Latihan Soal

A Gentle Tutorial on Statistical Inversion
using the Bayesian Paradigm
Tan Bui-Thanh∗
Institute for Computational Engineering and Sciences, The University of Texas at Austin
Contents
1 Introduction 1
2 Some concepts from probability theory 3
3 Construction of likelihood 10
4 Construction of Prior(S) 12
4.1 Smooth priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 “Non-smooth” priors . . . . . . . . . . . . . . . . . . . . . . . . . 17
5 Posterior as the solution to Bayesian inverse problems 19
6 Connection between Bayesian inverse problems and deterministic in-

verse problems 24
7 Markov chain Monte Carlo 27

7.1 Some classical limit theorems . . . . . . . . . . . . . . . . . . . . 27
7.2 Independent and identically distributed random draws . . . . . 31
7.3 Markov chain Monte Carlo . . . . . . . . . . . . . . . . . . . . . 36
8 Matlab codes 47
References 47
1 Introduction
In this note, we are interested in solving inverse problems using statistical
techniques. Let us motivate you by considering the following particular in-
verse problem, namely, the deconvolution problem. Given the observation
∗ tanbui@ices.utexas.edu
1
1. Introduction
signal g(s), we would like to reconstruct the input signal f (t) : [0, 1] → R,
where the observation and the input obey the following relation
Z1
(1) g(s j ) = a(s j , t) f (t) dt, 0 ≤ j ≤ n.
0
Here, a : [0, 1] × [0, 1] → R is known as the blurring kernel. So, in fact we

don’t know the output signal completely, but at a finite number of observa-
tion points. A straightforward approach you may think of is to apply some
numerical quadrature on the right side of (1), and then recover f (t) at the
quadrature points by inverting the resulting matrix. If you do this, you realize
that the matrix is ill-conditioned, and it is not a good idea to invert it. There
are techniques to go around this issue, but let us not pursue them here. In-
stead, we recast the deconvolution task into an optimization problem such as
 2
n Z1
(2) min ∑  g(s j ) − a(s j , t) f (t) dt .
f ( t ) j =0
0
However, the ill-conditioning nature of our inverse problem does not go away.
Indeed, (2) may have multiple solutions and multiple minima. In addition, a
solution to (2) may not depend continuously on g(s j ), 0 ≤ j ≤ n. So what is
the point of recast? Clearly, if the cost function (also known as the data misfit)
is a parabola, then the optimal solution is unique. This immediately suggests
that one should add a quadratic term to the cost function to make it more
like a parabola, and hence making the optimization problem easier. This is
essentially the idea behind the Tikhonov regularization, which proposes to solve
the nearby problem
 2
n Z1
κ 1/2 2

min ∑  g(s j ) − a(s j , t) f (t) dt + R f ,
f ( t ) j =0 2
0
where κ is known as the regularization parameter, and k·k is some appropriate

norm. Perhaps, two popular choices for R1/2 are ∇ and ∆, the gradient and
Laplace operator, respectively, and we discuss them in details in the following.
Now, in practice, we are typically not able to observe g(s j ) directly but its
noise-corrupted value
gobs (s j ) = g(s j ) + e j , 0 ≤ j ≤ n,
where e j , j = 0, . . . , n, are some random noise. You can think of the noise as
the inaccuracy in observation/measurement devices. The question you may
2
2. Some concepts from probability theory
ask is how to incorporate this kind of randomness in the above deterministic

solution methods. There are works in this direction, but let us introduce a
statistical framework based on the Bayesian paradigm to you in this note.
This approach is appealing since it can incorporate most, if not all, kinds of
randomness in a systematic manner.
A large portion of this note follows closely the presentation of two excellent
books by Somersalo et al. [1, 2]. The pace is necessary slow since we develop
this note for readers with minimal knowledge in probability theory. The only
requirement is to either be familiar with or adopt the conditional probability
formula concept. This is the corner stone on which we build the rest of the
theory. Clearly, the theory we present here is by no means complete since the
subject is vast, and still under development.
Our presentation is in the form of dialogue, which we hope it is easier
for the readers to follow. We shall give a lot of little exercises along the way
to help understand the subject better. We also leave a large number of side
notes, mostly in term of little questions, to keep the readers awake and make
connections of different parts of the notes. On the other hand, we often dis-
cuss deeper probability concepts in the footnotes to serve as starting points
for those who want to dive into the rigorous probability theory. Finally, we
supply Matlab codes at the end of the note so that the readers can use them
to reproduce most of the results and to start their journey into the wonderful
world of Bayesian inversion.
2 Some concepts from probability theory
We begin with the definition of randomness.
2.1 definition. An even is deterministic if its outcome is completely predictable.
2.2 definition. A random event is the complement of a deterministic event,

that is, its outcome is not fully predictable.
2.3 example. If today is Wednesday, then “tomorrow is Thursday” is deter-

ministic, but whether it rains tomorrow is not fully predictable.
As a result, randomness means lack of information and it is the direct

consequence of our ignorance. To express our belief1 on random events, we
use probability; probability of uncertain events is always less than 1, an event
that surely happens has probability 1, and an event that never happens has
0 probability. In particular, to reflect the subjective nature, we call it subjective
1 Different person has different belief which leads to different solution of the Bayesian inference
problem. Specifically, one’s belief is based on his known information (expressed in terms of σ-
algebra) and “weights” on each information (expressed in terms of probability measure). That is,
people working with different probability spaces have different solutions.
3
probability or Bayesian probability since it represents belief, and hence depend-

ing upon one’s experience/knowledge to decide what is reasonable to believe.
2.4 example. Let us consider the event of tossing a coin. Clearly, this is a
random event since we don’t know whether head or tail will appear. Nev-
ertheless, we believe that out of n tossing times, n/2 times is head and n/2
times is tail.2 We express this belief in terms of probability as: the (subjective)
probability of getting a head is 12 and the (subjective) probability of getting a
tail is 12 .
We define (Ω, F , P) as a probability space. One typically call Ω the sample

space, F a σ-algebra containing all events A ⊂ Ω, and P a probability measure
defined on F . We can think of an event A as information and the probability
that A happens, i.e. P [ A], is the weight assigned to that information. We
require that
Z
def
0 ≤ P [ A] = dω ≤ 1, P [∅] = 0, P [Ω] = 1.
A
2.5 example. Back to the tossing coin example, we trivially have Ω = { head, tail },
F = {∅, {head} , {tail } , Ω}. The weights are P [∅] = 0, P [{tail }] = P [{head}] =
2 , and P [{ head, tail }] = 1.
1
Two events A and B are independent3 if
P [ A ∩ B] = P [ A] × P [ B] .
One of the central ideas in Bayesian probability is the conditional probabil-

ity4 . The conditional probability of A on/given B is defined as5
P[ A ∩ B ]
(3) P [ A| B] = P[ B ]
,
This is the corner stone formula to build which can also be rephrased as the probability that A happens provided B has
most of results in this note, make sure that
you feel comfortable with it.
already happened.
2.6 example. Assume that we want to roll a dice. Denote B as the event of
getting of face bigger than 4, and A the event of getting face 6. Using (3) we
2 One can believe that out of n tossing times, n/3 times is head and 2n/3 times is tail if he uses
an unfair coin.
3 Probability theory is often believed to be a part of measure theory, but independence is where
it departs from the measure theory umbrella.

4 A more general and rigorous tool is conditional expectation, a particular of which is condi-
tional probability.
5 This was initially introduced by Kolmogorov, a father of modern probability theory.
4
have
1/6
P [ A| B] = = 1/2.
1/3
We can solve the problem using a more elementary argument. B happens
when we either get face 5 or face 6. The probability of getting face 6 when B
has already happened is clearly 12 .
The conditional probability can also be understood as the probability when

the sample space is restricted to B.
2.7 exercise. Determine P [ A| B] in Figure 1.
Ω Ω
B B
A A
(a) P [ A| B] = (b) P [ A| B] =
Figure 1: Demonstration of conditional probability.
2.8 exercise. Show that the following Bayes formula for conditional probabil-
ity holds
P[ B | A ]P[ A ]
(4) P [ A| B] = P[ B ]
.
By inspection, if A and B are mutually independent, we have
P [ A| B] = P [ A] , P [ B| A] = P [ B] .
The probability space (Ω, F , P) is an abstract object which is useful for

theoretical developments, but far from practical considerations. In practice, it
is usually circumvented by probability densities over the state space, which are
easier to handle and have certain physical meanings. We shall come back to
this point in a moment.
5
2.9 definition. The state space S is the set containing all the possible out-
comes.
In this note, the state space S (and also T) is the standard Euclidean space
Rn , where n is the dimension. We are in position to introduce the key player,
the random variable.
2.10 definition. A random variable M is a map6 from the sample space Ω to

the state space S
M : Ω 3 ω 7→ M (ω ) ∈ S.
We call M (ω ) a random variable since we are uncertain about its outcome.

