estimation of volatility for assets

1. On estimating the integrated co-volatility using
noisy high frequency data with jumps
Bing-Yi JINGa,b Cui-Xia LIa Zhi LIUc,∗
a
School of Mathematics and Statistics, Lanzhou University.
b
Department of Mathematics, Hong Kong University of Science and Technology,
c
Department of Statistics and Wangyanan Institute for Studies in Economics, Xiamen University.
Abstract: In this paper, we consider the estimation of covariation of two asset
prices which contain jumps and microstructure noise, based on high frequency data.
We propose a realized covariance estimator, which combines pre-averaging method
to remove the microstructure noise and the threshold method to reduce the jumps
effect. The asymptotic properties, such as consistency and asymptotic normality,
are investigated. The estimator allows very general structure of jumps, for example,
infinity activity or even infinity variation. Simulation is also included to illustrate the
performance of the proposed procedure.
Keywords: Ito semi-martingale; High frequency data; Microstructure noise; Co-
volatility; Jumps; Central limit theorem.
∗
The corresponding author
1

2. 1 Introduction
Suppose that we have p-multiple underlying log price processes of assets, X1t , ..., Xpt .
Denote Zt = (Z1t , ..., Zpt )τ for any generic processes Zit ’s. Then a widely used model
for Xt is the following semi-martingale:
Xt = Xc + Xd ,
t t (1.1)
where Xc and Xd are, respectively, the continuous and discontinuous components,
t t
whose forms are given in (2.3)-(2.4) later. From Protter (1990) and Jacod and
Shiryaev (2003), it is known that the covariance matrix [X, X]T has the following
decomposition:
[X, X]T = [Xc , Xc ]T + [Xd , Xd ]T .
Given discrete high frequency observations, Xti at 0 = t0 < t1 < ... < tn = T ,
both for pricing and hedging purposes and for financial econometrics applications,
it is important to separate the contributions of the diffusion part and jump part of
X (see Andersen et. al (2001); Barndorff-Nielsen and Shephard (2002b); Mancini
(2009)). Define the realized covariance matrix (RCV ) as
∑
n
RCV =: [X, X]T = ∆n X(∆n X)τ
i i
i=1
where ∆n X
i = Xti − Xti−1 is a column vector and Aτ defines the transpose of matrix
A. It is well known (see, e.g., Protter (1990)) that [X, X]T →P [X, X]T . Therefore,
it suffices to focus on the estimation of the continuous part [Xc , Xc ]T , which is the
center of our attention in the present paper.
However, it is widely accepted that the observed prices are contaminated by mi-
crostructure noise due to bid-ask spreads and/or rounding errors etc. Hence, instead
of observing Xti = (X1ti , ..., Xpti ), we observe Yti = (Y1ti , ..., Ypti ), where
Yti = Xti + ϵti , i = 0, 1, ..., n, (1.2)
2

3. where Xt are the latent semi-martingale price processes, the perturbation term ϵti =
(ϵ1ti , ..., ϵpti ) is the microstructure noise at time ti . Our objectives concern the in-
ference of [Xc , Xc ]T (and [Xd , Xd ]T ) for the latent price process X, based on the
contaminated observations Yti ’s.
The covariation matrix is of strong interest in financial applications, such as port-
folio risk and hedging of funds management; see, e.g., A¨
ıt-Sahalia, Fan, and Xiu
(2010), A¨
ıt-Sahalia, Mykland and Zhang (2011), Bai, Liu and Wong (2009), Zheng
and Li (2010) among others.
For p = 1, there has been a huge literature on the integrated volatility estimation.
In the absence of microstructure noise (ϵ = 0), we refer the reader to Andersen,
et al. (2001), Barndorff-Nielsen and Shephard (2002a, 2002b), Jacod and Protter
(1998), Mykland and Zhang (2006), Mancini (2009) among others. In the presence of
microstructure noise, the references include A¨
ıt-Sahalia, Mykland and Zhang (2005,
2011) and Bandi and Russell (2006), Zhang, Mykland, and A¨
ıt-Sahalia (2005) and
Zhang (2006), Fan and Wang (2007), Podolskij and Vetter (2009a, 2009b), Barndorff-
Nielsen et al. (2008), and Jacod, Podolskij and Vetter (2010), among others.
For p ≥ 2, there has been an increasing literature recently. For instance, Gobbi
and Mancini (2009, 2010) derived an asymptotically unbiased estimator of the con-
tinuous part of the covariation process as well as of the co-jumps, and considered
its asymptotic behavior as well. However, these works assume that there is no mi-
crostructure noise in the underlying price process Xt .
In this paper, we attempt to develop a procedure that gives a consistent estimator
of the integrated co-volatility in the simultaneous presence of microstructure noise
and jumps. The method combines the two approaches: the pre-averaging method and
threshold technique. The former is employed to reduce the effect of microstructure
noise while the latter is used to remove the jumps. The central limit theorem is also
developed. The study of the p × p covariation matrix has some distinct features,
3

