Computational Mechanics 24 (1999) 197±205 Ó Springer-Verlag 1999
A Differential Quadrature as a numerical method to solve
differential equations
T. Y. Wu, G. R. Liu
Abstract A Differential Quadrature proposed here can be
used to solve boundary-value and initial-value differential
equations with a linear or nonlinear nature. Unlike the
classic Differential Quadrature Method (DQM), the newly
proposed Differential Quadrature chooses the function
values and some derivatives wherever necessary as inde-
pendent variables. Therefore, the d-type grid arrangement
used in the classic DQM is exempt while applying the
boundary conditions exactly. Most importantly, the ex-
plicit weighting coef®cients can be obtained using the
proposed procedures. The present method is used to solve
two types of differential equations which are the single-
span Bernoulli±Euler beam's buckling equation and the
one-degree-of-freedom solid dynamic equation. Excellent
results were obtained.
1 junction with the quadrature analog of the boundary con-
Introduction
The Differential Quadrature Method (DQM) was proposed
by Bellman and Casti (1971) and has been employed re-
cently in the solution of solid mechanics problems by Bert
and Malik (1996a, b), Chen et al. (1997), Jang et al. (1989a,
b), Kang et al. (1995), Kukreti et al. (1992) and Striz et al.
(1988). Details on the development of the DQM and its
application to structural mechanics problems may be
found in an excellent review paper by Bert and Malik
(1996a). A d-point technique (Bert and Malik 1996a, b) has
been employed in the DQM's application to boundary-
value differential equations with multiple conditions. But
the initial-value differential quadrature method for struc-
tural dynamics has not been reported until now to the best
of the authors' knowledge.
1.1
The classic Differential Quadrature Method (DQM)
Consider a one-dimensional ®eld variable w…x† prescribed
in a ®eld domain a ˆ x1  x  xN ˆ b. Let wi ˆ w…xi † be
the function values speci®ed in a ®nite set of N discrete
points xi …i ˆ 1; 2; . . . ; N† of the ®eld domain in which the
end points x1 and xN are included. Next, consider the value
of the function derivative dr w=dxr at some discrete points
Received: 4 January 1999
T.Y. Wu (&), G.R. Liu
Department of Mechanical and Production Engineering,
National University of Singapore,
10 Kent Ridge Crescent, Singapore, 119260
197
xi , and let it be expressed as a linearly weighted sum of the
function values.
dr w…xi † X N
…r†
w…r† …xi † ˆ r
ˆ Aij wj …i ˆ 1; 2; . . . ; N†
dx jˆ1
…1†
where
…r†
Aij are the weighting coef®cients of the rth-order
derivative of the function w…x† associated with points xi .
Equation (1), the quadrature rule for a derivative, is the
essential basis of the Differential Quadrature Method. Thus
using Eq. (1) for various order derivatives, one may write a
given differential equation at each point of its solution
domain and obtain the quadrature analog of the differential
equation as a set of algebraic equations in terms of the N
function values. These equations may be solved, in con-
ditions, to obtain the unknown function values provided
that the weighting coef®cients are known a priori.
The weighting coef®cients may be determined by some
appropriate functional approximations; and the approxi-
mate functions are referred to as test functions. The pri-
mary requirements for the choices of the test functions are
of completeness in the same sense as one needs for the
interpolation functions in the ®nite element analysis
(Huebner 1975). Although there can be many choices of
the test functions, a convenient and most commonly used
choice in one-dimensional problems is the Lagrangian
interpolation shape functions lj …x†, where
X
N
w…x† ˆ lj …x†wj …2†
jˆ1
lj …x† are the monomials of the (N ÿ 1†th-order polyno-
mials. Note that the number of test functions is equal to
the number of the sampling points and, for completeness,
the number of sampling points should at least be equal to
one plus the order of the highest derivatives.
Substituting lj …x† of Eq. (2) into Eq. (1), it may be seen
that the weighting coef®cients can be easily obtained. The
detailed procedures can be found in references (Shu and
Richards 1992, Quan and Chang 1989).
1.2
The polynomial-test-function-based weighting coefficients
The accuracy of differential quadrature solution depends
on the accuracy of the weighting coef®cients. To obtain
accurate weighting coef®cients, Quan and Chang (1989)
derived explicit formulae of the Lagrangian-interpolation-
 function-based weighting coef®cients for the ®rst- and iÿ1
second-order derivatives. Shu and Richards (1992) gave a xi ˆ …i ˆ 1; 2; . . . ; N† …6†
N ÿ1
general recurrence relationship for any high-order deriv-
But the Differential Quadrature solutions usually deliver
atives. These formulae were obtained by considering the
more accurate results with unequally spaced sampling
test function in the Lagrangian interpolation process as in
points. A rational basis for the sampling points is provided
Eqs. (1) and (2). These explicit formulae's merit is that
by the zeros of the orthogonal polynomials. A well ac-
highly accurate weighting coef®cients may be determined
cepted kind of sampling points in the DQM is the so-called
for any number of arbitrarily spaced sampling points.
Gauss±Lobatto±Chebyshev points:
But both the explicit formulae and the recursive algo-
rithm mentioned earlier are not new. Villadsen and 1 ÿ cos‰…i ÿ 1†p=…N ÿ 1†Š
Michelsen (1978) and Quan and Chang (1989) have shown xi ˆ …i ˆ 1; 2; . . . ; N†
198 2
that the weighting coef®cients of rth-order derivatives of …7†
the Lagrangian interpolation test functions are
These normal serial Gauss±Lobatto±Chebyshev sampling
…r† dr
Aij ˆ lj …xi † …i; j ˆ 1; 2; . . . ; N† …3† points ± Eq. (7) are used here in the beam buckling
dxr analysis.
where:
Y
N 1.4
/…x†
lj …x† ˆ ; /…x† ˆ …x ÿ xm †; Inverse node numbering
…x ÿ xj †/…1† …xj † mˆ1 In the time coordinate, one usually numbers the discrete
YN time points from the beginning accordingly. For notation's
…1† d/…xj † convenience in the solid dynamics problems, these authors
/ …xj † ˆ ˆ …xj ÿ xm †
dx mˆ1;m6ˆj use the inverse node numbering as in space-rocket
launching. That is: one can use the initial time point as the
and xi 's are the locations of the grid points. N is the N point and the time domain end point as the ®rst point.
number of sampling points. Note that Eq. (3) is valid as These authors' numbering in dynamics problems is just
long as linearly independent polynomials are used as the contrary to the normal time-coordinate direction num-
trial functions and, thus, the values of the coef®cients are bering. The suggested numbering method is very conve-
affected only by the distribution of the grid points. Also nient for notation and programming. Please notice that in
the linearly independent polynomials should be complete. the Lagrangian interpolation process, the discrete point's
Note that the Lagrangian interpolation shape functions numbering can be arbitrary. Thus in the normalized time
lj …x† have the following properties coordinate s, s  [0,1], the inverse node numbering

