Keywords: Asymptotic normality, Consistency, Ergodic theorem, Method of moments, Tempered fractional Brownian motion
MSC Classification: 60G15, 60G18, 60G22, 62H12, 62F12
Introduction
Tempered fractional Brownian motions (tfBms) of first or second kind have been proposed by [17 , 26 ] as an extension of the fractional Brownian motion (fBm). They are constructed by exponentially tempering the power law kernel in the Mandelbrot and Van Ness representation of the fBm. A tfBm B H , λ B_{H,\lambda} B H , λ shares the Gaussianity and stationarity of increments with the fBm, but is characterized by a mean squared displacement which decays rapidly to a constant, i.e., Var ( B H , λ ( t ) ) = C + O ( t H − 1 / 2 e − λ t ) \operatorname{Var}(B_{H,\lambda}(t)) = C + O(t^{H-1/2}e^{-\lambda t}) Var ( B H , λ ( t )) = C + O ( t H − 1/2 e − λ t ) , as t → ∞ t \to\infty t → ∞ , for some constant C > 0 C > 0 C > 0 . The autocorrelation of its increments exhibits the semi-long range dependence, that is, it behaves like a power over moderate scales, but quasi-exponentially over large scales. TfBms are widely used to model wind speed data summarized by the so-called Davenport spectrum (e.g., [7 ] ). The latter is a modification of the classical Kolmogorov spectrum for the inertial range of turbulence at low frequencies. TfBms have also been used to model geophysical flow data [6 ] and in finance [30 ] .
The theoretical aspects of tfBm have attracted considerable research interest (see [1 , 16 , 17 , 18 , 27 ] and references therein). In contrast, literature on their statistical aspect remains limited to a few papers [5 , 6 ] , in which wavelet estimation procedures were performed. Recently, least-squares estimators for the unknown drift parameters within Vasicek models driven by tfBms have been proposed and examined in [19 , 20 ] . This work addresses the parameter estimation of a tfBm with drift of the form.
X t = μ t + σ B H , λ ( t ) (1) X_t = \mu t + \sigma B_{H,\lambda}(t) \tag*{(1)} X t = μ t + σ B H , λ ( t ) (1)
for all t ≥ 0 t \ge0 t ≥ 0 , where B H , λ B_{H,\lambda} B H , λ is a tfBm as defined below. Using the ergodic theorem and the method of moments (e.g., [2 ] ) we derive suitable estimators for model (1). We investigate both the weak and almost sure behaviour of the proposed estimators. Consequently, our results provide alternatives to those given in [5 , 6 ] . When H ∈ ( 0 , 1 ) H \in(0,1) H ∈ ( 0 , 1 ) , we apply Gladyshev’s results to construct some alternative estimators of the noise parameters H H H , σ 2 \sigma^2 σ 2 , λ \lambda λ and prove their consistency. The evaluation of their limiting distributions requires more advanced tools and is beyond the scope of this work.
The rest of the paper is organized as follows. In Section 2 we introduce some basic and preliminary results. Then we discuss the procedures of estimation to be used. In Section 3 we construct μ ^ \widehat{\mu} μ the estimator of the drift μ ∈ R \mu\in\mathbb{R} μ ∈ R and investigate its asymptotic behaviour (independently of the other parameters). Section 4 is devoted to the joint estimation of the noise parameters ( σ 2 , H , λ ) ∈ ( 0 , ∞ ) 3 (\sigma^2,H,\lambda) \in(0,\infty)^3 ( σ 2 , H , λ ) ∈ ( 0 , ∞ ) 3 by the method of moments (e.g., [2 ] ). In Section 5 we use Gladyshev’s results [12 ] to propose a multistage estimation for the noise parameters H H H , σ 2 \sigma^2 σ 2 , λ \lambda λ under the condition 0 < H < 1 0 < H < 1 0 < H < 1 . In this section, the consistency of the estimators of H H H and σ 2 \sigma^2 σ 2 holds almost surely, while the estimator of λ \lambda λ is consistent in probability. Section 6 concludes the paper. Some auxiliary and technical lemmas on the modified Bessel function K H ( ⋅ ) K_H(\cdot) K H ( ⋅ ) with the Birkhoff’s ergodic theorem are gathered in the Appendix.
In what follows we work on a probability space ( Ω , F , P ) (\Omega,\mathcal{F},\mathbb{P}) ( Ω , F , P ) and each process should be understood as being defined with respect to its own natural filtration. We are concerned with four modes of convergence: almost surely P -a.s. ⟶ \mathbb{P}\text{-a.s.}\longrightarrow P -a.s. ⟶ , in the mean square sense L 2 ( Ω ) ⟶ L^2(\Omega)\longrightarrow L 2 ( Ω ) ⟶ , in probability P ⟶ \mathbb{P}\longrightarrow P ⟶ and in law L ⟶ \mathcal{L}\longrightarrow L ⟶ . The following symbols O ( ⋅ ) O(\cdot) O ( ⋅ ) and o ( ⋅ ) o(\cdot) o ( ⋅ ) denote the usual “big-oh” and “small-oh” notations, respectively.
Preliminary results
Let ( B ( t ) : t ∈ R ) (B(t):t \in\mathbb{R}) ( B ( t ) : t ∈ R ) be a two-sided Brownian motion, that is, a process with stationary independent increments such that B ( t ) B(t) B ( t ) has a Gaussian distribution with mean zero and variance ∣ t ∣ |t| ∣ t ∣ for all t ∈ R t \in\mathbb{R} t ∈ R . Define an independently scattered Gaussian random measure B ( d x ) B(dx) B ( d x ) with control measure m ( d x ) = d x m(dx)=dx m ( d x ) = d x by setting B ( [ a , b ] ) = B ( b ) − B ( a ) B([a,b])=B(b)-B(a) B ([ a , b ]) = B ( b ) − B ( a ) for any a , b ∈ R a,b \in\mathbb{R} a , b ∈ R with a < b a < b a < b and then extending to all Borel sets with finite Lebesgue measure. It is well known that the Wiener integral ∫ f ( x ) B ( d x ) \int f(x)B(dx) ∫ f ( x ) B ( d x ) exists whenever ∫ f ( x ) 2 d x < ∞ \int f(x)^2 dx < \infty ∫ f ( x ) 2 d x < ∞ . Many famous fractional processes were constructed by choosing a suitable integrand f ( ⋅ ) f(\cdot) f ( ⋅ ) . For instance, the fractional Brownian motion B H B_H B H was introduced by setting f ( x ) = C H ( ( t − x ) + H − 1 / 2 − ( − x ) + H − 1 / 2 ) f(x)=C_H\left((t-x)_+^{H-1/2}-(-x)_+^{H-1/2}\right) f ( x ) = C H ( ( t − x ) + H − 1/2 − ( − x ) + H − 1/2 ) , where C H > 0 C_H>0 C H > 0 is some normalizing constant. In [17 ] , the integrand in the form of B H B_H B H is exponentially tempered to get the tfBm
B H , λ ( t ) = ∫ [ e − λ ( t − x ) + ( t − x ) + H − 1 / 2 − e − λ ( − x ) + ( − x ) + H − 1 / 2 ] B ( d x ) B_{H,\lambda}(t)=\int\left[e^{-\lambda(t-x)_+}(t-x)_+^{H-1/2}-e^{-\lambda(-x)_+}(-x)_+^{H-1/2}\right]B(dx) B H , λ ( t ) = ∫ [ e − λ ( t − x ) + ( t − x ) + H − 1/2 − e − λ ( − x ) + ( − x ) + H − 1/2 ] B ( d x )
for all t ∈ R t \in\mathbb{R} t ∈ R . B H , λ B_{H,\lambda} B H , λ exhibits many interesting properties such as stationarity of increments and semi-long range dependence. It is worth mentioning that this process is not self-similar, but instead the following scaling property: B H , λ ( c t ) = d c H B H , c λ ( t ) B_{H,\lambda}(ct)\overset{\mathrm{d}}{=}c^H B_{H,c\lambda}(t) B H , λ ( c t ) = d c H B H , c λ ( t ) holds for any c > 0 c>0 c > 0 . This property implies
E [ B H , λ ( ∣ t ∣ ) 2 ] = t 2 H E [ B H , λ t ( 1 ) 2 ] = : t 2 H C t 2 , ⟶ 2 Γ ( 2 H ) / ( 2 λ ) 2 H (2) \mathbb{E}\left[B_{H,\lambda}(|t|)^2\right]=t^{2H}\mathbb{E}\left[B_{H,\lambda t}(1)^2\right]=:t^{2H}C_t^2,
\qquad
\longrightarrow2\Gamma(2H)/(2\lambda)^{2H} \tag*{(2)} E [ B H , λ ( ∣ t ∣ ) 2 ] = t 2 H E [ B H , λ t ( 1 ) 2 ] =: t 2 H C t 2 , ⟶ 2Γ ( 2 H ) / ( 2 λ ) 2 H (2)
as t → ∞ t \to\infty t → ∞ , where
C t 2 = 2 Γ ( 2 H ) ( 2 λ ∣ t ∣ ) 2 H − 2 Γ ( H + 1 / 2 ) π ( 2 λ ∣ t ∣ ) H K H ( λ ∣ t ∣ ) C_t^2 = \frac{2\Gamma(2H)}{(2\lambda|t|)^{2H}} - \frac{2\Gamma(H+1/2)}{\sqrt{\pi}(2\lambda|t|)^H}K_H(\lambda|t|) C t 2 = ( 2 λ ∣ t ∣ ) 2 H 2Γ ( 2 H ) − π ( 2 λ ∣ t ∣ ) H 2Γ ( H + 1/2 ) K H ( λ ∣ t ∣ )
and C 0 2 = 0 C_0^2=0 C 0 2 = 0 ,
where K H K_H K H denotes the modified Bessel function of second kind. The covariance of B H , λ B_{H,\lambda} B H , λ can be expressed as
Cov ( B H , λ ( t ) , B H , λ ( s ) ) = 1 2 ( ∣ t ∣ 2 H C t 2 + ∣ s ∣ 2 H C s 2 − ∣ t − s ∣ 2 H C t − s 2 ) (3) \operatorname{Cov}(B_{H,\lambda}(t),B_{H,\lambda}(s)) = \frac{1}{2}\left(|t|^{2H}C_t^2+|s|^{2H}C_s^2-|t-s|^{2H}C_{t-s}^2\right) \tag*{(3)} Cov ( B H , λ ( t ) , B H , λ ( s )) = 2 1 ( ∣ t ∣ 2 H C t 2 + ∣ s ∣ 2 H C s 2 − ∣ t − s ∣ 2 H C t − s 2 ) (3)
for all t , s ≥ 0 t,s\geq0 t , s ≥ 0 . When H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) , the tfBm B H , λ B_{H,\lambda} B H , λ becomes a quasi-hé lix process in the sense of [15 ] , and the Hölder continuity follows with H H H as critical exponent. For H > 1 H>1 H > 1 , the sample paths of B H , λ B_{H,\lambda} B H , λ are continuously differentiable. For more details on properties of the tfBm we refer to [17 , 1 ] .
Given a function g : R → R g:\mathbb{R}\to\mathbb{R} g : R → R , we define its k k k th order increments with step h ∈ R h\in\mathbb{R} h ∈ R as
Δ h ( k ) g ( t ) : = ∑ j = 0 k ( − 1 ) k − j ( k j ) g ( t + h j ) \Delta_h^{(k)}g(t):=\sum_{j=0}^{k}(-1)^{k-j}\binom{k}{j}g(t+hj) Δ h ( k ) g ( t ) := j = 0 ∑ k ( − 1 ) k − j ( j k ) g ( t + hj )
for all t ∈ R t\in\mathbb{R} t ∈ R and k ∈ Z + k\in\mathbb{Z}_+ k ∈ Z + with the convention Δ h ( 0 ) g ( t ) = g ( t ) \Delta_h^{(0)}g(t)=g(t) Δ h ( 0 ) g ( t ) = g ( t ) . We say that a process Y = ( Y ( t ) : t ∈ R ) Y=(Y(t):t\in\mathbb{R}) Y = ( Y ( t ) : t ∈ R ) is k k k -stationary if for every h ∈ R h\in\mathbb{R} h ∈ R the processes ( Δ h ( r ) Y ( t ) : t ∈ R ) (\Delta_h^{(r)}Y(t):t\in\mathbb{R}) ( Δ h ( r ) Y ( t ) : t ∈ R ) , r = 0 , … , ( k − 1 ) r=0,\ldots,(k-1) r = 0 , … , ( k − 1 ) are non-stationary, but ( Δ h ( k ) Y ( t ) : t ∈ R ) (\Delta_h^{(k)}Y(t):t\in\mathbb{R}) ( Δ h ( k ) Y ( t ) : t ∈ R ) exhibits stationarity. The process X X X defined in (1) is 1-stationary and stationarity holds also for any increments of higher order k ≥ 2 k\geq2 k ≥ 2 . Consider the following statistics to be used later
X ‾ N p , q ( h ) : = 1 N ∑ k = 0 N − 1 ( Δ h ( p ) X t ∣ t = k h ) q , N = 1 , 2 , … , (4) \overline{X}_N^{p,q}(h):=\frac{1}{N}\sum_{k=0}^{N-1}\left(\left.\Delta_h^{(p)}X_t\right|_{t=kh}\right)^q,\qquad N=1,2,\ldots, \tag*{(4)} X N p , q ( h ) := N 1 k = 0 ∑ N − 1 ( Δ h ( p ) X t t = k h ) q , N = 1 , 2 , … , (4)
for fixed h > 0 h>0 h > 0 and p , q ∈ Z + ∗ p,q\in\mathbb{Z}_+^* p , q ∈ Z + ∗ . Then the following results are obtained.
