propagation of uncertainties in unfolding procedures
TRANSCRIPT
Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241
Propagation of uncertainties in unfolding procedures
Manfred Matzke*
Physikalisch-Technische Bundesanstal, Section 6.42, Bundesallee 100, D-38116 Braunschweig, Germany
Abstract
A review of methods currently used to unfold particle spectra from measured pulse height distributions or other
detector readings is given. It is pointed out that most of the measurements in particle spectrometry reveal ill-conditioned
or ill-posed problems. The presentation which is given here for examples of such inverse problems is focussed on the
algorithms used in the HEPRO unfolding program package of PTB. The question of uncertainty propagation is
discussed for least-squares algorithms as well as for those based on maximum entropy. A first attempt has been made to
quantify generally the ‘‘ambiguity’’ in ill-posed unfolding problems. The maximum entropy algorithm realized in the
MIEKE code allows a clear distinction to be made between two parts of uncertainty, one part coming from ambiguity
and one part coming from the usual uncertainty propagation. The resulting uncertainty matrix of the MIEKE code
provides these two parts. r 2002 Elsevier Science B.V. All rights reserved.
PACS: 29.30.Hs; 02.50.Ng; 29.85.+c
Keywords: Inverse problems; Spectrum unfolding; Maximum entropy; Unfolding codes; Neutron spectrometry; Uncertainty
propagation
1. Introduction
The determination of the spectral particle
fluence via unfolding of measured detector read-
ings has been investigated by many authors e.g.
[1–8]. Many papers are available, which refer to
the so-called inverse problem, where the existence
of solutions for various numbers of metrological
examples is investigated. The paper presented here
expands on the overview given in reference [1]. The
spectrometry performed by means of so-called
‘multi-channel’ or ‘few-channel’ measurements is
investigated and attention is given to the question
of uncertainty analysis and of propagation of
uncertainties in under- and overdetermined inverse
problems.
The evaluation of the spectral particle fluence
FEðEÞ from integrating measurements involves
solving the basic system of linear integral equa-
tions
z0i ¼ZRiðEÞFEðEÞ dE i ¼ 1;y;M ð1Þ
which represent the model of the measurement.
The vector zT0 ¼ ðz01;yz0i;yz0MÞ denotes the
expectation values of the (measured) readings of
the detector system, where the actual readings are
z00 ¼ z0 þ e with the statistically fluctuating quan-
tity e (T meaning transposition). It is assumed that
the uncertainty matrix (covariance matrix) Sz0 of
the vector z0 is known. The kernels RiðEÞ are the*Tel.: +49-531-592-6420; fax: +49 531-592-7015.
E-mail address: [email protected] (M. Matzke).
0168-9002/02/$ - see front matter r 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 8 - 9 0 0 2 ( 0 1 ) 0 1 4 3 8 - 3
response functions for the energy E of the Mdetector channels of the measuring system.
Mathematically, Eq. (1) is a degenerate case of
the Fredholm integral equation of the first kind. It
has no unique solution since a finite number of
discrete measurements cannot define a continuous
function FEðEÞ: Eq. (1) can be transformed at least
approximately into the discretised linear matrix
equation
z0 ¼ RU ð2Þ
with the group fluence vector UT¼ðF1; :::;Fn; :::;FNÞ: The N components Fn can be regarded as
average fluence values in the intervals between the
energies En and Enþ1: Eq. (2), too, has no unique
solution for N >M:In the overdetermined case of NoM (or for
N ¼M), the solution spectrum and its uncertainty
matrix is often calculated by the least-squares
method [see 1,2,5,7–10], where the quantity
w2 ¼ ðz0 � R � UÞT � S�1z0 � ðz0 � R � UÞ ð3Þ
is minimized, provided Sz0 is non-singular. The
solution of Eq. (3) is obtained from the so-called
normal equations
RT � S�1z0 � z0 ¼ RT � S�1
z0 � R � U ¼ B � U; ð4Þ
which have to be solved for U: Eq. (4) has a uniquesolution if the rank of the matrix
B ¼ RT � S�1z0 � R ð5Þ
is maximum, i.e. if it equals the number of fluence
groups N: Matrix B is sometimes called the
structure matrix [11].
In practice, it has been found that in many cases
the normal Eq. (4) are very close to singular
[1,3,9,12], even for a maximum rank of R: Thesituation is then similar to the underdetermined case
ðN >MÞ; and it must be concluded that, in practice,there is probably a manifold of possible solutions of
Eq. (4). This ambiguity may be extremely hard to
notice in complicated problems [12].
There are some possibilities of overcoming these
difficulties. First, all the information available on
the spectrum must be used within the unfolding
algorithm. Sometimes a pre-calculated fluence
vector U0 with uncertainty matrix SF0 is known,
sometimes a smoothing condition for the fluence
can be formulated. This a priori knowledge can be
used with Eq. (3) to construct a solvable system of
normal equations. However, if SF0 contains large
uncertainty components, non-physical results such
as negative elements of the fluence vector may be
obtained.
It is a conditio sine qua non for the resulting
fluence vector to be non-negative for all particle
energies. The discussion must therefore be ex-
tended to algorithms which include this condition,
such as the logarithmic least-squares method
[8,13–15] or, more general, to algorithms based
on the principle of maximum entropy with
constraints based on given information.
Computer programs based on this principle
have been developed at the PTB and are contained
in the HEPRO program system together with
other programs [1,2,4,13]. It is the aim of this
paper to review some algorithms used in particle
spectra unfolding (mainly those of the HEPRO
package) and to investigate how the propagation
of uncertainties can be performed.
