Phys Chem Minerals(1994) 21:3649
PHYSICSICHEMISTRY
Bayes Estimation: A Novel Approach to Derivation
of Internally Consistent Thermodynamic Data
for Minerals, their Uncertainties, and Correlations.
Part I: Theory
Walter Olbricht 1 *, Niranjan D. Chatterjee 2, Klaus Miller 1 **
Institute of Mathematics,Ruhr University,D-44780 Bochum,Germany
z Institute of Mineralogy,Ruhr University,D-44780 Bochum,Germany,Fax: 0234 7094 179
Received May 18, 1993/Revised, acceptedJanuary 22, 1994
Abstract. Computation of phase diagrams in mineral sys-
tems and quantitative geothermobarometry thrive on the
availability and accuracy of internally consistent thermo-
dynamic datasets for minerals. The prevailing two meth-
odologies applied to derive them, mathematical pro-
gramming (MAP) and least squares regression (REG),
have their very specific advantages and deficiencies
which are to some extent complementary. Bayes estima-
tion (BE), the novel technique proposed here for obtain-
ing internally consistent thermodynamic databases, can
combine the advantages of both MAP and REG but
avoid their drawbacks. It optimally uses the information
on thermochemical, thermophysical, and volumetric
properties of phases and experimental reaction reverals
to refine the thermodynamic data and returns their un-
certainties and correlations. Therefore, BE emerges as
the method of choice. The theoretical background of BE,
and its relation to MAP and REG, is explained. Al-
though BE is conceptually simple, it can be computa-
tionally demanding. Fortunately, modern computer
technology and new stochastic methods such as Gibbs
sampling help surmount those difficulties. The basic
ideas behind these methods are explored and recom-
mendations for their use are made using the A12SiO5
unary as an example. The potential of BE and its future
perspective for application to multicomponent-multi-
phase systems appear very promising. For the conve-
nience of readers not interested in the mathematical de-
tails of BE, an illustrative example is given in the Appen-
dix to promote an intuitive understanding of what BE
is all about.
Introduction
The availability of internally consistent thermodynamic
databases (Berman 1988; Holland and Powell 1990) has
Present addresses:
* Faculty for Mathematics and Physics, University of Bayreuth,
D-95440 Bayreuth,Germany
** Helene-Weber-Allee 18, D-80637 Munich,Germany
Correspondence to: W. Olbricht
]MIHERALS
@ Springer-Verlag1994
had an enormous impact both on the computation of
geological phase diagrams and on quantitative geother-
mobarometry (Powell 1985; Berman 1991). Regardless
of the methodologies employed to derive them, they are
invariably based on simultaneous treatment of
9 phase property (PP) data like thermochemical, ther-
mophysical, and volumetric properties of phases and
9 reaction property (RP) data based on the relevant re-
action reversal (bracketing) experiments.
The data on phase properties show normal probability
distributions; they can be expressed in terms of equali-
ties, with their associated standard deviations. The reac-
tion reversal experiments, by contrast, indicate inequali-
ties in energy difference, whose sign depends upon the
relative stability of the reactants or the products at the
T and P of the experiments. Such experiments help deli-
mit the feasible solution set (Z) for the reaction equilib-
ria. The phase properties and the reaction properties
are mutually linked through the stoichiometry of the
reaction.
Simultaneous treatment of PP and RP data has been
achieved thus far by two alternative methods. While Hol-
land and Powell (1990) used least squares regression
(REG) for that purpose, Berman (1988) applied the tech-
nique of mathematical programming (MAP). Both REG
and MAP 1 evolved over the years, the papers cited rep-
resenting a recent stage of development. The advantages
and limitations of both techniques have been acrimon-
iously debated and defended by their proponents (Ber-
man et al. 1986; Holland and Powell 1990; Engi 1992;
Powell and Holland 1993). The fundamental merit of
REG is that it provides a unique set of refined thermody-
namic data with their uncertainties and correlations,
making propagation of errors into the computed phase
diagrams straightforward. This is a very attractive fea-
ture indeed, especially when it comes down to geother-
mobarometric applications. Unfortunately, REG can
i Readers in search of simple worked examples of REG and MAP
may refer to Powell and Holland (1985) and Chatterjee (1990),
respectively

