Abstract
Asymptotic expansion is presented for an estimator of the Hurst coefficient of a fractional Brownian motion. We first derive the expansion formula of the principal term of the error of the estimator using a recently developed theory of asymptotic expansion of the distribution of Wiener functionals, and utilize the perturbation method on the obtained formula in order to calculate the expansion of the estimator. We also discuss some second-order modifications of the estimator. Numerical results show that the asymptotic expansion attains higher accuracy than the normal approximation.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Let \(B=(B_t)_{t\in [0,T]}\) be a fractional Brownian motion with Hurst coefficient \(H\in (0,1)\). For a fixed positive number T, let \({t_j}=t_j^n=jT/n\) for \(n\in {{\mathbb {N}}}\) and \(j\in \{0,1,...,n\}\). We are interested in estimation of the Hurst parameter H from the sampled data \((B_{t_j})_{j=0,1,...,n}\). The second-order difference \(d_{n,j} \) of B is denoted by
Note that \(t_j\) depends on n. We make the sum of squares \(V_{n,T}^{(2)}\) of \(d_{n,j}\) by
In Kubilius et al. (2018) this estimator is denoted by \(V_{n,T}^{(2)B}\). For estimation of H, Benassi et al. (1998) and Istas and Lang (1997) introduced the estimator
The estimator \(\widehat{H}_n^{(2)}\) is preferable since it is consistent and asymptotically normal as \(n\rightarrow \infty \). On the other hand, it is known that normal approximation to the distribution of a statistic is not necessarily satisfactory. Section 4 shows that the error of the normal approximation to the histogram of the estimator is not negligible numerically. In this paper, we will consider a higher-order approximation of the distribution of the error of \(\widehat{H}_n^{(2)}\) using the asymptotic expansion.
Asymptotic expansion is a standard concept in statistics and well developed for independent models. As for basic literature, we refer the reader to Cramér (1928), Cramér (2004), Gnedenko and Kolmogorov (1954), Bhattacharya (1971), Petrov (1975), Bhattacharya and Ranga Rao (1976), and to a recent textbook by Bhattacharya et al. (2016). The theory of asymptotic expansion has been extended to dependent models such as mixing Markov processes and martingales, and an amount of literature is available today. See Yoshida (2016) and references therein.
The theory of asymptotic expansion for Wiener functionals has been developed recently. In the central limit case, Tudor and Yoshida (2019a) derived the first-order expansion for vector-valued sequences of random variables, and Tudor and Yoshida (2019b) did the arbitrary order of asymptotic expansion for Wiener functionals. The latter has been applied to asymptotic expansion of the quadratic variation of a mixed fractional Brownian motion by Tudor and Yoshida (2020). In the mixed Gaussian limit case, Nualart and Yoshida (2019) presented asymptotic expansion of Skorohod integrals. This method is used in Yoshida (2020) for Brownian functionals having anticipative weights. Yamagishi and Yoshida (2022) provided order estimates of functionals related to fractional Brownian motion, with weighted graphs, and asymptotic expansion of the quadratic variation of fractional diffusions. In this paper, we will apply the scheme by Tudor and Yoshida (2019a) combined with a perturbation method.
The organization of this paper is as follows. Section 2 gives the main results, and Sect. 3 proves them. The real performance of asymptotic expansion is numerically studied in Sect. 4, and we verify the improvement produced by the correction term of the asymptotic expansion formula of \(\widehat{H}_n^{(2)}\). The technical part of this paper is Sect. 3. The technicalities are outlined at the beginning of Sect. 3.
2 Asymptotic expansion for \(\widehat{H}_n^{(2)}\)
To carry out the computation, we will work with the Malliavin calculus. Suppose that \(H\in (0,1)\). Denote by \(\mathcal{E}\) the set of step functions on [0, T]. Introduce an inner product in \(\mathcal{E}\) by
Let \(\Vert h\Vert _{{\mathfrak {H}}}=\langle h,h\rangle _{{\mathfrak {H}}}^{1/2}\) for \(h\in \mathcal{E}\). The closure of \(\mathcal{E}\) with respect to \(\Vert \cdot \Vert _{{\mathfrak {H}}}\) is denoted by \({{\mathfrak {H}}}\). It is a real separable Hilbert space with an extended inner product \(\langle \cdot ,\cdot \rangle _{{\mathfrak {H}}}\) and the corresponding norm denoted by \(\Vert \cdot \Vert _{{\mathfrak {H}}}\). With an isonormal Gaussian process \({{\mathbb {W}}}=({{\mathbb {W}}}(h))_{h\in {{\mathfrak {H}}}}\) on some probability space \((\Omega ,\mathcal{F},P)\), it is possible to realize a fractional Brownian motion B by \(B_t={{\mathbb {W}}}(1_{[0,t]})\). Denote by \(({{\mathbb {D}}}_{s,p},\Vert \cdot \Vert _{s,p})\) the Sobolev space of Wiener functionals with differentiability index \(s\in {{\mathbb {R}}}\) and integrability index \(p>1\). We write \({{\mathbb {D}}}_\infty =\cap _{s,p}{{\mathbb {D}}}_{s,p}\). The Malliavin derivative is denoted by D, and the divergence operator (Skorohod integral) by \(\delta \). The k-fold Wiener integral of \(h^{\otimes k}\) is denoted by \(I_k(h^{\otimes k})=\delta ^k(h^{\otimes k})\) for \(h\in {{\mathfrak {H}}}\). For basic elements in the Malliavin calculus, we refer the reader to Watanabe (1984), Ikeda and Watanabe (1989), Nualart (2006) and Nourdin and Peccati (2012).
