Problem 1 Nuclear reactor (20 credits)

A new design of a nuclear reactor is suggested for which it is claimed that a nuclear meltdown is impossible. In order
to verify this, a numerical code h (θ) has been developed which (perfectly) simulates the reactor under operating
conditions θ, where it is known that θ ∼ p (θ). The numerical code returns for each θ either 1 or 0 according to

(
1 if nuclear meltdown occurs
h (θ) = (1.1)
0 if NO nuclear meltdown occurs

Using this code a Monte Carlo simulation with N = 108 samples has been run in order to estimate the probability q
of a nuclear meltdown. This Monte Carlo estimate of the the failure probability q was exactly zero.

We introduce the following two hypotheses

on
• H1 : nuclear meltdown is impossible, i.e. q = 0
• H2 : nuclear meltdown is possible but rare, i.e. q = 10−10

We assume that the hypotheses are a-priori equally probable.

ti
0 a) Write down the Monte Carlo estimator used to obtain the numerical estimate
1
2
The Monte Carlo estimator is given by

lu
3
4
5
N
1 X  (i) 
Iˆ = h θ θ (i) ∼ p (θ)
N i=1
So
with θ (i) i.i.d samples from p (θ) and N = 108 .
e
pl
m
Sa

– Page 2 / 10 –
b) Assess the plausibility of H1 compared to H2 within the Bayesian framework given the results obtained during 0
the Monte Carlo study. Can you give preference to either of the hypotheses? 1
2
3
Due to a-priori equal probabilites of hypothesis H1 and H2 it follows 4
5
6
p (H1 |D) p (D|H1 ) 7
=
p (H2 |D) p (D|H2 ) 8
9
Note that for a Monte Carlo estimator to yield exactly zero probability it implies that there have been 10
N = 108 numerical simulations that predicted no occurence of a nuclear meltdown; as such the data is given 11
12
by D {0, 0, ..., 0} with |D| = 108 . We therefore obtain the Bayes factor 13
14
15
p (H1 |D) p (D|H1 ) 1

on
= = 108
≈ 1.0101
p (H2 |D) p (D|H2 ) (1 − q)

I.e. one can conclude that the simulations run for the Monte Carlo analysis do not offer any evidence for a
fail-safe design vs. a small failure probability of q = 10−10 .

ti
lu
So
e
pl
m
Sa

– Page 3 / 10 –
Problem 2 SIR Model (20 credits)
The SIR model is a simple epidemiological model for the spread of an infectious disease within a population of N
people. It is given by a set of ordinary differential equation

dS I ·S
= −β (2.1)
dt N
dI βI · S
= − γI (2.2)
dt N
dR
= γI (2.3)
dt
where

on
• β > 0, γ > 0 are model parameters,

• S : number of susceptible people (i.e. people who can get infected),

• I : number of infected people
• R : number of recovered people

ti
d(S+I+R)
• Note that S + I + R = N (constant) for all t (i.e. dt = 0).

The variables S, I, R depend on time and their values at t = 0 are known. A dataset DK has been collected which

lu
K
contains the number of infected people I1 , ..., IK at time instances t1 , ..., tK , i.e. DK = {(tk , Ik )}k=1 .

A colleague of yours has already written:

So
• a Matlab function prior which takes as input (β, γ) and returns the corresponding value of the prior p (β, γ)
• a Matlab function likelihood which takes as input (β, γ) and returns the likelihood of the data DK .

You are asked to:

e
pl
m
Sa

– Page 4 / 10 –
a) complete the function PandemicRisk whichtakes as inputs the function handles prior and likelihood and 0
β 1
returns an estimate of the probability q = P r γ > 1 (The ratio R0 = β/γ is known as the basic reproduction
2
number and when R0 > 1 the number of infected people grows at an exponential rate) 3
4
Note: Both the likelihood and prior implementations return probability zero if either β ≤ 0 or λ ≤ 0. Any 5
implementation that runs in finite time is acceptable - efficiency is no consideration. 6
7
8
9
1 function q = PandemicRisk(prior, likelihood) 10
2 11
3
12
13
4
14
5 % the first part of the problem consists of writing a standard MCMC 15

on
6 % sampler as introduced in the lecture and / or exercise
7

8 N = 10^8; stepsize = 1;
9 beta = 1; gamma = 1;
10 p = prior(beta,gamma) * likelihood(beta,gamma);

ti
11

12 posterior = @(beta, gamma) prior(beta,gamma)*likelihood(beta,gamma)

