Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Processes
T. W. Anderson; D. A. Darling
The Annals of Mathematical Statistics, Vol. 23, No. 2. (Jun., 1952), pp. 193-212.
Stable URL:
http://links.jstor.org/sici?sici=0003-4851%28195206%2923%3A2%3C193%3AATOC%22O%3E2.0.CO%3B2-R
The Annals of Mathematical Statistics is currently published by Institute of Mathematical Statistics.
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained
prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in
the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/journals/ims.html.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academic
journals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers,
and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community take
advantage of advances in technology. For more information regarding JSTOR, please contact support@jstor.org.
http://www.jstor.org
Wed Aug 29 17:43:32 2007
193
194
example, .Ol or .05). The main purpose of this paper is to give methods for finding the asymptotic distributions of W : and K., and, hence, approximate values
of the significance points, 21 and zz . We assume that the hypothetical distribution is continuous.
The fundamental ideas for tests of this nature are due to Kolmogorov [ll],
Smirnov [17], Cram& [2], and von Mises [19], and for large n certain tests have
been developed by them. The present paper treats these tests in somewhat
more detail, the analysis being greatly expedited by reducing the problems to
straightforward considerations in the theory of continuous Gaussian stochastic
processes. This reduction was developed by Doob [6], and used by him to give
a simplified proof of Kolmogorov's fundamental result.
The principal innovation in this paper is the incorporation of a weight function to allow more flexibility in the tests. Although we are able to make explicit
calculations for only a few simple types of weight functions, the principal mathematical problems are reduced to classical problems in the theory of differential
equations.
The function #(t), 0 S t $ 1, is to be chosen by the statistician so as to weight
the deviations according to the importance attached to various portions of the
distribution function. This choice depends on the power against the alternative
distributions considered most important. The selection of $(t) = 1 yields no2,
the criterion of von Mises, for w:, and the criterion of Kolmogorov for K , . For
to exist for all samples except a set with probability zero, it is necessary
and sufficient that the following integrals exist:
(2.6)
where
K, =
di
- ( j - 11, j - nF(xj)I I,
195
For (2.5) to hold the integrals &(t),42(t) must exist; for (2.6) to hold it is necessary and sufficient that
3. Reduction to a continuous stochastic process. Since F(x) is assumed continuous, we can make the transformation u = F(x). Then the observations are
ui = $'(xi) (i = I , 2, .. , n ) , and under the null hypothesis these can be considered as drawn from the uniform distribution between 0 and 1. Let G,(u) be
the empirical distribution derived from ul , . . , u, . Then ~ f and
, Kn are,
respectively,
(3.2)
Kn = sup
Ogu61
6I Gn(u) - u I m.
(3.4)
A.(z)
~r
B,(z)
Pr
{bl
1
I Z/$(u)
~:(u)$(u)du
5z
sup
osus1
Pr { W: 5 z ) ,
1 Yn(u)
5 zJ
= Pr
IK, 5 z},
we wish to calculate A(z) = Jim A&), n -+ GO, and B(z) = lim Bn(z),n --+ a,
if these limits exist.
For fixed ul , u2 , . . , uk the joint distribution of Yn(u1),Yn(zlz),. . , Y,(ut)
approaches a k-variate normal distribution as n -+ a. Thus the asymptotic
process is Gaussian (normal) and is specified by its mean and covariance functions. For finite n we have
E(Yn(u))= 0,
E(Yn(u)Y ,(v)) = min (u, v )
The limiting process is a Gaussian process' y(u), 0
- uv.
_I
u 5 1, for which
E(Y(~=
) ) 0,
E(y(u)y(v))= min (u,v ) - uv,
such that the probability is 1 that y(u) is continuous [6]. Putting
w.
196
T.
(3.8)
b(z)
A. DARLING
ANDERSON AND D.
P r ( sup
o$usl
I y(u) I
d#m5 z j ,
-+
lim Pr {FIYn(t)I 5 zj = Pr ( F [ y ( t ) ]5 z j .
