Computer Implementation of Finite Element Method
Computer Implementation of Finite Element Method
BY
J. ALAN GEORGE
STAN-CS-71-208
FEBRUARY, 1971
Raeproducd by
NATIONAL TECHNICAL
INFORMATION SERVICE
Springfield, Va. 22151
S'JUL14Q
J1
Stanford University
J. Alan George
C N O d9b. OTHER REPORT NOSt (Any other numbers thar may be assigned
this report)
10 DISTRIBUTION STATEMENT
13. ABSSTRACT
FDORMaV (PAGE 1
DNOV Ge 473 Unclassified
S/N 0101,807,6801 Securitv Classification
Security ClaaifLcation
much sound advice and consultation during the last six months of this
work.
only for his academic help but for his friendship, which, in numerous
ways, helped make life at Stanford for my wife, Doris, and we generally
typing.
iii
Table of Contents
page
CHAPTER 1: INTRODUCTION
2.1 Introduction 21
3.1 Introduction 52
iv
1
Page
REFERENCES 154
APPENDIX B: O/S 360 FORTRAN CODE FOR FINITE ELEMENT METHODS 164 -
I
V j
J
CHAPTER 1
INTRODUCTION
elliptic boundary value problems in the plane which are either formulated
majority of the program would remain unchanged for more general p-oblems.
I specific problem. Instead, we will adopt the attitude of one who must
important. Our study will include the problems of mesh generation and
Li
We will also be interested in the performance of finite element
will also compare different finite element methods; that is, finite
element methods using different bases. Our results should offer some
our problems carefully, almost any method can be made to look best. If
we have a specific problem that must be solved many times, then it may
very well be worthwhile to find the best method for that particular
of problems, and therefore unsuitable for the purposes we have set down
(c) Does the method rely on domain shape? (For example, does it
(d) Does -the method utilize a special technique which requires some L
information known only to an expert in the field? If so, can
. 2. L
(e) How generaly applicable is the method? For example, must the
depends upon individual needs and circumstances, but from our viewpoint
the comparisons are often made on the basis of (a) and perhaps (b), with
I: 3
2. The Variational Principle and a Brief Discussion of Ritz Methods
the function urV which satisfies the boundary value problem. This is
imposing the condition that the first variation of T[v] vanish [C2].
L
1~1
i "-
IN
(25) v E 1 x ,x *, , *j.Y d 1d
+2 * dx dy
and
next section.
5
li
Note 'that. for our chosen class of problems, it will always be
L
6 11
U
3. Essential Characteristics of Finite Element Methods
The concept was extended to continuous problems such as plate bending and
a common vertex or common side. The most common element shapes are
triangles and rectangles. Our attention will be devoted almost exclusively
Finite element methods are Ritz methods which use basis functions
having small support; that is, Ritz methods which make use of a so-called
this point we simply observe that finite element methods make use of
(5.1) NN
where
L
i L.
~7
I
iT
Figure 3.1-c
Support of *k
associated with an
interior node.
More than one basis function may be associated with a particular node, L_.
and because of the way the basis functions are chosen, the nodal
choice of these nodal parameters is done on the basis of (1) the number
L
_ L
requirements of v . Indeed, a common practice is not to consider the
element stencil
, or just
V V
II
LJ It is fairly easy to see that this amounts to using a "pyramid
i! 9
_ 11
/k
v, vv
Quadratic C abic
Note that in the last example, three basis functions will be associated
with each corner node, and the function associated with the interior node
10
L
A
I
values (and perhaps derivative values) that they assme at the node
will not (usually) interpolate the solut.on of our boundary value problem.
(i) Because each basis function vanishes over most of the domain, the
by stating that "by a suitable choice of the trial functions ... the
of taste; we prefer the former, and reserve the term "finite difference"
polynomial basis functions of low degree the two approaches sometimes yield
(ii) The value of finite element methods will obviously depend upon
how well the trial functions can approximate the true solution of our
boundary value problem. This problem has been studied for general
L Bramble and ZIlzal [B12], Zl~mal [Z4,Z5], leneck [Zl], Goal [G2],
}l
and others have proved convergence of the method and presented bounds
equivalent decrease in the error bound. That is, the error bounds are
of the form
Nll _ hd~l-q i l~
2 i 2
where lull =
1
l_<
i<
x ID U1
2
i = (ii
1 2 2
=
and
.
D1 U
_ _ u . Here h is the max-um length of any triangle side in
6x 1Z y
the mesh, and C is a constant which depends upon the sharpest angle in
the mesh and the polynomial basis (element) being employed. For specific
obtain will obviously depend upon our choice of *Is . Indeed, one of L
the problems of using the Ritz technique has been the numerical
problems would be much less troublesome for the finite element method
because the majority of the 's will be orthogonal (by virtue of having
disjoint support). Strang and Fix [S6] study this problem in depth for
12
uniform meshes by examining the condition number H(G) = JIG11 JIGh-l1
of the Gram matrix, whose entries are the inner products of the basis
trial fumctions yield a stable basis, where stability means that H(G)
we increase the degree of our polynomials. Note that these results only
Iapply for uniform meshes, and it is not knon how detrimental severe
13
4. Tensor Product Spaces
then given by
n
!'j=O ai
if *Ee[Dl1 ]
fi , i = 0,1,...,n , then *ij ee[D2 ] , ij = 1,...,n .
This is often done by taking a spline basis for the *'s . For
example, we can have vN EO2 (D2) by using the cubic spline basis [SI]. For
reason for these properties is the fact that the interelement boundaries
basis construction.
'14
BL
the union of rectangles. Furthermore, it is virtually impossible to
each coordinate, even though we only desire the fine mesh in the
fine.
problem; Bramble and Schatz [B111 and Babuska [Bl] have analyzed some
methods that. do not require that the basis functions satisfy any boundary
The boundary integral is designed so that its minimum occurs when the
some situations.
I ( 15
5. Review of the Thesis and Summary of Conclusions
and such a study leads to interesting practical probleas which are seldom
finite element spaces. We feel our work lies between these two extremes;
we are concerned with the actual implementation of finite element methods and
how they compare in practice with other methods for solving elliptic boundary
value problems.
for several reasons: (a) the required input for most domains is small,
(b) the method can easily be adapted for use with graphical display
(d) the net can be graded under control of the user, and (e) inter-element
16
* arbitrary simply connected domains and is designed to produce graded
some results obtained in Chapter 4., we compare the storage required for
required for the mesh is small compared to that required for just the
17
IL
and distinguish between graph methods, profile methods and band methods.
profile methods.
counts; (in Sections .1 and 4.2 we show that they will never be worse
reverse Cuthill-McKee ordering with some of the times required for the
true, but not all users are experienced, and more important, one must
L
.1
left to the computer. The code for doing the ordering appears as part
/ pase]1 in-Appendix B.
'of
storage, since we can allocate storage for the matrix as soon as the
-. following:
polynomial. This was true in all three examples, and because the
finite element solutions. Even for the problem where special fast
(c) Several methods for finding eigenvalues yielded more accurate numbers
than finite element methods (involving roughly the same cost), and
19
difficult to implement in a general code. Again we emphasize that
saying that they are less suitable than finite element methods as
segmented the code into modules, each one designed to carry out a specific
passing from one to the next via external storage media which can be disk,
d=n or tape. Our reasons for segmenting our code and attempting to
keep each segment itself modular are (a) to ease maintenance and/or
smaller machines than the one we used, and (c) to facilitate documentation
code itself.
20 .
~L
CHATER 2
1. Introduction
Figure 1.1
21
Section 4 is almost completely automatic. Section 5 contains a
with the problem of numbering the nodes (more precisely, the parameters
larger if each node has more than one parameter associated with it.
Since all the parameters associated with a particular node are connected
each other, little is lost by ordering the nodes. Many good ordering
of these algorithms and the criteria used to reduce storage and computational
17
U
22)
2. Review of Previous Work on Mesh Generation
the domain. Often it is fairly smooth over most of R , and varies rapidly
only over a small part of the domain, perhaps near a corner. For this reason
the grid can be coarse and areas where it should be fine. This grading
computer time.
provide for the grading of nets. This is due largely to the ease with which
array and the severe storage management problems which immediately result
when one departs from such a scheme. In the regular case, actual coordinates
LI
are much better fur regular nets, and the determination of the coefficients
time they are needed). Thus, finite difference programs usually make use of
uniform meshes, or meshes which are uniform in various parts of the region.
Boundary points that result when 6R intersects the mesh at a point other
(These special boundary points may cause storage problems even when the
simple two-dimensional array storage method is used; see Forsythe and Wasow
23
rather awkward to find the correct formula to preserve the order of
problem. The problems arise where the boundary (which can have more or
less arbitrary shape) intersects the regular mesh. Cryer [C3] treats this
here.
without cost provided we are going to store the node coordinates anyway.
It is often stated that irregular nets are expensive to use because the
coordinates of each net point must be stored, and for finite difference
the same), we will need 8n 2 words of storage.] However, for finite element L
methods, the number of node points will ordinarily be considerably fewer than
the number of parameters since each node will usually have derivative as
the basis functions increases, the storage required for the nodes quickly
becomes small compared with that required for the coefficient matrix. This
24
I
We will now review some methods appearing in the literature which have
by one family of straight lines (which do not intersect in the domain but
yield node points. The node points on adjacent lines are then joined in a
net is desired. This restriction can also lead to triangles with very
sharp angles.
of the diagonals are computed and the shorter is used to form the triangles.
of manually plotting the boundary of the domain and the node points (in the
25
L
of the points are automatically punched on cards which then serve as
necessarily a very good criterion alyway, and to number the nodes empirically
hard to decide where the next node woLuld be placed. If the triangles are
As the authors point out, however, the computer-based part of the procedure
eliminates the clerical errors which would inevitably result from tabulation
cathode ray display with a light pen. All the above objections could be
maps the given domain R onto the polygon, and then using the inverse of
the mapping so determined to find the image of the orthogonal grid in the t
26
polygon. The method obviously generates rectangles rather than triangles,
poor triangles. Als% the work involved in computing the mapping may be
substantial.
difference methods. The mesh lines are regarded as two intersecting sets
generated mesh varies smoothly over the entire domain, its relative grading
generated using the method. Although the examples reported are very nice,
they are for an extremely simple domain, and Winslow does not discuss the
(assuming that the program has the facility for handling one), and how to
easily input the boundary conditions (the ends of the potential lines). As
with most partial differential equation problems, the above tasks and the
associated data management problems are difficult to implement in general;
once done, the generation of the equations and their solutions are relatively
t He concedes that the method does not always work satisfactorily near re-
entrant corners, with node points outside the domain sometimes being
[V produced.
27
Reid and Turner [Ri] use the following scheme to generate nearly regular
intersect the boundary are called "boundary points", and node points of the
mesh closer than h/2 to a boundary point are moved to the boundary point
the same order of magnitude as (or smaller than) the shortest arc in 6R
requirement could force the mesh to be finer than otherwise necessary. This
scheme obviously assumes that the user desires a regular mesh, and this
Kamel and Eisenstein [KI] present a mesh generation scheme that is also
of arcs subdivided by nodes. First the authors find the "best" regular
mesh having the same number of boundary nodes as the given boundary 6R.
a node of a regular mesh and successively annexes rings of triangles (the last
28
ring may only be partially annexed) until the number of boundary nodes
number and relative positions of the triangles for the mesh. The
correct number of nodes are then placed inside R and the mesh is then
formula
(2.1) x. A / 1/(xi)
The authors caution that their procedure does not work well if the input
boundary has nodes with abrupt changes in spacing, or if the domain shape
is too complex. They imply that interaction with the algorithm using a
method.
I2
*1 29
3. A Serei-Automatic Mesh Generation Scheme
grading the net (i.e., making the net finer in selected areas of the domain)
from one region to another, and it may be desirable that triangle interfaces
positions.
compromise. The user must supply a very gross triangulation of the domain,
reflecting the desired grading of the net, and with triangle boundaries lying
in any desired position. This removes bcth of the problems raised above.
The large triangles can then be subdivided by the computer in the obvious
triangles having one curved side, the amount of input for most domains can
be kept small.
triangles, each congruent to the original large one. This has the
advantage that no sharp angles are generated; the smallest angle in the
30
For "curvilinear" triangles (having one curved side) the algorithm is
A A
IC C
B B
and AC as described above. The seven node points on the curve BC are
curve with lines perpendicular to the straight line BC and passing through
the node points on it. The node points on AB and AC are each joined to
the node points on the curve as in Fig. 3.1-b by straight lines, and their
31
ii
Below is aL example of the procedure:
(i
I-
1[.
32
This approach to mesh generation could very conveniently be adapted
to use with an interactive display system. The fact that the user has
control of the mesh while not being obliged to provide large amounts of
33
. ~An Automatic Two-Dimens:lonal Domain Triangulator
in Section 2.2, this method does not utilize a regular mesh; in fact, it
can be used for general simply connected domains, as the examples appearing
must form a closed loop, so for now we assume R has no holes. We then
triangles as possible consistent with the requirements that the mesh vary
smoothly and have no sharp angles or long sides. For example, for a unit
square domain with each side divided into segments of length 0.01 , we L
would like the generated mesh to be composed largely of triangles which
34
'(v
Fiue .
I3
will refer to this method as "notching".
below
P - a2 -
"'"al/ P5
P, c P
b!
L.
b22
b p L
P3
Figure 4.2
where
= L. /4
and 4o[
The first two terms are designed to make the (potential) triangles
56
lengths of the arcs of the current boundary, and the last term attempts to
either or both of the neighboring vertices bave angles less than 5Tr/ 6 ,
Now that we have a method for generating interior nodes, we can now
Step 1.
a b4
P3
Figure 4.3
St ep 2.
Find any consecutive vertices both having interior angles a1 and
choose the pair with the minimumn value of Ia1 - 2v/31 + Ia 2 - 2n/31 ,
and generate an interior node P as described above. We then have a
37
situation such as one of these below:
a, P a2 6
25
(a)2b
P1 bb PP
(. b c--- b
P3 3 '4
Figure 4.T
36
W - (1b 11 +
l-v 1dl -vl + 1b4-vlj + lb
3 (Ibl-v1 + 1al-, 51 + 1a
2 -v31 + lb5-v31)/4v,
Find any vertex having interior angle a < Tr/2 . If there are none,
Ii: P
1 b 1.\
I P
P3
SfU Figure
393
4.5
39
4?
by
v4 (b + d3 + b4 ) /5 ,
v5 . (bl+a/+a 2 +b 4)/ ,
repeat step 3 until a successful trim is achieved or until all the vertices
have been tested. If no trim can be made, proceed to step 4.
step 4i.
its incident bunldary arcs less than or equal to y , choose the node
Otherwise, go to (4d).
4o
Remarks:
relatively fine mesh. The averaging effect built into the node generator
combined with this restriction on the lengths of the arcs considered first
tends to fill in the domain near'the short boundary arcs first; the size of
(2) Steps 1, 2 and 3 are designed to remove any "protrusions" from the
forming any triangle wie check to make sure no nodes lie inside the triangle.
boundary arc and -the sine and cosine of each interior angle were also
41
I
compted once by the routines "trim" and "notch" which actually modify
the current boundary, and were then available as needed by steps 1-4.
(6) Figare 5.10 demonstrates the use of the algorithm when the domain
has a hole in it. We simply provide a "boundary arc" cutting through the
domain, joinirg the outer boundary to the inner one. The fact that the
closed loop forming the boundary overlaps itself and in some parts does
not really correspond to a boundary at all does not effect the algorithm.
The smoothing program (discussed below) does not move node points lying on
these psealdo boundary arcs; hence, this device can be used to force some
of the algorithm described above has been smoothed by carrying out three
The nodes on the curved portions of the boundary were obtained in the
It depends on the relative importance of (a) sharp angles (b) the total
42
-43
A -A
Figure h.8
I.