2. Methods, algorithms and codes
A number of methods have been developed for
solving the inverse problem; details may be found
elsewhere in textbooks. Only a small overview can
be given here, concentrated on the algorithms used
in the HEPRO program system [13]. During the
recent years, however, important progress has been
achieved in the development of other methods and
codes not included in the HEPRO package. Some
of these methods are mentioned here, too.
In reactor dosimetry for ‘few channel’ unfolding
the codes based on linear least-squares methods
were been successfully applied during the REAL-
84 exercise [16]. STAY’SL [7], LEPRICON [17],
LSL [14], DIFBAS [18] and MSITER and
MINCHI [19] are representative of a large number
of available least-squares adjustment codes using a
priori information on the fluence. These codes are
also successfully applied to the unfolding of the
measured data obtained by Bonner spheres. In a
second group, the codes excluding negative fluence
values, e. g. SAND-II [15], GRAVEL [13], LSL-
M2 [14], LOUHI [20], BUNKI [21], have to be
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241 231
noticed. These codes perform in principle a non-
linear least-square adjustment, with the constraint
of non-negative particle fluence. In this context the
approach of Schmittroth [8] has to be mentioned,
where a log-normal distribution of the variables is
used to avoid negative fluence values. Secondly
Schmittroth uses correlations in the a priori
spectrum as a means of ensuring smoothness.
Some of the least-squares codes use other smooth-
ing or a regularizing procedures to construct a
non-singular matrix B in Eq. (5), e.g. a Tikhonov
regularization [22] or a condition of small curva-
ture or small fluctuation, e. g. the FERDOR code
[23,24]. In this context, the singular value decom-
position [3,12] should be mentioned here.
Recently, genetic algorithms (e. g. those of
Freeman et al. [25] and Mukherje [26]) have been
applied in particle spectrometry, where an evolu-
tion model is used together with a ‘‘survival’’
strategy in order to solve complex mathematical
equations. Here, the optimisation criterion (defini-
tion of the fittest) is an important model para-
meter. Some of the codes mentioned here allow an
uncertainty analysis to be performed, namely
those codes based on the least-squares method.
Uptill now it has not been quite clear how to
propagate the uncertainties for the evolution
models or the models referring to neural networks.
Some progress has been made in this field by
introduction of Bayesian methods into neural
network theories [27–29].
Bayesian methods or methods based on the
maximum entropy method have been realized in
the MIEKE and UNFANA programs of the
HEPRO package [13] and in the MAXED code,
recently developed [38,39]. These codes make
consistent uncertainty propagation possible, how-
ever, it is not yet completely understood how the
uncertainty resulting from the ‘‘ambiguity’’ of the
solution (singularity of matrix B) can be taken into
account; a first trial is contained in the MIEKE
code (see below).
3. Uncertainties in least-squares adjustment
It is assumed that a pulse height spectrum or a
number of detector readings z0 with the uncer-
tainty matrix Sz0 have been determined. For Sz0;only a diagonal form is sometimes used with the
variances of z0 (often determined by Poisson
statistics). It is further assumed that pre-informa-
tion R0 on the response matrix is available (or that
some parameters p0 of the response matrix are
known before) with the uncertainty matrix SR0 or
Sp0: If, in addition, the particle fluence U0 was
previously calculated or measured by another
independent ‘‘measurement’’ (e.g. by a transport
code calculation) with the uncertainty matrix SF0;the general w2 expression:
w2 ¼ðz0 � zÞT � S�1z0 � ðz0 � zÞ
þ ðR0 � RÞT � S�1R0 � ðR0 � RÞ
þ ðU0 � UÞT � S�1F0 � ðU0 � UÞ ð6Þ
has to be minimized under constraint z ¼ RðpÞ � Uwith respect to z, R (or p) and U. Eq. (6) is the
general w2 expression for an adjustment procedure,when the parameters or the fluence are already
known to a certain extent and the new measure-
ment z0 is used for adjustment only. The various
algorithms used to find the minimum of Eq. (6)
make a distinction in the modelling of the last two
terms. There are only some codes, mainly used in
reactor dosimetry, where the a priori information
on the response matrix can be taken into account
(e.g. STAY’SL [7], LSL [14], LEPRICON [17],
MSITER and MINCHI of PTB [19]); in all other
cases, the second term in Eq. (6) is missing. If
nothing is known a priori on the fluence, the last
term in Eq. (6) may be replaced by a smoothing or
shape condition (see e.g. [20,23,24,30]) in order to
obtain a non-singular system of normal equations
for the solution as mentioned in the introduction.
With the constraint z ¼ R � U the minimization
of Eq. (6) leads to non-linear normal equations.
Since it can be assumed that the adjusted values
will not be too far from the a priori information
values, this constraint equation is replaced by a
Taylor approximation in the vicinity of R0 and U0
[13], leading to the linear expression
z � R � U ) z � z1 � zR � ðR � R1Þ � zF � ðU � U1Þ;
ð7Þ
where z1 ¼ R0 � U0; and zR and zF are matrix
derivatives of z at R0 and U0; where zF ¼ R0 and
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241232
zR ¼ U0: With Eq. (7), the linear least-squares
adjustment is performed. The solution for the
fluence vector and its uncertainty matrix is [13]:
U ¼ U0 þ SF0 � zF � W�1 � ðz0 � R0 � U0Þ;
SF ¼ SF0 � SF0 � zF � W�1 � zF � SF0:ð8Þ
with the weighting matrix
W ¼ Sz0 þ zR � SR0 � zR þ zF � SF0 � zF ð9Þ
(transposition not indicated here and in the
following).
