markov chain markov field dynamics: models and statistics
TRANSCRIPT
This article was downloaded by: [University of Otago]On: 06 October 2014, At: 04:55Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
Statistics: A Journal of Theoretical andApplied StatisticsPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/gsta20
Markov Chain Markov Field dynamics:Models and statisticsX. Guyon a & C. Hardouin aa SAMOS , Universite´ Paris 1 , FrancePublished online: 29 Oct 2010.
To cite this article: X. Guyon & C. Hardouin (2002) Markov Chain Markov Field dynamics: Modelsand statistics, Statistics: A Journal of Theoretical and Applied Statistics, 36:4, 339-363, DOI:10.1080/02331880213192
To link to this article: http://dx.doi.org/10.1080/02331880213192
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the“Content”) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoever as tothe accuracy, completeness, or suitability for any purpose of the Content. Any opinionsand views expressed in this publication are the opinions and views of the authors,and are not the views of or endorsed by Taylor & Francis. The accuracy of the Contentshould not be relied upon and should be independently verified with primary sourcesof information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever orhowsoever caused arising directly or indirectly in connection with, in relation to or arisingout of the use of the Content.
This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions
Statistics, 2002, Vol. 36(4), pp. 339–363
MARKOV CHAIN MARKOV FIELD DYNAMICS:MODELS AND STATISTICS
X. GUYON* and C. HARDOUIN
SAMOS – Universite Paris 1, France
(Received 14 March 2000; In final form 6 March 2001)
This study deals with time dynamics of Markov fields defined on a finite set of sites with state space E, focussing onMarkov Chain Markov Field (MCMF) evolution. Such a model is characterized by two families of potentials: theinstantaneous interaction potentials, and the time delay potentials. Four models are specified: auto-exponentialdynamics (E ¼ Rþ), auto-normal dynamics (E ¼ R), auto-Poissonian dynamics (E ¼ N) and auto-logisticdynamics (E qualitative and finite). Sufficient conditions ensuring ergodicity and strong law of large numbers aregiven by using a Lyapunov criterion of stability, and the conditional pseudo-likelihood statistics are summarized.We discuss the identification procedure of the two Markovian graphs and look for validation tests usingmartingale central limit theorems. An application to meteorological data illustrates such a modelling.
Keywords: Markov field; Markov Chain dynamics; Auto-model; Lyapunov stability criterion; Martingale CLTtheorem; Model diagnostic
AMS Classification Numbers: 62M40, 62M05, 62E20
1. INTRODUCTION
The purpose of this paper is to study time dynamics of Markov fields defined on a measur-
able state space E and a finite set of sites S. We present a semi causal parametric model called
MCMF, Markov Chain of Markov Field, defined as follows: X ¼ ðX ðtÞ; t 2 N�Þ is a Markov
chain on ES and X ðtÞ ¼ ðXiðtÞ; i 2 SÞ is a Markov field on ES conditionally to the past. We
study the properties of this model characterized by two families of potentials: instantaneous
interaction potentials and time-delay potentials.
Space time modelling has been considered in the literature by Preston (1974) for birth and
death processes, Durrett (1995) for particles systems, K€uunsch (1984) and Koslov and
Vasilyev (1980) for the study of reversible and synchronous dynamics, Pfeifer and
Deutsch (1980) for ARMA models in space. Among statistical studies and their applications,
some important contributions are those of Pfeifer and Deutsch (1980); Besag (1974, 1977);
* E-mail: [email protected] author. SAMOS, Universite Paris 1, 90 rue de Tolbiac 75634 Paris Cedex 13, France. E-mail:
ISSN 0233-1888 print; ISSN 1029-4910 online # 2002 Taylor & Francis LtdDOI: 10.1080/0233188021000010803
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
Keiding (1975) for birth and death process, Bennett and Haining (1985) for geographical
data, Chadoeuf et al. (1992) for plant epidemiology. The asymptotic properties of these meth-
ods are obtained in a standard way (see Amemiya, 1985; Dacunha-Castelle and Duflo, 1986;
Guyon, 1995; Bayomog et al., 1994).
The aim of this paper is to specify the structure of conditional Gibbs models according to
the kind of state space considered as well as to study ergodicity, identification, estimation and
validation of such models. For simplicity, we consider only time homogeneous dynamics (but
spatial stationarity is not assumed) and the one step Markovian property. Our results can be
generalized to a larger dependency with respect to the past as well as to time inhomogeneous
chains.
After the description of the probability transition Pðx; yÞ of the model in Section 2, we
study in Section 3 some properties about time reversibility, invariant probability measure,
and marginal distributions and we show that the MCMF dynamics is equivalent to a time
homogeneous space � time non causal Markov field representation.
Four examples are depicted in Section 4: the auto-normal dynamics ðE ¼ RÞ, the auto-ex-
ponential dynamics ðE ¼ RþÞ, the auto-Poissonian dynamics ðE ¼ NÞ, and finally, the auto-
discrete dynamics (E qualitative and finite). For each specification, we give sufficient condi-
tions ensuring the ergodicity of the chain with the help of a Lyapunov criterion. Further, we
specify the results for weak-reversible models, i.e., models such that the reverse (in time)
transition Qðx; yÞ belongs to the same auto-model family as Pðx; yÞ.
In Section 5, we summarize the main results on Conditional Pseudo-Likelihood (CPL) es-
timation: consistency and asymptotic (in time) normality of the estimator, tests of nested hy-
potheses based on the CPL-ratio.
Section 6 is devoted to the identification problem i.e., to the determination of the
dependency graphs G ¼ fG;Gg related to the MCMF dynamics. Graph G is undirected
and determines the instantaneous dependency, while graph G is directed and associated
to the time-delay dependency. In the Gaussian case, the characterization of G is given
through the partial-autocorrelations; for non-Gaussian models, we suggest a stepwise proce-
dure based on the CPL.
In Section 7, we propose two validation tests based on CLT for martingales. We conclude
this paper in Section 8 with the study of a real set of meteorological data to which we fit an
auto-logistic MCMF model. The data consist of daily pluviometric measures on a network of
16 stations in the Mekong Delta (Vietnam) during a three-month period (Tang, 1991). The
auto-logistic model allows for information to be gathered on spatial and temporal dependen-
cies, and the forecasting is relatively accurate for some sites. This study has to be considered
as a first illustrative attempt; there is no doubt that it will have to be refined in a precise
investigation and then compared to other space time models, like hidden Markov chains
or threshold models. It will, of course, be interesting to take into account the dual feature
of such data (it is not raining or the precipitation level is observed in Rþ�) and integrate it
in a Gibbs model. Such a generalization and comparisons to other models are subjected to
another study actually in progress.
2. NOTATIONS AND DESCRIPTION OF THE MCMF MODEL
Let S ¼ f1; 2; . . . ; ng be a finite set of sites. We call X an MCMF model if X is a Markov
Chain of Markov Fields (conditionally to the past). The latter are defined on a measurable
state space ðE; eÞ equipped with a positive s-finite measure v (v is usually the counting
measure if E is discrete, and the Lebesgue one if E � Rp). The product space is
340 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
(O;OÞ ¼ ðES; e�S) with the product measure vS ¼ v�S . We shall consider the case O ¼ ES
but all the following still holds for O ¼ Pi2SEi, with measures vi on measurable spaces
ðEi; eiÞ; i 2 S.
We use the following notations. Let A be a subset of S; we denote xA ¼ ðxi; i 2 AÞ and
xA ¼ ðxj; j 2 SnAÞ. Let G be a symmetric graph on S without loop: hi; ji denotes that i
and j are neighbours (in particular, i 6¼ j). The neighbourhood of A is @A ¼ fj 2 S : j=2A
s.t. 9i 2 A with hi; jig. We shall write xi ¼ xfig and @i ¼ @fig. With these notations, our
MCMF model X is the following:
� X ¼ ðX ðtÞ; t 2 N�) is a homogeneous Markov chain on O,
� X ðtÞ ¼ ðXiðtÞ; i 2 SÞ is, conditionally to X ðt 1Þ, a Markov field on S, with XiðtÞ 2 E.
We suppose that the transition probability measure of X has a density Pðx; yÞ w.r.t vS .
If E is discrete, Pðx; yÞ ¼ PðX ðtÞ ¼ yjX ðt 1Þ ¼ xÞ. We look at models for which Pðx; yÞ
is defined by an ðxÞ – a.s. admissible conditional energy U ðyjxÞ, i:e:, such that
Pðx; yÞ ¼ Z1ðxÞ exp U ðyjxÞ and
ZðxÞ ¼
ðO
exp U ðyjxÞ vsðdyÞ < 1 ð1Þ
ðZðxÞ ¼P
O exp U ðyjxÞ < 1 if E is discrete). Let X ðt 1Þ ¼ x. From the Moebius inver-
sion formula (Besag, 1974; Prum, 1986; Guyon, 1995), there exists a minimal family Cx
of non empty subsets of S, and conditional potentials F�W ð:jxÞ for each W 2 Cx such that
U ðyjxÞ ¼P
W2CxF�
W ðyjxÞ.
Throughout the paper, we suppose that Cx ¼ C does not depend on x. If we denote 0 a re-
ference state in O ð0 is 0 when E ¼ N;R or RþÞ, then, almost surely in x, we can choose (in a
unique way) potentials according to the ‘‘identifiability conditions’’: F�W ðyjxÞ ¼ 0 if for some
i 2 W ; yi ¼ 0. Besides, for each W in C, there exists a family CW of non-empty parts of S
such that the potential F�W can be written as:
F�W ðyjxÞ ¼ FW ðyÞ þ
XW 02CW
FW 0;W ðx; yÞ:
This means that the energy U is linked to two families of potentials: instantaneous interaction
potentials fFW ;W 2 Cg and ‘‘conditional’’ interaction potentials fFW 0;W ;W 2 C;W 0 2 CW g.