(3.9)
n-m
= a ( z ) and B ( z ) = b(z)'
T o handle more general weight functions for the case of integrals we want to
extend this result. We shall assume that +(u) is continuous in any interval
0 < ul 5 u $ u2 < 1. Secondly uTeassume that
(3.10)
lul
1
#(t)t log log - dt,
t
1-t
dt
+ t)y(t/(l + t))
X(t)
is the Wiener process which has the property (1121 p. 242 and p. 247)
1
there exists a to such that x 2 ( t ) S 2t log log - for 0 < t < to
t
This implies that
there exists a
(3.12)
~o
such that
1 - u
y 2 ( ~5
) 2u(1 - u) log log U
(3.13)
exists with probability 1. (taking the principal value when the integral is improper). A similar argument holds for the existence of
197
Thus ['$(t)$(t)
a(z).
and
(4.3)
the Kronecker delta. I n most cases k(0,O) = k(1, 1) = 0 ; hence f,(O) = fi(l) = 0.
Since k(s, t) is positive definite, X i > 0. The series (4.1) converges absolutely
and uniformly in the unit square.
Let X1 , X2 , . . . be independently, normally distributed with means zero and
variances 1. If k(t, t) < m , then we can define
the series converges in the mean and with probability one for each t. Then
z ( t ) is a Gaussian process with Ez(t) = , O and Ez(s)z(t) = k(s, t). Thus z(t)
y ( t ) when k(s, t) = 1/$(s)
gives the same stochastic process as
[min (s, t) - st]. From this it follows that with probability 1
d$(t)
d$(t)
198
T h e infinite product converges absolutely and uniformly for all real u, and in
general 1/X, = 0(l/n2).
We desire a more general result, however, because one weight function we
treat leads to a kernel that is not continuous a t (0, 0) and (1, 1). We use the
following theorem of Hammerstein [9]: Let k(s, t) be continuous in the unit square
except possibly at the corners of the square; let ak(s, t)/as be continuous in the
interior of both triangles in which the square is divided by the line between (0,O) and
(1, l ) , and let the partial derivative be bounded in the domain e $ s 5 1 - e and
0 5 t _I 1for each a(> 0). Then the series on the right of (4.1) converges uniformly
to k(s, t) in every domain in the interior of the unit square.
Since k(s, t) = d+(S)
z/@) [min (s, t) - st], the condition is that $(t) be
continuous for 0 < t < 1 and
be continuous for a 5 s $ t 5 1.
In this case (4.4) converges in the mean and with probability one for every
t(e S t 5 1 - a), and z(t) is the same process as x(t) in this interval.
If
J 1k(t,
t) dt
<
m,
EzlI/% <
lo
x:./Xj
For
small enough
ASYMPTOTIC
for any 6
tion of
I'
199
k(s, t )
zzlx:./Xjwith characteristic
i s continuous or if k(s, t ) i s continuous except at (0, 0 ) and (1, 1 ) with ak(s, t)/as
continuous for 0 < s, t < 1, s # t, and bounded in a I s S 1 - a, 0 5 t 6 1 for
every a ( > O ) then the characteristic function of W Zi s given by (4.6), where fAj) are
the eigenvalues of k(s, t ) defined by (4.2).
In our case the integral equation is
f (t) =
f(t)+-'(t)
0 when k(0, 0 ) =
0,
(4.15)
If +(t) is continuous ( 0 I t I l ) ,such a solution exists and h(t, A ) is continuous
in t(0 4 t 4 1). Since h(1, A) = 0 for A an eigenvalue of (4.13), the roots of
h(1, X) = 0 are the roots of the Fredholm determinant D(X) associated with
k ( s , t). I t can be shown that
The square root is taken so as to make (4.17) real and positive when the characteristic function is real and positive. The details of this proof are given in [8],
Section 605.