From Eq. (8) it is apparent that the introduction
of a priori information into the least-squares
formalism leads to normal equations, where only
an inversion of the M M matrix W instead of
the N N matrix B of Eq. (5) is required. Due to
the minus sign in the uncertainty equation, the
uncertainty is reduced after adjustment. From
Eq. (8) it is also apparent that the solutions and
their uncertainty matrices strongly depend on the a
priori information values U0 and the a priori
uncertainty matrix SF0: It is also seen that negativefluence values may appear here.
In practice, the a priori spectrum is often not
known on an absolute scale. In this case, the
scaling factor of the fluence has also to be
determined by the unfolding code. A non-linear
evaluation of this factor has been included in the
MSITER(MINCHI) program of PTB. It is not
included in the STAY’SL code.
From Eq. (6) it can be concluded that the a
priori information on both, the response matrix R0
and the fluence U0 is considered in the same way, i.
e. it is considered as if both quantities are intended
to be adjusted. This is the philosophy of the ISO
Guide [31], the German DIN Standard DIN1319
[32] and of the LEPRICON [17] methodology: all
measured quantities, those currently measured and
data quantified before have to be included into the
adjustment procedure, i.e. the ‘‘properties of the
entire world’’ are influenced by the current
measurement.
There is another way of adjustment not yet
established in particle fluence unfolding, namely to
consider the response function as a so-called
‘‘parameter’’ and applying the adjustment proce-
dure with this ‘‘parameter’’ being constant during
the unfolding procedure, and performing the
uncertainty propagation afterwards. In this case,
the second term of the right-hand side of Eq. (6) is
missing. The solution of this model is similar to
Eq. (8), but the term with SR0 is missing in the
expression for matrix W. From the uncertainty
propagation one finds (with new matrix W0):
SF ¼SF0 � SF0 � zF � W0�1 � zF � SF0
þ SF0 �zF �W0�1 �zR �SR0 � zR � W0�1 � zF � SF0:
ð10Þ
This model is different from that usually taken
in least-squares adjustment, and it has obviously
not been used to date in the standard software like
STAY’SL [7]; It offers advantages in cases where
the response functions are known with small
uncertainties.
Some remarks have to be made on the
construction of the matrices SF0 and SR0; whichare to date, in general, not available from the
transport code calculation used to derive U0 and
R0: In many cases, one has a rough idea of the
slope of the particle spectrum. In a reactor
moderator, for instance, a moderated fission
spectrum is expected, expressed as a sum of
parameterised subspectra with a number of kparameters (e. g. the temperature parameter of a
maxwell spectrum or the slope of a 1=E spectrum).
If the subspectra and the uncertainties of their
parameters are used to construct the matrix SF0; itturns out that SF0 has deficient rank rEk:Generally, S�1
F0 then is non-existent, however, a
solution vector U may be compiled from Eq. (8). It
can be shown that the possible solution spectra are
restricted to a very small subspace spanned by the
rEk eigenvectors of SF0; which means in principle
that the solution spectrum is a superposition of the
k subspectra defined a priori, thus putting severe
constraints on the solution spectrum by the special
choice of the a priori covariance matrix [37]. It is,
therefore, preferable to use a diagonal SF0 instead
in order that the entire solution space be available.
This in compliance with the German standard
DIN 1319 [32]: ‘‘If nothing is known on the
correlations, a diagonal uncertainty matrix must
be used’’.
In practical examples the uncertainty matrix SR0should be constructed in a way already outlined in
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241 233
the ENDF uncertainty file [33], and it is recom-
mended that the format used there should also be
used, for instance, for establishing the uncertainty
matrix of the Bonner spheres response function.
Apart from reactor dosimetry, where the use of a
full uncertainty matrix SR0 compiled from ENDF
data is considered to be the state of the art, only a
few attempts have been made in the past to include
response matrix uncertainties into the unfolding
process for other problems [34]. The HEPRO
package [13] provides the possibility of including
uncertainties of energy calibration and resolution
parameters in multichannel unfolding problems.
4. Non-linear least-squares methods
The linear least-squares method, well estab-
lished in the STAY’SL [7] code and the LEPRI-
CON methodology [17], is the unfolding algorithm
recommended when good a priori information and
consistent measurements are available. The dis-
advantage is that negative fluence values cannot be
excluded. To take into account the condition of
non-negative fluence, a method first realized in the
SAND-II code [15], later used in the LSL-M2 [14]
and GRAVEL [13] codes, can be applied. In this
case, not the group fluence elements Fn but their
logarithms ln ðFnÞ are determined by a special
iteration procedure which minimizes an expression
similar to Eq. (3) with the logarithms of the zoiinstead of zoi: A similar method is used in the
LOUHI code [20] where the unknown quantities
Fn are expressed as squares of real numbers.
The SAND-II (GRAVEL) solution is deter-
mined by a special gradient method [13]: iteration
is started with a spectrum Fð1Þn : Weights w
ð1Þin ¼
Rin � Fð1Þn =zð1Þi with z
ð1Þi ¼
Pn Rin � expðlnF
ð1Þn Þ are
calculated. In each iteration step the current
solution ðkþ 1Þ is obtained from the previous
solution ðkÞ by
lnFðkþ1Þm � lnFðkÞ
m ¼ lðkÞm �Xi
ðln z0i � ln zðkÞi Þ �
wðkÞim
r2i;
where lðkÞm ¼Xi
wðkÞim
r2i
!�1
: ð11Þ
The ri are the relative standard deviations of the
z0i; namely ri ¼ s0i=z0i which were not included in
the original SAND-II code.
