P. 1
Andrieu 2003 Intro MCMC Machine Learning

Andrieu 2003 Intro MCMC Machine Learning

|Views: 8|Likes:
Published by Nicholas J. Matzke

More info:

Published by: Nicholas J. Matzke on May 19, 2011
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

05/19/2011

pdf

text

original

Machine Learning, 50, 5–43, 2003

c 2003 Kluwer Academic Publishers. Manufactured in The Netherlands.
An Introduction to MCMC for Machine Learning
CHRISTOPHE ANDRIEU C.Andrieu@bristol.ac.uk
Department of Mathematics, Statistics Group, University of Bristol, University Walk, Bristol BS8 1TW, UK
NANDO DE FREITAS nando@cs.ubc.ca
Department of Computer Science, University of British Columbia, 2366 Main Mall, Vancouver,
BC V6T 1Z4, Canada
ARNAUD DOUCET doucet@ee.mu.oz.au
Department of Electrical and Electronic Engineering, University of Melbourne, Parkville, Victoria 3052, Australia
MICHAEL I. JORDAN jordan@cs.berkeley.edu
Departments of Computer Science and Statistics, University of California at Berkeley, 387 Soda Hall, Berkeley,
CA 94720-1776, USA
Abstract. This purpose of this introductory paper is threefold. First, it introduces the Monte Carlo method with
emphasis on probabilistic machine learning. Second, it reviews the main building blocks of modern Markov chain
Monte Carlo simulation, thereby providing and introduction to the remaining papers of this special issue. Lastly,
it discusses new interesting research horizons.
Keywords: Markov chain Monte Carlo, MCMC, sampling, stochastic algorithms
1. Introduction
Arecent survey places the Metropolis algorithmamong the ten algorithms that have had the
greatest influence on the development and practice of science and engineering in the 20th
century (Beichl &Sullivan, 2000). This algorithmis an instance of a large class of sampling
algorithms, known as Markov chain Monte Carlo (MCMC). These algorithms have played
a significant role in statistics, econometrics, physics and computing science over the last
two decades. There are several high-dimensional problems, such as computing the volume
of a convex body in d dimensions, for which MCMC simulation is the only known general
approach for providing a solution within a reasonable time (polynomial in d) (Dyer, Frieze,
& Kannan, 1991; Jerrum & Sinclair, 1996).
While convalescing from an illness in 1946, Stan Ulam was playing solitaire. It, then,
occurred to himto try to compute the chances that a particular solitaire laid out with 52 cards
would come out successfully (Eckhard, 1987). After attempting exhaustive combinatorial
calculations, he decided to go for the more practical approach of laying out several solitaires
at random and then observing and counting the number of successful plays. This idea of
selecting a statistical sample to approximate a hard combinatorial problem by a much
simpler problem is at the heart of modern Monte Carlo simulation.
6 C. ANDRIEU ET AL.
Stan Ulam soon realised that computers could be used in this fashion to answer ques-
tions of neutron diffusion and mathematical physics. He contacted John Von Neumann,
who understood the great potential of this idea. Over the next few years, Ulam and Von
Neumann developed many Monte Carlo algorithms, including importance sampling and
rejection sampling. Enrico Fermi in the 1930’s also used Monte Carlo in the calculation of
neutron diffusion, and later designed the FERMIAC, a Monte Carlo mechanical device that
performed calculations (Anderson, 1986). In the 1940’s Nick Metropolis, a young physicist,
designed new controls for the state-of-the-art computer (ENIAC) with Klari Von Neumann,
John’s wife. He was fascinated with Monte Carlo methods and this new computing device.
Soon he designed an improved computer, which he named the MANIAC in the hope that
computer scientists would stop using acronyms. During the time he spent working on the
computing machines, many mathematicians and physicists (Fermi, Von Neumann, Ulam,
Teller, Richtmyer, Bethe, Feynman, & Gamow) would go to him with their work problems.
Eventually in 1949, he published the first public document on Monte Carlo simulation with
Stan Ulam (Metropolis & Ulam, 1949). This paper introduces, among other ideas, Monte
Carlo particle methods, which form the basis of modern sequential Monte Carlo methods
such as bootstrap filters, condensation, and survival of the fittest algorithms (Doucet, de
Freitas, &Gordon, 2001). Soon after, he proposed the Metropolis algorithmwith the Tellers
and the Rosenbluths (Metropolis et al., 1953).
Many papers on Monte Carlo simulation appeared in the physics literature after 1953.
From an inference perspective, the most significant contribution was the generalisation of
the Metropolis algorithmby Hastings in 1970. Hastings and his student Peskun showed that
Metropolis and the more general Metropolis-Hastings algorithms are particular instances
of a large family of algorithms, which also includes the Boltzmann algorithm (Hastings,
1970; Peskun, 1973). They studied the optimality of these algorithms and introduced the
formulation of the Metropolis-Hastings algorithm that we adopt in this paper. In the 1980’s,
two important MCMC papers appeared in the fields of computer vision and artificial in-
telligence (Geman & Geman, 1984; Pearl, 1987). Despite the existence of a few MCMC
publications in the statistics literature at this time, it is generally accepted that it was only in
1990 that MCMC made the first significant impact in statistics (Gelfand & Smith, 1990). In
the neural networks literature, the publication of Neal (1996) was particularly influential.
In the introduction to this special issue, we focus on describing algorithms that we feel
are the main building blocks in modern MCMC programs. We should emphasize that in
order to obtain the best results out of this class of algorithms, it is important that we do not
treat themas black boxes, but instead try to incorporate as much domain specific knowledge
as possible into their design. MCMC algorithms typically require the design of proposal
mechanisms to generate candidate hypotheses. Many existing machine learning algorithms
can be adapted to become proposal mechanisms (de Freitas et al., 2001). This is often
essential to obtain MCMC algorithms that converge quickly. In addition to this, we believe
that the machine learning community can contribute significantly to the solution of many
open problems in the MCMCfield. For this purpose, we have outlined several “hot” research
directions at the end of this paper. Finally, readers are encouraged to consult the excellent
texts of Chen, Shao, and Ibrahim (2001), Gilks, Richardson, and Spiegelhalter (1996), Liu
(2001), Meyn and Tweedie (1993), Robert and Casella (1999) and review papers by Besag
INTRODUCTION 7
et al. (1995), Brooks (1998), Diaconis and Saloff-Coste (1998), Jerrumand Sinclair (1996),
Neal (1993), and Tierney (1994) for more information on MCMC.
The remainder of this paper is organised as follows. In Part 2, we outline the general
problems and introduce simple Monte Carlo simulation, rejection sampling and importance
sampling. Part 3 deals with the introduction of MCMC and the presentation of the most
popular MCMC algorithms. In Part 4, we describe some important research frontiers. To
make the paper more accessible, we make no notational distinction between distributions
and densities until the section on reversible jump MCMC.
2. MCMC motivation
MCMC techniques are often applied to solve integration and optimisation problems in
large dimensional spaces. These two types of problem play a fundamental role in machine
learning, physics, statistics, econometrics and decision analysis. The following are just some
examples.
1. Bayesian inference and learning. Given some unknown variables x ∈ X and data y ∈ Y,
the following typically intractable integration problems are central to Bayesian statistics
(a) Normalisation. To obtain the posterior p(x [ y) given the prior p(x) and likelihood
p(y [ x), the normalising factor in Bayes’ theorem needs to be computed
p(x [ y) =
p(y [ x) p(x)

X
p(y [ x
/
) p(x
/
) dx
/
.
(b) Marginalisation. Given the joint posterior of (x, z) ∈ X Z, we may often be
interested in the marginal posterior
p(x [ y) =

Z
p(x, z [ y) dz.
(c) Expectation. The objective of the analysis is often to obtain summary statistics of
the form
E
p(x[y)
( f (x)) =

X
f (x) p(x [ y) dx
for some function of interest f : X → R
n
f
integrable with respect to p(x [ y).
Examples of appropriate functions include the conditional mean, in which case
f (x) = x, or the conditional covariance of x where f (x) = xx
/
−E
p(x[y)
(x)E
/
p(x[y)
(x).
2. Statistical mechanics. Here, one needs to compute the partition function Z of a system
with states s and Hamiltonian E(s)
Z =
¸
s
exp
¸

E(s)
kT
¸
,
where k is the Boltzmann’s constant and T denotes the temperature of the system.
Summing over the large number of possible configurations is prohibitively expensive
(Baxter, 1982). Note that the problems of computing the partition function and the
normalising constant in statistical inference are analogous.
8 C. ANDRIEU ET AL.
3. Optimisation. The goal of optimisation is to extract the solution that minimises some
objective function from a large set of feasible solutions. In fact, this set can be contin-
uous and unbounded. In general, it is too computationally expensive to compare all the
solutions to find out which one is optimal.
4. Penalised likelihood model selection. This task typically involves two steps. First, one
finds the maximum likelihood (ML) estimates for each model separately. Then one uses
a penalisation term (for example MDL, BIC or AIC) to select one of the models. The
problem with this approach is that the initial set of models can be very large. Moreover,
many of those models are of not interest and, therefore, computing resources are wasted.
Although we have emphasized integration and optimisation, MCMC also plays a funda-
mental role in the simulation of physical systems. This is of great relevance in nuclear
physics and computer graphics (Chenney &Forsyth, 2000; Kalos &Whitlock, 1986; Veach
& Guibas, 1997).
2.1. The Monte Carlo principle
The idea of Monte Carlo simulation is to draw an i.i.d. set of samples {x
(i )
}
N
i =1
from a target
density p(x) defined on a high-dimensional space X (e.g. the set of possible configurations
of a system, the space on which the posterior is defined, or the combinatorial set of feasible
solutions). These N samples can be used to approximate the target density with the following
empirical point-mass function
p
N
(x) =
1
N
N
¸
i =1
δ
x
(i ) (x),
where δ
x
(i ) (x) denotes the delta-Dirac mass located at x
(i )
. Consequently, one can approx-
imate the integrals (or very large sums) I ( f ) with tractable sums I
N
( f ) that converge as
follows
I
N
( f ) =
1
N
N
¸
i =1
f

x
(i )

a.s.
−−−→
N→∞
I ( f ) =

X
f (x) p(x) dx.
That is, the estimate I
N
( f ) is unbiased and by the strong law of large numbers, it will
almost surely (a.s.) converge to I ( f ). If the variance (in the univariate case for simplicity)
of f (x) satisfies σ
2
f
E
p(x)
( f
2
(x)) − I
2
( f ) < ∞, then the variance of the estimator
I
N
( f ) is equal to var(I
N
( f )) =
σ
2
f
N
and a central limit theorem yields convergence in
distribution of the error

N(I
N
( f ) − I ( f )) =⇒
N→∞
N

0, σ
2
f

,
where =⇒ denotes convergence in distribution (Robert & Casella, 1999; Section 3.2).
The advantage of Monte Carlo integration over deterministic integration arises from the
fact that the former positions the integration grid (samples) in regions of high probability.
INTRODUCTION 9
The N samples can also be used to obtain a maximum of the objective function p(x) as
follows
ˆ x = arg max
x
(i )
;i =1,...,N
p

x
(i )

However, we will show later that it is possible to construct simulated annealing algorithms
that allow us to sample approximately from a distribution whose support is the set of global
maxima.
When p(x) has standard form, e.g. Gaussian, it is straightforward to sample from it using
easily available routines. However, when this is not the case, we need to introduce more
sophisticated techniques based on rejection sampling, importance sampling and MCMC.
2.2. Rejection sampling
We can sample from a distribution p(x), which is known up to a proportionality constant,
by sampling from another easy-to-sample proposal distribution q(x) that satisfies p(x) ≤
Mq(x), M < ∞, using the accept/reject procedure describe in figure 1 (see also figure 2).
The accepted x
(i )
can be easily shown to be sampled with probability p(x) (Robert &
Figure 1. Rejection sampling algorithm. Here, u ∼ U
(0,1)
denotes the operation of sampling a uniform random
variable on the interval (0, 1).
Figure 2. Rejection sampling: Sample a candidate x
(i )
and a uniform variable u. Accept the candidate sample if
uMq(x
(i )
) < p(x
(i )
), otherwise reject it.
10 C. ANDRIEU ET AL.
Casella, 1999, p. 49). This simple method suffers from severe limitations. It is not always
possible to bound p(x)/q(x) with a reasonable constant M over the whole space X. If M
is too large, the acceptance probability
Pr(x accepted) = Pr

u <
p(x)
Mq(x)

=
1
M
will be too small. This makes the method impractical in high-dimensional scenarios.
2.3. Importance sampling
Importance sampling is an alternative “classical” solution that goes back to the 1940’s;
see for example (Geweke, 1989; Rubinstein, 1981). Let us introduce, again, an arbitrary
importance proposal distribution q(x) such that its support includes the support of p(x).
Then we can rewrite I ( f ) as follows
I ( f ) =

f (x) w(x) q(x) dx
where w(x)
p(x)
q(x)
is known as the importance weight. Consequently, if one can simulate
N i.i.d. samples {x
(i )
}
N
i =1
according to q(x) and evaluate w(x
(i )
), a possible Monte Carlo
estimate of I ( f ) is
ˆ
I
N
( f ) =
N
¸
i =1
f

x
(i )

w

x
(i )

This estimator is unbiased and, under weak assumptions, the strong law of large num-
bers applies, that is
ˆ
I
N
( f )
a.s.
−→
N→∞
I ( f ). It is clear that this integration method can also be
interpreted as a sampling method where the posterior density p(x) is approximated by
ˆ p
N
(x) =
N
¸
i =1
w

x
(i )

δ
x
(i ) (x)
and
ˆ
I
N
( f ) is nothing but the function f (x) integrated with respect to the empirical measure
ˆ p
N
(x).
Some proposal distributions q(x) will obviously be preferable to others. An important
criterion for choosing an optimal proposal distribution is to find one that minimises the
variance of the estimator
ˆ
I
N
( f ). The variance of f (x)w(x) with respect to q(x) is given by
var
q(x)
( f (x)w(x)) = E
q(x)
( f
2
(x)w
2
(x)) − I
2
( f ) (8)
The second term on the right hand side does not depend on q(x) and hence we only need
to minimise the first term, which according to Jensen’s inequality has the following lower
INTRODUCTION 11
bound
E
q(x)
( f
2
(x)w
2
(x)) ≥

E
q(x)
([ f (x)[w(x))

2
=

[ f (x)[ p(x) dx

2
This lower bound is attained when we adopt the following optimal importance distribution
q

(x) =
[ f (x)[ p(x)

[ f (x)[ p(x) dx
The optimal proposal is not very useful in the sense that it is not easy to sample from
[ f (x)[ p(x). However, it tells us that high sampling efficiency is achieved when we focus
on sampling from p(x) in the important regions where [ f (x)[ p(x) is relatively large; hence
the name importance sampling.
This result implies that importance sampling estimates can be super-efficient. That is,
for a a given function f (x), it is possible to find a distribution q(x) that yields an estimate
with a lower variance than when using a perfect Monte Carlo method, i.e. with q(x) = p(x).
This property is often exploited to evaluate the probability of rare events in communication
networks (Smith, Shafi, & Gao, 1997). There the quantity of interest is a tail probability
(bit error rate) and hence f (x) =I
E
(x) where I
E
(x) =1 if x ∈ E and 0 otherwise (see
figure 3). One could estimate the bit error rate more efficiently by sampling according to
q(x) ∝I
E
(x) p(x) than according to q(x) = p(x). That is, it is wasteful to propose candidates
in regions of no utility. In many applications, the aim is usually different in the sense that
Figure 3. Importance sampling: one should place more importance on sampling from the state space regions that
matter. In this particular example one is interested in computing a tail probability of error (detecting infrequent
abnormalities).
12 C. ANDRIEU ET AL.
one wants to have a good approximation of p(x) and not of a particular integral with respect
to p(x), so we often seek to have q(x) . p(x).
As the dimension of the x increases, it becomes harder to obtain a suitable q(x) from
which to draw samples. A sensible strategy is to adopt a parameterised q(x, θ) and to
adapt θ during the simulation. Adaptive importance sampling appears to have originated
in the structural safety literature (Bucher, 1988), and has been extensively applied in the
communications literature (Al-Qaq, Devetsikiotis, & Townsend, 1995; Remondo et al.,
2000). This technique has also been exploited recently in the machine learning community
(de Freitas et al., 2000; Cheng & Druzdzel, 2000; Ortiz & Kaelbling, 2000; Schuurmans &
Southey, 2000). A popular adaptive strategy involves computing the derivative of the first
term on the right hand side of Eq. (8)
D(θ) = E
q(x,θ)

f
2
(x)w(x, θ)
∂w(x, θ)
∂θ

and then updating the parameters as follows
θ
t ÷1
= θ
t
−α
1
N
N
¸
i =1
f
2

x
(i )

w

x
(i )
, θ
t
∂w

x
(i )
, θ
t

∂θ
t
where α is a learning rate and x
(i )
∼ q(x, θ). Other optimisation approaches that make use
of the Hessian are also possible.
When the normalising constant of p(x) is unknown, it is still possible to apply the
importance sampling method by rewriting I ( f ) as follows:
I ( f ) =

f (x)w(x)q(x) dx

w(x)q(x) dx
where w(x) ∝
p(x)
q(x)
is now only known up to a normalising constant. The Monte Carlo
estimate of I ( f ) becomes
˜
I
N
( f ) =
1
N
¸
N
i =1
f

x
(i )

w

x
(i )

1
N
¸
N
j =1
w

x
(i )
=
N
¸
i =1
f

x
(i )

˜ w

x
(i )

where ˜ w(x
(i )
) is a normalised importance weight. For N finite,
˜
I
N
( f ) is biased (ratio of two
estimates) but asymptotically, under weak assumptions, the strong law of large numbers
applies, that is
˜
I
N
( f )
a.s.
−→
N→∞
I ( f ). Under additional assumptions a central limit theorem can
be obtained (Geweke, 1989). The estimator
˜
I
N
( f ) has been shown to perform better than
ˆ
I
N
( f ) in some setups under squared error loss (Robert & Casella, 1999).
If one is interested in obtaining M i.i.d. samples from ˆ p
N
(x), then an asymptotically
(N/M →∞) valid method consists of resampling M times according to the discrete distri-
bution ˆ p
N
(x). This procedure results in M samples ˜ x
(i )
with the possibility that ˜ x
(i )
= ˜ x
( j )
INTRODUCTION 13
for i ,= j . This method is known as sampling importance resampling (SIR) (Rubin, 1988).
After resampling, the approximation of the target density is
˜ p
M
(x) =
1
M
M
¸
i =1
δ
˜ x
(i ) (x) (13)
The resampling scheme introduces some additional Monte Carlo variation. It is, therefore,
not clear whether the SIR procedure can lead to practical gains in general. However, in
the sequential Monte Carlo setting described in Section 4.3, it is essential to carry out this
resampling step.
We conclude this section by stating that even with adaptation, it is often impossible to
obtain proposal distributions that are easy to sample from and good approximations at the
same time. For this reason, we need to introduce more sophisticated sampling algorithms
based on Markov chains.
3. MCMC algorithms
MCMC is a strategy for generating samples x
(i )
while exploring the state space X using a
Markov chain mechanism. This mechanism is constructed so that the chain spends more
time in the most important regions. In particular, it is constructed so that the samples x
(i )
mimic samples drawn from the target distribution p(x). (We reiterate that we use MCMC
when we cannot drawsamples from p(x) directly, but can evaluate p(x) up to a normalising
constant.)
It is intuitive to introduce Markov chains on finite state spaces, where x
(i )
can only take
s discrete values x
(i )
∈X = {x
1
, x
2
, . . . , x
s
}. The stochastic process x
(i )
is called a Markov
chain if
p

x
(i )

x
(i −1)
, . . . , x
(1)

= T

x
(i )

x
(i −1)

,
The chainis homogeneous if T T(x
(i )
[ x
(i −1)
) remains invariant for all i , with
¸
x
(i ) T(x
(i )
[
x
(i −1)
) = 1 for any i . That is, the evolution of the chain in a space X depends solely on the
current state of the chain and a fixed transition matrix.
As an example, consider a Markov chain with three states (s = 3) and a transition graph
as illustrated in figure 4. The transition matrix for this example is
T =
¸
¸
¸
0 1 0
0 0.1 0.9
0.6 0.4 0

If the probability vector for the initial state is µ(x
(1)
) = (0.5, 0.2, 0.3), it follows that
µ(x
(1)
)T = (0.2, 0.6, 0.2) and, after several iterations (multiplications by T), the product
µ(x
(1)
)T
t
converges to p(x) = (0.2, 0.4, 0.4). No matter what initial distribution µ(x
(1)
)
we use, the chain will stabilise at p(x) =(0.2, 0.4, 0.4). This stability result plays a funda-
mental role in MCMC simulation. For any starting point, the chain will convergence to the
14 C. ANDRIEU ET AL.
1
0.6
0.4
0.9
1
2
3
x
x
0.1
x
Figure 4. Transition graph for the Markov chain example with X = {x
1
, x
2
, x
3
}.
invariant distribution p(x), as long as T is a stochastic transition matrix that obeys the
following properties:
1. Irreducibility. For any state of the Markov chain, there is a positive probability of visiting
all other states. That is, the matrix T cannot be reduced to separate smaller matrices,
which is also the same as stating that the transition graph is connected.
2. Aperiodicity. The chain should not get trapped in cycles.
A sufficient, but not necessary, condition to ensure that a particular p(x) is the desired
invariant distribution is the following reversibility (detailed balance) condition
p

x
(i )

T

x
(i −1)

x
(i )

= p

x
(i −1)

T

x
(i )

x
(i −1)

.
Summing both sides over x
(i −1)
, gives us
p

x
(i )

=
¸
x
(i −1)
p

x
(i −1)

T

x
(i )
[ x
(i −1)

.
MCMC samplers are irreducible and aperiodic Markov chains that have the target distribu-
tion as the invariant distribution. One way to design these samplers is to ensure that detailed
balance is satisfied. However, it is also important to design samplers that converge quickly.
Indeed, most of our efforts will be devoted to increasing the convergence speed.
Spectral theory gives us useful insights into the problem. Notice that p(x) is the left
eigenvector of the matrix T with corresponding eigenvalue 1. In fact, the Perron-Frobenius
theorem from linear algebra tells us that the remaining eigenvalues have absolute value less
than 1. The second largest eigenvalue, therefore, determines the rate of convergence of the
chain, and should be as small as possible.
The concepts of irreducibility, aperiodicity and invariance can be better appreciated once
we realise the important role that they play in our lives. When we search for information on
INTRODUCTION 15
the World-Wide Web, we typically follow a set of links (Berners-Lee et al., 1994). We can
interpret the webpages and links, respectively, as the nodes and directed connections in a
Markov chain transition graph. Clearly, we (say, the random walkers on the Web) want to
avoid getting trapped in cycles (aperiodicity) and want to be able to access all the existing
webpages (irreducibility). Let us consider, now, the popular information retrieval algorithm
used by the search engine Google, namely PageRank (Page et al., 1998). PageRank requires
the definition of a transition matrix with two components T = L÷E. L is a large link matrix
with rows and columns corresponding to web pages, such that the entry L
i, j
represents the
normalised number of links from web page i to web page j . E is a uniform random matrix
of small magnitude that is added to L to ensure irreducibility and aperiodicity. That is, the
addition of noise prevents us from getting trapped in loops, as it ensures that there is always
some probability of jumping to anywhere on the Web. From our previous discussion, we
have
p

x
(i ÷1)