4. compared with that of variation in the one-dimensional case. As will be seen later,
its studies depend on the assumptions of how the jumps of components related to
each other, and yield different estimators under different assumptions.
The remainder of this paper is organized as follows. In Section 2, we give some
assumptions on the model. We introduce the asymptotic theorems and central limit
theorem in Section 3. Simulation study is put to Section 4. Section 5 features the
conclusions and all the technical proofs are postponed to the appendix.
2 Preliminaries
2.1 Model assumptions
Since there is no essential difference between the two cases: p = 2 and p > 2, so for
simplicity, we shall assume that p = 2 from now on. In this case, [Xc , Xc ]t involves
c c c c c c
three [X1 , X1 ]t , [X2 , X2 ]t , and [X1 , X2 ]t . The first two variation processes are well
c c
studied in the literature. Below, we will focus on the covariation process [X1 , X2 ]t .
c d
We now describe model (1.2) in details. Assume that Xrt = Xrt + Xrt , r = 1, 2,
defined on a stochastic basis (Ω, F, Ft , P ), are Itˆ semi-martingales of the form
o
∫ t ∫ t
c
Xrt = Xr0 + brt dt + σrt dWrt (2.3)
∫ t∫ 0 0
∫ t∫
d
Xrt = x(µr − νr )(ds, dx) + xµr (ds, dx), (2.4)
0 |x|≤1 0 |x|>1
where b and σ are locally bounded optional processes, µr is a jump measure compen-
sated by νr , and νr has the form dtFrt (dx), where Frt (dx) is a transition measure form
∫
Ω × R+ endowed with the predictable σ-field into R/0. If |x|≤1 |x|Frt (dx) < ∞, we
∫
say that Xr has finite variation. Let βr =: inf{s : |x|≤1 |x|s Frt (dx) < ∞}, which is
d
called jump activity index in the literature. Further Wr = (Wrt ) are standard Wiener
4

5. processes such that d[W1 , W2 ]t = ρdt. The integrated covariation is defined as
∫ T
c c
[X1 , X2 ]T = ρσ1s σ2s ds. (2.5)
0
We will impose some condition on the microstructure noise. For more details, see
Jacod, Podolskij and Vetter (2010).
Assumption 1 (Microstructure noise) The microstructure noise  = (ϵ1 , ϵ2 ) are
 ϵ
2
ω1 0
iid processes with E[ϵt ] = 0 and E[ϵt ϵ′t ] = ω, where ω =   . We further
2
0 ω2
assume X ⊥ ϵ (here, ⊥ denotes stochastic independence).
2.2 Notations
• Let g be a continuous weight function on [0,1] with a piecewise Lipschitz deriva-
∫1
tive g ′ with g(0) = g(1) = 0 and 0 g 2 (s)ds > 0. Define
∫1
– g (2) =
¯ 0
g 2 (s)ds,
∫1 ∫1
– h(s) = s
g(u)g(u − s)du, Φ = 0
h2 (s)ds.
• For a generic process Z = {Zt , t ≥ 0}, the one-step increment is ∆n Z = Zti −
i
Zti−1 , and kn -step pre-averaged increment is defined as (kn is an integer)
∑
kn −1
∆n n Z
i,k = g(j/kn )∆n Z.
i+j
j=1
• Yt = Xc + ϵt , Yt = Xd = Xt − Yt . ∆n = T /n and ∆s X = Xs − Xs− .
c
t
d
t
c
• ∥x∥ represents the Euclidean distance.
5

6. 3 Main results
c c
We now consider the estimation of the integrated covariation [X1 , X2 ]T in (2.5). The
d d
estimators depend on how the jumps from the two processes, X1t and X2t , are related
to each other. We consider two separated cases below.
Case I: Two jump processes are independent
First we suppose that jump parts of two processes are independent. This is appro-
priate when the jumps are mainly due to effects of individual stocks but not the
systematic influence. In this case, we define an estimator as
∑
n−kn +1
V1n (Y) = (∆n n Y1 )(∆n n Y2 ).
i,k i,k (3.6)
i=0
√
Theorem 1 Assume that E∥ϵ∥4 < ∞. Fix θ > 0 and choose kn such that kn ∆n =
1/4
θ + O(∆n ). Then, under Assumption 1, we have
∫ T
1
∆ V1n (Y) →
1/2 P
ρσ1s σ2s ds.
θ¯(2) n
g 0
Remark 1 Although V1n (Y) contains jumps from both processes, their effects disap-
pear in the limit. It is not surprising because two independent purely discontinuous
processes never jump together almost surely. It is therefore easy to understand why the
estimator takes the same form as that by Podolskij and Vetter (2009a), who consider
a similar problem but without jumps.
Case II: Two jump processes are dependent
We now consider a more general case, which allows the two jump process to be
correlated. The two individual stocks may jump together since the common news
6

