Professional Documents
Culture Documents
External Approximations by FE 2 PDF
External Approximations by FE 2 PDF
APPROXIMATIONS BY
FINITE ELEMENTS
© 2016 SIMSOLID Corporation All Rights Reserved. This document is an unpublished draft of
book “External Approximations by Finite Elements” by Victor Apanovitch. It is intended to
provide background information on the mathematical foundations used in the commericial
software application, SIMSOLID. Unauthorized use, distribution or duplication is prohibited.
Table of Contents
Preface
Introduction
5.1. Definitions
7.2.1. 2D elements for boundary value problems of second and fourth orders
7.2.2. 3D elements
8. Numerical experiments
8.1. Torsion
8.1.1. Formulation
8.2.1. Formulation
8.3.1. Formulation
8.6.1. Formulation
Reference list
Preface
The Finite Element Method proved to be a powerful analysis tool in many fields of engineering.
Its wide use and still great potential stimulate research and development of new effective finite
element based technologies. This book presents an innovative approach to the finite element
approximations. The approach possesses some new unique features and at the same time it
inherits the computational advantages of the conventional finite element analysis. For example:
finite elements of arbitrary shape with approximating functions of arbitrary type can be easily
formulated; the types of the approximating functions do not depend on the element’s geometry;
the element is not mapped onto a canonical shape; the matrix of the equation system is sparse,
symmetric, and positive definite; and others. I believe that the suggested concept will lead to new
Generality and constructiveness are the major features of the theory proposed. All the results are
obtained in abstract form under minimal restrictions on finite element geometry and contents of
approximating spaces. It provides wide applicability of the suggested theory to different areas of
analysis and simulation. It is described in detail how the methods of construction of the finite
elements and approximations follow from the basic theorems, and how they are implemented in
numerical algorithms. Numerous benchmark problems illustrate the applicability and accuracy of
the method.
Section 1 introduces the definitions of spaces and operators associated with a partition of the
domain into finite elements. Section 2 presents the necessary and sufficient sign, as well as two
sufficient signs of the external finite element approximations of Sobolev space H m () ( m 0 is
arbitrary integer) under minimal limitations on the shape of finite elements (it is assumed that the
boundary of the finite element K must be sectionally smooth only) and the contents of spaces of
approximating functions (it is assumed that the functions just belong to the space H m ( K ) ).
Herewith two types of the external approximations are considered: one is associated with a fixed
partition of the domain into finite elements, the other is associated with a family of partitions.
Section 3 addresses the formulations of different classes of finite elements for external
convergence properties on fixed and variable partitions of the domain. Section 5 develops the
spaces are used to construct external approximations by necessary and sufficient sign. Section 6
studies the external approximations of variational equations. Section 7 justifies some classic non-
conformed finite elements as well as presents examples of finite elements for external
problems.
This book should appeal to numerical analysts and software developers in different fields of
engineering such as structural and dynamic analysis, applied mechanics, multiphysics, and
computational mathematics.
The theory presented in the book was developed in 1981-1991 when I worked as a professor of
Elliptic boundary value problems can be stated in two equivalent forms: differential formulation
or weak formulation. The equivalency means that under the assumption of sufficient smoothness
of the input data (for example, the smoothness of the coefficients of the differential operator of
the problem and the smoothness of the boundary of the domain) both formulations can be
transformed one to another and their solutions coincide. The idea of a weak solution of a
In the weak form a boundary value problem is stated as follows: find element u V that
v V a u, v f v , (1)
where:
X h . Herewith the continuous variational problem (1) is approximated by the discrete one as
vh X h a uh , vh f vh . (2)
In result, the equation (1) is reduced to a system of linear algebraic equations from which the
In Finite Element Method (FEM) the space X h is constructed by partitioning the domain into
approximating functions (shape functions) that satisfy specific compatibility conditions on inter-
element boundaries. The fulfillment of the compatibility conditions is the major problem for the
requirements imposed onto the approximating functions from the space X h of finite elements. In
conventional FEM, the space X h is constructed in such a way that the following inclusion holds
Xh V (3)
If the spaces X h meet the condition (3) then the corresponding approximations of the space V
and equation (1) are called internal approximations. For the internal approximations the
closeness of the approximate and the exact solutions is evaluated by the norm of the space V .
Herewith the unique solvability of the continuous variational problem (1) guarantees the unique
solvability of the discreet variational problem (2), and the density of the family X h in the space
V guarantees the convergence of the approximate solutions to the exact one. Finite elements that
The construction of the conformed FE of high accuracy to approximate the Sobolev space
H m is not a simple task. When m 1 then the functions from the space X h must be
continuous, but their first derivatives can be discontinuous on inter-element boundaries. When
m 2 , the approximating functions must be continuous with their first derivatives. It
necessitates the use the complex structures of degrees of freedom even for finite elements of the
simplest form (triangle or quadrilateral for 2D problems). The problem becomes more
complicated in case of curved boundaries. Such boundaries have to be simulated with elements
that have non-planar faces. The curved elements are constructed through the mapping of some
initial FE of simple form. However, such mapping worsens the approximating properties of the
It is much simpler to construct FE which do not meet the condition (3). Such FE are called
introduce suitable spaces to compare the approximate uh and the exact u solutions (the
prove the existence and the uniqueness of the solution of the discreet variational problem
(2) (these features are not provided automatically like in the case of internal
approximations)
estimate the non-conformity errors and prove the convergence of the approximate
conducted, and the accuracy and robustness of the approximations are verified through numerical
28]. There had been developed several justification techniques. In 1965, in work [1] the patch
test was introduced as the criterion of convergence of non-conformed FE. The patch test has
been used to justify many non-conformed approximations applied to the specific problems [2 -
8]. In work [9] a variational treatment of the patch test was developed and a general estimate of
the non-conformity error of the approximations of the variational equations was derived. The
approximating functions from the space of FE at the inter-element boundaries. The proof of
non-conformity errors [5, 13, 14, 18 – 20, 29, 30]. In works [9, 12, 31] there was demonstrated
that for a particular type of problems, the success of patch testing is a sufficient condition for the
convergence of some non-conformed FE. Some of the non-conformed schemes were analyzed in
work [17]. The results of [17] were generalized in work [24] by establishing some sufficient
conditions of convergence. The iterative method of solution for schemes with non-conformed
In 1980, in work [23], there was developed an approximation that successfully passed the patch
test, but did not converge. Thus, it was demonstrated that, in general, the patch test is not a
sufficient condition for the convergence of non-conformed FE. The generalization of the patch
test was given in work [22], in which have been established the necessary and sufficient
conditions of convergence of the non-conformed approximations. The generalized patch test did
not depend on the type of the boundary value problem, but it was not constructive because in fact
it just stated that the error of non-conformity of the approximation must tend to zero when the
mesh of FE is refined. So, the justification of the specific non-conformed FE was performed in
work [22] for the model boundary value problems using the conventional technique for obtaining
In the above-mentioned works, the justification of the specific non-conformed FE was performed
for specific types of boundary value problems, and the properties of the corresponding equations
were extensively utilized. In such justifications, the proof technique usually assumed that a
subspace of conformed FE was present in the space of non-conformed FE. This assumption
reduced the applicability of the results obtained and prevented the development of general
only helped to find out what kind of non-conformed functions can be added to the basic
Another approach to treat non-conformed approximations is based on the concept of the external
approximations introduced in work [32]. The general theory of the external approximations of
Sobolev spaces and variational equations has been developed in work [33] where were also given
the examples of its application to the finite difference approximations and the approximations by
parts. In terms of the theory of external approximations some non-conformed FE have been
Let us note that the problem of justification of non-conformed FE is closely related to the
combinations have been studied in works [35, 36]. In work [37] it has been demonstrated that the
non-conformed basic functions can lead to a significant error in the solution. The combined
approximations are important when solving the boundary value problems in presence of
singularity in the solution, and for unbounded domains. For example, in works [38, 39] the
non-conformed combined approximation has been successfully applied for the solution of a
boundary value problem for a domain with a cut. Herewith in the circle that contained the tip of
the cut, the asymptotic representation of the solution for an infinite domain with a cut has been
used. On the rest of the domain the solution has instead been approximated using linear
triangular FE. In work [40], in terms of the theory of external approximations, it has been studied
the approximate fulfillment of the essential boundary conditions in a thin plate bending problem.
In work [41] it was shown that the methods of arbitrary accuracy could be developed through the
replacement of the interpolation of the boundary conditions with the orthogonalization of the
The external approximations of Sobolev spaces and variational equations by finite elements have
been systematically studied by the author in works [96-100]. This book generalizes the results of
these works.
1. Partitions, spaces, and boundary
operators
Basic definitions
Let us be an open bounded domain in R n . For any integer m 0 , the Sobolev space H m
1 ,..., n is a multi-integer, and a1 ... n ) belong to the space L2 of all functions
, m D D v L2
D v dx 1 vD dx
where dx dx1...dxn ; D is the space of all infinitely differentiable functions which have
compact supports in .
1
m 2
2
v D v dx
m, 0
1
2
2
v k,m D v dx
k
1 k m
The Sobolev space H 0m is defined as a closure of the space D by the norm m, .
A bounded domain in the Euclidean space R n is said to belong to the class C k , where k is a
non-negative integer, if there exists the finite number of local coordinate systems and local
xˆ r
Rn 1 : xnr
Here
xˆ r x1r , ..., xnr 1 , xˆ r
means that
xir , 1 i n 1
xˆ , x ;
R
r r
n xnr xˆ r , xˆ r
r 1
xˆ , x ; f xˆ x f xˆ ,
r r
n r
r r
n r
r
xˆ r
xˆ , x ; f xˆ x f xˆ ,
r r
n r
r r
n r
r
xˆ r
It is said that any domain of class C0 has a regular (or Lipshitz) boundary. The boundary (or
its part S) is called smooth if fr , 1 r R belong to the class C (i.e. if the boundary is a
called sectionally smooth if it can be divided into finite number of pieces i , of class C , so that
N
i .
i 1
partition of into a finite number of sub-domains K (finite elements). Let us the partition
K
K
for any K , domain K is closed and the set of its inner points K is not empty (i.e.
K 0 )
K1 K 2 0
V H K ,
K
m
K . (1.1)
Mutually orthogonal subspaces 0, 0, ..., H m
K , ..., 0 are contained in the space V . These
subspaces can be identified with the spaces H m K . Thus, the space V is a direct sum of its
V H K ,
K
m
K ,
v v
K
K
, K , (1.2)
The inner product and the norm in the space V are defined as follows:
u , v V u
K
K
, vK m, K
m
K
D u ,D v K
0, K
K 0
m
0 K
D u K
K
D v K dx
1
2
(1.3)
2
u u K
V K
m, K
1
m 2
2
K 0
D u K dx
1
m 2
2
K
0 K
D u dx
One can see that the norm m, is the restriction of the norm V on the space H m .
To approximate the space H m let us define a finite-dimensional space, also referred to as a
Xh P
K
h
K
V, K
where PhK are some finite-dimensional spaces of functions determined on the sub-domains
K .
For any FE K the space PhK approximates the space H m K ; the finite element space X h
approximates the space V . Approximating functions from the space X h are sufficiently smooth
on each FE of the partition ( PhK H m ( K ) ) but they are discontinuous across the inter-element
X h V L2 ,
and approximating functions from the space of FE do not belong to the space H m . The
functions are “external” with respect to the space H m and the concept of external
The formal definition of external approximations is as follows (see [33]). Assuming that:
h is a family parameter
u V u ph rh u V 0 .
Let us assume that V H m , and the space V , associated with the partition , is defined by
Vh R .
N h
Extension operators
ph : uh Vh phuh X h (1.4)
The primal goal is to establish conditions that must be met by the approximating functions
ph uh X h L2 () in order to restore the smoothness properties of the functions in limit, when
Let us there be given Hilbert spaces V and H , and the operator L V , T , such that:
maps V onto T
u, v a u, v v V0 .
The formal operator is a continuous linear operator from V into V0 . The domain of
V u V u H
1
2 2 2
u V u V
u H
.
Under the given assumptions there exists a unique operator such that the following Green’s
u V , v V a u, v u, v u , v , (1.5)
where , is the inner product defined on H H ; , is the duality pairing defined on T ' T
bilinear forms and spaces. Such formulas are widely used in the theory of elliptic boundary value
problems and in particular when proving the equivalence of weak and differential formulations
spaces let us introduce specific spaces, operators, and generalized Green’s formula associated
Let us partition the domain into FE, then let us determine the space V by the formula (1.1),
a u , v a u
K
K K
, vK
where
aK u K , v K u K
D v K dx
K
u K
u K
, v K
vK, m
The formal operator , associated with the form a u , v , is defined by the expression
u u
K
K K
, K
where
K u K 1 D u K ,
m
V H K V . m
K
Let us define a face K r of the FE K as a (n-1)-dimensional sub-domain of the boundary K ,
which either divides the adjacent elements or coincides with a patch of the boundary of the
M Kr ,l is the set of all smooth face sections K r , l of the face K r (in brackets a typical
In these notations symbols r and l identify the face and the smooth face section respectively
H L K L
2 2
K
V0 H K m
0 (1.6)
K
m 1
T H m j 1 2
K r , l
K K r K r , l j 0
0, 0, ..., H m j 1 2
Kr ,l , ..., 0
m 1
T H m j 1 2
Kr ,l
K Kr Kr , l j 0
and any element g T can be represented as follows
m 1
g g
K K r K r , l j 0
K
r ,l , j .
m 1
rK, l , j , (1.7)
K Kr Kr , l j 0
where rK,l , j is an operator of j-order differentiation along the outward normal to the smooth
the function u on K r , l .
According to the trace theorem from [33] each component rK,l , j of the operator is a
continuous linear mapping from the space H m K onto the space H m j 1 2 Kr , l . Hence the
operator maps the space V onto the space T . Since the space V is enclosed into the space H
with stronger topology, and the space V0 is a kernel of the operator that is dense in the space
H , then the assumptions of the Theorem 6.2.1 from [33] are met. So, for any , m there
exists a unique operator such that for any u , v V the following generalized Green’s
formula is valid
a u , v u , v u , v , (1.8)
where:
m
rK, l,,j
K K r K r , l j 2 m 1
K ,
r,l, j
L H m K , H m j 1 2 K r , l
m j 2m 1
Taking into account the representation (1.2), the formula (1.8) can be rewritten as:
D v K dx 1 D u
u , v V vdx u , u ,
K
u (1.9)
K K
where:
uK u K
; vK v K
.
The Sobolev space H m consists of functions v L2 such that all their generalized
D D
v dx 1 vD dx 0 ,
(1.10)
where D is the space of all infinitely differentiable functions that have a compact support in
.
Since the space D is dense in the space H 0m , it is possible to extend the expression
Let us v̂ denotes the function that is equal to the function v V in and that equals zero in a
function D R n from the space of infinitely differentiable functions, the expression below
holds
D vˆ dx 1
D vdx
K
ˆ 0, (1.11)
Rn K Rn
Because of the density of the space D R n in the space H m R n , the expression (1.11) can be
extended by continuity onto the space H m R n . Integrals in (1.11) are evaluated only on the
domain because vˆ 0 outside . It must be also noted that, since by assumption the domain
has a regular boundary, the space H m coincides with the space of restrictions on of all
Now let us consider the space V defined by (1.1). Spaces H m and H 0m are the closed
or H 0m readily follow from the formula (1.9). In fact, if the following inclusion holds
u u
K
K
H m
then
D u D u
K
K
.
Lemma 1.1
A function v V belongs to the space H m -and respectively to H 0m - if, and only if, for
any u H 0m -and respectively for any u H m - and for all , m , one of the two
u D v dx 1
D uvdx 0
K
K
u, v 0
Now, some properties of the boundary operators and that act in the spaces H m and
M r ,l is the set of common smooth face sections of all elements of the partition
For any r ,l M r ,l there exist the face sections Kr1,l and Kr2,l of adjacent FE K 1 and K 2
so that r ,l Kr1,l Kr2,l . The j-order derivative of a function u K H m K along the outward
j!
rK, l , j u K
D u
j
!
K
nr, l
where
The outward normals to the common face section of the adjacent elements K 1 and K 2 are
where
v1 v K1 , v2 v K 2 .
Here and below superscripts 1 and 2 denote the functions and the operators associated with the
Therefore, the operator maps the space H m onto the closed subspace T of the space T
characterized as follows
m 1
Tˆ g g K
r,l, j T r , l M r , l
K K r K r , l j 0 (1.12)
g 1r , l , j 1 g r2, l , j , 0 j m 1
j
Identically, considering the space H 0m it is apparent that the operator maps it onto the
closed subspace
Tˆ0 g Tˆ
g 0 . (1.13)
Lemma 1.2
Operator , m maps the space H 0m onto the orthogonal complement Tˆ of
◊ Proof – The operator maps the space H m onto T . According to Lemma 1.1:
, m
u H 0m , v H m u, v 0
u H m , v H 0m u, v 0
Now the characterizations of the spaces Tˆ and Tˆ0 can be established.It has been demonstrated
that:
f Tˆ , g Tˆ f, g 0. (1.15)
Taking into account the definition (1.13), the expression (1.14) can be rewritten as follows
m 1
f,g
K K r K r , l j 0
f rK, l , j , g rK, l , j
m 1
f r1, l , j , g 1r , l , j f r2, l , j , g r2, l , j (1.16)
r ,l j 0
m 1
where K , K r M K r , K r , l M K r , l , r , l M r , l .
Since (1.16) must hold for any g Tˆ , then the space Tˆ0 is defined by the characterization
Tˆ f Tˆ0 | K r, l M K r, l f r, l , j 0, 0 j m 1
where M K r, l is a set of smooth face sections of FE that coincide with the boundary of the
domain .
Figure legends for Section 1:
Fig. 1.1
Fig. 1.2
b) separated
2. The hallmarks of external
approximations by finite elements
Approximations on fixed and variable partitions
The first type is associated with a fixed partition of the domain . This type is an
The second type is associated with a family h of partitions. This type is an analogy of h-
version of FEM
Let us define a fixed partition of the domain and let us associate with it a family X h of FE
spaces
Xh P
K
h
K
, K
The absence of the parameter h in the partition designation indicates that each space X h is
associated with the same partition . The dependence between h and approximations must be
such that:
when h 0
dim PhK .
h max hK , hK diam K .
K
K h hK K (2.1)
where:
hK diam K ,
K sup diam S , S is a sphere from K
The family h is constructed through the partition refinement. Spaces PK of the approximating
functions are fixed; the number of FE - card h - tends to infinity when h tends to zero. The FE
Xh P
K
K
, K h .