Let \(1_j^n=1_{(t_{j-1}^n,t_j^n]}\). The function \(1_j^n\) will simply be denoted by \(1_j\). Then
By the product formula, we have
By definition,
Thus,
for \(j=1,...,n-1\). In particular,
The use of \(V_{n,T}^{(2)}\) for the estimator \(\widehat{H}_n^{(2)}\) of (1.1) is accounted by (2.1), that entails
as \(n\rightarrow \infty \).
We say \(X_n=o_M(n^\nu )\) [resp. \(X_n=O_M(n^\nu )\)] for a sequence \((X_n)_{n\in {{\mathbb {N}}}}\) of random variables and \(\nu \in {{\mathbb {R}}}\) if \(X_n\in {{\mathbb {D}}}_\infty \) and \(\Vert X_n\Vert _{s,p}=o(n^\nu )\) [resp. \(\Vert X_n\Vert _{s,p}=O(n^\nu )\)] as \(n\rightarrow \infty \) for any \(s\in {{\mathbb {R}}}\) and \(p\in (1,\infty )\). We shall need a high-order of stochastic expansion for \(\widehat{H}_n^{(2)}\) to go up to the asymptotic expansion from the central limit theorem.
Proposition 2.1
There exists a sequence \((\psi _n)_{n\in {{\mathbb {N}}}}\) of Wiener functionals satisfying the following conditions.
-
(i)
\(\psi _n:\Omega \rightarrow [0,1]\), \(\psi _n\in {{\mathbb {D}}}_\infty \) for \(n\in {{\mathbb {N}}}\), and \(\Vert \psi _n-1\Vert _{s,p}=O(n^{-L})\) as \(n\rightarrow \infty \) for every \((s,p,L)\in {{\mathbb {R}}}\times (1,\infty )\times (0,\infty )\).
-
(ii)
The sequence \((Z_n)_{n\in {{\mathbb {N}}}}\) of random variables defined by
$$\begin{aligned} Z_n= & {} (2\log 2)\sqrt{n}\big (\widehat{H}_n^{(2)}-H\big )\psi _n \end{aligned}$$(2.2)for \(n\in {{\mathbb {N}}}\) admits a stochastic expansion
$$\begin{aligned} Z_n= & {} M_n+n^{-1/2}N_n, \end{aligned}$$(2.3)where
$$\begin{aligned} M_n= & {} -n^{1/2}\bigg \{\bigg [\frac{V^{(2)}_{2n,T}}{E[V^{(2)}_{2n,T}]}-1\bigg ] -\bigg [\frac{V^{(2)}_{n,T}}{E[V^{(2)}_{n,T}]}-1\bigg ]\bigg \} \end{aligned}$$(2.4)and
$$\begin{aligned} N_n= & {} n\bigg \{ \frac{1}{2}\bigg [\frac{V^{(2)}_{2n,T}}{E[V^{(2)}_{2n,T}]}-1\bigg ]^2 {-}\frac{1}{2}\bigg [\frac{V^{(2)}_{n,T}}{E[V^{(2)}_{n,T}]}-1\bigg ]^2 \bigg \}-\frac{1}{2}+O_M(n^{-1/2}) \end{aligned}$$(2.5)as \(n\rightarrow \infty \). In particular, \(Z_n\in {{\mathbb {D}}}_\infty \) for \(n\in {{\mathbb {N}}}\).
See Sect. 3.3 for the proof of Proposition 2.1.
From now on, our approach will be as follows. We will first validate asymptotic expansion of the distribution of \(M_n\), and next apply the perturbation method to the stochastic expansion (2.3) with (2.4) and (2.5) in order to derive an Edgeworth expansion formula for \(\widehat{H}_n^{(2)}\). For positive numbers K and \(\gamma \), \(\mathcal{E}(K,\gamma )\) denotes the set of measurable functions on \({{\mathbb {R}}}\) such that \(|f(z)|\le K(1+|z|^\gamma )\) for all \(z\in {{\mathbb {R}}}\). The density function of the normal distribution with mean \(\mu \) and variance C is denoted by \(\phi (z;\mu ,C)\). The main interest of this paper is in the following theorem.
Theorem 2.2
There exist a positive constant \(v_H\) and an odd polynomial q(z) of degree 3 such that the function
satisfies
for every positive numbers K and \(\gamma \). See (3.38) of Sect. 3.5 for the definition of q(z) and \(v_H\).
The second-order modification of \(\widehat{H}_n^{(2)}\) is given by
where the function \(b:[0,1]\rightarrow {{\mathbb {R}}}\) is assumed to be smooth on (0, 1), and the symbols \(\wedge \) and \(\vee \) stand for \(\min \) and \(\max \), respectively. Asymptotic expansion of the distribution of \(\widehat{H}_n^{(b)}\) can be obtained easily by a trivial modification of that of \(\widehat{H}_n^{(2)}\). Define \(p_n^{(b)}(z)\) by
with
Theorem 2.3
For any positive numbers K and \(\gamma \),
as \(n\rightarrow \infty \).
As an application of Theorem 2.3, we obtain a second-order unbiased estimator \(\widehat{H}_n^{(*)}\) by choosing the function
as b(H), namely, \(\widehat{H}_n^{(*)}=\widehat{H}_n^{(b^*)}\). Then
as \(n\rightarrow \infty \). If to take the function
as b(H), then the estimator \(\widehat{H}_n^{(**)}=\widehat{H}_n^{(b^{**})}\) satisfies
and
as \(n\rightarrow \infty \). That is, the estimator \(\widehat{H}_n^{(**)}\) is second-order median unbiased. See Remark 3.15 about regularity of the functions \(b^*\) and \(b^{**}\).
The rest of this paper will be devoted to the proof of Theorems 2.2 and 2.3.