7. such as the government announcements may affect all the individual stocks. In this
case, the estimator V1n (Y) in (3.6) will not work as all the common jumps will be
included in the limit. We need to get rid of the effect of the jumps from V1n (Y).
For ease of exposition, let us take kn = θ n1/2 from now on unless otherwise stated.
After the pre-averaging procedure, the smoothed increments from the diffusion part
1/4
and the smooth noise part, are both of size ∆n , while the smoothed increments from
1/4
the jump part may still be larger than ∆n . Following the idea of Mancini (2009) or
Jacod (2008), we can propose the following threshold estimator of [X c , X c ]T :
∑
n−kn +1
V2n (Y) = (∆n n Y1 )(∆n n Y2 )1{|∆n (1) 1{|∆n (2) (3.7)
i,kn Y1 |≤un } i,kn Y2 |≤un }
i,k i,k
i=0
(r)
where, un satisfy
un /∆ϖ1 → 0, u(r) /∆ϖ2 → ∞, for some 0 ≤ ϖ1 < ϖ2 < 1/4, r = 1, 2.
(r)
n n n (3.8)
(r)
The threshold level un is chosen such that those (smoothed) increments larger than
(r)
un will be gradually excluded as n → ∞, and essentially only those increments due
to continuous part are included for the calculation of the integrated co-volatility since
we already smooth the data by pre-averaging. Hence we have the next theorem.
Theorem 2 Assume that E∥ϵ∥4 < ∞. Under Assumption 1, we have
∫ T
1
∆ V2n (Y) →
1/2 P
ρσ1s σ2s ds. (3.9)
θ¯(2) n
g 0
We now establish the central limit theorem. To do that, a structural assumption
on the volatility process σ is needed for technical reasons.
Assumption 2 The volatility process σ = {σt , t ≥ 0} satisfies the equation
∫ t ∫ t ∫ t
′ ′ ′ ′
σrt = σr0 + ars ds + σrs dWrs + νrs dBrs , (3.10)
0 0 0
where a′ , σ ′ and ν ′ are adapted c`dl`g processes, with a′ being predictable and locally
a a
bounded, and B ′ is a standard Brownian motion, independent of W , for r = 1, 2.
7

8. The concept of stable convergence is needed in the next theorem. A sequence
of random variables Xn converges stably in law to X defined on the appropriate
extension of the original probability space, written as Xn →S X, if
lim P (Xn ≤ x, A) = P (X ≤ x, A), for any A ∈ F and any x ∈ R.
n→∞
If Xn →S X, then for any F−measurable random variable σ, we have the joint
weak convergence (Xn , σ) =⇒ (X, σ). So stable convergence is slightly stronger than
convergence in law.
Theorem 3 Let X1t and X2t be given in (2.3) and (2.4). Under Assumptions 1-2,
E∥ϵ∥8 < ∞, and X1 and X2 are of finite variation, we have
d d
( ∫ T ) √ ∫ T
1 √
−1/4 S 2 θΦ
∆n ∆n V2n (Y) −
1/2
ρσ1s σ2s ds → 1 + ρ2 σ1s σ2s dWs′ ,
θ¯(2)
g 0 g (2) 0
¯
where W ′ is a standard Brownian motion which may be defined on an extension of
the original space and independent of F.
To do inference for the integrated covariance, we need to estimate the asymptotic
∫T
conditional variance 0 (1 + ρ2 )(σ1s σ2s )2 ds. Inspired by Mancini (2009), we have
∑
n−kn +1
Γ2
n = ∆n [(△n n Y1 )(△n n Y2 )]2 1{|∆n
i,k i,k (1) 1{|∆n (2)
i,kn Y1 |≤un } i,kn Y2 |≤un }
i=0
∫ T ( ∫ T ∫ T )
→P 3θ2 g (2)2
¯ (1 + ρ2 )σ1s σ2s ds + g (2)g ′ (2) ω2
2 2
¯ ¯ 2 2 2
σ1s ds + ω1 2
σ2s ds
0 0 0
¯ 2
g ′ (2) 2 2
+ 2 ω1 ω2 T.
θ
∑n
2
Remark 2 Zhang, et al. (2005) showed ωr = 1
2n
n
i=1 (△i Yr )
2
→p ωr when X has
2
continuous path. Jing, Liu and Kong (2010) showed that this is still true when X
contains jumps. Thus, we obtain a consistent estimator for integrated volatility as
∫ T
1 ¯
g ′ (2) 2
Σr = △ U (Yr , g)t − 2
1/2 n
ω T → P 2
σrs ds,
θ¯(2) n
g θ g (2) r
¯ 0
8

