For discussion at an Ordinary Meeting PaperNo.

6986
on Tuesday, 31 January, 1967 and
for subsequent written discussion*

DYNAMIC RELAXATION
by
Joseph Reuben Harry Otter, B.Sc.(Eng.), M.I.C.E.
Partner, Rendel, Palmer & Tritton; Visiting Professor, Imperial College
Alfred Carlo Cassell, Ph.D., D.I.C., A.M.I.C.E.
Senior Lecturer, Imperial College
Roger Edwin Hobbs, B.Sc.(Eng.)
Research Assistant, Imperial College

SYNOPSIS
The Paper describes a new numerical technique which has been used to
provide a full three-dimensional solution of the problemof an elastic arch dam.
The technique, dynamic relaxation, is an iterative method for use with a com-
puter, to solve the finite difference formulations of the equations of elasticity.
The methodis essentially a step-by-step integration of critically damped vibra-
tion, using viscous damping, toensure the attainmentof a steady-state solution.
The process is explained by the simple example of a bar under axial load.
Different variables are specified at different nodes to introduce the concept
of interlacing nets, which is developed inthe two- and three-dimensional
examples. A two-dimensionalexample of a planestrainproblem follows.
The applicationof the method toa three-dimensional problem is illustrated by
calculating the stresses in an arch dam under water, temperature and gravity
loading; the effect of varying Poissons ratio is also examined. The equations
are used in a separated form todeal with the mixed boundary conditions, and
the variables are specified on interlacing nets in space. The classical principle
of advancing to successively finer nets is followed, and the facility of dynamic
relaxation in solving the very large numbers of equations involved is demon-
strated. The Paper concludes with a discussion relating dynamic relaxation
with other methods of iteration such as successive over-relaxation.

INTRODUCTION
DYNAMIC RELAXATION IS A SYSTEM OF STRESS ANALYSIS for continuous structures
in one, two or three dimensions and it has been specifically designed for
numerical calculation using a high-speed digital computer. Its early history
and use for prestressed concrete pressure vessels have been describedby Otter1
and the principal objects of this Paper are togive a more precise formulation
of the mathematical basis of the system and to report results of calculations
made at ImperialCollege to determine the stresses in a cylindricalarch dam in

* Written discussion should reach the, Institution by 15 March, 1967, and will be
published in or after July 1967. Contrlbutions should not exceed 1200 words.
633
634 OTTER, CASSELL A N D HOBBS ON DYNAMIC RELAXATION

NOTATION
Spectral radius of a matrix
Cartesian co-ordinates
Stress components
Cylindrical polarco-ordinates of radius, angle anddepth
respectively
Stress components (e.g. rs denotes the componentwhere r is the
direction of resolution of the stress, and S the direction of the
normal to the surface on which the stress acts)
Youngs modulus
Poissons ratio
Mass density
Coefficient of thermal expansion
Displacement components along the r, S and z directions
Velocity components along the r, S and z directions
Viscous damping factor
Time and increment of time
Lame constants
Number of meshes from originin r, S and z directions re-
spectively
Mesh lengths in r, S and z directions respectively
Pressure wave velocity, celerity
Accelerating factor and optimum accelerating factor
Undamped and damped angular velocities
Convergence rate

a rigid valley due to hydrostatic and gravityloading, and to uniform and

differential temperature effects.
2. Dynamic relaxation is generally applicable to all forms of continuous
structures but in this Paper the exposition and applications are restricted to
those in which the number of independent displacements at any point is equal
to the number of spatial dimensions, namely, to a bar stressed along itslength,
to a flat plate stressed in its own plane and to a solid stressed in all three
dimensions. The governingelastic equations inallthese cases are second
order. There is no essential difference in the applicationof DR to theanalysis
of straight bars and flat plates subject to bending, nor of bent rods or thin
shells, all of which involve fourth-order equations; lack of space, however,
makes it necessary to defer discussion of these to a further paper.
3. The system is based onthedynamicequations for elasticisotropic
materials.
4. The derivation of these equations involves the separate consideration of
(a) the motion of an element of the body due to internal stresses and imposed
body forces and (b) the elastic relation between the stresses and displacements
during the course of the motion. The equations are used thus since the two
sets, in a finite difference form, are suitable for numerical solution by a digital
computer. This leads naturally to the use of centred finite differences in the
equations.
OTTER, CASSELL AND HOBBS
DYNAMIC
ON
RELAXATION 635
5 . The object of the calculations is not, however, to study the motion but
to determine the static stresses and displacements of a structure, the boundaries
of which are simply related to the co-ordinatesystem selected, under external
and internal forces. A viscous damping term is therefore introduced into the
dynamicequilibriumequations,thedamping being arrangedto be nearly
dead-beat.
6. The major features are as follows.
(a) Use is made of centred finite differences, which, in the spatial and time
dimensions, is equivalent to the use of interlacing nets.
(6) The elastic equations are in their simplest form; thatis, the stress-strain
and dynamic equilibrium equations are used separately.
(c) It is essentially an iterative solution ofthe static
elastic equations, which
is equivalent to the finite difference solution of the critically damped
wave equations of the system by an initial value boundary method.
7. The first part of the Paper deals with calculations in one, two and three
dimensions to demonstrate the method in detail and, in particular, to record
results of work carried out at Imperial College on a cylindrical dam. The
second part deals with the basic mathematics of the method and its'relation
to other iterative methods, and to thealternative matrix methodof formulation
of the problems.

