Professional Documents
Culture Documents
Estimation of Integrated Functionals of A Monotone Density
Estimation of Integrated Functionals of A Monotone Density
Monotone Density
Rajarshi Mukherjee and Bodhisattva Sen∗
Department of Biostatistics
arXiv:1808.07915v1 [math.ST] 23 Aug 2018
Harvard University
655 Huntington Avenue
Boston, MA 02115
e-mail: rmukherj@hsph.harvard.edu
Department of Statistics
Columbia University
1255 Amsterdam Avenue
New York, NY 10027
e-mail: bodhi@stat.columbia.edu
Abstract: In this paper we study estimation of integrated functionals of a
monotone nonincreasing density f on [0, ∞). We find the exact asymptotic
distribution of the natural (tuning parameter-free) plug-in estimator, based
on the Grenander
√ estimator. In particular, we show that the simple plug-in
estimator is n-consistent, asymptotically normal and is semiparametric ef-
ficient. Compared to the previous results on this topic (see e.g., Nickl (2008),
Jankowski (2014), and Söhl (2015)) our results holds under minimal assump-
tions on the underlying f — we do not require f to be (i) smooth, (ii) bounded
away from 0, or (iii) compactly supported. Further, when f is the uniform dis-
tribution on a compact interval we explicitly characterize the asymptotic dis-
tribution of the plug-in estimator — which now converges at a non-standard
rate — thereby extending the results in Groeneboom and Pyke (1983) for the
case of the quadratic functional.
Keywords and phrases: efficient estimation, empirical distribution func-
tion, Grenander estimator, Hadamard directionally differentiable, least con-
cave majorant, plug-in estimator, smooth functionals.
1. Introduction
where h : R → R is a known (smooth) function. Here, and in the rest of R the paper,
we use the operator notation, i.e., for any function g : R → R, P [g] := g(x)dP (x)
denotes the expectation of g(X) under the distribution X ∼ P . Prototypical ex-
amples of such integrated functionals include the (classical) quadratic functional
Another intuitive estimator of ν(h, f ) would be n1 ni=1 h(fˆn (Xi )). It is easy to
P
show that in this case both the above plug-in estimators are exactly the same (see
Lemma 2.1).
We characterize the asymptotic distribution of ν(h, fˆn ) for estimating ν(h, f ),
when h is assumed to be smooth (see√Theorem 2.2). Specifically, we show that
the simple plug-in estimator ν(h, fˆn ) is n-consistent, asymptotically normal with
the (semiparametric) efficient variance (see Corollary 2.5). Further, we can con-
sistently estimate this limiting variance which readily yields asymptotically valid
confidence intervals for ν(h, f ). Compared to the previous results on this topic
(see e.g., Nickl (2007), Nickl (2008), Jankowski (2014), and Söhl (2015)) our re-
sults holds under minimal assumptions on the underlying density (see conditions
(A1)-(A3) in Section 2.2) — in particular, we do not assume that f is smooth,
bounded away from 0, or compactly supported; in fact, our results also hold when
f is piecewise constant.
Finally, in the special case when f is constant (i.e., f is uniform on a com-
pact interval) we explicitly characterize the asymptotic distribution of the plug-in
estimator which now converges at a non-standard rate; see Theorem 2.6. This the-
orem extends the results in Groeneboom and Pyke (1983) beyond the quadratic
functional.
The rest of the paper is organized as follows. In Section 2.1 we collect a few use-
ful notations, definitions, and discussions which will help the presentation of the
rest of the paper. Subsequently, in Section 2.2, we define the class of integrated
functionals of interest along with our main assumptions. Section 2.3 discusses the
construction of the plug-in estimator, its equivalence with other natural estimators
in this context, as well as the main result of this paper which give an explicit char-
acterization of its asymptotic distribution. Moreover, we show the semiparametric
efficiency of the simple plug-in estimator. Since the case of the uniform density on
a compact interval deserves a finer asymptotic expansion, we devote Section 2.4
to this purpose. Subsequently, in Section 3 we present some numerical results to
2.1. Preliminaries
In this subsection we introduce some notation and definitions to be used in the rest
of the paper. Throughout we let R+ denote the (compact) nonnegative real line
[0, ∞]. The underlying probability space on which all random elements are defined
is (Ω, F, P). Suppose that we have nonnegative random variables
i.i.d.
X1 , . . . , X n ∼ P
For a nonempty set T , we let `∞ (T ) denote the set of uniformly bounded, real-
valued functions on T . Of particular importance is the space `∞ (R+ ), which we
equip with the uniform metric k·k∞ and the ball σ-field; see Pollard (1984, Chapters
IV and V) for background.
Following Beare and Fang (2017), we now define the notion of the least concave
majorant (LCM) operator. Given a nonempty convex set T ⊂ R+ , the LCM over T
is the operator MT : `∞ (R+ ) → `∞ (T ) that maps each ξ ∈ `∞ (R+ ) to the function
MT ξ where
n o
∞
MT ξ(x) := inf η(x) : η ∈ ` (T ), η is concave, and ξ ≤ η on T , x ∈ T.