9. ∑n−kn +1
where, U (Yr , g)n = (∆n n Yr )2 1{|∆n
. From these and using similar
(r)
i,kn Yr |≤un }
t i=0 i,k
∫T
procedure as theorem 2, we obtain a consistent estimator of 0 (ρσ1s σ2s )2 ds:
1 ¯
g ′ (2)2 2 2 ¯
g ′ (2)
2
Cn =: Γ2 − 4 ω1 ω2 T − 2 2
(ω 2 Σ1 + ω1 Σ2 )
3θ2 g (2)2 n 3θ g (2)2
¯ ¯ 3θ g (2) 2
¯
∫ T
→P (1 + ρ2 )σ1s σ2s ds.
2 2
0
Consequently, we have a studentized version of central limit theorem.
Corollary 1 Under the same assumptions as Theorem 3, we have
−1/4 ( ∫ T )
g (2)∆n
¯ 1
√ ∆1/2 V2n (Y) − ρσ1s σ2s ds →S N (0, 1),
2 θ¯(2) n
g 0
2 θΦCn
where N (0, 1) is a standard normal variable, independent of F.
4 Simulation Study
In the simulation, we take n = 23, 400. This corresponds to the number of transac-
tions observed every second within one day (T = 1) consisting of 6.5 trading hours.
The latent values are drawn from two Ornstein-Uhlenbeck processes with drift added
by two different symmetric stable Levy processes, respectively, namely,
∫t ∫t
• X1t = 0
cos(s)ds + 0
e−2(t−s) dW1s + X1t ,
d
∫t ∫t
• X2t = 0
cos(2s)ds + 0
e−3(t−s) dW2s + X2t ,
d
where W1s and W2s are two standard Brownian motions with correlation ρ, and
d
Xit ’s are symmetric βi -stable Levy process. We consider the weight function g(x) =
min{x, 1 − x}. The parameters will be given when we report the simulation results.
The microstructure noise ϵi ’s are independent and identical distributed N (0, ω) with
(r) 1/4
ω = 0.1, and ω = 0.2 for comparison. We take the threshold un = 5∆n , r = 1, 2.
9

10. The variance of noise is chosen to match the size of integrated co-volatility. The same
procedure is repeated 5000 times and results including relative biases, standard errors
and mean square errors are displayed in several tables, and some of the corresponding
histograms and QQ plots are displayed as well, which verify our central limit theorem.
We make the following observations from the simulation results.
(1). In all cases, as n increases, all the biases, standard errors and mean square
errors tend to decrease. This is in line with our theoretical results.
(2). The larger β is, the larger the biases and stand errors are. This is because, as β
gets larger, the jumps are more frequent and jump sizes are smaller, hence it is
more difficult to distinguish the increments of diffusion and β-stable process.
(3). The estimator is insensitive (robust) to the variance of noise.
Figure 1: Histogram and QQ plot of 5000 values of the estimator, with β1 = 1, β2 = 1,
VAR(ϵ) = 0.2, ρ = 0 and n = 23400.
QQ Plot of Sample Data versus Standard Normal
300 0.8
0.6
250
0.4
Quantiles of Input Sample
200 0.2
0
150
−0.2
100 −0.4
−0.6
50
−0.8
0 −1
−1 −0.5 0 0.5 1 −4 −2 0 2 4
Standard Normal Quantiles
10

11. Table 1: Simulation result: Var(ϵ) =0.1
β1 = 1
β2 = 1
ρ = 0.5 ρ=0 ρ = −0.5
n (relative bias, s.e., mse) (relative bias, s.e., mse) (relative bias, s.e., mse)
1170 (-0.1029, 0.1350, 0.0286) (0.0307, 0.1350, 0.0190) (0.0865, 0.1560, 0.0316)
4680 (-0.0670, 0.0970, 0.0138) (-0.0062, 0.0949, 0.0089) (0.0510, 0.1008, 0.0127)
7800 (-0.0573, 0.0814, 0.0098) (-0.0034, 0.0875, 0.0076) (-0.0491, 0.0926, 0.0109)
23400 (-0.0376, 0.0674, 0.0059) (-0.0021, 0.0595, 0.0036) (-0.0466, 0.0621, 0.0063)
β1 = 1.5
β2 = 1.5
ρ = 0.5 ρ=0 ρ = −0.5
n (relative bias,s.e.,mse) (relative bias, s.e., mse) (relative bias, s.e., mse)
1170 (-0.1236, 0.1695, 0.0437) (0.0372, 0.1809, 0.0338) (0.1251, 0.1936, 0.0528)
4680 (-0.1200, 0.1049, 0.0253) (-0.0199, 0.1170, 0.0139) (0.0919, 0.1202, 0.0227)
7800 (-0.0986, 0.1038, 0.0208) (-0.0056, 0.1128, 0.0126) (-0.0917, 0.1133, 0.0211)
23400 (-0.0825, 0.0792, 0.0130) (-0.0039, 0.0749, 0.0056) (-0.0820, 0.0841, 0.0137)
β1 = 1
β2 = 1.5
ρ = 0.5 ρ=0 ρ = −0.5
n (relative bias, s.e., mse) (relative bias, s.e., mse) (relative bias, s.e., mse)
1170 (-0.1090, 0.1528, 0.0350) (0.0292, 0.1564, 0.0251) (0.1091, 0.1693, 0.0403)
4680 (-0.0958, 0.1019, 0.0195) (-0.0095, 0.1047, 0.0109) (0.0747, 0.1098, 0.0175)
7800 (-0.0785, 0.0941, 0.0149) (-0.0089, 0.1013, 0.0102) (-0.0646, 0.0989, 0.0139)
23400 (-0.0562, 0.0692, 0.0079) (-0.0080, 0.0639, 0.0041) (-0.0612, 0.0696, 0.0090)
11

