Two adaptive rates of convergence in pointwise density estimation
Cristina BUTUCEA
Humboldt-Universitat zu Berlin, SFB373, Spandauer Strasse 1 D 10117 Berlin, Germany
Universite Paris VI, UPRES - A 7055 CNRS, 4, Place Jussieu, F 75005 Paris, France
(E-mail: butucea@wiwi.hu-berlin.de, butucea@ccr.jussieu.fr)
Abstract
We consider density pointwise estimation and look for best attainable asymptotic rates of convergence. The problem is adaptive, which means that the regularity pa- rameter, , describing the class of densities, varies in a setB. We shall consider, suc- cessively, two classes of densities, issued from a generalization ofL2 Sobolev classes:
W( pL) andM( pL).
Keywords: nonparametric density estimation, adaptive rates, Sobolev classes
1 Introduction
1.1 Adaptivity
We want to estimate the common probability density
f
:IR
!0 +1) ofn
independent, identically distributed random variablesX
1::: X
n, at a real pointx
0. We assume thatf
belongs to a large nonparametric class of functions,H
=H
(p L
), characterized by its smoothness (e.g., order of derivability),, a norm Lp,p >
1 and a radiusL >
0.For any estimator
f
bn off
, xedx
0 real andq >
1 we consider a sequence'
n of positive numbers and dene themaximal risk
over the classH
:R
n (f
bn'
nH
) = supf2H
'
;qnE
fhf
bn(x
0);f
(x
0) qi (1.1) whereE
f() is the expectation with respect to the distributionP
f ofX
1::: X
n, when the underlying probability density isf
.We say that
'
n is anoptimal rate of convergence
over the classH
=H
(p L
), if the maximal risk over this class stays positive, for all possible estimators, asymptotically and if there is an estimator whose maximal risk stays nite asymptotically. Minimax theory is concerned with nding the estimators attaining the optimal rates, which are given by minimizing the maximal risk, over all estimators.We are interested in adaptive estimation, which means that the regularity parameter
is supposed unknown within a given set. An estimatorf
b is calledoptimal rate adaptive
if, for the optimal rate of convergence over that class,
'
n , and a constantC >
0, we have limsupn!1
sup
2B
R
nf
bn'
nH
C <
1 (1.2) whereB
is a non-empty set of values.We shall prove here, that over two dierent classes of probability density functions, to be dened below, commonly denoted by
H
=H
(p L
), we can nd no optimal rate adaptive estimator. Similar results were obtained by Lepskii 20], Brown and Low 3] (on Holder classes of functions) and Tsybakov 27] (on L2 Sobolev classes), for the Gaussian white noise model. They are characteristic for the pointwise (not global) estimation. We shall then introduce the denition of adaptive rate of convergence, which is a modication by Tsybakov 27] of the denition of Lepskii 20] (see also Lepskii 21] and 22]). We also compute the adaptive rate over the same classes of functions as well as the corresponding rate adaptive estimators.More precisely, let us dene the considered classes of densities. At rst, we dene for integer
1,L >
0 and 1< p <
1 the class of functions inLpW
(p L
) =
f
:IR
!0 +1) :Z
IR
f
= 1Z
IR
f
( )(x
) pdx
L
p
where
f
( ), the derivative of order off
, is supposed to exist.Secondly, let us introduce for any absolutely integrable function
f
:IR
!0 +1) its Fourier transform F(f
)(x
) = RIRf
(y
)e
;ixydy
, for anyx
inIR
. We dene now for real>
1; 1p and 2p <
1 the class of absolutely integrable functions whose Fourier transforms belong to Lp andM
(p L
) =
f
:IR
!0 +1) :Z
IR
f
= 1Z
IR
jF(
f
)(x
)jpjx
jpdx
L
p:
From the results of optimal recovery of Donoho and Low 9], it is straightforward to obtain the optimal pointwise rates of convergence over these classes:
'
n (W
(p L
)) =1
n
;1=p 2( ;1=p)+1
and
'
n (M
(p L
)) =1
n
+1=p;1 2( +1=p);1
:
(1.3) In this paper, we prove that no optimal adaptive estimator can be found and we look for the adaptive rates of convergence on previously dened classesW
(p L
) andM
(p L
), for belonging to a setB
Nn to be dened for each class. We prove that the adaptive rates of convergence are within a factor logn
slower than the optimal rates.Remark 1.1
Ifp
0 is the conjugate ofp
(i.e. 1=p
+ 1=p
0 = 1), then the optimal rates of convergence (1:
3) coincide for integer , on the classesW
(p L
) andM
(p
0L
) (as well as the adaptive rates (2:
5) below). Moreover, by a result of Stein and Weiss 24] we have that for integer 1 and 1< p
2M
;p
0L
W
(p L
):
Thus, parts of our results on a scale of classes can be deduced from the results on the other scale, for certain values of the parameters. Nevertheless, our setups are considerably larger and the classes
W
andM
are not compatible except in the particular case described above. For these reasons, we prefer the above notation and give independent proofs for both setups.1.2 Previous results
The asymptotic study of minimax risks of estimators in the nonparametric framework, was developed considerably since the rst results of Stone 25] and 26], Bretagnolle et Huber 2], Ibragimov et Hasminskii 15] and 16]. Beside the density model, nonparametric regression and Gaussian white noise models were studied. Estimation was done over Holder, Sobolev or Besov classes. For an overview of the results in this area see Korostelev and Tsybakov 19] and Hardle, Kerkyacharian, Picard and Tsybakov 14].