THEOREM
4.2. Let +(t) be continuous for 0 I t I
1. Then the equation (4.14)
has a unique solution h(t, A) for every X > 0 satisfying (4.15). Then the characteristic function of wZi s
200
T. W. ANDERSON A N D D. A . DARLING
n
w
(4.19)
4(t) =
(1 - 2it/Xj)-',
9-1
Even without knowing the eigenvalues, the moments can be calculated in terms
of the iterates of the kernel k(s, t). Putting kl(s, t) = k(s, t) = (min (s, t) 1
st)
d m ) ,kn+l(s, 1) =
expansion
(4.22)
kn(s, t ) =
C X7"fi(~)fj(t)-
Hence,
and, in particular,
r
(4.24)
l1
k(s, S) ds =
6'
u2 = 2
s(l
- s)+(s) as,
(1
- sj2+(s) 0
k2(r, t) ds dt = 4
4x
h(1, A)
201
This expression was given by Smirnov [15] and later by von Mises [20] using
entirely different methods. A formal method for finding the distribution (by
inverting the Fourier transform) was given later by Smirriov [16], but his expression is not amenable to numerical calculation. The following procedure expresses al(z) = Pr [ w 25 z ] in terms of tabulated functions.
I t appears convenient to work with the Laplace transform. We have
for the cdf al(z). We wish to invert this Laplace transform. Now
where
(<')
= (- l)'I'(j
>
plex inversion formula can be used termwise here since the abscissa of convergence
of f(t)/t is Rit) = 0, and the above series coiirrergesabsolutely and uniformly in
the half plane R(t) 2 /? > 0.
Since
we have
where
202
dGr(3,4)ru4
Ae-A2/(8z)
lw
dm
2312d \/;
r(3/4)z3I4
- ( A ~ I ( S Z ) ) cash B
(sinh 8)' dB
zy-o
TABLE 1
a l ( z ) = lim Pr(nw2 5 z ]
n-m
F ( x ) equal weights. '4 statistician may prefer to use this weight function when
he feels that + ( t ) = 1 does not give enough weight to the tails of the distribution.
204
A.
DARLING
In this example
An analysis similar to that used in Example 1 shows that the cdf, a 2 ( z ) ,can be
expressed as
--
~/(8(~2+1))-((4j+1)~i~~u.~)'(Sz)
dw.
P r ( sup
osus1
dii / G,(u)
- u
1 d + j )5
2).
+(IL)
bounded we have
Pr ( K 5 z ) ,
ASYMPTOTIC
205
paths y(u) of the diffusing particle which do not get "absorbed into'' (i.e., intersect) the "barriers" y = &z(#(u))-'.
It is convenient to make a trailsformation due to Doob [6] which renders the
analysis simpler. If we put
it is easy to verify that X(t) is the Wiener-Einstein process; that is, X(t) is
Gaussian, X(0) = 0, E(X(t)) = 0, E(X(t)X(s)) = min (s, t). Then
where
Thus we have the absorption probability problem for the free particle with
barriers x = f[(t) for t 1 0.
The method of solution is to treat the corresponding diffusion problem as a
boundary value problem with the diffusiorl equation
associated with the region t 2 0, / x / $ [(t). In line with the preceding analogy
f(t, x) will be the "density" of paths X(ZL)which for 0 Iu It have not been
"absorbed" and for which X(t) = x; hence
(2)
lim
z-+?t{(t)
f (t, x) = 0,
lim
f(t, x)
1-10
g (7)
but if t(t) does not have a bounded derivative for t 2 0, (5.5) can no longer be
employed to determine b(z). However, if there are a finite number of intervals
in each of which [(t) has a bounded derivative and between which [(t) has a
simple jump discontinuity it is easy to modify the above result; in fact over
some of the intervals ((t) may be infinite. A piecewise determination can be
made and the solution can be continued to beyond the last discontinuity, and
then (5.5) can be used. Suppose the points of discontinuity of [(t) are
0 < tl < tz < . . . < t, and suppose t(t) is, say, left continuous. In the region
(0, tl) we have the solution go(t, x) = po(O, 0; t, x) by the above theorem. Now
if t(t1) < t(tl
0) we define g:(tl , x) by
(t, x)
g:(tl, Y ) ~(t1,
I Y; t, 2) d ~ .
iuI<E(11+0)
In the same way we can define a function g?(tz, x) which will yield a function
207
g2(t, x) for t2 < t < t3 . This process will ultimately yield a unique function
g,(t, 2) for t > t, . Finally
HZ)
lim
r-ta
((1)
-((t)
By choosing enough intervals, an arbitrary weight function can be approximated, in a manner of speaking.