For the iteration procedure in the SAND-II
(GRAVEL) code, a first input spectrum is needed
when the iteration is started. A solution always
exists [13], but the solution spectrum depends on
this input spectrum in a way which is not quite
transparent so that an uncertainty propagation
cannot be easily performed. It has been found [13]
that the matrix B of Eq. (5) controls the variety of
solutions; for non-singular B a unique solution
exists, for ill-conditioned B a manifold of solutions
may exist, depending on the rank of B.
For the uncertainty analysis, the method of
randomly chosen start spectra has been tried
[35,42], with final averaging over all solutions.
This gives a (model-dependent) plausible uncer-
tainty of the solution spectrum, including correla-
tions. This uncertainty represents the variety
(ambiguity) of the solutions and cannot take into
account the normal uncertainty propagation of the
uncertainty matrices Sz0 and SR0:The other non-linear least-squares methods
available may be considered as special unfolding
models which are used to regularize the matrix B
(e.g. [22–24]). Additional parameters (e.g. for
smoothing) are introduced to get a unique solu-
tion. Uncertainty propagation can be performed
by these algorithms but has obviously not yet been
established in the codes available. The uncertain-
ties of the solutions depend on the uncertainties of
the input data and, also, of the parameters
introduced.
5. Maximum entropy, MIEKE, UNFANA,
MAXED
The properties of the least-squares codes are less
satisfactory, when there is a lack of pre-informa-
tion on the particle spectrum. In this case, the
singularity or quasi-singularity of the matrix B of
Eq. (4) may lead to an ambiguity of the solutions,
even if the constraint of non-negative fluence is
included into the least-squares algorithm. For the
equivalent maximum likelihood method it follows
that a unique, most probable particle spectrum
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241234
might not exist. However, similar to the maximum
likelihood method, it is possible to construct a
posteriori probability density of the fluence using
Bayes’ theorem [36]. The solution spectrum can
then be defined as the expectation value of U;denoted by Uh i; which can be calculated from the
probability density. This method was used in Ref.
[9]; later, a more modern interpretation was given
in Refs. [2,4] using the principle of maximum
entropy.
From statistics and information theory it is
known that a statistical system of a multidimen-
sional random variable x prefers a probability
density PðxÞ; for which the expectation value of theso-called entropy [2,36]
S ¼ �ZPðxÞlogðPðxÞÞ dx ð12Þ
has a maximum value. When the constraining
conditions to be imposed are included, it is
possible to formulate an extremum principle for
the determination of the fluence probability
density PðUÞ: For the case realized in the MIEKE
code [1,13], two constraints are assumed to exist:
(a) all group fluence values should be non-
negative, and (b) the expectation value w2� �
of w2
(see Eq. (3)), should be equal to the number of
degrees of freedom involved, which is equal to the
number M of elements in z0: By means of the
extremalisation procedure, the probability density
PðUÞ ¼ C1 � exp �b2� w2ðUÞ
� for all FnX0 ð13Þ
is obtained, which is defined for all FnX0 and
vanishes for negative values due to condition (a).
C1 is the normalization constant. b is a
‘‘temperature’’ parameter to be determined from
condition (b). The probability density PðUÞrepresents a multivariate normal distribution with
a w2-exponent, truncated because of condition (a).
Although the exponent is degenerate in the case
N >M (or for ill-conditioned matrix B), the
distribution can be normalized and the expectation
value Uh i with its uncertainty matrix
SF ¼ UUh i � Uh i Uh i ð14Þ
can be calculated. For b ¼ 0:5; PðUÞ is equivalentto the likelihood expression used in Bayes’
theorem if a normal distribution is assumed for
the measured data.
The probability density of Eq. (13) is introduced
into the MIEKE code of the HEPRO package [13].
A Monte Carlo code with importance sampling is
used to calculate expectation values. For the
MIEKE code, the computing time turned out to
be very long, in particular for M5N: Weise [4]
therefore proposed an analytical approach to the
Monte Carlo results to reduce the computing time.
Weise replaced the distribution of Eq. (13) by the
simpler exponential ansatz [2,4]:
PðUÞ ¼ C1 � expð�bT � R � UÞ ð15Þ
and used the same constraints for the extremum
principle of maximum entropy. The parameters b
for maximum entropy can be obtained by a simple
solution of a non-linear matrix equation. Here, the
expectation values can easily be calculated. The
covariance matrix SF of the distribution turns out
to be diagonal and must be properly interpreted.
To obtain the uncertainty associated with the
expectation value Uh i due to the uncertainty
matrix Szo; the Gaussian law of uncertainty
propagation is used. The subroutine UNFANA
[4] of the SPECAN code uses Eq. (15) and
performs this analysis.
Weise [4] has shown that for N-N the
expectation values Uh i obtained are identical for
both distributions of Eqs. (13) and (15). It has
been found from a variety of examples that there is
good agreement for MEN; too.It should be noted that, in principle, a priori
information could be included in both distribu-
tions (Eqs. (13) and (15)) by using for w2 an
expression similar to Eq. (6) instead of Eq. (3). For
the algorithm used in the MAXED code [38], this a
priori information is already included into the
maximum entropy formulas. In MAXED, the
(normalized) spectral fluence is taken as a prob-
ability density, leading to very interesting equa-
tions. The uncertainty propagation [39] has been
implemented recently, the code runs within the
HEPRO package.
