Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science

6.432 Stochastic Processes, Detection and Estimation
Problem Set 3
Spring 2004
Issued: Thursday, February 19, 2004

Due: Thursday, February 26, 2004

Reading: This problem set: Chapter 2, Chapter 3 through Section 3.2.4

Next: Sections 1.7, 3.2.5, 3.3.1, 3.3.2

Problem 3.1
Suppose x and y are the random variables from Problem Set 2, problem 2.4. Their
joint density, depicted again below for convenience, is constant in the shaded area
and 0 elsewhere.

p x,y(x ,y )

-2

-1

2
x

-1
-2

(a) In the (PD , PF ) plane sketch the operating characeristic of the likelihood ratio
test for this problem. Also, indicate on this plot the region consisting of every
(PD , PF ) value that can be achieved using some decision rule.
(b) Is the point corresponding to PD = 23 , PF =
test that achieves this value. If not, explain.

5
6

in this region? If so, describe a

Problem 3.2
We observe a random variable y and have two hypotheses, H0 and H1 , for its prob
ability density. In particular, the probability densities for y under each of these two
hypotheses are depicted below:

py |H1 (y|H1 ) = 3y 2 , 0 y 1

py |H0 (y|H0 ) = 1, 0 y 1
3

y
1

(a) Find the decision rule that maximize PD subject to the constraint that PF 14 .
(b) Determine the value of PD for the decision rule specied in part (a).
Problem 3.3
Let k denote the uptime of a communications link in days. Given that the link is
functioning at the beginning of a particular day there is probability q that it will go
down that day. Thus, the uptime of the link k (in days) obeys a geometric distribution.
It is known that q is one of two values: q0 or q1 . By observing the actual uptime of
the link in an experiment, we would like to determine which of the two values of q
applies. The hypotheses are then:
H0 : pk|H [k|H0 ] = Pr [k = k|H = H0 ] = q0 (1 q0 )k k = 0, 1, 2,
H1 : pk|H [k|H1 ] = Pr [k = k|H = H1 ] = q1 (1 q1 )k k = 0, 1, 2, .
Suppose that q0 = 1/2, q1 = 1/4, and we observe k = k in our experiment.

. Find the decision rule H(k)

corresponding
(a) Let Pr [H = H0 ] = Pr [H = H1 ] = 1
2
to the minimum probability of error rule.
(b) Make an approximate labeled sketch of the operating characteristics for the
likelihood-ratio test (LRT).
2

(k) that maximizes PD = Pr H

= H1 H = H1 , sub(c) Find the decision rule H

ject to the constraint that PF 0.01, where PF = Pr H = H1 H = H0 .

What is the corresponding value of PD ?
Problem 3.4 (practice)
A 3-dimensional random vector y is observed, and we know that one of the three
hypotheses is true:
H1 :
H2 :
H3 :

y = m1 + w
y = m2 + w
y = m3 + w,

where

y1
y = y2 ,
y3

1
m1 = 0 ,
0

0
m2 = 1 ,
0

0
m3 = 0 ,
1

and w is a zero-mean Gaussian vector with covariance matrix 2 I.

(a) Let

Pr [H = H1 |y = y]
1 (y)
(y) = Pr [H = H2 |y = y] = 2 (y) ,
3 (y)
Pr [H = H3 |y = y]

and suppose that the Bayes costs are

C11 = C22 = C33 = 0,

C12 = C21 = 1,

C13 = C31 = C23 = C32 = 2.

(i) Specify the optimum decision rule in terms of 1 (y), 2 (y), and 3 (y).
(ii) Recalling that 1 + 2 + 3 = 1, express this rule completely in terms of
1 and 2 , and sketch the decision regions in the (1 , 2 ) plane.
(b) Suppose that the three hypotheses are equally likely a priori and that the Bayes
costs are

1, i = j
Cij = 1 ij =
.
0, i = j
Show that the optimum decision rule can be specied in terms of the pair of
sucient statistics
2 (y) = y2 y1 ,
3 (y) = y3 y1 .
3

Hint: To begin, see if you can specify the optimum decision rules in terms of
Li (y) =

py|H (y|Hi )
,
py|H (y|H1 )

for i = 2, 3.

Problem 3.5
A former 6.432 student, skilled in the methods of hypothesis testing, came upon a
professional gambler who specialized in betting on ips of a coin. Our friend knew
that there were three possibilities for the coin the gambler chose:
* The coin is fair. That is, each ip of the coin is independent of all other tosses
and has an equal probability of coming up heads or tails.
* The coin is biased towards heads. Specically, while successive ips of the coin
are independent, the probability that any individual ip comes up heads is 3/4.
* The successive tosses of the coin are not independent. Specically, while the
marginal probability of head (or tails) on any given ip is 1/2, the probability
that the next ip yields the same result as the preceding one is only 1/4. For
example, the probability that the next ip comes up heads given that the current
ip is a head is 1/4.
What our friend plans to do is observe two ips of the gamblers coin and then to
make a decision among these three possibilities regarding the type of coin the gambler
is using.
(a) Assume that our 6.432 experts prior assessment is that each of these three
possibilities is equally likely and that she views as equally bad any error she
might make (i.e. she simply wants to minimize the probability of error). De
termine the best decision she can make for each of the possible outcomes of the
gamblers pair of coin ips.
(b) Determine the probability of error associated with the decision rule determined
in part (a).
Problem 3.6
Suppose x and y are random variables. Their joint density, depicted below, is
constant in the shaded area and zero elsewhere.

y
1

-1

x
1

-1

(a) Determine xBLS (y ), the Bayes Least Squares estimate of x given y .

(b) Determine BLS , the error variance associated with your estimator in (a).
(c) Consider the following modied least squares cost function (corresponding to
the cost of estimating x as x):

(x x)2 ,
x <0
C(x , x) =
2
K(x x) , x > 0
where K > 1 is a constant.
Determine xMLS (y ), the associated Bayes estimate of x for this cost criterion.
(d) Give a brief intuitive explanation for why your answers to (a) and (c) are either
the same or dierent.
Problem 3.7
Let y be a discrete-valued random variable taking on nonnegative integer values,
y = 0, 1, 2, . . . , and let x be a continuous-valued random variable. Suppose that the
conditional probability mass function for y given x is given by a Poisson density
xy x
P r[y |x = x] = e ,
y!

y = 0, 1, 2, . . .

and suppose that x has the probability density

2e2x x > 0
px (x) =
0
x<0
Compute xM AP (y ), the MAP estimate of x based on y .
5

Problem 3.8 (practice)

When a light beam of known intensity v is shined on a photomultiplier the number
of photocounts observed in an interval of duration T is a Poisson random variable
with mean v, where > 0 is a known constant. Suppose we shine a light beam of
intensity v (v is a random variable) on a photomultiplier and measure

y1
y2

y = ..
.
yL

where yi is the number of counts in the ith of a set of L non-overlapping intervals

each of duration T . Conditioned on v , the numbers of photocounts observed in
non-overlapping intervals are statistically independent random variables. For each
1 i L,
(v)k ev
Pr [yi = k|v = v] =
,
k = 0, 1, 2, .
k!
(a) Suppose pv (v) = v 1 ev/v u(v). Determine vBLS , the Bayes least-squares esti
mate of v , and the resulting mean-square estimation error.
You may nd the following identity useful:

k!
xk eax dx = k+1
a
0
(b) As a check on your answer to part (a), verify that when v , the estimate
of part (a) reduces to
L

1
vBLS
yi + 1 .
L i=1