3 Proofs
This section proves Theorems 2.2 and 2.3. We first derive the asymptotic expansion formula of \(M_n\) in Sect. 3.4 (Theorem 3.14). In Sect. 3.5, using the perturbation method with the asymptotic expansion of \(M_n\) and the stochastic expansion of \(Z_n\) (Proposition 2.1, justified in Sect. 3.3), we obtain the asymptotic expansion formula of \(Z_n\), and hence of \(\widehat{H}_n^{(2)}\). To obtain the asymptotic expansion of \(M_n\) and to use the perturbation method, we need to know the asymptotic behavior of \(U_n=(M_n^{[1]},M_n^{[2]},n^{1/2}\widetilde{G}_n)\), which is defined by (3.20) below. Indeed, the fourth moment theorem will be applied to \(U_n\) in Proposition 3.12 (Sect. 3.2.3). In order to calculate the asymptotic covariance matrix of \(U_n\), we give technical lemmas in Sects. 3.1 and 3.2. The proof of Theorem 2.3 is given in Sect. 3.6.
The estimator \(\widehat{H}_n^{(2)}\) includes the second-order differences of the fBm B of two different time steps, namely \(d_{n,j}\) and \(d_{2n,j}\). The variance and covariance of \(d_{n,j}\) and \(d_{2n,j'}\) are represented by \(\widehat{\rho }\) and \(\widetilde{\rho }\) in Lemma 3.5(a) and (b), and 3.7(a). (See (3.5) and (3.13) for the definition of \(\widehat{\rho }\) and \(\widetilde{\rho }\).) Since the tail of \(\widehat{\rho }(j)\) and \(\widetilde{\rho }(j)\) (as \(j\rightarrow \infty \)) decay fast enough (Lemma 3.4 and 3.6), we can estimate the asymptotic behavior of the cumulants of \((M_n^{[1]},M_n^{[2]})\): Lemmas 3.5 and 3.7 for the second cumulants, Lemmas 3.8 and 3.9 for the third, and Lemmas 3.10 and 3.11 for the fourth. The basic idea of the estimates above are presented at Lemma 3.1. (See also Remark 3.2.)
3.1 Preliminary lemmas
Let \(k\in {{\mathbb {N}}}\) with \(k\ge 2\). Consider \(k+1\) sequences of positive integers
such that
for some constant \(p_\alpha \in (0,\infty )\) for every \(\alpha =1,...,k\).
For functions \(\varrho _1,...,\varrho _k:{{\mathbb {Z}}}\rightarrow {{\mathbb {R}}}\), let
and also let
For \({{\mathbb {I}}}=\{1,...,k-1\}\), let
Lemma 3.1
-
(a)
\(\displaystyle {{\mathbb {V}}}(\varrho _1,...,\varrho _k) \>\le \>{{\mathbb {V}}}(|\varrho _1|,...,|\varrho _k|) \>\le \>\big \Vert \varrho _1\big \Vert _{\ell _{\frac{k}{k-1}}}\cdots \big \Vert \varrho _k\big \Vert _{\ell _{\frac{k}{k-1}}} \).
-
(b)
$$\begin{aligned} A_n(\varrho _1,...,\varrho _k)= & {} \sum _{i_1,...,i_{k-1}\in {{\mathbb {Z}}}} \varrho _1(-i_1)\varrho _2(i_1-i_2)\varrho _{3}(i_2-i_3)\cdots \\{} & {} \cdots \varrho _{k-1}(i_{k-2}-i_{k-1})\varrho _k(i_{k-1}) p_n(i_1,...,i_{k-1}). \end{aligned}$$
-
(c)
Suppose that the condition (3.1) is satisfied. If \({{\mathbb {V}}}(|\varrho _1(-\cdot )|,|\varrho _2|,...,|\varrho _k|) <\infty \), then the limit
$$\begin{aligned} \lim _{n\rightarrow \infty }A_n(\varrho _1,...,\varrho _k)= & {} \bigg (\bigwedge _{\alpha =1,...,k} p_\alpha \bigg ) \sum _{i_1,...,i_{k-1}\in {{\mathbb {Z}}}} \varrho _1(-i_1)\varrho _2(i_1-i_2)\varrho _3(i_2-i_3)\cdots \nonumber \\{} & {} \times \varrho _{k-1}(i_{k-2}-i_{k-1})\varrho _k(i_{k-1}) \nonumber \\ \end{aligned}$$(3.2)exists, that is, the sum on the right-hand side of (3.2) is finite. In particular, if \(\big \Vert \varrho _\alpha \big \Vert _{\ell _{\frac{k}{k-1}}}<\infty \) for all \(\alpha \in \{1,...,k\}\), then the convergence (3.2) holds under (3.1).
-
(d)
Suppose that the functions \(\varrho _{1},...,\varrho _{k}:{{\mathbb {Z}}}\rightarrow {{\mathbb {R}}}\) satisfy
$$\begin{aligned} |\varrho _{\alpha }(j)|\le & {} C(1+|j|)^{-\gamma }\quad (j\in {{\mathbb {Z}}};\>\alpha =1,...,k) \end{aligned}$$for some constant \(\gamma >\frac{k}{k-1}\). Suppose that the condition (3.1) is satisfied. Then
$$\begin{aligned} {{\mathbb {V}}}(\varrho _{1},...,\varrho _{k}) \le {{\mathbb {V}}}(|\varrho _{1}|,...,|\varrho _{k}|) < \infty \end{aligned}$$In particular,
$$\begin{aligned} \sup _{n\in {{\mathbb {N}}}}\big |A_n(\varrho _{1},...,\varrho _{k})\big | \le \sup _{n\in {{\mathbb {N}}}}\big |A_n(|\varrho _{1}|,...,|\varrho _{k}|)\big |<\infty , \end{aligned}$$and the convergence (3.2) holds.
