Bending of Beam with Free Ends on Non-linear Subsoil

Ohyb volne polozeneho nosnku na nelinearnm podloz

Stanislav SYSALA
1
Abstract: A semi-coercive problem with a beam on a unilateral elastic subsoil of Winkler type is
investigated in the contribution. Firstly, the existence, the uniqueness and the continuous depen-
dence on data of the problem solution are discussed. Secondly, the problem is approximated by
the nite element method, where the subsoil is replaced by insulated springs due to a numerical
quadrature. The relations between the original problem and the family of approximated problems
are introduced. Thirdly, suitable numerical methods of Newton type are introduced. The conver-
gence analysis of the methods is investigated. The dependence of the methods on the discretization
parameter and the load is also discussed. And nally, some of the theoretical results are illustrated
on numerical examples.
Abstrakt: V prspevku je studovana modelova uloha ohybu volne polozeneho nosnku na jed-
nostranne pruznem podloz Winklerova typu. Nejprve jsou formulovany podmnky resitelnosti
ulohy kladene na zatzen nosnku a je analyzovana spojita zavislost resen na datech, zejmena na
zatzen. Na zaklade techto v ysledku jsou popsany tzv. nestabiln prpady zatzen.

Uloha je dale
aproximovana pomoc metody konecn ych prvku. Spojite podloz je pri tom nahrazeno bodov ymi
pruzinami a to na zaklade pouzite numericke kvadratury. Je odvozen vztah mezi resenm spojite a
aproximovane ulohy. Aproximovana uloha je nasledne resena pomoc nehladke Newtonovy metody
s tlumenm. Je studovana globaln a lokaln superlinern konvergence metody. Pro tzv. nestabiln
zatzen je navrzena modikace numericke metody za ucelem lepsch konvergencnch vlastnost.

Uloha je numericky ilustrovana na konkretnch prkladech potvrzujcch nektere teoreticke v ysledky.

Keywords: beam, subsoil, non-linearity, semi-coercivity, mathematical analysis, numerical analysis
Klcova slova: nosnk, podloz, nelinearita, semikoercivita, matematicka anal yza, numericka anal yza
1. Introduction
Bernoulli - Euler model of a beam on an elastic subsoil (foundation) of Winkler type is
often used in engineering, see [4, 5]. In this paper, we will consider several generalizations
of this linear model which can be useful in some particular problems.
The rst generalization connects with the response function of the subsoil which is usually
linear in models. However if a beam (or a thin plates) only laid on the subsoil then the subsoil
1
M.Sc. Stanislav SYSALA, Ph.D., Intitute of Geonics AS CR, Studentska 1768, 708 00 Ostrava-Poruba,
Czech Republic, Phone: +42 59 69 79 111, E-mail: sysala@ugn.cas.cz
is active only if the beam deects against it. In such cases, non-linear models of the subsoil
are more precise. The non-linear ones have been studied in [7, 8, 9, 14, 16]. Some related
problems were treated in [6, 13].
Here, we will consider a simple model of a unilateral elastic subsoil of Winkler type.
Other non-linear models can for example represent more-layered subsoils [14]. It is also
possible to use non-linear models for Pasternak type of the subsoils [8, 11].
The second generalization is connected with the coercivity of the problem. The models
with xed or supported beams are usually considered. Such problems are coercive and
uniquely solvable. Moreover, for bilateral elastic subsoil, the problem is coercive also for
Neumann boundary conditions. However, if we assume the unilateral elastic subsoil then
there also exist boundary conditions for which the problem is only semi-coercive. It happens
for example if only one end of the beam is supported, or if the beam has free ends. The
semi-coercive problems can be characterized such that the solvability of the problem and
the stability of the problem solution with respect to small changes of data are dependent
on the load. The semi-coercive problems with beams have been investigated for example in
[8, 14, 16].
Here we will consider the homogeneous Neumann boundary conditions (beams with free
ends). In this case, the admissible rigid motions of the beam are represented by polynomial
of the rst degree.
This contribution gives a survey of the main results from [16, 17, 18, 19]. Moreover we
add mechanical interpretations and discuss generalizations for some results.
In Section 2, we set the problem and discuss the existence, the uniqueness and the
continuous dependence on data of the problem solution. In Section 3, we approximate the
problem by the nite element method, where the subsoil is replaced by insulated springs
due to a numerical quadrature. The relations between the original problem and the family of
approximated problems are introduced. Then the problem is rewritten into the algebraical
form. In Section 4, the used numerical methods are introduced. Since the methods are
newly formulated and interpreted as Newton-like methods, the results from this section are
described in more detail in comparison with the other sections. The convergence analysis of
the methods is investigated. The dependence of the methods on the discretization parameter
and the load is also discussed. In Section 5, some of the theoretical results are illustrated on
numerical examples.
2. Mathematical Model of the Problem
The main goal of the section is to analyse the solvability of the problem and a continuous
dependence on data. First, we set the problem and its variational formulation. Then, we
derive necessary and sucient conditions for the existence and uniqueness of the problem
solution, and nally we investigate the continuous dependence of the problem solutions on
data and consequently describe some situations for which the solutions need not be stable
with respect to a small change of the load.
We consider a beam of the length l with free ends which is situated in the interval
= (0, l), and assume that the beam is supported by a unilateral elastic subsoil in the
interval
s
:= (x
l
, x
r
), 0 x
l
< x
r
l. Such a subsoil is active only if the beam deects
against it.
We use the Bernoulli - Euler model of a beam on a unilateral subsoil of the Winklers
type. The classical formulation of the problem (see [4, 10]) has the form of a non-linear
dierential equation of the fourth order with homogeneous Neumanns boundary conditions:

(E(x)I(x)(w

(x))

+ q(x)(w

(x) = f(x), x ,
(w

(0) = (w

(0) = (w

(l) = (w

(l) = 0,
(1)
where q = 0 in \
s
and w

(x) := min{0, w(x)} is the negative part of w. The functions

E, I and q represent the Youngs modulus of the beam material, the inertia moment of the
cross-section of the beam and the stiness coecient of the subsoil, respectively. The aim is
to nd the deection w

of the axes of the beam caused by the beam load. The situation is
depicted in Figure 1.
6
-
?????? ? ?
x
l
x
r l
q
f
y
x
hh
( (
hh
( (
hh
( (
hh
hh
( (
hh
( (
hh
( (
hh
hh
( (
hh
( (
hh
( (
hh
hh
( (
hh
( (
hh
( (
hh
hh
( (
hh
( (
hh
( (
hh
Fig. 1: Scheme of the subsoiled beam with axes orientation.
We will assume that the functions E, I, q are bounded in and there exist positive
constants E
0
, I
0
and q
0
such that
E(x) E
0
, I(x) I
0
, in , and q(x) q
0
in
s
.
The load density can be expressed in the form
f = f
1
+ f
2
,
where f
1
L
1
() (integrable in ) and f
2
represents generalized forces, i.e.
f
2
=

yX
P
P
y

y
+

yX
M
M
y

y
,
where X
P
, X
M
are nite sets of points belonging to , P
y
, M
y
R,
y
,

y
denote Dirac distri-
bution and its rst generalized derivative at a point y and P
y

y
, M
y

y
represent respectively
the point load and the moment at a point y.
Since the data and the solution need not be suciently smooth for the classical formu-
lation, we will work with the weak (variational) formulation of the problem. Let us dene
the bilinear and bounded forms
a(v
1
, v
2
) :=

EIv

1
v

2
dx and b(v
1
, v
2
) :=

s
qv
1
v
2
dx
on the Sobolev space H
2
() (the space of kinematically admissible functions). The forms
represent the work of the inner forces and the subsoil, respectively. The space of all contin-
uous and linear functionals dened on H
2
() will be denoted by V

and its corresponding

norm by .

. More information about function spaces can be found in [1]. The work of the
beam load will be represented by the functional L V

,
L(v) :=

f
1
(x)v(x)dx +

yX
P
P
y
v(y) +

yX
M
M
y
v

(y).
The weak formulation of the problem (denoted by (P)) means to solve the nonlinear
variational equation
(P) nd w

H
2
() : a(w

, v) + b((w

, v) = L(v) v H
2
(). (2)
The corresponding potential energy functional has the form
J(v) :=
1
2

a(v, v) + b(v

, v

L(v), v H
2
(). (3)
Since the beam does not have xed ends (it is only laid on the subsoil), the problem
solvability depends on the beam load (so-called semi-coercive problem). The existence and
uniqueness of the solution w

of the problem (P) are ensured by the condition

L(p) < 0 p P
1
, p > 0 in
s
, (4)
where the polynomials of the rst degree represent the rigid beam motions for which the
subsoil is not active, see [16, Theorem 3.1]. Notice that the functional J is coercive on H
2
()
if this condition holds. The condition (4) is also necessary for the uniqueness. The other
solvability condition is discussed in [16].
For other interpretation and analysis, it will be usefull to rewrite equivalently the condi-
tion (4) in the following way, see [16, Lemma 3.4]:
F < 0 and x
l
< T < x
r
, (5)
where
F := L(1) =

f
1
(x)dx +

yX
P
P
y
is the load resultant and
T := L(x)/L(1) =

f
1
(x)xdx +

yX
P
P
y
y +

yX
M
M
y

f
1
(x)dx +

yX
P
P
y
is the balance point of the load. The condition (5) means that the load resultant is situated
in
s
= (x
l
, x
r
) and oriented against the subsoil, which causes that the beam deection
activates the subsoil on the set M
s
with a positive one-dimensional Lebesgue measure,
i.e. w

< 0 in M. In addition, the balance point T lies in the convex closure of the set M,
see [16, Lemma 3.5].
The solvability condition (5) can be generalized for the problem with more parts of the
subsoils and also for the 2D case of thin elastic plates, see [9]. In such cases, the load
resultant also fullls F < 0 and the balance point T belongs to the interior of the convex
closure of subsoiled parts.
To determine the sensitivity of the solution on a change of the load, we will consider the
class S
,,
of the loads L V

such that T [x
l
+ , x
r
], F < 0 and L

,
with respect to positive parameters , , . If we will assume that S
,,
is non-empty then
there exists a positive constant c which depends on the loads from S
,,
only through the
parameters , , such that
w