[L ÷ E] = p(x
i
)
where, in this case, the invariant distribution (eigenvector) p(x) represents the rank of a
webpage x. Note that it is possible to design more interesting transition matrices in this
setting. As long as one satisfies irreducibility and aperiodicity, one can incorporate terms
into the transition matrix that favour particular webpages or that bias the search in useful
ways.
In continuous state spaces, the transition matrix T becomes an integral kernel K and
p(x) becomes the corresponding eigenfunction

p

x
(i )

K

x
(i ÷1)

x
(i )

dx
(i )
= p

x
(i ÷1)

.
The kernel K is the conditional density of x
(i ÷1)
given the value x
(i )
. It is a mathematical
representation of a Markov chain algorithm. In the following subsections we describe
various of these algorithms.
3.1. The Metropolis-Hastings algorithm
The Metropolis-Hastings (MH) algorithm is the most popular MCMC method (Hastings,
1970; Metropolis et al., 1953). In later sections, we will see that most practical MCMC
algorithms can be interpreted as special cases or extensions of this algorithm.
An MH step of invariant distribution p(x) and proposal distribution q(x

[ x) involves
sampling a candidate value x

given the current value x according to q(x

[ x). The Markov
chain then moves towards x

with acceptance probability A(x, x

) = min{1, [ p(x)q(x

[
x)]
−1
p(x

)q(x [ x

)}, otherwise it remains at x. The pseudo-code is shown in figure 5, while
figure 6 shows the results of running the MHalgorithmwith a Gaussian proposal distribution
q(x

[ x
(i )
) = N(x
(i )
, 100) and a bimodal target distribution p(x) ∝ 0.3 exp(−0.2x
2
) ÷
0.7 exp(−0.2(x − 10)
2
) for 5000 iterations. As expected, the histogram of the samples
approximates the target distribution.
16 C. ANDRIEU ET AL.
Figure 5. Metropolis-Hastings algorithm.
−10 0 10 20
0
0.05
0.1
0.15
i=100
−10 0 10 20
0
0.05
0.1
0.15
i=500
−10 0 10 20
0
0.05
0.1
0.15
i=1000
−10 0 10 20
0
0.05
0.1
0.15
i=5000
Figure 6. Target distribution and histogram of the MCMC samples at different iteration points.
The MH algorithm is very simple, but it requires careful design of the proposal distri-
bution q(x

[ x). In subsequent sections, we will see that many MCMC algorithms arise by
considering specific choices of this distribution. In general, it is possible to use suboptimal
inference and learning algorithms to generate data-driven proposal distributions.
The transition kernel for the MH algorithm is
K
MH

x
(i ÷1)

x
(i )

= q

x
(i ÷1)

x
(i )

A

x
(i )
, x
(i ÷1)

÷δ
x
(i )

x
(i ÷1)

r

x
(i )

,
INTRODUCTION 17
where r(x
(i )
) is the term associated with rejection
r

x
(i )

=

X
q

x

x
(i )

1 −A

x
(i )
, x

dx

.
It is fairly easy to prove that the samples generated by MH algorithm will mimic samples
drawnfromthe target distributionasymptotically. Byconstruction, K
MH
satisfies the detailed
balance condition
p

x
(i )

K
MH

x
(i ÷1)

x
(i )

= p

x
(i ÷1)

K
MH

x
(i )

x
(i ÷1)

and, consequently, the MH algorithm admits p(x) as invariant distribution. To show that
the MH algorithm converges, we need to ensure that there are no cycles (aperiodicity)
and that every state that has positive probability can be reached in a finite number of steps
(irreducibility). Since the algorithmalways allows for rejection, it follows that it is aperiodic.
To ensure irreducibility, we simply need to make sure that the support of q(·) includes the
support of p(·). Under these conditions, we obtain asymptotic convergence (Tierney, 1994,
Theorem 3, p. 1717). If the space X is small (for example, bounded in R
n
), then it is
possible to use minorisation conditions to prove uniform (geometric) ergodicity (Meyn &
Tweedie, 1993). It is also possible to prove geometric ergodicity using Foster-Lyapunov
drift conditions (Meyn & Tweedie, 1993; Roberts & Tweedie, 1996).
The independent sampler and the Metropolis algorithm are two simple instances of the
MH algorithm. In the independent sampler the proposal is independent of the current state,
q(x

[ x
(i )
) = q(x

). Hence, the acceptance probability is
A

x
(i )
, x

= min
¸
1,
p(x

)q

x
(i )

p

x
(i )

q(x

)
¸
= min
¸
1,
w(x

)
w

x
(i )

¸
.
This algorithm is close to importance sampling, but now the samples are correlated since
they result from comparing one sample to the other. The Metropolis algorithm assumes a
symmetric random walk proposal q(x

[ x
(i )
) = q(x
(i )
[ x

) and, hence, the acceptance ratio
simplifies to
A

x
(i )
, x

= min
¸
1,
p(x

)
p

x
(i )

¸
.
Some properties of the MH algorithm are worth highlighting. Firstly, the normalising
constant of the target distributionis not required. We onlyneedtoknowthe target distribution
up to a constant of proportionality. Secondly, although the pseudo-code makes use of a single
chain, it is easy to simulate several independent chains in parallel. Lastly, the success or
failure of the algorithmoften hinges on the choice of proposal distribution. This is illustrated
in figure 7. Different choices of the proposal standard deviation σ

lead to very different
results. If the proposal is too narrow, only one mode of p(x) might be visited. On the other
hand, if it is too wide, the rejection rate can be very high, resulting in high correlations. If all
the modes are visited while the acceptance probability is high, the chain is said to “mix” well.
18 C. ANDRIEU ET AL.
Figure 7. Approximations obtained using the MH algorithm with three Gaussian proposal distributions of dif-
ferent variances.
3.2. Simulated annealing for global optimization
Let us assume that instead of wanting to approximate p(x), we want to find its global
maximum. For example, if p(x) is the likelihood or posterior distribution, we often want
to compute the ML and maximum a posteriori (MAP) estimates. As mentioned earlier, we
could run a Markov chain of invariant distribution p(x) and estimate the global mode by
ˆ x = arg max
x
(i )
;i =1,...,N
p

x
(i )

.
This method is inefficient because the random samples only rarely come from the vicinity
of the mode. Unless the distribution has large probability mass around the mode, computing
resources will be wasted exploring areas of no interest. A more principled strategy is to
adopt simulated annealing (Geman & Geman, 1984; Kirkpatrick, Gelatt, & Vecchi, 1983;
Van Laarhoven & Arts, 1987). This technique involves simulating a non-homogeneous
Markov chain whose invariant distribution at iteration i is no longer equal to p(x), but to
p
i
(x) ∝ p
1/T
i
(x),
INTRODUCTION 19
Figure 8. General simulated annealing algorithm.
−10 0 10 20
0
0.1
0.2
i=100
−10 0 10 20
0
0.1
0.2
i=500
−10 0 10 20
0
0.1
0.2
i=1000
−10 0 10 20
0
0.1
0.2
i=5000
Figure 9. Discovering the modes of the target distribution with the simulated annealing algorithm.
where T
i
is a decreasing cooling schedule with lim
i →∞
T
i
= 0. The reason for doing this
is that, under weak regularity assumptions on p(x), p

(x) is a probability density that
concentrates itself on the set of global maxima of p(x). The simulated annealing involves,
therefore, just a minor modification of standard MCMCalgorithms as shown in figure 8. The
results of applying annealing to the example of the previous section are shown in figure 9.
To obtain efficient annealed algorithms, it is again important to choose suitable proposal
distributions and an appropriate cooling schedule. Many of the negative simulated annealing
20 C. ANDRIEU ET AL.
results reported in the literature often stem from poor proposal distribution design. In some
complex variable and model selection scenarios arising in machine learning, one can even
propose from complex reversible jump MCMC kernels (Section 3.7) within the annealing
algorithm (Andrieu, de Freitas, & Doucet, 2000a). If one defines a joint distribution over
the parameter and model spaces, this technique can be used to search for the best model
(according to MDL or AIC criteria) and ML parameter estimates simultaneously.
Most convergence results for simulated annealing typically state that if for a given T
i
,
the homogeneous Markov transition kernel mixes quickly enough, then convergence to the
set of global maxima of p(x) is ensured for a sequence T
i
= (C ln(i ÷T
0
))
−1
, where C and
T
0
are problem-dependent. Most of the results have been obtained for finite spaces (Geman
& Geman, 1984; Van Laarhoven & Arts, 1987) or compact continuous spaces (Haario &
Sacksman, 1991). Some results for non-compact spaces can be found in Andrieu, Breyer,
and Doucet (1999).
3.3. Mixtures and cycles of MCMC kernels
A very powerful property of MCMC is that it is possible to combine several samplers into
mixtures and cycles of the individual samplers (Tierney, 1994). If the transition kernels K
1
and K
2
have invariant distribution p(·) each, then the cycle hybrid kernel K
1
K
2
and the
mixture hybrid kernel νK
1
÷ (1 − ν)K
2
, for 0 ≤ ν ≤ 1, are also transition kernels with
invariant distribution p(·).
Mixtures of kernels can incorporate global proposals to explore vast regions of the
state space and local proposals to discover finer details of the target distribution (Andrieu,
de Freitas, & Doucet, 2000b; Andrieu & Doucet, 1999; Robert & Casella, 1999). This will
be useful, for example, when the target distribution has many narrow peaks. Here, a global
proposal locks into the peaks while a local proposal allows one to explore the space around
each peak. For example, if we require a high-precision frequency detector, one can use
the fast Fourier transform (FFT) as a global proposal and a random walk as local proposal
(Andrieu &Doucet, 1999). Similarly, in kernel regression and classification, one might want
to have a global proposal that places the bases (kernels) at the locations of the input data and
a local random walk proposal that perturbs these in order to obtain better fits (Andrieu, de
Freitas, & Doucet, 2000b). However, mixtures of kernels also play a big role in many other
samplers, including the reversible jump MCMC algorithm (Section 3.7). The pseudo-code
for a typical mixture of kernels is shown in figure 10.
Cycles allow us to split a multivariate state vector into components (blocks) that can be
updated separately. Typically the samplers will mix more quickly by blocking highly cor-
related variables. A block MCMC sampler, using b
j
to indicate the j -th block, n
b
to denote
the number of blocks and x
(i ÷1)
−[b
j
]
{x
(i ÷1)
b
1
, x
(i ÷1)
b
2
, . . . , x
(i ÷1)
b
j −1
, x
(i )
b
j ÷1
, . . . , x
(i )
b
n
b
}, is shown in
figure 11. The transition kernel for this algorithm is given by the following expression
K
MH-Cycle

x
(i ÷1)

x
(i )

=
n
b
¸
j =1
K
MH(j)

x
(i ÷1)
b
j

x
(i )
b
j
, x
(i ÷1)
−[b
j
]

where K
MH(j)
denotes the j -th MH algorithm in the cycle.
INTRODUCTION 21
Figure 10. Typical mixture of MCMC kernels.
Figure 11. Cycle of MCMC kernels—block MH algorithm.
Obviously, choosing the size of the blocks poses some trade-offs. If one samples the
components of a multi-dimensional vector one-at-a-time, the chain may take a very long
time to explore the target distribution. This problem gets worse as the correlation between
the components increases. Alternatively, if one samples all the components together, then
the probability of accepting this large move tends to be very low.
A popular cycle of MH kernels, known as Gibbs sampling (Geman & Geman, 1984), is
obtainedwhenwe adopt the full conditional distributions p(x
j
[ x
−j
) = p(x
j
[ x
1
, . . . , x
j −1
,
x
j ÷1
, . . . , x
n
) as proposal distributions (for notational simplicity, we have replaced the index
notation b
j
with j ). The following section describes it in more detail.
3.4. The Gibbs sampler
Suppose we have an n-dimensional vector x and the expressions for the full conditionals
p(x
j
[ x
1
, . . . , x
j −1
, x
j ÷1
, . . . , x
n
). In this case, it is often advantageous to use the following
22 C. ANDRIEU ET AL.
proposal distribution for j = 1, . . . , n
q

x

x
(i )

=
¸
p

x

j

x
(i )
−j

If x

−j
= x
(i )
−j
0 Otherwise.
The corresponding acceptance probability is:
A

x
(i )
, x

= min
¸
1,
p(x

)q

x
(i )

x

p

x
(i )

q

x

[x
(i )

¸
= min
¸
1,
p(x

) p

x
(i )
j

x
(i )
−j

p

x
(i )

p(x

j
[x

−j
)
¸
= min
¸
1,
p

x

−j

p

x
(i )
−j

¸
= 1.
That is, the acceptance probability for each proposal is one and, hence, the deterministic
scan Gibbs sampler algorithm is often presented as shown in figure 12.
Since the Gibbs sampler can be viewed as a special case of the MH algorithm, it is
possible to introduce MH steps into the Gibbs sampler. That is, when the full conditionals
are available and belong to the family of standard distributions (Gamma, Gaussian, etc.),
we will draw the new samples directly. Otherwise, we can draw samples with MH steps
embedded within the Gibbs algorithm. For n =2, the Gibbs sampler is also known as the
data augmentation algorithm, which is closely related to the expectation maximisation (EM)
algorithm (Dempster, Laird, & Rubin, 1977; Tanner & Wong, 1987).
Directed acyclic graphs (DAGS) are one of the best known application areas for Gibbs
sampling(Pearl, 1987). Here, a large-dimensional joint distributionis factoredintoa directed
graph that encodes the conditional independencies in the model. In particular, if x
pa( j )
Figure 12. Gibbs sampler.
INTRODUCTION 23
denotes the parent nodes of node x
j
, we have
p(x) =
¸
j
p

x
j

x
pa( j )

.
It follows that the full conditionals simplify as follows
p

x
j

x
−j

= p

x
j

x
pa( j )

¸
k∈ch( j )
p

x
k

x
pa(k)

where ch( j ) denotes the children nodes of x
j
. That is, we only need to take into account
the parents, the children and the children’s parents. This set of variables is known as the
Markov blanket of x
j
. This technique forms the basis of the popular software package for
Bayesian updating with Gibbs sampling (BUGS) (Gilks, Thomas, & Spiegelhalter, 1994).
Sampling from the full conditionals, with the Gibbs sampler, lends itself naturally to the
construction of general purpose MCMCsoftware. It is sometimes convenient to block some
of the variables to improve mixing (Jensen, Kong, & Kjærulff, 1995; Wilkinson & Yeung,
2002).
3.5. Monte Carlo EM
The EM algorithm (Baum et al., 1970; Dempster, Laird, & Rubin, 1977) is a standard
algorithm for ML and MAP point estimation. If X contains visible and hidden variables
x = {x
v
, x
h
}, then a local maximum of the likelihood p(x
v
[ θ) given the parameters θ can
be found by iterating the following two steps:
1. E step. Compute the expected value of the complete log-likelihood function with respect
to the distribution of the hidden variables
Q(θ) =

X
h
log( p(x
h
, x
v
[ θ)) p

x
h

x
v
, θ
(old)

dx
h
,
where θ
(old)
refers to the value of the parameters at the previous time step.
2. M step. Perform the following maximisation θ
(new)
= arg max
θ
Q(θ).
In many practical situations, the expectation in the Estep is either a sumwith an exponen-
tially large number of summands or an intractable integral (Ghahramani, 1995; Ghahramani
&Jordan, 1995; McCulloch, 1994; Pasula et al., 1999; Utsugi, 2001); see also Dellaert et al.
(this issue). A solution is to introduce MCMC to sample from p(x
h
[ x
v
, θ
(old)
) and replace
the expectation in the E step with a small sum over the samples, as shown in figure 13.
Convergence of this algorithmis discussed in Sherman, Ho, and Dalal (1999), while Levine
and Casella (2001) is a good recent review.
To improve the convergence behaviour of EM, namely to escape low local minima and
saddle points, various authors have proposed stochastic approaches that rely on sampling
from p(x
h
[ x
v
, θ
(old)
) in the E step and then performing the M step using these samples.
24 C. ANDRIEU ET AL.
Figure 13. MCMC-EM algorithm.
The method is known as stochastic EM (SEM) when we draw only one sample (Celeux
& Diebolt, 1985) and Monte Carlo EM (MCEM) when several samples are drawn (Wei
& Tanner, 1990). There are several annealed variants (such as SAEM) that become more
deterministic as the number of iterations increases (Celeux & Diebolt, 1992). The are also
very efficient algorithms for marginal MAPestimation (SAME) (Doucet, Godsill, &Robert,
2000). One wishes sometimes that Metropolis had succeeded in stopping the proliferation
of acronyms!
3.6. Auxiliary variable samplers
It is often easier to sample from an augmented distribution p(x, u), where u is an auxiliary
variable, than from p(x). Then, it is possible to obtain marginal samples x
(i )
by sampling
(x
(i )
, u
(i )
) according to p(x, u) and, subsequently, ignoring the samples u
(i )
. This very useful
idea was proposed in the physics literature (Swendsen & Wang, 1987). Here, we will focus
on two well-known examples of auxiliary variable methods, namely hybrid Monte Carlo
and slice sampling.
3.6.1. Hybrid Monte Carlo. Hybrid Monte Carlo (HMC) is an MCMC algorithm that
incorporates information about the gradient of the target distribution to improve mixing
in high dimensions. We describe here the “leapfrog” HMC algorithm outlined in Duane
et al. (1987) and Neal (1996) focusing on the algorithmic details and not on the statistical
mechanics motivation. Assume that p(x) is differentiable and everywhere strictly positive.
At each iteration of the HMC algorithm, one takes a predetermined number (L) of deter-
ministic steps using information about the gradient of p(x). To explain this in more detail,
we first need to introduce a set of auxiliary “momentum” variables u ∈ R
n
x
and define the
INTRODUCTION 25
Figure 14. Hybrid Monte Carlo.
extended target density
p(x, u) = p(x)N

u; 0, I
n
x

.
Next, we need to introduce the n
x
-dimensional gradient vector (x) ∂ log p(x)/∂x and
a fixed step-size parameter ρ > 0.
In the HMC algorithm, we draw a new sample according to p(x, u) by starting with
the previous value of x and generating a Gaussian random variable u. We then take L
“frog leaps” in u and x. The values of u and x at the last leap are the proposal candidates
in the MH algorithm with target density p(x, u). Marginal samples from p(x) are ob-
tained by simply ignoring u. Given (x
(i −1)
, u
(i −1)
), the algorithm proceeds as illustrated in
figure 14.
When only one deterministic step is used, i.e. L = 1, one obtains the Langevin algorithm,
which is a discrete time approximation of a Langevin diffusion process. The Langevin
algorithm is a special case of MH where the candidate satisfies
x

= x
0
÷ρu
0
= x
(i −1)
÷ρ

u

÷ρ

x
(i −1)

/2

with u

∼ N(0, I
n
x
).
The choice of the parameters L and ρ poses simulation tradeoffs. Large values of ρ
result in low acceptance rates, while small values require many leapfrog steps (expensive
computations of the gradient) to move between two nearby states. Choosing L is equally
problematic as we want it to be large to generate candidates far from the initial state, but
this can result in many expensive computations. HMC, therefore, requires careful tuning of
the proposal distribution. It is more efficient, in practice, to allow a different step size ρ for
each of the coordinates of x (Ishwaran, 1999).
26 C. ANDRIEU ET AL.
3.6.2. The slice sampler. The slice sampler (Damien, Wakefield, &Walker, 1999; Higdon,
1998; Wakefield, Gelfand, & Smith, 1991) is a general version of the Gibbs sampler. The
basic idea of the slice sampler is to introduce an auxiliary variable u ∈ R and construct an
extended target distribution p

(x, u), such that
p

(x, u) =
¸
1 if 0 ≤ u ≤ p(x)
0 otherwise.
It is then straightforward to check that

p

(x, u) du =

p(x)
0
du = p(x).
Hence, to sample from p(x) one can sample from p

(x, u) and then ignore u. The full
conditionals are of this augmented model are
p(u [ x) = U
[0, p(x)]
(u)
p(x [ u) = U
A
(x)
where A = {x; p(x) ≥ u}. If A is easy to identify then the algorithm is straightforward to
implement, as shown in figure 15.
It can be difficult to identify A. It is then worth introducing several auxiliary variables
(Damien, Wakefield, & Walker, 1999; Higdon, 1998). For example assume that
p(x) ∝
L
¸
l=1
f
l
(x),
where the f
l
(·)’s are positive functions, not necessarilydensities. Let us introduce L auxiliary
variables (u
1
, . . . , u
L
) and define
p

(x, u
1
, . . . , u
L
) ∝
L
¸
l=1
I
[0, f
l
(x)]
(u
l
).
x x
x u
(i+1)
(i)
(i+1)
f(x )
(i)
Figure 15. Slice sampling: given a previous sample, we sample a uniform variable u
(i ÷1)
between 0 and f (x
(i )
).
One then samples x
(i ÷1)
in the interval where f (x) ≥ u
(i ÷1)
.
INTRODUCTION 27
Figure 16. Slice sampler.
Then one can also check that

p

(x, u
1
, . . . , u
L
) du
1
. . . du
L
= p(x) as

p

(x, u
1
, . . . , u
L
)du
1
. . . du
L

L
¸
l=1
I
[0, f
l
(x)]
(u
l
) du
1
. . . du
L
=
L
¸
l=1
f
l
(x).
The slice sampler to sample from p

(x, u
1
, . . . , u
L
) proceeds as shown in figure 16. Al-
gorithmic improvements and convergence results are presented in Mira (1999) and Neal
(2000).
3.7. Reversible jump MCMC
In this section, we attack the more complex problem of model selection. Typical exam-
ples include estimating the number of neurons in a neural network (Andrieu, de Freitas,
& Doucet, 2001a; Holmes & Mallick, 1998; Rios Insua & M¨ uller, 1998), the number of
splines in a multivariate adaptive splines regression (MARS) model (Holmes & Denison,
this issue), the number of sinusoids in a noisy signal (Andrieu &Doucet, 1999), the number
of lags in an autoregressive process (Troughton & Godsill, 1998), the number of com-
ponents in a mixture (Richardson & Green, 1997), the number of levels in a change-
point process (Green, 1995), the number of components in a mixture of factor analy-
sers (Fokou´ e & Titterington, this issue), the appropriate structure of a graphical model
(Friedman & Koller, 2001; Giudici & Castelo, this issue) or the best set of input variables
(Lee, this issue).
Given a family of M models {M
m
; m = 1, . . . , N}, we will focus on constructing ergodic
Markov chains admitting p(m, x
m
) as the invariant distribution. For simplicity, we avoid
the treatment of nonparametric model averaging techniques; see for example (Escobar &
West, 1995; Green & Richardson, 2000).
Up to this section, we have been comparing densities in the acceptance ratio. However,
if we are carrying out model selection, then comparing the densities of objects in different
dimensions has no meaning. It is like trying to compare spheres with circles. Instead, we
have to be more formal and compare distributions P(dx) = Pr(x ∈ dx) under a common
measure of volume. The distribution P(dx) will be assumed to admit a density p(x) with
respect to a measure of interest, e.g. Lebesgue in the continuous case: P( dx) = p(x) dx.
The acceptance ratio will nowinclude the ratio of the densities and the ratio of the measures
(Radon Nikodym derivative). The latter gives rise to a Jacobian term. To compare densities
point-wise, we need, therefore, to map the two models to a common dimension as illustrated
in figure 17.
28 C. ANDRIEU ET AL.
1
1
*
*
2
1
1
Uniformly expanded density
x
Compare both
densities
point-wise
uniformly
*
2
1
Bivariate density
Univariate density
p(x ,x )
x
1
x
p(x ,x )
Proposex
x
p(x ) x
Figure 17. To compare a 1D model against a 2D model, we first have to map the first model so that both models
have common measure (area in this case).
The parameters x
m
∈ X
m
(e.g. X
m
= R
n
m
) are model dependent. Hence, to find the right
model and parameters we could sample over the model indicator and the product space
¸
M
m=1
X
m
(Carlin & Chib, 1995). Recently, Green introduced a strategy that avoids this
expensive search over the full product space (Green, 1995). In particular one samples on a
much smaller union space X