In other words, we admit our ignorance about M by calling it a random vari-
able. This ignorance is in turn a direct consequence of the uncertainty in the
outcome of elementary event ω.
The usual convention is to use lower case letter m = M (ω ) as an arbitrary
realization of the (upper case letter) random variable M, and we utilize this
convention throughout the note.
2.11 definition. The probability distribution (or distribution or law for short)
of a random variable M is defined as
h i
(5) µ M ( A) = P M−1 ( A) = P [{ M ∈ A}] , ∀ A ∈ S,
where we have used the popular notation7
def def
M −1 ( A ) = { M ∈ A } = { ω ∈ Ω : M ( ω ) ∈ A } .
From the definition, we can see that the distribution is a probability mea-
sure8 on S. In other words, the random variable M induces a probability mea-
sure, defined as µ M , on the state space S. The key property of the induced
probability measure µ M is the following. The probability for an event A in
the state space to happen, denoted as µ M ( A), is defined as the probability
for an event B = M−1 ( A) in the sample space to happen (see Figure 2 for
an illustration). The distribution and the probability density9 π X of M obey the
6 Measurable map is the rigorous definition, but we avoid technicalities here since it involves
operations on σ-algebra.
7 Rigorously, A must be a measurable subset of S.
8 In fact, it is the push-forward measure by the random variable M.
9 Here, the density is understood with respect to the Lebesgue measure on S = Rn . Rigorously,
π M is the Radon-Nikodym derivative of µ M with respect to the Lebesgue measure. As a result,

π M should be understood as equivalent class on L1 (S).
6
following relation
def R def R
(6) µ M ( A) = A π M (m) dm = { M∈ A} dω, ∀A ⊂ S .
where the second equality of the definition is from (5). The meaning of ran- Do you see this?
dom variable M (ω ) can now be seen in Figure 2. It maps the event B ∈ Ω
into the set A = M ( B) in the state space such that the area under the density
function π M (m) and above A is exactly the probability that B happens.
πM (m)
P[B]
S
Ω B A = M(B)
Figure 2: Demonstration of random variable: µ M ( A) = P [ B].
We deduce the change of variable formula
µ M (dm) = π (m) dm = dω .
We will occasionally simply write π (m) instead of π M (m) if there is no am-

biguity.
2.12 remark. In theory, we introduce the probability space (Ω, F , P) in order

to compute the probability of a subset10 A in the state space S, and this is
essentially the meaning of (5). However, once we know the probability den-
sity function π M (m), we can operate directly on the state space S without the
need for referring back to probability space (Ω, F , P), as shown in definition
(6). This is the key observation, a consequence of which is that we simply ig-
10 Again, it needs to be measurable.
7
nore the underlying probability space in practice, since we don’t need them
in computation of probability in the state space. However, to intuitively un-
derstand the source of randomness, we need to go back to the probability
space where the outcome of all events, except Ω, is uncertain. As a result,
the pair (S, π M (m)) contains complete information describing our ignorance
about the outcome of random variable M. To the rest of this note, we shall
work directly on the state space.
2.13 remark. At this point, you may wonder what is the point of introducing
the abstract probability space (Ω, F , P) to make life more complicated? Well,
its introduction is two fold. First, as discussed above, the probability space
not only shows the origin of randomness but also provides the probability
measure P for the computation of the randomness; it is also used to define
random variables and furnishes a decent understanding about them. Second,
the concepts of distribution and density in (5) and (6), which are introduced
for random variable M, a map from Ω to S, are valid for maps11 from an
arbitrary space V to another space W. Here, W plays the role of S, and V
the role of Ω on which we have a probability measure. For example, later in
Section 3, we introduce the parameter-to-observable map h (m) : S → Rr , then
S plays the role of Ω and Rr of S in (5) and (6).
2.14 definition. The expectation or the mean of a random variable M is the

quantity
Z Z
def def
(7) E [ M] = mπ (m) dm = M (ω ) dω = m,
S Ω
and the variance is

h i Z Z
def def def
Var [ M] = E ( M − m)2 = (m − m)2 π (m) dm = ( M (ω ) − m)2 dω.
S Ω
As we will see, the Bayes formula for probability densities is about the joint
density of two or more random variables. So let us define the joint distribution
and joint density of two random variables here.
2.15 definition. Denote µ MY and π MY as the joint distribution and density,

respectively, of two random variables M with values in S and Y with values in
T defined on the same probability space, then the joint distribution function
and the joint probability density, in the light of (6), satisfy
Z
def
(8) µ MY ({ M ∈ A} , {Y ∈ B}) = π MY (m, y) dmdy, ∀ A × B ⊂ S × T,
A× B
11 Again, they must be measurable.
8
where the notation A × B ⊂ S × T simply means that A ∈ S and B ∈ T.
We say that M and Y are independent if
µ MY ({ M ∈ A} , {Y ∈ B}) = µ M ( A) µY ( B) , ∀ A × B ⊂ S × T,
or if
π MY (m, y) = π M (m) πY (y) .
2.16 definition. The marginal density of M is the probability density of M

when Y may take on any value, i.e.,
Z
π M (m) = π MY (m, y) dy.
T
Similarly, the marginal density of Y is the density of Y regardless of M,

namely,
Z
πY ( y ) = π MY (m, y) dm.
S
Before deriving the Bayes formula, we define conditional density π (m|y)

in the same spirit as (6) as
Z
µ M|y ({ M ∈ A} |y) = π (m|y) dm.
A
Let us prove the following important result.
2.17 theorem. The conditional density of M given Y is given by
π (m,y)
π (m|y) = π (y)
.
Proof. From the definition of conditional probability (3), we have
µ M|Y ({m ∈ A} |y) = P [{ M ∈ A} |Y = y] (definition (5))

P [{ M ∈ A} , y ≤ y0 ≤ y + ∆y]
= lim (definition (3))
∆y→0 P [y ≤ y0 ≤ y + ∆y]
R
π (m, y) dm∆y
= lim A (definitions (6), (8))
∆y→0 π (y) ∆y
π (m, y)
Z
= dm,
π (y)
A
9
3. Construction of likelihood
which ends the proof.
By symmetry, we have
π (m, y) = π (m|y) π (y) = π (y|m) π (m) ,
from which the well-known Bayes formula follows
π (y|m)π (m)
(9) π (m|y) = π (y)
.
2.18 exercise. Prove directly the Bayes formula for conditional density (9)
using the Bayes formula for conditional probability (4).
2.19 definition (Likelihood). We call π (y|m) the likelihood. It is the proba-

bility density of y given M = m.
2.20 definition (Prior). We call π (m) the prior. It is the probability density
of M regardless of Y. The prior encodes, in the Bayesian framework, all infor-
mation before any observations/data are made.
2.21 definition (Posterior). The density π (m|y) is called the posterior, the
distribution of parameter m given the measurement y, and it is the solution of
the Bayesian inverse problem under consideration.
Can you write the conditional mean using 2.22 definition. The conditional mean is defined as
dω? Z
E [ M|y] = mπ (m|y) dm.
S
2.23 exercise. Show that

Z
E [ M] = E [ M |y] π (y) dy.
T
3 Construction of likelihood
In this section, we present a popular approach to construct the likelihood. We

begin with the additive noise case. The ideal deterministic model is given by
y = h (m) ,
where y ∈ Rr . But due to random additive noise E, we have the following

statistical model instead
(10) Y obs = h ( M) + E,
10
3. Construction of likelihood
where Y obs is the actual observation rather than Y = f ( M). Since the noise
comes from external sources, in this note, it is assumed to be independent of
M. In the likelihood modeling, we pretend to have realization(s) of M and the
task is to construct the distribution of Y obs . From (10), one can see that the
randomness in Y obs is the randomness
in E shifted by anamount h (m), see
Figure 3, and hence πY obs |m yobs |m = π E yobs − h (m) . More rigorously,
h(m)
Figure 3: Likelihood model with additive noise.
assume that both Y obs and E are random variables on a same probability space,
we have
Z
(5)

def
πY obs |m yobs |m dyobs = µY obs |m ( A) = µ E ( A − h (m))
A
Z Z
change of variable

= π E (e) de = π E yobs − h (m) dy, ∀ A ⊂ S,
A−h(m) A
which implies

πY obs |m yobs |m = π E yobs − h (m) .
We next consider multiplicative noise. The statistical model in this case
11
4. Construction of Prior(S)
reads
(11) Y obs = Eh ( M ) .
3.1 exercise. Show that the likelihood for multiplicative noise model (11) has
the following form

π E yobs /h (m)
πY obs |m yobs |m = , h (m) 6= 0.
h (m)

3.2 exercise. Can you generalize the result for the noise model e = g yobs , h ( x ) ?
4 Construction of Prior(S)
As discussed previously, the prior belief depends on a person’s knowledge

and experience. In order to obtain a good prior, one sometimes needs to per-
form some expert elicitation. Nevertheless, there is no universal rule and one
has to be careful in constructing a prior. In fact, prior construction is a sub-
ject of current research, and it is problem-dependent. For concreteness, let us
consider the following one dimensional deblurring (deconvolution) problem
Z1
g(s j ) = a(s j , t) f (t) dt + e(s j ), 0 ≤ j ≤ n,
0
where a (s, t) = √ 1 exp(− 2β1 2 (t − s)2 ) is a given kernel, and s j = j/n, j =

2πβ2
0, . . . , n the mesh points. Our task is to reconstruct f (t) : [0, 1] → R from
the noisy observations g s j , j = 0, . . . , n. To cast the function reconstruc-
tion problem, which is in infinite dimensional space, into a reconstruction
problem in Rn , we discretize f (t) on the same mesh and use simple rectan-
gle method for the integral. Let us define Y obs = [ g (s0 ) , . . . , g (sn )] T , M =
( f (s0 ), . . . , f (sn ))T , and Ai,j = a(si , s j )/n, then the discrete deconvolution
problem reads
Y obs = A M + E.
Here, we assume E ∼ N 0, σ2 I , where I is the identity matrix in R(n+1)×(n+1) .

Since Section 3 suggests the likelihood of the form

π yobs |m = N yobs − Am, σ2 I ,
12
the Bayesian solution to our inverse problem is, by virtue of the Bayes formula
(9), given by

(12) πpost m|yobs ∝ N yobs − Am, σ2 I × πprior (m) ,

where we have ignored the denominator π yobs since it does not depend on
the parameter of interest m. We now start our prior elicitation.
4.1 Smooth priors
In this section, we believe that the unknown function f (t) is smooth, which
can be translated into, among other possibilities, the following simplest re-
quirement on the pointwise values f (si ), and hence mi ,
1
(13) mi = ( m i −1 + m i +1 ) ,
2
that is, the value of f (s) at a point is more or less the same of its neighbor.
But, this is by no means the correct behavior of the unknown function f (s).
We therefore admit some uncertainty in our belief (13) by adding an innovative
term Wj such that
1
Mi = ( Mi−1 + Mi+1 ) + Wj ,
2
where W ∼ N 0, γ2 I . The standard deviation γ determines how much the

reconstructed function f (t) departs from the smoothness model (13). In terms
of matrices, we obtain
LM = W,
where L is given by
−1 −1
 
2
1 −1 2 −1 
L=  ∈ R(n−1)×(n+1) ,
 
2
 .. .. ..
. . . 
−1 2 −1
which is the second order finite difference matrix approximating the Laplacian
∆ f . Indeed,
(14) ∆ f (s j ) ≈ n2 ( LM ) j .
13
The prior distribution is therefore given by (using the technique in Section 3)

1
(15) πpre ∝ exp − 2 k LMk2 .
γ
But L has rank of n − 1, and hence πpre is a degenerate Gaussian density

in Rn+1 . The reason is that we have not specified the smoothness of f (s) at
the boundary points. In other words, we have not specified any boundary
conditions for the Laplacian ∆ f (s). This is a crucial point in prior elicitation
via differential operators. One needs to make sure that the operator is positive
definite by incorporating some well-posed boundary conditions. Throughout
the lecture notes, unless otherwise stated, k·k denotes the usual Euclidean
norm.12
Let us first consider the case with zero Dirichlet boundary condition, that
is, we believe that f (s) is smooth and (close to) zero at the boundaries, then
1 1
M0 = ( M−1 + M1 ) + W0 = M1 + W0 , W0 ∼ N 0, γ2
2 2
1 1
Mn = ( Mn−1 + Mn+1 ) + Wn = Mn−1 + Wn , Wn ∼ N 0, γ2 .
2 2
Note that we have extended f (s) by zero outside the domain [0, 1] since we
“know” that it is smooth. Consequently, we have L D M = W with
2 −1
 
 −1 2 −1 
 
1 −1 2 −1 
(16) L D =   ∈ R(n+1)×(n+1) ,
 
2 .. .. ..
 . . . 