not handle sets of inequalities and equalities with rigor.
To avoid that problem, it operates with arbitrarily paired
sets of reversals, making assumptions as to the nature
of the probability distribution between them. MAP, on
the other hand, does handle sets of inequalities and equa-
lities with rigor, but returns a range of solutions, out
of which an optimal solution must be derived by mini-
mizing a suitable, if somewhat arbitrarily chosen, objec-
tive function. Two cases arise during the process:
(a) A directly measured PP turns out to be an element
of Z, the feasible solution set. In this case the value of
the objective function is zero and the optimal solution
given by MAP is same as the measured phase property.
Whenever this happens, MAP makes no further use of
the information inherent to Z.
(b) In the majority of cases, however, a measured PP
is unlikely to be an element of the feasible solution set
Z. Consequently, the optimal solution obtained by mini-
mizing the objective function is an element on the
" b o u n d a r y " of the feasible solution set (cf. Chatterjee
1990, Fig. 7.9).
Either way, it seems that te optimal solution given by
MAP is not necessarily the most desirable one. Apart
from that, MAP has not so far led to error propagation
into computed phase diagrams in any straightforward
manner, which is indispensable for meaningful geother-
mobarometry.
The dichotomy of approach indicated above need not
persist, however. The objective of this paper is to draw
attention to the technique of Bayes estimation (BE),
which is capable of combining the advantages of both
R E G and MAP. BE is not only mathematically sound,
it also returns unique solutions for the refined thermody-
namic parameters, including their uncertainties and cor-
relations. Furthermore, the solution obtained from BE
is more appropriate than that from MAP, primarily be-
cause it is not an element on the boundary of the feasible
solution set Z. The virtue of such a solution becomes
quite clear when phase equilibria calculations are done
beyond the P - T - r a n g e s of the reversal brackets, a fre-
quent necessity in geothermobarometry. In an earlier
stage of development of the Bayes method, satisfactory
solutions were obtained in chemical systems comprising
just a few phases. 2 Development of faster processors and
more recent stochastic techniques made its application
to systems with a larger number of phases feasible. In
the first part of this paper (Part I), we deal with the
theoretical background of BE. In the second part (Part
II, Chatterjee et al. 1994), a worked example involving
22 phases of the system C a O - A 1 2 0 3 - - SiO2 - - H 2 0 will
be given. Further developments on theoretical and corn-
2 At that time the technique of restricted sampling, described later
on in the section "Computation of the Estimate and Uncertainty
Regions", was tested in an attempt to obtain internally consistent
thermodynamic datasets for the system AlzSiOs. The PP and RP
data were identical to those documented by Chatterjee (1990,
Chapter 7) in his worked example of MAP. Although the feasibility
of BE was established, the time required for restricted sampling
(on a mainframe then accessible to us) was too prohibitive for
its routine application to systems with a larger number of phases
37
putational fronts are underway to make this technique
more versatile than it is today.
The Thermodynamic Model
Derivation of internally consistent thermodynamic data-
sets requires that the underlying thermodynamic model
be specified. To achieve that, we start with the condition
of chemical equilibrium for a reaction written in terms
of phase components (solid or fluid),
vi #i---At p--O. (1)
t
In this equation v~ and #~ denote, respectively, the stoi-
chiometric coefficient and the chemical potential of the
i-th component. Furthermore, the products count posi-
tive and the reactants negative in the summation process,
specifying the meaning of Ar p, the difference of chemical
potentials due to a reaction. For the present purpose,
the chemical potential of the i-th component will be de-
fined as
#i-Aa G* +Gtr, i+Gds,i+ RTlnai. (2)
The first term on the RHS of (2), Aa G*, is the apparent
molar Gibbs energy of formaton of i at a given P and
T (as defined by Berman 1988) 3. The next two terms,
Gtr, t and Gds,~, refer to the Gibbs energy differences due
to phase transition and disordering, respectively, that
i may undergo between the standard pressure po (1 bar),
standard temperature T o (298.15 K) and P, T. And final-
ly, the R T In at-term includes the activity of i, at, referred
to the pure substance i having unit activity at P and
T. The Aa G*-term of Eq. (2) is obtained from the relation,
T
A~G*=AfH~ y Cp~(T)dr
TO
-T(S~ ~T~)dT)+ ~ V~,T(P)dP, (3)
T~ po
Af H/~ being the molar, enthalpy of formation of i from
the constituent stable elements at po, T 0, S Othe standard
molar entropy, Cpi(T) the molar heat capacity power
function, and V~(T,P) the T - P - d e p e n d e n t molar volume
of i. The Cpz(T)-function will be expressed as
Cpi(r)=ai§ T+ci T-2 § r ~ T 2 + f / T -3
q- gi T- l. (4)
Such a representation allows alternative use of the
Cpi(T) polynomials utilized by various authors includ-
ing Robie et al. (1979) and Berman and Brown (1985).
3 The use of molar Gibbs energy of i at P and T, G*, is more
deeply entrenched in thermodynamic literature. To grasp the differ-
ence between G* and the apparentmolar Gibbs energy of formation
of i at P and T, Aa G*, used here, readers may refer to Chatterjee
(1990, pp. 11-13)
38
Moreover, following Berman (1988), the V~(T,P)-terms
for the condensed phases are expressed as
V/(T, P ) = V/~[1 +vx,i(T- r~ - T~ 2
+ v3,i(P- pO) + v4,i(P _ po)2], (5)
Vii~ being the standard molar volume of i. In the event
of one or more phase transitions intervening between
p0, T o and P, Tthe Gtr-terms associated with them will
be evaluated in a manner described by Berman (1988,
pp. 448-451). We shall also stick with his strategy (Ber-
man 1988, pp. 451-452) to calculate Gas. Alternatively,
a Landau formalism will be used to compute those quan-
tities. Whenever that is done, the quantity of interest
will refer to the low-temperature low-symmetry phase
at T ~ pO. Consequently, the chemical potential of a con-
densed component s will follow from the relaton
ro r
P model. In future that will be necessary at least for those
+ ~ Vr, s(P )dP+ Gt~,s+ Ges,s+Rrln as. (6)
po
Aside from condensed phases, fluids also participate
in many equilibria to be handled by the Bayes method.
Choosing the hypothetical ideal gas having unit fugacity
at pO and T as standard state, the chemical potential
of a fluid component fl can be written as
af~=AfH}~+ S CPs~(r)dr-r S}~+ CPlz(r)dr
TO TO r relevant phases. Only when those prerequisites are met,
PT
f~'~_
+ R T In f~o,r -- R T in ail. (7)
f bracketing experiments.
Having outlined the requisite identities, let us return to
the model utilized to derive an internally consistent ther-
modynamic dataset. This demands that we specify which
of the terms contributing to the chemical potential of
solids [Eq. (6)] and fluids [Eq. (7)] must be refined first
and foremost; those are the "parameters" of our model.
The remaining terms will be handled as if they were
known with sufficient accuracy and need not be refined
anymore; they are the P - T - X~-dependent "constants"
for our model. Given the notorious difficulty of obtaining
sufficiently accurate calorimetric Af H ~ data for minerals,
it is evidently the prime candidate for refinement. Indeed,
Holland and Powell (1990) settled for refining only
Af H/~ thus their model had just one parameter per
phase. Our thermodynamic model will have two parame-
ters, Af H ~ and S ~ Though calorimetric measurements
of S o are, in general, quite reliable, constraining S o never-
theless appears desirable because
(a) in numerous cases an empirical "impurity correc-
tion" is applied to calorimetric data on chemically com-
plex naturally occurring minerals to derive the S o of
the relevant phase component (end member),
(b) the configurational contributions to S o of some
end members remain unknown, or
(c) an estimated value for S o has to be used because
no direct calorimetric measurement has been executed.
The flip side of treating S o as parameter is of course
an increased dimensionality of the correlation matrix (see
Appendix of Part II for an example), making data tabula-
tion an unwieldy affair for systems with a large number
of phases.
Two more statements must be made regarding our
model:
(1) For each oxide component of a system of interest,
we shall use its reference value of Af H ~ recommended
by the C O D A T A task group; it will not be subject to
any further refinement (see also Berman, 1988, p. 454).
To the extent that an oxide does not participate in any
of the reversed reactions or in any thermochemical cycle
used to derive the Af H ~ of the silicates, the Af H ~ of
an appropriate silicate will proxy as anchor. Note, that
S o of the reference phases will still be subject to refine-
ment by BE.
(2) At the present stage of application of BE to our
problem, Cp(T) is not treated as a "parameter" of the
phases for which the superambient heat capacity meas-
urements extend to barely a few hundred Kelvin and
whose Cp(T) must be extensively extrapolated in order
to handle the RP restrictions. Cases in point are such
common minerals as anthophyllite, aragonite, brucite,
calcite, diaspore, dolomite, kaolinite, lawsonite, pyro-
phyllite, talc, zoisite etc. (cf. Berman and Brown 1985,
Table 3). However, attempts to handle Cp(T) as a "pa-
rameter" will not be meaningful unless reliable experimen-
tal data on Af H ~ S ~ and V/(T,P) are available for the
the modified Cp(T) functions are likely to be useful for
predicting phase relations beyond the P - T range of the
Mathematical Background
Introductory Remarks
This section is intended to give a concise and yet fairly
precise description of the statistical techniques used in
the BE approach. Inevitably, this requires the use of
some mathematical terms and notation, which may not
be familiar to all readers. Although these methods as
such are not intrinsically complicated, their extensive use
here may make this section somewhat difficult to digest
at first sight even for some readers with a rather good
working knowledge of inference statistics. Those in need
of a refresher are referred to Box and Tiao (1973) for
an excellent introduction. Readers not interested in delv-
ing into the subtleties of BE will find an elementary ex-
ample in the Appendix, which seeks to provide an intui-
tive understanding of the essential issues. While pursuing
this section, readers may find it worthwhile to refer to
the Appendix and vice versa.

