Cartwright Longuet

Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

The statistical distribution of the maxima

of a random function
BY D. E. CARTWRIGHT AND M. S. LONGUET-HIGGINS
National Institute of Oceanography, Wormley

(Communicated by G. E. R. Deacon, F.R.S.-Received 14 April 1956)

This paper studies the statistical distribution of the maximum values of a random function
which is the sum of an infinite number of sine waves in random phase. The results are
applied to sea waves and to the pitching and rolling motion of a ship.

INTRODUCTION
Let f(t) denote a continuous, random function of the time t, representing, for
example, the height of the sea surface above a fixed point. It is interesting to inquire
into the statistical distribution of the heights of the maxima off(t).
There are two distinct problems. On the one hand we may consider the total wave
height 2a, being defined as the difference in level between a crest (maximum) and
the preceding trough (minimum). The statistical distribution of a is difficult to
determine in the general case, but whenf(t) has a narrow frequency spectrum it may
be shown that a is distributed according to a Rayleigh distribution
2a
p(a) = e-a/mo,
mo

where mo is the root-mean-square value off(t) (see Rayleigh x880). This distribution
has been compared with the observed distribution of the heights of sea waves and it
has been shown that many theoretical relations, for example the ratios of the mean
wave height to the mean of the highest one-third waves or to the mean of the
highest of N consecutive waves, are in close agreement with observation (Longuet-
Higgins I952). Application of the X2-test to some histograms of wave heights has
also indicated, apparently, no significant departure from the Rayleigh distribution
(Watters I953). It is certain, however, that for functions f(t) having a broad
frequency spectrum, the theoretical distribution of a must be different from the
Rayleigh distribution.
Alternatively, we may consider the difference in height 6 between a crest and the
mean level of the function f(t). Although in practice 6 may be less convenient to
measure than a (since the appropriate mean value is sometimes difficult to deter-
mine) the theoretical distribution of ? is easier to obtain, and has been found for
a wide class of random functions by Rice (1944, 945) in connexion with the analysis
of electrical noise signals. Rice's solution, which is only one out of many results in
a long paper, has not been fully discussed, and it is the purpose of the present paper
to examine the solution and to calculate some of the statistical parameters associated
with it. We shall also apply the results to ocean waves and to the motion of ships
at sea.
[ 212 ]
Statistical distribution of the maxima of a random function 213
In ? 1 we outline briefly Rice's derivation of the statistical distribution of the
maxima 6. The discussion shows that the distribution depends, surprisingly, on only
two parameters: the root-mean-square value off(t), which we denote by mj, and
a parameter e which, as we show in ? 2, represents the relative width of the frequency
spectrum of f(t). When e is small, the distribution of 5 tends to a Rayleigh dis-
tribution, as we should expect, and when e approaches its maximum value 1 the
distribution of 6 tends to a Gaussian distribution.
One of the main differences between the two variables 6 and a is that 6 may take
negative values (since some maxima may lie below the mean level) whereas a is
always positive. The proportion r of maxima that are negative can be readily
determined in practice, and in ? 3 we show that this proportion depends only upon e.
Hence if r is measured, e can be estimated.
In ?? 4-6 we calculate the moments of the distribution, the mean values of the
highest 1/nth of all the crest heights and the expectation of the highest in a sample
of N crest heights, and we show how these quantities depend upon e.
The distribution of crest heights, as measured from records of ocean wave
phenomena, is compared with the theoretical distribution in ? 7. No significant
difference is found. On the other hand, the crest-to-trough heights, examined in ? 8,
are found to depart significantly from the Rayleigh distribution.

1. THE DISTRIBUTION OF MAXIMA

The random function f(t) is represented as the sum of an infinite number of


sine-waves f(t)E cn cos (nt+e+), (1.1)
n

where the frequencies c- are distributed densely in the interval (0, oo), the phases
en are random and distributed uniformly between 0 and 27r,and the amplitudes cn
are such that in any small interval of frequency doc
o-+dcr
Cc = E(cr)doC, (1.2)
On--=o

where E(o) is a continuous function of o which will be called the energy spectrum
off(t) .The total energy per unit length of record is

O= fE(o-)dor. (1.3)

More generally we shall find it convenient to write

Mn= E(C) rondo- (1.4)

for the nth moment of E(o) about the origin.