L
)
L
4
L
V A
44'
-A
Figure li..9
104
4
Figure 4..10
145
me for Finite Element Meshes and Associated Bound Data
element mesh ill probably not be regular; thus the storage scheme for
general finite element meshes, and show that for most elements, the
We shalL see in Chapter 3 that the procedure for generating the finite
easily. On the other hand, we do not want to store copies of the node
more than one triangle. Another point is that we really only need to
the sides and in the interior of the triangle can be generated as needed,
{verticesl and boundary sides (verticesl will be denoted by SI(VI and 0B{V
IBtVB)
respectively. In [El] the following relations between these mesh parameters
are proved.
(5.1) N= (S 2 = VB+ 2VI - 2(H-l)
the figure below, where the domain has been covered by "3-10" elements
beginning with the vertex nodes, followed by the arc-midpoint nodes (see
below), followed by the nodes on the sides and interiors of the triangles.
circled n1ubers are boundary reference numbers which are associated with
6 and 7 are generated ard allowed for in the storage scheme so that some
A4
220
1 21 22 2 23 2h 3
Figure 5.1
The storage scheme is depicted schematically below. Note that the pointers
147
vertices, sides, and interior. Suppose the vertices are numbered
respectively, with vertex nodes, the node(s) on each triangle side (not
including the endpoints), and the interior of each triangle. For example,
148
Pointer List Node Coordinates
21
2
Triangle No. 1 22 Y3
15 x4 Y4
_9 5 y5
22 x6 Y6
4x7 Y7
Triangle No. 2 17
20
19
15
10
Triangle No. 3 2
13
18
17
12
triangle number- 1 1 2 3
Figure 5.2
449
The following table serves to make our point.
01l+02 1.. -m 1
Element+ a] v2 2 2 1 3
Table 5.1
rather than regular mesh for finite elemL methods does not in general
50
the definition of V , S , nV , nS and nI above, we have
(5.6) n
N nvV+ nsS + nN.
1 " n N
- 51
CHATTER 3
1. Introduction
(usually with regard to a specific problem and element); Zlamal [ Z5] has
although they give few details. Unfortunately, we feel that too little
are independent of others, and which ones are dependent only on particular
How much of the computation can be salvaged if only part of the problem is
cnanged and how can that amount be maximized for a given change? Answers
Lenerated
first phase is the mesh generation,
algebraic system.
and the third is the solution of the
V 52
J
As we stated in Chapter 1, the finite element method is a Ritz-
Galerkin method wherp the trial functions have small support. That is, I
the approximate solution is represented in terms of a local basis.
Once we have the basis for our approximate solution v(x,y) , the
next step is to carry out i;r~e integrations required to obtain the coefficients
of the linear system, as described in Chapter 1, Section 2. We emphasize
" V
that the computational procedure is considerably different from the formal
the coefficients are carried out element-by-element, and the actual basis
53
and the process of combining these -a'nctions into a single large one is
5),
2. Construction of Tnterpolating Polynomials
on triangles. However, the procedure and many of our remarks apply for a
the (x,y) plane with piecewise linear boundary 6R . Zlo'.mal [Z6] has
either a common vertex or a common side and with the union of the closed
RR
,
Figure 2.1
L
Our aim
il is to construct a piecewise polynomial of degree d on
55
derivatives at node points lying on their common boundary. We begin by
:LI L
~Y
Ti T 7
[T
Figure 2.2
~kv6k
k k
(2.2) -o s~ - , i 1,2 , <-P
56
which implies v(s) will be continuous along L if
6j V(Q) ~ 7
~I
(2.4) . i =1,2 , j< O-k,
(2.h) n 3n3
6n n~
and summing implies we need a(o+l)/2 additional "normal derivative"
)
parameters situated at nodes along L if v is to be of class C(
the inequality L
The factor 3
2
appears because a triangle has 3 sides and 3 vertices. The L
inequality (2.5) yields the conditions
57
I
triangle. For example, Goal [G2] begins with the 3-4 element (Appendix A)
L in the first derivatives along L . Zienkiewicz [Z3] and Clough and Tocher
'
[Cl] also present techniques for achieving the same goal. Irons [-1]
describes a method for constructing a quartic element generating a
for eliminating the side parameters on the 5-6 element by imposing the
side be a cubic rather than a quartic. Zl~mal [Z6 ] uses a similar technique
58
We now turn o the actubl construction of interpolating polynomials.
Let the number of nodes associated with each triangle be m = 3(m +)+m,
where m. > 0 is the number of nodes on each triangle side (not including
the endpoints), and let m I > 0 be the number of nodes in the interior
nodes (in no specific order). When m S > 0 we assume that the side nodes
Triangle TV V2
V 3
8L
59
~[1
Figure 2.3
To simplify the notaoion in the sequel, we will assume that v. = i
V ((d)
where (d is the n -dimensional column vector whose elements are the
d
monomial terms of the general d-th degree polynomial in two variables.
example,
()T
xy,,(,y ,x x yxyy )
and the vector of those parameters by q " Its j-th element will be
jL 6o
indicated by Figure 2.3, are then given by
(.) qT T T T
fy'Q 1/
LV by
(2.11) ( Zl- 1 , 3 ,. . ~L
:[T 1 ,,
CV aV
61
a
As we stated in Section 3.2, we are restricting our basis to be
T T -T
(2.14) pV(x,y) =aV = qV CV q
T
(2.15) qV (vl, vl, x vi, y, v2 , v2 ,x,v2 ,y v, v3.,x , v3,yV)
1 xX 22 35 2y xly 2 Y1
1 yo 1 x1y1 y 1 x.LJ. -
2 2
0 1 0 2~x y50 x,2x 1 yl y1
0 0 1 0 x1 2y, 0 2 2xy32
01 0 1 1 1Y
5y
x2 Y2 2 2 2 x2 5
2 2 x2Y2 y2 X2 x2Y2 x 2 Y2 y2
0 0 1 0 x3 2y3 o x 3y2 3 y
x4
2 x4 y4 y 42 5 x 24 y 4 x4 y42 4
4 x y4
62
We can write (2.14) in the form
nd d -T d
(2.16) p'V(xy) = k=1 qk (j=lE C(
vT =1q
kE
m Pi
i~ qi, (X Y)
j- 1 ,j=l
where *k4j is associated with the j-th parameter of node k . The 's
and are the members of the local basis in terms of which v(xy) is being
Item (ii) is supplied by the mesh, while (i) and (iii) can be specified by
2 >
the matrix terms are sinple monomial terms of the form x i l 2 0
63
Differentiation of them can be easily done symbolically, with obvious
to (d+l)(d+2)/2, the matrix CV will be non-singular and we can obtain the basis
-T
functions on TV in the factored form *Vi = [Cv q] Having pv(x,y)
equation holds
-T
(2.17) D'p)(x,y) qVC V D'p
T -1 -T
(2.18) q C [,'rq4D(p)T]CV qV
and again the matrix in the square brackets can be obtained syn, ol.caLy.
Note that the ease with which we can manipulate the basis functions
functions must be done numerically [HI] and/or carried out by hand and
64
I
programmed explicitly. This would not be particularly disadvantageous
functions can be done once and the results stored in a library. However,
from our point of view of designing a general a.urpose program we have
of monomial terms.
65
3. Generation of the Equations
2 2 2
(3.1) I~v]=S(a
IR[1VX v +aY V + aa
v+a2 vxy 4v+Y+v ) x
51
I
derivative of v on R
not concern ourselves with the range of boundary value problems that can
L
66
contributions from each triangle TV c R U R Thus we can write
NAN A
(3.4) I1v1 = f[Vj IN V]I IRV])
where IV[v] has the form (3.1) with the domain of integration replaced
by TV , and IVR[V] has the form (3.2) with the contour of integration
and we observed in Section 3.2 that D pV= qVc - D , where the operator L
D operates on the column vector (p term by term. Substituting (3.6)
into (3.5), we obtain the following expression for the first four
(3.7) .V
.V 4 d y C ~~~ap
TI
67
I
Kby
2'2 8RV . Then we have
and again using (3.6), we obtain the following quadratic iunction from
T -
(3.9) c fagqds6W CVqV~
We will denote the sum of the matrices in braces in (3.7) and (3.9)
-T -1
(3.10) A = C HC
T
and the quadratic terms of I[v] yield the function q% AVq.
T -T1
(3.11) qV y
a dxdy + j a7f ds
T -T T
(3.12) q CV wV =qV bV
we obtain finally
68
LN
N TT
(5.-15) I[v] (qV AV aV++qV bV)V
where we note th7.t there will be parameters qjV common to more than one
If we assume all our boundary conditions are natural (i.e., they are
computation can be done once and for all. The savings that can be realized L
depend rather heavily on whether the coefficients of the functional ai
69
{ vertices (xiyi) i = 12,3 is
i (3.15) (x (x
(x + I= +k9
J t
Q
Uy y 6Y1 -AY31/ X
/(3.16)
-Y 3 x 1 ,
in exactly the same way as their counterparts were defined in the x-y
Using (3.16) and (3.17), the integrals (3.5) and (3.8) can be
and
19 70
where w( ,q) = v(x( ,),y(t,q)) , and the g? I's will depend not only
on their respective ai Is, but also upon the mapping and the other terms
to collect the linear and quadratic terms together. Carrying out the above
(3.21)
(521
qC-r
-T -~T f (g
(-T (g1l-T
gV
g3IV
gV T
+ gV-
4 p)JIdtd
+ f(gV6 T ~jV)ds) C- q
C q
1-I
counterpart of AV is given by
(3.22) -T Hv C-i
or
iT -- T~ -T
(3.24) C w q b
71
(3.15), we can easily construct a block diagonal matrix K satisfying
(3.25) q=Kq'
from which we can get, by substitution of (3.25) into (3.22) and (3.24),
the following
T~ T~
(3.26) A' - Kv AK , b =- bV
constants (or at least constant over each triangle), then the corresponding
First we compate
for all i and j less than t , where p depends upon d and the terms
in the
~rlr functional. The components of the integral are then I 2 , where
L.
rI and r2 are simple integer functions. When gi% is not constant over
21
evaluate the expressions in the braces in (3.21) and (5.23). Even in this
instance, having the basis functions in the form (2.14) is still very
2 2-1
f21 2l -2 1 2
1-- 2 + gV ~ +jg V 2
(3.29) g1 4 -? 9 921
li-"2 > 0
(ii) The matrix C and its LU decomposition need only be computed once,
since C is independent of V
(iii) The computation done so far has been independent of the boundary
73
(iv) Consider the calculation represented by (3.22), and denote nd
require
3
2n + O(n I
2,
multiplicative operations, since we need to perform
The following technique has also been used in [p12] in connection with
procedure is
(b) Solve LU = WT
ii
Now consider step (b). We use the following notation to indicate partitions
T
of L U and C ,w here the upper left partition is k by k
L 7U
L
74
We will denote the i-th column of AV by a. , its first k elements
k k
by a , and its last n-k elements by a. . The first i elements of
the i-th row of W will be denoted by w.1 . Then step (b) can be described
k k k k k'
L 1 U 1 ak wk - C2 a k
The first step yields the last row and coll.-mn of A ; the next step yields
tne remaining unknown parts of the (n-l)-st row and column and so on. Note
that at each stage the vector ak has already been computed by previous
n 2 n_ n3 2
+ (n
i + i(n-i) = n i =
1~ i=l L
(n522 ) 0- + Ob= 7 n3
O + O (n 2
has been reduced from 2n"+ to
technique ill not be too important since the number of such congruence
L
75 L
E. St~
e1_ (Initialization)
Step 2
and bV
TI
76
JL
4. Assembly of the Equations
or V
N A
Combining the terms in (4.1), and renumbering the qIs and b.Vs from1
(4.2) Aq =b .L
the form (3.3) are imposed, then some of the ele:ents of q will be
single parameters, rather than specifying relations that must hold between
(4.3) (l(
A21 A2 2 q2 b2
77
As Felippa and Clough [FlI point ot, in order to avoid rearranging
form
(4.6) A'q' b,
Now suppose further that the boundary conditions impose some general
H y
x
Figure 4.1
Then at the point Q ,we want to impose the condition (4.7) on the derivative
L78
(4.7) -V(N)sin a + v(QI) cos a = g(Ql) .
(4.8) Qq' = c
The solution of (4.6) can be viewed as the point which minimizes the
quadratic function
QT)q ) b
(4.10) (At
d) Solve t = b,--QTX
A'q I
At first sight this algorithm appears expensive, since 2+2 solutions L
of systems of the form (4.6) are involved. However, if the coefficients
79
steps (b), (c) and (d) once the LU decomposicion of Y is available.
applied are relatively inexpensive and each one saves a solution of the
system (4.6). For very large systems, the difference between the two
the work required to decompose A' and that required for a back solution
is not as large for band systems as for dense systems, since the factor
is the bandwidth rather than N . Thus our remark above that A' need
such complexity pays off. For typical problems (and a one shot computation)
L the first approach can require twice as much ccmputation as the second.
L80
SL
5. Inclusion of Singular Functions in the Basis
Fix employed tensor product spaces rather than the interpolation method
two approaches was made in Chapter "). Thus, once he had designed the
them in the basis was straightforward. The extra terms were simply added
N
to the expansion l'or v
we wish to include one sin ular function '" in the basis, and assume
(5.1) fu 2dx dy .L
TV
-T
(5.2) *V~
= CV
i
8-L
The approximation to u on TV is thus given by
Sv~ =VC-T~
(5.4) v(x,y) = qV *V + q
T-
-T~ ~ IF 11 1
T 1
In this particular example, the stiffness matrix for T') will be (n+l)
by (n+l) rather than n by n . The extension to more than one singular
function is clear.
I-
82
jL
CHAPTER 4
SOLUTION OF FINITE ELIEMENT EQUATIONS
In this chapter we will stady the storage -id solution of finite element
interested in whether the matrix has exploitable structure rather than just
its spazseness. One of our aims in this chapter will be to study the structure
of finite element equations and to show how such structure can be utilized.
i algorithms (i.e., algorithms which order or reorder the rows and columns
'4 also present two efficient methods for storing sparse matrices.
(1) Storage is becoming increasingly abundant, and one of the prime reasons
for using iterative methods is that they generally require much less storage
than direct methods. Computer memories are steadily becoming larger, the
drums is improving rapidly, and large bulk core storage [F8] (which can be
use of virtual memory [Dl, M5] is another important Jevelopment. Under ideal
2)4
conditions, the user is allowed to address a very large memory (:
2 words
I8L
on the IBM 560/67) which need not exist physically but where addresses
today's computer systems allow the solution of large linear systems with
direct methods.
(2) Finite element methods tend to yield denser systems of equations than Li
relaxation (SOR) for some special tensor-product spaces, and their analysis
and numerical experiments suggest that SOR is more efficient for some L
problems, if N is large enough. Their conclusions are based on the
assumption that the equations have only one right side, and in many practical 1<
situations, this is unlikely. Also, their analysis is based solely on {
operation counts. For tensor-product bases such an analysis is reasonable,
since the structure of the grid and the coefficient matrix can conveniently
L.
be stored in two-aimensional arrays. The data management is no more complex
than -hat resulting from using a five point difference operator on a regular
mesh. However, for an arbitrary triangular mesh, A will not have such
symmetric and only its -upper or lower triangle will be stored; therefore,
to access lines of elements in both rows and columns of the upper (or lower)
in a specific column may require scanning several rows, and visa versa for
column-oriented schemes. By contrast, elimination schemes can be conveniently
this subject 4n detail later; our point is that data management can be
(3) Finally, and perhaps most important, a rather large amount of practical
engineering experience indicates that direct methods are preferable to
usual finite difference methods. This is due in part to the ease with
which we can grade the net (thus making efficient use of each degree
(ii) Direct methods allow the use of iterative refinement [F4, W2],
j 85
II
obtain using iterative methods. Since we d~o not know the true (discrete)
right sides.
11
The study of sparse matrix problems is a rapidly expanding field.
(See Willoughby [W3], and the extensive references therein.) in the sequel,
variational form yield this type of matrix. Following Rose (R31 and Cuthill
of non-zero elements in W L
Rose [R3] has given a detailed graph theoretic analysis of the Cholesky
86 L
Sai /-0 , then vertex i is joined to vertex j by an edge. (We
:1 0 X X0
X X X 0 X0
(1.1) A X 0 0 X 0 0 4
.. 0 X X 0 X
i oxoo x6
L.