M
m=1
{m} X
m
. The full target distribution defined in this
space is given by
p(k, dx) =
M
¸
m=1
p(m, dx
m
)I
{m}X
m
(k, x).
That is, the probability of k being equal to m and x belonging to an infinitesimal set centred
around x
m
is p(m, dx
m
). By marginalisation, we obtain the probability of being in subspace
X
m
.
Green’s method allows the sampler to jump between the different subspaces. To ensure
a common measure, it requires the extension of each pair of communicating spaces, X
m
and X
n
, to
¯
X
m,n
X
m
U
m,n
and
¯
X
n,m
X
n
U
n,m
. It also requires the definition of a
deterministic, differentiable, invertible dimension matching function f
n→m
between
¯
X
m,n
and
¯
X
n,m
,
(x
m
, u
m,n
) = f
n→m
(x
n
, u
n,m
) =

f
x
n→m
(x
n
, u
n,m
), f
u
n→m
(x
n
, u
n,m
)

.
INTRODUCTION 29
We define f
m→n
such that f
m→n
( f
n→m
(x
n
, u
n,m
)) = (x
n
, u
n,m
). The choice of the extended
spaces, deterministic transformation f
m→n
and proposal distributions for q
n→m
(· [ n, x
n
) and
q
m→n
(· [ m, x
m
) is problem dependent and needs to be addressed on a case by case basis.
If the current state of the chain is (n, x
n
), we move to (m, x
m
) by generating u
n,m

q
n→m
(· [ n, x
n
), ensuring that we have reversibility (x
m
, u
m,n
) = f
n→m
(x
n
, u
n,m
), and ac-
cepting the move according to the probability ratio
A
n→m
= min
¸
1,
p(m, x

m
)
p(n, x
n
)

q(n [ m)
q(m [ n)

q
m→n
(u
m,n
[ m, x

m
)
q
n→m
(u
n,m
[ n, x
n
)
J
f
n→m
¸
,
where x

m
= f
x
n→m
(x
n
, u
n,m
) and J
f
n→m
is the Jacobian of the transformation f
n→m
(when
only continuous variables are involved in the transformation)
J
f
m→n
=

det
∂ f
n→m
(x
m
, u
m,n
)
∂(x
m
, u
m,n
)

.
To illustrate this, assume that we are concerned with sampling the locations µ and number
k of components of a mixture. For example we might want to estimate the locations and
number of basis functions in kernel regression and classification, the number of mixture
components in a finite mixture model, or the location and number of segments in a segmen-
tation problem. Here, we could define a merge move that combines two nearby components
and a split move that breaks a component into two nearby ones. The merge move involves
randomly selecting a component (µ
1
) and then combining it with its closest neighbour (µ
2
)
into a single component µ, whose new location is
µ =
µ
1
÷µ
2
2
The corresponding split move that guarantees reversibility, involves splitting a randomly
chosen component as follows
µ
1
= µ −u
n,m
β
µ
2
= µ ÷u
n,m
β
where β is a simulation parameter and, for example, u
n,m
∼ U
[0,1]
. Note that to ensure
reversibility, we only perform the merge move if |µ
1
− µ
2
| < 2β. The acceptance ratio
for the split move is
A
split
= min
¸
1,
p(k ÷1, µ
k÷1
)
p(k, µ
k
)

1
k÷1
1
k

1
p(u
n,m
)
J
split
¸
,
where 1/k denotes the probability of choosing, uniformly at random, one of the k compo-
nents. The Jacobian is
J
split
=

∂(µ
1
, µ
2
)
∂(µ, u
n,m
)

=

1 1
−β β

= 2β.
30 C. ANDRIEU ET AL.
Figure 18. Generic reversible jump MCMC.
Similarly, for the merge move, we have
A
merge
= min
¸
1,
p(k −1, µ
k−1
)
p(k, µ
k
)

1
k−1
1
k
J
merge
¸
,
where J
merge
= 1/2β.
Reversible jump is a mixture of MCMC kernels (moves). In addition, to the split and
merge moves, we couldhave other moves suchas birthof a component, deathof a component
and a simple update of the locations. The various moves are carried out according to the
mixture probabilities (b
k
, d
k
, m
k
, s
k
, u
k
), as shown in figure 18. In fact, it is the flexibility
of including so many possible moves that can make reversible jump a more powerful
model selection strategy than schemes based on model selection using a mixture indicator
or diffusion processes using only birth and death moves (Stephens, 1997). However, the
problem with reversible jump MCMC is that engineering reversible moves is a very tricky,
time-consuming task.
4. The MCMC frontiers
4.1. Convergence and perfect sampling
Determining the length of the Markov chain is a difficult task. In practice, one often dis-
cards an initial set of samples (burn-in) to avoid starting biases. In addition, one can ap-
ply several graphical and statistical tests to assess, roughly, if the chain has stabilised
(Robert & Casella, 1999, ch. 8). In general, none of these tests provide entirely satisfactory
diagnostics.
Several theoreticians have tried to bound the mixing time; that is, the minimum number
of steps required for the distribution of the Markov chain K to be close to the target p(x).
(Here, we present a, by no means exhaustive, summary of some of the available results.) If
INTRODUCTION 31
we measure closeness with the total variation norm
x
(t ), where

x
(t ) =
¸
¸
K
(t )
(· [ x) − p(·)
¸
¸
=
1
2

K
(t )
(y [ x) − p(y)

dy,
then the mixing time is
τ
x
() = min{t :
x
(t
/
) ≤ for all t
/
≥ t }.
If the state space X is finite and reversibility holds true, then the transition operator
K (K f (x) =
¸
K(y [ x) f (y)) is self adjoint on L
2
( p). That is,
!K f [ g) = ! f [ Kg),
where f and g are real functions and we have used the bra-ket notation for the inner product
! f [ g) =
¸
f (x)g(x) p(x). This implies that K has real eigenvalues
1 = λ
1
> λ
2
≥ λ
3
≥ · · · ≥ λ
[X[
> −1
and an orthonormal basis of real eigenfunctions f
i
, such that K f
i
= λ
i
f
i
. This spectral
decomposition and the Cauchy-Schwartz inequality allow us to obtain a bound on the total
variation norm

x
(t ) ≤
1
2

p(x)
λ
t

,
where λ

= max(λ
2
, [λ
[X[
[) (Diaconis &Saloff-Coste, 1998; Jerrum&Sinclair, 1996). This
classical result give us a geometric convergence rate in terms of eigenvalues. Geometric
bounds have also been obtained in general state spaces using the tools of regeneration and
Lyapunov-Foster conditions (Meyn & Tweedie, 1993).
The next logical step is to bound the second eigenvalue. There are several inequalities
(Cheeger, Poincar´ e, Nash) from differential geometry that allows us to obtain these bounds
(Diaconis &Saloff-Coste, 1998). For example, one could use Cheeger’s inequality to obtain
the following bound
1 −2 ≤ λ
2
≤ 1 −

2
2
,
where is the conductance of the Markov chain
= min
0<p(S)<1/2;S⊂X
¸
x∈S,y∈S
c p(x)K(y [ x)
p(S)
Intuitively, one can interpret this quantity as the readiness of the chain to escape from any
small region S of the state space and, hence, make rapid progress towards equilibrium
(Jerrum & Sinclair, 1996).
32 C. ANDRIEU ET AL.
These mathematical tools have been applied to show that simple MCMC algorithms
(mostly Metropolis) run in time that is polynomial in the dimension d of the state space,
thereby escaping the exponential curse of dimensionality. Polynomial time sampling algo-
rithms have been obtained in the following important scenarios:
1. Computing the volume of a convex body in d dimensions, where d is large (Dyer, Frieze,
& Kannan, 1991).
2. Sampling from log-concave distributions (Applegate & Kannan, 1991).
3. Sampling from truncated multivariate Gaussians (Kannan & Li, 1996).
4. Computing the permanent of a matrix (Jerrum, Sinclair, & Vigoda, 2000).
The last problem is equivalent to sampling matchings from a bipartite graph; a problem
that manifests itself in many ways in machine learning (e.g., stereo matching and data
association).
Although the theoretical results are still far from the practice of MCMC, they will even-
tually provide better guidelines on how to design and choose algorithms. Already, some
results tell us, for example, that it is not wise to use the independent Metropolis sampler in
high dimensions (Mengersen & Tweedie, 1996).
A remarkable recent breakthrough was the development of algorithms for perfect sam-
pling. These algorithms are guaranteed to give us an independent sample from p(x) under
certain restrictions. The two major players are coupling from the past (Propp & Wilson,
1998) and Fill’s algorithm (Fill, 1998). From a practical point of view, these algorithms are
still limited and, in many cases, computationally inefficient. However, some steps are being
taken towards obtaining more general perfect samplers; for example perfect slice samplers
(Casella et al., 1999).
4.2. Adaptive MCMC
If we look at the chain on the top right of figure 7, we notice that the chain stays at each state
for a long time. This tells us that we should reduce the variance of the proposal distribution.
Ideally, one wouldlike toautomate this process of choosingthe proposal distributionas much
as possible. That is, one shoulduse the informationinthe samples toupdate the parameters of
the proposal distribution so as to obtain a distribution that is either closer to the target distri-
bution, that ensures a suitable acceptance rate, or that minimises the variance of the estimator
of interest. However, one should not allowadaptation to take place infinitely often in a naive
waybecause this candisturbthe stationarydistribution. This problemarises because byusing
the past information infinitely often, we violate the Markov property of the transition kernel.
That is, p(x
(i )
[ x
(0)
, x
(1)
, . . . , x
(i −1)
) no longer simplifies to p(x
(i )
[ x
(i −1)
). In particular,
Gelfand and Sahu (1994) present a pathological example, where the stationary distribution is
disturbed despite the fact that each participating kernel has the same stationary distribution.
To avoid this problem, one could carry out adaptation only during an initial fixed number
of steps, andthenuse standardMCMCsimulationtoensure convergence tothe right distribu-
tion. Two methods for doing this are presented in Gelfand and Sahu (1994). The first is based
on the idea of running several chains in parallel and using sampling-importance resampling
INTRODUCTION 33
(Rubin, 1988) to multiply the kernels that are doing well and suppress the others. In this
approach, one uses an approximation to the marginal density of the chain as proposal. The
secondmethodsimplyinvolves monitoringthe transitionkernel andchangingone of its com-
ponents (for example the proposal distribution) so as to improve mixing. A similar method
that guarantees a particular acceptance rate is discussed in Browne and Draper (2000).
There are, however, a fewadaptive MCMCmethods that allowone to performadaptation
continuously without disturbing the Markov property, including delayed rejection (Tierney
& Mira, 1999), parallel chains (Gilks & Roberts, 1996) and regeneration (Gilks, Roberts, &
Sahu, 1998; Mykland, Tierney, & Yu, 1995). These methods are, unfortunately, inefficient
in many ways and much more research is required in this exciting area.
4.3. Sequential Monte Carlo and particle filters
Sequential Monte Carlo (SMC) methods allow us to carry out on-line approximation of
probability distributions using samples (particles). They are very useful in scenarios involv-
ing real-time signal processing, where data arrival is inherently sequential. Furthermore,
one might wish to adopt a sequential processing strategy to deal with non-stationarity in
signals, so that information fromthe recent past is given greater weighting than information
from the distant past. Computational simplicity in the form of not having to store all the
data might also constitute an additional motivating factor for these methods.
In the SMC setting, we assume that we have an initial distribution, a dynamic model and
measurement model
p(x
0
)
p(x
t
[ x
0:t −1
, y
1:t −1
) for t ≥ 1
p(y
t
[ x
0:t
, y
1:t −1
) for t ≥ 1
We denote by x
0:t
{x
0
, . . . , x
t
} and y
1:t
{y
1
, . . . , y
t
}, respectively, the states and the ob-
servations up to time t . Note that we could assume Markov transitions and conditional inde-
pendence to simplify the model; p(x
t
[ x
0:t −1
, y
1:t −1
) = p(x
t
[ x
t −1
) and p(y
t
[ x
0:t
, y
1:t −1
) =
p(y
t
[ x
t
). However, this assumption is not necessary in the SMC framework.
Our aim is to estimate recursively in time the posterior p(x
0:t
[ y
1:t
) and its associated
features including the marginal distribution p(x
t
[ y
1:t
), known as the filtering distribution,
and the expectations
I ( f
t
) = E
p(x
0:t
[y
1:t
)
[ f
t
(x
0:t
)]
A generic SMC algorithm is depicted in figure 19. Given N particles {x
(i )
0:t −1
}
N
i =1
at
time t − 1, approximately distributed according to the distribution p(x
0:t −1
[ y
1:t −1
), SMC
methods allow us to compute N particles {x
(i )
0:t
}
N
i =1
approximately distributed according to
the posterior p(x
0:t
[y
1:t
), at time t . Since we cannot sample from the posterior directly,
the SMC update is accomplished by introducing an appropriate importance proposal dis-
tribution q(x
0:t
) from which we can obtain samples. The samples are then appropriately
weighted.
34 C. ANDRIEU ET AL.
Figure 19. In this example, the bootstrap filter starts at time t − 1 with an unweighted measure {˜ x
(i )
t −1
, N
−1
},
which provides an approximation of p(x
t −1
[ y
1:t −2
). For each particle we compute the importance weights using
the information at time t −1. This results in the weighted measure {˜ x
(i )
t −1
, ˜ w
(i )
t −1
}, which yields an approximation
p(x
t −1
[ y
1:t −1
). Subsequently, the resampling step selects only the “fittest” particles to obtain the unweighted
measure {x
(i )
t −1
, N
−1
}, which is still an approximation of p(x
t −1
[ y
1:t −1
). Finally, the sampling (prediction) step
introduces variety, resulting in the measure {˜ x
(i )
t
, N
−1
}, which is an approximation of p(x
t
[ y
1:t −1
).
Figure 20. Simple SMC algorithm at time t . For filtering purposes, there is no need for storing or resampling
the past trajectories.
INTRODUCTION 35
In generic SMC simulation, one needs to extend the current paths {x
(i )
0:t −1
}
N
i =1
to obtain
new paths { ˜ x
(i )
0:t
}
N
i =1
using the proposal distribution q( ˜ x
0:t
[y
1:t
) given by
q( ˜ x
0:t
[ y
1:t
)} =

q( ˜ x
0:t
[ x
0:t −1
, y
1:t
) p(x
0:t −1
[ y
1:t −1
) dx
0:t −1
.
To make this integral tractable, we only propose to modify the particles at time t , and leave
the past trajectories intact. Consequently
q( ˜ x
0:t
[ y
1:t
) = p(x
0:t −1
[ y
1:t −1
)q( ˜ x
t
[ x
0:t −1
, y
1:t
)
The samples from q(·), must be weighted by the importance weights
w
t
=
p( ˜ x
0:t
[ y
1:t
)
q( ˜ x
0:t
[ y
1:t
)
=
p(x
0:t −1
[ y
1:t
)
p(x
0:t −1
[ y
1:t −1
)
p( ˜ x
t
[ x
0:t −1
, y
1:t
)
q( ˜ x
t
[ x
0:t −1
, y
1:t
)

p (y
t
[ ˜ x
t
) p ( ˜ x
t
[ x
0:t −1
, y
1:t −1
)
q
t
( ˜ x
t
[ x
0:t −1
, y
1:t
)
. (22)
From Eq. (22), we note that the optimal importance distribution is
q( ˜ x
t
[ x
0:t −1
, y
1:t
) = p( ˜ x
t
[ x
0:t −1
, y
1:t
).
(When using this proposal, one might still encounter difficulties if the ratio of the first two
terms of Eq. (22) differs significantly from 1 (Andrieu, Doucet, & Punskaya, 2001; Pitt &
Shephard, 1999).) The optimal importance distribution can be difficult to evaluate. One can
adopt, instead, the transition prior as proposal distribution
q( ˜ x
t
[ x
0:t −1
, y
1:t
) = p ( ˜ x
t
[ x
0:t −1
, y
1:t −1
)
in which case the importance weights are given by the likelihood function
w
t
∝ p (y
t
[ ˜ x
t
) .
This simplified version of SMC has appeared under many names, including condensation
(Isard & Blake, 1996), survival of the fittest (Kanazawa, Koller, & Russell, 1995) and the
bootstrap filter (Gordon, Salmond, & Smith, 1993). The importance sampling framework
allows us todesignmore principledand“clever” proposal distributions. For instance, one can
adopt suboptimal filters and other approximation methods that make use of the information
available at time t togenerate the proposal distribution(Doucet, Godsill, &Andrieu, 2000; de
Freitas et al., 2000; Pitt & Shephard, 1999; van der Merwe et al., 2000). In fact, in some
restricted situations, one may interpret the likelihood as a distribution in terms of the states
and sample from it directly. In doing so, the importance weights become equal to the
transition prior (Fox et al., 2001).
After the importance sampling step, a selection scheme associates to each particle ˜ x
(i )
0:t
a number of “children”, say N
i
∈ N, such that
¸
N
i =1
N
i
= N. This selection step is what
36 C. ANDRIEU ET AL.
allows us to track moving target distributions efficiently by choosing the fittest particles.
There are various selection schemes in the literature, but their performance varies in terms
of var[N
i
] (Doucet, de Freitas, & Gordon, 2001).
An important feature of the selection routine is that its interface only depends on particle
indices and weights. That is, it can be treated as a black-box routine that does not require
any knowledge of what a particle represents (e.g., variables, parameters, models). This
enables one to implement variable and model selection schemes straightforwardly. The
simplicity of the coding of complex models is, indeed, one of the major advantages of these
algorithms.
It is also possible to introduce MCMC steps of invariant distribution p(x
0:t
[ y
1:t
) on each
particle (Andrieu, de Freitas, &Doucet, 1999; Gilks &Berzuini, 1998; MacEachern, Clyde,
& Liu, 1999). The basic idea is that if the particles are distributed according to the poste-
rior distribution p(x
0:t
[ y
1:t
), then applying a Markov chain transition kernel K(x

0:t
[ x
0:t
),
with invariant distribution p(· [ y
1:t
) such that

K(x

0:t
[ x
0:t
) p(x
0:t
[ y
1:t
) = p(x

0:t
[ y
1:t
), still
results in a set of particles distributed according to the posterior of interest. However, the
new particles might have been moved to more interesting areas of the state-space. In fact,
by applying a Markov transition kernel, the total variation of the current distribution with
respect to the invariant distribution can only decrease. Note that we can incorporate any
of the standard MCMC methods, such as the Gibbs sampler, MH algorithm and reversible
jump MCMC, into the filtering framework, but we no longer require the kernel to be
ergodic.
4.4. The machine learning frontier
The machine learningfrontier is characterisedbylarge dimensional models, massive datasets
and many and varied applications. Massive datasets pose no problem in the SMC context.
However, in batch MCMC simulation it is often not possible to load the entire dataset
into memory. A few solutions based on importance sampling have been proposed recently
(Ridgeway, 1999), but there is still great room for innovation in this area.
Despite the auspicious polynomial bounds on the mixing time, it is an arduous task
to design efficient samplers in high dimensions. The combination of sampling algorithms
with either gradient optimisation or exact methods has proved to be very useful. Gradient
optimisation is inherent to Langevin algorithms and hybrid Monte Carlo. These algorithms
have been shown to work with large dimensional models such as neural networks (Neal,
1996) and Gaussian processes (Barber & Williams, 1997). Information about derivatives of
the target distribution also forms an integral part of many adaptive schemes, as discussed
in Section 2.3. Recently, it has been argued that the combination of MCMC and variational
optimisation techniques can also lead to more efficient sampling (de Freitas et al., 2001).
The combination of exact inference with sampling methods within the framework of Rao-
Blackwellisation (Casella & Robert, 1996) can also result in great improvements. Suppose
we can divide the hidden variables x into two groups, u and v, such that p(x) = p(v [ u) p(u)
and, conditional on u, the conditional posterior distribution p(v [ u) is analytically tractable.
Then we can easily marginalise out v fromthe posterior, and only need to focus on sampling
from p(u), which lies in a space of reduced dimension. That is, we sample u
(i )
∼ p(u) and
INTRODUCTION 37
then use exact inference to compute
p(v) =
1
N
N
¸
i =1
p

v

u
(i )