To find the distribution of maxima off(t) we note that, iff(t) has a maximum in
the interval (t, t + dt), then in this intervalf'(t) must take values in a range of width
If"(t) Idt very nearly; and the probability of this occurrence, and off simultaneously
lying in the range (61,61+ d61), is

00

14 Vol. 237. A.
214 D. E. Cartwright and M. S. Longuet-Higgins
where p(6, 2,63) is the joint probability distribution of
(61, 2, 63) = (f,f,f"). (1.6)
The mean frequency of maxima in the range 61<f < i + d6l is therefore

F(61)d =
f- [p( o,0,63) | 63 1 d] d3, (1.7)
_-00

and the probability distribution of maxima is found by dividing this distribution


by the total mean frequency of maxima, which is

N, = 3 - 0oo0 00
(1,0,63) dgld3
I . (1.8)
Now from (1.6) we have
61=f(t) = ECc os (t+et+ ),
n

62 = f'(t) = - Ccn sin (nt + en), (9)


- OS +
93 = f"(t) = Cn).
y
n
1,g2,63are therefore each the sum of an infinite number of variables of zero
expectation and random phase. Therefore, by the central limit theorem in three
dimensions, the joint probability distribution of (6,, 25,3) is normal (under general
conditions assumed to be satisfied by the amplitudes cn; see Rice 1944, I945). The
matrix of correlations or statistical averages -Ei = gi6 is seen to be
Mo 0 -m2

() -= 0 m2 . (1.10)
-m2 0 m/
Hence
p(1 2, 3) = exp {- f~21'
m 2 + (m4g2 + 2m2 1~3 + m03)/A]},
(2T) (Am)
(1.11)
where A = mom4-m2. (1.12)
Substituting in (1.7) we have

F(61) (2=r)(Am2) exp{-?(m4~+2m2l13?+ m05)/A} j}lSi dS3. (1.13)

On evaluating the integral and writing


ll/m = , Ai/W/n2= (1-14)
we obtain
F 2 + (0/)J
( m=
7) m [ei2/&2 e-2 dx] . (1.15)
M
(27T)mi J -V14

The last integral can be expressed in terms of the known function

erfx = -2) eX d
d16) (1
\~/ Jo
Statistical distribution of the maxima of a randomfunction 215

rt

"0

Cn

I I
4-r

.i.
QS

- - 4
0

3 0

CD

CO
Pc

T"-

I4-2
216 D. E. Cartwright and M. S. Longuet-Higgins
The probability distribution of y is ml times the distribution of 6:

p(y) = mnp(g) = mF(6i)/N. (1.17)

From (1.8) we find N = (1.18)


27r\m2j
and so finally
(21
P()= [
+e1e-2/62+ -2)e-+2J e-dx , (1'19)
(27T)l L J -oo J

? 92 = A = mm4 4-2
m2
where 62 = (1.20)
1+ 2 m0m4 m0m4
The function f(t) is statistically symmetrical about the mean level t = 0. For, in
equation (1.1) each phase angle en might be increased or diminished by 7r without
affecting the random character of the phases; and this would merely reverse the
sign off(t). It follows that the statistical distribution of the minima is simply the
reflexion of (1-20) in the mean level y = 0.

2. DIsCUSSION
In equation (1.19) r denotes the ratio of the surface height to the r.m.s. height
mO.We see that the distribution of y depends only on the single parameter e. A simple
interpretation of e is as follows. From (1 12) we have

A = E E(r2) - do- d2.


m0m4 -m=f- E(r (-2 -2) (2.1)

On interchanging eo and r2, and adding, we have

2A
-f fE(ol)E(o2)(r2
-
o-2)2do-ldro2. (2.2)
JoJo
Since E(cr) is essentially positive, it follows that A > 0 and so
0<e<l. (2-3)
For a very narrow spectrum, with the energy grouped around o- = o, say, E(cr1)and
E(cr2) are small except when o- and a2 are both near to o-; but then the factor
(o21- 2)2 in (2.2) is small and so
e<l1 (2.4)
In general e is a measure of the r.m.s. width of the energy spectrum E.
Clearly e may take values indefinitely near 0. For a low-pass filter (E = Eo when
or< ao, and E = 0 when or> r0) we find
e= (2.5)
e may also take values indefinitely near 1. For suppose a proportion m of the energy
is at frequency Co= o1, and (1 -w) at o = r2; we have

m9 = mo{wrol + ( 1- ) (2r},
m4 = +
m0o{Joi (1- w)o}.J
Statistical distribution of the maxima of a random function 217
When cr2/0c -> oo we see that m2/mom4 - 1- and so

e62-> , (2.7)
which can be as near to unity as we please.
The first limiting case (e->0) gives the distribution for an infinitely narrow
spectrum. From equation (1-19) we have then

