• Tidak ada hasil yang ditemukan

1.9 Supplementary Appendix

1.9.3 Auxiliary Results

1.9.3.2 Auxiliary Lemmas

Lemma 1.7 Suppose that the assumptions of Theorem 1.3 hold. The function classes,G1-G6,Gb,Gη, andGϕ as defined in (1.9.19), (1.9.20),(1.9.22), (1.9.23), (1.9.27), (1.9.28), (1.9.24), (1.9.25), and (1.9.29) are of VC type with bounded envelop.

Proof of Lemma 1.7. We first identify the sub-classes that constitute the above functional classes and show that the uniform entropy condition is satisfied for each of these sub-classes. Then, we illustrate on how we use results on the

12A triangular array of stochastic process{fni(t):i=1, ...,n,tT}is separable if, for alln1, there exists a countable setTnTsuch that

E

"

1 (

sup

t∈Tsup

s∈Tn n i=1

(fni(s)−fni(t))2>0 )#

=0.

sub-classes to show that the functional classes in the theorem is of VC type. Define

M1≡ {y7→1{y≤t}:t∈T˜},

M2≡ {x17→K((x1γd−xγd)/h)1{|x1γd−xγd| ≤h}:(x,h)∈X ×H}, M3≡ {x17→K(1)((x1γd−xγd)/h)1{|x1γd−xγd| ≤h}:(x,h)∈X ×H}, M4,1≡ {y7→∂`Gd(y,xγd):`∈ {0,1,2, ...,s},(d,x)∈ {0,1} ×X}, M4,2≡ {y7→∂`Gd,1(y,xγd):`∈ {0,1,2, ...,s},(d,x)∈ {0,1} ×X}.

M4,3≡ {∂Gd(y,xγd)/∂y|y=t:(d,t,x)∈ {0,1} ×T˜×X}.

M4,4≡ {∂Gd,1(y,xγd)/∂y|y=t:(d,t,x)∈ {0,1} ×T˜×X}.

M5≡ {∂` fd(xγd):`∈ {0,1,2, ...,s},(d,x)∈ {0,1} ×X}.

By Lemma 19.15 in Van der Vaart (1998),M1is of VC type with the constant envelop. Under Assumption 1.8.2, bothK(·)andK(1)(·)are of bounded variation, Lemma 22(i) of Nolan and Pollard (1987) implies thatM2,1andM2,2

belong to the VC class with a constant envelop. Next, since∂v`FYd,Rd|D,Xγd(y,r|d,v),`=0, ...,s, is Lipschitz continuous with respect toxγd under Assumption 1.6.2(i), Lemma 2.13 of Pakes and Pollard (1989) impliesM4,1andM4,2are of VC type with bounded envelop functions. The proof forM4,3,M4,4, andM5follows the same arguments based on the Lipschitz continuity of∂yFYd,Rd|D,Xγd(y,r|d,v)with respect toyandv, and of∂vfd,γd(v)with respect tov, as implied by Assumption 1.6.2(iv), and 1.6.1(i), respectively.

Now we are ready to show why the functional classes in the lemma are of VC type. We illustrate onG1andGη. All others follow by same lines of reasoning.

We focus onG1first. Note that the class thatg11belongs to is a product of a finite set{(r1,d1)7→r11{d1=d},d∈ {0,1}},M1, andM2, and thus, it is of VC type by Corollary A.1 in Chernozhukov et al. (2014). Since all three sub- classes have finite envelops, their product also does. Regardingg12, we first show thatMφ ≡ {y7→φθ00(sd(y,xγd)): (x,θ)∈X ×Θ}is also a VC class with bounded envelop. For anyx1,x2∈X andθ12∈Θ, we have

φ

00

θ1(sd(y,x1γd))−φθ00

2(sd(y,x2γd))

≤ φ

00

θ1(sd(y,x1γd))−φθ00

2(sd(y,x1γd)) +

φ

00

θ2(sd(y,x1γd))−φ

00

θ2(sd(y,x2γd))

≤M11−θ2|+ sup

(θ,u)∈Θ×[υo,1]

φ

000

θ(u) sup

(y,x)∈T˜×X

Gd(y,xγd)

kx1−x2k kγdk

≤M11−θ2|+M2kx1−x2k ≤√

2 max{M1,M2}

1,x01)0−(θ2,x02)0 ,

whereM1andM2are positive constants. The second inequality is due to the Lipschitz continuity condition onφθ00, and the third follows becauseφθ000and∂Gdare uniformly bounded under Assumption 1.6.4, and 1.6.2. The last one is by H¨older’s inequality. Another application of Lemma 2.13 of Pakes and Pollard (1989) yields the desired result.