For the approximations associated with the fixed partition , definitions (1.1) and (1.6) of spaces
V , V0 , H , T contain finite products of spaces. For approximations associated with the family
h of partitions, the finite products in expressions (1.1) and (1.6) are substituted by numerable
products, because when h 0 then card h . In this case, for instance, the space V
u u1 , u 2 , ..., u K , ... , uK H m K
such that
2
.
2
u V
uK
m, K
K
In order to establish the signs of the external FE approximations of Sobolev spaces, the duality
m 1
HT L K (2.2)
K Kr Kr , l j 0
2
r ,l
1
m 1 m 1 2 2
g g K
HT g g rK, l , j .
K Kr Kr ,l
r,l, j HT
K r K r , l j 0 j 0
K
0, K r , l
Space T , defined by (1.6), is continuously and densely enclosed into the space HT . According
to the Theorem 2.1.5 from [33], T and HT can be identified with subspaces which are dense in
the conjugate space T ' : T H T T ' , and these inclusions are continuous and dense. Then
the duality pairing , on T ' T can be identified with the unique extension of the inner
product , H in HT . Further in the text the closures of the spaces Tˆ , Tˆ0 in the space HT are
T
designated as Ĥ and Ĥ 0 , and the orthogonal complements of these closures in HT as Ĥ and
Ĥ 0 correspondingly. Thus:
Hˆ f Hˆ 0 | f
0 (2.4)
Functions that are determined on the face K r can be represented through their components
f rK, j f rK, l , j L2 K r , l , 0 j m 1.
K r , l
It must be mentioned that, for the fixed partition of the domain, the space HT is a fixed separable
Hilbert space. Therefore, there exists a family of closed subspaces that is dense in HT . When the
accordance with (2.2), take place. When h 0 the product of finite number of spaces in (2.2)
becomes a product of numerable spaces. In this case a family of spaces Rh is called dense in
space HT if, for any sequence xh H Th , there exists such a sequence yh Rh that
lim xh yh 0.
h0 HTh
Some auxiliary assertions
Now let us consider two auxiliary lemmas.
Lemma 2.1
Let us M be a Hilbert space and N M is a closed subspace. A bounded sequence h M
N lim , h 0 (2.6)
h0
if, and only if, there exists the family of closed subspaces Gh N which is dense in the space
gh Gh gh , h 0 (2.7)
where N is an orthogonal complement of the subspace N in the space M . Then any element
h h1 h2 h1 N , h2 N .
For any h , let us Fh1 be a one-dimensional subspace generated by the element h1 , i.e.
f Fh1 f h1 , R.
Because the space Fh1 is one-dimensional, therefore the orthogonal complement Gh of the space
Fh1 in the space N does exist and has a co-dimension equal to one.
The space N can be represented as a direct sum of its mutually orthogonal subspaces, i.e.
Then
Thus Gh is the family of the subspaces that is dense in N and possesses the property (2.8).
Sufficiency - Let us the relationship (2.7) holds. Then for any element N
Lemma 2.2
Let us M , N and h fulfill the assumptions of Lemma 2.1. If there exists a sequence g h M
such that
N , g h 0 , (2.8)
lim h g h 0 , (2.9)
h 0
N lim , h lim , h g h .
h0 h0
Then (2.6) follows from (2.9) and the inequality
, h g h h g h .ٱ
The necessary and sufficient sign of external FE approximations of a Sobolev space is defined by
Theorem 2.1
The FE approximations Vh , ph , rh are external approximations of the space H m -and
space H 0m respectively- if, and only if, the following condition is met
g h Gh , ph uh X h gh , ph uh H T
0, (2.10)
Proof – Let us consider the approximations of the space H m . Let us assume that a
sequence ph uh weakly converges in the space V . It means that, for any function H m
u D dx u D dx
h 0
(2.11)
K
K D u h
dx u dx
where uh phuh u
K
k
h .
According to Lemma 1.1, in order that u0 H m , the fulfillment of the following condition
H 0m , , m
lim D uhK dx 1 u
D dx
K
h
K
h 0
K
u dx 1
u D dx 0
0
According to Lemma 1.2, the operator maps H 0m onto the space Tˆ . Thus the necessary
result follows from (2.12) and from Lemma 2.1, in which it is assumed:
M HT , N Hˆ , h ph uh .
Let us consider the approximations of the space H 0m . Let us the sequence ph uh weakly
converges in space V , i.e. relationship (2.11) holds. According to Lemma 1.1, in order that
, m, H m lim , ph uh 0. (2.13)
h 0 HT
Since the operator maps H m onto Tˆ0 (see Lemma 1.2), the necessary result follows
M HT , N Hˆ 0 , h ph uh . ٱ
The following theorems determine two sufficient signs of the external FE approximations of
Sobolev space.
Theorem 2.2
The FE approximations Vh , ph , rh are external approximations of the space H m -and space
ph uh X h lim ph uh ph vh HT
0 . (2.14)
h0
, m, ph vh Yh
, ph vh HT
0. (2.15)
Let us consider the approximations of the space H m . Let us phuh be a sequence that
converges weakly in the space V . Then, the sufficiency of the condition (2.14) follows from
M HT , h ph uh , gh ph vh .
When approximations of the space H 0m are considered, the sufficiency of condition (2.14)
m 1
j 0
j
where
m 1
j ij rK, l , i
K K r K r , l i 0
0 j m 1 (2.16)
1, i j
ij
0, i j
m
j 2 m 1
j , m
where
m
j ij rK, l,,i
K K r K r , l i 2 m 1
m j 2m 1 (2.17)
1, i j
ij
0, i j
m 1
HT H
j 0
T, j ,
where
m 1
HT , j ij L
2
Kr , l
K K r K r , l i 0
.
1, i j
ij
0, i j
m 1
Hˆ Hˆ
j 0
j
m 1
Hˆ 0 Hˆ
j 0
0, j
where
Hˆ 0, j f HT , j | r , l M r , l
f r1, l , j 1 f r2, l , j (2.18)
j
Hˆ j f Hˆ 0, j | f
0 (2.19)
Theorem 2.3
Let us Vh , ph , rh be a family of FE approximations of the space H m . Let us represent the
set I of values of the index j , 0 j m 1 as a sum of its not intersecting and not empty
subsets I I1 I 2 .
Approximations Vh , ph , rh are external approximations of the space H m -and the space
For any j I 2
where Ghj is a dense family of closed subspaces of the spaces Hˆ j -and Hˆ 0, j respectively-.
Here spaces Hˆ j and Hˆ 0, j are defined by the characterizations (2.19) and (2.18).
Proof – Let us consider approximations of the space H m . Let us the sequence phuh
converges weakly in the space V , i.e. the relationship (2.11) holds. According to Lemma 1.1, in
order that u0 H m , it is necessary and sufficient that the following condition is fulfilled
, m, H 0m
h 0
lim , ph uh HT
0. (2.22)
If the representations (2.16) and (2.17) are taken into account, the inner product in (2.22) can be
rewritten as follows
m 1
, ph uh HT
2 m j 1 , j ph uh HT
. (2.23)
j 0
h 0
lim 2m j 1 , j phuh HT
0. (2.24)
Let us j I1 . Because of the inclusion Yhj H j 1 , the following relationship holds
(2.25)
, j ph uh
2 m j 1 0
HT
Thus, when j I1 , then (2.24) follows from (2.20), (2.25) and Lemma 2.2 in which it is assumed
When approximations of space H 0m are considered, the result can be proven in a similar
way. ٱ
The Theorems 2.1-2.3 can be applied not only to justify well-known non-conformed finite
elements (in Section 7.1 it is done for the elements of Wilson, Crouzeix-Raviart, Morly, etc.), but
also to develop novel highly efficient finite elements and approximations. Some methods of
Let us consider how to construct FE that meet the necessary and sufficient sign established in
Theorem 2.1. There are two types of functions in the formulation of the theorem:
onto the FE K belong to the spaces H m ( K ) but the functions itself belong to the space
L2 () only
Let us Gh be a family of finite-dimensional spaces. These spaces will be referred to as the spaces
of boundary functions. They are elaborated in Section 5. Here, to satisfy the requirements of the
Theorem 1.1, the properties of the approximating functions from the spaces PhK H m K are
established. The types of finite elements that meet the conditions of the theorem are defined, as
g h Gh Hˆ , ph uh X n
m 1
g h , ph uh H T
K K r K r , l j 0 K r , l
g rK, l , j rK, l , j p K d (3.1)
m 1
g
r ,l j 0 r ,l
1
r,l, j r1, l , j p1 g r2, l , j r2, l , j p 2 d 0
where
K , K r M K r , K r , l M K r , l , r , l M r , l
p1 P1 H m K 1 , p2 P2 H m K 2 ,
and where
g rK, l , j is a restriction of a function from the space Gh onto the smooth face section K r , l
of the face K r
Because of the inclusion Gh Hˆ and of the characterization (2.4), restrictions of the functions
from the space Gh onto the boundary of the domain equal zero, and so are the
Let us assume that basic functions of the space Gh H have local supports - they are the sub-
empty multitude of internal points K r , j 0 (it is also possible that K r , j r ). Further it will
be shown that such choice of the basis in the space Gh leads to such basic functions of the finite
Let us GrK, j designates a finite-dimensional space generated by the basic functions grK,l ,t whose
When the approximations of the space H 0m are under consideration, then the inclusion
Gh Hˆ 0 takes place. In this case, due to the characterization (2.5), the boundary basic
functions with supports on the boundary of the domain must be added to the functions
pi Pi H m K i , g ri , j , t Gri , j , i 1, 2; 0 j m 1
where
rK, j
K r , l
K
r ,l , j .
Let us consider FE K 1 and adjacent elements K 2 , ..., K N (see Fig. 3.1). Let us G K
GK GrK, j , 0 j m 1, K r M K r .
j , Kr
pi Pi H m K i , g ri , j , t G i ,
1 i N , 0 j m 1, 2k N (3.4)
For (3.4) to hold there must exist such functions in the spaces Pi ,1 i N that meet the
g ri , j , t ri , j pi d r , j , t . (3.5)
K 1
For i 1 , the integrals in the left hand side of (3.5) are continuous linear forms (functionals) on
the space H m K 1 (further, the designation K is used instead of K 1 ). Let us denote these
i p i , 1 i M, (3.6)
Let us the functionals i ,1 i M from (3.5) be linearly independent on the space PK . Then:
P K P PZ , (3.7)
where
PZ p P K | i p 0, 1 i M (3.8)
condition
1, i k
i pk (3.9)
0, i k
M N M
p i p pi
i 1
p p
k 1
k
Z
k , (3.10)
of the space PZ .
F p 1 p , ..., M p
1, ..., M RM , 0
such that
M
p 0
i 1
i i
for any p P K , i.e. the functionals i are linearly dependent, which contradicts the initial
assumption.
equal to one and the other components equal to zero. Due to the surjectivity of the mapping F ,
F pk f k .
Then there exists a set of linearly independent elements p1 , ..., pM P K that satisfy the
condition (3.9).
Let us P designates a space generated by the elements pi , 1 i M . Then the subspace PZ
expressed as follows
p p1 p2 ,
where
M
p1 p p
i 1
i
i P
M
p2 p i p pi PZ
i 1
then
M
p2 c p
i 1
i
i
c p
M
k p2 i k
i
i 1
ck k p2 0, 1 k M
freedom that can be formulated as follows: in the space PK there exists at least one function that
satisfies the conditions (3.6) for any set of numbers i . When dim P K card K , the
-unique-solvency of the multitude K (see [10, 27]). When dim P K >card K the multitude
K is PK -poly-solvent and P -unique-solvent. In this case there can exist several subspaces
K is a closed domain in R n with not empty multitude of internal points and sectionally
domain K
freedom is PK -unique-solvent
Let us consider how to construct the basic functions of spaces P and PZ . Let us P H m H
be some finite-dimensional space, that below will be referred to as initial, and let us
N
p P p b p
k 1
k k , (3.11)
N
i p b p ,
k 1
k i k 1 i M
RbT T , (3.12)
where:
b is a vector of coefficients b bk k 1
N
where:
Let us the block R be non-singular. Such a block does exist if the rank of the matrix R equals
(3.13) from the left by the matrix R1 , the expression for the vector b through the vectors bZ
and is obtained
1 pk k 1 ; 2 pk k N M 1 .
M N
where
where:
One can see that the element p P P K has been represented by the elements of the new
M
pk pr
i 1
i ik , 1 k M
M M
p Zj p j M pl rlk rkjZ , 1 j N M
l 1 k 1
initial space P and then to linearly transform it. Herewith, the initial space P and the space PK
coincide.
The Theorem 3.1, that determines the structure of FE basis, is based on the assumption of linear
existence of the representation (3.10). The question remained is whether it is always possible to
Let us card K M , dim P K N . Further in the text, dimensions N of the space PK are
varied. For the purpose of differentiation between the spaces PK of different dimensions, the
parameter h is introduced into the designation: instead of PK one writes PhK , assuming that
h 1 N.
It is apparent that:
solvent even if h 0 . In other words, if there are no requirements to the content of the space
PhK , then however big is the dimension N of the matrix R , it does not guarantee the equality of
The following theorem defines a sufficient condition for the existence of such h0 beginning with,
Theorem 3.2
with some h h0 , it is sufficient that the family PhK is dense in the space H m K .
Proof - Let us i p , 1 i M be a multitude of the functionals (3.5) that are the boundary
degrees of freedom of FE. Let us show that these functionals are linearly independent on the
space H m K due to the linear independence of the basic functions g rK, j , t of the space G K of
Trace operators
rK, j
K r , l
K
r ,l , j , 0 j m 1
which is densely enclosed into the space L2 Kr . Let us assume that the functionals (3.5) are
r , j ,t R M , 0
m 1 m 1
r, j, t
j 0 K r
grK, j , t rK, j u K d
j 0 K r K r
r, j, t grK, j , t rK, j u K d 0 .
t K r t
t
r , j ,t g rK, j , t 0
That contradicts the assumption about linear independence of the basic functions g rK, j , t .
Now let us assume that the family of the spaces PhK is dense in H m K but there is no such h0
beginning with, for which the multitude of the functionals i p , 1 i M becomes linearly
independent on the spaces PhK . It means that for any h there exists a vector
M
p PhK p 0 .
i 1
i i (3.17)
From (3.17) and Theorem 1.2 from [33] it follows that the family PhK is not dense in the space
multitudes, where:
freedom is P -unique-solvent
When constructing basic functions of the element, according to the procedure described in
Section 3.1.2, it is necessary first to introduce some system of basic functions that generates the
initial space PK . Then one has to choose a subspace P P K on which the multitude K of
boundary degrees of freedom of FE is P -unique-solvent. But in the space PK there can exist a
K, G K
, P K , P .
If dim P K N, dim G K M and N M , then the maximum number of subspaces P
that can be extracted from the space PK is equal to the number of combinations
N!
CNM .
M ! N M !
M M from the matrix R in the system (3.12). It can be performed by the permutation of the
columns in the matrix R (and respectively the elements of vector of the initial system of
basic functions).
As it will be shown in Section 4, the convergence rate of the approximations is determined by the
approximation qualities of the space PK and does not depend on the choice of the subspace P .
The choice of P can only influence the magnitude of the constants in the estimates of the
approximation error. Thus, all conformed by subspace FE, characterized by the same multitudes
In relation to the above said, it makes sense to introduce the virtual FE as a triplet K , G K , P K ,
subspaces P .
p i P i , g ri , j , t G i , 1 i N, 0 j m 1
(3.18)
g ri , j , t ri , j p i d 0
K 1
From (3.18) follows that for the element K K 1 , the space PK is a finite-dimensional subspace
i p g
K
r , j ,t rK, j pd , 1 i M . (3.19)
K
A homogeneous FE is defined as a triplet K , G K , PZ
K
, where:
K is an element’s domain
(3.19)
Let us consider the procedure for the construction of the basic functions pkZ , 1 k N of a
pk k 1
N
homogeneous element. Let us choose a system of basic functions determined on the FE
K . The system generates the initial space P which is not necessarily a subspace of the kernel Z
, P Z . Any p P is expressed as
N
p b
k 1
k pk , (3.20)
b p 0,
k 1
k i k 1 i M,
RbT 0 , (3.21)
It is assumed that M N . If the rank r of the matrix R equals to N , then the system (3.21) has
the trivial zero solution only, and, therefore, the subspace PZK is empty. If r N then the
dimension of the space of solutions of the system (3.21) is N r . In this case a square non-
singular block R1 can be extracted from the matrix R by the permutation of the equations and
r N r.
Multiplying the equation (3.22) by the inverse matrix from the left the following complete
p
N
From the expression (3.25) it follows that the basic functions Z
k , N N r
k 1
r r
pkZ pk r pl rlj1rjk2 ,
l 1 j 1
where:
As one can see a homogeneous element has only internal degrees of freedom and does not have
any boundary degrees of freedom. It will be shown below that the homogeneous FE generate
variable partitions of the domain . However, these elements are useful when constructing the
K̂ K , (3.26)
Gˆ K G K , (3.27)
PˆZK PZ . (3.28)
K
Then the element K , G K , PC , P , where:
conformed by subspace FE. The equality (3.26) means the same geometry of the elements.
PˆZK P 0 . (3.30)
The properties (3.28) and (3.30) provide the expansion of the space PK due to the addition of
new functions from the space PˆZK . It can be useful for improvement of the approximation
qualities of FE.
support of every basic function g r , j , t , from the space Gh of the boundary functions, is either a
face K r coinciding with the boundary of the domain , i.e. supp gr , j , t K r (here and
satisfy the relation (3.2). Herewith, in accordance with Section 3.1, for the adjacent FE K 1 and
p1 P1 , p 2 P 2 , 0 j m 1
r1, j , t p1 r2, j , t p 2 ,
where:
r1, j , t p1 g
1
r1, j p1d
r , j ,t
r
(3.31)
r2, j , t p 2 g r2, j , t r2, j p 2 d
r
are the boundary degrees of freedom of the elements K 1 and K 2 respectively. Then the spaces
characterization
X h v pK PK | g r , j , t Gh Hˆ
K
K
r
g 1r , j , t r1, j p1 d g r2, j , t r2, j p 2 d ,
r
0 j m 1
X 0h v p K PK | g r , j , t Gh Hˆ 0
K
K
g p d g r2, j , t r2, j p 2 d ,
1 1 1
r, j,t r, j
r r
g r , j , t rK, j p K d 0, 0 j m 1
K r
Basis of the space Gh generates the multitude h of boundary degrees of freedom of the space
of FE. Every degree of freedom from the multitude h is a pair r1, j , t , r2, j , t whose elements
are determined by the relations (3.31). Any FE also possesses a possibly empty multitude of
B K i , 1 i N K M
N K dim P K
M card K dim P
Bh B K K , card Bh N h M h
K
where N h N
K
K
, M h dim Gh .