It follo~vsthat the problem will be essentially solved if we can determine the
208
functions pk(tk , y; t, 5) of Theorem 5.2. I n this case the function [(t) becomes,
by (5.21,
and we must find the solut'ion to equation (5.3) which satisfies the conditions
(2) and (3) of Theorem 5.1.
As before we put t k = uk/(l - uk), and it follows by a classical procedure of
superposing an infinite system of sources and sinks along the line t = tk that we
may get the Green's solution. I n fact, let us put a source a t t = t k , 5 = yj , of
strength sj , where
pk(tk g; t, 2) =
j-.-m
$ tk+l and I y
sj
42r(t
,-
I<(z/G)
(1
+ tk)we
t(z-~j)~/(f-fh)
- tk)
Now on putting
lim
gn(t, 2) dx.
t-oo
/rl<-t(l+t)
4Fi
The resulting functioll bl(z) is a multiply infinite sum of integrals of an nvariate Gaussian distribi~tionover an n-dimensional rectangle.
209
O S a < u S b S l ,
#(u) =
otherwise.
Thus the test of the hypothesis is confined to detecting discrepancies over only
a central portion of the interval [0, 11. Using the preceding notation we have
n = 2 and
and hence
Yj
= 2jz(tl
+ 1) + ( - l ) ' x i ,
sj = (- 1)' exp
( -2z2(tl
~ 2 0 22, 2 ; t, 2 ) =
+ 1)jZ- 2 z x ( -
I)'!,
e- i ( z - z 2 ) 2 / ( t - 1 2 )
4 2 ? r ( t - t2) '
a J
z(l+ty)
-z(i+ts) -z(l+tl)
210
means p l , p2 , variances a:
above integral
, a:
and correlation
where
There are tables available in which the function M is tabulated; see also P6lya
[14]. Also, if either a = 0 or b = 1 then p = 0 and the function can be calculated
with the ordinary univariate Gaussian tables. Putting a = 0, b = 1 simultaneously we obtain Kolmogorov's result
which has been tabulated [18]. I n the general case the convergence is very rapid
and good results can be obtained by using a few central terms (in (5.9) the terms
corresponding to f j are clearly equal).
The formula (5.9) is in disagreement with a recent announcement (without
proof) of Maniya [13]. Maniya's note appeared subsequent to a restricted paper
by the authors.
By using the general formula above it is possible to get, for example, a weight
function to test discrepancies over only the tails of the distribution, etc.
Ea-ample 2. We next investigate
0,
otherwise,
211
which is the weight function emsidered before with the w2test. By an earlier
remark we must have a > 0 and b < 1, else absorption is certain and b(z) is
degenerate. The transformation (5.2) yields
we obtain
1 u(t) 1
a
1 - a
S z, - log -2p
where D,(z) is the Weber function [23]. I t seems very difficult to get even any
qualitative information from this formula.,
REFERENCES
[I] R. BELLMAN
AND T. HARRIS,
"Recurreqce times for the Ehrenfest model," Pacific
Jour. Math., Vol. 1 (1951)) pp. 179-193.
C R A M ~"On
R ,
the composition of elementary errors," Skandinavisk Aktuarie[2] HARALD
lidskrift, Vol. 11 (1928), pp. 13-74, 141-180.
[3] G. DOETSCK,
"Les Bquations aux ddrivdes partielles du type parabolique," Enseignement Math., Vol. 35 (1936), pp. 43-87.
212
T. W. ANDERSON A N D D. A. DARLING