The Bayes Method
Using Eqs. (1), (6), and (7), the condition of chemical
equilibrium may be generalized as
Ar p = 0 = A r H ~ S~ + RS(P, T, X3. (8)
Here RS(P, T, Xi) is the sum of all P, T, X~-dependent
"constants" of our thermodynamic model. By contrast,
the terns Ar H ~ and Ar S O arise from summation of the
parameters of our model, Af//o and S ~ in keeping with
the sign convention implied in Eq. (1). Our objective is
to refine the latter two quantities by BE.
As indicated in Section 1, the experimental data fall
into two categories - phase property (PP) and reaction
property (RP) measurements. They yield two different
types of information: (PP): For k phases, all information
on Af H ~ and S o is conveniently described by a (2k)-
dimensional normal distribution with known expecta-
tion vector and covariance matrix on the basic parame-
ter space
0 : {(~f H 0, S 0, hf H ~ S ~ .. ., Af Hk,
o &)}
o =IR 2k (9)
The covariance matrix is diagonal, since the calorimetric
data for Af H ~ and S O are independent of each other.
(RP): Information on the sign of Arp of the reaction
reversal experiments. By virtue of Eq. (8), each reversal
corresponds to a linear inequality dividing O into a feasi-
ble half space and its complement. The intersection of
the feasible half spaces constitutes the feasible solution
set Z. Note that inconsistent reversals lead to an empty
feasible region Z. However, if Z is not empty, it is a
(not necessarily bounded) "polygon", because it is de-
fined by linear constraints. Note moreover that the rever-
sal experiments do not usually give more specific infor-
mation about the " t r u e " parameter values within Z (cf.
Demarest and Haselton 1981). Therefore, following ad-
justment of the reversals to account for the uncertainties
of measurement of T and P, it seems safe to assume
that all values of Z are equally likely. Thus, if we calcu-
late the probability of the observational data (the signs
of Ar #) as a function of the underlying parameter vector
0, we end up with a function which equals 1 on Z and
vanishes elsewhere. In statistical terminology, the so-
called likelihood function turns out to be the indicator
function of Z. If Z is bounded, the likelihood function
can be standardized, so that it integrates to one and
can then be regarded as a probability distribution on
Z. Thus, the standardized likelihood function corre-
sponds here to a (possibly improper) uniform distribu-
tion on Z.
The mathematical problem is to combine both types
of information, PP and RP, in a meaningful and exact
way. A simple example may serve to illustrate the situa-
tion. Suppose a regular die has been thrown once and
we are asked to guess the outcome. Since the die is regu-
lar, we know that all outcomes have equal probability;
thus, our initial state of knowledge is that the probability
for each of the possible outcomes, 0 = 1,..., 6 is ~, or
more formally, p(0)= 89 for 0 = 1, ..., 6. If new informa-
tion, y, is now added, asserting that the outcome is larger
39
than 2 but smaller than 6, then we shall of course update
our initial knowledge in the light of the new data y.
Ruling out 1, 2, and 6, and renormalizing, we end up
with the following conditional probabilities (for given
y)' p ( 0 t y ) = 0 for 0 = 1 , 2 , 6 and p(0iy)= 89 for 0 = 3 , 4 , 5 .
BE formalizes precisely this procedure. Our initial
knowledge about an unknown quantity 0 is described
by a probability distribution, the so-called prior proba-
bility distribution, p(0). Moreover, we assume that for
each value of 0 we can calculate the probability that
a certain data vector y is observed in a specified experi-
ment, i.e., we have p(y]0). Then if y is the observed
data v e c t o r - Bayes' theorem provides us the so-called
posterior distribution of 0 for given y,
p (0 ]y) - p (y [0). p(0), (10)
P(Y)
in other words, the posterior distribution is proportional
to likelihood times prior distribution. Note that the de-
nominator, p(y) is a normalizing constant (for further
explanation, see Appendix) which ensures that p(01y)
is again a probability distribution. The posterior distri-
bution is an update of the prior in the light of the ob-
served data and summarizes all that is known about
0; it combines prior knowledge and new data. For fur-
ther details of the Bayes method, the reader may refer
to the statistical literature, in particular, to the compre-
hensive treatise by Box and Tiao (1973).
In the specific situation under consideration, the prior
distribution is the normal distribution from PP and the
likelihood function is the indicator function on the feasi-
ble set Z from RP. Hence, by Bayes' theorem, Eq. (10),
the posterior distribution is the normal distribution from
PP truncated to the set Z from RP. In what follows,
we denote this conditional or truncated distribution by
F and its density (strictly speaking, with respect to the
Lebesgue measure) by 7 (0). We shall first describe F by
a few parameters in a manner reminiscent of the descrip-
tion of the normal distribution from PP by its expecta-
tion vector and its covariance matrix. Due to the unimo-
dality of F and the convexity of its support, the mean
of F in our situation has much to commend itself as
the best point estimator of 0 (cf. also remark (d) below).
It has the following desirable properties:
- it lies invariably in the interior of Z,
- it is the center of gravity of F, and
- it minimizes the posterior expected loss if a quadratic
loss function is employed (Box and Tiao 1973, p. 308).
Hence, we shall use
0.E = E ( F ) = ~ 0 d F ( 0 ) = ~ 0 7 ( 0 ) dO (11)
as point estimator of 0, E denoting the expectation vec-
tor.
The definition of a measure of uncertainty is some-
what less straightforward, though, depending upon the
purpose, several approaches are possible:
(U 1) The feasible region Z. The use of Z as a measure
of uncertainty of a computed equilibrium P - T - c u r v e
is conceptually simple; for example, at a specified T, we
40
may derive the range of P compatible with Z. Then the
resulting uncertainty bounds for the equilibrium curves
are the outer envelopes of the collection of all curves
stemming from points from Z. Clearly, this approach
is overly cautious and leads to a rather broad uncertainty
band.
(U2) An HPD-region (highest posterior density region)
of content 95%. Such a region M is defined by the prop-
erties
F(M)=0.95 and V01eM V02~M:y(01)>_y(02).
Uncertainties for a computed P - T - c u r v e may then be
obtained as for U 1. Though most attractive theoretical-
ly, such regions are difficult to determine in practice in
our specific situation and will not be used in the sequel.
Note, however, that the step from Z to an HPD-region
of content 95% corresponds exactly to the calorimetric
convention of quoting 2 a-uncertainty regions.
(U 3) An adjusted covariance matrix. Variances and co-
variances are the classical measures of uncertainty, but
their direct use in this case is hampered by the fact that
F is not a normal distribution and would invalidate the
intended probability statements such as 2 a ~ 9 5 % . As
a practical expedient, we.therefore suggest that the co-
variance matrix of F, coy(F), be modified by an adjust-
ment factor (a > 1) in such a way that the respective ellip-
soidal 50% and 95% regions (formed in the same way
as for the normal distribution) would contain 50% and
95% of the probability mass of F. Formally, let ~ (b,
c) denote the ellipsoid
e(b,c):={O~| (12)
and set
%=1
al ,=inf{b IF(~(b, ~2k(0.5))) > 0.5}
a2 :=inf{b IF(~(b, ~2k(0.05))) > 0.95},
with Z2(cQ denoting the c~-fractile (or 1 0 0 ( l - e ) th per-
centile) of the z2-distribution with n degrees of freedom.
Then the adjustment factor a is defined as
a : = m a x {ao, a 1, a2} (13)
and the adjusted covariance matrix is
I:Br:= a. coy (r). (14)
This adjusted covariance matrix can then be used in
an identical fashion as a classical covariance matrix to
compute ap or at-bands (or 2ap, 2at) on phase dia-
grams by error propagation. Its use implies that all cor-
relations are taken into account. The same procedure
can also be applied separately to the two-dimensional
marginal distributions of Af H ~ and S o for each phase
i(i--1, ..., k). This version may be attractive because it
permits an easier tabulation of the results. However, it
invariably leads to an unrealistically wide uncertainty
envelope because it ignores all correlations between
phases; we do not recommend its use. An alternative
way of using the adjusted covariance matrix is to first
delimit a 95%-ellipsoid in the basic parameter space |
and then derive the respective uncertainties for the equi-
librium curves in the same way as for U 1. Since the
use of error propagation is the conventional technique,
we shall use it here and in the follow-up paper (Part II).
Note, however, that the uncertainty envelope derived
from U3 by whatever method will also be called U 3
for the sake of simplicity. Furthermore, we wish to em-
phasize that all the above quantities are well-defined and
unique for the type of problem considered by us.
Prior to discussing their computation in practice, we
shall add a few comments concerning our approach:
(a) Two points, often made against the Bayesian ap-
proach, are of no concern in our context. First, the prior
distribution is objectively given and not subjectively -
let alone arbitrarily - chosen. Even if calorimetric data
on Af H ~ and S o are not available, a suitably chosen
uniform distribution is certainly defensible as prior. Sec-
ondly, a "degree-of-belief" interpretation of probability
and confidence intervals is generally accepted in this field
(cf. Demarest and Haselton 1981, p. 218). Thus, the Baye-
sian methodology seems to be quite appropriate here
(see also Kolassa 1991, p. 3547 for a discussion of some
difficulties associated with the frequentist approach).
(b) The methodology continues to remain valid if only
PP or RP type information is available. It also applies
to the special case where Af H ~ of certain phase (like
reference oxides or silicates) are anchored. The latter case
leads to a reduction of the dimension of |
(c) If the reaction reversal experiments yield more infor-
mation about the position of the equilibrium within the
reversal brackets (see K o h n and Spear 1991, p. 129), the
Bayes method is still applicable. K6nigsberger and
Gamsfiiger (1990) and K6nigsberger (1991) have used
Bayesian estimation for a special case with a normal
prior distribution and normally distributed data. They
obtained the Bayes estimate as solution of a general
minimization problem. However, for a full-fledged Baye-
sian analysis of the general situation the concept of the
posterior distribution, as in Eq. (10), is indispensable.
(d) The mode (i.e. the location of the maximum) of the
posterior distribution F, rather than its mean - as given
by Eq. (11)- may also appear worth considering in this
context. It is obvious that the mode coincides with the
solution obtained by MAP using the qudratic objective
function proposed by Berman et al. (1986, p. 1342). In
other words, MAP is a special case of BE. The mean
is a more logical choice here because the mode, the qua-
dratic optimality point, is either identical to the calori-
metric input data or lies on the boundary of the feasible
set Z. Nevertheless, our point of view has something
to offer to the MAP methodology: the uncertainty re-
gions U2 and U3 based on the posterior distribution
can be used in this context as well and could complement
the use of U 1 suggested, but not used by Berman et al.
(1986). (Note, incidentally, that U2 is not "symmetrical"
with respect to 0~E and that 0~E is not necessarily an
element of an HPD-region of content ( l - e ) - 1 0 0 % for
larger values of c~. With e = 0.05, this seems unlikely to
create any difficulty in practice, if it occurs at all.)
(e) Finally, it may be instructive to illustrate the metho-
dological differences between BE, MAP, and REG, using
a one-dimensional case as example. Suppose that calori-

