Econometric Theory, 7, 1991,519-529. Printed in the United Statesof America.

FROM CHARACTERISTIC FUNCTION

TO DISTRIBUTION FUNCTION:
A SIMPLE FRAMEWORK
FOR THE THEORY
N. G. SHEPHARD
London School of Economics

A unified framework is establishedfor the study of the computation of the distribution function from the characteristic function. A new approach to the
proof of Ourland's and Oil-Pelaez'sunivariate inversion theorem is suggested.
A multivariate inversion theorem is then derived using this technique.

1. INTRODUCTION

It is often easierto manipulatecharacteristicfunctions than distribution functions. If the characteristic function is known then we can compute the distribution function by using an inversion theorem. This paper reviews the
theoretical basis of inverting characteristic functions, presenting the work
within a unified framework based on the well-known results of Fourier
analysis.
Inverting the characteristic function to find the distribution function has
a long history (cf. Lukacs [16, chapter 2]). Levy's [15] result is the most famous of thesetheorems, although in this context its practical use is limited
to somespecialcasesunlessthe random variable of interest is alwaysstrictly
positive (seeHohmann [2] and Knott [14]). Ourland's [10] paper gavea more
useful inversion theorem, but it is the paper of Oil-Pelaez [9] which has provided the basisof most of the distributional work completedin this field (cf.
Davies [4,5] and Imhof [12]). Ourland's and Oil-Pelaez's results are almost
identical. Ourland's is basedon the principal value of a Lebesgueintegral,
while Oil-Pelaez removesthe needfor principal valuesby using a Riemann
integral. The univariate inversion has beenusedextensivelyin econometrics;
a short review is given in Phillips [17].
RecentlyShively [23] has generalizedOil-Pelaez'swork on Riemannintegrals to provide a bivariate inversion theorem, while Shively [24] used this
expressionto tabulate critical values of a statistic proposed by Watson and
Engle [25] for testing the stability of the parametersin a regressionmodel.
Only Ourland [10] has attempted to provide a multivariate inversion theoI thank Dr. M. Knott and D~. R.W. Farebrother

for their comments on earlier versions of this work. The

comments of Professor P.C.B. Phillips and the referees were also of considerable assistance.

@ 1991 Cambridge University Press

0266-4666/91$5.00 + .00

519

520

N.G. SHEPHARD

rem. Our results are slightly different from those obtained by Gurland. In
this paper we develop a framework for the analysis of univariate inversion
theoremswhich offers an easymultivariate generalization.The result, which
is given in Theorem 5, is an expressionthat involves terms that are straightforward to compute.
2. THE UNIVARIATE INVERSION THEOREM

Bohmann [1] studied inversion theorems using the results of Fourier analysis. His work, which relies on the properties of convolutions, is in keeping
with the discussionof characteristicfunctions given by Feller [8, chapterXV].
Although the subject matter of this paper is rather different from that considered by Bohmann, our general approach will be consistent with his.
To establishour notation we introduce somedefinitions. Let F denotethe
distribution function of interest. Supposeits correspondingdensity,f, is integrable in the Lebesguesense(written f E L, seefor example Rubin [19,
chapter 1] for an introduction to Lebesgueintegrals) and that its characteristic function is defined as rp(t) = f~ooeiIXf(x)dx. We supposethat rp is
known and we wish to compute F directly from it. The basic result we will
use to perform this calculation is the Fourier inversion theorem.
THEOREM 1 (Fourier inversion theorem). Supposeg, rpE L, and
<p(t)=

eitXg(x) dx,

(1)

then
e-i/Xtp(t) dt,

(2)

everywhere.

Proof. Cf. HewittandStromberg

[11,p. 409].

The other result which will be central to our developmentconcernsconvolutions.

THEOREM 2. Ifg,h E L, g * h(x) = eoog(x - y)h(y) dy, and g and h
have Fourier transforms q, and 1/;,then the Fourier transform of g*h(x) is
~(t) = q,(t)1/;(t).
Proof. Cf. Hewitt and Stromberg [11], Theorem (21.31) and Theorem
(21.41).

g*h is called the convolution of g and h. Thesetwo results will be nearly

sufficient to enableus to developall the resultswe want in this paper. A simple application of the Fourier inversion theorem givesus the following wellknown result.