Any element from the space X h of conformed by subspace FE can be uniquely defined by
The basis of the space X h consists of the basic functions of two types:
wrˆ, ˆj , tˆ X h
g
ri , j , t wrˆ, ˆj , tˆ i
r, j, t
ri , j wrˆ, ˆj , tˆ |K d
i (3.32)
r
1
i 1
when r rˆ, j ˆj , t tˆ, i 1, 2
0 otherwise
j , ˆj identify the orders of the derivatives that participate in the definition of boundary
degrees of freedom
faces r rˆ ,
orders of differentiation
Basic functions of the space X h are constructed from the basic functions of FE in the following
pi Pi , i 1, 2 are the basic functions from the spaces Pi of the elements K i that
pi x ,
x K i , i 1, 2
w
0, x Ki ,
is the basic function of the space X h corresponding to the boundary degree of freedom and
Let us p Z PZ designates a basic function from the space PZ of the element K corresponding
p x,
xK
Z
w
Z
(3.34)
0, xK
is the basic function (3.33) associated with the internal degree of freedom .
Thus, the support of any basic function wr, j , t X h consists of two adjacent FE (see Fig. 3.2). A
support of a function wkZ is a single FE. Functions wkZ and wr, j , t are linearly independent and
Let us now examine the spaces of homogeneous FE. For such FE the basic functions of the space
Gh of boundary functions do not generate boundary degrees of freedom of FE. The space X h of
X h v p K PZK | g r , j , t Gh Hˆ
K
K
r
g 1r , j , t r1, j p1 d g r2, j , t r2, j p 2 d 0,
r
0 j m 1
Any element from the space X h can be uniquely expressed through the multitude
degree of freedom b , then the basic function w Z X h is defined by conditions (3.34). Its
Therefore, the space X hC of the composite FE is a direct sum of the corresponding spaces
X hC X h X h . (3.35)
A function from the space X hC is uniquely defined through the multitudes Bh and Bh . Basic
functions from the space X hC are constructed from the basic functions of the conformed by
Fig. 3.1
Fig. 3.2
Let us
AK v K v SK v (4.1)
K : v p P is defined as follows
M
K v v p
i 1
i
i , (4.2)
S k : v p PZ
N M
SK v v p
i 1
i i
Z
is the best approximation in the space L2 K of the function v by the shift of the subspace
PZ P K , i.e.
p PZ v K v SK v, p 0, K 0. (4.3)
Because the subspace PZ has the basis p , 1 i N M ,
i
Z
then the equation (4.3) is
v K v SK v, piZ
0, K
0 1i N M (4.4)
The approximation (4.1) of the function v is unique by virtue of the P -unique-solvency of the
linear independence of the system of equations (4.4). Because the boundary degrees of freedom
p P K AK p p (4.5)
Let us space X h is the space of the conformed by subspace FE with the multitude of boundary
Ah v h v Sh v . (4.6)
Here
Mh
hv v w
j 1
j
j , (4.7)
where wj , 1 j M h are the basic functions (3.32), and i h , 1 j M h are the
v v S v, w
h h
Z
i 0,
0, 1 i Lh , (4.8)
j h v j v , 1 j M h .
Because the basic functions wiZ , 1 i M of the space X h coincide with the basic functions
piZ ,1 i N M of the spaces PZ P K , and the support of the function piZ is a single
element, then the system (4.8) is equivalent to the finite set of systems of sort (4.4) formulated
i Bh , Sh v X h i Sh v i v , 1 i Lh ,
j Ah v j v , 1 j Mh
i Ah v i v , 1 i Lh
The domain of determination of the operator Ah of X h -approximation is the space H m , i.e.
dom Ah H m .
Between the operators AK and Ah there exists the dependence determined by the following
theorem.
Theorem 4.1
Let us v be an arbitrary function from the space H m . Then restrictions v K belong to the
space H m K and
K Ah v K AK v K . (4.9)
Proof - Relations (4.9) follow from the definitions of the multitudes h and Bh .
AK* of PZK -approximation can be defined as an operator of orthogonal projection in the space
p PZK v A v, p
*
K 0, K
0. (4.10)
If p ,
Z
k 1 k N * is a basis of the space PZK , then the relation (4.10) is equivalent to the
where
N*
AK* v b v p
k 1
k
Z
k .
The uniqueness of PZK -approximation of the function v follows from the linear independence of
dom AK* L2 ( K ) H m K ,
p PZK AK* p p
Ah* : v Ah*v X h*
v A v, w
*
h
*
i 0, 0, 1 i Nh* ,
where wi* , 1 i N h* are the basic functions of the space X h* . Because the support of each
basic function wi* is a single FE, then any function Ah*v X h* is characterized by the following
property
bi Ah* v bi v , 1 i N h* .
Let us consider a composite FE K, G K
, PCK , P , where PCK P PZ PˆZK . The PCK
AKC v K v S KC v , (4.12)
S KC : v p PZ PˆZK
defines the best approximation in the space L2 K of the function v by the shifts of the space
PZ PˆZK , i.e.
p PZ PˆZK v K v SKC v, p
0, K
0. (4.13)
be a basis of the space PˆZK . Then the equation (4.13) is equivalent to the system of the equations:
v K v S KC v, piZ
0, K
0, 1 i N M
. (4.14)
v K v S K v, p j
C
ˆZ
0, K
0, 1 j Nˆ *
Herewith the PCK -approximant of the function v is unique because of the P -unique-solvency of
the multitude K of boundary degrees of freedom of FE and because of the linear independence
dom AKC H m K
(4.15)
p PCK AKC p p
Let us X hC be the space of composite FE. For an arbitrary function v H m , its X hC
-approximant is:
AhC v h v S hC v , (4.16)
v v S
h
C
h v, wiZ 0,
0, 1 i Lh
v v S
h
C
h v, wˆ iZ 0,
0, 1 i Nˆ h*
dim X hC Nh Nˆ h* .
It is easy to verify that the operators AhC and Ah* possess the property determined by the Theorem
rh : v H m vh 1 ,..., M h , 1 ,..., Lh R Nh Vh
rh* : v H m vh 1 ,..., N * R Nh Vh*
h
*
(4.17)
Theorem 4.2
with a fixed partition of the domain and built on the conformed by subspace FE.
Approximations Vh , ph , rh converge in H m if, and only if, for each FE K the family
Necessity - Let us assume that the external approximations Vh , ph , rh converge, i.e.
u H m lim u ph rh u V
0
h0
where ph rh Ah is the operator of X h -approximation. Then the density of the family PhK in the
spaces H m K , K follows from the definition (1.2) of the norm on the space V and from
K and for any h , the PhK -approximant of a function v K u K is defined according to (4.1)
Mh Nh M h
AK , h v K Ah u k i, h v K pi, h
i 1
i 1
i , h v K piZ, h , (4.18)
where:
From the family of spaces PhK let us chose such sequences of functions pˆ hK PhK that
K lim v K pˆ hK 0. (4.19)
h 0 m, K
These sequences exist because, according to the assumptions, the families of the spaces PhK are
v K AK , h v K v K pˆ hK pˆ hK Ak , h v K
m, K m, K m, K
which yields
u Ah u V
u uˆh V
uˆh Ah u V ,
here
uˆh pˆ
K
K
h , K .
lim pˆ hK AK , h v K 0. (4.20)
h 0 m, K
According to the Theorem 3.1, the element pˆ hK PhK can be uniquely expressed as follows
Mh Nh M h
pˆ hK i, h pˆ hK pi, h
i 1
i 1
i , h pˆ hK piZ, h . (4.21)
lim i , h v K pˆ hK 0
h0
Therefore
i , h pˆ hK i , h v K
i , h pˆ hK i , h v K
and the relation (4.20) is the implication of the representations (4.18) and (4.21).
Let us note that the density of the family PhK in the space H m K is the sufficient condition for
the PhK -multi-solvency of the multitude of the boundary degrees of freedom of FE (see Theorem
3.2), as well as it is also necessary and sufficient condition for the convergence of the
Theorem 4.3
External FE approximations Vh* , ph , rh* of the space H m associated with a fixed partition
Proof - Let us consider an element K . The family of subspaces PhK is dense in the space
H m K if, and only if, any continuous on H m K functional that equals zero on PhK , equals
zero on H m K also (see Theorem 1.2 from [33]). For homogeneous FE functionals (3.18)
equal zero on the elements from PhK . Obviously, functionals (3.18) do not turn to zero for any
function v H m K and, therefore, families of the spaces PhK are not dense in the spaces
H m K .
Theorem 4.4
with a fixed partition of the domain and built on composite FE. For the approximations
Vh
C
, ph , rhC to be converging in the space H m it is necessary and sufficient that families
PhK of the spaces of approximating functions of conformed by subspace FE contained in the
Proof – Proof of this theorem follows from (3.29), (3.35), Theorem 4.2 and from the definition
Let us h be a family of partitions of the domain , which satisfies the conditions of regularity
(2.1). In classic FEM, the partition h of the domain is usually conducted in the way that all
FE K h are linearly equivalent to some initial FE K̂ , also referred to as common FE (see, for
example, [10]). In this case any element K h can be obtained from the common element K̂
F : xˆ Rn F xˆ Bxˆ b Rn , (4.22)
then the following spaces can be associated with the spaces PK and G K
P*K p : K R
p pˆ F 1 , pˆ Pˆ ,
G*K g : K R g gˆ F 1 x , x K , gˆ Gˆ
where P̂ and Ĝ are the spaces of functions defined on the domain K̂ and on the boundary K̂
P*K P K , G*K G K .
So the relations between the points x̂ Kˆ and x K and between the functions p̂ Pˆ and
xˆ Kˆ x F xˆ K
pˆ Pˆ p pˆ F 1 P K (4.23)
gˆ Gˆ g gˆ F x G , x K
1 K
pˆ xˆ p x xˆ Kˆ , pˆ Pˆ ,
gˆ xˆ g x xˆ Kˆ , gˆ Gˆ
Let us Kˆ , Gˆ K , Pˆ K , Pˆ be some arbitrary conformed by subspace FE that is characterized by the
condition
diam Kˆ 1 . (4.24)
For any element K, G K
, P K , P , K h , through Kˆ , Gˆ K
, Pˆ K , Pˆ let us designate some
K F Kˆ
PK p : K R p pˆ F 1 , pˆ Pˆ K
(4.25)
GK g : K R g gˆ F 1 x , x K , gˆ Gˆ K
Moreover, let us the mapping (4.22) be a similarity mapping, when B is a diagonal matrix with
bii hK , 1 i n,
hk diam K .
Kj v hK j Kj vˆ
ˆ
d hKn 1 d ˆ (4.26)
0 j m 1
Let us modify the definition of the boundary degrees of freedom of FE by adding the multipliers
hK
n j 1
:
v hK n j 1 g K Kj vd ,
K r
where g K is a basic function of the space Gh of boundary functions. Let us note that the
multipliers do not change the definition of FE given in the Section 3.1.1. The modification
allows to prove the important relation between the operators of Pˆ K - and PK -approximations,
Theorem 4.5
K, G
Let us Kˆ , Gˆ K , Pˆ K , Pˆ , K
, P K , P be two similar conformed by subspace FE. If
pˆ i , 1 i M , pˆ Zj , 1 j N M
pi , 1 i M , p Zj , 1 j N M
are the basic functions of the element K , then for the operators of approximation AK and Aˆ K
AK v Aˆ K vˆ
^
vˆ dom Aˆ K v vˆ F 1 dom AK .
M N M
AK v i v pi
i 1
v p
j 1
j
Z
j ,
equations as in (4.4):
N M
v p
M
j 1
j
Z
j , ptZ
0, K
v
v p
i 1
i
i , ptZ ,
0, K
1t N M. (4.27)
Taking into account that F is the similarity mapping which maintains any normal orthogonal to
j 1
j
Z
j , pˆ tZ
0, Kˆ
vˆ
ˆ vˆ pˆ
i 1
i
i , pˆ tZ ,
0, Kˆ
1t N M. (4.29)
f , g 0, K ˆˆ
fgdx fgh
n
K dxˆ hKn fˆ , gˆ 0, Kˆ
K Kˆ
j 1
0, K
hK n v
v p
i 1
i
i , ptZ ,
0, K
1 t N M . (4.30)
j v ˆ j vˆ , 1 j N M .
Therefore
M N M
AK v ˆi vˆ pi ˆ vˆ p j
Z
j .
i 1 j 1
From this, using the relations (4.23) one can finally obtain
M N M
AK v ˆi vˆ pˆ i ˆ vˆ pˆ Aˆ K vˆ .
^ Z
j j
i 1 j 1
Error estimates for piecewise polynomial approximations
In the estimates below C designates a constant independent of the parameter h of the family of
Theorem 4.6
Let us Kˆ , Gˆ K , Pˆ K , Pˆ be a conformed by subspace FE. Let us m 1 be the highest order of
normal derivatives that participate in the definition of the boundary degrees of freedom of FE. If,
H k 1 Kˆ H m Kˆ , (4.31)
Pk Kˆ Pˆ K H m Kˆ , (4.32)
then there exists such a constant C that for all similar FE K , G K , P K , P and for any function
u H k 1 K
hKk 1
u AK u l , K C l u k 1, K , 0 l m, (4.33)
K
pˆ Pk Kˆ Aˆ K pˆ pˆ .
Let us û be a function from the space H k 1 Kˆ . Because of the inclusion (4.31), this function
also belongs to the space dom Aˆ K H m Kˆ . The linear mapping
AˆK : H k 1 Kˆ H l Kˆ
is continuous because of the inclusions (4.31) and (4.32). Herewith, the operators of Pˆ K - and
u dom AK AK u Aˆ K uˆ .
^
Then the necessary result follows from the Theorem 3.1.4 from [10].
Theorem 4.7
any FE from this family, inclusions (4.31) and (4.32) take place, in which K is substituted by K̂
. Then there exists such a constant C that for all FE from the family and for any function
u AK u l, K
ChKk 1 l u k 1, K , 0 l m. (4.34)
Proof - For any element K from the family it can be put in correspondence through the
similarity mapping some initial element K̂ that meets the relations (4.24) and (4.25). Then the
estimates (4.33), in which constants C vary for different elements, are true. Because the
constants C are bounded then the estimate (4.34) follows from the estimates (4.33) and the
Theorem 4.8
associated with a regular family of partitions h of the domain and built on the conformed by
subspace FE. Let us there exists such whole number k 0 when m k allowing the following
K h Pk K P K H m K .
Then, there exists such a constant C independent from h , that for any function u H k 1
1
2
Ch k 1 l u k 1, , 0l m
2
l, K
u Ah u
K
u Ah u V Ch k 1 m u k 1,
Using the relations (4.9) and the inequalities hK h, K h one obtains the following
1 1
2
2
u AK u u k 1, K
k 1 l
Ch Ch k 1 l u k 1,
2 2
K K
l, K
1
2
u AK u u Ahu
2
.
K
m, K V
Convergence of approximations
family h of partitions of the domain , i.e. the conditions when the following holds
u H m lim u ph rh u V
0,
h0
Theorem 4.9
Let us Vh , ph , rh be a family of external FE approximations of the space H m associated
with a regular family h of partitions of the domain and built on conformed by subspace FE.
If the inclusions
K h Pm K P K H m K
v H m 1 v ph rh v V
Ch v m 1, .
Therefore
lim v ph rh v V
0. (4.35)
h 0
u ph rhv V
uv V
v ph rhv V . (4.36)
By introducing the function u H m and the arbitrary number 0 , one can determine the
u v V
2
h h0 v ph rh v V
.
2
Therefore, taking into account the inequalities (4.36), it is obtained
u H m lim inf u ph vh V
0 .
h 0 v !h Vh
Theorem 4.10
regular family of partitions h of the domain and built on the homogeneous FE does not
Proof - In case of homogeneous FE for any member of the sequence ph uh X h* the following
g
K
Kj ph uh K d 0, 0 j m 1, (4.37)
K
where g K is a restriction on K of the basic function from the space Gh of boundary functions.
However, for any function u H m V the functionals (4.37) do not become zero.
Therefore, according to the Theorem 1.2 from [33], it is impossible to segregate the sequence
ph uh from the family of spaces X h* converging to the given function u H m in the norm
of the space V .
Theorem 4.11
Let us VhC , ph , rhC be a family of external FE approximations of the space H m associated
with a regular family h of partitions of the domain and built on composite FE. If the
K h Pm K PCK H m K
Proof – Proof of the Theorem is similar to the proof of Theorem 4.9 because the properties
(4.15) and (4.9) of the operators AhC and AKC of the approximation, determined by the relations
(4.12) and (4.16), are identical to the properties (4.5) and (4.9) of the operators Ah and AK for
families of spaces Gh of boundary functions that are dense in the boundary spaces Hˆ , Hˆ 0 (see
characterizations (2.5) and (2.6)). The approximations of the spaces Hˆ , Hˆ 0 by the spaces Gh
are referred to as boundary approximations, and they are examined in this Section.
Here and below C designates various constants independent of parameter h of the family of the
Let us R n be an open domain with sectionally smooth regular boundary , and let us
N
j, N
j 1
and let us assume that each of the domains j can be continuously and one-to-one mapped onto
x j t x , t t1 , ..., tn 1 D j R n 1 (5.1)
where
j t C0 Dj , 1 j N.
It is assumed that v H S if v j H S D j . Herewith:
If one map is sufficient to define the surface , then such surface is called elementary surface.
Local maps (5.1) can be built by straightening part j of the surface through the introduction
of new coordinates. Herewith the coordinates ti ti x are defined in the vicinity of point
x0 in such a way that within these coordinates, part of the boundary j has the equation
tn 0 and part of the domain , adjacent to it, is located in subspace tn 0 . The coordinates
ti ,1 i n can be introduced in the following way. Smooth section j is associated with the
Cartesian coordinate system y1 , ..., yn so that j can be defined through the equation
yn f y , y y1 , ... yn 1 (5.2)
and variables x are substituted by variables y . Then new variables t are introduced by the
formulas
ti yi , 1 i n 1,
tn yn f y .
Further, for simplicity, only elementary surfaces are considered. Norm and semi-norm on the
Dt v
2
1 grad f dt
2
v
s,
D s
1
(5.3)
2
v s, Dt v
2
1 grad f dt
2
D s
where:
The expression
d 1 grad f dt
2
When s 0 , the expression (5.3) defines the norm on space the L2 induced by the inner
product
u, v 0, uvd .