1
w

2,2
c(, , )L
1
L
2

L
1
, L
2
S
,,
, (6)
where w

i
= w

i
(L
i
) solves the problem (P) with respect to the load L
i
, i = 1, 2, see [16,
Theorem 3.3]. Notice that the constant c is independent of the load for coercive problems
(e.g. xed or supported beams).
Here, since the constant c depends on the parameters , , , the solution need not be
stable with respect to a small change of the load. In [19, Lemma 2.1], two types of the loads
(so called unstable loads), which cause solution instability, were described . The rst type
of the unstable load can happens if the balance point T is closed to the end points of the
subsoiled part
s
( 0). The second type can happen if the load resultant F is small with
respect to the size of the load ( 0 and L

0). Both types of the load instability will

be illustrated and interpreted on numerical example in Section 5. Notice that the unstable
load can also be characterized such that the corresponding beam deection activates only a
small part of the subsoil, see [19, Lemma 5.3]. Moreover the solutions of the problems with
unstable loads are sensitive to rigid motions.
The sensitivity of the solution on the material parameter is also analyzed in [16, Theorem
3.3].
3. Approximation of the problem
Since the solution of the problem (P) cannot be analytically found in general due to the
non-linearity, it is suitable to approximate the problem. Here, the nite element method
will be used. Moreover, it is not suitable to evaluate the term b(w

, v) directly due to the

non-linear term w

. Thus a numerical approximation of the form b will be used.

An interesting alternating approach is to work with the primal-dual formulation of the
problem where the non-penetration condition between the beam and the subsoil is unforced
in, see [11, 12, 17]. Then the corresponding problem can be approximated by the mixed
nite element method and solved by complementarity or augmented Lagrangian methods,
see [11, 12].
In this section, we set a nite element approximation of the problem (P), summarise
the relation between the original problem and its approximations and nally, the problem is
rewritten into the algebraical form.
Let us dene a partition
h
,
0 = x
0
< x
1
< . . . < x
N
= l, h := max
j=1,...,N
(x
j
x
j1
), h
min
:= min
j=1,...,N
(x
j
x
j1
)
of the interval = [0, l], with the nodal points x
j
, j = 0, 1, . . . , N, and with the parameters
h, h
min
> 0. With respect to a positive parameter , we will consider the system T

of such
partitions
h
for which the inequality h h
min
holds.
For a partition
h
T

with N + 1 nodal points, we will dene the function space

V
h
H
2
(), V
h
:= {v
h
C
1
() | v
h
|
(x
j1
,x
j
)
P
3
, j = 1, 2, . . . , N},
i.e. the space of continuously dierentiable and piecewise cubic functions.
For the sake of simplicity, we will assume that the function q, which represents the stiness
coecient of the subsoil, is piecewise constant in the interval
s
and that the considered
partition
h
T

take into account the points of discontinuity of q. Since the evaluation

of the term b(w

h
, v
h
), w
h
, v
h
V
h
, cannot be computed exactly due to the non-linear term
w

h
, an approximation of the form b must be used. The form b will be approximated by a
numerical quadrature on each subsoiled partition interval. Its approximation has the form
b
h
(v
1
, v
2
) :=
m(h)

i=1
r
i
v
1
(z
i
)v
2
(z
i
), v
1
, v
2
H
2
(), (7)
where z
i
, z
1
< z
2
< . . . < z
m(h)
, are the points of the numerical quadratures and the coe-
cients r
i
are equal to the products of the stiness coecients and weights of the numerical
quadrature. With respect to the assumption on
h
T

, it holds that there exists constants

c
1
, c
2
> 0 such that
c
1
q
0
h r
i
c
2
q

h, i = 1, 2, . . . , m(h). (8)
From a mechanical point of view, the subsoil is substituted by insulated springs. However
the number m and the stinesses r
i
of the springs depend on the discretization parameter h.
We will assume that the numerical quadrature is exact at least for polynomials of the rst
degree.
Now, we set the approximated problem (P
h
):
(P
h
) nd w

h
V
h
: a(w

h
, v
h
) + b
h
((w

h
)

, v
h
) = L(v
h
) v
h
V
h
. (9)
The corresponding potential functional has the form
J
h
(v
h
) :=
1
2
a(v
h
, v
h
) +
1
2
b
h
(v

h
, v

h
) L(v
h
).
The existence of the problem (P
h
) solution is ensured by the condition
F < 0 and z
1
< T < z
m(h)
, (10)
see [16, Lemma 4.7]. This condition also ensures the uniqueness of the solution for su-
ciently small h. Notice that if the condition (5) holds and the discretization parameter h is
suciently small, then the condition (10) also holds.
It holds that the set
A

h
:= {i {1, . . . , m(h)} | w

h
(z
i
) < 0} , (11)
which represents active springs is non-empty. In addition, the balance point T belongs to
the convex closure of the points {z
i
; i A

h
}.
It holds that w

h
w

in H
2
() for h 0. Moreover, if w

belongs to the suciently

wide class of the functions from H
4
() and the numerical quadrature is exact at least for
polynomial of the rst degree then the convergence order is two, see [16, Theorem 4.6].
The algebraic form of the problem (P
h
) will be denoted by (P) and has the form of the
non-linear system of equation
(P) nd w