metric data yield a normal distribution with mean 4, m,
and standard deviation, a, for Af H ~ of a certain phase
and that reaction reversal experiments constrain this
value to the interval [c, d] (c < d). Then the Bayes solu-
tion is the mean of the truncated normal distribution
on [c, d], that is
(t) - qo (u)
0BE = m-[ ( I ) ( u ) - (I)(l) "if' (15)
c -m d- m
where l.'= , u:= and ~0 and q~ are the density
O" o-
and the distribution function of the standard normal
distribution, respectively. The MAP solution with the
quadratic objective function leads to
d, if d<m
OMA P = m, if m E [c, d]
c, if m<c (16)
And finally, the R E G solution is to a first order approxi-
mation
c+d
0ReG-- 2 (17)
Thus, OMApalways ignores part of the reaction reversal
information whereas 0REG ignores the calorimetric infor-
mation. The meaning of the uncertainty regions U 1, U2,
and U 3 is obvious in this context (see Appendix).
Computation of the Estimates and Uncertainty Regions
The computation of OnE involves high-dimensional inte-
gration and is, in general, a challenging problem. Except
for trivial cases, explicit expressions are not available;
even attempts to resort to conventional numerical inte-
gration formulae falter. The only feasible approach ap-
pears to be the Monte Carlo technique, which determines
OnE by a statistical experiment on the basis of (pseudo-)
random numbers. However, its general formulation
leaves some latitude for a careful design of such an exper-
iment. We consider here two very different, if typical,
ideas, that is restricted sampling and the Gibbs sampler.
Restricted sampling is an almost foolproof, but poten-
tially extremely inefficient way of estimating OnE. In this
approach, a random sample of N vectors v~ .... , VN from
the normal distribution pertaining to PP is generated,
followed by elimination of all vj(j = 1, ..., N) which vio-
late at least one of the PR restrictions. The remaining
sample {wl .... , w.} comprises the vectors vj which lie
within Z,
{W 1 . . . . , Wn} : = {V1 . . . . , VN} (3 Z , (18)
and is a random sample from the posterior distribution
F. The mean of those vectors is a Monte Carlo estimate
4 The conventional statistical symbol for mean is p, which has been
replaced here by m to avoid any possible confusion with our symbol
for chemical potential
41
of OnE, that is
N
vj'lz(vj)
0hE= 1 ~ w j - j=IN ' (19)
J =' 2 lz" (vj)
j=l
where 1z is the indicator function of the set Z defined
by
lz (v) = {~', if w Z
if vCZ (20)
If F n denotes the empirical distribution of {w~, ..., w,}
1
(i.e. the probability distribution which places mass - on
n
each of the vectors wj(j = 1..... n)), Eq. (19) can be written
in analogy to Eq. (11) as
OnE= E (F~) = ~ w d F n (w). (21)
Likewise, we can obtain an estimate for U 3 by replacing
F with its empirical counterpart F n. The computation
of U 2 and U1 does not need to be based on a random
sample from F and will be briefly touched upon below.
The restricted sampling method has a serious draw-
back though. If Z is small or far apart from the PP
mean, which is true for most cases we need to explore,
then the probability that a vector vj fulfills all RP restric-
tions is low and most vectors vj will be rejected. In other
words, N may be so much larger than n that the proce-
dure becomes useless in practice. The real advantage of
this technique is that it produces n independent realiz-
tions from the posterior distribution F, which greatly
facilitates subsequent statistical treatment like estimation
and accuracy assessment. The method, however, is very
useful for simple problems involving few dimensions and
restrictions, for pilot studies (see Section 4), and for tests
and "calibrations" of more sophisticated techniques.
The Gibbs sampler is a Markov chain Monte Carlo
(MCMC) method for Bayesian computation. Although
it was applied in statistical physics for equations of state
calculations (Metropolfis et al. 1953) many years ago, the
method is currently in rapid development and an impres-
sive flow of papers has appeared in the last two years.
For a comprehensive survey of the state of the art and
a wealth of ideas on diverse aspects of M C M C methods
like convergence diagnostics and monitoring, selection
of starting values, and implementation details, reference
may be made to the papers presented at the Royal Statis-
tical Society meeting on "The Gibbs sampler and other
Markov chain Monte Carlo methods" (e.g. Smith and
Roberts 1993; Besag and Green 1993; Gilks et al. 1993
and the references therein). Smith (1991, p. 383) also dis-
cusses the relation between analytic approximation, nu-
merical integration, and Monte Carlo simulation for the
calculation of Bayesian integrals and gives some recom-
mendations as to the respective areas of application. Gel-
fand et al. (1990) present some illustrative examples for
Gibbs sampling.
42
The basic idea behind the M C M C methodology is
to use an iterative scheme that generates a discrete time
Markov chain whose state space is Z and whose equilib-
rium distribution is F. Simulated values from this chain
can then be used to calculate the desired quantities of
F. This is an appealing idea since we only generate vec-
tors in Z. On the other hand, successive realizations from
the Markov chain are clearly correlated such that the
generted random vectors do not represent a random
sample from F. The Gibbs sampler (Geman and Geman
1984) proceeds as follows. Given an arbitrary starting
vector t o =(t o .... , t~ from Z draw
tl from F(t, It~ ..., t~
t21from F(t2 Itl, t~ t~
i~k from r(t2k Itl, 9 . . , t2k-1), (22)
F(.13 denoting the respective full conditional distribu-
tion. This yields the first Gibbs vector t 1 =(t~, ..., t~k )
and by iterating the same cycle we obtain the sequence
t ~ t 1, ... which (under mild conditions) is a realization
of a Markov chain with the above properties.
The sequence t ~ t 1, ..., derived above, may be used
in two different ways. First, we could calculate the time
average as an estimate for 0~E. However, our experiences
with that type of ergodic averaging, recommended by
Ritov (1989), have been disappointing. Presumably, that
is due to the fact that for our type of problems, Z is
usually very "long and thin". Consequently, the chain
advances in minute steps and convergence to equilibrium
is very slow (cf. also Smith and Roberts 1993, p. 5). The
possible remedy - use of the correlation structure of F
- would unfortunately destroy the main asset of Gibbs
sampling: its ease of implementation.
The second way of using the Gibbs sampling output
is to mimic a random sample from F. Such a "pseudo-
independent" sample can then be used like a "genuine"
random sample from restricted sampling to compute 0B~
and U3; Eq. (21) remains valid for this case and no new
statistical considerations are required. The pseudo-inde-
pendent random sample can be obtained either by repli-
cate independent runs and taking the final states from
each or by subsampling from a single long run of the
chain with appropriately large spacings between the out-
comes that are to form the sample. The latter procedure
can be regarded as a version of the former if the final
state from the preceding subchain is chosen as the initial
value for the next one. We have opted for a compromise
idea in which the initial value is randomly chosen from
all the states assumed by the previous subchain. This
will hopefully further reduce dependence on the starting
value and also allow occasional larger steps lest the chain
stalls in a remote region of Z for too long. We shall,
however, always repeat the whole experiment several
times, starting each run near the quadratic optimality
point to ensure an element of genuine replication. The
quadratic optimality point itself is not suitable as a start-
ing point because it usually lies at the boundary of Z;
further details will be given in a subsequent publication
on the programming aspects of this study (Miller et al.
1993, in preparation). A yet more sophisticated proce-
dure might use different vertices of Z, but the use of
the quadratic optimality point can be justified on the
ground that it is the mode of F and hence, the probability
of its neigborhood is relatively high. In summary, we
thus have three "tuning parameters" characterizing our
simulation runs:
- the number of times the experiment is repeated,
starting near the quadratic optimality point,
- the number of Gibbs points that a single experi-
ment should yield or, in other words, the number of
subchains that are linked together in the manner de-
scribed above, and
- the length of each subchain.
The next section will give some details regarding our
choice of these tuning parameters. We emphasize that
no claim is being made concerning optimality or effi-
ciency of the above suggestions for a particular situation.
Our choices were chiefly guided by pragmatism, concep-
tual simplicity, and ease of implementation.
Let us turn once more to the computation of the
uncertainty regions. Calculation of U 3 based on the em-
pirical distribution Fn of a sample (or "pseudo-sample")
from F has been described above. Although we do not
intend to pursue this question in this paper any further,
we do wish to mention that Fn could be a makeshift
solution if direct handling of U2 happens to be intract-
able. For this purpose, one could trim off the 5% most
"distant" (in a sense to be specified) vectors from the
sample and use the convex hull of the rest - following
some adjustment - as an estimate of an HPD-region
of content 95%. An obvious choice for a "distance meas-
ure" is the so-called Mahalanobis distance based on the
PP-information, which is identical to the quadratic ob-
jective function used in MAP. And finally, uncertainty
envelopes due to U 1 are usually best computed directly
from RP. For a given temperature T, the extremal values
of RS(P, T, Xi) in Eq.(8) over Z can be calculated by
linear programming. If RS (P, T, Xi) is monotone in the
variable P, the pertaining lower and upper boundaries
for P can be readily obtained.
Implementation Issues
Some Computational Details
Aside from the methodological questions addressed in
Section 3, the sheer size of the input databases (PP and
RP) to be handled by the Bayes method poses special
problems that deserve attention. Efficient and skillful use
of the available resources becomes a necessity no matter
how powerful the available computing equipments are.
For example, elimination of the redundant RP-restric-
tions may easily help speed subsequent Monte Carlo
computations by a factor of 2 to 5 and should always
be the first step of the analysis. Restriction redundancy
can be checked by solving a linear programming prob-
lem, for which efficient algorithms are available (e.g. Gill
et al. 1983). Another crucial topic is the generation of
the required random numbers. Since our programs are
 43