Proof
(a) The system of linear equations
is solved by \(p=\frac{k}{k-1}\) and \(q=k\). Therefore, Young’s inequality yields
(b) By the change of variables
for given \(j_1\), we obtain
(c) Under the condition (3.1),
Then Lebesgue’s theorem implies the convergence (3.2).
(d) Use (a) and (c) to prove (d). \(\square \)
Remark 3.2
(i) We have \({{\mathbb {V}}}\big (\overbrace{|\rho _1(-\cdot )|,...,|\rho _k|}^{k}\big ) < \infty \) for \((\rho _1,...,\rho _k)\in \{\widehat{\rho },\widetilde{\rho }\}^k\) since \(4-2H>2>\frac{k-1}{k}\) for every \(k\ge 2\), thanks to the estimates (3.6) and (3.14) below. (ii) It is possible to strengthen the result (3.2) to a representation of the form \(A_n(\varrho _1,...,\varrho _k)\>=\>\text {constant}+O(n^{-1})\) by putting a more restrictive condition on \(\nu (n)\) and \(\nu _\alpha (n)\) than (3.1). In this paper, such a convergence is necessary only in the case \(k=2\), and we will give estimate for the error term directly without the help of Lemma 3.1.
Recall that \(1^n_j=1_{(t_{j-1}^n,t_j^n]}\). Let
for \(\alpha \in \{1,2\}\). Then \(M_n=I_2(f_n^{[1]})-I_2(f_n^{[2]})\) for \(M_n\) defined by (2.4). The operator \(L=-\delta D\) is the Malliavin operator (Ornstein-Uhlenbeck operator). It is a numeric operator such that \(LF=(-q)F\) for elements F of the q-th Wiener chaos. Since \( (-L)^{-1}M_n \>=\>2^{-1}I_2(f_n^{[1]})-2^{-1}I_2(f_n^{[2]}) \) and \( D(-L)^{-1}M_n \>=\>I_1(f_n^{[1]})-I_1(f_n^{[2]}) \), the second-order \(\Gamma \)-factor \(\Gamma _n^{(2)}(M_n)\) of \(M_n\) is given by
for \(n\in {{\mathbb {N}}}\). The product formula gives
where \(f_n^{[2]}\odot _1f_n^{[1]}\) is the symmetrized 1-contraction of \(f_n^{[2]}\otimes f_n^{[1]}\).
Backward shift operator \(\mathsf{{B}}\) is defined by \(\mathsf{{B}}\theta (x)=\theta (x-1)\) for \(x\in {{\mathbb {R}}}\) and a sequence \(\theta =(\theta (x+t))_{t\in {{\mathbb {Z}}}}\) of numbers. The following lemma is an exercise.
Lemma 3.3
Let \(x\in {{\mathbb {R}}}\), \(I\in {{\mathbb {N}}}\) and \(k_1,...,k_I\in {{\mathbb {N}}}\). Then
for any function \(f\in C^I\big ([x-\sum _{i=1}^Ik_i,x]\big )\), the set of functions of class \(C^I\) on \([x-\sum _{i=1}^Ik_i,x]\).
Let
Lemma 3.4
In particular, \(\sum _{i=1}^\infty |i|^3\widehat{\rho }(i)^2<\infty \).
Proof
With the backward shift operator \(\mathsf{{B}}\), we have
Therefore, with \(f(x)=|x+2|^{2H}\) we have
and hence (3.6) follows from Lemma 3.3. \(\square \)
Let
and let
Lemma 3.5
-
(a)
For \(j,j'\in \{1,...,n-1\}\),
$$\begin{aligned} \big \langle 1_{j+1}^n-1_j^n, 1_{j'+1}^n-1_{j'}^n\big \rangle _{{\mathfrak {H}}}= & {} \frac{T^{2H}}{n^{2H}}\widehat{\rho }(j-j'). \end{aligned}$$(3.8) -
(b)
For \(k,k'\in \{1,...,2n-1\}\),
$$\begin{aligned} \big \langle 1_{k+1}^{2n}-1_k^{2n}, 1_{k'+1}^{2n}-1_{k'}^{2n}\big \rangle _{{\mathfrak {H}}}= & {} \frac{T^{2H}}{(2n)^{2H}}\widehat{\rho }(k-k'). \end{aligned}$$(3.9) -
(c)
As \(n\rightarrow \infty \),
$$\begin{aligned} \Vert f_n^{[1]}\Vert _{{{\mathfrak {H}}}^{\otimes 2}}^2= & {} \frac{1}{2}\Sigma _{11}+O(n^{-1}). \end{aligned}$$(3.10) -
(d)
As \(n\rightarrow \infty \),
$$\begin{aligned} \Vert f_n^{[2]}\Vert _{{{\mathfrak {H}}}^{\otimes 2}}^2= & {} \frac{1}{2}\Sigma _{22}+O(n^{-1}). \end{aligned}$$(3.11)
Proof
(a) and (b): We will use an elementary formula
for \(a,b,c,d\in [0,T]\). We obtain (3.8) and(3.9) since
from (3.5). (3.9) is now trivial.
(c) and (d): For \(j,j'\in \{1,...,n-1\}\),
Therefore
With this result, we also obtain
\(\square \)
Let
for \(j\in {{\mathbb {Z}}}\). We will write
Then \(\widetilde{\rho }(k,j)=\widetilde{\rho }(k-2j)\).
Lemma 3.6
In particular, \(\sum _{j\in {{\mathbb {Z}}}}|i|^3\widetilde{\rho }(j)^2<\infty \).