P(Y)= {? o (^ (y ~0),)),}
(K (2.8)
which is the Rayleigh distribution, or the distribution of the envelope of the waves
(see Rice 1944, '945; Barber 1950; Longuet-Higgins 1952).
The second limiting case (e-> 1) can occur, as we have shown, when one wave of
high frequency and small amplitude is superposed on another disturbance of lower
frequency. The high-frequency wave forms a 'ripple' on the remaining waves, and
the distribution of maxima tends to the distribution of the surface elevation
(4l/mo) itself. On letting e tend to 1 in (1.19) we obtain

p() (2 e)I (2.9)

which, as we should expect, is a Gaussian distribution.


The distribution p(y) has been plotted in figure 1 for e = 0.0, 0-2, ..., 10. The
transition from the Rayleigh distribution to the Gaussian distribution can be clearly
seen.
3. THE PROPORTION OF NEGATIVE MAXIMA
This may be found by a simple geometrical argument as follows. Suppose that in
a certain interval of time, say (0, t), there are n+ zero up-crossings, at whichf passes
from negative to positive values, and similarly suppose that there are n- zero down-
crossings. Also let there be n+ positive maxima, n- negative maxima, n+ positive
minima and n- negative minima. Between a zero up-crossing and the next zero
down-crossing the function is always positive, and so the number of maxima exceeds
the number of minima by one. In other words, when n6 increases by 1, so also does
(nj - n+). Similarly, when no increases by 1, so does (n - nf). Therefore, if NO+,
No, N+j, Ni, N+, N- denote the average densities of zero up-crossings, etc., over
a long interval we have =
N+-
N NV - N.,
No = N,-NjJ (3.1)
Now since f(t) is statistically symmetrical about the mean level it follows that
N = N+
N = rN,
(32)
N_ = N+ = ( -r)N,
where N, denotes the total density of maxima, and r denotes the proportion of
negative maxima. So from (3-1)
N+- =N (l- 2r), (3-3)
or r = (1-No+IN+/ ). (3-4)
218 D. E. Cartwright and M. S. Longuet-Higgins
But from Rice (I944, 945) and equation (1.18) we have

1 1 (m4]~
2(
N+='= 2,' (m2]3 N1= (3.5)

So equation (3-4) can be written

r= 1-= t 1
[-(-e)] . (3.6)
21 (momM4)l
Hence the proportion of negative maxima increases steadily with the relative width
of the spectrum. Conversely, we have

62 = 1- ( - 2r)2. (3.7)

This relation provides us with a ready means of estimating e by simply counting the
numbers of positive and negative maxima in a length of record.

4. THE MOMENTSOF P(y)


The nth moment /t of the probability distribution p(y) taken about the origin,
is defined by
rUr
(Vpw
oo
p ^n
dV. (4-1)

The even moments (n = 2r) may be calculated by means of the moment-generating


function
Je-W(t)dp() d =' - 21! + 2 /.... (4-2)

On substituting from (1.19) and evaluating the integral we find

f f^ e-A(nt)'p(y)dy - (1 + e2t2)~(1 + t2)-1, (43)


ne-iW^f )d^= (l+e^2)a(l+^)-1 (4-3)

and so on, comparing coefficients of t2r in these two equations, we have

=
^ /. [1-2
P2r i2- 1.1 e 1.1.3... (2r- (4 4)
22.2! "2r.r!