For the uncertainty analysis it must be taken
into account that, in principle, two fluctuation
components exist, one resulting from the uncer-
tainties of the input quantities x0 (readings,
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241 235
response functions, a priori information on the
fluence), one resulting from the ambiguity of the
solution (compare, for example with the micro-
scopic and macroscopic fluctuations in particle
statistics). The fluctuation of the resulting fluence
can be considered as being composed of both
contributions:
dU ¼ oqUqx0
> dx0 þ dU1 ð16Þ
perpendicular to each other, i.e. dx0dU1h i ¼ 0:The resulting covariance matrix is thus the sum
of two components: ðq Uh i=qx0ÞSx0ðq Uh i=qx0Þ þSF1: For the MIEKE algorithm, the corresponding
formulas are quite simple to obtain from the
probability distribution. Considering only the
uncertainties Sz0 of z0, it can be shown
thatq Uh i=qz0 ¼ �bSF � R � S�1z0 ; and with the de-
finition of B (Eq. (5)) one finds SF1 ¼ SF � b2SF �B � SF: In linear least-squares methods b2SF ¼B�1; which holds for non-singular B and for a
Gaussian distribution. In this case, the solution is
unique and SF1 vanishes. The ambiguity part SF1
cannot be obtained so easily for the other
maximum entropy models. For UNFANA and
MAXED, only the conventional uncertainty pro-
pagation corresponding to z0 and (in principle to
R0) is included. It has been found that the part
from the uncertainty propagation of z0 for the
MIEKE code ðb2SF � B � SFÞ agrees well with the
corresponding result of UNFANA.
For the MIEKE algorithm one obtains in
obvious matrix notation (indices have to be chosen
carefully):
q Uh iqR
¼ bfSF � S�1z0 � z0
� UUUh i � Uh i UUh ið Þ � R � S�1z0 g
As is usual in uncertainty propagation, the
triple correlation function can be approximated
in Gaussian approximation: F1 � F1h ið ÞhF2 � F2h ið Þ F1 � F1h ið Þi ¼ 0: This means that
the usual propagation of uncertainties even for
the response matrix, can be compiled after an
unfolding run, provided the resulting covariance
matrix SF given by Eq. (14) is compiled by the
code.
It should be mentioned that the maximum
entropy algorithms allow other probability dis-
tributions to be constructed [40]. In the MIEKE
code the evaluation of PðH�Þ for the dose-
equivalent H� ¼P
n hFnFn ¼ hF � U is performed
(hF: fluence-to-dose conversion factor):
PðH�Þ ¼ dðH� �hF � UÞh i ð17Þ
The variance H�H�h i � H�h i H�h i ¼ hFSFhFof the distribution again includes contributions
from the uncertainty propagation and the ambi-
guity.
6. Results obtained from practical examples
Two groups of unfolding tasks are important in
neutron spectrometry: ‘‘few channel’’ and ‘‘multi
channel’’ measurements. As examples representa-
tive of ‘‘few channel’’ spectrometry, the measure-
ments in reactor dosimetry have to be mentioned
(e.g. [16]), where activities or reaction rates of
different irradiation probes are determined. It was
shown in the past that exact uncertainty propaga-
tion is possible in this field [1]. The other
important examples are from neutron dosimetry,
namely the superheated drop (bubble) detectors
(e.g. [41]) and the Bonner sphere detectors (e.g.
[42,43]).
6.1. ‘‘Few channel’’ analysis
The determination of neutron spectra from the
readings z0 of a Bonner sphere system is a
representative example of ‘‘few channel’’ unfold-
ing. Count rates forM detectors are measured (Mbeing of the order of 10), and it is intended to
evaluate the spectral neutron fluence or the dose
equivalent in 50–100 energy groups. Obviously,
the a priori information on the fluence is more
important here than in the multichannel case.
Adding of a priori information during unfolding is
equivalent to increasing the number of measure-
ments fromM toM þN: This means that a prioriinformation controls the results more than the
measured values. The results obtained in various
examples confirm this conjecture.
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241236
In the present version of the MIEKE and
UNFANA codes, a priori information cannot be
taken into account. If real a priori information is
available (no guess!), the only codes which should
be used are those based on Eq. (6), e.g. STAY’SL,
LEPRICON, DIFBAS, MSITER(MINCHI),
LSL. However, there is a multitude of examples
(e.g. Refs. [11,44]), where neutron spectra have
also been successfully unfolded using only guess
functions together with the SAND–II (GRAVEL)
program. Careful investigation of these methods
allows the conclusion to be drawn that a priori
information is introduced as well, although in an
interactive procedure, which is not easy to
quantify mathematically.
A lot of experience is needed for a successful
unfolding of neutron spectra employing the
SAND-II or the GRAVEL code. For instance,
the use of a thermal or a fission spectrum as the
starting guess spectrum in SAND-II or GRAVEL
means quantifying a priori information. On the
other hand, when codes are used which need no
pre-information at all, like MIEKE and UNFA-
NA, it is hardly possible to unfold a reasonable
neutron spectrum. However, due to negative
correlations between adjacent neutron groups of
the solution spectrum, resulting from the matrix
SF1; it has been shown that integrals over the
fluence, such as the total dose equivalent, can be
calculated with much smaller uncertainties [1,13].
The following example presents the results of
measurements with Bonner spheres performed in a
moderated 252Cf reference field [45]. Twelve
Bonner spheres were irradiated; the relative
uncertainties of the readings (uncertainty of a
common scaling factor of the response functions
included) were between 1.2% and 4%. A priori
information was available from MCNP calcula-
tions [46], the relative uncertainties of the MCNP
calculations were estimated to be 15% below
100 keV and 10% above this value.