We write M as shorthand for MR+ and refer to M as the LCM operator. Beare
and Fang (2017, Proposition 2.1) shows that the LCM operator M is Hadamard
directionally differentiable (see e.g., Beare and Fang (2017, Definition 2.2)) at any
concave function g ∈ `∞ (R+ ) tangentially to C(R+ ) (here C(R+ ) denotes the
collection of all continuous real-valued functions on R+ vanishing at infinity). Let
us denote the Hadamard directional derivative of M at the concave function g ∈
`∞ (R+ ) by M0g .
Let F̂n to be the LCM of Fn , i.e., F̂n is the smallest concave function that sits
above Fn . Let B(·) be the Brownian bridge process on [0, 1]. By Donsker’s theorem
(see e.g.,
√ van der Vaart (1998, Theorem 19.3)) we know that the empirical process
∞
Dn := n(Fn − F ) converges in distribution to G ≡ B ◦ F in ` (R+ ), i.e.,
√ d
Dn := n(Fn − F ) → G.
2.2. Assumptions
We are interested in estimating the functional ν(h, f ) defined in (1). Many com-
monly studied functionals of interest (see e.g., Bickel and Ritov (1988), Birgé and
Massart (1995)) can indeed be expressed in this form; in particular, the monomial
functionals Z
f p (x)dx for p ≥ 2.
Of particular significance among them is the quadratic functional corresponding
to p = 2 — being related to other inferential problems in density models such
as goodness-of-fit testing and construction of confidence balls (see e.g., Giné and
Nickl (2016) for more details). We now state the assumptions needed for our main
results.
(A1) The true underlying density f : [0, ∞) → [0, ∞) is non-increasing with
f (0+) := lim f (x) < ∞.
x↓0
where P̂n denotes the probability measure associated with the Grenander estimator
(i.e., P̂n has distribution function F̂n ).
The main result of this paper gives the exact asymptotic distribution for this
plug-in estimator. However, before going into the statement and discussion of this
result, we first contrast the plug-in estimator with two other natural estimators and
argue that the special structure of the Grenander estimator implies their equiva-
lence with the simple plug-in estimator (3). Observe that another natural estimator
of ν(h, f ) ≡ P [h ◦ f ] would be
n
1X ˆ
Pn [h ◦ fˆn ] := h(fn (Xi )).
n i=1
Indeed, Lemma 2.1 below shows that this natural estimator is exactly the same as
the plug-in estimator P̂n [h ◦ fˆn ].
Lemma 2.1. For any function h : R → R, P̂n [h ◦ fˆn ] = Pn [h ◦ fˆn ].
Remark 2.1 (One-step estimator). There is however another interesting by-product
of Lemma 2.1: The plug-in estimator fˆn is also equivalent to the classically stud-
ied one-step estimator which is traditionally efficient for estimating integrated
functionals such as ν(h, f ) over smoothness classes for f ; see e.g., Bickel et al.
(1993), van der Vaart (1998), van der Vaart (2002). In particular, a first order
influence function of ν(h, f ) at P is P [h(f ) + f h0 (f )] and consequently a one-step
estimator obtained as a bias-corrected version of a f˜-based plug-in estimator (here
f˜ is any estimator of f ) is given by
h i
˜ ˜ ˜ 0 ˜ ˜ ˜ 0 ˜
ν̃ := ν(h, f ) + Pn h(f ) + f h (f ) − P̃ [h(f ) + f h (f )]
where P̃ denotes the probability measure associated with f˜. However, Lemma 2.1
implies that when f˜ = fˆn (the Grenander estimator)
h i
Pn h(f˜) + f˜h0 (f˜) − P̃ [h(f˜) + f˜h0 (f˜)] = 0 ⇒ ν̃ = ν(h, fˆn ).
Before we discuss the many implications of the above result, we provide a very
brief high-level idea of the proof of Theorem 2.2 — mainly to isolate the crucial
components which will be the focus of our simulation studies in Section 3.
Observe that, using Lemma 2.1,
√
n{ν(h, fˆn ) − ν(h, f )}
√ √
= nPn [h(fˆn )] − nP [h(f )]
√ √ √
= n(Pn − P )[h(f )] + n(Pn − P )[h(fˆn ) − h(f )] + n(P [h(fˆn )] − P [h(f )])
= T1 + T2 + T3 , (6)
where we let
√
T1 := n(Pn − P )[h(f )],
√
T2 := n(Pn − P )[h(fˆn ) − h(f )], (7)
√
T3 := n(P [h(fˆn )] − P [h(f )]).
side of (4) is indeed 0. In the next subsection we deal with this scenario (i.e.,
when f is uniform) and obtain a non-degenerate distributional limit, after proper
normalization.
Remark 2.5 (Connection to estimation in√smoothness classes). For smoothness
classes of functions it is well-known that n-consistent estimation of ν(h, f ) is
possible if f has more than 1/4 derivatives (see e.g., Birgé and Massart (1995))
and a typical construction of such an estimator proceeds via a bias-corrected one-
step estimator; see e.g., Bickel and Ritov (1988), Robins et al. (2008), Robins et al.
(2017). Since (i) monotone densities have one weak derivative, and (ii) the one-
step bias-corrected estimator
√ in this scenario is equivalent to the plug-in estimator,
one can intuitively expect n-consistent estimation of ν(h, f ) without further as-
sumptions. Indeed Theorem 2.2 makes this rigorous.