Proof
We have
for the backward shift operator \(\textsf{B}\). Therefore we obtain (3.14). \(\square \)
Let
Lemma 3.7
-
(a)
For \(j\in \{1,...,n-1\}\) and \(k\in \{1,...,2n-1\}\),
$$\begin{aligned} \big \langle 1_{j+1}^{n}-1_j^{n}, 1_{k+1}^{2n}-1_k^{2n}\big \rangle _{{\mathfrak {H}}}= & {} \frac{T^{2H}}{n^{2H}}\widetilde{\rho }(k,j). \end{aligned}$$(3.15) -
(b)
As \(n\rightarrow \infty \),
$$\begin{aligned} \langle f_n^{[2]},f_n^{[1]}\rangle _{{{\mathfrak {H}}}^{\otimes 2}}= & {} \frac{1}{2}\Sigma _{12}+O(n^{-1}). \end{aligned}$$(3.16)
Proof
From (3.12), we have
This shows (3.15). By using (2.1) and (3.15), we obtain
and hence
with the help of Lemma 3.6. \(\square \)
3.2 A central limit theorem toward the asymptotic expansion
Let
From (3.4), Lemma 3.5 (c), (d) and Lemma 3.7 (b), we have
where
Write \( M_n^{[i]}=I_2(f_n^{[i]}) \) for \(i=1,2\). By the definition of \(f_n^{[i]}\) (\(i=1,2\)) given at (3.3), we have \(M_n=I_2(f_n^{[1]})-I_2(f_n^{[2]}) =M_n^{[1]}-M_n^{[2]}.\) We will derive a central limit theorem for the sequence
For \(f,g\in {{\mathfrak {H}}}^{\odot 2}\), symmetric tensors, we have
Thus,
for \(\alpha =1,2\).
3.2.1 Cubic formulas
Define \(\kappa (\alpha _1;\alpha _2,\alpha _2)\) (\(\alpha _1,\alpha _2\in \{1,2\}\)) as follows.
and
Lemma 3.8
As n tends to \(\infty \), the following estimates hold:
-
(a)
\(\displaystyle \big \langle f_n^{[1]},n^{1/2}f_n^{[1]}\otimes _1f_n^{[1]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (1;1,1)+o(1) \).
-
(b)
\(\displaystyle \big \langle f_n^{[2]},n^{1/2}f_n^{[1]}\otimes _1f_n^{[1]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (2;1,1)+o(1) \).
-
(c)
\(\displaystyle \big \langle f_n^{[1]},n^{1/2}f_n^{[2]}\otimes _1f_n^{[2]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (1;2,2)+o(1) \).
-
(d)
\(\displaystyle \big \langle f_n^{[2]},n^{1/2}f_n^{[2]}\otimes _1f_n^{[2]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (2;2,2)+o(1) \).
Proof
(a): By (3.3) and Lemma 3.5 (a), (b), we have
for \(\alpha =1,2\). Therefore, from (3.3), (3.22) and (2.1), we obtain
thanks to Lemma 3.1.
(b): Similarly to the proof of (a),
Indeed,
by Lemma 3.1.
(c): By using (3.3), (3.22) and (2.1), we obtain
In fact, thanks to Lemma 3.1,
and the last expression is equal to
(d): The proof of (d) is similar to that of (a). \(\square \)
Lemma 3.9
-
(a)
\(\displaystyle n^{1/2}\big \langle f_n^{[1]},f_n^{[2]}\odot _1f_n^{[1]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (2;1,1)+o(1)\).
-
(b)
\(\displaystyle n^{1/2}\big \langle f_n^{[2]},f_n^{[2]}\odot _1f_n^{[1]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (1;2,2)+o(1)\).
Proof
We have
The first equality stands since \(f_n^{[1]}\) is symmetric (i.e. \(\in \mathcal{H}^{\odot 2}\)). Hence we have (a) by Lemma 3.8 (b). Similar arguments and Lemma 3.8 (c) prove (b). \(\square \)
3.2.2 Fourth power formulas
Recall that
and
Let
Lemma 3.10
-
(a)
\(\displaystyle n\big \langle f_n^{[1]}\otimes _1f_n^{[1]},\> f_n^{[1]}\otimes _1f_n^{[1]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (1,1;1,1)+o(1)\).
-
(b)
\(\displaystyle n\big \langle f_n^{[1]}\otimes _1f_n^{[1]},\> f_n^{[2]}\otimes _1f_n^{[2]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (1,1;2,2)+o(1)\).
-
(c)
\(\displaystyle n\big \langle f_n^{[2]}\otimes _1f_n^{[2]},\> f_n^{[2]}\otimes _1f_n^{[2]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (2,2;2,2)+o(1)\).
Proof
It follows from Lemma 3.1 that
Thus, (a) has been obtained. Similarly, we can show (c).
Now we have
by the following argument:
and with Lemma 3.1, we see the gap between the above expression and the one below is of o(1):
Therefore we obtained (b). \(\square \)
Let
and
where \(\kappa (1,1;2,2)\) is defined at (3.25) and
Lemma 3.11
-
(a)
\(\displaystyle n\big \langle f_n^{[1]}\otimes _1f_n^{[1]},\> f_n^{[1]}\odot _1f_n^{[2]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (1,1;1,2)+o(1)\).
-
(b)
\(\displaystyle n\big \langle f_n^{[2]}\otimes _1f_n^{[2]},\> f_n^{[1]}\odot _1f_n^{[2]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (1,2;2,2)+o(1) \).
-
(c)
\(\displaystyle n\big \langle f_n^{[1]}\odot _1f_n^{[2]},\> f_n^{[1]}\odot _1f_n^{[2]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}} \>=\>\kappa (1,2;1,2)+o(1) \).