A FRAMEWORK FOR INVERSIONS

521

COROLLARY 1. If J,l(JE L, then

f(x)

= -1

oo

e-itxlp(t)dt.

211" -00

Proof. Thisfollowstriviallyfrom Theorem1.

Equally, following, for example, Feller, we can convolute Fwith the uniform distribution on [-h,h] and then useCorollary I to produce Levy's important theorem.

COROLLARY 2 (Levy [15]). If f,'p E L, then

oo sinht
-itx
F(x+h)-F(x-h)
2h
h e 'p( t )d t.

-~

211"

-00

Proof. By Corollary I, as the left-hand side is a density function.

This corollary has been of use to statisticians working in many fields, as

it allows the derivation of an algorithm for the calculation of the distribution function from the characteristic function when the random variable of
interestis strictly positive (cf. Bohmann [2, p. 238] and, independently,Knott
[14, p. 431]). However, as Oil-Pelaez [9] noted, when it is not positive this
expressioncannot be used for this purpose. As a result many writers have
abandonedthe idea of using convolutions. Oil-Pelaez employedthe notion
of a Riemannintegral (seethe proof in Kendall, Stuart and Ord [13, pp. 120121]), while Ourland used a similar idea, but his proof involved the manipulation of principal values of Lebesgueintegrals. Theorem 3 shows that
under weak regularity (which will be relaxed to some extent in Theorem 4)
this diversion was unnecessaryand so allows us to get away from a type of
derivation which". . . detracts from the logical structure of the theory"
(Feller [8, p. 511]). The theorem follows by using the Fourier inversion theorem on the convolution of f with a sign function defined on the range
[-h,h]. As h -+ 00 the Riemann-Lebesguetheorem gives the result.
THEOREM 3. Iff, 'p E L, then under the assumption of the existenceof
a meanfor the random variable of interest
F(x)

= ! - ~ ra>~
2

where Ll17(t)
t

211"

Jo

= 71(t) +

If'(t).e-;tX

dt,

(3)

It

71(-t).

Proof.Givenin theappendix.

The derivative of (3) gives the result in Corollary I immediately.

The requirement that 'p E L can be removed by using an additional convolution to improve the behaviorof the tails of the integrand. In the next theorem we do this by employingFejer's kernel (thereare, of course,many other
kernels we could have used, e.g., Abel's or Gauss's).

522

N.G. SHEPHARD

THEOREM 4. Iff E L, then under the assumption of the existenceof a

mean
F(x)

= -1 - - 1

Urn
211"n--+oo

[ 1 - -t

.:l

'P(t)riIX

lit.

(4)

it

Proof. Givenin theappendix.

The existenceof the meanis assumedto ensurethat 4(1()(t) exp( -itx)/it)

is uniformly bounded. Its existenceis a sufficient, although not a necessary
condition for this result to hold. It can be removed altogether by using the
principal value of the integral-see Gurland [10].
3. THE MULTIVARIATE INVERSION THEOREM
Supposewe now becomeinterestedin the multivariate generalizationsof the
above theorems. We supposef E LP and that
Ip(t)=

L~ ... L~ eit'Xf(x) dx,

(5)

where x = (XI,. . . ,xp)' and t = (II,. . . ,tp)' wherep is some positive integer. It is well known that ~heFourier inversion theoremand convolution theory go through to the multivariate case,so allowing us trivial proofs of the
following well-known corollaries.

f f

COROLLARY 3 (cf. Cramer [3J, eqn. 10.6.3). If/'

f(x) = ~
oo . . .
oo
1
(21r)

-00

I{)

e-it'XIp(t) dt.

E LP, then
(6)

-00

Proof. The proof follows using a generalization of the proof of Corol-

lary1.

Equality (6) is used extensivelyin the literature-see, for example, Phillips

[18] in his work on the sampling behavior of matrix quotients.
COROLLARY 4 (cf. Cramer [3], eqn. 10.6.2). If f,1p E LP, and we define an interval R by the inequalities Xj - h < Xj < Xj + h for j = 1,.. . ,Po

then

p
=- 1
Pr(R)
(2h)

(27r)P

f . . .f
oo

-00

oo

II
p

-00 j=l

sin htj
htj

e-it'xtp(t) dt.

