En , n = 1, . . . , N, respectively. Equation (1) can
be written in the following form
⎛ ⎞⎛ ⎞ ⎛ ⎞
Dd Ddb 0 qd gd
⎝ Dbd Db Dbs ⎠ ⎝ qb ⎠ = ⎝ gb ⎠ , (2)
0 Dsb Ds qs gs
where
D = K + M + R,
and the subscripts d, s, and b denote degrees of freedom which lie in deterministic
subsystems, in stochastic subsystems, and on the interfaces (boundaries) between
these two, respectively.
Finding numerically exact solutions in the stochastic components of the structure
is computationally expensive due to the large number of degrees of freedom neces-
sary to model these domains using FEM. Our aim is therefore to eliminate qs in
(2) by using statistical assumptions on the wave field in the stochastic subsystems.
A Hybrid Method 235
2 Direct Field
In our method we construct two different FE models for the direct and reverber-
ant field. The direct field describes the radiation from the deterministic subsystems
to stochastic ones. It can be shown that the direct field everywhere in the stochas-
tic subsystems could be described by the direct field propagators Ls , Lb [MaTa10]
which are defined as
(in) (out) (out)
qb = L b q b , q s = L s q b ,
(in) (out)
where qb , qb are the incoming and outgoing wavefield components on the in-
terfaces with respect to the deterministic subsystems. This makes it possible to elim-
inate qs from the equations of motion for the direct field which gives
Dd Ddb qd gd
= , (3)
Dbd Db qout gb
where
Ddb = Ddb Lb (4)
and
Db = Db Lb + Dbs Ls .
In order to compute the energy flux into the stochastic subsystems via the direct
(np) (np)
field one needs the direct field components qb and qs on the (np) interface
which can be obtained from (3). The direct field energy input into the pth stochastic
subsystem per period T = 2π /ω is then given as
ω N
∑ ℑ{(qs
(d) (np) † (np) (np) (np)
Pp = ) (Ksb + Msb )qb }, (5)
2 n=1
(np) (np)
where Ksb , and Msb are submatrices of Ksb and Msb which account for interac-
tions at the (np) interface.
We assume that it is not possible or even desirable to compute the reverberant field
numerically exact due to the uncertainties in the geometry of the stochastic subsys-
tems. However, if the stochastic reverberant field can be treated as a random diffuse
236 D.N. Maksimov and G. Tanner
field, one can describe the response of the deterministic subsystems to the rever-
berant field in the neighboring deterministic subsystems. The reverberant field in
subsystem p leads to stochastic field components on (np) interfaces which can be
modeled in terms of stochastic reverberant forces. These forces in turn excite vi-
brations in the deterministic subsystems and will lead to transport of vibro-acoustic
energy to other stochastic subsystems p connected to n. The reverberant field in the
pth stochastic subsystem gives rise to a response in the nth deterministic subsystem
(r) (r)
in terms of a wave field (qd , qb ) which solves the equation
= >= >
(n) (n) (r)
Dd Ddb qd 0
(n) (n) (n) (np) (r) = (np) . (6)
Dbd Db + Dbs L̃s qb f
The right hand side of (6) contains the stochastic reverberant force
(np) (r)
f(np) = −Dbs qs (7)
acting only on the boundary degrees of freedom on the (np) interface. The modified
(np)
direct field propagator L̃s is defined as [MaTa10]
1
(np n)
L̃s = Ls
(np) at (np ) interfaces, p = p (8)
0 at (np) interface.
Note that (9) is written for the sub-blocks of the total Green’s function of subsys-
tem n which account for the subsystems’ response to sources located on the (np)
(np) (pnp )
interfaces. The matrix Gb again consists of sub-blocks Gb which describe
the response of the interface (np ). For the energy input from the pth subsystem to
the p th subsystem via a reverberant field produced by f(np) in the nth deterministic
subsystem, we can write the equivalent of (5), that is,
(np n)
where Ls is the direct field propagator on the (np ) interface. One can also find
the total flux of energy flowing from the pth stochastic subsystem into the nth de-
terministic subsystem, that is,
ω (np) † (np) (np) (pnp)
Pp→n = ℑ{{qs } (Ksb + Msb )Gb f(np) }.
2
So far everything is exact. We proceed now by taking averages over the power
inputs in terms of the reverberant field components. That is, we consider averages
P by determining fluxes over an ensemble of stochastic subsystems with different
boundaries. Using (7) and (10), one can write
* +
ω (p→p ,n) ∗
Pp→p = ℑ ∑ Ti, j
(n)
(qis ) qsj ,
2 i, j
where
(pnp ) † (np n) † (np ) (np ) (pnp )
T (p→p ,n) = D†bs {Gb } {Ls } (Ksb + Msb )Gb Dbs .
The average energy influx from the pth to the nth subsystem can be written in the
form * +
ω i ∗ j
P p→n = ℑ ∑ Ti, j (qs ) qs ,
p→n
2 i, j
where
(np) (np) (pnp)
T p→n = (Ksb + Msb )Gb Dbs .
(r)
The field variables qs enter now only in from of the correlation functions
∗
(qis ) qsj . The latter can for diffuse fields be express in the following form
∗
(qis ) qsj = C2p F(ri , r j ),
(kin) ω 2 ρ pC2p S p
E p = ,
2
where S p is the area of subsystem p, and ρ p is the density. Now, applying the virial
theorem, one can express C2p through the mean energy stored in the reverberant field
(r)
of the pth subsystem E p , that is,
(r)
E p
C2p = .
ρpω 2Sp
238 D.N. Maksimov and G. Tanner
The problem is thus reduced to finding the correlation function F(ri , r j ). The
issue of random field correlation function was discussed in [WeLo04]. The central
result is that the correlation function is equal to the imaginary part of the Green’s
function. It may seem that this does not simplify the problem, because we do not
know the Green’s function for the whole system. However, we can approximate
the global solution locally using the free-space Green’s function. In fact this issue
was thoroughly investigated in wave chaos theory [TaSo07] and it turns out that
the imaginary part of the free-space Green’s function well describes correlations in
Gaussian random fields.
Now we can determine the total energy influx into the pth stochastic subsystem
through its interfaces, that is,
(r)
1 P A pp E p
∑
(d)
Ppin = Pp + ,
2ω p =1
ρ p S p
(d)
where the direct field contribution Pp is given in (5) and
N * +
(p →p,n)
A pp = ∑ℑ ∑ Ti, j F(|ri − r j |) ;
n=1 i, j
note that A pp is zero if the p th subsystem is not directly connected to the pth sub-
system via a deterministic subsystem.
The mean energy flowing out of the pth subsystem either due to damping and or
due to energy transfer into neighboring deterministic subsystems can be written as
(r) (r)
ωη p (d) ∗ (d) η p E p 1 B p E p
Ppout = (ψ p ) ψ p dS p + + , (11)
2 ρp 2ω ρ p S p
Sp
where * +
N
∑ℑ ∑
(p→n)
Bp = Ti, j F(|ri − r j |)
n=1 i, j
and η p is the damping factor. We assume here that η p is constant across each
stochastic subsystem. The first term in (11) is the energy lost to damping in the
direct field ψ (d) which can be found with the use of an appropriate direct field prop-
agator L. The second term is the energy lost in subsystem p to damping and the last
term gives the energy transferred into adjacent deterministic subsystems.
We finally obtain a set of SEA-like equations using the energy balance relations
Ppin = Ppout . That makes it possible to reduce the problem to a set of linear equa-
A Hybrid Method 239
tions ⎧ in
⎪
⎪ P − P1out = 0
⎨ 1in
P2 − P2out = 0
⎪
⎪ ...
⎩ in
PP − PPout = 0
(r) (r) (r)
which can be solved for the unknowns E1 , E2 , . . . , EP .
After rewriting the set of equations in matrix form for the unknowns
⎛ ⎞ ⎛ (r) ⎞ ⎛ ⎞
M11 − ηρ11 · · · M1P E1 Q1
⎜ ⎟⎜ . ⎟⎟ ⎜ . ⎟
⎠⎜
.. .. ..
⎝ . . . ⎝ .. ⎠ = ⎝ .. ⎠ ,
M1P · · · MPP − ηρPP (r)
EP QP
Finally, we can find the total energy in each subsystem summing the mean rever-
(d) (d)
berant energies with the energies stored in the direct field E1 , . . . , EP , which can
be found using the direct field propagator L.
5 Numerical Example
The method has been numerically tested for the structure shown in Fig. 1 driven by
a δ -function force in in subsystem n2 . We solved the wave problem both using the
hybrid approach and by treating the full system by FEM. To introduce randomness
into the latter, we slightly distorted the shape of the boundaries in the stochastic sub-
systems keeping the areas constant and find the FEM solution for every realisation.
Then we computed the response averaged over an ensemble of stochastic subsys-
tems (random boundaries). The computational results are shown in Fig. 2, where
one can see a good agreement between both approaches.
Acknowledgements We thank David Chappell and Brian Mace for helpful discussions. Fruitful
discussions with Frank Vogel and Cathleen Seidel, inuTech GmbH (Nuremberg), are also acknowl-
edged. We have received financial support through the European Community’s Seventh Framework
Programme (FP7/2007-2013) under grant agreement PIAP-GA-2008-230597.
240 D.N. Maksimov and G. Tanner
Fig. 1 The system is split into subsystems: p1 and p2 are stochastic subsystems, n1 and n2 are
deterministic subsystems
Fig. 2 The energy in subsystem p1 found with FEM using Monte Carlo sampling over randomized
boundaries compared to the results obtained with the hybrid method. Light grey: each Monte Carlo
realization; black: ensemble average; bold grey: hybrid method; η = 0.20
References
[LyDe95] Lyon, R.H., DeJong, R.G.: Theory and Applications of Statistical Energy Analysis,
Butterworth-Heinemann, Boston (1995).
[ShLa05a] Shorter, P.J., Langley, R.S.: Vibro-acoustic analysis of complex systems. J. Sound Vib.,
288, 669–699 (2005).
[MaTa10] Maksimov, D.N., Tanner, G.: A hybrid approach for predicting the distribution of
vibro-acoustic energy in complex built-up structures. arXiv:1009.3651v1 (2010).
[WeLo04] Weaver, R.L., Lobkis, O.I.: Diffuse fields in open systems and the emergence of the
Green’s function. J. Acoust. Soc. Am., 116, 2731–2734 (2004).
[TaSo07] Tanner, G., Søndergaard, N.: Wave chaos in acoustics and elasticity. J. Phys. A: Math.
Theor., 40, 443–509 (2007).
2-D and 3-D Elastodynamic Contact Problems
for Interface Cracks Under Harmonic Loading
1 Introduction
Achievements of material science, such as with the new high-tech materials, make
it possible to significantly increase the strength and stiffness of designed structures.
However, the cost of an unpredictable fracture is always enormously high. Apart
from the increased economic costs due to increased safety requirements, it is nec-
essary to remember that in extreme cases the material or structural fracture can put
human life at risk. Therefore, the ultimate milestone of modern fracture mechanics
is fracture control, which allows the prediction of the construction behavior and the
avoidance of sudden collapse.
The presence of structural defects considerably decreases the strength and the re-
liability of materials as well as essentially increases the cost of exploitation. It takes
on special significance in the case of high rate deformations, as found in impact
and high-frequency dynamics, which covers an extremely wide range of situations
and is of interest to researchers from a number of different disciplines. Therefore
it is necessary to ensure that the residual strength of the cracked structure will not
fall below an acceptable level over the required service life. A considerable body of
work is devoted to the solution of fracture mechanics problems for different classes
O. Menshykov
University of Aberdeen, UK,
e-mail: o.menshykov@abdn.ac.uk
M. Menshykova
University of Aberdeen, UK,
e-mail: m.menshykova@abdn.ac.uk
I. Guz
University of Aberdeen, UK,
e-mail: i.guz@abdn.ac.uk
V. Mikucka
University of Aberdeen, UK,
e-mail: r02vm9@abdn.ac.uk
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 241
DOI 10.1007/978-0-8176-8238-5_23, © Springer Science+Business Media, LLC 2011
242 O. Menshykov et al.
2 Methodology
where Δ is the Laplace operator, λ (m) and μ (m) are the Lamé elastic constants, ρ (m)
is the specific material density.
The standard conditions of continuity for displacements, u(1) (x,t) = u(2) (x,t),
and stresses, p(1) (x,t) = −p(2) (x,t), where p(m) (x,t) is the traction vector, are sat-
isfied at the bonding interface Γ ∗ = Γ (1) ∩ Γ (2) . At the crack surface the traction
vectors g(m) (x,t), caused by the external loading, are given.
It is also assumed that there are no initial displacements of the points of the body,
and the Sommerfeld radiation-type condition, which provides a finite elastic energy
of an infinite body, was imposed at infinity on the vector of displacements.
In order to include the contact interaction of the opposite crack’s faces into con-
sideration, the Signorini unilateral constraints must be imposed for the normal com-
ponents of the contact force and the displacement vectors
where [u(x,t)] = u(1) (x,t) − u(2) (x,t) is the displacement discontinuity vector; and
q(x,t) is the contact force that arises in the contact region, which is unknown be-
forehand, changes in time under deformation of the material and must be determined
as a part of solution. The contact region also depends on the frequency, magnitude
and direction of the external loading complicating the problem even more and mak-
ing it highly non-linear [Gu02, Me08a, Me08c, Me08d, Me10]. The constraints (2)
ensure that there is no interpenetration of the opposite crack faces; the normal com-
ponent of the contact force is unilateral and it is absent for any non-zero opening of
the crack. Note that, due to the contact interaction the traction vector at the crack
faces, p̃(m) (x,t), is the superposition of the initial traction caused by the incident
wave, g(m) (x,t), and the contact force, q(x,t).
It must also be noted that in reality the friction of the opposite crack faces would
also affect the solution of the problem. So, the opposite crack faces remain immov-
able with respect to each other in tangential direction while they are held by the
friction. On the other hand, when the magnitude of the contact forces reaches a
certain limit, depending on the crack friction coefficient, kτ , and the normal con-
tact forces, the crack faces begin to move and the slipping effect occurs. Thus the
tangential displacements and contact forces must satisfy the Coulomb friction law:
∂ Δ uτ (x,t)
|qτ (x,t)| < kτ qn (x,t) ⇒ = 0, (3)
∂t
∂ Δ uτ (x,t) qτ (x,t) ∂ Δ uτ (x,t)
|qτ (x,t)| = kτ qn (x,t) ⇒ =− , (4)
∂t |qτ (x,t)| ∂t
x ∈ Ω , t ∈ [0; T ].
The allowance for crack faces contact interaction makes the resulting process a
steady-state periodic, but not a harmonic one. Hence, the components of the stress–
strain state cannot be represented as a function of coordinates multiplied by an expo-
244 O. Menshykov et al.
nential function. Here all components of the solution are expanded into the Fourier
series 1 ?
+∞
f (· ,t) = Re ∑ f k (· )eiωk t , (5)
k=−∞
T
ω
f k (· ) = f (· ,t)e−iωk t dt. (6)
2π
0
For the general case of the dynamic loading of the bimaterial with interface crack
the Somigliana dynamic representations for the components of the displacement and
traction fields in the upper and lower half-spaces can be used in order to obtain the
necessary system of boundary integral equations corresponding to the considered
problem [Gu06, Me08b], see also [Me07, Gu09a, Me09a, Me09b]. Using the re-
sults presented by the authors of the current study in [Me10], the following system
of boundary integral equations for the Fourier coefficients of displacements and
tractions at the bonding interface and the crack faces can be obtained for the case of
harmonic loading:
k,(1) (1) 1 k,(1)
− p̃i (y)Ui j (x, y, ωk ) dy = − u j (x)
2
Γ (1)cr
k,(1) (1) (1)
− ui (y)Wi j (x, y, ωk ) dy + uik,∗ (y)Wi j (x, y, ωk ) dy
Γ (1)cr Γ∗
(1)
− pik,∗ (y)Ui j (x, y, ωk )dy, x ∈ Γ (1)cr , (7)
Γ∗
k,(2) (2) 1 k,(2)
− p̃i (y)Ui j (x, y, ωk ) dy = − u j (x)
2
Γ (2)cr
k,(2) (2) (2)
− ui (y)Wi j (x, y, ωk ) dy − uik,∗ (y)Wi j (x, y, ωk ) dy
Γ (2)cr Γ∗
(2)
+ pik,∗ (y)Ui j (x, y, ωk )dy, x ∈ Γ (2)cr , (8)
Γ∗
k,(1) (1) 1
− p̃i (y)Ui j (x, y, ωk ) dy = − uk,∗ (x)
2 j
Γ (1)cr
k,(1) (1) (1)
− ui (y)Wi j (x, y, ωk ) dy + uik,∗ (y)Wi j (x, y, ωk ) dy
Γ (1)cr Γ∗
(1) ∗
− i (y)Ui j (x, y, ωk )dy, x ∈ Γ ,
pk,∗ (9)
Γ∗
Contact Problems for Interface Cracks 245
k,(2) (2) 1
− p̃i (y)Ui j (x, y, ωk ) dy = − uk,∗ (x)
2 j
Γ (2)cr
k,(2) (2) (2)
− ui (y)Wi j (x, y, ωk ) dy − i (y)Wi j (x, y, ωk ) dy
uk,∗
Γ (2)cr Γ∗
(2) ∗
+ i (y)Ui j (x, y, ωk )dy, x ∈ Γ ,
pk,∗ (10)
Γ∗
where α = 2 in 2-D case and α = 4 in 3-D case, δi j is the Kronecker delta, r is the
distance between the observation point and the load point. The appropriate expres-
sions for functions ψ (m) and χ (m) for 2-D case are given as [BaSl89, AlRo91]:
(m)
!
(m) 1 (m) c2 (m)
ψ = K0 (l2 ) + (m) K1 (l2 ) − (m) K1 (l1 ) ,
(m)
(12)
l2 c1
= (m) >2
(m) c (m)
χ (m) = K2 (l2 ) − 2(m) K2 (l1 ). (13)
c1
Here Kn (· ) is the modified Bessel function of the second kind and order n;
(m) (m) (m) (m) (m)
< (m)
l1
< = iω k r/c1 and l2 = iω k r/c2 ; c1 = (λ (m) + 2μ (m) )/ρ (m) and c2 =
μ (m) /ρ (m) are the velocities of the longitudinal and the transversal waves in the
material.
In the 3-D case functions ψ (m) and χ (m) are [BaSl89, AlRo91, Gu02, Me07]
= > (m) = >2 = > (m)
e−l2 e−l1
(m)
1 1 c2 1 1
ψ (m)
= (m)
+ (m)
+1 − (m) (m)
+ (m)
, (14)
(l2 )2 l2 r c1 (l1 )2 l1 r
= > (m) = >2 = > (m)
e−l2 e−l1
(m)
3 3 c2 3 3
χ (m)
= (m)
+ (m)
+1 − (m) (m)
+ (m)
+1 .
(l2 )2 l2 r c1 (l1 )2 l1 r
(15)
246 O. Menshykov et al.
(m)
The expression for the integral kernel Wi j (x, y, ωk ) is obtained from the Green
(m)
fundamental displacement vector Ui j (x, y, ωk ) by applying the following differen-
tial operator:
∂ [· ] ∂ [· ] ∂ [· ]
Pik [· , (y)] = λ ni (y) + μ δik + nk (y) . (16)
∂ yk ∂ n(y) ∂ yi
(m)
Hence the integral kernel Wi j (x, y, ωk ) has the form
3 Iterative Algorithm
1. specify an initial distribution of the traction g(m) (x,t) at the crack surface;
2. compute the Fourier coefficients of the known traction using (6);
3. find out the Fourier coefficients of the unknown displacements and tractions by
solving the system (7)–(10) neglecting the constraints (2);
4. compute the physical values of displacements and tractions using (5);
Contact Problems for Interface Cracks 247
5. applying the constraints, (2), compute and correct the distribution of the contact
forces by using the operators of orthogonal projection on the sets of constraints
where *
0, qn (x,t) ≤ 0,
Pn [qn (x,t)] = (19)
qn (x,t), qn (x,t) > 0,
Kn is the iterative coefficients the value of which is chosen to satisfy the optimal
algorithm convergence;
6. compute the new traction vector p̃(m) (x,t) at the crack surface;
7. proceed to the next iteration step if necessary.
Below, this algorithm is applied to the problem for the bimaterial with an inter-
face crack under normally incident tension–compression wave.
4 Numerical Results
Both 2-D (a linear interface crack with the half-length R) and 3-D (a penny-
shaped interface crack with the radius R) cases are considered in the present
study. The materials of the upper and lower half-spaces have the typical proper-
ties of steel and epoxy: E(1) = 207 GPa, E(2) = 4.6 GPa; ν (1) = 0.28, ν (2) = 0.36;
ρ (1) = 7800 kg/m3 , ρ (2) = 1380 kg/m3 ; and steel and aluminium E(2) = 70.3 GPa,
ν (2) = 0.35; ρ (2) = 2700 kg/m3 .
A time-harmonic tension–compression wave propagates normally to the inter-
face. It is assumed that the incident wave can be described by the potential function
(1)
xβ −ω t)
Φ (x,t) = Φ0 ei(k1 (20)
where β = 2 in 2-D and β = 3 in 3-D case; Φ0 and ω are the amplitude and the
(1) (1)
frequency of the incident wave; k1 = ω /c1 is the generalized wave number. This
potential generates the displacement field u(x,t) = gradx Φ0 (x,t) and the normal
traction gn (x,t) caused by the incident wave is
(1)
gn (x,t) = −(k1 )2 Φ0 cos(ω t). (21)
The dynamic stress intensity factors (the opening and the transverse shear modes)
were computed in the vicinity of the crack front using the following asymptotic
expressions [Ch79, BaSl89, AlRo91]:
√ √
KI = max lim |p∗n (R + r,t)| 2π r, KII = max lim |p∗τ (R + r,t)| 2π r, (22)
t r→0 t r→0
where p∗n (R + r,t) and p∗τ (R + r,t) are the normal and tangential tractions at the
bonding interface; r is the distance from the crack front.
248 O. Menshykov et al.
The algorithm convergence for different wave numbers is given in Figs. 1–4 for
both 2-D and 3-D problems. The stress intensity factors (opening mode) as functions
of the wave number (normalized frequency of the incident wave) are presented in
Figs. 5 and 6. All results are normalized by the corresponding static stress intensity
factor. Similar distributions are obtained for the transverse shear mode of the stress
intensity factors.
Fig. 1 Stress intensity factor (opening mode) plotted against the number of iterations, 2-D, steel–
epoxy
Fig. 2 Stress intensity factor (opening mode) plotted against the number of iterations, 2-D, steel–
aluminium
The maximal values of the comparable stress intensity factors do not coincide
and are achieved at different wave frequencies. And, most importantly, the maxi-
mal value of the stress intensity factor evaluated taking into account the contact of
Contact Problems for Interface Cracks 249
Fig. 3 Stress intensity factor (opening mode) plotted against the number of iterations, 3-D, steel–
epoxy
Fig. 4 Stress intensity factor (opening mode) plotted against the number of iterations, 3-D, steel–
aluminium
crack faces is much smaller than previously predicted (i.e. disregarding the contact
interaction). At the same time, for structures working within certain ranges of the
external loading, the maximal value of the stress intensity factor can be considerably
higher than previously predicted.
5 Conclusion
The current paper is devoted to the study of linearly elastic bimaterials with interface
cracks under normally incident time-harmonic tension–compression wave. The 2-D
250 O. Menshykov et al.
Fig. 5 Stress intensity factor (opening mode) plotted against the wave number, 2-D
Fig. 6 Stress intensity factor (opening mode) plotted against the wave number, 3-D
and 3-D problems are solved by the boundary integral equations method and the
Somigliana dynamic identity for the displacement field.
The effects of the crack closure on the solution of the problem (particularly on
the distributions of the stresses and strains) are considered. The convergence of the
iterative algorithm is analyzed for different wave numbers; and the dynamic stress
intensity factors (the opening mode) are computed as functions of the frequency
of the incident wave and compared with those obtained neglecting the crack’s clo-
sure. Analysis of the results shows that the crack closure significantly changes the
solution and therefore must be taken into account.
Contact Problems for Interface Cracks 251
References
[AlRo91] Aliabadi, M.H., Rook, D.P.: Numerical Fracture Mechanics, Computational Mechanics
Publications and Kluwer Academic Publishers (1991).
[BaSl89] Balas, S., Sladek, J., Sladek, V.: Stress Analysis by Boundary Element Methods, Else-
vier, Amsterdam (1989).
[Ch79] Cherepanov, G.P.: Mechanics of Brittle Fracture, McGraw Hill, New York (1979).
[Ch05] Chen, Z.T.: Interfacial coplanar cracks in piezoelectric bi-material systems under pure
mechanical impact loading. Int. J. Solids. Struct., 43, 5085–5099 (2005).
[Co77] Comninou, M.: The interface crack. J. Appl. Mech., 44, 631–636 (1977).
[Di02] Dineva, P., Gross, D., Rangelov, T.: Dynamic behavior of a bi-material interface-
cracked plate. Eng. Fract. Mech., 69, 1193–1218 (2002).
[Ga07] Garcia-Sanchez, F., Zhang, Ch.: A comparative study of three BEM for transient dy-
namic crack analysis of 2-D anisotropic solids. Computat. Mech., 40, 753–769 (2007).
[Go76] Goldstein, R.W., Vainshelbaum, V.M.: Axisymmetric problem of a crack at the inter-
face of layers in a multi-layered medium. Int. J. Eng. Sci., 14, n. 4, 335–352 (1976).
[Guo09] Guo, S.: Elastic wave scattering from a penny-shaped interface crack in coated materi-
als. Arch. Appl. Mech., 79, 709–723 (2009).
[Gu02] Guz, A.N., Zozulya, V.V.: Elastodynamic unilateral contact problems with friction for
bodies with cracks. Int. Appl. Mech., 38, 895–932 (2002).
[Gu06] Guz, I.A., Menshykov, O.V., Menshykov, V.A.: Application of boundary integral equa-
tions to elastodynamics of an interface crack. Int. J. Fract., 140, 277–284 (2006)
[Gu09a] Guz, A.N., Guz, I.A., Menshykov, O.V., Menshykov, V.A.: Penny-shaped interface
crack under shear wave. Int. Appl. Mech., 45, n. 5, 534–539 (2009).
[Gu09b] Guz, A.N., Zozulya, V.V.: On dynamical fracture mechanics in the case of polyhar-
monic loading by P-waves. Int. Appl. Mech., 45, n. 9, 1033–1039 (2009).
[It80] Itou, S.: Transient analysis of stress waves around two coplanar Griffith cracks under
impact load. Eng. Fr. Mech., 13, 349–356 (1980).
[Ka93] Kachanov, M.: Elastic solids with many cracks and related problems, in: Advances
in Applied Mechanics, vol. 30 (Editors: J.W. Hutchinson, T.Y. Wu), Academic Press,
259–445 (1993).
[Ki06] Kilic, B., Madenci, E., Mahajan, R.: Energy release rate and contact zone in a cohesive
and an interface crack by hypersingular integral equations. Int. J. Solids Struct., 43,
1159–1188 (2006).
[Lo73] Loeber, J.F., Sih, G.C.: Transmission of anti-plane shear waves past an interface crack
in dissimilar media. Eng. Fract. Mech., 5, 699–725 (1973).
[Ma06] Martin, P.A.: Multiple Scattering. Interaction of Time-Harmonic Waves with N Obsta-
cles, Cambridge University Press (2006).
[Me07] Menshykov, O.V., Menshykov, V.A., Guz, I.A.: The effect of frequency in the problem
of interface crack under harmonic loading. Int. J. Fract., 146, 197–202 (2007).
[Me08a] Menshykov, O.V., Guz, I.A.: Elastodynamic problem for two penny-shaped cracks: the
effect of cracks’ closure. Int. J. Fract., 153, 69–76 (2008).
[Me08b] Menshykov, O.V., Guz, I.A., Menshykov, V.A.: Boundary integral equations in elasto-
dynamics of interface crack. Phil. Trans. R. Soc. A., 366, 1835–1839 (2008).
[Me08c] Menshykov, O.V., Menshykov, V.A., Guz, I.A.: The contact problem for an open
penny-shaped crack under normally incident tension–compression wave. Eng. Fract.
Mech., 75, 1114–1126 (2008).
[Me08d] Menshykov, O.V., Menshykova, M.V., Guz, I.A.: Effect of friction of the crack faces
for a linear crack under an oblique harmonic loading. Int. J. Eng. Sci., 46, 438–458
(2008).
[Me09a] Menshykov, O.V., Menshykov, V.A., Guz, I.A.: Elastodynamics of a crack on the bi-
material interface. Eng. Anal. Bound. Elem., 33, 294–301 (2009).
[Me09b] Menshykova, M.V., Menshykov, O.V., Guz, I.A.: Linear interface crack under plane
shear wave. CMES: Comput. Model. Eng. Sci., 48, n. 2, 107–120 (2009).
252 O. Menshykov et al.
[Me10] Menshykova, M.V., Menshykov, O.V., Guz, I.A.: Modelling crack closure for an inter-
face crack under harmonic loading. Int. J. Fract., 165, 127–134 (2010).
[My06] Mykhas’kiv, V., Zhang, Ch., Sladek, J., Sladek, V.: A frequency-domain BEM for non-
synchronous crack interaction analysis in elastic solids. Eng. Anal. Bound. Elem., 30,
167–175 (2006).
[Qu94] Qu, J.: Interface crack loaded by a time-harmonic plane wave. Int. J. Solids. Struct., 31,
n. 3, 329–345 (1994).
[Qu95] Qu, J.: Scattering of plane waves from an interface crack. Int. J. Eng. Sci., 33, n. 2,
179–194 (1995).
[Wa01] Wang, Y.S., Gross, D.: Interaction of harmonic waves with a periodic array of interface
cracks in a multi-layered medium: anti-plane case. Int. J. Solids Struct., 38, 4631–4655
(2001).
[Wü09] Wünsche, M., Zhang, Ch., Sladek, J., Sladek, V., Hirose, S., Kuna, M.: Transient dy-
namic analysis of interface cracks in layered anisotropic solids under impact loading.
Int. J Fract., 157, n. 1–2, 131–147 (2009).
[Zh89] Zhang, Ch., Achenbach, J.D.: Time-domain boundary element analysis of dynamic
near-tip fields for impact-loaded collinear cracks. Eng. Frac. Mech., 32, n. 6, 899–909
(1989).
[Zh91] Zhang, Ch.: Dynamic stress intensity factors for periodically spaced collinear antiplane
shear cracks between dissimilar media. Theor. Appl. Fract. Mech., 15, 219–227 (1991).
[Zo02] Zozulya, V.V., Men’shikova, M.V.: Study of iterative algorithms for solution of dy-
namic contact problems for elastic cracked bodies. Int. Appl. Mech., 38, n. 5, 573–577
(2002).
Asymptotic Behavior of Elliptic Quadratic
Algebraic Equations with Variable Coefficients,
and Aerodynamical Applications
A. Nastase
1 Introduction
The quadratic partial differential equations that govern mathematical physics can be
reduced to the study of equivalent quadratical algebraic equations (QAEs) with vari-
able coefficients. The qualitative analysis of the asymptotical behaviors of the QAEs
in the vicinity of their critical lines (surfaces or hypersurfaces) gives the possibility
to find the critical lines (surfaces or hypersurfaces) of the PDEs of mathematical
physics. Further, let us consider a QAE of elliptic or hyperbolic type:
!
M M
∑ ∑ ai j xi x j + 2ai,M+1 xi + aM+1,M+1 = 0. (1)
i=1 j=1
The discriminant δ and the great determinant Δ of this QAE are the following:
a11 a12 . . . a1M
a21 a22 . . . a2M
δ = . . ..
.. .. .
aM,1 aM,2 . . . aM,M
and
a11 a12 . . . a1,M a1,M+1
a21 a22 . . . a2,M a2,M+1
Δ = ... ..
.
..
.
..