12. Table 2: Simulation result: Var(ϵ) =0.2
β1 = 1
β2 = 1
ρ = 0.5 ρ=0 ρ = −0.5
n (relative bias, s.e., mse) (relative bias, s.e., mse) (relative bias, s.e., mse)
1170 (-0.0163,0.3263,0.1067) (0.0053,0.2517,0.0634) (0.0346, 0.2475, 0.0625)
4680 (0.0077,0.1933,0.0374) (-0.0034,0.1618,0.0262) (0.0126, 0.1616, 0.0263)
7800 (-0.0169,0.1640,0.0272) ( -0.0006,0.1319,0.0174) (0.0235, 0.1401, 0.0202)
23400 (-0.0097,0.1161,0.0136) (-0.0013,0.0982,0.0097) (0.0082,0.0990,0.0099)
β1 = 1.5
β2 = 1.5
ρ = 0.5 ρ=0 ρ = −0.5
n (relative bias,s.e.,mse) (relative bias, s.e., mse) (relative bias, s.e., mse)
1170 (-0.0192,0.3024,0.0918) (-0.0132,0.3078,0.0949) (0.0211,0.3297,0.1091)
4680 (-0.0053,0.2065,0.0427) (-0.0075,0.2025,0.0411) (0.0154,0.2109,0.0447)
7800 (-0.0198,0.1834,0.0340) (-0.0040,0.1763,0.0311) (0.0176,0.1813,0.0332)
23400 (-0.0135,0.1285,0.0167) (0.0035,0.1274,0.0163) (0.0092,0.1320,0.0175)
β1 = 1
β2 = 1.5
ρ = 0.5 ρ=0 ρ = −0.5
n (relative bias, s.e., mse) (relative bias, s.e., mse) (relative bias, s.e., mse)
1170 (-0.0354,0.2844,0.0822) (0.0085,0.2722,0.0742) (0.0132,0.2814,0.0793)
4680 (-0.0192,0.1881,0.0357) (-0.0099,0.1825,0.0334) (0.0236,0.1863,0.0353)
7800 (-0.0164,0.1513,0.0232) (-0.0008,0.1493,0.0223) (0.0106,0.1619,0.0263)
23400 (-0.0180,0.1133,0.0132) (0.0033,0.1115,0.0124) (0.0092,0.1112,0.0115)
12

13. Figure 2: Histogram and QQ plot of 5000 values of the estimator, with β1 = 1, β2 =
1.5, VAR(ϵ) = 0.2 and ρ = −0.5.
QQ Plot of Sample Data versus Standard Normal
250 0.8
0.6
200
Quantiles of Input Sample
0.4
150
0.2
0
100
−0.2
50
−0.4
0 −0.6
−1 −0.5 0 0.5 1 −4 −2 0 2 4
Standard Normal Quantiles
5 Conclusion
In this paper, we propose a threshold type estimator of integrated co-volatility in
the simultaneous presence of microstructure noise and jumps. The estimator can be
applied to a general semi-martingale which contains jump part, regardless of the types
of the jump and the result can be employed to deal with the statistical inference of co-
volatility. The future work includes estimation of regression and correlation between
two processes under the same settings as in this paper, and also the generalization of
proposed methodology to the non-synchronous data.
6 Appendix: Proofs
By a standard localization procedure, we can replace the local boundedness in assump-
tions by a boundedness assumption, and also assume that the process Yi , i = 1, 2,
13