(7)

Proof. The proof follows using a generalizati.00 of the proof of Corollary 2.

.
Little has been written about the theory of computing the distribution
function by inverting multivariate characteristicfunctions. Ourland [10] extended his univariate procedureto p dimensionsusing the notion of a prin-

A FRAMEWORK FOR INVERSIONS

523

cipal value of a Lebesgueintegral. His result is slightly different from the

result we presenthere. Further, his proof is much more complicatedthan the
one given here. RecentlyShively [23,24] has extendedOil-Pelaez's [9] result
to two variables in order to calculate some critical values of a test statistic
which arisesin econometrics.He usedRiemannintegralsin his derivation and
so his proof doesnot conform with modern work on integrals, as well as being rather obscure.
As we have seenin Section2 there is no needto usethesetechniques.The
advantageof using a convolution approach is that the multivariate generalizations now follow using standard results.

THEOREM5. If/' I() E LP, then under the assumptionof the existenceof
a mean, the following equality holds.
L
L
[ tp(t)e-ix't
-(-2)P
...
,:l,:l...,:l
] dt=ut(x),
(8)
""

(21r)P

""

t1 t2

tp it,it2...itp

where
ut(x) =2PF(x"...,xp)

- 2P-' [F(X2,X3,. . . ,xp) +... + F(x". . . ,Xp-2,Xp-,)]

+ 2P-2[F(X3,X4,'",xp) +... + F(x"... ,Xp-3,Xp-2)]

+...+(-l)p.
Proof. Givenin theappendix.

THEOREM 6. Under the existenceof a meanthefol/owing equality holds.

= 2ip-lJ1

.. .~ [ Ip(t)e-ix't
tp it 1it2 . . . it
II
p

.6

...J1 1m Ip(t)e-ix't
1\12' . . I

12

=2iPJ1...J1Re
12

'

ifpisodd

Ip

Ip(t )rix't
1\/2".1

Ip

] '

if p is even. (9)

Proof. Trivial, but extremelytediousmanipulationsof trigonometric func-

tions.

The assumptionof the integrability of I(Jcan be removedby following the

approach given in the proof of Theorem 4. This yields a straightforward, although rather ugly, proof of Theorem 7.
THEOREM 7. If f E LP, then under the assumption of the existenceof
a mean, the following equality holds.
(-2)P
-hm

(211")P n-+oo

i ... i II
n

n p

j=l

to ]

1-.1.n ~~
t,t2 ...~tp

[ Ip(t)e-x't
itlit2,..itp

] dt=ut(x)

(10)

524

N.G. SHEPHARD

where

ut(x)

= 2PF(xl>'"

,xp)

- 2P-l [F(XZ,X3' . . . ,Xp) + . . . + F(Xl, . . . ,Xp-Z,Xp-l)]

+ 2P-Z[F(X3,X4,' . . ,Xp) +. .. + F(XIo' .. ,Xp-3,Xp-Z)]

+ .. . + (-I)P.
4. EXAMPLES
Theorem 3 and the following examplesof Theorem 6 may be usedto successively generatethe distribution functions of F(xd, F(X),X2)' F(XIoX2,X3),
and so forth.
p

=2

1 1
""

~ 22
(21/")

""

A A [ ",,(t)e-ix"
..
II 12

1 1
""

=~ -23
(21/")

= 4F(XIoX2) ""

~-23

(21/")

""

ARe [ ",,(t)e-ix"
12

""

0
""

dt)dt2

t)t2

AAA. [ ",,(t)e-ix"
..
11 12 13

= ~ 24 1 1 1
(21/")

2 [F(xd + F(X2)] + 1.

L 1 1

p=3

""

dt) dt2

It)lt2

""

""

(11)

dt)dt2dt3

It)lt21tp

A A 1m [ ",,(t)e-ix"

12 13

= 8F(XIoX2,X3)
- 4[F(XIoX2)
+ 2[F(xd
p=4

+ F(X2) +

1 1 1 1
""

~ 24
(21/")

""

""

""

""

(21/")

""

+ F(XIoX3)
F(X3)] - 1.