Part 1. Applications to elastic problems

APPLICATION TO A ONE-DIMENSIONAL PROBLEM
8. A one-dimensional example provides a useful illustration of the method
although the problem selected is trivial, namely, that of a bar subject to an
axial load at its free end and fixed at the other.
9. The condition at time t = O is that the displacements W are zero every-
where, and the stresses zz are also zero except at the free end where z z = p is
suddenly applied.
10. Thesolutionrequiresthesimultaneous satisfaction of equations (1)
and (2) throughout the iteration until oscillations become insignificant. The
right hand side of equation (2) then tends to zero and equilibrium is satisfied.

1 1 . Finite difference forms of the f i s t derivatives in these equations are

subject to asmaller error when the displacement and stressvariables are
specified at alternate nodes. The bar is divided into eight elements thus:
- x o x o x o x o x 0
W1 zz1 W 2 zza W3 ZZQ W 4 zz4 W5 ( 2 2 5 )
The boundary conditions are +
w1= 0 and 2p = zzs zz4,where zzsis a fictitious
point.
12. Equations (1) and (2) in finite difference form are:
zzp =
E
zz
(WJ? - W;') . . . . . . .
636 OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION

whence, from equation (4)

13. The superscripts (r) and (r + 1) refer to successive iterations, and the
velocities are calculated at times A t / 2 before and after the stresses and dis-
placements. The principle of specifying variables on alternate nodesthere-
fore applies in time, as well as to the spatialdimension.
14. The initial conditions are that displacements and velocities are every-
where zero. The iteration begins at t = O with the substitution of zero dis-
placements in equation (3) to give the stresses. The stresses at time t = O are

TABLE1: ONE-DIMENSIONAL EXAMPLE

No. of iterations. . 1
Velocities . . . . 0 . 0. 0 .0 . 0.
Displacements . . 0. 0. 0. 0. 0.
Stresses . . . . 0 . 0 . 0. 0. -200.0
No. of iterations. . 2
Velocities . . . . 0. 0. 0. 0. -0.300
Displacements0 . . . 0. 0. 0. -45.
Stresses . . . . 0. 0. 0. -135.0 -65.0
No. of iterations. . 3
Velocities . . . . 0 .0 . 0. -0.202 -0.095
Displacements . . 0. 0. 0. - 30. -59.
Stresses . . . . 0. 0. -86.6
-91.1 -113.4
No. of iterations. . 4
Velocities . . . . 0. 0. -0.137 -0.128 -0.103
Displacements . . 0. 0. -21. - 50. -75.
Stresses . . . . 0. -61.5 -87.3 -75.5 -124.5
No. of iterations. . 5
Velocities . . . . 0. -0.092 -0,130 -0.068 -0.143
Displacements . . 0. - 14. -40. - 60. - 96.
Stresses . . . . -78.4
-41.5 -59,4 -109.2 -90.8
No. of iterations. . 10
Velocities . . . . 0 . -0.010 -0.020 -0.010 -0.046
Displacements . . 0. -37. -72. -110. -148.
Stresses . . . . -109.8 -106.9 -114.4 -113.7 -86.3
No. of iterations. . 20
Velocities . . . . 0 . 0.002 0.006 0.003 0.003
Displacements . . 0. -33. - 66. -99. -132.
Stresses . . . . -99.5
-99.0
-99.0
-99.2 -100.5
No. of iterations. . 30
Velocities . . . . 0 . -0.000 -0.001 -0.001 -0.000
Displacements . . 0. -33. -67. -100. -133.
Stresses . . . . -100.1 -100.1 -100.0 -99.9 -100.1
No. of iterations. . 39
-.
Velocities . . . . 0. 0.000 0.0000.000 0.000
Displacements . . 0. -33. -67. -100. -133.
Stresses . . . . -100.0 -100.0 -100.0 -100.0 -100.0
OTTER, CASSELL AND HOBBS ON DYNAMIC
RELAXATION 637
then used in equation ( 5 ) to give the new velocities at time (t+dt/2). The
new displacements at time ( t + d t ) are given by:
wf = w?+dt.6~+ . . . . . . (6)
15. The iterative step finishes with the calculation of zzg+l)from equation
(3).
16. The next iterative step begins with the substitution of these new stresses
in equation (5).
17. Stages in the calculations for a bar 4 in. long (i.e. A z = 1 in.), with an
applied compression p = 100 lb/sq. in., are given in Table 1.
18. The properties of the bar are E = 3 X 106 lb/sq. in., a=0.20, K=0.40,
p=& Ib/cu. in., and the time increment used is A t = 150 vs. The units for
the results are: stress (lb/sq. in.); velocity (in./s); displacement ( p in.).
19. It can be seen that by the twentieth iteration the displacements and
stresses are within 1 % of their final values. The iteration was continued until
successive iterations showed a change of less than 0.1% in the stresses.