R
n
: F(w

) = Kw

+B
T
D(Bw

= f. (12)
Here, a vector w R
n
, n = 2N + 2, contains the function values and the values of the rst
derivatives of the function w
h
at the nodal points x
j
, j = 0, 1, . . . , N. The symbol u

denotes
the negative part of a vector u, i.e. u

i
:= min{0, u
i
}, i = 1, 2, . . . , m, u R
m
. The symbol
K R
nn
denotes the stiness matrix of the beam, the vector f R
n
represents the load,
the diagonal matrix D R
mm
contains the coecients r
i
, i = 1, . . . , m, and the matrix
B R
nm
transforms the function values and the values of the rst derivatives at the nodal
points x
j
, j = 0, 1, . . . , N, onto the function values at the points z
i
, i = 1, . . . , m. Notice
that the the matrix B
T
DB is diagonal only if the quadrature points z
1
, . . . , z
m
coincide
with the partition points.
The potential functional has the form
J(w) :=
1
2
(Kw, w)
n
+
1
2

D(Bw)

, (Bw)

m
(f, w)
n
,
where (., .)
m
, (., .)
n
denote the Euclidean scalar products in R
m
and R
n
, respectively.
The below convergence estimates will be related to the energy norm of the beam on the
bilateral (linear) subsoil. The norm will be denoted by the symbol .
E
and is induced by
the scalar product
(v, w)
E
:= (Kv, w)
n
+ (DBv, Bw)
m
=

(K +B
T
DB)v, w

n
, v, w R
n
.
Since the evaluation in the energy norm is independent of the discretization parameter h, we
can decide whether the below convergence estimates related to the norm are independent of
discretization parameter h or not.
It will also be useful do introduce the matrix R R
n2
given by the functions 1 and x
to represent polynomials from P
1
(representation of rigid motions). Notice that KR = 0.
Let us denote
G := BR =

1 1 . . . 1
z
1
z
2
. . . z
m

T
, e := R
T
f = F

1
T

, (13)
where F, T are the load resultant and the balance point, respectively.
4. Semismooth Newton method with damping
Systems of non-linear equations are often solved by Newton-like methods. Notice that the
operator F : R
n
R
n
in the system (12) is not dierentiable in a classical way. Therefore it
is suitable to use some of the non-smooth variants of the Newton method, see e.g. [3, 15, 19].
We will solve the system of equations (12) by the so called semismooth Newton method
with damping, see [19]. In the article [19], the method is interpreted as a descent direction
method since we also minimize the functional J in the damping step. The method has also
been applied to elasto-plasticity [2, 20].
In this section, rstly, we describe the properties of the operator F. Then we introduce the
semismooth Newton method without and with damping. We summarise their convergence
results, advantages and disadvantages. Finally, we introduce the modication of the method
suitable for unstable loads and its interpretation.
The operator F fullls the estimate
0 (F(v) F(w), v w)
n
v w
2
E
v, w R
n
. (14)
Thus F is Lipschitz continuous and monotone in R
n
. Let v R
n
be such a vector that
(Bv)
i
= 0, i = 1, . . . , m. Then we can dene the derivative of F in the form
F

(v) R
nn
, F

(v) = K +B
T
DA(v)B,
with the diagonal matrix
A(v) R
mm
, (A(v))
ii
=

1, (Bv)
i
< 0
0, (Bv)
i
> 0
.
If we set for example (A(v))
ii
= 0 for (Bv)
i
= 0 we can dene the generalized derivative
F
o
(v) R
nn
of the function F in the form
F
o
(v) = K +B
T
DA(v)B
for any v R
n
. The operator F
o
represents the generalized derivative of F in the following
sense:
lim
w0
F(v +w) F(v) F
o
(v +w)w
n
w
n
= 0 (15)
for any v R
n
and suciently small w R
n
. The operator F
o
is the so called slant
function for F. The function F is semismooth (see [3]) with respect to the estimates (14)
and (15). Notice that the estimate (15) depends on the dimension n, i.e. on the discretization
parameter h.
The semismooth Newton iterates have the form
w
k+1
= w
k

F
o
(w
k
)

1
(F(w
k
) f), k = 0, 1, . . . . (16)
So we need to solve the linear systems with the matrix F
o
(v) = K +B
T
DA(v)B. Such a
matrix is symmetric and
0

(K +B
T
DA(v)B)w, w

n
w
2
E
v, w R
n
.
It means that the matrix F
o
(v) is only positive semi-denite in general. For example if the
beam deection represented by a vector v activates less than two springs, i.e. if there exist
less than two indices i with the property (Bv)
i
< 0, than the system with the matrix F
o
(v)
has not a unique solution and thus the Newton iterates (16) need not be well dened.
If there exist more or equal than two indices i with the property (Bv)
i
< 0 (a number
of active springs is more or equal than two), than the matrix F
o
(v) is positive denite and
there exists a positive constant c = c(h) such that
c(h)w
2
E

(K +B
T
DA(v)B)w, w

n
w
2
E
(17)
for any w R
n
and any v with the above property, see [19, Lemma 3.1]. The constant c
depends on the discretization parameter h due to the estimate (8).
To suppress the dependence of c on h in the estimate (17), we must assume that the
number of active springs is proportional to 1/h. Therefore we introduce the following nota-
tion:
A
h
(v) := {i {1, . . . , m(h)} | (Bv)
i
= v
h
(z
i
) < 0},
A

:=

h
{A
h
{1, . . . , m(h)} | card(A
h
) max{2, /h}} , > 0.
Then there exists a positive constant c = c() such that
c()w
2
E

(K +B
T
DA(v)B)w, w

n
w
2
E
v, w R
n
, A
h
(v) A

, (18)
see [19, Lemma 3.2]. Notice that the estimate (18) holds for coercive beam problems even
for any v R
n
and without the dependence on .
If we assume that the solvability condition (4) holds and the parameter h is suciently
small then there exists > 0 such that A
h
(w

) A

where the vector w

is the solution
of the investigated problem (P), see [19, Lemma 3.8]. It implies that the matrices F
o
(v),
with v in a neighborhood of the vector w

, fullls the uniform estimate (18) and hence the

semismooth Newton iterates (16) converge locally superlinearly to w

, see [3, Theorem 3.5].