Almost optimal rates of convergence in density pointwise estimation over Lp Sobolev classes,
W
(p L
), were obtained by Wahba 28], where technics of Farrel 11] for the proof of the lower bounds issued a rate of (1=n
)2( ;;11=s=s)+1, fors
=p
+"
," >
0 arbitrary small.Note that the optimal rate for
W
(p L
), as noted in (1:
3), is given by this expression with"
= 0.Technics of optimal recovery of Donoho 5], Donoho and Liu 8], Donoho and Low 9]
allow to compute optimal rates of convergence for dierent risks, in dierent setups. In these papers the classes
M
(p L
) and the corresponding rate in (1:
3) rst appear.Lepskii 20], Brown and Low 3] showed that for pointwise estimation on the Holder classes optimal rate adaptive estimators can not be found, both in Gaussian white noise and density models. In the Gaussian white noise model, Lepskii 20] rst considered the problem of nding the adaptive rates. He showed that a loss of logarithmic order is un- avoidable and introduced a procedure providing the adaptive estimator. For a detailed overview of adaptive rates of convergence we refer to Donoho, Johnstone, Kerkyacharian, Picard 6], Hardle, Kerkyacharian, Picard, Tsybakov 14] who give adaptive rates over Besov classes using the wavelet thresholding procedure. Lepski, Mammen and Spokoiny 23], Goldenshluger and Nemirovski 12], Juditsky 17] gave also adaptive rates of con- vergence using Lepski's scheme of adaptation. Most of these results are obtained for the Gaussian white noise model.
In density estimation, wavelet techniques were used in the minimax adaptive setup for Besov classes andLp risk, by Donoho, Johnstone, Kerkyacharian, Picard 7], Kerkyachar- ian, Picard and Tribouley 18] and Juditsky 17]. Sharp results, where the asymptotic value of the maximal risk was found explicitely, were obtained over L2 Sobolev classes in
L2 risk by Efromovich 10] and Golubev 13] and pointwise in Butucea 4].
In this paper, we are interested in adaptive rates in pointwise density estimation over
L
p Sobolev classes,
W
(p L
) andM
(p L
).2 Results
We consider adaptive density estimation problem, at xed real point
x
0, over the classesH
=H
(p L
), when belongs to the discrete setB
Nn =f1:::
Nng.Assumption (A)
The setB
Nn is such that 1< ::: <
Nn<
1, for a non-decreasing sequence of positive integersN
n. From now on, we shall consider two setups. WhenH
=W
(p L
) the setB
Nn contains positive integer values of (1 1) andp >
1, whileH
=M
(p L
) implies that can take real values, (1>
1;1=p
) andp
2. Moreover, we suppose that limn!1
Nn =1 and if n = mini=1::Nn;1j
i+1;ijwe assume that it satises
limsup
n!1
n
<
+1 (2.1)together with
lim
n!1
nlog
n
Nn2 log logn
=1:
(2.2)The following denition of an adaptive rate of convergence was introduced by Lepski, see Tsybakov 27]. The original denition of adaptive rate of convergence by Lepskii 20]
is not used here since it has a more special form.
