Journal of Nuclear Science and Technology

ISSN: 0022-3131 (Print) 1881-1248 (Online) Journal homepage: https://www.tandfonline.com/loi/tnst20

Neutron Spectra Unfolding with Maximum Entropy

and Maximum Likelihood

Shikoh ITOH & Toshiharu TSUNODA

To cite this article: Shikoh ITOH & Toshiharu TSUNODA (1989) Neutron Spectra Unfolding with
Maximum Entropy and Maximum Likelihood, Journal of Nuclear Science and Technology, 26:9,
833-843, DOI: 10.1080/18811248.1989.9734394

To link to this article: https://doi.org/10.1080/18811248.1989.9734394

Published online: 15 Mar 2012.

Submit your article to this journal

Article views: 299

View related articles

Citing articles: 4 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tnst20
journal of NucLEAR SciENCE and TECHNOLOGY, 26(9], pp. 833-843 (September 1989). 833

Neutron Spectra Unfolding with Maximum Entropy

and Maximum Likelihood
Shikoh ITOH and Toshiharu TSUNODAt
Department of Nuclear Engineering,
Faculty of Engineering, Nagoya University*

Received March 29, 1989

A new unfolding theory has been established on the basis of the maximum entropy principle
and the maximum likelihood method. This theory correctly embodies the Poisson statistics of
neutron detection, and always brings a positive solution over the whole energy range. More-
over, the theory unifies both problems of overdetermined and of underdetermined. For the
latter, the ambiguity in assigning a prior probability, i.e. the initial guess in the Bayesian sense,
has become extinct by virtue of the principle. An approximate expression of the covariance
matrix for the resultant spectra is also presented. An efficient algorithm to solve the nonlinear
system, which appears in the present study, has been established. Results of computer simula-
tion showed the effectiveness of the present theory.

KEYWORDS: neutron spectra unfolding, maximum entropy principle, maximum

likelihood method, overdetermined problem, underdetermined problem, Poisson sta-
tistics, positive solution, covariance matrix, algorithms, mathematical logics, com-
puter simulations

high accuracy in the unfolded results<IJ.

I. INTRODUCTION (2) The results obtained with the different
For neutron spectrometers observing recoil- codes in general exhibit significantly dif-
proton pulse-height distribution, it is indispen- ferent results. In examining the unfolded
sable to use unfolding procedure in restoring results obtained with the different codes,
an energy spectrum of neutrons incident on however, a judgement cannot be made that
the detector from a pulse-height distribution. one results or the other is the acceptable
As is well known, spurious oscillations and results< 2 J.
negative values have often appeared in the (3) This sort of data reduction process falls
unfolded results especially when the spectrum to a large extent within the category of
has a sharp dominant peak. To reduce or "art-forms." <SJ
overcome this difficulty, some authors add an As a first step to unravel such a difficulty,
artificial oscillation dumping term to the ob- an unfolding method had been proposed to
jective function to be minimized, and/or even correctly embody the Poisson statistics of neu-
presume a fictitious objective function. These tron detection based on the maximum likeli-
remedies, however, are waiting for clear logi- hood method< 4J. By using this new method,
cal supports if any. it was concluded, as a matter of course, that
In the following three passages which are the appearance of negative values in the un-
cited from literature, the present status of folded results essentially came from the statis-
unfolding techniques may be summarized ; tical fluctuations in pulse-height data, and that
(1) An abundant literature has been devel- the Gaussian approximation, i. e. the usage of
oped which describes a variety of unfolding * Furo-cho, Chikusa-ku, Nagoya 464.
methods for recoil proton data. It remains t Present address: Chubu Electric Power Co., Inc.,
unclear which of these methods achieves Toshin-cho, Higashi-ku, Nagoya 461-01.