Notice that we work with the semismooth Newton method in the same way as with the
standard Newton method from a numerical point of view since the set of non-dierentiable
points has the zeroth measure and thus it is unlikely to get such points in a computer
precision. The initial choice w
0
= 0 is the exception. In this case, it is suitable to specically
dene the initial value of the generalized derivative, see Algorithm 1 below.
The semismooth Newton method is theoretically correct only if the initial approximation
w
0
is closed to the solution w

. Moreover the local superlinear convergence depends on

the discretization parameter h due to the estimate (15). To suppress these drawbacks, the
semismooth Newton method with damping was proposed. The algorithm of the method is
following:
Algorithm 1
Initialization
w
0
= 0,
A
0
= I (the identity matrix).
Iteration k = 0, 1, . . .
s
k
, (K +B
T
DA
k
B)s
k
= f F(w
k
),

k
= arg min
01
J(w
k
+ s
k
),
w
k+1
= w
k
+
k
s
k
,
A
k+1
= A(w
k+1
).
Notice that if we omit the damping step and set
k
= 1 we obtain the above semismooth
Newton method. The method with damping is also a descent direction method since the
potential functional J is minimized in the damping step. The damping coecients
k
can
be computed for example by the regula-falsi method.
In comparison to the above method, the particular initial approximation is given. The
initial vector w
0
= 0 need not to belong to a suciently small neighborhood of the solution
w

.
If we assume that the solvability condition (4) holds and the parameter h is suciently
small then there exists > 0 such that
A
h
(w
k
) A

, k = 0, 1, . . . ,
see [19, Theorem 4.1]. It means that Algorithm 1 is well denite due to the estimate (18).
Further we can consequently derive the following estimates:
(J

(w
k
), s
k
)
n
cs
k

2
E
, (19)

k
c, if s
k
= 0, (20)
J(w
k+1
) J(w
k
)
1
2
c
2
k
s
k

2
E
, (21)
+

k=0
s
k

2
E

2(J(0) J(w

))
c
3
, (22)
+

k=0
w

w
k

2
E

2(J(0) J(w

))
c
5
, (23)
where the constant c = c() > 0 is given by the estimate (18), see [19, Theorem 4.2]
2
.
These estimates implies that the iterates w
k
converge to the solution w

. Moreover since
2
Notice that the standard norm of the Sobolev space was used in [19] instead of the energy norm in
(19)-(23) since Algorithm 1 was introduced for the variational formulation of the problem (9) therein.
the estimates are independent of the discretization parameter h, this convergence is uniform
with respect to h. The similar results have been proved for an elasto-plastic application
of the method, see [20, Theorem 5.2]. In [20, Theorem 5.2], there is also proved that the
damping coecients

k
1. (24)
The same result can also be expected here. It means that the damping eect is neglected in
the neighborhood of the solution and the method with damping has also local superlinear
convergence which is dependent on h in general. Moreover, in all the tested numerical
examples with initial choice w
0
= 0, the damping parameters
k
were even equal to one and
thus the methods with and without damping coincided.
The above results are conditioned by the assumption that the parameter h is suciently
small. Notice that the maximal possible size of h depends mainly on the load stability. It
is necessary to use a ner partition for unstable loads. This drawback can be suppressed by
the following modication of the method which also improve the convergence properties for
unstable loads:
Algorithm 2
Initialization
w
0
= Rc
0
, c
0
solves G
T
D(Gc
0
)

= e,
A
0
= A(w
0
),
Iteration k = 0, 1, . . .
s
k
, (K +B
T
DA
k
B)s
k
= f F(w
k
),

k
= arg min
01
J(w
k
+ s
k
),
w
k
= w
k
+
k
s
k
,
c
k
, c
k
solves G
T
D(Bw
k
+Gc
k
)

= e,
w
k+1
= w
k
+Rc
k
,
A
k+1
= A(w
k+1
).
Notice that Algorithm 2 contains the additional step where the system
G
T
D(Bv +Gc)

= e, c R
2
, (25)
of two non-linear equation of two unknowns is solved for a given v R
n
. This system
represents the following variational problem
nd p
h
= p
h
(v
h
) P
1
: J
h
(v
h
+ p
h
) J
h
(v
h
+ p) p P
1
,
or equivalently
nd p
h
= p
h
(v
h
) P
1
: b
h
((v
h
+ p
h
)

