5.2 Non-Centred Metropolis-Within-Gibbs Algorithm
5.2.2 Non-Centred Parametrisation
The reparametrisation adopted in this paper is based on a representation derived in Ferguson and Klass (1972). The use of this representation to derive a (partially) non-centred parametrisation for compound Poisson
5.2 Non-Centred Metropolis-Within-Gibbs Algorithm processes was suggested by Roberts et al. (2004) and such reparametrisa- tions were also extensively used by Griffin and Steel (2006) in the context of Lévy-driven stochastic volatility models. Alternative reparametrisa- tions were suggested in Roberts et al. (2004) but were found to be inferior in the presence of efficient updates of the latent point process.
The basic idea is to take points distributed according to a unit-intensity
ppp
on T Œ0;/N , denoted z, whereN is some value in Œ;1/. First,we discard those points whose second components exceedsand divide
the second component of the remaining points by. This is sometimes
calledthinningand leaves a set of points distributed according to a
ppp
onTŒ0;1with intensity measureLebjTŒ0;1. These points are then
transformed into realisations of points ˘ by applying the inverse
cumulative distribution function(
cdf
)-method to the second component.For completeness and to set up some notation, we outline the formal justification of this reparametrisation below.
Define the space
z ΨWD 1 [ kD0 .fkg TkŒ0;//N
and let˘ .z d /z be the distribution of points z D.K;z Sz1W zK;Ez1W zK/which
are generated by a unit-intensity
ppp
onT.0;N . Again these points aretaken to be ordered according to their first components. In the following, we describe how z is first thinned and then transformed to obtain the
desired points ˘.
WriteNn WD fk2N jkng. Let H WD˚
j 2NKz ˇˇEzj
be the indices of points in z whose second component does not exceed Dl. /and setK WD#H. Similarly, let
y H WD˚
j 2NKz ˇˇEzj > DNKznH
be the set of the remaining indices and setKy WD#Hy. Collect the elements
ofH andHy in vectorshDh1WK andhO D Oh1W OK, in increasing order. Note
Points ˘ are then the result of the one-to-one reparametrisation
.; /z ! .; ; /;y (5.2)
where the left hand side represents an
ncp
and the right hand side rep- resents acentred parametrisation(cp
), i.e.(1) z WD.K;z Sz1W zK;Ez1W zK/represents all the points under the
ncp
,(2) D.K; S1WK; E1WK/is a sample from the desired
ppp
under thecp
,(3) y WD.K;y Sy1W yK;Ey1W yK/are ‘artificial’ points added to under the
cp
.In the following, we specify the one-to-one transformation involved in Equation 5.2.
Forj 2 NKy, the points in y are related to those under the
ncp
via .Syj;Eyj/WD zShOj;EzhOj:
Forj 2 NK, the points in are related to those under the
ncp
via .Sj; Ej/D Szhj;Ezhj:
Above, letting Fx denote the (generalised) inverse of the
cdf
Fassociated with, the functionW T.0; !TEis defined by
.s;Q e/Q 7!.s;Q e/Q WDŒs;Q Fx.1 e=/:Q
This function transforms the thinned points from the unit-intensity
ppp
z(under the
ncp
) – specifically, those points whose second component does not exceedN – into the desired points which make up the compoundPoisson process (under the
cp
). SinceFx is the (generalised) inverse ofthe
cdf
associated with, the functioncan be interpreted as applyingthe inverse
cdf
method.Letting and z be related as in Equation 5.2, we obtain the following
proposition, stated here for completeness.
5.2 Non-Centred Metropolis-Within-Gibbs Algorithm
Proof. The fact that is the set of ordered points distributed according
to a
ppp
onTEfollows by the independence property and the MappingTheorem for
ppp
s (Kingman, 1992, p. 18). In particular, forA2B.TE/,letting D denote the Jacobian of, ŒLeb˝2jT.0;/ı. / 1.A/ D Z A jdet.D/../ 1.s; e//j 1dsde D ŒLebjT˝ .A/ D.A/:
Thus, the is the set of ordered points distributed according to a
ppp
onTEwith intensity measure.
5.2.3
Extended Target Distribution
In this subsection, we specify an extended distribution.Q dd /. First,
this distribution admits the distribution of interest, .d d /, as a
marginal. Second, a Gibbs sampler or Metropolis-within-Gibbs algorithm targeting this distribution makes use of an
ncp
based on the representa- tion outlined above.Centred Parametrisation. The original formulation of the model in
Equation 5.1 uses a
cp
, i.e. it implies prior dependence between and. To permit the reparametrisation from Equation 5.2, we augment the target distribution.dd /with
y y˘ WD z˘jT.;N :
Here,˘zjT.;N is the distribution of points – again ordered according to
the first component – that have been generated by a unit-intensity
ppp
onT.;N . We thus obtain a distributionO / O over the parameters
on the right hand side in Equation 5.2, defined by
O
Non-Centred Parametrisation. Recall thatyTrepresents the collection
of all available observations (in the intervalTDŒ0; T ). With the abuse
of notation induced by writing
g.yTj Q /Dg.yTj /
whenever and z are related according to Equation 5.2), we can equi-
valently define an extended target distributionQ / Q parametrised via
the left hand side in Equation 5.2, i.e. by
Q
.dd /Q D$ .d /˘ .z d /gQ .yTj Q /:
This parametrisation represents an
ncp
becauseand z are independenta-priori. If the observations are not too informative, then the dependence betweenand zunderQ should be smaller than the dependence between and underO. Hence, such a parametrisation is often beneficial in
the context of Gibbs sampling.
5.2.4
The Algorithm
A single iteration of the non-centred
mcmc
algorithm proposed by Roberts et al. (2004) is given in Algorithm 5.3.5.3 Algorithm (non-centred Metropolis-within-Gibbs).
(1) Update z by sampling // using the
cp
(i) (components of ) using a.d j /-invariant
mcmc
kernel, (ii) y O.dOj; /D z˘jT.;N .d /O .
(2) Updateusing a.Q dj Q /-invariant
mcmc
kernel. // using thencp
5.4 Remark. Note that Step 1 is independent of the previously sampled points in yas these are sampled again from their full conditional distribution. Furthermore, assuming that anMetropolis–Hastings(
mh
) kernel is used to update the parameters, Step 2 is independent of those points in y that lie in the setT._?;:N
Here,?Dl.?/, where?is the value forproposed as part of the