The odd moments (n = 2r+ 1) may be found in a similar way by means of the
moment-generating function
e0
o t3 t5
,te- t)2p(l) d ]/ t/;l -2./+ + ....) (4-5)
^ 2 2/ta
From (1.18) we have

f te-'(nt)p()d] = (r) (1 -2)t(1 + t2)-, (4-6)


a 00

and sc2r+1=(W77j( 1 1. 3.
3 5 ;(2r(7 + 1)
anclhence
hence =
rCC-r/ (le2)1 (1-7)
(r!)2.
Statistical distribution of the maxima of a randomfunction 219
In particularwe have = 1,
,6,=1-(ilt)i(1-e2)-^,p (4.8)
/4 = 2-e2,
/-1 = (1)i (1 - 62) .3

We see that the mean ju is a steadily decreasing function of e, the width of the
spectrum. A non-dimensional quantity depending on e is the ratio
'2 1_2
P Al~ ~o~(4.9) ~~~~~~~
The width of the spectrum is given in terms of p by the relation

e =7T-
4p (4.10)

12-

1-0--

0'8 -

0'6;--

6.1

0-2 -

0 0-2 0-4 0'6 0-8 1'0

FIGURE 2. Graphs of the mean /,u, variance /,2, skewness f, proportion r of


negative maxima, and p( =/i2/,o/#') as functions of e.

On the other hand, we have the following two quantities which are independent of e:

1,
,/Co-M_2v - /^ti = 3. (4-11)
The moments /u about the mean, which are defined by

/ fn= p(j/) (/-/ )~ dr, (4.12)


Goo

may be deduced immediately from the moments about the origin. In particular
we have from (4 8) =,
= v/o 1,
/1 = O,
- O(4-13)
da3
/62 = 1--( 7T- 1) (1-e2),

/b3 = (7T) (7T- 3)- ( _62).


220 D. E. Cartwright and M. S. Longuet-Higgins

Q-

Cf)
Statistical distribution of the maxima of a randomfunction 221
The coefficient of skewness is given by

/3= ( Tr)
(27- -3)[i- (-?)1 (4.14)
ft2

We see that the standard deviation /i| steadily increases as e increases. fi, on the
other hand, steadily decreases.
The mean Iu4,the variance #2, the skewness / and the ratios r and p are shown
graphically as functions of e in figure 2.
In some practical cases we may know the distribution of the maxima 61 ( = my^)
experimentally and wish to make an estimate of the mean energy mO. Let v' and
v, denote the nth moments, about the origin and about the mean, of the variate 1.
Then V=
1
m(n+l),n, vn = m4(n+l1, (4.15)
and so from (4.11)
V
2o -71 2= , v /v' = 3m0. (4.16)

By forming either of these quantities, therefore, we may estimate mo.

5. THE CUMULATIVE PROBABILITY

The cumulative probability q(v) may be defined as the probability of y exceeding


a given value: q
d= Pd.(5.1)
q(v/p(V)
Substituting from (1.19) we find
I1JX2 rjq2 00 2 dx + ( l-e2)I e- d-62)1/6
?,(I

q(v) =(2^-- \\e (5-2)


-X(
e-X' dxz.
02t
When e-> O,

q(v)- (5.3)
e-52
l_Jr
( > 0),
I r00
and whene- 1, q() -> e-2 dx. (5.4)

Graphs of q(Q) for these and intermediate values of e are shown in figure 3. The
proportion r of negative maxima is given by

r= p() dy = 1 -q(O), (5-5)

which from (5-2) is r = 1[1-(1-e2)l], (5-6)


in agreement with (3.6).
In some geophysical applications it is found convenient to consider only the higher
waves, say the highest 1/nth of the total number in a sample. The 1/nth highest
maxima correspond to those values of y greater than y', say, where

q(/')= fp(V) d = l/n. (5.7)


222 D. E. Cartwright and M. S. Longuet-Higgins
The average value of y for these maxima will be denoted by y(1/n), so that

V(l/n)= n p(y) d]. (5-8)

Clearly ((1)is the same as the mean ,4t. (lx/n)has been computed numerically for
n = 1, 2, 3, 5 and 10, and for different values of e. The results are shown in figure 4.
(li/n)is apparently a decreasing function of e. For small values of e, say e < 0 5, the
dependence of y(l/fl) on e is slight, but each curve gradually steepens, and it can be
shown that as e approaches 1 the gradient yI(1ln)/ls tends to -oo. Near e = 1 the
curves are all exactly similar in shape, being independent of n.

n=10
2.5--

20' 3

(ln)\ \

105
0.

0 0-2 0-4 0.6 0'8 1.0

FIGURE 4. Graphs of (lln), the mean height of the 1/nth highest maxima,
as a function of e, for n = 1, 2, 3, 5 and 10.