LetMx={x˜7→x˜`−x`:`=2, ...,k,x∈X}. SinceX is compact,Mxis a VC class becauseN (εsupx∈X x[−1]

, Mx,L2(Q))≤C(diam(X)/ε),for a positive constantCindependent ofε. Applying Corollary A.1 in Chernozhukov et al. (2014) again on the product ofM1, M3,Mφ,Mx, and the finite set{(y1,y2,d2)7→1{d2=d,y2≤y1},d∈ {0,1}}yields that the first half of g12 belongs to a VC class. Next, by Lemma 5 of Sherman (1994), we deduce that{y17→R1{d2=d,y2≤y1∧t}hK(1)(xγd,x02γd)(x2,l−xl)dF(w2):ω∈Ω}and{w27→Rg11(w1,ω)g12(w2,y1,ω) dF(w1):ω∈Ω}are both of the VC type. Since fd(xγd)is uniformly bounded away from 0,{1/fd(xγd)2:(d,x)∈ {0,1} ×X}admits a finite envelop. Applying Corollary A.1 in Chernozhukov et al. (2014) yet again concludes the proof.

Turning toGη, we first show that for a fixedx, the setMθ≡ {1/φθ0(sTd(t,xγd,θ)):(t,θ)×T˜×Θ}belongs to the VC class. Recall thatsTd(t,xγd,θ) =φθ−1

Rt

0φθ0(sd(y,xγd))sd,1(dy,xγd) . Hence,

1/φθ0

1(sTd(t1,xγd1))−1/φθ0

2(sTd(t2,xγd2))

≤n 1/φθ0

1(sTd(t1,xγd1))−1/φθ0

2(sTd(t1,xγd2))o +n

1/φθ0

2(sTd(t1,xγd2))−1/φθ0

2(sTd(t2,xγd2))o

≡∆1+∆2.

Decomposing the first term further into,

|∆1| ≤

1/φ˙θ−1

1

Z t

1

0

φ

0

θ1(sd(y,xγd))sd,1(dy,xγd)

−1/φ˙θ−1

1

Z t

1

0

φ

0

θ2(sd(y,xγd))sd,1(dy,xγd)

+

1/φ˙θ−1

1

Z t1

0

φ

0

θ2(sd(y,xγd))sd,1(dy,xγd)

−1/φ˙θ−1

2

Z t1

0

φ

0

θ2(sd(y,xγd))sd,1(dy,xγd)

≡∆11+∆12.

For the first term, we have

|∆11| ≤(1−υo) sup

(z,θ)∈[0,yo]×Θ

φ

00

θ−1

θ (z)) φ˙θ−1(z)3

sup

(u,θ)∈[υo,1]×Θ

φ

0

θ(u)

1−θ2|=M31−θ2|.

Under Assumption 1.10.(ii),∆12≤M41−θ2|.

|∆2| ≤ sup

(z,θ)∈[0,yo]×Θ

φ¨θ−1(z) φ˙θ−1(z)3

·

Z t1 0

φ

0

θ2(sd(y,xγd))sd,1(dy,xγd)− Z t2

0

φ

0

θ2(sd(y,xγd))sd,1(dy,xγd)

≤ sup

(z,θ)∈[0,yo]×Θ

φ¨θ−1(z) φ˙θ−1(z)3

· sup

(u,θ)∈[υo,1]×Θ

φ

0

θ(u) sup

(y,x)∈T˜×X

Gd,1(y,xγd)

|t1−t2|

=M5|t1−t2|.

where inequalities hold by the mean value theorem and under Assumptions 1.6.2, and 1.6.4. Combining the bounds and applying H¨older’s inequality, we conclude by Lemma 2.13 of Pakes and Pollard (1989) thatMθ is a VC class.