1 if i ˆ j equations which correspond to Eqs. (6) and (7) in the
lj …xi † ˆ dij ˆ …4† normal serial sampling points are
0 if i 6ˆ j
N ÿi
Using Eqs. (1), (2), and (3) based on Lagrangian inter- si ˆ …i ˆ 1; 2; . . . ; N† …8†
polation shape functions, Quan and Chang (1989) and Shu N ÿ1
and Richards (1992) obtained the following weighting 1 ÿ cos‰…N ÿ i†p=…N ÿ 1†Š
si ˆ …i ˆ 1; 2; . . . ; N†
coef®cients: 2
dlj …xi † /…1† …xi † …9†
…1†
Aij ˆ ˆ
dx …xi ÿ xj †/…1† …xj † This Gauss±Lobatto±Chebyshev inverse node numbering ±
Eq. (9) is used here in the dynamics problems.
…i; j ˆ 1; 2; . . . N; i 6ˆ j†
…rÿ1† ! 1.5
…r† dr lj …xi † …rÿ1† …1† Aij
Aij ˆ ˆ r Aii Aij ÿ Recent developments of the DQM
dxr …xi ÿ xj † Usually, the fourth-order differential equations in struc-
tural mechanics such as beam and plate's displacement,
…i; j ˆ 1; 2; . . . N; i 6ˆ j; r  2† buckling and free-vibration analysis have two boundary
…r† dr li …xi † equations at each boundary. Two conditions at the same
Aii ˆ point provoke a big and real challenge for the classic
dxr
Differential Quadrature Method, because in the classic
X N
…r† DQM we have only one quadrature equation at one point
ˆÿ Aij …i ˆ 1; 2; . . . ; N; r  1† …5†
but two boundary equations are to be implemented.
jˆ1;i6ˆj
Therefore, Bert and Malik (1996a, b), Jang et al. (1989a, b),
Kang et al. (1995), Kukreti et al. (1992) and Striz et al.
1.3 (1988) proposed the d-type grid arrangements, that is,
Normal serial sampling points besides the two boundary points, two additional adjacent
A convenient and natural choice for the sampling points is points with an order of 10ÿ5 (on a normalized spatial
that of the equally spaced sampling points. These are given variable) distance to the boundary points were also treated
in the normalized coordinate [0,1] by as boundary points. Therefore, there are two boundary
points at each boundary corresponding to their two re- The ®eld function's interpolation expression is
spective boundary conditions. constructed in the proposed Differential Quadrature
In solid dynamics problems, one has two initial con- just as in the numerical analysis and the ®nite element
ditions at the initial time, that is the initial displacement method. These authors use Hermite interpolation shape
and initial velocity. The same problems (two conditions at functions.
the same point) were also encountered. Therefore no one
N 
X
paper has appeared about solid dynamics problems solved …0† …1†
by the classic DQM. w…x† ˆ hj0 …x†wj ‡ hj1 …x†wj ‡



Although the d-type grid arrangements work well for jˆ1

some circumstances, this man-made boundary grid ar- …nj ÿ1†
‡hj…nj ÿ1† …x†wj
rangement is not mathematically sound and will some-
199
times cause ill-conditioned problems (Bert and Malik ˆ fh10 …x†; h11 …x†; . . . ; h1…n1 ÿ1† …x†; . . . ;
1996a). No matter how small the d distance is, it's still a T
domain point. A mathematically reasonable d distance is hN0 …x†; hN1 …x†; . . . ; hN…nN ÿ1† …x†
zero. This is the essence in these authors' proposed Dif- …0† …1† …n1 ÿ1† …0†
ferential Quadrature discussed below.  fw1 ; w1 ; . . . ; w1 ; . . . ; wN ;
Recently, Chen et al. (1997) and Wang and Gu (1997) …1†
wN ; . . . ; wN N
…n ÿ1†
g
presented a new idea about treating boundary conditions
in the DQM. These improved approaches eliminate the X
M