6. THE HIGHESTMAXIMUMIN A SAMPLEOF N

Suppose that a sample of N maxima is chosen at random; we wish to know the


average value of the highest of these, Ymax*.The problem has been considered in the
case e = 0 (Longuet-Higgins 1952) and the expectation Ymax.has been computed for
values of N up to 20. For values of N greater than 50 (in which we are usually
interested) it has been shown that the asymptotic formula
Vmax./l( #) - (ln N)l +iy(ln N)- (6.1)
is accurate to within 3 %. (Here y denotes Euler's constant, 0-5772....)
The formula (6*1) may be generalized to values of e between 0 and 1 as follows.
The probability distribution of Ymax.is given by*

a max.)s= d tmax'[1 - q ( qmax.)IN, (6s2)


* We follow here the same method as in the
paper just quoted. But a general study of
the limiting form of the distribution of the largest member of a sample has been made by
Fisher & Tippett (1928). For a more recent discussion see Gumbel (1954).
2

Statistical distribution of the maxima of a randomfunction 223


where q(y) is given by (5.1). Therefore we have

_
= -f
Vmax. - q(y)]Ndy. (6.3)
On separating the integral into two parts, from -oo to 0 and from 0 to oo, and
integrating by parts we find

Vrmax.[= l- q(y)]Nd+{1-[1-q(y)]N}dq. (6.4)

When N is large [1 _q(,)]N is very small unless q is of order 1/N. Now as x tends
to infinity we have 00
e-2 2dx=e +-i2-+0 (6-5)
and so from (5-2)

q(y) = (1-e62) e-2 + o e-42/2) (6.6)

for large values of q and when 0 < e < 1. If q is of order 1/N, V is of order (InV
N).
Therefore neglecting terms of order (ln N)- we have

q(]) = (1 -2)1 e-j2 - (1 -e2) e-0, (6-7)


where 0 =?V2. The first integral in (6-4) is negligible, and on substituting in the
second we have
iax. = )
"11^1"^16"^0"^-
{1- (1-6)e-
1- 0]N}0- d0. (6.8)
(6-8)
22.1o"0

Writing 0 =log[(l-e2))N], ' = 0-00, (6-9)


e_o,
and so e-o = (6-10)
(1-e2) N'

we have = I 1- 1-N- }(0o+ 9')-idO (6.11)

*-r (1-exp[-e-']) (0o+0')-d0,, (6.12)

with relative errors of order 1/N only. It may be shown (Longuet-Higgins 1952)
that when 00 is large the above integral equals

2[0 + yO-A + O((0-)]. (6.13)


Hence we have
max.-2!{[n (61- N] + ~y[ln (I - e2)' N],
2) ln- (614)
which can also be written
-- /!(, ) [ln (1 -62) Ni] +1[ln (1 - 2) Ni]- ~
~Vmax./(#2- n 1\-- .(6.15)
(1_-e 2)

When e - 0 this equation reduces to (6.1). The expression on the right-hand side of
(6-15) is an increasing function of e, when N is large. It follows that as the spectrum
broadens, the ratio of the greatest in a sample to the root-mean-square will tend
to increase.
224 D. E. Cartwright and M. S. Longuet-Higgins

When e approaches 1 (so that In (1 - e2)i N is not large compared with 1) the above
formula is no longer valid. The corresponding expression for the general case is
complicated and probably not of practical importance. We shall simply give the
limiting form when e- 1, and p(V) is normal (equation (2.9)). Fisher & Tippett
(1928) have shown that the average value of Vmx. in this case is given by

ym
Ymax. = m+ ^
+=M+^^2 (6-16)
(616)

approximately, where m is the mode of the distribution of ?max.,given by


(27r)m e-2 = N. (6-17)

From (6.17) we have m2 = In - lnm2, (6-18)


N27
19
and so mL n-( 27Tlnl) ( )] (6.19)
L 27r/J
The leading term in (6.16) is thus

ma= 2[In N (6.20)


However, Fisher & Tippett have shown (1928) that for the normal distribution the
limiting forms are approached exceptionally slowly. A table of the exact values of
rmax.computed for values of N up to 1000 is given by Tippett (I925).