The semi-causal representation is associated to G ¼ fG;Gg where G and G are defined
respectively by the instantaneous and time-delay dependencies:
hi; jiG , hðt; iÞ; ðt; jÞiG , 9W 2 C s:t: fi; jg � W and FW 6¼ 0
hj; iiG, hðt 1; jÞ; ðt; iÞiG , 9W 2 C;W 0 2 CW s:t:FW 0;W 6¼ 0;
i 2 W ; j 2 W 0
Note that G is a directed graph while G is not. Let us define C¼ [W2CCW ; then
U ðyjxÞ ¼P
W2C FW ðyÞ þP
W2C;W 02C FW 0;W ðx; yÞ with the understanding that FW 0;W � 0
if W 0=2CW .
Thus we finally write:
Pðx; yÞ ¼ Z1ðxÞ expnX
W
FW ðyÞ þX
W1;W2
FW1;W2ðx; yÞ
oð2Þ
where we have put:P
W forP
W2C andP
W1;W2for
PW12C
;W22C, with FW ðyÞ ¼ 0 (resp.
FW1;W2ðx; yÞ ¼ 0) if for some i 2 W (resp. i 2 W2), yi ¼ 0.
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 341
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
There are three components in the neighbourhood of a site i:
� @i ¼ fj 2 Snfig; hi; jiGg: the t-instantaneous neighbourhood of i
� @i ¼ fj 2 S; hj; iiGg: the ðt 1Þ-antecedent neighbourhood of i
� @iþ ¼ fj 2 S; hi; jiGg: the ðt þ 1Þ-successor neighbourhood of i
The semi-causal representation is related to @i and @i, while the non causal representation
that we will present in the next section depends on @i; @i and @iþ.
Therefore, for each t � 1 and A � SðA 6¼ ;Þ, the conditional distribution of XAðtÞ given the
past and X AðtÞ ¼ yA depends on X@Aðt 1Þ ¼ x@A and X@AðtÞ ¼ y@A only, where
@A ¼ fi 2 S : 9j 2 A s:t: hi; jiGg. The corresponding conditional energy is:
UAðyAjyA; xÞ ¼
XW :W\A6¼;
FW ðyÞ þX
W2:W2\A6¼;
nXW1
FW1;W2ðx; yÞ
o
3. SOME PROPERTIES OF AN MCMF
3.1. Time Reversibility, Invariant and Marginal Distributions
In this section we only consider potentials FW1;W2such that FW1;W2
ðx; yÞ ¼ 0 if for an i 2 W1,
xi ¼ 0 or for a j 2 W2; yj ¼ 0. When there is no ambiguity, 0 denotes also the layout with 0 in
any site of S.
The transition P is synchronous if we can write Pðx; yÞ ¼Q
s2S psðx; ysÞ: the values at all
sites s are independently and synchronously relaxed with distributions psðx; :Þ in s.
PROPOSITION 1
(i) The chain is time-reversible if and only if for all W1; W2; x; y : FW1;W2ðx; yÞ ¼
FW2;W1ðy; xÞ. In this case, P has an unique invariant probability measure given by:
pðyÞ ¼ pð0ÞPð0; yÞ
Pðy; 0Þ¼ pð0ÞZ1ð0ÞZðyÞ exp
XW
FW ðyÞ:
(ii) This invariant measure p is usually not a Markov field. If the transition P is synchronous
and reversible, then p has a Markov property.
Proof
(i) This can be derived from K€uunsch (1984), and Guyon (1995, Theorem 2.2.3.).
(ii) We give in Appendix 1 two examples of non Markovian p.
If P is synchronous, only singletons occur for W in FW and for W2 in FW1;W2. As the chain
is reversible, FW1;W2� 0 if either jW1j > 1 or jW2j > 1. This means that
PðxiðtÞjxiðtÞ; xðt 1ÞÞ ¼ PðxiðtÞjx@iðt 1ÞÞ, so that,
Pðx; yÞ ¼ Z1ðxÞ expnX
s2S
FsðysÞ þXhs0;siG
Ffs0g;fsgðxs0 ; ysÞ
o¼Ys2S
psðx; ysÞ
342 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
with
psðx; ysÞ ¼ Z1ðxÞ expnFsðysÞ þ
Xs0
Ffs0g;fsgðxs0 ; ysÞg
o:
Then we have:
pðyÞ ¼ pð0Þ Z1ð0Þ ZðyÞYs2S
expFsðysÞ:
Besides, if Usðzs; yÞ ¼ FsðzsÞ þP
s02@s Ffs0g;fsgðys0;zsÞ; ZðyÞ ¼Ð
expP
s2S
U ðZs; yÞgvSðdzÞ ¼
Ps2S
Ðexp U ðzs; yÞnðdzsÞ ¼ exp
Ps2S C@sðy@sÞ with C@sðy@sÞ ¼ ln
nPzs
exp U ðzs; yÞo
. u
Example 1 A synchronous and reversible transition.
Let E ¼ f0; 1g, S is the one dimensional torus; the transition Pðx; yÞ ¼
Z1ðxÞ expfaP
i2S yiðxi1 þ xiþ1Þg is reversible with invariant law pðyÞ ¼ pð0ÞZ1ð0Þ
expP
i2S Fi1;iþ1ðyÞ,where Fi1;iþ1ðyÞ ¼ ln 1 þ exp aðyi1 þ yiþ1Þ
. The conditional distri-
bution at site l depends on yl2 and ylþ2.
Marginal Distributions For A � S, A 6¼ S, the marginal distributions ðyAjxÞ, conditionally to
x, are generally not local in x, and not explicit and not local in y, except in specific cases as
the Gaussian case. This is illustrated in Appendix 2.
3.2. Non Causal Markov Field Representation
Let X be an MCMF with the semi-causal representation (2). We are going to show that there
is a unique equivalent time homogeneous space � time non-causal Markov field representa-
tion given by the bilateral transitions: for X ðt 1Þ ¼ x and X ðt þ 1Þ ¼ z,
Pðyjx; zÞ ¼ Z1ðx; zÞ exp
� XW1;W2
FW1;W2ðx; yÞ þ FW1;W2
ðy; zÞ
þX
W
FW ðyÞ
�ð3Þ
where the normalizing constant Zðx; zÞ is finite a.s. in ðx; zÞ. The time translation invariant
potentials on S ¼ S � Z are ~FFW�f0gðy; 0Þ ¼ FW ðyÞ and ~FFW1�f0g;W2�f1gððx; 0Þ; ðy; 1ÞÞ ¼
FW1;W2ðx; yÞ. The non-causal representation depends on all the three neighbourhoods @i,
@i and @iþ.
PROPOSITION 2 The representation (2) of MCMF dynamics with the neighbourhood system
f@i; @i�; i 2 Sg is equivalent to the S Space � Time Marko-Field representation (3) with the
neighbourhood system f@i; @i�; @iþ; i 2 Sg.
Proof
(i) It is easy to see that the chain is also a two-nearest neighbours Markov field in time.
Let t be the density of X ð0Þ; the likelihood of ðxð0Þ; xð1Þ; . . . ; xðT ÞÞ is tðxð0ÞÞQt¼1;T Pðxðt 1Þ; xðtÞÞ For 1 ! t ! T 1. the conditional density is:
PðxðtÞjxðt 1Þ; xðt þ 1ÞÞ ¼Pðxðt 1Þ; xðtÞÞ PðxðtÞ; xðt þ 1ÞÞÐ
O Pðxðt 1Þ; aðtÞÞ PðaðtÞ; xðt þ 1ÞvsðdaÞÞ:
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 343
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
Let us denote x ¼ xðt 1Þ, y ¼ xðtÞ, z ¼ xðt þ 1Þ. As ðX ðt þ 1Þj X ðt 1Þ ¼ xÞ admits an
a.s. finite density, Zðx; zÞ ¼ PðX ðt þ 1Þ ¼ zjX ðt 1Þ ¼ xÞ ¼ÐO Pðx; aÞ Pða; zÞvSðdaÞ, is
finite. We obtain from (2):
Pðyjx; zÞ ¼
exp
�XW1;W2
FW1;W2
ðx; yÞ þ FW1;W2ðy; zÞ
þX
W
fW ðyÞ þ FW ðzÞ
�ðO
exp
�XW1;W2
FW1;W2
ðx; aÞ þ FW1;W2ða; zÞ
þX
W
FW ðaÞ þ FW ðzÞ
�v sðdaÞ
:
This is nothing other than (3).
(ii) Conversely, let X be the space � time Markov field on S (3) with the neighbourhood
system f@i; @i; @iþg. The field being time-homogeneous, a direct computation shows
that its semi-causal representation is (2). u
We can derive easily from (2) or (3) the semi-causal or non-causal conditional distributions
at any point (i; t). Note that we have for each i; j : i 2 @j()j 2 @iþ.
Figure 1 shows an example of both representations. For the causal representation, we have
@i ¼ fj; kg; @i ¼ fi; jg (note that i 2 @lÞ; while for the non-causal representation,
@i ¼ fj; kg; @i ¼ fi; jg; @iþ ¼ fi; j; lg and now i 2 @l and l 2 @iþ.
The semi-causal conditional distribution at (i; t) is:
Pðyijy@i; x@iÞ ¼ Z1i ðy@i; x@iÞ exp
n XW1;W23i
FW1;W2ðx; yÞ þ
XW3i
FW ðyÞo: ð4Þ
The non-causal conditional distribution Pðyijy@i; x@i ; z@iþÞ at (i; t) has conditional energy:XW1;W23i
FW1W2
ðx; yÞ þ FW1;W2ðy; zÞ
þXW3i
FW ðyÞ:
FIGURE 1 AN EXAMPLE OF SEMI-CAUSAL AND NON-CAUSAL REPRESENTATIONS OF AN MCMF.
344 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
A time inhomogeneous Markov field on S ¼ S � T is not reducible to a semi-causal re-
presentation because @i and @iþ are strongly related: the Markov field (3) is very specific.
For the example considered in Figure 1, the non-causal representation without the dotted
arrow ði; t 1Þ ! ðl; tÞ cannot be reducible to a causal representation. In all the following,
we consider the time homogeneous framework.
