Bulletin of the Seismological Society of America. Vol. 60, No. 1, pp. 161-166.

February, 1970

A NOTE ON FAST COMPUTATION OF RAYLEIGH WAVE DISPERSION

IN THE MULTILAYERED ELASTIC HALF-SPACE

BY T. H. WATSON

ABSTRACT
Matrix formulas for modal solutions in the layered, elastic half-space are modi-
fied to give faster machine computation. Accuracy at high frequencies is also in-
cluded. Computing time is reduced by 30 per cent from the fastest program
previously reported.

INTRODUCTION
The application of matrix methods to the problem of surface wave dispersion in a
multilayered elastic half space has been widely discussed. Thomson (1950) laid the
theoretical groundwork later extended by Haskell (1953). Knopoff (1964), Dunkin
(1965), and Thrower (1965) presented equivalent formulations which made calcula-
tions possible for higher frequencies. Gilbert and Backus (1966) used a more general
approach which ultimately degenerated to the previous methods under certain con-
ditions.
Computer programs employing the Thomson-Haskell approach were written by
Press et al. (1961) and Dorman et al. (1960). More recently Thrower (1965) and Ran-
dall (1967) have reported success with the alternative formulation.
Since twenty or more homogeneous layers are often necessary to provide a realistic
earth model, the fastest possible method for calculating the dispersion relation (Period
versus Phase Velocity) is desirable. In this note we shall present a modification of the
alternative method which provides a reduction of about 30 per cent in computing
time.
A REVIEW OF MATRIX METHODS
For the case of n -- 1 homogeneous, isotopic elastic layers over a half-space we
define the matrix

J = ]~nAn-1 ... A2A 1

where A m is the 4 X 4 Haskell matrix for the m'th layer and ~n is the half-space in-
version matrix (appendix A). Then the secular function (dispersion relation) is written
as a second order subdeterminant of J.

A ( T , C ) = J I~ = JllJ22 - J~2J21 = O.

Therefore only 2 columns (or rows) of J are necessary for this result, and we require a
2 >( 4 matrix to store the product as each layer multiplication is computed.
But the Haskell matrix contains exponential terms which become very large at high
frequencies. Because certain of these terms should cancel one another when the final
subdeterminant is calculated, a severe loss of significant figures occurs in the secular
function.
Knopoff (1964), Dunkin (1965), and Thrower (1965) have offered alternative
methods which avoid this difficulty. By defining a new 6 X 6 layer matrix whose ele-
161
162 BULLETIN OF THE SEISMOLOGICAL SOCIETY OF AMERICA

ments are second order subdeterminants of the Haskell matrix, the offending terms are
eliminated analytically. In this case the secular function appears as a single element in
the product matrix, and thus a six dimensional column (or row) vector is sufficient to
carry the function through the n -- 1 layers. It should be mentioned that if particle
motion ratios are desired, the Haskell matrix is still required although it may be com-
bined with the subdeterminant formulas to insure accuracy at high frequencies (Dunkin
1965). Randall (1967) gives a detailed description of the application of Knopoff's
formulation to a program written for the I B M 7094 computer. His treatment includes
the factorization of the layer matrix into a matrix-vector product such that all fre-
quency dependent terms are excluded from the matrix. Thus b y considering frequency
as the dependent variable, the layer matrix is computed only once for each value of
phase velocity. But this introduces an additional matrix multiplication for each layer,
as well as the inconvenience of using frequency as the dependent variable in places
where the slope of the phase velocity curve is small.

REFORMULATION OF THE ALTERNATIVE ~V~ETHOD

For the subdeterminant formulas used in the following development, we refer to ap-
pendix B or Dunkin (1965). The 6 X 6 matrix B~q for the m'th layer is derived from the
second order subdeterminants A ~ I~ of the Haskell matrix by the following convention:

ij (kl) --> p (q)

12 1
13 2
14 3
23 4
24 5
34 6

Dunkin showed that these six ij (or kl) combinations were sufficient to describe all
possible unique second order subdeterminants for any 4 X 4 matrix.
The B matrix has several properties similar to A. All elements are either real or
imaginary (leaking modes excluded) and the same symmetry occurs about the second-
ary diagonal. Likewise the product of any two B matrices retains these features. Note
t h a t the particular convention chosen to define B is immaterial as long as it is con-
sistent. For this particular arrangement, the secular function is element 1, 1 in the
final product, and can be carried through as part of a simple column (or row) vector.
We initially define the six dimensional column vector

R i I ~- B ~ I .

The final product is then computed by repeated application of the formula

6
R¢m _- ~ R " ~ P m-1
j=l

for each successive layer m until we obtain R "~, the secular function. B ~ is the matrix
of subdeterminants of the half-space matrix E~, and consists of one row only.
But inspection of the B matrix reveals that
Bm m
81 ~ B41
 RAYLEIGH WAVE DISPERSION IN MULTILAYERED ELASTIC HALF SPACE 163

and therefore

R31 = R41.

From previous remarks, we see that for all values of m,

R3 m ~ R 4 m.