Next, following similar analysis as in the previous part, we deduce from Corollary A.1 of Chernozhukov et al.

(2014) that{w17→1{d1=d}(1{y1≤y} −Gd(y,x1γd))∂yGd,1(y,xγd):(d,y,θ)∈ {0,1} ×T˜×Θ}for a givenx∈X is a VC class with a finite envelop. Applying Lemma 5 of Sherman (1994), we get{w17→R1{y1≤t}1{d1=d} · (1{y1≤y} −Gd(y,x1γd))∂yGd,1(y,xγd)dy:(d,t,θ)∈ {0,1} ×T˜×Θ}also belongs to the VC class with an envelop Fη,1=Gd,1(yo∧yc,xγd). This is due to

Z 1{y1≤t}1{d1=d}(1{y1≤y} −Gd(y,x1γd))∂yGd,1(y,xγd)dy

≤2 Z t

0yGd,1(y,xγd)dy≤Gd,1(yo∧yc,xγd).

Analogous results can be established for the other two parts ofΨd.

Combining these results with the fact that {x17→K((x1γd−xγd)/h):(t,h)∈T˜×H} is VC with an envelop C1{|x1γd−xγd| ≤h}, we deduce thatGη is of the VC type, with the envelop given by∑d=0,1Cd1{|x1γd−xγd| ≤h}

whereC0andC1are positive constants. SettingHη,d(x1γd) =Cd1{|x1γd−xγd| ≤h}concludes the proof.

Lemma 1.8 Under the assumptions of Theorem 1.3, for anyδn=Op n−1/2 ,

sup

γd−γdk≤δn

sup

(t,x)∈T˜×X

Z t 0

φ¨d,γθ

d(y,x)

γd(y,xγ˜d)−∂γd(y,xγd) sd,1(dy,xγd)

=Op

(logn)1/2n−1/2h−5/2δn

+O(δn).

Proof of Lemma 1.8.Split the term inside the norm operator into

1(t,x,θ)≡ Z t

0

φ¨d,γθ

d(y,x)

γκˆd,y(xγ˜d)

d(xγ˜d) −∂γκˆd,y(xγd) fˆd(xγd)

sd,1(dy,xγd),

2(t,x,θ)≡ − Z t

0

φ¨d,γθ

d(y,x)

(κˆd,y(xγ˜d)∂γd(xγ˜d)

d2(xγ˜d) −κˆd,y(xγd)∂γd(xγd) fˆd2(xγd)

)

sd,1(dy,xγd).

Decomposing∆1and ignoring smaller order terms gives

1(t,x,θ) =fd(xγd)−1 Z t

0

φ¨d,γθ

d(y,x)

γκˆd,y(xγ˜d)−∂γκˆd,y(xγd) sd,1(dy,xγd) + fˆd(xγ˜d)−fˆd(xγd)

fd(xγ˜d)fd(xγd) Z t

0

φ¨d,γθ

d(y,x)∂γκˆd,y(xγd)sd,1(dy,xγd) + (s.o.).

We investigate the uniform rate of the first term only. The second term exhibits the same rate and is simpler. Define

11(t,x,θ,γ˜d)≡ Z t

0

φ¨d,γθ

d(y,x)E[∂γκˆd,y(xγ˜d)−∂γκˆd,y(xγd)]sd,1(dy,xγd),

12(t,x,θ,γ˜d)≡ Z t

0

φ¨d,γθ

d(y,x){∂γκˆd,y(xγ˜d)−∂γκˆd,y(xγd)

−E[∂γκˆd,y(xγ˜d)−∂γκˆd,y(xγd)]}sd,1(dy,xγd).