Let us be given a domain Rn with sectionally smooth regular boundary . Let us a
and H T (see characterizations (2.2), (2.4) and (2.5)) the following spaces are defined
m 1
W H
s j 1
K ,
r ,l s j 0,
K K r K r , l j 0
W0 f W | r ,l M r ,l f r1,l , j 1
j
f r ,l , j , 0 j m 1
2
(5.4)
W f W 0
| f
0
where K , K r M K r , K r , l M K r , l .
In the characterization (5.4), figures 1 and 2 indicate that the functions are defined on the inter-
1
m 1 2
g rK, l , j
2
g W K K K j 0 s j 1, K r , l
.
r r ,l
1
2
g j g rK, l , j
2
K K K s j 1, K r ,l
(5.5)
r r ,l
0 j m 1
Let us there be given some partition of the domain into FE K . In Section 3.1 it was
K r , j supp g r , j , t r , K r, j 0 (5.6)
or a connected sub-domain of the face K r* of FE that coincides with the boundary , i.e.
g r , j , t g 1r , j , t , g r2, j , t ,
Because of the inclusions Gh Hˆ Hˆ 0 the components of the pair satisfy the relation
Gh GK
K
, K (5.7)
where
m 1
GK G
K r j 0
K
r, j . (5.8)
The assumption (5.6) allows the construction of the spaces of FE whose basic functions have a
local support in the domain . With a specific choice of the functions g r , j , t , it also allows to
prove the density of the family Gh of the spaces of boundary functions in the space Ĥ 0 or Ĥ .
It will be shown in the Section 6 that, in order to obtain optimal error estimates of the
Let us designate through g r , l , j , t a basic function of the space Gh whose support is:
K r , l , j supp g r , l , j , t r , l , K r ,l , j 0
or a connected sub-domain of the smooth section Kr*, l of the face K r* that coincides
Kr ,l , j Kr* ).
m 1
GK G
Kr Kr , l j 0
K
r ,l , j , (5.9)
Let us designate through PrK, j (and through PrK,l , j correspondingly) the finite-dimensional spaces
of the basic functions g rK, j , t (and grK, l , j , t correspondingly) that have a common support K r , j (and
K r , l , j correspondingly). By definition, the spaces PrK, j and PrK,l , j are not empty. The domains K r , j
and K r , l , j , along with their spaces PrK, j and PrK,l , j , are referred to as surface elements. Therefore, a
Let us conduct the partitions r , j , 0 j m 1 of the face K r into the surface elements
Kr Kr , j , Kr , j r , j , 0 j m 1
Kr , j
for any K r , j r , j the domain K r , j is closed and the multitude of its internal points is not
empty: K r , j 0
for various arbitrary Kr1, j , Kr2, j r , j , K r1, j K r2, j
GrK, j P
Kr , j
K
r, j L2 K r Kr , j r. j , 0 j m 1 .
GrK, l , j P K
r,l, j L2 K r , l K r , l , j r , l , j , 0 j m 1.
Kr ,l , j
to define m layers of surface elements on the boundaries of FE. It is possible that partitions of
r,l, j r,l, k , r, j r, k
0 j m 1, 0 k m 1
0 j m 1
Bh : Hˆ 0 Gh
that can be represented as follows (depending on the patches of boundaries being partitioned into
the surface elements):
m 1
Bh B
K K r j 0
K
r, j (5.14)
or:
m 1
Bh B
K Kr Kr , l j 0
K
r ,l , j . (5.15)
The orthogonal projector Bh is referred to as the operator of Gh -approximation. Let us call the
m 1
BK B
K r j 0
K
r, j (5.16)
m 1
BK B
K r K r , l j 0
K
r ,l , j
as operators of G K -approximation.
From the definition of the operator of Gh -approximation and the operator of G K -approximation
K Bh g |K BK g . (5.17)
The relation (5.17) is similar to the relation (4.9) between the operators of PK - and X h
-approximation.
: D R n 1 R n
Since the partition of the domain is fixed, then the spaces Ĥ and Ĥ 0 are fixed as well.
The spaces L2 are the elementary components of the spaces Ĥ and Ĥ 0 . Dense in the
particular, for this purpose, complete systems of algebraic polynomials and other systems of
analytical functions can be used. Below are closely examined the approximations of the spaces
Let us define a family of partitions h of the surface into the surface elements
K h so that the partition requirements described in Section 5.2 are satisfied. Let us also
where:
hK diam K ,
K sup diam S , S is a sphere from K
The surface elements K h are scaled in the same way as FE K h in Section 4.3.1.
Let us designate for any surface element K h through K̂ the element that satisfies the
following condition
Herewith
K F ( Kˆ ) ,
bijection of the mapping F one can associate with the space PK the following space
P p : K R | p pˆ F1 , pˆ PˆK
where PˆK is a space of functions determined on the domain K̂ . Since F is a similarity
tˆ Kˆ t F tˆ K
pˆ PˆK p pˆ F1 PK
is the equality
Theorem 5.2
Let us Kˆ , Pˆ , K , P
K
K
be two similar surface elements. Then, for the orthogonal
projectors:
RK : L2 K PK
RˆK : L2 Kˆ PˆK
which is valid for all functions vˆ L2 Kˆ , v L2 K that are in the relation
vˆ L Kˆ v vˆ F
2
1
L2 K
◊ Proof – Let us pi , pˆ i , 1 i N are the basic functions of the spaces PK and PˆK
N
RK v v p ,
i 1
i i
algebraic equations
v p , p v, p j 0, K ,
N
i i j 0, K
1 j N. (5.21)
i 1
ˆ vˆ pˆ , pˆ vˆ, pˆ j 0, Kˆ ,
N
i i j 0, Kˆ
1 j N. (5.22)
i 1
p,p
i j 0, K
pi p j dt
K
pˆ i pˆ j hKn 1 dtˆ
Kˆ
hKn 1 pˆ i , pˆ j 0, Kˆ
ˆi vˆ i v , 1 i N
Theorem 5.3
Let us Kˆ , PˆK be a surface element. If for some whole integer s 0 , the following inclusions
hold
Ps Kˆ PˆK L2 Kˆ (5.24)
then, there exists such a constant C , that for all surface elements K , PK and any function
where RK g is the projection of the function g onto the space PK , and
hK diam K
K sup diam S , S is a sphere from K
◊ Proof - On the space PˆK , the orthogonal projector Rˆ K coincides with the identical operator,
i.e.
The required inequality (5.25) follows from (5.26) and the relation (5.20) between orthogonal
projectors RˆK , RK , and from the Theorem 3.4.1 from [10]. ٱ
Theorem 5.4
Let us there be a regular family of surface elements K , PK and for any element from the
Ps K PK L2 K , s 0.
Then, there exists such a constant C that for all elements from the family and for any function
◊ Proof - For every element K there can be put in correspondence some initial element K̂
which satisfies the relation (5.19). Herewith the space PK can be characterized as follows
PK p : K R | p pˆ F1 , pˆ PˆK .
Then, for any element K from the family, the inequality (5.25) holds. The inequality (5.27)
follows from the estimates (5.25), the conditions of regularity (5.18), and the bounded constants
Let us h be a regular family of the partitions of the surface patch into the surface
elements K h . Let us assume that for some whole number s 0 and for any
K h
Ps K PK L2 K ,
constant C that
where
BK h : L2 GK h
is the orthogonal projector from the space L2 onto the space of surface elements:
GK h P
K
K
L2 ; K h (5.29)
and:
◊ Proof - From the Theorem 5.4, for any K h the following inequality holds
g H s 1 K g RK h g ChKs 1 g s 1, K .
0, K
According to the representation (5.29) of the space GK h , the following relation exists
g L2 , K h B h g |
K
K RK g .
hK h, K h
one obtains:
g H s 1
1 1
2
2
g R g h g g
2 s 1
gB Ch Ch s 1 g
K K 2
0, s 1, K s 1,
K K
0, K
Let us s 0 then
g H 1 g BK h g Ch g 1, (5.30)
0,
and
lim g BK h g 0.
h 0 0,
The relation (5.28) follows from the relation (5.30) and the density of the enclosure
H 1 L2 .ٱ
To approximate the spaces H m , it is necessary to define m layers of the surface elements
on the boundaries of FE. The parameter of a regular family of partitions of the face
Kr M Kr (or smooth patch K r , l M K r , l of the face K r ) in general depends on the
patch under consideration and on the layer number. Let us designate through r , j hr , j the
family of partitions of the face K r for the j -layer of surface elements. Herewith
hr , j max diam K r , j .
Kr , j
Similarly through r , l , j hr , l , j , let us designate the family of partitions of a smooth patch K r , l
of face K r . Here
hr , l , j max diam K r , l , j .
Kr ,l , j
The parameter h of the family of boundary approximations is defined by one of the following
relations:
h max hr , j , (5.31)
K , Kr , j
or:
h max
K , K r , K r , l , j
h ,
r ,l , j (5.32)
where K , K r M K r , K r , l M K r , l , 0 j m 1 .
Theorem 5.6
Let us be a fixed partition of the domain into FE K . Herewith each face K r of any
Also, let us assume that for any surface element K r , j r , j hr , j the following enclosures hold
P0 K r , j PrK, j L2 K r , j , 0 j m 1
where PrK, j is the space of approximating functions of the surface element. Then, the family of
spaces Gh of boundary functions defined by relations (5.7) and (5.8) is dense in the space Ĥ 0 ,
i.e.
g Hˆ 0 lim g Bh g HT
0, (5.33)
h 0
where:
where operators BrK, j hr , j are defined by (5.11). Then for any K , Kr M Kr
g L2 K r lim g BrK, j hr , j g 0,
hr , j 0 0, K r
and the relation (5.33) follows from (5.31), (5.14), (5.16) and from the relation (5.17) between
Let us be a fixed partition of the domain into FE K and let us every patch K r , l of any
face of FE be an elementary surface. Let us assume that there are regular families
Ps j K r , l , j PrK, l , j L2 K r , l , j ,
where PrK,l , j are the spaces of approximating functions of the surface elements, and
s j 0, 0 j m 1 .
Then, the family of spaces Gh of boundary functions defined by (5.7) is dense in the space Ĥ 0 ,
i.e.
g Hˆ 0 lim g Bh g HT
0, (5.34)
h 0
where the operator Bh and the parameter h are defined by the expressions (5.15) and (5.32)
C j , 0 j m 1 that
m 1
C h
s j 1
g W0 g Bh g HT
j g j, (5.35)
j 0
where the space W0 is defined by the characterization (5.4); j are semi-norms (5.5)
where operators BrK, l , j hr , l , j are defined by the expression (5.13). Let us also
K , K r M K r , K r , l M K r , l
the equality
1
m 1 2
g BrK, l , j hr , l , j g
2
g Bh g (5.37)
HT j 0 K K K 0, K r , l
r r ,l
the enclosure
P0 K r , l , j Pk K r , l , j , k 0
K g BrK, l , j hr , l , j g
s j 1 s 1
g H r,l Chr ,jl , j g s j 1, K r , l
. (5.38)
0, K r , l
The required inequality (5.35) is the implication of (5.38), (5.37), and the definition (5.32) of the
parameter h .ٱ
Let us define a regular family of partitions h of the domain into FE K h . The parameter
h of the family has a meaning of the maximum diameter of FE. When refining the partition, then
h 0 and the number of finite elements tends to infinity: card h . In this case there are
For any FE K h let us designate through K̂ a similar FE that satisfies the condition
diam Kˆ 1 . (5.39)
Herewith:
K F Kˆ
diam K hK (5.40)
n
F : xˆ R F xˆ Bxˆ b R
n
where:
F is a similarity mapping
b is a vector of dimension n
Let us the patch ̂ of the surface of the element K̂ be defined by the map
ˆ : Dˆ R n 1 ˆ R n , (5.41)
where ̂ is a homeomorphism; D̂ is the domain of the parameters tˆ tˆ1 , ..., tˆn 1 of the map.
Similarity mapping (5.40) transforms the coordinates of the point Aˆ ˆ in the global and local
xi hK xˆi bi , 1 i n
t j hK tˆj , 1 j n 1
In this case the domain of parameters Dˆ Rn 1 of the map (5.41) is transformed into the similar
domain D R n 1 . Therefore, with the similarity mapping (5.40) is associated the similarity
mapping
F : tˆ R n 1 F tˆ Btˆ R n 1 , (5.42)
hK . Herewith D F Dˆ .
Let us establish the relations between the points tˆ Dˆ , t D and the functions
gˆ L2 Dˆ , g L2 D as follows:
tˆ Dˆ t F tˆ D (5.43)
gˆ : Dˆ R g gˆ F
1
: D R (5.44)
The equality
ĝ tˆ g t
holds for all points tˆ and tˆ in the correspondence (5.43) and for any functions ĝ and g in the
relation (5.44).
The mapping (5.42) associated with the mapping (5.40) establishes a one-to-one correspondence
between points of domains D and D̂ of patches and ̂ for similar elements K and K̂ . Then
P g:R| g gˆ F1 , gˆ PˆK
that, due to linearity of the mapping F , coincides with space PK , i.e. P PK .
Theorem 5.8
Let us PˆK and PK be the finite-dimensional spaces of functions defined on the boundary patches
RK : L2 PK
RˆK : L2 ˆ PˆK
R g RˆK gˆ
K ^
(5.45)
for all the functions gˆ L2 ˆ , g L2 from (5.44).
◊ Proof - The similarity mapping (5.40) of the element K̂ is associated with the similarity
mapping (5.42). Scales of these mappings are identical. Further course of proof of this theorem is
Theorem 5.9
Let us K̂ be a finite element, D̂ be the domain of parameters of the map of a boundary patch ̂
Ps Dˆ PˆK L2 Dˆ (5.46)
then, there exists such a constant C that for all elements K similar to K̂ and for any function
g H s 1 D
where:
hK diam K
mapping F : Dˆ D which has the same as F scale hK . On the space PˆK the orthogonal
projector
RˆK : L2 Dˆ PˆK
Then the inequality (5.47) follows from the relation (5.45) between orthogonal projectors Rˆ K
and RK and from the Theorem 3.1.4 from [10]. ٱ
Theorem 5.10
Let us there be given a family of finite elements, and any of the elements has a surface patch K
with a map
: DK R n 1 K R n .
Let us assume that the family of the domains DK of the patches is regular, i.e. there exists such a
constant that
where
If, for any FE from the family and for some whole number s 0 the following enclosures hold
Ps DK PK L2 DK ,
where PK is a finite-dimensional space of functions determined on the domain DK , then there
exists such a constant C , that for all elements from the family and for any function
g H s 1 DK
◊ Proof - To any element from the family, it can be put in correspondence some similar element
K̂ defined by the relations (5.39) and (5.40). Therefore, for every FE there exists such a constant
g H s 1 DK g RK g C hKs 1 Dl K g s 1, DK
. (5.50)
l , DK
The inequality (5.49) follows from the inequalities (5.50), the regularity conditions (5.48), and
Theorem 5.11
Let us define a regular family h of partitions of the domain into FE K h in such a way
that every face K r of any FE from the family is an elementary surface. Let us assume that for
elements K r , j r , j . Let us assume that for any surface element K r , j r , j , the spaces PrK, j of
P0 K r , j PrK, j L2 K r , j .
Then:
gh Hˆ 0h lim gh Bh gh HTh
0, (5.51)
h 0
for any:
K h , Kr, j r, j , 0 j m 1 ,
one obtaines:
m 1
Zh H K ,
K K r j 0
1
r K h , K r M K r
Taking into account the estimate (5.52), the definitions (5.14) and (5.16) of the orthogonal
projectors Bh and BK , and the inequality hK h , one obtains that for any function g h Z 0h
1
m 1 2
g h Rr , j g h
2
K
0, K r , j
K Kr j 0 Kr , j
1
(5.53)
m 1 2
g h Bh g h Ch g h
2
HTh K K j 0 K 1, K r , j
r r, j
Therefore
g h Z 0h lim g h Bh g h HTh
0.
h0
Then the relation (5.51) follows from the inequality (5.53) and the density of enclosure
Theorem 5.12
Let us define a regular family h of partitions of the domain into FE K h in such a way
assume that for each K h are defined the regular partitions r , l , j , 0 j m 1 of the smooth
patches K r , l into the surface elements K r , l , j r , l , j . Let us allow that for any surface element
Ps j K r , l , j PrK, l , j L2 K r , l , j
h max hK , K h
K
that
m 1
C h
s j 1
g h W0h g h Bh g h HTh
j gh j
(5.54)
j 0
one obtains
g H
s j 1
K
r,l, j g RrK, l , j g
0, Kr , l , j
ChKj
s 1
g s j 1, Kr ,l , j
,
where RrK, l , j is the orthogonal projector (5.12). From this, considering the definitions (5.7) and
(5.9) of the orthogonal projectors Bh and BK and the inequality hK h , it is obtained that for
1
m 1 2 m 1
C h
2 s j 1
g h Bh g h gh R K
gh gh j .
HTh K K K r,l, j 0, K r , l , j j
r r ,l j 0 Kr ,l ,l j 0
The relation (5.55) follows from the inequality (5.54) and the density of the enclosure
form
m
a u, v a x D
p , q 0
pq
p
uD q vdx ,
where a pq x L .
u, v a u, v v V0 ,
D q a pq x D p u .
m
u 1
q
p , q 0
V u u H m , u L2
Elliptic boundary value problems for the operator are equivalent to the variational equations
of this fact is based on the existence of Green formula that corresponds to the given boundary
value problem. For instance, for a boundary value problem with mixed boundary conditions the
W u H m | u 0
0
where is the trace operator (1.7); 0 is a bounded open sub-domain of the boundary .