de®ciencies of the d-type grid arrangements by applying ˆ hk …x†Uk …10†

kˆ1
the boundary conditions exactly. But Chen et al. (1997)
used the direct linear solver for the determination of where
weighting coef®cients. Wang and Gu (1997) only gave 3-,
4- and 5-point beam elements' explicit weighting coef®- X
N
cients. This paper will propose a Differential Quadrature to Mˆ nj ;
apply the multiple boundary or initial conditions exactly jˆ1
without using the d-point technique. The explicit weight- fUk g ˆ fU1 ; U2 ; . . . ; UM g
ing coef®cients are also presented through two types of
…0† …1†
differential equations. ˆ fw1 ; w1 ; . . . ; :
…n1 ÿ1† …0† …1† …n ÿ1†
2 w1 ; . . . ; wN ; wN . . . ; wN N g;
The newly proposed Differential Quadrature T
fhk g ˆ fh1 ; h2 ; . . . ; hM g
The newly proposed Differential Quadrature considers a 
ˆ h10 …x†; h11 …x†; . . . ; h1…n1 ÿ1† …x†; . . . ; hN0 …x†;
general situation. A one-dimensional ®eld variable w…x† is T
prescribed by a differential equation in a ®eld domain hN1 …x†; . . . ; hN…nN ÿ1† …x†
a ˆ x1  x  xN ˆ b and may also be constrained by a set
of given conditions at any points. The solution domain is The hj0 …x†; . . . ; hj…nj ÿ1† …x† …j ˆ 1; 2; . . . ; N† are Hermite
divided by points xi …i ˆ 1; 2; . . . ; N† that include all the interpolation shape functions. Their properties are listed
…k†
points with given conditions. Let wi ˆ w…k† …xi † in Table 1 when x ˆ xj , and its number's shape in Table 1
…k ˆ 0; 1; 2; . . .† be its kth-order derivatives. Of course, is identical to the identity matrix.
…0†
wi ˆ wi are the function values. Let ni denote the number When the discrete points are not at point xj , the values
of equations corresponding to the point xi . The biggest k of hj0 …x†; . . . ; hj…nj ÿ1† …x† and their derivatives of any pos-
corresponding to point xi is the number of equations sible order corresponding to that point are all zero. Notice
(which that point has) minus one (ni ÿ 1). These authors that the highest order of derivative corresponding to that
…k†
call wi ˆ w…k† …xi † …k ˆ 0; 1; 2; . . . ; ni ÿ 1) the indepen- point is that point's total number of independent variables
dent variables, which that point xi has. The independent minus one.
variables are chosen to be the function value and its de- Polynomial functions are often used in the FEM, but not
rivatives of possible lowest order wherever necessary. For restricted to, so are the test functions in the proposed
the examples discussed later, in the beam buckling anal- Differential Quadrature. In this paper, polynomial func-
ysis, n1 ˆ nN ˆ 2, n2 ˆ n3 ˆ


ˆ nNÿ1 ˆ 1. In the dy- tions are given as examples.
namics problem, nN ˆ 2, n1 ˆ n2 ˆ


ˆ nNÿ1 ˆ 1. From Eq. (10) one has

Table 1. The properties of …i† …i† …i†

Hermite interpolation shape hj0 …xj † hj1 …xj † ...... hj…nj ÿ1† …xj †
functions when x = xj
hj0 …xj † 1 hj1 …xj † 0 ...... hj…nj ÿ1† …xj † 0
…1† …1† …2†
hj0 …xj † 0 hj1 …xj † 1 ...... hj…nj ÿ1† …xj † 0
...... 0 ...... 0 ...... ...... 0
…n ÿ1† …n ÿ1† …n ÿ1†
hj0 j …xj † 0 hj1 j …xj † 0 ...... hj…nj j ÿ1† …xj † 1
 N  weighting coef®cient of the DQM is a matrix of N  N. But
dr w…xi † X …r† …0† …r† …1†
ˆ hj0 …xi †wj ‡ hj1 …xi †wj the resulting weighting coef®cient of the proposed Dif-
dxr jˆ1 ferential Quadrature is a matrix of N  M.
 To illustrate the generality and accuracy of the pro-
…r† …nj ÿ1†
‡


‡ hj…nj ÿ1† …xi †wj posed Differential Quadrature using the explicit weighting
n coef®cients, two examples are to be cited. One example is
…r† …r† …r†
ˆ h10 …xi †; h11 …xi †; . . . ; h1…n1 ÿ1† …xi †; . . . ; single-span Bernoulli±Euler beam's buckling analysis, and
oT the other example is one-degree-of-freedom dynamic
…r† …r† …r†
hN0 …xi †; hN1 …xi †; . . . ; hN…nN ÿ1† …xi † analysis.
n
…0† …1† …n ÿ1† …0†
 w1 ; w1 ; . . . ; w1 1 ; . . . ; wN ; 3
200
o Single-span Bernoulli±Euler beam's buckling analysis
…1† …n ÿ1†
wN ; . . . ; wN N
X
M 3.1
…11† Differential Quadrature expression
…r†
ˆ Eik Uk …i ˆ 1; 2; . . . ; N†
kˆ1 in this case
…r†
The governing equation of single-span Bernoulli±Euler
where Eik are called the weighting coef®cients of the rth- beam's buckling problem is
order derivative of the function at point xi .
Equation (11) is the expression of the newly proposed d4 w P d2 w
Differential Quadrature. Comparing the Eqs. (10) and (11), dx4 ÿ EI dx2 ˆ 0 x  ‰0; LŠ ˆ ‰0; 1Š …13†
one obtains
where w is the transverse displacement function in the y-
…r† …r†
Eik ˆ hk …xi † …12† direction, E and I denote the modulus of elasticity and
principal moment of inertia about the z-axis, respectively.
The most important viewpoint in the proposed Differential P is the compressive axial load.
Quadrature is that the total number of equations at a point The single-span Bernoulli±Euler beam has four boun-
is equal to the total number of the independent variables at dary conditions, two at each end. The beam is divided into
that point, and that independent variables are always the N ÿ 1 sections according to Eq. (7). The boundary con-
function value and its derivatives of possible lowest order ditions are usually the following forms in the buckling
wherever necessary. In most cases, most nj equal one, analysis
therefore only their function values are independent
variables. If all nj are one, the case is the classic Differential wi ˆ 0; w…1† i ˆ 0;
…2†
EIwi ˆ 0;
Quadrature Method. Therefore the classic DQM is a spe- …3†
ci®c case of the proposed Differential Quadrature. EIwi ˆ 0 …i ˆ 1 or N† …14†
This work also gives the explicit weighting coef®cients of
two types of differential equations through examples. The In this example, one has two boundary points x1 and xN .
polynomial test functions with reference to Lagrangian in- At point x1 one has two boundary …1†
conditions and thus two
terpolation shape functions are used to obtain the explicit independent variables w1 and w1 . At point xN one has
weighting coef®cients. Other problems' explicit weighting also two boundary conditions …1†
and then two independent
coef®cients can be obtained in a similar procedure. variables w N and w N . Therefore, PnN1 ˆ nN ˆ 2,
For points with more than one equation such as single- n 2 ˆ n 3 ˆ