Lemma 1. We have
X ‾ N p , q ( h ) → P -a.s. E ( Δ h ( p ) X t ∣ t = 0 ) q = : X ‾ ∞ p , q ( h ) \overline{X}_N^{p,q}(h)\xrightarrow{\mathbb{P}\text{-a.s.}}\mathbb{E}\left(\left.\Delta_h^{(p)}X_t\right|_{t=0}\right)^q=:\overline{X}_\infty^{p,q}(h) X N p , q ( h ) P -a.s. E ( Δ h ( p ) X t t = 0 ) q =: X ∞ p , q ( h ) as N → ∞ N\to\infty N → ∞ . In particular,
X ‾ ∞ 1 , 1 ( h ) = μ h , X ‾ ∞ 1 , 2 ( h ) = μ 2 h 2 + σ 2 h 2 H C h 2 , X ‾ ∞ 2 , 2 ( h ) = σ 2 h 2 H ( 4 C h 2 − 2 2 H C 2 h 2 ) . (5) \begin{aligned}
\overline{X}_\infty^{1,1}(h)&=\mu h,\qquad\overline{X}_\infty^{1,2}(h)=\mu^2h^2+\sigma^2h^{2H}C_h^2,\\
\overline{X}_\infty^{2,2}(h)&=\sigma^2h^{2H}\left(4C_h^2-2^{2H}C_{2h}^2\right). \tag*{(5)}
\end{aligned} X ∞ 1 , 1 ( h ) X ∞ 2 , 2 ( h ) = μ h , X ∞ 1 , 2 ( h ) = μ 2 h 2 + σ 2 h 2 H C h 2 , = σ 2 h 2 H ( 4 C h 2 − 2 2 H C 2 h 2 ) . (5)
Proof. Let Z h ( t ) : = Δ h ( 1 ) X t Z_h(t):=\Delta_h^{(1)}X_t Z h ( t ) := Δ h ( 1 ) X t . Because of stationarity of Z h Z_h Z h , we have
Cov ( Z h ( s + τ ) , Z h ( s ) ) = Cov ( Z h ( τ ) , Z h ( 0 ) ) = σ 2 E { B H , λ ( h ) ( B H , λ ( τ + h ) − B H , λ ( τ ) ) } = σ 2 2 ( ∣ τ + h ∣ 2 H C τ + h 2 + ∣ τ − h ∣ 2 H C τ − h 2 − 2 ∣ τ ∣ 2 H C τ 2 ) , (6) \begin{aligned}
\operatorname{Cov}(Z_h(s+\tau),Z_h(s))&=\operatorname{Cov}(Z_h(\tau),Z_h(0))\\
&=\sigma^2\mathbb{E}\{B_{H,\lambda}(h)(B_{H,\lambda}(\tau+h)-B_{H,\lambda}(\tau))\}\\
&=\frac{\sigma^2}{2}\left(|\tau+h|^{2H}C_{\tau+h}^2+|\tau-h|^{2H}C_{\tau-h}^2-2|\tau|^{2H}C_\tau^2\right), \tag*{(6)}
\end{aligned} Cov ( Z h ( s + τ ) , Z h ( s )) = Cov ( Z h ( τ ) , Z h ( 0 )) = σ 2 E { B H , λ ( h ) ( B H , λ ( τ + h ) − B H , λ ( τ ))} = 2 σ 2 ( ∣ τ + h ∣ 2 H C τ + h 2 + ∣ τ − h ∣ 2 H C τ − h 2 − 2∣ τ ∣ 2 H C τ 2 ) , (6)
= σ 2 Γ ( H + 1 / 2 ) π 2 H λ 2 H [ 2 g H ( λ τ ) − g H ( λ ( τ + h ) ) − g H ( λ ( τ − h ) ) ] ⟶ 0 (7) \begin{aligned}
&=\frac{\sigma^2\Gamma(H+1/2)}{\sqrt{\pi}2^{H}\lambda^{2H}}
\left[2g_H(\lambda\tau)-g_H(\lambda(\tau+h))-g_H(\lambda(\tau-h))\right]\\
&\longrightarrow0 \tag*{(7)}
\end{aligned} = π 2 H λ 2 H σ 2 Γ ( H + 1/2 ) [ 2 g H ( λ τ ) − g H ( λ ( τ + h )) − g H ( λ ( τ − h )) ] ⟶ 0 (7)
as τ → ∞ \tau\to\infty τ → ∞ , where g H ( z ) = ∣ z ∣ H K H ( ∣ z ∣ ) g_H(z) = |z|^H K_H(|z|) g H ( z ) = ∣ z ∣ H K H ( ∣ z ∣ ) . The above convergence is justified by the fact that g H ( z ) ∼ π 2 z H − 1 2 e − z g_H(z) \sim\sqrt{\frac{\pi}{2}}z^{H-\frac{1}{2}}e^{-z} g H ( z ) ∼ 2 π z H − 2 1 e − z , as z → ∞ z \to\infty z → ∞ [10 (A.5)] . For each p ∈ Z + ∗ p \in\mathbb{Z}_+^* p ∈ Z + ∗ the covariance of Δ h ( p ) X \Delta_h^{(p)}X Δ h ( p ) X is continuous and satisfies
Cov ( Δ h ( p ) X ( s + τ ) , Δ h ( p ) X ( s ) ) = ∑ j , r = 0 p − 1 d j , r Cov ( Z h ( τ + h ( j − r ) ) , Z h ( 0 ) ) (8) \operatorname{Cov}\left(\Delta_h^{(p)}X(s+\tau),\Delta_h^{(p)}X(s)\right)=\sum_{j,r=0}^{p-1}d_{j,r}\operatorname{Cov}\left(Z_h(\tau+h(j-r)),Z_h(0)\right) \tag*{(8)} Cov ( Δ h ( p ) X ( s + τ ) , Δ h ( p ) X ( s ) ) = j , r = 0 ∑ p − 1 d j , r Cov ( Z h ( τ + h ( j − r )) , Z h ( 0 ) ) (8)
as τ → ∞ \tau\to\infty τ → ∞ , where d j , r = ( − 1 ) j + r ( p − 1 j ) ( p − 1 r ) d_{j,r}=(-1)^{j+r}\binom{p-1}{j}\binom{p-1}{r} d j , r = ( − 1 ) j + r ( j p − 1 ) ( r p − 1 ) . Combining the previous asymptotics with Gaussianity and stationarity of the sequence { Δ h ( p ) X t ∣ t = k h : k = 0 , 1 , 2 , … } \{\Delta_h^{(p)}X_t|_{t=kh}: k=0,1,2,\ldots\} { Δ h ( p ) X t ∣ t = k h : k = 0 , 1 , 2 , … } yields its ergodicity (e.g., [11 Theorem 5, p. 144] ). Hence, by Lemma 4 we can justify the first part of Lemma 1. The proof of (5) follows by elementary calculus with the use of (2) .
Lemma 1 allows us to construct a vector of statistics (4) (say J N ∈ R 4 J_N \in\mathbb{R}^4 J N ∈ R 4 ) that converges towards a function of the parameters (say F ( μ , σ , H , λ ) ∈ R 4 F(\mu,\sigma,H,\lambda) \in\mathbb{R}^4 F ( μ , σ , H , λ ) ∈ R 4 ). Hence, an appropriate estimator of ( μ , σ , H , λ ) (\mu,\sigma,H,\lambda) ( μ , σ , H , λ ) can be defined as F − 1 ( J N ) F^{-1}(J_N) F − 1 ( J N ) . This approach has been successfully applied to various mixed models (see [25 , 24 , 8 ] ). However, finding such an inverse is often difficult, if not impossible. To overcome this problem we estimate the parameters in two steps: First, we estimate the drift parameter μ \mu μ by X ‾ N 1 , 1 ( h ) / h \overline{X}_N^{1,1}(h)/h X N 1 , 1 ( h ) / h . This estimator can be seen as a discrete form of μ ^ t : = X t / t \widehat{\mu}_t := X_t/t μ t := X t / t . Second, we follow [2 ] to perform the joint estimation of ( σ 2 , H , λ ) (\sigma^2,H,\lambda) ( σ 2 , H , λ ) by the method of generalized moments. Note that a joint estimation of all the parameters is not possible by this method, as the identifiability condition (see [2 ] ) does not hold. Finally, we apply Gladyshev’s results on limit theorems for processes with Gaussian increments to X X X and derive consistent (alternative) estimators of the noise parameters denoted in the following order as H ~ n \widetilde{H}_n H n , σ ~ m , n 2 \widetilde{\sigma}^2_{m,n} σ m , n 2 and λ ~ m , n , h \widetilde{\lambda}_{m,n,h} λ m , n , h . This order is crucial for the implementation issue. In fact, we can write the latter two estimators as σ ~ m , n 2 = F m ( H ~ n , X ) \widetilde{\sigma}^2_{m,n}=F_m(\widetilde{H}_n,X) σ m , n 2 = F m ( H n , X ) and λ ~ m , n , h = G h ( H ~ n , σ ~ m , n 2 , X ) \widetilde{\lambda}_{m,n,h}=G_h(\widetilde{H}_n,\widetilde{\sigma}^2_{m,n},X) λ m , n , h = G h ( H n , σ m , n 2 , X ) for some continuous functions F m F_m F m and G h G_h G h .
Estimation of the drift parameter $\mu`
Theorem 1. For a fixed h > 0 h>0 h > 0 , the following statistic μ ^ : = X ‾ N 1 , 1 ( h ) / h \widehat{\mu}:=\overline{X}_N^{1,1}(h)/h μ := X N 1 , 1 ( h ) / h converges towards μ \mu μ , almost surely and in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , as N → ∞ N\to\infty N → ∞ . Furthermore, μ ^ \widehat{\mu} μ is unbiased and satisfies the asymptotic normality i.e.,
N ( μ ^ − μ ) → L N ( 0 , 2 σ 2 Γ ( 2 H ) h 2 ( 2 λ ) 2 H ) , as N → ∞ (9) N(\widehat{\mu}-\mu)\xrightarrow{\mathcal{L}}\mathcal{N}\left(0,\frac{2\sigma^2\Gamma(2H)}{h^2(2\lambda)^{2H}}\right),\quad\text{as }N\to\infty \tag*{(9)} N ( μ − μ ) L N ( 0 , h 2 ( 2 λ ) 2 H 2 σ 2 Γ ( 2 H ) ) , as N → ∞ (9)
Proof. The unbiasedness of μ ^ \widehat{\mu} μ is trivial and its strong consistency follows immediately from Lemma 1. Let us evaluate the variance of μ ^ \widehat{\mu} μ . Consider Z h ( ⋅ ) Z_h(\cdot) Z h ( ⋅ ) as defined in the proof of Lemma 1 and set ϕ h ( τ ) : = ∣ h τ ∣ 2 H C h τ 2 \phi_h(\tau) := |h\tau|^{2H}C_{h\tau}^2 ϕ h ( τ ) := ∣ h τ ∣ 2 H C h τ 2 . In view of (6), we have
N 2 h 2 Var ( μ ^ ) = Var ( ∑ k = 0 N − 1 Z h ( h k ) ) = Cov ( ∑ k = 0 N − 1 Z h ( h k ) , ∑ j = 0 N − 1 Z h ( h j ) ) = E [ ∑ k = 0 N − 1 Z h ( h k ) ∑ j = 0 N − 1 Z h ( h j ) ] − ( E ∑ k = 0 N − 1 Z h ( h k ) ) ( E ∑ j = 0 N − 1 Z h ( h j ) ) = ∑ k , j = 0 N − 1 Cov ( Z h ( h ( k − j ) ) , Z h ( 0 ) ) , (By stationarity of Z h ), = σ 2 2 ∑ k , j = 0 N − 1 [ ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ] , (By (6) ), = σ 2 2 { ∑ k = j [ ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ] + ∑ k ≠ j [ ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ] } = σ 2 { ∑ k = 0 N − 1 ( ϕ h ( 1 ) − ϕ h ( 0 ) ) + ∑ j < k ( ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ) } = σ 2 ( N ϕ h ( 1 ) + ∑ k = 1 N − 1 ∑ r = 1 k ( ϕ h ( r + 1 ) − ϕ h ( r ) ) + ∑ k = 1 N − 1 ∑ r = 1 k ( ϕ h ( r − 1 ) − ϕ h ( r ) ) ) = σ 2 ( N ϕ h ( 1 ) + ∑ k = 1 N − 1 ( ϕ h ( k + 1 ) − ϕ h ( 1 ) ) + ∑ k = 1 N − 1 ( ϕ h ( 0 ) − ϕ h ( k ) ) ) = σ 2 ( N ϕ h ( 1 ) − ( N − 1 ) ϕ h ( 1 ) + ∑ k = 1 N − 1 ( ϕ h ( k + 1 ) − ϕ h ( k ) ) ) = σ 2 ∣ h N ∣ 2 H C h N 2 = σ 2 ( 2 Γ ( 2 H ) ( 2 λ ) 2 H − 2 Γ ( H + 1 / 2 ) π λ 2 H g H ( λ h N ) ) . \begin{aligned}
N^2h^2\operatorname{Var}(\hat{\mu}) &= \operatorname{Var}\left(\sum_{k=0}^{N-1} Z_h(hk)\right) = \operatorname{Cov}\left(\sum_{k=0}^{N-1} Z_h(hk),\sum_{j=0}^{N-1} Z_h(hj)\right) \\
&= \mathbb{E}\left[\sum_{k=0}^{N-1} Z_h(hk)\sum_{j=0}^{N-1} Z_h(hj)\right] - \left(\mathbb{E}\sum_{k=0}^{N-1} Z_h(hk)\right)\left(\mathbb{E}\sum_{j=0}^{N-1} Z_h(hj)\right) \\
&= \sum_{k,j=0}^{N-1}\operatorname{Cov}\left(Z_h(h(k-j)),Z_h(0)\right), \qquad\text{(By stationarity of $Z_h$),} \\
&= \frac{\sigma^2}{2}\sum_{k,j=0}^{N-1}\left[\phi_h(k-j+1)+\phi_h(k-j-1)-2\phi_h(k-j)\right], \qquad\text{(By (6) ),} \\
&= \frac{\sigma^2}{2}\left\{\sum_{k=j}\left[\phi_h(k-j+1)+\phi_h(k-j-1)-2\phi_h(k-j)\right]\right.\\
&\qquad\left. + \sum_{k\ne j}\left[\phi_h(k-j+1)+\phi_h(k-j-1)-2\phi_h(k-j)\right]\right\} \\
&= \sigma^2\left\{\sum_{k=0}^{N-1}\left(\phi_h(1)-\phi_h(0)\right)+\sum_{j<k}\left(\phi_h(k-j+1)+\phi_h(k-j-1)-2\phi_h(k-j)\right)\right\} \\
&= \sigma^2\left(N\phi_h(1)+\sum_{k=1}^{N-1}\sum_{r=1}^{k}\left(\phi_h(r+1)-\phi_h(r)\right)+\sum_{k=1}^{N-1}\sum_{r=1}^{k}\left(\phi_h(r-1)-\phi_h(r)\right)\right) \\
&= \sigma^2\left(N\phi_h(1)+\sum_{k=1}^{N-1}\left(\phi_h(k+1)-\phi_h(1)\right)+\sum_{k=1}^{N-1}\left(\phi_h(0)-\phi_h(k)\right)\right) \\
&= \sigma^2\left(N\phi_h(1)-(N-1)\phi_h(1)+\sum_{k=1}^{N-1}\left(\phi_h(k+1)-\phi_h(k)\right)\right) \\
&= \sigma^2|hN|^{2H}C_{hN}^2 = \sigma^2\left(\frac{2\Gamma(2H)}{(2\lambda)^{2H}}-\frac{2\Gamma(H+1/2)}{\sqrt{\pi} \lambda^{2H}}g_H(\lambda hN)\right).
\end{aligned} N 2 h 2 Var ( μ ^ ) = Var ( k = 0 ∑ N − 1 Z h ( hk ) ) = Cov ( k = 0 ∑ N − 1 Z h ( hk ) , j = 0 ∑ N − 1 Z h ( hj ) ) = E [ k = 0 ∑ N − 1 Z h ( hk ) j = 0 ∑ N − 1 Z h ( hj ) ] − ( E k = 0 ∑ N − 1 Z h ( hk ) ) ( E j = 0 ∑ N − 1 Z h ( hj ) ) = k , j = 0 ∑ N − 1 Cov ( Z h ( h ( k − j )) , Z h ( 0 ) ) , (By stationarity of Z h ), = 2 σ 2 k , j = 0 ∑ N − 1 [ ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ] , (By (6) ), = 2 σ 2 ⎩ ⎨ ⎧ k = j ∑ [ ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ] + k = j ∑ [ ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ] ⎭ ⎬ ⎫ = σ 2 ⎩ ⎨ ⎧ k = 0 ∑ N − 1 ( ϕ h ( 1 ) − ϕ h ( 0 ) ) + j < k ∑ ( ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ) ⎭ ⎬ ⎫ = σ 2 ( N ϕ h ( 1 ) + k = 1 ∑ N − 1 r = 1 ∑ k ( ϕ h ( r + 1 ) − ϕ h ( r ) ) + k = 1 ∑ N − 1 r = 1 ∑ k ( ϕ h ( r − 1 ) − ϕ h ( r ) ) ) = σ 2 ( N ϕ h ( 1 ) + k = 1 ∑ N − 1 ( ϕ h ( k + 1 ) − ϕ h ( 1 ) ) + k = 1 ∑ N − 1 ( ϕ h ( 0 ) − ϕ h ( k ) ) ) = σ 2 ( N ϕ h ( 1 ) − ( N − 1 ) ϕ h ( 1 ) + k = 1 ∑ N − 1 ( ϕ h ( k + 1 ) − ϕ h ( k ) ) ) = σ 2 ∣ h N ∣ 2 H C h N 2 = σ 2 ( ( 2 λ ) 2 H 2Γ ( 2 H ) − π λ 2 H 2Γ ( H + 1/2 ) g H ( λh N ) ) .