Remark 2.6 (Tuning parameter-free estimation). Finally, being based on the
Grenander estimator and the plug-in principle, the estimator ν(h, fˆn ) is completely
tuning parameter-free unlike those considered in Robins et al. (2008), Mukherjee
et al. (2016), Mukherjee et al. (2017) for the sake of adaptation over different
smoothness classes.
Although it is clear from Theorem 2.2 that the Grenander estimator based
plug-in
√ principle for estimating integrated functionals
√ of a monotone density is
n-consistent and possesses a weak limit in the n-scale, it is not immediately
obvious whether the limiting distribution is Gaussian with the efficient variance.
The rest of this subsection is devoted to making this connection. We also discuss
some further implications.
In particular, the description of the limiting behavior of the plug-in estimator
(properly normalized) in Theorem 2.2 involves the stochastic processes G and Ĝ.
where Tg,x = {x} ∪ Ug,x , and Ug,x is the union of all open intervals A ⊂ R+ such
that (i) x ∈ A, and (ii) g is affine over A.
The above proposition will help us demonstrate that indeed Y has a Gaussian
distribution with the desired efficient variance. However, to give more intuition to
the reader, we begin by discussing two simple scenarios.
Case (i): First, let F be strictly concave on R+ . From the above proposition
it follows that M0g is linear if and only if g is strictly concave on R+ (Beare and
Fang (2017)). Thus, M0F ξ = ξ for all ξ ∈ C(R+ ). Thus, if F is strictly concave
then Ĝ = G and consequently
Z
Y = G(x)d h0 (f (x))f (x) + h(f (x)) .
In this form, it is easy to see that Y is a centered Gaussian random variable. The
fact that its variance matches the efficiency bound will be demonstrated in the
proof of Corollary 2.5.
Case (ii): Next consider the other extreme case, i.e., F is piecewise affine. In
this case f is piecewise constant with jumps (say) at 0 < t1 < t2 < . . . < tk < ∞
and values v1 > . . . > vk , for some k ≥ 2 integer, i.e.,
k
X
f (x) = vi 1(ti−1 ,ti ] (x),
i=1
where
ai := {h(f (ti −)) − h(f (ti +))} and bi := {h0 (f (ti −))f (ti −) − h0 (f (ti +))f (ti +)}.
d
where = stands for equality in distribution.
Now, as an immediate corollary to Theorem 2.4 we get the following result
which expresses the right side of (4) in a more familiar form.
Corollary 2.5. Assume that conditions (A1)-(A3) hold. Then,
√
d
n ν(h, fˆn ) − ν(h, f ) → N (0, σeff
2
(f )), (8)
2 0
where, for X ∼ P , σeff (f ) := Var h (f (X))f (X) + h(f (X)) < ∞.
The proof of Corollary 2.5 invokes Theorem 2.2 followed by Theorem 2.4 and
involves some further tedious computations; see Section 4.4 for its proof.
Remark 2.7 (Asymptotic efficiency). Observe that the limiting variance in Corol-
lary 2.5 (see (8)) matches the semiparametric lower bound in this problem (Bickel
et al. (1993), van der Vaart (1998), van der Vaart (2002), Laurent (1996)). This
shows that the plug-in estimator ν(h, fˆn ) is asymptotically efficient when F is
strictly concave.
Remark 2.8 (Monomial functionals). Of special interest is the case when Rh(x) =
xp−1 , where p ≥ 2 is an integer. For example, the quadratic functional f 2 is
obtained when p = 2. In this case, Corollary 2.5 yields an asymptotically efficient
variance of
Z Z 2 !
2 p−1 2 2p−1 p
σeff (f ) = VarP pf (X) =p f (x)dx − f (x)dx , (9)
where h1 (x) = (h0 (x)x + h(x))2 and h2 (x) = h0 (x)x + h(x). As a result, if h ∈
C 3 (R+ ) then h1 , h2 ∈ C 2 (R+ ) and Corollary 2.5 implies that σ̂n2 := ν(h1 , fˆn ) −
(ν(h2 , fˆn ))2 is a consistent estimator of σeff
2
(f ). Consequently ν(h, fˆn ) ± zα/2 σ̂n
(with zα/2 being the 1 − α/2 quantile of the standard normal distribution) is an
asymptotically valid 100 × (1 − α)% confidence interval for ν(h, f ). Indeed, the
additional assumption of h ∈ C 3 (R+ ) compared to h ∈ C 2 (R+ ) in Corollary 2.5
can be reduced since we only demand consistency of σ̂n , instead of asymptotic
normality. This can be done by suitably modifying the control of T3 in Theorem
2.2. We omit the details for the sake of avoiding repetitions.
As mentioned in Remark 2.4, Theorem 2.2 does not give a non-degenerate limit
when f is the uniform distribution on the interval [0, c], for any c > 0. Indeed,
as we will see, in such a case ν(h, fˆn ) has a faster rate of convergence. In this
subsection we focus our attention to the case when f is the uniform distribution
on [0, 1]; an obvious scaling argument can then be used to generalize the result to
the case when f is uniform on [0, c], for any c > 0.
Suppose that P is the uniform distribution on [0, 1], i.e., f (x) = 1[0,1] (x), x ∈ R.
