Sie sind auf Seite 1von 11

J. Inst.

Maths Applies (1969) 5, 329-339

Properties of a Semi-discrete Approximation


to the Beam Equation t
ROLAND A. SWEET

[Received 19 January 1968, and in revised form 29 October 1968]


The solution of the equation
w(*)Hrt+[p(*)Jx = 0 ,
0 < x < L,
t>0,
where it is assumed that w and p are positive on the interval [0,L], is approximated by
using the method of straight lines. The resulting approximation is a linear system of
differential equations with coefficient matrix S. The matrix S is studied under very general
boundary conditions which result in a conservative system. In all cases the matrix S is
either an oscillation matrix or possesses nearly all the properties of an oscillation matrix.
1. Introduction
of small transverse (or lateral) vibrations of a beam with non-uniform
cross-sectional area and moment of inertia is of wide interest. Of particular importance
in the design of structures which possess the characteristics of a beam, for instance,
chimneys, are the natural frequencies and mode shapes. The equation describing the
motion is impossible to solve in most cases, hence information is obtained by numerical
approximating techniques. Linearly tapered cantilever beams have been studied by
Housner & Keightley (1963).
The general equation describing the free vibrations is
wix)ult+[pix)ixx\n
= 0,
0<x<L,
t>0.
(1)
In the case of forced vibrations the right side of (1) is not zero, but rather, some function
/ o f jcand t.
Throughout this paper we will assume that
THE STUDY

w(x) > 0 and p(x) > 0,

0 < x < L.

These conditions are physically meaningful and unrestrictive in vibration problems.


The primary goal of this paper is the study of the coefficient matrices which arise
through the use of the method of straight lines (a discretization of the space variable
to produce a linear system of ordinary differential equations). We prove that under
very general boundary conditions the coefficient matrices possess all the properties
of oscillation matrices (see Gantmakher & Krein, 1950).
2. Preliminaries
To solve equation (1) one must prescribe initial conditions

u(x,0) = 4>(x),

ut(x,0) = Mx),

(2)

t This research was supported by an NDEA Fellowship at Purdue University, West Lafayette,
Indiana.
329

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

Department of Computer Science,


Cornell University, Ithaca, New York, U.S.A.

330

ROLAND A. SWEET

and
B(U,X) = (>KXI)*"r-i'xx"x<-

The quantity E(u, t) gives the total energy (kinetic plus potential) of the system at the
time t. By the conservation of energy E is constant for all values of t. Therefore,
we must choose boundary conditions which make

f [ ( , ) ( ) ]

Jo

This may be done by choosing those conditions for which


5(II,0) = B(u,L) = 0.
(3)
Hence, we will consider only those boundary conditions which satisfy (3).
Table 1 lists the various boundary conditions which we have selected for consideration. We stress that the definition of the functions E and B depend on the type of
boundary conditions under consideration. The energy function for boundary condition 5-5, for instance, would be

Hence, equation (3) involves the new function


B(u,x) ^pu^U;,.
This paper is concerned with the approximations obtained by the method of straight
lines (see Berezin & Zhidkov, 1965).
Essentially the method is concerned only with finding the solution to equation (1)
at afinitenumber of points {xj}"^- For ease let us assume the points are chosen so that
where XQ = 0 and h = L\{n +1). Then xn+1 = L. Genin & Maybee (1966) develop this
method by integrating the equation over each interval [xl\h, xt+ih] and approximating all integrals not containing a partial derivative of x by the midpoint rule. Let
w(xD = wh X*i) = Pi> (' = >1 n+1).

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

for 0 < x < L, and four linearly independent boundary conditions. These boundary
conditions are usually given in the form of two conditions at the end x = 0 and two
at the end x = L, although this is not the only case which can occur (for example,
periodic boundary conditions can also occur). We will concern ourselves, however,
only with the former type.
We shall consider only those boundary conditions which correspond to conservative
problems. Multiplying (1) by ut, integrating over the rectangle 0 ^ x ^ L, 0 ^ t ^ T,
and using integration by parts twice we get
f r [L
0=
!>,+(pOr] dx dt
J oJ o
= E(u,T)-E(u,0)+ \T[B(u,L)-B(u,O)-]dt,
Jo
where
L
1 f
(".'i) = 2
[wuf+pull^dx