The ordering of the equations induces a corresponding ordering of the
1G i 1 by deleting vertex i and its incident edges and adding edges so that
the vertices of 7?i are pairwise adjacent. Using our example above,
we have:
41: 5
;iv 87
x
x
x x x
(1.2) L x x x
x xx xx
the fill-in, and is simply the difference oetween LZ and the number of
Rose [R3] points out that the fill-in will be zero iff for all
have this property (if we ignore the occurrence of accidental zeros), and
Then treating A as a dense band matrix, it is easy to show that the number
L8
of multiplicative operations required to canpute L is approximately
SNm(m+3) M3 ."bn
=
m@23- We will refer to the algorithm as the 'band
B 2 T
Cholesky (BC) decomposition algorithm".
of tSuppose now that F. < m for at least one i , and we take advantage
dense profiles.
Theorem 1.1
A A
N i-2
.= E: 23
2
(1.3)
i=2
-?
Proof:
i-8 f%
-k Z: i W / i
(1.5)
ki ki J=qik jf i
, 89
I:
which requires 8i(5i+1)/2 multiplicative operations and 5i(6i-l)/2
Proposition 1.1
is given by
N1 di(di+3)
(1.6) E 2
j=l
N-1 d.(d.+1)
An additionsa N square root operations and 21 addition
i=l 2
algorithm".
Now we must consider the tradeoff between the amount of computation and
L
90
Note that the graph theoretic analysis of elimination implicitly assumes that
elements within its profile, and it has been our experience that the L's
derived from finite element coefficient matrices do not have sparse profiles.
(See 2ection 4.5 for some numerical experiments in support of this claim.)
just how dense Pr(L) must be over all orderings appears to be an open
utriangular mesh.
So, in summary, we have chosen for various reasons to limit our
Within this framework, our goals are to reduce storage, reduce computation,
and to simplify data management. These gnals compete with one another, and
Pr(A) , B Q
9P and G are all f'unctions of the ordering a of A
91
2. Compact Storage Schemes for Sparse Matrices
will be more dense than that resulting from usual finite differerce schemes,
simple to al-low rapid row and/or column operations on the matrix. The
to have a good deal less uniformity in structure than those arising from
nets and the possibility of associating more than one parameter with each
locate element a.. In this section we describe two methods for efficienoly
92V
Method 1. Let V be a vector defined by
[ij A 0)
{I ( .=) . =
j=l 1
, i = 1,2, .. IN
N
Obviously, V Let
NV.0=*N 0. be define-d by
i1
(3.2) .= Vj
V. i 1.=1.. P
The non-zero elements of the i-th row of the lower triangle of A are then
8 1
6 4 3
8 9 5
1 4 6 8
9 12 SYNIvRIC 10
(2.3) 4 8 2 2 14
2 4 16
2 3 3 19
1 6 8 22
8 9 9 25
4 7 27
L 9 6 2 30
6 8 32
5 2 2 35
7 44 38
U
93
Here NO = 38 and the vectors S and. w are given by
8 0
6 1
4 0
8 1
9 0
1 2
4 ~ 1
6 0
( .4) S- 9 w= 3
12 0
4 5
8 3
1 1
2 0
At first glance, the overhead for this method appears prohibitively high
since each word stored requires an extra word to store its "offset" from
the diagonal. However, note that the elements of w will all be bounded
the overhead is only about 25 percent, and the total storage required is
the elements for i- < k < p, are ordered, a binary search can be
following:
1: Method 2. Let A and be as described above, and define the
~(2.5)
i
F)'-A
+ 1) , i -=
1,2,...N.
V j=l 1
Now define o by
i- 10 if a. J
We again use the example (2.3) to aid in understanding the scheme.
The arrays V and w are given by
VT = (1, 3, 5, 8, 12, 18, 21, 26, 30, 36, Ito 46, 50, 56, 60), and
T = (l
T I ] I ]0]
4 0llOolO lllol1 lO0 01000111)
row 12 3 4 5 6 7 8 9 10 11 12 13 14 15
Thus, the zero/non-zero structure of the i-th row of the lower triangle
Note that the storage required for jt and P becomes less significant
with increasing N and fixed N . The use of a bit array may cause some
95
retrieve element a.. (Note that i -i may be >> P -
mayb >i-
For example, using this method on an IBM 360 computer to store a dense
500 , 300 symmetric matrix in short precision requires less than 4 percent
when IPr(W) I; N ; that is, when there are few zero elements within
Pr(W) . As we mentioned before, it has been our experience that the L's
matrices, and it is primarily for this reason that we present it. The
overhead for this storage scheme is only the storage required for 11 . To
necessary.
Finally, we mention the most cemmonly used method for storing band
96
N1
i
I|
a2 1 a2 2
a31 a32 33
(2.8)
a ,1 a
i]. . in
Proposition 2,1
Proof:
N
= N + !Pr(L)I = N + = (5i+i)
N
=N + . + N
N
=(N+I)m + N - (m-F)
i N
'h B + N (m- i )
97
Thus, Proposition 2.1 says that for any ordering, the storage
required for method 3 cannot exceed that required for method 4 by more
Essentia.ly, (2.9) says that for a fixed m, IPr(A)l can vary by nearly
a factor of N.
98 L
II
98
3. Node Ordering for a Small Bandwidth
matrix are to reduce the storage and computation required to solve the
only if we plan to store and process the matrix as a dense band matrix. In
view of Prop. 1.1, Prop. 2.1 and (2.9), the only justification for ordering
section we discuss the reasons for bandwidth reduction and present some
algorithms for obtaining small bandwidth orderings. Note that the question
and positive definite, there sess to be no reason to use the BC rather than
However, the linea2 system we vant to solve may not always be positive
u origin which destrcy the positive definiteness o2 the system being solved [W2].
ri 99
situation m = mA. is important, since we can only guarantee that
mPA < mAA and the combined storage requirement for L and U (using
systems may be important in this regard since a shift of origin does not
small. Note that this does not preclude the possibility of using the PC
clear that bandwidth ordering makes no sense for iterative schemes that L
require only a residual calculation.
the minimum bandwidth is achieved, but we should get reasonably close to the
ascertained lower bound for the bandwidth (not necessarily attainable) can
!00
11
be obtained by finding [k/2] , where k is the largest munber of
use. They can be classed as direct (or one-pass) and iterative. The
direct schemes [R3, C5] usually work closely with the associated graph,
and proceed by successively removing (i.e., numbering) the nodes of the
of the graph. The iterative schemes, on the other hand. assume a given
the direct methods only need a single staiting node to begin, while the
use a direct method to obtain an initial orderine and then use a iterative
is discussed in Section 5.
increasing degree.
101
V
The algorithm is equivalent to finding a spanning tree (rooted at the
graph with N nodes arid N-1 edges. A spanning tree of the graph G is
vetices, and thus only increases linearly with N - Very good results are
and is as follows:
1. Set i = 1.
vertex satisfying
= min
i-I.o
yEX.i-
102
I
produces a larger bandwidth than the minimum degree algorithm, and even
with the above modifications the latter requires substantially more work
The general idea follows: Assume we are given an initial ordering yielding
1. Set max = m
Step 2.
values as follows:
may increase; after the first step repeat as long as the bandwidth
103
L
Step 4 has the effect of reordering the rows so that as nearly as
possible each row has The same number of non-zero elements on each side of
the diagonal element. It could be called the balanci sage. For matrices
that have a: imnerent band structure (as ours have), Step 4 does not have
much afect but for randomly sparse matrices Step 4 can improve the performance
[.
lo4.
2
2
104 L
4. Node Ordering to Reduce IPr(A)I
In the light of Prop. 1.1 and Prop. 2.1, it should be clear that if
(= IPr(L)I)
The term "near opzimal" as it appears in the literature [R3,T3]
have not tried to look for orderings to reduce it. Tacitly, we have
graph G by
of x. and its incident edges; no edges are added. This provides the
1. Set i = .
L 2. In the elimination graph G_ 1, choose x. to be any
1
ID(x i )I mm I(y)
min I
S105
where
- Gi_= (Xi_l,Ei) V
3. Set i = i+l.
In this direct algorithm the next node to be numbered is the one that 1.
will introduce the fewest non-zero elements when it is eliminated. It is
usually must be stored as a bit matrix, and few machines are bit-addressable,
these tests may involve considerable overhead. As with the minimum degree L~.
"
Step 2 to those nodes that have at least one numbered neighbor does not
hurt the ordering produced by the minimum deficiency method, and this
I Compute .*
Q=i(i N -fi)
v N
[=
The actual search for the best interchange (Step 3) is by far the most
expensive parL of the algorithm, and in a practical situation only those
rows with v.
1 greater than some threshold should be tested since the
Good results have been obtained with the parameter k mentioned in Step 3
set to 5.
11077
iii
k
the matrix arising from a finite difference method applied to the same
problem. First, the node points of the finite element mesh may not all
play the same role, and as a result have different connectivities. Whether
there is more than one parameter associated with the node will greatly
finite element mesh will very likely be graded, which also causes disorder
in the structure of A . I
Our aims in this section are
Pr(L) I'-
We will make use of the f'ollowing labels for the different algorithms
and quantities in this section. Some of them are repeated in other sections.
i08
CM -- Cuthill-McKee
B2 -- Bandwidth Reduction
I mprovem ent -,A-'gorit has
PR - Profile Reduction I
BC -- Band Cholesky
PC Profile Cholesky Decomposition Algorithms
B BC
multiplicative operation count for the [) decomposition algorithm
L"B
1/ -- storage required to store a symmetric or lower triangular
matrix using the band oriented method 4 (Section 4.2)
?(A)
PI -- profile of the matrix A
In order to keep the number and size of our tables at a level where the
obtain an initial ordering for the iterative improvement schemes (BR or PR).
We have limited our studies to tbe ordevings provided by CM, CM-PR, RCM and
MDF. [The hyphen should be read as "followed by". ] The application of the
BR algorithm to the CM and RCE orderings reduced m by only one or two, and
so the results are not included. The application of the PR algorithm to the RCM
and MDF ordering resulted in only a small reduction in Pr(A) , and was also
109
4.
not included. We have limited our studies to elements 1-5, 2-6 and 3-10
Figure 5.1
described in Section 2.3 in such a way that for a given domain each element
yielded the same N . The reported times are in seccnds for an IBM 360/91
computer. The values of 9 and V for each algorithm have been scaled by
the values for the CM ordering. The actual values for the CM ordering are
one of the two most widely separated triangles in the domain. For "long,
straight" domains this will obviously work well, but for U-shaped domains,
for example, it could lead to a bad choice. One should have the capability
for a good starting node would probably be more expensive than i-s ultimate
110
The results of the excerinents are contained in the following three
tables.
1-3
3-10
Time .25 9.9 .25 27.24
Z
m 63 81 63 8L,
9B ;i(406188) 1.55 1 1.4
N =241 9p 1(239873) .30 .21 .19
] ( 1 54 24 ) 1.28 1 1.3
'(B
: 49 .44
Z 3793 Vp a (10Y(2) .58
c (Pr (A)) .2O3 .3534 .428 .467f
&(Pr(L)) .979 .985 .990 .99,6
Fill-in 7707 36o6 2655 2 77
Table 5.1
11i.
Matrix and Elimination Statistics for Several Ordering
Algorithms for the Hollow Square Domain
1-3
Time .17 1.79 .17 9.2
m 16 16 16 20
0 1(34776) 1 1 !.5
BI
N = 252 p 1(26299) 1 .99 .95
(B 1(4284) 1 1 1.2
wA 1620 Vp 1(3598) 1 .99 .97
z
.B(Pr(A)) .272 .272 .273 .278
.&(Pr(L)) 1.000 1.000 1.000 1.000
Fill-in 2510 2505 2!,89 2421
2-6 V
Time .25 10.49 .23 18.3
m 36 46 36' 42
9B 2(155609) 1.56 1 1.32
N = 252 9p 1(95520) .35 .36 .33
VB 1(9324) 1.27 1 1.16
NA= 2628 V 1(6837) .60 .62 .59
zp
.&(Pr(A)) .219 .372 .362 .381
.&(Pr(L)) .982 .972 .991 .994
Fill-in 5029 2328 2493 2316
3-10
Time .32 1O.6 .32 25.1
m 71 78 71 63
@B j 1.17 1 .81
Table 5.2
L12
Matrix and Elimination Statistics fur Several
Ordering Algorithms for the Inverted T Domain
1-3
Time .22 2.52 .22 9.32
m 20 20 20 35
9 l(63283) 1 1 2.82
N = 301 9p 1(31496) 1 .88 .74
VB 1(6321) 1 1 2.52
.95
z z-1,,9 1(4258) 1 .88
'(Pr(A)) .280 .280 .296 .321
&(r(L1) 1.000 1.000 1.000 1.000
Fill-in 2852 2847 1886 2334
2-6
Time .30 12.36 .30 18.58
m 40 46 40 42
9B
1(230017) 1.3 1 1.1
N 301 p 1(103348) .34 .34 .28
YB 1(12341) 1.15 1 1.05
NA 2628 Vp 1(7422) .60 -59 .56
&(Pr(A)) .238 .406 .415 .441
3-10
Time .30 11.5 .30 22.2
m 57 72 57 65
Table 5.3
113
The information in the above tables leads us to the following
conclusions:
[We have computed the fill-in for some random orderingj as well, and
(2) The RCM algorithm seems to be easily the best algorithm. The ordering
not only supplies a near optimal bandwidth, but also yields a prfile almost
nodes. Secondly, we not only must test edges of the graph, but we also
must usually add edges as new elimination graphs are formed. This addition
of edges requires computer time, and also increases the degree of the nodes
sum of the squares of the degrees of the nodes being tested, these added edges
The reason that the RCM ordering is superior to the CM order (profile-
114
e
methods. The following table, which can be obtained from the tables
Domain I
a - Hollow Inverted
ElemSquare T
13 .098 .097
3-10
S.2 j.28 .28
Lo change much.
115
The Value of NA for Arbitrary Elements and Triangular or
Quadrilaterai Meshes
lying in the interior of the mesh. Let H be the number of holes in the
mesh (domain).
on each side, and the interior of eacn triangle. For example) element 3-10
Figure 6.1
116
EI
Lemma 6.1
Proof-
nodes) are not removed, and there are (n + 2nv 2 such non-zero elements.
S
This proves the lemma.
Lemma 6.2
17
Proof:
of the triangle sides and their incident vertex remain in the mesh, so the
assuming the equations in question are all grouped last in A and examining
Figure 6.2. The submatrices marked with an asterisk are the parts removed
I i I I 66
MatrixA I I
i--
I * parameters associated with the
internal vertex.
"sameters lying on an interior,
- -- - ieand
-- F],(it outside vertex/
internal internal
vertex-- sid( 7
S2 - outside vertex
2 01
.''. external side
Figure 6.2
118
Summing the elements in the marked submatrices yields (6.3).
Now suppose the mesh has a hole in it. Eventually we will reach a
situation where the hole is bounded at one placp by a single side such as
depicted below.
ij
V side to be removed
Figure 6.3
following
Lemma 6.3
Let nV , n. and n, be as above. Then the reinovwl of a connecting
The proof is siilar to that employed in Lemma 6.2 and we omit it.
119
Theorem 6.4
Proof:
removed during the reduction of the mesh, and this will reduce NA- bya H
These two forms of demolition account for the removal of 2(V-3)+H triangle
sides, and three sides remain in our final triangle. Hence, we must have [
removed S-3-2(V-3)-H triangles of type 1, accounting for a2 (S+3-2V-H)
I j
\ I
NAZ n2
(6.6) : a1 + a2 y2 + nn27.
jwhere
a1 = n (2nv+ns)
+
a3 = n(nI + n)- (n, + n8S 2 + 6(ns nr)2
n = 4(n+ns) + n,
271 + 2 = V-4
+ =
Y1 + Y2 Y3 -
I1 12
The numbers ' ?2 and y3 are, respectively, the number of
the fact that there are alternate ways to demoish the mesh, resulting in
have
2
=2 (v-4.) + a3 (2N-S+ 2) + n .