It has been shown in Groeneboom and Pyke (1983) (also see Groeneboom (1983))
that in this case
n fˆn2 (x) − 1 dx − log n d
R
√ → N (0, 1). (10)
3 log n
The above result yields the Rasymptotic distribution for the plug-in estimator of
the quadratic functional (as f 2 (x)dx = 1), properly normalized. In the following
theorem we extend the above result to general smooth functionals of the Grenander
estimator.
i.i.d.
Theorem 2.6. Suppose that h ∈ C 4 (R+ ) and X1 , . . . , Xn ∼ U [0, 1]. Then
ˆ
n ν(h, fn ) − ν(h, f ) − 21 h00 (1) + 2h0 (1) log n d
q → N (0, 1). (11)
3 00 (1) + 20 h(1) 2 log n
4
h
Remark 2.10. The proof of Theorem 2.6 above showsR that the result remains valid
for estimating a more general functional of the form g(f (x))dx, for g ∈ C 4 (R+ ),
3. Simulation study
1.4
Limit Limit
0.5
n = 1000 n = 1000
1.2
1.2
n = 10000 n = 10000
n = 50000 n = 50000
n = 200000 n = 200000
1.0
1.0
0.4
0.8
0.8
0.3
0.6
0.6
0.4
0.4
0.2
0.2
0.2
0.0
0.0
0.1
−1.0 −0.5 0.0 0.5 1.0 5000 10000 50000 2000000 −1.0 −0.5 0.0 0.5 1.0
Sample size
Fig 1: The left (right) plot shows kernel density estimates of the distribution of T1 (T3 )
as sample size n varies, along with the actual limiting normal distribution (in
red). Center: Box plots of the distribution of T2 (excluding outliers) as n varies.
We first consider the case when the true density f is strictly concave and thus,
the asymptotic distribution of ν(h, fˆn ) is given by (8) (in Corollary 2.5) where
2
σeff (f ) is given in (9) (with p = 2). Recall the decomposition given in (6). We
study each of the three terms T1 , T2 and T3 (see (7)) in this decomposition sep-
arately and illustrate their limiting behaviors. For the plots in Figure 1 we took
f to be the exponential density with parameter 1 for which ν(h, f ) = 1/2 and
2
σeff (f ) = 4/12. The plots show the limiting behaviors of T1 , T2 and T3 as the sam-
ple size increases (n = 1000, 10000, 50000, 200000); the estimated distributions are
computed from 1000 independent replications. We see that T1 , being just a simple
average (properly standardized), converges to a limiting normal distribution (in
this scenario it converges to the distribution N (0, 2/12)) by the central limit theo-
rem. The distribution of T2 , as n increases, seems to converge to 0 (in probability),
Limit Limit
n = 1000 n = 1000
n = 10000 n = 10000
0.4
0.4
n = 50000 n = 50000
n = 200000 n = 100000
0.8
0.3
0.3
0.6
0.2
0.2
0.4
0.1
0.1
0.2
0.0
0.0
0.0
Sample size
Fig 2: The same plots as in Figure 1 for the case when F is given in (13).
Let us now consider the case when f is piecewise √ constant. To fix ideas, as
in Beare and Fang (2017), let us take (with c := 1 − 1/ 2)
(
√x , x ∈ [0, c]
F (x) := 2−1√ √ (13)
2 − 2 + ( 2 − 1)x, x ∈ [c, 1].
Then, from Beare and Fang (2017, Proposition 2.1), it follows that, for any ξ ∈
C(R+ ), (
M[0,c] ξ(x), x ∈ [0, c]
M0F ξ(x) = (14)
M[c,1] ξ(x), x ∈ [c, 1].
Corollary 2.5 gives the asymptotic distribution of ν(h, fˆn ) in this case as well. Fig-
ure 2 shows the corresponding plots (the limiting behaviors of T1 , T2 and T3 as n in-
2
creases) when F is defined in (13) for which ν(h, f ) ≈ 1.828 and σPWA (f ) ≈ 3.314.
We see that T1 and T3 are virtually indistinguishable in this scenario and both
converge to a limiting normal distribution. In this case, the normal approximation
seems quite good even for moderately small sample sizes. The distribution of T2 ,
as n increases, seems to converge to 0 (in probability).
Let X(1) < X(2) < . . . < X(n) be the ordered data points and let
be the fitted values of the Grenander estimator at the ordered data points. A simple
characterization of θ̂ := (θ̂1 , . . . , θ̂n ) is given by
n 2
X 1
θ̂ = arg min − θi (X(i) − X(i−1) ),
θ∈C i=1
n(X(i) − X(i−1) )
where
C = {θ ∈ Rn : θ1 ≥ · · · ≥ θn },
and X(0) ≡ 0; see e.g., Groeneboom and Jongbloed (2014, Exercise 2.5). From the
characterization of projection onto the closed convex cone C it follows that for any
function h : R → R (see e.g., Robertson et al. (1988, Theorem 1.3.6))
n
X 1
− θ̂i (X(i) − X(i−1) )h(θ̂i ) = 0
i=1
n(X(i) − X(i−1) )
n n
1X X
⇔ h(θ̂i ) = h(θ̂i )θ̂i (X(i) − X(i−1) )
n i=1 i=1
4.2.1. Analysis of T2 :
First we claim that one can choose C, M > 0 large enough (depending on ) such
that for any n large enough (depending on )
To this end note that, by Woodroofe and Sun (1993, Theorem 2 and Remark 3),
ˆ (0+)
the sequence of random variables ffn(0+) is asymptotically P -tight and consequently,
∗
given any > 0 there exists M > 0 (depending on > 0) such that for n large
enough (depending on > 0)
P kfˆn k∞ > M ∗ f (0+) ≤ . (17)
4
Moreover, kfˆn k∞ ≤ M ∗ f (0+) implies that
2
√ ∗ 1
Z q
ˆ ˆ
2 2
p
kfn − f k2 ≤ 2( M + 1) f (0+) fn (x) − f (x) dx
2
√
= 2( M ∗ + 1)2 f (0+)h2 P̂n , P ,
where Z q 2
1
fˆn (x) −
2
p
h P̂n , P := f (x) dx
2
is the Hellinger divergence between P̂n and P . As assumptions (A1)-(A2) are sat-
isfied, by van de Geer (2000, Theorem 7.12), there exists a c > 0 such that for all
C 0 ≥ c,
C 02 n1/3
0 −1/3
P h P̂n , P > C n ≤ c exp − .
