PROBLEMS 85
so that
(3.25)
Note that P, has increased over the binary case (see (3.15)) since the detector must
decide among more hypotheses. The general M-ary case is discussed in Problem
3.20.
References
Kendall, Sir M., A. Stuart, The Advanced Theory of Statistics, Vol. 2, Macmillan,
New York, 1979.
Lehmann, E.L., Testing Statistical Hypotheses, J. Wiley, New York, 1959.
Problems
4
8.1 Determine the NP test for distinguishing between the hypotheses Ho :
versus 1; : jc = 1 based on the observed sample 2[0] ~ A (u, 1). Then, find
the Type I error (Pj-,) and the Type II error (Py = 1— Pp, where Py is the
probability of a miss). Finally, plot Py versus Ppa.
3.2 For the introductory example of Section 3.3 find two other critical regions to
satisly Pr, = 10-4.
3.3 Show that if g(x) is a monotonically increasing function of x, then L(x) > 7 if
and only if g(L(x)) > g(7). By monotonically increasing, we mean that the
function g(r) satisfies g(2) > g(1) if and only if x2 > x1
3.4 For the DC level in WGN detection problem assume that we wish to have Pra =
10-4 and Pp = 0.99. If the SNR is 10logyg A*/o? = —30 dB, determine the
necessary number of samples NV.
3.5 For the DC level in WGN detection problem consider the detector that decides
Hi ifz > 7. Since Fis an estimator of A or A = Z, a measure of the estimation
accuracy is £?(A)/var(A). Relate this quantity to the signal ENR.
3.6 Modify Example 3.2 so that now A < 0. Determine the NP detector and its
detection performance to show that it is the same as for A > 0. Hint: Use
Qa) =-Q(1—2).
3.7 We observe the ITD samples an) forn=0,1, N—1 from the Rayleigh PDF
z[n} 12 jn
p(a{nl) = 2) oxp (428) ’
a 2 a86 CHAPTER 3. STATISTICAL DECISION THEORY I
Derive the NP test for the hypothesis testing problem
Ho :0? = oy
Hh 10? =o? > 02.
3.8 Find the NP test to distinguish between the hypotheses that a single sample
{0] is observed from the possible PDFs
Ho: p(ao}) = 3 exp (—|x(0])
Hy : p(e(0]) = Ae exp (—42°[0))
‘Show the decision regions. Hint: You will need to solve a quadratic inequality.
3.9 For Example 3.3 let N = 2 and show that the detection performance is sum-
marized by s
Po = (Pra)?
Note that Pp is monotonically increasing with o?/a9.
8.10 For the DC level in WGN detection problem discussed in Example 3.2 show
that z is a sufficient statistic for A. Use the identity
Na na
DY (an) — 4? = YO 2? fn] -24Nz + NA?
m0 =
to effect the factorization.
3.11 The exponential family of PDEs is defined by
(20) = exp [A()B(x) + C(x) + DO)
where 0 is a parameter. Show that the Gaussian PDF with parameter js or
eek we ee
P(x; 4) = Tm? [ Jat — HY |
is a special case of this family. Next for the exponential family show that a
sufficient statistic for @ is T(x) = 2%! B(a\n|) based on the observed data
z|n) for n = 0,1,...,N —1. Using this result, find the sufficient statistic for
1 in the Gaussian PDF. The reader may also wish to refer to [Kay-I 1993,
Problems 5.14, 5.15).
3.12 Design a perfect detector for the problem
Ho : 2[0] ~Uf-ce]
Hii) ~Ull—ol +
where ¢ > 0 and U[a,6] denotes a uniform PDF on the interval [a,b], by
choosing ¢. A perfect detector has Pag =0 and Pp =1 ,PROBLEMS i
3.13 Prove that the ROC is a concave function over the interval [0,1]. A concave
function is one for which
ag(z1) + (1 a)o(22) < g(a: + (1 —a)z2)
for 0 < a < 1 and any two points x1 and x2. To do so consider two points
on the ROC (pr, Po(pi)) and (p2,Pp(p2)) and find Pp for a randomized test.
A randomized test: first ips a coin with Pr{head} = a. If the outcome is @
head, we employ the detector whose performance is (pi, Pp(p1)). Otherwise,
we employ the detector whose performance is (p2, Pp(p2)). We decide 1, if the
chosen detector decides 11. Hint: For a given Pr4 the detection performance
of the randomized detector mnst be less than or equal to that of the NP
detector.
3.14 Consider the hypothesis testing problem
m(ehtes) oo
«(| : Lo }) under #4,
where x = [x{0]r{1]]” is observed. Find the NP test statistic (do not evaluate
the threshold) and explain what happens if p = 0.
a
3.15 Consider the detection of a signal s[n] embedded in WGN with variance a?
based on the observed samples x(n] for n = 0,1,...,2N —1. The signal is
given by
A n=0,1,...,N-1
0 n=N,N+1,...,.2N-1
under Ho and by
Inj —{ A, PaO tne V1
SIN=\ 24 n=N,N41,...,2N-1
under #1). Assume that A > 0 and find the NP detector as well as its detection
performance. Explain the operation of the detector.
3.16 In Example 3.5 find the optimal detector if the prior probability P(71) is
arbitrary. For N = 1, A = 1, and o? = 1, find the detector, including the
threshold, for the prior probabilities P(Ho) = P(Hi) = 1/2 and also for
P(Ho) = 1/4, P(Hy) =3/4, Explain your results.
3.17 Assume that we wish to distinguish between the hypotheses Ho : x ~ N'(0, 01)
and Hy : x ~ N(u, 071) based on x = [2(0]-r{1]|7. If P(Ho) = P(Hs), find
the decision regions that minimize P,. Hint: Show that the decision region
boundary is a line that is the perpendicular bisector of the line segment from
O tow.88 CHAPTER 3, STATISTICAL DECISION THEORY I
3.18 Find the MAP decision rule for
Ho : x[0] ~.N(0,1)
Hy : 2/0] ~ (0.2)
if P(Ho) = 1/2 and also if P(Mo) =3/4. Display the decision regions in each
case and explain.
3.19 For the simple hypothesis testing problem of this chapter one can show that
an NP test based on the sufficient statistic is equivalent to an NP test based
on the original data [Kendall and Stuart 1979]. In this problem we show how
this result simplifies the derivation of a detector. Consider Example 3.6 and
recall the result of Problem 3.10 that the sufficient statistic for the DC level
is just the sample inean x. Find the ML detector based on the observed value
of the sample mean.
3.20 A general M-ary PAM communication system transmits one of M DC levels.
Let the levels be {0, +4, +2A,...,+(M —1)4/2} for M odd. The received
data x|n] for n = 0,1,...,N —1 will be one of the DC levels embedded in
WGN with variance 0”. Using the concept of sufficient statistics from Problem
3.19 find the ML detector. ‘Then, show that the minimum P, is
(F2)
3.21 Design a minimum P, detector to decide among the hypotheses whose PDFs
are
P=
e
for equal prior probabilities.
wa(|Ma) = exp (-fx(0] +1)
p([0]|Pt1)
pox (-H2(0l)
plelOlit) =f exp (-l2(0)—1))
assuming equal prior probabilities. Also, find the minimum P..