 −1 2 −1 
−1 2
which is the second order finite difference matrix corresponding to zero Dirich-
let boundary conditions. The prior density in this case reads

D 1 2
(17) πprior (m) ∝ exp − 2 k L D mk .
2γ
It is instructive to draw some random realizations from πprior D (we are

ahead of ourselves here since sampling will be discussed in Section 7), and
we show five of them in Figure 4 together with the prior standard deviation
curve. As can be seen, all the draws are almost zero at the boundary and the
Why are they not exactly zero? prior variance (uncertainty) is close to zero as well. This is not surprising since
our prior belief says so. How do we compute the standard deviation curve?
12 The `2 -norm if you wish
14
Well, it is straightforward. We first compute the pointwise variance as

 
def h 2 i Z −1
def
Var M j = E M j = e Tj  D
m2 πprior dm e j = γ2 e Tj L TD L D ej ,
Rn +1
where e j is the jth canonical basis vector in Rn+1 , and we have used the fact
that the prior is Gaussian in the last equality. So we in fact plot the square
−1
root of the diagonal of γ2 L TD L D , the covariance matrix, as the standard
deviation curve. One can see that the uncertainty is largest in the middle of the Do we really have the complete continuous
curve?
domain since it is farthest from the constrained boundary. The points closer
to the boundaries have smaller variance, that is, they are more correlated to
the “known” boundary data, and hence less uncertain.
30
20
10
−10
−20
−30
Standard deviation
−40
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
D
Figure 4: Prior random draws from πprior together with the standard deviation
curve.
Now, you may ask why f (s) must be zero at the boundary, and you are
right! There is no reason to believe that must be the case. However, we don’t
know the exact values of f (s) at the boundary either, even though we believe
that we may have non-zero Dirichlet boundary condition. If this is the case,
we have to admit our ignorance and let the data from the likelihood correct
us in the posterior. To be consistent with the Bayesian philosophy, if we do
not know anything about boundary conditions, let them be, for convenience,
15
Gaussian random variables such as

!
γ2 γ2

M0 ∼ N 0, 2 , Mn ∼ N 0, 2 .
δ0 δn
Hence, the prior can now be written as

1
R
(18) πprior (m) ∝ exp − 2 k L R mk2 ,
2γ
where
2δ0 0
 
 −1 2 −1 
 
1 −1 2 −1 
LR =   ∈ R(n+1)×(n+1) .
 
2 .. .. ..
 . . . 

 −1 2 −1 
0 2δn
A question that immediately arises is how to determine δ0 and δn . Since the

boundary values are now independent random variables, we are less certain
about them compared to the previous case. But to which uncertain level we
want them to be? Well, let’s make every values equally uncertain, meaning we
have the same ignorance about the values at these points, that is, we would
like to have the same variances everywhere. To approximately accomplish this,
we require
γ2 γ2 h i
2 T

T
−1
Var [ M0 ] = = Var [ M n ] = = Var M [ n/2 ] = γ e [n/2] L R L R e[n/2] ,
δ02 δn2
where [n/2] denotes the largest integer smaller than n/2. It follows that
1
δ02 = δn2 = −1 .
e[Tn/2] L TR L R e[n/2]
However, this requires to solve a nonlinear equation for δ0 = δn , since L R

depends on them. To simplify the computation, we use the following approx-
Is it sensible to do so? imation
1
δ02 = δn2 = −1 .
e[Tn/2] L TD L D e[n/2]
Again, we draw five random realizations from πprior R and put them to-
gether with the standard deviation curve in Figure 5. As can be observed, the
16
uncertainty is more or less the same at every point and prior realizations are
no longer constrained to have zero boundary conditions.
60
Standard deviation
50
40
30
20
10
−10
−20
−30
−40
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
R
curve.
4.1 exercise. Consider the following general scheme
mi = λi mi−1 + (1 − λi )mi+1 + Ei , 0 ≤ λi ≤ 1.
Convince yourself that by choosing a particular set of λi , you can recover all
the above prior models. Replace BayesianPriorElicitation.m by a generic
code with input parameters λi . Experience new prior models by using differ-
ent values of λi (those that don’t reproduce priors presented in the text).
4.2 exercise. Construct a prior with a non-zero Dirichlet boundary condition

at s = 0 and zero Neumann boundary condition at s = 1. Draw a few samples
together with the variance curve to see whether your prior model indeed
conveys your belief.
4.2 “Non-smooth” priors
We first consider the case in which we believe that f (s) is still smooth but may
have discontinuities at known locations on the mesh. Can we design a prior to
convey this belief? A natural approach is to require that M j is equal to M j−1
17
plus a random jump, i.e.,
M j = M j −1 + E j ,
where Ej ∼ N 0, γ2 , and for simplicity, let us assume that M0 = 0. The prior

density in this case would be

1
(19) πpren (m) ∝ exp − 2 k L N mk2 ,
2γ
where
 
1
 −1 1 
LN =  ∈ Rn × n .
 
.. ..
 . . 
−1 1
But, if we think that there is a particular big jump, relative to others, from 2

M j−1 to M j , then the mathematical translation of this belief is Ej ∼ N 0, γθ 2
with θ < 1. The corresponding prior in this case reads

O 1 2
(20) πprior (m) ∝ exp − 2 k JL N mk ,
2γ
with
 
J = diag 1, . . . , 1, |{z}

θ , 1, . . . , 1 .
jth index
O
Let’s draw some random realizations from πprior (m) in Figure 6 with n = 160,
j = 80, β = 1, and θ = 0.01. As desired, all realizations have a sudden jump
at j = 80, and the standard deviation of the jump is 1/θ = 100. In addition,
O
compared to priors in Figure 4 and 5, the realizations from πprior (m) are less
smooth, which confirms that our belief is indeed conveyed.
4.3 exercise. Use BayesianPriorElicitation.m to construct examples with

2 or more sudden jumps and plot a few random realizations to see whether
your belief is conveyed.
A more interesting and more practical situation is the one in which we

don’t know how many jump discontinuities and their locations. A natural
18
5. Posterior as the solution to Bayesian inverse problems
120
Standard deviation
100
80
60
40
20
−20
−40
−60
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
O
curve.
prior in this situation is a generalized version of (20), e.g.,

M 1 2
(21) πprior (m) ∝ C ( M ) exp − 2 k ML N mk ,
2γ
with
M = diag (θ1 , . . . , θn ) ,
where θi , i = 1, . . . , n, are unknown. In fact, these are called hyper-parameters

and one can determine them using information from the likelihood; the read- What is wrong with this?
ers are referred to [1] for the details.
4.4 exercise. Modify the scheme in Exercise 4.1 to include priors with sudden
jumps.
5 Posterior as the solution to Bayesian inverse problems
In this section, we explore the posterior (12), the solution of our Bayesian
problem, given the likelihood in Section 3 and priors in Section 4.
To derive results that are valid for all priors discussed so far, we work with
19
the following generic prior

1 2
1
πprior (m) ∝ exp − 2 Γ− 2 m ,
2γ
1
where Γ− 2 ∈ { L D , L A , HL N }, each of which again presents a different belief.
The Bayesian solution (12) can be now written as
 
 1 2 
1 − 12 2 

πpost m|yobs ∝ exp −
obs
− A m + m ,

2σ2 2γ2
y Γ
 
| {z }
T (m)
where T (m) is the familiar (to you I hope) Tikhonov functional; it is sometimes
called the potential. We re-emphasize here that the Bayesian solution is the
posterior probability density, and if we draw samples from it, we want to
know what the most likely function m is going to be. In other words, we
ask for the most probable point m in the posterior distribution. This point is
known as the Maximum A Posteriori (MAP) estimator/point, namely, the point
at which the posterior density is maximized. Let us denote this point as m MAP ,
and we have

def
m MAP = arg max πpost m|yobs = arg min T (m) .
m m
Hence, the MAP point is exactly the deterministic solution of the Tikhonov
functional!
Since both likelihood and prior are Gaussian, the posterior is also a Gaus-
This is fundamental. If you have not seen sian. For our case, the resulting posterior Gaussian reads
this, prove it!
2 !

obs
1 1 − 1 T obs
∝ exp −

πpost m|y m− 2H A y
2 σ
H

1 1 −1 T obs 1
= exp − m − 2 H A y , H m − 2 H −1 A T yobs
2 σ σ

def 1 1 1
= exp − m − 2 H −1 A T yobs , Γpost
−1
m − 2 H −1 A T yobs
2 σ σ
where
H= 1
σ2
A T A + γ12 Γ−1 ,
is the Hessian of the Tikhonov functional (aka the regularized misfit), and we
1 2
have used the weighted norm k·k2H = H 2 · .