At this stage we can assert that Var ( h N ( μ ^ − μ ) ) ∼ 2 σ 2 Γ ( 2 H ) / ( 2 λ ) 2 H \operatorname{Var}\left(hN(\hat{\mu}-\mu)\right)\sim2\sigma^2\Gamma(2H)/(2\lambda)^{2H} Var ( h N ( μ ^ − μ ) ) ∼ 2 σ 2 Γ ( 2 H ) / ( 2 λ ) 2 H , as N → ∞ N\to\infty N → ∞ . This implies the consistency of μ ^ \hat{\mu} μ ^ in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) . Now, due to Gaussianity of μ ^ \hat{\mu} μ ^ , equation (9) holds, and the proof is complete. □
Estimation of the noise parameters
Consider the process Y h = { Δ h ( 2 ) X t : t ≥ 0 } Y_h=\{\Delta_h^{(2)}X_t:t\ge0\} Y h = { Δ h ( 2 ) X t : t ≥ 0 } with h > 0 h>0 h > 0 to be specified later. Clearly, this process is stationary Gaussian with zero mean. It admits the following harmonizable representation (e.g., [17 (3.1)]
Y h ( t ) = σ ( B H , λ ( t + 2 h ) − 2 B H , λ ( t + h ) + B H , λ ( t ) ) = σ Γ ( H + 1 / 2 ) 2 π ∫ R e − i 2 ω h − 2 e − i ω h + 1 ( λ − i ω ) H + 1 / 2 e − i ω t B ~ ( d ω ) , ( i = − 1 ) , \begin{aligned}
Y_h(t)=\sigma\left(B_{H,\lambda}(t+2h)-2B_{H,\lambda}(t+h)+B_{H,\lambda}(t)\right) \\
&=\frac{\sigma\Gamma(H+1/2)}{\sqrt{2\pi}}\int_{\mathbb{R}}\frac{e^{-i2\omega h}-2e^{-i\omega h}+1}{(\lambda-i\omega)^{H+1/2}}e^{-i\omega t}\widetilde{B}(d\omega),\qquad(i=\sqrt{-1}),
\end{aligned} Y h ( t ) = σ ( B H , λ ( t + 2 h ) − 2 B H , λ ( t + h ) + B H , λ ( t ) ) = 2 π σ Γ ( H + 1/2 ) ∫ R ( λ − iω ) H + 1/2 e − i 2 ω h − 2 e − iω h + 1 e − iω t B ( d ω ) , ( i = − 1 ) ,
where B ~ ( d ω ) \widetilde{B}(d\omega) B ( d ω ) is a complex-valued Gaussian random measure satisfying B ~ ( d ω ) = B ~ ( − d ω ) \widetilde{B}(d\omega)=\widetilde{B}(-d\omega) B ( d ω ) = B ( − d ω ) and E ∣ B ~ ( d ω ) ∣ 2 = d ω \mathbb{E}|\widetilde{B}(d\omega)|^2=d\omega E ∣ B ( d ω ) ∣ 2 = d ω . We derive the spectral density function of Y h Y_h Y h [29 §. 9] as
f θ ( ω ) = 4 σ 2 Γ ( H + 1 / 2 ) 2 ( 1 − cos ( ω h ) ) 2 2 π ( λ 2 + ω 2 ) H + 1 / 2 , θ = ( σ 2 , H , λ ) . (10) f_\theta(\omega)=\frac{4\sigma^2\Gamma(H+1/2)^2(1-\cos(\omega h))^2}{2\pi(\lambda^2+\omega^2)^{H+1/2}},\qquad\theta=(\sigma^2,H,\lambda). \tag*{(10)} f θ ( ω ) = 2 π ( λ 2 + ω 2 ) H + 1/2 4 σ 2 Γ ( H + 1/2 ) 2 ( 1 − cos ( ω h ) ) 2 , θ = ( σ 2 , H , λ ) . (10)
Its autocovariance structure is described by
γ θ ( τ ) : = Cov ( Y h ( τ ) , Y h ( 0 ) ) = σ 2 2 { − ϕ ( τ + 2 h ) − ϕ ( τ − 2 h ) + 4 ϕ ( τ + h ) + 4 ϕ ( τ − h ) − 6 ϕ ( τ ) } (11) \begin{aligned}
\gamma_\theta(\tau):=\operatorname{Cov}(Y_h(\tau),Y_h(0)) \\
&=\frac{\sigma^2}{2}\left\{-\phi(\tau+2h)-\phi(\tau-2h)+4\phi(\tau+h)+4\phi(\tau-h)-6\phi(\tau)\right\} \tag*{(11)}
\end{aligned} γ θ ( τ ) := Cov ( Y h ( τ ) , Y h ( 0 )) = 2 σ 2 { − ϕ ( τ + 2 h ) − ϕ ( τ − 2 h ) + 4 ϕ ( τ + h ) + 4 ϕ ( τ − h ) − 6 ϕ ( τ ) } (11)
= c { g H ( λ ( τ + 2 h ) ) − 4 g H ( λ ( τ + h ) ) + 6 g H ( λ τ ) − 4 g H ( λ ( τ − h ) ) + g H ( λ ( τ − 2 h ) ) } , (12) =c\left\{g_H(\lambda(\tau+2h))-4g_H(\lambda(\tau+h))+6g_H(\lambda\tau)-4g_H(\lambda(\tau-h))+g_H(\lambda(\tau-2h))\right\}, \tag*{(12)} = c { g H ( λ ( τ + 2 h )) − 4 g H ( λ ( τ + h )) + 6 g H ( λ τ ) − 4 g H ( λ ( τ − h )) + g H ( λ ( τ − 2 h )) } , (12)
where ϕ ( τ ) = ∣ τ ∣ 2 H C τ 2 \phi(\tau)=|\tau|^{2H}C_\tau^2 ϕ ( τ ) = ∣ τ ∣ 2 H C τ 2 and
c = σ 2 Γ ( H + 1 / 2 ) π 2 2 H λ 2 H = : σ 2 κ and g H ( z ) = ∣ z ∣ H K H ( ∣ z ∣ ) . (13) c=\frac{\sigma^2\Gamma(H+1/2)}{\sqrt{\pi}2^{2H}\lambda^{2H}}=:\sigma^2\kappa\qquad\text{and}\qquad g_H(z)=|z|^H K_H(|z|). \tag*{(13)} c = π 2 2 H λ 2 H σ 2 Γ ( H + 1/2 ) =: σ 2 κ and g H ( z ) = ∣ z ∣ H K H ( ∣ z ∣ ) . (13)
Both (11) and (12)-(13) are of great importance. The latter is valid if all the arguments used in g H ( ⋅ ) g_H(\cdot) g H ( ⋅ ) are strictly positive. For (12) to be valid for any argument including the origin we shall add the convention: g H ( 0 ) = 2 H − 1 Γ ( H ) g_H(0)=2^{H-1}\Gamma(H) g H ( 0 ) = 2 H − 1 Γ ( H ) . Now, Instead of (3) the covariance can be rewritten as
Cov ( B H , λ ( t ) , B H , λ ( s ) ) = 1 2 { c 1 + c 2 [ g H ( λ ( t − s ) ) − g H ( λ t ) − g H ( λ s ) ] } , ∀ t , s ∈ R , \operatorname{Cov}(B_{H,\lambda}(t),B_{H,\lambda}(s))=\frac{1}{2}\left\{c_1+c_2\left[g_H(\lambda(t-s))-g_H(\lambda t)-g_H(\lambda s)\right]\right\},\qquad\forall t,s\in\mathbb{R}, Cov ( B H , λ ( t ) , B H , λ ( s )) = 2 1 { c 1 + c 2 [ g H ( λ ( t − s )) − g H ( λ t ) − g H ( λ s ) ] } , ∀ t , s ∈ R ,
where
c 1 = 2 Γ ( 2 H ) ( 2 λ ) 2 H and c 2 = 2 Γ ( H + 1 / 2 ) π 2 2 H λ 2 H . c_1=\frac{2\Gamma(2H)}{(2\lambda)^{2H}}\qquad\text{and}\qquad c_2=\frac{2\Gamma(H+1/2)}{\sqrt{\pi}2^{2H}\lambda^{2H}}. c 1 = ( 2 λ ) 2 H 2Γ ( 2 H ) and c 2 = π 2 2 H λ 2 H 2Γ ( H + 1/2 ) .
In particular, E ∣ B H , λ ( t ) ∣ 2 = c 1 − c 2 g H ( λ t ) \mathbb{E}|B_{H,\lambda}(t)|^2=c_1-c_2g_H(\lambda t) E ∣ B H , λ ( t ) ∣ 2 = c 1 − c 2 g H ( λ t ) , ∀ t ∈ R \forall t\in\mathbb{R} ∀ t ∈ R . Note that t = s = 0 t=s=0 t = s = 0 implies E ∣ B H , λ ( 0 ) ∣ 2 = c 1 − c 2 g H ( 0 ) = 0 \mathbb{E}|B_{H,\lambda}(0)|^2=c_1-c_2g_H(0)=0 E ∣ B H , λ ( 0 ) ∣ 2 = c 1 − c 2 g H ( 0 ) = 0 , which shows that B H , λ B_{H,\lambda} B H , λ starts from the origin. Note that ϕ ( 0 ) = ∣ 0 ∣ 2 H C 0 2 = 0 \phi(0)=|0|^{2H}C_0^2=0 ϕ ( 0 ) = ∣0 ∣ 2 H C 0 2 = 0 but g H ( 0 ) ≠ 0 g_H(0)\ne0 g H ( 0 ) = 0 . Let α > 0 \alpha>0 α > 0 (to be specified later) and set ϱ θ , 3 ( α ) : = ( γ θ ( 0 ) , γ θ ( α ) , γ θ ( 2 α ) ) ′ \varrho_{\theta,3}(\alpha):=(\gamma_\theta(0),\gamma_\theta(\alpha),\gamma_\theta(2\alpha))^\prime ϱ θ , 3 ( α ) := ( γ θ ( 0 ) , γ θ ( α ) , γ θ ( 2 α ) ) ′ . As the step h h h is arbitrary we simplify our computations by setting h = α h=\alpha h = α to obtain
ϱ θ , 3 ( α ) = c ( 2 { g H ( 2 λ α ) − 4 g H ( λ α ) } + 6 g H ( 0 ) g H ( 3 λ α ) − 4 g H ( 2 λ α ) + 7 g H ( λ α ) − 4 g H ( 0 ) g H ( 4 λ α ) − 4 g H ( 3 λ α ) + 6 g H ( 2 λ α ) − 4 g H ( λ α ) + g H ( 0 ) ) , θ = ( σ 2 , H , λ ) , \varrho_{\theta,3}(\alpha)=c\begin{pmatrix}
2\{g_H(2\lambda\alpha)-4g_H(\lambda\alpha)\}+6g_H(0)\\
g_H(3\lambda\alpha)-4g_H(2\lambda\alpha)+7g_H(\lambda\alpha)-4g_H(0)\\
g_H(4\lambda\alpha)-4g_H(3\lambda\alpha)+6g_H(2\lambda\alpha)-4g_H(\lambda\alpha)+g_H(0)
\end{pmatrix},\qquad\theta=(\sigma^2,H,\lambda), ϱ θ , 3 ( α ) = c 2 { g H ( 2 λ α ) − 4 g H ( λ α )} + 6 g H ( 0 ) g H ( 3 λ α ) − 4 g H ( 2 λ α ) + 7 g H ( λ α ) − 4 g H ( 0 ) g H ( 4 λ α ) − 4 g H ( 3 λ α ) + 6 g H ( 2 λ α ) − 4 g H ( λ α ) + g H ( 0 ) , θ = ( σ 2 , H , λ ) ,
with the Jacobian matrix with respect to θ \theta θ given as
∇ ϱ θ , 3 ( α ) = ( Σ 1 , 1 Σ 1 , 2 Σ 1 , 3 Σ 2 , 1 Σ 2 , 2 Σ 2 , 3 Σ 3 , 1 Σ 3 , 2 Σ 3 , 3 ) . (14) \nabla\varrho_{\theta,3}(\alpha)=
\begin{pmatrix}
\Sigma_{1,1} & \Sigma_{1,2} & \Sigma_{1,3}\\
\Sigma_{2,1} & \Sigma_{2,2} & \Sigma_{2,3}\\
\Sigma_{3,1} & \Sigma_{3,2} & \Sigma_{3,3}
\end{pmatrix}. \tag*{(14)} ∇ ϱ θ , 3 ( α ) = Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 Σ 1 , 2 Σ 2 , 2 Σ 3 , 2 Σ 1 , 3 Σ 2 , 3 Σ 3 , 3 . (14)
and
Σ 1 , 1 = 2 κ ( y − 4 x + 3 u 0 ) Σ 2 , 1 = κ ( z − 4 y + 7 x − 4 u 0 ) Σ 3 , 1 = κ ( t − 4 z + 6 y − 4 x + u 0 ) (15) \begin{aligned}
\Sigma_{1,1} &= 2\kappa(y-4x+3u_0)\\
\Sigma_{2,1} &= \kappa(z-4y+7x-4u_0)\\
\Sigma_{3,1} &= \kappa(t-4z+6y-4x+u_0) \tag*{(15)}
\end{aligned} Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 = 2 κ ( y − 4 x + 3 u 0 ) = κ ( z − 4 y + 7 x − 4 u 0 ) = κ ( t − 4 z + 6 y − 4 x + u 0 ) (15)
Σ 1 , 2 = 2 κ σ 2 [ ξ ( y − 4 x + 3 u 0 ) + ( B y − 4 A x + 3 u ∗ ) ] Σ 2 , 2 = κ σ 2 [ ξ ( z − 4 y + 7 x − 4 u 0 ) + ( C z − 4 B y + 7 A x − 4 u ∗ ) ] Σ 3 , 2 = κ σ 2 [ ξ ( t − 4 z + 6 y − 4 x + u 0 ) + ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) ] (16) \begin{aligned}
\Sigma_{1,2} &= 2\kappa\sigma^2\left[\xi(y-4x+3u_0)+(By-4Ax+3u^*)\right]\\
\Sigma_{2,2} &= \kappa\sigma^2\left[\xi(z-4y+7x-4u_0)+(Cz-4By+7Ax-4u^*)\right]\\
\Sigma_{3,2} &= \kappa\sigma^2\left[\xi(t-4z+6y-4x+u_0)+(Dt-4Cz+6By-4Ax+u^*)\right] \tag*{(16)}
\end{aligned} Σ 1 , 2 Σ 2 , 2 Σ 3 , 2 = 2 κ σ 2 [ ξ ( y − 4 x + 3 u 0 ) + ( B y − 4 A x + 3 u ∗ ) ] = κ σ 2 [ ξ ( z − 4 y + 7 x − 4 u 0 ) + ( C z − 4 B y + 7 A x − 4 u ∗ ) ] = κ σ 2 [ ξ ( t − 4 z + 6 y − 4 x + u 0 ) + ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) ] (16)
Σ 1 , 3 = 2 κ σ 2 [ ξ ( y − 4 x + 3 u 0 ) − 2 α ( b y − 2 a x ) ] Σ 2 , 3 = κ σ 2 [ ξ ( z − 4 y + 7 x − 4 u 0 ) − α ( 3 c z − 8 b y + 7 a x ) ] Σ 3 , 3 = κ σ 2 [ ξ ( t − 4 z + 6 y − 4 x + u 0 ) − 4 α ( d t − 3 c z + 3 b y − a x ) ] (17) \begin{aligned}
\Sigma_{1,3} &= 2\kappa\sigma^2\left[\xi(y-4x+3u_0)-2\alpha(by-2ax)\right]\\
\Sigma_{2,3} &= \kappa\sigma^2\left[\xi(z-4y+7x-4u_0)-\alpha(3cz-8by+7ax)\right]\\
\Sigma_{3,3} &= \kappa\sigma^2\left[\xi(t-4z+6y-4x+u_0)-4\alpha(dt-3cz+3by-ax)\right] \tag*{(17)}
\end{aligned} Σ 1 , 3 Σ 2 , 3 Σ 3 , 3 = 2 κ σ 2 [ ξ ( y − 4 x + 3 u 0 ) − 2 α ( b y − 2 a x ) ] = κ σ 2 [ ξ ( z − 4 y + 7 x − 4 u 0 ) − α ( 3 cz − 8 b y + 7 a x ) ] = κ σ 2 [ ξ ( t − 4 z + 6 y − 4 x + u 0 ) − 4 α ( d t − 3 cz + 3 b y − a x ) ] (17)
where ξ = − log ( 2 λ 2 ) + ψ ( H + 1 / 2 ) \xi=-\log(2\lambda^2)+\psi(H+1/2) ξ = − log ( 2 λ 2 ) + ψ ( H + 1/2 ) , ζ = − 2 H / λ \zeta=-2H/\lambda ζ = − 2 H / λ , u 0 = g H ( 0 ) u_0=g_H(0) u 0 = g H ( 0 ) , u ∗ = ∂ H g H ( 0 ) = ( log ( 2 ) + ψ ( H ) ) u 0 u^*=\partial_Hg_H(0)=(\log(2)+\psi(H))u_0 u ∗ = ∂ H g H ( 0 ) = ( log ( 2 ) + ψ ( H )) u 0 and
{ x = g H ( λ α ) , a = r H ( λ α ) , A = R H ( λ α ) y = g H ( 2 λ α ) , b = r H ( 2 λ α ) , B = R H ( 2 λ α ) z = g H ( 3 λ α ) , c = r H ( 3 λ α ) , C = R H ( 3 λ α ) t = g H ( 4 λ α ) , d = r H ( 4 λ α ) , D = R H ( 4 λ α ) , (18) \left\{
\begin{aligned}
x &= g_H(\lambda\alpha),\quad a=r_H(\lambda\alpha),\quad A=R_H(\lambda\alpha)\\
y &= g_H(2\lambda\alpha),\quad b=r_H(2\lambda\alpha),\quad B=R_H(2\lambda\alpha)\\
z &= g_H(3\lambda\alpha),\quad c=r_H(3\lambda\alpha),\quad C=R_H(3\lambda\alpha)\\
t &= g_H(4\lambda\alpha),\quad d=r_H(4\lambda\alpha),\quad D=R_H(4\lambda\alpha),
\end{aligned}
\right. \tag*{(18)} ⎩ ⎨ ⎧ x y z t = g H ( λ α ) , a = r H ( λ α ) , A = R H ( λ α ) = g H ( 2 λ α ) , b = r H ( 2 λ α ) , B = R H ( 2 λ α ) = g H ( 3 λ α ) , c = r H ( 3 λ α ) , C = R H ( 3 λ α ) = g H ( 4 λ α ) , d = r H ( 4 λ α ) , D = R H ( 4 λ α ) , (18)
with r H ( ⋅ ) r_H(\cdot) r H ( ⋅ ) and R H ( ⋅ ) R_H(\cdot) R H ( ⋅ ) being functions as defined in Lemma 3 and ψ ( ⋅ ) \psi(\cdot) ψ ( ⋅ ) stands for the digamma function. The next step is to introduce the concept of filters. For a given l , L ∈ Z + l,L\in\mathbb{Z}_+ l , L ∈ Z + a filter of length L + 1 L+1 L + 1 and order l l l is a sequence of reals A = ( a 0 , a 1 , … , a L ) \mathcal{A}=(a_0,a_1,\ldots,a_L) A = ( a 0 , a 1 , … , a L ) satisfying
{ ∑ q = 0 L a q q r = 0 , for all r = 0 , 1 , … , ( l − 1 ) ∑ q = 0 L a q q l ≠ 0 , \left\{
\begin{aligned}
\sum_{q=0}^{L}a_qq^r &= 0,\quad\text{for all }r=0,1,\ldots,(l-1)\\
\sum_{q=0}^{L}a_qq^l &\ne0,
\end{aligned}
\right. ⎩ ⎨ ⎧ q = 0 ∑ L a q q r q = 0 ∑ L a q q l = 0 , for all r = 0 , 1 , … , ( l − 1 ) = 0 ,
with the convention 0 0 = 1 0^0=1 0 0 = 1 . The filter with zero order (l = 0 l=0 l = 0 ) is defined by a 0 = 1 a_0=1 a 0 = 1 and a q = 0 a_q=0 a q = 0 for all 0 < q ≤ L 0<q\le L 0 < q ≤ L . Let A j \mathcal{A}_j A j , j = 1 , 2 , 3 j=1,2,3 j = 1 , 2 , 3 be three filters of length 4 defined as
A j : = ( A j ( 0 ) , A j ( 1 ) , A j ( 2 ) , A j ( 3 ) ) . \mathcal{A}_j := \left(\mathcal{A}_j(0),\mathcal{A}_j(1),\mathcal{A}_j(2),\mathcal{A}_j(3)\right). A j := ( A j ( 0 ) , A j ( 1 ) , A j ( 2 ) , A j ( 3 ) ) .