-15-
834 ]. Nucl. Sci. Techno[.,

the least-squares method, has little influence A diagonal matrix A formed by the ele-
on the results. .. ments of: a vector x. is expressed by
Turning our eyes to other fields of science
A=diag(x),
and engineering, the MEP (maximum entropy
principle) due to jaynes< 6 ><G> is being increas- A norm of vector x is denoted by
ingly important as a foundation to attack a
variety of "inverse problems," and bringing
II xll =(xt x) 112 •
many spectacular results in these fields, e. g. A vector whose elements are all unity is
seismology< 7>, optics<s>, astronomy< 9>, tomogra- expressed by
phy<'O><''', crystallography< 12 >, photography<' 3 '
u=(1, 1, ... , 1)t,
and so on. The prime superiority of the MEM
(maximum entropy method) is in its intrinsic where the length of u will be specified at used
nature of warranting the positiveness of solu- positions.
tion by nothing but objective facts. The expectation of a random variable x is
The purpose of the present paper is to denoted by putting it in brackets: <x>. The
introduce the MEM into the neutron spectra covariance matrix of x is written as
unfolding problem, making the procedure free
I,=((x-<x>)(x-<x>)t>=<ox·oxt>.
from the "art." In Chap. II, the unfolding
problem is formulated on the basis of the 2. Definitions
MEP. An approximate expression of the co- Let rp 1 the number of neutrons incident on
variance matrix for the unfolded result is also a detector and belonging j-th energy group,
presented. Chapter ill is devoted to the algo- and let c; the counts in i-th channel of a
rithm requisite to obtain the solutions of the pulse-height histogram of recoil-protons.
nonlinear equation system which appears in Then, the two quantities are related as
the present formulation of the problem. In
Chap. IV, results of computer simulations test-
ing the method's performance are shown.
Chapter V summarizes the present study. In or in matrix notation
APPENDIX 1, a concise summary of the MEM
C';::!Rrp, ( 1')
is given. In APPENDIX 2 are presented some
of formulae which might be helpful to handle where R; 1 is the probability that one neutron
vector matrix equations in this study. of j-th group brings a count in i-th channel.
Taking the expectation of both sides of Eq.
ll. FORMULATION OF PROBLEM ( 1') and denoting the total number of neu-
1. Notations trons incident on the detector by a, we have
In this paper, lightface letters indicate
a=Rp, (2)
scalar quantities, bold small letters vectors,
and bold capitals matrices. Where necessary, where a=<c>!a, ( 3)
vector matrix elements are indicated by sub-
p=(rp)/a, (4)
scripted corresponding letters.
Matrices At, A-' and A- denote the trans- a=ut<rp>, (5)
pose, inverse and generalized inverse (g-inverse)
utp=l. (6)
of A, respectively.
For notational convenience, the expressions 3. Introduction of MEM
Based on the MEP, we should choose the
energy distribution p which maximizes
H=-p 1 lnp, ( 7)

are employed. subject to the constraints of Eqs. ( 2) and ( 6)