APPLICATION TO A PLANE STRAIN PROBLEM

20. For purposes of comparison the problem taken was that dealt with by
Gilles2, that is a rectangular steel plate in a state of plane stress. The stress/
strain equations are:

x x = (A+2p)z+A-
au av
ay
1
l

and the dynamic damped equilibrum equations are:

21. The solution of Gilles problem by DR can be obtained quite simply

by using these equations alternately in the same manner as described for the
bar above. As a matter of interest the solution was obtained using the dis-
placementequations, obtained by the substitution of the stresses given in
equations (7) into the dynamic equations (8) and this results in the damped
dynamic form of the equations given by Gilles, as follows:

Where C 2= ( A + 2p)/p.
22. Gilles specified the variables on interlacing nets; i.e. as if the U variables
occur at the centresof the black squares and thev variables at the centres of
the white squares of a chess board. Interlacing nets are used as in Gilles
 FIG.1 : FINITE
DIFFERENCE GRID

paper and in the normal procedure for DR, giving a better finite difference
approximation for the shear stresses.
23. The values of the displacements calculated agree with those quoted by
Gilles, to the accuracy of his computations.

APPLICATION TO A THREE DIMENSIONAL PROBLEM

24. Using the equations of elasticity in the separated first order form, the
stress/strain equations, in cylindrical polar co-ordinates, are:
au 1 av aw
rr = (A+2p)-+A-+A--+h-
ar r r as az
. . . (11)
all 1 av
ss = hz+(h+2p):+(h+2p)--+A- aw . .' . (12)
r r 2s dz
au U 1 au
zz = h- +A-+A--+(h+2/4- . . . . (13)
ar r r as az
au aw
rz=p-+p-
az ar
. . . . . . . . . . . (14)

1 au v bv
ps =p-- -p-+p- . . . . . . . . . (15)
r as r ar
av 1 aw
zs = p-+p--
az r as
. . . .
OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION 639
and the equations of motion are:

ad K . ars l ass azs 2rs

p -+-v
( a t At ) =-+--+-+-
ar r as r az
. , . . (18)

(2
P -+---k -
arz lazs
azz
rz
. . . .
25. The iteration is between the previous two groups of equations, in a
manner similar totheiteration between equations (1) and (2) inthe first
example.
26. Figs 1 and 2 show the form of the seven interlacing space nets used in
the finite difference representation of the equations for the example selected,
an arch dam. In finite difference form equation (1 1) typically becomes:

/
/
/
t \

FIG.2: FINITE
DIFFERENCE BLOCK
640 OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION

DISTANCE FROM t OC
NR E S T Cmetrra)
3 1
0 20 30 40
I I I I I

30

5 20-

2
W

8
+ 10-
X
0
W
I

0-

FIG.3 : WATERFACE PRINCIPAL STRESSES

DISTANCE FROM c ON CREST Cmrtrrr)

0 10 20 30 40
I l I -7-

30-

: 20-
E
U

a
W

8
10-
+
Y
X

0-

LEGEND
- CONTOUR FOR M U . PRIKIPAL STRESS
-- - CONTOUR FOR MIN. PRINCIPAL STRESS) VALUES I N k9s9cm

FIG.4: WATERFACECONTOURS OF EQUAL PRINCIPAL STRESS

OTTER,
CASSELL AND HOBBS ON DYNAMIC RELAXATION 641
DISTANCE FROM c ON CREST omtm>
10 20 30 40
9 1 I I

LEGEND
- DIRECTION OF MU(. PRINCIPAL STRESS ( P I ) ,
__ - DIRECTION OF M I N .P R I N C I W
STRESS (PI)
STRESS VALUES ARE GIVEN ONLYFOR POINTS AT
WHICH EITHER Pl>+lSk~/sp.rmOR P2<-2Ok9/SqCm