c2
Consequently, for an appropriate C ∗ (depending on M ∗ ) and c̃, we have
P kfˆn − f k2 > C ∗ n−1/3 , kfˆn k∞ ≤ M ∗ f (0+) ≤ c exp −c̃n1/3 ≤ , (18)
4
imsart-generic ver. 2014/10/16 file: Draft_5.tex date: August 27, 2018
R. Mukherjee and B. Sen/Estimation of Functionals of a Monotone Density 17
for n large enough (depending on ). Now, (17) and (18) together imply (16) with
(C, M ) = (C ∗ , M ∗ ). The rest of the proof works with this choice of (C, M ). Let
δn := Cn−1/3 ,
and let Pmon denote the set of all nonincreasing densities on R+ . Subsequently,
note that
E kGn kFn (δn )
P |Gn (h(fˆn ) − h(f ))| > , E(C, M ) ≤ P kGn kFn (δn ) > ≤ ,
where
Fn (δn ) := {h(g) − h(f ) : g ∈ Pmon , kg − f k2 ≤ δn , kgk∞ ≤ M f (0+)} ,
and
kGn kFn (δn ) := sup |Gn (g)|.
g∈Fn (δn )
Note that here we do not mention measurability issues since this can be taken
care of in a standard way by working with the outer expectation instead (see e.g.
van der Vaart and Wellner (1996)) Thus, it is enough to show that
E kGn kFn (δn ) → 0 as n → ∞.
Now, note that for any g̃ := h(g) − h(f ) ∈ Fn (δn ), using the fact that h ∈ C 1 (R),
kg̃k∞ ≤ (M + 1)f (0+) sup |h0 (x)| = M 0 (say).
x∈[0,(M +1)f (0+)]
Moreover,
!2
2
|h0 (x)| δn2 = δ̃n2
P g̃ ≤ sup (say).
x∈[0,(M +1)f (0+)]
where
˜ ˜
J[ ] ≡ J[ ] δ̃n , Fn (δn ), L2 (P )
and for any class of functions F ⊆ L2 (P ), and δ > 0
Z δq
˜
J[ ] (δ, F, L2 (P )) = 1 + log N[ ] (u, F, L2 (P ))du,
0
We first finish the proof of Lemma 4.1 assuming the validity of Lemma 4.2.
Plugging in the estimate from Lemma 4.2 into (19) we get
q p ! q !
B δ̃n 0 B
E kGn kFn (δn ) ≤ B δ̃n 1 + √ M = B δ̃n 1 + 3/2 √ M 0
δ̃n2 n δ̃n n
q
≤ B 0 δ̃n ,
where the last inequality follows since δ̃n = O(n−1/3 ) and B 0 depends on M 0 , C, M, f (0+)
and supx∈[0,(M ∗ +1)f (0+)] |h0 (x)| which completes the proof of Lemma 4.1.
Note that h is uniformly Lipschitz on the interval [0, (M + 1)f (0+)]. If h were
monotone, the proof would immediately follow from the fact the ε-bracketing en-
tropy of all uniformly bounded monotone functions is of order 1/ε (which would
contain g ∈ Pmon , kgk∞ ≤ M f (0+)); see e.g., van de Geer (2000, Equation (2.5)).
To handle the case of a general h ∈ C 2 (R+ ) ⊂ C 1 (R+ ), we write h = h+ − h−
where h+ and h− are monotone nonincreasing (and Lipschitz) functions. Hence,
one can provide obvious modifications to obtain a bracketing number for Fn (δn )
which is similar in order to the case when h in monotone. This completes the proof
of the lemma.
where
Z
wn := (fˆn (x) − f (x))2 h00 (ξn (x))f (x)dx,
with |ξn (x) − f (x)| ≤ |fˆn (x) − f (x)|. In order to control wn , fix an > 0 and recall
the choice of (C, M ) which guarantees the validity of (16), (17), and (18). Thus,
with an appropriate choice of (C, M ), we have
P (|wn | > )
and
Z
(Pn − P )[h(f )] = − (Fn − F )(x)d(h(f (x))).
as n → ∞, where the last step follows from a version of the continuous mapping
d
theorem: Observe that (i) (Dn , D̂n ) → (G, Ĝ) in the product space `∞ (R+ ) ×
`∞ (R+ ); (ii) for any g1 , g2 : R+ → R uniformly bounded,
Z Z
(g1 , g2 ) 7→ g1 (x)d(h(f (x))) + g2 (x)d(h0 (f (x))f (x))
is a continuous function. This gives the convergence of (20) to (21) and completes
the proof of the theorem.