The equivalency means that if the function u V satisfies some boundary conditions and
u f , f L2 (6.1)
a u, v f v v W (6.2)
The external approximations of the equation (6.2) are constructed in the following way. Let us
conduct the partition of the domain into FE K and let us define the spaces
V , V0 , H , T by the relations (1.1) and (1.6). Let us introduce the continuous on V V bilinear
form
m
a u, v a x D
K p , q 0 K
pq
p
u K D q v K dx (6.3)
where u K u K
, v K v K . It is apparent that the form a u, v is the restriction of the form
N h
the determination of the element uh Vh R that satisfies discrete variational equation
vh Vh a ph uh , ph vh f ph vh , (6.4)
Theorem 6.1
Let us the element f V in the equations (6.2) and (6.4) is given and the following conditions
hold:
C1 0 v W a v, v C1 v
2
V
(6.5)
C2 0 v V a v , v C2 v
2
V
(6.6)
Then the variational problems (6.2) and (6.4) have unique solutions u -and uh correspondingly-
a u, ph wh f ph wh
u ph uh C inf u ph uh sup (6.7)
V
vh Vh V
wh Vh ph wh V
◊ Proof – The existence and the uniqueness of the solutions of problems (6.2) and (6.4) follows
from: the inequalities (6.5) and (6.6); the enclosure ph vh X h V for any h and for any
C2 ph uh ph vh a ph uh ph vh , ph uh ph vh
2
V
a u ph vh , ph uh ph vh f ph uh ph vh a u , ph uh ph vh
Then
f ph wh a u, ph wh
C2 ph uh ph vh V
C3 u ph vh V
sup .
wh Vh ph wh V
u phuh V
u phvh V
phuh phvh V
The first term in the right-hand side of the estimate (6.7) satisfies the following inequality
1
2
inf u ph vh u Ah u u AK u
2
, (6.8)
K
vh Vh V V m, K
determined in Section 4.1. Then one can see that the first term depends on the smoothness of the
exact solution u of the variational problem (6.2) and on the properties of the approximating
functions p P K of the finite elements. The magnitude of this term is referred to as an internal
error of approximation. For the approximations built on a family of partitions of the domain and
conformed by subspace FE, the estimates of the internal errors were obtained in Section 4.3.
dh a u, ph wh f ph wh .
a u, ph wh f ph wh
dh V
sup . (6.9)
wh Vh ph wh V
In order to obtain the estimate of the external errors let us derive the corresponding Green’s
Generated by the form (6.3) formal operator is defined by the following relations
u , v a u , v v V , 0
u u , K ,
K K
D q a pq x D p u K .
m
K uK 1
q
p , q 0
Let us be the trace operator (1.7) for the partition of the domain. For the spaces V , H , V0 , T
and for the operator hold the assumptions of the Theorem 6.2.1 from [33] and therefore, there
exists a unique operator such that it holds the following Green’s formula:
u , v V a u , v u , v u , v , (6.10)
m
rK, l , j , (6.11)
K K r K r , l j 2 m 1
rK, l , j L H m K , H
m j 1
2
K
r ,l
(6.12)
m j 2m 1
Theorem 6.2
Let us the variational problem (6.2) is approximated with FE that are built by the necessary and
sufficient sign of the external approximations established in the Theorem 2.1 (i.e. conformed by
subspace, homogeneous, or composite FE). Then there exists such independent of parameter h
constant C that the external error of the approximation satisfies the inequality
dh V
C u Bh u HT
, (6.13)
where:
a u, ph wh u, ph wh u, ph wh .
Let us multiply the equation (6.1) by the function ph wh and integrate it over the domain , so
u, ph wh f , ph wh .
The following equality holds
u, ph wh u, ph wh ,
and then
a u, ph wh f ph wh u, ph wh .
Let us identify the duality relation , on T T with the unique expansion of the inner product
g h Gh ph wh X h gh , ph wh H T
0,
u V , g h Gh
a u, ph wh f ph wh u , ph wh H u g h , ph wh H (6.14)
T T
u Bh u, ph wh H u Bh u HT
ph wh HT
T
Here, the operator is continuous and it maps the space V into the space HT . Therefore, there
v V v HT
C v V
,
and then
ph wh X h V ph wh HT
C ph wh V
. (6.15)
The inequality (6.13) follows from the inequalities (6.14) and (6.15). ٱ
by the discontinuity of the approximating functions from the space of finite elements on
problem
From the Theorem 6.2 it follows that, when the approximations are built by the necessary and
sufficient sign, then the external error depends on the smoothness of the element u H T and
Let us consider the approximations for which the spaces Gh are the spaces of surface elements
introduced in the Section 5. There were formulated two approaches to define the layers of the
surface elements:
in the first one it is assumed that the faces K r of FE undergo the partition into the
the partition into the surface elements (see Theorems 5.7 and 5.12)
By assumption, each patch K r , l is an infinitely differentiable manifold in R n of dimension (
n 1 ).
Let us the solution u of the variational problem (6.2) belongs to the space
H 2m k , m 0, k 0 .
Boundary operator defined by (6.11) and (6.12) maps the space H 2m k onto the product
of spaces
H 2 m k j 1 2 Kr , l (6.16)
K Kr Kr ,l j 2 m 1
Then, independently on smoothness of the exact solution u of the variational problem, the
functions:
u*j
K r , l
r ,l , j u, m j 2m 1
belong only to the space L2 Kr even if their components are smooth enough:
r , l , j u H 2 m k j 1 2 K r , l . (6.17)
Functions u*j , m j 2m 1 are approximated by the surface elements from the space Gh of
the boundary functions. If faces K r undergo the partition into the surface elements, then due to
the insufficient smoothness of the function u*j , the estimates (6.13) will not depend on the
smoothness of the solution u of the variational problem. Different situation occurs when into the
surface elements are partitioned smooth patches K r , l of faces. In this case the surface elements
approximate each component (6.17) of the product (6.16) that allows obtaining the estimates of
the external errors depending on the smoothness of the exact solution u and on the type of the
boundary functions.
Theorem 6.3
Let us the solution u of the variational problem (6.2) belongs to the space
H 2m k , m 0, k 0 . Let us assume that the external approximations are built with the
spaces Gh of the boundary functions that meet the conditions of Theorem 5.7 (or Theorem 5.12).
m 1
C h
s j 1
dh V ! j u j , (6.18)
j 0
rK, l , 2m j 1 L H 2m k K , H k j Kr , l , 0 j m 1.
Let us s j k j 1 then
L H 2 m k , W0 ,
where the space W0 is defined by the characterization (5.4). Then inequality (6.18) follows from
If the approximations of the variational problem (6.2) are built on FE that satisfy the sufficient
sign of the external approximations defined in Theorem 2.2, then the external error satisfies the
inequality
ph wh ph vh
u
HT
dh V
sup , (6.19)
HT
wh Vh ph wh V
where functions ph vh are defined according to the assumptions of the Theorem 2.2, i.e.
phvh Yh H m (6.20)
◊ Proof - From the enclosure (6.20) and Lemma 1.1, one has
ph vh Yh u, p v h h HT
0.
a u, ph wh f ph wh u, ph wh ph vh H u HT
ph wh ph vh HT
.ٱ
T
From the inequality (6.19) it follows that, if the approximations are built by the sufficient sign,
then the external error depends on the approximating quality of the spaces Yh H m of the
conformed approximating functions. Similarly one can obtain the estimate of the external error
for the approximations built by the sufficient sign established in the Theorem 2.3.
Theorem 6.5
Let us the conditions of the Theorem 6.1 hold. Let us the FE approximations of the space V are
where u and uh are the solutions of the variational problems (6.2) and (6.4) correspondingly.
◊ Proof - When h 0 , the internal and the external errors in the right-hand side of the
inequality (6.7) tend to zero, since by the assumption the approximations are external and
Let us analyze the solvability and the convergence of the approximate solutions of a discrete
variational problem in case when the solution of the continuous variational problem (6.2) exists
and is unique, but the expansion a u , v of the bilinear form a u, v onto the space V V
Theorem 6.6
Let us a partition of the domain into FE K be given and the following conditions hold:
boundary functions gˆ rK, j , t , 0 j m 1 that for any face Kr M Kr of any FE
W u V | r M r gˆ
1
r, j, t r1, j u1d gˆ
2
r, j, t r2, j u 2 d ;
r r
(6.22)
K , K r M K r M K r gˆ rK, j , t rK, j u K d 0
K r
1
2
uK
2
u W u (6.23)
K
W m, K
rK : u H m K rK u Pm 1 K
M
rK u u p
i 1
i
r r
i , (6.24)
p is some basis of the space Pm 1 K . Here the symbol r identifies the face
r M
where i i 1
Kr M Kr of FE. The representations (6.24) exist because of the linear independence of
Let us
u W u W
0.
Then on each FE
uK u K
Pm 1 K .
According to the characterization (6.22), for the projector that corresponds to the face
1r u K1 0,
which implies
u K1
0. (6.25)
Let us consider the element K 2 that has some common with the element K 1 face
K r2 K r1 r . From the equality (6.25) and the characterization (6.22) it follows that for this
face
gˆ
1
r, j, t r1, j u K1 d gˆ 2
r, j, t r2, j u K2 d 0.
r r
2r u K2 0,
and so
u K2
0.
Considering in a similar manner the rest of the finite elements of the partition, one comes to the
conclusion that
K u K
0,
and so u 0 because each FE of the partition has at least one face that it shares with the
Theorem 6.7
Let us the assumptions of the Theorem 6.1 hold. Let us define the enclosed subspace W V
with the norm (6.23) in accordance with the requirements of the Theorem 6.6. Let us the space
C 0 v W a v, v C v
2
W
.
Then the variational problems (6.2) and (6.4) have the unique solutions u and uh
a u, ph wh f ph wh
u ph uh C inf u ph vh sup (6.26)
W
vh Vh W
wh Vh ph wh W
◊ Proof - Proof of the Theorem is similar to the proof of the Theorem 6.1. ٱ
Theorem 6.8
Let us there be given a fixed partition of the domain . Let us the conditions of the Theorem
6.7 be satisfied. Let us the FE approximations of the space W be the converging external
lim u phuh V
0 (6.27)
h0
where u and uh are the solutions of the variational problems (6.2) and (6.4) correspondingly.
◊ Proof - Since approximations of the space W are the external approximations of the space W ,
then the internal error (6.8) and the external error (6.9) tend to zero when h 0 . Then relation
(6.27) follows from inequality (6.26) by taking into account the equivalence of the norms W
Theorem 6.9
Let us there be given a regular family h of partitions of the domain . For each partition from
the family let us be satisfied the conditions of the Theorem 6.7. Let us FE approximations of the
lim u phuh Wh
0 (6.28)
h0
where u and uh are the solutions of the variational problems (6.2) and (6.4) correspondingly.
◊ Proof – The relation (6.28) follows from the inequality (6.26) taking into consideration the
tendency towards zero of the internal and external errors of approximations. ٱ
The theorems show that even in the case when the bilinear form a u , v is not V -elliptic, it is
possible to define some minimal set of boundary degrees of freedom generated by the space Ĝ
which provides the unique solvability of the discrete variational problem at any partition of the
domain. If the approximations are built on a fixed partition of the domain, and if for any h ,
spaces Gh of the boundary functions contain the minimal subspace Ĝ defined in the Theorem
6.6, then takes place the convergence of the approximate solutions to the exact one, not only in
the norm W , that is a semi-norm on the space V , but in the norm V as well. In general, for
approximations built on a family of partitions of the domain, the convergence can be established
Many of classic non-conformed finite elements have been justified in works [9, 15 - 17, 22, 42 -
46]. Usually such justification is performed for a certain type of boundary value problem and the
features of the differential equations of the problem are essentially used in the justification.
Below, the theorems from the Section 2 are applied to prove that all of the examined FE generate
Spaces of the approximating functions of each of the elements under consideration contain the
spaces of the polynomials of m -degree. This feature provides the convergence of the
corresponding external approximations. Proofs of the convergence are not given here due to their
simplicity.
Wilson's element
Wilson's element is used for the approximation of boundary value problems of the second order,
ai , 1 i 4 n 1 .
The space PK of the approximating functions of FE can be defined by two equivalent ways:
n
P K P2 K L xi
i 1 (7.1)
P Q1 K L x , 1 i n
K 2
i
where:
i.e.
p Q1 K p
b1 ,... n x11 x22 ...xn n .
i 1
1 i n
p ai , 1 i 4 n 1
K 2 p .
x 2 dx, 1 i n
K i
Let us define a regular family h of partitions of the domain into FE. Let us introduce the
space X h of FE, whose functions ph vh X h are defined by their values in the vertices (nodes)
of the partition (here ph is an extension operator (1.4)) and by their average values, as follows
2 ph vh |K
xi2 dx, 1 i n, K h
K
where ph vh |K P K .
The functions from the space X h are discontinuous on the inter-element boundaries and are
Theorem 7.1
P K Q1 K (7.2)
K p ai , 1 i 4 n 1 (7.3)
Let us Yh C 0 be the space of the conformed FE defined by the relations (7.2) and (7.3), so
Yh H 1 . (7.4)
For any function phuh X h , let us designate through ph vh the unique function from the space
Yh that takes the same values as ph uh in the vertices of the partition h . Let us estimate the
magnitude of
phuh ph vh HT
,
phuh ph vh
K K r K r , l
K
r , l ,0 p K rK, l ,0 p*K ,
onto the face K r K r , l of FE. Due to the construction of the function ph vh , the function
) of the function rK, l ,0 p K P2 K r , l . Then, according to the Theorem 3.1.4 from [10], there
Therefore
lim phuh ph vh HT
0 (7.5)
h 0
The inclusion (7.4) and the property (7.5) make the conditions of the Theorem 2.2 fulfilled. ٱ
value problems of the second order. The domain K R 2 is a simplex with the vertices
p P K p |Kr ,l Pk K r , l (7.6)
where K r , l K r is a face of FE.
K p bi , 1 i 3k ,
7.1.c).
Theorem 7.2
Finite elements from Crouzeix-Raviart’s family generate the external approximations of the
space H 1 .
quadrature on the common face Kr1, l Kr2, l r , l of adjacent FE, the approximating functions
p1 P1 , p2 P2 p1 bi p2 bi 1 i k.
Since Gauss's quadrature of k -order is exact for polynomials of degree 2k 1 , and since the
r ,l
g p1
K r1, l d g p d
r ,l
2
Kr2,l
(7.7)
Let us associate with the family of partitions h the family of spaces Hˆ 0h defined by (2.3), and
Gh G
K K r K r , l
K
r,l, 0 Hˆ 0h ,
where GrK, l ,0 Pk 1 K r , l .
According to the Theorem 5.12, the family Gh is dense in the space Ĥ 0 where the space Ĥ 0 is
For the adjacent FE K 1 and K 2 , functions from the spaces GrK, l ,0 satisfy the relation
Using the designations introduced, the relation (7.7) can be rewritten as:
g h Gh , ph uh X h gh , ph uh H T
0,
Let us note that according to the Theorem 6.3, for sufficiently smooth solutions of variational
Morley's element
Morley's element is used to approximate the solutions of boundary value problems of fourth
order. The domain K R2 is a triangle with the vertices ai ,1 i 3 . The space of the
PK P2 K . (7.8)
p ai , 1 i 3
K p aij (7.9)
, 1 i, j 3
n
where
p aij
n
are the derivatives along the normal in the middles of the sides of the triangle (Fig. 7.1.d).
Let us define a regular family h of triangulations of the domain . The approximating
Derivatives along normal to the FE boundary are continuous in the middles of the sides of
Theorem 7.3
PK P1 K (7.10)
K p ai , 1 i 3 . (7.11)
The corresponding space of FE is designated through Yh0 . The element defined by (7.10) and
Yh0 H 1 (7.12)
Due to the characterizations (7.8) and (7.9) the inclusion Yh0 X h holds as well.
For each function phuh X h , through ph vh let us designate a function from the space Yh0 that
has the same values as ph uh in the vertices of the triangulation h . Let us introduce the
designations
p ph uh |K , p* ph vh |K .
rK, l , 0 p , so
and, therefore
ph uh X h lim 0 ph uh 0 ph vh HT
0, (7.13)
h0
where 0 is a trace operator of zero order for the partition of the domain. It is defined by the
expression (2.16).
Let us consider the adjacent FE K 1 and K 2 . The derivatives along normal to the common side of
ph uh |K1 ph uh |K 2
ph uh X h
n
a
ij
n
a .
ij
By designating through r ,l Kr1,l Kr2,l this common side, one can write:
where ri , l ,1 is the operator of differentiation of the first order along outward normal to FE
boundary; pi ph uh |K i , i 1, 2 .
The middle of the common side of the adjacent FE is a point of the Gauss's integration
quadrature of the first order that is exact for polynomials of the first degree. Since
p P K P2 K rK, l ,1 p P1 K r , l ,
Let us define the family of spaces of the boundary functions (see characterization (2.18))
1
Gh1
K Kr Kr , l j 0
1j GrK, l , j Hˆ 0,1
h , (7.16)
where GrK, l ,1 P0 K r , l .
According to the Theorem 5.12, the family Gh1 is dense in the space Ĥ 0,1 that is the limit space
for Ĥ 0,1 h when h 0 .
For the adjacent FE, due to the inclusion (7.16), the following equality holds:
and therefore
where 1 is the trace operator of the first order for the partition of the domain. The operator is
that the approximations built on Morley's FE are the external approximations of the space
H 2 . ٱ
Fraeijs de Veubeke’s element is used to approximate generalized solutions of the boundary value
problems of the fourth order. The domain K R 2 is a triangle and the space of the
PK p P3 K | f p 0 ,
where f p is some functional whose expression is not necessary here. The multitude of
p
K p ai , 1 i 3; p aij , 1 i j 3;
1
d ,
meas K r , l K n
r ,l
where:
ai , aij are the vertices and the middles of sides of the triangle correspondingly
p n is the derivative along normal to the side K r , l K r of the triangle (Fig. 7.1.d)
functions from the space X h of Fraeijs de Veubeke’s FE are continuous in the vertices and the
◊ Proof - Proof of the Theorem is similar to the proof of the Theorem 7.3. The only differences
are that the conformed space Yh0 has to be built on the quadratic triangular FE defined as follows:
P K P2 K ,
K p ai , 1 i 3; p aij , 1 i j 3
and that the relation (7.17) immediately follows from the definition of the degrees of freedom of
In the classic Finite Element Method, strict requirements are imposed on the element shapes, on
the location of the element nodes, and on the approximating functions of the element. This is
caused by the necessity to provide sufficient smoothness of the approximating functions from the
space of FE on the inter-element boundaries. The only requirement to the geometry of FE for
external approximations is that it must have sectionally smooth regular boundary. Such FE can
and they are defined independently on the types of approximating functions that represent field
variables inside FE. This property is reflected in the definitions of P -multi-solvency of the
multitude of boundary degrees of freedom of FE (see Section 3.1). Internal degrees of freedom
of FE are defined automatically after boundary degrees of freedom and a space of approximating
of FE and the approaches to the definition of boundary degrees of freedom are discussed.
k
p
b x ,
0
xK,
The number k is called the order of internal approximation on FE. Polynomial approximations
are convenient in computations and, as at will be demonstrated in Section 8, they provide high
It is assumed also that a face of the element consists of one smooth patch only, so K r K r , l .