By identifying “troublesome” variables and sampling them, the rest of the problem can
often be solved easily using exact algorithms such as Kalman filters, HMMs or junction
trees. For example, one can apply this technique to sample variables that eliminate loops in
graphical models and then compute the remaining variables with efficient analytical algo-
rithms (Jensen, Kong, &Kjærulff, 1995; Wilkinson &Yeung, 2002). Other application areas
include dynamic Bayesian networks (Doucet et al., 2000), conditionally Gaussian models
(Carter & Kohn, 1994; De Jong & Shephard, 1995; Doucet, 1998) and model averaging
for graphical models (Friedman & Koller, this issue). The problem of how to automatically
identify which variables should be sampled, and which can be handled analytically is still
open. An interesting development is the augmentation of high dimensional models with
low dimensional artificial variables. By sampling only the artificial variables, the original
model decouples into simpler, more tractable submodels (Albert &Chib, 1993; Andrieu, de
Freitas, & Doucet, 2001b; Wood & Kohn, 1998); see also Holmes and Denison (this issue).
This strategy allows one to map probabilistic classification problems to simpler regression
problems.
The design of efficient sampling methods most of the times hinges on awareness of
the basic building blocks of MCMC (mixtures of kernels, augmentation strategies and
blocking) and on careful design of the proposal mechanisms. The latter requires domain
specific knowledge and heuristics. There are great opportunities for combining existing
sub-optimal algorithms with MCMC in many machine learning problems. Some areas that
are already benefiting from sampling methods include:
1. Computer vision. Tracking (Isard & Blake, 1996; Ormoneit, Lemieux, & Fleet, 2001),
stereomatching(Dellaert et al., this issue), colour constancy(Forsyth, 1999), restoration
of oldmovies (Morris, Fitzgerald, &Kokaram, 1996) andsegmentation(Clark&Quinn,
1999; Kam, 2000; Tu & Zhu, 2001).
2. Web statistics. Estimating coverage of search engines, proportions belonging to specific
domains and the average size of web pages (Bar-Yossef et al., 2000).
3. Speech and audio processing. Signal enhancement (Godsill & Rayner, 1998; Vermaak
et al., 1999).
4. Probabilistic graphical models. For example (Gilks, Thomas, & Spiegelhalter, 1994;
Wilkinson & Yeung, 2002) and several papers in this issue.
5. Regression and classification. Neural networks and kernel machines (Andrieu, de
Freitas, & Doucet, 2001a; Holmes & Mallick, 1998; Neal, 1996; M¨ uller & Rios
Insua, 1998), Gaussian processes (Barber &Williams, 1997), CART(Denison, Mallick,
& Smith, 1998) and MARS (Holmes & Denison, this issue).
6. Computer graphics. Light transport (Veach & Guibas, 1997) and sampling plausible
solutions to multi-body constraint problems (Chenney & Forsyth, 2000).
38 C. ANDRIEU ET AL.
7. Data association. Vehicle matching in highway systems (Pasula et al., 1999) and mul-
titarget tracking (Bergman, 1999).
8. Decision theory. Partially observable Markov decision Processes (POMDPs) (Thrun,
2000; Salmond & Gordon, 2001), abstract Markov policies (Bui, Venkatesh, & West,
1999) and influence diagrams (Bielza, M¨ uller, & Rios Insua, 1999).
9. First order probabilistic logic. (Pasula & Russell, 2001).
10. Genetics and molecular biology. DNA microarray data (West et al., 2001), cancer gene
mapping (Newton & Lee, 2000), protein alignment (Neuwald et al., 1997) and linkage
analysis (Jensen, Kong, & Kjærulff, 1995).
11. Robotics. Robot localisation and map building (Fox et al., 2001).
12. Classical mixture models. Mixtures of independent factor analysers (Utsugi, 2001) and
mixtures of factor analysers (Fokou´ e & Titterington, this issue).
We hope that this review will be a useful resource to people wishing to carry out further
research at the interface between MCMC and machine learning. For conciseness, we have
skipped many interesting ideas, including tempering and coupling. For more details, we
advise the readers to consult the references at the end of this paper.
Acknowledgments
We would like to thank Robin Morris, Kevin Murphy, Mark Paskin, Sekhar Tatikonda and
Mike Titterington.
References
Al-Qaq, W. A., Devetsikiotis, M., & Townsend, J. K. (1995). Stochastic gradient optimization of importance sam-
pling for the efficient simulation of digital communication systems. IEEE Transactions on Communications,
43:12, 2975–2985.
Albert, J., & Chib, S. (1993). Bayesian analysis of binary and polychotomous response data. Journal of the
American Statistical Association, 88:422, 669–679.
Anderson, H. L. (1986). Metropolis, Monte Carlo, and the MANIAC. Los Alamos Science, 14, 96–108.
Andrieu, C., & Doucet, A. (1999). Joint Bayesian detection and estimation of noisy sinusoids via reversible jump
MCMC. IEEE Transactions on Signal Processing, 47:10, 2667–2676.
Andrieu, C., Breyer, L. A., & Doucet, A. (1999). Convergence of simulated annealing using Foster-Lyapunov
criteria. Technical Report CUED/F-INFENG/TR 346, Cambridge University Engineering Department.
Andrieu, C., de Freitas, N., &Doucet, A. (1999). Sequential MCMCfor Bayesian model selection. In IEEEHigher
Order Statistics Workshop, Caesarea, Israel (pp. 130–134).
Andrieu, C., de Freitas, N., & Doucet, A. (2000a). Reversible jump MCMC simulated annealing for neural
networks. In Uncertainty in artificial intelligence (pp. 11–18). San Mateo, CA: Morgan Kaufmann.
Andrieu, C., de Freitas, N., & Doucet, A. (2000b). Robust full Bayesian methods for neural networks. In S. A.
Solla, T. K. Leen, &K.-R. M¨ uller (Eds.), Advances in neural information processing systems 12 (pp. 379–385).
MIT Press.
Andrieu, C., de Freitas, N., & Doucet, A. (2001a). Robust full Bayesian learning for radial basis networks. Neural
Computation, 13:10, 2359–2407.
Andrieu, C., de Freitas, N., & Doucet, A. (2001b). Rao-blackwellised particle filtering via data augmentation.
Advances in Neural Information Processing Systems (NIPS13).
INTRODUCTION 39
Andrieu, C., Doucet, A., & Punskaya, E. (2001). Sequential Monte Carlo methods for optimal filtering. In A
Doucet, N. de Freitas, & N. J. Gordon (Eds.), Sequential Monte Carlo methods in practice. Berlin: Springer-
Verlag.
Applegate, D., & Kannan, R. (1991). Sampling and integration of near log-concave functions. In Proceedings of
the Twenty Third Annual ACM Symposium on Theory of Computing (pp. 156–163).
Bar-Yossef, Z., Berg, A., Chien, S., Fakcharoenphol, J., & Weitz, D. (2000). Approximating aggregate queries
about web pages via random walks. In International Conference on Very Large Databases (pp. 535–544).
Barber, D., & Williams, C. K. I. (1997). Gaussian processes for Bayesian classification via hybrid Monte Carlo.
In M. C. Mozer, M. I. Jordan, & T. Petsche (Eds.), Advances in neural information processing systems 9
(pp. 340–346). Cambridge, MA: MIT Press.
Baum, L. E., Petrie, T., Soules, G., & Weiss, N. (1970). A maximization technique occurring in the statistical
analysis of probabilistic functions of Markov chains. Annals of Mathematical Statistics, 41, 164–171.
Baxter, R. J. (1982). Exactly solved models in statistical mechanics. San Diego, CA: Academic Press.
Beichl, I., & Sullivan, F. (2000). The Metropolis algorithm. Computing in Science & Engineering, 2:1, 65–69.
Bergman, N. (1999). Recursive Bayesian estimation: Navigation and tracking applications. Ph.D. Thesis, Depart-
ment of Electrical Engineering, Link¨ oping University, Sweden.
Berners-Lee, T., Cailliau, R., Luotonen, A., Nielsen, H. F., & Secret, A. (1994). The World-Wide Web. Commu-
nications of the ACM, 10:4, 49–63.
Besag, J., Green, P. J., Hidgon, D., & Mengersen, K. (1995). Bayesian computation and stochastic systems.
Statistical Science, 10:1, 3–66.
Bielza, C., M¨ uller, P., & Rios Insua, D. (1999). Decision Analysis by Augmented Probability Simulation,
Management Science, 45:7, 995–1007.
Brooks, S. P. (1998). Markov chain Monte Carlo method and its application. The Statistician, 47:1, 69–100.
Browne, W. J., &Draper, D. (2000). Implementation and performance issues in the Bayesian and likelihood fitting
of multilevel models. Computational Statistics, 15, 391–420.
Bucher, C. G. (1988). Adaptive sampling—An iterative fast Monte Carlo procedure. Structural Safety, 5, 119–126.
Bui, H. H., Venkatesh, S., & West, G. (1999). On the recognition of abstract Markov policies. In National
Conference on Artificial Intelligence (AAAI-2000).
Carlin, B. P., &Chib, S. (1995). Bayesian Model choice via MCMC. Journal of the Royal Statistical Society Series
B, 57, 473–484.
Carter, C. K., & Kohn, R. (1994). On Gibbs sampling for state space models. Biometrika, 81:3, 541–553.
Casella, G., & Robert, C. P. (1996). Rao-Blackwellisation of sampling schemes. Biometrika, 83:1, 81–94.
Casella, G., Mengersen, K. L., Robert, C. P., & Titterington, D. M. (1999). Perfect slice samplers for mixtures of
distributions. Technical Report BU-1453-M, Department of Biometrics, Cornell University.
Celeux, G., & Diebolt, J. (1985). The SEM algorithm: A probabilistic teacher algorithm derived from the EM
algorithm for the mixture problem. Computational Statistics Quarterly, 2, 73–82.
Celeux, G., &Diebolt, J. (1992). Astochastic approximation type EMalgorithmfor the mixture problem. Stochas-
tics and Stochastics Reports, 41, 127–146.
Chen, M. H., Shao, Q. M., &Ibrahim, J. G. (Eds.) (2001). Monte Carlo methods for Bayesian computation. Berlin:
Springer-Verlag.
Cheng, J., &Druzdzel, M. J. (2000). AIS-BN: An adaptive importance sampling algorithmfor evidential reasoning
in large bayesian networks. Journal of Artificial Intelligence Research, 13, 155–188.
Chenney, S., &Forsyth, D. A. (2000). Sampling plausible solutions to multi-body constraint problems. SIGGRAPH
(pp. 219–228).
Clark, E., & Quinn, A. (1999). A data-driven Bayesian sampling scheme for unsupervised image segmentation.
In IEEE International Conference on Acoustics, Speech, and Signal Processing, Arizona (Vol. 6, pp. 3497–
3500).
Damien, P., Wakefield, J., & Walker, S. (1999). Gibbs sampling for Bayesian non-conjugate and hierarchical
models by auxiliary variables. Journal of the Royal Statistical Society B, 61:2, 331–344.
de Freitas, N., Højen-Sørensen, P., Jordan, M. I., & Russell, S. (2001). Variational MCMC. In J. Breese & D.
Koller (Eds.), Uncertainty in artificial intelligence (pp. 120–127). San Matio, CA: Morgan Kaufmann.
de Freitas, N., Niranjan, M., Gee, A. H., & Doucet, A. (2000). Sequential Monte Carlo methods to train neural
network models. Neural Computation, 12:4, 955–993.
40 C. ANDRIEU ET AL.
De Jong, P., & Shephard, N. (1995). Efficient sampling from the smoothing density in time series models.
Biometrika, 82:2, 339–350.
Dempster, A. P., Laird, N. M., & Rubin, D. B. (1997). Maximum likelihood from incomplete data via the EM
algorithm. Journal of the Royal Statistical Society Series B, 39, 1–38.
Denison, D. G. T., Mallick, B. K., & Smith, A. F. M. (1998). A Bayesian CART algorithm. Biometrika, 85,
363–377.
Diaconis, P., & Saloff-Coste, L. (1998). What do we know about the Metropolis algorithm? Journal of Computer
and System Sciences, 57, 20–36.
Doucet, A. (1998). On sequential simulation-based methods for Bayesian filtering. Technical Report CUED/F-
INFENG/TR 310, Department of Engineering, Cambridge University.
Doucet, A., de Freitas, N., & Gordon, N. J. (Eds.) (2001). Sequential Monte Carlo methods in practice. Berlin:
Springer-Verlag.
Doucet, A., de Freitas, N., Murphy, K., & Russell, S. (2000). Rao blackwellised particle filtering for dynamic
Bayesiannetworks. InC. Boutilier &M. Godszmidt (Eds.), Uncertainty inartificial intelligence (pp. 176–183).
Morgan Kaufmann Publishers.
Doucet, A., Godsill, S., &Andrieu, C. (2000). On sequential Monte Carlo sampling methods for Bayesian filtering.
Statistics and Computing, 10:3, 197–208.
Doucet, A., Godsill, S. J., & Robert, C. P. (2000). Marginal maximum a posteriori estimation using MCMC.
Technical Report CUED/F-INFENG/TR 375, Cambridge University Engineering Department.
Duane, S., Kennedy, A. D., Pendleton, B. J., & Roweth, D. (1987). Hybrid Monte Carlo. Physics Letters B, 195:2,
216–222.
Dyer, M., Frieze, A., & Kannan, R. (1991). A random polynomial-time algorithm for approximating the volume
of convex bodies. Journal of the ACM, 1:38, 1–17.
Eckhard, R. (1987). Stan Ulam, John Von Neumann and the Monte Carlo method. Los Alamos Science, 15,
131–136.
Escobar, M. D., & West, M. (1995). Bayesian density estimation and inference using mixtures. Journal of the
American Statistical Association, 90, 577–588.
Fill, J. A. (1998). An interruptible algorithm for perfect sampling via Markov chains. The Annals of Applied
Probability, 8:1, 131–162.
Forsyth, D. A. (1999). Sampling, resampling and colour constancy. In IEEE Conference on Computer Vision and
Pattern Recognition (pp. 300–305).
Fox, D., Thrun, S., Burgard, W., & Dellaert, F. (2001). Particle filters for mobile robot localization. In A. Doucet,
N. de Freitas, & N. J. Gordon (Eds.), Sequential Monte Carlo methods in practice. Berlin: Springer-Verlag.
Gelfand, A. E., & Sahu, S. K. (1994). On Markov chain Monte Carlo acceleration. Journal of Computational and
Graphical Statistics, 3, 261–276.
Gelfand, A. E., & Smith, A. F. M. (1990). Sampling-based approaches to calculating marginal densities. Journal
of the American Statistical Association, 85:410, 398–409.
Geman, S., &Geman, D. (1984). Stochastic relaxation, Gibbs distributions and the Bayesian restoration of images.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 6:6, 721–741.
Geweke, J. (1989). Bayesian inference in econometric models using Monte Carlo integration. Econometrica, 24,
1317–1399.
Ghahramani, Z. (1995). Factorial learning and the EM algorithm. In G. Tesauro, D. S. Touretzky, & J. Alspector
(Eds.), Advances in neural information processing systems 7 (pp. 617–624).
Ghahramani, Z., & Jordan, M. (1995). Factorial hidden Markov models. Technical Report 9502, MIT Artificial
Intelligence Lab, MA.
Gilks, W. R., & Berzuini, C. (1998). Monte Carlo inference for dynamic Bayesian models. Unpublished. Medical
Research Council, Cambridge, UK.
Gilks, W. R., & Roberts, G. O. (1996). Strategies for improving MCMC. In W. R. Gilks, S. Richardson, & D. J.
Spiegelhalter (Eds.), Markov chain Monte Carlo in practice (pp. 89–114). Chapman & Hall.
Gilks, W. R., Richardson, S., & Spiegelhalter, D. J. (Eds.) (1996). Markov chain Monte Carlo in practice. Suffolk:
Chapman and Hall.
Gilks, W. R., Roberts, G. O., & Sahu, S. K. (1998). Adaptive Markov chain Monte Carlo through regeneration.
Journal of the American Statistical Association, 93, 763–769.
INTRODUCTION 41
Gilks, W. R., Thomas, A., &Spiegelhalter, D. J. (1994). Alanguage and programfor complex Bayesian modelling.
The Statistician, 43, 169–178.
Godsill, S. J., & Rayner, P. J. W. (Eds.) (1998). Digital audio restoration: A statistical model based approach.
Berlin: Springer-Verlag.
Gordon, N. J., Salmond, D. J., & Smith, A. F. M. (1993). Novel approach to nonlinear/non-Gaussian Bayesian
state estimation. IEE Proceedings-F, 140:2, 107–113.
Green, P. J. (1995). Reversible jump Markov chain Monte Carlo computation and Bayesian model determination.
Biometrika, 82, 711–732.
Green, P. J., &Richardson, S. (2000). Modelling heterogeneity with and without the Dirichlet process. Department
of Statistics, Bristol University.
Haario, H., & Sacksman, E. (1991). Simulated annealing process in general state space. Advances in Applied
Probability, 23, 866–893.
Hastings, W. K. (1970). Monte Carlo sampling methods using Markov chains and their Applications. Biometrika
57, 97–109.
Higdon, D. M. (1998). Auxiliary variable methods for Markov chain Monte Carlo with application. Journal of
American Statistical Association, 93:442, 585–595.
Holmes, C. C., & Mallick, B. K. (1998). Bayesian radial basis functions of variable dimension. Neural Compu-
tation, 10:5, 1217–1233.
Isard, M., & Blake, A. (1996). Contour tracking by stochastic propagation of conditional density. In European
Conference on Computer Vision (pp. 343–356). Cambridge, UK.
Ishwaran, H. (1999). Application of hybrid Monte Carlo to Bayesian generalized linear models: Quasicomplete
separation and neural networks. Journal of Computational and Graphical Statistics, 8, 779–799.
Jensen, C. S., Kong, A., &Kjærulff, U. (1995). Blocking-Gibbs sampling in very large probabilistic expert systems.
International Journal of Human-Computer Studies, 42, 647–666.
Jerrum, M., & Sinclair, A. (1996). The Markov chain Monte Carlo method: an approach to approximate counting
and integration. In D. S. Hochbaum (Ed.), Approximation algorithms for NP-hard problems (pp. 482–519).
PWS Publishing.
Jerrum, M., Sinclair, A., & Vigoda, E. (2000). A polynomial-time approximation algorithm for the permanent of
a matrix. Technical Report TR00-079, Electronic Colloquium on Computational Complexity.
Kalos, M. H., & Whitlock, P. A. (1986). Monte Carlo methods. New York: John Wiley & Sons.
Kam, A. H. (2000). Ageneral multiscale scheme for unsupervised image segmentation. Ph.D. Thesis, Department
of Engineering, Cambridge University, Cambridge, UK.
Kanazawa, K., Koller, D., & Russell, S. (1995). Stochastic simulation algorithms for dynamic probabilistic net-
works. In Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence (pp. 346–351).
Morgan Kaufmann.
Kannan, R., & Li, G. (1996). Sampling according to the multivariate normal density. In 37th Annual Symposium
on Foundations of Computer Science (pp. 204–212). IEEE.
Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983). Optimization by simulated annealing. Science, 220, 671–
680.
Levine, R., & Casella, G. (2001). Implementations of the Monte Carlo EM algorithm. Journal of Computational
and Graphical Statistics, 10:3, 422–440.
Liu, J. S. (Ed.) (2001). Monte Carlo strategies in scientific computing. Berlin: Springer-Verlag.
MacEachern, S. N., Clyde, M., & Liu, J. S. (1999). Sequential importance sampling for nonparametric Bayes
models: The next generation. Canadian Journal of Statistics, 27, 251–267.
McCulloch, C. E. (1994). Maximum likelihood variance components estimation for binary data. Journal of the
American Statistical Association, 89:425, 330–335.
Mengersen, K. L., & Tweedie, R. L. (1996). Rates of convergence of the Hastings and Metropolis algorithms. The
Annals of Statistics, 24, 101–121.
Metropolis, N., & Ulam, S. (1949). The Monte Carlo method. Journal of the American Statistical Association,
44:247, 335–341.
Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H., & Teller, E. (1953). Equations of state
calculations by fast computing machines. Journal of Chemical Physics, 21, 1087–1091.
Meyn, S. P., & Tweedie, R. L. (1993). Markov chains and stochastic stability. New York: Springer-Verlag.
42 C. ANDRIEU ET AL.
Mira, A. (1999). Ordering, slicing and splitting Monte Carlo Markov chains. Ph.D. Thesis, School of Statistics,
University of Minnesota.
Morris, R. D., Fitzgerald, W. J., & Kokaram, A. C. (1996). A sampling based approach to line scratch removal
from motion picture frames. In IEEE International Conference on Image Processing (pp. 801–804).
M¨ uller, P., & Rios Insua, D. (1998). Issues in Bayesian analysis of neural network models. Neural Computation,
10, 571–592.
Mykland, P., Tierney, L., & Yu, B. (1995). Regeneration in Markov chain samplers. Journal of the American
Statistical Association, 90, 233–241.
Neal, R. M. (1993). Probabilistic inference using markov chain monte carlo methods. Technical Report CRG-TR-
93-1, Dept. of Computer Science, University of Toronto.
Neal, R. M. (1996). Bayesian learning for neural networks. Lecture Notes in Statistics No. 118. New York:
Springer-Verlag.
Neal, R. M. (2000). Slice sampling. Technical Report No. 2005, Department of Statistics, University of Toronto.
Neuwald, A. F., Liu, J. S., Lipman, D. J., & Lawrence, C. E. (1997). Extracting protein alignment models from
the sequence database. Nucleic Acids Research, 25:9, 1665–1677.
Newton, M. A., &Lee, Y. (2000). Inferringthe locationandeffect of tumor suppressor genes byinstability-selection
modeling of allelic-loss data. Biometrics, 56, 1088–1097.
Ormoneit, D., Lemieux, C., & Fleet, D. (2001). Lattice particle filters. Uncertainty in artificial intelligence. San
Mateo, CA: Morgan Kaufmann.
Ortiz, L. E., & Kaelbling, L. P. (2000). Adaptive importance sampling for estimation in structured domains. In
C. Boutilier, & M. Godszmidt (Eds.), Uncertainty in artificial intelligence (pp. 446–454). San Mateo, CA:
Morgan Kaufmann Publishers.
Page, L., Brin, S., Motwani, R., & Winograd, T. (1998). The PageRank citation ranking: Bringing order to the
Web. Stanford Digital Libraries Working Paper.
Pasula, H., & Russell, S. (2001). Approximate inference for first-order probabilistic languages. In International
Joint Conference on Artificial Intelligence, Seattle.
Pasula, H., Russell, S., Ostland, M., &Ritov, Y. (1999). Tracking many objects with many sensors. In International
Joint Conference on Artificial Intelligence, Stockholm.
Pearl, J. (1987). Evidential reasoning using stochastic simulation. Artificial Intelligence, 32, 245–257.
Peskun, P. H. (1973). Optimum Monte-Carlo sampling using Markov chains. Biometrika, 60:3, 607–612.
Pitt, M. K., & Shephard, N. (1999). Filtering via simulation: Auxiliary particle filters. Journal of the American
Statistical Association, 94:446, 590–599.
Propp, J., &Wilson, D. (1998). Couplingfromthe past: a user’s guide. InD. Aldous, &J. Propp(Eds.), Microsurveys
in discrete probability. DIMACS series in discrete mathematics and theoretical computer science.
Remondo, D., Srinivasan, R., Nicola, V. F., vanEtten, W. C., &Tattje, H. E. P. (2000). Adaptive importance sampling
for performance evaluation and parameter optimization of communications systems. IEEE Transactions on
Communications, 48:4, 557–565.
Richardson, S., & Green, P. J. (1997). On Bayesian analysis of mixtures with an unknown number of components.
Journal of the Royal Statistical Society B, 59:4, 731–792.
Ridgeway, G. (1999). Generalization of boosting algorithms and applications of bayesian inference for massive
datasets. Ph.D. Thesis, Department of Statistics, University of Washington.
Rios Insua, D., & M¨ uller, P. (1998). Feedforward neural networks for nonparametric regression. In D. K. Dey, P.
M¨ uller, & D. Sinha (Eds.), Practical nonparametric and semiparametric bayesian statistics (pp. 181–191).
Springer Verlag.
Robert, C. P., & Casella, G. (1999). Monte Carlo statistical methods. New York: Springer-Verlag.
Roberts, G., & Tweedie, R. (1996). Geometric convergence and central limit theorems for multidimensional
Hastings and Metropolis algorithms. Biometrika, 83, 95–110.
Rubin, D. B. (1998). Using the SIR algorithm to simulate posterior distributions. In J. M. Bernardo, M. H.
DeGroot, D. V. Lindley, & A. F. M. Smith (Eds.), Bayesian statistics 3 (pp. 395–402). Cambridge, MA:
Oxford University Press.
Rubinstein, R. Y. (Eds.) (1981). Simulation and the Monte Carlo method. New York: John Wiley and Sons.
Salmond, D., & Gordon, N. (2001). Particles and mixtures for tracking and guidance. In A. Doucet, N. de Freitas,
& N. J. Gordon (Eds.), Sequential Monte Carlo methods in practice. Berlin: Springer-Verlag.
INTRODUCTION 43
Schuurmans, D., & Southey, F. (2000). Monte Carlo inference via greedy importance sampling. In C. Boutilier,
& M. Godszmidt (Eds.), Uncertainty in artificial intelligence (pp. 523–532). Morgan Kaufmann Publishers.
Sherman, R. P., Ho, Y. K., & Dalal, S. R. (1999). Conditions for convergence of Monte Carlo EM sequences with
an application to product diffusion modeling. Econometrics Journal, 2:2, 248–267.
Smith, P. J., Shafi, M., & Gao, H. (1997). Quick simulation: A review of importance sampling techniques in
communications systems. IEEE Journal on Selected Areas in Communications, 15:4, 597–613.
Stephens, M. (1997). Bayesian methods for mixtures of normal distributions. Ph.D. Thesis, Department of Statis-
tics, Oxford University, England.
Swendsen, R. H., & Wang, J. S. (1987). Nonuniversal critical dynamics in Monte Carlo simulations. Physical
Review Letters, 58:2, 86–88.
Tanner, M. A., & Wong, W. H. (1987). The calculation of posterior distributions by data augmentation. Journal
of the American Statistical Association, 82:398, 528–550.
Thrun, S. (2000). Monte Carlo POMDPs. In S. Solla, T. Leen, & K.-R. M¨ uller (Eds.), Advances in neural
information processing systems 12 (pp. 1064–1070). Cambridge, MA: MIT Press.
Tierney, L. (1994). Markov chains for exploring posterior distributions. The Annals of Statistics, 22:4, 1701–1762.
Tierney, L., &Mira, A. (1999). Some adaptive Monte Carlo methods for Bayesian inference. Statistics in Medicine,
18, 2507–2515.
Troughton, P. T., &Godsill, S. J. (1998). Areversible jump sampler for autoregressive time series. In International
Conference on Acoustics, Speech and Signal Processing (Vol. IV, pp. 2257–2260).
Tu, Z. W., & Zhu, S. C. (2001). Image segmentation by data driven Markov chain Monte Carlo. In International
Computer Vision Conference.
Utsugi, A. (2001). Ensemble of independent factor analyzers with application to natural image analysis. Neural
Processing Letters, 14:1, 49–60.
van der Merwe, R., Doucet, A., de Freitas, N., & Wan, E. (2000). The unscented particle filter. Technical Report
CUED/F-INFENG/TR 380, Cambridge University Engineering Department.
Van Laarhoven, P. J., & Arts, E. H. L. (1987). Simulated annealing: Theory and applications. Amsterdam: Reidel
Publishers.
Veach, E., & Guibas, L. J. (1997). Metropolis light transport. SIGGRAPH, 31, 65–76.
Vermaak, J., Andrieu, C., Doucet, A., &Godsill, S. J. (1999). Non-stationary Bayesian modelling and enhancement
of speech signals. Technical Report CUED/F-INFENG/TR, Cambridge University Engineering Department.
Wakefield, J. C., Gelfand, A. E., & Smith, A. F. M. (1991). Efficient generation of random variates via the
ratio-of-uniforms methods. Statistics and Computing, 1, 129–133.
Wei, G. C. G., & Tanner, M. A. (1990). A Monte Carlo implementation of the EM algorithm and the poor man’s
data augmentation algorithms. Journal of the American Statistical Association, 85:411, 699–704.
West, M., Nevins, J. R., Marks, J. R., Spang, R., & Zuzan, H. (2001). Bayesian regression analysis in the “large
p, small n” paradigm with application in DNA microarray studies. Department of Statistics, Duke University.
Wilkinson, D. J., & Yeung, S. K. H. (2002). Conditional simulation from highly structured Gaussian systems,
with application to blocking-MCMC for the Bayesian analysis of very large linear models. Statistics and
Computing, 12, 287–300.
Wood, S., & Kohn, R. (1998). A Bayesian approach to robust binary nonparametric regression. Journal of the
American Statistical Association, 93:441, 203–213.