. .
aM,1 aM,2 . . . aM,M aM,M+1
aM+1,1 aM+1,2 . . . aM+1,M aM+1,M+1
A. Nastase
RWTH, Aachen University, Germany,
e-mail: nastase@lafaero.rwth-aachen.de
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 253
DOI 10.1007/978-0-8176-8238-5_24, © Springer Science+Business Media, LLC 2011
254 A. Nastase
Equation (1), after translation and rotation, is further written in the canonical form,
M
∑ λi xi2 + a = 0. (2)
i=1
The free term of this canonical equation is a = Δ /δ and the eigenvalues λi are ob-
tained as solutions of the characteristic equation Δc = 0. Here Δc is the characteristic
determinant of the QAE (2), namely
a11 − λ a12 . . . a1,M
a21 a22 − λ . . . a2,M
Δc = . .. .. .
.. . .
aM,1 aM,2 . . . aM,M − λ
The qualitative analysis is further performed for the QAE (1), written in the canon-
ical form (2).
In the further considerations it is supposed that the free term of the QAE, written
in the canonical form (2), is elliptic and has only a variable free term, namely, a =
aM+1,M+1 , which is varied from −∞ to ∞. For elliptic and hyperbolic QAEs the
discriminant δ = 0. For the elliptic QAEs the eigenvalues λi are all of the same sign.
The elliptic type of the QAE (1), its characteristic equation, its eigenvalues and its
discriminant remain unchanged during the variation of free term, because they do
not depend on it. Each elliptic QAE has one critical point, which is obtained for the
critical value of the free term a = ac . This critical value is obtained by canceling
of great determinant Δ . The QAE (2) can be always visualized in its principal two-
dimensional cuts, independent of the number M of its variables. The number of two-
dimensional principal cuts is Mc ≡ CM 2 = M(M − 1)/2, where C2 is a combinatorial
M
number. The critical point is located in the origin of coordinates.
• If a < ac , the QAE (2) is represented in each principal cut by coaxial ellipses,
which sizes decrease, when the values of the free term a of the QAE (1) increase
from −∞ to ac .
• If a = ac , the corresponding ellipses collapse in their critical point, which is
located in the origin of the coordinates of QAE (2).
• If a > ac , the QAE (2) has no more real solutions.
This behavior is independent of the number of independent variables M, which also
determines the dimension of the space of visualization. Let us further consider a
two-dimensional (M = 2) elliptic QAE with variable free term a33 = a, namely
This QAE takes, after translation and rotation, the following canonical form:
As an example of the theory, this QAE is taken in the following particular form:
This QAE is of elliptic type, because its two eigenvalues are positive, namely
λ1 = 1.764,
λ2 = 6.236.
The discriminant δ1 and the great determinant Δ1 of the QAE (4) are
δ1 = 11,
1
Δ1 = (44a − 135).
4
The critical value of the free term, which cancels the great determinant Δ1 , is
a ≡ ac = 135/44 = 3.068.
The visualization of the collapse of the coaxial ellipses corresponding to the QAE
(5) is presented in Fig. 1. Further visualizations for M = 3 and M = 4 are presented
in [Na96, Na97, Na07].
Fig. 1 Visualization of the collapse of the coaxial ellipses for the critical value a ≡ ac = 3.068 and
for M = 2
256 A. Nastase
In the further consideration it is firstly supposed that the two-dimensional QAE (3)
has two variable coefficients, namely the free term a33 = a and also the coefficient
of one of the free terms, namely a13 = d1 of x,
and the other coefficients of this QAE are supposed constants. This QAE has a
critical parabolic line obtained by canceling of its great determinant, namely Δ1 = 0.
The implicit and the explicit forms of this critical parabola are the following:
1
a= [a22 d12 − 2a12 a23 d1 + a11 a223 ]. (8)
δ1
The implicit equation (7) of the critical parabola, treated as a QAE of second order
with d1 as variable, has its following discriminant:
δ p = 4 δ1 (a22 a − a223 ).
For each value of a, for which δ p > 0, the QAE (6) has two real critical values for d1 ,
namely dc1 and dc2 (dc1 < dc2 ), which are the solutions of QAE (7). For the value
a = a0 , for which δ p = 0, a double solution for the critical point occurs, namely,
d0 = dc1 = dc2 . If the critical parabola is represented in a Cartesian plane, in which
the coefficients a and d1 are taken as variables, the point P(a0 ,d0 ) represents the
peak of critical parabola and the coordinates of the peak are the following:
a223
a0 = , (9)
a22
a12 a23
d0 = . (10)
a22
The critical parabola is tangent on its peak to the axis a = a0 and the line d = d0
is the axis of symmetry of this parabola. For the values of a, for which δ p < 0, the
QAE (6) has no more critical solutions. For the exemplification and visualization of
the critical parabola, the QAE (6) is taken in the following particular form:
The discriminant, the characteristic equation, the eigenvalues and the elliptic type
of (11) are the same as by QAE (3), because they do not contain the free term or the
coefficients of the linear terms. Only the behavior of the great determinant Δ1 , which
now depends on the both variable coefficients a and d1 , is changed. The implicit and
Quadratic Algebraic Equations with Variable Coefficients 257
explicit forms of the critical parabola of the QAE (11), are the following:
1
a= (20d12 + 24d1 + 27). (13)
44
The visualization of this critical parabola of elliptic QAE (11) is made in Fig. 2.
The coordinates of the peak P of critical parabola of elliptic QAE (11), computed
as in (9) and (10), are a0 = 0.45 and d0 = −0.6. The line d = d0 is the symmetry
axis of the parabola and the extremal value of the free term a33 = a = a0 is, for the
elliptic QAE (11), a minimum value of the critical parabola.
• If a < a0 , the QAE (11) has no critical points.
• If a = a0 , the QAE (11) has a double critical point located at the peak P(a0 , d0 )
of the critical parabola, corresponding to the double solution of the critical value
of d1 = a13 , dc1 = dc2 = d0 .
• If a > a0 , for each such value of a, there exist two critical points M 1 and M 2 ,
which correspond to the critical values dc1 and dc2 (dc1 < dc2 ) of the coefficient
a13 = d1 . For d1 < dc1 and d1 > dc2 and the elliptic QAE (11) has no critical
points. For the values of d1 between these two critical points (it is dc1 < d1 < dc2 )
the elliptic QAE (11) has no more real solutions.
The interior of the critical parabola is a black hole for the QAE (11). The QAE (11)
has in the black zone no more real solutions.
Now, in the QAE (3) beside the free term a = a33 , the coefficient of the other
linear term d2 = a23 (of y) is variable,
258 A. Nastase
and the other coefficients are taken constants. The discriminant, the characteristic
equation, the eigenvalues and the elliptic type of (11) are the same as by QAE (3),
because they do not contain the free term and also the coefficients of linear terms.
Only the behavior of the great determinant Δ1 , which now depends on the both
variable coefficients a and d2 , is changed. The QAE (14) has also a critical parabolic
line, obtained by canceling of its great determinant, namely Δ1 = 0. The implicit and
the explicit forms of this critical parabola are the following:
1
a= [a11 d22 − 2a12 a13 d2 + a22 a213 ]. (16)
δ1
The implicit equation (15) of the critical parabola, treated as a QAE of second order
with d1 as variable, has its following discriminant:
δ p = 4 δ1 (a11 a − a213 ).
For each value of a, for which δ p > 0, the QAE (14) has two real critical values for
d1 , namely dc1 and dc2 (dc 1 < dc 2 ), which are the solutions of QAE (15).
For the value a = a0 , for which δ p = 0, a double solution for the critical point
occurs, namely, d0 = dc 1 = dc 2 . If the critical parabola is represented in a Cartesian
plane, in which the coefficients a and d2 are taken as variables, the point P(a0 , d0 )
represents the peak of critical parabola and the coordinates of the peak are the fol-
lowing:
a2
a0 = 13 , (17)
a11
a12 a13
d0 = . (18)
a11
For the values of a for which δ p < 0 there are no critical points. Further, it is to
remark that the formulas (15)–(18) and of δ p can be also automatically generated
from the corresponding formulas (7)–(10) and δ p , by permutation of indices 1 and 2.
The critical parabola is tangent on its peak to the axis a = a0 and the line d2 = d0 is
the axis of symmetry of this parabola.
Now the free term a = a33 and the coefficients d1 = a13 and d2 = a23 , of the
both linear terms of the QAE (3), are simultaneous varied. The corresponding crit-
ical surface, in the space of variable coefficients, is obtained by canceling its great
determinant and is written in the explicit form:
4 Aerodynamical Applications
The author has reduced the study of the partial differential equations (PDEs) of the
Navier–Stokes layer (NSL) over a flying configuration (FC) in subsonic flow to the
study of elliptic QAEs with variable coefficients. Hybrid zonal solutions for these
PDEs are proposed under the following forms as in [Na07], it is:
N N N
uδ = ue ∑ ui η i , vδ = ve ∑ vi η i , wδ = we ∑ wi η i , (20)
i=1 i=1 i=1
where
(η = (x3 − Z(x1 , x2 ))/δ (x1 , x2 )).
Further, the spectral forms for the density function R = ln ρ and the absolute temper-
ature T are used:
N
R = Rw + (Re − Rw ) ∑ ri η i ,
i=1
N
T = Tw + (Te − Tw ) ∑ ti η i .
i=1
Here Z(x1 , x2 ) and δ (x1 , x2 ) are the equations of the surface of FC and of the thick-
ness of the NSL, Rw and Tw are the values of R and T at the wall and ue , ve , we ,
Re , Te are the edge values of uδ , vδ , wδ , R, T , which can be easily obtained from
the outer inviscid flow, at the NSL’s edge. For the determination of the pressure p,
the equations of perfect gas and an exponential law for the determination of the
viscosity μ as function of the absolute temperature T are used, it is
n1
T
p = Rg ρ T = Rg eR T, μ = μ∞ .
T∞
Let us consider now a set of QAEs defined by their following common properties:
they are of the same type (elliptic, hyperbolic or parabolic), have the same number
M of independent variables, the same coefficients variables and the other coeffi-
cients are constant. It is here supposed that all the QAEs belonging to this set have,
qualitatively, similar asymptotic behaviors in the vicinity of their singular points,
lines, surfaces or hypersurfaces. If the subsonic steady flow over FCs is considered,
the corresponding QAEs have elliptic character and each of the three equivalent im-
pulse QAEs leads to a critical hypersurface, which is obtained by canceling of its
corresponding great determinant. The cancelation of the great determinant leads to a
critical relation among the spectral coefficients of velocity’s components, which can
be visualized (in two-dimensional cuts) in the, here introduced, space of variable co-
efficients. The inside of this hypersurface is a hole for the elliptic QAE of impulse.
In this zone the elliptic QAE has no more real solutions. In vicinity of its critical
hypersurface the steady subsonic flow has three possibilities: to contour the hole by
detachment, to produce a reversal flow or to penetrate inside the hole by changing
the type of QAE from elliptic to hyperbolic type, by becoming non-stationary and
generating the transition. All these possibilities occur for the subsonic flow over
flying configurations.
5 Conclusions
The elliptic QAEs with variable coefficients present critical holes, which can be vi-
sualized in the space of variable coefficients introduced here. Inside these holes, the
elliptic QAEs have no more real solutions. The Navier–Stokes PDEs of impulse can
be written in a form of equivalent QAEs, with variable coefficients. The stationary
subsonic flow, governed by elliptic QAEs of impulse, presents holes, like each ellip-
tic QAE, which can be visualized in the space of variable coefficients. In the vicinity
of the critical hypersurface the transition can occur.
References
[Na96] Nastase, A., Stanciu, M., Purcaru, M.: The Three-Dimensional Visualization of the Qual-
itative Analysis of the Behaviors of Quadratical, Elliptical and Hyperbolical Algebraic
Equations and Its Applications to Aerodynamics, AG STAB Report, DLR–Goettingen
(1996).
[Na97] Nastase, A.: Qualitative Analysis of the Partial Differential Equations of the Three-
Dimensional Compressible Boundary Layer, via Spectral Solutions, Longman, UK
(1997).
[Na07] Nastase, A.: Computation of Supersonic Flow Over Flying Configurations, Elsevier, Ox-
ford (2007).
Artificial Neural Networks for Estimating
the Atmospheric Pollutant Sources
1 Introduction
F.F. Paes
Instituto Nacional de Pesquisas Espaciais (INPE), São José dos Campos, SP, Brazil,
e-mail: fabiana.paes@lac.inpe.br
H.F. de Campos Velho
Instituto Nacional de Pesquisas Espaciais (INPE), São José dos Campos, SP, Brazil,
e-mail: haroldo@lac.inpe.br
F.M. Ramos
Instituto Nacional de Pesquisas Espaciais (INPE), São José dos Campos, SP, Brazil,
e-mail: fernando@lac.inpe.br
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 261
DOI 10.1007/978-0-8176-8238-5_25, © Springer Science+Business Media, LLC 2011
262 F.F. Paes et al.
is tested using synthetic experimental data, where such data are obtained by run-
ning a Lagrangian atmospheric pollutant dispersion model: LAMBDA [FeAnBr95,
FeAn98a].
The Lagrangian particle model LAMBDA was developed to study the transport
process and pollutant diffusion, starting from Brownian random walk modeling
[FeAn98a, FeAn98b]. In the LAMBDA code, full-uncoupled particle movements
are assumed. Therefore, each particle trajectory can be described by the generalized
three dimensional form of the Langevin equation for velocity [Tho87]:
with
∂ φi ∂ PE ∂ (ui PE )
= − − ,
∂ ui ∂t ∂ xi (2)
φi → 0 when u → ∞
where cv = 1 for forward integration and cv = −1 for backward integration, PE =
P (x, u,t) is the non-conditional PDF of the Eurelian velocity fluctuations, and
Bi, j = 12 bi,k b j,k . Applying the backward integration, the time considered is t = −t,
and velocity U = −U, U being the mean wind speed. The horizontal PDFs are
considered as being Gaussian, and for the vertical direction the truncated Gram–
Charlier type-C of third order is employed [AnFe97].
Neural Networks Estimating Atmospheric Sources 263
The diffusion coefficients, bi j (x, u,t), for both forward and backward integration
are given by
2 1/2
σ
bi j = δi j 2 i (3)
τLi
where δi j is the Kronecker delta, σi2 and τLi are velocity variance at each component
and the Lagrangian time scale [De5co00], respectively. With the coordinates and the
mass of each particle, the concentration is computed—see (2) and (3).
The inverse problem here is to identify the source term S(t). As mentioned, a
source-receptor approach is employed for reducing the computer time instead of
running the direct model (1) for each iteration. This approach gives an explicit rela-
tion between the pollutant concentration of the ith receptor related the jth sources
NS
Ci = ∑ Mi j S j
j=1
where the matrix Mi j is the transition matrix, and matrix entry given by
⎧ V
⎨ VS, j Δt
N NR,i, j (forward)
Mi j = S, j
R,i
⎩ Δt N (backward)
NS, j S,i, j
with VR,i and VS, j being the volume for the ith receptor and jth source, respectively;
NS, j and NR,i are the number of particle realised by the jth source and ith sensor,
respectively; NR,i, j and NS,i, j are the number of particles released by the jth source
and detected by the ith receptor.
(l) (l−1)
The output y j (n) of neuron j is obtained from the inputs yi (n). The output
values of the neural network are compared with the target values.
(l)
The training algorithm is used to adjust the network weights w ji (n) and thresh-
olds, in order to minimize the difference between the output of neural network and
the target values. The use of a target value to be compared to the output value means
that the learning process is called supervised [Ha99]. In this study, a linear trans-
fer function was used for the input neurons, and the log-sigmoid function for the
neurons placed in hidden and output layers (function ϕ (.) in (4)). The error func-
tion was the sum-squared error, where the individual errors of output units on each
case were squared and summed together. The networks were trained using the back-
propagation algorithm [Bi95]. The correction of the weights Δ w was calculated ac-
cording to the following formula:
(l) (l−1)
Δ wi j (n + 1) = ηδ j (n) yi (n) + α [wi j (n) − wi j (n − 1)]
in which j is the index of the neuron in the current layer; the neuron in the upper
layer is indexed by i and its output by yi ; and the local error gradient is denoted
by δ j . There are two constants in this formula: the learning rate η , determining the
degree for changing the weights, and the momentum coefficient α , indicating the
degree of the previous adjustment is to be considered so as to prevent any sudden
changes in the direction in which corrections are made. The learning rate η and
momentum α were set to 0.1 and 0.5, respectively. They can be used to model
complex relationships between inputs and outputs or to find patterns in data.
Supervised neural networks have three stages for their application: the learning,
the activation, and the generalization steps. It is in the learning step that the weights
and bias corresponding to each connection are adjusted to some reference examples
(the input). In the activation phase, the output is obtained based on the weights and
bias computed in the learning phase [Ha99]. The transfer function selected was a hy-
perbolic tangent transfer function for the hidden layers and linear transfer function
for the output layer. The experimental data used in the learning step here were sim-
ulated by adding a random perturbation to the exact solution for forward problem
(LAMBDA):
I8 = Iexact (1 + σ μ )
where σ is the noise standard deviation and μ is a random variable taken from a
Gaussian distribution with zero mean and unitary variance. In all simulations we
used σ = 0.05 or σ = 0.10.
Neural Networks Estimating Atmospheric Sources 265
Overall, more than fifty pairs of rates of emission of pollutants and their concen-
trations are necessary for the process of inversion. Similar data sets were used for
the stages of activation and the general ANN.
4 Results
The area of the numerical experiment used is the same as that employed by Roberti
(2005) [Ro05] and Luz (2007) [Lu08]. The domain is divided into 25 sub-domains,
where each cell has the size: 300 m (width) × 200 m (length)—the volume of each
cell is 60.000 m3 (height = 1000 m) [Ro05]. Figure 1 shows the different sub-
domains of emissions of contaminants. In this figure, the positions of the sensors
in the area of study are represented by •.
Six sensors are used inside the domain, and the sensor size is a small volume:
0.1 m × 0.1 m × 0.1 m, positioned at a height of 10 meters, installed in the area,
according to Table 1.
Table 2 shows the meteorological data used by the LAMBDA model to sim-
ulate the dispersion of particles, taken from the Copenhagen experiment [Ro05,
FeAnBr95]. The meteorological data are speeds and direction averages of wind,
measured at three levels, for five different time periods measured on 19/10/1978
[Ro05].
The results obtained with noiseless data are in excellent agreement with the ex-
act model. The MLP network produced good estimation of the rate of emission of
266 F.F. Paes et al.
Table 3 Results of estimation of the rate of gas emission (gm−3 s−1 ), for noise level σ = 0.05
(ANN-A: 6:6:12:12, and ANN-B: 6:15:30:12)
Cell Exact Q-N (gm−3 s−1 ) PSO (gm−3 s−1 ) ANN-A ANN-B
(gm−3 s−1 ) (Roberti, 2005) (Luz, 2007) (gm−3 s−1 ) (gm−3 s−1 )
A2 10 9.82 09.34 09.92 09.79
A3 10 9.63 10.07 09.83 09.74
A4 10 11.26 11.26 09.86 04.81
A7 10 8.76 10.95 09.82 09.80
A8 10 11.06 10.93 09.71 09.73
A9 10 15.51 14.99 09.71 09.79
A12 20 20.12 20.79 20.94 20.61
A13 20 19.25 19.83 20.81 20.61
A14 20 11.52 13.06 20.93 20.60
A17 20 17.88 18.72 20.94 20.60
A18 20 23.82 22.76 20.96 20.61
A19 20 23.44 22.47 20.77 20.59
Table 4 Results of estimation of the rate of gas emission (gm−3 s−1 ), for noise level σ = 0.10
(ANN-A: 6:6:12:12, and ANN-B: 6:15:30:12)
Cell Exact Q-N (gm−3 s−1 ) PSO (gm−3 s−1 ) ANN-A ANN-B
(gm−3 s−1 ) (Roberti, 2005) (Luz, 2007) (gm−3 s−1 ) (gm−3 s−1 )
A2 10 8.97 09.83 10.33 10.11
A3 10 9.97 10.40 10.11 10.11
A4 10 12.52 10.79 10.12 10.20
A7 10 7.98 10.50 10.27 10.20
A8 10 10.14 12.06 10.24 10.06
A9 10 11.56 11.28 10.17 10.17
A12 20 13.84 14.56 21.21 20.97
A13 20 22.65 22.67 21.28 21.00
A14 20 14.14 15.85 21.32 20.95
A17 20 19.99 21.56 21.47 20.99
A18 20 21.17 20.05 21.47 20.99
A19 20 24.90 21.74 21.48 20.96
Neural Networks Estimating Atmospheric Sources 267
pollutants compared to the Quasi-Newton method [Ro05] and Particle Swarm Op-
timization (PSO) [Lu08]. The best results for noisy data (σ = 0.05) were obtained
for the MLP network with two layers (15 neurons for the first hidden layer, and 30
neurons for the second hidden layer).
In order to analyze the performance of the ANNs for the estimation of the rate
of gas emission, two experiments were performed. In the first experiment 5% of
white Gaussian noise (σ = 0.05) was added to the synthetic experimental data. In
the second one, 10% of noise is used to simulate the real experimental data. All
ANNs were trained with two hidden layers, varying the number of hidden neurons
and the database training.
Several tests were carried out to reach in a good neural network architecture.
After testing with different configurations (changing the number of neurons and/or
the number of hidden layers), it is shown the results for two MLP-ANN: (a) ANN-A:
two hidden layers, with 6 neurons for the first hidden layer, and 12 neurons for the
second hidden layer, (b) ANN-B: two hidden layers, with 15 neurons for the first
hidden layer, and 30 neurons for the second hidden layer. For all ANN, the number
of inputs is six, with 12 outputs (representing the emission of each sub-domain).
The ANN-B obtained the best result compared with the exact values. The topology
of the RNA is represented as (x1 : x2 : x3 : x4 ), where x1 : number of neurons in input
layer, x2 : number of neurons in the first hidden layer, x3 : number of neurons in the
second hidden layer, and x4 : number of neurons in output layer.
The training phase was carried out until a maximum number of iterations were
reached. Table 3 shows the exact results (LAMBDA), the results obtained with
regularized inversion (the optimization problem solved by quasi-Newton (Q-N,
deterministic) [Ro05] and particle swarm optimization (PSO, stochastic) [Lu08]
schemes), and the results obtained with ANN for δ = 0.05. Table 4 shows the same
results, but using σ = 0.10.
Figures 2 and 3 show the results of generalization tests in comparison with the
exact model, and the methods Q-N [Ro05] and PSO [Lu08].
5 Conclusion
The problem for identifying the emission rate of the minority gases for the system
ground-atmosphere is an important issue for the bio-geochemical cycle, and it has
being intensively investigated. This inverse problem has been previously solved us-
ing regularized solutions [Se00], Bayes estimation [En02, GiUl03], and variational
methods [El3co07] (this is approach started from the data assimilation studies).
Our first studies were initiated using generalized least square scheme, with en-
tropic regularization [Ro05, Ro3co05, Lu08, Lu3co07] (for the use of the maximum
entropy principle on this issue, see also [Bo05, DaBo07]).
ANNs were used as an effective tool for solving the inverse problem of esti-
mation of the rate of gas surface emission. The obtained reconstructions with the
ANN-MLP showed to be better than the ones obtained with regularization methods
268 F.F. Paes et al.
Fig. 2 Generalization results of rate of emission of pollutant (in gm−3 s−1 ) for noiseless data using
σ = 0.05. (a) True model; (b) Quasi-Newton [Ro05]; (c) PSO [Lu08]; and for ANN-MLP: (d) 6 and
12 neurons in the hidden layers; (e) 7 and 8 neurons hidden layers; (f) 15 and 30 neurons in the
hidden layers
Q-N [Ro05] and PSO [Lu08]. Another advantage of the use of neural network is
that, after the training phase, the reconstruction algorithm is much faster than the
regularized inversion methods. Neural inversion is a unique scheme in that it does
not need a solution of the associated forward problem.
In practice, operational inversion algorithms reduce the risk of being trapped in
local minima by starting the iterative search process from an initial guess solution
that is sufficiently close to the true profile. However, the dependence of the final
Neural Networks Estimating Atmospheric Sources 269
Fig. 3 Generalization results of rate of emission of pollutant (in gm−3 s−1 ) for noiseless data using
σ = 0.10. (a) True model; (b) Quasi-Newton [Ro05]; (c) PSO [Lu08]; and for ANN-MLP: (d) 6 and
12 neurons hidden layers; (e) 7 and 8 neurons hidden layers; (f) 15 and 30 neurons hidden layers
References
[Al74] Alifanov, O.: Solution of an inverse problem of heat conduction by iteration meth-
ods. Journal of Engineering Physics, 26, n. 11, 471–476 (1974).
[AnFe97] Anfossi, D., Ferrero, E.: Comparison among empirical probability density func-
tions of the vertical velocity in the surface layer based on higher order correlations.
Boundary-Layer Meteorol., 82, 193–218 (1997).
[Bi95] Bishop, C.M.: Neural Networks for Pattern Recognition, Oxford University Press,
Oxford (1995).
[Bo05] Bocquet, M.: Grid resolution dependence in the reconstruction of an atmospheric
tracer source. Nonlinear Processes in Geophysics, 12, 219–234 (2005).
[CaRa97] Campos Velho, H.F., Ramos, F.M.: Numerical inversion of two dimensional geolet-
ric conductivity distributions from electromagnetic ground data. Journal of Geoph-
ysis, 15, n. 2, 133–143 (1997).
[ChCa06] Chiwiaciwsky, L., Campos Velho, H.F.: Different approaches for the solution of a
backward heat conduction problem. Inverse Problems in Engineering, 11, n. 6, 471–
494 (2006).
[DaBo07] Davoine, X., Bocquet, M.: Inverse modelling-based reconstruction of the Chernobyl
source term available for long-range transport. Atmospheric Chemistry and Physics,
7, 1549–1564 (2007).
[De5co00] Degrazia, G.A., Anfossi, D., Carvalho, J.C., Mangia, C., Tirabassi, T., Campos
Velho, H.F.: Turbulence parameterization for PBL dispersion models in all stabil-
ity conditions. Atmos. Environ., 34, 3575–3583 (2000).
[El3co07] Elbern, H., Strunk, A., Schmidt, H., Talagrand, O.: Emission rate and chemical
state estimation by 4-dimensional variational inversion. Atmospheric Chemistry and
Physics, 7, 3749–3769 (2007).
[En02] Enting, I.G.: Inverse Problems in Atmospheric Constituent Transport, University
Press, Cambridge (2002).
[FeAnBr95] Ferrero, E., Anfossi, D., Brusasca, G.: Lagrangian particle model lambda: evaluation
against tracer data. International Journal of Environment and Pollution, 5, 360–374
(1995).
[FeAn98a] Ferrero, E., Anfossi, D.: Sensitivity analysis of Lagrangian stochastic models for
CBL with different pdf’s and turbulence parameterizations. Air Pollution Modelling
and its Applications, XII, 673–680 (1998a).
[FeAn98b] Ferrero, E., Anfossei, D.: Comparison of pdfs, closures schemes and turbulence pa-
rameterizations in Lagrangian stochastic models. International Journal of Environ-
ment and Pollution, 9, 384–410 (1998b).
[GaDo99] Gardner, M.W., Dorling, S.R.: Neural network modeling of hourly NOx and NO2
concentrations in urban air in London. Atmospheric Environment, 33, 709–719
(1999).
[GiUl03] Gimson, N.R., Uliasz, M.: The determination of agricultural methane emissions
in New Zealand using inverse modeling techniques. Atmospheric Environment, 37,
3903–3912 (2003).
[Ha99] Haykin, S.: Neural Networks: A Comprehensive Foundation, [S.l.], Prentice Hall
(1999).
[HiGm96] Hidalgo, H., Gmez-Trevin, E.: Application of constructive learning algorithms to the
inverse problem. IEEE T. Geosci. Remote, 34, n. 1, 874–885, (1996).
[IPCC07] IPCC. Climate Change 2007: The Physical Science Basis. Contribution of Work-
ing Group I to the Fourth Assessment Report of the Intergovernamental Panel on
Climate Change, [S.I.], Cambridge University Press (2007).
[Ku10co03] Kukkonen, J., et al.: Extensive evaluation of neural network models for the predic-
tion of NO2 and PM10 concentrations, compared with a deterministic modeling sys-
tem and measurements in central Helsinki. Atmos. Environ., 37, 4539–4550 (2003).
Neural Networks Estimating Atmospheric Sources 271
[LuHsCh06] Lu, H.C., Hsieh, J.C., Chang, T.S.: Prediction of daily maximum ozone concentra-
tions from meteorological conditions using a two-stage neural network. Atmos. Res.,
81, 124–139 (2006).
[Lu08] Luz, E.F.P.: Estimação de fonte de poluição atmosférica por enxame de partículas.
M.Sc. Thesis, Instituto Nacional de Pesquisas Espaciais (2008).
[Lu3co07] Luz, E.F.P., Campos Velho, H.F., Becceneri, J.C., Roberti, D.R.: Estimating atmo-
spheric area source strength through particle swarm optimization, in: Inverse Prob-
lems, Desing and Optimization Symposium IPDO (2007), April 16–18, Miami (FL),
USA.
[PeTi06] Pelliccioni, A., Tirabassi, T.: Air dispersion model and neural network: a new per-
spective for integrated models in the simulation of complex situations. Environmen-
tal Modelling & Software, 21, 539–546 (2006).
[Ro05] Roberti, D.R.: Problemas inversos em física da atmosfera, D.Sc. Thesis, Federal
University of Santa Maria (Centro de Ciencias Naturais e Exatas) (2005).
[Ro3co05] Roberti, D.R., Anfossi, D., Campos Velho, H.F., Degrazia, G.A.: Estimation of lo-
cation and strength of the pollutant sources. Ciencia e Natura, 131–134 (2005).
[Ro3co07] Roberti, D.R., Anfossi, D., Campos Velho, H.F., Degrazia, G.A.: Estimation of emis-
sion rate from experimental data. Nuovo Cimento della Società Italiana di Fisica. C,
Geophysics and Space Physics, 30, 177–186 (2007).
[Se00] Seibert, P.: Inverse modeling of sulfur emissions in Europe based on trajectories
inverse Methods in Global Biogeochemical Cycles, American Geophysical Union,
147–154 (2000).
[Se01] Seibert, P.: Inverse modelling Lagrangian particle dispersion model: application to
point releases over limited time intervals. Air Pollution Modeling and its Applica-
tion, XIV, 381–389 (2001).
[ShCaSi08] Shiguemori, E.H., Campos Velho, H.F., Silva, J.D.S.: Atmospheric temperature re-
trieval from satellite data: New non-extensive artificial neural network approach, in:
Proceedings of the 23rd Annual ACM Symposium on Applied Computing (Sympo-
sium on Applied Computing, Fortaleza), Association for Computing Machinery Inc,
III, New York, 1688–1692 (2008).
[Ti77] Tikhonov, A.N., Arsenin, V.Y.: Solutions of Ill-Posed Problems, Winston & Sons,
New York (1977).
[Tho87] Thomson, D.J.: Criteria for selection of stochastic models of particle trajectories in
turbulent flows. Journal of Fluid Mechanics, 180, 529–556 (1987).
[WeSuHa06] Wesolowski, M., Suchacz, B., Halkiewicz, J.: The analysis of seasonal air pollution
pattern with application of neural networks. Analytical and Bioanalytical Chemistry,
384, n. 2, 458–467 (2006).
[Wo00] Woodbury, K.: Neural networks and genetic algorithms in the solution of inverse
problems. Bulletin of the Brazilian Society for Computational Applied Mathematics
(SBMAC), (2000).
A Theoretical Study of the Stratified
Atmospheric Boundary Layer Through
Perturbation Techniques
1 Introduction
The structure of the stratified atmospheric boundary layer (ABL) over a flat uniform
terrain is fairly well understood, but many known results remain mathematically
unjustified. A number of factors contribute to this situation, the diversity of physi-
cal processes acting on the flow probably being the most important one. Much has
been accomplished in the past in the fields of meteorology and fluid mechanics by
the use of perturbation techniques. In this way, the atmospheric flow over small to-
pographic features was addressed by [Tow72]. The authors of [Kni77] studied the
turbulent flow over wavy surfaces and in [Jac75] a two-layer structure for the ABL
over a gentle hill was proposed. Another work [Syk80] used the asymptotic expan-
sion method to study the influence of small terrain elevations on the main flow and
on turbulence. The same idea was improved in [Hun88a, Hun88b] by introducing
different upwind profiles in the analysis and considering a stably stratified atmo-
sphere. In both articles a three layer structure was considered that shaped our un-
derstanding of the ABL up to present days. Further progress was made in [Wal92]
where matching techniques were used to propose a new ABL resistance law for
neutrally-stratified flow and in [Bau03] a parametrization of the effective roughness
length over flat terrain in neutral atmosphere was obtained. The authors of [Pel05]
introduced a modified logarithmic law for flows over gentle hills.