Let g(t) = t · h0 (t), for t ∈ R. Let us assume that the function g ◦ f is monotone
(otherwise we can split g ◦ f as the difference of two monotone functions and apply
the following to each part). Note that it is enough to show that for any ξ ∈ C(R+ ),
Z Z
0
MF ξ(x)d[g(f (x))] = ξ(x)d[g(f (x))].
It is easy to show that I is a Borel measurable set. Consider the collection {TF,x :
x ∈ I}, where TF,x is defined in Proposition 2.3 (also see Beare and Fang (2017,
Proposition 2.1)). First note that x ∈ TF,x , for every x ∈ I. We now claim that there
are at most countably many such distinct TF,x ’s, as x varies in I. This follows from
the fact that for any x in the support of F , F has a strictly positive slope on TF,x
(i.e., F cannot be affine with slope 0 on TF,x , as F is concave and nondecreasing)
which implies that we can associate a unique rational number in the range of F
to this interval TF,x . Let {TF,xi }i≥1 (for xi ∈ (0, ∞)) be an enumeration of this
countable collection {TF,x : x ∈ I}. Obviously, I ⊂ ∪∞ i=1 TF,xi .
Let ξ ∈ C(R+ ). Then,
Z Z
0
MF ξ(x)d[g(f (x))] − ξ(x)d[g(f (x))]
Z Z
≤ |M0F ξ(x) − ξ(x)|d[g(f (x))] + |M0F ξ(x) − ξ(x)|d[g(f (x))]
I [0,∞)\I
∞
XZ
≤ |MTF,xi ξ(x) − ξ(x)|d[g(f (x))]
i=1 TF,xi
where the last inequality follows from the fact that for x ∈ [0, ∞) \ I, M0F ξ(x) =
ξ(x) (see Beare and Fang (2017, Remark 2.2)) and the above integrals are viewed
as a Lebesgue-Steiljes. Notice now that on the open interval TF,xi , f is constant
(as F is affine on TF,xi ). Hence each of the integrals in the last display equals 0.
This completes the proof.
Recall the definition of Y in (4). Furthermore, note that by Beare and Fang (2017),
when F is strictly concave, we have Ĝ = G. Consequently,
Z
Y = − G(x)d(g(x)),
where
g(x) = h0 (f (x))f (x) + h(f (x)), x ∈ R.
It is easy to see that the limiting Y is a centered Gaussian random variable and
2
therefore we only show that Var(Y ) matches σeff . Further, we only prove the result
for the case when g(∞) := limt→∞ g(t) = 0, i.e., h(0) = 0. The other cases can be
derived by working with h∗ (x) := h(x) − h(0) instead of h(x).
d
where W is the standard Brownian motion on [0, 1] and = stands for equality in
distribution. Therefore,
Z Z
Var(Y ) = Var W ◦ F (x)dg(x) + Var W(1) F (x)dg(x)
Z Z
− 2 Cov W ◦ F (x)dg(x), W(1) F (x)dg(x)
Z 2 Z 2
=E W ◦ F (x)dg(x) + F (x)dg(x)
Z Z
− 2 F (x)dg(x)E W(1) W ◦ F (x)dg(x) .
and consequently it is enough to prove the next lemma to complete the proof of
Corollary 2.5.
Lemma 4.3. The following identities hold under the assumptions of Corollary 2.5:
Z Z
E W(1) W ◦ F (x)dg(x) = F (x)dg(x), and
Z 2 Z
E W ◦ F (x)dg(x) = g 2 (x)dF (x).
Proof. The first identity is obvious since under our assumptions one can inter-
change the integral and expectation and observe that
E(W(1)W(F (x))) = Cov (W(1), W(F (x))) = min{1, F (x)} = F (x).
For the second identity, note that
Z 2 Z Z
0 0
E W ◦ F (x)dg(x) = E W ◦ F (x)W ◦ F (x )dg(x)dg(x )
Z ∞ Z x0
=2 F (x)dg(x)dg(x0 ). (22)
0 0
R x0
Now, let v(x0 ) := 0
F (x)dg(x), for x0 ∈ R+ . Then by repeated integration by
parts
Z ∞ Z x0
F (x)dg(x)dg(x0 )
0 0
Note that for some ξn (x) lying between f (x) ≡ 1 and fˆn (x) one has by a four term
Taylor expansion (allowed by the assumption that h ∈ C 4 (R+ ))
n ν(h, fˆn ) − ν(h, f ) − 12 (h00 (1) + 2h0 (1)) log n
q
3
4
(h00 (1) + 2h0 (1))2 log n
n [g(fˆn (x)) − g(1)]dx − 12 g 00 (1) log n
R
= q
3[ 21 g 00 (1)]2 log n
n (fˆn (x) − 1)2 dx − log n
R
= √ (24)
3 log n
1 n g (3) (1)(fˆn (x) − 1)3 dx 1 n g (4) (ξn (x))(fˆn (x) − 1)4 dx
R R
+ q + q .