Reversed Dynamics We give in Appendix 3 an example where the time reversed process of
an MCMF is no more an MCMF.
4. ERGODICITY OF AUTOMODELS
Here we examine several examples of auto-models (see Besag, 1974; Guyon, 1995), and we
give conditions ensuring their ergodicity. We will use the Lyapunov Stability Criterion (see
e.g., Duflo, 1997, 6.2.2).
For n � 0, let F n ¼ sðX ðsÞ; s ! nÞ be the s-algebra generated by the X ðsÞ; s ! n. The Lya-
punov Stability Criterion is the following. Let us assume that a Markov chain defined on a
closed subset of Rd is strongly Feller, and that there exists a Lyapounov function V such that,
for n > 0, and for some 0 ! a < 1 and b < 1
E
V ðXnÞjF n1
�! aV ðXn1Þ þ b:
Then if there is at most one invariant probability measure, the chain is positive recurrent.
Besides, the following law of large numbers applies: 1=ðn þ 1ÞP
k¼0;n jðXkÞ !a:s:
mðjÞ for
any m-integrable function j, where m is the invariant measure of the chain. A difficulty is to
check whether j is m-integrable. Another useful result is available: we get the same law of
large numbers for any function j which is m-a.s. continuous and s.t. jjj ! aV þ b for
some constants a, b.
4.1. The Autoexponential Dynamics
Let us consider S ¼ f1; 2; . . . ; ng, and E ¼ Rþ. We suppose that for each i, conditionally to
ðX iðtÞ ¼ yi;X ðt 1Þ ¼ x) (later, we shall note this condition ðyi; xÞ), the distribution of XiðtÞ
is exponential with parameter liðyi; xÞ (in fact liðy@i; x@i )). From Arnold and Strauss (1988)
we can write:
U ðyjxÞ ¼Xi2S
aiðxÞyi þX
W :jW j�2
lW ðxÞyW
with yW ¼Q
i2W yi; aiðxÞ > 0 for any x and lW ðxÞ � 0. Then, the parameters are equal to
liðyi; xÞ ¼ aiðxÞ þ
PW3i lW ðxÞyWnfig. The ergodicity of the chain is obtained through the fol-
lowing assumption:
E1 (i) 8i 2 S; 8W ; x ! aiðxÞ and x ! lW ðxÞ are continuous.
(ii) 9a 2 ð0;1Þ; s.t. 8x 2 ðRþÞS; 8i 2 S : aiðxÞ � a:
PROPOSITION 3 Under assumption E1, the autoexponential dynamics is positive recurrent.
The strong law of large numbers holds for any integrable function and particularly for
functions x 7!f ðxÞ such that jf ðxÞj aVrðxÞ þ b (for some finite constants a and b) with
VrðxÞ ¼P
i2S xri ; r being any positive integer.
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 345
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
Proof The proof uses the Lyapounov stability criterion.
� The lower bound condition E1 (ii) says that exp U ðyjxÞ ! expaP
i2S yi, which is Le-
besgue integrable. Then, the chain is strongly Feller.
� As P is strictly positive, the chain is irreducible and there exists no more than one invariant
distribution (see Duflo, 1997, Proposition 6.1.9).
� On the other hand, for any positive integer r, we have for all x; y 2 E; i 2 S,
E�
XiðtÞrjx; yi
�¼
Gðr þ 1Þ
liðyi; xÞr!
Gðr þ 1Þ
ar:
It follows:
E�Vr
�X ðtÞ
���X ðt 1Þ ¼ x�!
Gðr þ 1Þ
arjSj < 1 ð5Þ
where jSj is the cardinal of S. u
Example 2 A case of ‘‘weak-reversibility’’
We assume that, conditionally to ðX iðt 1Þ ¼ xi;X ðtÞ ¼ yÞ, the reversed transition Qðy; xÞ
is also exponential with parameter miðxi; yÞ: Then, from Arnold and Strauss (1988), the joint
density for ðX ðt 1Þ ¼ x;X ðtÞ ¼ yÞ must be:
f ðx; yÞ ¼ C exp U ðx; yÞ;
with U ðx; yÞ ¼ X
W¼W1�W2�S2;W 6¼;
lW xW1yW2
with lW > 0 if jW j ¼ jW1j þ jW2j ¼ 1 and lW � 0 if jW j � 2. So E1 is satisfied.
Example 3 Besag’s conditional auto-models
We consider the case of conditional auto-models, i.e., for W � S, lW ¼ 0 if jW j > 2.
Therefore:
U ðyjxÞ ¼Xi2S
aiðxÞyi þXhi;ji
bijðxÞyiyj:
E1-(ii) is satisfied if aiðxÞ � a and bijðxÞ � 0 for all x, i, j. For example, if
U ðyjxÞ ¼Xi2S
diyi þXhi;jiG
bijyiyj þXhj;iiG
ajixjyi ð6Þ
with di > 0, bij and aij � 0, the distribution of XiðtÞ conditionally to ðyi; xÞ is exponential with
parameter liðyi; xÞ ¼ di þ
Pj2@i bijyj þ
Pl2@i alixl. The condition E1 is fulfilled.
4.2. The Autonormal Dynamics
Let E ¼ R. We assume that the conditional distribution of XiðtÞ given X iðtÞ ¼ yi and
X ðt 1Þ ¼ x is Gaussian with mean miðyi; xÞ and variance s2
i ðyi; xÞ. The principle of compat-
346 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
ibility requires (see Arnold and Press, 1989; Arnold et al., 1991) that the conditional energy
is of the following feature:
U ðyjxÞ ¼Xi2S
aiðxÞyi þ biðxÞy
2i
þ
XW s:t:jW j�2
gW ðxÞylWW
where ylWW ¼
Qi2W y
lii , li ¼ 1; 2; and the functions a; b; g ensuring that all the conditional var-
iances are positive and U ð(jxÞ is admissible.
Let us consider now Besag’s automodels. Then U ðyjxÞ ¼ P
i2SfaiðxÞyi þ biðxÞy2i g þP
hi;jiGfgijðxÞyiyj þ dijðxÞy
2i yj þ vijðxÞyiy
2j þ wijðxÞy
2i y2
j g. A typical example is:
Pðx; yÞ ¼ Z1ðxÞ exp
(Xi2S
yi
di þ
Xl2@i
alixl þ giyi
!þXhi;ji
bijyiyj
)ð7Þ
with gi > 0; bij ¼ bji such that the matrix Q ¼ ðQijÞ defined by Qii ¼ 2gi and Qij ¼ bij is
definite positive. The conditional distribution of XiðtÞ given ðyi; xÞ is Gaussian with mean
ðdi þP
l2@i alixl þP
j2@i bijyj�
2gi and variance ð2giÞ1.
Ergodicity We consider the model (7). Let us note D ¼ ðDilÞ;Dil ¼ ali; i; l 2 S, and
d ¼ ðdiÞ; a direct identification of the distribution of
ðY jxÞ ¼�X ðtÞ
��X ðt 1Þ ¼ x�
gives
ðY jxÞ ) N Sðm þ Ax;GÞ; with
m ¼ Q1d;A ¼ Q1D; and G ¼ Q1:
This can be written as an AR(1) process X ðtÞ ¼ m þ AX ðt 1Þ þ eðtÞ with a Gaussian white
noise e having covariance matrix G. Define t ¼ ðI AÞ1m when ðI AÞ is regular. Then the
zero-mean variable X �ðtÞ ¼ X ðtÞ t verifies X �ðtÞ ¼ AX �ðt 1Þ þ eðtÞ. Let rðAÞ be the
spectral radius of A (i.e., the greatest modulus of its eigen values).
PROPOSITION 4 If rðAÞ < 1, then (I � A) is regular and the chain is ergodic with a
Gaussian stationary measure m.
Proof The result is classic, given for example in Duflo (1997), Theorem 2.3.18. As e is
Gaussian, the stationary distribution for X � is the one ofP
k�0 AkeðkÞ, i:e:, N Sð0;SÞ;S being
the unique solution of G ¼ S ASðtAÞ, that is S ¼P
k�0 AkGðtAÞk . For X ðtÞ, we add the
mean t. As e has finite moments of all orders, we get the strong law of large numbers for any
continuous integrable function. u
4.3. The Auto-Poissonian Dynamics
Now E ¼ N (v is the counting measure) and we consider the dynamics associated with the
following conditional energy:
U ðyjxÞ ¼Xi2S
aiðxÞyi lnðyi!Þ
þXi 6¼j
bijðxÞyiyj
where bijðxÞ ! 0 for all i, j in order to make U admissible. Conditionally to ðyi; xÞ;XiðtÞ has a
Poisson distribution with parameter liðyi; xÞ ¼ expfaiðxÞ þ
Pj:j 6¼1 bijðxÞyjg. The ergodicity of
the chain is obtained through the following hypothesis:
P1: 9M < 1 such that 8i 2 S; supx aiðxÞ ! M
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 347
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
PROPOSITION 5 Under P1, the auto-Poissonian chain is positive recurrent. Besides, the
strong law of large numbers holds for any m-integrable function and for any functions f such
that jf ðxÞj aGuðxÞ þ b (for some finite constants a and b) where GuðxÞ ¼Q
i2S eui;xi , for
any fixed u 2 ðRþÞS.
Proof The proof follows the same lines as the one given for the auto-exponential dynamics.
P1 implies that liðyi; xÞ ! eM . Then the transition is strongly Feller. As this transition is
strictly positive, the invariant measure is unique.