By Fubini’s theorem and standard change of variables,

11(t,x,θ,γ˜d) =h−2 Z t

0

φ¨d,γθ d(y,x)·

E h

ρ1,1γ˜d(y,Xγ˜d)K(1)((Xγ˜d−xγ˜d)/h)−ρ1,1γd(y,Xγd)K(1)((Xγd−xγd)/h)i

sd,1(dy,xγd)

=h−1 Z t

0

φ¨d,γθ

d(y,x)· Z

R

K(1)(u)ρ1,1γ˜d(y,xγ˜d+uh)fd(xγ˜d+uh)du

− Z

R

K(1)(u)ρ1,1γd(y,xγd+uh)fd(xγd+uh)du

sd,1(dy,xγd)

= Z

R

uK(1)(u)du· Z t

0

φ¨d,γθ

d(y,x)·n

∂zρ1,1γ˜d(y,z)|z=x˜γdfd(xγ˜d) +ρ1,1γ˜d(y,xγ˜d)∂zfd(z)|z=x˜γd

∂zρ1,1γd(y,z)|z=xγdfd(xγd) +ρ1,1γd(y,xγd)∂zfd(z)|z=xγdo

sd,1(dy,xγd),

whereρ1,1γ is defined in (1.4.8). The second equality follows by Taylor expansion and the fact thatR[−1,1]K(1)(u)du=0.

By the Lipschitz continuity ofρ1,1γ (y,xγ),∂ρ1,1γ (y,xγ),fd(xγ), and∂fd(xγ), with respect toγas implied by Assump- tion 1.7.1, and by the fact thatkγ˜d−γdk ≤δn, we conclude that supγ

d−γdk≤δnsup(t,x,θ)∈T˜×X×Θk∆11(t,x,θ,γ˜d)k= O(δn).

The centered term∆12can be bounded using following empirical process

Gδ,n≡ {w7→gδ(w,ω,γ˜d):ω∈Ω,kγ˜d−γdk ≤δn},

wheregδ(W,ω,γ˜d)≡gδ,1(W,ω,γ˜d)−Rgδ,1(W,ω,γ˜d)dFW(W),and

gδ,1(W,ω,γ˜d)≡ Z t

0

φ¨d,γθ

d(y,x)1{D=d,Y ≤y}sd,1(dy,xγd)

·n

K(1)((Xγ˜d−xγ˜d)/h)−K(1)((Xγd−xγd)/h)o .

Applying similar lines of arguments as in Lemma 1.7, it is straightforward to show thatGδ,nis a VC type class with bounded envelop, for eachδn. From the continuous differentiability ofK(1)(·), we deduce by similar arguments to Lemma 8.4 in Maistre and Patilea (2019) that

K(1)((Xγ˜d−xγ˜d)/h)−K(1)((Xγd−xγd)/h)

≤δnh−1||K(2)((Xγd− xγd)/h)||+Cδn2h−2, for some positive constantC. Combine this fact with the uniform boundedness of ¨φd,γθ

d, andsd,1, and we find that supg

δGδE[g2

δ]is bounded from above at the rate ofO δn2h−1

. We then conclude from applying the maximal inequality in Lemma 1.5 that supγ

d−γdksup(t,x,θ)∈T˜×X×Θk∆12(t,x,θ,γ˜d)k=Op

(logn)1/2n−1/2h−5/2δn

. For∆2, we have

2(t,x,θ,γ˜d) =−fd−2(xγ˜d) Z t

0

φ¨d,γθ

d(y,x)

κˆd,y(xγ˜d)−κˆd,y(xγd) ∂γd(xγ˜d)sd,1(dy,xγd)

−fd−2(xγd) Z t

0

φ¨d,γθ

d(y,x)κd,y(xγd)

γd(xγ˜d)−∂γd(xγd) sd,1(dy,xγd) +(fd(xγ˜d) +fd(xγd))(fˆd(xγ˜d)−fˆd(xγd))

fd−2(xγ˜d)fd−2(xγd)

· Z t

0

φ¨d,γθ

d(y,x)κd,y(xγd)∂γd(xγd)sd,1(dy,xγd).

Arguing as in the case of∆1, one finds that the second term in the above display dominates the other two with a uniform rate ofOp

(logn)1/2n−1/2h−5/2δn

+O(δn). Gathering results on∆1and∆2completes our proof.