6

C. ANDRIEU ET AL.

Stan Ulam soon realised that computers could be used in this fashion to answer questions of neutron diffusion and mathematical physics. He contacted John Von Neumann, who understood the great potential of this idea. Over the next few years, Ulam and Von Neumann developed many Monte Carlo algorithms, including importance sampling and rejection sampling. Enrico Fermi in the 1930’s also used Monte Carlo in the calculation of neutron diffusion, and later designed the FERMIAC, a Monte Carlo mechanical device that performed calculations (Anderson, 1986). In the 1940’s Nick Metropolis, a young physicist, designed new controls for the state-of-the-art computer (ENIAC) with Klari Von Neumann, John’s wife. He was fascinated with Monte Carlo methods and this new computing device. Soon he designed an improved computer, which he named the MANIAC in the hope that computer scientists would stop using acronyms. During the time he spent working on the computing machines, many mathematicians and physicists (Fermi, Von Neumann, Ulam, Teller, Richtmyer, Bethe, Feynman, & Gamow) would go to him with their work problems. Eventually in 1949, he published the first public document on Monte Carlo simulation with Stan Ulam (Metropolis & Ulam, 1949). This paper introduces, among other ideas, Monte Carlo particle methods, which form the basis of modern sequential Monte Carlo methods such as bootstrap filters, condensation, and survival of the fittest algorithms (Doucet, de Freitas, & Gordon, 2001). Soon after, he proposed the Metropolis algorithm with the Tellers and the Rosenbluths (Metropolis et al., 1953). Many papers on Monte Carlo simulation appeared in the physics literature after 1953. From an inference perspective, the most significant contribution was the generalisation of the Metropolis algorithm by Hastings in 1970. Hastings and his student Peskun showed that Metropolis and the more general Metropolis-Hastings algorithms are particular instances of a large family of algorithms, which also includes the Boltzmann algorithm (Hastings, 1970; Peskun, 1973). They studied the optimality of these algorithms and introduced the formulation of the Metropolis-Hastings algorithm that we adopt in this paper. In the 1980’s, two important MCMC papers appeared in the fields of computer vision and artificial intelligence (Geman & Geman, 1984; Pearl, 1987). Despite the existence of a few MCMC publications in the statistics literature at this time, it is generally accepted that it was only in 1990 that MCMC made the first significant impact in statistics (Gelfand & Smith, 1990). In the neural networks literature, the publication of Neal (1996) was particularly influential. In the introduction to this special issue, we focus on describing algorithms that we feel are the main building blocks in modern MCMC programs. We should emphasize that in order to obtain the best results out of this class of algorithms, it is important that we do not treat them as black boxes, but instead try to incorporate as much domain specific knowledge as possible into their design. MCMC algorithms typically require the design of proposal mechanisms to generate candidate hypotheses. Many existing machine learning algorithms can be adapted to become proposal mechanisms (de Freitas et al., 2001). This is often essential to obtain MCMC algorithms that converge quickly. In addition to this, we believe that the machine learning community can contribute significantly to the solution of many open problems in the MCMC field. For this purpose, we have outlined several “hot” research directions at the end of this paper. Finally, readers are encouraged to consult the excellent texts of Chen, Shao, and Ibrahim (2001), Gilks, Richardson, and Spiegelhalter (1996), Liu (2001), Meyn and Tweedie (1993), Robert and Casella (1999) and review papers by Besag

INTRODUCTION

7

et al. (1995), Brooks (1998), Diaconis and Saloff-Coste (1998), Jerrum and Sinclair (1996), Neal (1993), and Tierney (1994) for more information on MCMC. The remainder of this paper is organised as follows. In Part 2, we outline the general problems and introduce simple Monte Carlo simulation, rejection sampling and importance sampling. Part 3 deals with the introduction of MCMC and the presentation of the most popular MCMC algorithms. In Part 4, we describe some important research frontiers. To make the paper more accessible, we make no notational distinction between distributions and densities until the section on reversible jump MCMC. 2. MCMC motivation MCMC techniques are often applied to solve integration and optimisation problems in large dimensional spaces. These two types of problem play a fundamental role in machine learning, physics, statistics, econometrics and decision analysis. The following are just some examples. 1. Bayesian inference and learning. Given some unknown variables x ∈ X and data y ∈ Y, the following typically intractable integration problems are central to Bayesian statistics (a) Normalisation. To obtain the posterior p(x | y) given the prior p(x) and likelihood p(y | x), the normalising factor in Bayes’ theorem needs to be computed p(y | x) p(x) p(x | y) = . p(y | x ) p(x ) d x X (b) Marginalisation. Given the joint posterior of (x, z) ∈ X × Z, we may often be interested in the marginal posterior p(x | y) =
Z

p(x, z | y) dz.

(c) Expectation. The objective of the analysis is often to obtain summary statistics of the form E p(x|y) ( f (x)) =
X

f (x) p(x | y) d x

for some function of interest f : X → Rn f integrable with respect to p(x | y). Examples of appropriate functions include the conditional mean, in which case f (x) = x, or the conditional covariance of x where f (x) = xx −E p(x|y) (x)E p(x|y) (x). 2. Statistical mechanics. Here, one needs to compute the partition function Z of a system with states s and Hamiltonian E(s) Z=
s

exp −

E(s) , kT

where k is the Boltzmann’s constant and T denotes the temperature of the system. Summing over the large number of possible configurations is prohibitively expensive (Baxter, 1982). Note that the problems of computing the partition function and the normalising constant in statistical inference are analogous.

one finds the maximum likelihood (ML) estimates for each model separately. f (x) p(x) d x.s. the space on which the posterior is defined. Moreover. 1999.) converge to I ( f ). This task typically involves two steps. then the variance of the estimator f I N ( f ) is equal to var(I N ( f )) = Nf and a central limit theorem yields convergence in distribution of the error √ N (I N ( f ) − I ( f )) =⇒ N 0. or the combinatorial set of feasible solutions). 3. In general. 1997). computing resources are wasted. Kalos & Whitlock. many of those models are of not interest and.g. the estimate I N ( f ) is unbiased and by the strong law of large numbers. Penalised likelihood model selection. The Monte Carlo principle N The idea of Monte Carlo simulation is to draw an i.i. The goal of optimisation is to extract the solution that minimises some objective function from a large set of feasible solutions. one can approximate the integrals (or very large sums) I ( f ) with tractable sums I N ( f ) that converge as follows IN ( f ) = 1 N N i=1 f x (i) − − I ( f ) = −→ N →∞ a.8 C. If the variance (in the univariate case for simplicity) of f (x) satisfies σ 2 E p(x) ( f 2 (x)) − I 2 ( f ) < ∞. Although we have emphasized integration and optimisation. MCMC also plays a fundamental role in the simulation of physical systems. f N →∞ σ2 where =⇒ denotes convergence in distribution (Robert & Casella. Section 3. The advantage of Monte Carlo integration over deterministic integration arises from the fact that the former positions the integration grid (samples) in regions of high probability. it will almost surely (a. . σ 2 . ANDRIEU ET AL. Then one uses a penalisation term (for example MDL. These N samples can be used to approximate the target density with the following empirical point-mass function p N (x) = 1 N N δx (i) (x). this set can be continuous and unbounded. 4. 2.2). The problem with this approach is that the initial set of models can be very large. it is too computationally expensive to compare all the solutions to find out which one is optimal. 2000.1. Optimisation. set of samples {x (i) }i=1 from a target density p(x) defined on a high-dimensional space X (e. In fact.s. X That is. Veach & Guibas. BIC or AIC) to select one of the models. Consequently.d. therefore. First. 1986. i=1 where δx (i) (x) denotes the delta-Dirac mass located at x (i) . the set of possible configurations of a system. This is of great relevance in nuclear physics and computer graphics (Chenney & Forsyth.

Rejection sampling: Sample a candidate x (i) and a uniform variable u. e.1) denotes the operation of sampling a uniform random variable on the interval (0. The accepted x (i) can be easily shown to be sampled with probability p(x) (Robert & Figure 1. using the accept/reject procedure describe in figure 1 (see also figure 2). we need to introduce more sophisticated techniques based on rejection sampling... Rejection sampling We can sample from a distribution p(x). it is straightforward to sample from it using easily available routines. Figure 2. 2. we will show later that it is possible to construct simulated annealing algorithms that allow us to sample approximately from a distribution whose support is the set of global maxima. Here.N p x (i) However. otherwise reject it. u ∼ U(0. Accept the candidate sample if u Mq(x (i) ) < p(x (i) ). However. by sampling from another easy-to-sample proposal distribution q(x) that satisfies p(x) ≤ Mq(x). . When p(x) has standard form. Gaussian. M < ∞. when this is not the case.g. 1).i=1.INTRODUCTION 9 The N samples can also be used to obtain a maximum of the objective function p(x) as follows ˆ x = arg max x (i) .. Rejection sampling algorithm.. which is known up to a proportionality constant. importance sampling and MCMC.2.

It is clear that this integration method can also be interpreted as a sampling method where the posterior density p(x) is approximated by N N →∞ ˆ p N (x) = i=1 w x (i) δx (i) (x) ˆ and I N ( f ) is nothing but the function f (x) integrated with respect to the empirical measure ˆ p N (x). This makes the method impractical in high-dimensional scenarios. a possible Monte Carlo estimate of I ( f ) is N ˆ I N( f ) = i=1 f x (i) w x (i) This estimator is unbiased and.s. An important criterion for choosing an optimal proposal distribution is to find one that minimises the ˆ variance of the estimator I N ( f ). under weak assumptions. Then we can rewrite I ( f ) as follows I( f ) = f (x) w(x) q(x) d x p(x) where w(x) is known as the importance weight. If M is too large. 2. again. bers applies.d. 1981).10 C. 1989. p. Importance sampling Importance sampling is an alternative “classical” solution that goes back to the 1940’s. The variance of f (x)w(x) with respect to q(x) is given by varq(x) ( f (x)w(x)) = Eq(x) ( f 2 (x)w 2 (x)) − I 2 ( f ) (8) The second term on the right hand side does not depend on q(x) and hence we only need to minimise the first term.i. 1999. Some proposal distributions q(x) will obviously be preferable to others. that is IˆN ( f ) −→ I ( f ). which according to Jensen’s inequality has the following lower . an arbitrary importance proposal distribution q(x) such that its support includes the support of p(x).3. This simple method suffers from severe limitations. Rubinstein. the acceptance probability Pr(x accepted) = Pr u < p(x) Mq(x) = 1 M will be too small. It is not always possible to bound p(x)/q(x) with a reasonable constant M over the whole space X . Casella. if one can simulate q(x) (i) N N i. ANDRIEU ET AL. Consequently. the strong law of large numa. samples {x }i=1 according to q(x) and evaluate w(x (i) ). see for example (Geweke. Let us introduce. 49).

INTRODUCTION 11 bound 2 Eq(x) ( f 2 (x)w 2 (x)) ≥ Eq(x) (| f (x)|w(x)) 2 = | f (x)| p(x) d x This lower bound is attained when we adopt the following optimal importance distribution q (x) = | f (x)| p(x) | f (x)| p(x) d x The optimal proposal is not very useful in the sense that it is not easy to sample from | f (x)| p(x). i. That is. it tells us that high sampling efficiency is achieved when we focus on sampling from p(x) in the important regions where | f (x)| p(x) is relatively large. for a a given function f (x). In this particular example one is interested in computing a tail probability of error (detecting infrequent abnormalities). Shafi. with q(x) = p(x). There the quantity of interest is a tail probability (bit error rate) and hence f (x) = I E (x) where I E (x) = 1 if x ∈ E and 0 otherwise (see figure 3). One could estimate the bit error rate more efficiently by sampling according to q(x) ∝ I E (x) p(x) than according to q(x) = p(x). That is. it is possible to find a distribution q(x) that yields an estimate with a lower variance than when using a perfect Monte Carlo method. This property is often exploited to evaluate the probability of rare events in communication networks (Smith. However. & Gao. the aim is usually different in the sense that Figure 3. 1997). it is wasteful to propose candidates in regions of no utility. Importance sampling: one should place more importance on sampling from the state space regions that matter. This result implies that importance sampling estimates can be super-efficient. In many applications. hence the name importance sampling. .e.

θ) ∂θ and then updating the parameters as follows θt+1 = θt − α 1 N N f 2 x (i) w x (i) . 1995. ˜ applies. the strong law of large numbers a.. 2000. Remondo et al. and has been extensively applied in the communications literature (Al-Qaq. θ ). 1999).s. one wants to have a good approximation of p(x) and not of a particular integral with respect to p(x). 2000). 2000. This procedure results in M samples x (i) with the possibility that x (i) = x ( j) . 1988). under weak assumptions. Cheng & Druzdzel. For N finite. θ ) ∂w(x. 1989). samples from p N (x). ANDRIEU ET AL. (8) D(θ ) = Eq(x. it is still possible to apply the importance sampling method by rewriting I ( f ) as follows: I( f ) = f (x)w(x)q(x) d x w(x)q(x) d x p(x) where w(x) ∝ q(x) is now only known up to a normalising constant. Ortiz & Kaelbling. it becomes harder to obtain a suitable q(x) from which to draw samples.i.. This technique has also been exploited recently in the machine learning community (de Freitas et al. A sensible strategy is to adopt a parameterised q(x.θ) f 2 (x)w(x. Devetsikiotis. Other optimisation approaches that make use of the Hessian are also possible. The estimator I˜ N ( f ) has been shown to perform better than ˆ I N ( f ) in some setups under squared error loss (Robert & Casella. that is I N ( f ) −→ I ( f ). then an asymptotically (N /M → ∞) valid method consists of resampling M times according to the discrete distriˆ ˜ ˜ ˜ bution p N (x). When the normalising constant of p(x) is unknown. Schuurmans & Southey. As the dimension of the x increases.12 C. θt i=1 ∂w x (i) . I˜ N ( f ) is biased (ratio of two ˜ estimates) but asymptotically. ˆ If one is interested in obtaining M i. so we often seek to have q(x) p(x). θ ) and to adapt θ during the simulation. A popular adaptive strategy involves computing the derivative of the first term on the right hand side of Eq. Under additional assumptions a central limit theorem can N →∞ be obtained (Geweke. 2000. θt ∂θt where α is a learning rate and x (i) ∼ q(x. The Monte Carlo estimate of I ( f ) becomes 1 N 1 N N i=1 I˜ N ( f ) = f x (i) w x (i) N j=1 N w x (i) = i=1 f x (i) w x (i) ˜ where w(x (i) ) is a normalised importance weight. 2000). Adaptive importance sampling appears to have originated in the structural safety literature (Bucher.d. & Townsend.

9 0. For this reason. 0.INTRODUCTION 13 for i = j. it follows that µ(x (1) )T = (0. after several iterations (multiplications by T ).6.4. xs }.4). consider a Markov chain with three states (s = 3) and a transition graph as illustrated in figure 4. The transition matrix for this example is    T =  0 0.6 0. it is often impossible to obtain proposal distributions that are easy to sample from and good approximations at the same time. No matter what initial distribution µ(x (1) ) we use. but can evaluate p(x) up to a normalising constant. it is essential to carry out this resampling step. the evolution of the chain in a space X depends solely on the current state of the chain and a fixed transition matrix.4 0 If the probability vector for the initial state is µ(x (1) ) = (0. For any starting point. not clear whether the SIR procedure can lead to practical gains in general. 3. the product µ(x (1) )T t converges to p(x) = (0. MCMC algorithms MCMC is a strategy for generating samples x (i) while exploring the state space X using a Markov chain mechanism. x2 . However.) It is intuitive to introduce Markov chains on finite state spaces.2) and.3). the chain will stabilise at p(x) = (0.2. 0. 0. This mechanism is constructed so that the chain spends more time in the most important regions. therefore. x (1) = T x (i) x (i−1) . (We reiterate that we use MCMC when we cannot draw samples from p(x) directly. That is. . 1988). As an example. This method is known as sampling importance resampling (SIR) (Rubin. in the sequential Monte Carlo setting described in Section 4. . 0. The chain is homogeneous if T T (x (i) | x (i−1) ) remains invariant for all i. We conclude this section by stating that even with adaptation.2. 0.2. where x (i) can only take s discrete values x (i) ∈ X = {x1 .4. . This stability result plays a fundamental role in MCMC simulation. . .3. After resampling. . It is.5. we need to introduce more sophisticated sampling algorithms based on Markov chains. The stochastic process x (i) is called a Markov chain if p x (i) x (i−1) .4). 0.1 0. with x (i) T (x (i) | x (i−1) ) = 1 for any i.2. . the approximation of the target density is ˜ p M (x) = 1 M M δx (i) (x) ˜ i=1 (13) The resampling scheme introduces some additional Monte Carlo variation. In particular. 0. it is constructed so that the samples x (i) mimic samples drawn from the target distribution p(x). . the chain will convergence to the 0 1 0  . 0.

x2 . therefore. In fact. invariant distribution p(x). and should be as small as possible. The second largest eigenvalue. determines the rate of convergence of the chain. most of our efforts will be devoted to increasing the convergence speed.4 x1 0. there is a positive probability of visiting all other states. the Perron-Frobenius theorem from linear algebra tells us that the remaining eigenvalues have absolute value less than 1. Aperiodicity. That is.6 Figure 4. gives us p x (i) = x (i−1) p x (i−1) T x (i) | x (i−1) . Spectral theory gives us useful insights into the problem. The concepts of irreducibility. but not necessary. condition to ensure that a particular p(x) is the desired invariant distribution is the following reversibility (detailed balance) condition p x (i) T x (i−1) x (i) = p x (i−1) T x (i) x (i−1) . When we search for information on . aperiodicity and invariance can be better appreciated once we realise the important role that they play in our lives. the matrix T cannot be reduced to separate smaller matrices. ANDRIEU ET AL. However. 2. which is also the same as stating that the transition graph is connected.1 x2 1 0. as long as T is a stochastic transition matrix that obeys the following properties: 1.14 C. The chain should not get trapped in cycles. For any state of the Markov chain. 0. Summing both sides over x (i−1) .9 0. x3 }. A sufficient. x3 Transition graph for the Markov chain example with X = {x1 . Indeed. it is also important to design samplers that converge quickly. Notice that p(x) is the left eigenvector of the matrix T with corresponding eigenvalue 1. Irreducibility. MCMC samplers are irreducible and aperiodic Markov chains that have the target distribution as the invariant distribution. One way to design these samplers is to ensure that detailed balance is satisfied.