, p) = L(p) p P
1
, (26)
for a given function v
h
. It can be interpreted so that we are looking for a rigid motion of
the deformed beam which minimizes the energy functional. Especially, for the initial choice
v
h
= 0, we are looking for the deection of a rigid beam on a unilateral elastic subsoil. If
we take into account that the solutions of the problems with unstable loads are sensitive to
rigid motions, it is clear that the additional step stabilises the algorithm.
The solvability of (26) is ensured by the condition (4). This problem can be solved by
the algorithm which is introduced in [19].
Algorithm 2 also represents a descent direction method and the estimates (19)-(24) hold,
see [19, Theorem 4.3]. Therefore the method has the same convergence properties as the
previous one.
The additional step can also be interpreted as a projection step. Let us dene the set
:= { R
m
| 0, G
T
D = e}.
The set represents the admissible Lagrange multipliers from the dual formulation of the
problem, see [17, Section 5.2], where the non-penetration condition between the beam and
the subsoil is unforced. The set is closed, convex and bounded in R
n
, see [19, Lemma 5.1].
The diameter of depends mainly on the load stability. The diameter is small for unstable
loads, see [19, Lemma 5.3]. If the solvability condition (4) does not hold, then the set is
either empty or degenerate into one point = 0. If the solution w

exists then the vector

(Bw

representing the deformation of the subsoil belongs to . It can be easily proved if

we multiply from the left the equation (12) by the matrix R
T
.
Since the set is closed, convex and non-empty, we can uniquely dene the projection
P of the space R
m
onto the set with respect to the scalar product (D., .)
m
in R
m
, i.e.
(D( P()), P())
m
0 . (27)
Let v R
n
and let c = c(v) R
2
solve the system (25). Then the vector (Bv + Gc)

belongs to and
P(Bv) = (Bv +Gc)

.
Hence the vectors w
k
, k 0, generated by Algorithm 2, also fullls (Bw
k
)

. Since
the diameter of is small for unstable loads, the vectors (Bw
k
)

are closed to the vector

(Bw

, which means that the vectors Bw

k
have the similar set of the active springs as
the vector Bw

and consequently the iterates w

k+1
are closed to w

. Therefore we can
expect better convergence properties for Algorithm 2 than for Algorithm 1 for such loads.
It will also be demonstrated on numerical examples in the next section. Notice that the
additional (projection) step is important mainly in the initial part of Algorithm 2.
5. Numerical Examples
In this section, some theoretical results of the problem and the numerical methods will
be illustrated on numerical examples. In all the examples, we will consider the beam of the
length l = 1 m and the equidistant partition with 10 2
j
, j = 1, 2, . . ., elements. We use the
following stopping criterion:
r
k

n
f
n
10
8
, r
k
:= f Kw
k
B
T
D(Bw
k
)

and the numerical quadratures

1
1
() d 2(0),

1
1
() d (1) + (1),

1
1
() d (

3/3) + (

3/3),
which are denoted by NQ
1
, NQ
2
, NQ
3
, respectively.
Comparison of the algorithms. Let EI = 5 10
5
Nm
2
, x
l
= 0.1 m, x
r
= 0.9 m and
q = 2 10
7
Nm
2
. At the end points 0, l of the beam, we will consider the point loads
0 1m 0.1 T
1
T
2
0.9
F
0
F
l
?
?
Fig. 2: Scheme of the tested problem.
F
0
and F
l
. In Example 1, we choose F
0
= 5000 N, F
l
= 5000 N and F
0
= 5000 N,
F
l
= 1000 N in Example 2. The situation is depicted in Fig. 2.
The load is stable in Example 1, since the balance point T
1
of the load lies in the central
part of the subsoil, contrary to Example 2, see Fig. 2. The dependence of the number
of Newton iterations on the number of elements of the partition is shown in Tab. 1. The
quadrature NQ
3
is used.
Ex. 1 40 80 160 320 640 1280
ALG1 4 3 4 4 4 4
ALG2 3 3 3 3 3 3
Ex. 2 40 80 160 320 640 1280
ALG1 6 6 7 8 7 8
ALG2 2 2 2 2 2 2
Tab. 1: Numbers of Newton iterations in dependence on the numbers of elements.
Notice that the number of Newton iterations does not depend on renement of the
partition for both algorithms. In Example 1, the numbers of Newton iterations are practically
the same for both algorithms, contrary to Example 2, which justies the theoretical result.
The graphs of the approximated solutions for both examples are depicted in Fig. 3.
0 0.2 0.4 0.6 0.8 1
2
1.5
1
0.5
0
0.5
1
x 10
4
x
w
*
0 0.2 0.4 0.6 0.8 1
2.5
2
1.5
1
0.5
0
0.5
1
x 10
4
x
w
*
Fig. 3: The approximated deections of the beam axis w

for Examples 1, 2.
Dependence of the solution on the load. We investigate the inuence of the small change
of the load on the change of the solution of the problem (P) in three examples. The load is
stable in Example 1 and unstable in Examples 2 and 3. In particular, we will evaluate the
ratio
c :=
w

1
w

E
w

E
f
1
f
2
n
f
1
n
,
where the vectors f
1
, f
2
represent the loads and the vectors w