C.C. Pellegrini
Universidade Federal do São João del-Rei, MG, Brazil,
e-mail: pelle@ufsj.edu.br
M.T. Vilhena
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: vilhena@pq.cnpq.br
B.E.J. Bodmann
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: bardo.bodmann@ufrgs.br
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 273
DOI 10.1007/978-0-8176-8238-5_26, © Springer Science+Business Media, LLC 2011
274 C.C. Pellegrini et al.
the conservation of mass, momentum and energy for the ABL for two-dimensional
flows are taken from [Ary01] with the variables, u, w the mean field values for hori-
zontal and vertical velocity, u , w its fluctuations, (p0 , ρ0 , θ0 ) the state of hydrostatic
equilibrium, (p1 , ρ1 , θ1 ) its deviation from hydrostatic equilibrium and αh = ρkc p is
the thermal molecular diffusivity. The Coriolis force, the adiabatic heat and the ra-
diation divergence have been neglected, as usual in ABL studies. Upon substituting
the variables, X = Lxx , Z = Lzz , τ = ttc , U = Uug , W = Wwc , U = uu∗ , W = ww∗ , P1 = ρ pU1 2 ,
0 g
Θ = θθ , Θ = θθ∗ the original equations of [Ary01] are rendered dimensionless. Here
Lx is the horizontal micro-meteorological disturbance scale, Lz the ABL depth, Ug
1 1
the geostrophic velocity, u∗ = (u w )s2 , w∗ = ( θ wθ gzi )s3 the friction velocity scales,
θ∗us = wwθ∗ , θ∗st = wuθ∗ the friction temperature for unstable and stable ABL and Wc ,
tc is characteristic velocity and time, resp. The subscript s refers to values at the
surface and zi is the ABL depth.
∂U ∂U
∂X + ∂∂WZ = 0 and ∂X + ∂∂WZ = 0, (1)
∂ U U
εt ∂∂Uτ +U ∂∂ UX +W ∂∂UZ = − ∂∂PX1 − ε∗2 + ∂ U∂ ZW + εεR2 εL2 ∂∂ XU2 + ∂∂ ZU2 ,
2 2
∂X L
(2)
∂ W U ∂ W W
εL2 εt ∂∂Wτ +U ∂∂W
X +W ∂W
∂Z = − ∂∂PZ1 − εL2 ε∗2 + ∂X ∂Z
+εR εL2 ∂∂ XW2 + ∂∂ ZW2 + Ribs ,
2 2
(3)
εt ∂Θ ∂Θ ∂Θ
= −ε∗θ ε∗ ∂ U∂ XΘ + ∂ W∂ ZΘ + εεP2 εL2 ∂∂ XΘ2 + ∂∂ ZΘ2 .
2 2
∂ τ +U ∂ X +W ∂ Z (4)
L
the next section we propose a way to systematize the calculations and perform the
matching using IVT.
Varying ε in [0, 1] in the mass conservation equation, (1), and recalling that the
shallow continuity form of the mass conservation holds, shows that both terms in (1)
must be of the same order of magnitude. This conclusion applied to (2)–(4) shows
that their advective terms are always of the same order of magnitude. Further, in
order to maintain the mass equation free of contradiction we assume that Wc → 0 as
ε → 0.
For the analysis of the x-momentum equation a relation between ε∗ εR and εL is
needed. For the neutral ABL, where εL = Os(1) [Pel05] showed that εR ε∗4 , i.e.
the relation shall hold for large enough Re. Here we extend this idea to the non-
neutral ABL with the typical values [Hol92, Par93, Stu97] shown in Table 1 and
2
values for Ug = 10 ms , Lx = 1000 m and ν = 1.5 × 10−5 ms .
In Table 1, the stable regime was divided into weakly, moderately and strongly
stable stratifications. There is, however, some uncertainty in the typical values pro-
posed for the weakly and moderately stable regimes and the distinction between
them is not rigid. They were established by supposing monotone behavior of the
variables between the neutral and strongly stable regimes. The leading order terms
of (2) are Os (1), Os (ε∗2 ε −1 ) and Os (εR εL−2 ε −2 ), considering that εt = o(1) is van-
ishing small. The values of ε that make those terms of the same order of mag-
nitude define the scaling where they change dominance. Thus the three possi-
ble combinations of these terms shall be investigated, i.e. Os (1) with Os (ε∗2 ε −1 ),
Os (1) with Os (εR ε −2 εL−2 ) and Os (ε∗−2 ε −1 ) with Os (εR ε −2 εL−2 ). However, not all
of these combinations lead to a useful scaling. For example, comparing Os (1) with
Os (εR ε −2 εL−2 ) leads to a balance between advective and viscous terms, but when
such balance is achieved, turbulence is the dominating effect. This can be antici-
pated observing that viscous terms are proportional to ε −2 , turbulent terms to ε −1
and advective terms to ε 0 . Therefore, in the limit ε → 0 the advective term domi-
nates over the turbulent and, finally, the viscous one. Hence the significant heights
A Theoretical Study of the Stratified Atmospheric Boundary Layer 277
may be found comparing only the first and third combination, respectively, resulting
in the limits ε1 = Os (ε∗2 ) and ε2 = Os (εR ε∗−2 εL−2 ).
Table 1 shows for the neutral case (where εL = Os (1)) that ε1 ε2 because ε∗4
εR , confirming the dominance of terms in the ABL proposed by [Pel05], which is
also the case for the unstable ABL. For the stable ABL, however, we may have either
of the three relations ε1 ε2 , ε1 = Os (ε2 ) or ε1 ε2 . In those cases the analysis
shall consider possibility that defines one stability sub-range. Here we define the
weakly stable case by ε1 ε2 and thus ε∗4 εL2 εR . The moderately stable case is
defined by ε1 = Os (ε2 ) or ε∗4 εL2 = Os (εR ) and the strongly stable case by ε1 ε2 or
ε∗4 εL2 εR .
Upon varying ε in the interval [0, 1] in (2) yields a set of non-dimensional
equations for each flow regime. The results are presented below in dimensional
form (see Tables 2, 3, 4). With the adopted values of Ug and Lx the value for
εt = LxUg−1tc−1 is Os (1) only if tc = Os (100 s), which is too short a period for most
micro-meteorological applications. This result indicates that the storage term must
be considered only in the transition period of sunrise and, possibly, sunset.
Table 2 Unstable, neutral & weakly stable ABL (εR = o(ε∗4 εL2 ))
Layer Condition Equation
∂ p1
Supra-outer ε1 ε ≤ 1 u ∂∂ ux + w ∂∂ uz = − ρ10 ∂x
1 ∂ p1
Outer ε = Os (ε1 ) u ∂∂ ux
+ w ∂∂ uz = − ρ0 ∂ x − ∂ u∂ zw
Overlap ε2 ε ε1 0 = ∂ u∂ zw
0 = − ∂ u∂ zw + ν ∂∂ z2u
2
Inner ε = Os (ε2 )
0 = ∂∂ z2u
2
Sub-inner ε ε2
Typical values for the variables involved in the evaluation of the z-momentum
equation are shown in Table 5. Again, there is some uncertainty in the values
adopted in the intermediate cases. Virtual temperatures were θ0 = 300 K and
θ1 = −1 K for the most stable case and θ1 = 2 K for the most unstable.
The scaling where terms change dominance can be obtained comparing O(εL2 )
with Os (ε −1), Os (εL2 ε∗2 ε −1 ), Os (εR ε −2 ) and Os (Ribs ). Again, not all possible com-
binations need to be considered. The turbulence term is always smaller than the
pressure gradient term because εL2 ε∗2 1 even for the most unstable case. Also, ad-
vective or buoyancy terms (which are independent of ε ) never dominate together
with viscous terms. The comparisons to be considered are
ε1 = Os (εR ), ε2 = Os (Ri−1
bs ), ε3 = Os (εL−2 ).
Comparing the magnitudes of the leading order terms in (4), i.e., Os (1) with
Os (ε∗θ ε∗ ε −1 ) and Os (εP ε −2 εL−2 ) yields
Our principal results are illustrated in Figs. 1, 2 and 3. Recalling that we used a
variety of related parameters that in some way hide some geometrical properties
of the ABL. For example, the rescaled height Z̃ = Zε = Os (1) with Z = Lzz implies
z = Os (Lz ) because of ε = Os (1). Thus z appears in all figures as a horizontal line,
but in fact represents the stratification dependent ABL height. All other limits have
to be interpreted accordingly.
For the x-momentum equation Fig. 1 shows two regions for the ABL under un-
stable up to moderately stable conditions, where the inner region is governed by
the vertical flux of turbulent and viscous x-momentum, and an outer region, driven
by vertical flux of turbulent x-momentum, by a pressure gradient and by advection.
Between the two regions there is an overlap domain where the turbulent flux domi-
nates. The innermost layer has a flow that is governed by viscous forces and above
the outer layer advection and pressure forces dominate. This five-region description
is similar to the classical picture as presented in [Mah99]. Our findings differ from
those due to the fact that in the approach of [Mah99] the layer structure is defined
in terms of similarities.
The overlap region where the turbulent flux dominates yields the logarithmic law
if a mixing length turbulence closure is used. The validity of this domain does not
extend to the surface and if the equation for this region is to be integrated, knowledge
of an adequate boundary condition at some height from the surface is required. As
this condition is generally not known, the solution is generally assumed to be true
A Theoretical Study of the Stratified Atmospheric Boundary Layer 281
Fig. 2 Dominating terms on the z-momentum equation. Terms as in Fig. 1 and B = buoyancy
within the sub-inner layer leading to the appearance of the roughness length z0 in
the resulting logarithmic law due to the use of ū(z0 ) = 0.
The inner layer also offers an explanation for the traditional hypothesis of con-
stant stress near the surface. Upon multiplication of the inner layer equation by ρ
results in a relation for the turbulent and viscous stresses (τt and τl ), ∂ (τ∂t +z τl ) = 0.
Integrating this equation to the surface shows that the total shear stress is indeed
constant and equal to ρ u∗2 in first order approximation. This result implies that the
most accepted definition for the surface layer is identical to our definition for the
inner layer, but a discrepancy appears concerning the region where the logarithmic
law is valid.
Our equation for the sub-inner layer shows that the flow there is dominated by
viscous forces as expected. Because turbulence is a second order influence in this
282 C.C. Pellegrini et al.
region, it is often referred to as the laminar sub-layer [Sch93]. A solution for its
equation yields a velocity profile varying linearly with height. However, we believe
that this layer is only rarely observed in nature. Its simple structure is probably de-
stroyed by the relatively large surface roughness and by the movement of leaves
and branches induced by the wind at the top of canopies. For our analysis to hold,
the characteristic length Lx had to be modified to account for the detailed geome-
try of the roughness elements and therefore x-derivatives would no longer appear
as second order corrections in the analysis. This reasoning permits us to identify
our sub-inner layer with the roughness sub-layer in [Mah99], where the flow varies
spatially with the scale of the roughness elements. In order to observe the linear
wind profile, one must resort to wind tunnel experiments over dynamically flat sur-
faces. Our analysis suggests that the linear profile may be observed in nature above
boundaries by water, bare soil or rock.
Above the outer layer, there is a region where pressure gradients and advection
dominate. A solution for this layer obviously yields the potential flow solution found
also by the authors of [Syk80] using first order approximation in their asymptotic
expansions. This layer can be thought of as an overlap region for the layer outside
the ABL where advection, pressure gradient and Coriolis effects govern the flow,
yielding the gradient wind balance.
A different picture appears for strongly stable atmospheric stratification. In this
case just one layer can be identified in the structure, with pressure gradient, advec-
tion and viscous forces in balance. Above and below this layer, the potential flow and
viscous dominating layers of the moderately stable ABL appear again. The fact that
the turbulent flux never appears in our first order approximation indicates that this is
the regime where turbulence is weak, intermittent or totally absent. Also, as the only
constant stress layer is the very thin viscous layer, surface stress is not expected to
be an important scaling parameter for most of the ABL and, thus, Monin–Obukhov
similarity is not expected to hold. This conclusion is independent of the choice of
the turbulence closure and agrees with interpretations pointed out in [Mah99]. For
the limiting case of moderately stable stratification, a vanishingly thin intermediate
layer appears, with advection, pressure, turbulence and viscous forces in approxi-
mate balance, limited from above by the potential flow layer and from below by the
viscous layer.
The x-momentum structure detailed above suggests a shape of the velocity profile
in some regimes. In weakly stable up to strongly unstable situations, the dominant
viscous forces at the sub-inner layer convert kinetic energy into heat thus generating
sub-geostrophic winds there. As viscosity is the dominating mechanism, velocity
decays to zero at the surface. The overlap region above transports this destruction of
x-momentum upwards via turbulent vertical flux maintaining the conversion of ki-
netic energy into heat, which keeps the velocity sub-geostrophic. At the supra-outer
layer, the influence of turbulence finally turns into a second order effect allowing a
geostrophic balance to be established. In case of the existence of a horizontal tem-
perature gradient, the geostrophic wind may increase with decreasing height, but
this is generally not sufficient to generate a maximum in the velocity profile at some
A Theoretical Study of the Stratified Atmospheric Boundary Layer 283
distance from the ground, as turbulence mixing is very efficient in vertical momen-
tum destruction.
In the strongly stable regime turbulence disappears from the intermediate layer,
and that tends to confine the momentum destruction to the sub-intermediate region.
In this case, a sub-geostrophic low level jet may form above the intermediate layer.
In fact, any reason that causes the horizontal velocity to increase with decreasing
height (see [Stu97]) may cause a low level jet, once the mixing effect of turbulence
is not dominant in this regime in the intermediate layer. Super-geostrophic jets are
also suggested by our results once some of its known causes (such as, for example,
inertial oscillation and a terrain with slope) strongly suppress turbulence for highly
stable cases.
Typical values of z1 and z2 in Table 1 were obtained from ε1 and ε2 using Z̃ = Zε =
OS (1) and Z = Lzz . Note that the numerical values of z1 and z2 are evidently too low,
which we associate with Ug which may not be the most adequate scaling for U in a
general sense. For the outer layer, u = OS (Ug ) but nearer to the surface u = o(Ug ),
indicating that a local velocity should be used instead. The calculated values for
z1 and z2 may be regarded as associated values to the real heights of the layers, in
the same sense that z0 is associated to the mean height of the roughness elements.
Following that reasoning, our calculations confirm that the viscous layer is indeed
very thin and the larger part of the ABL is dominated by potential flow. An analysis
of the relation between stability conditions and the related z1 and z2 values also
show that the region dominated by viscous effects gets thinner as stability increases,
as expected. This explains why the influence of individual roughness elements gets
more pronounced as stability increases, as was found in [Mah99].
Analysis of the z-momentum equation shows that a two-layered structure exists
only in the most unstable cases. From the stable to the moderately unstable regimes,
only one layer exists that is governed by pressure gradients and viscous forces. This
region is limited from above and from below by layers where only pressure forces
and viscous forces dominate, respectively. Even considering that the heights z1 , z2
and z3 (see Table 5) do not correspond to the real heights of the layers as pointed out
before, the very low value of z1 indicates that most of the ABL is governed by pres-
sure gradients, a conclusion dating back to Prandtl’s original studies of boundary
layers. This result implies that the horizontal variation of the geostrophic pressure is
communicated almost unchanged to the surface. The existence of the viscous sub-
inner layer implies that w(z) as well as u(z) are also linear very close to the surface.
In the strongly unstable regimes, the outer layer of the z-momentum equation is
governed by advection, pressure and buoyancy and the inner layer by pressure and
viscosity. The values of z2 and z3 suggest that a considerable part of the ABL is still
governed by pressure gradients alone in this regime, whereas the sub-inner layer is
dominated by viscous forces alone. The energy equation shows a behavior similar
to the x-momentum equation. A two-layered structure appears in the neutral and un-
stable regimes, with an outer layer governed by advection and turbulent vertical flux
of heat, and an inner layer governed by diffusion and turbulent vertical flux of heat.
Between these two, an overlap region, which is predominantly turbulent, appears
where the logarithmic temperature profile holds for a mixing length turbulence clo-
284 C.C. Pellegrini et al.
sure. The supra-outer layer is dominated by pressure gradient and advection and the
sub-inner layer by molecular diffusion of heat alone. In the latter region, velocities
are so small that air behaves practically as a static fluid, yielding a linear temperature
profile.
For moderately to strongly stratified regimes, a one-layered structure appears,
with advection and diffusion dominating the energy equation in first order approx-
imation. Above and below that layer, only advection and only diffusion, respec-
tively, characterize these layers. Turbulence is absent in the whole vertical extent of
the ABL in the first order approximation. For the limiting moderately stable case,
a spurious intermediate layer is seen, with balance between advection, turbulence
and diffusion. This layer is limited from above and from below by layers where ad-
vection and diffusion dominates. The shape of the temperature profile is also well
accounted for by the structure proposed here, as is the case of the u velocity profile.
4 Conclusions
In this contribution we studied the stratified Atmospheric Boundary Layer over flat,
vegetated surfaces, by means of perturbation techniques. More specifically, we re-
sorted to the Intermediate Variable Technique. In agreement with other approaches
known from the literature this procedure results in a layered structure with features
in agreement with phenomenology. Moreover, the present analysis showed up new
properties of the ABL as for instance the existence of a dividing condition (the mod-
erately stable stratification) where turbulence is found to be no longer a first order
magnitude effect. Furthermore the ABL is divided in an outer region dominated by
advection and pressure, and an inner domain dominated by viscosity.
The equations obtained from first order perturbation theory that describe the over-
lap layers coincide with those obtained from matched asymptotic expansions and
thus may be understood as a validation of the approach. Consequently a uniformly
valid solution for the ABL shall consider the overlap layers that emerged from our
perturbation scheme. It is noteworthy that the IVT is less restricted than matched
asymptotic expansion technique, where the focus is put only on the overlap domains
of the problem. Nevertheless, we are aware of the fact that we have excluded from
our considerations cases where dominant contributions could come from higher or-
der terms in the perturbation expansion. In this sense our discussion is not complete
which we hope to extend at later time. Validity of our consideration is limited to the
cases where the first order approximation does contain the essential physics of the
problem and where higher order corrections do not qualitatively alter our findings.
For weakly stable to strongly unstable stratification we conclude: The ABL
presents a two-layered structure, both for x-momentum and heat. Near the sur-
face, molecular and turbulent effects dominate. Higher up in the ABL, turbulence,
horizontal pressure gradients (in the x-momentum equation) and advection are the
principal effects. The definition of the stratification regime varies depending on the
equation in consideration. In the z-momentum equation, strongly unstable means
A Theoretical Study of the Stratified Atmospheric Boundary Layer 285
εL2 = Os (Ri−1
bs ), while for the energy conservation equation ε∗θ ε∗ εL εP holds. The
2 2 2
most accepted definition for the surface layer coincides with our inner layer, which
is a constant stress layer. The analysis also showed a constant heat flux layer to exist.
According to our analysis the layers are not in the same regions where logarithmic
laws for velocity and temperature are valid. Furthermore, the analysis gives mathe-
matical support for the dominance of pressure effects in the z-momentum equation.
An additional finding is the existence of a region very close to the surface where
molecular effects are dominant.
For strongly stable stratification we find that the ABL presents a one-layered
structure, both for x-momentum and heat. Concerning momentum, there is a balance
between pressure, advection and viscous effects. The energy equation is character-
ized by a balance between advection and diffusion. For the case where the surface
layer is a constant stress layer, our analysis shows that such a layer shall exist only
very close to the surface (the viscous layer). Turbulent fluxes do not appear in our
first order approximations, indicating that turbulence is weak and appears only as a
second order correction or effectively suppressed. Turbulence might also appear in-
termittent, however, this case has to be excluded from our analysis, since the adopted
time scale is inappropriate. Monin–Obukov similarity does not hold in the expected
range of height but is valid only very close to the surface.
As already mentioned a number of specific situations were not taken into con-
sideration here, but will be accounted for in a future work. Besides higher order
contributions which may turn significant also the time scale used will be changed in
order to probe fast-changing scenarios as for instance the sunset period.
References
[Ary01] Arya, S.P.: Introduction to Micro-meteorology, 2nd ed., Academic Press, 420 pp.
(2001).
[Bau03] Baudauf, M., Fiedler, F.: A parametrization of the effective roughness length over in-
homogeneous flat terrain. Bound.-Layer Meteor., 106, 189–216 (2003).
[Hin95] Hinch, E.J.: Perturbation Methods, Cambridge Univ. Press, Cambridge, 160 pp.
(1995).
[Hol92] Holton, J.R.: An Introduction to Dynamic Meteorology, Academic Press, London, 511
pp. (1992).
[Hun88a] Hunt, J.C.R., Leibovich, S., Richards, K.J.: The turbulent shear flow over low hills.
Quart. J. Roy. Meteor. Soc., 114, 1435–1470 (1988).
[Hun88b] Hunt, J.C.R., Richards, K.J., Brighton, P.W.M.: Stably stratified shear flow over low
hills. Quart. J. Roy. Meteor. Soc., 114, 859–886 (1988).
[Jac75] Jackson, P.S., Hunt, J.C.R.: Turbulent wind flow over a low hill. Quart. J. Roy. Meteor.
Soc., 106, 929–955 (1975).
[Kap67] Kaplun, S.: Fluid Mechanics and Singular Perturbations, Academic Press, San Diego,
366 pp. (1967).
[Kni77] Knight, D.: Turbulent flow over a wavy boundary. Bound.-Layer Meteor., 11, 205–222
(1977).
[Lag72] Lagerstron, P.A., Casten, R.G.: Basic concepts underlying singular perturbation tech-
niques. SIAM, 14, 63–120 (1972).
286 C.C. Pellegrini et al.
[Mah99] Mahrt, L.: Stratified atmospheric boundary layer. Bound.-Layer Meteor., 90, 375–396
(1999).
[Par93] Parlange, M.B., Brütsaert, W.: Regional shear stress of broken forest from radiosonde
wind profiles in the unstable surface layer. Bound.-Layer Meteor., 64, 355–368 (1993).
[Pel05] Pellegrini, C.C., Bodstein, G.C.R.: A modified logarithmic law for neutrally stratified
flow over low-sloped hills. J. App. Meteor., 44, n. 6, 900–916 (2005).
[Pie02] Pielke, R.A.: Mesoscale Meteorological Modeling, 2nd ed., Academic Press, London,
675 pp. (2002).
[Sch93] Schetz, J.A.: Boundary Layer Analysis, Ed. Prentice Hall, New Jersey, 586 pp. (1993).
[Stu97] Stull, R.B.: An Introduction to Boundary Layer Meteorology, 6th ed., Kluwer Ac. Publ.,
Dordrecht, 670 pp. (1997).
[Syk80] Sykes, R.I.: An asymptotic theory of incompressible turbulent boundary layer flow over
a small hump. J. Fluid. Mech., 101, 647–670 (1980).
[Tow72] Towsend, A.A.: Flow in a deep turbulent boundary layer over a surface distorted by
water waves. J. Fluid Mech., 55, 719–735 (1972).
[Wal92] Walmsley, J.L.: Proposal for new PBL resistance laws for neutrally-stratified flow.
Bound.-Layer Meteor., 60, 271–306 (1992).
Integro-Differential Equations for Stress
Analysis in the Bridged Zone of Interface Cracks
M. Perelmuter
1 Basic Equations
The plane elasticity problem for two semi-infinite elastic plates jointed together and
with a straight crack (|x| ≤ ) placed on the material interface (y = 0) is considered.
It is assumed that the uniform normal (σ0 ) and shear (τ0 ) loading are applied at
infinity of these half-planes in directions normal and parallel to a crack line, respec-
tively. The surfaces of a crack are connected by bonds or fibers in some regions
starting from the crack tips, − d ≤ |x| ≤ , (bridged zones), see Fig. 1. The sizes
of these zones can be comparable to the whole length of the crack. The spring-like
quasilinear bonds deformation law are assumed as a simple mathematical model
of the crack surfaces interaction at the bridged zones. Denote as Q(x) the tractions
arising in the bonds at the crack bridged zone due to the external loading:
where qy (x) and qx (x) are the normal and shear components of the bond tractions.
Due to the superposition principle the following boundary condition on the surfaces
of the crack |x| < are considered (qy (x) = qx (x) = 0 for |x| < − d):
u(x) = uy (x) − iux (x) = cy (x)qy (x) − icx (x)qx (x) (2)
where uy , ux are the projections of the crack opening on the coordinate axes and the
functions cy , cx could be considered as the effective bonds compliance in the direc-
M. Perelmuter
A. Ishlinsky Institute for Problems in Mechanics, Russian Academy of Science, Moscow, Russia,
e-mail: perelm@ipmnet.ru
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 287
DOI 10.1007/978-0-8176-8238-5_27, © Springer Science+Business Media, LLC 2011
288 M. Perelmuter
where u∞ and uq are the crack opening caused by the external loads (σ0 or τ0 )
and by the bonds tractions Q (x) respectively. The actions of the external normal
load and external shear load are considered separately as two independent problems.
According to [Sl81] the relation for u∞ can be written as
−iβ
−x
u∞ (x) = u∞,y (x) − iu∞,x (x) = 2A α (2 − x2 ) (σo − iτo ) (5)
+x
where
μ 1 + μ 2 κ1 ln α 0.25 k1 + 1 k2 + 1
α= , β= , A= + (6)
μ 2 + μ 1 κ2 2π (1 + α ) μ1 μ2
and κ1,2 = 3 − 4ν1,2 or κ1,2 = (3 − ν1,2 )/(1 + ν1,2 ) for the plane strain and the plane
stress, respectively, ν1,2 and μ1,2 are the Poisson ratios and shear modulus of the
joint materials 1 and 2. For piece-wise media the convenient way to analyze the
crack opening and bonds tractions in the crack bridged zone is the transformation
of the relation (4) into differential equations (see [GoPe99])
H ∂ ∂ uq (s) ∂ u∞ (s)
[ φ1 (s)qy (s) − iφ2 (s)qx (s)] − Eb = Eb . (8)
∂s ∂s ∂s
The right side of this relation can be obtained from the relation (5), hence, it is the
given function of the coordinate, material parameters and the external loads. The
∂ u (s)
derivative of the crack opening caused by the bond tractions ( ∂qs ) was obtained in
[GoPe99]. Assuming that the external loading has a strong influence on the solution
of (8) and taking into account (2) and (5) we will search for the tractions Q (s) in
the following form:
Integro-Differential Equations for Stress Analysis 289
−iβ
1−s
qy (s) − iqx (s) = (py (s) − ipx (s)) , (9)
1+s
where qy (s), py (s) are even functions and qx (s), px (s) are odd ones for the normal
loading (problem 1) and vice versa for the shear loading (problem 2).
The derivative of the crack opening due to tractions applied to the interface
crack surfaces at the bridged zone is derived for tension loading in [GoPe99] (for-
mula (15)) with the accounting of tractions’ symmetry and the statement (9). After
the similar modification for the shear loading case it can be written as
∂ uq (s)
= A(1 − α )(qx (s) + iqy (s) )
∂s
1
2A(1 + α ) 1 − s −iβ
− √ (R1 (s,t) py (t) − iR2 (s,t) px (t))dt. (10)
π 1 − s2 1 + s
1− d
The kernels R1,2 (s,t) of the relation (10) depend on the load type: for problem 1
where √
1 − t2
R(s,t) = 2 2 .
t −s
Substituting the expression (10) to (8) one can obtain the system of two singular
integral-differential equations relative to the bond tractions qy (s) and qx (s):
∂ πε (1 − α )
[ φ1 (s)qy (s) − i φ2 (s )qx (s)] − (qx (s) + iqy (s))
∂s 2(1 + α )
1
ε 1 − s −iβ
+√ [R1 (s,t)py (t) − iR2 (s,t)px (t)] dt
1 − s2 1 + s
1−d/
√
π ε R∞ α 1 − s −iβ
= √ (2 β i − s) (13)
(1 + α ) 1 − s2 1+s
where
290 M. Perelmuter
B
R∞ = Σ σo2 + τo2 , Σ = σo (for problem 1), Σ = −iτo (for problem 2).
Equation (13) is the system of two integral-differential equations of the second order
Fredholm type with singular Cauchy-type kernels. This system with the boundary
condition qx,y (1) = 0 can be solved only numerically. The parameter ε in (13) char-
acterizes the relative stiffness of the bond at the bridged zone of the crack
Eb k1 + 1 k2 + 1 H
ε= + , co = . (14)
2π co μ1 μ2
It should be noted that the similar parameter for a homogeneous media was in-
troduced in [En68]. The parameter co in (14) for given mechanical properties of
materials and bonds can be considered as the relative compliance of the bond at the
bridged zone of the crack.
2 Numerical Algorithm
For the numerical solution of (13) a piece-wise quadratic approximation of the un-
known bond tractions is used. This approximation in conjunction with the partition
scheme similar to one-dimensional finite elements is enabled to reduce the system
of the integro-differential equations (SIDE) (13) to a system of the linear algebraic
equations.
First, the SIDE (13) is √transformed taking into account the asymptotic behavior
the
√ crack opening u (s) ∼ 1 − s for s → 1, [HuMeRi87]. Therefore, the multiplier
1 − s in the representation of the unknown functions px (s), py (s) is extracted as
follows ( f1,2 are new unknown functions):
√
py,x (s) = f1,2 (s) 1 − s. (15)
1
d f j (s)
Ti j (s) +Wi j (s) f j (s) + ε Gi j (s,t) f j (t)dt = Zi (s), i, j = 1, 2 (17)
ds
1−d/
The functions Ti j (s), Wi j (s), Gi j (s) and Zi (s) are given in [GoPe99] for the ten-
sion loading. For the shear loading the function Gi j (s) and Zi (s) are modified ac-
cording to the relations (11) and (12).
Sampling in (17) is performed by the following nodes at the bridged zone:
sn = λ + h(n − 1), n = 1, 2, . . . , M,
1
∂ uq (s)
u(s) = u∞ (s) + ds; (20)
∂s
x0 /
where the derivative of the crack opening due to tractions applied to the crack sur-
faces is defined as (10).
292 M. Perelmuter
Having the solution of system (17) the stress intensity factors (SIF) KI , KII can be
calculated (see [HuMeRi87]) as follows:
√
KI + iKII = lim 2πδ (σyy (δ ) + iσxy (δ )) δ −iβ , (21)
δ →0
where σyy (δ ), σxy (δ ) are the stresses ahead of the crack tip caused by the external
loads and bond stresses and δ = x − is the distance to the crack tip, x > .
Denote by σyy ext , σ ext and σ int , σ int the stresses caused ahead the crack tip by the
xy yy xy
external loads, σo or τo , and bond tractions, respectively. Then
iβ
cosh(πβ ) x−
σyy
int
(x) + iσxy
int
(x) = √
π x 2 − 2 x+
+ < 2
− ξ 2 + ξ iβ
× (qy (ξ ) + iqx (ξ ))d ξ . (23)
ξ −x −ξ
−
By incorporating the formula (23) and using the relations (9), (21) and (22) the
following formulas can be obtained:
√ 1
!
σo π 2 cosh(πβ ) (py (t) + it px (t))
KI + iKII = (1 + 2iβ ) − √ dt ,
(2)iβ π 1 − t2
1−d/
√ 1
!
τo π 2 cosh(πβ ) (t py (t) + ipx (t))
KI + iKII = ( i − 2β ) − √ dt
(2)iβ π 1 − t2
1−d/
Integro-Differential Equations for Stress Analysis 293
for problems 1 and 2, respectively. The bonding in the crack bridged zone reduces
the SIF. As a measure of that reduction we will use the relative value of the SIF
modulus
Kr (d/) = Ko / (KIext )2 + (KIIext )2 ,
(24)
Ko = KI2 + KII2 .