1994). such that the entry L i. 1998).INTRODUCTION 15 the World-Wide Web.. In later sections. the invariant distribution (eigenvector) p(x) represents the rank of a webpage x. In continuous state spaces. the addition of noise prevents us from getting trapped in loops. in this case.1. 1970. The pseudo-code is shown in figure 5. respectively. we have p x (i+1) [L + E] = p(x i ) where. the random walkers on the Web) want to avoid getting trapped in cycles (aperiodicity) and want to be able to access all the existing webpages (irreducibility). the popular information retrieval algorithm used by the search engine Google. Metropolis et al. as it ensures that there is always some probability of jumping to anywhere on the Web. That is. now. otherwise it remains at x.. PageRank requires the definition of a transition matrix with two components T = L + E. as the nodes and directed connections in a Markov chain transition graph. one can incorporate terms into the transition matrix that favour particular webpages or that bias the search in useful ways. We can interpret the webpages and links. we (say. .3 exp(−0. the transition matrix T becomes an integral kernel K and p(x) becomes the corresponding eigenfunction p x (i) K x (i+1) x (i) d x (i) = p x (i+1) . j represents the normalised number of links from web page i to web page j. E is a uniform random matrix of small magnitude that is added to L to ensure irreducibility and aperiodicity. As long as one satisfies irreducibility and aperiodicity. L is a large link matrix with rows and columns corresponding to web pages. Let us consider. [ p(x)q(x | x)]−1 p(x )q(x | x )}. The kernel K is the conditional density of x (i+1) given the value x (i) . 100) and a bimodal target distribution p(x) ∝ 0. while figure 6 shows the results of running the MH algorithm with a Gaussian proposal distribution q(x | x (i) ) = N (x (i) . In the following subsections we describe various of these algorithms.2x 2 ) + 0. An MH step of invariant distribution p(x) and proposal distribution q(x | x) involves sampling a candidate value x given the current value x according to q(x | x).. namely PageRank (Page et al. Clearly.2(x − 10)2 ) for 5000 iterations. The Markov chain then moves towards x with acceptance probability A(x. 3. we will see that most practical MCMC algorithms can be interpreted as special cases or extensions of this algorithm. the histogram of the samples approximates the target distribution. As expected. Note that it is possible to design more interesting transition matrices in this setting.7 exp(−0. 1953). From our previous discussion. It is a mathematical representation of a Markov chain algorithm. The Metropolis-Hastings algorithm The Metropolis-Hastings (MH) algorithm is the most popular MCMC method (Hastings. x ) = min{1. we typically follow a set of links (Berners-Lee et al.

we will see that many MCMC algorithms arise by considering specific choices of this distribution.05 0 −10 0 10 20 0 −10 0 10 20 0. In general.1 i=500 0.1 i=1000 0.05 0.15 0. but it requires careful design of the proposal distribution q(x | x). 0. Metropolis-Hastings algorithm.1 i=100 0.05 0. .05 0 −10 0 10 20 0 −10 0 10 20 Figure 6.16 C.15 0.15 0. it is possible to use suboptimal inference and learning algorithms to generate data-driven proposal distributions. In subsequent sections.15 0. The MH algorithm is very simple. Target distribution and histogram of the MCMC samples at different iteration points.1 i=5000 0. ANDRIEU ET AL. The transition kernel for the MH algorithm is K MH x (i+1) x (i) = q x (i+1) x (i) A x (i) . Figure 5. x (i+1) + δx (i) x (i+1) r x (i) .

By construction. the acceptance ratio simplifies to A x (i) . We only need to know the target distribution up to a constant of proportionality. only one mode of p(x) might be visited. This algorithm is close to importance sampling. it follows that it is aperiodic. bounded in Rn ). The independent sampler and the Metropolis algorithm are two simple instances of the MH algorithm. Hence. consequently. This is illustrated in figure 7. we obtain asymptotic convergence (Tierney. we need to ensure that there are no cycles (aperiodicity) and that every state that has positive probability can be reached in a finite number of steps (irreducibility). Since the algorithm always allows for rejection. 1993. It is also possible to prove geometric ergodicity using Foster-Lyapunov drift conditions (Meyn & Tweedie. the chain is said to “mix” well. if it is too wide. q(x | x (i) ) = q(x ). Some properties of the MH algorithm are worth highlighting. Secondly. If the space X is small (for example. p(x )q x (i) p x (i) q(x ) = min 1. If all the modes are visited while the acceptance probability is high. we simply need to make sure that the support of q(·) includes the support of p(·). The Metropolis algorithm assumes a symmetric random walk proposal q(x | x (i) ) = q(x (i) | x ) and. 1996). On the other hand. x dx . but now the samples are correlated since they result from comparing one sample to the other. If the proposal is too narrow. the MH algorithm admits p(x) as invariant distribution. 1717).INTRODUCTION 17 where r (x (i) ) is the term associated with rejection r x (i) = q x X x (i) 1 − A x (i) . It is fairly easy to prove that the samples generated by MH algorithm will mimic samples drawn from the target distribution asymptotically. x = min 1. the normalising constant of the target distribution is not required. Under these conditions. the success or failure of the algorithm often hinges on the choice of proposal distribution. K MH satisfies the detailed balance condition p x (i) K MH x (i+1) x (i) = p x (i+1) K MH x (i) x (i+1) and. x = min 1. then it is possible to use minorisation conditions to prove uniform (geometric) ergodicity (Meyn & Tweedie. . the rejection rate can be very high. hence. Different choices of the proposal standard deviation σ lead to very different results. w(x ) w x (i) . 1994. Lastly. To ensure irreducibility. Roberts & Tweedie. 1993). although the pseudo-code makes use of a single chain. To show that the MH algorithm converges. Theorem 3. Firstly. In the independent sampler the proposal is independent of the current state. p. resulting in high correlations. p(x ) p x (i) . the acceptance probability is A x (i) . it is easy to simulate several independent chains in parallel.

Approximations obtained using the MH algorithm with three Gaussian proposal distributions of different variances. Van Laarhoven & Arts. if p(x) is the likelihood or posterior distribution.. .i=1. As mentioned earlier. we often want to compute the ML and maximum a posteriori (MAP) estimates. Unless the distribution has large probability mass around the mode. Figure 7. Gelatt.2. 3. ANDRIEU ET AL.. but to pi (x) ∝ p 1/Ti (x). 1987). A more principled strategy is to adopt simulated annealing (Geman & Geman.N This method is inefficient because the random samples only rarely come from the vicinity of the mode.. x (i) . we could run a Markov chain of invariant distribution p(x) and estimate the global mode by ˆ x = arg max p x (i) . & Vecchi. Simulated annealing for global optimization Let us assume that instead of wanting to approximate p(x).. 1983.18 C. For example. This technique involves simulating a non-homogeneous Markov chain whose invariant distribution at iteration i is no longer equal to p(x). Kirkpatrick. 1984. computing resources will be wasted exploring areas of no interest. we want to find its global maximum.

p ∞ (x) is a probability density that concentrates itself on the set of global maxima of p(x). The simulated annealing involves.INTRODUCTION 19 Figure 8. therefore.2 0.2 0. it is again important to choose suitable proposal distributions and an appropriate cooling schedule. Many of the negative simulated annealing . 0.1 i=1000 0. under weak regularity assumptions on p(x).1 i=5000 0 −10 0 10 20 0 −10 0 10 20 Figure 9. The reason for doing this is that.1 i=500 0 −10 0 10 20 0 −10 0 10 20 0. General simulated annealing algorithm. The results of applying annealing to the example of the previous section are shown in figure 9.2 0. where Ti is a decreasing cooling schedule with limi→∞ Ti = 0.2 0. Discovering the modes of the target distribution with the simulated annealing algorithm.1 i=100 0. just a minor modification of standard MCMC algorithms as shown in figure 8. To obtain efficient annealed algorithms.

Breyer. Robert & Casella. . . . 2000a). 2000b). ANDRIEU ET AL. mixtures of kernels also play a big role in many other samplers. . de Freitas. The transition kernel for this algorithm is given by the following expression nb K MH-Cycle x (i+1) x (i) = j=1 (i+1) (i) (i+1) xb j . Some results for non-compact spaces can be found in Andrieu. one might want to have a global proposal that places the bases (kernels) at the locations of the input data and a local random walk proposal that perturbs these in order to obtain better fits (Andrieu. a global proposal locks into the peaks while a local proposal allows one to explore the space around each peak. . For example. 1994). xb j−1 . . & Doucet. . using b j to indicate the j-th block. for example. one can even propose from complex reversible jump MCMC kernels (Section 3.7). when the target distribution has many narrow peaks. x−[b j ] K MH(j) xb j where K MH(j) denotes the j-th MH algorithm in the cycle. Typically the samplers will mix more quickly by blocking highly correlated variables. Here. Most convergence results for simulated annealing typically state that if for a given Ti . A block MCMC sampler. 1999). xbn }. and Doucet (1999). This will be useful. one can use the fast Fourier transform (FFT) as a global proposal and a random walk as local proposal (Andrieu & Doucet. The pseudo-code for a typical mixture of kernels is shown in figure 10.7) within the annealing algorithm (Andrieu. 1987) or compact continuous spaces (Haario & Sacksman. 1991). Cycles allow us to split a multivariate state vector into components (blocks) that can be updated separately. Mixtures of kernels can incorporate global proposals to explore vast regions of the state space and local proposals to discover finer details of the target distribution (Andrieu. 1984. then the cycle hybrid kernel K 1 K 2 and the mixture hybrid kernel ν K 1 + (1 − ν)K 2 . then convergence to the set of global maxima of p(x) is ensured for a sequence Ti = (C ln(i + T0 ))−1 .20 C. de Freitas. de Freitas. 1999). including the reversible jump MCMC algorithm (Section 3. Similarly. . results reported in the literature often stem from poor proposal distribution design. & Doucet. Andrieu & Doucet. the homogeneous Markov transition kernel mixes quickly enough.3. where C and T0 are problem-dependent. Most of the results have been obtained for finite spaces (Geman & Geman. However. are also transition kernels with invariant distribution p(·). this technique can be used to search for the best model (according to MDL or AIC criteria) and ML parameter estimates simultaneously. If the transition kernels K 1 and K 2 have invariant distribution p(·) each. xb2 . 1999. xb j+1 . Mixtures and cycles of MCMC kernels A very powerful property of MCMC is that it is possible to combine several samplers into mixtures and cycles of the individual samplers (Tierney. is shown in b figure 11. In some complex variable and model selection scenarios arising in machine learning. if we require a high-precision frequency detector. If one defines a joint distribution over the parameter and model spaces. & Doucet. 3. n b to denote (i+1) (i+1) (i+1) (i+1) (i) (i) the number of blocks and x−[b j ] {xb1 . . Van Laarhoven & Arts. for 0 ≤ ν ≤ 1. in kernel regression and classification. 2000b.

. This problem gets worse as the correlation between the components increases. . . .4. 1984). Alternatively. Obviously. The following section describes it in more detail. choosing the size of the blocks poses some trade-offs. Cycle of MCMC kernels—block MH algorithm. we have replaced the index notation b j with j). . In this case. x j+1 . . . x j+1 . . x j−1 .INTRODUCTION 21 Figure 10. A popular cycle of MH kernels. the chain may take a very long time to explore the target distribution. if one samples all the components together. . x j−1 . The Gibbs sampler Suppose we have an n-dimensional vector x and the expressions for the full conditionals p(x j | x1 . Typical mixture of MCMC kernels. . is obtained when we adopt the full conditional distributions p(x j | x− j ) = p(x j | x1 . xn ). . xn ) as proposal distributions (for notational simplicity. If one samples the components of a multi-dimensional vector one-at-a-time. . Figure 11. known as Gibbs sampling (Geman & Geman. 3. . . it is often advantageous to use the following . . . then the probability of accepting this large move tends to be very low.

For n = 2. Since the Gibbs sampler can be viewed as a special case of the MH algorithm. p x (i) p(x j |x− j ) p x− j (i) p x− j = 1. Laird. we can draw samples with MH steps embedded within the Gibbs algorithm. the acceptance probability for each proposal is one and. . Directed acyclic graphs (DAGS) are one of the best known application areas for Gibbs sampling (Pearl. hence. x = min 1. q x 0 Otherwise. ANDRIEU ET AL. That is. The corresponding acceptance probability is: p(x )q x (i) x p x (i) q x |x (i) (i) p(x ) p x (i) x− j j A x (i) . n x (i) = (i) p x j x− j (i) If x− j = x− j C. we will draw the new samples directly. . = min 1. 1987). In particular.). if x pa( j) Figure 12. a large-dimensional joint distribution is factored into a directed graph that encodes the conditional independencies in the model. = min 1. 1977. & Rubin. Here. Otherwise. . . Gibbs sampler. it is possible to introduce MH steps into the Gibbs sampler. which is closely related to the expectation maximisation (EM) algorithm (Dempster.22 proposal distribution for j = 1. 1987). Tanner & Wong. etc. the deterministic scan Gibbs sampler algorithm is often presented as shown in figure 12. Gaussian. when the full conditionals are available and belong to the family of standard distributions (Gamma. the Gibbs sampler is also known as the data augmentation algorithm. . That is.

Thomas. 1970. and Dalal (1999). (this issue). Sampling from the full conditionals. That is. Compute the expected value of the complete log-likelihood function with respect to the distribution of the hidden variables Q(θ ) = log( p(x h . then a local maximum of the likelihood p(xv | θ ) given the parameters θ can be found by iterating the following two steps: 1.INTRODUCTION 23 denotes the parent nodes of node x j . Monte Carlo EM The EM algorithm (Baum et al. Perform the following maximisation θ (new) = arg maxθ Q(θ ). 1994. we only need to take into account the parents. the expectation in the E step is either a sum with an exponentially large number of summands or an intractable integral (Ghahramani. 1995. with the Gibbs sampler. . various authors have proposed stochastic approaches that rely on sampling from p(x h | xv . 1977) is a standard algorithm for ML and MAP point estimation. & Spiegelhalter. If X contains visible and hidden variables x = {xv . lends itself naturally to the construction of general purpose MCMC software. 2002). while Levine and Casella (2001) is a good recent review. θ (old) ) and replace the expectation in the E step with a small sum over the samples. θ (old) d x h . Dempster. we have p(x) = j p x j x pa( j) . Wilkinson & Yeung. M step. It follows that the full conditionals simplify as follows p x j x− j = p x j x pa( j) k∈ch( j) p xk x pa(k) where ch( j) denotes the children nodes of x j . Xh where θ (old) refers to the value of the parameters at the previous time step. McCulloch. see also Dellaert et al. Pasula et al. 2001). Kong. Ho. A solution is to introduce MCMC to sample from p(x h | xv . as shown in figure 13. θ (old) ) in the E step and then performing the M step using these samples. It is sometimes convenient to block some of the variables to improve mixing (Jensen. 1999.. Laird. This technique forms the basis of the popular software package for Bayesian updating with Gibbs sampling (BUGS) (Gilks.. & Kjærulff. the children and the children’s parents. This set of variables is known as the Markov blanket of x j . In many practical situations. 1995. To improve the convergence behaviour of EM.5. Ghahramani & Jordan. 1994). E step. 2. & Rubin. xv | θ)) p x h xv . namely to escape low local minima and saddle points. 1995. Utsugi. Convergence of this algorithm is discussed in Sherman. 3. x h }.

There are several annealed variants (such as SAEM) that become more deterministic as the number of iterations increases (Celeux & Diebolt. MCMC-EM algorithm. Godsill. At each iteration of the HMC algorithm. Figure 13. Then. 1987). (1987) and Neal (1996) focusing on the algorithmic details and not on the statistical mechanics motivation. 1985) and Monte Carlo EM (MCEM) when several samples are drawn (Wei & Tanner. This very useful idea was proposed in the physics literature (Swendsen & Wang. 2000).6. Here. we will focus on two well-known examples of auxiliary variable methods. To explain this in more detail. 1990). u (i) ) according to p(x. We describe here the “leapfrog” HMC algorithm outlined in Duane et al. u) and. ignoring the samples u (i) . we first need to introduce a set of auxiliary “momentum” variables u ∈ Rn x and define the . ANDRIEU ET AL.6. & Robert. Assume that p(x) is differentiable and everywhere strictly positive. Hybrid Monte Carlo. than from p(x). 3. namely hybrid Monte Carlo and slice sampling. it is possible to obtain marginal samples x (i) by sampling (x (i) . Auxiliary variable samplers It is often easier to sample from an augmented distribution p(x. where u is an auxiliary variable.24 C. The are also very efficient algorithms for marginal MAP estimation (SAME) (Doucet. The method is known as stochastic EM (SEM) when we draw only one sample (Celeux & Diebolt.1. One wishes sometimes that Metropolis had succeeded in stopping the proliferation of acronyms! 3. subsequently. u). one takes a predetermined number (L) of deterministic steps using information about the gradient of p(x). Hybrid Monte Carlo (HMC) is an MCMC algorithm that incorporates information about the gradient of the target distribution to improve mixing in high dimensions. 1992).

one obtains the Langevin algorithm. When only one deterministic step is used. which is a discrete time approximation of a Langevin diffusion process. The choice of the parameters L and ρ poses simulation tradeoffs. Choosing L is equally problematic as we want it to be large to generate candidates far from the initial state. u (i−1) ). u) = p(x)N u. HMC. Marginal samples from p(x) are obtained by simply ignoring u. In x . u). to allow a different step size ρ for each of the coordinates of x (Ishwaran. L = 1. Given (x (i−1) . In x ). while small values require many leapfrog steps (expensive computations of the gradient) to move between two nearby states. extended target density p(x. The Langevin algorithm is a special case of MH where the candidate satisfies x = x0 + ρu 0 = x (i−1) + ρ u + ρ x (i−1) /2 with u ∼ N (0. i. in practice. u) by starting with the previous value of x and generating a Gaussian random variable u. We then take L “frog leaps” in u and x. In the HMC algorithm. the algorithm proceeds as illustrated in figure 14.INTRODUCTION 25 Figure 14. therefore. 0. Next. 1999). . Large values of ρ result in low acceptance rates. Hybrid Monte Carlo.e. requires careful tuning of the proposal distribution. we need to introduce the n x -dimensional gradient vector (x) ∂ log p(x)/∂ x and a fixed step-size parameter ρ > 0. but this can result in many expensive computations. It is more efficient. The values of u and x at the last leap are the proposal candidates in the MH algorithm with target density p(x. we draw a new sample according to p(x.

Hence. & Smith. Wakefield. u L ) and define L p (x. The basic idea of the slice sampler is to introduce an auxiliary variable u ∈ R and construct an extended target distribution p (x. Slice sampling: given a previous sample. . u 1 . . 3. The full conditionals are of this augmented model are p(u | x) = U[0. . It is then straightforward to check that p (x.6. we sample a uniform variable u (i+1) between 0 and f (x (i) ). Higdon. The slice sampler. fl (x)] (u l ). 1999. . u) = 1 0 if 0 ≤ u ≤ p(x) otherwise. p(x) ≥ u}. as shown in figure 15. . One then samples x (i+1) in the interval where f (x) ≥ u (i+1) . & Walker. f(x(i) ) u (i+1) x (i+1) x (i) x Figure 15. It can be difficult to identify A. For example assume that L p(x) ∝ l=1 fl (x). u). Gelfand. ANDRIEU ET AL. 1998). such that p (x. 1991) is a general version of the Gibbs sampler. not necessarily densities.26 C. 1998. p(x)] (u) p(x | u) = U A (x) where A = {x. 1999. The slice sampler (Damien. . to sample from p(x) one can sample from p (x. u L ) ∝ l=1 I[0. Let us introduce L auxiliary variables (u 1 . It is then worth introducing several auxiliary variables (Damien. . Wakefield. If A is easy to identify then the algorithm is straightforward to implement. where the fl (·)’s are positive functions. Higdon. . Wakefield. u) du = 0 p(x) du = p(x). & Walker. . u) and then ignore u.2.

therefore. Up to this section. . Giudici & Castelo. . e. the number of u splines in a multivariate adaptive splines regression (MARS) model (Holmes & Denison. Algorithmic improvements and convergence results are presented in Mira (1999) and Neal (2000). 1997). . m = 1. u 1 . The distribution P(d x) will be assumed to admit a density p(x) with respect to a measure of interest.7. Then one can also check that p (x. . 1998). u L ) du 1 . . The acceptance ratio will now include the ratio of the densities and the ratio of the measures (Radon Nikodym derivative). we have to be more formal and compare distributions P(d x) = Pr(x ∈ d x) under a common measure of volume. du L ∝ l=1 I[0. Green & Richardson. Holmes & Mallick.INTRODUCTION 27 Figure 16. . . The slice sampler to sample from p (x. this issue). du L = l=1 fl (x). the number of sinusoids in a noisy signal (Andrieu & Doucet. . this issue). Typical examples include estimating the number of neurons in a neural network (Andrieu. 1998. Rios Insua & M¨ ller. the appropriate structure of a graphical model e (Friedman & Koller. . we attack the more complex problem of model selection. 1998). . 1999). The latter gives rise to a Jacobian term. N }. 1995. . . Instead. we have been comparing densities in the acceptance ratio. u L ) proceeds as shown in figure 16. . 2000). see for example (Escobar & West.g. . . & Doucet. the number of components in a mixture of factor analysers (Fokou´ & Titterington. if we are carrying out model selection. u 1 . However. . It is like trying to compare spheres with circles. . . fl (x)] (u l ) du 1 . Given a family of M models {Mm . Slice sampler. 2001. . xm ) as the invariant distribution. the number of levels in a changepoint process (Green. . we will focus on constructing ergodic Markov chains admitting p(m. 3. the number of components in a mixture (Richardson & Green. this issue). To compare densities point-wise. 1995). we need. to map the two models to a common dimension as illustrated in figure 17. the number of lags in an autoregressive process (Troughton & Godsill. For simplicity. . u 1 . 2001a. Lebesgue in the continuous case: P( d x) = p(x) d x. . du L = p(x) as L L p (x. we avoid the treatment of nonparametric model averaging techniques. Reversible jump MCMC In this section. this issue) or the best set of input variables (Lee. . u L )du 1 . then comparing the densities of objects in different dimensions has no meaning. de Freitas.