1
, w

2
the corresponding solu-
tions.
Let EI = 20 Nm
2
, x
l
= 0.2 m, x
r
= 0.8 m and q = 1.2 10
6
Nm
2
. We use the
equidistant partition with 80 elements and the numerical quadrature NQ
3
. We consider
three constant loads P
1
, P
2
, P
3
, which are situated in the intervals (0 m, 0.1 m), (0.4 m, 0.6 m)
and (0.9 m, 1 m), respectively, in all the examples. The particular values of the load and
their small changes (in brackets) are in Tab. 2 and the schemes of Examples 2,3 are depicted
in Fig. 4, 5.
P
1
[Nm
1
] P
2
[Nm
1
] P
3
[Nm
1
]
Ex. 1 -100 (-1) 0 -100 (-1)
Ex. 2 -100 0 -450 (-1)
Ex. 3 100 -110 100 (+1)
Tab. 2: Loads and their changes.
0 1m 0.1 0.2
T
0.8 0.9
P
1
P
3 ??? ? ?
? ? ???
Fig. 4: Scheme of Example 2
0 1m 0.1 0.2 T
1
T
2
0.8 0.9 0.4 0.6
P
1
P
3
P
2
666 6 6
??????????
6 6 666
Fig. 5: Scheme of Example 3
In Example 1, the ratio c = 1.00 and the dierence between the solutions w

1
, w

2
is not
even visible graphically, see Fig. 6.
In Example 2, the ratio c = 25.83 which is signicantly greater than in stable Example
1. The dierence between the solutions w

1
, w

2
is visible graphically, see Fig. 7. The load is
unstable here since the balance point T
1
= 0.79 is closed to end points of the subsoil. It is
clear from Fig. 7 that the other moving of the balance point T
1
to the point x
r
= 0.8 can
turn over the beam from the subsoil.
In Example 3, the ratio c = 44.53 which is also signicantly greater than in stable
Example 1. The dierence between the solutions w

1
, w

2
is visible graphically, see Fig. 8.
The load is unstable here since the load resultant (approximately 2 N) is small in comparison
with the size of the load. Due to this fact, the compression of the springs is minimal and the
problem behaves similar to the Neumann problem without subsoil. It is well-known that the
dierence between solutions of the Neumann problem can be described by polynomials of
the rst degree. Thus also the compared solutions diers approximately by a polynomial of
the rst degree. This polynomial not vanishes since the shape of the solutions is convex and
the small change of the load meant a signicant moving of the balance point T
1
, see Fig. 5.
Convergence properties of the numerical quadratures. Let EI = 1 Nm
2
, x
l
= 0.2 m,
x
r
= 0.8 m, q = 1 Nm
2
and let the load density have a form
f(x) =

384

x
1
2

2
+ 32, x (0,
1
5
) (
2
5
,
3
5
) (
4
5
, 1),

16
15

x
1
2

6
+
4
3

x
1
2

4
385

x
1
2

2
+
148816
15000000
+ 32, x (
1
5
,
2
5
) (
3
5
,
4
5
).
0 0.2 0.4 0.6 0.8 1
10
8
6
4
2
0
2
4
x 10
3
x
w
*
w
1
*
w
2
*
Fig. 6: Dependence of the solution on the load - Example 1.
0 0.2 0.4 0.6 0.8 1
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
x
w
*
w
1
*
w
2
*
Fig. 7: Dependence of the solution on the load - Example 2.
0 0.2 0.4 0.6 0.8 1
5
0
5
10
15
20
x 10
3
x
w
*
w
1
*
w
2
*
Fig. 8: Dependence of the solution on the load - Example 3.
Then the exact solution of the problem (P) is a function
w

(x) =
16
15

x
1
2

6
+
4
3

x
1
2

x
1
2

2
+
148816
15000000
.
Notice that w

H
4
() and thus the convergence order is equal to 2. Therefore, we will
evaluate the terms
w

2,2
/h
2
, v
2
2,2
:=

1
0
(v

)
2
dx+

4/5
1/5
v
2
dx and |w

h
|
0,2
/h
2
, |v|
2
0,2
:=

1
0
v
2
dx
for numerical quadratures NQ
1
, NQ
2
, NQ
3
. The approximated values of the terms are
described in Table 3.
w

2,2
/h
2
10 20 40 80 160
NQ1 1.0422 1.0606 1.0651 1.0663 1.0666
NQ2 1.0423 1.0606 1.0652 1.0663 1.0664
NQ3 1.0422 1.0606 1.0651 1.0663 1.0666
|w

h
|
0,2
/h
2
10 20 40 80 160
NQ1 0.0575 0.0577 0.0577 0.0577 0.0585
NQ2 0.1559 0.1351 0.1289 0.1302 0.1388
NQ3 0.0005 0.0001 0.0000 0.0000 0.0017
Tab. 3: Error estimates of the quadratures in dependence on the number of elements.
The results conrm that the order of convergence is two for this example for all the
quadratures. The quadratures NQ
1
, NQ
2
are exact for polynomials of the rst degree, the
quadrature NQ
3
is exact for polynomials of the third degree. While the values of the rst
term are quite similar for all the quadratures, the values of the second term are low order
for NQ
3
than for NQ
1
or NQ
2
.
The graphical comparison of the dierences w

h
and (w

(w

h
)

for the investigated

quadratures is depicted in Fig. 9. It is considered 20 nite elements.
0 0.2 0.4 0.6 0.8 1
4
3
2
1
0
1
2
x 10
4
x
w