If our mesh has H holes in it, and we rename the a3 of Lemma 6.3
(6.9) NA = a -4) +
A -S(2N-s+2+H) + - + n
2.
know the mesh and the characterization of the polynomial on each element.
It is also useful in checking that our mesh is consistent and our program
is working correctly.
122
IL
I
(6.10) O(A) -- 2
f(n, + ;7 n + . nslN~]
L r(nV,s,h)
where
(6.11) f'nV,nS,n1 ) 12
2(n + 1 nV + 3 nsl
Lhe average number of non-zero elements per row of the coefficient matrix
Some typical values of P and average number of non-zero elemen65 per row
I'
j Table 6.1
123
L
iv
Model Problem
profile methods rather than band methods. The mesh we consider is obtained
by subdividing a unit square into p small squares of side 1/p , and then
subdividing each small square into two right triangles. An example with
p 6 is given below.
J\K..1..
side, and the interior of each triangle. We number the nodes diagonal by
124
Ft
diagonal, beginning at the lower left hand corner, and considering nodes
29
/1
/ \2Z
12\2
,
.t >\ \
AU
13 \2
:"s" ',\ rw
\7 15 ~2
\\ \o3, 2
ii '\ \ 16 17 "2 "
tIt
12
Figure 7.2
125
I
As usual, we denote our symmetric positive definite coefficient matrix
by A , with Cholesky factorization LLT . Making use of (6.1), and
2
(7.2) N = (n+ 3n S + 2n)p + O(p)
B O
PB P B 23
= Thus) 9=-p2 and "B P-
B _
Theorem 7.1
For a p xp regular rectangular grid, the coefficient matrix A
126 1
1Proof:
HConsider'.ng the first node point as diagonal 1, and recalling the
A
5 =0
A
2
5 -2, 54 2
5A 2 8A 2
3-4
8A 3 6A ~ 8
A
(7.6) N- ' (r-1)(r-2)
2 _< 1_ < r(r-1)
2 -1, 3 <- r < p+l
(
Using the formula IPr(A)j = N 5A + 1) along with (7-5) and (7.6), we
have
p 2
+ 2.3
+ + = + p2 2
Thus, using the profile storage scheme rather than the band storage
127
[1
straightforward (but tedious) to show that for a general stencil;
with pp B "
N F) +
A A.I 3) L
(7.8) .p'L 2
following
Theorem 7.2 [
Let the EC algorithm be applied to A . Then the number o2
Proof:
1I2
128
2jI
p 3 + i2 + 7P2
__ (i 3 +6i + 7i) + p
L~ 2
1 p 4 +3p2+ - -P
(7.10) Pi P
L by using profile instead of band methods. Note that we did not prejudice
[1
129
V
8. Miscellaneous Topics and Concluding Remarks
in the form
jV BVVb
B2 1 B2 q2
(8.2) by eliminating q:
BVBV-V V V V1
112 B2 2 2 1 q I b1 L 22 2
In this way, the dimension of Al, (see Section 3.4) can be reduced by
problem discussed in Section 4.g to show the savings possible by using this
technique To simplify the analysis we will consider the use of thd band
9B 125p
N
4
is about 5p'- , yielding
L
(8.3)
130
I
F1
Now consider the corresponding quantities if we apply static condensation
2 2
and eliminate . variables before assembly. The bandwidth m is now
pi2
only about 3p, and N !3 . Thus
(8.5) 9 32p ,
S-
L for this particular element, problem, and solution method for all p . Of
course, its use might be justified for storage reasons alone, even if it
32
fl (8.6) [ -6-+ 3n (ni l)(nS+n V) + 9ni(nv+n) 2 N
of the coefficients depend on more than one element) and the elimination.
point that is usually made in favor of these schemes is that variables are
131
eliminated as soon as possible, rather than in a predetermined order.
complexity, and the question of ordering has really only been moved a
level higher. The problem of optimal equation ordering has been replaced
high premium.
132
CHAPTER 5
1. Introduction
in this chapter have been produced by a general program. No -.,e was mrde
I would reasonably expect. For example, we graded the net small near the
the problems below are upper bounds for the true eigenvalues.
133
2. The L-Shaped Membrane Eigenvalue Problem
authors. For background materiel, see Forsythe and Wasow [F5] and Moler
[M3], and for various special computational methods, see Reid and Walsh
[R21, Fix [F2), Schwartz [1S7] and Fox, Henrici, and Moler [F61. The
find the stationary values X (O <\% <k2 <_ < ... ) of the
fu-nctional:
(2.1) Ilul=
u2+ u2 )d
u
y
/ u
2
dy
U
R R
134
NI
I.
Figure 5.1
LWe are obviously making use of the symmetry of the l)rst eigenfunction
here and have graded the net appropriately near the corner. In the
tables below k is the factor by which the mesh of Figure 5.1 was
to the last digit. Set-up tLae includes the time required to generate
'the mesh and order the nodes as well as the time required to generate
I i and assemble the equations. The missing times in the table were so small
that they were meaningless. All times are in seconds on an IBM 360/91.
L
L15
[.U
I
1 5 3 .33 -- 3.4003
2 22 6 .57 .05 1.0089
3 51 1P .62 .10 .4605
4 92 12 .93 .17 .264o
5 145 15 1.33 .30 .1718
6 210 18 2.34 .55 .1210
7 287 21 2.73 .90 .0901
8 376 24 3.53 1.28 .0699
Table 2.1
*
The rate of convergence of the ,computed k I to kI as k - o is
dominates the solution time in all cases. This is due in part to the
rapidly and efficie u],m . The set-up 9:_jcedure, on the other hand, requires
Another reason for the relatively large set-up time is that we are using
136
(Set-up time)/(Solution time) is steadily (if slowly) decreasing.
Our second experiment again makes use of the mesh of Figure 5.1,
they can be. Table 2.2 contains results for polynomials of degrees .
through 6 ; in all cases the original mesh was used. Our inverse
[%
l-X I for Piecewise Polynomials of Degrees 1 to 6
1 5 3 .53 -- 3.4003
2 22 11 .43 .1 .3720
3 51 24 .70 .25 .o16o
4 92 42 .87 .95 .0063
5 145 65 1.7 3.02 .0034
6 210 101 2.68 6.02 .0021
Table 2.,2
L3
137
Li
were virtually the same; we required 15385 words (including the storage
of A and B of the generalized eigenvalue problem Ax = %Bx ).
units is about 100, while the effective memory speed ratio is about 10.
The ratio of times for other operations lie somewhere between these two
extremes. We feel we can safely say that the finite element method is
problem.
have used a method due to Peters and Wilkinson [P1] which essentially finds the
larger than for inverse iteration, the required storage for our quintic
deflation would have to be done implicitly which implies that the eigen-
We feel that the ability of the high order finite element methods to
138
Ii
apply well known, dependable methods for finding the eigenvalues of the
H discrete problem.
Ii methods are the best ones to use for solving this particular problem.
[I Indeed, the method proposed by Fox. Henrici and Moler [F6] is probably
the best known method for finding the eigenvalues of the L-shaped
which may only be .own to an expert in the field, and the utilization
B
LI
Ii
LJ I139
LI
U
139
3. Eigenvalues of Rhombical Domains
by Moler [M4], Birkhoff and Fix [B7] and Stadter [S41. Moler obtains
we will show that with finite element formulations having relatively few L
parameters we can get close to or within the bounds produced by the methods
xi
((, 0Ii
Figure 5.2
14o
Our first experiment takes no account of symmetry, and the results
1 are compared with some of the bounds presented by Moler [M4]. These
factor by which the input mesh (indicated by the dashed line in Figure 5.2)
-. calculations were done on an IBM 360/91. Our set-up times (for each
U example) and solution times (for each eigenvalue) have been included in
Table 3.1 for comparison purposes. The 360/67 and 360/91 have radically
U1 different design features and a comparison between the two machines is
difficult. The largest ratio of execution times this author has
11
[ encountered between identical prcgrams run on the two machines is 15, and
n] that was for a verj special program. Usually the ratio is from three to
U
ii 141
JLl
00
%_ x* Set-up and
, 34 Solution times
m = 121,.1
k=3,
N =121, 2.52284 5.33339 7.26653 8.49424 .69
m =43., .95
d=4
I k= 4, 13
N 22.52279 5.35134 7.26611 8.49374 36
j.m =65, C52933343.7
d=4
k 2, 7.26651 8.49420 .
N 81
N 1 2.52284 5.35' 1.2
m =55, .4
d5 _
Table 3.1
i 142
iii
- Case 2: Rhombical Membrane Eigenvalues: 9 =
*:ji
"3. 4
N = 49,
m=22, 2.01232 4.90567 5.16407 8.00979
N = 121,
4 4 5=.
d 2.01226
N =81,
35 ,
Iid m
= 5,
md35, 2.01226 4.9040.8 5.15730 7.99851
k=2
N = 196,
m =.69., 2.01225 4.90389 5.15705 7.99308
d = 51
_k= Ik'3
Table 3.2
L- 143
Our first observation is that again the higher degree polynomials
pronounced, although still apparent. We point out that our numbers are
particular solutions.
Moler's results are for moderate values of G , and only for the
fixed membrane problem. We now wish to make some comparisons with the
results of Stadter [IS4] and Birkhoff and Fix [B7]. They report bounds
for 9 = 300 (150) 750 for the rhombus fixed at all edges, and Stadter
reports bounds for the rhombus fixed at two opposite edges and free on
We begin with the fixed membrane problem. The bounds reported are
shown below: [
L
N L
l14h
L
11,
-x
9r9
Figure 3.1
j
H
The boundary coiidition un = 0
the corners. In both cases, only the even-even symmetry class was sought.
the number of intermediate problems used, and is the size of the two (dense)
L U 145
The eigenvalues X2,k
n below are our finite element results for
problems.
"nuisance factor". All our finite element computations were done with a
each angle on the IBM 360/91 was about 0.6 seconds. About 0.2 seconds
[Since it appears that the major portion of the time used was for the
generation of the eigenproblem rather than its solution, the fact that L
Birkhoff and Fix used a method yielding all the eigenvalues of the discrete
146
I
Symmetric Eigenvalues for the Fixed Rhombical Domain
n XB L,50O50 %U
r, n n n
Case l: 0 = 3e
Case 2. = 450
Case 3: Q = 60
I; 1 6.3238 6.3598 6.3217 6.3485
2 14.968 15.088 14.958 15.005
3 25.333 25.571 25.202 25.338
14 38.064 38.981 37.436 37.774
5 43.581 43.717 47,.480 44.013
6 5-4.267 56.379 51.883 52.575
0
Case 4: Q = 75
1 20.194 20.283 20.185 20.407
2 36.373 36.452 36.301 36.617
3 53.596 53.562 52.794 53.499
4 76.746 80.125 70.951 72.660
5 110.20 111.52 90.964 94.982
6 154.89 144.38 112.87 121.75
L
Table 3.3
4 1147
We offer the following observations:
(1) The remarks of Birkhoff and Fix suggesting that their Rayleigh-Ritz
methods yield much more accurate upper bounds than the method of
(2) The upper bounds produced by the finite element method appear to be
fully competitive with the XBIs , and are appreciably better for the
n
lower eigenvalues.
(a) information about the behavior of the solution near the corners
of the domain
(b) the fact that the domain is affinely equivalent to one in which
problems.
1h8
was not convenient for us to restrict our problem correspondingly, so
4)15 and 4 l15
we solved the "full" problem. We report results for an 3
1 1 1 3 3
L.
L4
I-
[
4. A Dirichlet Problem
2
-u+
2
u = 0 on R F
(li..l .,.2
u =e cos y on S UT
Ts
zI.
The inp t mesh is indiceted by the dashed lines in the diagram above.
the factor by which the input mesh was subdivided. The profile Cholesky
algorithm and the RCM ordering (see Chapter 4) were employed in all
cases.
We begin by comparing different element/mesh combinations which
150
iI
Operations
* Set-up Solution for Storage
Time Time Solution Error
Element N k (Seconds) (Seconds) X (10 + 3 ) x(lO- 5 ) L A
VTable 4.1
polynomials for solving this problem. Set-up times, solution times and
U We now present some experiments using the initial mesh and varying
the degree.
I-
Operations
for
Solution Solution Storage
Set-up
Element N Time Time (X 103 ) Error L A
Table 4.2
151
Again the case for higher degree polynomials is apparent. Compare, for
example, the third entries in Tables 4.1 and 1.2. Their demands on
system resources are about the same, but the error for the quintic
which makes use of very fast direct methods for solving the discrete
has been found to be superior to SOR or ADI solution times (by factors
152
i
~fl Each solution of the 127 x 127 rectangular problem requires about
. seconds on the IBM 360/91. Thus, even asing the iterative scheme
L (based on fast direct methods) proposed by George [GI which avoids ,e
in overall time (solution and set-up time) with the last entry in
requirements, and the last entry in Table 4-.2 requires only 5805 words.
1.16x lO - 8
The observed error for the sixth degree polynomial was
6
kl compared to 7 x10 for the difference equations.
Again we should point out that there are still better ways to solve
7' this problem if we are prepared to take advantage of its p,rticular 2hara2 teI-
Idiscret iet of points on the boundary SUT. The least squares solation of
a 26 x 15 problem was all that was required and the program was only a few
L
L
IL
I
IL
I
Ref c r nce s
I
p.101-122.
155
Dl1 Robert C. Daley and Jack B. O)rinis, "Virtu-i1 'enory,
processes, -ind sIhar ii in tMult ic -," Co.-,.,-. Assoc.
Comput. Mach., 11 (1969), op. 35-3112.
D2 F. W. Dorr, "The dIirect solutio.-i of t.,2 liscroto
Poisso.i equation on a rectan,;1E," SIV~ Rhv., 12 (197n),
pp. 24~8-263.
D3 Goor;e Dupuis and Jeail-JacqUOS- 'o l, "Elements finis
raf fins enI elIasti,-i tV I)id.i eS iO PIe1," Z. 4~w
El D. J.
"Rules
F. Ewing, A. J. Fawikos, an,! J. R. Griffiths,
;overning the numher of io-ies an1 l ?lornents in a
finite elem-2nt mesh," Internat. J. Iunier. MAeth.
1
Engr;.,2 (1970), op. 597-611f0.
'AE2 Es'eIi
"1 "Desip-n and
,m, I m nt aetio- -)f in al-ehr-iic
processor," HabilIit-itionsschrift, Institut f'ir
Angewiandte Mathemnatik -4er ETH, Zuricl', 196F.
F1 Carlos Fel ippa and Ray V1. Cloij-!, "The finite e1.-'int
mecthod in sol id m-chanics," po. 219-252 of B9.
F2 Gcor,-e Fix, "Hi-~h or-Ir Raylrei -h-Ri tz ap-)ro i:mt ions,"
J. tiath. Ilech., 18 (1969), -op. 5L4 5-657.
15(L
F8 Kurt Fuchel and Sidney Heller, "Considerations in the
design of a multiple computer system with an extended
core storage," Comm. Assoc. Comput. W1ach., 11 (1969),
pp. 334-340.
157
Ki i. A. Karnel an.1 H. K. Eisenstein, "Auto'iatiz : shi
generation in two and three dimensionail interconnecte'
'omains," Symposiumi on HiF'i Speed CoirnutinTT of Elastic
Structurcs, Liege, Belg~ium, 1970. I
K2 Ian P. King,a "An automatic reorderinF scl-;-in for
simultaneous equations derived frori network systcP's,"
Internat. J. Numer. Meth. E. 2 M171) .
523-533.
K3 L. V. Kantorovich and V. 1. Krylov, Approxiimate HiQtho- s
of Higher Analysis, P. 'oorIlhoff Ltd., The 1.ethiirlan-!s,
1958.
Mi 1 .S. Ma~rtin and J. 1i.WilIkinson, "Solution of
symmetric an,, unsymmetric ban.] ecuations and
calculation of ei-,envectors of band marcs" Numer.
Math., 9 (1967), pp. 279-301.
158L
I
[1
2L
S7 Blair K. S;.1artz, "Exr'! icit O(h )bourn-i; - th'.
07?. 40-59.
1.2
1
L 161.