From Fig. 1 it is apparent that the MINCHI
result is nearly the a priori spectrum with the
exception of the thermal part, where the MCNP
results are uncertain because of the not well-
known components of reflecting concrete. The
SAND-II result [45] was obtained by an educated
guess using various combinations of thermal,
intermediate and fission spectra. From the UN-
FANA result the spectrum can hardly be recog-
nized, since available a priori information was not
included. Nevertheless, it was found that integral
values such as the total fluence or the total dose
equivalent agree within the uncertainties.
This also holds for the integral values obtained
with the MIEKE code. The uncertainty of the dose
equivalent is the sum of two nearly equal parts
(from ambiguity matrix SF1 and from uncertainty
propagation) which can be seen in Fig. 2, where
the probability distribution of H� (as obtained by
MIEKE) is given.
Considering the other examples of Bonner
spheres unfolding given in the literature, it must
be concluded that in moderated fission spectra like
reactor surroundings the fluence can in general be
Fig. 1. Spectral fluence of a moderated 252Cf neutron field
unfolded from the readings of 12 Bonner spheres.
Fig. 2. Probability density of H� in a moderated 252Cf neutron
field due to Eq. (17). There are two contributions to the
uncertainty coming from ambiguity and from the usual
propagation of uncertainties. The peak is asymmetric.
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241 237
determined with a relative uncertainty of about
5% and the total dose equivalent with an
uncertainty of less than 15%.
6.2. ‘‘Multichannel’’ analysis
Successful unfolding of a measured pulse height
spectrum requires that the particle response func-
tion for each particle energy and each channel be
known. The progress made over the past years in
Monte Carlo calculations of response functions
has induced progress in unfolding techniques, too.
As regards gas-filled neutron detectors, the GNSR
code [47] and the SPHERE code [54] have to be
mentioned which calculate the neutron response
function in various neutron beam configurations
for cylindrical and spherical detectors. Included
are recoil detectors filled with H2,3He, 4He, and
mixtures with quenching gases commonly used.
For organic scintillators, the NRESP4 [48], EGS4
[49], and MCNP [46] codes have been used to
calculate neutron and photon response functions
for NE-213 [50]. In practice, only the so-called
ideal response function for an energy En of an
incoming particle is calculated by determining the
distribution of energy En deposited within the
detector. The other parameters, referring to energy
calibration and resolution, have to be determined
separately by the experimenter [13].
For practical unfolding work the question
arises, how many multichannel and fluence groups
should reasonably be chosen. It has been found in
Ref. [1] that the bin width of the pulse height
channels should be of the order of 1/5 of the
FWHM; the same holds for the corresponding
fluence groups.
The organic liquid scintillator NE-213 is very
well suited for measurements in mixed neutron-
photon fields because of its excellent neutron–
photon discrimination capabilities. Photons
produce scintillations by the compton and pair-
production interactions, which, together with wall
effects, lead to a broad response function for
monoenergetic incoming photons [51]. In the
following, the photon unfolding example of a
cylindrical NE-213 detector (5 cm in radius, 5 cm
in height) irradiated in the high-energy reference
photon field of the PTB [50] is considered. Details
of the experiment can be found in Refs. [50–52].
The response functions for monoenergetic
photons were calculated by Novotny [51] using
the EGS4 code. There is a non-linear relationship
between the measured pulse height and the
corresponding calculated light output. This rela-
tionship was experimentally determined together
with the resolution parameters from irradiations in
reference photon fields (see also Ref. [52]). In the
photon spectrum, apart from a peak at 0.511MeV,
three lines were expected between 1.2 and 1.5MeV
and, in addition, lines at 6.1, 6.9 and 7.1MeV.
The upper part of Fig. 3 shows the measured
pulse height distribution. The resolution for a
pulse height at the COMPTON edge amounts to
about 400 keV (6%) [51] (full width at half
maximum). The unfolded results are shown in
the lower part of Fig 3, where it can be seen that
lines of FWHM of about 50 keV (0.8%) could be
resolved in the unfolded photon spectrum.
The resolution of the unfolded peaks of Fig. 3
depends strongly on the uncertainties of the
measured pulse height spectrum and on the
resolution of the detector used. The spectra (a)
and (b) of Fig 4 were obtained from the same pulse
height spectrum of Fig. 3 but with an uncertainty
of the multichannel counts assumed to be two or
four times higher. In the lower curve (c) the results
Fig. 3. Experimental pulse height spectrum (a) with properly
scaled abscissa and unfolded photon spectrum (b) in the high-
energy photon field of the PTB.
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241238
of a simulation are shown, where the resolution of
the detector was assumed to be twice the resolu-
tion of the detector from Fig. 3.
There was good agreement between the un-
folded photon spectra and the spectra determined
previously with a germanium detector [50]. In-
tegrals over peak areas were reproduced within the
range of uncertainties [51].
The uncertainty analysis with the MIEKE code
results in large uncertainties of the fluence in the
individual energy groups, but the results of
adjacent groups are strongly anti-correlated, as a
result of the ambiguity part SF1; this leads to muchsmaller uncertainties of fluence integrals over
broader energy groups.
7. Eigenvalues and limits of unfolding
It has been shown in previous papers [1,53] that
there is a criterion to quantify the amount of
information available in an unfolding process. The
number of non-vanishing eigenvalues of the matrix
B (Eq. (5)) determines the amount of ambiguity,
or, which is equivalent, the number of independent
parameters available in the solution spectrum.
Values near zero of the ratio r (ratio of an
individual eigenvalue to the largest one) cause
high uncertainties in the solution spectrum and are
related to the ambiguity involved. Non-negative
eigenvalues with low ratio r near the floating pointprecision of the computer may in addition lead to
numerical errors in unfolding.