-16-
 Vol. 26, No. 9 (Sep. 1989) 835

The results is violate the MEP, since the exact number

of "'/. 2 is never known for each experiment.
p= Z~x) exp(Rtx), ( 8) A reasonable approach to the present prob-
lem may be the one in which Eq. ( 8 ) is con-
with the partition function sidered to be the prior probability governing
the appearance of c. For more concretely,
Z(x)=ut exp(Rtx). ( 9)
maximize the likelihood relevant to the Poisson
The real vector x is the Lagrange multiplier statistics of neutron detection<'l:
to be determined from the constraints ( 2)
Lv=ct ln(c)-ut(c) (11)
which is equivalent to
subject to the constraints of Eqs. ( 3 ) and ( 8 ),
a
a=- -In Z(x) (10) or equivalently, find out the unconstrained
ax maximum of
(see APPENDIX 1.) L=ct ln(c)-ut(c)
As is obvious from Eqs. ( 8 ) and ( 9 ), thus
+((c)-aRp)t).+(f(x)-p)tp (12)
obtained solution p has non-negative compo-
nents. We must however aware of the fol- with respect to (c), p, a and x, where
lowing facts ;
(1) What is obtained by an experiment is c f(x)= Z~x) exp(Rtx), (13)
but not a which is never attained within
a reasonable nor in a finite time. In addi- ). and p are Lagrange multipliers. Different-
tion, while the value of rp had already iating L with respect to these variables, and
fixed after the experiment once performed, equating the results to zero respectively, we
we could expect many different c's for have
just the same rp. To use Rrp=c for a
(D)- 1c-u+l=O, (14)
substitute of Eq. ( 2) means to let us ex-
press the statistical fluctuations in c with aRtl+p=O, (15)
a vector rp having non-negative compo-
ptRtl=O, (16)
nents. Thus obtained solution still has the
same functional form as Eq. ( 8 ), but some R(Q- ppt)).=O, (17)
components in x, and then in p, will be-
where (D)=diag((c)), (18)
come complex numbers due to possible
contradictions in Rrp=c for the cases where Q=diag(p). (19)
m~n. Such cases where m<n are even
Eliminating). and p from Eqs. (141 to (17), we
more dangerous in practice because the
have
contradictions may be no longer obvious
mathematically. ilL
g=-
(2) It involves the same difficulty as the - ox'
above to employ a least-squares or maxi-
=RQRt A -'(c-aa)=O, (20)
mum likelihood solution of Eq. ( 1 ) for a
substitute of Eq. ( 2 ). Although these solu- (21)
tions are self-consistent, they admit nega-
where A=diag(a), (22)
tive components for some cases, apparently
contradicting with PJ~O Ci=1, ···, n) and (23)
bringing complex solution.
(24)
(3) On the other hand, some authors have
proposed the constraint which equates the Incidentally, with p=O, i.e. not introducing
"'/. statistic to its expectation<'Hl2)(I.;J. With
2
the MEP, we have a likelihood equation for
this type of constraints, we are forced t the present problem<'l:
-17-
836 ]. Nucl. Sci. Techno!.,

as a variable and maximize the likelihood L(y)

p=R 1A- 1(c-aa)=O. (25)
with respect to y, to obtain
4. Solution to Problem
g(y*)+H(y*)·oy=O, (33)
Our problem is now to solve the nonlinear
system (20) for the vector p which is included as a substitute of Eq. (26), where
in Q, A, a and a.
The equation system (20) seems to be what
a
g(y)== ay L(y)=Qp, (34)
is called of underdetermined, when the number
n of elements in p is larger than m of those a
H(y)== ay g(y)
in c. The latter number m is also that of
mutually independent equations. Equation (20)
=QM+QRt(___!!_uu 1 -A-'DA-')RQ.
can, however, be solved for x of length m, a tot
and then p can be uniquely calculated with (35)
Eq. ( 8 ). 5. Error Estimation
Applying Newton's method to Eq. (20), we In this section, we will investigate the
have an iterative scheme of change orp in the resultant spectrum rp due to
the statistical fluctuations in data c, namely
g(x)~g(x*)+H(x*)·ox=O, (26)
we are going to evaluate the covariance ma-
where x* is the Taylor's series expansion point trix I 'P of rp.
and Taking a total derivative of rp=ap, then
x=x*+ox. (27) ap (aa)t
orp=a ax ox+p ax ox+p
(aa)t
ac oc
Then,
1 1
ox=- n-'(x*). g(x*)' (28) =a(In- - -ps 1)QR 1ox+ - -pu 1oc,
atot atot

where
a
H(x)=. ax g(x)
(36)
where In is the identity matrix of order n,
=RQMR 1 +RQR1 and