Api-tiJ.Lx A: S ;:' Represent itive Triangular Elements
a
3L.A1d. Tne two-part hyphenated name refers respectively to the degree
,f t'ie pulynumial and th, nmber of nodes associated with the element.
-3
2-6x i
V, V,V
V,V v ,v
x y
3-10
-- V ,
x
V ,VV VxVY
4-6 vx y x Y
L
Element Name Stencil
Vx vylxx xy Vyy
V V
n ')xn
5-6nn
V vyvxx vxy) yy vn x y xx xy yy
5-21
6-280 0
'4p
Appendix B: O/S 360 Fortran Code for Finite Element Methods
The codes in this appendix are all written in O/S 360 Fortran. There
are five separate programs whose relation is depicted in the following chart:
10
Li
Solve Ax kBx using
inverse iteration with
ni(userc indefiniteopinslvr
symnmetri band linear equation
164
L
'B
The transmission of data from one program to the next is clone via Fortran
units 1, 2 (and 3 if an eigenvalue problem must be solved). All data sets read
and written are sequential, so the program would work without alteration whether
the storage devices are disks, drums, or tape units. Only changes in the job
Il[v]/J[v], where
I [v] = LX
c4 v dx dy,
R
J = ff v2 dxdy
Rv]
jJ R
Here cl, c2 , and c3 are constants, and c4 is a function supplied by the
U user in the subroutine FUNC. For further details and sample input see the
With minor changes in the mainline of PHASE 2, other terms can be included
inI1 and I2, and with somewhat more substantial changes variable coefficient
quadratic terms could be handled. Note that phases 1 and 3 would not need. to be
altered.
and first-derivative parameters can be selected by the user and are automatically
The choice of method for solving the generalized eigenvalue problem depends
L on the relative size of the number of equations and the bandwidth, as discussed
in section 4.3. Both programs assume that the initial shift (SHIFT) supplied
L by the user is a good one; the decomposition of A-SHIFT*B is done only once
165
Phase 1
of the Nodes
I-. 0- wu<
> -i 0- - L V, &
4x CW 5% cce- -
ILJI.
~ ~
M
W U. % - Z
1 - - w L'- H
>W =)- Lu
LL. -4
V, ) 4
LL> z -a
X
U,- tn
M3 -a- 3% %L 0 1 0 1
c - -n LU VS W w~
LA) x. 3) MWO
'C' %<L V, 4 Is
3% 0 wI -J, P
c- u CL LL II;:
"o.I
ea-C)..
U. j 3 ~ .I ZS z U
w UJ w X2 % H-JUS8 d'L..JO -
LI 7
3: i aI w I. > ka ii-i.-
A % <
- - ---- - U
I
I
-
'~~~~
0
Z~~ cc
LU
0
I
ccH--
.1* *0 LU A
AI. J -jW -
* 3~ -U IA cr. US
NX L w.LL
Lwt L4 u zC 4 A. wt u
WLU
U) LI Si
WC 0 ' LIJ H- VS a.
Q6 55 Z J cc2I*
z z 0 LU f. w >. .4 4 V--,.
U fa. c
*~~L 3 .'C- - - 4 L lUX LU SSj (x V) -a. IA..
* L J C SC
Pt2I 0 % 34 4 I-L- 4 2 C
C2 C0 0
VS L.
LI
:-L
w
LI.
-
-,A
0J
w
ce.
H-LL
w-
Li
2-
Cl LI L LI
U~
LL w V, ~ t.~V
cn:) 0- wI CD 13.
C *- cc 14 -< z z -1J 4 zA zJ 2>. G- <
7 A
* C Lu 55.
tL -2 - f A 7 1 .LW C -C ZL ., 'IN
t U.
* 0 LI. LLU)
Z I -
.
4 S
0
4*
54- -
C '
t
LU tA
LU
CC
H .I
D. 5
I II
1.4 C aZ
9X
-
H-S
.a(
S ~ A
A
- C C C- L, - - . 4 IC S t z I
- .I -, I
U. cc
C.. A) 1 > - .? 7A z r, 7
Z. m U,
I j.x x
In>4
[1
.0 N N
o
IS -.w* w
C- .)
Z ~~ J CJ.
un
wg II
0 U - z C;
OW~w -I XP
- . tj U . L3 U , M 0 a
C, T- I c T- 0 z C
z .00 - -V
q
00 CjL 1-
- IT a -
- - - - - - - - - - - - -
z 0'A0
1 0 0
I SI Z )
I Ix T -aC
- I I T, I
IZl,
Ic a
I a~rQ
I4 TN I~ 1 0 C I
*~ U - Ca -- C
IT 5 1 I
11l TJ~
SJI ,U4Oi'~0
* 32u r0
* *'Z I u.' A 2. C IU
4CC lea 9 1. 1-a. . U. w
*u U. I U L ~ a .- >
*o' I WU - or U C CL2 - 0 I
42
6w
CC
Q
C.
~ I-
't.
wC
W~IIIU~4..C~.
"'.-. 41ZI
.
u ALI
4,- C, I 3C.X-~ IU U U 4 -/ L Wilt )UCZ . I4
c* -.
/ C I P two I' S - I-U-I . Z UI- U.)I to
II 5I Iz U I.. It j I Oz x o
*l I r- 'Aa
4t I I,
t I, I41u
t, U. 41 c1
I. D
cc 4x a ;-
4AC el < 4l u
D VIZ24
2.3 4. WCC
x1 vN I-5 **)
o1 o I a3
1S .1 w 21/I v L l, 4 eC Li
In U!C * 3 -
4 1/lI m~i c%
0 C 143-20 t "'I1
I 4.3W r^*
U.0 *1C (5 w C S 4n I ppp.: x4
c~~ I toI- t
7f on mo 3U - '. ;9 (II
>
1, I. QU o-0 1 V S c I l oto1
.. 6
'Ut
9
- c
4
a aV,
2-a a, 4 cc
ZZU0. * 4 -
az
cc
,aC
U. 00
ccaW- ~ a.
C-:
L)-ta..aU e w
Ix
w.J
.4
6. x a j Cc-
a- aL Q.--
It
a a. w. .. a4U
ICL a) a
CL' u a
U
atu C. I -.a.Ua, r
~~~LC~~a.
-
.4 a ~
U
a
.
~ua -
. ~
a.
.4- ~ -
-C
U)
-. UO,2zv
a- a-U
aJ0 .-...- C *-
>a. U. *>
U a.oz
SI-
Oa. aC a- U).-)2 .4 NOW
Z2.0U)CC
S. a .a.-. a4UNa) -. a- - 4 ua
t.
4a.~~~
U) U - P - -J .. a.. . ua.. a > .4a t..a
L4UU ~a~in~.- U) UP.4~L da.,c~Ua.'am-a 4 P ~.aa~
cc C * O .a
cc-a : . 4
- * t= -wI.
- - -a, Cna. N .4 a-
a-U as n - a- i C . 2
at at- M U c- > w 0 a-
7a. -Id
170,
a aI Cc a a I tz
wI a ac
# 1 w a 1
aL ui ar IaI a C-
a~~~~~N a a -a I za
ccga
a .1 a. 11 a
ai z
Z.I- N a Z -
a; 2 a1 Q a Ia a Z a ua. ~ Qa
a a a' ~ a
* Ia~
al~ I
a-I a. a a ~ a
a-
a
aL.a
a C
-
a
a~uaL" a1a - aa~
a~~~~~~~
* a L
o
a
z It a
:u~
a a ~ a
a
. C'*
g a.
~a-a
aa
Z.
a
-a
-n
a-
I z
Z a
I n z a.l
U. a-a.- .. q -a
. a aa-a.0--
I&Z z 2" arz i
If -a a aa-~ z ju a'a> :. a wn ma a -0' a -> aa -L . a
a ar
1an
1 a aI. -A :L a - aI a a .- a -a Iava
ap-~ l- aIa
01 1" va-~ ma -i a3L..Ca- 73 M
ra x CL i . a za-aaa . -w a
C.3 .. m . m 0 a I W I zz
aw. a u -a a.
'-a sal a. aUaa. vw ve a-
a'. aI
a~~ - ..
.a -
-r a- a-a al
. ale1
a a a -
1 a
.a~--a
a a
a
a
an ca a
a. a - i
~~a
ay
a laI.,
x~ "1 a ;
6'
-C
a a ~~ a a
.aao aauc
. L.
C~n. . a.
ia
l.7a ~ a~
) -
a LZa aa
a a - a a -
-l aj a a. a a a
* a aD ," at -j a al
VM >
a-~~l aa aa M u.a
-~w ac aa-a- 9- 1-
-~~~~ - a- aa
:z I I I
ccII C z
L t- a
C. a 0-m* x I
ZL > -'A . - * IE
a * oo a
U 4 . I - , I
[IC
"D -0 m <4I ma
U~ * I I 01 L7 I
, C, ix a. an m 4 4 *z
w. 4u.m *.
~
4, ~ ~ O 0,:..4 * 'I 0
N,, ., ~ I- CI Iv aI - .
N: t: N: -t
- ~ - c
- N
1
-
A4
1 N 2t 1 4 W
0
1
I '
1.
7
Of -. - - ~ 0. * . - 'A
,4 * 1
0Z. t- N 1 . - * t .
44~L 4 , *1 . I a..,.
o* N a, .. 0 . Z70
U I'.0Z *C - I U
4 41 , 0.0 - 'A 4N U *N.I 04I7 I
C
- -- C. - '. It It r , 1 V
- 4, - ~ 0 'A - 04, X - - *4I~ .INN I
-t1'
- ).4 4 ' KU - ~ 1.. UI4' 4,
- v 0" - .- 44 , 0. - s--I.x u
U1,
ft _j * * '
o CLIC' Vtj - % t ;j C, 11
N ft I CILW Z S- Z. CLu i
0~~~ *C 40 fl - ~ c C.)N4
I.. 2 dVgw -2 I V
t x~* zLL
P._Z IZV
40
I0
N *_
ft
fft
, M
.12
x IJ
I
O I" Lfu
W.-1
w
WI 2
NLCLtI43
Lux.ui 0
4
_ 1 0 z--
44
In1
C I
AV
* tI 4d W L, t (D 40 NI !!M -" x V.
ftflI4 cvA<Z oI - w IV
x N L Q ft 1 4) V m3 d N 4 ,.. - I - V, V
- *. ',I L 0 xI
L 4
In~L2r. 11
x 1 xaA
.34 -i U
I Iz cI I
Z X C It Q
ft I. u L
Itf 4 ,
cu. I I 2
go I I~ In I 4 c .1
.1 N ft 1431 I" -
X~ -. t 4 cI I2 > 2 I
.. i Zt
a3.i IN le 1 $_ 2. ft w O NIw
.N.. uf t - I *4V1 ft I- 5' *. x4
r C .U11 ft. I - It, fi4 f0 I -
0.4 - - 44J
It I LL. I f I -- N
o 1 1 _J0I
x.fte- R 4 t 4
j -; - Q It 4
u.I- T 4; 4-I
2-I q I -
-Vim 'I ~It'4 L
1 4
4 N -1. 1Ni z
ft ftVI t Z I 7 I _t I
cc 1~4 1I-J-I 4 a01 0-I A. w I VIc
3
41
L Lij
43.1A ~~~~~~U
4L
I
ft
1 w.
I N4
1 4
%
2 1
12
.. f-'fwVI
-r 0W 0* I.
II N
Z
%J
N 'U2
01 c
.
I/ I I Eif
- ft5024-0N
-
Z I N I
'C4
l.U *j.
4417 f~i4
t I I
44-
U I
..
I
~l L
47 7 US I-'V
L . tI C 4U.f4 I 2
47j474
.i ~ I 'U
..
~ 0N.173*
t
I I I c
* aa
04 z a I
i
*
x
a
I
w
[
I x
* I I , 0. - U-a I
II z0 7 r.. I *'L 2 I1
U
Qw ai - lNi c
.4 ai -Q C ~ x
-- - - --- -a a-- a -
i ( . c U
a.L cc. cl4
42. e z a-Ca~ * a.a ~ -It
;a a 49a-
* ~ ~~~~ aof-
-* Iu a -a -
a- 7 2.49 z~
4 7
CZ
>8 -0,i
a N.I 9 lna 1a ..
o. -.Zaaaa
a1
Y - QC9 * aa4a-
-P 2I ia. 4Ca-Q ka a~ a Z7
V
aaa V ~ -a 1.. -dt.1 4a 1a -n .~
-a~~~~~ 7 9844 UL I79 9 I Z
a
U
~5U _jaZ
'4
~ V
V
9l
9
0 Y
aa*
7 1-
a~
4w- r ~ N a
a-
aL
e
*
..
2 t 0.
a4 I-9 . a a ,' . . ~ ln I ..
Lakd2 ~ a.-.
a-C~n
4 .'
*~~~~~~~~~~~~~~ ~ aaa4
L
c
-- av .
-a--
U .2 ~ J - ~ a-
4.9
c
9
aa
2
-
C
i-
1
*
w
~~ ~ ~ . - I
t, * I
I W I
II w .) crI
*a 4
~
Ic I
C, Z
*c c I Dc
*t It -u91< I o- 141 w
cc IC I SUn ~ .
I. a 4
-l 4 14 4e
C A 9 3C M 1 I I
ti
4 -
I.Iu I I I. . . V I Z
I C C 9.41. aI a
* 14;. ir
C 4 II cJ
CS
Oi.
I
I 0T
. U U0 41Il I
=U 91 cW
I' 0 49
u-S I I
I- 4 a.)U4I
4 * ~LUL.II-
I 4 4 lS I U..
~ 4 ~.. I 4175
I
Phase 2
L
L
L
FW
2: v = QZ'0.a.1
"- , t-
e'rI
W 'L 0 44 .
0). 22
nZ O COO OC' -j 44%u -C In 2' 40 x
.000 n r *2: 0% -: 2.W .
U. 0 1- C~~-4l.~
%n
U ,:
< C.
Z
2:U I -U W~ 1-
n 4U- W4U%.co UGO0 U N
~ golV-4SOUN 4l
1 - - W
-V 0
D U Cu
Z
2:
=-Z O n O
u.-
~ 4 Ix -
o . u4-
ccI
aZ-~ I Mw0 1214.I 4-v'
Mr <:. - + 01-D w22 -OM-z-
CL
ICI B wooW*.2X
'Io IDIn C, a 0 Iwv JUDU
C)1 S. . W4 I
4 -1a
Il
*..,2. %L
nJ2.
S., ~ 2 - 1 . w
04
I.
KICIS
oU4a#-ZI-
2 I- 51L 4
0. 0.
i4n
0r
~ 42
C~
0I
. .a.
: ~
I~4IW
W- Z:0W2
2IJ-4?.0
0
UO-U.
v MV9DM W c W 0 du
C40C C. 3 X5 Q X a:
- - - - C - - -. -. -S - - - - - - - - - - - - 7 7
u z 1 Z u WI Jxn IUc
I IIn a 1-
2 _: u.
a- W-31*,uj 2
*~I
ft~~ ~ I
~~~
Ig I
*~~~~~~U n 0I
4
QD
2
.
2t
:
I-L
. 1 1
8.00
0.I
LvI2
I~
0
4 I U. W I- I- 2 2: 2 I -12
U
W W- U.
0Zi
-L
LI
'A 26 (6 a2 L 2
. n 0e 4
W 4
ne I - -00W
5 w.05
1*
2 6e .40 '
44 21 w - .
4 1- I
w74 w2 u CC L W.' 1 ) 0~
111 2: I w4 : LI
1z 6 3 c~!- W1 Z: zU 1.0
. U. 2 O'Z- .. In u Z~ U
*: 4r C, IL U I ' .L
57~I. CU- 0 C ..2: 41-
*OS,~UL ~~~U CU tL
0.
1^.1OJ<
u4 -a
0
>-
WO
0 J-
U 105
I aW 0 I-- 7 a 0 L)
1I) ;- Ml C L' C; T L -: > 7 0 U.
-
U.