FIG.5 : AIRFACE PRINCIPAL STRESSES

0 1
0 X) 30 40
DISTANCE FROM C O N CREST metres

LEGEND.
- CONTOUR FOR
MAX PRINCIPAL
___- CONTOUR FORMIN. a~

FIG.6: AIR FACE CONTOURS OF EQUAL PRINCIPAL STRESS

642 OTTER,
CASSELL A N D HOBBS ON DYNAMIC RELAXATION

The twosets of equations arelinked by displacements obtained from equations

(22)--(24)
U ( ' + 1 ) = U(T)fti('+l)dt . . . . . . (22)
V('+') = ucr)+zYr+l)dt . . . . . . (23)
+1) = + ,p + l ) d t . . . . . . (24)
21. The iterations continue until the accelerations and velocities diminish
and the right hand sides of equations (17), (18) and (19) approximate to zero;
thus satisfying equilibrium. Theiteration is thus a stiffness method which
always satisfies compatibility and ultimately achieves equilibrium.
28. The example chosen is an elastic cylindrical dam of constant thickness
set in a rigid valley. The properties anddimensions of the dam are asfollows.
Young's modulus ( E ) = 200 000 kg/sq. cm
Poisson's ratio (U) = 0.15
mass density = l / g .2400 kg/cu. m
coefficient of expansion = 0~000009/"C
height = 30m
thickness = 3m
radius to air face = 43.25 m
crest length = 80m
and crest angle = 106"

29. The symmetrical valley is then defined by the shape of the developed
surface of the air face (Fig. 1). The dam has been analysed for the following
loadings :
(i) water to crest level,
(ii) uniform temperature drop of 10C over the whole dam,
(iii) uniform temperature gradient through the thickness of the dam, the
air face being at datum temperature and thewater face 10" cooler,
(iv) gravity acting on the completed structure as a whole,
(v) water acting on the dam with Poisson's ratio=0.45.

Boundary conditions
30. The boundary conditions used in these analyses do not involve stress
or displacement derivatives, and thus the mixed boundary conditions present
no problem as seen in the following examples.
31. Rigid valley ut base of dam. Reference to Fig. 1 shows that the nodes
for W occur in the boundary, and thesedisplacements are set to zero; U and v
occur at nodesabovethe base,accordinglyfictitiousdisplacements are
specified at an equal distance below the base. These are made equal and
opposite to those occurring justabovethe base foreachiteration.Thus
U, U, W are made zero at the base.
32. Water face of dum. The rs and rz nodes occur on the face and these
stresses are set to zero. The hydrostaticpressure (load case 1 and 5) is applied
by means of fictitious rr nodes, placed so that the water face is equidistant
from them and the rr nodes nearest the face. The sum of rr stresses at the
real and fictitious points i s then set to be twice the water pressure at each
iteration. For load cases (ii)-(iv) this sum is made zero. Fictitious ss values
are obtained by linear extrapolation of the real variables.
OTTER,
CASSELL AND HOBBS O N DYNAMIC RELAXATION 643
33. Air face of dam. This differs from the water face only in setting the
sum of the rr stresses at the real and fictitious points on theface to be zero for
all load cases.
34. Body forces. The gravity loading of case (iv) is applied as a body force
to equation (19) thus:

33. Initial conditions. It is possible to start the iteration from any com-
patible set of displacements, but for convenience the initial displacements and
the internal stresses were chosen as zero which are obviously compatible for
all cases. Inload cases (ii) and (iii) forthetemperature loading, normal
stresses proportional to thelocal lack of fit (equation (26)) were immediately
added to theinitial values of the stresses:
rr = ss = zz = E(1-02)a.AB . . . . (26)
The subsequent iterations restoring the stress compatibility converged to the
stresses rising from the given temperature distribution.
36. Time of iterative step. Using c = pressure wave velocity, then Forsythe
and Wasow3 haveshownthatfor m Cartesian co-ordinatesthe stability
criterion is
1 {( 1 )2
A t < c G +...+ -
(1 )Z}-
. . . .
A xm
Where A x , . . . Axm are the respective mesh lengths.

