647

(1985), 72, 3, pp. 647-52

Printed in Great Britain

Tests of fit for the von Mises distribution

BY
Simon

A. LOCKHART AND M . A. STEPHENS

Department of Mathematics and Statistics,
University,
British Columbia, V5A 1S6, Canada

Critical points for the asymptotic distribution of Watson's U2 statistic are given
for testing the hypothesis that a sample comes from the two-parameter von Mises
distribution on the eircle when one or more of the parameters is unknown. Monte Carlo
results show that convergence to the asymptotic distribution is quick. Dependence of the
points on the shape parameter is discussed and seen not to cause any problems.
Some key words: Empirical distribution function test; Watson's

1. INTRODUCTION

The von Mises distribution is often used to describe unimodal data on the circumference of a circle. I n this paper we give a
test based on the empirical
2
distribution function statistic U for the hypothesis that a sample comes from this
distribution with unknown parameters to be estimated from the data.
Suppose the circle has centre 0 and radius one, and let a radius OP be measured in
radians by the polar coordinate 0 from some fixed reference radius ON. Let
be the
coordinate of a radius OA, and let be a positive constant. The von Mises density is
0 < 2n).
Here
is the imaginary
function of order zero. The distribution has a mode
along OA, that is, at 9 =
and is symmetric about OA; for = 0 the distribution
reduces to the uniform distribution around the circle. Suppose a random sample of points
on the circumference of the circle has coordinates
We discuss the test
of the hypothesis
that
random sample of
comes from the von Mises
distribution (1). Four cases may be distinguished according to which parameters, if any,
are known:
Case 0:
and are known, so that f(9;
Case 1:
is unknown and is known.
Case 2:
is known and K is unknown.
Case 3: Both
and K are unknown.

is completely

Maximum likelihood estimates of

and are found as follows. Let be the length of
the vector sum or resultant of the vectors
(i =
n). The estimate of
is the
direction of
the vector sum, and the estimate k of is the solution to
= R/n,

(2)

where
is the imaginary Bessel function of order one. Tables for solving (2) are given,
for example, by Pearson & Hartley (1972, p. 364) and Mardia (1972, p. 297). When OA,
that is
is known, let X be the component of the vector sum along OA; then the

648
estimate of

R. A .

AND M . A . STEPHENS

now satisfies
= X/n.

(3)

See 4 for a discussion of negative estimates of in (3).

The tost statistic to be discussed is Watson's statistic U2 defined as follows. Let
be the empirical distribution function of the
values 0,, that is,
is the proportion of
0, for 0 0
Define

For ease of notation let F(0)

with slight changes in notation,

The definition of U2 is, from Watson (1961)

(4)

Clearly U2 measures the discrepancy between F and F. The statistic is an adaptation of

the well-known Cram r -von Mises statistic W2; the adaptation is necessary to make U2
independent of the choice of origin for 0 around the circle.
In 2 , we set out the test procedure. I n 3 we present asymptotic theory and the
results of Monte Carlo studies. I n 4 , wo comment on the dependence of the points on
Particular attention is given to large and small K.

2. T H E

A test of

TEST

is made as follows:

(i) for the appropriate case, estimate unknown parameters as described above;
(ii) calculate
=
where
and K are replaced by their estimates i f
necessary;
(iii) put the
in ascending order to obtain
< ... <
(iv) calculate U2 as
U2 =

+ l/(12n),

where z =
is the sample
of the z values;
(v) for Case 0 calculato U*
(vi) for Case 0 refer 17* to Table 1; for other cases
to the part of
1
appropriate to the given
using the value of K, or its estimate if necessary, to
enter the tablo; reject
at approximate significance level a i f U* or
exceeds
the point given for level a.
For Case 0 exact points for finite n havo been
by Stephens (1964). These
load to the modification U* which, together with the asymptotic points, gives
an exact test. For other cases the levels for U2 are approximate since only asymptotic
points are given; however, these will be accurate for practical purposes for
20. This
has been verified by Monte Carlo studies for finite n and is consistent with the behaviour
of the null U2 distribution in many other situations;
for example, Stephens (1974,
1977). The asymptotic points for U2 have been obtained from the theoretical calculations of 3 .

Tests of fit for von Mises

T a b l e 1.

distribution

649

Upper tail percentage points for

015

Significance level a
010
0025

001
0 005
(a) Case 0: both parameters known
All
0069
0105
0131
0152
0187
(b)
1: shape porameter known
00
0071
0089
0105
0133
0048
0-072
0091
0107
0135
100
0051
0095
0111
0169
0139
0-080
0053
0-115
0144
0173
200
0082
400
0124
0086
0153
0-183
oo
0-059
0089
0127
0157
(c) Case 2:
known
00
0071
0089
0105
0133
0163
0-204
050
0091
0135
0165
100
0076
0139
0053
0080
0100
0116
0144
200
0082
0103
0119
400
0085
0106
0122
0181
0151
oo
0105
0-085
0122
0180
(d) Case 3:
parameter known
0-030
0040
0052
0-069
0081
0031
0050
0100
100
0035
0049
0066
0092
0122
0056
0067
0077
0092
0108
0039
0128
200
0074
0043
0061
0084
0101
0142
400
0-047
0067
0082
0132
0158
0178
oo
0 048
0069
0084
0137
For K > 4 use linear interpolation in
For Cases 2 and 3 enter the table at the estimate

3. THEORY OF T H E TESTS

The statistic

may be written as

where the integral is over

Here

,
whore F is the empirical distribution function of the z,. The
behaviour of
has been studied extensively; wo givo only the details needed for the present calculations. Durbin (1973)
the theoretical development for both unknown and known
parameters.
As n
oo, Y
weakly to a Gaussian process Y with mean 0 and covariance
function p(s, t) which
on the case. For Case 0,
p(s,t)
For other cases p{s,t) =
Case 2,
+
Set M(K)

t)R(s,t),
where R is
for Case 3, and
N(K) =

for Case 1,
as follows.

for

650

R.