Lemma 1.9 Suppose the conditions of Theorem 1.5 hold. Then

sup

(t,θ)∈T˜×Θ

n−1/2h1/2

n i=1

gd,γd,`(Xi,x)Ψˆdd,γˆd,i,Eˆd,1,γˆd,i

(t,x,θ)−Ψd Ed,γd,i,Ed,1,γd,i

(t,x,θ)

=Op

(logn)1/2n−1/2h−(2`+1)/2 ,

wheregd,γ,`(X,x) =h−(`+1)K(`)(xγd,Xγd)/f(xγd,d), for`=0,1.

Proof of Lemma 1.9. Defineη3,1(t,xγ)≡φθ0 sTd(t,xγ,θ)

3,2(t,xγ)≡φθ00(sd(t,xγ)),η3,3(W,t,γ)≡1{D=d} · (1{Y ≤t} −Gd(t,Xγ)), η3,4(t,xγ)≡sd,1(t,xγ), η3,5(W,t,γ)≡1{D=d} R1{Y ≤t} −Gd,1(t,Xγ)

,and for `= 1, ...,5, let the estimator ofη3,`be denoted by ˆη3,`. Their definitions should be apparent. Index onθis suppressed.

From Theorem 1.3 and Lemma 1.4, we have

sup

(t,θ)∈T˜×Θ

ηˆ3,`(t,xγd)−η3,`(t,xγd) =Op

(logn)1/2n−1/2h−1/2 ,

for`=1,2,4.

Given these notations, we divideΨdinto

Ψd,1(W,t,xγ) = 1 η3,1(t,xγ)

Z t 0

η3,2(y,xγ)η3,3(W,y,xγ)η3,4(dy,xγ), Ψd,2(W,t,xγ) = −1

η3,1(t,xγ)η3,4(t,xγ)η3,5(W,t,xγ), Ψd,3(W,t,xγ) = 1

η3,1(t,xγ) Z t

0

η3,2(y,xγ)η3,5(W,y,xγ)η3,4(dy,xγ),

and thusΨd(Ed,γ,Ed,1,γ) =∑3`=1Ψd,`. We illustrate onΨd,1, since the other two terms share a similar structure. From tedious manipulation, it can be shown that ˆΨd,1(W,t,xγˆd)−Ψd,1(W,t,xγd) =∑10`=1A3,`(W,t,x), where

A3,1(W,t,x) =−ηˆ3,1(t,xγˆd)−η3,1(t,xγd) η3,1(t,xγd)ηˆ3,1(t,xγˆd)

Z t

0 η3,2(y,xγd3,3(W,y,γd3,4(dy,xγd), A3,2(W,t,x) = 1

η3,1(t,xγd) Z t

0

(ηˆ3,2(y,xγˆd)−η3,2(y,xγd))η3,3(W,y,γd3,4(dy,xγd), A3,3(W,t,x) = 1

η3,1(t,xγd) Z t

0

η3,2(y,xγd)(ηˆ3,3(W,y,γˆd)−η3,3(W,y,γd))η3,4(dy,xγd), A3,4(W,t,x) = 1

η3,1(t,xγd) Z t

0

η3,2(y,xγd3,3(W,y,γd)(ηˆ3,4(dy,xγˆd)−η3,4(dy,xγd)), A3,5(W,t,x) =−ηˆ3,1(t,xγˆd)−η3,1(t,xγd)

η3,1(t,xγd)ηˆ3,1(t,xγˆd)

· Z t

0

(ηˆ3,2(y,xγˆd)−η3,2(y,xγd))η3,3(W,y,γd3,4(dy,xγd), A3,6(W,t,x) =−ηˆ3,1(t,xγˆd)−η3,1(t,xγd)

η3,1(t,xγd)ηˆ3,1(t,xγˆd)

· Z t

0

η3,2(y,xγd)(ηˆ3,3(W,y,γˆd)−η3,3(W,y,γd))η3,4(dy,xγd), A3,7(W,t,x) =−ηˆ3,1(t,xγˆd)−η3,1(t,xγd)

η3,1(t,xγd)ηˆ3,1(t,xγˆd)

· Z t

0 η3,2(y,xγd3,3(W,y,γd)(ηˆ3,4(dy,xγˆd)−η3,4(dy,xγd)),

A3,8(W,t,x) = 1 η3,1(t,xγd)

Z t 0

(ηˆ3,2(y,xγˆd)−η3,2(y,xγd))