For each filter A j \mathcal{A}_j A j we associate a vector B j = ( B j ( 0 ) , B j ( 1 ) , B j ( 2 ) , B j ( 3 ) ) \mathbb{B}_j=(\mathbb{B}_j(0),\mathbb{B}_j(1),\mathbb{B}_j(2),\mathbb{B}_j(3)) B j = ( B j ( 0 ) , B j ( 1 ) , B j ( 2 ) , B j ( 3 )) , where
{ B j ( 0 ) = ∑ q = 0 3 A j 2 ( q ) B j ( k ) = 2 ∑ q = 0 3 − k A j ( k + q ) A j ( q ) \left\{
\begin{aligned}
\mathbb{B}_j(0)&=\sum_{q=0}^{3}\mathcal{A}_j^2(q)\\
\mathbb{B}_j(k)&=2\sum_{q=0}^{3-k}\mathcal{A}_j(k+q)\mathcal{A}_j(q)
\end{aligned}
\right. ⎩ ⎨ ⎧ B j ( 0 ) B j ( k ) = q = 0 ∑ 3 A j 2 ( q ) = 2 q = 0 ∑ 3 − k A j ( k + q ) A j ( q )
for all k = 1 , 2 , 3 k=1,2,3 k = 1 , 2 , 3 . For practical purposes we choose A 1 \mathcal{A}_1 A 1 to be of order one, and the others A 2 , A 3 \mathcal{A}_2,\mathcal{A}_3 A 2 , A 3 to be of order two. Their numerical values are
\begin{pmatrix}
\mathcal{A}_1\\
\mathcal{A}_2\\
\mathcal{A}_3
\end{pmatrix}
=
\begin{pmatrix}
0 & 1 & 0 & -1\\
0 & 1 & -2 & 1\\
1 & -1 & -1 & 1
\end{pmatrix}. \tag*{(19)}\end{pmatrix}
After some elementary computations, we obtain the following associated 3 × 4 3 \times4 3 × 4 -matrix
B = ( B 1 B 2 B 3 ) = ( 2 0 − 2 0 6 − 8 2 0 4 − 2 − 4 2 ) (20) \mathbb{B}=\begin{pmatrix}\mathbb{B}_1\\ \mathbb{B}_2\\ \mathbb{B}_3\end{pmatrix}=\begin{pmatrix}2&0&-2&0\\6&-8&2&0\\4&-2&-4&2\end{pmatrix} \tag*{(20)} B = B 1 B 2 B 3 = 2 6 4 0 − 8 − 2 − 2 2 − 4 0 0 2 (20)
Let us assume that observations of the model (1) at equidistant points t k = α k t_k=\alpha k t k = α k , k = 0 , 1 , 2 , … k=0,1,2,\ldots k = 0 , 1 , 2 , … are available. Using the method of generalized moments as proposed in [2 ] , we define our estimator θ ^ = ( σ 2 ^ , H ^ , λ ^ ) \widehat{\theta}=(\widehat{\sigma^2},\widehat{H},\widehat{\lambda}) θ = ( σ 2 , H , λ ) by minimizing the contrast function
Q ^ N ( θ ) = ∑ l = 1 3 ( 1 N − 2 ∑ j = 3 N g l ( t j , θ ) ) 2 , θ = ( σ 2 , H , λ ) , \widehat{Q}_N(\theta)=\sum_{l=1}^{3}\left(\frac{1}{N-2}\sum_{j=3}^{N}g_l(t_j,\theta)\right)^2,\qquad\theta=(\sigma^2,H,\lambda), Q N ( θ ) = l = 1 ∑ 3 ( N − 2 1 j = 3 ∑ N g l ( t j , θ ) ) 2 , θ = ( σ 2 , H , λ ) ,
where g l ( t , θ ) = φ l 2 ( t ) − V l ( θ ) g_l(t,\theta)=\varphi_l^2(t)-V_l(\theta) g l ( t , θ ) = φ l 2 ( t ) − V l ( θ ) and
{ φ l ( t ) : = ∑ q = 0 3 A l ( q ) Y α ( t − α q ) , for all t ≥ 0 , V l ( θ ) : = ∑ q = 0 3 B l ( q ) γ θ ( α q ) . \begin{cases}
\varphi_l(t):=\sum_{q=0}^{3}A_l(q)Y_\alpha(t-\alpha q),&\text{for all }t\ge0,\\
V_l(\theta):=\sum_{q=0}^{3}\mathbb{B}_l(q)\gamma_\theta(\alpha q).
\end{cases} { φ l ( t ) := ∑ q = 0 3 A l ( q ) Y α ( t − α q ) , V l ( θ ) := ∑ q = 0 3 B l ( q ) γ θ ( α q ) . for all t ≥ 0 ,
where γ θ ( ⋅ ) \gamma_\theta(\cdot) γ θ ( ⋅ ) , A j A_j A j and B \mathbb{B} B are given respectively in (11)-(12), (19) and (20) . The estimator θ ^ \widehat{\theta} θ is formally defined by
θ ^ N = arg min θ ∈ Θ Q ^ N ( θ ) , (21) \widehat{\theta}_N=\underset{\theta\in\Theta}{\arg\min}\widehat{Q}_N(\theta), \tag*{(21)} θ N = θ ∈ Θ arg min Q N ( θ ) , (21)
where Θ ⊂ R 3 \Theta\subset\mathbb{R}^3 Θ ⊂ R 3 is the parameter space. Note that this estimator depends on the step of observations α > 0 \alpha>0 α > 0 to be fixed. The following results are obtained.
Theorem 2. Consider the estimator defined by (21) and assume that the true parameter θ = ( σ 2 , H , λ ) \theta=(\sigma^2,H,\lambda) θ = ( σ 2 , H , λ ) belongs to the interior of the compact parameter space
Θ ⊂ { ( σ 2 , H , λ ) ∈ R 3 : σ 2 > 0 , H > 1 / 2 and 0 < λ < e ( ψ ( H + 1 / 2 ) + ψ ( H ) ) / 2 } \Theta\subset\left\{(\sigma^2,H,\lambda)\in\mathbb{R}^3:\ \sigma^2>0,\ H>1/2\ \text{and}\ 0<\lambda<e^{(\psi(H+1/2)+\psi(H))/2}\right\} Θ ⊂ { ( σ 2 , H , λ ) ∈ R 3 : σ 2 > 0 , H > 1/2 and 0 < λ < e ( ψ ( H + 1/2 ) + ψ ( H )) /2 } with rectangular form, then there exists some α ∗ > 0 \alpha^*>0 α ∗ > 0 depending only on Θ \Theta Θ for which θ ^ N \widehat{\theta}_N θ N is a consistent estimator of θ \theta θ almost surely and in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) -sense, for every fixed α ∈ ( α ∗ , ∞ ) \alpha\in(\alpha^*,\infty) α ∈ ( α ∗ , ∞ ) .
Remark 1. (i) The range ( α ∗ , ∞ ) (\alpha^*,\infty) ( α ∗ , ∞ ) can be improved by adjusting δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) in (61) and optimizing the numerical constants ( M j , L j , j = 0 , 1 , 2 , … ) (M_j,L_j,j=0,1,2,\ldots) ( M j , L j , j = 0 , 1 , 2 , … ) appearing in the proof of Lemmas 6 and 7 to obtain sharp bounds. The case where H ∈ ( 0 , 1 / 2 ) H\in(0,1/2) H ∈ ( 0 , 1/2 ) can be treated by similar procedures. The main task is to verify the identifiability condition, specifically, to check whether the mapping θ ↦ ϱ θ , 3 ( α ) \theta\mapsto\varrho_{\theta,3}(\alpha) θ ↦ ϱ θ , 3 ( α ) is globally injective (or univalent) on Θ \Theta Θ . For that purpose, one can rely on the existing global univalence results related to (weak almost) P-matrices [9 , 23 ] or (weak almost) N-matrices [21 ] , or PN-matrices [14 ] . In general, due to the complexity of the functions involved in ∇ ϱ θ , 3 ( α ) \nabla\varrho_{\theta,3}(\alpha) ∇ ϱ θ , 3 ( α ) , it is not clear how to analytically verify such a condition for a given α \alpha α . Nevertheless, this condition holds in the high-frequency estimation regime, that is, when the observation step α \alpha α is small (α → 0 \alpha\to0 α → 0 ). This can be established by evaluating the principal minors given in Lemma 7 by means of Taylor expansions. (ii) In the case H = 1 / 2 H = 1/2 H = 1/2 and λ = 0 \lambda= 0 λ = 0 , the process Y h Y_h Y h reduces to the second order increments of the Brownian motion whose autocovariance is linear with respect to σ 2 \sigma^2 σ 2 . Hence, the identifiability condition becomes trivial.
Proof of Theorem 2. Here we adopt notations of [2 ] . Clearly, the matrix B \mathbb{B} B defined by (20) is of full rank 3. There remains to verify the identifiability condition [2 Assumption 2.1] for any α \alpha α chosen in the range ( α ∗ , ∞ ) (\alpha^*,\infty) ( α ∗ , ∞ ) , where α ∗ \alpha^* α ∗ is the numerical constant due to Lemma 7. Let U ⊂ R 3 \mathcal{U} \subset\mathbb{R}^3 U ⊂ R 3 be an open rectangular region such that Θ ⊂ U \Theta\subset\mathcal{U} Θ ⊂ U and
σ 2 > 0 , H > 1 / 2 and 0 < λ < e ( ψ ( H + 1 / 2 ) + ψ ( H ) ) / 2 , for all ( σ 2 , H , λ ) ∈ U . (22) \sigma^2 > 0,\quad H > 1/2\quad\text{and}\quad0 < \lambda< e^{(\psi(H+1/2)+\psi(H))/2},\quad\text{for all }(\sigma^2,H,\lambda)\in\mathcal{U}. \tag*{(22)} σ 2 > 0 , H > 1/2 and 0 < λ < e ( ψ ( H + 1/2 ) + ψ ( H )) /2 , for all ( σ 2 , H , λ ) ∈ U . (22)
Let F : U ⟶ R 3 F:\mathcal{U}\longrightarrow\mathbb{R}^3 F : U ⟶ R 3 be the mapping defined by F ( θ ) = Λ ϱ θ , 3 ( α ) F(\theta)=\Lambda\varrho_{\theta,3}(\alpha) F ( θ ) = Λ ϱ θ , 3 ( α ) , where Λ = diag ( − 1 , 1 , 1 ) \Lambda=\operatorname{diag}(-1,1,1) Λ = diag ( − 1 , 1 , 1 ) is a diagonal matrix. Clearly, F F F is of class C 1 C^1 C 1 with the Jacobian matrix
∇ F ( θ ) = ( − Σ 1 , 1 − Σ 1 , 2 − Σ 1 , 3 Σ 2 , 1 Σ 2 , 2 Σ 2 , 3 Σ 3 , 1 Σ 3 , 2 Σ 3 , 3 ) . \nabla F(\theta)=
\begin{pmatrix}
-\Sigma_{1,1} & -\Sigma_{1,2} & -\Sigma_{1,3}\\
\Sigma_{2,1} & \Sigma_{2,2} & \Sigma_{2,3}\\
\Sigma_{3,1} & \Sigma_{3,2} & \Sigma_{3,3}
\end{pmatrix}. ∇ F ( θ ) = − Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 − Σ 1 , 2 Σ 2 , 2 Σ 3 , 2 − Σ 1 , 3 Σ 2 , 3 Σ 3 , 3 .
Taking into account that (22) holds on U \mathcal{U} U and using Lemma 7, we assert that for any fixed α ∈ ( α ∗ , ∞ ) \alpha\in(\alpha^*,\infty) α ∈ ( α ∗ , ∞ ) the Jacobian matrix ∇ F ( θ ) \nabla F(\theta) ∇ F ( θ ) has negative diagonals: ( ∇ F ( θ ) ) 1 , 1 = − ⟨ 2 , 3 ⟩ Σ (\nabla F(\theta))_{1,1}=-\langle2,3\rangle_{\Sigma} ( ∇ F ( θ ) ) 1 , 1 = − ⟨ 2 , 3 ⟩ Σ , ( ∇ F ( θ ) ) 2 , 2 = ⟨ 1 , 3 ⟩ Σ (\nabla F(\theta))_{2,2}=\langle1,3\rangle_{\Sigma} ( ∇ F ( θ ) ) 2 , 2 = ⟨ 1 , 3 ⟩ Σ and ( ∇ F ( θ ) ) 3 , 3 = ⟨ 1 , 2 ⟩ Σ (\nabla F(\theta))_{3,3}=\langle1,2\rangle_{\Sigma} ( ∇ F ( θ ) ) 3 , 3 = ⟨ 1 , 2 ⟩ Σ , while the remaining proper principal minors: ⟨ 3 ⟩ ∇ F ( θ ) = − ⟨ 3 ⟩ Σ \langle3\rangle_{\nabla F(\theta)}=-\langle3\rangle_{\Sigma} ⟨ 3 ⟩ ∇ F ( θ ) = − ⟨ 3 ⟩ Σ , ⟨ 2 ⟩ ∇ F ( θ ) = − ⟨ 2 ⟩ Σ \langle2\rangle_{\nabla F(\theta)}=-\langle2\rangle_{\Sigma} ⟨ 2 ⟩ ∇ F ( θ ) = − ⟨ 2 ⟩ Σ and ⟨ 1 ⟩ ∇ F ( θ ) = ⟨ 1 ⟩ Σ \langle1\rangle_{\nabla F(\theta)}=\langle1\rangle_{\Sigma} ⟨ 1 ⟩ ∇ F ( θ ) = ⟨ 1 ⟩ Σ are negative. Furthermore, its determinant ∣ ∇ F ( θ ) ∣ = − ⟨ ∅ ⟩ Σ |\nabla F(\theta)|=-\langle\varnothing\rangle_{\Sigma} ∣∇ F ( θ ) ∣ = − ⟨ ∅ ⟩ Σ is positive. This means that ∇ F ( θ ) \nabla F(\theta) ∇ F ( θ ) is a weak almost N-matrix, and according to [21 Theorem 4.1] with the associated remarks [21 p. 123] F F F is globally injective (or univalent) on U \mathcal{U} U . Hence, the mapping θ ↦ ϱ θ , 3 ( α ) \theta\mapsto\varrho_{\theta,3}(\alpha) θ ↦ ϱ θ , 3 ( α ) is globally injective on Θ \Theta Θ as well. Moreover, the mapping θ ↦ B ϱ θ ( α ) \theta\mapsto\mathbb{B}\varrho_{\theta}(\alpha) θ ↦ B ϱ θ ( α ) is injective (see Appendix § 8.4 ). At this stage we can assert that all the conditions of [2 Theorem 2.1 and Theorem 2.2] are met, and the consistency of θ ^ N \widehat{\theta}_N θ N holds almost surely and in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) -sense. □
Theorem 3. Under the assumptions of Theorem 2 we have
N ( θ ^ N − θ ) → L N ( 0 , C ( θ ) Γ ( θ ) C ( θ ) ′ ) \sqrt{N}(\widehat{\theta}_N-\theta)\xrightarrow{\mathcal{L}}\mathcal{N}\left(0,C(\theta)\Gamma(\theta)C(\theta)'\right) N ( θ N − θ ) L N ( 0 , C ( θ ) Γ ( θ ) C ( θ ) ′ ) as N → ∞ N\to\infty N → ∞ , where
( Γ ( θ ) ) i , j = 2 ∑ p ∈ Z ( ∑ k = 0 3 b k i , j γ θ ( α ( p + k ) ) ) 2 , for all i , j = 1 , 2 , 3 , b k i , j : = ∑ q = 0 3 − k A i ( q ) A j ( q + k ) + ∑ q = 0 3 − k A j ( q ) A i ( q + k ) , k = 0 , 1 , 2 , 3 \begin{aligned}
(\Gamma(\theta))_{i,j} &= 2\sum_{p\in\mathbb{Z}}\left(\sum_{k=0}^{3}b_k^{i,j}\gamma_{\theta}(\alpha(p+k))\right)^2,\qquad&&\text{for all }i,j=1,2,3,\\
b_k^{i,j} &:= \sum_{q=0}^{3-k}\mathcal{A}_i(q)\mathcal{A}_j(q+k)+\sum_{q=0}^{3-k}\mathcal{A}_j(q)\mathcal{A}_i(q+k), &&k=0,1,2,3
\end{aligned} ( Γ ( θ ) ) i , j b k i , j = 2 p ∈ Z ∑ ( k = 0 ∑ 3 b k i , j γ θ ( α ( p + k )) ) 2 , := q = 0 ∑ 3 − k A i ( q ) A j ( q + k ) + q = 0 ∑ 3 − k A j ( q ) A i ( q + k ) , for all i , j = 1 , 2 , 3 , k = 0 , 1 , 2 , 3 C ( θ ) = − [ ( ∇ ϱ θ ( α ) ) ′ B ′ B ∇ ϱ θ ( α ) ] − 1 ( ∇ ϱ θ ( α ) ) ′ B ′ , C(\theta)=-\left[(\nabla\varrho_{\theta}(\alpha))'\mathbb{B}'\mathbb{B}\nabla\varrho_{\theta}(\alpha)\right]^{-1}(\nabla\varrho_{\theta}(\alpha))'\mathbb{B}', C ( θ ) = − [ ( ∇ ϱ θ ( α ) ) ′ B ′ B ∇ ϱ θ ( α ) ] − 1 ( ∇ ϱ θ ( α ) ) ′ B ′ , ∇ ϱ θ ( α ) \nabla\varrho_{\theta}(\alpha) ∇ ϱ θ ( α ) is the Jacobian of the vector ϱ θ ( α ) : = ( γ θ ( 0 ) , γ θ ( α ) , γ θ ( 2 α ) , γ θ ( 3 α ) ) ′ \varrho_{\theta}(\alpha):=(\gamma_{\theta}(0),\gamma_{\theta}(\alpha),\gamma_{\theta}(2\alpha),\gamma_{\theta}(3\alpha))' ϱ θ ( α ) := ( γ θ ( 0 ) , γ θ ( α ) , γ θ ( 2 α ) , γ θ ( 3 α ) ) ′ with γ θ ( ⋅ ) \gamma_{\theta}(\cdot) γ θ ( ⋅ ) being defined by (11) –(12) and A j \mathcal{A}_j A j , B \mathbb{B} B are given in (19) –(20) .