13

lu
14 beta_samples = zeros(1,N);
15 gamma_samples = zeros(1,N);
16

17 for n=1:N
So
18

19 beta_proposed = beta + stepsize*randn();

20 gamma_proposed = gamma + stepsize*randn();
21 p_proposed = posterior(beta_proposed, gamma_proposed);
22

23 if rand() < p_proposed / p

24 p = p_proposed;
e

25 beta = beta_proposed;
26 gamma = gamma_proposed;
pl

27

28

29 beta_samples(n) = beta; gamma_samples(n) = gamma;

30
m

31 end
32

33 % the second part of the problem implies a Monte Carlo estimator

34 % using the samples created by MCMC
Sa

35 q = mean((beta_samples/gamma_samples) > 1);

36

37

38 end

– Page 5 / 10 –
0 b) describe or mark within your code of the PandemicRisk function where the numerical expense of solving the
1 ordinary differential equations (i.e. Eqs. (2.1) - (2.3)) occurs
2
3
4 The numerical burden of solving the ODE occurs in Line 21, where the posterior and thus the likelihood of
5 the proposed β, γ values are evaluated.

on
ti
lu
So
e
pl
m
Sa

– Page 6 / 10 –
Problem 3 Linear latent variable model (15 credits)
A set of random vectors x(i) ∈ R4 is generated as follows:

x(i) = Ly (i) + γ y (i) ∼ N (0, Σ) i = 1, ..., N (3.1)

where:
β1−1
 
0 0 0
 0 β2−1 0 0 
• y (i) ∈ R4 and Σ =   with β1 < β2 < β3 < β4 (given).
 0 0 β3−1 0 
0 0 0 β4−1

• L ∈ R4×4 (given) and orthogonal, i.e. LLT = I

on
• γ ∈ R4 (given)
 N
Given the data D = x(i) i=1 we use an Expectation-Maximization algorithm to find the maximum likelihood

estimate of the parameters W , b, σ 2 of a linear latent variable model

ti
 ∼ N 0, σ 2


x = Wz + b +  (3.2)

where

• z ∈ R2 and z ∼ N (0, I).

• W ∈ R4×2 , b ∈ R4
lu
So
As N → ∞ (i.e. as the number of data increases), what would be the maximum likelihood estimates of the following
parameters (briefly justify / explain your answer).

a) Maximum likelihood estimate of parameter b ∈ R4 0

1
2
With E [x] = γ we will obtain b = γ 3
4
e

5
pl
m

b) Maximum likelihood estimate of parameter W ∈ R4×2 0

Sa

1
2
Since L defines a rotation it is clear based on the ordering of the eigenvalues that the principal directions, i.e. 3
span (W ), will be given by the span of the two first column vectors of L. Given the spherical prior on z and 4
the rotation-invariance in the latent space, no further statements can be made. 5

– Page 7 / 10 –
0 c) Maximum likelihood estimate of parameter σ 2 ∈ R+
1
2
3 The MLE of σ 2 depends on the smallest eigenvalues of Σ, since the eigenvalues of the covariance of x are
2
4 unaffected by L and γ corresponding to a rotation or shift. Given the ordering of βi therefore σM LE only
1 −1 −1
2


5 depends on β3 and β4 , defininig the largest precision values, i.e. σM LE = 4−2 β3 + β4

on
ti
lu
So
e
pl
m
Sa

– Page 8 / 10 –
 Problem 1 Auxiliary Variable Markov Chain Monte Carlo (15 credits)
We want to make use of Markov Chain Monte Carlo to sample from a posterior distribution π (x), with x ∈ RD .
To this end we consider a MCMC Method where the proposal mechanism makes use of an auxiliary variable
u ∼ U [0, 1], as explained in the code below. The function takes as input a function handle posterior (which returns
π (x)) and the dimension D = dim (x).