ˆ n Nÿ1 ˆ 1, M ˆ jˆ1 nj ˆ N ‡ 2. Now
span beam's end points, the more than one independent Eq. (11) adopts the following expression:
variable is introduced to implement the same number of
X N
equations. Then the de®ciencies of the d-type grid ar- w…r† …xi † ˆ
…r† …r† …1†
hj0 …xi †wj ‡ h11 …xi †w1 ‡ hN1 …xi †wN
…r† …1†
rangements are eliminated, and the boundary conditions jˆ1
are applied directly.
The proposed Differential Quadrature can be extended X
N ‡2
…r†
to multi-dimensional problems in a similar way as the ˆ Eij Uj …15†
DQM. But it must be borne in mind that the total number jˆ1
of independent variables at a point must be equal to the
total number of equations from that point. The two-di- where
mensional problems have been discussed in another paper. …r† …r† …r† …r†
In conclusion, the DQM has only the function values as fEij g ˆ fh10 …xi †; h11 …xi †; h20 …xi †; . . . ;
the independent variables. Therefore at one point only one …r† …r†
hN0 …xi †; hN1 …xi †g
differential quadrature analog can be implemented. But in
the newly proposed Differential Quadrature one has the …r† …r† …r† …r†
ˆ fh1 …xi †; h2 …xi †; . . . ; hN‡1 …xi †; hN‡2 …xi †g
function value and its derivatives wherever necessary as 
…1† …1†
the independent variables. Thus at one point, more than Uj ˆ fw1 ; w1 ; w2 ; . . . ; wN ; wN g
one differential quadrature analog can be implemented in
the proposed Differential Quadrature. The resulting ˆ fU1 ; U2 ; . . . ; UN‡2 g
3.2 where
Interpolation shape function 1 ÿ…x1 ‡ xN † x1 xN
According to the above-de®ned properties of the inter- a11 ˆ ; b11 ˆ ; c11 ˆ
polation shape functions, h10 …x† should have the following x1 ÿ xN x1 ÿ xN x1 ÿ xN
properties
The other interpolation shape functions can be obtained in
…1† …1†
h …x † ˆ 1; h …x † ˆ 0; h …x † ˆ 0; an identical manner. Their properties are as follows:
10 1 10 1 10 N
…1† …1†
h10 …xj † ˆ 0; …j ˆ 2; 3; . . . ; N† …16† hj0 …xj † ˆ 1; hj0 …x1 † ˆ 0; hj0 …xN † ˆ 0; hj0 …xi † ˆ 0
The following h10 …x† satis®es the fourth equation in …i ˆ 1; 2; . . . ; N; j ˆ 2; 3; . . . ; N ÿ 1; i 6ˆ j† …23†
Eq. (16) 201
…1† …1†
2
hN0 …xN † ˆ 1; hN0 …xN † ˆ 0; hN0 …x1 † ˆ 0;
h10 …x† ˆ …a1 x ‡ b1 x ‡ c1 †l1 …x† …17†
hN0 …xi † ˆ 0 …i ˆ 1; 2; . . . ; N ÿ 1† …24†
Using the ®rst three equations in Eq. (16), the unknown …1† …1†
constants a1 , b1 and c1 in Eq. (17) can be obtained. Notice hN1 …xN † ˆ 1; hN1 …x1 † ˆ 0;
l1 …xN † ˆ 0, l1 …x1 † ˆ 1 from Eq. (4), then one has hN1 …xj † ˆ 0; …j ˆ 1; 2; . . . ; N† …25†
8 2 2
< …a1 x1 ‡ b1 x1 ‡ c1 †l1 …x1 † ˆ a1 x1 ‡ b1 x1 ‡ c1 ˆ 1
> Using the identical way as mentioned above, one obtains,
…1†
…2a1 x1 ‡ b1 †l1 …x1 † ‡ …a1 x21 ‡ b1 x1 ‡ c1 †l1 …x1 † ˆ 0 respectively:
>
: …1† ÿ