Results
37. Because of the very large number of variables involved, only a selection
of the results for each loading case are presented here. However, the full
results are available at Imperial College.
38. Load case (i). The stresses inthedamunder hydrostatic loadare
given in Figs 3-6 and 10-13.
39. Load cases (ii), (iii) and (iu). Thetemperatureand gravityloading
results are shown in Figs 7, 8 and 9. The stresses in the crest and crown
cantilever and the crown cantilever displacements are given as an illustration
of the structural action produced.
40. Effect of mesh size. The error in the finite difference approximation
is reduced with a decrease in the mesh size. The effect of varying the mesh
size is shown inFigs 10-13. The difference between the results given by
the 16 X 16 X 10 and 24 X 24 X 10 grids is very small, and it is concluded that
the first order finite differences adoptedareadequatefor thisproblem.
Figs 3-6 are based on the fine mesh solution. Because the 16 X 16 X 6 mesh
gave a reasonably accuratesolutionit was used fortheotherload cases
(Figs 7-9 and 14-17), to reduce computer time.
41. Load case (v). The results are shown in Figs 14-17, where they are
compared with the 0.15 Poissons ratio solution and with other solutions.
Snell, Brown and Peatfield4 used a photoelastic technique and Severn, Smith
and Taylor5 a rubber model, with U = 0.48 and 0.5 respectively. The United
States Bureau of Reclamation used the trial load method and U = 0.15 in an
unpublished report to the Civil Engineering Research Association.
644 OTTER, CASSELL AND HOBBS
DYNAMIC
RELAXATION
ON

z
r!t'1 ..
YJ

o_

0
44

0
OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION 645

FJ: 0
0
0 P 0
646 OTTER. CASSELL AND HOBBS ON DYNAMIC RELAXATION

t
OTTER,
CASSELL A N D HOBBS ON DYNAMIC RELAXATION 647
648 OTTER, CASSELL A N D HOBBS ON DYNAMIC RELAXATION

E
5

c.
0
i
OTTER, CASSELL AND HOBBS O N DYNAMIC RELAXATION 649

t
650 OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION

Discussion of arch dam solution

42. Load case (i). The effect of decreasing the mesh size (Figs 10-13)
indicatesno significant errordue to the finite differences. Manual calcu-
lations showed that parts of the dam selected arbitrarily were in equilibrium.
The trial load solution is approximate and based on a method which enforces
compatibility of displacement only at the intersections of the notional arches
and cantilevers; however, the agreement (Figs 14-17) between the two solu-
tions is good.
43. Loadcases (ii), (iii) and (iv). Thereareno publishedresults for
comparison with these solutions. It is interesting to notice that stresses as
large as those producedby water load are induced by temperature changes of
15"-20"C. The gravity load problem was included to demonstrate that the
method dealswithbodyforce problems as well. The dotted linein Fig. 9
shows thevertical stresses corresponding to the height and density of dam above
any level.
44. Load case (v). The results (Figs 14-17) do not agree well with the
experimental values, but the experiments used valleys that were not com-
pletely rigid and the experimental Poisson's ratios exceeded 0.45. Neverthe-
bss, the dynamicrelaxation solution shows the same trendof behaviour. The
change in Poisson's ratio altered the sign of the hoop stress on the air face
(Fig. 15); this suggests that frozen stress photoelasticity is not always a valid
experimental technique for determining stresses in prototypes madeof materials
with a low Poisson's ratio.
45. The work reported above confirms the conclusions reached by Otter'
that the system gives results of acceptable accuracy for practical engineering
problems. N o difficulties have arisen in obtaining satisfactory convergence,
even in a case with highPoisson's ratio inwhich serious convergencedifficulties
had previously been encountered at Imperial College.
46. It seems clear that compatibility and equilibrium are satisfied in the
final results of the iteration and that the use of first order finite differences
is adequate to give acceptable accuracy.

General
47. The DR method has provided accurate solutions to previously intract-
able problems of arch dam behaviour. Many other three-dimensional elastic
problemsshould be equallymanageable. Themethod is economical in
computer time; even the fine mesh solution only took 38 mins on the Imperial
College IBM 7090 and the practically adequate mesh (16 X 16 X 6) only 8 mins.