SOLUTION OF THE BEAM EQUATION

331

Then, if we denote the approximation of u(x,, t) by ufj), one has


Cxt + ik

0=

fxi + i*

w(x)uttdx+\

(5)
TABLE 1
Boundary Conditions

Case

Condition (at x = 0)

3
4
5t

xx(0,0 = CP"x*)x(0,')=0
ux(0, t) = 0 ) ^ ( 0 , 0 = 0
uJ0,t) = 0

Condition (at x = L)

ux{L, t) = (pu^L,
ux(L,t) = 0

6t
au(0,t) = GM^CXXXO,/)

7t

-bu^Tt^ujltt)

8t

(p$x(lojUf

9t

a(0,r) = (pu^JP,!)
cujp,t) = u^O.t)

bu(L,t) = (puxJx(L,t)
-dux(L,t) = Uxx(L,t)

t Note: we assume a > 0, b > 0, c > 0, and d > 0.

If an equation of the type of (5) is written for each unknown , one obtains as the
approximation, a linear system of coupled ordinary differential equations. The system
has the form
Hi)+/t^SU = F,
(6)
where H is a diagonal matrix with positive elements, 5 is a pentadiagonal matrix,
F is a vector whose components depend on the nature of the boundary conditions,
and U is the vector whose ith component is the function ut).
The precise forms of H and S for various boundary conditions will be exhibited
in the next section. All the boundary conditions we consider yield an F which is
zero, hence we shall solve (6) in its homogeneous form. To do this we seek a solution
of the form
U(f) = V(a sin VA /+P cos JX t),
(7)
where V is a vector and a, /?, A, are constants to be determined. Substituting (7) into (6)
and abbreviating H^S to just S we arrive at the condition
(SV-AHV)(a sin Jk t+0 cos VA 0 = 0.

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

= hwfit + [(pOx]x'-xt-i*
(4)
If we now replace all partial derivatives of x by central difference approximations,
we obtain from (4)
0 = AW|fi(+A

332

ROLAND A. SWEET

For this to be true for all values of /, it must be that


that is, k must be an eigenvalue of H-'S and V a corresponding eigenvector. If S and
H are of order m, then there are m eigenvalues X\,X2,...,Xm and (we assume) m corresponding eigenvectors V\, Vz,...,Vm. Then for each Xk and Vk (7) is a solution of (6).
Hence,

is a solution of (6), the coefficients ak and pk (k = l,2,...,m) being determined by the


initial conditions (2), i.e.
U(0) = <D,
0(0) = ,
where
<S>k = <l>(xk) and ^ = ^ 0 0
(fc = 0,l,2,...,n + l).
Therefore, knowledge of the structure of the eigenvalues and eigenvectors of the
matrix H~iS is desirable. To prepare for this let us state some definitions and main
results of Gantmakher & Krein (1950) for the arbitrary matrix A of order m with
elements a,j (1 < i, j < m).
We will use the notation d(A) to denote the determinant of A. The minors of A
are all possible numbers

,Jl

)l

where 1 < i\ < ii < ... < ip < m, and 1 ^j\<

ji < ... < jp ^ m. Also, we will

denote the pth leading principal minor by Ap, i.e.


Ap = A

2 ...

[l

Definition. The matrix A is said to be totally non-negative (totally positive) if all


its minors are non-negative (positive).
Definition. The matrix A is said to be an oscillation matrix if A is totally nonnegative, and if there exists a positive integer/) such that A* is totally positive.
Definition. The matrix A* is defined to be the matrix of order m with elements
ij

ij*

At this point let us note that the transformation of A into A* is a similarity transformation by the orthogonal matrix
D = diag[(l) 1 ^I) 2
(1)].
Hence, A and A* have the same eigenvalues. Furthermore, if the elements of A satisfy
then A* = |A|, where the elements of |A| are just \atJ\.
THEOREM

1. An oscillation matrix A has only simple, positive eigenvalues:


Ai > Xi> ... > km> 0.