14. d
and thus the jump process Xit , are bounded. That is, for all results which need the
assumption about volatility and L´vy measure, we may assume, almost surely,
e
max{|bit |, |σit |, |Xit |} ≤ C, for some constant C > 0
Proof of Theorem 1. The proof is similar to the case without jump compo-
nent. Rewrite (3.6) as
∑
n−kn +1 ∑
n−kn +1
∆1/2 V1n (Y)
n = ∆1/2
n (∆n n Y1c )(∆n n Y2c )
i,k i,k + ∆1/2
n (∆n n Y1c )(∆n n X2 )
i,k i,k
d
i=0 i=0
∑
n−kn +1 ∑
n−kn +1
+∆1/2
n (∆n n X1 )(∆n n Y2c ) + ∆1/2
i,k
d
i,k n (∆n n X1 )(∆n n X2 )
i,k
d
i,k
d
i=0 i=0
=: T1 + T2 + T3 + T4 .
Firstly, from Podolskij and Vetter (2009a), we have
∫ t
T1 → θ¯(2)
P
g ρσ1s σ2s ds.
0
Secondly, by the independence of Yic and Xj for i ̸= j, we can show easily Tj →L 0,
d 2
hence Tj →P 0 for j = 2, 3. Finally, as X1 and X2 are independent, by Proposition
d d
5.3 of Cont and Tankov (2004), there is no intersection between the supports of two
∑
jump measures. Hence T4 →P d d
s≤t C(θ, g, s)∆s X1 ∆s X2 = 0, where C(θ, g, s) is a
bounded functional depending only on g, θ, and s. ∆s Z = Zs − Zs− denotes jump
size of Z at s.
Proof of Theorem 2 When Yr , (r = 1, 2) have continuous path, the proofs
of un-truncated version have been already finished. Therefore it suffices to show that
∆1/2 (V2n (Y) − V1n (Yc )) →P 0.
n (6.11)
Rewrite the left hand side as
∆1/2 (V2n (Y) − V1n (Yc ))
n
14

15. ∑
n−kn +1
= ∆1/2 (∆n n Y1c )(∆n n Y2c )1{|∆n (1) (2) − (∆n n Y1c )(∆n n Y2c )
i,kn Y1 |≤un ,|∆i,kn Y2 |≤un }
n i,k i,k n i,k i,k
i=0
∑
n−kn +1
+∆1/2 (∆n n Y1c )(∆n n Y2d )1{|∆n (1) (2)
i,kn Y1 |≤un ,|∆i,kn Y2 |≤un }
n i,k i,k n
i=0
∑
n−kn +1
+∆1/2 (∆n n Y1d )(∆n n Y2d )1{|∆n (1) (2)
i,kn Y1 |≤un ,|∆i,kn Y2 |≤un }
n i,k i,k n
i=0
∑
n−kn +1
+∆1/2 (∆n n Y1d )(∆n n Y2c )1{|∆n (1) (2)
i,kn Y1 |≤un ,|∆i,kn Y2 |≤un }
n i,k i,k n
i=0
=: A1 + A2 + A3 + A4 .
Since the continuous and discontinuous components are independent, it is easy to
show that A2 →P 0 and A4 →P 0.
Next, we show A3 →P 0. For any arbitrarily small ϵ > 0, there exists an integer
N1 , as long as n > N1 , we have
∑
n−kn +1
|A3 | ≤ ∆1/2
n |(∆n n Y1d )(∆n n Y2d )|1{|∆n n Y1 |≤ϵ/2,|∆n n Y2 |≤ϵ/2} .
i,k i,k i,k i,k
i=0
By the Levy law for modulus of continuity of Brownian motion’s paths (see, Theorem
9.25, Karatzas and Shreve, 1999) and time changed Brownian motion (Theorems
1.9-1.10, Revuz and Yor, 2001), there exists an integer N2 , when n > N2 , we have
∆n n Xr
i,k
c
sup √ ≤ Λ(ω),
i ∆1/4 1
2 log ∆n
n
for r = 1, 2, and some constant Λ only depending on ω. On the other hand, the
central limit theorem implies that
√ 1 1 √
P (|∆n n ϵr | ≤ 1/kn log
i,k ) = P (|∆n n ϵr | ≤ ∆1/4 log
i,k n ) = 1 − o(1/ kn ).
∆n ∆n
We take N3 = max(N1 , N2 ) and let n > N3 , we get
∑
n−kn +1
|A3 | ≤ ∆1/2
n |(∆n n Y1d )(∆n n Y2d )|1{|∆n n Y1 |≤ϵ,|∆n n Y2 |≤ϵ} + oP (1)
i,k i,k i,k i,k
i=0
∑
→ P
|h(g, θ, s)(∆s Y1 )(∆s Y2 )|1{|∆s Y1 |≤ϵ,|∆s Y2 |≤ϵ} ,
s≤t
15