0 Implement the accept-reject step (line 20 - 38) to obtain a valid Metropolis-Hastings algorithm. Do not alter the
1 code outside these lines, and do not alter the way in which proposals are generated.
2
3
4
1 function X = MCMC(posterior, D)
5
6 2

7 3 % The array X needs to contain valid samples from the posterior when returned
8 N = 10^8; X = zeros(N,D); x = ones(D,1);

n
4
9
10 5 % function handle takes D-dimensional vectors and returns posterior probability
11 6 p = posterior(x);

tio
12 7
13
14
8 for n=1:N
15 9

10 % propose (using auxiliary variable u)

11 u = rand();

lu
12 if u < 0.75
13 y = x + randn(D,1);
14 else
15 y = x + sqrt(10)*randn(D,1);
So
16 end
17

18 % ====================== IMPLEMENT ACCEPT-REJECT STEP ============================

19 % Implement a accept-reject step that is valid for the proposal used above
20

21 % q(y|x) is symmetric, and therefore the solution is

22 % is the default Metropolis accept-reject step
e

23 if rand() < posterior(y) / posterior(x)

24 y = x
pl

25 end
26

27

28
m

29

30

31

32
Sa

33

34

35

36

37 % ===============================================================================
38 X(n,:) = x;
39 end

Reminder: rand() returns a sample from U (0, 1), randn() returns a sample from N (0, 1).

– Page 2 / 12 –
Problem 2 Buckling Mode (10 credits)

We consider a beam under vertical and horizontal loading. Buckling is assumed to occur
if the horizontal force FH exceeds 10% of the vertical force FV , which we simplify to be
FV
given by the event FH > 0.1 · FV . It is known that FH and FV are jointly Gaussian

 
FH
∼ N (µ, Σ)
FV

where the parameters µ and Σ are fully defined by the marginal distributions

FH
FH ∼ N (1, 0.5) FV ∼ N (10, 2)

n
and the correlation coefficient ρ ∈ [0, 1] of FH and FV .

tio
Derive the probability q for buckling to occur. 0
1
Note: You may expess your answer wrt. the CDF Φ (·) (or its inverse) of the distribution N (0, 1). 2
3
4
The probability q is found to be independent of ρ, since 5
6

lu
7
 
FH 1 8
q=p > 0.1 = p (FH > 0.1 · FV ) = p (FH − 0.1 · FV > 0) = 9
FV 2
10
So
The last step follows by noting that Z := FH − 0.1 · FV is zero-mean Gaussian, i.e. E [FH − 0.1 · FV ) = 0.
e
pl
m
Sa

– Page 3 / 12 –
 Problem 3 Statistical Independence (5 credits)
Consider a random vector x ∼ N (0, K ) where K is:

 
1 0.5 0 0 0

 0.5 1 0.5 0 0 

K =
 0 0.5 1 0.5 0 

 0 0 0.5 1 0.5 
0 0 0 0.5 1

0 Are x1 , x5 statistically independent? Justify your answer.

1
2
3 The marginal p (x1 , x5 ) is given by a Gaussian

n
4
5  
x1
∼ N (0, I)

tio
x5

and therefore can be factorized as p (x1 , x5 ) = p(x1 )p(x2 ), implying x1 ⊥ x5 .

lu
So
Problem 4 Linear Kernel (5 credits)
5
We wish to use Gaussian Process regression to infer a mapping implied by the dataset {xn , fn }n=1 , where fn = f (xn ).
e

For this purpose we introduce a zero-mean Gaussian Process GP (0, C ) with a stationary covariance function
defined by
pl

C xi , xj = xiT xj + δij σ 2

with σ 2 ∈ R+ . We are interested in predicting f ∗ = f (x ∗ ).

m

0 Specify in which circumstance you will obtain p (f ∗ |x ∗ , D) = p (f ∗ |x ∗ ), i.e. the posterior is equal to the prior.
1
2
3 This is the case when the covariance between the observed data points and f ∗ is zero. We therefore require
Sa

4 x ∗ T xn = 0 ∀n ∈ {1, 2, 3, 4, 5}, i.e. all xn in the dataset need to be orthogonal to x ∗ (leading to exclusively
5 zero off-diagonal entries and no coupling).