7. APPLICATIONS
It is interesting to verify that the distribution just discussed is applicable to
records of sea waves and of associated phenomena. In this section we shall consider
five such examples: a record of wave pressure at a fixed point on the sea bed; two
continuous records of wave height made at sea by a shipborne instrument; one
record of the angle of pitch of the ship, and one of the angle of roll. The widths of the
corresponding Fourier spectra are fairly representative of the possible range 0 < e < 1.
Typical sections of the records are shown in figure 5 (a) to (e). Each complete
record lasted from 12 to 20 min and contained about 100 maxima and 100 minima.
In order to increase the amount of data both maxima and minima were included in
the sample. The analysis was carried out as follows. The ordinates An of all the
stationary points in the record, measured from some common baseline, were
numbered consecutively from 1 to N so that the maxima, say, corresponded to even
values of n and the minima to odd values of n. The zero of the record was taken to
be the mean of A: 1
A=- A. (7.1)
The distribution of the variate

Xn =(-1)(An-A) (7-2)
was then studied. The histograms corresponding to the distribution of Xn are shown
in figure 6 (a) to (e).
Statistical distribution of the maxima of a random function 225
To obtain the parameters for the theoretical distribution a harmonic analysis of
the original record was made by means of the N.I.O. Fourier analyser (see Darby-
shire & Tucker I953). The range of frequency was divided into a number of equal

aft.e
water (a)

mil
1 mil
I

(b)

l1 miin

A
20ft.[ I (c)

1Iin

10 [ A .1 i i.k AA i AA^ A A IA A AA
a (d)

min
I min

204[ jjl (e)

Imin
FIGURE 5. Typical short sections of the five records chosen for analysis. (a) pressure on the
sea bed off Pendeen, Cornwall, 08.00 to 08.20, 15 March 1945; (b) wave height in the Bay
of Biscay, 19.00 to 19.12, 11 November 1954; (c) wave height in the Bay of Biscay,
02.00 to 02.12, 12 November 1954; (d) angle of pitch of R.R.S. Discovery II, in North
Atlantic, 13.21 to 13.33, 25 May 1954; (e) angle of roll of R.R.S. Discovery II, in North
Atlantic, 14.05 to 14.17, 21 May 1954

narrow ranges each containing about 10 harmonics of the length of the record, and
the energy SEc2 was summed for each interval. The energy spectra are shown in
figure 7 (a) to (e). The moments m0, m2 and m4 of the distribution were then calculated
by multiplying the energy in each small range of frequency by 1, ar2 and oa4
226 D. E. Cartwright and M. S. Longuet-Higgins

m~
1e-

it

0 ?
0
__ =

o
I . I o
I I *J ^
c
^c<

0
c)
, |

. I I I N
b
O

I I I I
o 0
*^f cs

to
III i,
0

I I I I EI (,- .I

E (f.28s) E (deg.2s)
no. of maxima no. of crests no
per ft. of pressure per ft. of height per f

4<
(b)

excess pressure above mean height above mean


(ft. of water) level (ft.)

no. of pitches no. of rolls


per degree per degree

60

N (d)

40
j 7

e=0-49
i ,1
-2 0 2 4 6
angle of pitch (deg.) angle of roll (deg.)
228 D. E. Cartwright and M. S. Longuet-Higgins
respectively. From these three moments the parameter e defined by equation (1.20)
was calculated. The corresponding curves of probability p(y), multiplied by the
total number N in each sample, are shown in figure 6 (a) to (e).
In constructing the histograms the horizontal scale has been divided, not into
equal intervals, but into intervals such that the expected numbers of maxima in
each interval (according to the theoretical distribution) are equal. The purpose is
to avoid the small classes that must otherwise occur at the two ends of the dis-
tribution, and which make the application of the X2significance test unsatisfactory
unless the classes are amalgamated in some arbitrary way. The vertical scale is so
chosen that, for each separate subclass, a rectangle whose height indicated the
expected frequency of maxima would enclose the same area as is enclosed by the
curve of theoretical frequency. The width of the two outermost rectangles is chosen
quite arbitrarily, but this does not affect in any way the application of the X2 test.
Some relevant data concerning the five records are given in table 1. The first record
is of wave pressure measured on the sea bed in a depth of 110 ft. of water by a power-
phone pressure recorder, in March 1945 (described by Barber & Ursell, I948). The
section of record in figure 5 (a) indicates a long, regular swell with a fairly narrow
spectrum (e = 0.41). However, it contains a certain amount of energy outside the
main frequency band.