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

U(0 = t=if (ottsin W + ^ c o s W ) ^

SOLUTION OF THE BEAM EQUATION

333

3. Let the matrix A of class d+ have the sign pattern ek = 1 (k = 1,2, ...,d),
and let its eigenvalues, X\, k2,..., Am, be arranged in order of decreasing moduli. Then

THEOREM

(1) h > h > ... > h > IVil > IJrfd > - ^ I-U > 0,
(2) the kth eigenvector corresponding to Xk has exactly k\ sign changes
(k = \,2,...,d). The nodes of two successive eigenvectors uk and uk+i (k = l,2,...,d 1)
alternate.
3. The Structure of the Coefficient Matrix
Let us consider the boundary condition 9-9, i.e. condition 9 of Table 1 applying
at x = 0 and condition 9 applying at x = L. We can write these conditions in the form
Clux(0,t)

= dlUJ0,t)
a2u(L,t) = hipu^ULj)
-c2ux(L,t) = d2uJ<L,i),

where it can be assumed that the constants ah b,, ct, and d( (i = 1 , 2 ) are positive.
Using central difference approximations we may replace (8) by
!,*) =

P1Uxx(Xl,t)+-y-U0

Let us definethe quantities


2iclh+2dl)'
2

*>

c2h

aDd

2
e

a2h3
=

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

The eigenvector corresponding to the kth eigenvalue has exactly k\ sign changes and
the nodes of two successive eigenvectors alternate.
THEOREM 2. A totally non-negative matrix A is oscillatory if and only if
(1) d(A) > 0, and
(2) au+i > Oandai+i,t > 0 (i =
1,2,...,m-1).
A generalization of totally non-negative and oscillation matrices has been made
by Gantmakher & Krein (1950). We state now one of their results.
Definition. A matrix A is called a fixed-sign matrix of class d if for any p ^ d, all
the non-zero minors of order p have the same sign ep. If for any p ^ d, all the pth
order minors are different from zero and have the same sign ep, than A is called strictly
fixed-sign. A fixed-sign matrix A of class d is called a matrix of class d+, if a positive
integer k exists such that A* is strictly fixed-sign of class d.

334

ROLAND A. SWEET

Then using equations (5) and conditions (9) we can write the equations for the unknowns uo, ui, un and n +i.
-iwouo =
h-*[(e21+4poel+p1)uo-2(2poel+p1)u1+plu2]

-wjin =
2(Pn+2pn+ie23)un+1']

-2i2poel+Pl)
S=

o
Defining the numbers
c( = V>i.
/ = 0,1,2,..., + 1 ,
it is easy to see that S can be factored as the product
S = AA r
where A is the (+2) x (n+4) matrix
2coe2

A =

0
0
0

-2coe2
0
0

ci

2ci
-ci
0

-c2
2c2
-c2
-c.
2c.

0
-2c+ie3

We observe that A is a tridiagonal matrix J bordered by two columns (one on each


side). If we define the orthogonal matrices
and
C =
then we have
S* = DSD = DAAT) = ( D A Q ^ A 7 ! ) ) = (DACXDACF =
where F is the matrix obtained by replacing J by J* in A.
The matrix J* has the following properties. It is non-negative and

d(J,*) = 2e 2 'fl c, > 0,


1

p = 1,2,...,

d(J*) = d(J? + 2 ) = 0.

Hence, by a lemma of Gantmakher & Krein (1950) J* is totally non-negative.

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

Hence, the matrices of (6) are


H = diag[iwo,H'i,...>M'11,iH'B+i]
and
e21+4p0e22+p1
-2(2poel+p1)
Pi

SOLUTION OF THE BEAM EQUATION

335

Therefore, we have that all minors of F not involving the first row and column
or the last row and column are non-negative. But those minors which do involve the
above-mentioned rows and columns can be reduced to a minor of J* multiplied by
a positive number {a, e*, or eie*). Hence, we have that F is totally non-negative.
By the Cauchy-Binet formula for the determinant of a product of two matrices,
we have that S* is also totally non-negative. By the same formula S* is non-singular
since

fl (~cd2 > 0.
i-0

Hence, by Theorem 2 S* is an oscillation matrix.


We are really interested in the structure of H~iS, but since H is a diagonal matrix
of positive elements it is clear that H = H* and that H -1 S* is oscillatory if S* is.
We have proved:
THEOREM 4. The coefficient matrix of the linear system of equations generated by the
boundary conditions of case 9-9 is similar to an oscillation matrix.
This most general set of boundary conditions was considered first for the simple
reason that all other cases follow from it almost immediately. For choosing appropriate
values of at, bh c,, and d, (i = 1,2) we can obtain any set of the boundary conditions
which we have listed in Table 1.
TABLE 2
Effects of boundary conditions