A . L O O K H A R T AND M . A .

STEPHENS

Then
ro

J
+

where 0 is related to s by s =
0,
The derivation follows the analysis of Stephens
(1976) for the normal and exponential distributions.
The asymptotic distribution of U2 is the same as that of S =
where
are
independent
variables and
... are the eigenvalues of the Fredholm equation
f(s)
Effectively, therefore, the problem is to find the eigenvalues
for the different cases and
then caloulate the distribution of S.
The eigenvalues
are found (Stephens,
by Fourier expansion as follows. Let
2
(2nj) (j =
the
eigenvalues are one set of the
and the set of roots
of the equation
S(y)

= 0,

(5)

where
sin

af =

=
Jo

For Case 2 the

are one set of the

= 0,

\
Jo

{2njF(9;
and the

of roots of
0,

(6)

where
=
=

P(K)

0,

0, K)d9,
= 0.

For Case 3 the eigenvalues are the two sets

I n practice we calculate af and ft; by numerical integration for j =
where we
find J = 50 is a suitable limit. Equations (5) and (6) are then solved numerically for
The
40 works well.
the eigenvalues for a particular
case have been found the percentage points of S are found by Imhof's method, as
adapted for an infinite sot of eigenvalues by Durbin & K n o t t (1972).
Since the computations use many approximations i t is important to have
on
their numerical accuracy. The cumulants S wore calculated in several ways; details are
available from the authors. Having correct values for the cumulants in hand, we fitted
Pearson curves to check the accuracy of Table 1.

Tests of

for von Mises

distribution

651

4. INFLUENCE OF T H E SHAPE PARAMETER

2

Since U is location invariant, its distribution does not depend on

i t does, however,
depend on even asymptotically, except in Case 0 where the test is distribution free. I n
Oases 2 and 3, where is unknown, an estimate k must be used to onter Table 1. This
introduces an extra approximation beyond the asymptotic one since the Table may be
ontored at the incorrect line.
the effect is negligible since k is consistent
and the points depend continuously on Even for finite samples simulations show that
the points depend only slightly on for reasonably large I n fact, then, little error will
result for such unless k is seriously in error. For small values of the points depend
more strongly on x; however in a practical situation in which i t appears that x is small
ono would probably first test for randomness of the points on the circle. For this test
simply use Case 0 above with x =
= 0, calculate U* as described above, and refer to
Table 1.
The levels of the tests outlined here for Case 3 were checked by Monte Carlo simulation
based on 10,000 trials for x = 0 5, 2, 5 and 10 and n = 20 and 40 for tests at a nominal
level of 0 05, using an expanded version of Table 1. I n all oases the attained significance
level was between 0 045 and 0 050.
For n 20 the tests were found to bo useful in practice with the attained significance
levels boing slightly lower than the nominal
for small x. A similar pattern prevails
for other cases.
I n Tablo 1 points are provided for x 0 and x = oo. These points deserve special
mention. I n each case
x
oo the covariance function p converges in L 2 to the
covariance function of the corresponding case for the normal distribution; see Stephons
(1976). The critical points thus also converge. The point x = 0 is actually part of the
parameter space but when x = 0,
is unidentifiable. For Case 1 the points given
correspond to the covariance function obtained as x
0. For Cases 2 and 3 the
parameter space is identified with the plane via
=
=
I n Case 3 the
problem is then regular at T =
and the asymptotic theory of Durbin (1973) applies.
I n Case 2 the problem is regular at x =
only if negative estimates of x are allowed as
solutions of (3). Thus the problem is regular i f the parameter space is regarded as the line
making an angle
with ON but irregular at
i f the parameter space is taken as a
half line making this angle. I n practice thon when, in Case 2, a negative estimate k solves
(3) we recommend replacing
by
+ n and x by x*
x. Now follow the procedure as
usual using
to onter Table 1.

ENTS

This work was supported in part by the Natural Soience and Engineering Research
Counoil of Canada, and the U.S. Office of Naval Research.

REFERENCES
J . (1973). Distribution Theory for Tests Based on the Sample Distribution Function. Regional
conference series in
Math.,
Philadelphia: S 1 A M .
J . & K N O T T , M.
Components of
Mises statistics I . J.
Statist,
B 34,
K . V . (1972). Statistics of Directional Data. London: Academic Press.
PBABSON, E . S . & H A R T L E Y , H. O .
Tables for Statisticians, 2. High
STEPHENS, M. A. (1964). The distribution of the goodness-of-fit statistic
Biometrika

Griffin.
393-7.

652

R.

STEPHENS, M . A. (1974).

A.

AND M . A . S T E P H E N S

statistics for goodness-of-flt and some comparisons. J. Am. Statist. Assoc. 6 9 ,

730-7.

STEPHENS, M . A. (1976). Asymptotio rasults for goodness-of-flt statistics with unknown parameters. Ann.
4, 3 5 7 - 6 9 .

STEPHENS, M . A. (1977). Goodness-of-flt for the extreme-value distribution. Biometrika 6 4 , 5 8 3 - 8 .

WATSON, G. S .
Goodness-of-flt tests on a cirole. Biometrika 4 8 ,

[Received November 1984. Revised March 1985]