For the conditional moment generating function, we have, for all s > 0, CXiðsÞ ¼
E½esXiðtÞjyi; x, ¼ eliðyi;xÞðes1Þ ! expðeM ðes 1ÞÞ. Let us set u ¼ ðui; i 2 SÞ 2 ðRþ
ÞS; Ku ¼
maxi2S expðeM ðeui 1ÞÞ and GuðxÞ ¼Q
i2S euixi . Then, conditionally to X ðt 1Þ ¼ x,
EhGu
�X ðtÞ
�jxi¼ E
" Yi2Snfjg
euiXiðtÞE�eujXjðtÞ
��y j; x���x#
! KuE
" Yi2Snfjg
euiXiðtÞ��x#:
Taking successive conditional expectations, one obtains:
E
"Yi2S
euiXiðtÞ��X ðt 1Þ
#! K jSj
u < 1:
u
For example, P1 is satisfied for
aiðxÞ ¼ di þXl2@i
alixl and bijðxÞ ¼ bij ¼ bji ð8Þ
with ali ! 0 for any i, l, and bij < 0 if hi; ji and 0 else.
4.4. The Auto-discrete Dynamics
Let E be a finite qualitative set. The conditional energy of the auto-model is:
U ðyjxÞ ¼Xi2S
aiðyi; xÞ þXhi;ji
bijðyi; yj; xÞ:
As aið(Þ and bijð(Þ are finite conditional potentials, the ergodicity is ensured without any
restrictions on the parameters. For instance, the autologistic dynamics is defined for
E ¼ f0; 1g, and U ðyjxÞ ¼P
i2S aiðxÞyi þP
hi;ji bijyiyj; XiðtÞ has a conditional Bernouilli
distribution
with parameter
pi yi; x� �
¼exp di yi; x
� �1 þ exp di yi; xð Þ
with di yi; x� �
¼ aiðxÞ þXj2@i
bijyj: ð9Þ
348 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
For E ¼ f0; 1; . . . ;mg, the autobinomial dynamics is given by U ðyjxÞ ¼Pi2S aiðxÞyi þ
Phi;ji bijyiyj. Conditionally to yi; x
� �;XiðtÞ has a Binomial distribution
B m; yi yi; x� �� �
with parameter yi yi; x� �
¼ ð1 þ expfaiðxÞ þP
j2@i bijyjgÞ1 if we take the
counting measure with densitym
yi
� �as reference measure.
5. CONDITIONAL PSEUDO-LIKELIHOOD STATISTICS
5.1. Parametric Estimation
We suppose that the transition probabilities of the MCMF depend on an unknown parameter
y, y lying in the interior of Y, a compact subset of Rd . When ergodicity holds, we can obtain
the asymptotic properties of the estimators derived from the classic estimation methods (max-
imum likelihood, pseudo maximum likelihood, maximum of another ‘‘nice’’ objective func-
tion) in a standard way. An analytical and numerical difficulty inherent to the maximum
likelihood procedure is the complexity of the normalizing constant ZyðxÞ in the likelihood;
one can use stochastic gradient algorithms to solve the problem (see Younes, 1988); another
numerical option is to compute the log likelihood and its gradient by simulations via a Monte
Carlo algorithm (see Geyer 1999, Chapter 3; and Geyer and Thompson, 1992). A third alter-
native (see Besag, 1974) is to consider the Conditional Pseudo-Likelihood (CPL); in the pre-
sence of strong spatial autocorrelation this method performs poorly, and we then have to use
the previous procedures. In the absence of strong dependency, it has good asymptotic proper-
ties, the same rate of convergence as the maximum likelihood estimator with a limited loss in
efficiency (see Besag, 1977; Guyon, 1995; Guyon and K€uunsch, 1992). The asymptotic beha-
viour follows in a standard way (see Amemiya, 1985; Dacunha-Castelle and Duflo, 1986, for
general theory – Besag, 1984; Guyon and Hardouin, 1992; Guyon, 1995, for Markov field
estimation – Bayomog, 1994; Bayomog et al., 1996, for field dynamics estimation). We
briefly recall the main results.
We assume that the chain is homogeneous and that for all i 2 S; x; y 2 E; y 2 Y, the
conditional distribution of XiðtÞ given X iðtÞ ¼ yi and X ðt 1Þ ¼ x is absolutely continuous
with respect to v, with positive conditional density fiðyi; yi; x; yÞ (which is in fact
fiðyi; y@i; x@i; yÞ).Let y0 be the true value of the parameter, and P0 be the associated transition. The process
is observed at times t ¼ 0; . . . ; T . Let us denote yyT ¼ arg miny2YUT ðyÞ the conditional pseu-
do-conditional likelihood estimator (CPLE) of y, a value minimizing the opposite of the Log-
CPL:
UT ðyÞ ¼ 1
T
XT
t¼1
Xi2S
ln fiðxiðtÞ; xiðtÞ; xðt 1Þ; yÞ:
The following conditions C and N ensure the consistency and the asymptotic normality of yyT
respectively.
Conditions for consistency (C)
C1: For y ¼ y0, the chain X is ergodic with a unique stationary measure m0.
C2: (i) For all i 2 S; x; y 2 E; y 7! fiðyi; yi; x; yÞ is continuous.
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 349
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
(ii) There exists a measurable m0 � P0-integrable function h on E � E such that for
all i 2 S; y 2 Y; x; y 2 E; j ln fiðyi; yi; x; yÞ� ln fiðyi; yi; x; y0Þj hðy; xÞ.C3: Identifiability: if y 6¼ y0 then
Pi2S m0 fx s:t: fið�; �; x; y0Þ 6¼ð fið�; �; x; yÞgÞ > 0.
Let fð1Þ
i ðyÞ and fð2Þ
i ðyÞ stand for the gradient and the Hessian matrix of fiðyi; yi; x; yÞ with
respect to y. We define the following conditional (pseudo) information matrices: for
x; y 2 E; i; j 2 S; i 6¼ j; y 2 V0, a neighbourhood of y0;V0 � Y:
Iijðyfi;jg; x; yÞ ¼ Ey0
"fð1Þ
i ðyÞf ð1Þj ðyÞ0
fiðyÞfjðyÞ
�����X fi;jgðtÞ
¼ yfi;jg;X ðt 1Þ ¼ x
#;
Iijðx; yÞ ¼ Ey0Iij X fi;jgðtÞ;X ðt 1Þ; y� �
X ðt 1Þ ¼ x��� �
:
If Zi ¼ ð@=@yÞ ln fi XiðtÞ;X iðtÞ;X ðt 1Þ; y� �
jy¼y0, then Iij yfi;jg; x; y
� �is the covariance matrix
of ðZi; ZjÞ given X fi;jg ¼ yfi;jg and X ðt 1Þ ¼ x.
Conditions for asymptotic normality (N)
N1: For some V0 � Y., a neighbourhood of y0; y 7! fi yi; yi; x; y� �
is two times continuously
differentiable on V0 and there exists a measurable, m0 � P0-square integrable function H
on E � E such that for all y 2 V0, x; y 2 E; 1 ! u; n ! d:
1
fi
@
@yu
fi yi; yi; x; y� ����� ���� and
1
fi
@2
@yu@ynfi yi; yi; x; y� ����� ���� ! Hðy; xÞ:
N2: I0 ¼P
i2S Em0IiiðX ðt 1Þ; y0Þ½ , and J0 ¼
Pi;j2S Em0
IijðX ðt 1Þ;�
y0Þ, are positive definite. >
PROPOSITION 6 (Bayomog et al., 1996; Guyon and Hardouin, 1992; Guyon, 1995)
(i) Under assumptions C; yyT
P0
T !1�����! y0
(ii) Under assumptions C and N,ffiffiffiffiT
pðyyT y0Þ
Dd
T !1�����!N dð0; I1
0 J0I10 Þ
Identifiability of the model and regularity of I0
We give here sufficient conditions ensuring both (C3) and the regularity of the matrix I0. We
suppose that each conditional density belongs to an exponential family
fi yi; yi; x; y� �
¼ KiðyiÞ exp tygi yi; yi; x� �
Ci y; yi; x� �
; ð10Þ
and set the hypotheses:
(H): there exists ðiðkÞ; xðkÞ; yðkÞ; k ¼ 1; dÞ s.t. g ¼ ðg1; g2; . . . ; gdÞ is of rank d, where
we denote gk ¼ giðkÞ yiðkÞðkÞ; yiðkÞðkÞ; xðkÞ� �
. We strengthen (H) in (H0):
ðH01Þ: for each i 2 S; giðyi; yi; xÞ ¼ hiðyiÞGiðyi; xÞ with hið�Þ 2 R
ðH02Þ: 9a > 0 s.t. for each i; x; yi;VarfhiðYiÞjyi; xÞg � a > 0
350 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
ðH03Þ: 9fðiðkÞ; xðkÞ; yiðkÞðkÞÞ; k ¼ 1; dg s.t. G ¼ ðG1;G2; . . . ;GdÞ is of rank d where
Gk ¼ GkðyiðkÞðkÞ; xðkÞÞ. Obviously, ðH0Þ implies ðHÞ.
PROPOSITION 7 We suppose that the conditional densities ffi �; yi; x; yð Þ; i 2 Sg belong to the
exponential family (10).
(i) Under ðHÞ, the model related to this family of conditional densities is identifiable.
(ii) I0 is regular under ðH0Þ.
The proof is given in Appendix 4. A sufficient condition ensuring that J0 is positive
definite can also be obtained using a strong coding set; this idea is developed in Jensen
and K€uunsch (1994). It is sketched in the same Appendix 4. Many models fulfil ðHÞ or
ðH0Þ, for instance log-linear models or automodels. We give explicit conditions ðCÞ and
ðNÞ for the autoexponential dynamics in Section 5.3.
5.2. Testing Submodels
It is now possible to test the submodel ðHqÞ : y ¼ jðaÞ; a 2 Rq; q < d;j : Rq! Rd such
that:
� j is twice continuously differentiable in a bounded open set L of Rq, with jðLÞ � Y, and
there exists a0 2 L such that jða0Þ ¼ y0.
� R ¼ ð@j=@aÞja¼a0is of rank q.
Let �yyT ¼ arg mina2LUT ðjðaÞÞ be the CPLE of y under ðHqÞ, and �II0 ¼ �II ða0Þ the associated
information matrix I . If A is a positive definite matrix, with spectral decomposition
A ¼ PDP0, P orthogonal, we take Að1=2Þ ¼ PDð1=2ÞP0 as a square root of A.