·(___!!_uu 1 -A-'DA-')RQR 1 ,
(37)
a to,
is the efficiency vector of the detector. The
(29)
relation (36) shows that ocpi=O, thereby (I rp)ik
and where M=.diag(p), (30) =<ocprocpk>=O, for Pi==Qii=O. Based on the
above observations, let p and Q be of reduced
D=.diag(c). (31)
size in which zero elements are omitted, and
In getting the right-hand side of Eq. (29), the let R is also of reduced one in which columns
relations given in APPENDIX 2 might be help- corresponding to zero elements in p are
ful. omitted.
For the cases where m>n, i.e. overdeter- For the cases when m<n the matrix RQR 1
mined problem, the number m of elements in can be considered regular, consequently Eq.
x is equal to that of equations. Nevertheless, (20) reduces to
the value of x which satisfies Eq. (20) cannot
c-aRp=O, (38)
be uniquely determined, since the rank of
mXm matrix RQR 1 is n at most. It is suf- obtaining a relation of
ficient, however, to know the value of R 1x,
1 1
but not of x itself, for determining p. Hence oc=aR(In- - -ps 1)QR 1ox+ - -autoc
atot atot
we may choose
(39)
(32)
=Rorp. (40)

-18-
 Vol. 26, No. 9 (Sep. 1989) 837

Rearranging Eq. (39), we have RtA-'R/a from the fact that

1 2"'(R 1A-'R)2"'=a2"', (51)
( A-'- - -uut)oc
Utot
and Eq. (48)<' 6 J.
(41) For the cases where m> n, we start with
g(y)=O in which R and Q are also interpreted
Here, note the fact that as of reduced size ;
(52)
getting
(42) R 1 A-'oc=(RtA-'DA-'R-~ee 1)op
Utot

i.e. A is a g-inverse of the matrix in the 1

+-eu oc 1
(53)
right-hand side of Eq. (42): Utot

1 instead of Eq. (39). To take a step forward,

( A-'- -Utot uut)- =A. (43) we may approximate D in Eq. (53) by D~
<D>=aA to have
From Eqs. (41) and (42),
1
1 1
R A-'oc=aR (A-'- - -uu )Rop
1
1 Utot
a(A-'---uut)RQRt2xRQRt
Utot 1
+-eu 1oc. (54)
1
1 1 Utot
·(A- - --uut)=(A-'---uu 1) (44)
Utot Utot , Next, premultiplying Eq. (36) by R 1A_, R, we
where <oc·oc )=aA is used. Equation (44)
1 have
together with Eq. (43) tells us that
aRQR 12xRQR 1 =A, (45)
and then
By substituting Eq. (54) into Eq. (55), we have
(46)
R 1A-'Rof!J=R 1A-'oc, (56)
On the other hand, from Eq. (40) we have
consequently
(47)
(57)
Put Eq. (47) between R A-' and A-'R, then
1
since R 1A-'R is regular for the present case.
(R 1 A-'R)2"'(R 1A-'R)=aR 1A-'R. (48)
ill. ALGORITHM
Hereupon 2"' is a g-inverse of R 1A-'R/a. By
In proceeding the iterative scheme given
the same operation on Eq. (45) as the above,
in Sec. ll-4, there are several cruxes com-
we have
monly to either case where m<n, or m>n.
a(R 1A_, R)(QR 12 xRQ)(R 1A_, R) So, we will discuss only about the former
=R 1 A-'R. (49) case in the following ;

Compared Eq. (48) with Eq. (49), it may be (58)

seen that one choice of (59)
2'"=aQR (RQR
T
1 1
)-
1
A(RQR 1
)-' RQ. (50) where g<kJ=g(x<kl) and H<kJ=H(x<kl).
Note that 2 9 is a reflexive g-inverse of The first difficulty is that the Hessian rna-
-19-
838 ]. Nucl. Sci. Techno!_,

trix H is almost always of ill-conditioning, as 0 MeV and continued through 17.0 MeV.
recognized through the form of Eq. (29). To 2. Response Matrix
overcome this difficulty, we employed the in- The response functions used, i.e. column
complete (or modified) Cholesky method<' 7 > in vectors in R, are of idealized rectangular
obtaining the correction ox< k>. shape which are the same as used in our
Next, we need to adjust Rtx(=y), to avoid previous study<•> ;
exponential overflows in calculating p by Eq. 1/w, l~i~[w],
( 8 ), as