written in Turbo-Pascal, version 6.0, we have utilized Table 1. Results of the experiment E 1. Entries are means and stan-
its standard generator R A N D O M . Theoretical and em- dard deviations of Af H 0 (J/mol) of andalusite based on 5000 sam-
pirical tests conducted by D o h m a n n et al. (1991) have ples points for ten runs. The last line indicates the overall values
on the basis of all 50000 sample ponts
not detected any deficiencies of this generator. It is a
linear congruential generator with a period length 232, Run Mean Af H ~ (J/tool) of andalusite Standard deviation
but uses only the integers 0 .... ,231 and assumes each
of the values between 1 and 231-1 twice before the se- 1 - 2 589948.54 529.36
quence restarts. The uniformly distributed random 2 - 2 589 952.20 530.25
numbers generated by R A N D O M were mostly con- 3 -2589 946.21 533.78
verted to normally distributed ones by the Box-Muller 4 - 2589960.94 527.65
5 -2589959.08 522.64
method; only if the truncation interval for the one-di- 6 -2589951.14 536.05
mensional conditional normal distributions in Eq. (22) 7 -2589962.15 517.76
was too far from the mean, an acceptance-rejection meth- 8 -2589946.44 510.52
od was utilized. Two further points are worth mention- 9 -2589938.85 528.37
ing. First, if we carry out replicate runs of an experiment 10 - 2589949.73 523.68
(with unchanged tuning parameters), we would like to overall -2589951.53 526.06
be able to combine the results afterwards. However, if
a single experiment already consumes a substantial por-
tion of the random number sequence, there is a non-
negligible danger of " overlap ". As a precaution, we have
always linked repetitions of the same experiment and inclusion of the latter drastically increased the run times
employed the last random number from the preceding needed for a sufficiently large restricted sampling, which
one to initialize R A N D O M for the next run. Secondly, is a prerequisite for a broad range of comparison with
the whole sequence may occasionally be used up. For Gibbs samples with different subchain lengths. This arti-
instance, the ten replicates of the first experiment de- ficial limitation will, however, be lifted in a subsequent
scribed in the following subsection required a total of step. Consequently, the results obtained in the two steps
2517246835 random numbers, roughly half of the period will be slightly different. It must be clearly understood
(232), and about as many as the total number of distinct that exclusion of the restrictions due to the andalusite-
values that R A N D O M can produce. Even though this = sillimanite equilibrium has no consequence whatsoev-
would be inappropriate for the simulation of stochastic er for the methodological aspects of M C M C emphasized
phenomena, it seems still possible for Monte Carlo inte- in this subsection. Note moreover that for the sake of
gration by simple averaging, as in Eq. (19). The reason brevity, we discuss the results of the pilot study in terms
is that the lattice structure which is produced by linear of Af H ~ of andalusite only; however, the final results
congruential generators then makes the procedure given in Tables 3 and 5 list the values of all six variables
border on to a "numerical quadrature". For the experi- of the unary.
ment in question, which serves as a control experiment, The simulation for our pilot study consists of five
this was a desired effect, but in general, one should be parts, each of which comprises 10 replicates of the follow-
wary about exhausting the random number sequence. ing experiments:
Further details about the generation and testing of ran- E 1: generation of a random sample of size 5000 by re-
dom numbers and their use in Monte Carlo studies can stricted sampling,
be found in Ripley (1987) or in Rubinstein (1981) and E2: generation of a random sample of size 500 by re-
the references listed there. stricted sampling,
E3: generation of a Gibbs sample of size 500 with sub-
chain length 500,
A Pilot Study and Some Considerations Regarding E4: generation of a Gibbs sample of size 500 with sub-
the Choice of the Tuning Parameters chain length 1000, and
E5: generation of a Gibbs sample of size 500 with sub-
In this subsection we shall demonstrate the accuracy chain length 1500.
and reliability of the Monte Carlo method and give some For each part, the random number generator was initia-
guidelines for the choice of the tuning parameters and lized with a seed value from a table of random numbers;
the design of the simulation experiments. To facilitate the ten replicates within each part were sequentially
a broad range of comparison, we chose the AlzSiO 5 un- "linked" in the way described earlier to prevent overlaps.
ary - of perennial interest to petrologists - as an exam- Note, however, that each of the ten replicates in E3,
ple. The PP and RP input data used for this purpose E4 and E 5 starts near the quadratic optimality point.
are documented in Part II of this paper (Chatterjee et al. Table 1 reproduces the ten mean values and standard
1994). deviations of Af H ~ for andalusite obtained from ten rep-
The circumstances required that the simulation exper- lications of E 1. The overall mean and standard devia-
iments proceed in two steps. In a preliminary step, we tion,
processed only the restrictions for the equilibria kyani-
te=andalusite and kyanite=sillimanite, but not those Af H/~ (andalusite) = -- 2 589 951.53 J/mol,
for andalusite = sillimanite. Early on it was observed that (andalusite) = 526.06 J/mol, (23)
44

Table 2. Summary statistics for the 10 replicates of the experiments that a r a n d o m sample of 500 points leads to satisfactory
E I-E 5. The four successive entries for each experiment list mean, results and that a subchain length of 500, even more
standard deviation, minimum, and maximum for the ten determina- so 1000, ensures already that the Gibbs sample is suffi-
tions of the mean (left column) and the standard deviation (right
column) of Af H ~ (J/mol) for andalusite. Note that the results for ciently close to a r a n d o m sample. In E5, the mean,
E 1 reproduce the values already given in Table 1 in this condensed -2589940.26, may at first sight appear to be somewhat
form. The results for E2-E5 are analogous summaries for those different, but is in fact still within the range of what
experiments, whose individual values were not presented for brevity is expected. To grasp this, consider that this mean is
based on 5000 points like any other value in Table 1,
Experiment Af H ~ (J/tool) for andalusite Standard deviation and that the others vary to this extent too. Also the
slightly low standard deviation in E 5 is no reason for
E1 - 2589951.53 526.01
7.35 7.63 concern: considering the multitude of comparisons that
-2589962.16 510.52 we carry out, some minor fluctuations are bound to oc-
-2589938.85 536.05 cur. Incidentally, we have not followed the suggestion
to omit the initial part of each simulation run. While
E2 -2589945.36 530.78
16.91 15.65 such a strategy might appear promising and indeed re-
- 2589967.29 506.69 duce dependencies due to the identical starting point,
-2589919.53 550.73 it could on the other hand render detection of insufficient
E3 - 2589958.53 526.97 subchain lengths more difficult and thereby avoid one
17.23 12.75 specific s y m p t o m rather than cure the disease. In summa-
-2589980.41 506.03 ry, we conclude that for this example
- 2589923.62 545.46 - The M o n t e Carlo method works well and a sample
E4 -2589955.54 524.13 size of 500 is adequate for satisfactory results.
28.39 21.51 A subchain length of 500 or better 1000 seems
-2589998.85 492.96 to ensure that a Gibbs sample mimics a r a n d o m sample.
- 2589904.11: 561.26 We have outlined these arguments in some, but by no
E5 -- 2589940.26 533.76 means in exhaustive, details to demonstrate that even
11.01 19.42 elementary considerations suffice to make a headway to-
--2589954.96 508.62 wards a judicious choice of the tunning parameters.
--2589920.31 574.50 Note, in particular, that we are not at all helpless if
only Gibbs sampling works in a specific situation: our
conclusions would still remain valid if only E 3-E 5 were
at our disposal. The following suggestions m a y serve
are based on all 50000 sample points and should be as m i n i m u m requirements for simulation experiments of
this type:
good approximations to the "true values". E 1 was in-
cluded in this study to provide " a l m o s t true values" 1. Conduct pilot studies and replicates, if possible, using
to be used as a yardstick for later comparisons and, different methods and different values of the tuning pa-
rameters.
as indicated above, uses a brute force approach which
2. Carefully analyze the results applying cross-checks.
comes close to numerical quadrature. Table 1 shows a
remarkable stability of the results and illustrates that If unexpected phenomena are observed and they appear
to be attributable to a particular set up - such as the
for sufficiently large samples, the M o n t e Carlo method
method, r a n d o m number generator, tuning parameter
yields highly reliable and reproducible results, despite
- make sure this does not distort the final results.
its inherent randomness.
3. More specifically, for Gibbs sampling the subchain
Table 2 lists similar information for all five parts of
length ought to be sufficiently large so that no influence
the study in a condensed form. Instead of the ten values,
of this parameter on the final results is perceptible. A
we only reproduce the mean, the standard deviation as
reasonably stable estimate of the standard deviation
well as the minimum and the m a x i m u m over the ten
of F can then be used to determine the sample size
values. Note that the overall standard deviation in Eq.
(number of Gibbs points) required to achieve a desired
(23) or in Table 1 is not strictly identical to the mean
standard error using the relation
of the standard deviations of the ten runs from E l . If
everything were "ideal", Eq. (23) and the central limit standard error = o-/(sample size) ~ (24)
theorem would imply that the ten results for Af H ~ (anda-
lusite) from E2 to E5 would in each case behave like
Because of the central limit theorem, standard errors
r a n d o m samples of size 10 from a normal distribution
will be meaningful quantities for any reasonable sample
with mean value - 2 5 8 9 9 5 1 . 5 3 and standard deviation
size even though F m a y be far from normal. This stan-
526.06/5~=23.53. Likewise, the results f r o m E1 dard error describes the accuracy with which Eq.(21)
should also behave like a r a n d o m sample of size 10 from estimates 0BE and is, of course, conceptually different
a normal distribution with mean value - 2 5 8 9 9 5 1 . 5 3 from the uncertainty as measured by "uncertainty re-
and standard deviation 526.06/ 51/5~= 7.44, and the gions". If our suggestions regarding U 1 to U 3 are
reader m a y verify that they indeed do so. Evidently, E 2 adopted, there is no need to incorporate it explicitly in
through E4 pass the test with flying colors, indicating them.
 45