Proof. We recall that the filters used in our framework are at most of order 2 (l 1 = 1 l_1=1 l 1 = 1 and l 2 = l 3 = 2 l_2=l_3=2 l 2 = l 3 = 2 ). According to [2 Theorem 2.3] the estimator θ ^ N \widehat{\theta}_N θ N is asymptotically normal if the conditions of Theorem 2 are fulfilled and R ( ω ) : = ( 1 ∧ ∣ ω ∣ 3 ) f ‾ θ ( ω ) ∈ L 2 ( − π , π ) \mathcal{R}(\omega):=(1\wedge|\omega|^3)\overline{f}_{\theta}(\omega)\in L^2(-\pi,\pi) R ( ω ) := ( 1 ∧ ∣ ω ∣ 3 ) f θ ( ω ) ∈ L 2 ( − π , π ) , where
f ‾ θ ( ω ) : = ∑ p ∈ Z f θ ( ω + 2 p π α ) , ω ∈ [ − π , π ] \overline{f}_{\theta}(\omega):=\sum_{p\in\mathbb{Z}}f_{\theta}\left(\frac{\omega+2p\pi}{\alpha}\right),\qquad\omega\in[-\pi,\pi] f θ ( ω ) := p ∈ Z ∑ f θ ( α ω + 2 p π ) , ω ∈ [ − π , π ]
with f θ ( ⋅ ) f_{\theta}(\cdot) f θ ( ⋅ ) being the spectral density function given in (10) and θ \theta θ denotes the true parameter. Set Ψ ( ω ) = ( 1 ∧ ∣ ω ∣ 3 ) 1 ( − π , π ) ( ω ) \Psi(\omega)=(1\wedge|\omega|^3)1_{(-\pi,\pi)}(\omega) Ψ ( ω ) = ( 1 ∧ ∣ ω ∣ 3 ) 1 ( − π , π ) ( ω ) and let ∥ Ψ ∥ \|\Psi\| ∥Ψ∥ be its L 2 L^2 L 2 -norm. Straightforward computations lead to
π 2 16 σ 4 Γ ( H + 1 / 2 ) 4 ∫ − π π ∣ R ( ω ) ∣ 2 d ω ≤ α 4 H + 2 ∫ Ψ 2 ( ω ) [ ∑ p ∈ Z ( ( λ α ) 2 + ( ω + 2 p π ) 2 ) − H − 1 / 2 ] 2 d ω \frac{\pi^2}{16\sigma^4\Gamma(H+1/2)^4}\int_{-\pi}^{\pi}|\mathcal{R}(\omega)|^2\,d\omega
\leq\alpha^{4H+2}\int\Psi^2(\omega)\left[\sum_{p\in\mathbb{Z}}\left((\lambda\alpha)^2+(\omega+2p\pi)^2\right)^{-H-1/2}\right]^2\,d\omega 16 σ 4 Γ ( H + 1/2 ) 4 π 2 ∫ − π π ∣ R ( ω ) ∣ 2 d ω ≤ α 4 H + 2 ∫ Ψ 2 ( ω ) p ∈ Z ∑ ( ( λ α ) 2 + ( ω + 2 p π ) 2 ) − H − 1/2 2 d ω
≤ 3 α 4 H + 2 { ∫ Ψ 2 ( ω ) ( ( λ α ) 2 + ω 2 ) 2 H + 1 + ∫ Ψ 2 ( ω ) [ ∑ p ≥ 1 ( ( λ α ) 2 + ( ω + 2 p π ) 2 ) − H − 1 / 2 ] 2 d ω + ∫ Ψ 2 ( ω ) [ ∑ p ≤ − 1 ( ( λ α ) 2 + ( ω + 2 p π ) 2 ) − H − 1 / 2 ] 2 d ω } \leq3\alpha^{4H+2}\left\{\int\frac{\Psi^2(\omega)}{\left((\lambda\alpha)^2+\omega^2\right)^{2H+1}}+\int\Psi^2(\omega)\left[\sum_{p\geq1}\left((\lambda\alpha)^2+(\omega+2p\pi)^2\right)^{-H-1/2}\right]^2\,d\omega
+\int\Psi^2(\omega)\left[\sum_{p\leq-1}\left((\lambda\alpha)^2+(\omega+2p\pi)^2\right)^{-H-1/2}\right]^2\,d\omega\right\} ≤ 3 α 4 H + 2 ⎩ ⎨ ⎧ ∫ ( ( λ α ) 2 + ω 2 ) 2 H + 1 Ψ 2 ( ω ) + ∫ Ψ 2 ( ω ) [ p ≥ 1 ∑ ( ( λ α ) 2 + ( ω + 2 p π ) 2 ) − H − 1/2 ] 2 d ω + ∫ Ψ 2 ( ω ) [ p ≤ − 1 ∑ ( ( λ α ) 2 + ( ω + 2 p π ) 2 ) − H − 1/2 ] 2 d ω ⎭ ⎬ ⎫
≤ 3 α 4 H + 2 { ∥ Ψ ∥ 2 ( λ α ) 4 H + 2 + ( ∑ p ≥ 1 ∣ 2 p − 1 ∣ − 2 H − 1 ) 2 ∥ Ψ ∥ 2 π 4 H + 2 + ( ∑ p ≤ − 1 ∣ 2 p + 1 ∣ − 2 H − 1 ) 2 ∥ Ψ ∥ 2 π 4 H + 2 } < ∞ . \leq3\alpha^{4H+2}\left\{\frac{\|\Psi\|^2}{(\lambda\alpha)^{4H+2}}+\frac{\left(\sum_{p\geq1}|2p-1|^{-2H-1}\right)^2\|\Psi\|^2}{\pi^{4H+2}}+\frac{\left(\sum_{p\leq-1}|2p+1|^{-2H-1}\right)^2\|\Psi\|^2}{\pi^{4H+2}}\right\}
<\infty. ≤ 3 α 4 H + 2 ⎩ ⎨ ⎧ ( λ α ) 4 H + 2 ∥Ψ ∥ 2 + π 4 H + 2 ( ∑ p ≥ 1 ∣2 p − 1 ∣ − 2 H − 1 ) 2 ∥Ψ ∥ 2 + π 4 H + 2 ( ∑ p ≤ − 1 ∣2 p + 1 ∣ − 2 H − 1 ) 2 ∥Ψ ∥ 2 ⎭ ⎬ ⎫ < ∞.
This implies R ( ω ) ∈ L 2 ( − π , π ) \mathcal{R}(\omega)\in L^2(-\pi,\pi) R ( ω ) ∈ L 2 ( − π , π ) . In the second inequality above we split the series inside the integrand as ∑ p = 0 + ∑ p ≥ 1 + ∑ p ≤ − 1 \sum_{p=0}+\sum_{p\geq1}+\sum_{p\leq-1} ∑ p = 0 + ∑ p ≥ 1 + ∑ p ≤ − 1 and used the fact that ( u + v + w ) 2 ≤ 3 ( u 2 + v 2 + w 2 ) (u+v+w)^2\leq3\left(u^2+v^2+w^2\right) ( u + v + w ) 2 ≤ 3 ( u 2 + v 2 + w 2 ) , for all u , v , w ∈ R u,v,w\in\mathbb{R} u , v , w ∈ R .
Alternative estimators of σ 2 \sigma^2 σ 2 , H H H and λ \lambda λ
As the drift μ \mu μ can be estimated independently of the noise parameters, we restrict ourselves to the model (1) with μ = 0 \mu=0 μ = 0 . We denote the resulting process as X \mathbf{X} X , that is,
X t = σ B H , λ ( t ) (23) X_t=\sigma B_{H,\lambda}(t) \tag*{(23)} X t = σ B H , λ ( t ) (23)
for all t ≥ 0 t\geq0 t ≥ 0 . Based on a discrete observations of X \mathbf{X} X , we aim to estimate σ 2 \sigma^2 σ 2 , H H H and λ \lambda λ . Given a stochastic process ξ = { ξ t : t ≥ 0 } \xi=\{\xi_t:t\geq0\} ξ = { ξ t : t ≥ 0 } and for n ∈ Z + n\in\mathbb{Z}_+ n ∈ Z + we write
V n ( 2 ) ( ξ ) : = ∑ k = 1 2 n ( ξ k / 2 n − ξ ( k − 1 ) / 2 n ) 2 V_n^{(2)}(\xi):=\sum_{k=1}^{2^n}\left(\xi_{k/2^n}-\xi_{(k-1)/2^n}\right)^2 V n ( 2 ) ( ξ ) := k = 1 ∑ 2 n ( ξ k / 2 n − ξ ( k − 1 ) / 2 n ) 2
The following results are obtained.
Theorem 4. Consider the model (23) with H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) . The following statements hold.
(i) The statistic
H ~ n = 1 2 ( 1 − log 2 V n ( 2 ) ( X ) n ) (24) \widetilde{H}_n=\frac{1}{2}\left(1-\frac{\log_2 V_n^{(2)}(\mathbf{X})}{n}\right) \tag*{(24)} H n = 2 1 ( 1 − n log 2 V n ( 2 ) ( X ) ) (24)
is a strongly consistent estimator of H H H . Furthermore, we have almost surely H ~ n − H = O ( 1 / n ) \widetilde{H}_n-H=O(1/n) H n − H = O ( 1/ n ) , as n → ∞ n\to\infty n → ∞ .
(ii) Under the condition m / n = o ( 1 ) m/n=o(1) m / n = o ( 1 ) , as m , n → ∞ m,n\to\infty m , n → ∞ , the parameter σ 2 \sigma^2 σ 2 can be strongly estimated by
σ ~ m , n 2 = 2 π sin ( π H ~ n ) Γ ( 2 H ~ n + 1 ) Γ ( H ~ n + 1 / 2 ) 2 ⋅ 2 m ( 2 H ~ n − 1 ) − 1 V m ( 2 ) ( X ) (25) \widetilde{\sigma}_{m,n}^{2}=\frac{2\pi\sin(\pi\widetilde{H}_n)\Gamma(2\widetilde{H}_n+1)}{\Gamma(\widetilde{H}_n+1/2)^2}\cdot2^{m(2\widetilde{H}_n-1)-1}V_m^{(2)}(X) \tag*{(25)} σ m , n 2 = Γ ( H n + 1/2 ) 2 2 π sin ( π H n ) Γ ( 2 H n + 1 ) ⋅ 2 m ( 2 H n − 1 ) − 1 V m ( 2 ) ( X ) (25)
Remark 2. If H ∈ ( 0 , 1 / 2 ) H\in(0,1/2) H ∈ ( 0 , 1/2 ) , then the same results of Theorem 4 hold based on the observations of the model (1). This is due to the fact that
2 n ( 2 H − 1 ) V n ( 2 ) ( X ) = 2 n ( 2 H − 1 ) V n ( 2 ) ( X ) + o ( 1 ) 2^{n(2H-1)}V_n^{(2)}(X)=2^{n(2H-1)}V_n^{(2)}(\mathbf{X})+o(1) 2 n ( 2 H − 1 ) V n ( 2 ) ( X ) = 2 n ( 2 H − 1 ) V n ( 2 ) ( X ) + o ( 1 )
as n → ∞ n\to\infty n → ∞ with probability one.
Proof of Theorem 4. It is well known that X \mathbf{X} X is centered Gaussian with stationary increments, and admits the spectral density function
f X ( ω ) = σ 2 Γ ( H + 1 / 2 ) 2 2 π ( λ 2 + ω 2 ) H + 1 / 2 = σ 2 Γ ( H + 1 / 2 ) 2 2 π ω 2 H + 1 ( 1 − O ( ω − 2 ) ) \begin{aligned}
f_{\mathbf{X}}(\omega)=\frac{\sigma^2\Gamma(H+1/2)^2}{2\pi(\lambda^2+\omega^2)^{H+1/2}} \\
&=\frac{\sigma^2\Gamma(H+1/2)^2}{2\pi\omega^{2H+1}}\left(1-O\left(\omega^{-2}\right)\right)
\end{aligned} f X ( ω ) = 2 π ( λ 2 + ω 2 ) H + 1/2 σ 2 Γ ( H + 1/2 ) 2 = 2 π ω 2 H + 1 σ 2 Γ ( H + 1/2 ) 2 ( 1 − O ( ω − 2 ) )
as ω → ∞ \omega\to\infty ω → ∞ . So, we apply Lemma 5 with c = σ 2 Γ ( H + 1 / 2 ) 2 / ( 2 π ) > 0 c=\sigma^2\Gamma(H+1/2)^2/(2\pi)>0 c = σ 2 Γ ( H + 1/2 ) 2 / ( 2 π ) > 0 and α = 2 H + 1 ∈ ( 1 , 3 ) \alpha=2H+1\in(1,3) α = 2 H + 1 ∈ ( 1 , 3 ) to obtain the following almost sure convergences
2 − 1 n log 2 V n ( 2 ) ( X ) ⟶ α , as n → ∞ , (26) 2-\frac{1}{n}\log_2 V_n^{(2)}(\mathbf{X})\longrightarrow\alpha,\quad\text{as }n\to\infty, \tag*{(26)} 2 − n 1 log 2 V n ( 2 ) ( X ) ⟶ α , as n → ∞ , (26)
− cos ( α π 2 ) Γ ( α ) 2 m ( α − 2 ) − 1 V m ( 2 ) ( X ) ⟶ c , as m → ∞ . (27) -\cos\left(\frac{\alpha\pi}{2}\right)\Gamma(\alpha)2^{m(\alpha-2)-1}V_m^{(2)}(\mathbf{X})\longrightarrow c,\quad\text{as }m\to\infty. \tag*{(27)} − cos ( 2 α π ) Γ ( α ) 2 m ( α − 2 ) − 1 V m ( 2 ) ( X ) ⟶ c , as m → ∞. (27)
The first part of (i) follows immediately from (26) with the use of elementary computations. For the second part, note that results of Lemma 5 is just an application of [12 Theorems 1 and 2] with γ = 3 − α \gamma=3-\alpha γ = 3 − α and g ( t ) = − 2 c / ( cos ( α π 2 ) Γ ( α ) ) ≠ 0 g(t)=-2c/\left(\cos\left(\frac{\alpha\pi}{2}\right)\Gamma(\alpha)\right)\ne0 g ( t ) = − 2 c / ( cos ( 2 α π ) Γ ( α ) ) = 0 . So, revising the proof, one can readily derive a more precise statement showing the almost sure rate of convergence of H ~ n \widetilde{H}_n H n , as given in (i)-Theorem 4. The function g ( t ) g(t) g ( t ) chosen here is well defined and nonnegative for H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) . Clearly, this choice is not suitable for the boundary values of H H H . However, if H = 1 H=1 H = 1 , one can still estimate σ 2 \sigma^2 σ 2 by a statistic with a reduced form. In fact, one has (see [12 Theorems 4] )
2 m + 2 m log ( 2 ) V m ( 2 ) ( X ) → P − a . s . σ 2 , as m → ∞ . \frac{2^{m+2}}{m\log(2)}V_m^{(2)}(\mathbf{X})\xrightarrow{\mathrm{P-a.s.}}\sigma^2,\quad\text{as }m\to\infty. m log ( 2 ) 2 m + 2 V m ( 2 ) ( X ) P − a.s. σ 2 , as m → ∞.