Even with the value of p thus obtained,

(60)

where
Rii=

m
!1-[w]/w,
0,
.
w=nXJ,
i=[w]+l,
[w]+2~i~m,

j=l, ··· ,n,

(65)

(66)
an absurd value of £<k>, and also of JJg<k>J[,
will usually be encountered. Even if not so, and the brackets denote the Gauss notation.
the new vector x<k+l> does not satisfy neither With reference to commonly used matrix
sizes<' 9 >< 20 l, the number m of pulse-height bins
(61) was chosen as 104 for the problems of over-
(62) determined, and as 52 for those of under-
determined, respectively, while the number n
in almost all iterative steps. Then the intro- of energy groups was fixed to 77.
duction of a step size s<k>, which adjusts the 3. Recoil-proton Pulse-height Data
magnitude of ox<k>, become essential to relay
To simulate experimental data, recoil-proton
on Newton's method; spectra c of mean Rv; were computed using
ox<k>~s<k>ox<k)' (63) the computer subroutine "RANP2" which gen-
erates Poisson random numbers for a given
where 0<s<k>=s,f3 1, (64)
mean value< 2 '>.
and where s 0 and {3<1 are scalar constants 4. Initial Guess Spectra
prescribed. With Armijo's method the step As is obvious from the results in Chap. II,
size was determined in every k<'s>. unfolded spectra by the present method cannot
In earlier iterations, we must use Eq. (61) be affected by an arbitrary choice of initial
as a criterion to determine the order l of {3, guess, but it may be better for faster con-
to reach the global maximum of L, while vergence of the iterative scheme to employ
avoiding trapped by one of local extrema. an appropriate initial guess spectrum of v;.
Following this scheme, however, the sequence At first, calculate
{x<k>} converges to a point, while g<k> is far
(67)
away from 0. By changing the criterion to
Eq. (62) at this point, {g<k>} converges to 0 where R+ is the Moore-Penrose g-inverse of
and satisfactory solution can be obtained. R<''>. Equation (67) gives the least-squares
solution without weight for the cases when
W. NUMERICAL TEST m>n, or the minimum norm one for m<n.
1. Test Spectra "Next, smoothen f/Jinit, fix up negative por-
In order to test the present method, we tions in it to some small constants, e. g.
applied it to four types of assumed spectra: machine epsilon, and normalize the result so
(1) two kinds of Gaussian shape spectra with that the sum of its elements is unity to make
average energies of 14.8 and 8.5 MeV, respec- an initial guess p< 0 > with which the iterative
tively, whose standard deviations were as- scheme starts.
sumed to 1. 0 MeV ; (2) a monotonically de- 5. Error Magnification Rate
creasing (ramp-down) ; (3) a monotonically in- As a measure for overall relative error in
creasing (ramp-up); and (4) a fiat spectrum. an unfolded result, we may choose a ratio of
Excepting case (1) these spectra started at norms:
-20-
Vol. 26, No. 9 (Sep. 1989) 839