Condition
at end

Col. 1/
Col. H + 4

Col. 2/
Col. n+3

1
2
3
4
5
6
7
8
9

X
X
X
X

X
X

Rowl/
ROW/J+1

X
X

X
X

Excess of
cols over rows
1
0
-1
0
2
0
1
0
2

In Table 2 we have tabulated the effects on the matrix A of the various boundary
conditions considered. An "X" placed in the table means that the particular row
or column listed at the top of the column is removed from A. The meaning of the last
column of the table will be explained later.
In each case A always has the form of a tridiagonal matrix, whose star-matrix is
totally non-negative, possibly bordered by a column or row. The only non-zero element
in this column or row is always on the main diagonal, and hence, in the matrix F
it appears as a positive element. For this reason F is always totally non-negative.
22

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

2 ... n+2\ Fr / l 2 ... n+2\


... n + 2) {l 2 ...
l)

336

ROLAND A. SWEET

2c2
F =

2c-1
c
c-i
2c
0
c
which is of order (n + l)x. Therefore, the rank of S* must be no greater than n.
Since the first n rows and columns of S is just the matrix obtained in case 2-2, we
have by the above work that
(11)
2 ::: ) > 0 Hence, the rank of S* is precisely n. Furthermore, the last n rows and columns of S
is the matrix which arises in case 1-3 which we have shown to be non-singular. So

> 0.

Conditions (11) and (12) and the nature of the elements of S* allow us to prove the
following theorem.
THEOREM

5. Let A = (oy)7 be a totally non-negative matrix of order m and rank

m-l.If
(1) a/,+i > 0 andai+i,t > 0 (i = l,2,...,m-l), and
2 ... m - l \ .(2 3 ... m\

2 ... m-l)A{2
+

then A is of class (m l) .

3 ...

mj>0'

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

Therefore, to check the oscillation of the matrix S* it is sufficient to check that in


each case S* is non-singular.
There are cases when we remove from A a sufficient number of columns to make its
rank smaller than the order of S. In these cases S must be singular since it is the product
of A and A r . Hence, it cannot be oscillatory. However, practically all of the structure
of the eigenvalues and eigenvectors remains.
The last column of Table 2 expresses the excess of the number of columns over the
number of rows (for one end condition). In considering case / - / we add the two
numbers in rows / and / of the last column, and if the number is negative then the
coefficient matrix must be singular.
Immediately we see that cases 2-3, 3-4, 3-6, and 3-8 give singular matrices. So
does the case 3-3, but the treatment of this case will be taken up later. Singular matrices
arise from the cases 4-4, 4-8, 6-8, and 8-8 also. This is due to the fact that the determinant of the coefficient matrix contains (10) as a multiplicative factor.
By the same type of argument used to establish the non-singularity of S* in case 9-9,
it can be shown that all other cases are non-singular. Using these results we can show
that the singular matrices listed above retain most of the structure of oscillation
matrices. We use now the generalization of oscillation matrices developed by
Gantmakher & Krein (1950).
To illustrate let us examine case 2-3. Here we have

SOLUTION OF THE BEAM EQUATION

337

Before we prove this theorem let us make a definition.


Definition. The minor
W 2 - JpJ
of the matrix A is said to be a quasi-principal minor if

... ;,

which, using the Cauchy-Binet formula m2 times, is just the expression for the
minor

Ui h

there is at least one positive term.


Gantmakher & Krein (1950) show that it is always possible to choose the sets of
indices (k[ k^)),...,(k^'~2\...,k^'~2'>) in such a way that for one term each minor
in the product is a quasi-principal minor of A. Therefore, it suffices to show that
conditions (1) and (2) will guarantee that all quasi-principal minors of order p < m ~ 1
are positive.
Let us prove this by induction on p. For p = 1, the quasi-principal minors are
just the elements in (1) and the entries on the main diagonal. From Gantmakher &
Krein (1950) we have that a totally non-negative matrix A satisfies
/I 2 ... m\
fl 2 ... p\ Jp+1 p+2 ... m>
A
Vl 2 ... m)^A\l
2 ... p)A\p+l
p+2
for any p < m. Using this fact and condition (2), we have
0 < A

<; I::: ::

and
. (2
(2 3 ... m\
m\ .
II < a 22 a 33
0 < A(( 2 3
from which we conclude that all diagonal entries are positive. Hence, the statement
is proved for p = 1.
Assume that the statement is true for all quasi-principal minors of order k < p.
Further, assume that the quasi-principal minor