The field of the total stresses near the crack tip is characterized by the SIF modulus
Ko .
c0 = 0.1 and different values of the relative bridged zone size. The crack surfaces
displacement jumps are decreasing with increasing bridging zone size. Note that the
normal component of the jump at the crack center decreases almost linearly with
increasing bridging zone size. The normal component exceeds the shear one by an
order of magnitude.
The dependence of the modulus of the vector
σ (s) = q2y (s) + q2x (s)
of the bond tractions at the trailing edge of the bridged zone on the relative size of
the zone is given in Fig. 6 for different values co . Maximum of tractions is attained
at the smaller values (d/) when the relative bond compliance, co , is decreased.
For instance, the maximum value is equal to (σ /σo ) = 10.136 at sm ≈ 0.0654 if
co = 0.025. Note that the more compliant bonds then the less significant maximum
of the bond tractions (Fig. 6).
Almost uniform distribution of the bond tractions occurs if the bonding zone oc-
cupies the whole crack (crazelike defect). The size of the small regions of nonuni-
form bond tractions near the crack tips decreases with diminishing the bond com-
pliance.
Note that the existence of the absolute maximum of the bond tractions and dis-
placement distributions in the case of the linear bonds represents the most interesting
feature of these distributions. This feature was not mentioned earlier.
The dependencies of the relative modulus of the SIF,
Kr = (Ko /Koext )
(see formula (24)) on the relative bridged zone size are given in Fig. 7 at the dif-
ferent values of the relative bond compliance, co . If the relative bond compliance is
decreasing then the shielding effect by bonds will increase. After the certain value of
bridged zone size the relative SIF does not vary very much if the bond compliance
is rather small (co < 0.1).
5 Conclusion
Acknowledgements The support of this research by Russian Foundation for the Basic Research,
the number of the grant is 11-08-01243 is gratefully acknowledged.
Integro-Differential Equations for Stress Analysis 295
Fig. 1 Crack at the interface under the normal and shear loads
Fig. 2 The dimensionless normal tractions over the crack bridged zone
296 M. Perelmuter
Fig. 3 The dimensionless shear tractions over the crack bridged zone
Fig. 7 The dimensionless stress intensity factor, Kr , vs. the bridges zone size
References
[GoPe99] Goldstein, R.V., Perelmuter, M.N.: Modelling of bonding at an interface crack. In-
ternational J. of Fracture, 99, n. 1–2, 53–79 (1999).
[Sl81] Slepjan, L.I.: Crack Mechanics, Sudostroenie Publishing, Leningrad (1981).
[En68] Entov, V.M., Salganik, R.L.: Prandtl model of brittle fracture. Inzhenernyi Zhurnal,
MTT, 6, 87–99 (1968).
[HuMeRi87] Hutchinson, J.M., Mear, M.E., Rice, J.R.: Crack parallel an interface between dis-
similar materials. J. of Applied Mechanics, 54, 828–832 (1987).
[GoPe00] Goldstein, R.V., Perelmuter, M.N.: An interface crack with bonds between the sur-
faces. Mechanics of Solids, 36, n. 1, 77–92 (2000).
[Pe08] Perelmuter, M.N.: A criterion for the growth of cracks with bonds in the end zone.
J. of Applied Mathematics and Mechanics (PMM), 71, n. 1, 137–153 (2007).
Design and Performance of Gas–Liquid
Cylindrical Cyclone/Slug Damper System
1 Introduction
E. Pereyra
The University of Tulsa, OK, USA,
e-mail: ep@utulsa.edu
L. Gómez
The University of Tulsa, OK, USA,
e-mail: lgm@utulsa.edu
R. Mohan
The University of Tulsa, OK, USA,
e-mail: ram-mohan@utulsa.edu
O. Shoham
The University of Tulsa, OK, USA,
e-mail: ovadia-shoham@utulsa.edu
G. Kouba
Chevron Energy Technology Company, Houston, TX, USA,
e-mail: genekouba@chevron.com
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 299
DOI 10.1007/978-0-8176-8238-5_28, © Springer Science+Business Media, LLC 2011
300 E. Pereyra et al.
In order to minimize the effect of large flow variations on the GLCC operation,
several control strategies have been developed, depending on the operational con-
ditions and the flow characteristics at the inlet. The control strategies typically aim
at maintaining the liquid level and pressure in the GLCC. As a result, even with
control, the downstream devices will experience the flow variations occurring at the
GLCC inlet.
To avoid this problem, preconditioning flow devices can be utilized at the inlet of
the GLCC. One of these devices is the Slug-Damper (SD), which can be connected
at the inlet of the GLCC. A schematic of the GLCC-SD is given in Fig. 1. As can
be seen, the slug-damper consists of two legs. The lower leg is equipped with a flow
restriction, namely, a segmented orifice (the opening is at the bottom). When a slug
hits the damper, the liquid accumulates in the lower leg due to the flow restriction
provided by the segmented orifice. Thus, the slug is dampened and a reduced and
fairly constant liquid flow rate passes through the orifice opening into the GLCC.
As can be seen in Fig. 1, the lower leg of the SD is connected to the GLCC lower
inlet while the upper leg is connected to the upper inlet of the GLCC, forming a
dual inlet configuration. Mainly separated gas flows through the upper leg into the
GLCC.
There are no previous studies of the GLCC-SD system, but there are related
studies of the individual devices. A brief summary of these studies is presented
next. Refer to [Pe05] for further details.
Studies of SD performance were presented by [ReGo08] and [Ko02], in which
the SD device was investigated theoretically and experimentally. [Wa00] developed
a dynamic model for the analysis of different control strategies for the GLCC, based
on classical control theory. Utilizing several sets of experimental data available in
the literature for swirling flows and CFD simulations, [Ma98] proposed a model for
Design and Performance of GLCC-SD 301
predicting the flow field in the GLCC. In a later study, a novel mechanistic model to
characterize two-phase swirling flow in a GLCC separator was proposed by [Go01].
The goal of this study is to analyze, for the first time, the GLCC-SD system.
2 Slug-Damper Model
A description of the SD system is given in the introduction along with its schematic
in Fig. 1. The objective of this model is to develop a relationship for the liquid level
as a function of time, input flow rates and SD geometry. A schematic of the lower
leg of the SD with all the variables involved is presented in Fig. 2. The proposed
model assumes no accumulation of gas in the SD.
For incompressible liquid flow, an overall liquid mass balance on the SD yields
dVL
= qLin (t) − qDLo (t). (1)
dt
The flow through the segmented orifice (qDLo ) can be predicted as the flow
through an underflow gate, following [MuYo02], namely
<
qDLo (t) = Cd(t) A0 2 g HD (t) (2)
where Cd is the discharge coefficient (which varies typically between 0.5 and 0.6)
that is a function of the depth ratio (HD /hDO ). Correlating the data presented by
[MuYo02], the following expression was obtained for a rectangular channel:
HD (t)/hDO
Cd(t) = < .
0.708 HD (t)/hDO + 1.506 HD (t)/hDO − 0.066
302 E. Pereyra et al.
HDCRIT = dD cos(θ ).
For Case 2 (HD < HDCRIT ), the liquid level forms a truncated cylinder with volume
of
dD 1
VL = (VL1 +VL2 +VL3 ) , (4)
48 sin(θ ) cos(θ )
where
and
2 HD HD [dD cos(θ ) − HD ]
Fa = − 1, Fb = .
dD cos(θ ) dD2 cos2 (θ )
A simplification of the SD model can be carried out by assuming a constant dis-
charge coefficient Cd ∼ 0.6. Additionally, the relationship between the SD liquid
level and liquid volume can be obtained from (3). Combining (1) and (3), yields
where
qLin sin(θ ) <
CA = , CB = 0.6 β 2 2 g sin(θ ).
AD
Design and Performance of GLCC-SD 303
A simplified dynamic model of the GLCC is presented, consisting of mass and pres-
sure balances on the GLCC.
The rate of liquid level change as a function of time is given by an overall liquid
mass balance in the GLCC, resulting in
Combining the gas continuity, gas equation of state and the overall gas volume bal-
ance equations, yields the rate of change of pressure as follows:
where
Applying a pressure balance on the liquid section (GLCC body and liquid leg) yields
C
[ pS (t) − pLo (t)] + ρL g/gC [HoL − hL (t)]
qLoL (t) = ,
KL (t)
where
!
γL 8 ρL n fL(i) L(i) m k( j)
KL (t) =
−3
(2.228 × 10 CvL (t))2
+ 2
π gc ∑ 5
d(i)
+∑ 4
d(i)
.
i=1 i=1
304 E. Pereyra et al.
Similarly, applying pressure balance on the gas section (GLCC body and gas leg)
yields
!
Δ pa n f
G(i) L(i)
m k
/ ∑ +∑ 4 ,
2 ( j)
qGoG (t) =
8/π 2 i=1
5
d(i) i=1 d(i)
where
1 " !?
180C1 520 γG 3600 qGoG (t)
Δ pa = pS (t) 1 − sin−1 − pGo (t).
π 3417 T 14.7C1CvG (t)
4 Control Strategies
The proposed control strategy is the utilization of two proportional plus integral plus
derivative controllers, one for the liquid level utilizing a liquid control valve (LCV),
and the other for the separator pressure utilizing a gas control valve (GCV).
The LCV is utilized for controlling the liquid level in the GLCC, as proposed by
[Wa00]. The difference between the actual GLCC liquid level and the set point is
given by the error signal, which is the input to the controller module. The controller
uses the error signal for determining the control action, as a pressure signal (pCLCV ),
which is sent through a pneumatic line to the LCV. The pressure in the control valve
(pCLCV ) experiences a delay, due to the compressibility of the pneumatic line. The
valve pressure pushes the stem and spring of the actuator, closing or opening the
flow orifice, which determines the flow rate across the LCV. Thus, the action of the
controller is defined by
The error signal for this case represents the difference between the actual GLCC
pressure and the set point pressure. The controller can multiply, integrate or differ-
entiate the error signal for determining the control action to be taken. This is demon-
strated by the change of the pressure in the pneumatic line of the GCV (pCGCV ).
Similar to the liquid level control, the controller action is defined by
where
pSSET − pS (t) d pS d 2 pS
eG = − − tdGCV 2 .
trGCV dt dt
d pV pC (t) − pV (t)
= . (11)
dt τ0
The time constant (τ0 ) is a function of the length of the pneumatic line (Lpnl),
which can be predicted by means of the correlation
Lpnl
τ0 = . (12)
518 − 0.175 Lpnl − 0.0009 Lpnl2
The movement of the valve stem, promoted by the change of the diaphragm pressure
(pV ) also depends on the spring characteristics, the mass of the stem and the friction
in the packer. The valves used for the GLCC system are chosen to be air-to-close
valves.
Based on Newton’s law and neglecting the acceleration of the stem, [Ge85] pro-
posed the following equation for predicting the dynamic behavior of the actuator of
the valve
dx pMAX − pV (t) 1
= 100 − x(t) . (13)
dt pMAX − pMIN CO
Assuming that the time required for the stem position to reach an arbitrarily cho-
sen valve opening (xTEST ) is tTEST , Co can be calculated from the equation
306 E. Pereyra et al.
tTEST
CO = .
100
ln 100−xTEST
The control valve function is to vary the fluid flow through the valve by changing the
opening of the valve. The relationship between the flow through the valve and the
valve stem position (x) is called the valve characteristics. Only linear characteristics
is analyzed in this study, which can be predicted using the following equation:
CvMAX
Cv = x .
100
The objective of the pipeline slugging model is to provide an appropriate input for
the developed models. The slug flow model published by [TaBa90] for equilibrium
liquid film condition is adopted in this study. A train of slugs is generated using
random number, following the slug length distribution proposed by [AlSa03]. An
important feature of this model is the prediction of the maximum and minimum
bubble and droplet size using [Hi55] and [Ko03], respectively. These two parameters
enable prediction of the fluid particle size distribution at the inlet of the separator.
The overall GLCC-SD system model consists of nine (9) non linear ordinary dif-
ferential equations. The state variables and their corresponding equations are given
below:
1. SD volume (VL ), given by (1) (with auxiliary (2) through (6))
2. GLCC liquid level (hL ), (7)
3. GLCC pressure (pS ), (8)
4. LCV controller pressure (pCLCV ), (9)
5. GCV controller pressure (pCGCV ), (10)
6. LCV diaphragm pressure, (11)
7. GCV diaphragm pressure, (11)
8. LCV stem position, (13)
9. GCV stem position, (13)
Design and Performance of GLCC-SD 307
The solution of this model gives the evolution of the GLCC-SD system behavior
with time. The boundary conditions are the pressures in the gas and liquid outlet
lines and the instantaneous gas and liquid flow rate into the system.
A comparison between the SD model performance and [Ko03] data, and an example
of SD design are presented in this section. Figure 3 presents a comparison between
model predictions and the [Ko02] data for the SD outlet flow rate, corresponding
to a step liquid flow rate at the SD inlet. As can be seen, outlet flow rate increases,
due to the increase of the liquid potential energy in the SD. The model and the
experimental results present good agreement.
Simulations are carried out for the GLCC only (without the SD) and the GLCC-
SD system, for the same input flow rate conditions, in order to demonstrate the
advantage of the GLCC-SD system. The input data used for this example correspond
to a slug train obtained in a horizontal 3 in. pipe, operating at 350 psia and 100◦ F,
transporting 25◦ API gravity oil and 0.8 specific gravity gas. The GLCC is 6 in. in
diameter and 11 ft tall. Figure 4 presents the input and output flow rate of the GLCC
when no SD is connected. Usually, under normal operation, the GLCC dampens the
308 E. Pereyra et al.
outlet flow rate. However, as can be observed, for this case, where the GLCC and the
upstream pipeline have the same diameter, and as a result of fast liquid level control,
the outlet liquid flow rate follows closely the input liquid flow rate, exhibiting large
fluctuations. This can cause operational problems in downstream equipment, such
as flow meters.
Figure 5 presents the input liquid flow rate, as well as the GLCC-SD system
outlet flow rate. The same GLCC is considered with an attached SD, 6 in diameter
and 9.4 ft long. As can be seen, the input liquid flow rate exhibits large fluctuations,
corresponding to the large slugs produced. However, the output liquid flow rate from
the GLCC-SD is quite uniform, showing the strong damping effect of the system on
the input slugs. This flow rate is much smoother than the input liquid flow rate,
varying between 1100 to 2200 barrels per day (BPD), as compared to 500 to 3500
BPD for the input flow rate. Also, when the SD-GLCC system is used, the liquid
level in the GLCC is fairly constant as compared to its value for the GLCC alone.
Fig. 4 GLCC Input and output flow rates as functions of time (without SD)
9 Conclusions
A dynamic model and a simulator have been developed for the Gas–Liquid Cylin-
drical Cyclone/Slug-Damper (GLCC-SD) system, operating under slug flow con-
ditions. The separate models of the GLCC and the SD units are integrated to-
gether with a slug generator model/simulator into an overall model/simulator for
the GLCC-SD system. It is demonstrated that an installation of a SD at the inlet of
the GLCC can provide damping capability to the compact GLCC separator, ensur-
ing low and fairly constant liquid flow rates out of the systems, which is required to
protect downstream equipment, such as flow meters.
Nomenclature
A = Area, L2 , ft2
Cd = discharge coefficient (-)
Co = control valve actuator time response constant (-)
Cv = valve coefficient (-)
d = diameter, L, ft; derivative
f = friction factor (-); frequency
g = acceleration due to gravity, L/t 2 , ft/s2
gc = conversion of units parameter, gc = 32.2 lbm ft
lbf s2
H = damper height, L, ft
Hs = separator total height, L, ft
L = hydrostatic pressure correction, L, ft
h = height, L, ft
L = length, L, ft
Lpnl = length of pneumatic line, L, ft
p = pressure, M/Lt2 , psi
pC = control pressure, M/Lt2 , psi
pS = separator pressure, M/Lt2 , psi
pV = valve pressure, M/Lt2 , psi
q = volumetric-flow rate, L3 /t, ft3 /s
Ts = separator temperature, T, ◦ F
t = time, t, s
V = volume, L3 , ft3
x = valve opening (%)
310 E. Pereyra et al.
Greek Letters
Subscripts
D = damper; drift
DO = damper orifice
G = gas
GOG = output gas flow rate through the gas leg
GOL = output gas flow rate through the liquid leg
in = inlet
INI = initial
L = liquid
LOL = output liquid flow rate in the liquid leg
LOG = output liquid flow rate in the gas leg
MAX = maximum
MIN = minimum
O = orifice
o = outlet
S = slug; separator
SET = set point
Abbreviations
References
[Wa00] Wang, S.: Dynamic simulation, experimental investigation and control system design
of gas–liquid cylindrical cyclone separators. Ph.D. Dissertation. The University of
Tulsa (2000).
[ReGo08] Reinoso, A., Gomez, L., Wang, S., Mohan, R., Shoham, O.: Design and performance
of slug damper. Energy Resour. Technol., 130, n. 4, (2008).
[Ko02] Kouba, G.: A slug damper for compact separators, in: ETCE2002/PROD-29116, Pro-
ceedings of ASME Engineering Technology Conference on Energy, Houston, TX (Feb.
4–5, 2002).
[Ma98] Mantilla, I.: Bubble trajectory analysis in GLCC separators. M.S. Thesis. The Uni-
versity of Tulsa (1998).
[Go01] Gomez, L.: Dispersed two-phase swirling flow characterization for predicting gas
carry-under in gas–liquid cylindrical cyclone compact separators. Ph.D. Dissertation.
The University of Tulsa (2001).
[Av01] Avila, C.: Mathematical modeling for integrated three-phase compact separators.
M.S. Thesis. The University of Tulsa (2003).
[MuYo02] Munson, B., Young, D., Okiishi, T.: Fundamentals of Fluid Mechanics, Fourth edition,
John Wiley and Sons (2002).
[Go98] Gomez, L.: A state-of-the-art simulator and field application design of gas–liquid
cylindrical cyclone separators. M.S. Thesis. The University of Tulsa (1998).
[Pe05] Pereyra, E.: Transient mechanistic model for slug damper/Gas–Liquid Cylindrical
Cyclone (GLCC) compact separator system. M.S. Thesis. The University of Tulsa
(2005). (http://www.tustp.org/publications)
[Ge85] Genceli, H.: Dynamic simulation of slug catcher behavior. M.S. Thesis. The Univer-
sity of Tulsa (1985).
[TaBa90] Taitel, Y., Barnea, D.: Two-phase slug flow, in: Advances in Heat Transfer, vol. 20,
Academic Press Inc., San Diego, CA, 83–132 (1990).
[AlSa03] Al-Safran, E., Sarica, C., Zhang, H., Brill, J.: Probabilistic/mechanistic modeling of
slug length distribution in a horizontal pipeline, in: SPE 84230 (2003).
[Ko03] Kouba, G.: Mechanistic model for droplet formation and breakup, in: Proceedings of
ASME/JSME (2003).
[Hi55] Hinze, J.: Fundamentals of the hydrodynamic mechanism of splitting in dispersion
processes. AIChE J., 1, 289 (1955).
On Quasimodes for Compact Operators
and Associated Evolution Problems
E. Pérez
1 Introduction
E. Pérez
Universidad de Cantabria, Santander, Spain,
e-mail: meperez@unican.es
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 313
DOI 10.1007/978-0-8176-8238-5_29, © Springer Science+Business Media, LLC 2011
314 E. Pérez
tial and time variable in a separate way, and, by analogy with the wave equations,
are referred to as standing waves (see [Pe08] and [LoPe10]) for comparison). In
Sect. 3, we use the explicit solution of nonhomogeneous abstract evolution equa-
tions for bounded operators on Hilbert spaces, in terms of the semigroup (cf. (24)),
to obtain the estimates mentioned above. The bounds in these estimates depend on
t, μ , r, r∗ , the norm of A, and the inverse of the spectral distance d(μ , σ (A))−1 .
Comparing bounds in Sects. 2 and 3, those in Sect. 3 seem to be simpler and
they might be used in an alternative way. However, when applying the results, the
spectral distance might be unknown and very small: see Remark 2 in connection
with this.
satisfying
I(r∗ )
2r
u − u∗ H = u − ∑ αk ui(r∗ )+k ≤ ∗ . (2)
k=1
r
H
I(r∗ ) I(r∗ )
Here {ui(r∗ )+k }k=1 are the eigenfunctions of A associated with {λi(r∗ )+k }k=1 which
we assume to be orthonormal in H.
In order to introduce a general framework for first and second order evolution
problems associated with the operator arising in Lemma 1, throughout the section
we consider A a linear, compact, self-adjoint and nonnegative operator on the sepa-
rable Hilbert space H. Let us start by considering first-order evolution equations.
Let ω ± > 0 and ξ0± > ω ± be the constants, depending on the sign accompanying
the operator A, such that
and
R(±A + ξ0± I) = H . (4)
Here, each one of the signs ± implies a chosen sign ± accompanying A, and as is
well known (cf. [Pe11] for instance), it suffices to take ω + = 0 and ω − = AL (H)
to get (3), while ξ0+ > 0 and ξ0− > AL (H) give (4).
For any ϕ ∈ H, let us consider the evolution problem
⎧
⎨ du±
± Au± = 0 (5)
⎩ dt
u± (0) = ϕ
with ω + = 0.
The following theorem states the relationship between quasimodes (u, μ ) of the
operator A and standing waves of the type e±μ t u approaching solutions of (5) (see
[Pe11] for the proof).
for the negative sign. Here, C0 and ω ± are constants independent of λ , r, r∗ and t,
with ω ± appearing in (6). The same bounds hold for
∗
I(r )
±λ t
∑ αk e
±λi(r∗ )+k t
ui(r∗ )+k − e u
k=1
H
For the same operator A introduced above, let us consider the bounded operator
A defined on H = H × H by = >
0 , ±I
(9)
A,0
316 E. Pérez
has a unique solution ū(t), ū ∈ C1 ([0, T ]; H) for any positive T , which satisfies
satisfying (1)–(2). Let us assume that r∗ > r and λ − r∗ > 0. Then the following
assertions hold:
(i) For ϕ = 0, ψ = u, and for any t > 0, the solution u(t) of (12) satisfies
sin(√λ t)
√ u − u(t)
λ
H
√
0 2r 2r 1 √ ∗ r∗
≤ C max ∗ (CA + t), √ , √ r t+ √ √ (14)
r r∗ λ λ λ − r∗
and
√ √
cos( λ t)u − du (t) ≤ C0 max 2r (CA + t) , 2r , r∗ t . (15)
dt H r∗ r∗
Also, the bounds (14) and (15) hold for the discrepancy
I(r∗ ) sin( λi(r∗ )+k t) √
sin( λ t)
∑ αk ui(r∗ )+k − √
λ
u
k=1 λi(r∗ )+k
and
√ √
λ sin( λ t)u + du (t)
dt H
0 2r 2r √ √ √ ∗ √
≤ C max ∗ (CA + t), ∗ λ , λ r t+ r ∗ . (17)
r r
Also, the bounds (16) and (17) hold for the discrepancy
I(r∗ ) √
∑ αk cos( λi(r∗ )+k t) ui(r∗ )+k − cos( λ t)u
k=1
⎪ u(0) = ϕ (18)
⎪
⎪
⎪
⎩ du
(0) = ψ
dt
for initial data (ϕ , ψ ) ∈ H × H. Problem (18) reduces to (10) for the operator A
defined on H = H × H by (9), with the positive sign accompanying I, and for ϕ̄ =
(ϕ , −ψ ). In addition, the unique solution u(t) of (18), u ∈ C2 ([0, T ]; H) for any
positive T , satisfies the inequality (13) with F(t) = 2eω t defined in (11) (see [Pe11]).
Theorem 3. Let (u, λ ) be a quasimode with remainder r of the operator A arising
in (18). Let {(λi(r∗ )+k , ui(r∗ )+k )}k=1,2,...,I(r∗ ) be the eigenelements of the operator A
satisfying (1)–(2). Let us assume that r∗ > r and λ − r∗ > 0. Then, the solution u(t)
of (18) satisfies: √
(i) For (ϕ , ψ ) = (u, − λ u) and any t > 0,
√
√
− λt 2r √ √
e u − u(t) ≤ C0 max (1 + λ ) + r ∗ e ωt
, r ∗ t (19)
H r∗
and
√ −√λ t du
λe u + (t)
dt H
√
2r √ 2r √ √ √ √ √
≤ C0 max (1 + λ ) + r ∗ eω t , λ , ( λ + r ∗ ) r∗ t + r∗ .
r∗ r∗
(20)
√
(ii) For (ϕ , ψ ) = (u, λ u) and any t > 0,
√
λt
e u − u(t)
H
√
2r √ 2r √λ t √ ∗ (√λ +√r∗ )t
≤ C0 max ∗
(1 + λ ) + r∗ eω t , e , r te (21)
r r∗
and
√ √λ t √ √
λ e u − du (t) ≤ C0 max 2r
(1 + λ ) + r ∗ eω t ,
dt H r∗
2r √ √λ t √ √ √ √ √ √ √
∗ ) r ∗ te( λ + r∗ )t + r ∗ e λ t . (22)
λ e , ( λ + r
r∗
In addition, estimates (19) and (20) also hold for the norms in H of
I(r∗ ) √ √
∑ αk e− λi(r∗ )+k t
ui(r∗ )+k − e− λt
u
k=1
Quasimodes and Evolution Problems 319
and its time derivative, while estimates (21) and (22) hold for the norms in H of
I(r∗ ) √ √
∑ αk e λi(r∗ )+k t
ui(r∗ )+k − e λt
u
k=1
As is well known, the solution of (10) (cf. also (5), (12) and (18)) and, more pre-
cisely, of the nonhomogeneous equation
⎧
⎨ d ū
+ A ū = f¯(t)
(23)
⎩ dt
ū(0) = ϕ̄
In this section we show that (24) allows us to obtain certain bounds for the dis-
crepancies between the solutions ū(t) of (10) when the initial data are linear com-
binations of eigenfunctions (or quasimodes) and the standing waves constructed
from the quasimode (u, λ ) in Theorems 1–3. The bounds that we obtain in this sec-
tion are different from those in Sect. 2; they are simpler but they involve, among
other parameters, the inverse of the distance from λ to the spectrum of A, that is,
d(λ , σ (A))−1 : see Remark 2 in this connection.
Problem (5), which is the simplest one, is considered in Sect. 3.1. As a matter of
fact, bounds (7) and (8) in Theorem 1 are replaced by (25) in Theorem 4. We proceed
in a similar way for the second-order evolution problems (12) and (18) in Sect. 3.2.
Let us refer to Sects. III.6.2–III.6.3 and V.3.8 in [Ka66] for the general theory and
formulas that we use throughout this section. See also Sect. III.8 of [SaSa89] in this
connection.
Proof. As is well known, u± (t) = e∓At u. Let us consider the function e∓λ t u which
satisfies
d ∓λ t
e u ± Ae∓λ t u = ±(Au − λ u)e∓λ t
dt
and takes the value u at t = 0. Let us denote by w the “rest”, that is, w = Au − λ u
with a norm in H bounded by r. Then, the function e∓λ t u − u± (t) is the solution of
problem (23) for f = e∓λ t (±w) and ϕ = 0.
Consequently, formula (24) reads
t t
e∓λ t u − u± (t) = e−(±A)(t−s) e∓λ s (±w)ds = e∓At e±(A−λ I)s ds (±w)
0 0
5t
= e∓At (A − λ I)−1 [e±(A−λ I)s 0 w = (A − λ I)−1 (e∓λ It − e∓At )w ,
(27)
It suffices to realize that it satisfies (23) for the same f (t) above, namely, f (t) =
I(r∗ )
±(Au − λ u)e∓λ t , and for ϕ = u − ∑k=1 αk ui(r∗ )+k with a norm less or equal than
I(r∗ )
2r(r∗ )−1 . Let us denote w∗ ≡ u − ∑k=1 αk ui(r∗ )+k . Therefore, (24) takes the form
I(r∗ ) t
e ∓λ t
u− ∑ αk e∓λi(r∗ )+k t ui(r∗ )+k = e∓At w∗ + 0
e−(t−s)(±A) e∓λ s (±w)ds
k=1
Throughout this section we consider the case of the second-order equation, namely,
(23) for the operator A in (9), depending on the sign accompanying I. For the sake
of simplicity, throughout the section we perform the proof for problem (18) and we
state the results for problem (12) without proof (cf. comments after the proof of
Theorem 5).
Theorem 5. Let us consider the hypotheses in Theorem 3 on the operator A, the
quasimode (u, λ ), the eigenelements {(λi(r∗ )+k , ui(r∗ )+k )}k=1,2,...,I(r∗ ) , and the num-
√
bers r, r∗ and λ . Then, for ϕ = u, ψ = ± λ u, and for any t > 0, the solution u(t)
of (18) and its derivative satisfy
√ √ √
u(t) − e± λ t u + − du (t) ± λ e± λ t u
H dt
√ H √
≤ (1 + 2 λ + λ ) d(λ , σ (A))−1 + 1 e± λ I t − e−A t r . (28)
L (H)
and the same bound holds for the time derivative of the discrepancy above. Here,
A is the operator defined by (9) for the positive sign accompanying I, and I is the
identity operator on H.
Proof. We follow the notations used in the proof of Theorem 4 for w and w∗ . That
is, w = Au − λ u, while w∗ denotes the discrepancy between u and the linear com-
bination of eigenfunctions u∗ arising in the definition of quasimode (1)–(2), namely
w∗ = u − u∗ . They satisfy wH ≤√r and w∗ H ≤ 2r(r∗ )−1 .
Let us consider the function e± λ t u and u(t) the solution of (18) in the statement
of the theorem. We compute
2 √ √ √
d ± λt ± λt ± λt
− A e u = (−Au + λ u)e = −e w.
dt 2
Here, and in what follows, the superscript τ denotes the transpose vector.
Performing the integration in (30), we have
√ √
ū(t) = (A ± λ I )−1 e± λ I t − e−A t (0, −w)τ
and, therefore,
√ √
ū(t)H ≤ (A ± λ I )−1 L (H) e± λ I t − e−A t L (H) r . (31)
√
Let us obtain a bound for (A ± λ I )−1 L (H) in terms of (A − λ I)−1 L (H)
and consequently, in terms of d(λ , σ (A))−1 . As is known,
√
√ (A ± λ I )−1 w̄H
−1
(A ± λ I ) L (H) = sup . (32)
w̄∈H w̄H
w̄=0
√
For w̄√= (w1 , w2 )τ , let us analyze (A ± λ I )−1 w̄H ,√ by considering
(A ± λ I )−1 w̄ = v̄ ≡ (v1 , v2 )τ . That means that w̄ = (A ± λ I )v̄ and con-
sequently, (w1 , w2 ) satisfy
√ √
w 1 = ± λ v1 + v2 and w2 = Av1 ± λ v2 .
√ √ √
Thus, w2 = Av1 ± λ (∓ λ v1 + w1 ), or equivalently, Av1 − λ v1 = w2 ∓ λ w1 ,
which determines v1 and v2 in terms of the resolvent of A at λ . Namely,
√ √ √
v1 = (A − λ I)−1 [w2 ± λ w1 ] and v2 = w1 ∓ λ ( (A − λ I)−1 [w2 ± λ w1 ] ).
√
Therefore, from (32), (A ± λ I )−1 L (H) is bounded by
√ √
(1 + λ )(A − λ I)−1 w2 H + ( λ + λ )(A − λ I)−1 w1 H + w1 H
sup
w̄∈H w̄H
w̄=0
The same bound is obtained for the time derivative of the discrepancy. Hence, from
(33) and the bound above, (29) holds, and the theorem is proved.
Let us state the main results for the case of the second-order evolution equation
arising in (12). We avoid here the proofs which imply extending real Hilbert spaces
and operators acting on these spaces to complex Hilbert spaces and associated op-
erators (cf. Sect. VI.I in [Mi69] for the kind of construction).
Below, we consider the assumptions in Theorem 2. Let us consider the solution
u(t) of (12) for the initial data ϕ = α u, ψ = β u for the value of the constants α and
β either 0 or 1.