C 1 I U iL
z0 'ZI" 2: 0 l0 64 '0-
I~ I-.. 1 2: W a.a I' IW SL LUL eI , l
.1W
- 62 c L c or : 02: I, I:OS2:7
xn I. C - -j c a. 0 I 0
*~~~
U.In-S : a. ~7 7
C-. - - a.7
C, - 2 C- U
U7 I
0 0 ID a 4 C, a 4I 1
77) z77
cc.
N -4
0 I z
VI of-
I. ICE
I - e-Z - 1 -
4. Z+ 2 W
2~~~ OO. oz~ C,33
c
-Y! I I IeN C O.
0~~L -i I I It z4~.
2'~~~~~
4I'T* -- X ~ '
, *LU
1. 0- I- z- cc x 0I
-S
* -- 0
q
'-W ' *i- ;ZZ N ,. - . - C - - 1 O
f ..
N
N*~~~~~~~~~~ U'l.N" 4e U *0 .. I.,. ~ O CU
*~ ~ 2~ ~ ~0
L
0
-T
3
4) I
Z,.e~
- -
Of
u
~ -
z'~.!-C Z
U, 0
l..~ 44 z N0 l 22 c
Iti
~~isZZ3 I4
~ - 1 l x x22 2 ~
Z.J I JI
2' - is-
I
-
'-TI.
n02'~
-
s'
1 7,U.NI4~~
- -
w~e
-
''
~ .
.,
.
'
ee
3
j..i
..J**,*J a4 aI.4
~ *.J~
.
. U 4
X C4
0~~~~~~~~~~~~~
O . O O i -2 It0~ L .4 . .
.4 4 ~ ~~ J J ~ I ~
J . ~ L ~ ~1 ~.~ III ~ 111
Q 1 0Z. 1 4 1
1784.
11II
cc I Al cc I I 4
* a aa * a a a a a *
*~t a .u aI I <
*~~~~~ x I
* a 4 a
, a * a a a . a .cc
01 1 ai 1- 4 I I I I I x gc
c I
'OL24
44a ~ *~ a ~a ~*-4 ~4at a.~ -
~a a o I-~*
-
a-~
aa a ' ~ *4.~.-It? Z
..
)a
*4..a.~" .j ~ aK 4
)-a a .. 47a'a a a.
~
~~~~~N~ ~
"~4 a aOaW 4,aW a - ,aoa 4~aa. .~I
4 a 4-a - a a4l-I 1 a4 0- X
a a
O'~~~~~~~~~~~~~4~4 + 4-w " 1aa~
4 'aa a -a 41ZaU-
.. 4a X
*~~~~~~a~l
aa 0
a
awN
~~
1a- 10 .1za a
4a
.a
a1 Nj
4 .
.w
a W~-
1a4
K;~~~
to
0
4"1
0
W2.1
*
4wa
2-
a
-C..-Ia
a.
aa-a
WIz
.-.
In
a
aa
~~~~~~~~
4
aI
a
*
a.a.
:j
a ~I
.- al-.-~a-a
0
a .
cc
a
a
aa
a
ala
a
a
a
a4
a
a .
-0a
-4
z
t..0a.a~.-0
1a
z~w4
~0a~.-.a0
Eaaa
Iu*. 4 a
1
a
2
.0-ZC
a
~
~ c.~.~fIO.W
-
4
*
1
z
04~Z-0-
a
0
a
a
a-
c
at
-
+UI
a
a
a
a
Z
a4
2
'uQ
+
a
an
-~C a,
a a a
_j *t a I
-a
a 4 .aJ aLor
0 0 m . IJ
a u Qa 0 U
ata aj L
- - Z a179
I 0x -
uo 0. a a-C 4 a
-of i -aw
I a
CL=IC a
* z cca CLa
ki
7-. . -a i 74i u-~
ki Z I. U. u0
-A 4
.O~a- -Uj..-
- -- - - -" - - -7 7
*W *.0. 0. . - I I U I u
~ 7a
z ~ -~a
z
10. 0.
I~ T ' 10
0It
is:
cc
w~ .Ia.
.iC4 7
IIa
:.j
-
1
1a
a
~ 1
.. a.
0 *.i
- ODa.~-
l
a a
I~l.. C~0
-
Z1tfla
* Z
I
IJJ4J.In
%~..J
ac 1
01-.10
~..JOIa 0J
I-.
-*
0.
~3 .
*
U.....1-.
.. 0. 0 I
U
Z...0.
In -OI 4.a4A10 S 0 401 -4 -> ' -~L C).a.Ji L l
~ ~~~~~~~~a'%NZa.t. +1
-- 7~~~~-Z a *~.w z o ~ a w -
* a-1 a- -0 Caz
.c 7, r 11 auz z~ Tl'-0 -- -
C0 C- I
: . ~
a -
aVolII
pcc
I" C, -, 0.8u
ZU U U
C,
.. _
- C,
C ,
- 71 -- - ------------
0. 0. I-C LC
alt ~ o - . ~ ~
I In2
C .LL 0
.14~
.0
2 2
040.
1-
W
0
. 0
0,
U
c.rs
C.
C,1
N W 4.
t.,
'At~0
,U
01't
cc IC N.4
II. I
I II.4
M I C
0 UUC.
*.. C I **
91-9
0 ) I
94
I
I
... I 11..UI
~ ,
I*'
~ ~ n~t* 1
- 10 I7..1 1- 0
U , 9U ,
C ~ Z9 9 9 * ~ U - N ~ -of
aC.
4-z
L17
a. ILu 0
x =a -
0W 12CW
cU, 01 a-.-I
4 - . a-a w a -0 It - It
a.Is4- a I,
C, _j -, cc 0 >.
:0:1 1
a- a a; 1 * 'V :~ * Z,-
i-
X.>-'
1a- as
1a 0 I W ca- ) ,a'- a
4t Va). I~ i 4 . j~ saId al--
~adaa~..
a- a-J
.s~C 4 ad 4 a> a~jaaU 4 a
3 ~ -a..
x' 1
.~
aJI~~a-N7~
*g -ax Kw Wa
-a O -~ a Zj..-
a- a" us )' --- O >-.a
*.On74a,-
-
aa~~
n' ~ C .aKa-
Z
*.a
a.~~ *Wa---i kn
2Ka--., - 11
>Z..uWx x >-.
aa a -~
a ) 'C:
.4a-. aaa. Ca.
,
5
a a~a . a .aa
a aa N a a as
a~
a a aa a
~~~ a a
I
- a ~~~ ~ ~ aI aa a.1
-~~~i aa a a
Lt,
0 a a ao
'a. a aa a.aa a a
a,
aa~~W
a aa
a a a
a a a I) .a >-.a~.,
N
Ct a V N
ki a a a a
z~~~~~~,j C a a aa ~4 -u 1 ~
N~-j o
a- * a aa a a a a a a-sa182
II
III
ax a a a * I :
fta e u It
a I S 6f
I 'A 4 I ft I I It a I
it I z CC aC a I I a ft a a
ft a, aaf
I7 a af
Ita..- 7 Iaf ft a a a
ft~~~~~~I t a a af Iz~a
fttn i ao
%,-w a4
Itz 0
ft aI
a
w u
~ Wa ad
.-
t
t
a *
a aaJ af
aa
f t ao a I
I za - _i a - f a c I
I a f
fta a~ft 31ao a a -
cr -
ft-aw-~~~~1
a
a 1 * .
.Iw,* aa
-, -a
- I 2f
& -f
Caa
ft~aaZ- a -aL. aa
I ;: t
ft awufwaa a -a -a 4 a-4am ta
a
It X 0 a Y. 9-Z ft -a a -- da
-J - a0 Ia of a
aa
ftof~~a ajj1
co I 80
ftw a 8. j m I
L i14-
ftra- a 3z za ft-l-=a1
a o a
ft-aI a 0Cc
--.
~ aIItaf 4 1a
-1
u u a Q
ta
11waa
a
I.xt
taca.n-a 4 u
ft
fttaz ac- a-z W a:
ftWV -'J0 -E a z z ft- a a *
0 z.Va.-a.
-f~o8a4
Iua- a o -f~~8a4 >(J
a ~~
aa~ ~~ t a a af
a, a a Nl ft I ~fs
It a a'' a a<
ft 18a~~3 ~
Lii
cci i Ix,>
9 I ~u. w .~ >1,~ .4
-. ~~ cc v -4.
7w Q "
0 l g U. D - 4( 0
~7 4
A~ x~ a4 Z <U'.J. 34 -
I n.- 7Q~ . 4ZU
-~L A
CL 0. x 0.f ) 4 -. Z z3>
3~ - 0 - .w * 0
0- 3-.44
< L
(n 06 zs -j z
*XC.43CO z.u- wo w~ 7 7
o~O . . . .>-. u 0X m I3
9..I 130L3 3 %
4
- - - - .0- -> - . -3 -3 -
-7 -34- 3
0 - ~ . -0 3 ** * *4.....3 . 9 s W. 0 IZA*It
-0~ ~~~~
...7 3 ~
3* ~ ..
00)4
. . .
1 10,
.7 3. .
7
4
I-4
3
.
Z ~ .3 4 3-
.
4.
0
3I
~ ~~ ~"I-~It ~ ~ ~j 0~ ~ *~~~~~L
.4743
-4
I
4 -,
37
4
K. 43K-.04
3 9
3 .L
- 0
300
- 4 -
j
0
33*
33I
0
oo 1.0 0
0 -
CL.-9.3 3- - 3- 7 ~ 1
6 A3 7 3. 7 X3. 0 > 1
Z 0 + + +
- *n IN493
*: !, ' ;z 1
943m 0 I I x w
9 3 7 I W (3 I m
11 u vv IS33 vN3-4
cc Cf I Iu
0. aU Ln.. b
-~ -. *. I I LL.
CL -fl * Z -J Z 2
- ...... 1 3 ~ ! In4
4 Z o n 70
us- : 0.
* aN w be * 1 .
0.~77 0 .0 4u 1O >
w C) 0. X. 0. W. I- N atOf.Iwno .5 I X 0
W CL .. 0- us a-Wo -Iu...
- , '.. I * -I : sz.. I
zx LL I.z 0j In I I-Ioo0~
-o0- - W z *
)CU.
II w0 *i Z1.*
4 - W.
a 0 PI C- a
a I4 - 14 i I
*u I
a IcN ~I
- ~ ~
aLL0 ~ - W
If m -C _j . I
I 55. .i S W
2
It% -j _jw.J
m a CL 4-I 14
L 4 4T eN cJ w1.g
U0. I x I
I I zu.
, co 3
I -a-
*0L ~1 T cf w I 0 no C
A 0 WI -5 a acsg Io'I-0 <
x I IL * 4 I-U..JI. d
a 0. 1 -C))
#~I
- 14 11 w 5 0 1 I . 14Qt.~0 a Cs. W
a, >x
0.* , ow01 1, C. 0, MD. 05 a,I I
U. Lu t" 41 0 0 0 U I>VC I
WXWW4 j .- j I)~j 4 * . Z .1..I
I'w 7 4
* ~ ~ ~ t u~ u u*~I~'
1~
5*!
I I
Phase 3
cc ~ C,(
U, U
ci
Lu 1(2 ~
c- c U. orU... <un1( .1
a. . - ~~~c
- I 0 0 ucc >
Zrf
>I
( < 2.. I u.-Z- I ..
7Ici
CC u2.. ;,. C C 7 9
.- c-*-~~~~~~~~~~
i c - CU U 7
Tc UUC I U .. *-. .I C
* IC,
*. . C-
c C< C4)
I 0.. C C w ci. - - 7a
*I
* ~~~~~~
~ ~
CL
L0
t.U)CX0
. . 0 CCC
-C .
)?
0
C.
C C
- cA
u
1.) 7C- a0 Q
Lu) U. u CC.O I-7 U.
I C3
ox4
-ofI t.
U. w u ol -- 3-a
I- 4Z --
co' . .. J .. .4
1 z
0M ~
.4
~
-- -C
d.0.
w
c12I
W1
1 C.-.
w x Ca.-
.J44V
CL-.
o +.If 1
7
If1 Q.. w
zoo
12
x4
. - CC 4
c0
70 "1.
.. W
-
. -oC.. a--
ZW
I-
cc % .
U..
~ 4
-o .1
C
U-
cc
Q u4 (4 2, 7<_f>'1
-1 -j -'.C j
~
7 1. 0L. U t Z* C..
-I"-'-
U4
-~
2 ~ ..--.
(3 -
~Kj It ~t.C t ~2 7Z~~f~ !
eIU'-.4 12 a-
4 InN.. 4 0. a* *) *N~to
Li
4 04 U. Uto
C, l' % 1 '-I -.
" 0 120 tW t - 4
I I 4o
1 1, toU 7
"O I -Aa1
voI x V .1 -a W .
w-* O ) 2U -a- o
* I UJ a')189
'.10
t12
U, .4
CL a .1 W. a
* I * I z 0 - .
I tI-1 w - I -
T C. Tc - U
>~c . a o a.
4
* 1 C, C.... - cc I z 1
X*
.s yl 1.
6 .4 C2 m
In n I' Z . I 4Z _ 0.2.
- ,C :5- ,~ . .CI-- . C
* --
-~~~~~~
..
lZ*~?..
. c.II
. )
Z.
~
.
-
.
!4 W0
4 ~r
-._j
C*( 1
<
>)
*.44.1 . .. *..Ir
4~~~~~InI~~~C
1
-
~0~ IL
. .< .(* n
-
~ 0
4
~
I* (C ~
I ~~
~
~~..I-
I<'WI
0 1~C
a
I
4
a
U.~7
*C".jIn.
C.
- 01
) ,
I.
~ IC 14 .* 4
*InZ.0:~~~~~~~ouO I 12 * V ~ 1 . - I. t .
Qv-'
- - - - - - - - *I- .
----.. 4 - - - - - - - - - - - - - - - - - - -
__ I.
>0. - 10
*r IL
c - tn 0
* l I
I* ,*- z 1 Z I
I z-03 I
* I uI1 0 c I+>
c* It~~~I
*W
i0,. a 0 %C
ICIO MN . W. ax
4x0 le3U - rn I-.
*~~ ~ ~~~~ N ... W- x
410
lu O Il 3L nf L;
of1
4>~~~~I a IaI * .. 3I->. I- 4 I *4 1
I JI..I-f 7I - 0cWC'.
0. ~
~
- ~ ~ w -I-UI U~- Q- Q ~'.-
NO I.. --.-~ - -A.. - -. -~ - -. - 4 ~.. *
4 ~~ ~~ ~ ~I ~~~t
7. -3-~ *0. I I
4- 07 C1(f0 I- IL.uu CL cc
I .4 ; !: - .4 10 IA
I~~a.
I~IL4lI-2'-a
.6 40;r ua - I 6.e -4 . > e ~ -- g
tIC*3 cc
NN n
2
4IIA0 cc C IaI >IU C
~ - :~:-w::1Ia~L
0 - * a a . . a
~ 44.41C1->
LL:. u* -2-I -'' - - a O ~
-. A . I .~ r ) IX- U- . ' . . 3 4
C,
41- "- . . 0w.0 ,- U - C . . cc40
N ' u
40~11 ~ .L ~ . u Y.
-a.
~ ~ ~~ ~ ~W.'al~ ~ ~ ~~ ~CC ~ ~
~ SM I *
~ ~
Li-
44- 40coa I C - 44. .. S I
44 03L" 1 - IA >I0.
AUS C() . 4 4
L0 I I C . . 4C I Z 40, * -,0 ~ II '. . IA A 4 a U -
u~
0 wa -w~
: : Z 1
9 5 1
,~C
O
~
C:!~CC
A
u
S -
*
-
.3 *
-
1 -
1
~
4-~W~4
I A
0 - .
*
I-
~ I I
C
(~_4IAS0
. a - N
*- ar
~
>>y
S i
IA
-
7-* p~c 53-3 U.I IN 7 U,34 a~~
0 .a A .4 U j .Y1
: ceT~
7 .. UC0..
~c1
U
4 . >> >
* I SeI
U U USJ U U et.
cc .~a *'
IA 1114 14 ON
7 4 I~L T I.'CJ
*a ~ Ia~a~-. I...9 C
ja I
ILi
z & - :- -
- 0 S
14 1. N
- -o - - - -N
C,
IrW I3-
3.)
w. 1
.0.-~ 1. >1 Z 0. 7 'A e
cc I
7 00 In
CLL
Is .
o~ ~~ - ~.
cu 4 0
L'.