PROPOSITION 8 CPL ratio test (Bayomog, 1994; Bayomog et al., 1996; Guyon, 1995) If
UT ðyÞ and UT ðjðaÞÞ satisfy assumptions C and N, then, under ðHqÞ, we have as T ! 1:
DT ¼ 2T UT�yyT
� � UT yyT
! "! "!D Xdq
i¼1
liw21;i
where the w21;i’s are independent w2
1 variables and the li; i ¼ 1; d q, are the ðd qÞ strictly
positive eigenvalues of G0 ¼ J1=20 I1
0 R�II10 R0
� �J
1=20 .
Let C be a coding subset of S, i.e., for all i; j 2 C; i 6¼ j; i and j are not (instantaneous)
neighbour sites. We can define coding estimators as previously, but in the definition of the
coding contrast UCT ðyÞ, the summation in i is then restricted to C. For those estimators, we
have IC0 ¼ J C0 ; so the asymptotic variance forffiffiffiffiT
pðyyC
T y0Þ is ðIC0 Þ1, and the former statistic
has a w2dq asymptotic distribution.
5.3. An Example: The Autoexponential Model
Here we look at these various conditions for the model given in Section 4.1. The assumption
E1 ensures the ergodicity. For a positive integer r and VrðxÞ ¼P
i2S xri , we define the follow-
ing property for a function f :
There exist two finite constants a and b such that
f ðxÞ ! aVrðxÞ þ b: ð11Þ
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 351
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
We add the following hypotheses to obtain the consistency and the asymptotic normality of
the CPL estimator.
E2: For all i 2 S;W 2 C; y 2 Y and x 2 E; aiðx; yÞ and lW ðx; yÞ satisfy (11).
E3: If y 6¼ y0, then there exists A � ðRþÞS; lðAÞ > 0, such that for one i 2 S and all
x 2 A; lðfy 2 Ejliðyi; x; yÞ 6¼ liðy
i; x; y0ÞgÞ > 0.
E4: The functions y ! aiðx; yÞ; y ! lW ðx; yÞ are twice continuously differentiable for all
x; i; j;W , and for all 1 ! u; n ! d, and the absolute values of their first and second order
derivatives satisfy (11).
E5: I0 and J0 are positive definite.
E6: ðH0Þ is satisfied and J0 is positive definite.
PROPOSITION 9 Let yyT be the CPLE of y for the autoexponential model. Then:
(i) under assumptions E1, E2, and E3, byyT is consistent. If we add E4 and E5, yyT is
asymptotically normal.
(ii) under assumptions E1, E2, E4 and E6, yyT is asymptotically normal.
Proof E1 implies C1 and E4 implies C2-(i). E3 implies C3. Then we just have to show that
E2 implies C2-(ii). The conditional density of XiðtÞ is given here by
ln fiðyi; yi; x; yÞ ¼ ln liðyi; x; yÞ � yiliðyi; x; yÞ. Then j ln fiðyi; yi; x; yÞ � ln fiðyi; yi; x; y0Þj j lnðliðyi; x; yÞ=liðyi; x; y0ÞÞjþ yijliðyi; x; yÞ� liðyi; x; y0Þj. As j lnðx=yÞj jx � yjðx�1 þ y�1Þfor all x; y > 0, we obtain j ln fiðyi; yi; x; yÞ � ln fiðyi; yi; x; y0Þj ðð2=aÞ þ yiÞ jliðyi; x; yÞ �liðyi; x; y0Þj. Finally, there exist two integers r and r0 and four constants a1; a2; a3; a4 such
that we can take hðy; xÞ ¼P
i2S 2ðð2=aÞ þ yiÞ fa1VrðxÞ þ a2 þ ða3Vr0 ðxÞ þ a4ÞP
j 6¼i yjg.
On the other hand, E4 implies that the modulus of the first and second derivatives of
liðyi; x; yÞ with respect to y are bounded by a square integrable function of x and y,
and this ensures N1. Finally, E5 is N2. In another hand, E6 ensures E3 and E5 (see
Proposition 7). u
Example 4 The conditional-exponential dynamics.
For the particular model (6), all conditions E are fulfilled without any assumption on the
parameters. Besides, E6 is satisfied with d ¼ n þ ð1=2ÞP
i2S j@ij þP
i2S j@ij; liðyi; x; yÞ ¼t
ygiðyi; xÞ;t y ¼t ððdiÞi2S; ðbijÞhi;jiG ; ðajiÞhj;iiG Þ 2 Rd .
6. MODEL IDENTIFICATION
We suppose that we want to fit a semi-causal MCMF dynamics model; first, we have to
determine the graphs G and G; then, we will estimate the parameters, and lastly validate
the model. There are two possible strategies for the identification procedure, the choice
depending on the complexity of the problem and on the number of sites. The first one is
global: we could try to determine globally the graphs by CPL maximization, joined with a
convenient ‘‘AIC’’ penalization criterion.
On the other hand, a less expensive procedure is to work locally site by site, providing
estimations for @i and @i for each site i 2 S. For this, we maximize the likelihood of
XiðtÞ; t ¼ 1; T , conditionally to X iðtÞ ¼ yi (jSj 1 sites) and X ðt 1Þ ¼ x (jSj sites) (the
conditional distribution of XiðtÞ depends on (2jSj 1) sites) using a backward procedure;
352 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
we choose a small signification level for the adequate statistics in order to keep only the
most significant variables. This can be associated with a forward procedure if we want to
take into account a particular information on the geometry of S. Further, we have to harmo-
nize the instantaneous neighbourhood relation to get a symmetric graph G: if j 2 @@i, we
decide i 2 b@@j. Thus we generally get an overfitted model, and the next step is to reduce it
by progressive elimination using a descending stepwise procedure. Finally, if we have got
two or more models, we choose the one which minimizes the AIC (or BIC) criterion.
In a Gaussian context, the computation is particularly easy and fast, because it is linear and
explicit. Using the partial correlation rP, we have the characterizations (see Garber, 1981;
Guyon, 1995 x1.4.): rPðXj;XijXLiÞ ¼ 0 , j=2Li ¼ @i [ @i. This equation allows us to deter-
mine Li by a fast linear stepwise procedure of the regression of XiðtÞ on the ð2jSj 1Þ other
variables.
For other log linear models, the conditional log likelihood is concave so we can get the
CPLE using a gradient algorithm. We then consider the general procedure given previously.
An alternative is to follow the Gaussian approach, even if the model is not Gaussian. This
can be done when the dimension of the parameter becomes large enough to make the
results suspicious and slow progressing. This procedure is used in Section 8.
7. MODEL VALIDATION
7.1. Some Central Limit Theorems
In the case E � R, validation tests are based on the estimated conditional residuals. Let us
denote y0 and yyT the true value of the parameter and its CPLE respectively. Let us set:
eit ¼ XiðtÞ mit and eeit ¼ XiðtÞ mmit ð12Þ
where mitðyÞ ¼ E½XiðtÞjXiðtÞ;X ðt 1Þ; y,; mit ¼ mitðy0Þ; mmit ¼ mitðyyT Þ is explicit in y@i; x@i
and yyT .
For an autodiscrete dynamics on a K-states space E, we will have an expression equivalent
to (12), but with the ðK 1Þ-dimensional encoded variable ZiðtÞ related to XiðtÞ (see Section
7.2.4).
From (12), we propose a validation statistic; we derive its limit distribution using a Central
Limit Theorem for martingales (Duflo, 1997; Hall and Heyde, 1980).
We build up two tests, based on the estimated residuals eeit and on the squared residuals ee2it.
The latter allows for the detection of possible variance deviation. Anyway, those tests are
more useful to reject a model than to select the best one.
7.1.1. CLT for the Residuals eit
We denote A ? B if the variables A and B are independent. Let C be a coding subset of S (for
G), eit ¼ ðeit=sðeitÞ and ~eet ¼P
i2C eit :
(i) eit is zero mean and of variance 1;
(ii) eit ? XjðtÞ if i 6¼ j and eit ? Xjðt 1Þ for all j 2 S.
Besides, for all i and j which are not neighbour sites, XiðtÞ and XjðtÞ are independent,
conditionally to the past and the neighbourhood values on @i; @j. Then eit ? ejs if t 6¼ s,
for all i; j 2 S, and conditionally to X ðt 1Þ and X �CCðtÞ; eit ? ejt for all i; j 2 C; i 6¼ j;
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 353
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
(iii) as E eitjF t1j ¼ E E eitjyi; x
� �F t1
� ��¼ 0; ðeitÞt�0 is a square integrable martingale dif-
ference sequence w.r.t. the filtration ðF t ¼ sðX ðsÞ; s ! tÞ; t � 0Þ. Then ~eet ¼P
i2C eit is
also a square integrable martingale difference sequence.
We can thus apply a martingale’s CLT (Duflo, 1996; Corollary 2.1.10). For a square integ-
rable martingale ðMtÞ, let ðhMitÞ be its increasing process defined by:
hMit ¼ hMit1 þ E jjMt Mt1jj2jF t1
� �for t � 1; and hMi0 ¼ 0:
The first condition to check is:
hMiT
T
P
T !1�����! G; for some positive G: ð13Þ
We set Ms ¼Ps
t¼1 ~eet; we then have:
hMit hMit1 ¼Xi2C
E e2itjF t1
� �¼Xi2C
E E e2itjy
i; x� �
jF t1
� �¼ jCj:
The second condition is the Lindeberg condition. It is implied by the following one:
9 a > 2 such that E j~eetjajF t1½ , is bounded: ð14Þ
Therefore, (13) is fulfilled with G ¼ jCj, whatever the model, while (14) has to be checked in
each particular case.