(ii) First, equation (24) can be rewritten as σ ~ m , n 2 = φ m ( H ~ n ) V m ( 2 ) ( X ) \widetilde{\sigma}_{m,n}^2=\varphi_m(\widetilde{H}_n)V_m^{(2)}(\mathbf{X}) σ m , n 2 = φ m ( H n ) V m ( 2 ) ( X ) , where
φ m ( x ) = 2 π sin ( π x ) Γ ( 2 x + 1 ) Γ ( x + 1 / 2 ) 2 ⋅ 2 m ( 2 x − 1 ) − 1 , for all x ∈ ( 0 , 1 ) . (28) \varphi_m(x)=\frac{2\pi\sin(\pi x)\Gamma(2x+1)}{\Gamma(x+1/2)^2}\cdot2^{m(2x-1)-1},\quad\text{for all }x\in(0,1). \tag*{(28)} φ m ( x ) = Γ ( x + 1/2 ) 2 2 π sin ( π x ) Γ ( 2 x + 1 ) ⋅ 2 m ( 2 x − 1 ) − 1 , for all x ∈ ( 0 , 1 ) . (28)
Performing some trigonometric computations with the use of (27) , we can write
σ ~ m , n 2 = { φ m ( H ) V m ( 2 ) ( X ) } ( φ m ( H ~ n ) φ m ( H ) ) ∼ σ 2 ( φ m ( H ~ n ) φ m ( H ) ) , P -a.s. \begin{aligned}
\widetilde{\sigma}_{m,n}^2=\left\{\varphi_m(H)V_m^{(2)}(\mathbf{X})\right\}\left(\frac{\varphi_m(\widetilde{H}_n)}{\varphi_m(H)}\right) \\
&\sim\sigma^2\left(\frac{\varphi_m(\widetilde{H}_n)}{\varphi_m(H)}\right),\quad\mathbb{P}\text{-a.s.}
\end{aligned} σ m , n 2 = { φ m ( H ) V m ( 2 ) ( X ) } ( φ m ( H ) φ m ( H n ) ) ∼ σ 2 ( φ m ( H ) φ m ( H n ) ) , P -a.s.
To complete the proof, it suffices to prove that φ m ( H ~ n ) / φ m ( H ) → P − a . s . 1 \varphi_m(\widetilde{H}_n)/\varphi_m(H) \xrightarrow{\mathrm{P-a.s.}} 1 φ m ( H n ) / φ m ( H ) P − a.s. 1 as m , n → ∞ m,n \to\infty m , n → ∞ . By straightforward computations we have
φ m ( H ~ n ) φ m ( H ) = sin ( π H ~ n ) Γ ( 2 H ~ n + 1 ) sin ( π H ) Γ ( 2 H + 1 ) ⋅ ( Γ ( H + 1 / 2 ) Γ ( H ~ n + 1 / 2 ) ) 2 ⋅ 2 2 m ( H ~ n − H ) . (29) \frac{\varphi_m(\widetilde{H}_n)}{\varphi_m(H)}
=
\frac{\sin(\pi\widetilde{H}_n)\Gamma(2\widetilde{H}_n+1)}
{\sin(\pi H)\Gamma(2H+1)}
\cdot
\left(\frac{\Gamma(H+1/2)}{\Gamma(\widetilde{H}_n+1/2)}\right)^2
\cdot2^{2m(\widetilde{H}_n-H)}. \tag*{(29)} φ m ( H ) φ m ( H n ) = sin ( π H ) Γ ( 2 H + 1 ) sin ( π H n ) Γ ( 2 H n + 1 ) ⋅ ( Γ ( H n + 1/2 ) Γ ( H + 1/2 ) ) 2 ⋅ 2 2 m ( H n − H ) . (29)
By means of continuous mapping theorem, both the first and the second factors on the right hand side (RHS) of (29) converge almost surely to 1 as n → ∞ n \to\infty n → ∞ . Now, using the last statement in (i)-Theorem 4 with the hypothesis m / n = o ( 1 ) m/n=o(1) m / n = o ( 1 ) yields m ( H ~ n − H ) = o ( 1 ) m(\widetilde{H}_n-H)=o(1) m ( H n − H ) = o ( 1 ) as m , n → ∞ m,n \to\infty m , n → ∞ . Hence, the last factor on the RHS of (29) converges as well to 1, with probability one.
Consider the estimators H ~ \widetilde{H} H and σ ~ 2 \widetilde{\sigma}^2 σ 2 as defined in Theorem 4 and the average
X ‾ n 1 , 2 ( h ) = 1 n ∑ k = 0 n − 1 ( X h ( k + 1 ) − X h k ) 2 , \overline{X}_n^{1,2}(h)=\frac{1}{n}\sum_{k=0}^{n-1}\left(X_{h(k+1)}-X_{hk}\right)^2, X n 1 , 2 ( h ) = n 1 k = 0 ∑ n − 1 ( X h ( k + 1 ) − X hk ) 2 ,
then the following results are obtained.
Theorem 5. Consider the model (23) with H ∈ ( 0 , 1 ) H \in(0,1) H ∈ ( 0 , 1 ) . The following statistic
λ ~ m , n , h = 1 2 exp ( log ( 2 σ ~ 2 Γ ( 2 H ~ ) ) − log X ‾ n 1 , 2 ( h ) 2 H ~ ) \widetilde{\lambda}_{m,n,h}
=
\frac{1}{2}\exp\left(
\frac{\log\left(2\widetilde{\sigma}^{2}\Gamma(2\widetilde{H})\right)-\log\overline{X}_n^{1,2}(h)}
{2\widetilde{H}}
\right) λ m , n , h = 2 1 exp 2 H log ( 2 σ 2 Γ ( 2 H ) ) − log X n 1 , 2 ( h ) converges in probability to λ \lambda λ as m , n , h → ∞ m,n,h \to\infty m , n , h → ∞ .
Proof. First, let us prove that X ‾ n 1 , 2 ( h ) → L 2 ( Ω ) 2 σ 2 Γ ( 2 H ) / ( 2 λ ) 2 H \overline{X}_n^{1,2}(h)\xrightarrow{L^2(\Omega)}2\sigma^2\Gamma(2H)/(2\lambda)^{2H} X n 1 , 2 ( h ) L 2 ( Ω ) 2 σ 2 Γ ( 2 H ) / ( 2 λ ) 2 H as n , h → ∞ n,h \to\infty n , h → ∞ . For that purpose, we write
X ‾ n 1 , 2 ( h ) − 2 σ 2 Γ ( 2 H ) ( 2 λ ) 2 H = ( X ‾ n 1 , 2 ( h ) − σ 2 ϕ ( h ) ) ⏟ = : F n , h + σ 2 ( ϕ ( h ) − 2 Γ ( 2 H ) ( 2 λ ) 2 H ) . (30) \overline{X}_n^{1,2}(h)-\frac{2\sigma^2\Gamma(2H)}{(2\lambda)^{2H}}
=
\underbrace{\left(\overline{X}_n^{1,2}(h)-\sigma^2\phi(h)\right)}_{=:F_{n,h}}
+
\sigma^2\left(\phi(h)-\frac{2\Gamma(2H)}{(2\lambda)^{2H}}\right). \tag*{(30)} X n 1 , 2 ( h ) − ( 2 λ ) 2 H 2 σ 2 Γ ( 2 H ) = =: F n , h ( X n 1 , 2 ( h ) − σ 2 ϕ ( h ) ) + σ 2 ( ϕ ( h ) − ( 2 λ ) 2 H 2Γ ( 2 H ) ) . (30)
where ϕ ( h x ) = ∣ h x ∣ 2 H C h x 2 = : ϕ ( x ) \phi(hx)=|hx|^{2H}C_{hx}^2=:\phi(x) ϕ ( h x ) = ∣ h x ∣ 2 H C h x 2 =: ϕ ( x ) . In view of (2), the last term on the RHS of (30) converges towards 0 as h → ∞ h \to\infty h → ∞ , while the first term can be expressed as
F n , h = 1 n ∑ k = 0 n − 1 [ ( Z ~ h ( k h ) ) 2 − E ( Z ~ h ( k h ) ) 2 ] , F_{n,h}
=
\frac{1}{n}\sum_{k=0}^{n-1}
\left[
\left(\widetilde{Z}_h(kh)\right)^2
-
\mathbb{E}\left(\widetilde{Z}_h(kh)\right)^2
\right], F n , h = n 1 k = 0 ∑ n − 1 [ ( Z h ( k h ) ) 2 − E ( Z h ( k h ) ) 2 ] ,
where Z ~ h ( k h ) = X h ( k + 1 ) − X h k \widetilde{Z}_h(kh)=X_{h(k+1)}-X_{hk} Z h ( k h ) = X h ( k + 1 ) − X hk . Clearly, E ( F n , h ) = 0 \mathbb{E}(F_{n,h})=0 E ( F n , h ) = 0 and
Var ( F n , h ) = n − 2 Var { ∑ k = 0 n − 1 ( ( Z ~ h ( k h ) ) 2 − E ( Z ~ h ( k h ) ) 2 ) } = n − 2 Cov { ∑ k = 0 n − 1 ( ( Z ~ h ( k h ) ) 2 − E ( Z ~ h ( k h ) ) 2 ) , ∑ j = 0 n − 1 ( ( Z ~ h ( j h ) ) 2 − E ( Z ~ h ( j h ) ) 2 ) } = n − 2 ∑ k = 0 n − 1 ∑ j = 0 n − 1 Cov ( ( Z ~ h ( k h ) ) 2 , ( Z ~ h ( j h ) ) 2 ) = 2 n 2 ∑ k = 0 n − 1 ∑ j = 0 n − 1 Cov 2 ( Z ~ h ( k h ) , Z ~ h ( j h ) ) . = 2 n 2 ∑ k = 0 n − 1 ∑ j = 0 n − 1 Cov 2 ( Z ~ h ( h ( k − j ) ) , Z ~ h ( 0 ) ) , (by stationarity of Z ~ h ), = σ 4 2 n 2 ∑ k , j = 0 n − 1 ( ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ) 2 , (by (6)), = σ 4 2 n 2 ( ∑ k = j + 2 ∑ k = j − 1 + 2 ∑ k < j − 1 ) ( ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ) 2 ≤ σ 4 2 n 2 { 4 n ϕ h ( 1 ) 2 + 2 n ( ϕ h ( 2 ) − 2 ϕ h ( 1 ) ) 2 + 2 ∑ k < j − 1 ( ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ) 2 } (31) \begin{aligned}
\operatorname{Var}(F_{n,h})
=
n^{-2}\operatorname{Var}\left\{
\sum_{k=0}^{n-1}
\left(
\left(\widetilde{Z}_h(kh)\right)^2
-
\mathbb{E}\left(\widetilde{Z}_h(kh)\right)^2
\right)
\right\} \\
&=
n^{-2}\operatorname{Cov}\left\{
\sum_{k=0}^{n-1}
\left(
\left(\widetilde{Z}_h(kh)\right)^2
-
\mathbb{E}\left(\widetilde{Z}_h(kh)\right)^2
\right),
\sum_{j=0}^{n-1}
\left(
\left(\widetilde{Z}_h(jh)\right)^2
-
\mathbb{E}\left(\widetilde{Z}_h(jh)\right)^2
\right)
\right\} \\
&=
n^{-2}\sum_{k=0}^{n-1}\sum_{j=0}^{n-1}
\operatorname{Cov}\left(
\left(\widetilde{Z}_h(kh)\right)^2,
\left(\widetilde{Z}_h(jh)\right)^2
\right)
=
\frac{2}{n^2}\sum_{k=0}^{n-1}\sum_{j=0}^{n-1}
\operatorname{Cov}^2\left(\widetilde{Z}_h(kh),\widetilde{Z}_h(jh)\right). \\
&= \frac{2}{n^2}\sum_{k=0}^{n-1}\sum_{j=0}^{n-1}\operatorname{Cov}^{2}\left(\widetilde{Z}_h(h(k-j)),\widetilde{Z}_h(0)\right), \quad\text{(by stationarity of } \widetilde{Z}_h\text{),} \\
&= \frac{\sigma^4}{2n^2}\sum_{k,j=0}^{n-1}\left(\phi_h(k-j+1)+\phi_h(k-j-1)-2\phi_h(k-j)\right)^2, \quad\text{(by (6)),} \\
&= \frac{\sigma^4}{2n^2}\left(\sum_{k=j}+2\sum_{k=j-1}+2\sum_{k<j-1}\right)\left(\phi_h(k-j+1)+\phi_h(k-j-1)-2\phi_h(k-j)\right)^2 \\
&\le\frac{\sigma^4}{2n^2}\left\{4n\phi_h(1)^2+2n\left(\phi_h(2)-2\phi_h(1)\right)^2
+2\sum_{k<j-1}\left(\phi_h(k-j+1)+\phi_h(k-j-1)-2\phi_h(k-j)\right)^2\right\} \tag*{(31)}
\end{aligned} Var ( F n , h ) = n − 2 Var { k = 0 ∑ n − 1 ( ( Z h ( k h ) ) 2 − E ( Z h ( k h ) ) 2 ) } = n − 2 Cov { k = 0 ∑ n − 1 ( ( Z h ( k h ) ) 2 − E ( Z h ( k h ) ) 2 ) , j = 0 ∑ n − 1 ( ( Z h ( j h ) ) 2 − E ( Z h ( j h ) ) 2 ) } = n − 2 k = 0 ∑ n − 1 j = 0 ∑ n − 1 Cov ( ( Z h ( k h ) ) 2 , ( Z h ( j h ) ) 2 ) = n 2 2 k = 0 ∑ n − 1 j = 0 ∑ n − 1 Cov 2 ( Z h ( k h ) , Z h ( j h ) ) . = n 2 2 k = 0 ∑ n − 1 j = 0 ∑ n − 1 Cov 2 ( Z h ( h ( k − j )) , Z h ( 0 ) ) , (by stationarity of Z h ), = 2 n 2 σ 4 k , j = 0 ∑ n − 1 ( ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ) 2 , (by (6)), = 2 n 2 σ 4 k = j ∑ + 2 k = j − 1 ∑ + 2 k < j − 1 ∑ ( ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ) 2 ≤ 2 n 2 σ 4 ⎩ ⎨ ⎧ 4 n ϕ h ( 1 ) 2 + 2 n ( ϕ h ( 2 ) − 2 ϕ h ( 1 ) ) 2 + 2 k < j − 1 ∑ ( ϕ h ( k − j + 1 ) + ϕ h ( k − j − 1 ) − 2 ϕ h ( k − j ) ) 2 ⎭ ⎬ ⎫ (31)
≤ σ 4 2 n 2 { 4 n ϕ h ( 1 ) 2 + 2 n ( ϕ h ( 2 ) − 2 ϕ h ( 1 ) ) 2 + 2 c 2 2 ∑ k < j − 1 [ 2 g H ( λ h ∣ k − j ∣ ) − g H ( λ h ∣ k − j + 1 ∣ ) − g H ( λ h ∣ k − j − 1 ∣ ) ] 2 } (32) \le\frac{\sigma^4}{2n^2}\left\{4n\phi_h(1)^2+2n\left(\phi_h(2)-2\phi_h(1)\right)^2
+2c_2^2\sum_{k<j-1}\left[2g_H\left(\lambda_h|k-j|\right)-g_H\left(\lambda_h|k-j+1|\right)-g_H\left(\lambda_h|k-j-1|\right)\right]^2\right\} \tag*{(32)} ≤ 2 n 2 σ 4 ⎩ ⎨ ⎧ 4 n ϕ h ( 1 ) 2 + 2 n ( ϕ h ( 2 ) − 2 ϕ h ( 1 ) ) 2 + 2 c 2 2 k < j − 1 ∑ [ 2 g H ( λ h ∣ k − j ∣ ) − g H ( λ h ∣ k − j + 1∣ ) − g H ( λ h ∣ k − j − 1∣ ) ] 2 ⎭ ⎬ ⎫ (32)
where c 2 = 2 Γ ( H + 1 / 2 ) / ( π 2 H λ 2 H ) c_2=2\Gamma(H+1/2)/(\sqrt{\pi}2^H\lambda^{2H}) c 2 = 2Γ ( H + 1/2 ) / ( π 2 H λ 2 H ) and g H ( ⋅ ) g_H(\cdot) g H ( ⋅ ) being defined in the appendix. In (30) we used the fact that Cov ( ξ 2 , ζ 2 ) = 2 Cov 2 ( ξ , ζ ) \operatorname{Cov}(\xi^2,\zeta^2)=2\operatorname{Cov}^2(\xi,\zeta) Cov ( ξ 2 , ζ 2 ) = 2 Cov 2 ( ξ , ζ ) , for any Gaussian vector ( ξ , ζ ) (\xi,\zeta) ( ξ , ζ ) . For the inequality (31), we refer to (7). Now, by virtue of (i)-Lemma 2, the series appearing in (31) can be evaluated as
∑ k < j − 1 [ 2 g H ( λ h ∣ k − j ∣ ) − g H ( λ h ∣ k − j + 1 ∣ ) − g H ( λ h ∣ k − j − 1 ∣ ) ] 2 ≤ A δ , H 2 ∑ j = 0 n − 1 ∑ i = j − 2 n − 1 ( 2 e − λ δ h ∣ k − j ∣ + e − λ δ h ∣ k − j + 1 ∣ + e − λ δ h ∣ k − j − 1 ∣ ) 2 \begin{aligned}
\sum_{k<j-1}\left[2g_H\left(\lambda_h|k-j|\right)-g_H\left(\lambda_h|k-j+1|\right)-g_H\left(\lambda_h|k-j-1|\right)\right]^2 \\
&\le A_{\delta,H}^2\sum_{j=0}^{n-1}\sum_{i=j-2}^{n-1}\left(2e^{-\lambda\delta h|k-j|}+e^{-\lambda\delta h|k-j+1|}+e^{-\lambda\delta h|k-j-1|}\right)^2
\end{aligned} k < j − 1 ∑ [ 2 g H ( λ h ∣ k − j ∣ ) − g H ( λ h ∣ k − j + 1∣ ) − g H ( λ h ∣ k − j − 1∣ ) ] 2 ≤ A δ , H 2 j = 0 ∑ n − 1 i = j − 2 ∑ n − 1 ( 2 e − λ δ h ∣ k − j ∣ + e − λ δ h ∣ k − j + 1∣ + e − λ δ h ∣ k − j − 1∣ ) 2
≤ 3 A δ , H 2 ∑ j = 2 n − 1 ∑ r = 2 j ( 4 e − 2 λ δ h r + e − 2 λ δ h ( r − 1 ) + e − 2 λ δ h ( r + 1 ) ) ≤ 18 n A δ , H 2 1 − e − 2 λ δ h . \le3A_{\delta,H}^2\sum_{j=2}^{n-1}\sum_{r=2}^{j}\left(4e^{-2\lambda\delta hr}+e^{-2\lambda\delta h(r-1)}+e^{-2\lambda\delta h(r+1)}\right)
\le\frac{18nA_{\delta,H}^2}{1-e^{-2\lambda\delta h}}. ≤ 3 A δ , H 2 j = 2 ∑ n − 1 r = 2 ∑ j ( 4 e − 2 λ δ h r + e − 2 λ δ h ( r − 1 ) + e − 2 λ δ h ( r + 1 ) ) ≤ 1 − e − 2 λ δ h 18 n A δ , H 2 .