– Page 4 / 12 –
Problem 5 Linear elastic bar (21 credits)
We consider a linear elastic bar of length l = 1 and cross-sectional area Acs = 1
l=1

Acs = 1

The displacement of the linear elastic bar shown is given by

n
u (x) = θx

where θ is the reciprocal of the elastic modulus E > 0, i.e. θ = 1/E . Suppose θ is unknown and we try to determine

tio
it by measuring the displacement at x = 1. Suppose we obtain a noisy measurement u1 which is assumed to relate
to u(x = 1) as follows:

 ∼ N 0, σ 2


u1 = u (x = 1) + 

We assume that a priori θ ≥ 0 follows an exponential distribution, i.e. p (θ) = λ exp (−λθ), where λ > 0 and σ 2 > 0

lu
are given.

a) write the likelihood p (u1 |θ) 0

1
So
2
With u1 = u (x = 1) +  = θ +  the likelihood follows as 3
4
5
p (u1 |θ) = N u1 θ, σ 2

alternatively, written out

e

 
1 1
p (u1 |θ) = √ exp − (u1 − θ)2
2πσ 2 2σ 2
pl
m

b) write the posterior (up to a multiplicative constant) 0

1
Sa

2
Noting that the prior p (θ) only has support for θ ≥ 0 3
4
5
p (θ|u1 ) ∝ p (u1 |θ) p (θ) 6
 
1
∝ exp − 2 (u1 − θ)2 exp (−λθ) · Iθ≥0 (θ)
2σ

with the indicator function

(
1 if θ ≥ 0
Iθ≥0 (θ) =
0 else

– Page 5 / 12 –
0 c) determine the MAP estimate of θ (for σ 2 = 1, λ = 1/2)
1
2
3  
4 1 1
5 p (θ|u1 ) ∝ exp − (u1 − θ)2 − θ · Iθ≥0 (θ)
2 2
 
1 2 2

= exp − u1 − 2θu1 + θ + θ · Iθ≥0 (θ)
2
   
1 2 1
∝ exp − θ − 2θ u1 − · Iθ≥0 (θ)
2 2
  2 !
1 1
∝ exp − θ − u1 − · Iθ≥0 (θ)
2 2

n
Therefore θMAP = max (u1 − 0.5, 0), i.e.

tio
(
u1 − 0.5 if u1 > 0.5
θMAP =
0 else

0
1
2
lu
d) determine the Laplace approximation of the posterior if u1 = 1.5 (and σ 2 = 1, λ = 1/2)

From the result of the previous problem it follows trivially that the Laplace approximation (for u1 > 0.5) is given
So
3
by N (θ|u1 − 0.5, 1), i.e. N (θ|1, 1). Alternatively one may obtain the result by taking the second derivative at
the mode.
e
pl
m
Sa

0 e) comment on the accuracy of the Laplace approximation

1
2
Required answer: The Laplace approximation introduces an error, since it places non-zero probability mass
on θ < 0 compared to the posterior.

– Page 6 / 12 –
Problem 6 Monte Carlo estimator (24 credits)
We wish to estimate the value of the integral

Z+∞
2
I= x 3 e −2x dx
−∞

with a Monte Carlo estimator of the form

N
1 X
h xi xi ∼ p xi



Î =
N
i=1

n


where x i are independent and identically distributed random numbers drawn from a PDF p x i .


a) specify a valid choice for the form of h (·) and the distribution p x i 0

tio
1
2
1
With σ = 2
3
4
5
Z∞ Z∞ √   6
3 2 3 2πσ 2 1
exp − 2 x 2


I= x exp −2x dx = x √ dx 7

lu
2πσ 2 2σ 8
−∞ −∞
Z∞ √ h √ i
= x 3 2πσ 2 N (0, 0.25) dx = EN (0,0.25) x 3 2πσ 2
So
−∞

this implies

√
h (x) = x 3 0.5π p x i = N (0, 0.25)


e
pl
m

b) explain why for the choices above the resulting Monte Carlo estimator wil converge to I
Sa

0
1
2
Explanation should contain
  
• Ep (x i ) h(x i ) = I and Var Î → 0

• reference to law of large numbers or central limit theorem

– Page 7 / 12 –


0 c) implement the function MonteCarlo which returns the estimate Î (using your previous choice of h (·) , p x i ).
1
2
3
1 function I_hat = MonteCarlo()
4
2
5
6 3 % number of Monte Carlo samples
4 N = 10^8;
5

6 I_hat = 0;
7

8 for n = 1:N
9 % sample x_i ~ N(0, 0.25)
10 x_i = 0.5*randn();

n
11 % additional term due to normalization of N(0, 0.25)
12 I_hat = I_hat + (x^3)*sqrt(0.5*pi);
13 end

tio
14

15 I_hat = I_hat / N;
16

17

18

lu
19

20

21

22
So
23

24

25

26

27

28
e

29

30

31
pl

32

33

34

35 end
m
Sa

Reminder: rand() returns a sample from U (0, 1), randn() returns a sample from N (0, 1).