16. where h(g, θ, s) only depends on g, θ and s. Now, taking ϵ → 0 yields A3 →P 0.
1/2 ∑n−kn +1 i
Finally, we show that A1 =: ∆n i=0 A1 →P 0. We consider the following
disjoint cases. Note that l1 , l2 , r1 , r2 denote any positive real numbers.
(r)
• If |∆n n Yrc | ≥ un /2, for r = 1, 2, we have for some appropriate constant K,
i,k
K|∆n n Y1c |1+l1 |∆n n Y2c |1+l2
i,k i,k
|Ai |
1 ≤ (1) (2)
. (6.12)
(un )l1 (un )l2
(r)
• If |∆n n Yrc | ≤ un /2, for r = 1, 2, similarly, we have,
i,k
K|∆n n Y1c ||∆n n Y2c ||∆n n Y1d |r1 |∆n n Y2d |r2
i,k i,k i,k i,k
|Ai |
1 ≤ (1) (2)
. (6.13)
(un )r1 (un )r2
(1) (2)
• If |∆n n Y1c | ≥ un /2, |∆n n Y2c | ≤ un /2, we have,
i,k i,k
 K|∆n n Y1 |1+l1 |∆n n Y2 ||∆n n Y2 |r2
c c d


i,k i,k i,k
, if |∆i,kn Y2 | ≥ un
(2)


(1) (2)
(un )l1 (un )r2
|Ai | ≤ 0,
(2)
if |∆i,kn Y2 | ≤ un and |∆n n Y1c | ≤ un ,
(1)
1


i,k

 K|∆n n Y1 |1+l1 |∆n n Y2 |
c c
(2) (1)
i,k i,k
(1) , if |∆i,kn Y2 | ≤ un and |∆n n Y1c | ≥ un .
i,k
(un )l1
(6.14)
(1) (2)
• The case of |∆n n Y1c | ≤ un /2, |∆n n Y2c | ≥ un /2 is similar to (6.14).
i,k i,k
Next, we estimate |∆n n Y c | and |∆n n Y d |. By Holder’s and Birkholder’s inequalities,
i,k i,k
we have

 E(|∆n Y d |2 ) = ∫ (i+kn )∆n ∫ g 2 ( 1 ⌊ s−i∆n ⌋)x2 F (dx)ds ≤ K(k ∆ ),
i,kn i∆n R kn ∆n t n n
(6.15)
 E(|∆n Y c |l ) ≤ K (k ∆ )l/2 for l > 0.
i,kn l n n
Without loss of generality, we let l1 = l2 = r1 = r2 = 1. Then we deduce from above
3/2
1/2
inequalities and estimations that ∆n E(|Ai |) ≤ K(
1
∆n
(1) (2) ). In view of (3.8), we get
un un
A1 →P 0.
16

17. Proof of Theorem 3. By Theorem 4.1 of Jacod et al. (2010), it suffices to show
∑
n−kn +1
∆1/4 (V2n (Y)
n − V1n (Y )) =:
c
B1 →P 0.
i
(6.16)
i=0
d d
By assumption, X1 and X2 are of finite variation, i.e., β1 < 1 and β2 < 1. Therefore
c d
we have the following decomposition: Xr (t) = Xr (t) + Xr (t) for r = 1, 2, where,
∫t ∑
Xr (t) = brs − 0 xFrs (dx) and Xr (t) = s≤t ∆s Xr .
c d
We have the following estimation in this case:

 E(|∆n X d |) = ∫ (i+kn )∆n ∫ g( 1 ⌊ s−i∆n ⌋)|x|F (dx)ds ≤ K(k ∆ ),
i,kn i∆n R kn ∆n t n n
(6.17)
 E(|∆n Y c |l ) ≤ K (k ∆ )l/2 for l > 0.
i,kn l n n
Then we have for some constant K,
( l +l −2 (l +l )ϖ r1 +r2 (r +r )ϖ
)
1 2 + 1 2 2
2 + 4 − 1 22 2
1
E|B1 | ≤ K∆n ∆n
i 8
+ ∆n 8
=: K∆n (∆c1 + ∆c2 ).
n n
Then we can choose l1 , l2 , r1 , r2 and ϖ2 such that, c1 > 0, c2 > 0; see A¨
ıt-Sahalia and
Jacod (2010) or Jing, Liu and Kong (2010) on how to select these values. This proves
(6.16), hence Theorem 3.
Proof of Corollary 1 The corollary is a consequence of Theorem 3 and
Slutsky Theorem for stable convergence.
Acknowledgements
The authors would like to thank the Editor, an Associate Edition and two
referees for their constructive suggestions that improved this paper considerably.
JING’s research is supported by Hong Kong RGC grants HKUST6011/07P and
HKUST6015/08P, and in part by Fundamental Research Funds for the Central Uni-
versities, and Research Funds of Renmin University of China (Grant No. 10XNL007)
and by NSFC (Grant No. 71071155). LIU’s research is supported in part by Fu-
jian Key Laboratory of Statistical Sciences, China, and in part by NSFC (Grant No.
71171103).
17