Denition 2.1
The sequence n is anadaptive rate of convergence
over the scale of classes fH
2B
Nng, if1. There exists an estimator
f
n, independent of overB
Nn, which is calledrate adaptive estimator
, such thatlimsup
n!1
sup
2B
N
n
R
n (f
n nH
)<
1 (2.3) 2. If there exist another sequence of positive reals n and an estimatorf
n such thatlimsup
n!1
sup
2B
Nn
R
n (f
n nH
)<
1 and, at some 0 inB
Nn, n 0n 0
!
n!1
0, then there is another
0 0 inB
Nn such that
n 0
n 0
n 00
n 00
!
n!1
+1
:
Note that condition (2
:
3) introduces a wide class of rates. We choose between those rates by a criterion of uniformity over the setB
Nn, expressed in the second part of De- nition 2:
1. If some other rate satises a condition similar to (2:
3) and if this rate is faster at some point0 then the loss at some other point0 0 has to be innitely greater for large sample sizesn
.Remark 2.2
If an optimal adaptive estimator exists, it is also rate adaptive.Indeed, an optimal adaptive estimator satises (2
:
3) by denition, for the optimal rate of convergencen ='
n . We can easily verify that in this case condition 2 in Denition 2:
1 is redundant, since such a sequence n can not exist.In what follows we assign to any
inB
Nn the value e=e(H
) =;1
=p
+ 1=
2, ifH
=W
+ 1=p
;1=
2, ifH
=M
(2.4)where equalities
H
=W
andH
=M
denote the cases when we consider the scales of classesfW
(p L
) 2B
NngorfM
(p L
) 2B
Nng, respectively. We remark that in both setups:>
e 1=
2.Let us dene
B
;=B
Nn nfNng and n =n (H
) =8
<
:
(log
n=n
)e;21e=2, if 2B
;(1
=n
)e;21e=2, if =Nn:
(2.5) Then the rate n (H
) is slower than the optimal rate of convergence, except for the last point Nn. As by our hypothesis limn!1
Nn =1, this asymptotic phenomenon is not characteristic and we can use the setB
; instead ofB
Nn.2.1 The adaptive procedure
Let us proceed to the construction of the estimator
f
n called adaptive estimator. We start for each inB
Nn with the corresponding kernel estimatorf
n (x
0) = 1nh
nn
X
i=1
K
X
i;x
0h
n
:
(2.6)Here the kernel
K
is dened in the next section (dierently for each setup) and the bandwidth is in both problemsh
n =log
n n
1
2e, if
2B
; andh
n Nn =1
n
1 2eNn
where
e=e(H
) and eNn =eNn;H
Nn in (2:
4). We shall evaluate the regularity of the estimated density and replace it into the kernel estimatorf
n in order to obtainf
n, the adaptive estimator, in the spirit of Lepskii 20].More precisely, let
a >
0 be a suciently large constant and n =a
log
n n
e
;1=2 2e
:
Then, we dene b=b(H
) = maxf2B
Nn :jf
n (x
0);f
n(x
0)jn 8<
2B
Nng:
In the sequel, e (appearing inn) is dened as in (2:
4). Finally,f
n(x
0) =f
nb(x
0):
(2.7)2.2 Statement of results
Theorem 2.3
In both pointwise density estimation problems described above, we can nd no optimal rate adaptive estimators (see De nition 1:
2) over the scale of classesf
H
(p L
) 2B
g, as soon asB
has at least two dierent elements andB
B
Nn, whereB
Nn satis esAssumption (A)
.Theorem 2.4
The estimatorf
n(x
0) off
(x
0), in (2:
7), is rate adaptive estimator and n (H
) in (2:
5) is the adaptive rate of convergence in the sense of De nition 2:
1, over the scaleH
(p L
)B
, where the setB
satis esAssumption (A)
.The proof is organized as follows. In Section 3 we prove that