– Page 8 / 12 –
d) consider an alternative estimator (for I) of the form 0
1
N
2
1 X 3
h xi + a xi − µ




Ĩ = 4
N
i=1 5
 
6
where µ = Ep (x i ) x i and x i are again i.i.d. samples from p x i . 7
8
For which value of a will the estimator above converge the fastest? Provide a numerical value.

• You mase use: the first four moments of a zero-mean Gaussian Y ∼ N 0, σ 2 are given by

n
E Y 2 = σ2 E Y3 = 0 E Y 4 = 3σ 4
     
E [Y ] = 0

• You can reuse known results from the lecture or exercise to expedite the solution

tio
This is a Monte Carlo estimator making use of a control variate. As discussed e.g. in problem sheet 6 the
optimal value is given by

a ∗ = −Cov [h, x] /Var [x]

Since E [x] = E [h] = 0 (h is an asymmetric function), we get

lu
So
E [x · h]
a∗ = −
Var [x]
√   √
Note that E [x · h] = 2πσ 2 EN (0,0.25) x 4 = 2πσ 2 · 3σ 4 and Var [x] = σ 2 ; therefore (with σ = 0.5)

√
e

∗ 2πσ 2 · 3σ 4 √
a =− = − 2π · 3σ 3 ≈ −0.94
σ2
pl
m
Sa

– Page 9 / 12 –
 Problem 1 Monte Carlo (12 credits)
A rod in a truss system experiences significant loading on any given day with probability q = 0.1 due to wind. If the
truss is under signficant loading due to wind, the maximum stress it experiences on this day can be modeled as a
Gaussian (in MPa)

X ∼ N µ = 40, σ 2 = 9

0 Complete the Matlab function to return a numerical estimate of the expected number of days within a year (365
1 days), for which the stress exceeds 60 MPa.
2
3
4
1 function D = MonteCarlo()
5

n
6 2 % D: expected number of days for which stress exceeds 60MPa (in a year)
7 3
8 4 N = 1e6;

tio
9
10 5 mu = 40;
11 6 sigma = 3;
12 7

8 I = 0;
9

lu
10 for n=1:N
11 days_threshold_exceeded = 0;
12 for d=1:binornd(365,0.1)
13 S = mu + sigma*randn();
So
14

15 if S > 60
16 days_threshhold_exceeded = days_threshold_exceeded + 1;
17 end
18 end
19
e

20 end
21

22 D = days_threshold_exceeded / N;
pl

23

24 end
25

26
m

27

28

29
Sa

30

31

32 end

Reminder: rand() returns a sample from U (0, 1), randn() returns a sample from N (0, 1), binornd(n,p) returns a
sample from Binomial(n, p).

– Page 2 / 12 –
Problem 2 Model Selection (14 credits)
We consider a bar of length ` = 1 and cross-sectional area Acs = 1 subjected to a force F
`=1

Acs = 1

The following two models which predict the displacement of the bar along the x-direction are proposed:

n
√
M1 : u (x) = α x M2 : u (x) = β x

with the priors on the parameters α, β.

tio
p (α) = N (1, 0.01) p (β ) = N (1, 0.01)

Suppose we obtain a noisy measurement û of the displacement at x = 0.64 as follows:

lu
û = u (x = 0.64) +   ∼ N (0, 0.01)

p(M1 |û)
Find the evidence ratio , if a-priori both models are considered equally plausible and û = 0.70. 0
So
p(M2 |û)
1
Hint: there is a way to do this without solving an integral. 2
3
4
Provide a numerical value of the evidence ratio for û = 0.70. 5
6
With p (M1 ) = p (M2 ) = 0.5 7
8
e

9
p (M1 |û) p (û|M1 ) p (M1 ) p (û|M1 ) 10
R= = = 11
p (M2 |û) p (û|M2 ) p (M2 ) p (û|M2 )
pl

