1.3 State-Space Extension and Reduction
1.3.4 Examples
A main thread throughout this work is that most seemingly-complicated Monte Carlo algorithms can be viewed as special cases of
is
on a suitably extended space (and thus as instances oftheMonte Carlo method). Forexample, it is well known thatsequential importance sampling (the idea
of which dates at least as far back as (Hammersley & Morton, 1954; M. N. Rosenbluth & Rosenbluth, 1955) and special cases of it such asannealed importance sampling (Jarzynski, 1997b, 1997a; Neal, 2001) may be viewed
as standard
is
.In this subsection, we show, by example, that this also applies to many other algorithms. First, as shown in Example 1.8,rejection sampling (also
referred to asaccept–reject method) first described in Kahn (1949), may
be viewed as a special case of (self-normalised)
is
on an extended space. This was pointed out in Y. Chen (2005), for instance.1.8 Example (rejection sampling). Rejection sampling is usually viewed as generating a random number of
iid
samples from some distribution2M1.X/, as follows.
(1) Propose
iid
samplesX1; : : : ; XN from some distribution 2M1.X/ satisfying .(2) Assume there existsz>0such thatwWDzd=d 1and such that
wcan be evaluated.
(3) Forn 2 NN WD fn 2 N j n Ng, independently ‘accept’Xnwith probabilityw.Xn/and setK WD fn2 NN jXnis ‘accepted’g. Then marginally,.Xn/n2K is an
iid
sample from.Rejection sampling thus entails generating
iid
samplesXx1; : : : ;XxN from an extended proposal distributionN WD ˝UnifŒ0;1 2M1.Xx/;
whereXxWDXŒ0;1. These proposals are then used to form an
is
approx- imationNis;N of the extended measureN
WD ˝L2M.xX/;
where L.x;dz/ WD 1Œ0;w.x/.z/dz. The importance weights are therefore defined byw.x x/N WDŒd =N d .x; z/N D1Œ0;w.x/.z/. The resulting marginal
1.3 State-Space Extension and Reduction
self-normalised
is
approximation of can then easily be seen to beN WD.#K/ 1X n2K
•Xn;
where#K denotes the cardinality of the setK. In particular,Nis;N.1/is an unbiased estimate of the marginal acceptance probability,z.
Finally, a standard
is
approximation of Dz on the marginal spaceX (with proposal distribution ) can be viewed as a Rao–Blackwellisation of the rejection-sampling approximation. That is,is;N.f /D 1 N N X nD1 wf .Xn/DENis;N.f ˝1Œ0;1/ ˇ ˇX :
Note that we are fixing the number of proposed samples, N, in the rejection-sampling scheme. A Rao–Blackwellisation in the case where rejec- tion sampling is performed until a certain number of acceptedsamples has been obtained was developed in Casella and Robert (1996).
Many other algorithms that seem to be generalisations of
is
, at a first glance, can actually also be viewed as standardis
on an extended space, as shown in Examples 1.9 and 1.10.1.9 Example (generalised importance sampling). Let x 2 K1.X;X/ be some-invariant stochastic kernel, i.e. such that x D. As shown in MacEachern, Clyde and Liu (1999, Theorem 6.1) it is possible to apply such a kernel to the weighted sample used to construct an
is
approximation of without having to adjust the weights.Even though this procedure is sometimes referred to as ‘generalised’ im- portance sampling (e.g. Robert & Casella, 2004, Section 14.2), as pointed out in Doucet and Johansen (2011) (see also Del Moral, Doucet & Jasra, 2006b), it may be viewed as standard importance sampling on the extended space
x
X WDX2(i.e.ZDXin the notation of this section), with extended proposal distribution N WD ˝ and extended target measureN WD˝˘, where
˘.x0;dx/ WD d .x; /
d .x
0/.dx/ D d .x; /
d .x
0/.dx/
represents thetime-reversal kernelof associated with. Indeed writing
x
X D .X; X0/, the weightsw.x x/N D Œd =N d .N x/N D Œd=d .x/ do not depend on the second component.
1.10 Example (dynamic weighting). Thedynamically weighted Monte
Carlo-framework (Wong & Liang, 1997; Liu, Liang & Wong, 2001; Liang, 2002) designs an extended measure
Q
.dxdv/ WDvg.dx;dv/;
on Xz WD X Œ0;1/. Here, g 2 P.zX/, Xz WD XŒ0;1/, is said to be
correctly weighted with respect to ifQ admits as a marginal, i.e. if
.A/D Q .AŒ0;1//for anyA2B.X/. In this case, an
iid
samplez
X1; : : : ;XzN Q WDg;
whereXznD.Xn; Vn/, can be used to approximate by standard
is
. The method is called ‘dynamic’ importance sampling because the nth importance weight, w.z Xzn/ D Vn, is not necessarily deterministic givenXn. It is also referred to as ‘generalised’ importance sampling in Liu (2001,
pp. 36–37), Liang (2002) because taking
g.dxdv/WD .dx/•w.x/.dv/;
wherew WDd=d , leads back to a direct
is
approximation of the marginalusing as a proposal distribution.
However, this approach is clearly no more than standard
is
on an extended space. Indeed, let N 2 M.xX/be some other extended measure on a spacex
X DXZsuch that (1)N admits as a marginal, (2)N has a densitywx with respect to some probability measure N 2M1.Xx/. Using the proposal distribution N, we can then construct an
is
approximationN is;N WD 1 N N X nD1 x w.Xxn/•Xxn
ofN and hence obtain an approximationN of the marginal.
Note, however, thatZ and thusXxD XZis often a high-dimensional space which can render the preceding