M-780 computer in the Nagoya University

llo- 10 1 =[trace II' ] 1 ' 2 =F I 1- ( 68 ) Computation Center. A stopping criterion of
119' I 9't9' I' V Ctot'
I g; I <ePs for all i was imposed on the itera-
where (a 10 )i=(~ 10 )}?, tive scheme, where eps is the machine's epsilon
in double precision calculation.
and F =[ atot·T ]1/2 (69) In Fig. l(a)"-'(e), examples of unfolded re-
I' ptp '
sults are compared with assumed originals.
T-{ trace(RtA-'R)-t, for m>n, All of the results have no negative component
- trace QRt(RQRt)- 1 A(RQRt)- 1 RQ, at all, as pointed out theoretically. The orig-
for m<n. (70) inals were restored within error bars at about
Similarly for input data c, with the approxi- 80% of energy points. For the Gaussian
mations of (c)~aa and Ic~aA, shapes, however, some lower bunds of error
bars extend to negative values, but this can-
llucll . F I. 1- (71) not be helped since the present expression
ll<c>ll=:= c VCtot for I I' were obtained through a linearization
of basic equations.
Table 1 presents an example of the per-
(72) formance of the present method combined with
the algorithm proposed in Chap. ill. The val-
Hereupon, we are naturally led to an idea of ues of Ctot in the third column may be seen
error magnification rate<•l : too large to be realistic in practical experi-
ments, but they are indispensable for attaining
F= llo- 10 11/ llucll =F IF=[ ata ·_I__J112 the accuracies in the fourth column. This
- 119'11 ll<c>ll 10
c ptp atot •
situation will not be improved for actual re-
(73)
sponse matrices in considering the fact pointed
Note that F'P, Fe and F are functionals of out in the last paragraph of Sec. N-5. In this
the shapes of R and 9', not depending on respect, the fifth column gives a guide to plan
their absolute values, i.e. on a (detector effi- experiments. The values in the seventh column
ciency) and a (total number of neutrons incident show that the present method requires accept-
on the detector.) able computation time in spite of the complexity
6. Unfolded Results of equations to be solved.
All calculations were done on the F ACOM

Table 1 Example of performance of present method on M-780

computer of 20 MFLOPS in double precision calculation
Number of Total Overall Error Total
Type of pulse- Number of relative magnifi- Number of CPU No. of
spectrum height recoil error in cation iterations time figure
bins protons 9>t (%) ratett (s)
Gaussian 104 l.OxlO' 2.49 10.7 41 7.8 1(a)-(A)
8.5 MeV 52 l.OxlO' 3.02 18.3 92 12.2 1( a )-(B)
Gaussian 104 l.Ox10' 3.36 11.0 39 8.4 l(b)-(A)
14.8 MeV 52 l.Oxl0 7 3.81 17.7 115 14.3 1(b)-(B)
104 1. 0 X lOB 2.57 44.5 9 3.3 1(c)-(A)
Ramp-down 52 l.OxlOB 2.9 1(c)-(B)
1. 64 40.2 14
Ramp-up 104 2.0xlOB 3.13 49.8 21 4.1 l(d)-(A)
52 1. 0 X lOB 2.83 45.1 11 1.6 l(d)-(B)
104 4.0xlOB 2.09 56.9 20 3.9 1( e )-(A)
Flat 52 2. Ox lOB l(e)-(B)
1.89 51.5 18 2.5

t Defined by Eq. (68). tt Defined by Eq. (73).

-21-
840 ]. Nucl. Sci. Techno/.,

•••
IAI 181
•••
"'>
...a:
.. ...
~

>

.. ...
~
~
z
0
...::>
"'zo.r

... ,______1 \---~

••• ,.. ....
NEUTRON ENE ROY I "EY I "'' ... i·D
'
ao.o
NEUTRON ENERGY I "EY I
., ..
(a)

•••

...> •••
IAI 181
"
...a:
..
~a .
•
•

.. D
j \. .
.t 1
. \
••• s.o ao.o
NEUTRON ENERGY I "EY I
11-0 ... ••• ••••
NEUTRON ENEROY I "EY I
••••
(b)

...> •••
...
a:

"'~~-·
...
>
X:
;:;;
z
0
....
::::>
...z Q.l

...
••• ,.. ....
NEUTRON ENERGY I "EY I
... ••• ••••
NEUTRON ENE ROY I "EY I
JI,Q

( c)
Fig. l(a)-(e) Comparison of original spectrum (histogram) and unfolded spectrum
(dots with error bars) ; (A) for m=104 (number of pulse-height bins)
and n=77 (number of energy groups), (B) for m=52 and n=77
Vol. 26, No. 9 (Sep. 1989) 841