12
13
The remaining terms involve a (tractable) integral 14

Z Z
m

p (û|M1 ) = p (û|α, M1 ) p (α) d α p (û|M2 ) = p (û|β , M2 ) p (β ) d β

Note however that we can stidestep this integration due to the properties of a Gaussian being closed under
liner transformations and addition. With x = 0.64 we obtain for
Sa

√
Model 1: û = α x + ε

 √  √
E [û] = E α x + ε = 0.64 · E [α] + E [] = 0.8
 √ 
Var [û] = Var α x + ε = x · Var [α] + Var [] = 0.64 · 0.01 + 0.01 = 0.0164

Model 2: û = β x + ε

E [û] = E [β x + ε] = x E [β ] + E [ε] = 0.64

Var [û] = Var [β x + ε] = x 2 Var [β ] + Var [ε] = 0.642 · 0.01 + 0.01 = 0.014096

From which follows p (û|M1 ) = N (0.80, 0.0164) and p (û|M2 ) = N (0.64, 0.014096). For û = 0.70 this yields

2.2966
R≈ ≈ 0.7766
2.9574

– Page 3 / 12 –
Problem 3 Probability Density Function (20 credits)
We consider the following PDF


0
 for x < a,
 2(x −a)

(b −a)(c −a)
for a ≤ x ≤ c,
fX (x) = 2(b −x)
 (b −a)(b −c)

 for c < x ≤ b,

0 for x > b

with parameters a = 0, b = 2, c = 1.

a) Find the variance Var [X ]. Provide a numerical value. 0

1
2

n
For the given parameters, fX (x) defines a simple triangular distribution with the mean obviously E [X ] = 1. 3
4
5

tio


0 for x < 0, 6
 7
x for 0 ≤ x ≤ 1,
fX (x) = 8

2−x for 1 < x ≤ 2, 9


0 for x > 2 10

 
Making use of Var [X ] = E X 2 − E2 [X ]

E X
 2

=
Z+∞
2
x fX (x) dx =
Z1
3
x dx +
lu
Z2
x 2 (2 − x) dx
So
−∞ 0 1
 1  2
1 4 2 3 1 4
= x + x − x
4 0 3 4 1
 
1 16 2 1 7
= + −4− + =
4 3 3 4 6
e

And hence we find Var [X ] = 61 .

pl
m
Sa

– Page 5 / 12 –
 0 b) Complete the
Matlab function sample() to generate and return a single sample from the conditional distribution
1 fX |X ≤1 x |x ≤ 1 . The only random number generator you are allowed to use is rand(), which returns a sample
2 from the uniform distribution U (0, 1). Provide a short derivation / justification for your solution.
3
4
5
6 1 function x = sample()
7 2

8 3 u = rand();
9
4
10
5 % solution:
6 x = sqrt(u)
7

n
9

10

11

tio
12

13

14

15

16 end

Space for notes / derivations:

lu
With the PDF symmetric around 1, it follows p (X ≤ 1) = 0.5. Therefore the conditional PDF
So
follows immediately as

(
2x 0≤x≤1
fX |X ≤1 (x) =
0 else

With FX |X ≤1 (x) = x 2 (for 0 ≤ x ≤ 1) it follows that u ∼
U (0, 1) will yield the desired distribution
√
if mapped via x = u, since d /dxF −1 (x) = d /dx x 2 = 2x (again, for 0 ≤ x ≤ 1).
e
pl
m
Sa

– Page 6 / 12 –
Problem 4 Elastic Rod (10 credits)
A rod made of material with yield strength Y modeled with a Gaussian

Y ∼ N (100, 5)

is subjected to stress only known stochastically as (in MPa)

X ∼ N (70, 10)

Obtain the probability that the yield strength will be exceeded. 0

1
Note: Express your answer w.r.t. the CDF Φ (·) of the standard Normal distribution N (0, 1). 2

n
3
4
We are interested in the probability that X exceeds Y , i.e. 5

tio
6
7
8
p (X > Y ) = p (X − Y > 0) = 1 − p (X − Y ≤ 0)
9
10
Let D := X − Y with D ∼ N (70 − 100, 10 + 5) = N (−30, 15), then with Z ∼ N (0, 1)

lu
   
30 30
p (X > Y ) = 1 − p (D ≤ 0) = 1 − p Z≤ √ =1−Φ √ ≈ 1 − Φ (7.746)
15 15

Or equivalently (starting fromY − X > 0, or by property of Φ (·))