Part 2 Generalmathematical basis

INTRODUCTION
48. This part of the Paper extends the comparison given by Otter.6 DR
is an iterative method of solving certain types of partial differential equations,
which when expanded infinite difference form canbe expressed as symmetrical
positive definite matrices in terms of displacements. The basic equations may
be of the second order as inextensional problems, or of the fourth order asin
flexural problems; both types can be handled by this method.
49. To illustrate the mathematical basis, the Laplace or Poisson equation
will be used. Consider the Laplace equation in two dimensions:
OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION 65 1

(28)
Subject to the boundary conditions, the equation becan
expressed as a damped
wave equation:

or
Using centred differences in the timedimension so that ;(r) occurs at timedr/2
before U ( ' ) as in equation ( 3 , then equation (30) becomes:
; ( r + l ) = {(l - K / 2 ) i ( " + d t . C'F:)}/(l +K/2) . . . (31)
Using 6u=ridt then
S U ( ' + ~ )= { ( l - K / 2 ) 6 ~ ' " f ( d t ) ~ . C ~ F : " } / ( 1 + K / 2 .) . (32)
50. Taking the equality in equation (27) to eliminate (dQ2C2gives:

i
F$)
6u" 1) = (1 - K/2)6U'" + +
/(l K / 2 ) . . (33)
(ti'.id;)'l
+

51. In a normal engineeringcalculation d x fdy, butfor simplicity in

exposition and to comparewith most references on iterative methods, dx and
d y are taken to be equal; this causes no serious loss in generality.
52. The spatial function F, is now formulated in finitedifferenceterms,
giving:
a U l(,r, + l ) = {(l-K/2)6~I:j++(ul?1.j+ui'l1.j+ul::+1+~',3-1
- 4~$:])}/(1+ K/2) . . . . . . . . . (34)
53. It is important to note that the values of the dependent variable U in
F* at different mesh points areall taken at thesame iteration stage and that the
iteration is thereforesimultaneous.Afterall ~ u ' , I , + have
~ ) been calculated
using the previously determined U!:;, the next set of values for ZI are then cal-
culated from :
u(r + 1) =
t.1
+ sui;$ 1) . . . . . . (35)
54. In thepractical application of the algorithmcomprised in equations (34)
and (35), the initial values of u i , ] that is, the ujpj' are taken to be zero at all
points except those on the boundaries which aretaken to have the given
boundary values.
55. Comparingthealgorithm developed above with that used inthe
dynamicrelaxationmethodit is interesting tonotethatthere is now no
explicit reference to any specific value of the time interval, since the criterion
for numerical stability has been built into the'equations.
56. It will also be noted that the iteration is carried out directly on the
second order spatial derivative compared with the alternateuse of the equiva-
lent first-order derivative equations adopted in dynamic relaxation, and that
the mesh is non-interlacing. The iteration is now expressed in a form which
can be compared directly with other iterative methods; it may be that it would
be better to refer to the present iteration form as 'dynamic iteration' leaving
the name 'dynamic relaxation' for the complete system developed for the
stress analysis of elastic structures.
652 OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION

COMPARISON OF DYNAMIC RELAXATION AND FRANKEL'S ITERATION

57. As previously described, D R was developed as a system of numerical
analysis from earlier work on tidal analysis, and included centred differences
in time and space, separated stress/strain and equilibrium equations, and an
iteration based on damped wave equations using near-critical damping for
optimum speed of convergence. Theboundaryconditions were entered in
asimple manner which in general appeared to require little mathematical
justification. When the system came to be compared with other methods of
numerical analysis questions arose as to the theoretical basis of the method,
particularly with regard to its convergence to the accurate solution of the
elastic problems to which the system was being applied. Whenever it had
been possible to check the results against analytical solutions they had been
confirmed, nevertheless the derivation of the iteration was heuristic rather
than rigorous.
58. Rigorous treatments of a number of methods of iteration, particularly
those of Jacobi, Gauss-Seidel and successive over-relaxation, were based
mainly on matrix methods of analysis. At first there seemed little in common
between DR and other methods,since matrix methods resulted in the use of
eigenvalues of the various matrices to determine convergence rates and opti-
mum convergence conditions, in contrast toD R which uses the fundamental
vibration frequency of the elastic system and a numerical stability criterion.
However, it eventually became clear that the basis of D R is a second-order
Richardsoniteration,and is essentially thesameasFrankel'smethod
(Franke17 and Stiefel*).
59. After somerearrangement,andinthenotation of thisPaper, the
equations for Frankel's method (Stiefel, p. 31) become:

where a and b are a lower bound for the minimum and an upper bound for the
maximum eigenvalues of the matrix [R], and where b>>a.
60. The generalsimilarity between this equationandequations such as
(32) and (34) is immediately apparent. The connexion can be found from the
relationships of the numerical stability criterion andthe fundamental vibration
frequency to the eigenvalues of the matrix [R].
61. The equivalence between Frankel's method and the optimized form
of D R means that the mathematicalbasis of D R has, in consequence, already
been justified.