Proof
(a): We have
The last equality can be verified by
(b): We will consider the product
Define \(k^0\) and \(k^1\) by
respectively for an integer k. The sum \(\widetilde{\sum }_{k_1,k_2,k_3}\) should read according to a proper configuration of \(k^0_i\) and \(k^1_i\), \(i=1,2,3\). To proceed,
and by Lemma 3.1, the above sum can be written as follows:
Therefore
(c): For the product \(n\big \langle f_n^{[1]}\odot _1f_n^{[2]},\> f_n^{[1]}\odot _1f_n^{[2]}\big \rangle _{{{\mathfrak {H}}}^{\otimes 2}}\), we have
Now the convergence in (c) is verified as follows. The first sum is the half of the sum (3.26). Hence it converges to \(\frac{1}{2}\kappa (1,1;2,2)\). For the second sum,
\(\square \)
3.2.3 Central limit theorems
We are now on the point of getting a central limit theorem for \(U_n\). Recall the notation
Denote by \(N_k(0,C)\) the k-dimensional centered normal distribution with covariance matrix C.
Proposition 3.12
The random vector \(U_n\) is asymptotically normal, that is,
as \(n\rightarrow \infty \), where \({{\mathfrak {U}}}=({{\mathfrak {U}}}_{ij})_{i,j=1}^3\) is a symmetric matrix with components
Proof
The Cramér-Wold device is used to show the three-dimensional central limit theorem. It follows from Lemmas 3.8, 3.9, 3.10 and 3.11 with the representations (3.21) and (3.19) that the asymptotic covariance matrix of \(v\cdot U_n\) is \(^tv{{\mathfrak {U}}}v\) for every \(v\in {{\mathbb {R}}}^3\). To apply the fourth moment theorem (Nualart and Peccati 2005), what we need to show is \(\text {Var}\big [|Dv\cdot U_n|_{{\mathfrak {H}}}^2\big ]\rightarrow 0\) as \(n\rightarrow \infty \). We will demonstrate \(\text {Var}[|DF_n|_{{\mathfrak {H}}}^2]\rightarrow 0\) only for the component \(F_n=2^{-1}n^{1/2}I_2\big (f_n^{[1]}\otimes _1f_n^{[1]}\big )\), because the proof of the convergence of the other components is quite similar. For example, we consider a component \(F_n=2^{-1}n^{1/2}I_2\big (f_n^{[1]}\otimes _1f_n^{[1]}\big )\). By definition of \(F_n\), we have
and
Therefore
We have \({{\mathbb {V}}}\big (\overbrace{\widehat{\rho },...,\widehat{\rho }}^{k}\big ) < \infty \) since \(4-2H>2>\frac{k-1}{k}\) for every \(k\ge 2\). Therefore
by Lemma 3.1. \(\square \)
A central limit theorem for \( T_n\>=\>\big (M_n,n^{1/2}\widetilde{G}_n\big ) \) follows immediately from Proposition 3.12 since \(M_n=M_n^{[1]}-M_n^{[2]}\).
Corollary 3.13
\( T_n \rightarrow ^d N_2(0,{{\mathfrak {T}}}) \) as \(n\rightarrow \infty \), where \({{\mathfrak {T}}}=({{\mathfrak {T}}}_{ij})_{i,j=1}^2\) is a symmetric matrix with components
We remark that
where \(G_\infty \) was defined by (3.17). Indeed, from Proposiion 3.12, \((M_n^{[1]},M_n^{[2]})\rightarrow ^d(M_\infty ^{[1]},M_\infty ^{[2]})\), a two-dimensional Gaussian random vector with covariance matrix \((\Sigma _{ij})_{i,j=1,2}\).
We applied the Schwarz inequality to obtain (3.27).
3.3 Proof of Proposition 2.1
Let \(\psi :{{\mathbb {R}}}\rightarrow [0,1]\) be a smooth function satisfying \(\psi (x)=1\) on \(\{x;|x|\le 1/2\}\), and \(\psi (x)=0\) on \(\{x;|x|\ge 1\}\). We define \(\psi _n\) by \(\psi _1=0\) and
for \(n\ge 2\). Whenever \(\psi _n>0\),
where \(\eta _0\) is a positive constant. The properties of the truncation functional \(\psi _n\) in (i) of Proposition 2.1 are easy to verified. We choose a sufficiently small \(\eta _0\) and a positive integer \(n_0\) both depending on H so that \(\widehat{H}^{(2)}_n\in (0,1)\) whenever \(\psi _n>0\) and \(n\ge n_0\). We will only consider \(n\ge n_0\) in what follows. On the event \(\{\psi _n>0\}\), we have
where
Since the family
of Wiener functionals is bounded in \({{\mathbb {D}}}_\infty \), as is known by the hypercontractivity and stability of the \({{\mathfrak {F}}}\) under the Malliavin operator (i.e., \(-2^{-1}L\) is the identity on the second chaos), we see \(R_n\psi _n = O_M(n^{-1.5})\) with the help of (3.28). \(\square \)
3.4 Asymptotic expansion of \(M_n\)
We will apply Theorem 3 of Tudor and Yoshida (2019a) to the sequence \((M_n)_{n\in {{\mathbb {N}}}}\) by checking Conditions [C1]-[C3] therein. Condition [C1] is satisfied by Corollary 3.13. Conditions [C2] and [C3] for boundedness of \(\{M_n\}_{n\in {{\mathbb {N}}}}\) and \(\{n^{1/2}(G_n-G_\infty )\}_{n\in {{\mathbb {N}}}}\) are verified with the hypercontractivity and stability of a fixed chaos under the Malliavin operator, as mentioned in the last part of the proof of Proposition 2.1. Remark that \(n^{1/2}(G_n-G_\infty )=n^{1/2}\widetilde{G}_n+O(n^{-1/2})\); \(n^{1/2}\widetilde{G}_n\) is in the second chaos and the error term is deterministic.