In particular one samples on a M much smaller union space X m=1 {m} × Xm . the probability of k being equal to m and x belonging to an infinitesimal set centred around xm is p(m. to Xm. we obtain the probability of being in subspace Xm . The parameters xm ∈ Xm (e. d x) = m=1 p(m. Recently. Xm ¯ ¯ and Xn .m . differentiable.n Xm × Um.m ) .n ¯ and Xn. d xm ). 1995). d xm )I{m}×Xm (k. Xm = Rn m ) are model dependent. Hence.n ) = f n→m (xn . It also requires the definition of a ¯ deterministic. . invertible dimension matching function f n→m between Xm. The full target distribution defined in this space is given by M p(k. x1 Compare both densities point-wise Univariate density p(x 1) p(x 1. u n. u n. To ensure a common measure.m ) = f n→m (xn . Green introduced a strategy that avoids this expensive search over the full product space (Green.g.x 2 ) x2 C. That is. it requires the extension of each pair of communicating spaces.m ). x u (xm . u n.x* ) Uniformly expanded density x* x1 Proposex * uniformly x1 Figure 17.28 Bivariate density p(x 1. f n→m (xn .m . By marginalisation.m Xn × Un. we first have to map the first model so that both models have common measure (area in this case). u m. x). to find the right model and parameters we could sample over the model indicator and the product space M m=1 Xm (Carlin & Chib. ANDRIEU ET AL. To compare a 1D model against a 2D model. Green’s method allows the sampler to jump between the different subspaces.n and Xn. 1995).

m ∼ U[0.INTRODUCTION 29 We define f m→n such that f m→n ( f n→m (xn . whose new location is µ= µ1 + µ2 2 The corresponding split move that guarantees reversibility.m ).n ) To illustrate this. we only perform the merge move if µ1 − µ2 < 2β. u n.n ) = f n→m (xn . deterministic transformation f m→n and proposal distributions for qn→m (· | n. uniformly at random.n | m. we move to (m.m β µ2 = µ + u n.m ). Note that to ensure reversibility.n ) . u n. p(n. p(k + 1. p(u n. The choice of the extended spaces. and accepting the move according to the probability ratio An→m = min 1.m ) where 1/k denotes the probability of choosing. For example we might want to estimate the locations and number of basis functions in kernel regression and classification. one of the k components. we could define a merge move that combines two nearby components and a split move that breaks a component into two nearby ones.m | n. for example. u m. xn ) x where xm = f n→m (xn . xn ). ensuring that we have reversibility (xm . xm ) by generating u n.m ∼ qn→m (· | n.1] . u m. xm ) × × × J fn→m .m β where β is a simulation parameter and. µk ) 1 k+1 1 k × 1 × Jsplit . u n. Here. xm ) q(n | m) qm→n (u m. assume that we are concerned with sampling the locations µ and number k of components of a mixture. xm ) is problem dependent and needs to be addressed on a case by case basis. xn ) q(m | n) qn→m (u n. involves splitting a randomly chosen component as follows µ1 = µ − u n. u n. The Jacobian is Jsplit = 1 ∂(µ1 . the number of mixture components in a finite mixture model. xn ) and qm→n (· | m. ∂(xm . or the location and number of segments in a segmentation problem. xn ). If the current state of the chain is (n.m ) −β 1 = 2β.m ) and J fn→m is the Jacobian of the transformation f n→m (when only continuous variables are involved in the transformation) J fm→n = det ∂ f n→m (xm . p(m. β . u m.m )) = (xn . u n. The merge move involves randomly selecting a component (µ1 ) and then combining it with its closest neighbour (µ2 ) into a single component µ. µk+1 ) × p(k. µ2 ) = ∂(µ. The acceptance ratio for the split move is Asplit = min 1. u n.

1997). However. In general. that is. death of a component and a simple update of the locations. m k .30 C. In fact. as shown in figure 18. Similarly. one often discards an initial set of samples (burn-in) to avoid starting biases. µk ) 1 k−1 1 k × Jmerge . ch. by no means exhaustive. The various moves are carried out according to the mixture probabilities (bk . Convergence and perfect sampling Determining the length of the Markov chain is a difficult task. one can apply several graphical and statistical tests to assess. In addition. The MCMC frontiers 4. time-consuming task.1. where Jmerge = 1/2β. 4. if the chain has stabilised (Robert & Casella. In practice. to the split and merge moves. roughly. p(k − 1. the problem with reversible jump MCMC is that engineering reversible moves is a very tricky. u k ). we present a. 8). sk . Several theoreticians have tried to bound the mixing time. 1999. summary of some of the available results. Figure 18. Reversible jump is a mixture of MCMC kernels (moves). none of these tests provide entirely satisfactory diagnostics. we could have other moves such as birth of a component.) If . the minimum number of steps required for the distribution of the Markov chain K to be close to the target p(x). In addition. (Here. µk−1 ) × p(k. Generic reversible jump MCMC. it is the flexibility of including so many possible moves that can make reversible jump a more powerful model selection strategy than schemes based on model selection using a mixture indicator or diffusion processes using only birth and death moves (Stephens. for the merge move. we have Amerge = min 1. ANDRIEU ET AL. dk .

1993). 1998). 1996). . Geometric bounds have also been obtained in general state spaces using the tools of regeneration and Lyapunov-Foster conditions (Meyn & Tweedie. we measure closeness with the total variation norm x (t) where = K (t) (· | x) − p(·) = 1 2 K (t) (y | x) − p(y) dy. If the state space X is finite and reversibility holds true. |λ|X | |) (Diaconis & Saloff-Coste. then the transition operator K (K f (x) = K (y | x) f (y)) is self adjoint on L 2 ( p). Jerrum & Sinclair. 1998. Nash) from differential geometry that allows us to obtain these bounds e (Diaconis & Saloff-Coste. hence. ≤ √ 2 p(x) where λ = max(λ2 . Poincar´ .S⊂X p(S) Intuitively. is the conductance of the Markov chain = min x∈S. This spectral decomposition and the Cauchy-Schwartz inequality allow us to obtain a bound on the total variation norm x (t) 1 λt . This classical result give us a geometric convergence rate in terms of eigenvalues. then the mixing time is τx ( ) = min{t : x (t )≤ for all t ≥ t}. such that K f i = λi f i . one can interpret this quantity as the readiness of the chain to escape from any small region S of the state space and. where f and g are real functions and we have used the bra-ket notation for the inner product f |g = f (x)g(x) p(x). There are several inequalities (Cheeger. The next logical step is to bound the second eigenvalue. 1996). make rapid progress towards equilibrium (Jerrum & Sinclair. one could use Cheeger’s inequality to obtain the following bound 1−2 where ≤ λ2 ≤ 1 − 2 2 . This implies that K has real eigenvalues 1 = λ1 > λ2 ≥ λ3 ≥ · · · ≥ λ|X | > −1 and an orthonormal basis of real eigenfunctions f i . That is. K f | g = f | Kg . For example.y∈S c p(x)K (y | x) 0< p(S)<1/2.INTRODUCTION 31 x (t).

Frieze.32 C. From a practical point of view. in many cases. & Vigoda. To avoid this problem. one should not allow adaptation to take place infinitely often in a naive way because this can disturb the stationary distribution.. some results tell us. & Kannan. In particular. That is. 1999). 1996).. Sampling from log-concave distributions (Applegate & Kannan. Computing the permanent of a matrix (Jerrum. The first is based on the idea of running several chains in parallel and using sampling-importance resampling . we notice that the chain stays at each state for a long time. that ensures a suitable acceptance rate. these algorithms are still limited and. for example. 3. we violate the Markov property of the transition kernel. This tells us that we should reduce the variance of the proposal distribution. A remarkable recent breakthrough was the development of algorithms for perfect sampling. x (1) . ANDRIEU ET AL. that it is not wise to use the independent Metropolis sampler in high dimensions (Mengersen & Tweedie. The two major players are coupling from the past (Propp & Wilson. thereby escaping the exponential curse of dimensionality. That is. Polynomial time sampling algorithms have been obtained in the following important scenarios: 1. . a problem that manifests itself in many ways in machine learning (e. stereo matching and data association). and then use standard MCMC simulation to ensure convergence to the right distribution. The last problem is equivalent to sampling matchings from a bipartite graph. . 1996). However. p(x (i) | x (0) . Already. one should use the information in the samples to update the parameters of the proposal distribution so as to obtain a distribution that is either closer to the target distribution. x (i−1) ) no longer simplifies to p(x (i) | x (i−1) ). one could carry out adaptation only during an initial fixed number of steps. These mathematical tools have been applied to show that simple MCMC algorithms (mostly Metropolis) run in time that is polynomial in the dimension d of the state space. . However. where the stationary distribution is disturbed despite the fact that each participating kernel has the same stationary distribution. computationally inefficient. or that minimises the variance of the estimator of interest. Computing the volume of a convex body in d dimensions. Ideally. 1991). Gelfand and Sahu (1994) present a pathological example.2. This problem arises because by using the past information infinitely often. some steps are being taken towards obtaining more general perfect samplers.g. for example perfect slice samplers (Casella et al. where d is large (Dyer. Adaptive MCMC If we look at the chain on the top right of figure 7. . These algorithms are guaranteed to give us an independent sample from p(x) under certain restrictions. Although the theoretical results are still far from the practice of MCMC. Sampling from truncated multivariate Gaussians (Kannan & Li. 1998) and Fill’s algorithm (Fill. Two methods for doing this are presented in Gelfand and Sahu (1994). 4. Sinclair. one would like to automate this process of choosing the proposal distribution as much as possible. 4. 1998). 2000). 2. they will eventually provide better guidelines on how to design and choose algorithms. 1991).

p(xt | x0:t−1 . the SMC update is accomplished by introducing an appropriate importance proposal distribution q(x0:t ) from which we can obtain samples. y1:t−1 ) p(yt | x0:t . In the SMC setting. & Yu. the states and the observations up to time t. Sequential Monte Carlo and particle filters Sequential Monte Carlo (SMC) methods allow us to carry out on-line approximation of probability distributions using samples (particles). Computational simplicity in the form of not having to store all the data might also constitute an additional motivating factor for these methods. Furthermore. approximately distributed according to the distribution p(x0:t−1 | y1:t−1 ). & Sahu. Our aim is to estimate recursively in time the posterior p(x0:t | y1:t ) and its associated features including the marginal distribution p(xt | y1:t ). however. yt }. unfortunately. However. so that information from the recent past is given greater weighting than information from the distant past. They are very useful in scenarios involving real-time signal processing. where data arrival is inherently sequential. a dynamic model and measurement model p(x0 ) p(xt | x0:t−1 . Tierney. There are. The samples are then appropriately weighted. . 1999). xt } and y1:t {y1 .INTRODUCTION 33 (Rubin. one might wish to adopt a sequential processing strategy to deal with non-stationarity in signals. A similar method that guarantees a particular acceptance rate is discussed in Browne and Draper (2000). one uses an approximation to the marginal density of the chain as proposal. 1998. . y1:t−1 ) = p(yt | xt ). this assumption is not necessary in the SMC framework. . known as the filtering distribution. The second method simply involves monitoring the transition kernel and changing one of its components (for example the proposal distribution) so as to improve mixing. y1:t−1 ) for t ≥ 1 for t ≥ 1 We denote by x0:t {x0 . including delayed rejection (Tierney & Mira. 1988) to multiply the kernels that are doing well and suppress the others. parallel chains (Gilks & Roberts. 4. at time t. respectively. SMC (i) N methods allow us to compute N particles {x0:t }i=1 approximately distributed according to the posterior p(x0:t |y1:t ).3. inefficient in many ways and much more research is required in this exciting area. 1995). a few adaptive MCMC methods that allow one to perform adaptation continuously without disturbing the Markov property. Since we cannot sample from the posterior directly. . . 1996) and regeneration (Gilks. Mykland. Given N particles {x0:t−1 }i=1 at time t − 1. In this approach. . . These methods are. Roberts. . . and the expectations I ( f t ) = E p(x0:t |y1:t ) [ f t (x0:t )] (i) N A generic SMC algorithm is depicted in figure 19. Note that we could assume Markov transitions and conditional independence to simplify the model. we assume that we have an initial distribution. y1:t−1 ) = p(xt | xt−1 ) and p(yt | x0:t .

N −1 }. This results in the weighted measure {˜ t−1 . Finally.34 C. For filtering purposes. Simple SMC algorithm at time t. Figure 19. ANDRIEU ET AL. which yields an approximation x ˜ p(xt−1 | y1:t−1 ). the bootstrap filter starts at time t − 1 with an unweighted measure {˜ t−1 . the sampling (prediction) step (i) introduces variety. there is no need for storing or resampling the past trajectories. x which provides an approximation of p(xt−1 | y1:t−2 ). wt−1 }. Subsequently. N −1 }. x (i) Figure 20. which is still an approximation of p(xt−1 | y1:t−1 ). which is an approximation of p(xt | y1:t−1 ). resulting in the measure {˜ t . In this example. N −1 }. . the resampling step selects only the “fittest” particles to obtain the unweighted (i) measure {xt−1 . For each particle we compute the importance weights using (i) (i) the information at time t − 1.

1999.INTRODUCTION 35 (i) N In generic SMC simulation. one might still encounter difficulties if the ratio of the first two terms of Eq... Pitt & Shephard. van der Merwe et al. 1995) and the bootstrap filter (Gordon. (22). In doing so. one needs to extend the current paths {x0:t−1 }i=1 to obtain (i) N ˜ ˜ new paths {x0:t }i=1 using the proposal distribution q(x0:t |y1:t ) given by ˜ q(x0:t | y1:t )} = ˜ q(x0:t | x0:t−1 . Consequently ˜ ˜ q(x0:t | y1:t ) = p(x0:t−1 | y1:t−1 )q(xt | x0:t−1 . such that i=1 Ni = N . must be weighted by the importance weights wt = ˜ ˜ p(x0:t−1 | y1:t ) p(xt | x0:t−1 . (22) differs significantly from 1 (Andrieu. y1:t−1 ) in which case the importance weights are given by the likelihood function ˜ wt ∝ p (yt | xt ) . (When using this proposal. ˜ (i) After the importance sampling step. 1999). and leave the past trajectories intact. we note that the optimal importance distribution is ˜ ˜ q(xt | x0:t−1 . de Freitas et al. y1:t ) p(x0:t−1 | y1:t−1 ) d x0:t−1 . one can adopt suboptimal filters and other approximation methods that make use of the information available at time t to generate the proposal distribution (Doucet. 2001. y1:t−1 ) ∝ . y1:t ). This selection step is what . in some restricted situations. To make this integral tractable. 2000). & Smith. y1:t ) = p (xt | x0:t−1 . 2001). For instance. 2000. ˜ qt (xt | x0:t−1 . y1:t ) p(x0:t | y1:t ) = ˜ ˜ q(x0:t | y1:t ) p(x0:t−1 | y1:t−1 ) q(xt | x0:t−1 .. One can adopt. Pitt & Shephard. In fact. Doucet. & Punskaya. y1:t ) The samples from q(·). Salmond. instead. y1:t ) (22) From Eq. including condensation (Isard & Blake. The importance sampling framework allows us to design more principled and “clever” proposal distributions. 2000. y1:t ) = p(xt | x0:t−1 . we only propose to modify the particles at time t. y1:t ) ˜ ˜ p (yt | xt ) p (xt | x0:t−1 . survival of the fittest (Kanazawa.) The optimal importance distribution can be difficult to evaluate. one may interpret the likelihood as a distribution in terms of the states and sample from it directly. This simplified version of SMC has appeared under many names. say Ni ∈ N. 1993). the importance weights become equal to the transition prior (Fox et al. & Andrieu. a selection scheme associates to each particle x0:t N a number of “children”. 1996). Koller. Godsill. & Russell. the transition prior as proposal distribution ˜ ˜ q(xt | x0:t−1 .

1996) can also result in great improvements. in batch MCMC simulation it is often not possible to load the entire dataset into memory. the conditional posterior distribution p(v | u) is analytically tractable. & Doucet. 2001).. 4. ANDRIEU ET AL. 1996) and Gaussian processes (Barber & Williams. as discussed in Section 2. Then we can easily marginalise out v from the posterior. Gradient optimisation is inherent to Langevin algorithms and hybrid Monte Carlo. such as the Gibbs sampler. 1998. The machine learning frontier The machine learning frontier is characterised by large dimensional models. massive datasets and many and varied applications. MH algorithm and reversible jump MCMC. into the filtering framework. the new particles might have been moved to more interesting areas of the state-space. 1999). Despite the auspicious polynomial bounds on the mixing time. Information about derivatives of the target distribution also forms an integral part of many adaptive schemes. An important feature of the selection routine is that its interface only depends on particle indices and weights. There are various selection schemes in the literature. 1997). and only need to focus on sampling from p(u). we sample u (i) ∼ p(u) and .. MacEachern.g. allows us to track moving target distributions efficiently by choosing the fittest particles. such that p(x) = p(v | u) p(u) and. the total variation of the current distribution with respect to the invariant distribution can only decrease. That is. In fact. The combination of sampling algorithms with either gradient optimisation or exact methods has proved to be very useful. u and v. then applying a Markov chain transition kernel K (x0:t | x0:t ). it has been argued that the combination of MCMC and variational optimisation techniques can also lead to more efficient sampling (de Freitas et al. Gilks & Berzuini. 1999). The simplicity of the coding of complex models is. This enables one to implement variable and model selection schemes straightforwardly. Recently. de Freitas. It is also possible to introduce MCMC steps of invariant distribution p(x0:t | y1:t ) on each particle (Andrieu. That is. Note that we can incorporate any of the standard MCMC methods. but there is still great room for innovation in this area. These algorithms have been shown to work with large dimensional models such as neural networks (Neal. Clyde. indeed. parameters. & Liu. 1999. conditional on u. 2001). & Gordon. However.36 C. but their performance varies in terms of var[Ni ] (Doucet. variables. it can be treated as a black-box routine that does not require any knowledge of what a particle represents (e. The basic idea is that if the particles are distributed according to the posterior distribution p(x0:t | y1:t ). Suppose we can divide the hidden variables x into two groups. but we no longer require the kernel to be ergodic. A few solutions based on importance sampling have been proposed recently (Ridgeway. one of the major advantages of these algorithms. still results in a set of particles distributed according to the posterior of interest. However.3. models). The combination of exact inference with sampling methods within the framework of RaoBlackwellisation (Casella & Robert. Massive datasets pose no problem in the SMC context. by applying a Markov transition kernel.4. de Freitas. with invariant distribution p(· | y1:t ) such that K (x0:t | x0:t ) p(x0:t | y1:t ) = p(x0:t | y1:t ). it is an arduous task to design efficient samplers in high dimensions. which lies in a space of reduced dimension.

. 2002) and several papers in this issue. conditionally Gaussian models (Carter & Kohn. 1996. 1998) and MARS (Holmes & Denison. Some areas that are already benefiting from sampling methods include: 1. Speech and audio processing. 1994. 1997). restoration of old movies (Morris. The problem of how to automatically identify which variables should be sampled. Neural networks and kernel machines (Andrieu. Computer vision. By sampling only the artificial variables. Fitzgerald. 5. the rest of the problem can often be solved easily using exact algorithms such as Kalman filters. 2001b. see also Holmes and Denison (this issue). There are great opportunities for combining existing sub-optimal algorithms with MCMC in many machine learning problems. 1996. 6. 1999). 1995. Wilkinson & Yeung. stereo matching (Dellaert et al. Signal enhancement (Godsill & Rayner. Probabilistic graphical models. De Jong & Shephard. Doucet. & Doucet. 1999. and which can be handled analytically is still open. 2001a. Lemieux. 1993.. For example (Gilks. Web statistics. Tracking (Isard & Blake. the original model decouples into simpler. 1997) and sampling plausible solutions to multi-body constraint problems (Chenney & Forsyth. 1998.INTRODUCTION 37 then use exact inference to compute p(v) = 1 N N p v u (i) i=1 By identifying “troublesome” variables and sampling them. 2001). Tu & Zhu. 3. 2000). This strategy allows one to map probabilistic classification problems to simpler regression problems. Neal. 1998). & Doucet. proportions belonging to specific domains and the average size of web pages (Bar-Yossef et al. 1998. Regression and classification. 2000). & Smith. augmentation strategies and blocking) and on careful design of the proposal mechanisms. this issue). Andrieu. An interesting development is the augmentation of high dimensional models with low dimensional artificial variables. CART (Denison. this issue). Estimating coverage of search engines. 2. & Fleet. Light transport (Veach & Guibas. Thomas. HMMs or junction trees. & Spiegelhalter. Mallick. Other application areas include dynamic Bayesian networks (Doucet et al. & Kjærulff. 2002).. Wilkinson & Yeung. The design of efficient sampling methods most of the times hinges on awareness of the basic building blocks of MCMC (mixtures of kernels. 1995. The latter requires domain specific knowledge and heuristics. 1996) and segmentation (Clark & Quinn. 1994. Vermaak et al. 2000). Computer graphics. 2000. 4.. For example. de Freitas. Kong. & Kokaram. M¨ ller & Rios u Insua. one can apply this technique to sample variables that eliminate loops in graphical models and then compute the remaining variables with efficient analytical algorithms (Jensen.. colour constancy (Forsyth. 1998). 1999). 1998) and model averaging for graphical models (Friedman & Koller. Ormoneit. Wood & Kohn. this issue). Kam. Holmes & Mallick. 2001). de Freitas. more tractable submodels (Albert & Chib. Gaussian processes (Barber & Williams.