Theorem 6. Let us consider the hypotheses in Theorem 2 on the operator A, the
I(r∗ )
quasimode (u, λ ), the eigenelements {(λi(r∗ )+k , ui(r∗ )+k )}k , and the numbers r,
r∗ and λ . Then, for ϕ = α u, ψ = β u, with the value of the constants α and β either
0 or 1, and for any t > 0, the solution u(t) of (12) satisfies
= √ >
√ sin ( λ t)
u(t) − α cos ( λ t) + β √ u
λ
H
du √ √ √
+ (t) + α λ sin ( λ t)u − β cos ( λ t)u
dt H
√ √
β
≤ α+√ 1 + 2 λ + λ ) d(λ , σ (A))−1 + 1 e−i λ I t − e−A t r.
λ L (H∗ )
depending on α and β .
324 E. Pérez
and the same estimate holds for its time derivative. Also, in the case where α = 0
and β = 1, the inequality above holds for
I(r∗ ) sin( λi(r∗ )+k t) √
sin( λ t)
∑ αk ui(r∗ )+k − √
λ
u
k=1 λi(r∗ )+k
Acknowledgements This work has been partially supported by the Spanish grant MICINN:
MTM2009-12628.
References
[Ka66] Kato, T.: Perturbation Theory for Linear Operators, Springer-Verlag, New York
(1966).
[La99] Lazutkin, V.F.: Semiclassical asymptotics of eigenfunctions, in: Partial Differential
Equations V (Editor: M.V. Fedoryuk), Springer-Verlag, Heidelberg, 133–171 (1999).
Quasimodes and Evolution Problems 325
[LoPe10] Lobo, M., Pérez, E.: Long time approximations for solutions of wave equations as-
sociated with Steklov spectral homogenization problems. Math. Meth. Appl. Sci., 33,
1356–1371 (2010).
[Mi69] Mikhlin, S.G.: Mathematical Physics, an Advanced Course, North-Holland, Leningrad
(1969).
[Pe08] Pérez, E.: Long time approximations for solutions of wave equations via standing
waves from quasimodes. J. Math. Pure Appl., 90, 387–411 (2008).
[Pe11] Pérez, E.: Long time approximations for solutions of evolution equations from quasi-
modes: perturbation problems. Math. Balkanica (N.S.), 25 (Fasc. 1–2), 95–130 (2011).
[SaSa89] Sanchez-Hubert, J., Sanchez-Palencia, E.: Vibration and Coupling of Continuous Sys-
tems. Asymptotic Methods, Springer, Heidelberg (1989).
Error Estimation by Means of Richardson
Extrapolation with the Boundary Element
Method in a Dirichlet Problem for the Laplace
Equation
S. Pomeranz
1 Introduction
2 Problem Statement
The problem to be solved is the interior Dirichlet problem for Laplace’s equation on
:
D, a bounded domain in the plane, i.e., to find u ∈ C(D) C2 (D) such that
!u(P) = 0, P ∈ D,
u(P) = u0 (P), P ∈Γ, (1)
S. Pomeranz
The University of Tulsa, OK, USA,
e-mail: pomeranz@utulsa.edu
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 327
DOI 10.1007/978-0-8176-8238-5_30, © Springer Science+Business Media, LLC 2011
328 S. Pomeranz
Once (2) is solved for q ≡ ∂ u/∂ n, the unknown normal boundary flux, then u, the
unknown potential in the interior of domain D, can be determined using
1 ∂ ln |P − Q|
u(P) = u0 (Q) − q(Q) ln |P − Q| dSQ , P ∈ D.
2π Γ ∂ nQ
where A denotes a constant of proportionality and the higher-order error terms are
assumed negligible. Taking a linear combination of (3) and (4) and solving for q,
yields
h p q1 − h1p q2
q ≈ q̃ ≡ 2 p .
h2 − h1p
The Richardson extrapolation result is given by q̃.
If the value of p, the order of the dominant error term, is unknown, then three
approximations, using grid spacings h1 , h2 , and h3 , respectively, can be used to
obtain
Error Estimation by Means of Richardson Extrapolation 329
q1 q3 − q22
q ≈ q̃ ≡ , (5)
q1 − 2q2 + q3
where it is assumed that h1 /h2 = h2 /h3 ≡ c, for some constant c > 1. Further, the
order of the dominant error term can be approximated by
ln qq23 −q
−q2
1
p≈ . (6)
ln c
In the numerical example investigated in Sect. 5, the value of p is not known.
Richardson extrapolation is applied to the model problem. Using [BuFa05, #16,
p. 87], the Richardson extrapolation convergence can be described as if it were con-
vergence from Aitken’s !2 method. That convergence result is stated now in Theo-
rem 1:
Theorem 1. Suppose that the sequence of approximations {qn }∞
n=1 converges to the
limit q such that
qn+1 − q
0 ≤ lim ≡ λ < 1,
n→∞ qn − q
qn qn+2 − q2n+1
q̃n ≡ , n = 1, 2, . . . ,
qn − 2qn+1 + qn+2
q̃n − q
lim = 0.
n→∞ qn − q
Now consider approximating the unknown quantity, for example, the normal
boundary flux, q, (or the interior potential) using three successive grid spacings,
h1 > h2 > h3 , chosen sufficiently small. The approximations q1 , q2 , and q3 are
computed using h1 , h2 , and h3 , respectively, and are such that (q2 − q)/(q1 − q) ≈
(q3 − q)/(q2 − q). Further assume that the errors qi − q, for i = 1, 2, 3, all have the
same sign. Theorem 1 states that if q1 , q2 , and q3 belong to the sequence converging
to q as described, then Richardson extrapolation based on q1 , q2 , and q3 will give a
more accurate approximation than that given by q1 , q2 , or q3 .
Recall that the original problem to be solved is the interior Dirichlet problem for the
:
Laplace equation (1), to find u ∈ C(D) C2 (D) such that
330 S. Pomeranz
!u(P) = 0, P ∈ D,
u(P) = u0 (P), P ∈Γ,
where Γ is the boundary of domain D and u0 ∈ C(Γ ) is a given function. For the
numerical experiments, the domain was selected to be the unit square,
Therefore, the boundary of D is not smooth, and, consequently, the theory for do-
mains with smooth boundaries does not apply. Nevertheless, it is still of interest to
apply Richardson extrapolation (5) to boundary element results.
Let p be the order of the dominant error term, as defined in (3) and (4). One
issue that should be resolved is how to determine if the Richardson extrapolation
numerical results are valid. Numerical results for the p value estimates (6) can be
used to predict those grid points at which the Richardson extrapolation results are
not accurate. The p values should be positive numbers. The p value estimates that
are complex numbers (with nonzero imaginary parts), negative, or too small (rela-
tive to the neighboring values) can be interpreted as warning flags, i.e., indicators
that predict the grid points at which the Richardson extrapolation results are not ac-
curate (relative to the neighboring results). This provides a rejection criterion: The
corresponding Richardson extrapolation results should be rejected. These values are
not accurate either due to the boundary element method approximations themselves
being inaccurate (for example, near corners of the domain) or due to the Richardson
extrapolation being inaccurate (for example, if the grid spacings are too large). We
cannot say with certainty that some Richardson extrapolation results are good, but
we can say with certainty that specific Richardson extrapolation results are bad.
For the purposes of describing the numerical experiments reported here, the fol-
lowing terminology is used. A “bad Richardson extrapolation value” at a grid point
(either on the boundary, when the normal boundary flux is of interest; or in the inte-
rior, when the potential (primary unknown) is of interest) is one for which the fine
grid result is more accurate than the Richard extrapolation result. This means that
the error magnitude for the fine grid result is smaller than that for the Richardson
extrapolation result, which should not be the case.
Similarly, a “bad p value” at a grid point (either on the boundary or in the interior)
is one that is complex (with nonzero imaginary part), negative, or too small. Since
the p value estimate is the numerically approximated order of the dominant error
term, it should be a positive number, and not too small in regions for which the
solution is reasonably smooth.
The bad Richardson extrapolation values are known in Sect. 5, since a test prob-
lem is used with known exact solutions for both the normal boundary flux and the
interior potential. However, in applications the bad Richardson extrapolation values
would not be known, and the bad p values, which are known, would be used as
warning flags, i.e., predictors of points with poor Richardson extrapolation results.
Error Estimation by Means of Richardson Extrapolation 331
5 Numerical Experiments
The numerical example chosen for the model problem (1) has the exact solution
1
u(x, y) = π ey cos(x − π /7) + e(1−π x) cos(π y − π /2) + π e5x cos(5y − π /2).
100
The square domain is D = {(x, y)| 0 < x < 1, 0 < y < 1}. The Dirichlet boundary
data used were those given on the boundary by the exact solution.
A Mathematica© notebook implementing the boundary element method was
written by the author of this paper. The notebook implemented collocation in the
classical direct boundary element method [GaKoWa03]. Piecewise constant (dis-
continuous) elements (basis functions) for the primary unknown and its normal
boundary flux were used. The collocation points were selected as the midpoints of
each element (subinterval) in a uniform element grid on the boundary of the square.
Note that the geometry nodes that delineate the end points of each boundary element
are different from the collocation nodes, which are at the element midpoints. The
boundary nodes were numbered sequentially, counterclockwise from the lower left
vertex.
Initially, a relatively coarse uniform geometry grid was used on the boundary.
The number of geometry x-nodes on a horizontal edge of the square domain was set
equal to the number of geometry y-nodes on a vertical edge, denoted by nxnodes
and nynodes, respectively. The boundary grid was then uniformly refined in such a
way so as to have the collocation nodes in the coarse boundary grid remain as collo-
cation nodes in the two other grids, the refined grids, that were used to construct the
Richardson extrapolation results. The numbers of boundary geometry nodes were
taken successively as nxnodes = nynodes = 10, 28, and 82, respectively, for the
coarse, intermediate, and fine boundary grids.
The interior grid was similarly refined for Richardson extrapolation of the interior
potential values, where there was a 9 × 9 interior coarse grid (81 interior nodes).
The 81 interior nodes were numbered sequentially from the lower left, from left to
right and then from bottom to top.
The focus of the numerical results that will now be described is (i) to use Richard-
son extrapolation to improve the numerical results for the normal outward boundary
flux and, separately, (ii) to use Richardson extrapolation to improve the numerical
results for the interior potential. A main interest in both (i) and (ii) is to be able to
get information on the validity of these Richardson extrapolation results, especially
since precise values of the parameters involved in the error terms are not known.
The numerical results for normal boundary flux errors computed with a coarse grid
(nxnodes = nynodes = 10), intermediate grid (nxnodes = nynodes = 28), fine grid
332 S. Pomeranz
(nxnodes = nynodes = 82), and associated Richardson extrapolation and the result-
ing numerical estimates for the p values are given in Table 1.
These numerical errors for the normal boundary flux obtained from executing the
Mathematica boundary element notebook three times, using the coarse, intermedi-
ate, and refined boundary grids, are compared with the Richardson extrapolation
error that is computed using these three data sets. Specifically, the Richardson ex-
trapolation errors should be compared node-by-node with the fine grid errors to see
which are smaller in magnitude. In particular, see the data in bold font in Table 1,
which will now be discussed.
The numerical justification for application of Richardson extrapolation, with the
expectation that Richardson extrapolation will improve the numerical results, is pro-
vided most clearly by Table 1 and Fig. 1. The Richardson extrapolation results
should be better than the fine grid results. However, the Richardson extrapolation
results at the nine boundary nodes numbered 10, 17, 18, 21, 22, 23, 27, 28, and 36
are not as accurate as the corresponding fine grid results. This is known here since
a test problem with a known solution is used. However, in an actual application,
this information would be desired, but not available. The utility of the numerical p
value estimates can now be demonstrated. These a posteriori estimates are available
and predict the locations of bad Richardson extrapolation results. In Fig. 1, it can
be observed that there are bad p values at the ten boundary nodes numbered 1, 9,
10, 17, 18, 19, 21, 27, 28, and 36. The p values predict the locations at which the
Richardson extrapolation results are bad.
There are discrepancies, though, at the following boundary nodes. Boundary
nodes numbered 1, 9, and 19 are detected as bad by the p value estimates but are not
bad Richardson nodes. However, this is desirable because, although the magnitudes
of the Richardson extrapolation errors are smaller than the magnitudes of the fine
grid errors at these boundary nodes, both the fine grid results and the Richardson ex-
trapolation results are very inaccurate. Note from Fig. 1 that these boundary nodes
that have been detected are at or near corners of the domain, and these are points at
which the boundary element results themselves are inaccurate. Therefore, it is good
that the p values detected these boundary nodes, and local grid refinement could
be performed near these nodes in order to improve the accuracy of the boundary
element results.
On the other hand, in Fig. 1, it can be observed that at boundary nodes num-
bered 22 and 23 the p value estimates do not predict anything problematic, however
these are bad Richardson extrapolation nodes. This is not troublesome because both
the fine grid normal boundary flux results and the Richardson extrapolation normal
boundary flux results are relatively accurate and the differences in relative accuracy
are not significant.
Error Estimation by Means of Richardson Extrapolation 333
The numerical results for interior potential errors computed with a coarse grid (nxn-
odes = nynodes = 10), intermediate grid (nxnodes = nynodes = 28), fine grid (nxn-
odes = nynodes = 82), and associated Richardson extrapolation and the resulting
numerical estimates for the p values are given in Table 2.
Fig. 1 Boundary flux nodes at which Richardson extrapolation normal boundary flux is bad (R),
compared with boundary nodes at which the p values predict bad Richardson extrapolation re-
sults (p)
Mathematica graphics of the solution for the interior potential obtained from
the Richardson extrapolation interior potential results and the exact solution are
displayed in Figs. 2 and 3.
These numerical errors for the interior potential obtained from executing the
Mathematica boundary element notebook three times, using the coarse, interme-
diate, and refined boundary grids, are compared with the Richardson extrapolation
error that is computed using these three data sets at the end of the calculations (i.e.,
the Richardson extrapolation for the normal boundary flux is not used here). Specif-
ically, the Richardson extrapolation errors should be compared node-by-node with
334 S. Pomeranz
Table 2 (Continued)
Interior Potential Errors p Estimate
node nxnodes = 10 nxnodes = 28 nxnodes = 82 Richardson
# Extrapolation
41 −3.23 × 10−3 −3.17 × 10−4 −3.37 × 10−5 −3.24 × 10−6 2.12
42 −3.16 × 10−3 −3.09 × 10−4 −3.28 × 10−5 −3.23 × 10−6 2.12
43 −3.26 × 10−3 −3.18 × 10−4 −3.37 × 10−5 −3.40 × 10−6 2.13
44 −3.69 × 10−3 −3.59 × 10−4 −3.80 × 10−5 −3.80 × 10−6 2.13
45 −3.67 × 10−3 −4.73 × 10−4 −5.01 × 10−5 1.44 × 10−5 1.84
46 −4.06 × 10−3 −5.29 × 10−4 −5.68 × 10−5 1.61 × 10−5 1.83
47 −4.34 × 10−3 −4.32 × 10−4 −4.62 × 10−5 −3.90 × 10−6 2.11
48 −3.78 × 10−3 −3.74 × 10−4 −4.00 × 10−5 −3.56 × 10−6 2.11
49 −3.39 × 10−3 −3.37 × 10−4 −3.60 × 10−5 −3.17 × 10−6 2.11
50 −3.09 × 10−3 −3.08 × 10−4 −3.29 × 10−5 −2.83 × 10−6 2.11
51 −2.76 × 10−3 −2.76 × 10−4 −2.95 × 10−5 −2.51 × 10−6 2.11
52 −2.29 × 10−3 −2.28 × 10−4 −2.43 × 10−5 −2.18 × 10−6 2.11
53 −1.55 × 10−3 −1.49 × 10−4 −1.58 × 10−5 −1.72 × 10−6 2.14
54 −5.30 × 10−4 −4.21 × 10−5 −4.15 × 10−6 −9.48 × 10−7 2.32
55 −3.24 × 10−3 −4.09 × 10−4 −4.37 × 10−5 1.05 × 10−5 1.86
56 −3.79 × 10−3 −3.72 × 10−4 −3.96 × 10−5 −3.86 × 10−6 2.12
57 −3.51 × 10−3 −3.46 × 10−4 −3.70 × 10−5 −3.43 × 10−6 2.12
58 −3.25 × 10−3 −3.25 × 10−4 −3.48 × 10−5 −2.90 × 10−6 2.10
59 −2.96 × 10−3 −3.00 × 10−4 −3.23 × 10−5 −2.37 × 10−6 2.09
60 −2.52 × 10−3 −2.59 × 10−4 −2.80 × 10−5 −1.77 × 10−6 2.08
61 −1.62 × 10−3 −1.69 × 10−4 −1.85 × 10−5 −9.52 × 10−7 2.06
62 2.51 × 10−4 2.43 × 10−5 2.40 × 10−6 4.83 × 10−8 2.12
63 2.55 × 10−3 3.89 × 10−4 4.18 × 10−5 −2.44 × 10−5 1.67
64 −2.51 × 10−3 −2.51 × 10−4 −2.62 × 10−5 −1.44 × 10−6 2.10
65 −3.20 × 10−3 −3.02 × 10−4 −3.19 × 10−5 −4.10 × 10−6 2.16
66 −3.20 × 10−3 −3.15 × 10−4 −3.36 × 10−5 −3.22 × 10−6 2.12
67 −3.11 × 10−3 −3.14 × 10−4 −3.38 × 10−5 −2.54 × 10−6 2.09
68 −2.97 × 10−3 −3.08 × 10−4 −3.34 × 10−5 −1.90 × 10−6 2.07
69 −2.68 × 10−3 −2.85 × 10−4 −3.12 × 10−5 −1.09 × 10−6 2.04
70 −1.87 × 10−3 −2.11 × 10−4 −2.34 × 10−5 3.59 × 10−7 1.99
71 3.48 × 10−4 1.66 × 10−5 9.35 × 10−7 1.61 × 10−7 2.78
72 3.27 × 10−3 6.20 × 10−4 6.65 × 10−5 −7.99 × 10−5 1.42
73 −2.20 × 10−4 −1.81 × 10−4 −1.59 × 10−5 −2.32 × 10−4 −1.31
74 −2.42 × 10−3 −2.69 × 10−4 −2.82 × 10−5 2.20 × 10−6 1.99
75 −2.38 × 10−3 −2.92 × 10−4 −3.12 × 10−5 6.06 × 10−6 1.89
76 −2.40 × 10−3 −3.07 × 10−4 −3.32 × 10−5 7.91 × 10−6 1.85
77 −2.46 × 10−3 −3.27 × 10−4 −3.57 × 10−5 1.04 × 10−5 1.81
78 −2.52 × 10−3 −3.54 × 10−4 −3.90 × 10−5 1.45 × 10−5 1.76
79 −2.49 × 10−3 −3.81 × 10−4 −4.24 × 10−5 2.23 × 10−5 1.67
80 −1.09 × 10−3 −3.51 × 10−4 −3.99 × 10−5 1.83 × 10−4 0.79
81 1.34 × 10−3 1.55 × 10−4 1.55 × 10−5 −3.32 × 10−6 1.94
Error Estimation by Means of Richardson Extrapolation 337
Fig. 4 Interior nodes at which Richardson extrapolation potential is bad (R), compared with inte-
rior nodes at which the p values predict bad Richardson extrapolation results (p)
the fine grid errors to see which are smaller in magnitude. In particular, see the data
in bold font in Table 2, which will now be discussed.
The numerical justification for application of Richardson extrapolation, with the
expectation that Richardson extrapolation will improve the numerical results, is pro-
vided most clearly by Table 2 and Fig. 4. The Richardson extrapolation results
should be better than the fine grid results. However, the Richardson extrapolation
results at the five interior nodes numbered 1, 9, 72, 73, and 80 are not as accurate
338 S. Pomeranz
as the corresponding fine grid results. This is known here since a test problem with
a known solution is used. However, in an actual application, this information would
be desired, but not available. Again, this demonstrates the utility of the numerical p
value estimates. These a posteriori estimates are available and predict the locations
of bad Richardson extrapolation results. In Fig. 4, it can be observed that there are
bad p values at the five interior nodes numbered 1, 9, 72, 73, and 80. The p values
predict the locations at which the Richardson extrapolation results are bad.
For this numerical example, the interior nodes at which the Richardson extrapola-
tion values are bad and the interior nodes at which the p value estimates predict bad
Richardson extrapolation results turn out to coincide. Further, these interior nodes
are all near corners of the domain, where the boundary element method performs
poorly. Therefore, the p values again can serve as a posteriori warning flags, i.e.,
indicators that the Richardson extrapolations results are suspect. In such cases the
grid can be locally refined or some other corrective actions can be taken.
Acknowledgements The author acknowledges support from a University of Tulsa Faculty Sum-
mer Fellowship.
References
[AtHa01] Atkinson, K., Han, W.: Theoretical Numerical Analysis–A Functional Analysis
Framework, Sections 11.1–11.2, 342–362; Chapter 12, 405–435, Springer-Verlag,
New York (2001).
[BuFa05] Burden, R.L., Faires, J.D.: Numerical Analysis, Eighth edition, Section 2.5, 83–87;
Section 4.2, 179–186, Thomson–Brooks/Cole (2005).
[Co00] Constanda, C.: Direct and Indirect Boundary Integral Equation Methods, Chapter
1, 1–53, Chapman & Hall/CRC, Boca Raton (2000).
[GaKoWa03] Gaul, L., Kogl, M., Wagner, M.: Boundary Element Methods for Engineers and
Scientists, Springer, Berlin (2003).
[KaNi87] Kaitai, L., Ningning, Y.: The Extrapolation Method for Boundary Finite Elements,
IMA Preprint Series, http://ima.umn.edu/preprints/Jan87-Dec87/287.pdf.
[NiKa89] Ning, Y., Kai-tai, L.: An extrapolation method for bem. J. Comput. Math., 7, n. 2,
217–225 (April 1989).
[Pa06] Panahi, M.: The boundary element method for numerical solution of the Laplace
equation. Int. J. Appl. Math., 19, n. 4, 403–410 (2006).
[RüZh98] Rüde, U., Zhou, A.: Multi-parameter extrapolation methods for boundary integral
equations. Adv. Comput. Math., 9, 173–190 (1998).
[Sm87] Smith, G.D.: Numerical Solution of Partial Differential Equations: Finite Differ-
ence Methods, Third edition, Oxford University Press, Oxford, 249 (1987).
[XuZh96] Xu, Y., Zhao, Y.: An extrapolation method for a class of boundary integral equa-
tions. Math. Comp., 65, n. 214, 587–610 (April 1996).
[YaSl88] Yan, Y., Sloan, I.H.: On integral equations of the first kind with logarithmic kernels.
J. Integral Equations Appl., 1, 549–579 (1988).
[Ya90] Yan, Y.: The collocation method for first-kind boundary integral equations on
polygonal regions. Math. Comp., 54, n. 189, 139–154 (Jan. 1990).
Convergence of a Discretization Scheme Based
on the Characteristics Method for a Fluid–Rigid
System
1 Preliminaries
In this chapter, we present our latest results concerning the convergence of a numeri-
cal method to discretize the equations modeling the motion of a rigid solid immersed
into a viscous incompressible fluid using the characteristics technique.
Before stating these results, let us introduce the continuous model of our prob-
lem. We assume that the fluid–rigid system occupies a bounded and regular domain
O ⊂ R2 and that the solid is a ball of radius 1 whose center, at time t, is denoted by
ζ (t). The fluid fills the part Ω (t) = O \ B(ζ (t)) at time t. The velocity field u(x,t)
and the pressure p(x,t) of the fluid, the center of mass ζ (t) and the angular velocity
ω (t) of the ball satisfy the following Navier–Stokes system coupled with Newton’s
laws:
∂u
ρf + (u · ∇)u − μΔ u + ∇p = ρ f f, x ∈ Ω (t),t ∈ [0, T ], (1)
∂t
div u = 0, x ∈ Ω (t),t ∈ [0, T ], (2)
u = 0, x ∈ ∂ O,t ∈ [0, T ], (3)
u = ζ (t) + ω (t)(x − ζ (t))⊥ , x ∈ ∂ B(ζ (t)),t ∈ [0, T ], (4)
mζ (t) = − σ n d Γ + ρs f(x,t)dx,t ∈ [0, T ], (5)
∂ B(ζ (t)) B(ζ (t))
J. San Martín
Universidad de Chile, Santiago, Chile,
e-mail: jorge@dim.uchile.cl
J.-F. Scheid
Université Henri Poincaré, Nancy, France,
e-mail: scheid@iecn.u-nancy.fr
L. Smaranda
Universitatea din Piteşti, Romania,
e-mail: smaranda@dim.uchile.cl
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 339
DOI 10.1007/978-0-8176-8238-5_31, © Springer Science+Business Media, LLC 2011
340 J. San Martín et al.
J ω (t) = − (x − ζ (t))⊥ · σ n dΓ + ρs (x − ζ (t))⊥ · f(x,t)dx,t ∈ [0, T ]. (6)
∂ B(ζ (t)) B(ζ (t))
In the above system, σ = −pId + 2μ D(u) denotes the Cauchy stress tensor with
D(u) = (∇u + ∇uT )/2 and ∇uT means the transpose of ∇u. The positive constant
μ is the dynamic viscosity of the fluid and the constants m and J are the mass and
the moment of inertia
of the rigid
body.
Throughout this chapter, we will use the
notation x⊥ = −x x1
2 for all x = x1 ∈ R2 . The system (1)–(6) is completed with
x2
initial conditions:
u(x, 0) = u0 (x), x ∈ Ω (0), (7)
ζ (0) = ζ 0 ∈ R2 , ζ (0) = ζ 1 ∈ R2 , ω (0) = ω0 ∈ R. (8)
One important hypothesis of our problem is that the density ρ f of the fluid and the
density ρs of the solid are constant, but different, that is,
ρ f = ρs .
In the remainder of this chapter we assume that any velocity field in K (ζ ) will be
extended by zero outside of O.
According to Lemma 1.1 of [Te83, pp. 18], for any u ∈ K (ζ ), there exist lu ∈ R2
and ωu ∈ R such that
u(y) = lu + ωu (y − ζ )⊥ ∀y ∈ B(ζ ).
The spaces (10), (11) are specific to our problem. In fact, if the solution u of (1)–(8)
is extended by
then we easily see that u(t) ∈ KD(ζ (t)). In the remainder of this chapter, the solution
u of (1)–(8) will be extended as above.
342 J. San Martín et al.
Remark 1. If
then by using a classical result of Liouville (see, for instance, [Ar92, p. 251]), we
see that
det Jψ0 = 1,
where we have denoted by Jψ0 = ( ∂∂ ψy ji )i, j the jacobian matrix of the transformation
0
0 (y).
y → ψ
Let us now state the weak formulation of the system (1)–(8), which we use to
discretize the problem in time.
Then (u, p, ζ , ω ) is the solution of (1)–(8) if and only if for all t ∈ [0, T ], u(·,t) ∈
K (ζ (t)), p(·,t) ∈ M(ζ (t)) and (u, p) satisfies
d 4 5
ρ u ◦ψ 0 (t), ϕ + a(u, ϕ ) + b(ϕ , p) = (ρ f(t), ϕ ) ∀ϕ ∈ K (ζ (t)), (13)
dt
b(u, q) = 0 ∀q ∈ M(ζ (t)), (14)
where the bilinear forms a(·, ·) and b(·, ·) are defined as follows:
a(u, v) = 2μ D(u) : D(v) dx ∀u, v ∈ H 1 (O)2
O
Discretization Scheme for a Fluid–Rigid System 343
and
b(u, p) = − div(u)p dx ∀u ∈ H 1 (O)2 , ∀ p ∈ L02 (O).
O
and
dist (B(ζ (t)), ∂ O) > 0 ∀t ∈ [0, T ]. (17)
Remark 2. The hypotheses (16) and (17) imply the existence of η > 0 such that
By using the weak formulation (13), (14), let us derive a semi-discrete version of
our system. For N ∈ N∗ we denote Δ t = T /N and tk = kΔ t for k = 0, . . . , N. Denote
by (uk , ζ k ) ∈ KD(ζ k ) ∩C0 (O)2 × O the approximation of the solution of (1)–(8) at
the time t = tk . From now on we shall use the notation
0
X(x) 0 (tk ;tk+1 , x) ∀x ∈ O.
=ψ
We approximate the position of the rigid ball at instant tk+1 by ζ k+1 which is
defined by
ζ k+1 = ζ k + uk (ζ k )Δ t.
We then define the characteristic function ψ associated with the semi-discretized
velocity field as the solution of
⎧
⎨ d ψ (t;t , x) = uk (ψ (t;t , x)) − uk (ζ k ) ∀t ∈ [tk ,tk+1 ],
k+1 k+1
dt (18)
⎩
ψ (tk+1 ;tk+1 , x) = x − uk (ζ k )Δ t
344 J. San Martín et al.
and we denote
k
X (x) = ψ (tk ;tk+1 , x) ∀x ∈ O. (19)
In (18), the velocity field is extended by zero outside of the domain O.
uk
We next define uk+1 ∈ KD(ζ k+1 ) as the solution of the following Stokes type
system:
k+1 − uk ◦ Xk
k+1 u
ρ , ϕ + a(uk+1 , ϕ )
Δt
= (ρ k+1 fk+1 , ϕ ) ∀ ϕ ∈ KD(ζ k+1 ), (20)
Equation (20) can be rewritten by using a mixed formulation. It is clear that (20)
is equivalent to the following system:
k
uk+1 − uk ◦ X
ρ k+1 , ϕ + a(uk+1 , ϕ ) + b(ϕ , pk+1 )
Δt
= (ρ k+1 fk+1 , ϕ ) ∀ ϕ ∈ K (ζ k+1 ), (21)
Then there exist two positive constants C and τ ∗ not depending on Δ t such that
for all 0 < Δ t τ ∗ the solution (uk , pk , ζ k ) of the semi-discretization problem (21),
(22) satisfies
sup |ζ (tk ) − ζ k | + u(tk ) − uk L2 (O)2 CΔ t.
1kN
The key features used in the proof of the above theorem are some of the properties
of the characteristic functions associated with the semi-discretized velocity field
which are given in the following lemma (more details and the complete proof of this
result could be found in [SMSS10b]):
Lemma 2. For any k ∈ {0, . . . , N}, the characteristic function ψ defined in (18),
(19) satisfies the following properties:
k
i) X B(ζ k+1 ) = B(ζ k );
ii) If we extend by ρ f the density field ρ k outside of O, we have
k
ρ k+1 = ρ k ◦ X ;
In order to discretize the problem (21), (22) with respect to the space variable, let
us introduce two families of finite element spaces which approximate the spaces
K (ζ ) and M(ζ ) defined in (10) and (12). To this end, we consider the discretization
parameter 0 < h < 1.
Let Th be a quasi-uniform triangulation of the domain O. We denote by Wh
the P1 -bubble finite elements space associated with Th for the velocity field in the
Stokes problem and by Eh the P1 -finite elements space for the pressure. Then, we
define the following finite elements spaces for a conforming approximation of the
fluid–rigid system:
Kh (ζ ) = Wh ∩ K (ζ ) ∀ζ ∈ O,
Mh (ζ ) = Eh ∩ M(ζ ) ∀ζ ∈ O.
Rh (ζ ) = {∇⊥ ϕh : ϕh ∈ Fh , ϕh = 0 on ∂ O} ∩ K (ζ ) ∀ζ ∈ O,
346 J. San Martín et al.
− ∂ ϕ /∂ y
where ∇⊥ ϕh = ∂ ϕ h/∂ x .
h
We denote P(ζ ) the orthogonal projection from L2 (O)2 onto Rh (ζ ), i.e. if u ∈
L (O)2 then P(ζ )u ∈ Rh (ζ ) such that (u − P(ζ )u, rh ) = 0 for all rh ∈ Rh (ζ ).
2
by
ζ k+1
h = ζ kh + ukh (ζ kh )Δ t.
We now consider the approximated characteristic function ψ kh defined as the so-
lution of
⎧
⎨ d ψ k (t;tk+1 , x) = P(ζ k )uk (ψ k (t;tk+1 , x)) − P(ζ k )uk (ζ k ) ∀t ∈ [tk ,tk+1 ],
dt h h h h h h h
(23)
⎩ k
ψ h (tk+1 ;tk+1 , x) = x − uh (ζ h )Δ t
k k
and we define
k
Xh (x) = ψ kh (tk ;tk+1 , x) ∀x ∈ O. (24)
We observe that since div (P(ζ kh )ukh (ψ kh (t;tk+1 , ·))−P(ζ kh )ukh (ζ kh )) = 0 and ∇(x−
ukh (ζ kh )Δ t) = Id,
we get det Jψ k = 1.
h
We now split the mesh into the union of four different types of triangle’s subsets.