…2a1 xN ‡ b1 †l1 …xN † ‡ …a1 x2N ‡ b1 xN ‡ c1 †l1 …xN † ˆ 0 hj0 …x† ˆ aj x2 ‡ bj x ‡ cj lj …x† …j ˆ 2; 3; . . . ; N ÿ 1†
…18† …26†
…1†
Notice that l1 …xN †
can be obtained from Eq. (5) and isn't 1
always zero. The third of Eq. (18) is satis®ed only if: aj ˆ 2
xj ÿ xj …x1 ‡ xN † ‡ x1 xN
a1 x2N ‡ b1 xN ‡ c1 ˆ 0 ÿ…x1 ‡ xN †
bj ˆ2 ÿ x …x ‡ x † ‡ x x
From this equation and the other two equations in Eq. (18), x j j 1 N 1 N
one obtains x1 xN
cj ˆ 2
ÿ1
…1†
ÿl1 …x1 † xj ÿ xj …x1 ‡ xN † ‡ x1 xN
a1 ˆ ‡ ÿ

…x1 ÿ xN †2 …x1 ÿ xN † hN0 …x† ˆ aN x2 ‡ bN x ‡ cN lN …x† …27†
1 …1†
b1 ˆ ÿ a1 …x1 ‡ xN † ÿ1 lN …xN †
…x1 ÿ xN † a N ˆ ‡
…x1 ÿ xN †2 …x1 ÿ xN †
…1†
2x1 …x1 ‡ xN †l1 …x1 † ÿ1
ˆ 2‡ bN ˆ ÿ aN …x1 ‡ xN †
…x1 ÿ xN † …x1 ÿ xN † …x1 ÿ xN †
c1 ˆ 1 ÿ a1 x21 ÿ b1 x1 2xN
…1†
…x1 ‡ xN †lN …xN †
…1† ˆ ÿ
…xN ÿ 2x1 †xN ÿx1 xN l1 …x1 † …x1 ÿ xN †2 …x1 ÿ xN †
ˆ 2 ‡ …19†
…x1 ÿ xN † … x1 ÿ x N † 2
cN ˆ 1 ÿ aN xN ÿ bN xN
The h11 …x† should have the following properties …1†
…x1 ÿ 2xN †x1 x1 xN lN …xN †
ˆ ‡
…1† …1†
h11 …x1 † ˆ 1; h11 …xN † ˆ 0; …x1 ÿ xN †2 …x1 ÿ xN †
ÿ

h11 …xj † ˆ 0; …j ˆ 1; 2; . . . ; N† …20† hN1 …x† ˆ aN1 x2 ‡ bN1 x ‡ cN1 lN …x† …28†
The following h11 …x† satis®es the third equation in Eq. (20) a ˆ ÿ1 x1 ‡ xN ÿx1 xN
N1 ; bN1 ˆ ; cN1 ˆ
h11 …x† ˆ …d1 x ‡ e1 †…x ÿ x1 †l1 …x† …21† x1 ÿ xN x1 ÿ xN x1 ÿ xN

Using the ®rst two equations in Eq. (20), d1 and e1 can be

3.3
obtained in a similar way as in Eq. (18).
Explicit weighting coefficients
1 ÿxN From Eqs. (12), (15), (17), (22), (26), (27) and (28), one
d1 ˆ ; e1 ˆ
x1 ÿ xN x1 ÿ xN gets
Therefore …r† dr
Eij ˆ hj …xi †
x ÿ xN dxr
h11 …x† ˆ …x ÿ x1 † l1 …x†
x1 ÿ xN …i ˆ 1; 2; . . . ; N; j ˆ 1; 2; . . . ; N ‡ 2; r ˆ 1; 2; 3; 4†
ÿ

ˆ a11 x2 ‡ b11 x ‡ c11 l1 …x† …22† …29†
   
The above shape functions hj …x† …j ˆ 1; 2; . . . ; N ‡ 2† have ‰Sbb Š ‰Sbd Š fUb g
the following form
‰Sdb Š‰Sdd Š fUd g
F…x† ˆ …ax2 ‡ bx ‡ c†lj …x† …j ˆ 1; 2; . . . ; N†   
‰0Š ‰0Š fUb g
ÿk ˆ0 …33†
Thus ‰Qdb Š ‰Qdd Š fUd g
…1† where the subscripts b and d indicate the grid points used
F …1† …x† ˆ …ax2 ‡ bx ‡ c†lj …x† ‡ …2ax ‡ b†lj …x†
for writing the quadrature analog of the boundary condi-
…2†
F …2† …x† ˆ …ax2 ‡ bx ‡ c†lj …x† tions and the governing differential equation, respectively.
…1† n o
‡ 2…2ax ‡ b†lj …x† ‡ 2alj …x† …1† …1†
fUb g ˆ fU1 ; U2 ; UN‡1 ; UN‡2 g ˆ w1 ; w1 ; wN ; wN :
202 …3†
F …3† …x† ˆ …ax2 ‡ bx ‡ c†lj …x† …30† fUd g ˆ fU3 ; U4 ; . . . ; UN g ˆ fw2 ; w3 ; . . . ; wNÿ1 g:
…2† …1†
‡ 3…2ax ‡ b†lj …x† ‡ 6alj …x†
By matrix substructuring of Eq. (33), one has the following
…4†
F …4† …x† ˆ …ax2 ‡ bx ‡ c†lj …x† two equations
…3† …2†
‡ 4…2ax ‡ b†lj …x† ‡ 12alj …x† ‰Sbb ŠfUb g ‡ ‰Sbd ŠfUd g ˆ 0 …34†
In Eq. (30), if x is assigned a different xi …i ˆ 1; 2; . . . ; N†, ‰Sdb ŠfUb g ‡ ‰Sdd ŠfUd g ÿ k …‰Qdb ŠfUb g ‡ ‰Qdd ŠfUd g† ˆ 0
…r†
the weighting coef®cients Eij in Eq. (29) can be explicitly …35†
…1† …2† …3†
obtained. Notice that lj …xi †, lj …xi †, lj …xi †,
…4†
lj …xi †…i; j ˆ 1; 2; . . . ; N† have been obtained in Eq. (5). In fact, Eq. (34) is a combination of boundary condition
Because this governing equation is a fourth-order differ- equations, and it does not contain k. From Eq. (34) one
ential equation, only the 1±4th-order weighting coef®- obtains:
cients are needed. fUb g ˆ ÿ‰Sbb Šÿ1 ‰Sbd ŠfUd g …36†
Now the differential quadrature analog of the governing
Eq. (13) according to Eq. (15) is Back-substituting Eq. (36) into Eq. (35), one gets
ÿ