w
h
20 elements
NQ
1
NQ
2
NQ
3
0 0.2 0.4 0.6 0.8 1
8
6
4
2
0
2
4
6
8
x 10
5
x
w

h
20 elements
NQ
1
NQ
2
NQ
3
Fig. 9: Comparison of the quadratures NQ
1
, NQ
2
, NQ
3
for the function values and the
values of the rst derivatives.
Notice that the values of the rst derivatives have the similar behavior for all the numer-
ical quadratures, in comparison to the function values. However the better approximation
of the quadrature NQ
3
for the function values was not conrmed for another loads.
6. Conclusion
The semicoercive model of a beam on a unilateral elastic subsoil of the Winklers type has
been investigated. The model can be more accurate in some situations than the model with
the standard linear subsoil. The article summarized some of the results from mathematical
and numerical modeling of the problem.
The main results of the work are following: setting the necessary and sucient conditions
for existence and uniqueness of the solution, sensitive analysis of the problem, setting a priori
estimates between the solutions of the continuous and approximated problems, convergence
analyses of the investigated numerical methods and the investigation of the convergence
properties for the unstable loads. Some of the results were mechanically interpreted and
numerically illustrated.
Many of the results simplies if we consider coercive problems with xed or supported
beams. On the other hand, some of the results can be generalized for other problems, for
example for a system with more layers of the subsoils or thin plates. The numerical methods
can be used for another non-linear problems. Especially, the Newton methods with and
without damping have been used for an elasto-plastic application.
Acknowledgements
This work was supported by the Academy of Sciences of the Czech Republic, Institutional
Research Plan No. AVOZ 30860518.
References
[1] Adams, R.A.: Sobolev Spaces, Academic Press, New York, 1975.
[2] Byczanski, P., Sysala, S.: Modied semismooth Newton method: Numerical exam-
ple, In: Proceedings of the seminar SiMoNA 2009, TU Liberec, 2009, pp. 24-30, ISBN
978-80-7372-543-3.
[3] Chen, X., Nashed, Z., Qi, L.: Smoothing methods and semismooth methods for
nondierentiable operator equations, SIAM J. Numer. Anal. 38 (2000), pp. 1200-1216.
[4] Frydr ysek, K.: Nosnky a ramy na pruznem podkladu 1, V

SB - TU Ostrava, 2006, in
Czech, ISBN 80-1244-4.
[5] Frydr ysek, K., Janco, R.: Nosnky a ramy na pruznem podkladu 2, V

SB - TU
Ostrava, 2008, in Czech, ISBN 978-80-248-1743-9.
[6] Hlavacek, I., Lovsek, J.: Optimal Control of Semi-coercive Variational Inequalities
with Application to Optimal Design of Beams and Plates, Zeitschr. Angew. Math. Mech.
78, 1998, pp. 405-417.
[7] Horak, J.V., Netuka, H.: Mathematical models of non-linear subsoils of Winklers
type, In: Proceedings of 21
st
Conference Computational Mechanics 2005, Z

CU v Plzni,
pp. 235-242, 431-438, in Czech.
[8] Horak, J.V., Netuka, H.: Mathematical Model of Pseudointeractive Set: 1D Body
on Nonlinear subsoil, I. Theoretical Aspects, Engineering Mechanics, 14, 2007, pp. 311-
325.
[9] Horak, J.V., Sysala, S.: Modelovan ohybu desky na jednostranne pruznem podloz,
In: Proceedings of 13
st
Conference Modern matematicke metody v inzen yrstv, J

CMF,
V

SB TU Ostrava, 2004, pp. 51-55, in Czech.

[10] Necas, J., Hlavacek, I.: Mathematical Theory of Elastic and Elasto-Plastic Bodies.
An Introduction, Elsevier, Amsterdam, 1981.
[11] Netuka, H.: A new approach to the problem of an elastic beam on a nonlinear foun-
dation. Part 1: Formulations, Applied and Computational Mechanics, in print.
[12] Netuka, H., Machalova, J.: A new approach to the problem of an elastic beam
on a nonlinear foundation. Part 2: Solution, Applied and Computational Mechanics,
submitted.
[13]

Simecek, R.: Optimalizace nosnku na jednostranne pruznem podloz: Existence
resen, In: Proceedings of Seminar ODAM, 2009, pp. 68-84, in Czech.
[14] Svobodova, I.: Cone Decomposition of Hilbert Space: Existence of Weak Solution for
1D Problem, 2005, http://homel.vsb.cz/ svo19/, pp. 19.
[15] Svobodova, I.: Numerical Approximation of the Non-linear Fourth-Order Boundary-
Value Problem, In: Proceedings of Seminar Programs and Algorithms of Numerical
Mathematics 14, Doln Maxov, 2008, pp. 201-206.
[16] Sysala, S.: Unilateral Elastic Subsoil of Winklers type: Semi-Coercive Beam Problem,
Applications of Mathematics, 53, 2008, pp. 347-379.
[17] Sysala, S.: Modelovan nelinearn semikoercivn ulohy nosnku na podloz, Ph.D. The-
sis, UP Olomouc, 2008, in Czech.
[18] Sysala, S.: Numerical Illustration of Theoretical Results for Non-linear Semi-coercive
Beam Problem, In: Proceedings of Seminar on Numerical Analysis - SNA08, TU
Liberec, pp. 110-114.
[19] Sysala, S.: Numerical Modelling of Semi-coercive Beam Problem with Unilateral Elas-
tic Subsoil of Winklers Type, Application of Mathematics, 55, 2010, pp. 151-187.
[20] Sysala, S.: Application of a modied semismooth Newton method to some elasto-
plastic problems, submitted to Mathematics and Computers in Simulations.
Reviewer: doc. RNDr. Josef Malk, CSc., Institute of Geonics AS CR, Czech Republic.