f
n(x
0) in (2:
7) satises, for a constantC >
0,limsup
n!1
sup
2B
Nn
R
n (f
n nH
)C <
1:
(2.8) This result will be called the upper bound. Section 4 is devoted to the proof of the lower boundliminf
n!1
inf
f
n
sup
2f g
R
n(f
n nH
)c >
0where
and are inB
Nn, arbitrary chosen elements such that<
,c >
0 and the inmum is taken over all possible estimatorsf
noff
. These relations, Theorem 2:
3 (proved in Section 5) and the fact that n (H
) is the adaptive rate of convergence over the setB
Nn (see also Section 5) imply Theorem 2:
4.3 Upper bounds
We shall prove that the estimator
f
n, independent of inB
Nn, dened in (2:
7), is such that the upper bound (2:
8) holds. Throughout this section,C
,c
i andC
i,i
= 1 2:::
, denote positive constants, depending possibly on xedq
,1 andL
.3.1 Auxiliary results
Denition 3.1
Let the densityf
belong to the classH
=H
(p L
). De ne for any kernel estimatorf
n off
(see(2:
6)), with , inB
Nn such that its bias termB
n =B
n(x
0H
) =jE
ff
n(x
0)];f
(x
0)j and its stochastic termZ
n =Z
n(x
0H
) =jf
n(x
0);E
ff
n(x
0)]j:
Besov, Il'in and Nikol'skii 1], Theorem 15.1 implies the following:Lemma 3.2
Let , be integers and 0<
, 1p
0p
1p
1,>
1=p
. If there exists 2(=
1) such thatp
10 ; = (1;) 1p
1 +1p
; (3.1) then any functionf
2L 1 (IR
) withf
( )p<
1 satis es
f
()p0
C
kf
k1p;1f
( )p
where
C
is a constant that depends only onp
0,p
1,p
, , .Lemma 3.3
There exists a nite constant depending onL
, andp
only such that supf2H( pL)k
f
k1:
Proof.
Forf
2W
(p L
), we apply the previous result with = 0,p
0 =1,p
1 = 1.Then (3
:
1) takes the form0 = (1;
) +1p
;which implies
= 1=
(+ 1;1=p
). Then 2(=
1) if>
1=p
which holds by hypoth- esis. Thus, we apply the previous result, Lemma 3:
2, and getk
f
k1C
kf
k11;f
( )p
CL
for allf
inW
(p L
).If
f
2M
(p L
), then k F(f
)k11 sincef
is a density. We havek
f
k1 21Z
IR
jF(
f
)(y
)jdy
= 12
Z
IR
jF(
f
)(y
)j1 +jy
jdy
1 +jy
j
21
Z
IR
jF(
f
)(y
)jp1 +jy
j pdy
1=p0
B
@ Z
IR
dy
1 +j
y
j p01
C
A
1=p0
where 1
=p
+ 1=p
0 = 1. This is less than a constant (L p
)>
0, forf
in the classM
(p L
). 2Lemma 3.4
Iff
2H
(p L
) and is inB
Nn such that<
, thenf
2H
(p L
0), whereL
0>
0 depends only uponL
andp
.Proof.
For classesW
(p L
) putp
0 =p
,p
1 = 1 in the auxiliary Lemma 3:
2. Then (3:
1) takes form1
p
;=1;e+e1
p
;
which gives
e= ( + 1;1=p
)=
(+ 1;1=p
) and thus e2(=
1) if>
1=p
(true, by hypothesis). By Lemma 3:
2 we get
f
()p
C
ekf
k11;ef
( )ep
CL
e efor all
f
inW
(p L
).For
f
2M
(p L
), asp >
1 and kF(f
)k11, we writeZ
IR
jF(
f
)(y
)jpjy
jpdy
Zjyj1j F(
f
)(y
)jpdy
+Z
jyj>1jF(
f
)(y
)jpjy
jpdy
1 +L
p:
Lemma 3.5
If and are inB
Nn such that and iff
belongs toH
=H
(p L
) then there existsb
(H
)>
0 (given in the proof and depending also onL
andp
), such thatB
n(x
0H
)b
(H
)h
n;1=p, ifH
=W
(p L
),B
n(x
0H
)b
(H
)h
n;1+1=p, ifH
=M
(p L
), andE
fZ
n(x
0H
)]2 kK
k22nh
n Def=s
2n. (3.2)Moreover, for the kernels f
K
2B
Nng used in the proof, we can nd constantsK
max,k
min,k
max andb
max depending possibly on xedp
and 1, such thatk
K
k1K
maxk
minkK
k2k
maxfor all
inB
Nn andb
(H
)b
max for all and inB
Nn such that .Remark 3.6
From now on, e = e(H
) is obtained as in (2:
4). Then Lemma 3:
5 says thatB
n(x
0H
)b
h
ne(H );1=2.Proof.