C
3.. C 0(3J
U 2 7
N
3.)
-*~~~I
4 -
~-
cc -0
.
- .. ;,-
1 - - Z 3.2
3..
*J
ujn
NOI~~~~r I-
.--
- .. CZ ~ 3
'14 c
3
N3..
- - .4- 33 - 0 2. ~ 0 .. - -
V Vu
0NZ-~-
192:
- 3N
C-: V,
[1;
x .- U)t I U I w I
w4 4 1 a..xUa
I i -cc I
.- a13a
.n a- c . -
WC W .I* I -Ul a I m a-I
it. w I Z aw-. - I
0.
W
a->
Cc
In.
4
L7
4 -
x
'A.a1aa
.* ~
.- IC
:.a
.)
I"Z
,a
~aa -
uj u- a- - . I
I ? 11a ~.
L7 - aa u._- -- x 4- 'a a.
4~~~ -o
a/7C .M * - -.- a-'a I
4 Ct * I- LL a U. q. ~
I-~ I- f aa~?~
1- -
~a- a- a-a
a------ n c~ -
--- a, a a. u~ - a a
--------
~~~~~~~~a.~ c 4naaan~
aZa .a I.~ aa -
U.~
- u~~~~~~~~~~
7- 4. aa aaC aa 13a0
'U a a -a
L - Laa,.)-a..a~n
aVa aI . - . a U aC
w)- Va>0 - j Ua~l U. u. *4 a. *.I- U 4CIL In-
- J a--a a -.. a - a . a-
aaa-lnau*o
&. a- .aa ~ 4..~a - .. a..a C* 04a.- a a
a0!
N.-
za0
'A
a- .* 7 4 ~ *
.lia
.a ~ .-. *
aa4a . > Ia..
C I aa
a Iaa--? Oa~
x r, nC0
I-a
a nN
a-----
U . .a'
.-
- ... IU ~a taa~
a I-.
a a
.U.L a'
Ti
~1.14 az j.7a
U -,a nn
aa-.
-C
4
.Q3 4C'a
c a :' -. C 14
a .J-..aaaa a-a11.a- .I
11 a 3.
I U aU U.
u~M~,~i~~Z i .>n a 44 U 4 aL- -a
>e I .
u n~ .> -~ > - a,~a- 4 aa3 a-a 21 .)~ .
ar a Va a
t a a 4 .
CY0 z a aa
cc a- a) -
r. a -aua
C, 1a a
U.aa-.a 4z
C.~~~
--aa.J~n 4 a . 4
13 a a~a-.a
a~~ ~~ uNa UQU u 4 ao a
I Cl U >.
I*~ 0 * -
.. <CcI
44 .J I .t *2..1.4
-. 0
444 I 4 4-V oC
28IC "
*C I 1-71 -. .IWo-
44 ~~ ~ ~ ~ ~ ~ 727I~ ~ aD ~ I 2l4~4I4
4 I U
4.I
04
~ I
i
i141g
I
* ~~ ~.- ~~f
4: 44 0I
*~44I.... 4
.04.
*4
- 4
'I7~.4-..-Iet
4 4 I
1
*
.. 2D
4~4n44 340 I 0 -4 - I " I 4 ?
44 I -'I*I 4 l4.el4'4Icc
4~ Ij4'~l OQ - 4. -. 44.I
~l.4l44.
-. I Li 4C~4l..4OI ~ :cc
co I 4' l I~*. I "' 4'4
4 *Se-~I~4 4
I-Sle *-I 4-
cc 20,T3l4..5 I42.
c I c
4 - * Ccii
4"~5 ~ 34.. 1~ N .4 5
* .4.4
"' 444.Jl
0 *4' ~ 5 4.4 V ~ ) I~l~.?..4L
m - I I I
IL -W 'C
IZ I 4-o
0-1 C; . . I
1K I N +
C, 4L I, "
1J 1 r V *
W 0 10. I - - 2-
40 t
%.I 0 I:
o I I Ic I 00 *I, I K
* W II. I4 WO0 In
N h kK II Ix o 1
a I ON In -l x
*~~~~ I*1K 4 1 -
CCW Vl 4V I. I- C I 00 I * 0 . I
w -j -j * II Im
Go~
2KQ: ~~ WC CC .z t * I
: .. -KI I I
I of I I
iou C *
* I
I ~. I-rc It
w
4c a I-
I I
I
I C.
-%J u N vuu-
- -- - - - - - - - - a ---
I U, e V
J? I ~C Nu
- I
c II* r
c- VI UN
0
I1 -7
21 C. C
04 IS w C. ft W
L I O, f* f. :S
C C3 C C C - 1
A IfI (I K Lir
I -. I ~
- Ia * -~ .
I V) . ... Z. K. - W- 195 I
4 1 Q to 9 9
4~~~ 9 I
cc I I-. 'y
<0 I cc I C-
9s z c9 2 Z' 9
99- . e to 4 I
N 4 99 9 9 9
C~~
-L ~ I x93
4 ~ 9
9
9
94 0. 9
9
c C m - IC cc
W- It-9. 1 641Iz1
-~ ~~~
. ~ ~~~~ ~to 9U
l.9
1.
I ~ C94
,:9to9 W
99
I ~ C
97~~~~- N, ~9 >97 - 9 9
C I . CN9 a I. y
9 0 -e N > . j oz N49~ 9.~
-
C. 9 U . 9 4 I. I
MM- . 0 - .. l '9 .
N~Q 9 W 9 N .k 9 9 7N 9 j
I CC
Ix
It
I UZ - 4Lii P I I
U, )
L4X W
* 1--of Z Z
ft - -oU. x
ft IKX L
ft1 2 WW +Z
gis. .9 I- W
I)wt IU a) I I
0 -- .N - Z ccI
11 U!U14 - A Iu I u
N " o-
CN'4. g . c
'A I li UJ-J -0 LL
ft...S.Z~.Jl w - S
W t- s-s ei ni
*4 .Wo~. IL 4. It5
t.-iS *W0 kn C,1I -1a11. -t
*I10 j.j'0W1 0 f iI
D 99 c.I W 4 I ~ I -a -- Z . U L -L
*7 0'~ 0s v~ Q. u V -------
N'10.
*Z-~I
ft~ITW CjI110
4Z *N- * S-) *
I C. C.
It
40. C ~~~ V - ~ 4 ~ ~ o 1.C. - 1 - *..
II~~~~~~~I U I N-- I uu - I N
~ftSW) ~ .- ~~ O- ~~~~~~~~~~
-4
-~ W
.S " 1
I I ~~ 0
Li
Z I ~ *
* *~5
*~ -~
Li 5*N- 0,
xt4
* .~s.-W
S IAO.j -~ .~O i- Ui I - * .
a e.c .I a .4 O f C 0 1 at L I?
u ft ~ Oi- O -A
NL La NI
If C, t"I r 01 1 j7> S1
I~e c
* U
N IC
1975
L
-
N
L
1.
"-N.
-J
o
I 7
-
F
a -
4
N -
- 7
0
0
- ~
X 4
- co,~
-a, -. -o
- 4
-I-
o
'a
~
7
~--
~
N
NO-
.l~
-~
~
~
N
L
C*- ~C~0 1-0
0~ Cs'.. w~
c~w
0 0 'a
* I I
e - N l'~~ I
N N l~ I I
* I.. I
4 30 I
* ma I I
* I I I
* 64 I
.-. 4 I I
N l~ I 9
IA* * I-.l I
~N * I Al I
0~.. * lau'l I
'U. 4 l~Q.l I
- * I *AI I
I
.1 4 lW~l I
-,
N
(.~ a
T~
.)J
~.0
w
4 -
4I l'(0I
*Xl-~.,,l
* ,--~m
Al I
I
L
a ~ ~ * -m~ in-. -~
a, 4..Jl .s". *...
0 0 4 a~ 4 ~ I -~ '! I -
- I- -r 'a 4 I *a I
g
a
* C.)
- -~ 4 ~lN~I~
4am aI I - - -
a~ .... 3 + NI. 'lla,,'l 0 N
4 l.)C I I I. * I 0
L
a, a, N
-. ..l~ -a * *I '-1'- I - - - --
0 a, - - a~ C.. a. 4 7 I ~A V I ... I * N C )N
N - ca C C.~. 4-la I * I - - CI I -- -
'-I a -- *Q~
"'C *aI..J-I'1
N>I>Xl- -I ~-------- I + -. -a II _ ''~
*lC(I~.''I * 0~ yN
a a' -l N4 ?~u.
- 'C 4a.IAl..I~'..I N
'.*1 N 0 4~3 43 ~I N.' 7 *+ -l N N *4 7.. 'a
a-va l~ N~ (Na,~ *VI~.~ I I - 0 - a
A 0 a, 4 I? l I
C~4C ~'l7 - w -- ~ 4a,3~'7I-' I * ~ .- 'c. 11-,, I- I- N -1- 4
- 'N N *~ N7IlI.~.e.I - 'C... N. N -
-" N '~Nl-. ~e 71a. -N:: 4''I al $43 N' ~~1.'' 'NI'
V" 0 0 LO .'. UJ~ e~ 403 ~l4.I a, II 0 l-.7 -- a, U -
"7 -0 '') 'D-! N'IC,~I ,mI ''
U~-'W 0 fl ao '-7 -. 1- ' N~lv'~lrIk I N'I-. -z r nuu-. r
CI-
oz 11
.~f Z CN~I'-'l'~~l
*~INA4l~,~I------~ 0 ~I'Z C
0 0 --
".C >
a 4 0'-'-~.
)IA-I<' r
0 KU. YU.ail,.,~ l.(1 0 '0
C ~C- l~ZN~l~Ctl'a!I o.~C.
4 I I
4 N I I I
'I ~ 0 4 I I I N'J ~
C C. N I I I
N - I I
I I
0
0 0 0 0 0 0 00'..,0 0000 IN 0 000
fri"
198
-N,
44
--N 'I
b.I
I
IV
it
~
I ~I!
ILl
1K
rrt 1?
t It
3j
L3I
3 le
3 3
3 3;
17
3 3u
U. co cc I. .
I3nx -1 4w -C
3U - 3 IL 3 - N L 3-
-3 4 '. 4 1 *xx
1
-3~ - -J -3 - - - -. 0
3t 3* U)
03 3 1 N U - ,
3u I -
Ng.-g
W
I. ~ *0 * 4
-j 0f
4
CO
.N
-4 13.
.3m
3A
Z*C .'A Z. Z3N
0> . 3 I cc - C3. ZN . 4 fC3. 4 Z C . 4U .uN 3'I. ~ 3~
3~~~~~~~~~3~~r . 4)( 4 4 ' O .- C f I m.
- z~~f
I.
343~C
U -N.
*- t, 03 * Z 3
* i Y z3 0 c . r
k w Z3 v 7. -
.4 03 r 0 Nl
r3 I LJ
-7
Ic ~3 I)W>3
I 8
*- U3 Li Q. 0 Ni V
N~44I
3 * 40U U.t~~3 ...b *200 -
10 UZ I I
'A I I I
U, I 0 0 1t o
o t I " I
I wWc c
1 wft 2
U~ I Z *
-_ I *-1 I.
_. x39I ct I. - 2_
'A Xt0 a W!
i . :~Z_ IWOw
ft.IJ
0A -A Ij lea
.
w
W- ft,
I" uI i. + CL
z f.0 2I - - .4IO
at ftIt
O-* 0.
IC Ex*.
U1 .*Z ~ 8 IC*
**ftI JlCL I*4 T *L-N~
~ o.ZSr4
4~~~~~~~~ W40f~WI(~f t iW .
IunX U
0. 1 I- w T_ I - 1
rj)0 t
a.0
N Lt
Z0 1'
NC
f
I
ft
x
ft I
ft -- c .n-. I
ft I :f
~_IIIU i- z
ft..I .of1- I w
x W I!;- ' -cej44
ftU110
4 00
1 -
ft1 IA. J & 01 be J , 0
ca l . c _. 0 . -4 . ..
m or,- NI 0- -- .
... .. TI x p9z
ft < 2I - I ,( %) w
_0 Ij N Qt
fWt M
I Q u
ft I owu
I I01I
I'CI
4 2 . - 4 S I I
*~ 3. " 3 W
CL
3- I - - . IU
-z Wc 3? 0 ce0
IL Ii "rI
*x 3j 03. cc 30 43 w
ou o 3 ow u .1,1 uu u
3 3 4 13 I~2
-. l '
3343 -338.i It
4 3.3 .. - II> CL =0 nz If
~Z
Li 03
CiW 3 CL-rlv0L
-o p
* 3 3 4- 3
* j Q
3 u
* 3 3 3#43202
' i
l
Iu IL
I. U)E
0
a. I zj
41 z IL C0 Z
CC
00 c 2
w .U> CC(%
IL !: C ) CC w )
- - to - - --
I~ 0' 0 m
* - - - - --- - - - - - - - - - - -
e I CC I
~ I - 'i~ -. L i
.1 I -4- 4 2 A.' . .L
I
CL1 CC U. z.N i
orI I 4 2.91 01
%
Il-I U. Vi U) x C. 1)-
,-,-i 9., .*NU A Ii Li I z
~N I -4 -j 2. - 4 w U0
aI I Ci Vi CLiN- Ui Ic 'A
C"5.2. U) 0 .4 x- - - -4 11 W-U
iCIU) Vi a, IU. Q~.U .I tv
4~~~~~III~~~~~ . . 2 - N ) U)
%iUU w. L V. i U N W. - *
* I I
9. 41 1'~
9~ I ~ WO.
* 1 I Iz~w a . a
- -- J XI
---- ' --
9i 9 Wg 1.1'~.
za , - 0 I
OL
I UI Cc
70 11U.,L OI.J IlI
Q L
I' IUI I-
U.W
21~01 I tkd I
0 - u 0 L) 3 6
IIA17ccIZ
IV*J V I.
11 - U I~
w
1 .1 eU I
le- 'A ;I 1 7
eI.I I
U* I, 0
I4
9.304 1u.Z!
w- l
I - ~ I - 9 I*J
I j 9.IJ
90105
I-Z
Ui
0 11
zm U. 41 1
0.~~C
U -~ 4 I-
~
a,-*I
I
x- 0-L0 I~
*c I~
- .0a I I
'A 1 Ib
IL Ian wm I
X*2 4
WON so
14 1,
3
ga
::-. " P.3
- * LI -. 0 a a mL ~ ~ l
.4-~~ ~~ 0 . ;i
U ~~~~ 4-4--II M )I II
U --
0 C- 0 0 4 ; 0 ,I 44 - L
C; U 4 A 4 I U
-
I--94 0 U
C) C,
.J~~~~~~~~%
'.*4I
4 .0-n "w
. U d
'. I
.
tuL 9-
.J..I1
W I
~. 0
0 .>
~~ ~.
- 44 *~~ 4-.J I, 4 - U .~a-0.Z
' *aC.I 1 5
IL
-
N..
U
-~
4
. 4-I....
9
-
- 4.
-U
~~~
IIL
-
01
C
-I
-
II
49-XVIL
40 .
44 N
-
L
W
U
.4~a
3
U
~ 4-
*V IW I
C
11.
Is,~~~~~~~e I. 1 X U1 4 - U N- m*- 4j I -~
U
NW . U 44 . -' ~ I~ 9 4 21 - N1f
*4.
N4- 4~ 4 4-1 ~ 4.0O ..IAI 4U2061
I
:1 1 1 '14
III
4 1 4X~-6 I 4 3r
I~c *1 4 I4
44 -CL j 4 I.
W o c * I
* I ~~~1 LIX 13 1-
** xI %w I
I I W 19 IL Z 1 1
* I LI ~U4 0 - I
0
ii-,
* ?
It
I; I
I
a
I&
iGi!i~
4
c
f
42.1 Ili4
0
4 I, a I
10
xI r 2
IC 4
)
'a
W.I
X L
-~
a.1
,I
%J
I..