Table 3. Final results for the A12SiO5 unary based on 50000 points Table 4. Maximum computing times over the ten replicates of the
of the experiment E 1. The six parameters indicated are At H ~ (J/ simulation experiments E I-E 5
tool) and So (J/K-tool) for each of the phases andalusite (And),
kyanite (Ky), and sillimanite (Si). The entries refer to (a) the mean Experiment Maximum run time (minutes)
values and the adjusted 2a values and (b) the correlation matrix
(up to two decimal places). Note that At H ~ and So of the phases E1 253
have been abbreviated as And H, And S etc. E2 27
E3 14
(a) The mean values (with adjusted 2~ in parentheses: E4 28
adjustment factor = 1.0170804) E5 41

And H -2589952 (1061)

And S 91.396 (0.140)
Ky H -2594198 (1059) Table 5. Refined thermodynamic data for the AI2SiO5 polymorphs
Ky S 82.309 (0.131) (andalusite, And; kyanite, Ky; sillimanite, Si) based on 5000 Gibbs
points obtained in the experiments E6 nd E7 described in the
Si H -2585686 (1081) text. The entries refer to (a) AfH ~ (J/mol) and So (J/K.mol) for
Si S 95.757 (0.371) each phase and (b) the correlation matrix (up to two decimal places)

(b) The correlation matrix (a) The mean values (with adjusted 2or in parentheses;
adjustment factor = 1.0845265)
And H And S Ky H Ky S Si H Si S
And H -2589782 (1063)
And H 1.00 And S 91.398 (0.143)
And S 0.07 1.00 Ky H -2594077 (1057)
Ky H 0.98 -0.03 1.00 Ky S 82.338 (0.132)
Ky S - 0.02 0.00 0.07 1.00 Si H -2585991 (1062)
Si H 0.91 -0.03 0.92 -0.05 1.00 Si S 95.493 (0.255)
Si S -0.13 -0.00 -0.13 0.02 0.23 1.00

(b) The correlation matrix

We wish to emphasize that more sophisticted conver- And H And S Ky H Ky S Si H Si S

gence diagnostics and intrinsic convergence monitoring
And H 1.00
is possible and strongly encourage use of such tech-
And S 0.11 1.00
niques; for further details, the reader may consult refer-
Ky H 0.99 -0.01 1.00
ences on Gibbs sampling cited earlier. A word of caution
Ky S -0.07 -0.01 0.05 1.00
is, however, in order: by their very nature these ideas
Si H 0.97 0.01 0.97 -0.01 1.00
cannot prove that the method works, but only establish
Si S -0.08 0.06 -0.08 0.29 0.13 1.00
a lack of evidence for its failure.
Table 3 displays the results for all six parameters of
the A12SiO5 unary based on the 50000 points from E 1.
The high correlation between Af H ~ of the three phases E6: generation of a Gibbs sample of size 500 with sub-
can be regarded in hindsight as a justification for display- chain length 1000 and
ing the results for andalusite only; the corresponding E7: generation of a Gibbs sample of size 500 with sub-
tables for kyanite and sillimanite would look very simi- chain length 1500,
lar. As in the previous tables, we have retained more linking each run to the previous one to avoid overlap
digits than necessary to facilitate re-calculations. On the in the random number sequence and also overlap be-
whole, numerical accuracy is not a matter of concern tween E6 and E7. A careful perusal of the results de-
in this approach; the formulae are far too simple to tected no inadequacy of the study, nor relevant differ-
create problems. However, time may be the impediment. ences between E6 and E7, so that the samples could
Table 4 lists the maximum run times required over the be pooled. Table 5 lists the results, including the correla-
10 replicates for the simulations E l - E 5 on a PC tion matrix, based on all 5000 Gibbs points.
equipped with a 80486 processor. With 64 bit processors A cursory look at the refined thermodynamic data
in the offing, and considering that the technique lends for the A12SiO5 polymorphs, reproduced in Tables 3 and
itself to all sorts of parallelization, its prospect for appli- 5, may not reveal much difference. However, the subtle
cation to substantially larger number of parameters ap- difference between them translates to a remarkable dif-
pears to be bright. ference in the computed phase diagrams. Figures 1 a and
In the final step of our pilot study, we processed the 1 b depict, respectively, the A12SiO5 phase diagram gen-
complete set of PP and RP restrictions for the A12SiOs erated from the data in Tables 3 and 5. Note that the
unary; that is, the RP restrictions due to the andalusite phase diagrams differ both with regard to the location
=silimanite equilibrium were no l o n g e r e x c l u d e d from of the andalusite = sillimanite curve and its U 3 uncer-
the input database. Thus time, five replicates for each tainty. This is due to our ignoring the RP restrictions
of the following experiments were conducted: for that equilibrium in deriving the datasets of Table
46
AI2SiO5
I I I I I I I I
16
WithoutAnd= Si ,~
Restrictions 2;,.
-O
12
9 ,,!;;;;!;;iii
i!!!!!iiiii'#'#'
8 13-
4-
Andalusite-~"l
~';'" i i "4 \l i i
400 600 800 1000
a T [~
Fig. la-b. Computed phase diagram for the A12SiO5 unary based on the refined thermodynamic datasets given in Table 3 (a) and
Table 5 (b). The uncertainty envelope corresponds to U 3 in both cases. For further details, see text
3. Clearly, the database of Table 5 is more refined, be-
cause it was subjected to all relevant restrictions. The
A12SiOs invariant point, based on the comprehensive
dataset of Table 5, is located at 509~ and 3808 bar,
in near-agreement with that given by Holdaway and
Mukhopadhyay (1993). This is not fortuitous; it is be-
cause our PP and RP input data are very similar to
theirs. In Part II of this paper (Chatterjee et al. 1994),
an example of refining the thermodynamic data for 22
phases of the C a O - A 1 2 O s - S i O 2 - H 2 0 system will be
given. That system also comprises the three AI2SiO s
polymorphs. The refined thermodynamic data for those
three phases will be even further constrained because
we shall have to consider additional RP restrictions in-
volving kyanite and andalusite within the framework of
the C a O - A 1 2 0 3 - S i O 2 - H 2 0 system. Owing to that
we may anticipate a marginally different set of refined
values for the A12SiOs polymorphs (see Chatterjee et al.
1994, Table 4).
Conclusions and Future Perspectives
A close analysis of the two types of information, PP
and RP, to be combined for deriving internally consistent
thermodynamic data, almost compellingly suggests that
BE is the method of choice. It is conceptually simple,
mathematically sound, and makes full and adequate use
of all available information. It retains the advantages
of both MAP and REG while avoiding their respective
limitations and thus reconciles two objectives - strict
consistency and calculation of meaningful uncertainties
- long regarded as conflicting or even incompatible. In
multicomponent-multiphase systems, MCMC methods
AI2SiOs
I I [ I I I I I
16 i AllRestrictions
/
I Included , /
128 , ,ss*'S*""/Z"#/"
4! Kyanite ,'," ''J' '
n'.~//,- i i ',i'-4 i i i
400 600 800 1000
b T [~
like Gibbs sampling provide a feasible tool to calculate
the required quantities with adequate accuracy.
An increased computational efficiency can be ex-
pected in future by pursuing the following strategies:
- Use of reparametrization and/or correlation structure.
That could considerably accelerate and facilitate Gibbs
sampling. Likewise, random direction sampling schemes
or random choice of orthogonal directions may lead to
potential gains (see Smith 1984).
- Sophisticated intrinsic convergence monitoring and
convergence diagnostics could alleviate the necessity of
time consuming pilot studies or repeat experiments.
- O t h e r techniques of multiple integration may also
prove useful. The reader may find the articles in Fluor-
noy and Tsutakawa (1989) useful in this regard.
Appendix: An Illustrative Example of Bayes Estimation
In Eq. (9), we are dealing with a 2k-dimensional space,
which refers to two unknowns (Af H ~ and Sd) for each
of k phases. In general, we shall have calorimetrically
measured values for them, which we seek to refine by
BE. To make the situation intuitively understood, we
simplify matters by considering one dimension only; in
our case, this will refer to the standard molar enthalpy
of formation of phase 1, AfH ~ Consequently, this is
now our parameter 0, and the corresponding parameter
space | is the real line (real numbers) IR. Suppose, more-
over, that the calorimetrically obtained value for Af H ~
is tabulated as -3000 kJ/mol with a 2~-uncertainty of
4- 2 kJ/mol. That is, Af H ~ lies with a probability of 0.95
in the interval [-3002, -29981. In other words, our
prior knowledge about 0 (identical to Af H ~ is expressed
 47