We first introduce Ah as the union of all triangles intersecting the ball B(ζ kh ), i.e.
;
Ah = T.
T ∈Th
◦ ◦
T ∩B(ζ k )=0/
h
We denote by Qh the union of all triangles such that all their vertices are contained
in Ah . The triangles of Th are then split into the following four categories:
• F1 is the subset of Th formed by all triangles T ∈ Th such that T ⊂ B(ζ kh ).
• F2 is the subset formed by all triangles T ∈ Th \ F1 such that T ⊂ Qh .
• F3 is the subset formed by all triangles T ∈ Th such that T ∩ Qh = 0/ and T ⊂ Qh .
• F4 = Th \ (F1 ∪ F2 ∪ F3 ).
We introduce two approximated density functions ρhk and ρ kh as follows:
1 1
ρs if x ∈ B(ζ kh ), ρs if x ∈ Qh ,
ρh (x) =
k
ρ h (x) =
k
ρ f if x ∈ O \ B(ζ h ),
k ρ f if x ∈ F4 .
With these notations, we consider the following mixed variational fully discrete
h ) ∈ Kh (ζ h ) × Mh (ζ h ) such that
formulation: Find (uhk+1 , pk+1 k+1 k+1
Discretization Scheme for a Fluid–Rigid System 347
k
k+1 uh − uh ◦ Xh
k+1 k
ρh , ϕ + a(uk+1
h , ϕ ) + b(ϕ , ph )
k+1
Δt
= (ρ k+1
h fh , ϕ ) ∀ϕ ∈ Kh (ζ h ), (25)
k+1 k+1
b(uk+1
h , q) = 0 ∀q ∈ Mh (ζ k+1
h ), (26)
where fk+1
h is the L2 (O)2 -projection of fk+1 = f(tk+1 ) on (Eh )2 .
We can now state the second result of this chapter concerning the convergence of
the fully discrete scheme (25), (26) (for the proof of this result, we refer the reader
to [SMSS10b]):
Theorem 2. Let O be a convex domain with a polygonal boundary. Suppose that
f and u0 satisfy the conditions (15) and that (u, p, ζ , ω ) is a solution of (1)–(8)
satisfying the regularity properties (16) and such that (17) holds. Let C0 > 0 be a
fixed constant.
Then there exist two positive constants C and τ ∗ independent of h and Δ t such
that for all 0 < Δ t ≤ τ ∗ and for all h ≤ C0 Δ t 2 we have
sup |ζ (tk ) − ζ kh | + u(tk ) − ukh L2 (O)2 ≤ CΔ t.
1≤k≤N
The key to the proof of the previous convergence result are the properties of the
characteristic functions associated with the fully discretized velocity field given in
the following lemma (the proof of this result is analogous to the proof of Lemma 2
and could be found in [SMSS10b]):
Lemma 3. For any k ∈ {0, . . . , N} and h ∈ (0, 1), the characteristic function ψ kh
defined in (23), (24) satisfies the following properties:
k
(i) Xh B(ζ hk+1 ) = B(ζ kh );
k
(ii) If we extend by ρ f the density field ρhk outside of O, then ρhk+1 = ρhk ◦ Xh ;
(iii) For any f ∈ L2 (R2 ) such that f = 0 in R2 \ O, we have
f ◦ ψ k (t;tk+1 , ·) 2 2 ≤ f 2 2 ∀t ∈ [tk ,tk+1 ].
h L (O) L (O)
Acknowledgements J. San Martín was partially supported by Grant Fondecyt 1090239 and
BASAL-CMM Project. J.-F. Scheid gratefully acknowledges the Program ECOS-CONICYT (Sci-
entific cooperation project between France and Chile) through grant C07-E05. L. Smaranda was
partially supported by Grant RP-2, no. 6/01.07.2009 of CNCSIS-UEFISCSU.
References
[AcGu00] Achdou, Y., Guermond, J.-L.: Convergence analysis of a finite element projec-
tion/Lagrange–Galerkin method for the incompressible Navier–Stokes equations.
SIAM J. Numer. Anal., 37, 799–826 (2000).
348 J. San Martín et al.
1 Introduction
In the last few years special attention has been devoted to searching analytic so-
lutions for the diffusion equation. We are aware of literature for this sort of so-
lution for specialized topics dealing with the simulation of pollutant dispersion in
the atmosphere. For illustration we cite the works of [BuViMoTi07, BuViMoTi07a,
MoViBuTi09]. On the other hand, the literature is scarce regarding analytical so-
lutions for the neutron diffusion equation, except for very specialized problems
[MaSaCaRoAnOlLe07]. Work on analytical solutions to the one-dimensional and
two-dimensional two-group neutron diffusion equation for either homogeneous or
heterogeneous sheets by the well-known GITT technique [MoViBuTi09] has re-
cently emerged in the literature. The key feature of this methodology is that is it
uses an expansion of the fast and thermal fluxes in a series written in terms of a set
of orthogonal eigenfunctions. Replacing these expansions in the original equation
and taking moments, yields a second-order matrix differential equation, known in
the framework of this methodology as the GITT transformed problem.
At this point, we shall recall that the solution of this sort of problem is given
by a linear combination of the sine and cosine functions of the square root of a
M. Schramm
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: marceloschramm@hotmail.com
C.Z. Petersen
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: claudiopetersen@yahoo.com.br
M.T. Vilhena
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: vilhena@pq.cnpq.br
B.E.J. Bodmann
Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil,
e-mail: bardo.bodmann@ufrgs.br
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 349
DOI 10.1007/978-0-8176-8238-5_32, © Springer Science+Business Media, LLC 2011
350 M. Schramm et al.
matrix. This means a function composed of two matrix functions. Efficient algo-
rithms to evaluate this type of solution are not available in the literature except for
very specific problems, mainly the ones requiring low-order matrices. Therefore
the standard procedure which is usually adopted to solve this second-order matrix
differential equation has been to reduce the order of this equation to a system of
first-order matrix differential equations which has a well-known solution. For more
details see elsewhere [MoVa78]. The drawback of this procedure relies on the fact
that the resulting matrix has its order doubles. This imposes a limitation on the abil-
ity of this technique to handle problems requiring high-order matrices. In order to
circumvent this difficulty, we report an efficient approach from the computational
point of view. After diagonalizing the matrix, we define a new variable in such a
way that the matrix appearing in the new transformed equation is diagonal. As a
consequence this new system of equations is a set of uncoupled equations which
can be solved using standard results for second-order linear differential equations
with constant coefficients. Once the solution of the transformed problem is known,
the solution of the neutron diffusion equation for a homogeneous rectangle can be
determined. This solution technique can be extended for the heterogeneous sheet,
following the idea in the work of [BoViFeBa10].
Finally, we complete our analysis by reporting on a criterion of convergence for
the proposed solution using the Cardinal Theorem of Interpolation Theory, which
permits us to obtain results with any prescribed accuracy. Moreover the Cauchy–
Kowalewski theorem guarantees the existence and uniqueness of the obtained solu-
tion, so that a finite number of terms in the series solution gives a suitable and robust
algorithm to generate exact results to within a chosen accuracy. We shall remark that
this sort of solution is appropriate for validating the physical model. The question of
benchmarks for numerical approaches becomes obsolete once an analytical solution
is available.
2 Mathematical Formulation
∂ 2 φg (x, y) ∂ 2 φg (x, y)
−Dg − Dg + ΣRg φg (x, y)
∂x 2 ∂ y2
2 2
1
= χg ∑ νΣ f g φg (x, y) + ∑ Σgg φg (x, y) , (1)
keff g =1 g =1
g =g
φg (0, y) = φg (M, y) = 0,
φg (x, 0) = 0,
∂ φg φg (x, L)
(x, L) = − ,
∂y d
where d is the extrapolated distance. Here g is the index that denotes the energy
group (1 for fast, 2 for thermal); φg (x, y) is the neutron scalar flux of group g of
energy; Dg is the diffusion coefficient of group g of energy; ΣRg is the macroscopic
removal cross section of group g of energy; Σag is the macroscopic absorption cross
section of group g of energy; Σgg is the macroscopic scattering cross section of
group g of energy to group g of energy; Σ f g is the macroscopic fission cross section
of group g of energy; keff is the effective multiplication factor; ν is the average
number of neutrons liberated by fission; χg is the integrated fission spectrum of
group g of energy and ΣRg = Σag + Σgg .
In order to solve (1) by the GITT approach, we initially expand the fast and the
thermal flux in a series expansion in terms of a set of orthonormal eigenfunctions
ψi (x) = sin(λi x) with the respective eigenvalues λi = iπ /M for i = 1, . . . , Nmax :
Nmax
ψi (x)ϕgi (y)
φg (x, y) = ∑ 1 . (2)
i=1 Ni2
which is known as the GITT transformed equation. Here Y (y) is the column vector
Y (y) = (ϕ11 (y) . . . ϕ1N (y)ϕ21 (y) . . . ϕ2N (y))T and the matrix U has the form
⎛ ⎞
A − λ12 0 B 0
⎜ 0 A − λ22 0 B ⎟
⎜ ⎟
⎜ .. .. ⎟
⎜ . . ⎟
⎜ ⎟
⎜ A − λQ2 B ⎟
U =⎜
⎜ C
⎟,
⎟
⎜ 0 D − λ12 0 ⎟
⎜ 0 C 0 D − λ22 ⎟
⎜ ⎟
⎜ .. .. ⎟
⎝ . . ⎠
C D − λQ2
νΣ νΣ
where A = D1 kf 1 − ΣDR11 , B = D1 kf 2 , C = ΣD122 and D = − ΣDR22 . Recalling that the eigen-
eff eff
values of matrix U are distinct since the operator associated to the neutron diffusion
equation is self-adjoint, the next step is to diagonalize the matrix U to obtain
352 M. Schramm et al.
where P in the matrix of the eigenvectors of the U matrix, P−1 is its inverse and D
is the diagonal matrix of the eigenvalues of the U matrix. Defining a new available
R(y) = P−1Y (y), (3) becomes
Recalling that the matrix D is diagonal, the matrix equation (4) reduces to a set of
uncoupled linear second-order differential equations with constant coefficients,
rgi (y) + γi2 rgi (y) = 0
for g taking values 1 and 2 and for i = 1, . . . , Nmax . The well-known solution to this
equation is
where αgi and βgi are the integration constants. Applying the same series expansion
for the boundary conditions, we come out with identical boundary conditions for
the ϕgi (y). Therefore, the solutions of the neutron diffusion equation for the fast and
thermal neutron are completely determined by the application of the boundary con-
ditions and evaluating Y (y) as Y (y) = PR(y). Although this solution is determined
for homogeneous rectangle, its extension to heterogeneous problem can be easily
done following the procedure of [BoViFeBa10].
To show the efficiency of the proposed algorithm and to evaluate the solution,
in what follows we discuss a criterion for convergence with a prescribed accuracy.
In fact, the cardinal theorem of interpolation theory states: “A square-integrable
function
ω = ψi (x)dx ∈ L2
r
3 A Case Study
In Table 2, we report the fast and thermal neutron fluxes values given by the
proposed solution with an relative error of < 1% and by summing just two terms
in the solution series at symmetric points in the interior of rectangle including its
center. In Fig. 1 we display the fast neutron flux which shows, as expected, that the
value of the neutron flux takes its maximum at the center and decreases to near to
zero at the boundary.
Table 2 Values of the fast and thermal neutron fluxes at some points of the domain
Point of the domain Fast neutron flux Thermal neutron flux
(x, y) [N/cm2 s] [N/cm2 s]
(50, 50) 0.197212 0.167536
(50, 100) 0.268498 0.228091
(50, 150) 0.168486 0.142740
(100, 50) 0.278900 0.236932
(100, 100) 0.379713 0.322570
(100, 150) 0.238275 0.201865
(150, 50) 0.197212 0.167536
(150, 100) 0.268498 0.228091
(150, 150) 0.168486 0.142740
4 Concluding Remarks
This paper presented an efficient method which generates analytical solutions for
neutron diffusion in two dimensions. In the future these solutions will be incorpo-
354 M. Schramm et al.
rated into a program library and serve as generic solutions for multi-group neutron
flux in a multi-region geometry. The principal steps employed are the generalized in-
tegral transform technique and diagonalizing the equation system. We note that exis-
tence and uniqueness are guaranteed by the Cauchy–Kowalewsky theorem. Further,
the convergence of the solution is under control by a new interpretation of the Cardi-
nal Theorem of Interpolation theory, where the macroscopic cross section plays the
role of a sampling density and the reconstruction of the solution follows the com-
mon procedure in signal processing as originally introduced by Shannon [Go05].
Thus it is fair to say that we have found an analytical solution to the problem. For
any numerical implementation one only needs the desired precision which may be
immediately transformed into a truncation index of the solution series.
The only task to be which needs to be carried out for applications is to deter-
mine numerically the GITT eigenvalues and to substitute the physical parameters
and boundary conditions into the stored solutions which may then be calculated di-
rectly. In order to get comparable precision, numerical or stochastic procedures will
be more time consuming because they have to execute a numerical algorithm for
each energy group and region, respectively. This is a clear advantage, if the influ-
ences of modifications in geometry and material composition on the solution are to
be examined, which in the present procedure may be carried out in an analytical
fashion.
References
1 Introduction
There have been many conventional approaches to discretize and solve time-depend-
ent wave equations on Eulerian grids. For short pulses or waves, these are, of course,
limited by the requirement that a sufficient number of grid cells must span the pulse
to accurately solve the equations, and that the propagation distance cannot be very
long compared to the pulse width. (Our method applies equally well to pulses and
waves and we will use the term “pulse width” to also denote the wave length.)
These approaches to the wave equation problem involve, as usual, formulating the
governing partial differential equations (pde’s), discretizing them and solving them
as accurately as possible on feasible computational grids, assuming smooth enough
solutions. For smooth, long pulses that span over many grid cells, these methods
are well known to converge to the correct solution as the number of points across
the pulse becomes large [1]. Even then, solutions degrade over moderate propaga-
tion distances (a number of pulse widths). As a result, they will not be feasible for
most problems involving thin pulses convecting over moderate distances. Further,
adaptive, unstructured grids cannot significantly improve the solution for realistic
problems with many thin, time-dependent pulses (or intersecting waves). In such
cases, adding grid points in many locations will only result in long computation
time and complexity.
J. Steinhoff
University of Tennessee Space Institute, Tullahoma, TN, USA,
e-mail: jsteinho@utsi.edu
S. Chitta
Flow Analysis, Inc., Brooklyn, NY, USA,
e-mail: subha@flowanalysis.com
P. Sanematsu
University of Tennessee Space Institute, Tullahoma, TN, USA,
e-mail: paulasan@gmail.com
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 357
DOI 10.1007/978-0-8176-8238-5_33, © Springer Science+Business Media, LLC 2011
358 J. Steinhoff et al.
Another approach that has been widely used for this purpose is Lagrangian ray
tracing, which requires complex interpolation routines to accommodate the addition
of markers (when wave surfaces propagate and expand, reducing resolution). Un-
like ray tracing, our new Eulerian finite difference method, “Wave Confinement”
(WC), uses a continuous phase field, which can automatically compute multiple
waves. WC allows propagation of short pulses as coherent phase surfaces, for in-
definitely long distances, without numerical dissipation. Such accurate propagation
will greatly extend the potential of WC to simulate many important wave propaga-
tion cases (with minimum computational cost and effort), including over the horizon
(OTH) radar and cell phone communications, with effects such as reflections from
complex terrain and buildings, as well as ionospheric and atmospheric refraction.
Also, it will allow for the inclusion of turbulence models, interference and disper-
sion for synthetic aperture radars (SAR) propagation.
WC is based on a previously successful technique, “Vorticity Confinement”,
which allows the propagation of thin, concentrated vortices over arbitrarily long
distances, and keeps many of the Eulerian finite difference properties of the origi-
nal fluid dynamic solution method [2, 3]. Both methods employ nonlinear solitary
waves based on a regular lattice and use the same new idea to generate discrete so-
lutions for these thin features. The basic scheme involves an “inner” and an “outer”
solution as shown in Fig. 1. The inner solution represents the very thin, physical
pulse. (It could be only several centimeters wide.) This obviously could not be rep-
resented, in general, on a computational grid spanning several thousand kilometers,
nor does it have to. We take an approach similar to shock capturing in compressible
flow. There, the actual shock can be only millimeters wide, but the computational
shock is several grid cells wide. In our case, the computational pulse may be sev-
eral grid cells wide, and hence span several kilometers which is not a problem as
long as it is very thin compared to the overall dimension of the problem, (∼1000
km), and as long as it carries the correct total energy (and other important integrated
quantities) and the arrival time (or speed) is correct.
As in shock capturing, there are a few cases (such as combustion in fluid dy-
namics and nonlinear effects in waves) where details of the internal structure must
be modeled or solved for, but for most cases the computed “captured” shock is ad-
equate. Similar scenarios occur in pulse propagation where nonlinear effects may
Nonlinear Localized Dissipative Structures 359
be important. For many problems where thin, concentrated pulses or wave fronts
must be numerically propagated over long distances, the main interest is in the far
field, where the amplitude integrated normally through the pulse at each point along
the pulse surface, together with the motion of the centroid surface (which we use
to represent the wave fronts) are important, rather than the details of the internal
structure.
The basic idea of WC is to treat thin features, such as short waves, as a type of
“weak solutions” of the governing pde. Basically, the features are treated as multi-
dimensional nonlinear solitary waves that “live” on the computational lattice, while
preserving the essential physics of the pulse. This approach has been taken in con-
densed matter physics for some years [4] where the computational objects have
these necessary properties, and are known as “nonlinear diffusive solitary waves”
[5]. The basic point is that if we can formulate a pde such that the solutions are
qualitatively what we desire (thin surfaces), propagate with correct speed, conserve
the desired quantities and are stable to the inevitable perturbations from the discrete
computation, then they will serve as useful “basis functions”.
As a simple example, we start with the 1-D scalar wave equation with constant
wave speed, c,
∂t2 φ = c2 ∂x2 φ . (1)
Evidently, when the above equation is numerically solved using conventional
discretization schemes based on Taylor expansions, there will be numerical errors,
which will grow for linear discretization. When we confine the pulse solution to 2–3
grid cells, which is our goal, the derivatives of φ and hence these “errors” will be
large. Also, corresponding to the small number of grid points within the pulse, there
will be only a small number of quantities that we can conserve. However, this is
sufficient for our goal. Adding a term
∂t ∂x2 F(φ )
to (1) that vanishes at the boundaries, along with sufficient derivatives, will not affect
the conservation of these quantities, which include the total amplitude and centroid
speed. We then have
∂t2 φ = c2 ∂x2 φ + ∂t ∂x2 F. (2)
To preserve the essential physics of (1) for a short convecting pulse, F should be
homogeneous of degree one, like (1), so that it does not depend on the magnitude
of φ . This is an important distinction of the WC equation. Many nonlinear equations
use non-homogeneous terms for a nonlinear term [6]. In fact, Cahn and Hilliard also
used a nonlinear term under a second derivative as in our equation, but one that
360 J. Steinhoff et al.
was not homogeneous, and represented a model for the free energy for multi-phase
phenomena [7].
Equation (2) produces a stable solution, which can be easily derived in the con-
vecting frame of the pulse. The pde then becomes the heat equation
∂t φ = ∂x2 F (3)
and
α 4 2 5
F= ∂ ψ −λψ (4)
ψ2 x
where ψ = φ −1 . The above form of F is proven to produce stable solutions and is
described in [8, 9]. When (3) converges, the pulse relaxes to the form
φ → φ0 sech [γ (x − ct)]
√
where γ = λ and φ0 is an arbitrary constant.
One possible discretized formulation of the pde given in (2) can be written as
δn2 φ = ν 2 δ j2 φ + a δn δ j2 F (5)
The above sum is over a set of N neighboring grid nodes near and including the
node where Φ is computed. A small positive constant (δ ∼ 10−8 ), is added with the
Nonlinear Localized Dissipative Structures 361
sign of φ n to prevent problems due to finite precision. The above term can be used
in multi-dimensional formulations with the sum in (6) taken over nearest neighbors.
If ρ = 1, Φ is the harmonic mean of φ on the local stencil. Other forms could also
be used, however, we use only ρ = 1 or ρ = 2.
The two parameters, ε and μ (> 0), are determined by the two small scales of
the computation, h and Δ t, since we want the small features to relax to their solitary
wave shape in a small number of time steps and to have a support of a small number
of grid cells. Thus, even though h may be small, the Laplacian will be large and so
will the total effect. At convergence, the solution to the above equation that vanishes
in the far field is then (modulo a small variation), φ → φ0 sech [γ ( j − j0 ± ν n)], where
j0 is an arbitrary constant (since (5) is translation invariant). With propagation in a
smooth external field, this relation is still approximately satisfied, as verified by
computations and heuristic arguments [11].
The main constraint on the confinement term, F, as in advection, is that it forces
an initially isolated, short-range pulse that propagates with a single maximum to
remain short range and also not develop any additional maxima. The behavior of a
solution computed using (5) is shown in Fig. 2. When μ = 0 and ε = 0, the solution
becomes unstable. A small amount of diffusion (μ = 0.2) will prevent the unstable
behavior but the pulse quickly spreads, rendering the method useless except for short
distance computation, especially with multiple waves. This problem can be resolved
by adding the confinement term. The dissipative/dispersive effects are automatically
balanced with those of the nonlinearity to produce stable localized structures [12].
The pulse then remains thin for a range of values of ε .
Nonlinear equations would ordinarily introduce phase shifts into the solution
when the waves pass through each other. However, a very important feature of WC
is that the waves do not suffer a “phase shift” when they pass through each other
in spite of its nonlinearity. This is an obvious requirement for the equation we want
to simulate—the linear wave equation. Even though the phase shift is a 1-D effect,
362 J. Steinhoff et al.
such a phase shift would generally show up as a kink in two waves in 2 or 3 dimen-
sions that are passing through each other. Results for the centroid trajectories for
two pulses passing through each other (in 1-D) are presented in Fig. 3. There, the
computed centroids are plotted for forward and backward propagating waves with
periodic boundary conditions. It can be seen that there is no phase shift, to plottable
accuracy.
This lack of nonlinear interaction is due to the fact that both the Laplacian and
time derivative operator operate on the nonlinear term which cause the change of
integrals of interest (amplitude and centroid) to vanish in the interaction region.
For 3-D, we simply substitute a multi-dimensional Laplacian into the original wave
equation and use a multi-dimensional harmonic mean, where we sum inverse val-
ues of φ over each point and the N immediately neighboring grid points on the
multi-dimensional grid. An expanding spherical wavefront in a cube with periodic
Nonlinear Localized Dissipative Structures 363
3 Carrier Function
Developments that allow wave fronts to capture and propagate other variables, in ad-
dition to total amplitude and centroid speed, are described in this section. WC can be
used to generate constant arrival time (eikonal phase) surfaces accurately by storing
the centroid arrival time at each grid point. These variables can be required to com-
pute interference. Also, multiple arrival times can be easily accommodated, which
is difficult using eikonal schemes. Some recent developments in eikonal methods
[13] can treat multiple arrival times, but these methods require extra independent
variables and complex data management schemes are used to control memory re-
quirements. For WC, it is required that the wave front pulse has almost completely
passed the grid point to accurately compute these properties. For very short pulses,
or wavelengths, it becomes difficult to accommodate multiple arrival times when the
waves are close to each other. To overcome this problem, a scalar field, φ , is used as
a carrier function that can act as wave packet, which carries the required details of
the propagating quantity. It is then not necessary for a pulse to entirely pass a grid
point to capture the properties. The “computed” scalar pulse, which propagates ac-
cording to the actual wave equation without numerical errors, can be used to “carry”
(not propagated by themselves, but as a product of actual scalar and the variable of
interest) an array of variables which will have the same propagation path. These
variables can be used to describe additional properties of the pulse or wave, such as
the propagation direction vector, actual (sub-grid scale) thickness, etc. Propagation
of directions has proven to be very beneficial in capturing diffraction.
For example, to propagate the directions in 2D, the wave equation is solved for
three quantities, φ , φ1 and φ2 , as
δn2 φln = ν 2 ∇2 φln + μδn− ∇2 φln − εδn− ∇2 Φln
where
φ1 = φ , φ2 = sx φ , φ3 = sy φ ,
and sx and sy are the normalized components of the propagating direction. The up-
dated propagation vector is then
φ2n+1 φ3n+1
sxn+1 = 2 2 , syn+1 = 2 2 .
φ2n+1 + φ3n+1 φ2n+1 + φ3n+1
In the test computation described, the initial conditions for φ are given as
364 J. Steinhoff et al.
A0
φi,n=0
j = A0 sech [γ ri, j ] , φi,n=1
j = sech [γ (ri, j − ν )]
A1
with initial directions,
i − i0 j − j0
sx = , sy =
ri, j ri, j
where the radius of the wave is ri, j = (i − i0 )2 + ( j − j0 )2 with origin at the center
of the domain (i0 , j0 ). The parameters ν , μ , and ε are in the same range for all
three quantities. Here, ν = 0.23, μ = 0.2, and ε = 0.3. φ1 and φ2 behave as the
scalar function itself during wave collision. In Fig. 5, directions for the thin wave
are shown at four different time steps. During interaction, directions temporarily are
the mean values of the waves that are together. Then, they subsequently take their
original forms after a short relaxation time.
4 Continuous Waves
WC has previously proven to be efficient for pulse-like wave propagation and has
been recently extended to accurately simulate propagation of continuous waves
(CW’s). A constant is first added to the amplitude so that it is non-negative. This
addition simplifies the algorithm to study periodic waves with constant and varying
amplitude. As a first example, we take an initial condition
Nonlinear Localized Dissipative Structures 365
1
2π j
2 + sin , 5λs ≤ j ≤ 15λs
φ j0 = φ j1 = λs
2, otherwise
where λs is the wave length. The wave was constrained to 5λs ≤ j ≤ 15λs because
of the use of periodic boundary conditions. When the wave spans the entire domain,
we cannot clearly see the interactions between the two traveling waves because they
result in a single standing wave. It can be seen in Fig. 6 that WC propagates these
periodic waves for many time steps with no plottable numerical dissipation and no
phase shift.
Fig. 6 Propagation and interaction of sinusoidal wave with λs = 50, ν = 0.23, μ = 0.2, and ε =
0.205
To avoid effects due to the symmetry of waves with constant amplitude, a case
was run with periodic waves of varying amplitude. Results are shown in Fig. 7 for
the initial condition
366 J. Steinhoff et al.
1
2π j 2π j
2 + sin 3
+ 1
cos , 5λs ≤ j ≤ 15λs
φ j0 = φ j1 = λs 4 4 λc
2, otherwise
where λs and λc correspond to the wave length of the sine and cosine waves, respec-
tively.
An important characteristic of WC for continuous waves is that the wave main-
tains its initial shape, which can further help in capturing interference and disper-
sion effects. As well as for applications for single pulse propagation, WC uses a
very simple scheme to propagate continuous waves over long distances with min-
imal dissipation. This can be a major advantage over higher order schemes when
simulating propagation of electromagnetic waves in satellite, cell phone and OTH
communications.
Fig. 7 Propagation and interaction of periodic wave with varying amplitude and λs = 25, λc = 125,
ν = 0.23, μ = 0.2, and ε = 0.205
Nonlinear Localized Dissipative Structures 367
addition of new rays and interpolation. A comparison is shown in Fig. 8 in which the
smooth contours are calculated by the confinement method and compared with ray
tracing (depicted as circular “blobs”). It can be seen that in the far field, ray tracing
techniques cannot continue to describe the wave as a smooth surface without adding
markers and interpolation. Parabolic equation solvers are widely used to propagate
flow over both ionospheric layers and atmospheric ducts. These solvers are usually
paraxial and can only be used to capture waves that are very close to the propagation
direction. Many developments have been made to allow the paraxial equation to
include much wider angles, but they involve complicated logic [14]. In such cases,
WC can easily accommodate propagation through any angle in one computation and
does not need any complex logic to include multiple waves.
References
1. Anderson, D., Tannehill, J., Pletcher, R.: Computational Fluid Mechanics and Heat Transfer,
Hemisphere, New York (1984).
2. Steinhoff, J., Lynn, N., Wang, L.: Large eddy simulation using vorticity confinement, in: Im-
plicit Large Eddy Simulation: Computing Turbulent Flow Dynamics (Editors: F.F. Grinstein,
L.G. Margolin, W.J. Rider), Cambridge University Press (2006).
3. Steinhoff, J., Lynn, N., Wang, L., Wenren, Y., Fan, M.: Turbulent flow simulations using vor-
ticity confinement, in: Implicit Large Eddy Simulation: Computing Turbulent Flow Dynamics
(Editors: F.F. Grinstein, L.G. Margolin, W.J. Rider), Cambridge University Press (2006).
4. Bishop, A., Krumhansl, J., Trullinger, S.: Solitons in condensed matter: A paradigm. Physica
D: Nonlinear Phenomena, 1, n. 1, 1–44 (1980).
5. Kuramoto, Y., Mori, H.: Dissipative Structures and Chaos, Springer-Verlag (1998).
6. Rosenau, P., Hyman, J., Staley, M.: Multidimensional compactons. Physical Review Letters,
98, n. 2, 24101 (2007).
7. Cahn, W., Hilliard, J.: Free energy of a nonuniform system. I. Interfacial free energy. Journal
of Chemical Physics, 28, n. 2, 258 (1958).
8. Steinhoff, J., Chitta, S.: Long-time solution of the wave equation using nonlinear solitary
waves, in: Integral Methods in Science and Engineering: Computational Aspects (Editors:
C. Constanda, M. Perez), Springer Verlag (2009).
9. Steinhoff, J., Chitta, S.: Long distance wave computation using nonlinear solitary waves. Jour-
nal of Computational and Applied Mathematics, 234, n. 6, 1826–1833 (2010).
10. Peyrard, M., Kruskal, K.: Kink dynamics in the highly discrete sine-Gordon system. Physica
D: Nonlinear Phenomena, 14, n. 1, 88–102 (1984).
11. Steinhoff, J., Haas, S., Xiao, M., Lynn, N., Fan, M.: Simulating small scale features in fluid dy-
namics and acoustics as nonlinear solitary waves, in: Proceedings of the 41st AIAA Aerospace
Sciences Meeting and Exhibit, (January 2003).
12. Lax, P.: Hyperbolic systems of conservation laws II. Communications on Pure and Applied
Mathematics, 10, n. 2, 537–566 (1957).
13. Vinje, V., Iversen, E., Gjoystdal, H.: Traveltime and amplitude estimation using wavefront
construction. Geophysics, 58, n. 8, 1157–1166 (1993).
14. Bamberger, A., Engquist, B., Halpern, L., Joly, P.: Higher order paraxial wave equation ap-
proximations in heterogeneous media. SIAM Journal on Applied Mathematics, 48, n. 1, 129–
154 (1988).
Semianalytical Approach to the Computation
of the Laplace Transform of Source Functions
1 Introduction
The method of “Sources and Sinks” has has found wide application to the solution
of the diffusion equation for linear conduction of heat in infinite, semi-infinite and
finite solids; see, for example, [Carslaw88]. Since reservoir flow of a slightly com-
pressible, constant viscosity fluid is governed by the diffusion equation, fundamen-
tal solutions of the one-dimensional diffusion equation are routinely used to build
solutions to three-dimensional flow problems using the Newmann product method
[Newmann:36]. These solutions can be rapidly generated for constant-rate produc-
tion (or injection) to a uniform-flux line or plane in an anisotropic homogeneous
reservoir; (see for example, [Grin:73]). However, for variable-rate, variable-flux or
complex-geometry (heterogeneous reservoir) problems, solutions are most easily
generated by manipulating Laplace transforms of the product solutions and invert-
ing to the time domain using a numerical inversion procedure such as the Stehfest
algorithm [STEHFEST:70]. Raghavan and Ozkan [RagOzkan94] presented com-
plete analytical expressions for the Laplace transforms of the 3-D source functions.
