S.T (Tripos)

MATHEMATICAL TRIPOS Part III
Tuesday, 15 June, 2021 12:00 pm to 2:00 pm
PAPER 223
ROBUST STATISTICS
Before you begin please read these instructions carefully
Candidates have TWO HOURS to complete the written examination.
Attempt no more than THREE questions.

There are FOUR questions in total.
The questions carry equal weight.
STATIONERY REQUIREMENTS SPECIAL REQUIREMENTS

Cover sheet None
Treasury tag
Script paper
Rough paper
You may not start to read the questions

printed on the subsequent pages until
instructed to do so by the Invigilator.
2
1 Changing neighborhoods.
Consider the minimax bias problem
min max b({Tn }, F ).

{Tn }⊆T F ∈PK (Φ)∩M
Here, we write
PK (Φ) := F : sup |F (t) − Φ(t)| 6
t∈R
to denote the Kolmogorov -neighborhood of the standard normal distribution, and let M
denote the class of distributions with a finite variance and a probability density function
that is nonzero on all of R. We write T to denote the class of translation-invariant
estimators for which the asymptotic
bias b({Tn }, F ) = |limn→∞ EF (Tn )| is well-defined for
1
all F ∈ M. Suppose ∈ 0, 2 .
(a) Show that when {Tn } corresponds to the sample median, we have the upper bound

−1 1
max b({Tn }, F ) 6 Φ + := b1 .
F ∈PK (Φ)∩M 2

[Hint: You may use, without proof, the fact that b({Tn }, F ) = F −1 12 for F ∈ M.]
(b) Suppose we can construct symmetric distributions F+ , F− ∈ PK (Φ) ∩ M, centered

at ±b1 , such that F− (t) = F+ (t + 2b1 ). Show that this implies the lower bound
min max b({Tn }, F ) > b1 .

{Tn }⊆T F ∈PK (Φ)∩M
Thus, the median is minimax optimal.
(c) Now find a construction of F+ and F− according to the prescription in part (b).
Part III, Paper 223

3
2 Influence vs. breakdown.

For 0 6 α < 12 , consider the α-trimmed mean, defined as
n−m
X
1
Tn (x1 , . . . , xn ) := x(i) ,
n − 2m
i=m+1
where m = bαnc and x(i) denotes the ith order statistic. If we define the functional
Z 1−α
1
T (F ) = F −1 (s)ds,
1 − 2α α
P i.i.d.
it can be shown (under appropriate regularity conditions) that Tn → T (F ) when xi ∼ F .
(a) Compute the influence function IF (x; T, F ) when F is a differentiable, strictly

monotonic cumulative distribution function of a distribution with a probability
density function which is symmetric around 0.
[You may assume that interchanging derivatives and integrals is allowed, and also
use the fact that
d −1 s − ∆x (F −1 (s))
Ft (s) = ,
dt t=0 F 0 (F −1 (s))
when Ft = (1 − t)F + t∆x , without proof.]
[Hint: Your answer should be the influence function of the Huber M -estimator with
parameter k = −F −1 (α). You may find it useful to note that F −1 (1 − t) = −F −1 (t)
−1
and F 0 is even. Also recall the formula dF dt (t) = F 0 (F 1−1 (t)) .]
1
(b) Show that the breakdown point of the trimmed mean is n bαnc.
Part III, Paper 223 [TURN OVER]

4
3 Median as a scale M -estimator.

Consider the normal-scale family, where Fθ is the cdf of a N (0, θ2 ) distribution. Let
Tn denote the sample median of {|x1 |, . . . , |xn |} (defined in the usual way as the average
of the two middle order statistics when n is even).
i.i.d.
(a) Suppose the |xi |’s are unique (which happens with probability
P 1 when xi ∼ Fθ ).
Show that Tn is a solution to the estimating equation n1 ni=1 ψ xti = 0, where
ψ(u) = sign(|u| − 1), and sign(u) is defined in the usual way to be ±1 for ±u > 0
and 0 for u = 0. Thus, Tn is a scale M -estimator.
i.i.d.
(b) What is the asymptotic distribution of Tn when xi ∼ Fθ ? Use the result to derive
an asymptotically valid level-α hypothesis test for
H0 : θ2 = 1 vs. H1 : θ2 > 1
based on Tn .
[Hint: One way to approach this problem is to use the general theorem about location
M -estimators, which states that

√ d σ 2 (t0 )
n(Tn − t0 ) → N 0, 0 ,
(λ (t0 ))2
where t0 is a root of λ(t) = EF [ψ(xi − t)] and σ 2 (t) = EF [ψ 2 (xi − t)] − λ2 (t).]
Part III, Paper 223

5
4 Not quite an M -estimator.

Suppose {xi }ni=1 are i.i.d. with µ := E(xi ) and E(x2i ) 6 σ 2 < ∞. Let ψ be a
non-decreasing function satisfying

t2 t2
− log 1 − t + 6 ψ(t) 6 log 1 + t + .
2 2
(a) Show that for any θ > 0, we have

" n
!# n
!
X θ2 X
E exp (ψ(θxi ) − θE(xi )) 6 exp 2
E(xi ) ,
2
i=1 i=1
" n
!# n
!
X θ2 X
E exp (θE(xi ) − ψ(θxi )) 6 exp E(x2i ) .
2
i=1 i=1
[Hint: The inequality 1 + x 6 exp(x) for all x ∈ R may be helpful.]

1 Pn
bθ = nθ
(b) Let µ i=1 ψ(θxi ). Use the inequalities in part (a) to show that
!
1 Xn θ2 σ2 n

P (ψ(θxi ) − θE(xi )) > t 6 2 exp −θt + ,
θ 2
i=1
√
2 log(2/δ)
for any θ, t > 0. Conclude that for δ > 0, taking θ = √
σ n
and
p
t = σ 2n log(2/δ) gives
r !
2 log(2/δ)
µθ − µ| > σ
P |b 6 δ.
n
[Hint: Use Markov’s inequality after exponentiating the appropriate quantities.]
END OF PAPER
Part III, Paper 223

S.T (Tripos)

Uploaded by

Copyright:

Available Formats

S.T (Tripos)

Uploaded by

Document Information

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

S.T (Tripos)

Uploaded by

Copyright:

Available Formats

MATHEMATICAL TRIPOS Part III

Tuesday, 15 June, 2021 12:00 pm to 2:00 pm

Before you begin please read these instructions carefully

Candidates have TWO HOURS to complete the written examination.

Attempt no more than THREE questions.

STATIONERY REQUIREMENTS SPECIAL REQUIREMENTS

You may not start to read the questions

min max b({Tn }, F ).

(b) Suppose we can construct symmetric distributions F+ , F− ∈ PK (Φ) ∩ M, centered

min max b({Tn }, F ) > b1 .

Thus, the median is minimax optimal.

Part III, Paper 223

2 Influence vs. breakdown.

(a) Compute the influence function IF (x; T, F ) when F is a differentiable, strictly

Part III, Paper 223 [TURN OVER]

3 Median as a scale M -estimator.

Part III, Paper 223

4 Not quite an M -estimator.

(a) Show that for any θ > 0, we have

[Hint: The inequality 1 + x 6 exp(x) for all x ∈ R may be helpful.]

[Hint: Use Markov’s inequality after exponentiating the appropriate quantities.]

Part III, Paper 223

You might also like

(b) Suppose we can construct symmetric distributions F+ , F− ∈ PK (Φ) ∩ M, centered