for some constant A δ , H > 0 A_{\delta,H}>0 A δ , H > 0 and δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) . Therefore, combining this with the fact that
lim h → ∞ ϕ h ( x ) = lim h → ∞ ϕ ( h x ) , for any x ≠ 0 = lim h → ∞ ∣ h x ∣ 2 H C h x 2 = 2 Γ ( 2 H ) / ( 2 λ ) 2 H , \begin{aligned}
\lim_{h\to\infty}\phi_h(x)=\lim_{h\to\infty}\phi(hx), \quad\text{for any }x\ne0 \\
&=\lim_{h\to\infty}|hx|^{2H}C_{hx}^2=2\Gamma(2H)/(2\lambda)^{2H},
\end{aligned} h → ∞ lim ϕ h ( x ) = h → ∞ lim ϕ ( h x ) , for any x = 0 = h → ∞ lim ∣ h x ∣ 2 H C h x 2 = 2Γ ( 2 H ) / ( 2 λ ) 2 H ,
the inequality (31) becomes
Var ( F n , h ) ≤ 2 σ 4 n ( 4 ϕ h ( 1 ) 2 + 2 ( ϕ h ( 2 ) − 2 ϕ h ( 1 ) ) 2 + 36 c 2 2 A δ , H 2 1 − e − 2 λ δ h ) ⟶ 0 , \operatorname{Var}(F_{n,h})\le\frac{2\sigma^4}{n}\left(4\phi_h(1)^2+2\left(\phi_h(2)-2\phi_h(1)\right)^2+\frac{36c_2^2A_{\delta,H}^2}{1-e^{-2\lambda\delta h}}\right)
\longrightarrow0, Var ( F n , h ) ≤ n 2 σ 4 ( 4 ϕ h ( 1 ) 2 + 2 ( ϕ h ( 2 ) − 2 ϕ h ( 1 ) ) 2 + 1 − e − 2 λ δ h 36 c 2 2 A δ , H 2 ) ⟶ 0 ,
as h , n → ∞ h,n\to\infty h , n → ∞ . At this stage we can claim that X ‾ n 1 , 2 ( h ) → L 2 ( Ω ) 2 σ 2 Γ ( 2 H ) / ( 2 λ ) 2 H \overline{X}_n^{1,2}(h)\xrightarrow{L^2(\Omega)}2\sigma^2\Gamma(2H)/(2\lambda)^{2H} X n 1 , 2 ( h ) L 2 ( Ω ) 2 σ 2 Γ ( 2 H ) / ( 2 λ ) 2 H as n , h → ∞ n,h\to\infty n , h → ∞ . Now, the continuous mapping theorem with the convergence of H ~ , σ ~ 2 \widetilde{H},\widetilde{\sigma}^2 H , σ 2 and X ‾ n 1 , 2 \overline{X}_n^{1,2} X n 1 , 2 can be properly used to conclude the proof.
Conclusions
In this paper, we have investigated the problem of parameter estimation for a tempered fractional Brownian motion (tfBm) with drift of the form X t = μ t + σ B H , λ ( t ) X_t = \mu t + \sigma B_{H,\lambda}(t) X t = μ t + σ B H , λ ( t ) , where B H , λ ( t ) B_{H,\lambda}(t) B H , λ ( t ) denotes the tfBm. By combining the method of moments with the ergodic theorem, we derived consistent and asymptotically normal estimators for the drift and noise parameters.
First, we showed that the drift estimator μ ^ = X N 1 ( h ) / h \widehat{\mu} = X_N^{1}(h)/h μ = X N 1 ( h ) / h is unbiased, strongly consistent, and asymptotically normal. Next, by applying the generalized method of moments, we constructed joint estimators for the parameters ( σ 2 , H , λ ) (\sigma^2, H, \lambda) ( σ 2 , H , λ ) , proving their strong consistency and asymptotic normality under appropriate regularity conditions. To address identifiability limitations, we adopted a two-step estimation strategy, separating the drift and noise estimation.
Additionally, using Gladyshev’s limit theorems for Gaussian increment processes, we proposed alternative, computationally efficient estimators for the parameters ( H , σ 2 , λ ) (H, \sigma^2, \lambda) ( H , σ 2 , λ ) and established their strong or weak consistency. These results extend existing methods for fractional and tempered fractional models by providing a moment-based estimation framework that avoids heavy reliance on spectral or wavelet techniques.
Future research directions include the derivation of explicit asymptotic distributions for the alternative estimators, the development of efficient numerical algorithms for their implementation, and the extension of the proposed methodology to higher-order or multivariate tempered fractional models.
Technical lemmas
In this section, we present some auxiliary lemmas needed to prove our main results. Lemma 4 is well known [28 p. 413] , while Lemma 2 plays a vital role in our proofs, especially (i) which improves the upper bound given in [10 p. 383] . Consider the function g H g_H g H defined on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) by g H ( x ) = x H K H ( x ) g_H(x)=x^H K_H(x) g H ( x ) = x H K H ( x ) , where K H ( ⋅ ) K_H(\cdot) K H ( ⋅ ) is the modified Bessel function. The norm in the space of bounded functions f : ( 0 , ∞ ) ⟶ R f:(0,\infty)\longrightarrow\mathbb{R} f : ( 0 , ∞ ) ⟶ R is denoted as ∥ f ∥ 0 , ∞ = sup x > 0 ∣ f ( x ) ∣ \|f\|_{0,\infty}=\sup_{x>0}|f(x)| ∥ f ∥ 0 , ∞ = sup x > 0 ∣ f ( x ) ∣ . The following results are obtained.
Lemma 2. For any H > 0 H>0 H > 0 the following statements hold.
(i) For any δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , there exists a constant A δ , H > 0 A_{\delta,H}>0 A δ , H > 0 such that
∣ g H ( x ) ∣ ≤ A δ , H e − δ x |g_H(x)|\leq A_{\delta,H}e^{-\delta x} ∣ g H ( x ) ∣ ≤ A δ , H e − δ x for all x > 0 x>0 x > 0 . Moreover, A δ , H ≤ 2 H − 1 Γ ( H ) ( 1 − δ ) − H A_{\delta,H}\leq2^{H-1}\Gamma(H)(1-\delta)^{-H} A δ , H ≤ 2 H − 1 Γ ( H ) ( 1 − δ ) − H .
(ii) We have 0 < g H ( y ) g H ( x ) ≤ ( y x ) H e − ( y − x ) 0<\frac{g_H(y)}{g_H(x)}\leq\left(\frac{y}{x}\right)^H e^{-(y-x)} 0 < g H ( x ) g H ( y ) ≤ ( x y ) H e − ( y − x ) , for all y ≥ x > 0 y\geq x>0 y ≥ x > 0 .
Remark 3. The statement (ii) is known since 1973 (e.g, [3 ] and references therein). We provide a different and simple proof of it. By using the identity K − H ( x ) = K H ( x ) K_{-H}(x)=K_H(x) K − H ( x ) = K H ( x ) we deduce that (ii)-Lemma 2 holds for H < 0 H<0 H < 0 as well.
If we combine (i) with the lower bound of g H g_H g H given in [10 p. 383] , we obtain
2 H − 1 Γ ( H ) e − x < g H ( x ) ≤ A δ , H e − δ x , for all x > 0. (33) 2^{H-1}\Gamma(H)e^{-x}<g_H(x)\leq A_{\delta,H}e^{-\delta x},\quad\text{for all }x>0. \tag*{(33)} 2 H − 1 Γ ( H ) e − x < g H ( x ) ≤ A δ , H e − δ x , for all x > 0. (33)
If these inequalities are properly used with (ii)-Lemma 2, then for every q > 1 q>1 q > 1 one has
2 H − 1 Γ ( H ) A δ , H e − ( q − δ ) x < g H ( q x ) g H ( x ) < q H e − ( q − 1 ) x , for all x > 0. (34) \frac{2^{H-1}\Gamma(H)}{A_{\delta,H}}e^{-(q-\delta)x}<\frac{g_H(qx)}{g_H(x)}<q^H e^{-(q-1)x},\quad\text{for all }x>0. \tag*{(34)} A δ , H 2 H − 1 Γ ( H ) e − ( q − δ ) x < g H ( x ) g H ( q x ) < q H e − ( q − 1 ) x , for all x > 0. (34)
Proof of Lemma 2. (i) Let p > 1 p>1 p > 1 , a ( p ) : = ( p − 1 ) / p a(p):=(p-1)/p a ( p ) := ( p − 1 ) / p and observe that [22 Eq 10.32.9]
K H ( x ) = ∫ 0 ∞ e − x cosh t cosh ( H t ) d t K_H(x)=\int_0^\infty e^{-x\cosh t}\cosh(Ht)\,\mathrm{d}t K H ( x ) = ∫ 0 ∞ e − x c o s h t cosh ( H t ) d t
for all H , x > 0 H,x>0 H , x > 0 . Therefore, for all x > 0 x>0 x > 0
∣ g H ( x ) ∣ = ∣ x ∣ H ∫ 0 ∞ e − x a ( p ) cosh t ( e − x p cosh t cosh ( H t ) ) d t ≤ ∣ x ∣ H e − x a ( p ) K H ( x p ) ≤ p H ∥ g H ∥ 0 , ∞ e − x a ( p ) ≤ 2 H − 1 p H Γ ( H ) e − x a ( p ) . \begin{aligned}
|g_H(x)|&=|x|^H\int_0^\infty e^{-xa(p)\cosh t}\left(e^{-\frac{x}{p}\cosh t}\cosh(Ht)\right)\,\mathrm{d}t\\
&\leq|x|^H e^{-xa(p)}K_H\left(\frac{x}{p}\right)\leq p^H\|g_H\|_{0,\infty}e^{-xa(p)}\\
&\leq2^{H-1}p^H\Gamma(H)e^{-xa(p)}.
\end{aligned} ∣ g H ( x ) ∣ = ∣ x ∣ H ∫ 0 ∞ e − x a ( p ) c o s h t ( e − p x c o s h t cosh ( H t ) ) d t ≤ ∣ x ∣ H e − x a ( p ) K H ( p x ) ≤ p H ∥ g H ∥ 0 , ∞ e − x a ( p ) ≤ 2 H − 1 p H Γ ( H ) e − x a ( p ) .
In the last inequality above we used the fact that ∣ g H ( x ) ∣ < 2 H − 1 Γ ( H ) |g_H(x)|<2^{H-1}\Gamma(H) ∣ g H ( x ) ∣ < 2 H − 1 Γ ( H ) , ∀ H , x > 0 \forall H,x>0 ∀ H , x > 0 (e.g., [10 p. 383] ). Due to the arbitrariness of p p p , we can choose its value so that a ( p ) = δ a(p)=\delta a ( p ) = δ . Hence, the statement is achieved. Once again, start from [22 Eq 10.32.9] and write
0 < g H ( y ) = ∣ y ∣ H ∫ 0 ∞ e − ( y − x ) cosh t ( e − x cosh t cosh ( H t ) ) d t ≤ ∣ y x ∣ H e − ( y − x ) g H ( x ) . 0<g_H(y)=|y|^H\int_0^\infty e^{-(y-x)\cosh t}\left(e^{-x\cosh t}\cosh(Ht)\right)\,\mathrm{d}t\leq\left|\frac{y}{x}\right|^H e^{-(y-x)}g_H(x). 0 < g H ( y ) = ∣ y ∣ H ∫ 0 ∞ e − ( y − x ) c o s h t ( e − x c o s h t cosh ( H t ) ) d t ≤ x y H e − ( y − x ) g H ( x ) .
This proves (ii), and the proof is complete.
Lemma 3. For any H > 0 H>0 H > 0 the following statements hold true.
(i) The partial derivatives of g H ( z ) g_H(z) g H ( z ) are given as ∂ H { g H ( z ) } = R H ( z ) g H ( z ) \partial_H\{g_H(z)\}=R_H(z)g_H(z) ∂ H { g H ( z )} = R H ( z ) g H ( z ) and ∂ z { g H ( z ) } = − r H ( z ) g H ( z ) \partial_z\{g_H(z)\}=-r_H(z)g_H(z) ∂ z { g H ( z )} = − r H ( z ) g H ( z ) , where r H ( z ) : = K H − 1 ( z ) K H ( z ) r_H(z):=\dfrac{K_{H-1}(z)}{K_H(z)} r H ( z ) := K H ( z ) K H − 1 ( z ) and
R H ( z ) : = log ( z ) + 2 H ( ∫ z ∞ I H ( s ) K H ( s ) s d s − I H ( z ) K H ( z ) ∫ z ∞ K H 2 ( s ) s d s ) . R_H(z):=\log(z)+2H\left(\int_z^\infty\frac{I_H(s)K_H(s)}{s}\,\mathrm{d}s-\frac{I_H(z)}{K_H(z)}\int_z^\infty\frac{K_H^2(s)}{s}\,\mathrm{d}s\right). R H ( z ) := log ( z ) + 2 H ( ∫ z ∞ s I H ( s ) K H ( s ) d s − K H ( z ) I H ( z ) ∫ z ∞ s K H 2 ( s ) d s ) . (ii) r H ( z ) = 1 + o ( 1 ) r_H(z)=1+o(1) r H ( z ) = 1 + o ( 1 ) , as z → ∞ z\to\infty z → ∞ .