O
1.O p(ylO) , .o

0.5 0.5

I I I I I
--3002 --3000 --2998 --3002 --3000 --2998
0=AtH? 0=thH ?

c ~(OlY) 1.o
1 .ol

, \ o. 1- 0.5

i i L ~ I' I
--3002 --3000 --2998 --3002 --3000 0B E --2998
O=AtH ~ 0 = A f H I~

f
1 .0 1.0
i:!~:
~p(01y)
i:/:iii::ii!i~.
iiiiiiiiiiiiiiii~:. Fig. A. The prior p(0) (a), the likelihood
iiii::i::ii::iii!::~.. 0.5 0.5 p(yl0) (b), the combination of prior and
likelihood (c), and the resulting posterior
iiiiiiiiiiiiiiiiiiiiiiiiiiii::~: distribution p(01Y)for the simple (one-di-
mensional) example in the Appendix. Fig-
ure (d) shows the mean of the posterior dis-
i i I t li:i:i:i:i:i:i:i:~;i:2:!:!:i:!:!:!l! I I J J 1 tribution, 0~E, superimposed on the posteri-
--3002 --3000 -299E1 --3002 - - 3 0 0 0 OB E - - 2 9 9 8 or distribution. In each case, the densities
O:At HO '
i
I
O=AtH ~ i of the respective distributions have been de-
U2i ,I I picte& Finally, (e) and (f) show the uncer-
i l I
tainty regions U 1, U 2, and U 3. For further
UIi i U3 ~ J : discussion, see text

by a normal distribution with expectation value (i.e. not give more specific information about the location
mean value) - 3 0 0 0 kJ/mol and standard deviation (a) of 0 (or Af H ~ within the feasible solution set Z. In sum-
1 kJ/mol. The probability density, p(0), of this so called mary, the likelihood function p(yl0), which gives the
prior distribution is depicted in Fig. A(a). probability for the actually observed data y as a function
Next we consider the second type of information at of the parameter 0 is equal to one on the feasible solution
our disposal, which has been derived from reaction rever- set Z = [ - 2 9 9 9 . 7 , - 2 9 9 8 ] and vanishes elsewhere as in-
sal experiments. It puts constraints on the range of Af H ~ dicated in Fig. A(b). In mathematical language, such a
values that are compatible with the reaction reversals function, which basically describes the set Z, is often
and defines in that way the feasible solution set Z. Sup- called "indicator function" of Z.
pose that this range is [ - 2 9 9 9 . 7 , - 2 9 9 8 1 ; it is based The essential ingredient of Bayes estimation is the
on two half-brackets. For any value outside of this inter- combination of the above two types of information. It
val, the observed signs of Ar It (denoted as y in Eq. (10)) is mathematically achieved via Eq. (10). We arrive at the
would have had probability zero of occurring. For any intuitively expected result: first, values outside Z are ig-
value within that interval, this probability is one; this nored, and second, the area under the curve p(0) over
is true if we assume that the reversal experiments do Z is renormalized such that it becomes one again (see
48
Fig. A(c)). The ensuing new distribution p(01y) is the
posterior distribution of 0 in the light of the new data
y (cf. Fig. A(d)). Note that the denominator in Eq. (10),
p (y), is nothing but
p (y) = ~ p (yl0) p (0) d 0, (A 1)
i.e., in our example the area under p(0) over the interval
Z.
For mathematical convenience, we denote the posteri-
or distribution by F and its desnity by y(0). Equation
(11) then gives three equivalent expressions for the mean
of the posterior distribution. This means (or expectation
value), indicated in Fig. A(d), is the Bayes estimator 0BE.
Applying Eq. (15) to our example, 0B~=
-2999.089 kJ/mol. By contrast, the MAP solution with
quadratic objective function is that value from Z which
is "closest" to the expectation value of the prior distribu-
tion ( - 3 0 0 0 kJ/mol in our example), thus OMAP is
-2999.7 kJ/mol. And finally the R E G solution, 0REG,
is to a first order approximation -2998.85 kJ/mol, the
midpoint of the interval Z = [ - 2999.7, -2998]. In other
words, 0REGsimply ignores the calorimetric information.
On the other hand, 0~tae makes no use of the right end-
point of the interval Z. If the expectation value of the
prior distribution is within the interval Z, OMAe would
even go so far as to ignore both endpoints of the interval
Z. It should be clearly understood that making partial
use of available information has nothing to do with our
specific example, it is intrinsic to the R E G and MAP
methodololgy as is more apparent from our comment
(e) in the section "The Bayes Method".
We now focus on the derivation of the uncertainty
regions in our example. U 1 is simply the feasible interval
Z = [ - 2999.7, - 2998] itself (Fig. A (e)). U 2 is obtained
by gradually "lowering" a threshold line until the area
of the part of the density that "sticks out" carries 95%
of the total probability mass (the stippled area in Fig.
A (e)); numerically, we get U 2 = [ - 2999.7, - 2998.26].
Note that U 2 is not necessarily symmetrically disposed
around 0B~. The adjusted variance from U3, however,
yields symmetrical intervals around 0B~ of the type
0B~ _+2 l/a- a (A 2).
Here a is an adjustment factor and a the standard devia-
tion of the posterior distribution; in our example,
a = 0.43. The adjustment factor a_> 1 ensures that proba-
bility statements associated with expressions like (A2)
are still meaningful and not overly optimistic even
though the posterior distribution can be far from normal.
Here it is determined in such a way that the interval
0B~_+0.67. a ~ a carries at least 50% and that the inter-
val 0Be _+ 1.96. a ~ s. a carries at least 95 % of the probabil-
ity mass of the posterior distribution F. Note that we
invariably set a > 1 (cf. Eq. (13)) and that uncertainty re-
gion of the type (A2) may extend beyond the boundaries
of the feasible solutions set Z (Fig. A(f)). In our example,
a = 1.37, and U3 (for 20) is [-3000.10, -2998.07].
Unfortunately, explicit formulae (like our Eq. (15)) for
the calculation of 0B~ are generally not available. In such
t~
t~
t~ tl tl
Fig. B. A schematic illustration of Gibbs sampling. One full cycle
leading from the initial point t o to the first pont t 1 is indicated
cases, an estimate 0BE of 0BE can be obtained by Monte
Carlo methods. For restricted sampling, we would first
generate a sequence of random numbers from our prior
distribution (Fig. A(a)), e.g.:
- 2999.50, - 2999.72, - 2999.87, - 3000.42,
- 3001.60, - 3000.53, -2998.89, - 3000.19,
- 2998.63, - 2998.08.
Following the process illustrated in Figs. A(a) to (d), we
would then omit all such numbers which are outside
of the interval [-2999.7, -2998]. The remaining
numbersconstitute a sample from the posterior distribu-
tion and their mean is an estimate of 0B~. Thus,
0B~ = ( - 2999.50 - 2998.89 - 2998.63 - 2998.08):4
= - 2998.775 kJ/mol.
Although this tiny sample based on merely 4 numbers
yields admitedly a very crude estimate compared to the
value of 0B~= -2999.089 kJ/mol indicated in Fig. A(d),
it serves well to illustrate the procedure. Note that in
this case, we have 4 '"hits'' and six "misses" out of ten
random numbers. If we so desire, we may improve our
estimate of 0BE by using a sample of 1000 random
numbers. Results, not reproduced here, show that we
then obtain 361 "hits" and 0B~ becomes -2999.093 kJ/
mol, in near-agreement with 0BE given earlier. Unfortu-
nately, the the ratio of "hits" to "misses" plummets with
increasing number of dimensions and the situation is
further exacerbated by concomitant "narrowing" of the
set Z, making restricted sampling of little use for the
majority of our applications. In such cases, we recom-
mend resorting to Gibbs sampling. The basic idea of
the latter is schematically shown in Fig. B for a two-
dimensional setup. We start with an arbitrary point t o
= (t ~ t ~ inside of Z. First, we keep t o fixed and choose
a new t o in accordance with Eq. (22); this yields (tl, t2~
Next, we keep tl fixed and alter t o such that we end
up with the next Gibbs point t 1 =(t~, t~). It is clear that
t 1 will strongly depend on the initial point t ~ but after