I.Q

IAI

.. o.•

..
>
....
.;:1 •.•
•
. ..
>
z:

......•
;,
z
~

zo.:t

Q.Q ••• 10·0 ••• o a.o 1·0 10.0 11·0

NEUTRON ENEROT I HEY I NEUTRON ENEROT I HEY I

Fig. l(d)

IAI 181
D·l
w
>

o.o

Q.Q ••• 10-0 o.o &.0 10.0

NEUTRON ENERGY I HEY I NEUTRON ENE ROT I HEY I

Fig. l(e)

mathematical rigorousness ; the proofs of con-

V. CONCLUDING REMARKS vergency of the iterative scheme and unique-
A new theory of neutron spectra unfold- ness of the solution are left for a future study.
ing, which guarantees positivity of resultant As observed in Sec. N-6, some lower bounds
solutions and correctly embodies the Poisson of error bars extended to negative values
statistics of neutron detection, has been estab- around the base of Gaussian spectra. This
lished based on the maximum entropy principle fact clearly violates the contention of the MEP
and the maximum likelihood method. Both which does not allow negative value. Hereafter,
problems of overdetermined and of under- such an expression of covariance matrix that
determined have been unified in the light of confines the lower bounds of solutions within
the principle without any artificial requirement positive region must be searched for.
nor hypothesis. Our FORTRAN program materializing the
In developing the method for solution, pri- present theory has shown satisfactory per-
ority has been given to the establishment of formance for a variety of assumed neutron
practically useful algorithm and not to its spectra.
-23-
 842 ]. Nucl. Sci. Techno!.,

