5.2. Análisis e interpretación de los resultados por categorías y
5.2.1. Procesamiento y análisis de la información:
More complicated
mcmc
kernels are constructed by combining element- arymcmc
kernels. Indeed, let .Pzm/m2M be a countable collection of-invariant kernels withPzmD zP ..m; /; /2K1.X;X/, form2 MWD
NM. Additionally, letˇ 2K1.M;M/be another stochastic kernel which
will be used to determine the choice of kernelPm.
In this case, we can include the indicesM into the state space to justify
kernels of the form
P ..m; x/;dm0dx0/WDˇ.m;dm0/P ..mz 0; x/;dx0/:
This implies amixtureof (elementary) kernels on the marginal spaceX.
As a special case, we can justifycompositionsof kernels (on the marginal
space),P DP1 PM;by setting
3.4 Example (Gibbs sampler). Assume that we can decompose the state space as X DX
1WJ for some J 2 N and write X D X1WJ. Define Sets J1; : : : ; JM NJ such thatS
m2MJmDNJ. Furthermore, define themth Gibbs kernel via
z
P ..m; x/;dy/WD•xJm.dyJm/˘m.y Jm;dyJm/;
where XJ
m WD .Xj/j2Jm,X Jm WD.Xj/j2JnJm and where˘m.x Jm; /
denotes the full conditional distribution ofXJ
m under. In this case, the
Markov chain induced byP Dˇ˝ zP is called aGibbs sampler(S. Geman & Geman, 1984). In particular, it is called a
random-scanGibbs sampler ifˇ.m;/is non-degenerate,
deterministic-scanGibbs sampler ifˇis as given in Equation3.4.
3.5 Example (partially-collapsed Gibbs sampler).Deterministic-scan Gibbs samplers are often presented as updating every component ofX ex- actly once per iteration, i.e. taking.Jm/m2Mto be a partition ofNJ. However, large reductions in the asymptotic variance are achievable ifJm\JmC1 ¤ ;, and in this case, the Gibbs sampler is often referred to as apartially-collapsed Gibbs sampler (Van Dyk & Park, 2008).
3.2.3
Generic
mcmc
Kernel
Clearly, sampling from a tractable distribution on a sufficiently small, finite state space is feasible and we would not need
mcmc
algorithms for this task. However, the onlymcmc
kernel capable of sampling (approximately) from a higher-dimensional distribution on general state space is theGibbs-sampling kernel presented in Example 3.4. Unfortunately, this kernel requires sampling from full conditional distributions under
which are often intractable.
In this subsection, we construct a generic-invariant
mcmc
kernelwhich does not require sampling from full conditional distributions under
. Unsurprisingly, this generic kernel is based around a state-space
extension pioneered by Tjelmeland (2004). That is, it can be viewed as being invariant with respect to a distribution on an extended space which (1) admits as a marginal, (2) has full conditional distributions from
3.2 Generic
mcmc
Kernel as a Gibbs sampler targeting this extended distribution. As we shall see, this generic kernel admits most knownmcmc
kernels as a special case.The generic
mcmc
kernel is again based around (a normalised version of) the extended measureN D N = .N 1/2 M1.Xx/from Chapter 1. RecallthatXxDXKZ, whereK DNN andZDXN Y, whereYis the set
of values taken by some auxiliary variables,Y.
Note that in the previous section, we used the framework from Chapter 1 to interpret the entire
mcmc
algorithm as a special case ofis
. Here, we employ the same framework again but at a lower level to construct the-invariant kernel,P.
Further Extended Target Distribution. Recall that the extended dis-
tribution can be seen as a distribution over some auxiliary variablesY, a
pool ofN candidates forX,X DX1WN, and an indexK. It is construc-
ted in such a way that ifXx D .X; K;Z/ D .X; K; X1WN; Y / N then XK DX (Assumption 1.12). Approximations of were previously
constructed by integrating out.X; K/and thus averaging over all possible
candidates in an
is
scheme.Here, instead of integrating.X; K/out, we generate a new value for
the indexK, denotedKz, and a ‘new’ candidate,Xz DXKz, in such a way
that again,Xz under the extended distribution defined below. More
precisely, writingXz WD.Xx;K;z X /z and recalling thatZx D .K;Z/, the
generic
mcmc
kernel targets the extended distributionQ
.dxQ/WD N.dxN/ .xN;dkQdx/Q
D.dx/˘ .x;x dzN/ .xN;dkQdx/Q
onXz WD xXKX. Here, 2 K1.Xx;KX/is chosen such that z
X Q ) XKz D zX : (3.5)
Note that such a kernel exists because we can always (and will often) take
.xN;dkQdx/Q D Nc.z;dkQdx/Q , whereNc.z; /is the full conditional
distribution of.K; X /underN.
Dominating Measure. Throughout the remainder of this chapter, we
assume that N has a density wx with respect to a suitable dominating
measure N which admits the factorisation N
where 2 M¢.Z/and 2 K¢.Z;K/. This factorisation is required to
obtain the following explicit representation ofNc. ForxN D.x; k;z/, write wk.z/WD xw.x/ .N z;fkg/. We can then represent the full conditional dis-
tribution of.X; K/underN asNc.z;dkdx/ D Nz.dk/•xk.dx/, where
N z.fkg/WD wk.z/ PN nD1wn.z/ :
Finally, note the measure does not depend onk. It therefore gen-
eralises the symmetric dominating measure required in Green (1995) as discussed in the Subsection 3.3.2,
Summary. By Equation 3.5, a-invariant kernel is given by
P .x; A/ WD Z x ZKA x ˘ .x;dzN/ .xN;dkQdx/;Q
for all.x; A/2 XB.X/. It is summarised in Algorithm 3.6.
3.6 Algorithm (generic
mcmc
kernel). GivenX , (1) sampleZx x˘ .x; /and setXx WD.X;Zx/,(2) sample.K;z X /z .xN; /, (3) outputX WD zX(DXKz).
Note that the generic
mcmc
kernel is just a Gibbs kernel targeting the extended distributionQ. The advantage of this state-space extension isthat the kernels˘x andcan often be chosen in such a way that sampling
from them – and thus sampling from this Gibbs kernel – is feasible even if sampling from full conditional distributions under is not.
Finally, we note that the generic
mcmc
kernel is-invariant by con-struction. In particular, this insight immediately shows that all the
mcmc
kernels mentioned in this chapter are-invariant (since they are all spe-
cial cases of this approach) without appealing (explicitly) to sufficient conditions such asdetailed balance. However, we reiterate the fact that
-invariance is not sufficient for obtaining useful
mcmc
kernels. Indeed,takingq WD in Example 1.14 can be viewed as initialising the Markov
chain from and then applying the trivial kernelP .x;/WD •x. This
3.2 Generic
mcmc
Kernel For later use, we define the following terminology, borrowed from Nicholls, Fox and Watt (2012), which will be justified in the next section.3.7 Definition (randomised