PROPOSITION 10 (Duflo, 1997; Hall and Heyde, 1980) Let eit be defined by (12), C a coding
subset such that ~eet fulfill (14). Then,
1ffiffiffiffiffiffiffiffiffijCjT
pXT
t¼1
~eetD
T !1�����!Nð0; 1Þ
7.1.2. CLT for the Squared Residuals
Let us define
wit ¼e2
it 1
sðe2itÞ
: ð15Þ
The wit’s have the same properties of independency as the eit’s. Hence (wit) is a martingale
difference sequence in t, and so is ~wwt ¼P
i2C wit, for C a coding subset of S. Let
NT ¼P
t¼1;T ~wwt. As hNit hNit1 ¼P
i2C E w2itjF t1
� �¼ jCj, the first condition (13) for
the CLT is fulfilled. The second condition will be ensured under:
9a > 2 such that E j ~wwtjajF t1½ , is bounded : ð16Þ
PROPOSITION 11 Let C be a coding subset of S, wit be defined by (15) s.t. ~wwt fulfills (16).
Then,
1ffiffiffiffiffiffiffiffiffijCjT
pXi2C
XT
t¼1
witD
T !1�����!Nð0; 1Þ:
As we do not know y0, we apply the previous results to the residuals calculated with yyT .
When CPLE is convergent, standard manipulations show that Propositions 10 and 11 still
remain valid for the estimated residuals.
354 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
7.2. Applications
We give the explicit results for the auto-models studied in Section 4. The proofs of the con-
ditions (14) or (16) are given in Appendix 5.
7.2.1. The Autoexponential Dynamics
We suppose that we are in the framework of (6), assuming E1;E2;E3. Then yyT is convergent
and we have:
1ffiffiffiffiffiffiffiffiffijCjT
pXi2C
XT
t¼1
beeit
D
T !1�����!Nð0; 1Þ and
1
2ffiffiffiffiffiffiffiffiffiffiffi2jCjT
pXi2C
XT
t¼1
bee2it 1
� � D
T !1�����!Nð0; 1Þ:
7.2.2. The Autopoissonian Dynamics
We consider the framework of (8) and we suppose that the CPLE yyT is consistent. Then for
li ¼ li X iðtÞ;X ðt 1Þ� �
; g2 ¼P
i2C E½li, and s2 ¼P
i2C E½liðli þ 1Þðl2i þ 5li þ 1Þ,, we
have:
1ffiffiffiffiT
pXi2C
XT
t¼1
eeit
D
T !1�����!Nð0; g2Þ and
1ffiffiffiffiT
pXi2C
XT
t¼1
ee2it liðy
i; x� � D
T !1�����!Nð0; s2Þ:
7.2.3. The Autologistic Dynamics
For E ¼ f0; 1g, we consider the framework of (9) and we assume:
B1: For y 6¼ y0; 9 x; y 2 E s.t., for an i 2 S; piðyi; x; yÞ 6¼ piðy
i; x; y0Þ.
Under assumption B1, we have:
1ffiffiffiffiffiffiffiffiffijCjT
pXi2C
XT
t¼1
eeit
D
T !1����!Nð0;1Þ and
1ffiffiffiffiffiffiffiffiffijCjT
pXi2C
XT
t¼1
bwwit
D
T !1����!Nð0;1Þ:
7.2.4. The Autodiscrete Dynamics
More generally, if E is qualitative, E ¼ fa0; a1; . . . ; aK1g, we consider the encoded ðK 1Þ-
dimensional variable Zit 2 f0; 1gK1 linked to the XiðtÞ by:
Zitl ¼ 1 if XiðtÞ ¼ al; Zitl ¼ 0 elsewise; 1 ! l ! K 1:
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 355
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
We suppose that the conditional energy is given by U ðztjzt1Þ ¼P
i2S
Pl¼
1;K 1zitl di;l þP
k¼1;K1 ai;lk ziðt1Þk
� �þP
hi;ji
Pl¼1;K1
Pk¼1;K1 bi;j;klzitlzjtk . We have:
1ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðK 1ÞjCjT
p Xi2C
XT
t¼1
!teitC
1it eit ðK 1Þ
"D
T !1�����!Nð0; 1Þ
where Cit ¼ ðCitÞkl; 1 ! k; l ! K 1, with ðCitÞkk ¼ pikðzit; zt1Þ
ð1 pikðzit; zt1ÞÞ; ðCitÞkl ¼ ðCitÞlk ¼ pik zi
t; zt1
� �pil zi
t; zt1
� �if l 6¼ k, and
pil zit; zt1
� �¼
exp yitl
1 þP
l¼1;K1 exp zitlyitl
ð17Þ
where yitl ¼ di;l þP
k¼1;K1 ai;lk ziðt1Þk þP
j2@i
Pk¼1;K1 bi;j;klzjtk :
7.2.5. The Autonormal Dynamics
In the framework of (7), we suppose that the ergodicity condition is fulfilled. We have:
1ffiffiffiffiffiffiffiffiffijCjT
pXi2C
XT
t¼1
eitD
T !1�����!Nð0; 1Þ and
Xi2C
XT
t¼1
2gie2it
D
T !1�����! w2
jCjT
and so,
1
2ffiffiffiffiffiffiffiffiffijCjT
pXi2C
XT
t¼1
ð2gie2it 1Þ
D
T !1�����!Nð0; 1Þ:
A third result holds, which uses the joint distribution on S of all the residuals. The eit’s covar-
iances are Covðeit; ejtÞ ¼ ð1=2giÞ if i ¼ j ¼ ðbij=2gigjÞ if j 2 @i, and 0 else. Let Ce be the
n � n covariance matrix of et ¼ ðeit; i 2 SÞ. ThenP
t¼1;TtetC
1e et ) w2
nT and thus
1ffiffiffiffiffiffinT
pXT
t¼1
tetC1e et
D
T !1�����!Nð0; 1Þ:
Let yyT be a consistent estimator of y0 and CCe is the estimate of Ce obtained by replacing the
parameters by their estimates. Then, the three statistics T1; T2; T3 defined below are asymp-
totically Gaussian Nð0; 1Þ and can be used for validation tests of the model:
T1 ¼1ffiffiffiffiffiffiffiffiffijCjT
pXi2C
XT
t¼1
eeit; T2 ¼1
2ffiffiffiffiffiffiffiffiffijCjT
pXi2C
XT
t¼1
ð2ggiee2it 1Þ;
T3 ¼1ffiffiffiffiffiffinT
pXT
t¼1
t eetCC1e eet:
8. MCMF MODELLING OF METEOROLOGICAL DATA
We have tried an MCMF modelling on a real set of meteorological data. The data comes from
the study of Tang (1991) and consists of daily pluviometric measures on a network of 16 sites
356 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
in the Mekong Delta (Vietnam). We have retained a period of 123 consecutive days from July
to October, 1983. Geographically, the 16 meteorological stations, situated at latitude 9.17 to
10.8 and longitude 104.48 to 106.67, are not regularly located (see Tab. I).
The data have been previously studied by Tang (1991); he proposed different models for
each site, each being free of the other sites; this seems to be insufficient and hardly satisfac-
tory.
We study an MCMF autologistic model; we consider the binary f0; 1g data where 1 stands
for rain and 0 for no rain. The results lead to an average of exact prediction of about 77%,
which is rather satisfactory noting that the random part is large in this kind of meteorological
phenomenon. We compare this model with a site-by-site Markov Chain, for which the pre-
diction results are bad.
Of course, we do not pretend that our models are definitive and we know they need to be
refined for effective forecasting. An interesting study would be to fit other competitive
models of more or less the same dimension, as an Hidden Markov Chain (see e:g., Zucchini
and Guttorp, 1991; MacDonald and Zucchini, 1997), the hidden states standing for the under-
lying climatic stage, or threshold models, the dynamics at the time t depending on the posi-
tion of the state at the time t 1 with respect to thresholds (see Tong, 1990 and references
herein). Also, these models should take into account the dual character of such a data, i:e:,the binary feature f0g [ Rþ� of the space state. This is work in progress.
The first task is to identify the two dependency graphs. Considering the large number of para-
meters involved in the CPL procedure, we use the Gaussian linear procedure with the original
data (see Section 6); we fit a regression model on each site, with respect to the 15 other variables
at the same time and all the 16 variables at the previous time. We then select the neighbours as the
variables giving the best multiple correlation coefficient R2 in a stepwise procedure, taking into
account the symmetry of the instantaneous graph together with a principle of parsimony. We
give in Table II the neighbourhoods and the R2 obtained from all the 31 variables (denoted
R231) and the R2 calculated on the selected neighbours (denoted R2
@i;@i ). The instantaneous
graph has 25 links and the time-delay graph 17 links. We see that for each site, the instantaneous
neighbourhood may contain few sites, while there are at most 3 sites making up the time delay
neighbourhood. We note that if we draw the directed graph G on a geographical map, there is a
main direction of the arrows, which is roughly S-N.