For flow in rectangular parallelepiped reservoirs, their expressions involve triple in-
finite series; for example, the Laplace transform of the dimensionless fundamental
solution (γ̄ ) for flow in a sealed rectangular parallelepiped reservoir was given as
(see Section 2.4, Eq. 1 of [RagOzkan94])
∞ ∞ ∞
γ̄ = ∑ ∑ ∑ {S1,1,1 + S2,1,1 + S1,2,1
k=−∞m=−∞n=−∞
+S2,2,1 + S1,1,2 + S2,1,2 + S1,2,2 + S2,2,2 } (1)
L.G. Thompson
The University of Tulsa, OK, USA,
e-mail: lgt@utulsa.edu
G. Zhao
University of Regina, SK, Canada,
e-mail: gang.zhao@uregina.ca
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 369
DOI 10.1007/978-0-8176-8238-5_34, © Springer Science+Business Media, LLC 2011
370 L.G. Thompson and G. Zhao
where
√
exp − u (x̃Di − 2kxeD )2 + (ỹD j − 2myeD )2 + (z̃Dl − 2nzeD )2
Si, j,l = (2)
(x̃Di − 2kxeD )2 + (ỹD j − 2myeD )2 + (z̃Dl − 2nzeD )2
for i, j, l = 1 or 2. Here u is the dimensionless Laplace variable, and the terms with
subscript D represent various dimensionless distances in the x-, y- and z-directions.
Note that if we were to use the Stehfest algorithm to numerically invert (1) to the
(dimensionless) time domain, we would have to evaluate (2) at real values of the
Laplace variable that could vary widely in magnitude. In particular, for very small
values of u, many terms would have to be evaluated for the triple series of (1)
to converge. In order to enable greater efficiency in their numerical computation,
Raghavan and Ozkan simplified their triple infinite sums to equivalent expressions
involving the sums of single and double infinite series. Even with these simplifica-
tions, it is not clear that the resulting series always converge rapidly. In this work,
we explore an alternative method for efficiently computing the Laplace space func-
tions that involves only the computation of three (single) infinite series. The number
of actual terms involved in the evaluation of each series is small (typically less than
20 terms), so the method is attractive for application to large systems of equations
in Laplace Space.
and
Table 1 Instantaneous 1-D point source functions for various boundary conditions
Reservoir Boundary Conditions Source Function
2
(x−x )
Infinite reservoir = √ 1 exp − 4η (t−τ )
2 πη (t−τ )
2
2
(x−x ) (x+x )
Semi-infinite with constant pressure at x = 0 = √ 1 exp − 4η (t−τ ) − exp − 4η (t−τ )
2 πη (t−τ )
2
2
(x−x ) (x+x )
Semi-infinite with sealing fault at x = 0 = √ 1 exp − 4η (t−τ ) + exp − 4η (t−τ )
2 πη (t−τ )
2
2
n=∞
Semianalytical Laplace Transforms
∞
= ∑ εn cos (2nπ z) exp −π 2 n2t , (4b)
n=0
respectively, where *
1 for n = 0
εn =
2 for n = 1, 2, 3 . . .
We will denote the leading term of (3a) and (4a) (i.e., when n = 0) as θL (z,t) ,
2
1 z
θL (z,t) = √ exp − . (5)
πt t
In terms of these functions, the point source solutions of Table 1 can be written as
shown in Table 2.
Equations (3a), (3b) and (4a), (4b) are each expressed in alternative equivalent
forms. A question now arises as to which form should be used at any particular
value of z and t. For small values of the time group t, the series given by (3b) (or
(4b)) converges very slowly; for small values of n, the exponential term in of (3b)
(or (4b)) is close to 1, and trigonometric terms oscillate in value between ±1. Under
this condition, literally millions of terms may be required to obtain accurate values
of this form of the series. On the other hand, for small values of t, successive terms
in the series given by (3a) (or (4a)) approach zero rapidly as n increases.
The
situation
at large values of t is reversed; for large t, successive values of
exp −π 2 n2t → 0 rapidly as n increases, whereas n must become very large before
2
exp(− (z+n)
t ) → 0.
This behavior suggests the following algorithm for deciding on which series rep-
resentation to use for the source function at a given value of the time group, t. We
would like to use the series representation whose terms disappear at the smallest
value of n. In each of the series forms, the argument of the exponential terms gov-
erns the rate of disappearance of successive terms in the series. For the series given
by (3a) or (4a), the series will converge when
= >
(z + n̂)2
exp − ≤ ε,
t
Similarly, the series given by (3b) (or (4b)) will converge for values of n such that
exp −π 2 ñ2t ≤ ε .
In well testing applications, we are most interested in finding the Laplace Transform
with respect to time of the time integral of products of the previously defined source
functions. That is, a typical unit constant-rate pressure drop solution assumes the
form
t
5.615
Δ pcu (x, y, z,t) = Δ psx Δ psy Δ psz d τ ,
φ ct
0
Each of the source functions in (6) approach a unique functional behavior at late
times. Table 3 summarizes the late-time behaviors of each of the instantaneous
source functions and the time range for which this late-time behavior is valid.
Equation (6) is written as
∗
t
5.615 e−uτ
Δ pcu = Δ psx Δ psy Δ psz dτ +Ψ , (7)
φ ct u
0
where
∞
5.615 e−uτ
Ψ= Δ psx Δ psy Δ psz dτ .
φ ct u
t∗
The value of t ∗ is determined as follows. If the system contains at least one con-
stant pressure boundary in any coordinate direction, then t ∗ is the minimum time at
which any of the source functions approaches zero. If there are no constant pres-
sure boundaries, t ∗ is the minimum time at which all of the source functions attain
their late-time functional forms. In (7), the first integral is determined numerically
Semianalytical Laplace Transforms 375
Table 3 Late time behavior of 1-D point source functions for various boundary conditions
Reservoir Boundary Conditions Late Time For t >
f (t)
2
(x−x )
Infinite system √1
2 πηx t − 4η ln(1−ε )
2 2
(x−x ) (x+x )
Semi-infinite with constant pressure at x = 0 0 max − 4ηx ln(1−ε ) , − 4ηx ln(1−ε )
2 2
(x−x ) (x+x )
Semi-infinite with sealing fault at x = 0 √1
πηx t max − 4ηx ln(1−ε ) , − 4ηx ln(1−ε )
−4 Lπx2lnη ε
2
Constant pressure at x = 0 and x = Lx 0
x
− Lπx2lnη ε
2
1
Sealing faults at x = 0 and x = Lx Lx x
−4 Lπx2lnη ε
2
Constant pressure at x = 0, sealing fault at x = Lx 0
x
−4 Lπx2lnη ε
2
Sealing fault at x = 0, constant pressure at x = Lx 0
x
Table 4 Late time behavior of 1-D point source functions for various boundary conditions
x-Boundaries y-Boundaries z-Boundaries Ψ
E1 (ut ∗ )
Both infinite Both infinite Both finite sealed 5.615 √
φ ct Lx Ly Lz 4π u Λx Λy
E1 (ut ∗ )
Semi-infinite, sealed Both infinite Both finite sealed 5.615 √
φ ct Lx Ly Lz 2π u Λx Λy
E1 (ut ∗ )
Both infinite Semi-infinite, sealed Both finite sealed 5.615 √
φ ct Lx Ly Lz 2π u Λx Λy
E1 (ut ∗ )
Semi-infinite, sealed Semi-infinite, sealed Both finite sealed 5.615 √
φ ct Lx Ly Lz π u Λx Λy
√
∗
5.615 erf c( ut )
Semi-infinite, sealed Both finite sealed Both finite sealed φ ct Lx Ly Lz √ 3
Λx u 2
5.615 exp(−ut ∗ )
Both finite sealed Both finite sealed Both finite sealed φ ct Lx Ly Lz u2
∞
2
e−t dt.
2
erf c (z) = √
π z
where tmin is the value of time before which all of the source functions in the prod-
uct are practically zero. Considering the leading term in the early-time form of the
elliptic theta functions (5), this can only be nonzero on a computer if
2
z
exp − > εmin
t
where εmin is the minimum floating point number that can be represented on a com-
puter (approximately 10−308 on a PC). Thus, each source function will only be com-
putationally nonzero if
−z2
t> . (9)
ln (εmin )
We set tmin to the maximum value of the right side of (9) for each of the source
functions that comprise the integrand.
Since the integrand in (8) exhibits large variations in value over the interval, we
split the interval into log cycles and evaluate the integral as
∗ i
t ncycles 10tmin
e−uτ e−uτ
Δ psx Δ psy Δ psz
u
dτ = ∑ Δ psx Δ psy Δ psz
u
dτ
tmin i=1
10i−1 tmin ∗
t
e−uτ
+ Δ psx Δ psy Δ psz dτ ,
n
u
10 cycles tmin
where
t∗
ncycles = int log10 ;
tmin
here int (x) is the integer part of x without rounding.
Semianalytical Laplace Transforms 377
On each integration interval, we approximate the integrand and the integral us-
ing Chebyshev polynomials (see [Press92]); i.e., if Tn (x) represents the Chebyshev
polynomial of degree n,
where
c j−1 − c j+1
Cj = for j > 0,
2j
and C0 is an integration constant. Details are given in [Press92].
We obtained excellent results using Chebyshev polynomials of maximum degree
32, and using the Stehfest algorithm with Stehfest parameter equal to 16 for the nu-
merical inversion. Example applications of the method can be found in [Zhao:2002].
4 Summary
Acknowledgements This work was sponsored by the Tulsa University Petroleum Reservoir
Exploitation Projects (TUPREP). The authors wish to thank the supporting member companies.
378 L.G. Thompson and G. Zhao
References
[Stegun:72] Abramowitz, M., Stegun I.: Handbook of Mathematical Functions with Formu-
las, Graphs, and Mathematical Tables, United States Department of Commerce,
National Bureau of Standards Applied Mathematics Series 55 (1972).
[Carslaw88] Carslaw, H., Jaeger, J.: Conduction of Heat in Solids, Claredon Press, Oxford
(1988).
[FUNLIB:76] Wayne Fullerton, L.W.: Portable special function routines, in: Portability of Nu-
merical Software (Editor: W. Cowell), Springer-Verlag, New York, 452–483
(1976).
[Grin:73] Gringarten, A.C., Ramey, H.J.: Use of source and Greens functions in solving
unsteady flow problems in reservoirs. Society of Petroleum Engineers Journal,
13, 285–296 (1973).
[Newmann:36] Newmann, A.B.: Heating and cooling rectangular and cylindrical solids, Ind. &
Eng. Chem., 28, 545–548 (1936).
[OBER:73] Oberhettinger, F., Badii, L.: Table of Laplace Transforms, Springer-Verlag, New
York (1973).
[Ohaeri:91] Ohaeri, C.U., Vo, D.T.: Practical solutions for interactive horizontal well test
analysis, Paper SPE 22729 presented at the 66th Annual Technical Conference
and Exhibition of the Society of Petroleum Engineers held in Dallas, TX, Octo-
ber 6–9, (1991).
[Press92] Press, W.H., Teukolsky, S.A., Vetterling, W.T., Flannery, B.P.: Numerical Recip-
ies in c – The Art of Scientific Computing, Cambridge University Press (1992).
[RagOzkan94] Raghavan, R., Ozkan, E.: A Method for Computing Unsteady Flows in Porous
Media, Pitman Research Nodes in Mathematics Series, Longman Scientific &
Technical (1994).
[STEHFEST:70] Stehfest, H.: Numerical inversion of Laplace transforms. Communications of
ACM, 13, n. 1, 47–49 (1970).
[Thompson91] Thompson, L., Manrique, J., Jelmert, T.: Efficient algorithms for computing the
bounded reservoir horizontal well pressure response, Paper SPE 21827, Pre-
sented at the Rocky Mountain Regional Meeting and Low-Permeability Reser-
voirs Symposium Held in Denver Colorado, (April 15–17, 1991).
[Zhao:2002] Zhao, G., Thompson, L.G.: Semi-analytical modeling of complex geometry
reservoirs, SPEREE, ASME (2002).
Asymptotic Analysis of Singularities
for Pseudodifferential Equations in Canonical
Non-Smooth Domains
V.B. Vasilyev
”Although this may seem a paradox, all exact science is dominated by the idea of
approximation.”
Bertrand Russell
1 Introduction
V.B. Vasilyev
Bryansk State University, Russia,
e-mail: vbv57@inbox.ru
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 379
DOI 10.1007/978-0-8176-8238-5_35, © Springer Science+Business Media, LLC 2011
380 V.B. Vasilyev
thor hopes that in the future it will be possible to obtain a “non-smooth” version of
Sternin’s results [St66] by the wave factorization method.
2 Two-Dimensional Case
We will begin with the two-dimensional case. The problem is to find out what does
an operator, which we multiply with a characteristic function which is positive on
the y-axis, represent in Fourier images. Analytically, we have
*
1, x = 0, y > 0,
m (x, y) =
0, otherwise.
It is clear that such a multiplier in Fourier images is convolution with a distribution,
which is homogeneous of degree −2, as long as the function m (x, y) is homogeneous
of degree 0 ([Mi77, Mi62, Mi86, Ma62, CaZy57]), and, consequently, we can think
of it as a function, defined on a unit circle, which is touched by the y-axis at a single
point, the “north pole”.
We can use the standard method: the multiplier m (x, y) “blurs” the angle with
size α , then we find an appropriate distribution (this function has been found already
[BoMa48, Vl64]) and pass to a limit with α → 0.
Remark 1. In the author’s work [Va98] the Gakhov operator G was introduced,
which was defined in the entire complex plane. Of course, it is desirable to de-
fine this operator in real variables, and for this we must cut out the singularity using
a special method. In a Hilbert transform the singularity is cut off by “cross”. Us-
ing a two-dimensional Gakhov operator we probably should use some “hyperbolic”
cut off. This work will not focus on this (Gakhov’s operator works well); maybe
N. Kasumov will deal with this problem as he has some experience [Ka92].
The angle of α is the set {(x, y) : y > a |x| } , a = cot α , and thus the asymptotic
needed is (α → 0), i.e. a → ∞. The distribution, appropriate to this angle multiplier
is the function
1
δ (ξ ) + Ka (ξ1 , ξ2 ) ,
2
a 1 (1)
Ka (ξ1 , ξ2 ) = 2 ,
2π ξ12 − a2 (ξ2 + i 0)2
where ξ = (ξ1 , ξ2 ), and δ (ξ ) is the Dirac’s mass.
Now we just have to find lim 2 πa 2 ξ 2 −a1 2 ξ 2 in distribution sense. Let ϕ (ξ ) ∈
2 x→∞ 1 2
S R (Schwartz class of infinitely differentiable, rapidly decreasing at infinity,
functions)
a ϕ (ξ1 , ξ2 ) d ξ1 d ξ2 a ϕ (ξ1 ,t/a)
= 2 d ξ1 dt.
2π 2 R2 ξ1 − a ξ2
2 2 2 2π R2 ξ12 − t 2
Pseudodifferential Equations in Canonical Non-Smooth Domains 381
Now calculate the inner integral. The function under the integral sign is even, so
+∞ +∞
+∞ +∞
dt dt 1 dt dt
=2 = +
−∞ ξ12 − t 2 0 ξ12 − t 2 ξ1 0 ξ1 − t 0 ξ1 + t
1 ξ1 + t ∞ 1 πi
= ln = ln (−1) = .
ξ1 ξ1 − t 0 ξ1 ξ1
where the designation of the functional P is taken from Vladimirov’s books [Vl81,
Vl79], and ⊗ means direct multiplication of distributions.
Distribution (4) corresponds to a semi-infinite crack (of course, with the addition
of 12 δ (ξ )). In order to control this, find another asymptotic of the function (1) when
a → 0. What we obtain corresponds to the half-plane y > 0 (instead of the “point”
this will be the “upper unit half-circle”). We now have what is needed to compare
to this result.
If we take b = 1/a, then we can write
1 ϕ (t/b, ξ2 ) d t d ξ2
lim
b→0 2π 2 R2 t 2 − ξ22
+∞
+∞
1 dt
= 2 ϕ (0, ξ2 ) d ξ2
2π −∞ −∞ t 2 − ξ2
2
which is the same result, as the one we had in half-plane (see [ViEs65, Es81]).
382 V.B. Vasilyev
3 Multidimensional Case
aΓ (m/2) 1
m+2 m/2 , (6)
2π 2
|ξ | − a (ξm + i0)
2 2 2
Here we should distinguish between the cases where m is odd and even. First con-
sider the odd sequence m = 3, 5, 7, . . .. With the help of a formula from [GrRy71]
we can write
+∞
+∞ k
dt 1 (m−3)/2 C(m−3)/2 t 2k+1
=
|ξ |m−1
∑ 2k+1
0 |ξ |2 − t 2 m/2 k=0
(2k + 1) |ξ | − t
2 2 2
0
(m−3)/2
1 1
=
i |ξ |m−1
∑ Ck
2k + 1 (m−3)/2
.
k=0
a Γ (m/2) 1
lim m+2 m/2
a→∞ 2π 2 (|ξ |2 −a2 ξm2 )
Γ (m/2) bm 1
= P ⊗ δ (ξm ) , m = 3, 5, 7, . . . . (8)
|ξ |m−1
m+2
iπ 2
1
The meaning of P will be explained separately.
|ξ | m−1
Now consider the even sequence m = 2, 4, 6, . . .. Using the formula [GrRy71] and
taking m/2 = k, we have
+∞
+∞
dt t
k =
| − t 2 k−1
0
|ξ | −t
2 2 2 (k − 1) |ξ |2
| ξ 2
0
+∞
2 (k − 1) − 1 dt
+ , k = 2, 3, . . . ,
2 (k − 1) |ξ |2 0 |ξ |2 − t 2 k−1
Consequently,
+∞
dt 2·2−3 πi π
2 = · = ,
0
| ξ |2 − t 2 2 (2 − 1) |ξ |3 2 4 |ξ |3
+∞
dt (2 · 3 − 3) (2 · 2 − 3) 1 π i
3 = · · ,
2 (3 − 1) 2 (2 − 1) |ξ |5 2
0
| ξ |2 − t 2
+∞
dt (2 · 4 − 3) (2 · 3 − 3) (2 · 2 − 3) 1 π i
4 = · · ,
2 (4 − 1) 2 (3 − 1) 2 (2 − 1) |ξ |7 2
0
| ξ |2 − t 2
or, in distributions,
a Γ ( m2 ) 1
lim m+2 m/2
a→∞ 2π 2 (| ξ |2 −a2 ξm2 )
Γ m2 cm i 1
= P ⊗ δ (ξm ) , m = 4, 6, 8, . . . . (12)
(2π ) m/2
|ξ |m−1
At last, it is possible to calculate the distribution for the half-space xm > 0 (a → 0):
a Γ ( m2 ) ϕ (ξ , ξm ) d ξ d ξm
lim m+2 m/2
a→0 2π 2 Rm
| ξ |2 − a2 ξm2
Γ ( m2 )bm−1 ϕ (ξ ) d ξ
= lim m+2 m/2
b→∞ 2π 2 Rm
| bξ |2 − ξm2
ϕ tb1 , . . . , m−1
t
Γ ( m2 ) b , ξm d t1 . . . d tm−1 d ξm
= lim m+2 m/2
b→∞ 2π 2 Rm
|t |2 − ξm2
⎛ ⎞
+∞
Γ ( m2 ) ⎜ d t ⎟
= m+2 ϕ (0, . . . , 0, ξm ) ⎝ m/2 ⎠ d ξm ,
2π 2 −∞ Rm−1
|t |2 − ξm2
where ωm−1 is the surface area of the unit sphere in (m − 1)-dimensional space;
m−2
2π 2
ωm−1 = Γ ( m−2
(see, for example, [Fi68]).
2 )
Using the relevant formula from [GrRy71] we obtain (m ≥ 4)
+∞ +∞
rm−2 d r 1 rm−2 d r
=
0 (r2 − ξm2 )m/2 im(ξm2 − r2 )m/2
0
= +∞ >
rm−3 m − 3 +∞ rm−4 d r
= i−m m−2 −
(m − 2) (ξm2 − r2 ) 3 0 m − 2 0 (ξ 2 − r2 ) m−2
m
2
+∞
m−3 rm−4 d r
= i−m .
m−2 0 (ξm2 − r2 )
m−2
2
Hence, if we set
+∞
rm−2 d r >1.
m−2 ≡ Im (ξm ) ,
0 (ξm2 − r2 ) 2
then
m−3
Im (ξm ) = − Im−2 (ξm ) . (13)
m−2
For m = 3,
+∞
+∞ −1/2
rdr 1
=− 3 = −c2
0 (r2 − c2 )
3
2 (r2 − c2 ) 2
0
1 i 1
= = − ; I3 (c) =
ic c c
(for c is the same as c + i0). For m = 4 use the recurrent formula (13),
+∞
r2 d r 1 πi
I4 (c) = = − · I2 (c) = −
0 (r2 − c2 )2 2 2c
+∞ r2 d r
(we recall that I2 (c) = 0 c2 −r2
= 2π ci ). Further, using the recurrent formula (13),
we see that
2 21
I5 (c) = − I3 (c) = − ,
3 3c
3 3 πi 3 1 π
I6 (c) = − I4 (c) = − − =− − i ,
4 4 4c 4 4 c
4 4 2 1
I7 (c) = − I5 (c) = − − ,
5 5 3 c
386 V.B. Vasilyev
5 5 3 πi
I8 (c) = − I6 (c) = − − − .
6 6 4 4c
Consequently, we can write (m ≥ 5)
* m−5 2 1
m−2 − m−4 · · · − 3 c ,
Im (c) = − m−3 m odd
πi
Im (c) = − m−2 − m−4 · · · − 4 − 4c , m even
m−3 m−5 3
where ⎧
⎪ −1,
⎪
⎨ πi
m = 3,
− , m = 4,
dm = 2m−3 m−5 2
⎪
⎪ − − · · · − , m ≥ 5 odd,
⎩ m−2 m−4 33 π i
− m−3
m−2 − m−5
m−4 · · · − 4 − 4 , m > 5 even.
The formula (14) for distributions is
a Γ (m/2) Γ (m/2) dm 1
lim 1
= m−1 δ ξ ⊗ P . (15)
a→0 2π
m+2
2 (| ξ |2 −a2 ξm2
m/2
) π 3/2 im Γ 2
ξm
Now we can see, that (15), with the accuracy of a constant, gives us the known
formula for half-space. Consequently
Γ (m/2) dm 1
m−1 = . (16)
π 3/2 im Γ 2
2π i
4 The Pyramid
In this section we talk about asymptotics to a polyhedral angle, with its simplest
variant being a pyramid.
The equation of such a pyramid x3 > a |x1 | + b | x2 | , which contains two parame-
ters a, b, that correspond to the length of the relevant axis. If we set these parameters
to 0 or ∞, then we can obtain other types of singularities.
The distribution, which corresponds to a such pyramid, is [Va98]
Pseudodifferential Equations in Canonical Non-Smooth Domains 387
4iab ξ
Ka,b (ξ1 , ξ2 , ξ3 ) = 2 3 2 .
(2π ) ξ1 − a ξ3 ξ2 − b2 ξ32
3 2 2
and hence
+∞
4i a b ξ3 ϕ (ξ1 , ξ2 , ξ3 ) d ξ1 d ξ2 d ξ3 i ϕ (0, ξ2 , 0)
lim = d ξ2 .
b→∞ (2π )3 R3 x 2π −∞ ξ2
For distributions,
388 V.B. Vasilyev
4 i a b ξ3 i 1
lim 3 2 = δ (ξ1 ) ⊗ P ⊗ δ (ξ3 ) .
b→∞ (2 π ) ξ1 − a2 ξ32 ξ22 − b2 ξ32 2π ξ2
4 i a b ξ3 i 1
lim 2 = P ⊗ δ (ξ2 ) ⊗ δ (ξ3 ) .
a→∞ (2 π )3 ξ12 − a2 ξ32 ξ2 − b ξ3
2 2 2 π ξ1
that is,
4iab ξ
lim 2 3 2 = δ (ξ2 ) ⊗ Ka (ξ1 , ξ3 )
b→0 (2 π ) ξ1 − a ξ3 ξ2 − b2 ξ32
3 2 2
4iab ξ
lim 3 2 = δ (ξ1 ) ⊗ Kb (ξ2 , ξ3 ) .
a→0 (2 π ) 3 ξ12 − a2 ξ32 ξ2 − b2 ξ32
4iab ξ 1 1
lim 2 3 2 = δ ξ ⊗P , ξ = (ξ1 , ξ2 ) .
a→0
b→0
(2 π ) ξ1 − a ξ3 ξ2 − b ξ3
3 2 2 2 2 2π i ξ3
References
[Vl81] Vladimirov, V.S.: Equations of Mathematical Physics, Nauka, Moscow (1981) (Rus-
sian).
[Vl79] Vladimirov, V.S.: Distributions in Mathematical Physics, Nauka, Moscow (1979)
(Russian).
[ViEs65] Vishik, M.I., Eskin, G.I.: Convolution equations in bounded domains. Russian Math.
Surveys, 20, 89–152 (1965) (Russian).
[Es81] Eskin, G.: Boundary Value Problems for Elliptic Pseudodifferential Equations, AMS,
Providence, RI (1981).
[Va00] Vasil’ev, V.B.: Wave Factorization of Elliptic Symbols: Theory and Applications,
Kluwer Academic Publishers, Dordrecht–Boston–London (2000).
[Mi77] Mikhlin, S.G.: Linear Partial Differential Equations, Vysshaya Shkola, Moscow
(1977) (Russian).
[Mi62] Mikhlin, S.G.: Multidimensional Singular Integrals and Integral Equations, Fizmat-
giz, Moscow (1962) (Russian).
[Mi86] Mikhlin, S.G., Prössdorf, S.: Singular Integral Operators, Akademie-Verlag, Berlin
(1986).
[Ma62] Malgrange, B.: Singular integrtal operators and uniqueness of the Cauchy problem.
Matematika, 6, 87–129 (1962) (Russian).
[CaZy57] Calderon, A.P., Zygmund, A.: Singular integral operators and differential equations.
Amer. J. Math., 79, 901–921 (1957).
[Va98] Vasilyev V.B.: Regularization of multidimensional singular integral equations in non-
smooth domains. Trans. Moscow Math. Soc., 59, 73–105 (1998) (Russian).
[BoMa48] Bochner, S., Martin, W.T.: Several Complex Variables, Princeton University Press,
Princeton, NJ (1948).
[Vl64] Vladimirov, B.S.: Methods of Function Theory of Several Complex Variables, Nauka,
Moscow (1964) (Russian).
[Ka92] Kasumov, N.: Calderon–Zygmund theory for kernels with non-point singularities.
Mat. Sb., 183, 89–104 (1992) (Russian).
[GrRy71] Gradstein, I.S., Ryzhik, I.M.: Handbook of Integrals, Sums, Series and Products,
Nauka, Moscow (1971) (Russian).
[Fi68] Fikhtengol’ts, G.M.: Calculus, vol. 3, Nauka, Moscow (1968) (Russian).
[Ho83] Hörmander, L.: Analysis of Partial Differential Operators, vols. 1–4, Springer-Verlag,
Berlin (1986–1988).
[Tr83] Treves, F.: Introduction to Pseudodiferential and Fourier Integral Operators, vols. 1,
2, Plenum Press, New York–London (1982).
[Ta83] Taylor, M.: Pseudodifferential Operators, Princeton Univ. Press, Princeton, NJ
(1981).
[ReSc83] Rempel, S., Schulze, B.-W.: Index Theory of Elliptic Boundary Value Problems,
Akademie-Verlag, Berlin (1982).
[Bo71] Boutet de Monvel, L.: Boundary problems for pseudodifferential operators. Acta
Math., 126, 11–51 (1971).
[St66] Sternin, B.Yu.: Elliptic and parabolic problems on manifolds with boundary consisting
of components of different dimension. Trans. Moscow Math. Soc., 15, 346–382 (1966)
(Russian).
Optimizing Water Quality in a River Section
1 Introduction
Since early times, rivers have been not only sources of life but also water discharge
receivers (both from industrial and urban origin) from the human settlements on
their banks. This fact brings with it that pollutant matter concentration surpasses
healthy levels in some sections of the rivers. In our paper, we use mathematical
modeling and optimal control theory to simulate one of most common strategies
in the pollution reduction of a river section: clear water injection into the channel
from a nearby reservoir. In this process of increasing the river flow by controlled
releases of water from reservoirs, the main problem consists (once the injection
point has been chosen by geophysical reasons) of finding the minimum quantity
of water which needs to be injected into the river section in order to purify it to a
desired level.
We formulate this environmental problem as a hyperbolic optimal control prob-
lem with control constraints. The state system is given in terms of section-averaged
velocity of water and wet section by the 1D inviscid shallow water system cou-
pled with the pollutant concentration equation. In this work, we deal with a unique
generic pollutant: pathogenic microorganisms or chlorides (because in an urban sce-
M.A. Vilar
Universidad de Santiago, Lugo, Spain,
e-mail: miguel.vilar@usc.es
L.J. Alvarez-Vázquez
Universidad de Vigo, Spain,
e-mail: lino@dma.uvigo.es
A. Martínez
Universidad de Vigo, Spain,
e-mail: aurea@dma.uvigo.es
M.E. Vázquez-Méndez
Universidad de Santiago, Lugo, Spain,
e-mail: miguelernesto.vazquez@usc.es
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 391
DOI 10.1007/978-0-8176-8238-5_36, © Springer Science+Business Media, LLC 2011
392 M.A. Vilar et al.
nario they are the most commonly used pollutants for water quality analysis), and
can be easily scaled to several pollutants. In our case, we take the flux of injected
water as the control parameter of the optimal control problem, and the constraints
correspond to technological bounds, while the objective function is related to the
total quantity of released water and to the fulfillment of water quality constraints.
For the numerical solution we combine finite difference and finite element tech-
niques for the time-space discretization, and a gradient-free method (the Nelder–
Mead algorithm) to solve the discrete optimization problem. Finally, computational
results for a realistic problem are provided.
in [0, L], and where δ (x − b) denotes the Dirac measure at point b ∈ [0, L]; for j =
1, . . . , O, e j ∈ (0, L) is the point where the mouth of the jth tributary is located,
q j (t) is the corresponding flow rate, U j (t) is its velocity, α j is the angle between
the jth tributary and the main river, and n j (t) is its mass pollutant flow rate; for
k = 1, . . . ,V , vk ∈ (0, L) is the point where the kth waste-water discharge is located,
pk (t) is the corresponding flow rate, Vk (t) is its velocity, βk is the angle between the
kth discharge and the river, and mk (t) is its mass pollutant flow rate; p ∈ (0, L) is the
point where clear water is discharged, Q(t) is the corresponding flow rate (which
will be our control), W (t) is its velocity, and γ is the angle between the discharge
and the river (it is worthwhile remarking here that, since we are injecting clear water,
this term does not appear in the second member of the pollutant equation); g stands
for the acceleration due to gravity; S f denotes the bottom friction stress, which can
be given, for instance, by the Chézy law; η (x,t) = H(x,t) + b(x) is the height of
water with respect to a fixed reference level, where H(x,t) represents the height of
the water column and b(x) describes the geometry of the river bottom; and k(x,t) is
the pollutant loss rate.
At first sight we can see four unknowns in state system (1)–(3): A(x,t), q(x,t),
η (x,t), and c(x,t). However, it is obvious that, if the river geometry is known,
A(x,t) can be derived from η (x,t). In effect, for each x ∈ [0, L], the geometry of
the river gives us a smooth, strictly increasing and positive function S(., x) satisfy-
ing S(0, x) = 0 and S(H(x,t), x) = A(x,t) in [0, L] × [0, T ]. Specific characterizations
of S for particular geometries can be found, for instance, in [Be06]. So, since we are
dealing with a system of balance laws whose conservative variables are A, q and c, if
we write η in terms of A, the non-conservative unknown η in (1) can be suppressed.
394 M.A. Vilar et al.
Explicitly, we have
∂η ∂
(x,t) = [B(A(x,t), x)] + b (x), (4)
∂x ∂x
where, for each x ∈ [0, L], B(., x) denotes the inverse of the function S(., x) (i.e.
B(A(x,t), x) = H(x,t)), which is also a smooth, strictly increasing and positive func-
tion.