6 1 00 2
3[ 2 g (1)] log n 24 1 00 2
3[ 2 g (1)] log n
since kfˆn k∞ = OP (1) by Woodroofe and Sun (1993) and g (4) is bounded on com-
pact intervals. Therefore, modulo the claim above, one readily obtains the desired
result (11) from (24) and (10).
n (fˆn (x)−1)k dx
R
The next two subsections are devoted towards understanding √
log n
for
k = 3, 4. In the analysis, we crucially use Groeneboom and Pyke (1983, Theorem
2.1) and consequently need the following notation to proceed. Let
n Nj
n X
X X
Tn := jNj , Sn := Sji ,
j=1 j=1 i=1
where {Sji : i ≥ 1, j ≥ 0} are independent random variables with Sji ∼ Gamma(j, 1),
for j ≥ 1, and S0i ∼ Gamma(1, 1), and Nj ∼ Poisson(1/j) for j ≥ 1 and N0 = 1
a.s.; further, {Nj } and {Sji } are independent. Also let,
n Nj
n X
1X −1/2
X
Wn = jNj , Vn = n (Sji − j).
n j=1 j=1 i=1
n (fˆn (x)−1)k dx
R
With this notation we are ready to analyze √
log n
for k = 3, 4.
n (fˆn (x)−1)3 dx
R
4.5.1. Proof of √
log n
= oP (1)
By an argument similar to that in Groeneboom and Pyke (1983, Section 3), it can
be seen that Ln has the same distribution as that of
n Nj n Nj
1 X 3X 1 3 X 2X 1
L∗n := j − j + 2 {Tn = n, Sn = n}.
2
n j=1 i=1 Sji n j=1 i=1 Sji
The following lemma, crucial to our proof, is the analogue of Groeneboom and
Pyke (1983, Lemma 3.1).
√ d
Lemma 4.4. Let cn = log n, γn = 0 and c0 = −1. Then (Un , Vn , Wn ) →
(U, V, W ) where U = δ0 , the dirac measure at 0, and (V, W ) has the infinitely
R1 2 2
divisible characteristic function φV,W (s, t) = exp 0 (eitu−s u /2 − 1)u−1 du .
Proof. With the choice of c0 , cn , γn provided in the statement of the theorem,
P
denote Un ≡ Un (c0 , cn , γn ). We only prove the fact that Un → 0. The subsequent
conclusion follows verbatim from the proof of Groeneboom and Pyke (1983, Lemma
Γ(j−k)
−k
3.1). In the following analysis we repeatedly use the fact that E Sji = Γ(j) for
j > k.
First, define Un∗ to be Un without its first 3 summands in j, i.e.,
Un∗
n Nj n Nj X n X Nj
1 X X 1 1 X X 1 1
:= j3 2
− 2 −3 j2 − − (Sji − j) .
cn j=4 i=1 Sji j j=4 i=1
Sji j j=4 i=1
P P
We will show that Un∗ → 0 and as Un − Un∗ → 0 (since cn → ∞), this will establish
the desired result. Note that,
n n
∗
X
3 Γ(j − 2) 1 X
2 Γ(j − 1) 1
cn E (Un ) = j E(Nj ) − 2 −3 j E(Nj ) −
j=4
Γ(j) j j=4
Γ(j) j
n
X 4
= .
j=4
(j − 1)(j − 2)
and consequently, lim supn→∞ Var(E(cn Un∗ |N1 , . . . , Nn )) < ∞. Next, note that
n
∗
X 1 1
Var(cn Un |N1 , . . . , Nn ) = Nj j 6 Var( 2 ) + 9j 4 Var( ) + Var(Sj1 )
j=4
Sj1 Sj1
5 1 1 3 1 1
−6j Cov( 2 , ) − 2j Cov( 2 , Sj1 ) + 6Cov( , Sj1 )
Sj1 Sj1 Sj1 Sj1
n
X 15j 5 + 129j 4 − 404j 3 + 116j 2 + 48j
= Nj 2 (j − 2)2 (j − 3)(j − 4)
.
j=4
(j − 1)
Consequently,
lim sup E(Var(cn Un∗ |N1 , . . . , Nn ))
n→∞
n
X 15j 4 + 129j 3 − 404j 2 + 116j + 48
= lim sup < ∞.
n→∞
j=4
(j − 1)2 (j − 2)2 (j − 3)(j − 4)
n (fˆn (x)−1)4 dx
R
4.5.2. Proof of √
log n
= oP (1)
(fˆ (x)−1)3 dx
R
n
The proof technique is similar to that of the case √
log n
albeit with much
more cumbersome details and a different choice of c0 . As before, let us start by
considering the distribution of
Z Z Z Z
4 3 2
Ln := (fˆn (x) − 1) dx = fˆn (x)dx − 4 fˆn (x)dx + 6 fˆn (x)dx − 3.
4
Once again we have a crucial lemma which is the analogue of (Groeneboom and
Pyke, 1983, Lemma 3.1).
√ d
Lemma 4.5. Let cn = log n, γn = 0 and c0 = 1. Then (Un , Vn , Wn ) → (U, V, W )
where
RU = δ0 and (V, W ) has the infinitely divisible characteristic function φV,W (s, t) =
1 itu−s2 u2 /2
exp 0 (e − 1)u−1 du .