TABLE I 16 Meteorological stations in the Mekong Delta
Site number Name Latitude Longitude
1 Tan Chau 10.80 105.252 My Tho 10.35 106.373 Chau Doe 10.70 105.014 Can Tho 10.03 105.785 Soc Tran 09.60 105.976 Vinh Long 10.25 105.977 Sa Dec 10.30 105.758 Go Cong 10.35 106.679 Ca Mau 09.17 105.1710 Long Xuyen 10.40 105.4211 Rach Gia 10.00 105.0812 Ha Tien 10.38 104.4813 Cao Lanh 10.47 105.6314 Moc Hoa 10.75 105.9215 Vi Thanh 09.77 105.4516 Tan An 10.53 106.40
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 357
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
Going back to the binary data, we estimate the autologistic model (9): for each site i, XiðtÞ
has a Bernouilli distribution of parameter piðtÞ, conditionally to X@iðtÞ and X@iðtÞ, with
piðtÞ ¼ ð1 þ expliðtÞÞ1 and liðtÞ ¼ di þ
Pj2@i bijXjðtÞ þ
P‘2@i aliXlðt 1Þ. The dimen-
sion of the model is 16 þ 25 þ 17 ¼ 58. We first estimate the parameters site-by-site, max-
imizing the log conditional pseudo-likelihoods; then, we calculate the validation statistics
based on the square residuals (see Section 7.2): Vi ¼ ð1=ffiffiffiffiT
pÞPT
t¼1 wwit for each site i. Sec-
ondly, we proceed to the global estimation on the basis of the global pseudo-likelihood;
we take as initial values the first site-by-site estimates for the parameters di and ali, and
the means of the two previous estimations bij and bji (theoretically equal) for the instanta-
neous interaction parameters. We do not give here the results for the 58 parameters for
sake of place, but we summarize the results: concerning the statistics Vi, each local model
is accepted. The final (global) estimation provides results close to the individual ones. The
TABLE II Neighbourhoods and R2 statistics. R231: regression on all vari-
ables. R2@i;@� : regression on the neighbourhoods variables
Site i @i @i� R231 R2
@i;@�
1 f2; 3; 8; 14g 0.611 0.4962 f1; 5; 6; 7; 8; 15; 16g f14g 0.648 0.5743 f1; 12g f1g 0.741 0.6384 f6g f7; 10g 0.319 0.1695 f2; 9g f4g 0.297 0.1396 f2; 4; 7; 10; 14g f7g 0.589 0.4847 f2; 6; 13; 15; 16g f6g 0.633 0.5318 f1; 2; 15g 0.592 0.3709 f5; 11g f4; 9; 12g 0.589 0.48110 f6; 12; 13g f4; 5g 0.463 0.34811 f9; 15g f11g 0.596 0.45112 f3; 10g f11g 0.648 0.52813 f7; 10g f4g 0.463 0.32114 f1; 6; 16g f4g 0.441 0.26415 f2; 7; 8; 11g f10g 0.568 0.47616 f2; 7; 14g 0.527 0.279
TABLE III Percentages of similarity between the realdata and the predicted values for the autologistic modeland the site by site Markov chain
SiteAutologistic model from
global estimates Markov chain
1 79.59 45.902 84.43 35.253 72.95 49.184 84.43 52.465 82.79 82.796 81.97 81.977 82.79 59.028 71.31 56.569 77.05 57.3810 68.03 54.9211 78.69 51.6412 72.13 57.3813 81.15 50.0014 78.69 51.6415 72.13 50.0016 74.59 57.38
358 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
validation statistic V ¼ ð1=ffiffiffiffiffiffiCT
pÞP
i2C
PTt¼1 wwit computed for the coding set
C ¼ f3; 4; 5; 7; 8; 10; 11; 14g (of maximal size 8) is equal to 0:1186 and we do not reject
the model.
Another way to validate the model is to compare the true data with what is predicted. So
we compute the predicted values (maximizing the local conditional probabilities), first site by
site (with the local estimates) and then globally (with the global estimates). Table III gives the
percentages of similarity which spread out from 68.03% (site 10) to 84.43% (site 2), with a
mean of 77.36% (for the global parameter’s estimate).
We compare our model with a site by site Markov chain. Of course, this alternative is very
poor; as expected, the predictions are bad, spreading out between 35.25% and 59.02% with a
mean equal to 51.84% (see Tab. III). In conclusion, the autologistic model leads to a rela-
tively correct forecasting; it could be better if we increase the number of links or the depen-
dence in time, at the expense of the rising parametric dimension.
References
Amemiya, T. (1985). Advanced Econometrics (Blackwell).Arnold, B. C. and Press, J. (1989). ‘‘Compatible conditional distributions’’, JASA 84(405), 152–156.Arnold, B. C. and Strauss, D. (1988). ‘‘Bivariate distribution with exponential conditionals’’, JASA 83(402), 522–
527.Arnold, B.C., Castillo, E. and Sarabia, J. M. (1991). ‘‘Conditionally specified distributions’’, L.N.S. 74 (Springer).Bayomong, S. (1994). Modelisation et analyse des donnees spatio-temporelles, PhD thesis, Univ. d’Orsay, France.Bayomong, S., Guyon, X., Hardouin, C. and Yao, J. F. (1996). ‘‘Test de difference de constraste et somme ponderee
de khi-deux’’, Can. J. Stat. 24(1), 115–130.Bennett, R. J. and Haining, H. (1985). ‘‘Spatial structure and spatial interaction: Modelling approaches to the
statistical analysis of geographical data’’, JRSS 148, 1–36.Besag, J. (1974). ‘‘Spatial interaction an the statistical analysis of lattice systems’’, JRSS B, 36, 192–236.Besag, J. (1974). ‘‘On spatial temporal models and Markov fields’’, 7th Prague Conference and of European Meeting
of Statisticians, pp. 47–55.Besag, J. (1977). ‘‘Efficiency of pseudo likelihood estimation for simple Gaussian fields’’, Biometrika 64, 616–618.Chadoeuf, J., Nandris, D., Geiger, J.P., Nicole, M. and Piarrat, J. C. (1992). ‘‘Modelisation spatio temporelle d’une
epidemie par un processus de Gibbs: Estimation et tests’’, Biometrics 48, 1165–1175.Dacunha-Castelle, D. and Duflo, M. (1986). Probability and Statistics, Vol. 2 (Springer, Verlag).Duflo, M. (1977). Random Iterative Models (Springer).Durrett, R. (1995). Ten Lectures on Particles Systems, Saint-Flour 1993. L. N. M. 1608 (Springer), pp. 97–201.Garber, D. D. (1981). Computational models for texture analysis and texture synthesis, PhD thesis, Univ South
California.Geyer, C. (1999). ‘‘Likelihood inference for spatial point processes’’, In: Barndorff-Nielsen, Kendall and Van
Lieshout (Eds.), Stochastic Geometry, Likelihood and Computation, Chapter 3 (Chapman and Hall).Geyer, C. J. and Thompson, E. A. (1992). ‘‘Constrained Monte Carlo maximum likelihood for dependent data (with
discussion)’’, JRSS, B 54, 657–699.Guyon, X. (1995). Randon Fields on a Network: Modelling, Statistics and Applications (Springer).Guyon, X. and Hardouin, C. (1992). ‘‘The chi2 difference of coding test for testing Markov Random Field
hypothesis’’, L.N.S. 74 (Springer), pp. 165–176.Guyon, X. and Kunsch, H. R. (1992). ‘‘ Asymptotic comparison of estimators in the Ising model’’, L.N.S. 74
(Springer), pp. 177–198.Hall, P. and Heyde, C. C. (1980). Martingale Limit Theory and its Applications (Acad. Press).Jensen, J. L. and Kunsch, H. R. (1994). ‘‘On asympotic normality of pseudo-likelihood estimate for pairwise
interaction processes’’, Ann. Inst. Statist. Math. 46, 475–486.Keiding, N. (1975). ‘‘Maximum likelihood estimation in the birth and death process’’, Ann. of Stats. 3, 363–372.Koslov, O. and Vasilyev, N. (1980). ‘‘Reversible Markov chain with local interaction’’, In: Dobrushin and Sinai
(Eds.), Multicomponent Random Systems (Dekker).Kunsch, H. R. (1984). ‘‘Time reversal and stationery Gibbs measures’’, Stoch. Proc. and Applications 17, 159–166.Pfeifer, P. E. and Deutsch, S. J. (1980). ‘‘Identification and interpretation of first order for space–time ARMA
models’’, Technometrics 22, 397–408.Pfeifer, P. E. and Deutsch, S. J. (1980). ‘‘A three stage iterative procedure for space time modelling’’, Technometrics
22, 35–47.Preston, C. (1974). Gibbs States on Countable Set, Cambridge tracts in math 68.Prum, B. (1986). Processus sur un reseau et mesure de Gibbs Applications, (Masson).Ripley, B. (1986). Spatial Statistics (Wiley).
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 359
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
Tang van Tham (1991) Rainfall greenhouse effect in the delta of Mekong, PhD thesis, Univ. of Cantho., Vietnam.Tong, H. (1990). Non Linear Time Series. A Dynamical System Approach (Clarendon Press, Oxford).Younes, L. (1988). ‘‘Estimation and annealing for Gibssian fields’’, Ann. I.H.P. 2, 269–294.Zucchini, W. and Guttorp, P. (1991). ‘‘A hidden Markov model for space time precipitation’’, Water Resour. Res. 27,
1917–1923.MacDonald, I. and Zucchini, W. (1997). ‘‘Hidden Markov and other models for discrete valued time series’’,
Monographs on Statistic and Applied Probability (Chapman and Hall).
APPENDIX 1
The Invariant Law p of an MCMF is Generally Not Markovian
We illustrate this with two examples.
(1) The one dimensional marginal of a two dimensional Gaussian Markov field is not
anymore a Markov process. Let us consider a centered Gaussian Markov isotropic field
with respect to the four-nearest neighbours over Z2 (see Besag, 1974; Ripley, 1981 and
Guyon, 1995, x1.3.4.)
Xst ¼ aðXs1;t þ Xsþ1;t þ Xs;t1 þ Xs;tþ1Þ þ est; jaj <1
4
with E½Xstes0t0 , ¼ 0 if ðs; tÞ 6¼ ðs0; t0Þ. The spectral density of X is
f ðl; mÞs2eð1 2aðcos lþ cos mÞÞ1. The one index field ðXs;0; s 2 ZÞ has spectral density
FðlÞ ¼ 2
ðp0
f ðl; mÞ dm ¼ 2ps2e ð1 2a cos lÞ2 4a2� �1
2
which cannot be written ð1=QðeilÞÞ or ð1=jPðeilÞj2Þ as for a Markovian or an AR process.
The distribution p of ðXs;0; s 2 ZÞ is not Markovian.
(2) Conditional Ising model: O ¼ f0; 1gS; S ¼ f1; 2; . . . ; ng is the one dimensional torus
with the agreement n þ 1 � 1, and Pðx; yÞ ¼ Z1ðxÞ expfaP
i2S xiyi þ bP
i2S yiyiþ1g.
This chain is reversible and the invariant distribution can be written
pðyÞ ¼ CZðyÞ expfgP
i2S yi þ bP
i2S yiyiþ1g where ZðyÞ is the normalizing constant of
Pð:jyÞ. It is easy to see that log pðyÞ has a non zero potential FS on the whole set S.