TABLE 1. DATA FOR THE RECORDS IN FIGURES 5 TO 7

(from energy e e
example N spectrum) P(X2) (from r) (from p)
(a) 164 0.41 0'60 0-31 0'37
(b) 220 0.57 0.62 0.58 0.66
(c) 270 0.67 0.55 0.68 0.69
(d) 180 0.48 0.67 0.44 0.45
(e) 250 0.20 0.12 0.26

The second and third records are of waves in deep water (Bay of Biscay) measured
by the shipborne wave recorder installed in R.R.S. Discovery II. The instrument has
been described by Tucker (1952). The two records are somewhat more irregular than
the pressure record and have correspondingly broader spectra (e = 0-57 and
e = 0-67 respectively). This is due partly to the fact that the records of wave height
contain more energy of higher frequency than the record of pressure.
The last two records are of the pitching and rolling motion of R.R.S. Discovery II
in a seaway in the North Atlantic. The angles of pitch and roll were measured in the
conventional manner by gyroscopes. The roll, in particular, has a very narrow
spectrum (e = 0-20) and the record is correspondingly regular. This is as we should
expect, since the rolling motion of a ship is only lightly damped, and is tuned sharply
to oscillations having a period close to its period of free motion.
For each of the above records the quantity X2 was calculated, and also the
probability of X2 exceeding this value. Since two parameters have been estimated
from the sample (the mean height and the total frequency) x2 has in each case
8 degrees of freedom. From table 1 it will be seen that for none of the records is the
probability of x2 significantly small.
Statistical distribution of the maxima of a random function 229
For each measured sample of Xn the quantities r (the proportion of negative
maxima) and p (= /il2/1/tl) have been found, and from the relations (3.7) and
(4 10) two independent estimates of chave been made. These are also given in table 1.
It will be seen that in examples (b), (c) and (d) the values of e are in good agreement
with that derived from the moments of the energy function E(C). In examples
(a) and (e) the estimate derived from r is not in such good agreement, but this is
hardly surprising, since the number of negative maxima on which the estimate is
based is rather small. In example 5, the estimate derived from p gives a small
negative value for 62, which is of course impossible. In all the other cases the
alternative estimates of e are so close to the original estimate as to make no
significant difference to the probability of X2.

8. CREST-TO-TROUGH WAVE HEIGHTS

In view of the agreement of the observed distributions of the heights of crests


with the theoretical distribution it is interesting to study also the distribution of the
crest-to-trough wave heights in the same records.
The local crest-to-trough wave amplitude a, may be defined as half the absolute
difference in height between a crest and the preceding trough, or between a trough
and the preceding crest. Thus
an = (Xn+Xn_). (8.1)

The statistical distribution of an is more difficult to obtain theoretically than that


of Xn for general values of e. However, when e < 1 the functionf(t) is a regular sine-
wave with slowly varying phase and amplitude, so that an = Xn very nearly. So we
may expect an to be distributed according to the Rayleigh distribution (2-8). By
considering a disturbance consisting of a small ripple superposed on a long wave
(e 1) it can be seen that the distribution of an must in general be different from the
Rayleigh distribution, though not necessarily by very much. The general distribu-
tion no doubt depends on other parameters besides e. Yet it is reasonable to expect
that for small values of e the observed distribution of an will be in better agreement
with the Rayleigh distribution than for larger values of e.
In figure 8 are shown the observed distributions of an in the five examples
discussed in ?7, together with the corresponding Rayleigh distributions

p(a) = e-a2

where a is the root-mean-square wave amplitude. The values of X2 and P(X2) are
given in table 2. (X2again has 8 degrees of freedom, since two parameters-in this
case the total number in the sample and the root-mean-square amplitude-have
been estimated.)
The table shows that the records with the smallest value of e (examples (a), (d)
and (e)) do not give significantly small values of P(X2). On the other hand, those
with the two largest values of e give very significant values of P(X2). This verifies
our expectation that the observed distribution departs more from the Rayleigh
distribution as the width of the energy spectrum increases.

I5 Vol. 237. A.
230 D. E. Cartwright and M. S. Longuet-Higgins
From figure 8 it will be seen that the records with the two broad spectra deviate
especially from the Rayleigh distribution for low values of the wave amplitude,
having relatively more waves in that range. It appears that the mode of the dis-
tribution has a tendency to move to the left in the broader spectra.