While the present paper has been written R. H.: Neutron response matrix for unfolding
in a form close to recoil-proton data unfolding, NE-213 measurements to 21 MeV, ORNL;
RSIC-40, (1976).
the theory can find broad applications, since
tzl) FORTRAN SSL-li User's Guide, No. 99sp-0050-
the response matrices have been symbolically 4, Fujitsu Inc.
treated but not characterized for specific prob-
lem. Applications to Bonner detector data [APPENDIX]
and activation foil data may be especially in- 1. Brief Review of Maximum
teresting. Entropy Method
Consider a random variable x which can
ACKNOWLEDGMENT take on the values (x 11 X 2 , .. • , Xn). Then, the
The authors wish to express their thanks quantity
to Prof. T. Katoh of Department of Nuclear
Engineering, Nagoya University, for his en- (Al)
couragements in preparing the present paper
and for his thorough review of the manuscript. is known as Shannon's information entropy,
where p1 is the probability of occurrence of
-REFERENCE8- event x 1 :
(1) SLAUGHTER, D., STROUT, IT, R.: Nucl. Instrm.
Methods, 198, 349 (1982). (A2)
(2) MILLER, W. H., MEYER, W.: ibid., 205, 185
(1983).
and K is a positive constant. It had been
(3) CHAMBLES, D. A., BROADWAY, ].A.: ibid., 179,
563 (1981). proved in many textbooks on the information
(4) !ToH, S.: ibid., A251, 144 (1986). theory that the functional form of Eq. (Al) is
(5) ]AYNES, E. T.: Phys. Rev., 106, 620 (1957). the unique rational measure for the amount
(6) idem; IEEE, ssc-4[3), 227 (1968). of uncertainty represented by a discrete prob-
(7) BuRG, ]. P. : "Modern Spectrum Analysis," (ed.
ability distribution(l•l. We can make K to
CHILDERs, D. G.), (1978), IEEE Press.
(8) FRIEDEN, B. R.: ]. Opt. Soc. Am., 62[4), 511 unity hereafter with all results remaining un-
(1972). changed.
(9) GuLL, S. F., DANIELL, G.].: Nature, 272[20), The maximum entropy principle due to
686 (1978). jaynes< 5 ' asserts that we should choose just
M MINERBO, G.: Comp. Graphics Image Process.,
one probability distribution which has the
10, 48 (1979).
(ll) NILAND, R. A., ScHMIDT-HARMS, C. A.: Nucl. greatest entropy out of many different feasible
Instrm. Methods, A241, 221 (1985). ones which are all consistent with given in-
(l~ WILKINS, S. W., VARGHESE, J. N., LEHMANN, complete information. In other words, while
M.S. : Acta Cryst., A39, 47 (1983). some amount of uncertainty remains for un-
M BuRCH, S. F., GuLL, S. F., SKILLING, ]. : Comp.
available data, one should leave the uncertainty
Vision Graphics Image Process, 23, 113 (1983).
M For example, IHARA, S.: "Kakuritsu Katei to .for subsequent data, and should not try to
Entropy (Random Process and Entropy)," (in reduce it with any other proposal or hypothesis
Japanese), (1984), lwanami. than the facts we have.
(l$ ABLEs, ].G.: Astron. Astrophys. Suppl., 15, Let us consider for an example the situation
383 (1974).
where the available informations about x take
M RAo, C. R.: "Linear Statistical Inference and
Its Application," (2nd Ed.), (1973), john Wiley the form of mean values:
& Sons.
M GILL, P. E., GoLuB, G. H., MuRRAY, W., SAUN-
DERS, M.A.: Math. Comp., 28, 505 (1974).
(!.$ ARMIJO, L.: Pacific ]. Math., 16, 1 (1966).
where m < n. On the basis of this partial in-
M KENDRICK, H., SPERLING, S.M.: An introduc-
tion to the principle and use of the FERDOR formation, what is th_; inference on the prob-
unfolding code, GA-9882, (1970). ability distribution p1 ? The solution Pi> which
tzc) INGERSOLL, D. T., WEHRING, B. H., JoHNsoN, is of maximal entropy subject to the con-
-24-
 Vol. 26, No. 9 (Sep. 1989) 843

straints of Eqs. (A2) and (A3), is determined are obtained :

using the method of Lagrange multipliers, to
be of the form a
-(rv)=v (--
ar )t (A9)
ax ax '
ax ar)t v'
a (vtf)= ( ax (AlO)

where Z(A 1 , ••• , Am) is the partition function

j_(Bf)=Bar (All)
ax ax.
( 4 ) Practical Relations
and the values A; are to be determined from Used in Main Text
the constraints (A3), which reduce to the Notations used below are the same as those
relations in the main text. Apply the above general
rules to the specific forms of p, a, p and g,
a
a;= aA; ln Z(AI, ... ' Am). (A6) then the following relations can be obtained.

Note that positivity of the solution is obvious (A12)

from Eqs. (A4) and (A5), provided with real
{At}. aa
-=--·-(ua)
a a t
2. Some Useful Relations to Handle ax ax
atot
Vector Matrix Equations
( 1 ) Nomenclature
B: m X n constant matrix
v: n x 1 constant vector a
x: m x 1 variable vector =--RQs+aa,
a tot
(Al3)
r(x): scalar function of x
f(x) : n X 1 vector function of x.

( 2) Definitions
The symbols ar ;ax and ar;ax are used as
usual:

11_ =(_§]_ _§]_ ··· _§J_)t (A7) =-RtDA- 2 R(ap)+~sst(ap)

ax - ax! ' axz' ' axm ' ax atot ax '
(A14)
ar =(jh_ ~ ... aJn )t (AS)
ax - ax ' ax ' ' ax . au a ap ap
-a
X
=-aX CRQp)=RQ-aX +RM-aX.
( 3 ) General Relations
From the above definitions, the followings (Al5)

-25-