By the induction hypothesis


k2k3 ... kj

\kik2 ...

kp.J

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

IW*I < 1
(k= 1,2,...,/>).
Proof of Theorem 5. To show that A is of class (m1)+, it sufBces to show that every
minor of A1""1 of order p ^ m1 is positive. For this it is sufficient to show that in
the sum

338

ROLAND A. SWEET

Let B be the matrix composed of rows /i,ii + l,...,i p and columns ki,ki + l,...,kp
of A. Then by a lemma proved by Gantmakher & Krein (1950), B must have rank
p l. Let h = max (/i,A:i). Then
h,ki < h,h+p-l

< ip,kp.

Therefore, the minor


h+l
h+l

... h+p-l\_
(h h + l
... h+p-l) ~ \ h h + l

... h + p-l\
... h + p-lj

is of order p, and so must be zero.


We have h > l . I f A = l.then
h+p-l

< l+m-l-l

= m-l.

Using (13) we get


(I
[l

2 ... m-l\
(h h + l ... h + p-l\
2 ... m-l)^A\h
h + l ... h + p-l)

(h+p ...
\h + p ...

m-l\_
m-l}~"'

which contradicts assumption (2). If h > 1, then


(2 3 ... m\
(2 ... h-l\ (h h + l ... h+p-l\ (h + P ... m\_
\2 3 ... m)**A\2 ... h-l)X{h h + l ... h + p-l)A\h + p ... m)-"'

which also contradicts assumption (2).


Hence (14) is not true. In other words, all quasi-principal minors of order p are
positive as was to be shown.
The matrix S* of order n +1 and rank n arising from case 2-3 satisfies condition (1)
and (2) of Theorem 1, so it is of class n+. Then S* satisfies Theorem 3.
In the same manner it can be shown that all other singular matrices belong to class
r+, where r is one less than the order of the matrix, except the matrix of case 3-3
which belongs to class (r1)+. Again, it should be noted that S* in the class r+
implies (H -1 S)* is also in the class r+ since H is a diagonal matrix with positive
diagonal elements.

4. Summary
For each possible set of boundary conditions we have investigated the nature of
the solutions of the linear system of order m
SV = (;j^)HV,

(15)

where S and H have been defined by (6). Let p be the number of zero eigenvalues and
order the eigenvalues as
0 = Ai = h = ... = kp < ;. p + i < ... < lm.
Let V t be the eigenvector associated with Xk.
Table 3 summarizes the relevant properties of the eigenvalues for each set of
boundary conditions. The number of sign changes in an eigenvector V t corresponding
to a positive eigenvector is equal to kp l.

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

(h
\h

SOLUTION OF THE BEAM EQUATION

339

TABLE 3

Properties of Matrix S

Cases

Number of
Number of
positive distinct
zero eigenvalues
eigenvalues

1-3, -4, -5, -6, -8, -9;


2-4, -5, -6. -8, -9;
3-7; 4-7; 5-7; 6-7;
7-8, -9

n+1

2-3

n+1

3-5,-9:4-5,-6,-9;-)
5-5, -6, -8, -9;
}
6-6, 6-9; 8-9; 9-9 J

n+2

n+2

n+2

n+1

n+2

1-1

n+1

3-4, - 6 , -8 ;)

4-4,-8;
6-8; 8-8

[
J

3-3

The author would like to thank Professor John S. Maybee for suggesting this
problem and to the referees who offered many helpful criticisms.

REFERENCES
BEREZIN, I. S. & ZHTDKOV, N. P.
GANTMAKHER, F. R. & KRETN,

1965 Computing Methods, IL Oxford: Pergamon Press.


M. G. 1950 Oscillation Matrices and Kernels and Small
Vibrations of Mechanical Systems. Moscow: State Publishing House for Technical
Theoretical Literature. (An English translation available from Office of Tech. Serv.,
Dept. of Commerce.)
GENIN, J. & MAYBEE, J. S. 1966 / . Inst. Maths Applies, 2 (4), 343-357.
HOUSNER, G. W. & KOGHTLEY, W. O. 1963 Trans. Am. Soc. Civ. Engrs 128 (1), 1020-1054.

Downloaded from imamat.oxfordjournals.org at University of South Carolina - Columbia on November 24, 2010

1-1,-2,-7;-)
2-2,-7;
\

Order
of S

Das könnte Ihnen auch gefallen