20
5.1 exercise. Show that the posterior is indeed a Gaussian, i.e.,

2 !

obs
1 1 −1 T obs
πpost m|y ∝ exp − m − 2 H A y
.
2 σ
H
The other important point is that the posterior covariance matrix is pre-
cisely the inverse of the Hessian of the regularized misfit, i.e.,
Γpost = H −1 .
Last, but not least, we have showed that the MAP point is given by
−1
1 1 1 T 1
m MAP = 2 H −1 A T yobs = 2 A A + 2 Γ −1 A T yobs ,
σ σ σ2 γ
which is, again, exactly the solution of the Tikhonov functional for linear in-
verse problem.
5.2 exercise. Show that m MAP is also the least squares solution of the follow-
ing over-determined system
" #
1
A
1 obs
σ y
1 − 12 m= σ
γ Γ 0
5.3 exercise. Show that the posterior mean, which is in fact the conditional
mean, is precisely the MAP point.
Since the covariance matrix, generalization of the variance in multi-dimensional

spaces, represents the uncertainty, quantifying the uncertainty in the MAP es-
timator is ready by simply computing the inverse the Hessian matrix. Let’s us
now numerically explore the Bayesian posterior solution.
We choose β = 0.05, n = 100, and γ = 1/n. The truth underlying function
that we would like to invert for is given by

f (t) = 10(t − 0.5) exp −50(t − 0.5)2 − 0.8 + 1.6t.
The noise level is taken to be the 5% of the maximum value of f (s), i.e. σ =
0.05 maxs∈[0,1] | f (s)|.
D
We first consider the belief described by πprior in which we think that
f (s) is zero at the boundaries. Figures 7 plots the MAP estimator, the truth
function f (s), and the predicted uncertainty. As can be observed, the MAP is
in good agreement with the truth function inside the interval [0, 1], though
it is far from recovering f (s) at the boundaries. This is the price we have to
pay for not admitting our ignorance about the boundary values of f (s). The
21
likelihood in fact sees this discrepancy in the prior knowledge and tries to
make correction by lifting the MAP away from 0, but not enough to be a good
reconstruction. The reason is that our incorrect prior is strong enough such
that the information from the data yobs cannot help much.
0.8
0.6
0.4
0.2
−0.2 MAP
truth
−0.4 uncertainty
−0.6
−0.8
−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 7: The MAP estimator, the truth function, and the predicted uncertainty
D .
(95% credibility region) using πprior
5.4 exercise. Can you make the prior less strong? Change some parameter to
make prior contribution less! Use BayesianPosterior.m to test your answer. Is
the prediction better in terms of satisfying the boundary conditions? Is the
uncertainty smaller? If not, why?
On the other hand, if we admit this ignorance and use the corresponding
D , we see much better reconstruction in Figure 8. In this case, we in
prior πprior
fact let the information from the data yobs determine the appropriate values
for the Dirichlet boundary conditions rather than setting them to zero. By
doing this, we allow the likelihood and the prior to be well-balanced leading
to good reconstruction and uncertainty quantification.
5.5 exercise. Play with BayesianPosterior.m by varying γ, the data misfit

(or the likelihood) contribution, and σ, the regularization (or the prior) contri-
bution.
22
0.8
0.6
0.4
0.2
−0.2 MAP
truth
−0.4 uncertainty
−0.6
−0.8
−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
A .
5.6 exercise. Use your favorite deterministic inversion approach to solve the
above deconvolution problem and then compare it with the solution in Figure
8.
Now consider the case in which the truth function has a jump discontinuity
at j = 70. Assume we also know that the magnitude of the jump is 10. In
particular, we take the truth function f (s) as the following step function

0 if s ≤ 0.7
f (s) = .
10 otherwise
Since we otherwise have no further information about f (s), let us be more

O
conservative by choosing γ = 1 and θ = 0.1 at j = 70 in πprior as we discussed
in (20). Figure 9 shows that we are doing pretty well in recovering the jump
and other parts of the truth function.
A question you may ask is whether we can do better? The answer is yes by
taking smaller γ if the truth function does not vary much everywhere except at Why?
the jump discontinuity. We take this prior information into account by taking
γ = 1.e − 8, for example, then our reconstruction is almost perfect in Figure
10.
23
6. Connection between Bayesian inverse problems and deterministic
inverse problems
14
12 MAP
truth
uncertainty
10
−2
−4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
O .
5.7 exercise. Try BayesianPosteriorJump.m with γ deceasing from 1 to 1.e −

8 to see the improvement in quality of the reconstruction.
5.8 exercise. Use your favorite deterministic inversion approach to solve the
above deconvolution problem with discontinuity and then compare it with
the solution in Figure 10.
6 Connection between Bayesian inverse problems and

deterministic inverse problems
We have touched upon the relationship between Bayesian inverse problem
and deterministic inverse problem in Section 5 by pointing out that the po-
tential of the posterior density is precisely the Tikhonov functional up to a
constant. We also point out that the MAP estimator is exactly the solution of
the deterministic inverse problem. Note that we derive this relation for a lin-
ear likelihood model, but it is in fact true for nonlinear ones (e.g. nonlinear
Can you confirm this? parameter-to-observable map Am).
Up to this point, you may realize that the Bayesian solution contains much
more information than its deterministic counterpart. Instead of having a point
estimate, the MAP point, we have a complete posterior distribution to explore.
In particular, we can talk about a simple uncertainty quantification by exam-
24
inverse problems
12
MAP
10 truth
uncertainty
−2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 10: The MAP estimator, the truth function, and the predicted uncer-
O .
tainty (95% credibility region) using πprior
ining the diagonal of the posterior covariance matrix. We can even discuss
about the posterior correlation structure by looking at the off diagonal ele-
ments, though we are not going to do it here in this lecture note. Since, again,
both likelihood and prior are Gaussian, the posterior is a Gaussian distribu-
tion, and hence the MAP point (the first order moment) and the covariance
matrix (the second order moment) are the complete description of the poste-
rior. If, however, the likelihood is not Gaussian, say when the Am is nonlinear,
then one can explore higher moments.
We hope the arguments above convince you that the Bayesian solution pro-
vide information far beyond the deterministic counterpart. In the remainder
of this section, let us dig into details the connection between the MAP point
and the deterministic solution, particularly in the context of the deconvolution
problem. Recall the definition of the MAP point

1 obs 2 1 σ 2 1 2
def 2 −2
m MAP = arg min T (m) = σ y − Am + Γ m

m 2 2 γ2
2 1 1 2
1 obs
= arg min T (m) = σ2 − y + κ R 2 m ,

2 2
y
m
1 1
where we have defined κ = σ2 /γ2 , R 2 = Γ− 2 , and y = Am.
We begin our discussion with zero Dirichlet boundary condition prior
25
inverse problems
D
πprior (m) in (17). Recall in (14) and (16) that L D M is proportional to a dis-
cretization of the Laplacian operator with zero boundary conditions using
second order finite difference method. Therefore, our Tikhonov functional is
in fact a discretization, up to a constant, of the following potential in the infi-
nite dimensional setting
1 2 1
2
T∞ ( f ) = y − yobs + κ k∆ f k L2 (0,1) ,

2 2
def R1
where k·k2L2 (0,1) = 0 (·)2 ds. Rewrite the preceding equation informally as
1 2 1
T∞ ( f ) = y − yobs + κ f , ∆2 f 2 ,

2 2 L (0,1)
and we immediately realize that the potential in our prior description, namely
k L D mk2 , is in fact a discretization of Tikhonov regularization using the bihar-
monic operator. This is another explanation for the smoothness of the prior
realizations and the name smooth prior, since biharmonic regularization is
very smooth. 13
The power of the statistical approach lies in the construction of prior
R
πprior (m). Here, the interpretation of rows corresponding to interior nodes
s j is still the discretization of the biharmonic regularization, but the design
of those corresponding to the boundary points is purely statistics, for which
we have no corresponding deterministic counterpart (or at least it is not clear
how to construct it from a purely deterministic point of view). As the results
in Section 5 showed, πprior R (m) provided much more satisfactory results both
in the prediction and in uncertainty quantification.
As for the “non-smooth”priors in Section 4.2, a simple inspection shows
that L N m is, up to a constant, a discretization of ∇ f . Similar to the above dis-
cussion, the potential in our prior description, namely k L D mk2 , is now in fact
a discretization of Tikhonov regularization using the Laplacian operator.14 As
a result, the current prior is less smooth than the previous one with harmonic
operator. Nevertheless, all the prior realizations corresponding to πpren (m)
13 From a functional analysis point of view, k∆ f k2L2 (0,1) is finite if f ∈ H 2 (0, 1), and by Sobolev
imbedding theorem we know that in fact f ∈ C1,1/2−ε , the space of continuous differential
functions whose first derivative is in the Hölder space of continuous function C1/2−ε , for any
0 < ε < 21 . So indeed f is more than continuously differentiable.
14 Again, Sobolev embedding theorem shows that f ∈ C 1/2−ε for
k∇ f k2L2 (0,1) to be finite. Hence,
all prior realizations corresponding to πpren (m) are at least continuous. The prior πpriorO (m) is
different, due to the scaling matrix J. As long as θ stays away from zero, prior realizations are
still in H 1 (0, 1), and hence continuous though having steep gradient at s j as shown in Figures 9
and 10. But as θ approaches zero, prior realizations are leaving H 1 (0, 1), and therefore may be no
1
longer continuous. Note that in one dimension, H 2 +ε is enough to be embedded in the space of
C ε -Hölder continuous functions. If you like to know a bit about the Sobolev embedding theorem,
see [3].
26
7. Markov chain Monte Carlo
are at least continuous, though may have steep gradient at s j as shown in Fig-
ures 9 and 10. The rigorous arguments for the prior smoothness require the
Sobolev embedding theorem, but we avoid the details.
For those who have not seen the Sobolev embedding theorem, you only
O
loose the insight on why πprior (m) could give very steep gradient realizations
(which is the prior belief we start with). Nevertheless, you still can see that
O
πprior D
(m) gives less smooth realizations than πprior (m) does, since, at least,
O
the MAP point corresponding to πprior (m) only requires finite first deriva-
tive of f while second derivative of f needs to be finite at the MAP point if
D
πprior (m) is used.
7 Markov chain Monte Carlo
In the last section, we have shown that if the paramter-to-observable map is

linear, i.e. h (m) = Am, and both the noise and the prior models are Gaus-
sian, then the MAP point and the posterior covariance matrix are exactly the
solution and the inverse of the Hessian of the Tikhonov functional, respec-
tively. Moreover, since the posterior is Gaussian, the MAP point is identically
the mean, and hence the posterior distribution is completely characterized. In
practice, h (m) is typically nonlinear. Consequently, the posterior distribution
is no longer Gaussian. Nevertheless, the MAP point is still the solution of the
Tikhonov functional, though the mean and the convariance matrix are to be
determined. The question is how to estimate the mean and the covariance
matrix of a non-Gaussian density.
We begin by recalling the definition the mean
m = E [ M] ,
and a natural idea is to approximate the integral by some numerical inte-

gration. For example, suppose S = [0, 1] and then we can divide S into N
intervals, each of which has length of 1/N. Using a rectangle rule gives
( M1 + . . . + M N )
(22) m ≈ .
N
But this kind of method cannot be extended to S = Rn . This is where the
central limit theorem and law of large numbers come to rescue. They say that
the simple formula (22) is still valid with a simple error estimation expression.
7.1 Some classical limit theorems
7.1 theorem (central limit theorem (CLT)). Assume that real valued random vari-
ables M1 , . . . are independent and identically distributed (iid), each with expectation
27
m and variance σ2 . Then