Moreover, we can write
∂η ∂
A(x,t) (x,t) = [G(A(x,t), x)] − F(A(x,t), x) + A(x,t)b (x), (5)
∂x ∂x
where
B(A,x)
G(A, x) = S(r, x) dr,
0
B(A,x) A
∂S ∂B
F(A, x) = (r, x) dr = − (s, x) ds.
0 ∂x 0 ∂x
Then, the second equation of (1) can be rewritten as
∂ q ∂ q2 ∂
+ +g G(A, x) − F(A, x) + Ab (x)
∂t ∂x A ∂x
= QW cos(γ )δ (x − p) + g2 . (6)
Since B(·, x) is a strictly increasing function and B(0, x) = 0, it is easy to prove that
G(·, x) is strictly increasing on [0, ∞) and G(0, x) = 0. On the other hand, realistic
solutions correspond to A ≥ 0. Hence, we can replace G by G 0 defined by
⎧
⎨ G(A, x) if A > 0,
0 x) = (−∞, 0] if A = 0,
G(A,
⎩
Ø if A < 0.
since we are just injecting (not extracting) clear water, and the quantity of injected
water must be bounded.
In order to formulate the control problem we consider as the cost functional the
total amount of clear water injected through the point p together with a measure (in
the region of the river starting from point p) of the contaminant concentration which
remains higher than the fixed threshold cmax . Thus, we define the cost function:
T T L
ε μ
J(Q) = Q(t)2 dt + (c(x,t) − cmax )2+ dx dt (9)
2 0 2 0 p
where ε and μ are two weight parameters, and (c − cmax )+ denotes the positive part
of c − cmax , that is, (c − cmax )+ = max{c − cmax , 0}.
So the problem, denoted by (P), of the optimal water injection for the purifi-
cation of a polluted section in a river consists of finding the control flux Q ∈ Uad
of injected clear water in such a way that the state system (1)–(3) is satisfied and
which minimizes the cost function J given by (9). Thus, the problem can be written
in short as
For technical reasons (flow control mechanisms cannot act upon water flow in a
continuous way, but discontinuously at short time periods) we look for admissible
controls Q in the piecewise-constant L2 (0, T ) functions. So, for the time interval
[0, T ] we choose a number K ∈ N, consider the time step Δ τ = T /K, and define
the discrete times τm = mΔ τ , for m = 0, . . . , K. Thus, a function Q ∈ Uad which
is constant at each subinterval determined by the grid {τ0 , τ1 , . . . , τK } is completely
fixed by the set of values QΔ τ = (Q0 , Q1 , . . . , QK−1 ) ∈ [0, Qmax ]K ⊂ RK , where Qm =
Q(τm ), m = 0, . . . , K − 1.
This discretization leads to a time discretization of the cost function J and the
state system (1)–(3). In order to solve accurately this state system we discretize the
problem. Thus, for given N ∈ N (usually a multiple of K), we define Δ t = T /N and
take tn = nΔ t, for n = 0, . . . , N. First and third equations of system (1) are discretized
in an implicit way, but for the second one we use the method of characteristics that
stems from considering the following identity:
396 M.A. Vilar et al.
D(V q) ∂q ∂ (uq)
(x,t) = (x,t) + (x,t), (10)
Dt ∂t ∂x
∂
for the total derivative D(VDt (x,t) = ∂ τ [V (x,t; τ )q(X(x,t; τ ), τ )]|τ =t , where X(x,t; τ )
q)
is the characteristic line (providing the position at time τ of the particle that occupied
the position x at time t), which is the unique solution of the following ordinary
differential equation:
⎧
⎨ dX
(x,t; τ ) = u(X(x,t; τ ), τ ),
dτ (11)
⎩
X(x,t;t) = x,
and V (x,t; τ ) is the evolution of the element of volume, which is given by the solu-
tion of the following ordinary differential equation:
⎧
⎨ dV ∂u
(x,t; τ ) = (X(x,t; τ ), τ )V (x,t; τ ),
dτ ∂x (12)
⎩
V (x,t;t) = 1.
2
Expression (10) and the equalities uq = Aq q = qA change (7) into
D(V q) ∂ζ
(x,t) + g (x,t) − F(A(x,t), x) + A(x,t)b (x)
Dt ∂x
= Q(t)W (t) cos(γ )δ (x − p) + g2 (x,t).
Then, if we denote X n (x) = X(x,tn+1 ;tn ) and V n (x) = V (x,tn+1 ;tn ), the state
system (1) can be approximated by the following semi-discrete system: for n =
0, . . . , N − 1 find functions An+1 (x), qn+1 (x), cn+1 (x), defined in (0, L), such that
∂ qn+1
A n+1
= A (x) + Δ t Q(tn+1 )δ (x − p) + g1 (x,tn+1 ) −
n
(x) , (13)
∂x
qn+1 (x) ∂ ζ n+1
+g (x) − F(An+1 (x), x) + An+1 (x)b (x)
Δt ∂x
qn (X n (x))V n (x)
= + Q(tn+1 )W (tn+1 ) cos(γ )δ (x − p) + g2 (x,tn+1 ), (14)
Δt
ζ n+1 (x) ∈ G̃(An+1 (x), x), (15)
cn+1 (x) − cn (x) ∂ qn+1 cn+1
+ (x) + k(x,tn+1 )cn+1 (x)
Δt ∂x An+1
= g3 (x,tn+1 ). (16)
The admissible set Uad is approached by Uad Δ τ , the set of controls in U which
ad
are piecewise-constant in the partition of the time interval [0, T ] given by the time
step Δ τ . Finally, for any given control QΔ τ ∈ UadΔ τ , we use the following discrete
ε K−1
(Qm )2 + (Qm+1 )2
J Δ t (QΔ τ ) = Δ τ ∑
2 m=0 2
L n
μ N−1 (c (x) − cmax )2+ + (cn+1 (x) − cmax )2+
+ Δt ∑ dx. (17)
2 n=0 p 2
To evaluate this cost function we need to solve the semi-discrete system (13)–(16)
for n = 0, . . . , N − 1. Since variable cn+1 is not coupled with the first three equations
of the system, we will proceed to solve it sequentially.
It has been proved in [Be06] that this variational problem has, at least, one so-
lution. In order to obtain it, we choose Λh = {x0 = 0, x1 , . . . , xM = L} a partition
of interval [0, L] in M subintervals Ik = [xk−1 , xk ], k = 1, . . . , M, such that there ex-
ists P ∈ {1, . . . , M − 1} satisfying xP = p. In association with this, we consider the
following finite dimensional vector spaces:
where Xhn and Vhn are the numerical solutions of (11) and (12), respectively. We
resolve this nonlinear discretized system doing an implicit discretization of the op-
erator F, and using the Bermúdez–Moreno iterative algorithm [Be81] for dealing
with operator G.
Equation (16) can be now solved by using an implicit upwind finite difference
scheme. In order to do it, because of the Dirac measures characterizing the sources,
we consider the following approximations: for each k = 0, . . . , M, we define δhk :
[0, L] → [0, +∞) by
⎧
⎪ b − xk−1
⎪
⎪ if b ∈ [xk−1 , xk ],
⎪
⎨ k − xk−1 )2
(x
δhk (b) = xk+1 − b
⎪
⎪ if b ∈ [xk , xk+1 ],
⎪
⎪ (x − xk )2
⎩ k+1
0 otherwise,
qn+1
h (xk ) n+1 qn+1
h (xk−1 ) n+1
ckn+1 − cnk ck − ck−1
An+1
h (xk ) An+1
h (xk−1 )
+ + k(xk ,tn+1 )ckn+1
Δt xk − xk−1
Optimizing Water Quality in a River 399
E V
= ∑ n j (tn+1 )δhk (e j ) + ∑ m j (tn+1 )δhk (v j ).
j=1 j=1
where
ε K−1
(Qm )2 + (Qm+1 )2
JhΔ t (QΔ τ ) = Δ τ ∑
2 m=0 2
xk+1 n
μ N−1 M−1 (ch (x) − cmax )2+ + (cn+1
h (x) − cmax )+
2
+ Δt ∑ ∑ dx. (18)
2 n=0 k=P xk 2
4 Numerical Optimization
where the parameter β > 0 determines the relative contribution of the objective
function and the penalty terms. Function J˜ is an exact penalty function in the sense
400 M.A. Vilar et al.
that, for sufficiently large β , the solutions of our constrained problem (PhΔ t ) are
equivalent to the minimizers of function J˜ in RK .
For computing the minimum of this penalty function J˜ we use a direct search al-
gorithm: the Nelder–Mead simplex method [Ne65]. This is a gradient-free method,
which merely compares function values where the values of the objective function
are taken from a set of sample points (simplex). For the interested reader, a detailed
description of the above algorithm can be found, for instance, in the paper of the au-
thors [Al02]. Although the Nelder–Mead algorithm is not guaranteed to converge in
the general case, it presents good convergence properties in low dimensions (which
is our case).
5 Computational Experiments
Both initial
√
and boundary conditions were taken as constant, particularly, AL (t) =
A0 (x) = 3 m2 , q0 (t) = q0 (x) = 1 m3 s−1 , c0 (t) = c0 (x) = 0 um−1 . The time
4 125
interval to control the pollution was T = 3600 s. Moreover, the loss rate for pollutant
was considered constant (k(x,t) = 10−4 s−1 ), and the bottom friction stress was
neglected (S f = 0).
Optimizing Water Quality in a River 401
Fig. 3 Uncontrolled (left) and controlled (right) quantity of pollutant at three significant times
around points p = 700 and e3 = 850
References
[Al02] Alvarez-Vázquez, L.J., Martínez, A., Rodríguez, C., Vázquez-Méndez, M.E.: Numerical
optimization for the location of wastewater outfalls. Comput. Optim. Appl., 22, 399–417
(2002).
[Al06] Alvarez-Vázquez, L.J., Martínez, A., Vázquez-Méndez, M.E., Vilar, M.A.: Optimal lo-
cation of sampling points for river pollution control. Math. Comput. Simul., 71, 149–160
(2006).
[Al09] Alvarez-Vázquez, L.J., Martínez, A., Vázquez-Méndez, M.E., Vilar, M.A.: An applica-
tion of optimal control theory to river pollution remediation. Appl. Numer. Math., 59,
845–858 (2009).
[Be81] Bermúdez, A., Moreno, C.: Duality methods for solving variational inequalities. Comput.
Math. Appl., 7, 43–58 (1981).
[Be06] Bermúdez, A., Muñoz-Sola, R., Rodríguez, C., Vilar, M.A.: Theoretical and numerical
study of an implicit discretization of a 1D inviscid model for river flows. Math. Models
Meth. Appl. Sci., 16, 375–395 (2006).
[Br73] Brezis, H.: Operateurs Maximaux Monotones el Semi-groupes de Contractions dans les
Espaces de Hilbert, North-Holland, Amsterdam (1973).
[Ne65] Nelder, J.A., Mead, R.: A simplex method for function minimization. Computer J., 7,
308–313 (1965).
Boundary Integral Equations for Arbitrary
Geometry Shells
V.V. Zozulya
1 Introduction
V.V. Zozulya
Centro de Investigacion Cientifica de Yucatan A.C., Mérida, Mexico,
e-mail: zozulya@cicy.mx
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 403
DOI 10.1007/978-0-8176-8238-5_37, © Springer Science+Business Media, LLC 2011
404 V.V. Zozulya
mental solutions are available only for some special cases of shell’s geometry and
in particular for plates [BrTW84].
The application of the BIE method to the analysis of shells with arbitrary geom-
etry has been studied by many researchers [ArGr02, Be91, Pa91]. In some of those
publications, integral equations are applied without the construction of the funda-
mental solution for the shell’s differential operator. In spite of the difference in the
approaches to the problem, all of these publications have a common disadvantage.
The integral equations obtained in those publications contain unknown functions
not only on the contour of the shell but also in the surface integrals.
A theory of shells based on expansion of the 3-D equations of elasticity into a
series of Legendre polynomials has been developed in [Ve82]. The equations ob-
tained are based on a strong mathematical foundation and have many applications
in engineering practice [Ho83-85, Zo89, Zo06b, Zo07]. In [Zo97] integral identities
of Somigliana type and also fundamental solutions for the coefficients of a Legen-
dre polynomial expansion have been obtained based on the approach developed in
[Ve82].
In this paper we extended and generalize our previous result related to the BIE
application for analysis of the arbitrary geometry shells earlier published in [Zo97,
Zo98]. An integral representations of Somigliana type which allow us to transform
the problem of analysis of the shells with arbitrary geometry to a 1-D BIE over the
contour of the shell is obtained. Thus an approach which transforms analysis of the
3-D problem of elasticity to analysis of the 1-D BIE for arbitrary geometry shells is
developed. We also show here that all the kernels presented in the BIE for shells of
Vekua type [Ve82] can be calculated analytically.
where the summation convention over repeated indices has been used (and will use
from now on). More detailed information related to the tensor notations used here,
and its application to shell theory, may be found in [Ve78].
The mapping (1) and (2) are also called coordinate transformations which relate
differentials
dzi = Aij dx j and dxi = Bij dz j .
Denote the basis vectors in Cartesian coordinates by ei and in curvilinear by Ri , they
are related in the following way:
ei = Bij R j , Ri = Aij e j .
Covariant, contravariant and mixed components of the metric tensor are scalar prod-
ucts of the basic vectors
Ri · R j = gi j , Ri · R j = gi j , Ri · R j = gij . (3)
σ = σ i j Ri ⊗ R j , ε = ε i j Ri ⊗ R j , u = ui Ri , p = pi Ri , b = bi R i .
Covariant and mixed components of the related tensors and vectors can be found in
the standard way by raising and lowering indexed using metric tensors (3). These
quantities are not independent as they are related by the equations of linear elasticity.
If the displacements and their gradients are small, the Cauchy relations can be used:
1 1 1
ε= (∇u)T + ∇u , εi j = (∇i ui + ∇ j ui ) = (∂i ui + ∂ j ui ) − Γi kj uk . (4)
2 2 2
Here ∇i are the covariant derivatives, Γi kj are the Christoffel symbols and ∂i =
@
∂ ∂ xi are partial derivatives with respect to the space variables xi .
The equations of equilibrium have the form
The stress and strain tensors are related by the generalized Hook’s law
vector form
A · u = b, A = (λ + μ )∇∇ + μ ∇ · ∇ = Ai j Ri ⊗ R j ,
∇∇u = gk j Ri ∇i ∇k u j , ∇ · ∇ u = Δ u = Ri ∇k ∇k ui = gk j Ri ∇ j ∇k ui .
Ai j u j (x) + bi (x) = 0, ∀x ∈ V,
(7)
Ai j = [μ gi j gkl + (λ + μ )gil g jk ]∇ k ∇l , ∇i u j = ∂i u j − Γi kj uk .
Boundary conditions in shell theory are usually stated in the following form. On the
outer surfaces Ω+ and Ω− the traction vectors are prescribed
P · u = p+ , ∀x ∈ Ω+ , P · u = p− , ∀x ∈ Ω−
or
Pi j u j (x) = pi+ (x) , ∀x ∈ Ω+ , Pi j u j (x) = pi− (x) , ∀x ∈ Ω− . (8)
On the sheer side S = S p ∪ Su , S p ∩ Su = 0/ the mixed boundary conditions
P·u = ψ, ∀x ∈ S p u = ϕ , ∀x ∈ Su
or
Pi j u j (x) = ψ i (x) , ∀x ∈ S p ; ui (x) = ϕi (x) , ∀x ∈ Su (9)
are prescribed. The traction operator in (8) and (9) in curvilinear coordinates has the
form
P = Pi j Ri ⊗ R j = λ n · ∇ + μ n · (∇ )T + ∇ ,
(10)
Pi j = λ ni g jk + μ (nk gi j + n j gik ) ∇k .
The boundary value problem (7)–(9) may be transformed to a system of BIE using
the Somigliana’s identity. For this purpose the work that the body forces and the
surface tractions perform on the corresponding displacements is considered and is
given by
1 1
L(ui , pi , bi ) = bi (x)ui (x)dV + pi (x)ui (x)dS. (11)
2 2
V ∂V
Together with ui , pi , bi let us consider another system of forces, tractions and dis-
placements ūi , p̄i , b̄i . From the reciprocal theorem of Betty [Ki63] it follows that
Boundary Integral Equations for Arbitrary Geometry Shells 407
(3 − 4ν )δi j + ∂i R∂ j R
Ui j (z(x) − z(y)) = , (14)
16π μ (1 − ν )R
<
where R = (z1 (x) − z1 (y))2 + (z2 (x) − z2 (y))2 + (z3 (x) − z3 (y))2 is the Euclidean
distance between points x and y.
The fundamental tensor of traction Wi j (z(x), z(y)) in Cartesian coordinates may
be found by applying the operator Pi j from (10) in the form
Pi j = λ ni ∂k + μ (δik ∂n + nk ∂i )
In (14)
< and (15) zi (x) and zi (y) are the Cartesian coordinates of the points x and y,
R = (z1 (x) − z1 (y))2 + (z2 (x) − z2 (y))2 + (z3 (x) − z3 (y))2 is the Euclidean dis-
tance between points x and y and υ is Poisson’s ratio.
From (11) and (12) and taking into account that
The fundamental tensors of the displacements and tractions (14), (15) in curvilinear
coordinates using coordinates transformations (1) and (2) may be presented in the
408 V.V. Zozulya
form
Ui j (x − y) = Aki (x)Alj (y)Ukl (z(x) − z(y)),
W ji (x, y) = Bik (x)Alj (y)Wlk (z(x), z(y)).
The integral representation (16) may be used for further development of the BIE
for 3-D elasticity in arbitrary curvilinear coordinates.
As mentioned in [ArGr02] and [Be91] the application of the 3-D BIE for elasticity
to problems in the theory of shells has some methodological and computational
difficulties. In order to avoid those problems an approach based on the reduction of
a 3-D problem to a 2-D one has been developed here.
For convenience we transform above equations of elasticity taking into account
that the radius vector R(x) of any point in domain V , occupied by material points of
the shell, may be presented as
or in a coordinate form
zi (x) = zi (xα ) + x3 ni3 (xα ),
where r(xα ) is the radius vector of the points located on the middle surface of the
shell, xα = (x1 , x2 ) are curvilinear coordinates associated with the middle surface,
n(xα ) is a unit vector normal to the middle surface.
The basis vectors in curvilinear coordinates are related to the middle surface of
the shell and, taking (17) into account, are presented in the form
Here and in (19), (20) aαβ = rα · rβ and bαβ = −rα · nβ are coefficients of first and
second quadratic forms of the surface.
The differential equations of equilibrium for the displacements (7) are
The next step in the transformation of the 3-D equations of elasticity into 2-D
equations of shells consists of the following. Assume that the components of stress
σ αβ and strain εαβ tensors; and displacements ui , body forces bi , and surface trac-
tions pi vectors are sufficiently smooth functions of coordinate x3 and may be ex-
panded into a series of Legendre’s polynomials. Then using the approach developed
in [Ve82], they can be expressed as
∞ n n h
2n + 1
σ (x) =
ij
∑σ ij
(xα )Pn (ω ) , σ (xα ) =
ij
2h
σ i j (xα , x3 )Pn (ω )dx3 ,
n=0
−h
∞ h
2n + 1
∑ εi j (xα )Pn (ω ) ,
n n
εi j (x) = εi j (xα ) = εi j (xα , x3 )Pn (ω )dx3 ,
n=0 2h
−h
∞ h
2n + 1
∑ ui (xα )Pn (ω ) ,
n n
ui (x) = ui (xα ) = ui (xα , x3 )Pn (ω )dx3 , (24)
n=0 2h
−h
∞ n n h
2n + 1
pi (x) = ∑ pi (xα )Pn (ω ) , i
p (xα ) = pi (xα , x3 )Pn (ω )dx3 ,
n=0 2h
−h
∞ n n h
2n + 1
bi (x) = ∑ b (xα )Pn (ω ) ,
i i
b (xα ) =
2h
bi (xα , x3 )Pn (ω )dx3 ,
n=0
−h
@
where ω = x h is a dimensionless coordinate.
3
Substituting the series expansions (24) into the equations of elasticity (19) and
(20) we get the corresponding relations for the coefficients of the expansion (24).
The Cauchy relations have the form
n 1 n n 3 n 1 n n β n n n
εαβ = (∇α uβ +∇β uα ) − bαβ u3 , εα 3 = ∇α u3 + ui −bα uβ , ε33 = ui , (25)
2 2 − −
410 V.V. Zozulya
where
n 2n + 1 n+1 n+3
ui = ( u3 + u3 + · · · ).
− h
The equations of equilibrium are
n n n n n
γ
∇β σ αβ −2bαβ σ 3β −bγ σ 3β + σ α 3 + f α = 0,
−
n n n n n (26)
γ 33
∇β σ 3β +bαβ σ αβ −bγ σ + σ 33 + f 3 = 0,
−
where
n n
2n + 1 i3
f i (xα ) = bi (xα ) + (σ+ (xα ) − (−1)n σ−i3 (xα )),
h
n
2n + 1 n+1 n+3
σ α3 =
(σ i3 + σ i3 + · · · ).
− h
The generalized Hook’s law for the coefficients of the expansion (24) has the same
form as (6) after considering (21).
Now with taking into account (25), (26) the differential equations for coefficients
of expansion (24) of the displacements has the form
n n n 0 1 n
Aαβ uβ (xα ) + Aα 3 u3 (xα ) + Lα (ui , ui , . . . , ) + f α (xα ) = 0
n n ∀xα ∈ Ω, (27)
n n 0 1
A3β uβ (xα ) + Aα 3 u3 (xα ) + L3 (ui , ui , . . . , ) + f 3 (xα ) = 0
n
where Ai j are the differential operators defined in (22), (23), Li is the operator that
depend on all members of the series expansion (24) and include first order differen-
tial operators.
The boundary conditions (9) for the coefficients of expansion (24) have the form
n
Pαβ uβ (xα ) = pα+ (xα ) ,
n
∀xα ∈ ∂ Ω p ,
(28)
uα (xα ) = pi− (xα ) , ∀xα ∈ ∂ Ωu .
Substituting the series expressions (24) into the equation for the work that the body
forces and the surface tractions perform on the corresponding displacements (11),
we have
Boundary Integral Equations for Arbitrary Geometry Shells 411
∞ n ∞
1
∑ bi (xα )Pn (ω ) ∑ ui (xα )Pm (ω )dV
m
L(ui , pi , bi ) =
2
V n=0 m=0
∞ n ∞
1
∑ pi (xα )Pn (ω ) ∑ ui (xα )Pm (ω )dS.
m
+ (29)
2 n=0 m=0
∂V
Here, the orthogonal property of Legendre’s polynomials is used together with the
fact that Pn (1) = 1 and Pn (−1) = (−1)n , and boundary conditions given in (8).
Substituting expressions for the volume (30) and surface integral (31) into (29)
results in
n n ∞ n ∞ n
1 1
∑ ∑ pi (xα ) ui (xα )dl,
n n n
L(ui , pi , bi ) = i
f (xα ) ui (xα )d Ω + (32)
2 n=0 2 n=0
Ω ∂Ω
where
n n
f i (xα ) = bi (xα ) + (pi+ (xα ) − (−1)n pi− (xα )).
From (32), similar to 3-D elasticity, a corresponding reciprocal theorem for the
coefficients of the expansion into the Legendre’s polynomial series may be obtained
in the form
n n
− − n n n
n −
L(ui , pi , bi ) = L(ui , pi , bi ), (33)
which is an analog of Betty’s reciprocity theorem for (12).
In [Zo97] it has been proved that if
n n
− nnm − nm
−
b = δi δ
i j nm
(xα − yα ), ui = Ui j (xα , yα ), pi = W ji (xα , yα ),
where
h h
nm 2n + 1 2m + 1
Ui j (xα , yα ) = · U ji (x − y)Pn (ω )Pm (ω )dx3 dy3 ,
2h 2h
−h −h
412 V.V. Zozulya
nm h h
2n + 1 2m + 1
W ji (xα , yα ) = · Wi j (x, y))Pn (ω )Pm (ω )dx3 dy3 ,
2h 2h
−h −h
(34)
h h
δ nm
(xα − yα ) = δ (x − y))Pn (ω )Pm (ω )dx dy 3 3
−h −h
put in (33) the integral representation for the expansion of the coefficients of the dis-
placement vector components into a Legendre polynomial series may be obtained.
These integral representations are presented in the form
= >
∞ nm n nm
∑ Ui j (xα , yα ) pi (xα ) − W ji (xα , yα ) ui (xα )
m n
ui (yα ) = dl
n=0
∂Ω
∞ nm n
+ ∑ Ui j (Xα , yα ) f i (xα )dS / ∂ Ω (m = 0, 1, 2, . . . , ∞). (35)
∀yα ∈
n=0
Ω
This equation is the analog of the Somigliana identity in the theory of elasticity
[Ki63].
Somigliana’s identity (16) is used for formulation of the BIE in theory of elas-
ticity [BrTW84]. In the same way the BIE for shells of arbitrary geometry may
be obtained using the analogy of the Somigliana’s identity (35). For that purpose
the point yα must be brought to a position on the boundary ∂ Ω through a limiting
process. The boundary properties of the integral operator in (35) depends on the
boundary properties of the elastostatic fundamental solutions (14) and (15). These
properties are well known (see for example [BrTW84]). In [Pa91] it was shown that
the boundary properties of the integral operators in (35) are expressed by
* n + n
nm nm
lim Ui j (xα , yα ) pi (xα )dl = Ui j (xα , yα ) pi (xα )dl,
yα → ∂ Ω
∂Ω ∂Ω
* nm + nm (36)
n 1 n n
lim i
W j (xα , yα ) ui (xα )dl = ∓ ui (yα ) + W ji (xα , yα ) ui (xα )dl.
yα → ∂ Ω 2
∂Ω ∂Ω
1 1
(ξ 3 − ζ 3 )a (ξ 3 )b (ζ 3 )c
dξ 3dζ 3,
−1 −1
(r2 (x α , yα ) + h2 (ξ 3 − ζ 3 )2 )k/2 (38)
k = 1, 3, 5; a = 0, 1, 2; b = 0, 1; c = 0, 1,
1 1
dξ 3dζ 3 2
= r(x ,
α αy ) − r2 (xα , yα ) + 4h2
R2 (xα , yα , ξ 3 , ζ 3 )1/2 h3
−1 −1
+ h ln(h r (xα , yα ) + 4h + 2h) − h ln(h r (xα , yα ) + 4h − 2h) ,
2 2 2 2
1 1
(ξ 3 − ζ 3 )2 d ξ 3 d ζ 3
R2 (xα , yα , ξ 3 , ζ 3 )5/2
−1 −1
= >
4 r2 (xα , yα ) + 2h2 1
= 3 < − . (39)
3h r2 (xα , yα ) r2 (xα , yα ) + 4h2 r(xα , yα )
Some integrals given here are singular and need special consideration. For these
the approach developed in [Zo06a] can be used. We compare results of analytical
414 V.V. Zozulya
and numerical methods for evaluating the double integrals in (38). Calculations us-
ing analytical expressions for (39) have been almost more than ten times faster than
using numerical quadrature formulas.
References
[ArGr02] Artyuhin, Yu.P., Gribanov, A.P.: Solution of the Problems of Nonlinear Deformation of
Plates and Shallow Shells by Boundary Element Method, Kazan University Publisher
House, Kazan (2002) (Russian).
[BrTW84] Brebbia, C.A., Telles, J.C.F., Wrobel, L.C.: Boundary Element Techniques. Theory
and Applications in Engineering, Berlin (1984)
[Be91] Beskos, D.E.: Boundary Element Analysis of Plates and Shells, Springer, Berlin, 93–
140 (1991).
[GuBL78] Guliaev, V.I., Bajenov, V.A., Lizunov, P.P.: Nonclassical Shell Theory and Its Appli-
cation for Engineering Problems Solution, Vyshcha shkola, Lviv (1978) (Russian).
[Ho83-85] Hormander, L.: The Analysis of Linear Partial Differential Operators, vols. I–IV,
Springer Verlag, Berlin (1983–1985).
[Ki63] Kil’chevski, N.A.: Foundation of Analytical Theory of Shells, Publisher House of
ANUkrSSR, Kiev (1963) (Russian).
[Pa91] Paymushin, V.I., Sidorov, I.N.: Variant of boundary integral equation method for solu-
tion of static problems for isotropic arbitrary geometry shells. Mech. Solids, 1, 60–169
(1991).
[Ve78] Vekua, I.N.: Fundamental of Tensor Analysis and Theory of Covariants, Nauka,
Moskow (1978) (Russian).
[Ve82] Vekua, I.N.: Some General Methods of Construction Various Variants of Shell Theory,
Nauka, Moskow (1982) (Russian).
[Zo89] Zozulya, V.V.: The combined problem of thermoelastic contact between two plates
though a heat conducting layer. J. App. Maths. and Mech., 53, n. 5, 722–727 (1989).
[Zo97] Zozulya, V.V.: Somigliano identity and fundamental solutions for arbitrary geometry
shells. Doclady Akademii Nauk of Ukraine, 6, 60–65 (1997) (Russian).
[Zo98] Zozulya, V.V.: The boundary integral equations for the shells of arbitrary geometry.
Inter. App. Mech., 34, n. 5, 79–83 (1998).
[Zo06a] Zozulya, V.V.: Regularization of the divergent integrals. I. General consideration.
Electr. J. Boun. Elem., 4, n. 2, 49–57 (2006).
[Zo06b] Zozulya, V.V.: Laminated shells with debonding between laminas in temperature field.
Inter. App. Mech., 42, n. 7, 135–141 (2006).
[Zo07] Zozulya, V.V.: Mathematical modeling of pencil-thin nuclear fuel rods, in: Structural
Mechanics in Reactor Technology (Editor: A. Gupta), SMIRT, Toronto, C04–C12,
(2007).
Index
A homogenization, 161
integral equation, 79, 403
a posteriori pointwise estimator, 329 method, 129, 147, 403
absolute minimizers, 181 maps, 71
absorbing boundary conditions, 236 boundary–domain integral equations
adaptive particle filter, 47 localized, 91
adjoint state technique, 395 segregated, 109
advection–diffusion equation, 25 Bragg curve, 19
advective terms, 276 Brownian random walk, 262
aerodynamical applications, 259 bubble behavior, 147
algebraic equation, elliptic quadratic
qualitative analysis of, 256 C
algebraic equations, elliptic quadratic, 253
qualitative analysis of, 254 captured shock, 358
almost periodicity, 141 cardinal theorem of interpolation theory, 352
asymptotic analysis of singularities, 379 carrier function, 363
asymptotics, 161, 386 Cauchy
atmospheric and ionospheric ducts, integral equations, 1
evaporative, 367 principal value, 2
atmospheric pollutant relations, 405
dispersion model, 262 Cauchy–Kowalewski theorem, 32, 350
sources, 261 change of variable methods, 207
characteristic
B function, 342
method, 339
Banach principle, 143 Chebyshev polynomial, 377
Bayes theorem, 48 basis, 73
Bayesian approach, 47 Christoffel symbols, 405
Bermúdez–Moreno iterative algorithm, 398 co-normal derivative, 115
Bernoulli equation, 149 operators, 93
Bessel potential spaces, 110 collocation, 195
bimaterial, linearly elastic, 242 method, 1
Boltzmann equation, 17 conjugate gradient methods, 80
Boltzmann–Gibbs–Shanon formula, 51 contact problems, elastodynamic, 241
bootstrap filter, 55 continuous waves, 364
boundary contraction mapping, 143
element method, 327 control valve model, 305
C. Constanda, P.J. Harris (eds.), Integral Methods in Science and Engineering, 415
DOI 10.1007/978-0-8176-8238-5, © Springer Science+Business Media, LLC 2011
416 Index
Q thermal
conductivity, 130
quadratic forms of a surface, 409 expansion coefficient, 130
quadrature methods, 204 neutron flux, 61
quasimode with remainder, 313 thermostatistics, 52
toroidal bubble model, 154
R trace operators, 132
transfer operator, 69
Rayleigh–Plesset bubble, 152 transformation parameter, 131
regularization procedure, 1 transgenic mouse model, 214
reservoir flow, 369 transmission problem
Index 419