Proof. With the above choice of c0 , cn , γn , denote by Un ≡ Un (c0 , cn , γn ). We only
P
prove the fact that Un → 0. The subsequent conclusion follows verbatim from the
proof of Groeneboom and Pyke (1983, Lemma 3.1), as in the proof of Lemma 4.4.
First, define Un∗ to be Un without it’s first 4 summands in j, i.e.,
n Nj n Nj
∗ −1
X
4
X 1 1 X
3
X 1 1
Un =cn j 3
− 3 −4 j 2
− 2
j=5 i=1
S ji j j=5 i=1
S ji j
n Nj X n X Nj
X X 1 1
+6 j2 − + (Sji − j) .
j=5 i=1
S ji j j=5 i=1
P P
We show that Un∗ → 0 and indeed by Un − Un∗ → 0 (since cn → ∞) we establish
the desired claim. Note that,
n n
∗
X
4 Γ(j − 3) 1 X
3 Γ(j − 2) 1
cn E (Un ) = j E(Nj ) − 3 −4 j E(Nj ) − 2
j=5
Γ(j) j j=5
Γ(j) j
n
X Γ(j − 1) 1
+6 j 2 E(Nj ) −
j=5
Γ(j) j
n
X 2(j − 9)
= .
j=5
(j − 1)(j − 2)(j − 3)
and consequently, lim supn→∞ Var(E(cn Un∗ |N1 , . . . , Nn )) < ∞. Next note that
1 1 1 1 1
− 8j 7 Cov(
3
, 2 ) + 12j 6 Cov( 3 , ) + 2j 4 Cov( 3 , Sj1 )
Sj1 Sj1 Sj1 Sj1 Sj1
5 1 1 3 1 2 1
− 48j Cov( 2 , ) − 8j Cov( 2 , Sj1 ) + 12j Cov( , Sj1 )
Sj1 Sj1 Sj1 Sj1
n
X 96j 7 + 4297j 6 − 14259j 5 − 28198j 4 + 102180j 3 − 33336j 2 − 4320j
= Nj .
j=5
(j − 1)2 (j − 2)2 (j − 3)2 (j − 4)(j − 5)(j − 6)
Consequently,
(fˆn (x)−1)4 dx
R
P
for our choice of γn = 0. Consequently, √
log n
→ 0.
Beare, B. K. and Z. Fang (2017). Weak convergence of the least concave majorant
of estimators for a concave distribution function. Electron. J. Stat. 11 (2), 3841–
3870.
Bickel, P. J., C. A. Klaassen, P. J. Bickel, Y. Ritov, J. Klaassen, J. A. Wellner, and
Y. Ritov (1993). Efficient and adaptive estimation for semiparametric models.
Johns Hopkins University Press Baltimore.
Bickel, P. J. and Y. Ritov (1988). Estimating integrated squared density deriva-
tives: sharp best order of convergence estimates. Sankhyā Ser. A 50 (3), 381–393.
Bickel, P. J. and Y. Ritov (2003). Nonparametric estimators which can be
“plugged-in”. Ann. Statist. 31 (4), 1033–1053.
Birgé, L. (1987). On the risk of histograms for estimating decreasing densities.
Ann. Statist. 15 (3), 1013–1022.
Birgé, L. and P. Massart (1995). Estimation of integral functionals of a density.
Ann. Statist. 23 (1), 11–29.
Cai, T. T. and M. G. Low (2003). A note on nonparametric estimation of linear
functionals. Ann. Statist. 31 (4), 1140–1153.
Cai, T. T. and M. G. Low (2004). Minimax estimation of linear functionals over
nonconvex parameter spaces. Ann. Statist. 32 (2), 552–576.
Cai, T. T. and M. G. Low (2005). Nonquadratic estimators of a quadratic func-
tional. Ann. Statist. 33 (6), 2930–2956.
Donoho, D. L., R. C. Liu, and B. MacGibbon (1990). Minimax risk over hyper-
rectangles, and implications. Ann. Statist. 18 (3), 1416–1437.
Donoho, D. L. and M. Nussbaum (1990). Minimax quadratic estimation of a
quadratic functional. Journal of Complexity 6 (3), 290–323.
Dümbgen, L. and K. Rufibach (2009). Maximum likelihood estimation of a log-
concave density and its distribution function: basic properties and uniform con-
sistency. Bernoulli 15 (1), 40–68.
Fan, J. (1991). On the estimation of quadratic functionals. Ann. Statist. 19 (3),
1273–1294.
Giné, E. and R. Nickl (2016). Mathematical foundations of infinite-dimensional
statistical models. Cambridge Series in Statistical and Probabilistic Mathemat-
ics, [40]. Cambridge University Press, New York.
Grenander, U. (1956). On the theory of mortality measurement. II. Skand. Aktu-
arietidskr. 39, 125–153 (1957).
Groeneboom, P. (1983). The concave majorant of Brownian motion. Ann.
Probab. 11 (4), 1016–1027.
Groeneboom, P. (1985). Estimating a monotone density. In Proceedings of the
Berkeley conference in honor of Jerzy Neyman and Jack Kiefer, Vol. II (Berke-
ley, Calif., 1983), Wadsworth Statist./Probab. Ser., Belmont, CA, pp. 539–555.
Wadsworth.
Groeneboom, P. and G. Jongbloed (2014). Nonparametric estimation under shape