IfH
=W
=W
(p L
), let us introduce a kernelK
of order , in the expression of the kernel estimator (2:
6). Such a kernel must be bounded uniformly in (kK
k1K
max, for all inB
Nn), absolutely integrable, with a bounded L2 norm (k
min kK
k2k
max, for all inB
Nn), such thatRIRK
(y
)dy
= 1, RIRy
jK
(y
)dy
= 0 forj
= 1:::
;1 andZ
IR
j
K
(y
)jjy
j;1=pdy
L
0<
1 (3.3) whereL
0 depends only on xedp
and 1. It is not dicult to nd examples of such kernels. For example, the kernelK
having Fourier transform F(K
)(u
) = 1=
(1 +ju
jp) satisfy these conditions and the proofs are given later on.From now on we denoteR =RIR. Then the bias can be bounded as follows
B
n(x W
) =Z
K
(y
)f
(x
+yh
n);f
(x
)]dy
Z
K
(y
);X1j=1
(
yh
n)jj
!f
(j)(x
)dy
+Z
K
(y
)Z
x+yhn
x
(
x
+yh
n;u
);1(
;1)!f
()(u
)dudy
;1
X
j=1
h
jnj
!f
(j)(x
)Z
y
jK
(y
)dy
+ +Z
j
K
(y
)jf
()p
j
yh
nj;1=p(
;1)!((;1)p
0+ 1)1=p0dy
where the rst term is zero by hypotheses on the kernel and we applied the Holder in- equality with 1
=p
+ 1=p
0 = 1 for the second term. This givesB
n(x W
)L
0(
;1)!h
;n1=p ((;1)p
0+ 1)1=p0Z
j
K
(y
)jjy
j;1=pdy
b
(W
)h
en(W );1=2 whereb
(W
) =L
0 (;1)!R
j
K
(y
)jjy
j;1=pdy
((;1)p
0+ 1)1=p0:
We can also see that
b
(W
)b
max,b
max depending only onp
,L
and1, for all and inB
Nn, .If
H
=M
=M
(p L
), let us choose the kernelK
dened by its Fourier transform as followsF(
K
)(u
) = 1 1 +ju
jp:
This kernel has, by Plancherel's formula:k
K
k2 = 1p2k F(K
)k2 = 1p2Z
du
(1 +ju
jp)2
1
p2
Z
juj1
du
1 +j
u
jp 12 =k
min(p
1) alsok
K
k2 1 + 1p2Z
juj>1
du
1 +j
u
jp 12 =k
max(p
1) andk
K
k1 21Z
j F(
K
)(u
)jdu
1 + 12Z
juj>1
du
1 +j
u
jp 1 =K
max(p
1) sincep
1>
1, in our setting. Then the bias isB
n(x M
) = Z 1h
nK
y
;x h
n
f
(y
)dy
;f
(x
)= 12
Z
F(
f
)(y
)e
ixyF(K
)(h
ny
);1]dy
21
Z
jF(
f
)(y
)j jh
ny
jp 1 +jh
ny
jpdy:
Then we apply Holder's inequality for 1
=p
+ 1=p
0 = 1 as followsB
n(x M
)h
n2
Z
j F(
f
)(y
)jjy
j jh
ny
j(p;1) 1 +jh
ny
jpdy
L
0h
;n1=p0 2
Z
j
y
jp (1 +jy
jp)p0dy
!1=p0
=
b
(M
)h
en(M);1=2 whereL
0 is the constant from Lemma 3:
4 andb
(M
) =L
0 2
Z
j
y
jp (1 +jy
jp)p0dy
!1=p0
:
This term is bounded as followsb
(M
)L
0 20
B
@ Z
jyj1
dy
1 +j
y
jp 1p0 +Z
jyj>1
dy
j
y
jp0 11
C
A
1=p0
=
b
max(p L
1):
Let us check at last that condition (3:
3) is fullled:Z
IR
j
K
(y
)jjy
j;1=pdy
Zjyj1j
K
(y
)jdy
+Zjyj>1j
K
(y
)jjy
j;1=pdy
K
max+
Z
jyj>1j
K
(y
)j2jy
j2dy
!1=2Z
jyj>1j
y
j;2=pdy
!1=2
L
0(p
1):
For the variance term, we write, using Lemma 3
:
3E
fZ
n(x H
)]2 1nh
nZ 1
h
nK
2y
;x h
n
f
(x
)dx
kK
k22nh
n:
2
Let us recall the following inequalities (see e.g. Hardle, Kerkyacharian, Picard, Tsy- bakov 14]).