1 m√
ZN = √ ( M1 + M2 + · · · + M N ) − N
σ N σ
converges, in distribution15 , to a standard normal random variable. In particular,
Zm
t2

1
(23) lim P [ ZN ≤ m] = exp − dt
N →∞ 2π 2
−∞
Proof. The proof is elementary, though technical, using the concept of charac-
teristic function (Fourier transform of a random variable). You can consult [4]
for the complete proof.
7.2 theorem (Strong law of large numbers (LLN)). Assume random variables
M1 , . . . are independent and identically distributed (iid), each with finite expectation
m and finite variance σ2 . Then
1
(24) lim S N = ( M1 + M2 + · · · + M N ) = m
N →∞ N
almost surely16 .
Proof. A beautiful, though not classical, proof of this theorem is based on

backward martingale, tail σ-algebra, and uniform integrability. Let’s accept it
in this note and see [4] for the complete proof.
7.3 remark. The central limit theorem says that no matter what the underly-
ing common distribution looks like, the sum of iid random variables, when
properly scaled and centralized, converges in distribution to a standard nor-
mal distribution. The strong law of large numbers, on the other hand, states
that the average of the sum is, as expected in the limit, precisely the mean of
the common distribution with probability one.
Both the central limit theorem (CLT) and the strong law of large numbers
(LLN) are useful, particularly LLN, and we use them routinely. For example,
if you are given an iid sample { M1 , M2 , · · · , M N } from a common distribu-
tion π (m), the first thing you should do is to compute the the sample mean
S N to estimate the actual mean m. From LLN we know that the sample mean
can be as close as desired if N is sufficiently large. A question immediately
arises is whether we can estimate the error between the sample mean and the
15 Convergence in distribution is also known as weak convergence and it is beyond the scope of
this introductory note. You can think of the distribution of Zn is more and more like the standard
normal distribution as n → ∞, and it is precisely (23).
16 Almost sure convergence is the same as convergence with probability one, that is, the event
on which the convergence (24) does not happen has zero probability.
28
truth mean, given a finite N. Let us first give an answer based on a simple
application of the CLT. Since the sample { M1 , M2 , · · · , M N } satisfies the con-
dition of the CLT, we know that ZN converges to N (0, 1). It follows that, at
least for sufficiently large N, the mean squared error between z N and 0 can be
estimated as
h i
def def
kz N − 0k2L2 (S,P) = E (z N − 0)2 = Var [ ZN − 0] ≈ 1,
which, after some simple algebra manipulations, can be rewritten as
def σ2
(25) kS N − mk2L2 (S,P) = Var [S N − m] ≈
N
7.4 exercise. Show that (25) holds.
√ the error of the sample mean S N in the

7.5 remark. The result (25) shows that
L (S, P)-norm goes to zero like 1/ N. One√should be aware of the popular
2
statement that the error goes to zero like 1/ N independent of dimension is

not entirely correct because the variance σ2 , and hence the standard deviation
σ, of the underlying distribution π (m) may depend on the dimension n.
If you are a little bit delicate, you may not feel completely comfortable with
the error estimate (25) since you can rewrite it as
σ
kS N − mk L2 (S,P) = C √ ,
N
and you are not sure how big C is and the dependence of C on N. Let us try
to make you happy. We have
" ! !#
N N
1
2
kS N − mk L2 (S,P) = 2 E ∑ ( Mi − m ) ∑ M j − m

N i =1 j =1
" !#
N
1 1 N σ2
= 2E
N ∑ ( Mi − m ) 2 = 2 ∑ σ2 = ,
N i =1 N
i =1
where we have used m = N1 ∑iN=1 m in the first equality, E ( Mi − m) M j − m =

0 if i 6= j in the second equality since Mi , i = 1, . . . , N are iid random variables,

and the definition of variance in the third equality. So in fact C = 1, and we
hope that you feel pleased by now.
In practice, we rarely work with M directly but indirectly via some map-
ping g : S → T. We have that g ( Mi ), i = 1, . . . , N are iid17 if Mi , i = 1, . . . , N
are iid.
17 We avoid technicalities here, but g needs to be a Borel function for the statement to be true.
29
7.6 exercise. Suppose the density of M is π (m) and z = g (m) is differentially

invertible, i.e. m = g−1 (z) exists and differentiable, what is the density of
g ( M )?
Perhaps, one of the most popular and practical problems is to evaluate the
mean of g, i.e.,
Z
def
(26) I = E [ G ( M )] = g (m) π (m) dm,
S
which is an integral in Rn .
7.7 exercise. Define z = g (m) ∈ T, the definition of the mean in (7) gives
Z
def
E [ G ( M )] ≡ E [ Z ] = zπ Z (z) dz.
T
Derive formula (26).
Again, we emphasize that using any numerical integration methods that

you know of for integral (26) is either infeasible or prohibitedly expensive
when the dimension n is large, and hence not scalable. The LLN provides a
reasonable answer if we can draw iid samples { g ( M1 ) , . . . , g ( M N )} since we
know that
1
lim ( g ( M1 ) + . . . + g ( M N )) = I
N →∞ N
| {z }
IN
with probability 1. Moreover, as showed above, the mean squared error is

Do you trivially see this? given by
h i Var [ G ( M )]
k IN − I k2L2 (T,P) = E ( IN − I )2 = .
N
√
Again, the error decreases to zero like 1/ N “independent” of the dimension
of T, but we need to be careful with such a statement unless Var [ G ( M )]
DOES NOT depend on the dimension.
A particular function g of interest is the following
g (m) = (m − m) (m − m) T ,
whose expectation is precisely the covariance matrix

h i
Γ = cov ( M ) = E ( M − m) ( M − m) T = E [ G ] .
30
The average IN in this case is known as the sample (aka empirical) covariance
matrix. Denote
N
1
Γ̂ =
N ∑ ( mi − m ) ( mi − m ) T
i =1
as the sample covariance matrix. Note that m is typically not available in prac-
tice, and we have to resort to a computable approximation
N
1
Γ̂ =
N ∑ (mi − m̂) (mi − m̂)T ,
i =1
with m̂ denoting the sample mean.
7.2 Independent and identically distributed random draws
Sampling methods discussed in this note are based on two fundamental iid
random generators that are available as built-in functions in Matlab. The first
one is rand.m function which can draw iid random numbers (vectors) from
the uniform distribution in [0, 1], denoted as U [0, 1], and the second one is
randn.m function that generates iid numbers (vectors) from standard normal
distribution N (0, I ), where I is the identity matrix of appropriate size.
The most trivial task is how to draw iid samples { M1 , M2 , . . . , M N } from a
multivariate Gaussian N (m, Γ). This can be done through a so-called whiten-
ing process. The first step is to carry out the following decomposition
Γ = RR T ,
which can be done, for example, using Cholesky factorization. The second
step is to define a new random variable as
Z = R −1 ( M − m ) ,
then Z is a standard multivariate Gaussian, i.e. its density is N (0, I ), for which Show that Z is a standard multivariate
Gaussian.
randn.m can be used to generate iid samples
{ Z1 , Z2 , . . . , ZN } = randn(n,N).
We now generate iid samples Mi by solving
Mi = m + RZi .
7.8 exercise. Look at BayesianPriorElicitation.m to see how we apply the

above whitening process to generate multivariate Gaussian prior random re-
alizations.
31
You may ask what if the distribution under consideration is not Gaussian,
which is true for most practical applications. Well, if the target density π (m)
is one dimensional or multivariate with independent components (in this case,
we can draw samples from individual components separately), then we still
can draw iid samples from π (m), but this time via the standard uniform
distribution U [0, 1]. If you have not seen it before, here is the definition: U [0, 1]
has 1 as its density function, i.e.,
Z
(27) µU ( A) = ds, ∀ A ⊂ [0, 1] .
A
Now suppose that we would like to draw iid samples from a one dimensional
(S = R) distribution with density π (m) > 0. We still allow π (m) to be zero,
but only at isolated points on R, and you will see the reason in a moment.
Define the cumulative distribution function (CDF) as
Zw
(28) Φ (w) = π (m) dm,
−∞
Why? then it is clearly that Φ(w) is non-decreasing and 0 ≤ Φ(w) ≤ 1. Let us define
a new random variable Z as
(29) Z = Φ ( M ) .
Our next step is to prove that Z is actually a standard uniform random vari-
able, i.e. Z ∼ U [0, 1], and then show how to draw M via Z. We begin by the
following observation
−1 ( a )
ΦZ
h i
(30) P [ Z < a] = P [Φ ( M ) < a] = P M < Φ−1 ( a) = π (m) dm,
−∞
where we have used (29) in the first equality, the monotonicity of Φ ( M) in the
second equality, and the definition of CDF (28) in the last equality. Now, we
can view (29) as the change of variable formula z = Φ (m), then combining
this fact with (28) to have
dz = dΦ (m) = π (m) dm, and z = a when x = Φ−1 ( a) .
Do you see the second equality? Consequently, (30) becomes
Za
P [ Z < a] = dz = µ Z ( Z < a) ,
−∞
32
which says that the density of Z is 1, and hence Z must be a standard uniform
random variable. In terms of our language at the end of Section 7.1, we can
define M = g ( Z ) = Φ−1 ( Z ), then drawing iid samples for M is simple by first
drawing iid samples from Z, then mapping them through g. Let us summarize
the idea in Algorithm 1.
Algorithm 1 CDF-based sampling algorithm