: D R Kr ,l R2 , (7.19)
where D is the domain of determination of the parameter t . Let us assume that the mapping
(7.19) is a polynomial of l -degree. When l 1 , the patches K r , l are the curves of second,
third, and higher orders. In general, the parametric equations of a curve of l -order can be written
in the form
l 1
xi f t x ,
j 1
j i
j
1 i 2, (7.20)
where xi , 1 i 2, 1 j l 1 are the coordinates of the nodal points of the curves;
j
Below there are the expressions for the functions when l 3, t D 1,1 (see Fig. 7.2):
l 1 (mapping 1 ):
f1 1 2 1 t
f 2 1 2 1 t
l 2 (mapping 2 ):
f1 1 2 t t 1
f 2 1 2 t 1 t
f3 1 t 1 t
l 3 (mapping 3 ):
f1 1 16 1 t 10 9 t 2 1
f 2 1 16 1 t 10 9 t 2 1
f 3 9 16 1 t 2 1 3t
f 4 9 16 1 t 2 1 3t
From practical point of view the expressions above are sufficient. More detailed description of
the boundary of the domain is best done through increase the number of smooth patches that
approximate the boundary instead of through the increase of the order l of the curves.
coordinates of their nodal points. Let us recall, however, that in the suggested method, the
degrees of freedom are not associated with the nodal points. Nodal points, in this case, are only
Let us consider how to construct the spaces G K of the boundary functions of the element.
belong to the space H m . Herewith the order of normal derivatives that appear in the
necessary to define spaces PrK,l , j of approximating functions. According to the Theorems 5.7 and
5.12, the sufficient condition of convergence of the approximations is the following inclusion
P0 K r , l , j PrK, l , j , 0 j m 1 .
Here and below Pi designates the space of polynomials of i -order defined on the
Ps j K r , l , j PrK, l , j , 0 j m 1
then, as it is shown in the Section 6, the external error of approximation of the variational
s min s j 1 , 0 j m 1 .
j
In the benchmark problems of the Section 8, the following space of polynomials of s j -order is
PrK, l , j Ps j K r , l , j , 0 j m 1 .
The number s j is called the order of the boundary approximation of j -layer of the surface
elements. Herewith the space G K of the boundary approximating functions of the element is
defined as follows
m 1
GK G
K r K r , l j 0
K
r,l, j , (7.21)
where
GrK, l , j P K
r ,l , j L2 K r , l
Kr ,l , j
Let us also point out one important class of the spaces G K characterized by that the patches
K r , l , j D, card r , l , j 1, 0 j m 1 . (7.22)
Herewith:
GrK, l , j Ps j D , 0 j m 1 .
Basic functions of the space G K generate boundary degrees of freedom of FE. According to the
Ps j K r , l , j Ps j t ,
t Kr , l , j , 0 j m 1
The simplest system of the basic functions of the space Ps j t is the system of monomials
t
sj
i
. When the relations (7.22) hold, the functions f j ,1 i l 1 from (7.20), that describe
i 0
the shape of FE, can be used as the basic functions of the spaces Ps j t as well.
Let us:
g t
i sj
j i 0
be some basis of the space Ps j t . Then the corresponding boundary degrees of freedom are
defined as follows
p P K ij p g ij t rK, l , j pd ,
Kr ,l , j (7.23)
0 j m 1, 0 i s j
When j 0 , the function rK, l ,0 p is the restriction of the function p onto the patch K r , l :
rK, l ,0 p p K .
r ,l
When j 1 , then
p p
rK, l ,1 p n1 n2 ,
x1 x2
where n1 , n2 are cosines of the outward normal to the boundary of FE. The surface measure is:
1
dx1 2 dx2 2 2
d dt ,
dt dt
3D elements
: D R2 Kr ,l R3 ,
In Fig. 7.3 are shown the patches of faces of FE whose parametric equations have the form
N
xi f t x ,
j 1
j i
j
1 i 3, (7.24)
the patches K r , l . Here are the expressions for the functions f j t ,1 j N for the patches
degree, N 6 (mapping 2 ):
f1 t1 2t1 1
f 2 t2 2t2 1
f3 1 t1 t2 1 2t1 2t2
f 4 4t1t2
f5 4t2 1 t1 t2
f 6 4t1 1 t1 t2
f1 1 4 1 t1 1 t2
f 2 1 4 1 t1 1 t2
f3 1 4 1 t1 1 t2
f1 1 4 1 t1 1 t2
f1 1 4 1 t1 1 t2 t1 t2 1
f 2 1 4 1 t1 1 t2 t1 t2 1
f3 1 4 1 t1 1 t2 t1 t2 1
f 4 1 4 1 t1 1 t2 t1 t2 1
f5 1 2 1 t12 1 t2
f 6 1 2 1 t22 1 t1
f 7 1 2 1 t12 1 t2
f8 1 2 1 t22 1 t1
Domains D of the parameters of maps shown on Fig. 7.3 have simple shapes and that simplifies
t ,
sj ,
t t1 , t2 , 1 , 2
can be chosen as the system of the basic functions of the surface element K r , l , j . The system of
functions f j t , 1 j N from the parameterization (7.24) also can be used for the same
purpose.
d g12 g 22 g32 2 dt
1
where:
dt dt1dt2
x2 x3 x2 x3
g1
t1 t2 t2 t1
x3 x1 x1 x3
g2
t1 t2 t1 t2
x1 x2 x2 x1
g3
t1 t2 t1 t2
x xi i 1 R n , i.e.
n
U x ui x i 1 ,
n
(7.25)
and let us assume that each component ui ,1 i n of the function belongs to the space
a space of vector functions. For example, when solving the problems of Theory of Elasticity in
displacements, the function (7.25) is a displacement vector and V H 1 .
n
using the FE described above. Therefore, for each FE are constructed the independent
k
uiN
b x ,
0
i
1 i n.
There can be another choice of boundary degrees of freedom of FE different from that defined by
(7.23). It can be caused by the following reasons. Often the boundary conditions are imposed not
onto the components ui ,1 i n of vector in global coordinate system, but onto the
components ui ,1 i n 1 of the vector in natural coordinate system, that has coordinate axes
normal and tangent to the boundary of the domain. For example, when solving the problems of
Structure Mechanics, one may need to fulfill the boundary condition when the normal to the
boundary component of the displacement vector equals to zero: un 0 and the tangent
displacements ui can be arbitrary (slippery support). Similar situation occurs when it is
necessary to take into account the cyclic symmetry of the structure. Besides, if the
systems, then when forming the system of equations it is necessary to bring the local boundary
One of the possible ways of effective solution of the above mentioned problem is the
as follows
N
UN b
i 1
i i , (7.26)
Let us, for example, n 2 . Then the simplest choice of representation (7.26) is
2 k 1i x
UN x
b i
i 1 0 2i
where
0, i j
ij
1, i j
P K Pi K , 1 i n
and, as the boundary degrees of freedom, it can be taken the integrals from the normal and
When n 2 , taking into account the relations (7.23), the boundary degrees of freedom can be
written as follows
p p1 , p2 P K P1K , P2K
0i n p g 0i t rK, l , 0 p1 n1 rK, l , 0 p2 n2 d (7.27)
K r , l ,0
0i p g 0i t rK, l , 0 p2 n1 rK, l , 0 p1 n2 d
K r , l ,0
where n1 , n2 are the cosines of the outward normal to the boundary of FE. Elements with the
boundary degrees of freedom of type (7.27) are referred to as vector FE. In the Section 8 the
vector FE are used when solving the 2D and the axisymmetrical problems of the Theory of
Elasticity.
Axisymmetrical elements
Let us assume that the domain R 3 is a body of revolution. In this case, taking into account
the axial symmetry, the geometry of the domain can be described by its axial cut in the
cylindrical coordinate system. If the boundary conditions of the boundary value problem and the
right-hand side of the corresponding differential equation are such that the solution of the
problem does not depend on the angle coordinate, then the problem is called axisymmetrical.
From the mathematical point of view, axisymmetric problems are 2D problems because the
sought functions depend only on the coordinates x1 , x2 . Herewith the analysis domain is a
radial component u1 and the axial component u2 . On the axis of symmetry the radial component
of displacements turns into zero, therefore the vector U belongs to the product of spaces V1 V2
where:
V1 u1 H 1 | u1 0 when x1 0
(7.28)
V2 H 1
Let us partition the domain into FE. The representation of the displacement vector of FE can be
in the form (7.26) and the boundary degrees of freedom in the form (7.27). Herewith in the
1
dx1 2 dx2 2 2
d 2 x1 dt
dt dt
If FE does not touch the axis of symmetry OX 2 , then the space is:
P K Pk K
2
If FE touches the axis of symmetry then, due to the characterization (7.28), the
approximant of the radial component of the displacement vector has to satisfy the
condition:
The condition (7.29) holds if it assumed that u1N x1 Pk ( K ) . Therefore, for the FE touching the
axis of symmetry OX 2 :
PK Pk K x1 Pk (K ) .
In classic FEM, analysis of stress concentration around holes, notches, and other small details of
structures, requires significant mesh refinement in domains with high stress gradients. Herewith,
the dimension of the discrete problem increase, and the conditioning of the equation system
worsens. Here, the approach developed in the previous sections is used to construct special
The FE is intended to simulate stress concentration domains. Unlike some other approaches [53 -
55], that also use local asymptotic representations of solutions of boundary value problems, in
external approximations it is not necessary to construct any particular functionals that take into
of the system of complex basic functions in the domain of special FE and the fulfillment of the
general criteria of the convergence of the external approximations provide the convergence of
approximate solutions to the exact. This is supported by the results of the numerical experiments
described in Section 8.
Let us consider a polygonal FE meant for the simulation of displacement and stress field in the
vicinity of an elliptic hole or a notch. The FE domain can be either double connected, when the
hole is located completely within the FE (Fig. 7.4.a), or single-connected when a part of the
contour of the element is an elliptical arch (Fig. 7.4.b). Lengths of the semi-axis of the ellipse
can be arbitrary, which in combination with the possibility of changing the FE connectivity,
Let us use the methods of the theory of functions of complex variable z x1 ix2 to obtain the
local asymptotic representation of displacements and stresses that exactly satisfies boundary
conditions on the contour of the hole as well as the equilibrium equations. Using two analytical
complex conjugate.
Any FE with an elliptical hole can be obtained from some initial FE with a round hole by the
conformal mapping
z C m 1 ,
where:
ab ab
i ; C ; m
2 ab
z z 2 4C 2 m
1
2
2C
maps the element with an elliptical hole onto the element with a round hole (see Fig. 7.4.a).
Conformal mapping allows to construct the approximation on the complex plane instead of z
and, therefore, the problem of boundary conditions fulfillment at the contour of the elliptic hole
is replaced by a simpler problem of doing the same thing for the unit round hole.
z 1
X z X X 1
Then:
1
z
X1
Xz
1 1
(7.31)
z
3
X 1 X 1
X z
3
Substituting expressions (7.31) into (7.30) and leaving out the index 1 one obtains:
X
2G u1 iu2 C1
11 22 4 Re
11 22 i 2 12 (7.32)
2
3
X X
a0 j j j j
,
j 1
expansion
N
N 0 j
j
j j
j 1
N 2 N (7.33)
X N 0 j
j j
j 1
j
j 1
Let us name the number N here as the order of internal approximation on complex FE. Unlike
the work [56], in (7.33) are used the segments of complete Laurent expansion and all the
X 0 .
(7.34)
Substituting the equation (7.33) into (7.34) and reducing at the same degrees of , the equations
0 C 0 m 2 2 2
1 C m1 1 1 m1
2 C m 0 2 2m 2
j C m j 2 j j 2 j 2 mj j ,
(7.35)
3 j N 2
j C m j 2 j j 2 j 2 mj j ,
1 j N
N 1 N 2 N 1 N 2 0
Making the reverse substitution of equations (7.35) into (7.33) it is obtained the following
j 1
X N j C j j 2 mj j
N
j 1
j C m j 2 j
N
(7.36)
j 0
j C j j 2 mj j
N
j 1
j C m j 2 j
N
j 1
j j1 i j 2
j j1 i j 2
where j1 , j 2 , j1 , j 2 are real numbers. Then relations (7.36) can be rewritten in the form:
N
N 01 02i j j 2i j
j1
j 1
j1 j 2i
j j
X N 01C m 1 02iC 1 m
2 2
j1C j j 2 mj j j m j 2
N
(7.37)
j 1
j 2iC j j 2
mj m
j j j 2
j1C m j 2 j j 2 mj j j
j 2iC j j j 2 mj j m j 2
Let us denote the complex basic functions in (7.37) through j and j
correspondingly, for 1 j 4N 2 , and let us denote the real factors of these functions
F
X
N 4N 2
FN
X b F ,
j j (7.38)
N j 1
where:
j
Fj
j
F
4N 2
are vector complex basic functions. Basic functions j j 1
generate the
F on FE.
Substituting (7.38) into (7.32) and taking into account that 1 z it is obtained the
representations for the displacement vector of the element on the -plane and on the z -plane:
u1N 4N 2 0j
U N N
u
bj 0
j
2 j 1
u1N z 4 N 2 j z
0 1
UN z N bj (7.39)
j
u z 0
1
2 j 1 z
4N 2 ˆ j z 4 N 2
bj b jU j z
ˆ j z
j 1 j 1
necessary to build basic functions of complex FE corresponding to its boundary and internal
degrees of freedom. These functions generate the space PK of the approximating functions of
FE. To construct the basic functions of FE it is necessary to determine the explicit form of
calculate stiffness matrix of FE that does not require explicit expressions for the basic functions
of FE.
Due to one-to-one correspondence of the elements from spaces K and PK , for each function
known. Therefore, instead of finding the representation U N z for the displacement vector of
the element through its boundary and internal degrees of freedom, it is sufficient to find such
representation for the vector potential FN . In accordance with the procedure described in
Section 3.1, it is necessary first - to build the system (3.12) of equations for the factors
b j ,1 j 4 N 2 of the initial representation (7.38) of the vector potential, and second - to
Let us consider how the elements of the matrix of the system (3.12) are calculated in this case.
Referring to the Section 7.2, boundary degrees of freedom of FE on z -plane are as follows:
0i n g 0i t n1u1N z t n2u2N z t d
K r , l ,0
(7.40)
0i g 0i t n1u2N z t n2u1N z t d
K r , l ,0
where:
From expressions (7.39) and (7.40) it follows that to calculate the matrix of the system (3.12)
t2 t2
g0 t j z t d , g t z t d ,
i 0 i 0
0 j
t1 t1
that can be performed numerically (for example, by Gauss’s quadrature). To calculate values of
necessary to know the explicit expressions for these functions. It can be done by the formula
1 ˆ j ˆ
ˆ j z iˆ j z C1 j
j
2G
ˆ ˆ
that is evaluated in the points of -plane:
r 1 z tr .
Representations (7.30) for displacements and stresses satisfy the equilibrium equations in the
domain of special FE when the mass forces are absent. In this case the corresponding expression
for bilinear form in the variational equation can be transformed by Green's formula
2 2
ij u ij v dx u n v d
ij j i
K i , j 1 K i , j 1
where u u1 , u2 , v v1 , v2 .
Therefore, when calculating the stiffness matrix, integration is performed along the contour and
not on the domain of special FE. It reduces significantly the amount of calculations.
Figure legends for Section 7:
Fig. 7.1
Examples of non-conformed FE
Fig. 7.2
Fig. 7.3
Fig. 7.4
All the experiments are performed using the FE described in Section 7.2. Orders k of the
s j , 0 j m 1 of the boundary approximation are assumed to be the same for all surface
elements of j -layer. Identical partitions of smooth patches of faces of FE into the surface
r ,l , j r ,l ,i , 0 i m 1, 0 j m 1.
In the figures referring to the problems, boundaries of the surface elements are shown in dashes
that are perpendicular to the contour of FE. Smooth patches of faces, on which the surface
elements are not defined, are marked by dash line placed inside FE.
In Section 3 it was shown that there are two types of the basic functions from the space of
Functions associated with boundary degrees of freedom of FE. The support of the
Functions associated with internal degrees of freedom of FE. The support of the function
is a single FE
Because of this property, the internal degrees of freedom can be computed immediately after the
stiffness matrix and right-hand side of FE has been built. So they are not included into the final
equation system. In conventional FEM this procedure is called the “condensation” of degrees of
freedom. In the tables below N is the dimension of final equation system assembled from the
condensed stiffness matrices of FE, so in fact it is the number of boundary degrees of freedom in
the system.
Torsion
Formulation
Torsion problem consists in finding shear stresses 23 and 13 in cross-section of a bar when it
is in twist (Fig. 8.1). This problem can be formulated as a Dirichlet’s problem in two equivalent
forms:
w f (8.1)
0w 0 (8.2)
where:
is Laplace’s operator
2
2
2
i 1 xi
and
V w w H 01 , w L2
2
w v
x
i 1 i xi
dx fvdx
v H 01 (8.3)
For torsion problems f 2 in the right-hand side of (8.1) and the function w is called a stress
function. Shear stresses in the section of the bar are determined by the formulas
w w
13 G , 23 G ,
x2 x1
In the external approximation of the equation (8.3) space Gh of the boundary functions consists
Model problem
Let us consider a model problem when the domain is a square 2 2 and f 1 . The exact
solution of the given problem is known [47]. Due to the symmetry it is sufficient to analyze 1 8
of the square.
In the first experiment the convergence on a family of partitions is under consideration. In Fig.
8.2 are shown the partitions of the domain into triangular FE. For all the partitions and any FE,
the order of the internal approximation is k 1 and the order of the boundary approximation is
s 0 . Let us note that for the given problem the external error of approximation is caused not
only by the discontinuity of the approximating functions at the inter-element boundaries, but also
In Table 8.1 is given the comparison of the approximate and the exact solutions for the partitions
shown in Fig. 8.2. Analyzing the data in Table 8.1 one can see that the convergence rate is not
high and in general corresponds to the theoretically predicted order O h . When partitioning
1 8 of the square into 16 FE, the maximum deviation of the function w from zero on the
boundary is approximately 6.6 % of its maximum value in the center of the square. For the
alternative partitions 5 and 6 the results almost coincide. The results illustrate the convergence of
h -version of MEFEA when the mesh of FE is refined without changing the type of FE.
Herewith, both the external and internal errors of approximation simultaneously decrease.
In the second experiment the convergence on a fixed partition of the domain is analyzed.
Herewith, the problem is solved on the partition 2(see Fig. 8.2) for different orders s 0;1; 2 of
the boundary approximation, while the order of the internal approximation for each FE of the
partition is kept unchanged ( k 4 for all values s ). In this experiment the order of the internal
error of approximation is fixed, but the external error is decreased through the refinement of the
boundary approximation.
The results given in Table 8.2 show how approximating qualities of the spaces of the boundary
functions affect the accuracy of the solution both within and on the boundary of the domain.