COMPARISON WITH OTHER ITERATIVE METHODS

62. It will be noted at once that if the damping factor K is taken as equal
to 2, equation (34) reduces to the form:
SU&+ 1) 4(uII:1.,+u~~1.,+ul~~+l+uI~:-~)-u!~~
= 1 . . (37)
which is the iteration formulaof the Jacobi'iteration for thetwo-dimensional
Laplace equation. The Jacobi iteration is of course a simultaneous iteration,
as is dynamic iteration, and it is appropriate to refer to it as thebasic simul-
taneous iteration. The Jacobi iteration, which is of the first order in time, can
OTTER, CASSELL AND HOBBS
DYNAMIC
RELAXATION
ON 653
beconsidered as the finite difference equivalent of the parabolic diffusion
equation :

63. It will be found that the Jacobi iteration converges more slowly than
the dynamic iteration, and can be considered as the limiting case of a much
overdamped wave equation. The choice of the damping factor, K, less than
2 in dynamic iteration accelerates the convergence as an overdamped system,
until K falls to a critical value, after which the solutions execute a damped
oscillation about the analytical static solution. Dynamic iteration may thus
be referred to as anaccelerated simultaneous iteration. It will be shown later
than when K is selected to give critical damping, the speed of convergence is
greatest, and the value of K may then be considered as the optimumaccelera-
tion factor.
64. The other mainiterative methods are theGauss-Seidel iteration and the
method of successive over-relaxation (SOR), both of which are successive
iterations inwhich newly calculated values of the dependent variable at a mesh
point are immediately used for the calculation of the value at the next mesh
point.
65. The Gauss-Seidel iteration is represented by the algorithm:
&&+ 1) = L +
& i + l( 7. j) +
ul-+1:: a:;:+ 1 +&+P) -U::: . . (39)
This, except for the use of the ( r + 1)th iterationvalues at certain mesh points,
is the same as the Jacobi iteration,may andbe referredto as the basic successive
iteration. Its speed of convergence is greater than that of the Jacobi as will
be seen later.
66. Successive over-relaxation, which has often been the method of choice
for numerical solutions of elliptic equations, is represented by:
+ +
SujYj = w/~(u?:1.9 ~:??j U;:: + 1 + uiTT22 - 4~:;)) . . (40)
where W is the accelerating factor. This may be called the accelerated suc-
cessive iteration, and, as in thecase of dynamic iteration, thereis an optimum
value of W which gives the fastest speed of convergence (see for example,
Varga,Q Forsythe and Wasow3). It is shown below that for solutions of the
Laplace equation, there is a simple relation between the optimum accelerating
factors of the accelerated successive and simultaneous iterations. Whilst the
derivation of the optimum accelerating factor for SOR is usually obtained by
matrix methods, it is interesting to note that GarabedianlO has shown that
SOR can be expressed as the finite difference solution of a critically damped
wave equation in a certain space-time co-ordinate system, and he develops
the optimum accelerating factor in a form closely reminiscent of that which is
used below for dynamic iteration.
67. The comparisons given above may be summarized as follows:

I Simultaneous
l
I Successive
Basic I Jacobi I Gauss-Seidel
Accelerated 1 Dynamic
iteration 1 SOR
654 OTTER, CASSELL AND HOBBS O N DYNAMIC RELAXATION

SPEED OF CONVERGENCE OF ITERATIONS

68. Much study has been given to the speed of convergence of different
forms of iteration and it is known for the Jacobi, Gauss-Seidel and SOR
iterations that after certain initial effects have died down, they all converge
at a uniform (but different) rate in which l)/Sdr),in each case, is a constant
+

(m).
69. In the case of the SOR solutions of matriceshaving Forsythe and
Wasow's property 'A'(p. 243, ref. (3)) it can be shown that the optimum
accelerating factor is given by the equation:
wb = 2/(1 +(l -7")""j . . . . . . (41)
where 7 is the spectral radius of the point Jacobimatrix. It is also shown that
for such matrices the maximum eigenvalue of the Gauss-Seidel matrix (h)
is equalto 7" andthatfor SOR theiteration is alwaysconvergent for
O<w,<2. t
70. The corresponding relations for dynamic iteration can be developed
using the damped wave equations.
71. It is unnecessary in the determinationof the critical damping factor to
consider modes of vibration other than the fundamental,since this determines
the speed of convergence after the disappearance of the initial effects of the
other modes by overdamping, although it has been found that for damping
factors near critical, the theoretically smallest number of iterations based on
the fundamental mode is not obtained, because of contributions from the
other modes.
72. Inthelater stages of convergence fordynamic iteration, the time
solution in the overdamped case can be written as:
(U) = (uf){l -e-at(cosh w , t + a / w , sinh w 3 t ) } . . (42)
where U,"= (a" - W ; ) , and w o / 2 r is the undamped fundamental frequency.
73. Differentiating once and twice with respect to time gives:

li = u,e-at (<) 2
sinh w,t . . . . . (43)
and

74. In the later stages of the iteration cosh w,t-tsinh w,t++em.8t when:
-$+m) = (-.+us) . . . . . . (45)
U

75. It is knownthat when this stage is reached theratio of successive

increments in U is constant, that is

76. In finite difference terms this is

OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION 655
from which
2(1-m)
-- ii
- --At = (a-wJAt from (49,
l+m U

therefore
2 -(a -w,)At -
- 2 - (a - d a 7 ) A t
m=
2+(a--w,Mf 2+(a- d / c ( a ) A t
. . . (48)

77. The maximum rate of convergence is obtained when the expression

under the root sign is zero, that is when
w 0 = a = - Kc r
2At
. . . . . . . (49)
when

78. For the Jacobi iteration a A t = 1 and the Jacobi rate of convergence is
then m,, = 4 1 --:At2. It is known that mJ=q. Then woAt= d m which
in turn is equal to twice the optimum damping factor K , ,
K C ,= 2d- . . . . . . (51)
79. The optimum convergence rate for D R is given by
2- d/1-172 . . . .
mDR =

80. Substitutinginequation (52) it will beseen thattherelationship

between the optimum acceleration factorw b for the SOR iteration for matrices
with property A and the optimum damping factor for D R is:
2
w b = -
1 + KcrP
. . . . . . . (53)
and the optimum convergence rate for SOR is known to be:

81. The convergence rate for SOR is therefore faster than that for DR.

CONCLUSION
82. Thus D R is not as efficient as SOR when that method is tridiagonalized
with a n optimum acceleration factor, and even less efficient than the elimina-
tion method, and this must be true also for dynamic iteration, as is to some
extentshown inthecomparison of dynamiciterationand SOR forthe
Laplace equation. The virtues of D R as a system depend rather on the pos-
sibility of its use in relatively complicated problems when the band width of
the two-dimensional matrix is so large that storage problems arise for the
elimination method, and the tridiagonalization for theSOR solution presents
difficulties. The use of three-dimensional arrays in D R gives a dense matrix
instead of a sparse two-dimensional matrix; the simplicity of the separated
equations and the resultant simplified boundary conditions greatly ease the
656 OTTER, CASSELL AND HOBBS ON DYNAMIC RELAXATION

programming of the problem. In all problemstackled to date the cost of

computer runs has not been the governing consideration and has always been
economically acceptable.
ACKNOWLEDGEMENT
83. This work would nothave been possible withoutthe generousprovision
of computer facilities by Professor S. Gill and the Imperial College Computer
Unit.

REFERENCES
1. OTTERJ. R. H. Computations for prestressed concrete reactor pressure vessels
using dynamic relaxation. Nucl. Struct. Eng, 1965, l, Amsterdam.
2.GILLES D. C.Theuse of interlacingnets for theapplication of relaxation
methods to problems involving two dependent variables. Proc. Roy. Soc. A.,
1948, 193,407-433.
3. FORSYTHE G. E. and WASOWW. R. Finite-difference methods for partial
differentialequations. JohnWiley and Sons,1960,381-382.
4. SNELL C.. BROWNJ. M. and PEATFIELD A. M. Thethree-dimensionalwhoto-
elastic stress analysis of a constant thickness cylindrical dam under hydraulic
loading. CERA, TR/AD/l, 1965.
5. SEVERN R. T., SMITH N. A. F. and TAYLOR P. R. Theoretical and rubber mode
studies on design shape 1 arch dam under hydrostatic load. CERA, TR/AD/4,
1965.
6. OTTERJ. R. H. Dynamicrelaxationcomparedwithotheriterativefinitedif-
ferencemethods. Nucl. Eng. Des., 1966, 3, 1, 183-185.
7. FRANKEL S. P. Convergence rates of iterative treatments of partial differential
equations. Math. Tables Aids Cornput., 1950, 4, 65-75.
8. ENGELI M. et al. Refined iterative methods for computation of the solution and
eigenvalues of self-adjoint boundary value problems. Mitt. Inst. ange. Math.,
1959, 8, 30-3 1.
9. VARGAR.S. Matrix iterative analysis. PrenticeHall, London, 1962.
10. GARABEDIAN P. R. Partial differentialequations. John Wiley and Sons,1964.