Lemma 3.7 Rosenthal's inequality
: Letq
2 andY
1::: Y
n be independent random variables such thatE Y
i] = 0,E
jY
ijq]<
1. Then there existsC
(q
) a constant depending onq
such thatE
"
n
X
i=1
Y
iq
#
C
(q
)8
<
: n
X
i=1
E
jY
ijq] +
n
X
i=1
E
Y
i2!
q=29=
:
Bernstein's inequality
: LetY
1::: Y
n be i.i.d. random variables such that jY
ijM
,E Y
i] = 0 and denoteb
2n=Pni=1E
Y
i2. Then for any>
0,P
"
n
X
i=1
Y
i#
2exp
;
2 2(b
2n+M=
3)
:
Lemma 3.8
Iff
belongs toH
=H
(p L
) and<
, ifK
is the kernel function andZ
n(x
0H
) = 1nh
nn
X
i=1
K
X
i;x
0h
n
;
E
fK
X
i;x
0h
n
then for any
u >
0P
fZ
n(x
0H
)u
]2exp
;
u
2 2s
2n(1 +c
0u
)
where
c
0>
0 does not depend on .Proof.
Indeed, we can apply Bernstein's inequality for=nu
and the i.i.d., centered variablesY
i = 1h
n
K
X
i;x
0h
n
;
E
fK
X
i;x
0h
n
bounded as follows: j
Y
ij 2kK
k1=h
n. Thenb
2ns
2n = kK
k22=
(nh
n) by (3:
2) and, by Lemma 3:
5, 2kK
k1=
kK
k222K
max=
;k
2min=c
0, which does not depend on .2
Remark 3.9
Forq >
1, we can nd a constantc
(q
)>
0 such that the stochastic term of the kernel estimator satis esE
fZ
n(x
0H
)]qc
(q
)s
qn where we denoteds
2n =kK
k22=
(nh
n).Indeed, for
q >
2, we apply Rosenthal's inequality to the previous centered variablesY
i, bounded as follows: jY
ij2kK
k1=h
n. Then we can nd a constant depending onq
,c
0(q
), such thatE
f"
n
1n
X
i=1
Y
i#
q
c
0(q
)(
2k
K
k1nh
n
q;2 1
nE
fY
21+1
nE
fY
21q=2)
and this leads to our result for some constant
c
(q
), because of the inequality (3:
2). We can easily deduce this result by standard convexity inequalities, for 1< q
2, from (3:
2).Let us introduce the sequence
n2 =C
qs
2n21e
; 21e
log
n
where<
are inB
Nn,C
>
0 ande and eare dened by (2:
4).Lemma 3.10
1. If the setB
Nn satis es conditions (2:
1) and (2:
2), then logn
eN n!1! 1, logeNs
eNlog
n
n!1! 0 and log 1ns
eNlog
n
n!1! 0 where eN =eNn is de ned by the transformation (2:
4).2. If
, are inB
Nn such that<
then there exist constantsC
1,C
2 depending only on previously xed constants such thatsup
f2H
B
nq (x
0H
) +s
qn qnC
1sup
f2H
B
nq (x
0H
) +nq qnC
2plogn
1
n
;12 12e;1 2e
: Proof.
1. The limits are easy consequences of hypotheses (2:
1) and (2:
2).2. By Lemma 3
:
5, there existb
max andk
max not depending on, such thatb
b
maxand k
K
k2k
max, for any inB
Nn. Thus, for 2B
; and<
:B
n (x
0H
) nb
max,s
n nk
maxs log
n
andB
n(x
0H
) nb
maxlog
n n
;12 12e;1 2e
:
Finally, n nqC
a
e1q
eNlog
n n
;12 12e;1 2e
:
Because
eN=
logn
!0 whenn
!1 we get the lemma for 2B
;. For the case =Nn, denoted N:B
n N(x
0H
) n Nb
max,s
n N n Nk
maxp
:
Moreover,B
n(x
0H
) n Nb
maxplogn
n
1
;1212e; 1 2eN
n n N 2
qC
a
s
eNnlog
n
1
n
;12 12e; 1 2eN
:
2