X
N‡2 X
N‡2 ‰Sdd Š ÿ ‰Sdb Š‰Sbb Šÿ1 ‰Sbd Š fUd g
…4† …2†
Eij Uj ÿ k Eij Uj ˆ 0 ÿ

jˆ1 jˆ1
ÿ k ‰Qdd Š ÿ ‰Qdb Š‰Sbb Šÿ1 ‰Sbd Š fUd g ˆ 0 …37†
…i ˆ 2; 3; . . . ; N ÿ 1† …31† In short notation, one obtains

where k ˆ P=EI. The differential quadrature analogs of the …‰SŠ ÿ k‰QŠ†fUd g ˆ 0 …38†
boundary condition Eq. (14) are This is a generalized eigenvalue equation and can be re-
XN‡2 duced to standard eigenvalue equation. By the procedure
…1† …2† proposed here, one obtains the normalized critical buck-
wi ˆ 0; wi ˆ 0; EI Eij Uj ˆ 0;
jˆ1 ling axial load k of the beam with various boundary con-
ditions. The calculated k is compared with analytic results
X
N ‡2
…3† in Table 2. Good agreements were obtained. When more
EI Eij Uj ˆ 0 …i ˆ 1 or N† …32†
sampling points are employed, Table 2 shows that the
jˆ1
convergence rate is very rapid.
This work will give the normalized critical buckling load k
of the different boundary conditions such as pinned-pin- 4
ned, ®xed-®xed and ®xed-pinned ends in Table 2. For One-degree-of-freedom dynamics
different boundary conditions one has a proper combi-
nation of boundary Eq. (32). By rearranging Eqs. (31) and 4.1
(32), the assembled form, which is similar to the expres- The normalized dynamics equation
sions in reference (Bert and Malik 1996a), is The simplest dynamics equation is the one-degree-of-
freedom vibration equation
Table 2. Comparison of beam normalized critical buckling load
k under various boundary conditions d2 y dy
m 2
‡ 2c ‡ ky ˆ q sin…pt † …39†
dt dt
N Pinned±Pinned Fixed±Fixed Fixed±Pinned
where m is the mass, c the damping value, k the elastic
Analytic 9.869604 39.47842 20.19073 coef®cient. q and p are the exciting force's amplitude and
6 9.867287 40.44472 20.17477 frequency, respectively, and d2 y=dt 2 ,dy=dt and y the ac-
7 9.869683 39.37706 20.18902 celeration, velocity and displacement, respectively. One
8 9.869631 39.48238 20.19110 usually uses the following standard form
9 9.869604 39.47825 20.19075
10 9.869604 39.47845 20.19072 d2 y dy q sin…pt†
11 9.869604 39.47842 20.19073 2
 2y ˆ
‡ 2g ‡ x …40†
dt dt m
where g is the damping ratio and x  is the natural fre- 4.3
quency. Interpolation shape functions
In the differential quadrature method, one usually em- According to the above-de®ned properties of the inter-
ploys the normalized coordinate [0,1]. Here again, these polation shape functions, hj …s†…j ˆ 1; 2; . . . ; N ÿ 1† should
authors also use the normalized coordinate [0,1]. Suppose have the following properties
that s ˆ t=T, T is the time length of solution domain. Using
…1†
this normalized time coordinate s ˆ t=T, one obtains hj …sj † ˆ 1; hj …sN † ˆ 0; hj …si † ˆ 0
d2 y dy 2 T 2 q sin…pTs† …j ˆ 1; 2; . . . ; N ÿ 1; i ˆ 1; 2; . . . ; N; i 6ˆ j† …45†
‡ 2gT ‡ … 
xT † y ˆ …41†
ds2 ds m
hN …s† and hN‡1 …s† should have the following properties,
where g ˆ gT is the normalized damping ratio, x ˆ xT  respectively. 203
the normalized natural frequency, p ˆ pT the normalized
…1†
exciting force's frequency, and q ˆ T 2 q=m the normalized hN …sN † ˆ 1; hN …sN † ˆ 0;
exciting force's amplitude. Therefore …46†
hN …si † ˆ 0 …i ˆ 1; 2; . . . ; N ÿ 1†
d2 y dy …1†
2
‡ 2g ‡ x2 y ˆ q sin…ps† …42† hN‡1 …sN † ˆ 0; hN‡1 …sN † ˆ 1;
ds ds …47†
hN‡1 …si † ˆ 0 …i ˆ 1; 2; . . . ; N ÿ 1†
In this example the units are omitted or taken as inter-
national standard (SI) units. Besides the initial displace- The interpolation shape functions
…1†
ment y0 ˆ 1 and initial velocity y0 ˆ 1, one has the hj …s†…j ˆ 1; 2; . . . ; N ‡ 1† can be obtained in an identical
following data manner as in the aforementioned beam problem. Here
 ˆ 1;
x p ˆ 2; q=m ˆ 1; T ˆ p=2; g ˆ 0:05 their results are given and their properties can easily be
veri®ed. Notice that lj …s†…j ˆ 1; 2; . . . ; N† is de®ned in Eq.
The analytic solution of Eq. (42) is (3).
y ˆ eÿgs …A sin x1 s ‡ B cos x1 s† ‡ C cos ps ‡ D sin ps s ÿ sN
hj …s† ˆ lj …s† …j ˆ 1; 2; . . . ; N ÿ 1† …48†
…43† sj ÿ sN
h i
where …1†
hN …s† ˆ lN …sN †…sN ÿ s† ‡ 1 lN …s† …49†
2 2
ÿ2pqg q…x ÿ p † hN‡1 …s† ˆ …s ÿ sN †lN …s† …50†
Cˆ 2 ; Dˆ ;
…x2 ÿ p2 † ‡4p2 g2 …x2 ÿ p2 †2 ‡4p2 g2
p …1†
Ty0 ‡ gB ÿ Dp
2
B ˆ y0 ÿ C; x1 ˆ x ÿ g ; A ˆ 2 4.4
x1 Explicit weighting coefficients
From Eqs. (12), (44), (48), (49) and (50), one obtains
4.2 …r† dr
Differential Quadrature expression Eij ˆ hj …si † …i ˆ 1; 2; . . . ; N;
in this case dsr
In this case, the time domain t  ‰0; TŠ is normalized to j ˆ 1; 2; . . . ; N ‡ 1; r ˆ 1; 2† …51†
s  ‰0; 1Š and then divided into N ÿ 1 sections according
to Eq. (9). At the initial time point sN , one has two initial The above shape functions hj …s†…j ˆ 1; 2; . . . ; N ‡ 1† are of
conditions ± the initial displacement and initial velocity. the following form
Thus at point sN one has two independent variables yN and F…s† ˆ …as ‡ b†lj …s† …j ˆ 1; 2; . . . ; N†
…1†
yN . Therefore, nN ˆ 2, n1 ˆ n2 ˆ