To quantify the resulting precision of an
unfolding process, the number nl of eigenvalues,
with a ratio r above a certain limit (e. g. r > 10�4),
should be compiled. nl might be compared with
the maximum possible rank nr (number of neutrongroups) of the matrix B. For the Bonner sphere
example of Fig. 1, nl ¼ 5 ðnr ¼ 52Þ has been
obtained, which means that in principle only five
‘‘parameters’’ of the solution spectrum can be
determined and that a priori information is
required for successful unfolding.
For the NE-213 example, strong dependency of
nl on the resolution function was found. The data
used for Figs. 3 and 4 lead to nl ¼ 70 ðnr ¼ 278Þfor unfolding with the ‘‘ideal’’ response function
(ideal resolution), to nl ¼ 45 ðnr ¼ 278Þ for the
experimental resolution available (used in Fig. 3)
and to nl ¼ 26 ðnr ¼ 278Þ for the resolution used inFig 4c. nl does not depend on the uncertainties of
the measured values z0: The loss of resolution in anunfolding process due to ‘‘poor statistics’’ is a
consequence of the principle of maximum entropy
(from all possible solutions, those of minimum
curvature are taken).
The ambiguity part SF1 always leads to large
uncertainties but strong anti-correlations of the
fluence values in adjacent energy groups so that
integral values can be obtained with much lower
uncertainty. Recently attempts have been made to
unfold the full energy- and angle-dependent
differential neutron fluence from the measure-
ments of six electronic dosimeters mounted on a
sphere [53]. High uncertainties were obtained for
the unfolded fluence, since nl ¼ 13 ðnr ¼ 450Þ wasfound. [53]. However, integral values like the
personal dose equivalent could be compiled after
unfolding with a reasonably small uncertainty [55].
8. Summary
In the present paper, it has been tried to review
some of the unfolding algorithms used in neutron
Fig. 4. (a) and (b): Broadening of the lines due to higher
uncertainties, (c): due to increased resolution parameters.
Results obtained with the MIEKE code.
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241 239
physics and to quantify the amount of information
involved in an unfolding process. As a first attempt
the ambiguity part due to the underdetermination
was analysed by means of the MIEKE code. On
the basis of the results of the discussion of the
algorithms and the examples presented here, the
following statements can be made:
Unfolding of measured detector readings in
particle spectrometry often constitutes a system of
ill-conditioned normal equations, even in the case
M > N: The matrix to be inverted might be
numerically unstable, in particular for Gaussian
broadened response functions.
Unfolding codes should not be used as ‘‘black
boxes’’. Some experience is required, and it is
recommended to use more than one of the codes
mentioned here. In the HEPRO package [13], the
GRAVEL, MIEKE and UNFANA codes may be
used in succession.
References
[1] M. Matzke, G. Vourvopoulos (Ed.), International Con-
ference: Neutrons in Research and Industry, Crete, Greece,
1996, Proc. of SPIE 2867 (1997), 598–607, ISBN 0-8194-
2263-0.
[2] K. Weise, M. Matzke, Nucl. Instr. and Meth. A 280 (1989)
103–112.
[3] J.T. Routti, J.V. Sandberg, Rad. Prot. Dosim. 10 (1985)
103–110.
[4] K. Weise, Report PTB-N-24, Physikalisch-Technische
Bundesanstalt, Braunschweig, 1995.
[5] J. Pulp!an, M. Kr!al!ık, Nucl. Instr. and Meth. A 325 (1993)
314–318.
[6] V.B. Anykeyev, A.A. Spiridonov, P. Zhigunov, Nucl.
Instr. and Meth. A 325 (1993) 314–318.
[7] F.G. Perey, Report ORNL/TM-6062, Oak Ridge National
Laboratory, Oak Ridge, 1978.
[8] F. Schmittroth, Nucl. Sci. Eng. 72 (1979) 19–34.
[9] M. Matzke, K. Weise, Nucl. Instr. and Meth. A 234 (1985)
324–330.
[10] M. Tichy, Nucl. Instr. and Meth. A 267 (1988) 502–508.
[11] A.V. Alevra, B.R.L. Siebert, A. Aroua, M. Buxerolle, M.
Grecescu, M. Matzke, C.A. Perks, H. Schraube, D.J.
Thomas, H.L. Zaborowski, Report PTB-7.22-90-1, Physi-
kalisch-Technische Bundesanstalt, Braunschweig, 1990.
[12] W.H. Press, B.P. Flannery, W.T. Vetterling, Numerical
Recipes, Cambridge University, Cambridge, 1986.
[13] M. Matzke, Report PTB-N-19, Physikalisch-Technische
Bundesanstalt, Braunschweig, 1994.
[14] F.W. Stallmann, Report NUREG/CR-4349, ORNL/TM-
9933, Oak Ridge National Laboratory, Oak Ridge, 1986.
[15] W.N. McElroy, S. Berg, T. Crockett, R.G. Hawkins,
Report AFWL-TR-67-41, US Air Force Weapons La-
boratory, 1967.
[16] W.L. Zijp, E.M. Zsolnay, H.J. Nolthenius, E.J. Szondi,
Report ECN-212; IAEA-NDS-212, Netherlands Energy
Research Foundation, Petten, 1988.
[17] R.E. Maerker, B.L. Broadhead, J.J. Wagschal, Nucl. Sci.
Eng. 91 (2) (1986) 137–170.
[18] M. Tichy, Report PTB-7.2-93-1, Physikalisch-Technische
Bundesanstalt, Braunschweig, 1993.
[19] N. Kocherov, Neutron metrology file NMF-90, Report
IAEA-NDS-171, International Nuclear Energy Agency,
Vienna, 1996 (the MINCHI code is a revised version of
MSITER).