When s 2 , the maximum deviation of the function w from zero on the boundary is 1.8% of
the function's value in the center of the square. At the transition from s 1 to s 2 , the error at
the center of the square is decreased by 15.5 times while at the transition from s 1 to s 2 it
is decreased by 1.6 times only. This indicates that when s 0 , due to the low order of the
boundary approximation, the major contributor to the total error is the external error. When
s 1 both the external and the internal errors become commensurable so that further refinement
of the boundary approximation does not lead to a valuable improvement of the solution inside
the domain but it significantly decreases the error in the boundary condition.
Comparison of the Tables 8.1 and 8.2 shows that, for a given number of equations in the final
system, the accuracy of MEFEA, for high orders of the internal and the boundary
approximations, exceeds the accuracy when using the simplest FE with linear approximating
functions.
The ratio of the diameter of the bar to the radius of the notch is 2R r 10 (see Fig. 8.3). Due to
the symmetry one can consider only a half of the cross-section by assigning to the symmetry axis
w
0. (8.4)
n
The domain is simulated with two FE; the contour of the domain is approximated with inscribed
broken line. The boundary between the elements is a line that is inscribed into the circumference
with radius 2r . Unlike the previous problem, the orders k of the internal approximation are
For the problem, only smooth patches of the inter-element boundary have boundary degrees of
freedom that are not equal to zero. For the smooth patches that coincide with the approximated
boundary of the domain, the values of the boundary degrees of freedom are equal to zero due to
the homogeneous boundary condition (8.2). It is not necessary to create the boundary degrees of
freedom on faces where natural boundary condition (8.4) is prescribed. These patches are
marked in Fig. 8.3 with a dash line in accordance with the accepted rule. The number of the
equations in the final system is equal to 5; it is defined by the number of the inter-element
patches. The minimum order k of the internal approximation that provides P -multi-solvency of
In Table 8.3 are given the errors in shear stresses calculated by the formula
0
100 % ,
0
where is the calculated shear stress and 0 is the exact value from [48].
Table 8.3 shows that at an increase of the orders of the internal approximation on both FE from
k 4 to k 6 , the error in stress in the area of its concentration (point A in Fig. 8.3) decreases
from 57% to 2.7%. The refinement of the solution is achieved through the increase of the number
of the internal degrees of freedom of FE, while the number of the boundary degrees of freedom
remains unchanged. However, to get better solution there is no need for a simultaneous increase
of the order of the internal approximation on all FE of the partition. It is sufficient to do that only
for those FE located in the area of the stress concentration. Thus, if the order of the internal
approximation on the first FE is equal to 6, then the high accuracy is achieved independently of
the order on the second FE (see Table 8.3). It happens because the high stress gradient is located
in the domain of the first FE, and on the second FE the stresses field is smooth.
It is also necessary to note that the convergence in stresses is the convergence of the first
rather rough partition of the domain into FE, as well as the approximation of the boundary of
A great attention is paid to the development of FE for the problem of thin plates bending. It is
sufficient to mention a large catalogue [49] of FE used in various software systems (88
elements). The construction of FE for plate bending is more difficult because the differential
equation that describes the bending has the fourth order. Bilinear form of the corresponding
variational equation contains derivatives of the second order and therefore, the internal
approximations of the problem require inter-element continuity not only of the approximating
functions but of their first derivatives as well. This results in complex structures of nodal degrees
of freedom of FE, and in high orders of approximating polynomials even for the elements of the
q
2w (8.5)
D
where:
2 is a bi-harmonic operator
4 4 4
4 2 2 2 4
2
D EH 3 12 1 2 ,
Boundary conditions can be of different types. In case when on the patch 0 of the plate
w
w 0, 0 , (8.6)
0
n 0
and on the patch 1 a lateral load of intensity Q is applied, then solution of the equation
(8.5) is equivalent to finding the function w W that satisfies the variational equation:
2 w 2v 2 w 2v 2 w 2v
v W
D wv 1 2 2 2
dx
x1 x2 x1 x2 x 2
1 x2
2
x2 x1
qvdx Qvd
1
The bending moments and the torque are calculated correspondingly by the formulas:
2 w 2 w
M 11 D 2 2
x1 x2
2 w 2 w
M 22 D 2 2
x2 x1
2 w
M 12 D 1
x1x2
For bending problems m 2 and it is necessary to define two layers of surface elements on FE
boundary.
Long cantilever plate is loaded on the edge with uniform lateral load intensity Q (Fig. 8.4.a).
ratio 0.3 . The dimensions of the plate allow to consider it like a beam, and so to take the
polynomials of the first order are the spaces of boundary functions for both layers of
surface elements)
The results of MEFEA presented in Fig. 8.4.b practically coincide with the solution of the Beam
Theory. It is expected, since the hypothesis of flat section in the Beam Theory is a special case
of the hypothesis of straight normal in Kirchhoff’s Thin Plate Theory. Despite discontinuity of
approximating functions there are no jumps of deflections and bending moments on the
It is interesting to know how the complexity of FE shape affects the accuracy. The necessity to
use FE of a complicated shape can be caused by the geometry of the initial problem as well as by
the need to change the geometry of the domain at iterative computations (for example, when
non-linear problems are under consideration). In Section 8.1 it was demonstrated that the use of
FE of complex shape provides a high accuracy. However, the problem of torsion is the boundary
value problem of the second order. The main purpose of the following experiment is to study the
influence of the FE shape on the accuracy of solution of boundary value problems of the fourth
order.
Let us consider the following benchmark problem of bending of the square plate: dimensions
;Poisson's ratio 0.3 . On the contour of the plate is assigned either hinge support condition
Due to the symmetry only 1 / 8 of the plate is simulated. The symmetry condition is that the
derivative of the deflection function along the normal to the axis of symmetry being equal to
zero. There are three partitions of the domain organized by the degree of their irregularity (Fig.
8.5):
The second partition is created by the distortion of the first partition through
displacement of the common point of FE to the centre of the plate. In result, one FE of
The third partition contains three FE of the swallow-tail type and one FE of irregular
hexagon
For all the partitions and for any FE, the order of internal approximation is k 5 . The orders of
for hexagonal FE of the third partition, the order of internal approximation is increased to k 6
for the clamped plate problem. One can see that the degree of irregularity of the partition
insignificantly affects the accuracy. For all the partitions, when s 0 , the error in deflection at
the center of the plate is within the range 30 – 60%; when s 1 the error decreases to 0.1 –
0.2%. So the order of the boundary approximation mainly affects the accuracy: increasing s
from s 0 to s 1 , the error in deflections in the center of the plate decreases by two orders of
magnitude.
Similar results take place in the case of hinge supports of plate edges (Table 8.5). The deviation
from zero of the bending moment M 22 on the edge with the hinge is approximately 2.5% of the
on the order of the boundary approximation. For s 1 it is 0.5 – 3% of the maximum moment.
When s 1 , deflection w is calculated with high accuracy for both types of constraints. Its
jump on the inter-element boundary is 0.1 – 0.5% of the maximum deflection. The degree of
2D elasticity problems
Formulation
Equations that describe equilibrium inside an elastic body R n have the form:
E E
,
1 1 2 2 1
Deformations and stresses are expressed through the components of the displacement vector as
follows:
1 ui u j
ij u ,
2 x j xi
n
ij u ji u kk u ij 2 ij u ,
k 1
1 i n, 1 j n
where:
1, i j
ij
0, i j
is Kronecker’s delta.
Let us consider a boundary value problem with the following boundary conditions:
u 0 (8.8)
1
n
ij u n j gi , 1 i n (8.9)
j 1 2
where:
1 2
V H1
n
Y L2
n
V v v V , v Y
W v v V, v 0
1
The problem of finding the function u V that satisfies the relations (8.7) - (8.9) is
equivalent to the following variational problem: find the function u W such that
n n n
v W ij u ij v dx fi vi dx
ij 1 i 1
g v d
2 i 1
i i , (8.10)
where
n
1
f Y, g H 2 2 .
The external approximations of the equation (8.10) are constructed on vector FE described in the
Section 7.2.3.
Extension of a square plate
The load of intensity q is uniformly distributed along the edges of the plate (Fig. 8.6). The exact
values of stresses are 11 12 0, 22 q . Taking into account the symmetry, only 1 4 of
the plate is analyzed at different orders of the internal k and boundary s approximations. The
exact solution of the problem is obtained for all combinations of k and s changing in the range
boundaries. Therefore, in the example, the external and the internal errors of approximation are
The input data: cross-section of the bar H H 0.01 0.01 ;radius R 0.2 ; E 0.1 ;Poisson’s
ratio 0 . One end of the bar is loaded with total force P 105 , and the other end is
The curved contour of the bar is approximated by two inscribed broken lines with 16 straight
the first partition contains 8 FE, each is a fragment of arch (Fig. 8.7.b)
the second and third partitions are constructed by dividing each FE of the first partition
approximation is variable: k 2, 3 . The bandwidth of the equation system for all the partitions
is M 7 .
MEFEA solution is compared to the solution of the Theory of Circular Bars in [50]. The results
of the comparison are summarized in Table 8.6. One can see that, for the first partition at k 2
the error is significant (up to 80% in stress). It is expected, because stresses in the arched FE with
22.5o span (see Fig. 8.7.b) cannot be satisfactory simulated by the polynomial of the first order.
Increasing the order of the internal approximation to k 3 , the error in both the displacements
and stresses decreases to 10 – 15%. Decreasing the size of FE, the error decreases and for the
One can see that the accuracy is improved either through an increase in the number of FE (i.e.
decrease of the internal error) or through an increase in the order of internal approximation. The
bar is simulated by one layer of elements only. So the increase in the number of FE at fixed s
does not reduce the external error since the size of the inter-element boundaries remains
constant. The results indicate that for s 1 the external error, which is caused by the
displacement discontinuity on the inter-element boundaries, is small. The main contribution into
the error is made by the internal error. This is supported by the fact that displacements are
practically continuous. The discontinuity in stresses does not exceed 0.1% of their average value
The exact solution of the problem of stress concentration around the hole in the infinite plate is
known [48]. In computations, the infinite plate is replaced with a square plate with the ratio
between the side length and the diameter of the hole a d 8 (Fig. 8.8).
If the classic FEM is applied to a stress concentration problem, then the width of the surrounding
the hole layer of elements must be 0.05 – 0.1 of the radius of the hole [51, 52]. Let us use the
rather rough partition of 1 4 of the plate (see Fig. 8.8). In this partition, the width of the layer of
the adjacent to the hole FE is 1.25 of the radius of the hole. As it follows from the analytical
solution of the problem, the area of stress concentration spreads approximately onto just a half of
At a fixed partition of the domain into FE, the number of equations of the system is defined only
by the order s of the boundary approximation, and for s 0 the system contains 60 equations
s 1 the bandwidth is M 55 .
The results for the stress concentration factor are detailed in Table 8.7 for various combinations
of the approximation parameters. Increasing the orders of the internal approximation for the
adjacent to the hole FE without increasing the order of the boundary approximation leads to the
region, where the stress reaches its maximum). Because of this, the error in stress does not
decrease (see Table 8.7). Instead, at the transition from s 0 to s 1 the error significantly
decreases to 1.4%.To increase the accuracy in stress at a fixed partition of the domain it is
necessary to increase the orders of both internal and boundary approximations, which agrees
This Section contains the results of numerical experiments with the special element developed in
Orientation of long semi-axis of the ellipse, relatively to the direction of the extension, is defined
Square plate has ratio 2a L 0.032 between the long axis and the side of the plate. The plate is
partitioned into 9 FE, one of which is a regular octagon with an elliptic hole (Fig. 8.9.b). At one
side of the plate is applied a uniformly distributed load while the other side is constrained against
vertical displacements.
The order of internal approximation on the complex FE with the hole is k 5 . For the other
Computations are performed for 30 , 60 , 90 . Orientation of the hole changes without the
change in the partition of the domain. For all the difference between the computed contour
stress and the exact stress from the analytical solution for the infinite plate [48] does not exceed
1 – 2%.
The input data for the problem as shown in Fig. 8.10, are: b d 0.4 ; R b 0.125 ; L d 3 ;
d 0.2 ; the thickness of the strip is 0.006;modulus of elasticity is E 19.6 104 ;Poisson's ratio
Taking into account the symmetry, 1/4 of the strip is analyzed by partitioning it into 13 FE (Fig.
8.10.b). The layer of the adjacent to the hole elements, which width is b 8 , includes:
two polynomial elements which simulate straight patches of the contour of the hole
Let us denote through n, k the orders of the internal approximation for the complex and
polynomial FE respectively. Values k are assumed to be the same for all polynomial FE.
The results for contour stress, for two versions of the approximation, are shown in Fig. 8.11
where, for comparison purpose, is also given the experimental data from [56]. From Fig. 8.11
follows that the result at n 4, k 5, s 1 almost coincides with the experimental data.
Despite the fact that only one FE simulates straight patches of the contour of the hole, the
calculation proves the experimentally observed fact of the stress decrease in the center of the
vertical patch of the contour ( x1 b 2 , x2 0 ). The results nicely agree with the data from
work [56], in which 1/4 of the strip is simulated by one complex FE and 81 iso-parametric FE.
distribution of stresses on contour remains, but the peak stress decreases approximately by 15% .
The effect of the stress decrease in the center of the vertical sides of the contour is not observed
The perforation of the plate corresponds to a riveted or bolted joint (Fig. 8.12.a). The input data:
the diameter of the holes d 0.02 ; the intensity of the load 1 ; modulus of elasticity
with the holes and 12 polynomial FE (Fig. 8.12.b). The order of the internal approximation is
n 5 for the complex FE and k 4 for the polynomial FE, the order of the boundary
approximation is s 1 .
The stress distribution at contours of the holes is shown in Fig. 8.12.c. From Fig. 8.12.c it
follows that the mutual influence of the holes, that are in the central area of the plate, causes
some shift of the stress peak towards the neighbor hole. A similar effect is observed for an
infinite plate with two rows of holes [57]. However, assuming K max 4.85 the stress
concentration factor for the infinite plate, the actual value for the finite plate is in the range 5.40
– 5.85. For the holes close to the edge of the plate, the shifts of the stress peaks do not occur
For the problem, the order of the equation system is N 130 , and the bandwidth is M 51 .
The input data: the ratio of the radius of the hole to the width of the strip is R b 0.275
extension (Fig.8.13.b)
lateral bending (Fig. 8.13.c)
The partition of the strip includes three FE: one FE is complex (element with the hole), and the
other two are polynomial (see Fig. 8.13.b). The number of Gauss’s integration points on the
A
K ,
Bnom
where: A is the normal stress at point A (see Fig. 8.13.b) calculated by MEFEA; Bnom is the
nominal stress at point B calculated using the elementary formulas of the Theory of Strength of
In Table 8.8 are given the stress results for the extension problem depending on the orders of
internal and boundary approximations. For comparison purpose, in Table 8.8 are also given the
results obtained by classic FEM with 8-node iso-parametric elements (MIFE), by the method of
large elements (MLE), as well as experimental data of photo-elasticity method (MPE) [54].
From Table 8.8 it follows that the best conformity of the results of MEFEA with the results of
other methods takes place when s 1 . Herewith, the number of elements in MEFEA is 3 times
less than in MLE and 90 times less than in MIFE. Similar results take place for the lateral
bending (see Table 8.9) and pure bending (see Table 8.10).
Extension of a strip with a cutout
This problem illustrates the effectiveness of the complex FE with an elliptic arch (see Fig. 7.4.b).
The input data (see Fig. 8.14.a): the ratio of the depth of the cutout to the width of the strip is
Variants of the partitions of the strip are shown in Fig. 8.14.b.c. The data detailed in Table 8.11
indicate that the results of MEFEA are similar to the results of other methods [54]. Herewith the
characterize the stress field at the tip of a crack. They are used to forecast survivability of the
structures containing cracks as well as in fatigue analysis. Accurate simulation of the stress field
at the tip of the crack is critical to Fracture Mechanics. For a qualitative approximation of the
solution in the vicinity of the crack special FE are required that have the singularity in stresses of
type r 1 2 [58 - 68]. The problem of construction of such elements is twofold: on one hand, it is
necessary to simulate the singularity of a certain type on the element, on the other hand it is
necessary to provide the continuity of the approximating functions (including singular) through
various ways:
by use of the local asymptotic representations of solutions for the elements with
More difficult problem is to provide continuity of the approximating functions through the
boundary of singular FE. This problem is solved successfully when using modified
iso-parametric FE. However, as it has been demonstrated (see [79 - 81]), the size of the element
has a major influence on the accuracy, and it leads to the necessity of a mesh refinement near the
tip of the crack. Thus, for example, in work [69] when solving the problem of inclined crack the
size of the smallest iso-parametric element was 1 250 of the length of the crack.
Discontinuity of the approximating functions can be handled with special buffer elements that
surround the element with singularity [82], or by developing singular asymptotic representations
that turns to zero on the boundary [83, 84], or by hybrid schemes of FEM. Hybrid FE are widely
used when solving the problems of fracture mechanics. They are based on independent
approximation of singular field of stresses and displacements of the boundary of FE [54, 59, 85 -
90]. Complexity of the construction of conformed singular FE makes some authors use the non-
benchmark problems. However, the benchmark problems cannot cover all the diversity of
practical cases and the successful solution of a problem does not guarantee the success in all
situations.
Below, the special FE described in the Section 7.2.5 is used to solve fracture mechanics
problems. Herewith the length of one of the semi-axes of the elliptic hole is equal to zero, and
the space PK of the approximating functions of FE is generated by the terms of the complete
asymptotic representation of the exact solution of the problem with a crack in infinite domain.
Let us notice that the FE allows the crack orientation change and growth without re-partitioning
of the domain. It is convenient when solving the problems of the crack propagation.
The exact stress intensity factors for the inclined crack in the infinite plate are calculated by the
formulas
K I l sin 2
K II l sin cos
where: is the stress at the infinity; l is the semi-length of the crack; is the orientation angle
(Fig. 8.15.a).
In the computations, the infinite plate is replaced with a square plate with a ratio of the length of
the crack to the side of the square of 2l L 0.032 (Fig.8.15.b). Uniformly distributed load
1 is applied at the upper edge, while the opposite edge is constrained in normal direction.
The partition includes one complex FE with the crack and 8 polynomial FE (see Fig. 8.15.b). The
The crack orientation is changed without re-partitioning the domain. The orders of the internal
approximation are: n 5 for the complex FE, and k 3 for all polynomial FE. The order of the
M 24 .