1. Draw z ∼ U [0, 1],
2. Compute the inverse of the CDF to draw m, i.e. m = Φ−1 (z). Go back
to Step 1.
The above method works perfectly if one can compute the analytical in-
verse of the CDF easily and efficiently; it is particularly efficient for discrete
random variables, as we shall show. You may say that you can always compute
the inverse CDF numerically. Yes, you are right, but you need to be careful
about this. Note that the CDF is an integral operation, and hence its inverse
is some kind of differentiation. The fact is that numerical differentiation is an
ill-posed problem! You don’t want to add extra ill-posedness on top of the
original ill-posed inverse problem that you started with, do you? If not, let us
introduce to you a simpler but more robust algorithm that works for multi-
variate distribution without requiring the independence of individual compo-
nents. We shall first introduce the algorithm and then analyze it to show you
why it works.
Suppose that you want to draw iid samples from a target density π (m),
but you only know it up to a constant C > 0, that is, you only know Cπ (m).
(This is perfect for our Bayesian inversion framework since we typically know
the posterior up to a constant as in (12).) Assume that we have a proposal distri-
bution q (m) at hand, for which we know how to sample easily and efficiently.
This is not a limitation since we can always take either the standard normal
distribution or uniform distribution as the proposal distribution. We further
assume that there exists D > 0 such that
(31) Cπ (m) ≤ Dq (m) ,
then we can draw a sample from π (m) by the rejection-acceptance sampling

Algorithm 2.
In practice, we carry out Step 3 of Algorithm 2 by flipping an “α-coin”. In
particular, we draw u from U [0, 1], then accept m if α > u. It may seem to be
magic to you why Algorithm 2 provides random samples from π (m). Let us
confirm this with you using the Bayes formula (9).
7.9 proposition. Accepted m is distributed by the target density π.
33
Algorithm 2 Rejection-Acceptance sampling algorithm

1. Draw m from the proposal q (m),
2. Compute the acceptance probability
Cπ (m)
α= ,
Dq (m)
3. Accept m with probability α or reject it with probability 1 − α. Go back

to Step 1.
Make sure you understand this proof since

we will reuse most of it for the Metropolis- Proof. Denote B as the event of accepting a draw q (or the acceptance event).
Hastings algorithm! Algorithm 2 tells us that the probability of B given m, which is precisely the
acceptance probability, is
Cπ (m)
(32) P [ B|m] = α = .
Dq (m)
On the other hand, the prior probability of m in the incremental event dA =

[m0 , m0 + dm] in Step 1 is q (m) dm. Applying the Bayes formula for conditional
probability (4) yields the distribution of a draw m provided that it has been
already accepted
P [ B|m] q (m) dm
P [m ∈ dA| B] = = π (m) dm,
P [ B]
where we have used (32) and P [ B], the probability of accepting a draw from
q, is the following marginal probability
C C
Z Z
P [ B] = P [ B|m] q (m) dm = π (m) dm = .
D D
S S
Note that
P [ B, m ∈ dm] = π ( B, m) dm = P [ B|m] πprior (m) = P [ B|m] q (m) dm,
an application of (3), is the probability of the joint event of drawing an m from

q (m) and accept it. The probability of B, the acceptance event, is the total of
accepting probability, which is exactly the marginal probability. As a result,
we have
Z
P [m ∈ A| B] = π (m) dm,
A
34
which, by definition (6), says that the accepted m in Step 3 of Algorithm 2 is

distributed by π (m), and this is the desired result.
Algorithm 2 is typically slow in practice in the sense that a large portion
of samples is rejected, particularly for high dimensional problem, though it
provides iid samples from the true underlying density. Another problem with
this algorithm is the computation of D. Clearly, we can take very large D and
the condition (31) would be satisfied. However, the larger D is the smaller the
acceptance probability α, making Algorithm 2 inefficient since most of draws
from q (m) will be rejected. As a result, we need to minimize D and this could
be nontrivial depending the complexity of the target density.
7.10 exercise. You are given the following target density
m2

g (m)
π (m) = exp − ,
C 2
where C is some constant independent of m, and

1 if x > a
g (m) = , a ∈ R.
0 otherwise
Take the proposal density as q (m) = N (0, 1).

1. Find the smallest D that satisfies condition (31).
2. Implement the rejection-acceptance sampling Algorithm 2 in Matlab and
draw 10000 samples, by taking a = 1. Use Matlab hist.m to plot the
histogram. Does its shape resemble the exact density shape?
3. Increase a as much as you can, is there any problem with Algorithm 2?
Can you explain why?
7.11 exercise. You are given the following target density

q 2 !
1 1
π (m) ∝ exp − 2 m21 + m22 − 1 − 2 ( m2 − 1)2 ,
2σ 2δ
where σ = 0.1 and δ = 1. Take the proposal density as q (m) = N (0, I2 ),

where I2 is the 2 × 2 identity matrix.
1. Find a reasonable D, using any means you like, that satisfies condition
(31).
2. Implement the rejection-acceptance sampling Algorithm 2 in Matlab and
draw 10000 samples. Plot a contour plot for the target density, and you
should see the horse-shoe shape, then plot all the samples as dots on top
of the contour. Do most of the samples sit on the horse-shoe?
35
7.3 Markov chain Monte Carlo
We have presented a few methods to draw iid samples from a target distri-
bution q (m). The most robust method that works in any dimension is the
rejection-acceptance sampling algorithm though it may be slow in practice.
In this section, we introduce the Markov chain Monte Carlo scheme which
is the most popular sampling approach. It is in general more effective than
any methods discussed so far, particularly for complex target density in high
dimensions, though it has its own problems. One of them is that we no longer
have iid samples but correlated ones. Let us start the motivation by consider-
ing the following web-page ranking problem.
Assume that we have a set of Internet websites that may be linked to the
others. We represent these sites as nodes and mutual linkings by directed ar-
rows connecting nodes such as in Figure 11. Now you are seeking sites that
1 4
Figure 11: Five internet websites and their connections.
contains a keyword of interest for which all the nodes, and hence websites,
contain. A good search engine will show you all these websites. The question
is now which website should be ranked first, second, and so on? You may
guess that node 4 should be the first one in the list. Let us present a proba-
bilistic method to see whether your guess is correct or not. We first assign the
36
network of nodes a transition matrix P as

 
0 1 0 1/3 0
 1/2 0 0 1/3 0 
 
(33) P =  0
 0 0 0 1/2 .

 1/2 0 1/2 0 1/2 
0 0 1/2 1/3 0
The jth column of P is contains the probability of moving from the jth node
to the rest. For example, the first column says that if we start from node 1, we
can move to either node 2 or node 4, each with probability 21 . Note that we
have treated all the nodes equally, that is, the transition probability from one
node to other linked nodes is the same (a node is not linked to itself in this
model). Note that the sum of each column is 1, meaning that a website must
have a link to a website in the network.
Assume we are initially at node 4, and we represent the initial probability
density as
 
0

 0 

π0 = 
 0 ,

 1 
0
that is, we are initially at node 4 with certainty. In order to know the next node
to visit, we first compute the probability density of the next state by
π1 = Pπ0 ,
then randomly move to a node by drawing a sample from the (discrete) prob-
ability density π j (see Exercise 7.12). In general, the probability density after
k steps is given by
(34) πk = Pπk−1 = . . . = Pk π0 ,
where the jth component of πk is the probability of moving to the jth node.
Observing (34) you may wonder what happens if k approaches infinity.
Assume, on credit, the limit probability density π∞ exists, then it ought to
satisfy
(35) π∞ = Pπ∞ .
It follows that π∞ is the “normalized” eigenvector of P corresponding to unity

eigenvalue. Here, normalization means taking that eigenvector and then di-
viding by the sum of its components so that the result is a probability density.
37
Figure 12 shows the visiting frequency (blue) for each node after N = 1500
moves. Here, visiting frequency of a node is the number of visits to that node
divided by N. We expect that numerical visiting frequencies approximate the
visiting probabilities in the limit. We confirm this expectation by also plotting
the components of π∞ (red) in Figure 12. By the way, π1500 is equal to π∞ up
to machine zero, meaning that a draw from π N , N ≥ 1500, is distributed by
the limit distribution π∞ . We are now in the position to answer our ranking
question. Figure 12 shows that node 1 is the most visited one, and hence
should appear at the top of the website list coming from the search engine.
0.35
visiting frequency
first eigenvector
0.3
0.25
0.2
0.15
0.1
0.05
0
1 2 3 4 5
Figure 12: Visiting frequency after N = 1500 moves and the first eigenvector.
7.12 exercise. Use the CDF-based sampling algorithm, namely Algorithm 1,

to reproduce Figure 12. Compare the probability density π1500 with the limit
density, are they the same? Generate 5 figures corresponding to starting nodes
1, . . . , 5, what do you observe?
7.13 exercise. Using the above probabilistic method to determine the proba-
bility that the economy, as shown in Figure 13, is in recession.
The limit probability density π∞ is known as the invariant density. Invari-

ance here means that the action of P on π∞ returns exactly π∞ . In summary,
we start with the transition matrix P and then find the invariant probability
density by taking the limit. Drawings from πk are eventually distributed as the
invariant density (see Exercise 7.12). A Markov chain Monte Carlo (MCMC)
38
Figure 13: Economy states extracted from

http://en.wikipedia.org/wiki/Markov_chain
method reverses the process. In particular, we start with a target density

and look for the transition matrix such that MCMC samples are eventually
distributed as the target density. Before presenting the popular Metropolis-
Hastings MCMC method, we need to explain what “Markov” means.
Let us denote m0 = 4 the initial starting node, then the next move m1 is
either 1 or 2 or 5 since the probability density is
(36) π1 (m1 |m0 ) = [1/3, 1/3, 0, 0, 1/3] T ,
where we have explicitly pointed out that π1 is a conditional probability den-