ˆ nNÿ1 ˆ 1. Now
Eq. (11) adopts the following expression: Thus
X
N
…r† …r† …1†
X
N ‡1
…r† …1†
y…r† …si † ˆ hj0 …si †yj ‡ hN1 …si †yN ˆ Eij Uj F …1† …s† ˆ …as ‡ b†lj …s† ‡ alj …s†
jˆ1 jˆ1 …52†
…2† …1†
F …2† …s† ˆ …as ‡ b†lj …s† ‡ 2alj …s†
…44†
where In Eq. (52), if s is assigned a different si …i ˆ 1; 2; . . . ; N†,
…r†
…r† …r† …r† …r† …r† the weighting coef®cients Eij in Eq. (51) can be explicitly
fEij g ˆ fh10 …si †; h20 …si †; . . . ; hN0 …si †; hN1 …si †g …1† …2†
obtained. Notice that lj …si †, lj …si † …i; j ˆ 1; 2; . . . ; N†
…r† …r† …r† have been obtained in Eq. (5). Because this governing
ˆ fh1 …si †; h2 …si †; . . . ; hN‡1 …si †g equation is a second-order differential equation, only the
…1† ®rst- and second-order weighting coef®cients are needed.
fUj g ˆ fy1 ; y2 ; . . . ; yN ; yN g
Now according to Eq. (44), the differential quadrature
ˆ fU1 ; U2 ; . . . ; UN ; UN‡1 g: analog of the governing Eq. (42) is
 Table 3. Comparison of displacements between the calculated Table 4. Comparison of displacements between the calculated
results and analytic results in dynamics problems to check con- results and analytic results in dynamics problems to check sta-
vergence for the ®rst step bility

N si Analytic Relative error (%) Time step si Analytic Relative error (%)