APPENDIX 2
Two Examples Where Marginals in y are Not Local in x
(1) Binary state space. Let us consider a binary chain on the one dimensional torus
S ¼ f1; 2; :::; ng (with n þ 1 � 1) with the transition
Pðx; yÞ ¼ Z1ðxÞ exp aXn
i¼1
xiyi þ bXn
i¼1
yiyiþ1 þ gXn
i¼1
yi
( ):
Then PiðyijxÞ ¼P
yi Pðx; yi; yiÞ ¼ ðeyiðaxiþgÞP
yj;j 6¼i ebyiðyj1þyjþ1ÞþAiÞ= ðP
yj;j 6¼i eAi þ eðaxiþgÞPyj;j 6¼i ebðyj1þyjþ1ÞþAiÞ, with Ai ¼ b
Pj 6¼i;i1 yjyjþ1 þ
Pj 6¼iðaxj þ gÞyj. This distribution
depends on x on all sites.
360 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
(2) A Gaussian example. We take Pðx; yÞ ¼ Z1ðxÞ exp U ðyjxÞ with U ðyjxÞ ¼ ðy mðxÞÞ0
�Qðy mðxÞÞ, where Q is symmetric, definite positive, and E½Y jx, ¼ mðxÞ. For
U ðyjxÞ ¼Pn
i¼1ðgiy2i þ diyiÞ þ
Pni¼1 aixiyi þ
Pni;j¼1;hi;ji bijyiyj; mðxÞ ¼ ðQ1=2ÞðdþaxÞ
where d ¼ ðdiÞi2S; ax ¼ ðaixiÞi2S , and Var½Y jx, ¼ ð1=2ÞQ1 : V ðY jxÞ1¼ 2Q is local in x
but it is easy to find parameters for which each component mlðxÞ of mðxÞ depends on all
xi; i 2 S.
APPENDIX 3
An Example of a Reverse MCMF Non MCMF
We consider S ¼ f1; 2; . . . ; ng; X ¼ X ð0Þ; Y ¼ X ð1Þ and we suppose that Z ¼ ðX ; Y Þ is
a 2n-dimensional centered Gaussian variable with covariance S ¼D BtB D
� �. Then,
ðY jxÞ ) N nðAx;QÞ, where A ¼ tBD1 and Q ¼ D tBD1B. We fix D and look at B as a
parameter (s.t. S is definite positive).
If we want conditional independence for fðXið1ÞjX ð0ÞÞ ¼ xÞ; i ¼ 1; ng, we force Q to be
diagonal; this involves ðnðn 1Þ=2Þ conditions on B, for n2 degrees of freedom.
On the other hand, ðX jyÞ ) N nðTy;RÞ where T ¼ BD1 and R ¼ D BD1tB. We can
chose B such that for some i;Rij 6¼ 0 for any j. In such a case, ðXið0Þjxi; yÞ depends on xi
on whole Snfig and the reverse chain is not an MCMF.
For example, this happens for n ¼ 2 if we take D ¼1 1=2
1=2 1
� �, B ¼
1=4 3=4
1=2 0
� �.
APPENDIX 4
Parametric Identifiability; Regularity of I0 and J0
First we take n�S ¼ �i2SKin as the reference measure, where Ki is defined in (10).
Identifiability. We suppose that the conditional densities are equal for y and y0. Then, for
each i 2 S; x and yi we have
t y y0ð Þgi yi; yi; x� �
¼ C y; yi; x� �
C y0; yi; x� �
But the right member is 0 as gið0; yi; xÞ ¼ 0. Identifiability follows because the d � d matrix
g is regular.
Regularity of I0. First we note that for the stationary distribution
I0 ¼ EX ðt1Þ;X iðtÞ
Xi2S
VarXiðtÞ gi XiðtÞ;X iðtÞ;X ðt 1Þ� � " #
:
Then, under (H0), we haveXi2S
Var gi Yi; yi; xÞjyi; x� �� �
� aXi2S
Gi x; yi� �
t Gi x; yi� �
:
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 361
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
As the density of ðX ðt 1Þ;X ðtÞÞ is strictly positive anywhere under C1 (see Duflo (1997),
Proposition 6.1.9), I0 is regular.
Regularity of J0. Let us recall that J0 ¼ VaryfgðX ; Y Þg where ðX ; Y Þ ¼ ðY ð0Þ;Y ð1ÞÞ; gðx; yÞ ¼
Pi2S giðx; yÞ; giðx; yÞ ¼ log fiðyi; yi; x; yÞ,ð1Þy . We follow an idea given by
Jensen and K€uunsch (1994). We suppose that there exists a ‘‘strong coding subset’’ C � S
in the following sense:
(i) there exists a partition fSj; j 2 Cg of S s. t. j 2 Sj.
(ii) For j 2 C, let us define Gj ¼P
i2Sjgi and let F be the s-field generated by X ¼ Y ð0Þ and
fYið1Þ; i =2 Cg. Then, conditionally to F , the variables fGj; j 2 Cg are independent.
A sufficient condition for (ii) is that for each j; l 2 C; j 6¼ l; l=2Sj [ @Sj. For example, for the
2-dimensional torus T ¼ f1; 2; . . . ; 3Kg2 and the 4-nearest neighbours vicinity,
C ¼ f3ðm; nÞ;m; n ¼ 1;Kg is a strong coding subset and Sm;n ¼ fði; kÞ : ji mj þ
jk nj ! 2g.
As g ¼P
j2C Gj, and J0 ¼ Var gðX ; Y Þ � EF ðVarfgðX ; Y ÞjFgÞ, we have, as a conse-
quence of (ii):
J0 �Xj2C
EF ðVarfGjðX ; Y ÞjFgÞ ¼ G0:
Then a sufficient condition ensuring that J0 is p. d. is that G0 is p. d. Such a verification has to
be done. For example, if S is the 1-dimensional torus with n ¼ 3K sites, with energy
U ðyjxÞ ¼ aP
i2S yini; ni ¼ xi þðyi1 þ yiþ1Þ; yi 2 f0; 1g, we can take C ¼ f3j; j ¼ 1;Kg and
Sj ¼ f3j 1; 3j; 3j þ 1g. As the model is homogeneous in the space, and as
ðg2 þ g3 þ g4jF Þ is never constant, G0 � EF ðVarðg2 þ g3 þg4ÞjF Þ > 0.
APPENDIX 5
Validation Tests for MCMF Models
The Autoexponential Dynamics (6)
eit ¼ liðyi; xÞXiðtÞ 1 and wit ¼ 23=2ðe2
it 1Þ. Next, for any a > 2, as ða þ bÞa !
2aðaa þ baÞ, we have
E jeitjajyi; x
� �! 2aliðy
i; xÞaGðaþ 1Þ
liðyi; xÞaþ liðy
i; xÞa� �
! 2aðGðaþ 1Þ þ 1Þ
and (14) is fulfilled.
Besides, E jwitjajyi; x
� �! 2ða=2Þ 22aðGð2aþ 1Þ þ 1Þ þ 1
and we get (16).
The Autopoissonian Dynamics (8)
The conditional distribution of XiðtÞ is Poisson with mean liðyi; xÞ ¼ expfdi þP
l2@i alixl þP
j2@i bijyjg. We look at conditions (13) and (14) for the eit . First we have
362 X. GUYON AND C. HARDOUIN
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014
hMit hMit1 ¼P
i2C liðXiðtÞ;X ðt 1ÞÞ ! eM jCj. Next, for any integer a, and some con-
stants Ck;a,
E eaitjyi; x
� �! 2a
Xk¼1;a
Ck;ali yi; x� �k
þli yi; x� �a !
! 2aX
k¼1;a
Ck;aeMk þ eMa
!:
Next, we consider the squared residuals nit ¼ e2it liðy
i; xÞ and ~nnt ¼P
i2C nit. As previously,
we can bound E½je2it liðy
i; xÞjajyi; x, for any a > 2, which implies (14). And we have (16)
with hNit hNit1 ¼ E½j~nntj2jF t1, ! jCjðe4M þ 6e3M þ 6e2M þ eM Þ. We get the announced
result.
The Autologistic Dynamics (9)
eit ¼ ðXiðtÞ piðyi; xÞÞ=
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffipiðyi; xÞð1 piðyi; xÞÞ
p. We can easily check that E½jeitj
ajyi; x,
as E½jwitjajyi; x, are bounded for any a > 2. For example, if the conditional energy
is U ðyjxÞ ¼P
i2S yiðdi þ aixiÞ þP
hi;ji bijyiyj, then piðyi; xÞ ¼ ð1 þ exp ½di þ aixi þP
j2@i bijyj,Þ1 and B1 is satisfied.
The Autodiscrete Dynamics
For each l; Zitl has a conditional Bernouilli distribution of parameter pilðyi; xÞ given by (17).
So, eit ¼ Zit E½Zitjyi; x; y0, has conditional variance Cit . Conditionally to the past and
X �CCðtÞ; eit ? ejt0 for all i; j 2 S if t 6¼ t0; and eit ? ejt if i; j are elements of a coding subset C
of S; thus feitgt is a martingale’s increment. Besides, E½teitC1it eitjy
i; x, ¼ K 1.
Finally, the state space E being finite and eit belonging to f0; 1gK1, all the moments of eit
are bounded and the condition on the increasing process is fulfilled.
The Autonormal Dynamics
We consider (7). The conditional distribution of XitðtÞ is Gaussian Nðmit; ð1=ð2giÞÞ with
mit ¼ ð1=ð2giÞÞðdi þP
l2@i alixl þP
j2@i bijyjÞ. We immediately deduce: eit ¼ ðXiðtÞ mitÞ=ffiffiffiffiffiffi2gi
p) Nð0; 1Þ, and e2
it ¼ 2gie2it ) w2
1.
MARKOV CHAIN MARKOV FIELD DYNAMICS: MODELS AND STATISTICS 363
Dow
nloa
ded
by [
Uni
vers
ity o
f O
tago
] at
04:
55 0
6 O
ctob
er 2
014