I
a!I"
4u
1-I4
U
0
II
I*U I W D1
I
I
1
9
0
0 -
-
L
Li
0 Q.
4
*WIE. I
*LII4
*-aia
~
L11M m
4Iu 04 0
10
kIIiaw
~ a at
a~--
N.g.
i
w
-
zL..
I UI I
*
L~
I - I
Ou
LiuQl.
-
* *WIW
u w 0 u.i u. w Ili 10 Q u0.u u0 *I uI u-
------ - --------- 1 4 64
-- C6 'I .-- --------.. II
*I* 04 41-I
0 II.~ . ~ *1- .- P4I
~~*ff.I4UwIW~jI
N..o..
0~4-L
N=-i.LO . :0I L
4
3cc 4 0 I I 4 I
II~~W 'A w
tL CL ft 0z ci 1- -
W 2 . ... 0 .
> * R-
v I I-W>
U. H 01 ~2
w~ x I, Io- 0-
Wt 0 L) --
0 V, Ap( 0
2C ~ a -- -- a -- -a. - - -207-
1112
4 ci I I
I I
0 IL,
'9 i I
m I.. exI
I!
The following pages contain deck set-ups and the output of the resulting
runs for a sample problem. The runs were made on an IBM 360/91 at the Stanford
Linear Accelerator Center. All the cards with "//" or "/*" in columns 1-2 are
OS/360 job control language cards, and do not change from problem to problem.
I Thus the actual required input is rather small. For information about the in-
put parameters, see the comments at the beginning of each of the program modules.
Extensive use is made of the namelist fEature of the IBM Fortran language to
Object modules for Phase 1, Phase 2, and Phase 3 are contained in the data
2 2
u = x + y on x = O0,l, 0 < y<,
y=
yO'l, 0< x<l1.
The solution to this problem is u = x2 + y2
The first run solves the problem using piecewise quadratics (element 2-6),
and the second run uses piecewise cubics (element 3-4). In both cases the
The final two pages of this Appendix contain a deck set-up for an eigen-
value problem. The deck listed is the one used to produce the quintic entry
in Table 5.2.2.
209
Deck Set-up for Sample Problemi.
210,
17.r-C
a a.
. in W) In in W/2 1/
o 0 0 0 0
j U U UU
C. C',
I.-/ In n
C., C.0
0 , 0 16J
cn l'- C, I-.
LI U-S
I
wj
I-
ki
m
u
0 0)
o ~~ 0 4
- lot ) a
-Ju
C- 4 4
.- .. 0
0. it A In in
in 0
LUU. u 'W- W W
C-~ C00
tA
V
0. NI In ./2 tip In
A0: 1
1/1 to. u.' / ~
:: 7:
.0
j It 0.
21
IL
Ii
1
C..
.11
* a
LI V~ Vi Vi Vi
I-. ci 0
IL 0. Z ~ Z
- C 0 0
Vi - Li Li Li
LI Vi Vi Vi
f~ 0
4- IIi
4 0 - N
a * *
C,
4.~ C 0~
ViVi U. ..
Vi
U. II .L U.
~T 4- 4- 4-
* 4 4 4
'0 .J ..l ..J
o .) ~
4-
U.
Z r
* 4
Li .
4Li
'.)
4
Li
Li
1
I. -
I
ci
Vi 1-
4 4
Z Li
C.. (C LI
C Z * *
U. - Vi Vi Vi
*
-
a
-a
Vi
ci
Vi
I
I
* .- Vi O~ .t ~
* 0. Ci
.2' I-
*.
IL
I
Li
.1.
-
Vi
C
Vi
~
**-
4Li
- Li
~4
-J
Vi
U
Vi
U
I
*
E
-. C., - - ).4 - -
~ C.' C' C.- .. it . 4-
- Ci~ CLI
- CI.
.44
-~ '~ ~.
-'-1
~
II -
4
-
-I
Li
~i U J.J.
.J -, - - 4-
- - 4
- I/i 2 Ii I.- iLl -
1. ). 0.0..IJ .~ .I!C ~ W
'is .C
~..A '~
0
.~
Li
- ~
212
~I
z z z x z Z
* 0 0 0 0 0 c 0 0
*N ) Wi U. L. U. UL. U. U.
V) V) Sfl
Wn SA UnSnS
in ID -
on V%
4nC
w '
14 W
Li ~z z V.Z
S SS vi S
Li S
V) n
In SN Sn l Sn
0. - 0 Sn In 0 . 0~
-1 A z 0 0 0.
e. it- . -
I n
j U,
4n-A
U.
5
US. L- I-2 I I Ln In
LL wn U. U- W 'AZu *i
3 55 0 * W VI
/5~L 0 n0 n
5'-k
- . ) i
v. 14O
- CS.C, C' 7 c 2 V J.
1.'.~u 55 i * iW Z 0 z
-j0 2 0U. 'n -, z 2
z :-2
cc,_O 0 -0 Sn7 W.10
4 0 L N4 . '"S3
45- m0 -T z r A 0. 0n tkU L Lu Li U
w.N"5,,a.a.a ji 2 4 I- & US'UW ' Sn
A 0j
3 CL U 2. I. WU
~-3Q
J. .. JSJo :0 Sn
'A L:'a.: 0 J10 An 0
2135
c- C. C C ,c C-
c-
C.. Ccoc. C. C
, CC, c-C.- C C
CC,.
r
(C. cC. ,C. C CC-
CC.
c.. C. C. C C.CQ.C.
C C C. c.C C C C,
r- cC, GC
c c -c-
0 00 eC,C cooa
C50. m mCmm
f 6 LC
w4 , c-cc.
6. W.c . V.c.c
C cC - CcC- cGC. 0
C, C,Cc. c-C,. r- C, wt
C, C. c- 0C.- C c- C.
N~U N N- 4'
z' 14 U
In-o
wWL.* WWW
6L U
m D. 0> m D -)
ao- C,0
T C II CorIII I00
0 ---0 C -- ,C' - C
.bbIt- C C (d C I. -C - - C-
C, Cr
c'c-o n c c IA ~,U
Q0 Q-n
o C V f
4 4
- 0-A A0
INCO4'(. ~ A
I
-. ~
ne'
. 3
NA~ .
03Z
LL .0
ILN 4
C u - o 0w .VC
' 0
C. Cd) t :Af -
-
~~~~3~
~ CJC..O
~ ~. .~ ~C.C,-CC4CU
CCC
C rC. L,C
CCC.C
'4
C.CC
. 0
C-~c" A..CC.CC(f.C.tcC. C IN 0 XC,
9C.C. 0" C. C, C. W CC 1C' 0 C ) C
f JC.A
pr CC I C,C.
0j C' I- -
A
C -~C (-C' LC- C Q C -c
0 (C C'NJC 0 -
c C. . 4 C,
C C-Cr ,C, -c : i
.4
... C. C .. L..CA Cn OC C C ' (0.4 I
.4 .
'.
~~~~~
.4~4.
j
-3 3, U.JU
3. 3. 34 .~. 3. 3. 1 44CC C
.4 44~4
.~ f44 444. 444' 44 2144IL
These cards and input solve the sample problem
using piece-wise cubics (element 3-l4). The output
fromr this run appears on the following pages.
II/JAGX~xTST JOB 'U.AG$CG%,54,CtASS=E,RECiONt=300K
//STP1 EXEC LOADGO,PARH.GO='StZE=2?S00O'
//GO.SYSLIN2 DO DSNAlKE=PUB. .AG.P01,DISP=O-D,(NIT=23141,
I II V0LUME=SER=PUBOO1
//GO.FT01F0O1 DD DSNAMiE=JAGCG.OUjT1,UNIT=SYSRA,D!S3P=(MIEU,PASS)),
II SPACE=(CYL,(1,1),RLSE)
//GO.SYSIN DP
FPARMCS NODIVS=3, NI'S0,NCEN=1, LAST= 1, 1IBUG=19, &END
&POINTS PT(I)=(0,0), PTC2)=(1,0),PTC3)=(1,1),PT(4)=(0,1) E-END
fTR NODES=1,2,3, 12IND(1=1, BN1P(2)=2 &ENP
t NODES=1, 3,4i, BMD( 2)=3, BND( )=Ii, ENDTR=T AENP
~&TR
IISTP2 EXEC LOADGC,PARM.G0'SIZE=288000'
//GO.SYSLIN2 DD DSNAE PUB. JAG. PO 2, DISP=OL-P,
/1UN IT= 23 1,VOL1f,:E=SER=P1JB0O3
//GC. FT01FOO1 DD DSNAtE=JACG.OUT1,DISP(OLP),PIASS),ULit =SYSDA
//GO.FTO2FOO1 DD DSNAIE=JAGCG.OUT2,DISP(Nr'_i,PI'SS),UI.'IT=SYSP,.A,
II SPACE=(CYL,(2,1),RLSE)
//GO.systrN DD
&PPRI-IS IBUG=0,IDEG=3,NCP3,ICP=1,2,3,
IIIHS=2,t X2 =l,UY2--l,t2O0,IEIG=0 HIND
//STP3 EXEC FORTHLG
//LKED. JAGP03 DR RSNIA1'E=PUP.JAG.Tk.P,DISPQI.-D,L'NIT=2314i,
t II VOLUI-,ESER='PUB001
//LKED.SYSIN DD
INCLUDE JAGPO3
//GC.FTO1FOO1 RD DSNAIE=JAGCG.OUT1,UNITSYSr)AJISP=(OL,,PELLETE)
//GO.FT0J2F0O1 OD DONAMVE=JAGCG.OUT2,UNIT=SYSRA,DISP=(OLD,DELETE)
//GO. FTO3FOO01 DO DSNA E=JAGCG.OUT3,UNIT=SYS)A,I)SP=IIE:, PASS),
// SPACE=(CYL,(1;4),RLSE)
//GO.SYSIN DRD
&PARf.S NBNDS= 4, 1PRIMT= 1,1 OU(1) 10, 1s.2)=11, ISOU(3)1, EMl
1 10 11 11
2 10 11 11
3 10 11 11
4~ 10 11 11
215
000 f
U 0 0, 0 0 0 0
uj U 4. Ul u LA
. in in n 'A V) wn
- ~ 0'0
r *
00 0 0a
1w
0.0 0 W . UW W
78 0 0 I . . S.
8 4 u .
* 80 C, C
00 C4
A . .. . *
uW , w IU W
r- 0CU S.- ul
In-
Lw
in uuiZ ot
- ~ m X In 0.
Uj X
cci C~a~ 0
fy M MIV 216
z E -L
'Iz
* In
LA '
LU Z LU w
-. 0 0
L LU -1 W
I .
w >-
LUL
Ni
z
L
L* U- w U L
LL -
*L (Xa z .00
*~ 0.wU
U -_ __ _ ___- -- -~-- -El
ii
%A *n'
r z z z
0
3 Rs %j11
wU w w w
z9 -
4LL
wj v u v
u, 0. w
-X -A In .* qV In w j LA
ui ~
I-a ~ ~ 4 . Av
*j a% w w U. w~ 4 A w a
0. z uJ . W - U - , UI
-~ -4 -Z 7crf 0 0
7 -- j Om- W, CCol O
CL a- Q-
PU w. V%
x " 0 0.s x A 1.. z z * 7c
2L Z 0 CU w
VA
Z.0 4D .4U CO
-- *
PU 0 0 N 9
218 V
00 00, 0 0 00 0 00C0000000 00000 00000
00
00
00
00r.
00
00c
0
0
C
0
Q
0
0
00
:
00
0ol
0
0
0000(0C00
000000000
c 0
ao 0v
00000
0000
vC
01.000
C.,0CI,
00000
0000
00000
00 00 0 0 00 0 0000000-00 0 ,0 00 00 0 00
N-tn ~ 0N ^j 4t
0 &
0 I
-NNJ' -I- U z. - 1 L' o -f
00 a 0 0r 0 000000000 0 0%00m 0 I0 0
CIO
0 0 0 00 0 00-000000 0000- 00000O
UW LU U
Wt LU tu ILIUW u LU
U t L
0 iWLUUWOW
)UW WIU j wWU. LU)W u UW
a It I I It I l~ ofl~ III111
1
4
In C, (10 03 . .J
mZI 04 40 0 mC'r0e ZO' 'C .0 0 ml -Nm 4.' M. NZ, on (r44N
'A 4,1 C' C'.0 U 24 D - , ,4rI mO rl- C C, c N.0men U N.4 .. N. 04. n ='344
c0 wtf~- 0
mF4. 400 QlNlo (Ct-C IO 00 0m m m".-4N4 In'.tf4Z
0m Cl-. a- .44 co O 4.'
'I. ml- '.,j 0 e,
0 . ?-. c %J %h~' .j') t0: 1 C.O %% 1.0 ll hI. J ) . . 1 .1
f.
.J 000 -. 00 Nc OC 0 .- N 0.44 0 0.4 p4 0 0C, 0 0 0 0 0 I 00 0 c
7
'a
0C
T
4~0 '
10
~ -
(n%
)
m CCN 4-0 ODU'4
r C-1
c F ir0
au
-
q0 0.
fr-,-
tf40 n~t~e,
.0
a10
0 -"s-
I-4
,
en
.flten1
'AC0 -Z 0fmt 0~ wc 0o P-A~Z UN~f
t''M .Lf10 4
014mC In Q
l410' CO ct.4 NO %n~4 N ',
.J0~4 W0U.t '
0. j 0O 000 OC 00 00 00 0
COO 0- U' C00 0
.1UjluLU
U) "Uw LuW ,w Uj
u219w JLu Uu WL, luL
LU w
000
000
Ui LU
I-r. Of
.0L
LC CLN
L LC .I
.00
w C, 0-
O 90-b- 0 W ..
[
4c
MM= M~m -J
-j iO
<
>>>> 44
.
C-. CL 0.
0- CNN -.
(L L
C.-. cCN CS
S,~ cc-
0 cc
0 .0 - q
22
These cards and !nput produced the quintic entry in
table 5.2.2. Note that the object decks for the
Ti
inverse iteration code using Bunch's symmetric solver
are stored In the data set PUB.JAG.INV.
//JAGXXHL5 JOB 'JAG$CG',54,CLASS=E,REGION=300K
//STP1 EXEC LOADGO,PARV,.G0'1S1ZE=288000'
//GQ.SYSLiN2 DD DSNAME=PUB.JAG.PO1,DISP=OLD,UNIT=231i,
/1 VOLUME=SER=PUB001
//GO.FTO1FGO1 DD DSNAME=JAGCG.OUT1,UNIT=SYSDA,DISP=(NEWI,PASS),I
/1SPACE=( CYL, (1,1), RLSE)
//GO.SYSIW DD *
&PARMS NDIVS=1, NPS=4, NCEN=6, LAST1, IBUG=O, &END
L 221
//STP3 EXEC FORTHLG
//LKED.JAGPO3 DO DSNAME= PUB. JAG. TMP, DISP=OLP,UN IT=2314~,
II
/ILKEO.SYSIN DOD
INCLUDE JAGPO3
VOLUIIE=SER=PUBOO1
1
//GO.FTO1FOO1 DD DSNAM1E=JAGCG.OUT1,UNIT=SYSDA,DISP=COLD,DELETF)
//GO.FTQ2FOO1 DD DSNAf4E=JAGCG.OUT2,UNIT=SYSDAOISP=(OLD,DELETE) [
//GO.FTO3FOO1 DD DSNAI14E=JAGCG.OUT3,UNIT=SYSPA,DISP=(NEWV,PASS),
/1
//GO.SYSIN DD *
SPACE= (CYL, (1,1), RLSE) I
&PARNS NBNDS=1, IBUG=O &END
/1 5 1
IISTP5 EXEC FORTHLG
//LKED.JAGP4i DD DSNAME=PUB.JAG.INV,DISP=OLD,UINIT=2314,
//LED.YSN 0IIVOLUtME=SERPUBOO1 i
INCLUDE JAGP4
//GO.FTO2FOO1 DD DSNA1ME=JAGCG.OUT3,UNIT=SYSDA,DISP=(OLD,DELETE)
//GO.SYSIN Of) *
E
&PARMS SHIFT=9.6 &END
&PARMS SHIFT=-1 &END v
122