a sufficiently large n u m b e r , say k of r e p e t i t i o n s of t h a t
process, the last p o i n t of the chain, t k, will be " a l m o s t "
i n d e p e n d e n t of t o. F u r t h e r m o r e , the m e c h a n i s m of
Eq. (22) implies t h a t t k also lies within Z a n d t h a t it
b e h a v e s a s y m p t o t i c a l l y like a r a n d o m v e c t o r f r o m the
p o s t e r i o r d i s t r i b u t i o n . Therefore, b y i t e r a t i n g the w h o l e
p r o c e d u r e , we e v e n t u a l l y o b t a i n a p s e u d o - s a m p l e f r o m
F, f r o m which a n e s t i m a t e 0BE m a y be c o m p u t e d .
Acknowledgments. We are highly indebted to the Deutsche For-
schungsgemeinschaft, Bonn, for financially supporting this work
for 18 months under the grant Ch 46/20-1. For occasional help
with programming, we wish to thank C. Sykes and R. Kriiger.
We are grateful to H.W. Day, Davis, CA, and an anonymous re-
viewer for their perceptive reviews which helped improve the pre-
sentation. In particular, they have been instrumental in our adding
the Appendix despite an increased length of this paper, a move
kindly accepted by our executive editor K. Langer, Berlin.
References
Berman RG (1988) Internally-consistent thermodynamic data for
minerals in the system N a 2 0 - K 2 0 - C a O - M g O - F e O
- F e 2 0 3 - A1203- SiO2- T i O 2 - H2 O - CO2. J Petrol 29:445-
522
Berman RG (1991) Thermobarometry using multi-equilibrium cal-
culations: a new technique, with petrological applications. Can-
ad Mineral 29:833-855
Berman RG, Brown TH (1985) Heat capacity of minerals in the
system N a z O - K 2 0 - - C a O - M g O - FeO-- F e 2 0 3 - - A120 a
- SiO 2 - TiO 2 - H 2 0 - CO 2 : representation, estimation, and
high temperature extrapolation. Contrib Mineral Petrol
89:168-183
Berman RG, Engi M, Greenwood HJ, Brown TH (1986) Derivation
of internally consistent thermodynamic data by the technique
of mathematicl programming: a review with application to the
system M g O - S i O 2 - H 2 0 . J Petrol 27:1331-1364
Besag J, Green PJ (1993) Spatial statistics and Bayesian computa-
tion. J Roy Statist Soc B55:25-37
Box GEP, Tiao GC (1973) Bayesian Inference in Statistical Analy-
sis. Addison-Wesley, Reading, Mass
Chatterjee ND (1990) Applied Mineralogical Thermodynamics: Se-
lected Topics. Springer, Berlin Heidelberg New York
Chatterjee ND, Miller K, Olbricht W (1994) Bayes estimation: a
novel approach to derivation of internally consistent thermody-
namic data for minerals, their uncertainties, and correlations.
Part II. Application. Phys Chem Minerals 21:50-62
Demarest HH, Haselton HT (1981) Error analysis for bracketed
phase equilibrium data. Geochim Cosmochim Acta 45:217-224
Dohmann B, Falk M, Lessenich K (1991) The random number
generators of the Turbo Pascal family. Comp Statist Data Anal-
ysis (Statistical Software Newsletter) 12:129-132
Engi M (1992) Thermodynamic data for minerals: a critical assess-
ment. In: Price GD, Ross NL (eds) The Stability of Minerals.
Chapman and Hall, London, pp 267-328
Flournoy N, Tsutakwa RK (eds) (1991) Statistical Multiple Integra-
tion. Contemporary Mathematics 115. Am Math Soc, Provi-
dence, RI
Gelfand AE, Hills SE, Racine-Poon A, Smith AFM (1990) Illustra-
tion of Bayesian inference in normal data models using Gibbs
sampling. J Am Statist Assoc 85:972-985
49
Geman S, Geman D (1984) Stochastic relaxation, Gibbs distribu-
tions and the Bayesian restoration of images. IEEE Trans Pattn
Anal Mach Intell 6:721-741
Gilks WR, Clayton D J, Spiegelhalter D J, Best NG, McNeil A J,
Sharpies LD, Kirby AJ (1993) Modelling complexity: Applica-
tions of Gibbs sampling in medicine. J Roy Statist Soc B 55:39
52
Gill PE, Murray W, Saunders MA, Wright MH (1983) User's guide
for SOL/QPSOL: a FORTRAN package for quadratic pro-
gramming. Technical Report SOL 83-7, Department of Opera-
tions Research, Stanford University, Stanford, California
Holdaway M J, Mukhopadhyay B (1993) A reevaluation of the sta-
bility relations of andalusite: Thermochemical data and phase
diagram for the aluminum silicates. Am Mineral 78:298-315
Holland TJB, Powell R (1990) An enlarged and updated internally
consistent thermodynamic dataset with uncertainties and corre-
lations: the system K20 -- Na20 -- CaO - MgO -- MnO
- F e O - F e z O 3 - A 1 2 0 3 - T i O 2 S I O 2 - - C - H 2 - O 2 . J Meta-
morphic Geol 8:8%124
K6nigsberger E (1991) Improvement of excess parameters from
thermodynamic and phase diagram data by a sequential Bayes
algorithm. CALPHAD 15:69 78
K6nigsberger E, GamsjS.ger H (1990) Analysis of phase diagrams
employing Bayesian excess parameter estimation. Monatsh
Chemie 121:119 127
Kohn MJ, Spear FS (1991) Error propagation for barometers: I.
Accuracy and precision of experimentally located end-member
reactions. Am Mineral 76:128 137
Kolassa JE (1991) Confidence intervals for thermodynamic con-
stants. Geochim Cosmochim Acta 55:3543 3552
Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, Teller
E (1953) Equations of state calculations by fast computing ma-
chines. J Chem Phys 21:1087 1091
Miller K, Olbricht W, Chatterjee ND (1994) Bayes estimation of
internally consistent thermodynamic data for minerals: some
computational and programming aspects (in preparation)
Powell R (1985) Geothermometry and geobarometry: a discussion
J Geol Soc London 142:29 38
Powell R, Holland T (1993) The applicability of least squares in
the extraction of thermodynamic data from experimentally
bracketed mineral equilibria. Am Mineral 78:107 112
Powell R, Holland TJB (1985) An internally consistent thermody-
namic dataset with uncertainties and correlations: I. Methods
and a worked example. J MetamOrphic Geol 3 : 327-342
Ripley BD (1987) Stochastic Simulation. Wiley, New York
Ritov Y (1989) Monte Carlo computation of the mean of a function
with convex support. Comp Statist Data Analysis 7:269-277
Robie RA, Hemingway BS, Fisher JR (1979) Thermodynamic
properties of minerals and related substances at 298.15 K and
1 bar (10 s Pascals) pressure and at higher temperatures. US
Geol Surv Bull 1452
Rubinstein RY (1981) Simulation and the Monte Carlo Method.
Wiley, New York
Smith AFM (1991) Bayesian computational methods. Phil Trans
Roy Soc London A337:369-386
Smith AFM, Roberts GO (1993) Bayesian computation via the
Gibbs sampler and related Markov chain Monte Carlo meth-
ods. J Roy Statist Soc B55:3 23
Smith RL (1984) Efficient Monte Carlo procedures for generating
points uniformly distributed over bounded regions. Operations
Res 32:1296 1308