AAAA

""

""

+ F(X2,X3)]
(12)

[ ",,(t)e-ix"
. . .. ] dt)dt2dt3dt4

11 12 13 14

= ~ 2S 1 1 1 L

dt)dt2dt3

t)t2tp

It) lt21t3 lt4

AAARe
12 13 14

[ ",,(t)e-iX"

dt)dt2dt3dt4

t)t2t3t4

= 16F(XIoX2,X3,X4)

- 8[F(x),X2,X3)

+F(X),X2,X4)

+F(X),X3,X4)

+F(X2,X3,X4)]

+ 4[F(XIoX2) + F(XIoX3) + F(XIoX4) + F(X2,X3)

+ F(X2,X4) + F(X3,X4)]

2[F(xd

+ F(X2) + F(X3) + F(X4)] + 1.

(13)

Numerical integration rules for the evaluationof theseintegralsare derived

in Shephard [20] using Riemann sumswith the step sizesautomatically cho-

A FRAMEWORK FOR INVERSIONS

525

sen.As suchthey provide multivariate generalizationsof Davies' [5,6] result.

The bivariate result is usedby Shephard [22] to tabulate Farebrother's [6,7]
test for linear restrictions in a heteroskedasticregression model and by
Shephard [21] to tabulate the distribution function of the maximum likelihood estimator of a noninvertible moving averageprocess.
5. CONCLUSION
This article provides a unified framework for the study of inverting characteristic functions in order to compute the distribution function. The theory
is basedon the familiar ideasof convolutions, Lebesgueintegrals, and Fourier inversions.
The multivariate inversion formula can be applied to a number of important problems in econometrics. The use of Riemann sums to evaluate
the integrals allows a straightforward and comparatively cheap method of
numerically implementing the theory.
REFERENCES
I. Bohmann, H. Approximate

Fourier analysis of distribution functions. Arkiv fur Matematik

4 (1961): 99-1157.
2. Bohmann, H. A method to calculate the distribution function when the characteristic function is known. Nordisk Tidskr. Informationsbehandling
(BIT) 10 (1970): 237-242.
3. Cramer, H. Mathematical methods of statistics. Princeton: Princeton University Press, 1946.
4. Davies, R.B. Numerical inversion of a characteristic function. Biometrika 60 (1973): 415-417.
5. Davies, R.B. AS 155: The distribution of a linear combination of x2 random variables. Applied Statistics 29 (1980): 323-333.
6. Farebrother, R.W. Testing linear restrictions with unequal variances, a problem. Econometric
Theory 4 (1988): 349.
7. Farebrother, R. W. Testing linear restrictions with unequal variances, a solution. Econometric
Theory 5 (1989): 324-326.
8. Feller, W. Introduction to Probability Theory and Its Applications, Volume Two. 2nd ed.
New York: Wiley, 1971.
9. Gil-Pelaez, J. Note on the inversion theorem. Biometrika 37 (1951): 481-482.
10. Gurland, J. Inversion formulae for the distribution of ratios. Annals of Mathematical Statistics 19 (1948): 228-237.
11. Hewitt, E. & K.R. Stromberg. Real and Abstract Analysis. New York: Springer-Verlag, 1965.
12. Imhof, J.P. Computing the distribution of quadratic forms in normal variables. Biometrika
48 (1961): 419-426.

13. Kendall, M.G., A. Stuart & J.K. Ord. Advanced Theory of Statistics, Vol. 1. 5th ed.
London: Griffin, 1987.
14. Knott, M. The distribution of the Cramer-von Mises sta~isticfor small samplesizes.Journal of the Royal Statistical Society, SeriesB 36 (1974): 430-438.
15. Levy, P. Calcul desprobabilites. Paris: Gauthier-Villars, 1925.
16. Lukacs, E. Characteristic Functions. London: Griffin, 1970.
17. Phillips, P.C.B. Exact small sample theory in the simultaneous equations model. In
llandbookof Econometric5,
VolumeI, Z. Griliche5& M.D. Intriligator (ed5.),Arn5tenlam;
North-Holland Publishing Company, 1983.