·(ηˆ3,3(W,y,γˆd)−η3,3(W,y,γd))η3,4(dy,xγd), A3,9(W,t,x) = 1

η3,1(t,xγd) Z t

0

η3,2(y,xγd)(ηˆ3,3(W,y,γˆd)−η3,3(W,y,γd))

·(ηˆ3,4(dy,xγˆd)−η3,4(dy,xγd)), A3,10(W,t,x) = 1

η3,1(t,xγd) Z t

0

(ηˆ3,2(y,xγˆd)−η3,2(y,xγd))

·η3,3(y,xγd)(ηˆ3,4(dy,xγˆd)−η3,4(dy,xγd)), A3,11(W,t,x) =−ηˆ3,1(t,xγˆd)−η3,1(t,xγd)

η3,1(t,xγd)ηˆ3,1(t,xγˆd)

· Z t

0

(ηˆ3,2(y,xγˆd)−η3,2(y,xγd))(ηˆ3,3(W,y,γˆd)−η3,3(W,y,γd))η3,4(dy,xγd), A3,12(W,t,x) = 1

η3,1(t,xγd) Z t

0

(ηˆ3,2(y,xγˆd)−η3,2(y,xγd))

·(ηˆ3,3(W,y,γˆd)−η3,3(W,y,γd))(ηˆ3,4(dy,xγˆd)−η3,4(dy,xγd)), A3,13(W,t,x) =−ηˆ3,1(t,xγˆd)−η3,1(t,xγd)

η3,1(t,xγd)ηˆ3,1(t,xγˆd)

· Z t

0

(ηˆ3,2(y,xγˆd)−η3,2(y,xγd))η3,3(W,y,γd)(ηˆ3,4(dy,xγˆd)−η3,4(dy,xγd)), A3,14(W,t,x) =−ηˆ3,1(t,xγˆd)−η3,1(t,xγd)

η3,1(t,xγd)ηˆ3,1(t,xγˆd)

· Z t

0

η3,2(y,xγd)(ηˆ3,3(W,y,γˆd)−η3,3(W,y,γd))(ηˆ3,4(dy,xγˆd)−η3,4(dy,xγd)), A3,15(W,t,x) =−ηˆ3,1(t,xγˆd)−η3,1(t,xγd)

η3,1(t,xγd)ηˆ3,1(t,xγˆd) · Z t

0

(ηˆ3,2(y,xγˆd)−η3,2(y,xγd))

·(ηˆ3,3(W,y,γˆd)−η3,3(W,y,γd))(ηˆ3,4(dy,xγˆd)−η3,4(dy,xγd)).

Following the same type of analysis we have used so far, namely performing Taylor expansion, integration by parts, and applying the maximal inequality from Lemma 1.5 whenever appropriate, we get

sup

(t,θ)∈T˜×Θ

n−1/2h1/2

n

i=1

gd,γd,`(Xi,x)A3,`1(Wi,t,x)

=Op

(logn)1/2n−1/2h−(2`+1)/2 ,

sup

(t,θ)∈T˜×Θ

n−1/2h1/2

n i=1

gd,γd,`(Xi,x)A3,`2(Wi,t,x)

=Op

logn·n−1h−(`+1) ,

sup

(t,θ)∈T˜×Θ

n−1/2h1/2

n i=1

gd,γd,`(Xi,x)A3,`3(Wi,t,x)

=Op

(logn)3/2n−3/2h−(2`+3)/2

,

sup

(t,θ)∈T˜×Θ

n−1/2h1/2

n i=1

gd,γd,`(Xi,x)A3,15(Wi,t,x)

=Op

(logn)2n−2h−(`+2) .

for`=0,1,`1=1,2,3,4,`2=5,6,7,8,9,10, and`3=11,12,13,14. As a result,

sup

(t,θ)∈T˜×Θ

n−1/2h1/2

n i=1

gd,γd,`(Xi,x) Ψˆd,1(Wi,t,xγˆd)−Ψd,1(Wi,t,xγd)

=Op

(logn)1/2n−1/2h−(2`+1)/2 .

Analogous results hold forΨd,2andΨd,3, concluding the proof.