So obviously we are doing unnecessary calculations to compute both these elements for
every layer. The solution offered here is to define a modified 5 × 5 layer matrix/)
such that
-Bll B12 2B1~ B15 B161
!
B21 B~2 2B23 B25 B26|!
/~ = B31 B~2 (2B33--1) B35 B36 I"
!

B51 B~ 2B53 B55 B56 I

I
B~I B62 2B63 B65 B~d
R~ is thus reduced from a six dimensional vector to five. One may easily verify that/~
is equivalent to the previous method, and is also suitable for use in the particle motion
formulas of Dunkin (1965). This reformulation is simply the result of a degeneracy in
columns (and rows) three and four of our arbitrary arrangement of the Dunkin sub-
determinants. No physical interpretation is immediately apparent.
C O M P U T E R APPLICATIONS

In order to compare the different methods for efficient use of computer time, we
first investigate the number of floating point multiplications required. We know that
when any two matrices of order n are multiplied, each element of the product matrix
is the sum of n pairs of products. Thus to compute one element in the ttaskell matrix
product requires four multiplications and additions. Since we must carry along eight
elements, thirty-two such "terms" are computed for each layer in the model. The
Knopoff-Dunkin-Thrower method reduces the number of elements carried in the prod-
uct to six. However, each is calculated by summing six terms, resulting in a total of
thirty-six for every layer. By the same reasoning, the number is reduced to twenty-five
by using our modified matrix method.
Each time the secular function is tested for a particular root, new values must be
calculated and stored in the layer matrices. Examination of the formulas in the ap-
pendices reveals 10 independent elements in each Haskell matrix. This compares with
16 and 15 for the B and modified (B) matrices respectively. Thus the Haskell method
should have a slight advantage over the B matrices in total computation time re-
quired.
Obviously there are many other factors which affect the time necessary to compute
a whole dispersion curve. The efficiency of the curve following routine and the accuracy
to which a root is determined are two examples. This discussion is restricted to the
mechanics of computing the secular function, for which our method offers both speed
and accurate results at high frequencies.
The/~ matrix method has been incorporated into a dispersion curve program written
in Fortran II for the IBM 7090 computer. The program uses period as the independent
164 BULLETIN OF THE SEISMOLOGICAL SOCIETY OF AMERICA

variable and determines a phase velocity root to the desired accuracy. Five point
difference formulas are employed both to predict the next root and to compute group
velocity.
In order to test the program we ran the seventeen layer model of Randall (1965).
He reported a computation rate of 15.3 layer points per second (lps) on the I B M 7094,
versus 3.3 lps for the program described b y Press et al. (1961). A section of the Ray-
leigh wave phase velocity curve was computed over the 100-500 second period range.
Forty-one points were determined to an accuracy of .01 per cent in thirty-one seconds
execution time. This gives a speed of 22.4 lps on a machine that is considerably slower
than the 7094.
From this result and the previous discussion it seems the main advantage of the
Randall program is that only one column is carried through the matrix multiplication.
I t would appear that the convenience of separating the frequency dependent terms is
offset b y the additional computation required when the product is taken. The speed of
the program of Press et al. (3.3 lps on the 7090) is somewhat less than would be antici-
pared. However, we suspect that straight 4 X 4 matrix multiplication was used,
doubling the calculation time by carrying two unnecessary columns in the product
matrix.
Our program thus appears to offer the fastest method reported to compute Rayleigh
wave dispersion curves. This technique should also be useful for the calculation of
partial derivatives with respect to the various layer parameters, with a similar re-
duction in computation time.

ACKNOWLEDGMENT
This research was s u p p o r t e d by the National Science F o u n d a t i o n under grant GA-4201.

APPENDIX A

Notation employed:
= angular frequency
c = phase velocity
k = w / c = angular-wave number.
For layer m:
pm = density
d m = thickness
am = compressional-wave velocity
~m -- shear-wave velocity
w-- 2(~/c) 2
r.m = f'+[C2/otm
~ -- 1] I/~ c > am
\-i[1 - c V ~ J ] "~ ~ < ~
r ~ = ~+[C2/~m 2 -- 1]112 C > ~,~
\--i[1 -- c2/flm2] '/2 c < ~,~
P m = kr.,~dm
Qm = kr~,~ d.~.
The Haskell matrix components for layer m are:
A~ = A~ = vmCP- (w- 1)CQ
AI"~ = A3~ = i [ ( w - 1 ) r : ~ S P + 7mr~mSQ]
A I~ = A m
24 = - ( Pmc2 ) -1 ( C P - C Q )
m • 2 --1 --1
A~4 = ~(p~c ) (r . r a S P + r~mSQ )
m
A~I = ~
A48 = - i [ T m r . ~ S P + ( 7 ~ - 1)r~.,SQ]
-1
 RAYLEIGH WAVE DISPERSION IN MULTILAYERED ELASTIC HALF SPACE 165

Am
23 = A ~m = -('~m -- 1)CP + ~.~CQ
m • 2 --1 --1
A~ = *(OmC ) (r~mSP + r~mSQ)
2
A ~ = A4% = OmC'~m('~m- 1 ) ( C P - - CQ)
m • 2 2--1
A~ = zpmc [ ( ' ~ -- 1 ) r.mSP + ~/m r~mSQ]
A'~ = ipmc2[~m:r.mSP + ('Win -- 1)er2~SQ]
where

S P = sin Pm SQ = s i n Qm

CP = cos P m CQ = c o s Qm.