526

N.G. SHEPHARD

18. Phillips, P.C.B. The distribution of matrix quotients. Journal of Multivariate Analysis 16
(1985): 157-161.
19. Rubin, W. Real and Complex Analysis. New York: McGraw-Hili,
1970.
20. Shephard, N .G. Numerical integration rules for multivariate inversions. Journal of Statistical
Computation and Simulation 39 (1991): 37-46.
21. Shephard, N.G. Evaluating the distribution function of the maximum likelihood estimator of a first order moving average process and a local level model. Working paper, London
School of Economics, 1990.
22. Shephard, N.G. Tabulation of Farebrother's test for linear restrictions in linear regression
models under heteroscedasticity. Working paper, London School of Economics, 1990.
23. Shively, T.S. Numerical inversion of a bivariate characteristic function. Working paper,
University of Texas at Austin, 1988.
24. Shively, T.S. An analysis of tests for regression coefficient stability. Journal of Econometrics
39 (1988): 367-386.
25. Watson, M.W. & R.F. Engle. Testing for regression coefficient stability with a stationary
AR(1) alternative. Review of Economics and Statistics 67 (1985): 341-346.

APPENDIX
Proof of Theorem 3. The function g(y) = sign(y), y E [-h,h], hasthe transform

'Ph(t)=

i:

eity sign(y)

dy = 2(cos ht it

which is bounded for all

1)
'

(A.I)

t.

The convolution, written Uh(X), of g(y) with the continuous density f(x) is
2F(x) - F(x + h) - F(x - h) which, although bounded, is not integrable as h -+ 00.
The convolution has the transform 2\O(t)(cos ht - l)/it, which is integrable as
\0 E L. Hence, for fixed h we can use the inversion theorem to give the equality
00
(cos - 1) 'P(t)e-ixt dt,

211"

~t

-00

=- 2

It

""

(cosht-l)A

211" 0

[ fII(t)e-iXf
.
] dt=Uh(X).
t

It

(A.2)

As h increases,cos ht moves increasing rapidly and so the part of the integrand

which includes cos ht makesa diminishing contribution to the integral. A sufficient
condition to uniformly bound ~p(t)exp( -ixt)/it)
is that <p(t) is differentiable
t
near zero, which will be true if the mean exists. The consequenceof this bounding

is that the integralof this functionwill exist.As a result,in the limit ash -+ 00 the
left-hand side of the (A2) can be reducedto, using the Riemann-Lebesguetheorem
(cf. Feller [8, p. 513])

-2
d
[
27rJo t
(""

"'U).e-ilX

It

dt

= 2F(x)

- 1.

A FRAMEWORK FOR INVERSIONS

527

Proof of Theorem 4. Recall the proof of Theorem 3. To improve the behavior of

the tails of Uh(X) we convolute it with Fejer's kernel
[ Sin(nXI2) ] 2
(nx/2)
.

= -.!...n
211"

kn(x)

(A.3)

It is importantto note that

L:kn(x) dx = 1.

(i)

(A.4)

(ii) The transform of Fejer's kernel is

'Pn(t) = [1
~

1:1]1(1/1 < n),

(A.S)

where I( .) is an indicator function.

For fixed n the convolution of Uh(x) with kn (x), and its transform, are integrable, as Ip(t) is bounded. Hence we can usethe inversion theorem to give the equality

f [
""

-2
271"

-""

= -2

Itl

l--/(Itl<n)(cosht-l)

""

[1-- Itl ]/(t < n)(cosht

271" 0

",(t)e-iX'
It

-1).1

dt

",(t)e-iX'
.
It

dt

= Uh* kn(x).

(A.6)

When h -+ 00 the left-hand side of (A.6) can be manipulated using the RiemannLebesguetheorem (cf. Feller [8, p. 513]) becausefor fixed n
[1

I~I ]/((

< n)~[ ~(t~;-ixt]

(A.7)

is integrable, as A (",,(t)exp( -ixt)/it) is bounded due to the assumption of the exI

istenceof a mean for the densityf(x) -see the proof of Theorem 3. Hence the lefthand side of (A.6) reducesto

--2

CO

1 - - ] I(t < n)~ [ ",(t)e-ixt

.