[20] J.T. Routti, J.V. Sandberg, Comput. Phys. Commun. 21
(1980) 119–144.
[21] K.A. Lowry, T.L. Johnson, NRL Report 5340, Naval
Research Laboratory, Washington DC, 1984.
[22] J. Weese, Comput. Phys. Commun. 69 (1992) 99–111.
[23] W.R. Burrus, J.D. Drischler, Report ORNL-4154, Oak
Ridge, 1965.
[24] R.H. Johnson, B.W. Wehring, Report ORNL/RSIC-40,
Oak Ridge, 1976, pp. 33–39.
[25] D.W. Freeman, D.R. Edwards, A.E. Bolon, Nucl. Instr.
and Meth. A 425 (1999) 549–576.
[26] B. Mukherjee, Nucl. Instr. and Meth. A 432 (1999) 305–
312.
[27] D.J.C. MacKay, Nucl. Instr. and Meth. A 354 (1995)
73–80.
[28] E. Cordes, G. Fehrenbacher, R. Schuetz, M. Sprumck, K.
Hahn, R. Hofmann, J.P. Biersack, W. Wahl, IEE Trans.
Nucl. Sci. 45 (1998) 1464–1469.
[29] C.C. Braga, M. S. Dias, Nucl. Instr. and Meth. A 476
(2002) 252, these proceedings.
[30] C.A. Oster, in: Current Status of Neutron Spectrum
Unfolding, Proc. of a Technical Committee Meeting,
Oak Ridge 1977, Report IAEA-TECDOC-221, Vienna
1979, pp. 29–44.
[31] ISO Guide, Guide to expression of uncertainty in
measurement, International Organisation for Standardiza-
tion, Geneva, 1995.
[32] DIN 1319-4 Grundlagen der Messtechnik, Beuth Verlag,
Berlin, 1999.
[33] P.F. Rose, C.L. Dunford, ENDF-102, Brookhaven Na-
tional Laboratory, 1990.
[34] M. Matzke, W.G. Alberts, E. Dietz, in: H. Farrar, E.P.
Lippincott, J.G. Williams, D.W. Vehar (Eds.), Symposium
on Reactor Dosimetry, ASTMSTP1228, American
Society for Testing and Materials, Philadelphia, 1994,
pp. 310–319.
[35] R. Sanna, K. O’Brien, Nucl. Instr. and Meth. 91 (1971)
573–576.
[36] H. Jeffreys, Theory of Probability, Clarendon Press,
Oxford, 1967.
[37] M. Matzke, Report PTB-FMRB-108, Physikalisch-Tech-
nische Bundesanstalt, Braunschweig, 1986.
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241240
[38] M. Reginatto, P. Goldhagen, Report EML-595, Environ-
mental Measurements Laboratory, New York, 1998.
[39] M. Reginatto, P. Goldhagen, S. Neumann, Nucl. Instr.
and Meth. A 476 (2002) 242, these proceedings.
[40] M. Matzke, Rad. Prot. Dosim. 23 (1988) 297–300.
[41] F. d’Errico, Rad. Prot. Dosim. 84 (1999) 55–62.
[42] A.V. Alevra, Radioprotection 34 (1999) 305–333.
[43] B. Wiegel, A. V. Alevra, Nucl. Instr. and Meth. A 476
(2002) 36, these proceedings.
[44] H. Klein, L. Lindborg (Eds.), SSI-Rapport 95-15, Swedish
Rad. Protection Institute, Stockholm, 1995.
[45] H. Kluge, A.V. Alevra, S. Jetzke, K. Knauf, M. Matzke,
K. Weise, J. Wittstock, Rad. Prot. Dosim. 70 (1997) 327–
330.
[46] J. F. Briesmeister (Ed.), Report LA-12625-M, Los Alamos,
1993.
[47] M. Weyrauch, A. Casnati, P. Schillebeecks, Nucl. Instr.
and Meth. A 403 (1998) 442.
[48] G. Dietze, H. Klein, NRESP4 and NEFF4 Monte Carlo
Codes for the Calculation of Neutron Response Functions
and Detection Efficiencies for NE-213 Scintillation Detec-
tors Report PTB-ND-22, PTB Braunschweig, 1982.
[49] W.R. Nelson, H. Hirayama, D.W.O. Rogers, SLAC
Report No 256, Stanford, 1985.
[50] L. B .uermann, S. Ding, S. Guldbakke, H. Klein, T.
Novotny, M. Tichy, Nucl. Instr. and Methods A 332
(1993) 483–492.
[51] T. Novotny, S. Guldbakke, H. Klein, in: G. Vourvopoulos
(Ed.), International Conference: Neutrons in Research and
Industry, Crete, Greece, 1996, Proc. SPIE 2867 (1997),
627–630, ISBN 0-8194-2263-0.
[52] S. Neumann, R. Boettger, S. Guldbakke, M. Matzke, W.
Sosaat, Nucl. Instr. and Meth. A 476 (2002) 353, these
proceedings.
[53] M. Matzke, F. d’Errico, O. Hecker, M. Luszik-Bhadra,
Rad. Prot. Dosim. 85 (1999) 93–97.
[54] K. Knauf, M. Weyrauch, F. Simon, N. Kaldune, Report
PTB-6.42-98-1, Physikalisch-Technische Bundesanstalt,
Braunschweig, 1998.
[55] M. Luszik-Bhadra, F. d’Errico, O. Hecker, M. Matzke,
Nucl. Instr. and Meth. A 476 (2002) 457, these proceed-
ings.
M. Matzke / Nuclear Instruments and Methods in Physics Research A 476 (2002) 230–241 241