0 400 m
P.,
(a)

g
4
"S 200

6
/i? _. pressure amplitude
0 025 0-5 0-75 1'0 (ft. of water)

40 '
.4 4z
2'55 50 7'5100(b)
CD

20
4-0 0
0 .
o .4
/I I I 1 wave amplitude (ft.)
0 25 50 75 100

C*

O t
I 0 cd
I)E

60-

(d)
,O C
C) ;w4

I I I
amplitude of pitch
3 4 5 6 (deg.)

O W
30
(e)
I4
O t
20
2o
Cd

10

amplitude of roll
I\--t-4?
0 10 15 (deg.)

FIGURE 8. The statistical distribution of the crest-to-trough amplitudes


for the five records shown in figure 5.
Statistical distribution of the maxima of a random function 231
Our conclusions may be compared with those of Watters (I953) who studied
histograms of wave heights of 109 records, and compared 38 of these with the
corresponding Rayleigh distributions (with variance chosen so as to give the best
fit). Although some of the values of P(X2) were low (as small as 0-05) the values taken
as a whole did not show a significant departure from the Rayleigh distributions.
There are two possible explanations for this. First, the intervals of wave height
were equal, and so there were many classes containing only very few heights. In
applying the test these classes were arbitrarily pooled, and it can be shown that in
several cases pooling the classes in a different way would have resulted in much lower
values of X2. (The difficulty is avoided by our present method of making the
theoretical classes of uniform size.) Secondly, the widths of the energy spectra of the
records studied by Watters were probably less than in examples (b) and (c) of the
present paper, which were in fact chosen on account of their exceptional breadth.

TABLE 2. DATA FOR THE DISTRIBUTIONS OF FIGURE 8

example e P(X2)
(a) 0.41 0.33
(b) 0.57 0.001
(c) 0.67 0.000
(d) 0.48 0.55
(e) 0.20 0-51

9. CONCLUSIONS
If gdenotes the height of a maximum of the random functionf(t) above the mean
=
level, and if ml is the r.m.s. value off(t), then the statistical distribution of (6/m( )
is a function only of y and one other parameter e, which defines the relative width
of the energy spectrum off(t). e lies between 0 and 1. When e-> 0, p(v) tends to
a Rayleigh distribution; when e- 1, p(s) tends to a Gaussian distribution. As
e increases from 0 to 1, the mean of p(,) gradually decreases, the variance increases
and the shewness decreases. The proportion of maxima that are negative steadily
increases. The mean height of the highest 1/1th of the waves varies little for small
values of e, but tends always to decrease. The highest maximum in a sample of
N maxima tends to decrease relative to m, but to increase relative to the r.m.s.
height of the maxima.
The records of ocean waves and of ship motion which are discussed in the present
paper show good agreement with the theoretical distributions, for various values
of e ranging from 0.20 to 0.68.
The theoretical distribution of crest-to-trough heights is known only for a narrow
spectrum (e = 0), when it is a Rayleigh distribution. In three of the examples in
this paper, for which e < 0-5 and the total number in the sample was less than 300,
there was no significant departure from the Rayleigh distribution. On the other
hand, the examples with the broadest spectra (e = 0.57 and e = 0.67) did show
significant departures.
This indicates the need for a theoretical derivation of the crest-to-trough height
distribution when e > 0. Meanwhile, for the purpose of practical prediction, it would
15-2
232 D. E. Cartwright and M. S. Longuet-Higgins
be better to deal with crest-heights rather than with crest-to-troughheights as is
customary at present.

REFERENCES
Barber, N. F. I950 Ocean waves and swell. Lecture, Instn Elect. Engrs, Lond.
Barber, N. F. & Ursell, F. I948 Phil. Trans. A, 240, 527.
Darbyshire, J. & Tucker, M. J. I953 J. Sci. Instrum. 30, 212.
Fisher, R. A. & Tippett, L. H. C. I928 Proc. Camb. Phil. Soc. 24, 180.
Gumbel, E. J. I954 Appl. Math. ser. U.S. Bur. stand. no. 33.
Longuet-Higgins, M. S. I952 J. Mar. Res. 11, 245.
Rayleigh, Lord I880 Phil. Mag. 10, 73.
Rice, S. 0. 1944 Bell Syst. Tech. J. 23, 282.
Rice, S. 0. 1945 Bell Syst. Tech. J. 24, 46.
Tippett, L. H. C. I925 Biometrika, 17, 364.
Tucker, M. J. I952 Nature, Lond., 170, 657.
Watters, J. K. I953 N.Z. J. Sci. Tech. B, 34, 409.

You might also like