sity given known initial state m0 . Similarly, (34) should have been written as
πk (mk |m0 ) = Pπk−1 (mk−1 |m0 ) = . . . = Pk π0 .
Now assume that we know all states up to mk−1 , say mk−1 = 1. It follows that
πk−1 (mk−1 |mk−2 , . . . , m0 ) = [1, 0, 0, 0, 0] T ,
since we know mk−1 for certain. Consequently,
πk (mk |mk−1 , . . . , m0 ) = Pπk−1 (mk−1 |mk−2 , . . . , m0 ) = P [1, 0, 0, 0, 0] T
39
regardless of the values of the states mk−2 , . . . , m0 . More generally,
π k ( m k | m k −1 , . . . , m 0 ) = π k ( m k | m k −1 ) ,
which is known as Markov property, namely, “tomorrow depends on the past

only through today”.
7.14 definition. A collection {m0 , m1 , . . . , m N , . . .} is called Markov chain if the

distribution of mk depends only on the immediate previous state mk−1 .
Next, let us introduce some notations to study MCMC methods.
7.15 definition. We call the probability of mk in A starting from mk−1 as the

transition probability and denote it as P (mk−1 , A). With an abuse of notation,
What does P (mk−1 , dm) mean? we introduce the transition kernel P (mk−1 , m) such that
Z Z
def
P ( m k −1 , A ) = P (mk−1 , m) dm = P (mk−1 , dm) .
A A
R
Clearly P (m, S) = S P (m, p) dp = 1.
7.16 example. Let mk−1 = 4 in our website ranking problem, then the proba-
What is the transition probability bility kernel P (mk−1 = 4, m) is exactly the probability density in (36).
P (mk−1 = 4, mk = 1)?
7.17 definition. We call µ (dm) = π (m) dm the invariant distribution and

π (m) invariant density of the transition probability P (mk−1 , dm) if
Z
(37) µ (dm) = π (m) dm = P ( p, dm) π ( p) dp.
S
7.18 example. The discrete version of (37), applying to our website ranking
problem, reads
5
π∞ ( j) = ∑ P ( j, k) π∞ (k) = P( j, :)π∞ , ∀ j = 1, . . . , 5,
k =1
which is exactly (35).
7.19 definition. A Markov chain {m0 , m1 , . . . , m N , . . .} is reversible if
(38) π (m) P (m, p) = π ( p) P ( p, m) .
The reversibility relation (38) is also known as detailed balanced equation.

You can think of the reversibility saying that the probability of moving from
m to p is equal to the probability of moving from p to m.
40
7.20 exercise. What is the discrete version of (38)? Does the transition matrix
in the website ranking problem satisfy the reversibility? If not, why? How
about the transition matrix in Exercise 7.13.
The reversibility of a Markov chain is useful since we can immediately

conclude that π (m) is its invariant density.
7.21 proposition. If the Markov chain {m0 , m1 , . . . , m N , . . .} is reversible with re-

spect to π (m), then π (m) is the invariant density.
Proof. We need to prove (37), but it is straightforward since

Z Z
reversibility
π ( p) P ( p, dm) dp = π (m) dm P (m, p) dp = π (m) dm.
S S
The above discussion shows that if a Markov chain is reversible then even-
tually the states in the chain are distributed by the underlying invariant dis-
tribution. A question you may ask is how to construct a transition kernel such
that reversibility holds. This is exactly the question Markov chain Monte Carlo
methods are designed to answer. Let us now present the Metropolis-Hastings
MCMC method in Algorithm 3.
Algorithm 3 Metropolis-Hastings MCMC Algorithm

Choose initial m0
for k = 0, . . . , N do
1. Draw a sample p from the proposal density q(mk , p )

2. Compute π ( p), q(mk , p), and q( p, mk )
3. Compute the acceptance probability
π ( p)q( p, mk )

α(mk , p) = min 1,
π (mk )q(mk , p)
4. Accept and set mk+1 = p with probability α(mk , p). Otherwise,

reject and set mk+1 = mk
end for
The idea behind the Metropolis-Hastings Algorithm 3 is very similar to

that of rejection-acceptance sampling algorithm. That is, we first draw a sam-
ple from an “easy” distribution q (mk , p), then make correction so that it is
distributed more like the target density π ( p). However, there are two main
41
differences. First, the proposal distribution q (mk , p) is a function of the last

state mk . Second, the acceptance probability involves both the last state mk
and the proposal move p. As a result, a chain generated from Algorithm 3 is
in fact a Markov chain.
What remains to be done is to show that the transition kernel of Algorithm
3 indeed satisfies the reversibility condition (38). This is the focus of the next
proposition.
7.22 proposition. Markov chains generated by Algorithm 3 are reversible.
Proof. We proceed in two steps. In the first step, we consider the case in which
the proposal p is accepted. Denote B as the event of accepting a draw q (or the
acceptance event). Following the same proof of Proposition 7.9, we have
P [ B | p ] = α ( m k , p ),
Make sure you see this! leading to
π ( B, p) = P [ B| p] πprior ( p) = α(mk , p)q (mk , p) ,
which is exactly P (mk , p), the probability density of the joint event of drawing
p from q (mk , p) and accept it, starting from mk . It follows that the reversibility
holds since
π ( p)q( p, mk )

π (mk ) P (mk , p) = π (mk ) q (mk , p) min 1,
π (mk )q(mk , p)
= min {π (mk )q(mk , p), π ( p)q( p, mk )}
π (mk )q(mk , p)

= min , 1 π ( p)q( p, mk )
π ( p)q( p, mk )
= π ( p) P ( p, mk ) .
In the second step, we remain at mk , i.e., mk+1 = mk , then the reversibility

Explain why is trivially satisfied no matter what the transition kernel P (mk , p) is. This is
the end of the proof.
7.23 exercise. What is the probability of staying put at mk ?
As you can see the Metropolis-Hastings algorithm is simple and elegant,

but provides us a reversible transition kernel, which is exactly what we are
looking for. The keys behind this are Steps 3 and 4 in Algorithm 3, known
as Metropolized steps. At this point, we should be able to implement the
algorithm except for one small detail: what should we choose for the proposal
42
density? Let us choose the following Gaussian kernel

1 1
q (m, p) = p exp − 2 km − pk2 ,
2πγ2 2γ
which is the most popular choice. Metropolis-Hastings algorithm with above

isotropic Gaussian proposal is known as Random Walk Metropolis-Hastings (RWMH)
algorithm. For this particular method, the acceptance probability is very sim-
ple, i.e.,

π ( p)
α = min 1, .
π (m)
We are now in the position to implement the method. For concreteness, we

apply the RWMH algorithm on the horse-shoe shape in Exercise 7.11. We take
the origin as the starting point m0 . Let us first be conservative by choosing a
small proposal variance γ2 = 0.022 so that the proposal p is very close to the
current state mk . In order to see how the MCMC chain evolves, we plot each
state mk as a circle (red) centered at mk with radius proportional to the number
of staying-puts. Figure 14(a) shows the results for N = 1000. We observe that
the chain takes about 200 MCMC simulations to enter the high probability
density region. This is known as burn-in time in MCMC literature, which tells
us how long a MCMC chain takes to start exploring the density. In other
words, after the burn-in time, a MCMC begins to distribute like the target
density. As can be seen, the chain corresponding to small proposal variance
γ2 explores the target density very slowly. If we approximate the average
acceptance rate by taking the ratio of the number of accepted proposal over
N, it is 0.905 for this case. That is, almost all the proposals p are accepted, but
exploring a very small region of high probability density.
1.5 1.5 1.5
1 1 1
0.5 0.5 0.5
0 0 0
−0.5 −0.5 −0.5
−1 −1 −1
−1.5 −1.5 −1.5

−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5
(a) γ = 0.02 (b) γ = 5 (c) γ = 0.5
Figure 14: RWMH with different proposal variance γ2 .
Let us now increase the proposal stepsize γ to 5, and we show the corre-
sponding chain in Figure 14(b). This time, the chain immediately explores the
43
target density without any burn-in time. However, it does so in an extremely

slow manner. Most of the time the proposal p is rejected, resulting in a few
big circles in Figure 14(b). The average acceptance rate in this case is 0.014,
which shows that most of the time we reject proposals p.
The results in Figures 14(a) and 14(b) are two extreme cases, both of which
explore the target density in a very lazy manner since the chain either accepts
all the small moves with very high acceptance rate or rejects big moves with
very low acceptance rate. This suggests that there must be an optimal accep-
tance rate for the RWMH algorithm. Indeed, one can show that the optimal
acceptance rate is 0.234 [5]. For our horse-shoe target, it turns out that the
corresponding optimal stepsize is approximately γ = 0.5. To confirm this, we
generate a new chain with this stepsize, again with N = 1000, and show the
result in Figure 14(c). As can be seen, the samples spread out nicely over the
horse-shoe.
We have judged the quality and convergence of a MCMC chain by looking
at the scatter plot of the samples. Another simple approach is to look at the
trace plot of components of m. For example, we show the trace plot of the
first component in Figure 15 for the above three stepsizes. The rule of thumb
is that a Markov chain is considered to be good if its trace plot is close to
Plot a trace plot for a one dimensional a white noise one, a “fuzzy worm”, in which all the samples are completely
Gaussian white noise to see how it looks
like!
uncorrelated. Based on this criteria, we again conclude that γ = 0.5 is the best
compared to the other two extreme cases.
Nevertheless, the above two simple criteria are neither rigorous nor possi-
ble in high dimensions. This observation immediately reminds us the strong
law of large number in computing the mean and its dimension-independent
error analysis using the central limit theorem. Since the target is symmetric
about the vertical axis, the first component of the mean must be zero. Let us
use the strong law of large number to estimate the means for the above three
stepsizes and show them as cross signs in Figures 14(a), 14(b), and 14(c). As
can be seen and expected, the sample mean for the optimal stepsize is the most
accurate, though it is not exactly on the vertical axis since m1 = −0.038. This
implies that the sample size of N = 1000 is small. If we take N = 10000, the
Take N = 100000 for the optimal step- sample mean gives m1 = 0.003, signifying the convergence when N increases.
size case, and again compute the sample
mean using BayesianMCMC.m. Is the sam- However, the application of LLN and CLT is very limited for Markov
ple mean better? If not, why? chains since they don’t provide iid samples. Indeed, as in the above Markov
chain theory, the states of the chain eventually identically distributed by π (m),
but they are always correlated instead of independent since any state in the
chain depends on the previous one. What we could hope for is that the cur-
rent state is effectively independent from its kth previous state. In that case,
the effective number of iid samples √ is N/k, and the mean square error, by the
central limit theorem, decays as k/N. As the result, if k is large, the decay
rate is very slow. How to estimate k is the goal of the autocorrelation study, as
we now discuss.
44
1.5
0.5
−0.5
−1
−1.5
0 1 2 3 4 5 6 7 8 9 10
4
x 10
(a) γ = 0.02
1.5
0.5
−0.5
−1
−1.5
0 1 2 3 4 5 6 7 8 9 10
4
x 10
(b) γ = 5
1.5
0.5
−0.5
−1
−1.5
0 1 2 3 4 5 6 7 8 9 10
4
x 10
(c) γ = 0.5
Figure 15: Trace plots of the first component of m with different γ2 .
45
We shall compute the autocorrelation for each component of m separately,