211" 0

dl.

(A.8)

II

Now think of the right-hand side of (A6) as h -+

00,

Rememberit is

L: [2F(x - y) - F(x - y - h) - F(x - y + h)]kn(y) dy.

(A.9)

As 2F(x - y) - F(x - Y - h) - F(x - y + h) is uniformly bounded and kn(y) is

integrable (see(A.4 we can employ the Lebesgue-dominatedconvergencetheorem
(seeHewitt and Stromberg [II, Section 12.34]) to imply
lim Uh * kn(x)
h-oo

"" ut * kn(x),

(A.tO)

528

N.G.SHEPHARD

= 2F(x) -

where ut(x)

=-2 1

""

1--

211" 0

1. Thus,

[(1<"'.

>

(t)e-ixt

dt=ut*kn(x).

It

(A.ll)

We now becomeinterestedin the behavior of this equality as n -+ 00. Proposition

1 dealswith the right-hand side. This is a well-known theory (cf. Hewitt and Stromberg [11, Section 21.42]) although we are using slightly different assumptionsfrom
those usually employed and so we include its proof in this appendix. Hence

i [ ][

n
t
--2 Hrn
1 - -.i
2... n-oo 0
n I

]dt = 2F(x)- 1.

<p(t)e-ixl

It

Proof of Theorem5. Define

g(ylo.

. . ,yp) = IT sign (Yj) ,

j=l

where

Yj E [-h,h],

(A.12)

so it has the transform

'Ph(t)

= fr. 2(costjh
j=l

Uh(X)=

itj

j h ... j

hj(X)-y),...,.Xp-Yp)g(Y)dY,

-h

- 1)

(A.t4)

-h

L
a..az

Vthal.haz
ap

,hap'

(A.tS)

where the summation is taken over all the values (OJ = ::t 1), with
Vth..hz

hp=

..bp

(-l)bt+"'+bpF(x)-h)b),...,xp-hpbp).

(A.t6)

where the summation is taken over the binary numbers (hj = 0,1) andF denotesa
generic distribution function. By the (multivariate) Fourier inversion theorem,

1 ..,1
co

2P

-co

(2"Y

2P

(2r)P

= -

.p(t)e-tx" fI

co

-co

eo

...

[cos I~~ -

J=1
eo

IIJ
1p(t)e-Ix'I

.<:1.<:1..,.<:1.. .
'112

Ip

Itllt2"'ltp

] dl,

]IT (cost}h-l)dt=Uh(X),
P

}=1

(A.17)
Allowing h -+

the left-hand side can be simplified by exploiting the Riemann-

00,

Lebesgue theorem. The result is that we have

-(-2)P
(2?r)P

""

." L

""

.1.1 "'.1

0 '1'2

.p(t)e-ix't

Ip it1itz...itp

] dt = ut(x).

. (A.II)

A FRAMEWORK FOR INVERSIONS

PROPOSITION

529

1.

Hrn ut * kn(x) = ut(x).

n-oo

Proof.

ut * kn(x) -

ut(x)

i:

(ut(x-

t) - ut(xkn(t)dt

= l'"' (3(t)kn(t) dt,

(A.19)

where (3(t) = ut(x+ t) + ut(x- t) - 2ut(x). As ut(x) = 2F(x) - 1, (3(t) cannot

be integrable for its tails do not go to zero as t -+ 00, even though it is uniformly
1

bounded. Choose EI >

0, thenusingthe continuityof the distributionfunctionwe

Ii"" kn(t)(3(t)dtl

If

canchooseE > oso small that I(3(t)1< EI vt E [O,E).Sincekn(t):S 2hrnt2, we have

~.
:S

:S E(

kn(t)(3(t)dtl + 11"" kn(t)(3(t)dtl

+ s~PI(3(t)111""kn(t)dtl

:S EI +

:S EI +

-2

supl(3(t)1

1 ""

~n

-2

supl(3(t)I.

E~n

"2 dt
t

(A.20)

so for n > 2 supll3(t)I/EEI 11"

we have the required result by choosing EI sufficiently

small.

.
Ii