According to Corollary 3.13, \(T_n\rightarrow ^dT_\infty \) as \(n\rightarrow \infty \), where \(T_\infty =(T_{\infty ,1},T_{\infty ,2})\) is a centered two-dimensional Gaussian random variable defined on some probability space with variance matrix \({{\mathfrak {T}}}\). Then
where \(\theta ={{\mathfrak {T}}}_{12}/{{\mathfrak {T}}}_{11}\). Suggested by Formula (23) of Tudor and Yoshida (2019a), we define a symbol \({{\mathfrak {S}}}_n(\texttt{i}z)\) by
We note that introducing the parameter \(\theta \) makes the resulting formula look slightly simple, that involves many \({{\mathfrak {T}}}_{11}^{-1}\) or \(G_\infty ^{-1}\) by derivatives. An approximate density \(p_n^M(z)\) is then defined by
In other words,
The function \(p_n^M(z)\) is the Fourier inversion of the function
Recall that \(\mathcal{E}(K,\gamma )\) is the set of measurable functions on \({{\mathbb {R}}}\) such that \(|f(z)|\le K(1+|z|^\gamma )\) for all \(z\in {{\mathbb {R}}}\). Theorem 3 of Tudor and Yoshida (2019a) gives the following result.
Theorem 3.14
For any positive numbers K and \(\gamma \),
as \(n\rightarrow \infty \).
3.5 Proof of Theorem 2.2
In this section, we will combine the asymptotic expansion of \(M_n\) with the perturbation method to give asymptotic expansion for \(Z_n\). The perturbation method was used in Yoshida (1997), Yoshida (2001), Yoshida (2013). For this method, we refer the reader to Sakamoto and Yoshida (2003) and Yoshida (2020). Following the formulation in Sakamoto and Yoshida (2003), it is possible to derive asymptotic expansion of \({{\mathbb {S}}}_n={{\mathbb {X}}}_n+r_n{{\mathbb {Y}}}_n\) from the joint convergence of \(({{\mathbb {X}}}_n,{{\mathbb {Y}}}_n)\) once that of \({{\mathbb {X}}}_n\) was obtained. We set \({{\mathbb {X}}}_n=M_n\), \({{\mathbb {Y}}}_n=N_n\), \({{\mathbb {S}}}_n=Z_n\) and \(r_n=n^{-1/2}\). Let
Then,
whenever \(|\xi _n|<1\). Therefore \(\sup _nE\big [1_{\{|\xi _n|<1\}}|DM_n|_{{\mathfrak {H}}}^{-p}\big ]<\infty \) for any \(p>1\). Moreover, \(P\big [|\xi _n|>1/2\big ]=O(n^{-L})\) for every \(L>0\) since \(\widetilde{G}_n\) is bounded in \(L^{\infty \text {--}}=\cap _{p>1}L_p\). It is easy to see \(\xi _n\), as well as \(M_n\) and \(N_n\), is bounded in \({{\mathbb {D}}}_\infty \). On the other hand, since
from Proposition 3.12, we obtain
Therefore
with the coefficient
From (3.30) and (3.32), we obtain the expansion formula for \(Z_n\) as
where
recall \(\theta ={{\mathfrak {T}}}_{12}/{{\mathfrak {T}}}_{11}\). That is,
as \(n\rightarrow \infty \). However, since \(\widehat{H}_n^{(2)}\in [0,1]\) by definition, and the family \({{\mathfrak {F}}}\) of (3.29) is bounded in \(L^{\infty \text {--}}\), we know
as \(n\rightarrow \infty \), for every \(L>0\), where \(c=2\log 2\). Consequently
as \(n\rightarrow \infty \).
Finally we rescale the argument of \(p_n^Z(z)\) by c to obtain the approximate density \(p_n(z)\) for \(\widehat{H}_n^{(2)}\) appearing in (2.6) with
This completes the proof of Theorem 2.2. \(\square \)
3.6 Proof of Theorem 2.3
The effect of the modification in (2.8) appears as the shift of \(\widehat{H}_n^{(2)}\) by the constant \(-b(H)/n\) in the asymptotic expansion. So, the estimate (2.10) is almost obvious if we replace the function f(z) by \(f\big (z-n^{-1/2}b(H)\big )\) in (2.7) with some modified \((M,\gamma )\) and next by expansion after change of variables. Rigorous justification does not matter. We may cut off the event \(\{\widehat{H}_n^{(2)}\not \in (H/2,(H+1)/2)\}\) since \(\big \Vert \sqrt{n}\big (\widehat{H}_n^{(b)}-H\big )\big \Vert _\infty =O(n^{1/2})\). Just start with \(N_n\) including \(-cb\big (\widehat{H}_n^{(2)})\). Then a modification is adding \(-cb(H)\) in (3.31) to the limit of \(N_n\). This causes addition of \(-cb(H)G_\infty ^{-1}z\phi (z;0,G_\infty )\) to (3.32) to give the same result as the one by the above intuition. \(\square \)
Remark 3.15
It is not difficult to show that \(b^*\) and \(b^{**}\in C^\infty (0,1)\) by using the representation of \(\Sigma _{ij}\) (\(i,j=1,2\)) in H, and the representation (3.38) of q(z) depending on H, as well as the uniform positivity (3.27) of \(G_\infty \) depending on H.
4 Numerical study
We give some results of simulation. The following two figures compare the histogram of \(\widehat{H}_n^{(2)}\) with normal approximation and asymptotic expansion, in the case of \(H=0.5\) and \(n=32\). The left figure plots the histogram of \(\widehat{H}_n^{(2)}\) and the right one the histogram but without cutoff as (1.1), i.e., \(\widehat{H}_n^{(2)'}:= \frac{1}{2}-\frac{1}{2\log 2}\log \big ({V_{2n,T}^{(2)}}/{V_{n,T}^{(2)}}\big )\). The solid lines are obtained by the asymptotic expansion, and the dashed ones by the normal approximation. Each of the following numerical results is based on \(10^5\) paths.