therefore, without loss of generality, assume that m ∈ R1 and that the Markov
N
chain m j j=0 has zero mean. Consider the following discrete convolution
quantities
N −k
ck = ∑ m j+k m j , , k = 0, . . . , N − 1,
j =0
and define the autocorrelation of m with lag k as

ck
ĉk = , k = 0, . . . , N − 1.
c0
If ĉk is zero, then we say that the correlation length of the Markov chain is
approximately k, that is, any state m j is considered to be insignificantly corre-
lated to m j−k (and hence any state before m j−k ), and to m j+k (and hence any
state after m j+k ). In other words, every kth sample point can be considered to
be approximately independent. Note that this is simply a heuristic and one
should be aware that independece implies un-correlation but not vice versa.
Let us now approximately compute the correlation length for three Markov
chains corresponding to γ = 0.02, γ = 0.5, and γ = 5, respectively, with
N = 100000. We first subtract away the sample mean as
N
1
zj = mj − ∑ m.
N + 1 i =0 i
Then, we plot the autocorrelation functions ĉk for each component of the zero
N
mean sample z j j=0 in Figure 16. As can be observed, the autocorrelation
length for the chain with optimal stepsize γ = 0.5 is about k = 100, while the
others are much larger (not shown here). That is, every 100th sample point can
be considered to be independent for γ = 0.5. The case with γ = 0.02 is the
worst, indicating slow move around the target density. The stepsize of γ = 5
is better, but so big that the chain remains at each state for a long period of
time, and hence autocorrelation length is still significant relatively to that of
γ = 0.5.
Extensive MCMC methods including improvements on the standard RWMH
algorithm can be found in [6]. Let us introduce two simple modifications
through the following two exercises.
7.24 exercise. Consider the following target density

1 2 1 2
(39) π (m) ∝ exp − 2 kmk − 2 ky − h (m)k ,
2δ 2σ
46
8. Matlab codes
where
m21 − m2

−0.2
h (m) = , y= .
m2 /5 0.1
Take δ = 1 and σ = 0.1.
1. Modify BayesianMCMC.m to simulate the target density in (39) with N =

5000.
2. Tune the proposal stepsize γ so that the average acceptance probability

is about 0.234. Show the scatter, trace, and autocorrelation plots for the
optimal stepsize.
7.25 exercise. So far the proposal density q (m, p) is isotropic and indepen-
dent of the target density π (m). For anisotropic target density, isotropic pro-
posal is not a good idea, intuitively. The reason is that the proposal is dis-
tributed equally in all directions, whereas it is not in the target density. A
natural idea is to shape the proposal density to make it locally resemble the
target density. A simple idea in this direction is to linearize h (m), and then
define the proposal density as

1 2 1 2
q (mk , p) ∝ exp − 2 k pk − 2 ky − h (mk ) − ∇h (mk ) ( p − mk )k ,
2δ 2σ

1. Determine H (mk ) such that q (mk , p) = N mk , H (mk )−1 , by keeping
only the quadratic term in p − mk .
2. Modify BayesianMCMC.m to simulate the targetdensity in (39) using the

proposal density q (mk , p) = N mk , H (mk )−1 . Show the scatter, trace,
and autocorrelation plots. Is it better than the isotropic proposal density?
7.26 exercise. Another idea to improve the standard RWMH algorithm is by

adaptation. Let’s investigate a simple adaptation strategy. Use the resulting
sample in Exercise 7.24 to compute the empirical covariance Γ̂, then use it to
construct the proposal density q (m, p) = N m, Γ̂ . Show the scatter, trace,

and autocorrelation plots. Is it better than the isotropic proposal density?
8 Matlab codes
A set of Matlab codes that can be used to reproduce most of the figures in the
note can be downloaded from
http://users.ices.utexas.edu/~tanbui/teaching/Bayesian
47
References
References
[1] D. Calvetti and E. Somersalo. Introduction to Bayesian Scientific Computing:
Ten Lectures on Subjective Computing. Springer, New York, 2007.
[2] Jari Kaipio and Erkki Somersalo. Statistical and Computational Inverse Prob-
lems, volume 160 of Applied Mathematical Sciences. Springer-Verlag, New
York, 2005.
[3] Todd Arbogast and Jerry L. Bona. Methods of Applied Mathematics. Univer-
sity of Texas at Austin, 2008. Lecture notes in applied mathematics.
[4] Rick Durret. Probability: theory and examples. Cambidge University Press,
2010.
[5] Gareth O. Roberts and Jeffrey S. Rosenthal. Optimal scaling for various
metropolis-hastings algorithms. Statistical Science, 16(4):pp. 351–367, 2001.
[6] Christian P. Robert and George Casella. Monte Carlo Statistical Methods
(Springer Texts in Statistics). Springer-Verlag New York, Inc., Secaucus, NJ,
USA, 2005.
48
References
Sample Autocorrelation Function (ACF)

1
0.9
0.8
0.7
Sample Autocorrelation
0.6
0.5
0.4
0.3
0.2
0.1
1st component
2nd component
0
10 20 30 40 50 60 70 80 90 100
Lag
(a) γ = 0.02

1
1st component
2nd component
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
10 20 30 40 50 60 70 80 90 100
Lag
(b) γ = 5

1
1st component
2nd component
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
10 20 30 40 50 60 70 80 90 100
Lag
(c) γ = 0.5
Figure 16: Autocorrelation function plot for both components of m with dif-
ferent γ2 .
49

Latihan Soal

Uploaded by

Copyright:

Available Formats

Latihan Soal

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Latihan Soal

Uploaded by

Copyright:

Available Formats

A Gentle Tutorial on Statistical Inversion

using the Bayesian Paradigm

2 Some concepts from probability theory 3

5 Posterior as the solution to Bayesian inverse problems 19

6 Connection between Bayesian inverse problems and deterministic in-

7 Markov chain Monte Carlo 27

Here, a : [0, 1] × [0, 1] → R is known as the blurring kernel. So, in fact we

where κ is known as the regularization parameter, and k·k is some appropriate

ask is how to incorporate this kind of randomness in the above deterministic

2 Some concepts from probability theory

We begin with the definition of randomness.

2.1 definition. An even is deterministic if its outcome is completely predictable.

2.2 definition. A random event is the complement of a deterministic event,

2.3 example. If today is Wednesday, then “tomorrow is Thursday” is deter-

As a result, randomness means lack of information and it is the direct

probability or Bayesian probability since it represents belief, and hence depend-

We define (Ω, F , P) as a probability space. One typically call Ω the sample

Two events A and B are independent3 if

One of the central ideas in Bayesian probability is the conditional probabil-

it departs from the measure theory umbrella.

The conditional probability can also be understood as the probability when

2.7 exercise. Determine P [ A| B] in Figure 1.

Figure 1: Demonstration of conditional probability.

By inspection, if A and B are mutually independent, we have

The probability space (Ω, F , P) is an abstract object which is useful for

2.10 definition. A random variable M is a map6 from the sample space Ω to

We call M (ω ) a random variable since we are uncertain about its outcome.

where we have used the popular notation7

π M is the Radon-Nikodym derivative of µ M with respect to the Lebesgue measure. As a result,

Figure 2: Demonstration of random variable: µ M ( A) = P [ B].

We deduce the change of variable formula

We will occasionally simply write π (m) instead of π M (m) if there is no am-

2.12 remark. In theory, we introduce the probability space (Ω, F , P) in order

2.14 definition. The expectation or the mean of a random variable M is the

and the variance is

2.15 definition. Denote µ MY and π MY as the joint distribution and density,

11 Again, they must be measurable.

where the notation A × B ⊂ S × T simply means that A ∈ S and B ∈ T.

We say that M and Y are independent if

π MY (m, y) = π M (m) πY (y) .

2.16 definition. The marginal density of M is the probability density of M

Similarly, the marginal density of Y is the density of Y regardless of M,

Before deriving the Bayes formula, we define conditional density π (m|y)

Let us prove the following important result.

2.17 theorem. The conditional density of M given Y is given by

Proof. From the definition of conditional probability (3), we have

µ M|Y ({m ∈ A} |y) = P [{ M ∈ A} |Y = y] (definition (5))

which ends the proof.

π (m, y) = π (m|y) π (y) = π (y|m) π (m) ,

from which the well-known Bayes formula follows

2.19 definition (Likelihood). We call π (y|m) the likelihood. It is the proba-

2.23 exercise. Show that

In this section, we present a popular approach to construct the likelihood. We

where y ∈ Rr . But due to random additive noise E, we have the following

Figure 3: Likelihood model with additive noise.

We next consider multiplicative noise. The statistical model in this case

As discussed previously, the prior belief depends on a person’s knowledge

where a (s, t) = √ 1 exp(− 2β1 2 (t − s)2 ) is a given kernel, and s j = j/n, j =