C. N. Technical Report CUED/F-INFENG/TR 346. Venkatesh.. 1997) and linkage analysis (Jensen. Sequential MCMC for Bayesian model selection. Neural Computation. Caesarea. Bayesian analysis of binary and polychotomous response data. 2359–2407. Advances in neural information processing systems 12 (pp. San Mateo. C. Robust full Bayesian methods for neural networks. 7. In IEEE Higher Order Statistics Workshop. Andrieu. 1999) and multitarget tracking (Bergman. Andrieu. A. 2667–2676. DNA microarray data (West et al. L. A.. 11–18). First order probabilistic logic.. A.. de Freitas. 1995). de Freitas. & Doucet. M¨ ller (Eds. 8. & Kjærulff.. Robust full Bayesian learning for radial basis networks. IEEE Transactions on Communications. 2000). & Doucet. T.. (1995). de Freitas. (1999). 669–679. A. In S. K. Solla. 2001). including tempering and coupling. For conciseness. 1999). (2000a). Leen. 2001). Sekhar Tatikonda and Mike Titterington. J. 2975–2985. Monte Carlo. we advise the readers to consult the references at the end of this paper. K. Anderson. Kong. Vehicle matching in highway systems (Pasula et al. & K. Cambridge University Engineering Department. C. M.. Israel (pp. A. Acknowledgments We would like to thank Robin Morris. u MIT Press. Metropolis.... Journal of the American Statistical Association. Breyer. (Pasula & Russell. 88:422. Convergence of simulated annealing using Foster-Lyapunov criteria.. (1999). 1999). Decision theory. u 9. 11. Salmond & Gordon. 2001) and mixtures of factor analysers (Fokou´ & Titterington. Andrieu. C. 2000. de Freitas. we have skipped many interesting ideas. 379–385). and the MANIAC. (1986). Mixtures of independent factor analysers (Utsugi. Mark Paskin.). H. 2001). Andrieu. In Uncertainty in artificial intelligence (pp. Genetics and molecular biology. . 12. & Doucet. For more details.. ANDRIEU ET AL. Robot localisation and map building (Fox et al. & Doucet. N. 130–134). Data association. 96–108. (2001b).38 C.. A. & Rios Insua. cancer gene mapping (Newton & Lee. Andrieu. Advances in Neural Information Processing Systems (NIPS13). C.. M¨ ller. 47:10.. A. 1999) and influence diagrams (Bielza. & West. this issue). 10. 2001).. A. A. Andrieu. (1999). A. C. Joint Bayesian detection and estimation of noisy sinusoids via reversible jump MCMC. References Al-Qaq. (1993). N. CA: Morgan Kaufmann. de Freitas. Rao-blackwellised particle filtering via data augmentation. & Chib. Partially observable Markov decision Processes (POMDPs) (Thrun. Kevin Murphy. e We hope that this review will be a useful resource to people wishing to carry out further research at the interface between MCMC and machine learning. (2000b). C. & Doucet.-R. & Townsend. protein alignment (Neuwald et al. W.. (2001a). abstract Markov policies (Bui. Andrieu. 13:10. Reversible jump MCMC simulated annealing for neural networks. J.. N. & Doucet. L. Stochastic gradient optimization of importance sampling for the efficient simulation of digital communication systems. 43:12. 14. & Doucet. Devetsikiotis. Classical mixture models. Los Alamos Science. IEEE Transactions on Signal Processing.. Albert. Robotics. N. S..

de Freitas. 120–127). Statistical Science. Rao-Blackwellisation of sampling schemes. Communications of the ACM. Jordan. P. M. 995–1007. Gee. J.. 541–553. Shao. Celeux. (1995).. C. H. G. J. H. J. Annals of Mathematical Statistics. 955–993. C. Sweden. (2000). Doucet. 73–82. de Freitas.. 6. (1999). G. Decision Analysis by Augmented Probability Simulation.. The Metropolis algorithm. Cheng. 69–100. (1985).. R. J. Robert. (1997). N. Bergman.. Perfect slice samplers for mixtures of distributions. (1970). Jordan.. Bayesian Model choice via MCMC. The SEM algorithm: A probabilistic teacher algorithm derived from the EM algorithm for the mixture problem. Berlin: Springer-Verlag. S. 2. & Chib. (1994). I.. Journal of Artificial Intelligence Research..INTRODUCTION 39 Andrieu. (1996). (1982).. Sequential Monte Carlo methods for optimal filtering. M. 127–146. 535–544). Stochastics and Stochastics Reports. (Eds. Speech. & West. P.. Recursive Bayesian estimation: Navigation and tracking applications. B.. K. D. A data-driven Bayesian sampling scheme for unsupervised image segmentation. AIS-BN: An adaptive importance sampling algorithm for evidential reasoning in large bayesian networks. & N. Gordon (Eds. Casella. Department of Electrical Engineering. Celeux. A. Brooks. N. C. 83:1. & Robert. Z.. Venkatesh. Petrie. N. 49–63. 3497– 3500). In J. J.. Green. A. Gibbs sampling for Bayesian non-conjugate and hierarchical models by auxiliary variables. Advances in neural information processing systems 9 (pp. Sequential Monte Carlo methods to train neural network models. u Management Science. (1999). 61:2. & Diebolt. S.). Department of Biometrics. P. H. (1998). Besag. 391–420.. Fakcharoenphol. S. D. M¨ ller. J.. o Berners-Lee. Højen-Sørensen... Mozer. A. (2000). I. (2000). & Kannan. Sequential Monte Carlo methods in practice. Hidgon. D. Exactly solved models in statistical mechanics. Soules. Sampling plausible solutions to multi-body constraint problems. 57. & Druzdzel.. (2000). The World-Wide Web.. G. Carlin. Neural Computation.D. G.. & Titterington. 331–344.. (2001)... Monte Carlo methods for Bayesian computation. I. Beichl. Computing in Science & Engineering. & Draper. & Kohn. Ph. 12:4.. 41. Computational Statistics. Carter.. 45:7. 15.. & Russell. 47:1. In International Conference on Very Large Databases (pp. Biometrika... Damien. G. & Mengersen. Computational Statistics Quarterly. D. & Sullivan. 219–228). H.. S. Koller (Eds. J. 10:4. The Statistician. In IEEE International Conference on Acoustics. Applegate. L. M.. D.. In National Conference on Artificial Intelligence (AAAI-2000). Arizona (Vol.. (1988).. M. . C. F. Cailliau. 156–163). 340–346). Journal of the Royal Statistical Society Series B. & T. Structural Safety. 5. M. G. MA: MIT Press. Adaptive sampling—An iterative fast Monte Carlo procedure. On Gibbs sampling for state space models. & Williams. W. J. J. Implementation and performance issues in the Bayesian and likelihood fitting of multilevel models. J. & Diebolt. Journal of the Royal Statistical Society B. N.) (2001).). Q. D. (2001). 3–66. Breese & D. F. H. Wakefield. & Secret. Bui. E. 164–171. M. N. K. A. S. 81:3.. & Rios Insua. 65–69. 13. P. 41. Markov chain Monte Carlo method and its application. D. T. G. (1994).). Barber. & Quinn. A. 155–188. Uncertainty in artificial intelligence (pp.. R. 2:1. C. 473–484. A stochastic approximation type EM algorithm for the mixture problem. A. Biometrika. J. Chien. 10:1. 119–126.. R. CA: Morgan Kaufmann. Browne. In M.. (1991). Clark. Casella. Baxter. Baum. Sampling and integration of near log-concave functions. Chenney. J. San Diego. pp.. (1992). Bucher. A. Petsche (Eds.. & Ibrahim. In Proceedings of the Twenty Third Annual ACM Symposium on Theory of Computing (pp. Bar-Yossef. SIGGRAPH (pp.. L. G. C. Thesis. & Walker. Link¨ ping University. Nielsen. Bielza. D. de Freitas. P. Cambridge. CA: Academic Press. S. A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains. Gaussian processes for Bayesian classification via hybrid Monte Carlo. & Forsyth. Bayesian computation and stochastic systems. Berlin: SpringerVerlag. I. (1999). Luotonen. (1999). Approximating aggregate queries about web pages via random walks. Chen. (1999). Mengersen. P. K. A. K. On the recognition of abstract Markov policies. R. Cornell University. Technical Report BU-1453-M. Berg. S. P. E. M. and Signal Processing. T. & Punskaya.. (1999). C. (2000). P. Niranjan.. 81–94. Variational MCMC. San Matio. (1995). E. In A Doucet. & Doucet. (2000). & Weitz... & Weiss. C.

Kennedy. (Eds... A. Advances in neural information processing systems 7 (pp. K. E.. R. D. Thrun. (2000). A.. E. 6:6.). (2000). 20–36. Rao blackwellised particle filtering for dynamic Bayesian networks. On Markov chain Monte Carlo acceleration. & Jordan. D. Cambridge University Engineering Department. 39. J. Efficient sampling from the smoothing density in time series models. Gilks. 176–183). N.. 24.. & Sahu. & Geman. In A. Journal of the ACM.. Doucet. 763–769. 216–222. Sequential Monte Carlo methods in practice.. 197–208. A. P. D. Bayesian density estimation and inference using mixtures. 339–350. A. 90. Fox. Gordon (Eds. C. Physics Letters B. J.. Statistics and Computing. 93. Geman.. Monte Carlo inference for dynamic Bayesian models. J... P. G. On sequential simulation-based methods for Bayesian filtering. D. An interruptible algorithm for perfect sampling via Markov chains. K. Gilks. W. 85. & D. (2001). N.) (2001). & Spiegelhalter. & Roberts. & Kannan. Marginal maximum a posteriori estimation using MCMC. N.. & Sahu... R.. (1997). Chapman & Hall. 3. On sequential Monte Carlo sampling methods for Bayesian filtering. Godsill. 57. Richardson. P. 1–17. S. B.. (1994).. 1:38.). In IEEE Conference on Computer Vision and Pattern Recognition (pp.. T. Technical Report 9502. In G. 721–741. Journal of the Royal Statistical Society Series B. 195:2. (1995). S. A. G.. Escobar. (1998). (1996). M.. Stochastic relaxation. R. Gilks. W. (1987). Hybrid Monte Carlo. 398–409. (1990). J. Berlin: Springer-Verlag. F.. MA. A. W. Eckhard. In W. Medical Research Council. Doucet. K. & Saloff-Coste. (1989). D. M. Dyer. Godsill. & Dellaert. 300–305). R. J. J. ANDRIEU ET AL. Z. Doucet. Gilks. Markov chain Monte Carlo in practice. & Andrieu. A. Gilks. D. Unpublished. & Roweth. Doucet. De Jong. W. (1999). & J. Adaptive Markov chain Monte Carlo through regeneration. Technical Report CUED/FINFENG/TR 310. C. resampling and colour constancy. L. Biometrika. What do we know about the Metropolis algorithm? Journal of Computer and System Sciences. Touretzky. & West. Gelfand. N. Gelfand. J.. R. Boutilier & M. Berlin: Springer-Verlag. In C. Frieze. A. The Annals of Applied Probability. Alspector (Eds. Uncertainty in artificial intelligence (pp. Suffolk: Chapman and Hall... G. Stan Ulam. & Russell. A. O. Journal of the American Statistical Association. Spiegelhalter (Eds. Gibbs distributions and the Bayesian restoration of images.. N. K. S. 1317–1399. Particle filters for mobile robot localization. Fill. 10:3. (1995). (1998). Technical Report CUED/F-INFENG/TR 375.. Factorial hidden Markov models. O. (Eds. 1–38. 15. Cambridge. 363–377. IEEE Transactions on Pattern Analysis and Machine Intelligence. (1987). Geweke. Markov chain Monte Carlo in practice (pp. (1991). S. Department of Engineering. & N. Roberts.). (1995). R. (1984). . Econometrica.. Laird. Factorial learning and the EM algorithm. de Freitas. Pendleton. Ghahramani. Maximum likelihood from incomplete data via the EM algorithm. S. P. R. B. Murphy. Godszmidt (Eds.. UK. 261–276. 8:1. Biometrika. Z. Bayesian inference in econometric models using Monte Carlo integration. M. 577–588. C. 85:410. F. Diaconis. D. 82:2. M. Tesauro. A Bayesian CART algorithm. A. & Berzuini. M. MIT Artificial Intelligence Lab. S. Burgard. & Robert. D. Dempster. A. S.40 C. 131–162. S. & Smith. A. Sampling. & Rubin. Journal of Computational and Graphical Statistics. (1995). B. Strategies for improving MCMC. M. Cambridge University. Ghahramani. Sampling-based approaches to calculating marginal densities. (1998). Doucet. 617–624). D. W. & Gordon.) (1996). S. M. (1998). F..). 89–114). (2000).. Richardson. Los Alamos Science. (1998). Journal of the American Statistical Association. de Freitas. & Smith. N. & Shephard. (1998). Mallick. S. J. A. de Freitas. S. Forsyth. A. Morgan Kaufmann Publishers.. A random polynomial-time algorithm for approximating the volume of convex bodies.. Doucet. D. Journal of the American Statistical Association. Denison.. John Von Neumann and the Monte Carlo method. Duane. Sequential Monte Carlo methods in practice. 131–136.

& Whitlock. (2000). S. Contour tracking by stochastic propagation of conditional density. Hastings. M. M. 204–212).. Canadian Journal of Statistics. E. A language and program for complex Bayesian modelling. H. 585–595. Ishwaran. Implementations of the Monte Carlo EM algorithm. B. 42. IEEE. 779–799. Metropolis. The Annals of Statistics. A. Sinclair. W.. H. (1996).) (1998).. Technical Report TR00-079. W. Monte Carlo strategies in scientific computing. R. . 343–356). A.. In European Conference on Computer Vision (pp.. Hochbaum (Ed. A general multiscale scheme for unsupervised image segmentation. (2000). E. Jerrum. A. 346–351). Approximation algorithms for NP-hard problems (pp. 866–893. S. & Teller. Cambridge. (1998). Metropolis. S. P. Rosenbluth.. (2001). (1970). Journal of Chemical Physics. J. K. 711–732. Kirkpatrick. G. 8. M. Green. S.. & Casella. Rosenbluth. Bayesian radial basis functions of variable dimension.. H. R.. UK. E. 671– 680. P. & Mallick. 21. Journal of Computational and Graphical Statistics. In D. C. M. P. Electronic Colloquium on Computational Complexity. Rates of convergence of the Hastings and Metropolis algorithms.. (1996).. W. In 37th Annual Symposium on Foundations of Computer Science (pp. D.). International Journal of Human-Computer Studies. Markov chains and stochastic stability. C. Teller. Morgan Kaufmann. (1998). Koller. Equations of state calculations by fast computing machines. K. The Markov chain Monte Carlo method: an approach to approximate counting and integration. R. J. Salmond. 97–109. J. 23. Neural Computation. A. U. Kong. Monte Carlo sampling methods using Markov chains and their Applications. Reversible jump Markov chain Monte Carlo computation and Bayesian model determination.INTRODUCTION 41 Gilks. Sequential importance sampling for nonparametric Bayes models: The next generation. & Tweedie. (Ed. 107–113. (1991). 335–341. The Statistician. S. Cambridge University.. Cambridge. (1949). 647–666. E. Berlin: Springer-Verlag. Levine. Maximum likelihood variance components estimation for binary data. Thesis. 251–267. A. & Blake. N.. & Kjærulff. (1953). Jerrum.. 24. 10:3. M. A... Optimization by simulated annealing. (2000). 1217–1233. J. New York: Springer-Verlag.D. (1993). Sampling according to the multivariate normal density. N. M. K. D. (1986). D. J. J. S. & Smith. (1994). S. (1994). & Sinclair. Ph. & Sacksman. 220. Berlin: Springer-Verlag. Gordon. & Vigoda. H. G. Journal of the American Statistical Association. McCulloch. Green. S. 27. 43. 482–519). (1995). (1995). Modelling heterogeneity with and without the Dirichlet process. & Vecchi. M. J. 101–121. J. & Rayner. Kannan. Bristol University. F. IEE Proceedings-F. Kalos.. A. R.. Godsill. & Li. & Russell. L. In Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence (pp. 422–440. C. Thomas. Holmes. N... Novel approach to nonlinear/non-Gaussian Bayesian state estimation. (1995).. M. Higdon. New York: John Wiley & Sons. Clyde. Biometrika 57. P. (1983). N. UK. Stochastic simulation algorithms for dynamic probabilistic networks.) (2001). Department of Engineering. 82. & Liu. Kanazawa. N. D. 89:425. & Tweedie. Gelatt. D.. (1999). Department of Statistics. R. 93:442. 140:2. PWS Publishing. Monte Carlo methods. Application of hybrid Monte Carlo to Bayesian generalized linear models: Quasicomplete separation and neural networks. A polynomial-time approximation algorithm for the permanent of a matrix. (1996). Journal of American Statistical Association. Journal of the American Statistical Association. H.. The Monte Carlo method. Journal of Computational and Graphical Statistics. S.. Jensen. Science. & Ulam. Blocking-Gibbs sampling in very large probabilistic expert systems. P. Digital audio restoration: A statistical model based approach. L.. & Richardson. Liu. Mengersen. (1996).. MacEachern. J. Meyn. Auxiliary variable methods for Markov chain Monte Carlo with application. A. Simulated annealing process in general state space. (1999). & Spiegelhalter. L. Biometrika. (Eds. 169–178. C.. Kam.. C. (1993). K. Isard. M. W. 330–335. A. A. Haario.. 1087–1091. 10:5. S. P. 44:247. Advances in Applied Probability.. S.

Journal of the Royal Statistical Society B. . M. Lipman. & Rios Insua. Russell. In A. & Gordon. J.). R. New York: Springer-Verlag. & Shephard. Rios Insua. Gordon (Eds. Adaptive importance sampling for performance evaluation and parameter optimization of communications systems. Department of Statistics. F.. S.. & Russell. H. 181–191). (1998). (2001). Lattice particle filters. D. Microsurveys in discrete probability. In J. Using the SIR algorithm to simulate posterior distributions. E. & Tattje. Uncertainty in artificial intelligence (pp.. Richardson.). 90. Coupling from the past: a user’s guide. M. D. Artificial Intelligence. D. A. Lindley.). Smith (Eds. (1997). M¨ ller. 2005. University of Toronto. H. S. J.. Sequential Monte Carlo methods in practice. D. P. P. 1665–1677. 95–110. University of Toronto. T. M.D. Practical nonparametric and semiparametric bayesian statistics (pp.. Generalization of boosting algorithms and applications of bayesian inference for massive datasets. u M¨ ller. Ormoneit. Ordering.D. Monte Carlo statistical methods. A sampling based approach to line scratch removal from motion picture frames. & D.. Technical Report CRG-TR93-1. Seattle. Liu. On Bayesian analysis of mixtures with an unknown number of components.. 245–257. (1999). 94:446. 56. G. D. Cambridge. Dept. 59:4. Y.. Bayesian statistics 3 (pp. Pearl. G. B. Peskun. H. (1997).. (2000). Srinivasan. P. In C. DeGroot.. Department of Statistics. 60:3.. Mykland.. (2001). University of Washington. Geometric convergence and central limit theorems for multidimensional Hastings and Metropolis algorithms. Roberts. L. Journal of the American Statistical Association. Adaptive importance sampling for estimation in structured domains. (1999). (1999). In D. M. ANDRIEU ET AL.. Berlin: Springer-Verlag. F.42 C. In IEEE International Conference on Image Processing (pp. Uncertainty in artificial intelligence. Godszmidt (Eds. L. R. (1998). (1996). N. (2000). CA: Morgan Kaufmann Publishers. F. Neuwald. R. Ostland. K. C. Simulation and the Monte Carlo method.. & A. J. M. & M. van Etten. C. Tierney... A. In D. slicing and splitting Monte Carlo Markov chains. (1999). A. In International Joint Conference on Artificial Intelligence. E.. P. S. School of Statistics. J. In International Joint Conference on Artificial Intelligence. New York: John Wiley and Sons. R. J.. P. Stockholm. E. Technical Report No. R. H. 32. 801–804). & Yu. P. L. R. & Ritov. C. R. San Mateo. & J. J. Pitt. Biometrika. 590–599. MA: Oxford University Press. Motwani. Rubinstein.). New York: Springer-Verlag. Thesis. Probabilistic inference using markov chain monte carlo methods. Biometrics. (1973). 25:9.). San Mateo. 731–792. Pasula. 83. C.. (1998). Neural Computation. (1999). & Green. & M¨ ller. D. u Springer Verlag. (1998). G. Neal. & Fleet. (1996). Y. Ortiz. Doucet.. of Computer Science. Rubin. & Wilson. Issues in Bayesian analysis of neural network models. de Freitas. CA: Morgan Kaufmann. 607–612. 395–402). L. (1998). Tracking many objects with many sensors. (1996). R. P. N. & Tweedie.) (1981). & Lee. C... V. M. Dey. Newton.. M. (Eds. M. H.. Neal. Propp (Eds. Brin. S. S. Inferring the location and effect of tumor suppressor genes by instability-selection modeling of allelic-loss data. N. (2000). Regeneration in Markov chain samplers. Ph. Mira. 1088–1097. & Kokaram. 233–241. B. Aldous.. IEEE Transactions on Communications. University of Minnesota. (1993). Pasula. 118. 557–565. Nucleic Acids Research. K. Propp. Y. Salmond. Robert. Page. V. 48:4. Ridgeway.. (1995). Lemieux. Evidential reasoning using stochastic simulation. Ph. Optimum Monte-Carlo sampling using Markov chains. P. Extracting protein alignment models from the sequence database. 446–454). Boutilier. The PageRank citation ranking: Bringing order to the Web. D. D. Fitzgerald. (2000).. W. Particles and mixtures for tracking and guidance. & Casella. Journal of the American Statistical Association. Bernardo. Remondo. Morris. & Lawrence. Filtering via simulation: Auxiliary particle filters. (1987). Lecture Notes in Statistics No. u 10. A. 571–592. M.. D. & Kaelbling. Slice sampling. D. (2001). Approximate inference for first-order probabilistic languages. Feedforward neural networks for nonparametric regression.. Sinha (Eds. Nicola. D. Thesis. & N. P. Bayesian learning for neural networks. DIMACS series in discrete mathematics and theoretical computer science.. J. Biometrika. Stanford Digital Libraries Working Paper. W. & Winograd. Neal.

C. (1987). England.. A. (1997). 287–300. Monte Carlo inference via greedy importance sampling. MA: MIT Press. Godszmidt (Eds. (2000).. Metropolis light transport. S. (2001)... 93:441. & Guibas. G. P. Simulated annealing: Theory and applications. R. Morgan Kaufmann Publishers. 82:398. In S. A. Speech and Signal Processing (Vol. Amsterdam: Reidel Publishers. 58:2. In C. R. Sherman. 14:1. J. Y. small n” paradigm with application in DNA microarray studies. 248–267.. C. 1701–1762. H. (2000). D. J. (1998). 129–133. (1998). & Yeung. K. Leen. A reversible jump sampler for autoregressive time series. K... Gelfand.. & Smith. & Wang. Veach. Tu. Doucet. S. Non-stationary Bayesian modelling and enhancement of speech signals. & Kohn. J. R. (2001). Tierney. (1991). Marks. F. . (1987).. 65–76. S. S. & Arts. R. W. (1997). (1987). A Bayesian approach to robust binary nonparametric regression.. 699–704. 85:411. H.). Bayesian regression analysis in the “large p.. In International Computer Vision Conference. A.). (2001). IV.. & Tanner. J. Conditional simulation from highly structured Gaussian systems. A. Physical Review Letters. Solla. C.-R. Journal of the American Statistical Association. Ph. Nevins. West. Statistics and Computing. Markov chains for exploring posterior distributions.. Quick simulation: A review of importance sampling techniques in communications systems. R. Technical Report CUED/F-INFENG/TR. A. C. 15:4. Boutilier.. The unscented particle filter. Thrun. 597–613. (1994). (1997). P. S. Swendsen. SIGGRAPH. 523–532). H. Advances in neural u information processing systems 12 (pp. M. Bayesian methods for mixtures of normal distributions. Cambridge. Wakefield. Tierney. L. Journal of the American Statistical Association. 12. van der Merwe. & Godsill. Nonuniversal critical dynamics in Monte Carlo simulations. & Wan. (1999)... P. 49–60. E. Van Laarhoven.. W. R. Vermaak. H. P. Statistics in Medicine. (1999). & Wong. 1064–1070). Troughton. Uncertainty in artificial intelligence (pp.. M¨ ller (Eds. T. Tanner. J. Ho. Image segmentation by data driven Markov chain Monte Carlo. Andrieu. Technical Report CUED/F-INFENG/TR 380. Department of Statistics. Smith. 203–213. Neural Processing Letters. J. & Southey. J. 1. 2257–2260). Thesis. The calculation of posterior distributions by data augmentation. Z.INTRODUCTION 43 Schuurmans. H. Some adaptive Monte Carlo methods for Bayesian inference. R. Stephens. M. D. Wood. R. pp. de Freitas. 31. E.. L.. A Monte Carlo implementation of the EM algorithm and the poor man’s data augmentation algorithms. G. Doucet. J. & Zhu. Wilkinson. with application to blocking-MCMC for the Bayesian analysis of very large linear models. Utsugi. L. (1990).. E. & Zuzan. S. Spang. N. (2000). The Annals of Statistics. Department of Statistics. M. 86–88. In International Conference on Acoustics.. H.. & K. Monte Carlo POMDPs. & Mira.. 528–550. Ensemble of independent factor analyzers with application to natural image analysis.. L. Oxford University. M. F. S. & M. J. T. 18. Conditions for convergence of Monte Carlo EM sequences with an application to product diffusion modeling. & Dalal. Journal of the American Statistical Association. E. (2002). J. IEEE Journal on Selected Areas in Communications. 2:2.D. A. Cambridge University Engineering Department. Statistics and Computing. M. A. A. Efficient generation of random variates via the ratio-of-uniforms methods. Shafi. Wei. J. M. 2507–2515. 22:4. & Godsill.. Duke University. Cambridge University Engineering Department. S. & Gao. (1999).. Econometrics Journal.

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->