T h e h a l f - s p a c e m a t r i x is

I 2 2 2 2 2 --1
2 --1
- - (~'~ -- 1 ) / % r ~ . l l l ( p n c'y. r ~ ) l(p. c ~)- J
APPENDIX B
Formulas for the second order subdeterminants o f t h e H a s k e l l l a y e r m a t r i x are:

Am l~ = A'~ ]~4 = --2q'm(~'m - - 1) + [2~'m2 - - 2~'m + 1]CPCQ

2 2 --1 p
-- [~'mr~mr~m + (~',~ -- 1) (r~mr~m) ]S SQ
• 2 p
A m]~ = A m[i~ = ~(pmc)[r..~S C Q + r ~ m- 1S Q C p
]
~ m ta2 = A m a2 = A m 14 A m 23 = - ( p m c )2 -1 [(2~m
L14 123 1"4
:
134 - 1)(1 -- CPCQ)

+ ( (~,,~ -- 1 ) (r.mr~m)-I + "ymr.mrOm)SPSQ]

• 2 --1 --1
A "~ [~ = A m ]~ = --,(pmc ) [r.mS p CQ + ra.~.~QCP]

A m 1~] = (pmc2 )-212 (1 - CPCQ) + (r.mr~m + (r.mram)-l)SPSQ]

Am i~a
j~2 = A m I~ = ~• (pmC)[
2 (~/m
-- 1) 2r -.1~ S P C Q + "Ym2r ~ S Q C p ]

Am 118
[13 = A m 124
24 = CPCQ

A m [13 A m 13 A m 14 Am [23 --1

1~4 = [2~ = ]24 = ~ 124 = i['ymr~mSQCP + ('Ym -- 1 )r.mSPCQ]
--1
Am I~] = r.mr~mS p SQ

Am]~ = Am]~ = Am]]~ = A m ] ~ = (pmC2)[~'m(~/m- 1 ) ( 2 5 ' m - 1)(1 - - CPCQ)

+ ((%~ - - 1 ) 3 ( r . m r ~ ) -1 + 7mSr.mr~m)SPSQ]
-1 p
Am I~ = Am [~ = A ~ ]~ = A ~ I~ = -i[(~ - 1)r~:~QC + .ymr.~SPCQ]

Aml~ = Aml~ = 1 + 2~m(~m - - 1 ) ( 1 - - CPCQ)

+ [('y2r.mram + ('Ym -- 1)2(r.mr~m)-l]SPSQ

166 BULLETIN OF THE SEISMOLOGICAL SOCIETY OF AMERICA

Am 1~ = A ~ I~ = A " 112 - 1

= (pmc)[7.~ r.mSPCQ + ('Ym -- 1)2r-~iSQCP]

= r~mra,~St SQ

(p,~)[(2,y~(,y,. 1)~)(1 CPCQ)

F o r m u l a s for the s u b d e t e r m i n a n t s of t h e half-space m a t r i x are:

~: = -2(~/~,~) ~
Oln2~[n ran ?'fin

~ n ] 12 2 --1
~3 = - (p~ ~ , ~ ~,~ r ~ . )

1 - - ~Yn
pn Oln2~n ran rSn

24 : (p. otn ~n r . ~ ) -~

34 Pn C20~n2"~n ran r~n "

REFERENCES
Dorman, J., M. Ewing and J. Oliver (1960). Study of shear velocity distribution by mantle Ray-
leigh Waves. Bull. Seism. Soc. Am. 50, 87-115.
Dunkin, J. (1965). Computation of modal solutions in layered, elastic media at high frequencies,
Bull. Seism. Soc. Am. 55,335-358.
Gilbert, F. and G. Backus (1966). Propagator matrices in elastic wave and vibration problems,
Geophysics 31,326-332.
Haskell, N. A. (1953). The dispersion of surface waves on multilayered media. Bull. Seism. Soc.
Am. 43, 17-34.
Knopoff, L. (1964). A matrix method for elastic wave problems, Bull. Seism. Soc. Am. 54,431-438.
Press, F., D. Harkrider and C. A. Seafeldt (1961). A fast, convenient program for computation of
surface wave dispersion curves on multilayered media, Bull. Seism. Soc. Am. 51,495-502.
Randall, M. J. (1967). Fast programs for layered half-space problems, Bull. Seism. Soc. Am. 57,
1299-1315.
Thomson, W. T. (1950). Transmission of elastic waves through a stratified solid medium, J.
Appl. Phys. 21, 89-93.
Thrower, E. N. (1965). The computation of the dispersion of elastic waves in layerea media, J.
Sound Vib. 2,210-226.

DEPAI~TMENT OF EARTH AND PLANETARY SCIENCES

UNIVERSITY OF PITTSBURGH
PITTSBURGH~ PENNSYLVANIA
CONT~IBVTION No. DEPS-69-215

Manuscript received July 19, 1969.