![figure a](http://media.springernature.com/lw685/springer-static/image/art%3A10.1007%2Fs11203-023-09298-8/MediaObjects/11203_2023_9298_Figa_HTML.png)
When n increases, the atoms become smaller as well as the difference between the asymptotic expansion and the normal approximation decreases.
![figure b](http://media.springernature.com/lw685/springer-static/image/art%3A10.1007%2Fs11203-023-09298-8/MediaObjects/11203_2023_9298_Figb_HTML.png)
Since the precision of approximation of the atoms is determined by that of density, we will leave the curves and omit to draw atoms by the approximations, in the following several plots for different values of H. In practical use, the density should be integrated on the tails to approximate the distribution of \(\widehat{H}_n^{(2)}\) of (1.1), that has atoms on the endpoints 0 and 1.
![figure c](http://media.springernature.com/lw685/springer-static/image/art%3A10.1007%2Fs11203-023-09298-8/MediaObjects/11203_2023_9298_Figc_HTML.png)
References
Benassi A, Cohen S, Istas J, Jaffard S (1998) Identification of filtered white noises. Stochast Process Appl 75(1):31–49
Bhattacharya RN (1971) Rates of weak convergence and asymptotic expansions for classical central limit theorems. Ann. Math. Stat. 42:241–259
Bhattacharya RN, Ranga Rao R (1976) Normal approximation and asymptotic expansions, p. 274. Wiley, New York. Wiley Series in Probability and Mathematical Statistics
Bhattacharya R, Lin L, Patrangenaru V (2016) A course in mathematical statistics and large sample theory. Springer, Berlin
Cramér H (1928) On the composition of elementary errors: first paper: mathematical deductions. Scand Actuarial J 1928(1):13–74
Cramér H (2004) Random variables and probability distributions, vol 36. Cambridge University Press, Cambridge
Gnedenko B, Kolmogorov A (1954) Limit distributions for sums of independent random variables. Cambridge University Press, Cambridge
Ikeda N, Watanabe S (1989) Stochastic differential equations and diffusion processes, 2nd edn. North-Holland Mathematical Library, vol. 24, p. 555. North-Holland Publishing Co., Amsterdam
Istas J, Lang G (1997) Quadratic variations and estimation of the local hölder index of a gaussian process. In: Annales de l’Institut Henri Poincare (B) probability and statistics, vol 33, pp 407–436. Elsevier
Kubilius K, Mishura Y, Ralchenko K (2018) Parameter estimation in fractional diffusion models, vol 8. Springer, Berlin
Nourdin I, Peccati G (2012) Normal approximations with Malliavin calculus: from Stein’s method to universality, vol 192. Cambridge University Press, Cambridge
Nualart D (2006) The Malliavin calculus and related topics, 2nd edn. Probability and its Applications, Springer, Berlin, p 382
Nualart D, Peccati G (2005) Central limit theorems for sequences of multiple stochastic integrals. Ann Probab 33(1):177–193
Nualart D, Yoshida N (2019) Asymptotic expansion of skorohod integrals. Electron J Probab 24
Petrov VV (1975) Sums of independent random variables, p. 346. Springer, Translated from the Russian by A. A. Brown, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band, p 82
Sakamoto Y, Yoshida N (2003) Asymptotic expansion under degeneracy. J Jpn Stat Soc 33(2):145–156
Tudor CA, Yoshida N (2019) High order asymptotic expansion for Wiener functionals. arXiv preprint arXiv:1909.09019
Tudor CA, Yoshida N (2019) Asymptotic expansion for vector-valued sequences of random variables with focus on Wiener chaos. Stochast Process Appl 129(9):3499–3526
Tudor CA, Yoshida N (2020) Asymptotic expansion of the quadratic variation of a mixed fractional Brownian motion. Stat Infer Stoch Process 23(2):435–463
Watanabe S (1984) Lectures on stochastic differential equations and malliavin calculus (Notes by Nair, M Gopalan and Rajeev, B) vol. 73. Springer
Yamagishi H, Yoshida N (2022) Order estimate of functionals related to fractional brownian motion and asymptotic expansion of the quadratic variation of fractional stochastic differential equation. arXiv preprint arXiv:2206.00323
Yoshida N (1997) Malliavin calculus and asymptotic expansion for martingales. Probab Theory Relat Fields 109(3):301–342
Yoshida N (2001) Malliavin calculus and martingale expansion. Bull Sci Math 125(6–7):431–456
Yoshida N (2013) Martingale expansion in mixed normal limit. Stochast Process Appl 123(3):887–933
Yoshida N (2016) Asymptotic expansions for stochastic processes. In: Denker M, Waymire E (eds) Rabi N. Bhattacharya. Springer, pp 15–32
Yoshida N (2020) Asymptotic expansion of a variation with anticipative weights. arXiv:2101.00089
Funding
Open access funding provided by The University of Tokyo. The first author was partially supported by the Swedish Foundation for Strategic Research grant Nr. UKR22-0017. This work was in part supported by Japan Science and Technology Agency CREST JPMJCR2115; Japan Society for the Promotion of Science Grants-in-Aid for Scientific Research Nos. 17H01702 and 23H03354 (Scientific Research); and by a Cooperative Research Program of the Institute of Statistical Mathematics.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
On behalf of all authors, the corresponding author states that there is no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Mishura, Y., Yamagishi, H. & Yoshida, N. Asymptotic expansion of an estimator for the Hurst coefficient. Stat Inference Stoch Process 27, 181–211 (2024). https://doi.org/10.1007/s11203-023-09298-8
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11203-023-09298-8