In Table 8.12 are given the errors in stress intensity factors K I and K II at the tip of the crack for
different orientations. Despite the rather coarse partition of the domain and the low order of the
boundary approximation, a high accuracy in fracture parameters is achieved. For small , the
error for K I and K II is rather lower than for large . Thus, at 15 , the error is 2.8% for
K I and 0.8% for K II ; at 30 the error is 0.5% for K I and 0.2% for K II . Values of K I
and K II , calculated for both tips of the crack, coincide up to three digits. Error in the load
The analysis of stress field shows that on the contour of the complex FE takes place the one-axial
stress state that corresponds to the extension of the plate without the crack. The domain of high
stress gradients is located completely inside the complex FE. This is caused by that the ratio
0.125 between the length of the crack and the diameter of the complex FE is rather small. Inter-
element boundaries are outside the area of stress disturbance and this is the reason of high
In the case when the length of the crack is of the same order of magnitude as the width of the
strip, the stress intensity factors cannot be calculated by the formulas for the crack in the infinite
domain. The finite width of the strip is usually accounted by the correlation factor, i.e. the actual
K I FK I
The data to the problem: the ratio of the width to the length of the strip is a L 5 18 ;module of
elasticity is E 19.6 104 ;Poisson’s ratio is 0.3 ;one edge of the strip is fixed while at the
The strip is simulated by three FE, the element with the crack is complex,the other two elements
are polynomial (Fig. 8.16).The order of the internal approximation on the polynomial FE is
follows that the results of calculation of the correlation factor conform to the data from other
In polar coordinate system x1 , x2 , x3 , the equilibrium equations (8.7) have the form
11 12 11 33
f1 0
x1 x2 x1
12 22 12
f2 0
x1 x2 x1
u1 u2 u1
22 D
1 x1 x2 1 x1
u1 u2 u1
33 D
1 x1 1 x2 x1
1 2 u1 u2
12 D
2 1 x2 x1
E 1
D
1 1 2
here: E is the modulus of elasticity; is Poisson's ratio; u1 , u2 are, correspondingly, radial and
axial displacements; f1 , f 2 are the radial and axial components of the volume force.
The variational problem (8.10) is transformed as follows: find the function u W that satisfies
the equation
3
*
i 1
ii u ii v 12 u 12 v dx
2 2
(8.11)
f v dx g v d
i i i i v W
* i 1 2 i 1
where:
vi
ii v , i 1, 2
xi
v1
33
x1
v1 v2
12 v
x2 x1
W V1 V2
V1 u1 H 1
u1 0 for x1 0
V2 H 1
and where is the half of the meridian section of the structure located in the semi-plane x1 0 .
External approximations of the equation (8.11) are built with axisymmetrical elements described
in Section 7.2.4.
The data to the problem: the radius of the sphere is R 0.1 ; the thickness of the wall is
H 0.04 ; the internal pressure is P 1 ; the material properties are E 2 105 , 0.3 .
For this problem displacements and stresses are functions of the distance from the center of the
sphere. Therefore, for the analysis of the sphere, one can consider just a fragment of its meridian
section that is bounded by the central angle of 15 . The domain is the polygon ACDFB (Fig.
8.18.a) with its lines AC and DFB inscribed into the contour of the meridian section of the
sphere. The displacements along the normal to sides AB and CD of the fragment ACDBF equal
zero. The domain is partitioned into one and two FE (Fig. 8.18.a.b).
The results detailed in Table 8.14 indicate the high accuracy achieved even at the partition of the
domain into one FE. Thus, at s 1, k 4 , the error in maximum radial stress 11 is 1.1% and
in maximum circumference stress 33 is 6.4%. The radial stress at the outer surface where it
must equal zero is 3% of the maximum. The error in circumference stresses on the outer surface
is 0.05%. Partitioning the domain into two FE (Fig.8.18.b), the accuracy increases.
The infinite domain around the cavity (Fig. 8.19.a) is simulated by a finite cylinder (Fig. 8.19.b).
The data to the problem: the radius of the cylinder is R ; the height of the cylinder is 2R ; the
ratio between the radius of the cylinder and the radius of the cavity is R r 8 ;Poisson's ratio is
0.3 ; the uniformly distributed load of unit intensity is applied to the ends of the cylinder
are performed for two partitions of the domain that differ by the width a of the layer of the
adjacent to the cavity elements (Fig. 8.19.b). For the first partition a r 1.25 , and for the
The order of the boundary approximation is increased from s 0 to s 2 . The order of the
The results for stresses on the equator and in the pole of the spherical cavity (points A and B in
Fig. 8.19.b) are given in Table 8.15 for different partitions and orders of the boundary
approximation. One can see that the width of the boundary layer of FE significantly influences
the accuracy. Thus, for a r 1.25, s 2 , the error in maximum stress is 17.8% and, for
a / r 0.35, s 2 , it is 0.68%. The error rapidly decreases by increasing the order of the
boundary approximation that proves that for k 4 the internal error of approximation is rather
small and the accuracy is, in general, determined by the quality of the boundary approximation.
The data for the problem: the radius of the plate is R 0.1 ; the thickness is H 0.01 ; the
increased from 2 to 4.
From Table 8.16 it follows that increase in the order of the internal approximation from k 2 to
k 3 provides sufficient accuracy of the solution. The further increase to k 4 practically does
not cause any changes in the displacements and stresses. So the internal error is significant for
s 1, k 2 . For s 1, k 3 the internal and external errors become of the same order of
magnitude and the increase to k 4 does not lead to a valuable change in the accuracy. For
k 4 , the difference between the calculated stress and the Theory of Thin Plates Bending is
approximately 0.5% while the difference in the displacements in the center of the plate is 4.1%.
The data to the problem: the radius of the middle surface of the dome is R 2.28 ; the thickness
The dome is partitioned uniformly with spacing 3.5 (Fig. 8.21). Three variants of the
1. k 2 for all FE
2. k 3 for all FE
3. For the FE located at the edge of the dome k 4 while for the others k 3
The order of the boundary approximation is s 1 .
In Table 8.17 by the asterisk is marked the data that correspond to the third variant of the
approximation. Here the stresses on the external and internal surfaces of the dome are marked by
(+) and (-) respectively. The data in Table 8.17 show that the results in 3D formulation well
agree with the theory of thin shells. The increase of the order of the internal approximation for
the FE adjacent to the edge of the dome (third variant of the approximation) does not lead to the
change in stresses at the center of the dome. At the edge of the dome the stresses also change
Equations of the middle surface of the shell in the polar coordinate system (Fig. 8.22.a) are as
follows:
x1 R f 1 cos
l
x2
2
where is a parameter.
Considering the periodic symmetry of the structure, only one semi-wave of the shell is analyzed
R 0.225 ; f 0.025 ; l 0.1 ; H 0.0025 ;the material properties E 2 105 , 0.3 ;the
pressure q 1 .
For better approximation of the geometry of the shell, the semi-wave of is partitioned into 18
elements (Fig. 8.22.b). The order of the internal approximation is k 2 for all FE,the order of
the boundary approximation is s 1 . The number of the equations in the system is N 58 , the
bandwidth is M 7 .
In Fig. 8.23 are shown the results for meridian 1 and circumference 2 stresses on external (+)
and internal (-) surfaces. For the comparison purposes, in Fig. 8.23 are also given the results
obtained by Timoshenko's model [94] as well as by the Theory of Thin Shells [95]. From Fig.
8.23 it follows that the calculations by different models give similar results.
Shell compensator
In Fig. 8.24 is shown a semi-wave of the compensator and its partition into 14 axisymmetric
elements. The semi-wave consists of two torus segments I and III , and a circular plate II .
Coordinate is measured along the meridian of the compensator. The modulus of elasticity is
The order of the internal approximation for all FE is k 3 , the order of the boundary
approximation is s 1 .
there are no restrictions for the displacements of the rim. The condition of symmetry u2 0 is
defined at x2 0 . In Fig. 8.25 are shown the graphs for the meridian stress 1 and for the
circumference stress 2 , both on the internal (continuous lines) and on the external (dash line)
From Fig. 8.25 it follows that the maximum circular stress occur on the upper rim of the
compensator, and the stresses at the internal and external surfaces are practically identical.
Meridian stress has two obvious peaks: the first peak is located in the second half of the torus
segment I while the second peak is located at the center of the torus segment III . Herewith, the
amplitude of the second peak exceeds the amplitude of the first peak approximately by 50%.
Case 2 - Often the compensator system consists of few connected shell compensators. The
calculations are made for the compensator that is one wave of the compensator system, at the
given axial convergence of its rims. Boundary conditions of the problem are defined as follows:
u2 0 at x2 0
The stress graphs are given in Fig. 8.26. One can see that the stress picture significantly depends
on the type of the boundary conditions at the upper rim. For a given convergence of the rims,
maximum circumference stress occurs at the tips of the torus segments. At the upper rim the
meridian stress is significant. The second peak of the meridian stress is approximately equal to
the amplitude of the first peak and it is located at the center of the torus segment III . The stress
graphs shown in Fig. 8.25 well conform to the results for such compensators from [94].
Figure legends for Section 8:
Fig. 8.1
Fig. 8.2
Partitions
Fig. 8.3
Fig. 8.4
Fig. 8.5
Fig. 8.6
Fig. 8.7
Bending of a circular bar:
Fig. 8.8
Fig. 8.9
Fig. 8.10
Fig. 8.11
A perforated plate:
(b) partition
(c) Stress distribution at the contour of the holes (circles show the points where the stress
Fig. 8.13
(a) geometry
(b) partition
Fig. 8.14
(a) geometry
Fig. 8.15
Fig. 8.17
Fig. 8.18
(a)(b) partitions
Fig. 8.19
Fig. 8.20
Fig. 8.21
Fig. 8.22
(a) geometry
(b) partition of a semi-wave
Fig. 8.23
Stress distribution:
Fig. 8.24
Fig. 8.25
Fig. 8.26
2. Amores D., Deserable D., Martinez I., Surry C., “ Elements finis conformes on non
33-34.
3. Bergan P.G., Nygard M.K., “Finite elements with increased freedom in choosing shape
4. Fortin M., Soulie M., “A nonconforming piecewise quadratic finite element on triangles”,
5. Irons B.M., Razzaque A., “Experiens with the patch test for convergence of finite
elements”, Proc. Symp. on the Mathematical Foundations of the Finite Element Method
with Applications to Partial Differential Operators, Baltimore; N.Y., 1972, pp. 557-587.
6. Loikkanen M.J., Irons B.M., “An 8-node brick finite element”, Int. J. Numer. Meth. Eng.,
7. Zienkiewicz O.C., Qu S., Taylor R.L., Nakazawa S.,“The patch test for mixed
formulations”, Int. J. Numer. Meth. Eng., 1986, 23, 10, pp. 1873-1883.
9. Gilbert Strang, George J. Fix,An Analysis of the Finite Element Method, Prentice-Hall,
1973, 349 p.
10. Philippe Ciarlet,The Finite Element Method for Elliptic Problems, North-Holland
p.
12. Arantes e Oliveira E.R., “ The patch test and general convergence criteria of the finite
element method”, Int. I. of Solids and Structures, 1977, 13, pp. 159-178.
13. Janovsky V., Prochazka P., “Convergence analysis of a nonconforming finite element
method solving a plate with ribs”, Apl. mat., 1978, 23, 1, pp. 9-30.
14. Janovsky V., Prochazka P,.” The nonconforming finite element method in the problem of
clamped plate with ribs”, Apl. mat., 1976, 21, 4, pp. 273-289.
15. Lascaux P., Lesaint P., “Some nonconforming finite elements for the plate bending
problem”, Rev. franc. automat., inform., rech. oper., 1975, 9, R1, pp. 9-53.
16. Lesaint P., Zlamal M., “Convergence of the nonconforming Wilson element for arbitrary
17. Nitsche J.A., “Convergence of nonconforming methods”, Math. Aspects Finite Elements
18. Rannacher R., “L - Feherabschatzung fur eine nicht Ronforme Finite-Elemente-Methode
beim Plattenproblem”, Z. agnew. Math. und Mech., 1977, 57, 5, pp. T247-T249.
19. Rannacher R., “Nonconforming finite element methods for eigenvalue problems in linear
20. Rannacher R., “On nonconforming and mixed finite element methods for plate bending
problems. The linear case”, RAIRO. Anal. Numer., 1979, 13, 4, pp. 369-387.
21. Stummel F., “Basic compactness properties of nonconforming and hybrid finite element
22. Stummel F., “The generalized patch test”, SIAM J. Numer. Anal., 1979, 16, 3, pp. 449-
471.
23. Stummel F., “The limitations of the patch test”, Int. J. Numer. Meth. Eng., 1980, 15, pp.
177-188.
27. Корнеев В.Г., Схемы метода конечных элементов высоких порядков точности, Л.,
1977, 206 с.
с.
29. Barnhill R.E., Brown J.H., “Nonconforming finite elements for curved regions”, Lect.
30. Goldstein C.I., “Variational crimes and Lerror estimates in the finite element method”,
32. Ce’a., “Approximation variationelle des problemes aux limites”, Ann. Inst. Fourier.,
34. Glowinski R., “Approximations extrenes, par elements finis de Lagrange d’ordre un et
resolution des problemes approaches”, Top. Numer. Anal., L.;N.Y., 1973, pp. 123-171.
36. Филин А.П., Коносов Г.И.,“Решение краевой задачи при новом ( смешанном )
37. Li Zi-Cai, ”On the reduced rate of convergence for a nonconforming combined method”,
38. Li Zi-Cai, ”A nonconforming combined method for solving Laplace’s boundary value
39. Li Zi-Cai, “An approach for combining the Ritz-Galerkin and finite element methods”, J.
41. Blair J.J., “Higher order approximations to the boundary conditions for the finite element
43. Kikuchi F., “Convergence of the ACM finite element scheme for plate bending
problems”, Publs Res. Inst. Math. Sci., 1975, 11, pp. 247-265.
45. Shi Zhong-ci., „A convergence condition for the quadrilateral Wilson element”, Numer.
49. Hrabok M.M., Hrudey T.M., “A review and catalogue of plate bending finite elements”,
method solving a plate with ribs”, Apl. mat., 1978, 23, 1, pp. 9-30.
52. Сахаров А.С. и др. Метод конечных элементов в механике твердых тел, Киев,
53. Delves L.M., Hall C.A., “An implicit matching principle for global element calculations”,
54. Jirousek J., “Implementation of local effects into conventional and nonconventional finite
55. Jirousek J., Teodorescu P., “Large finite elements method for the solution of problems in
the theory of elasticity”, Comput. and Struct., 1982, 15, 5, pp. 575-587.
56. Moriya K., “Finite element analysis of stress concentration around elliptic holes, fillets
and cracks in plates under stretching and out-of-plane bending”, Bull. JSME., 1984, 27,
58. Chen E.P., “Finite element analysis of a bimaterial interface crack”, Theor. and Appl.
59. Kuna Meinhard, “Behandlung raumlicher Ribprobleme mit der Methode der finiten
60. Lin K.Y., Mar J. W., “Finite element analysis of stress intensity factors for cracks at a bi-
61. Ogen G., Schiff B., “Constrained finite elements for singular boundary value problems”,
finite element method”, Eng. Fract. Mech., 1984, 20, 1, pp. 169-177.
63. Wahba N.N., “On the use of singular displacement finite elements for cracked plate in
64. Wait R., “Singular isoparametric finite elements”, J. Inst. Math. and Appl., 1977, 20, 1,
pp. 133-141.
65. Ye T.Q., Gallagher R.H., “A singular finite element for analysis in plate bending problem
12, c.91-96
1984, 7, c.12-17.
69. Морозов Е.М., Никишков Г.П. Метод конечных элементов в механике разрушения,
70. Saouma Victor E., Schwemmer David, “Numerical evaluation of the quarterpoint crack
tip element”, Int. J. Numer. Meth. Eng., 1984, 20, 9, pp. 1629-1641.
71. Staab G.H., “Estimating stress intensity factors with singular components of the total
finite element solution”, Int. J. Numer. Meth. Eng., 1982, 18, 7, pp. 1063-1076.
72. Марчук Г.И., Агошков В.И. Ввведение в проекционно-сеточные методы, М., 1981,
216 с.
73. Blum H., Dobrowolski M., “On finite element methods for elliptic equations on domains
74. Destuynder P., Djaona M., “Estimation de l’erreur sur le coefficient de la singularite de la
solution d’un probleme elliptique sur un ouvert avec coin”, RAIRO. Anal. numer., 1980,
75. Yamamoto Y., Tokuda N., Sumi Y.,”Finite element treatment of singularities of
boundary value problems and its application to analysis of stress intensity factors”,
Theory and Pract. Finite Elem. Struct. Anal., Tokyo, 1973, pp. 75-90.
76. Hendry J., “Singular problems and the global element method”, Comput. Meth. Appl.
77. Kermode M., Mc Kerrel A., Delves L.M., “The calculation of singular coefficients”,
Comput. Meth. Appl. Mech. and Eng., 1985, 50, 3, pp. 205-215.
78. Tolley M.D., “Torsion des barres polygonales”, Bull. cl. sci. Acad. rog. Belg., 1977, 63,
79. Harrop L.P., “The optimum size of quarter-point crack TIP elements”, Int. J. Numer.
80. Jehia Nabil A.B., Shephard Mark S., “On effect of quarter-point element size on fructure
81. Yunus S.M., Shephard M.S., “A comment on the effectiveness of uniform finite element
83. Li Yao-Chen, “The finite element method by employing the singular element with
concordant displacement at the crack tip”, Eng. Fract. Mech., 1984, 19, 5, pp. 959-972.
84. Stern Morris, Becker Eric B., “A nonconforming crack tip element with quadratic
variation in the singular fields”, Int. J. Numer. Meth. Eng., 1978, 12, 2, pp. 279-288.
85. Chen Wen-Hwa, Chen Pei-Yen, “A hybrid-displacement finite element model for the
bending analysis of thin cracked plates”, Int. J. Fract., 1984, 24, 2, pp. 83-106.
86. Orringer O., “K-solutions with assumed-stress hybrid elements”, J. of the Structural
Division, Proc. Of the Amer. Soc. Of Civil Engin., 1977, 103, 2, pp. 321-334.
88. Piltner R., “Special finite elements for an appropriate treatment of local effects”, Local
89. Piltner R., “Special finite elements with holes and internal cracks”, Int. J. Numer. Meth.
90. Wang S.S., Yuan F.G., “A hybrid finite element approach to composite laminate
elasticity problems with singularities”, Trans. ASME : J. Appl. Mech., 1983, 50, 4a, pp.
835-844.
91. Byskov Esben, “The calculation of stress intensity factors using the finite element
method with cracked elements”, Int. J. Fract., 1984, 26, 4, pp. 329-337.
92. Staab G.H., “A Variable power singular element for analysis of fracture mechanics
93. Сиратори М., Миеси Т., Мацусита Х.,Вычислительная механика разрушения, М.,
1986, 334 с.
172 с.
с.66-72