18. References
[1] A¨
ıt-Sahalia, Y., Fan, J., and Xiu, D. (2010). High frequency covariance esti-
mates with noisy and asynchronous aata. Journal of the American Statistical
Association. 2010, 105, 15041517.
[2] Ait-Sahalia, Y and Jacod, J. (2010). Is Brownian motion necessary to model high
frequency data? Annals of Stastistics. 38(5), 3093-3128.
[3] A¨ıt-Sahalia, Y., Mykland, P., Zhang, L. (2005). How often to sample a
continuous-time process in the presence of market microstructure noise. Review
of Financial Study 18, 351-416.
[4] A¨ıt-Sahalia, Y., Mykland, P. A., and Zhang, L. (2011). Ultra high frequency
volatility estimation with dependent microstructure noise, Journal of Economet-
rics, 160, 190-203.
[5] Andersen, T., Bollerslev, T., Diebold, F., Labys, P. (2001). The distribution of
realized exchange rate volatility. Journal of the American Statistical Association
96, 42-55.
[6] Bai, Z., Liu, H., and Wong, W. (2009). Enhancement of the applicabiity of
Markowitz’s portfolio optimization by utilizing random matrix theory. Math. Fi-
nance, 19, 639-667.
[7] Bandi, F.M. and Russell, J.R. (2006). Separating microstructure noise from
volatility. Journal of Financial Econometrics 1, 1-48.
[8] Barndorff-Nielsen, O.E. and Shephard, N. (2002a). Econometric analysis of re-
alized volatilty and its use in estimating stochastic volatility models. Journal of
the Royal Statistical Society, B 64, 253-280.
[9] Barndorff-Nielsen, O.E. and Shephard, N. (2002b). Estimating quadratic varia-
tion using realized variance. Journal of Applied Econometrics 17, 457-477.
[10] Barndorff-Nielsen, O.E., Hansen, P.R., Lunde, A. and Shephard, N. (2008). De-
signing realised kernels to measure the ex-post variation of equity prices in the
presence of noise. Econometrica, 76(6), 1481–1536.
[11] Cont, R. and Tankov, P. (2004). Financial modelling with jump processes. Chap-
man and Hall-CRC.
[12] Daniel, R. and Marc, Y. (2001). Continuous martingale and Brownian Motion.
Springer, New York.
[13] Gobbi, F., Mancini, C. (2009). Identifying the diffusion covariation and the co-
jumps given discrete observations. Econometric theory, Forthcoming.
18

19. [14] Gobbi, F. and Mancini, C. (2010). Diffusion covariation and co-jumps in bi-
dimensional asset price processes with stochastic volatility and infinite activity
L´vy jumps. Arxiv, Jan, 2007.
e
[15] Jacod, J. (2008). Asymptotic properties of realized power variation and associated
functions of semi-martingale. Stochastic Processes and their Application 118,
517-559.
[16] Jacod, J., Protter, P. (1998). Asymptotic error distributions for the Euler method
for stochastic differential equations. Annals of Probability 26, 267-307.
[17] Jacod, J., Li, Y., Mykland, P.A., Podolskij, M. and Vetter, M.(2009). Microstruc-
ture noise in the continuous case: the pre-averaging approach. Stochastic Pro-
cesses and Their Applications 119, 2249-2276.
[18] Jacod, J., Podolskij, M. and Vetter, M. (2010). Limit theorems for moving aver-
ages of discretized processed plus noise. Annals of Statistics 38, 1478-1545.
[19] Jacod, J and Shiryayev, A. N. (2003). Limit Theorems for Stochastic Processes.
Springer, New York.
[20] Jing, B., Liu, Z., and Kong, X. (2010). Inference on integrated volatility with
jumps present in the latent price process under noisy observations. Working pa-
per.
[21] Karatzas. I. and Shreve, S. E. (1999). Brownian Motion and Stochastic Calculus.
Springer, New York.
[22] Mancini, C. (2009). Nonparametric threshold estimation for models with stochas-
tic diffusion coefficient and jumps. Scandinavian Journal of Statistics, 36(2),
270-296.
[23] Podolskij, M. and Vetter, M. (2009a). Estimation of volatility functionals in the
simultaneous presence of microstructure noise and jumps. Bernoulli 15, 634-658.
[24] Podolskij, M. and Vetter, M (2009b). Bi-power-type estimation in noisy diffusion
setting. Stochastic processes and their applications 119, 2803-2831.
[25] Protter, P. (1990). Stochastic integration and differential equations. Springer:
Berlin.
[26] Zhang, L. (2006). Efficient estimation of volatility using noisy observations.
Bernolli, 12(6) 1019-1043.
[27] Zhang, L., Mykland, P., A¨ ıt-Sahalia, Y. (2005). A tale of two time scales: Deter-
mining integrated volatility with noisy high-frequency data. Journal of the Amer-
ican Statistical Association 100, 1394-1411.
[28] Zheng, X. and Li, Y. (2010). On the estimation of integrated covariance matrices
of high dmensional diffusion processes. Manuscript.
19