1 1.59965496 )0.720 1 0.042823156 0.157e-2

2 1.64322220 )0.694 2 )0.00811146 )0.804e-2
5 3 1.51950632 )0.681 3 )0.14379374 )0.257e-3
4 1.19930126 )0.328 58 4 )0.28119012 )0.112e-3
5 )0.29884991 0.758e-5
1 1.61144389 0.116e-1
6 )0.18738275 0.219e-3
2 1.64051469 0.114e-1
204 7 )0.06143679 0.769e-3
3 1.64587199 0.111e-1
7 4 1.53005910 0.885e-2 1 )0.0221246 0.306e-2
5 1.31944193 0.620e-2 2 )0.02955779 )0.219e-2
6 1.09938225 0.262e-2 3 0.166658201 )0.223e-3
223 4 0.304214792 )0.882e-4
1 1.61125535 )0.995e-4
5 0.319368312 0.291e-4
2 1.62979823 )0.984e-4
6 0.203391381 0.243e-3
3 1.65470856 )0.965e-4
7 0.073272567 0.766e-3
9 4 1.62896462 )0.885e-4
5 1.52992250 )0.795e-4 1 0.02212457 0.306e-2
6 1.37690546 )0.577e-4 2 )0.02955780 )0.219e-2
7 1.20325105 )0.340e-4 3 )0.1666582 )0.223e-3
8 1.05785891 )0.135e-4 388 4 )0.30421477 )0.882e-4
5 )0.31936827 0.291e-4
6 )0.20339133 0.243e-4
7 )0.07327251 0.766e-3
X
N‡1
…2†
X
N‡1
…1†
Eij Uj ‡ 2g Eij Uj ‡ x2 yi ˆ q sin…psi †
jˆ1 jˆ1
the analytic results are compared in Tables 3 and 4. From
…i ˆ 1; 2; . . . ; N ÿ 1† …53†
Table 3 one can see that the convergence is very fast,
The initial conditions' quadrature analogs in the normal- when the second-order time division N is increased.
ized coordinate are From Table 4, it's shown that the stability of this method
is also very good. Note that double precission Fortran is
…1† dy …1† employed in this work.
yN ˆ y…s ˆ 0† ˆ y0 ; yN ˆ …s ˆ 0† ˆ Ty0 …54†
ds 5
Matrix expression of Eq. (53) is Conclusion
  A new concept of the Differential Quadrature is pro-
fyd g
‰‰Sd Š ‰Sb ŠŠ ˆ fqd g …55† posed here. The proposed method can be extended to
fyb g
solve boundary-value and initial-value differential
where equations with a linear or nonlinear nature. Besides
n o n o applying the boundary conditions exactly, the d-type
…1† …1†
fyb g ˆ yN ; yN ˆ y0 ; Ty0 ; grid arrangement used in the classic DQM is exempt in
the newly proposed method. Any ®nite boundary dif-
fyd g ˆ fy1 ; y2 ; . . . ; yNÿ1 g ˆ fU1 ; U2 ; . . . ; UNÿ1 g ferential equation with ®nite function values, their de-
rivatives and their combinations within its domain can
By matrix substructuring, Eq. (55) is rewritten as
be solved using the proposed method. The proposed
‰Sd Šfyd g ˆ fqd g ÿ ‰Sb Šfyb g …56† Differential Quadrature can be extended to multi-di-
mensional problems in the similar way as the DQM.
Therefore, every point's displacement can be obtained But the following main differences between the classic
from Eq. (56), and thus the velocity and acceleration can DQM and the proposed method must be borne in
be obtained from the Differential Quadrature Eq. (44). The mind.
inverse node numbering's convenience is now seen. One In the DQM one has only the function values as the
need not rearrange the matrix in Eq. (55), and its sub- independent variables. Therefore at one point only one
structuring is straightforward. differential quadrature analog can be implemented. But in
For the dynamics Eq. (42), many time steps are cal- the newly proposed Differential Quadrature one has the
culated to check the stability of the proposed method. function values and their derivatives wherever necessary
The time step is Dt ˆ T. Because one also divides the as the independent variables. The resulting weighting co-
time domain in Dt, one can call them the ®rst-order time ef®cients of the DQM is a matrix of N  N. But the re-
division and the second-order time division, respectively. sulting weighting coef®cients of the proposed Differential
In the second-order division of the time domain Dt, Quadrature is a matrix of N  M. When M ˆ N, the
N ˆ 8 is always used in Table 4. The calculated dis- proposed Differential Quadrature is reduced to the classic
placements by the proposed Differential Quadrature and DQM.
 As examples, the Differential Quadrature expressions
and their explicit weighting coef®cients have been pre-
sented for the fourth-order boundary-value differential
equation and the second-order initial-value differential
equation. Good results were obtained in the two examples
as compared with the analytic results. Most importantly,
the explicit weighting coef®cients of the proposed Differ-
ential Quadrature for higher-order differential equations
with multiple given conditions can be obtained using this
paper's procedures. These highly accurate weighting co-
ef®cients may be determined for any number of arbitrarily
spaced sampling points.
References
Bellman R, Casti J (1971) Differential quadrature and long term
integration. J. Math. Anal. Appl. 34:235±238
Bert CW, Malik M (1996a) Differential quadrature method in
computational mechanics: a review. Appl. Mech. Rev. 49:1±27
Bert CW, Malik M (1996b) Free vibration analysis of thin cylin-
drical shells by the differential quadrature method. J. Pres- Villadsen JV, Michelsen ML (1978) Solution of differential
sure Vessel Technol. 118:1±12
Chen W, Striz AG, Bert CW (1997) A new approach to the dif-
ferential quadrature method for fourth-order equations. Int. Wang X, Gu H (1997) Static analysis of frame structures by the
J. Num. Meth. Eng. 40:1941±1956
Huebner KH (1975) The ®nite element method for engineers.
John Wiley & Sons, New York
Jang SK, Bert CW, Striz AG (1989a) Nonlinear bending analysis
of orthotropic rectangular plates by the method of differential
quadrature. Comput. Mech. 5:217±226
Jang SK, Bert CW, Striz AG (1989b) Application of differential
quadrature to static analysis of structural components. Int. J.
Num. Meth. Eng. 28:561±577
Kang K, Bert CW, Striz AG (1995) Vibration analysis of shear
deformable circular arches by the differential quadrature
method. J. Sound Vib. 183:353±360
Kukreti AG, Farsa J, Bert CW (1992) Fundamental frequency of
tapered plates by differential quadrature. J. Eng. Mech.
118:1221±1238
Quan JR, Chang CT (1989) New insights in solving distributed 205
system equations by the differential quadrature method ± I:
analysis. Comput. Chem. Eng. 13:779±788
Shu C, Richards BE (1992) Application of generalized differential
quadrature to solve two-dimensional incompressible Navier-
Stokes equation. Int. J. Num. Meth. Fluids 15:791±798
Striz AG, Jang SK, Bert CW (1988) Nonlinear bending analysis of
the thin circular plates by differential quadrature. Thin
Walled Struct. 6:51±62
equation models by polynomial approximation. Prentice-Hall
Englewood Cliffs, New Jersey
differential quadrature element method. Int. J. Num. Meth.
Eng. 40:759±788