So
 
30
p (X > Y ) = Φ − √ ≈ Φ (−7.746)
15
e
pl
m
Sa

– Page 7 / 12 –
 Problem 5 Bayesian Inference (20 credits)
The measured displacements Y1 , Y2 of a material at two different locations are assumed to relate to an externally
applied force X as follows:

Y1 = 2X + W1
Y2 = X + W2

with W1 ∼ N (0, 2) and W2 ∼ N (0, 4) assumed independentent, and the a-priori belief X ∼ N (0, 1).

We have observed the data

Y1 = 5 Y2 = 4

n
0 a) Derive the maximum likelihood estimate of X given Y1 , Y2

tio
1
2
3 The likelihood factorizes
4
5
6 L = p (y1 , y2 |x) = N (y1 |2x, 2) · N (y2 |x, 4)
7

lu
8 working with the log-likelihood
9
10
11 11
log L = − (y1 − 2x)2 − (y2 − x)2 + const.
So
22 24
we take the derivative

d 1
log L = y1 − 2x + (y2 − x)
dx 4
9 !
=6− x =0
4
e

From this follows the maximum likelihood estimate x ∗ = 24

9
≈ 2.66.
pl
m
Sa

– Page 8 / 12 –
b) Complete the MCMC function below such that it returns N = 5000 samples from the a-posteriori distribution of X 0
using a Random Walk Metropolis Hastings algorithm. 1
2
3
4
1 function X = MCMC()
5
2 6
3 % init 7
4 N = 5000; 8
9
5 X = zeros(1,N); 10
6 y1 = 5; y2 = 4;
7 x = 0;
8

9 % note that normpdf takes sigma, not sigma^2 as input

n
10 posterior = @(x) normpdf(y1, 2*x, sqrt(2))*normpdf(y2, x, 2)*normpdf(x,0,1);
11

12 for n=1:N

tio
13

14 % propose
15 x_proposed = x + randn();
16

17 if rand() < posterior(x_proposed)/posterior(x)

lu
18 x = x_proposed;
19

20 X(n) = x;
21
So
22

23

24

25

26 end
27

end
e

28

Space for notes

pl

Identifying the posterior follows straightforwardly from Bayes’ theorem:

m

p (x |y1 , y2 ) ∝ p (y1 , y2 |x) · p (x) p (x)

= N (y1 |2x, 2) · N (y2 |x, 4) N (0, 1)

Alternatively of course p (y1 , y2 |x) can also be implemented using the joint distribution. This
Sa

is identical to the exercise, and merely the posterior needs to be adapted to the specific
case.

Reminder: rand() returns a sample from U (0, 1), randn() returns a sample from N (0, 1)

– Page 9 / 12 –
 Problem 6 Gaussian Process Regression (14 credits)
We wish to solve a regression problem using the Gaussian Process prior f ∼ GP (0, C ), where we assume a linear
kernel

C xi , xj = xi xj
N
The observed values {yn }n=1 are corrupted by additive Gaussian noise

yn = f (xn ) + n n ∼ N (0, 1)

Our dataset D consists out of two pairs (N = 2)

n
(x1 , y1 ) = (0, 1) (x2 , y2 ) = (1, 1)

tio
0 What can you say about the value of the function f at x3 = 2 based on this data?
1
2 Provide numerical values for the parameters of the PDF.
3
4
5 Let f3 = f (x3 ). The vector z = [y1 , y2 , f3 ] is jointly Gaussian with zero mean and covariance matrix following

lu
6 from Cij = xi xj , and - for fy , y2 , the additive Gaussian noise (+δij ).
7
8
9  
1 0 0
10
So
11
C= 0 2 2 
12 0 2 4
13
14 which implies that we can disregard the obserd value y1 due to independence. We can instead consider

     
y2 0 2 2
∼N ,
f3 0 2 4
e

from the lecture / exercise it is known that the conditional distribution follows as
pl

p (f3 |D) = N µ, σ 2

with
m

µ = 2 · 2−1 (1 − 0) = 1.0 σ 2 = 4 − 2 · 2−1 · 2 = 2

Sa

– Page 10 / 12 –