(iii) R H ( z ) = log ( z ) + H 2 z + o ( 1 z ) R_H(z)=\log(z)+\frac{H}{2z}+o\left(\frac{1}{z}\right) R H ( z ) = log ( z ) + 2 z H + o ( z 1 ) as z → ∞ z\to\infty z → ∞ .
(iv) ∣ R H ( q z ) − R H ( z ) ∣ ≤ log ( q ) + 3 M ( 2 H + 1 ) |R_H(qz)-R_H(z)|\leq\log(q)+3M(2H+1) ∣ R H ( q z ) − R H ( z ) ∣ ≤ log ( q ) + 3 M ( 2 H + 1 ) , for all q , z ∈ ( 1 , ∞ ) q,z\in(1,\infty) q , z ∈ ( 1 , ∞ ) .
(v) R H ( z ) ≥ log ( z ) − M R_H(z)\geq\log(z)-M R H ( z ) ≥ log ( z ) − M , for all z ∈ ( 1 , ∞ ) z\in(1,\infty) z ∈ ( 1 , ∞ ) .
(vi) ∣ R H ( z ) ∣ ≤ log ( z ) + M ( 6 H + 1 ) |R_H(z)|\leq\log(z)+M(6H+1) ∣ R H ( z ) ∣ ≤ log ( z ) + M ( 6 H + 1 ) , for all z ∈ ( 1 , ∞ ) z\in(1,\infty) z ∈ ( 1 , ∞ ) .
Proof. (i) It is well-known that the modified Bessel function of the second kind K H ( z ) K_H(z) K H ( z ) is differentiable with respect to H H H and z z z . Hence, g H ( z ) g_H(z) g H ( z ) is differentiable as well.
Moreover, by virtue of [10 (A.13)] one has ∂ z g H ( z ) = − z H K H − 1 ( z ) = − r H ( z ) g H ( z ) \partial_z g_H(z)=-z^H K_{H-1}(z)=-r_H(z)g_H(z) ∂ z g H ( z ) = − z H K H − 1 ( z ) = − r H ( z ) g H ( z ) . On the other hand, taking derivatives with respect to H H H with the use of [13 (75)] yields
∂ H { g H ( z ) } = ∂ H { z H K H ( z ) } = log ( z ) z H K H ( z ) + 2 H z H ( K H ( z ) ∫ z ∞ I H ( s ) K H ( s ) s d s − I H ( z ) ∫ z ∞ K H 2 ( s ) s d s ) = g H ( z ) [ log ( z ) + 2 H ( ∫ z ∞ I H ( s ) K H ( s ) s d s − I H ( z ) K H ( z ) ∫ z ∞ K H 2 ( s ) s d s ) ] = : g H ( z ) [ log ( z ) + 2 H ( U H ( z ) − I H ( z ) K H ( z ) V H ( z ) ) ] = R H ( z ) g H ( z ) \begin{aligned}
\partial_H\{g_H(z)\}&=\partial_H\{z^H K_H(z)\}\\
&=\log(z)z^H K_H(z)+2Hz^H\left(K_H(z)\int_z^\infty\frac{I_H(s)K_H(s)}{s}\,\mathrm{d}s-I_H(z)\int_z^\infty\frac{K_H^2(s)}{s}\,\mathrm{d}s\right)\\
&=g_H(z)\left[\log(z)+2H\left(\int_z^\infty\frac{I_H(s)K_H(s)}{s}\,\mathrm{d}s-\frac{I_H(z)}{K_H(z)}\int_z^\infty\frac{K_H^2(s)}{s}\,\mathrm{d}s\right)\right]\\
&=:g_H(z)\left[\log(z)+2H\left(U_H(z)-\frac{I_H(z)}{K_H(z)}V_H(z)\right)\right]\\
&=R_H(z)g_H(z)
\end{aligned} ∂ H { g H ( z )} = ∂ H { z H K H ( z )} = log ( z ) z H K H ( z ) + 2 H z H ( K H ( z ) ∫ z ∞ s I H ( s ) K H ( s ) d s − I H ( z ) ∫ z ∞ s K H 2 ( s ) d s ) = g H ( z ) [ log ( z ) + 2 H ( ∫ z ∞ s I H ( s ) K H ( s ) d s − K H ( z ) I H ( z ) ∫ z ∞ s K H 2 ( s ) d s ) ] =: g H ( z ) [ log ( z ) + 2 H ( U H ( z ) − K H ( z ) I H ( z ) V H ( z ) ) ] = R H ( z ) g H ( z )
say. The statement (ii) is an immediate result of [10 (A.5)] . By using the asymptotic expansions [13 (67)-(68)] of I H I_H I H and K H K_H K H , as z → ∞ z\to\infty z → ∞ we deduce that
K H ( z ) 2 ∼ π 2 z − 1 e − 2 z , I H ( z ) K H ( z ) ∼ 1 2 z (35) K_H(z)^2\sim\frac{\pi}{2}z^{-1}e^{-2z},\qquad I_H(z)K_H(z)\sim\frac{1}{2z} \tag*{(35)} K H ( z ) 2 ∼ 2 π z − 1 e − 2 z , I H ( z ) K H ( z ) ∼ 2 z 1 (35)
and
I H ( z ) K H ( z ) ∼ e 2 z π (36) \frac{I_H(z)}{K_H(z)}\sim\frac{e^{2z}}{\pi} \tag*{(36)} K H ( z ) I H ( z ) ∼ π e 2 z (36)
as z → ∞ z\to\infty z → ∞ . Applying the L’Hopital’s rule with the use of (34) gives U H ( z ) ∼ 1 2 z U_H(z)\sim\frac{1}{2z} U H ( z ) ∼ 2 z 1 and V H ( z ) ∼ π 4 e − 2 z z V_H(z)\sim\frac{\pi}{4}\frac{e^{-2z}}{z} V H ( z ) ∼ 4 π z e − 2 z as z → ∞ z\to\infty z → ∞ . Combine these results with (35) and perform some elementary computations to get the desired result (iii). Let q , z ∈ ( 1 , ∞ ) q,z\in(1,\infty) q , z ∈ ( 1 , ∞ ) then write the difference and rearrange the integrals as
R H ( q z ) − R H ( z ) = log ( z ) + 2 H ( ∫ q z ∞ I H ( s ) K H ( s ) s d s − I H ( q z ) K H ( q z ) ∫ q z ∞ K H 2 ( s ) s d s ) − 2 H ( ∫ z ∞ I H ( s ) K H ( s ) s d s − I H ( z ) K H ( z ) ∫ z ∞ K H 2 ( s ) s d s ) = log ( q ) − 2 H ∫ z q z I H ( s ) K H ( s ) s d s − 2 H ∫ q z ∞ K H 2 ( s ) s d s ( I H ( q z ) K H ( q z ) − I H ( z ) K H ( z ) ) + 2 H I H ( z ) K H ( z ) ∫ z q z K H 2 ( s ) s d s = : log ( q ) − J 1 − J 2 + J 3 . \begin{aligned}
R_H(qz)-R_H(z)&=\log(z)+2H\left(\int_{qz}^\infty\frac{I_H(s)K_H(s)}{s}\,\mathrm{d}s-\frac{I_H(qz)}{K_H(qz)}\int_{qz}^\infty\frac{K_H^2(s)}{s}\,\mathrm{d}s\right)\\
&\quad-2H\left(\int_z^\infty\frac{I_H(s)K_H(s)}{s}\,\mathrm{d}s-\frac{I_H(z)}{K_H(z)}\int_z^\infty\frac{K_H^2(s)}{s}\,\mathrm{d}s\right)\\
&=\log(q)-2H\int_z^{qz}\frac{I_H(s)K_H(s)}{s}\,\mathrm{d}s-2H\int_{qz}^\infty\frac{K_H^2(s)}{s}\,\mathrm{d}s\left(\frac{I_H(qz)}{K_H(qz)}-\frac{I_H(z)}{K_H(z)}\right)+2H\frac{I_H(z)}{K_H(z)}\int_z^{qz}\frac{K_H^2(s)}{s}\,\mathrm{d}s\\
&=:\log(q)-J_1-J_2+J_3.
\end{aligned} R H ( q z ) − R H ( z ) = log ( z ) + 2 H ( ∫ q z ∞ s I H ( s ) K H ( s ) d s − K H ( q z ) I H ( q z ) ∫ q z ∞ s K H 2 ( s ) d s ) − 2 H ( ∫ z ∞ s I H ( s ) K H ( s ) d s − K H ( z ) I H ( z ) ∫ z ∞ s K H 2 ( s ) d s ) = log ( q ) − 2 H ∫ z q z s I H ( s ) K H ( s ) d s − 2 H ∫ q z ∞ s K H 2 ( s ) d s ( K H ( q z ) I H ( q z ) − K H ( z ) I H ( z ) ) + 2 H K H ( z ) I H ( z ) ∫ z q z s K H 2 ( s ) d s =: log ( q ) − J 1 − J 2 + J 3 .
Applying [4 (1.2)-Theorem 1] we claim that I H ( s ) K H ( s ) ≤ M s − 1 / 3 I_H(s)K_H(s)\le Ms^{-1/3} I H ( s ) K H ( s ) ≤ M s − 1/3 , for all s > 0 s>0 s > 0 . Here M M M is the numerical constant defined as
M : = 2 π 3 / 2 3 Γ ( 2 3 ) Γ ( 5 6 ) 2 3 ⋅ sup s ∈ R + s 3 J 0 ( s ) (37) M:=\frac{2\pi^{3/2}}{\sqrt{3}\Gamma\left(\frac{2}{3}\right)\Gamma\left(\frac{5}{6}\right)\sqrt[3]{2}}\cdot\sup_{s\in\mathbb{R}_+}\sqrt[3]{s}J_0(s) \tag*{(37)} M := 3 Γ ( 3 2 ) Γ ( 6 5 ) 3 2 2 π 3/2 ⋅ s ∈ R + sup 3 s J 0 ( s ) (37)
and J ν J_\nu J ν stands for the Bessel function of the first kind. It follows that
∣ J 1 ∣ = 2 H ∫ z q z I H ( s ) K H ( s ) s d s ≤ 2 M H ∫ z q z s − 4 / 3 d s \lvert J_1\rvert=2H\int_z^{qz}\frac{I_H(s)K_H(s)}{s}\,\mathrm{d}s\le2MH\int_z^{qz}s^{-4/3}\,\mathrm{d}s ∣ J 1 ∣ = 2 H ∫ z q z s I H ( s ) K H ( s ) d s ≤ 2 M H ∫ z q z s − 4/3 d s
≤ 6 M H ( 1 − q − 1 / 3 ) z − 1 / 3 < 6 M H . \le6MH(1-q^{-1/3})z^{-1/3}<6MH. ≤ 6 M H ( 1 − q − 1/3 ) z − 1/3 < 6 M H .
and
∣ J 2 ∣ = 2 H ∫ q z ∞ K H 2 ( s ) s d s ∣ I H ( q z ) K H ( q z ) − I H ( z ) K H ( z ) ∣ ≤ 2 H ( I H ( q z ) K H ( q z ) + I H ( z ) K H ( z ) ) ∫ q z ∞ g H 2 ( s ) s 2 H + 1 d s , ( g H ( s ) = s H K H ( s ) ) , ≤ 2 H g H 2 ( q z ) ( I H ( q z ) K H ( q z ) + I H ( z ) K H ( z ) ) ∫ q z ∞ s − 2 H − 1 d s ≤ 2 H g H 2 ( q z ) ( I H ( q z ) K H ( q z ) + I H ( z ) K H ( z ) ) ⋅ ( q z ) − 2 H 2 H = K H 2 ( q z ) ( I H ( q z ) K H ( q z ) + I H ( z ) K H ( z ) ) ≤ I H ( q z ) K H ( q z ) + ( K H ( q z ) K H ( z ) ) 2 [ I H ( z ) K H ( z ) ] ≤ I H ( q z ) K H ( q z ) + ( g H ( q z ) q H g H ( z ) ) 2 [ I H ( z ) K H ( z ) ] ≤ M ( q z ) − 1 / 3 + q − 2 H ⋅ q 2 H e − 2 ( q − 1 ) z ⋅ M z − 1 / 3 < 2 M . (38) \begin{aligned}
|\mathcal{J}_2| &= 2H \int_{qz}^{\infty} \frac{K_H^2(s)}{s}\,ds \left|\frac{I_H(qz)}{K_H(qz)}-\frac{I_H(z)}{K_H(z)}\right| \\
&\le2H\left(\frac{I_H(qz)}{K_H(qz)}+\frac{I_H(z)}{K_H(z)}\right)\int_{qz}^{\infty}\frac{g_H^2(s)}{s^{2H+1}}\,ds,\qquad(g_H(s)=s^H K_H(s)),\\
&\le2H g_H^2(qz)\left(\frac{I_H(qz)}{K_H(qz)}+\frac{I_H(z)}{K_H(z)}\right)\int_{qz}^{\infty}s^{-2H-1}\,ds \\
&\le2H g_H^2(qz)\left(\frac{I_H(qz)}{K_H(qz)}+\frac{I_H(z)}{K_H(z)}\right)\cdot\frac{(qz)^{-2H}}{2H}
=K_H^2(qz)\left(\frac{I_H(qz)}{K_H(qz)}+\frac{I_H(z)}{K_H(z)}\right)\\
&\le I_H(qz)K_H(qz)+\left(\frac{K_H(qz)}{K_H(z)}\right)^2[I_H(z)K_H(z)]\\
&\le I_H(qz)K_H(qz)+\left(\frac{g_H(qz)}{q^H g_H(z)}\right)^2[I_H(z)K_H(z)]\\
&\le M(qz)^{-1/3}+q^{-2H}\cdot q^{2H}e^{-2(q-1)z}\cdot Mz^{-1/3}<2M. \tag*{(38)}
\end{aligned} ∣ J 2 ∣ = 2 H ∫ q z ∞ s K H 2 ( s ) d s K H ( q z ) I H ( q z ) − K H ( z ) I H ( z ) ≤ 2 H ( K H ( q z ) I H ( q z ) + K H ( z ) I H ( z ) ) ∫ q z ∞ s 2 H + 1 g H 2 ( s ) d s , ( g H ( s ) = s H K H ( s )) , ≤ 2 H g H 2 ( q z ) ( K H ( q z ) I H ( q z ) + K H ( z ) I H ( z ) ) ∫ q z ∞ s − 2 H − 1 d s ≤ 2 H g H 2 ( q z ) ( K H ( q z ) I H ( q z ) + K H ( z ) I H ( z ) ) ⋅ 2 H ( q z ) − 2 H = K H 2 ( q z ) ( K H ( q z ) I H ( q z ) + K H ( z ) I H ( z ) ) ≤ I H ( q z ) K H ( q z ) + ( K H ( z ) K H ( q z ) ) 2 [ I H ( z ) K H ( z )] ≤ I H ( q z ) K H ( q z ) + ( q H g H ( z ) g H ( q z ) ) 2 [ I H ( z ) K H ( z )] ≤ M ( q z ) − 1/3 + q − 2 H ⋅ q 2 H e − 2 ( q − 1 ) z ⋅ M z − 1/3 < 2 M . (38)
Equations (37) and (38) are justified by the fact that g H ( ⋅ ) g_H(\cdot) g H ( ⋅ ) is decreasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and (33), respectively. We proceed in a similar way to get
∣ J 3 ∣ = 2 H I H ( z ) K H ( z ) ∫ z q z K H 2 ( s ) s d s ≤ 2 H g H 2 ( z ) I H ( z ) K H ( z ) ∫ z q z s − 2 H − 1 d s ≤ 2 H z 2 H K H 2 ( z ) I H ( z ) K H ( z ) ⋅ z − 2 H − ( q z ) − 2 H 2 H = ( 1 − q − 2 H ) I H ( z ) K H ( z ) ≤ M ( 1 − q − 2 H ) z − 1 / 3 < M . (39) \begin{aligned}
|\mathcal{J}_3| &= 2H\frac{I_H(z)}{K_H(z)}\int_z^{qz}\frac{K_H^2(s)}{s}\,ds\le2Hg_H^2(z)\frac{I_H(z)}{K_H(z)}\int_z^{qz}s^{-2H-1}\,ds\\
&\le2Hz^{2H}K_H^2(z)\frac{I_H(z)}{K_H(z)}\cdot\frac{z^{-2H}-(qz)^{-2H}}{2H}
=(1-q^{-2H})I_H(z)K_H(z)\\
&\le M(1-q^{-2H})z^{-1/3}<M. \tag*{(39)}
\end{aligned} ∣ J 3 ∣ = 2 H K H ( z ) I H ( z ) ∫ z q z s K H 2 ( s ) d s ≤ 2 H g H 2 ( z ) K H ( z ) I H ( z ) ∫ z q z s − 2 H − 1 d s ≤ 2 H z 2 H K H 2 ( z ) K H ( z ) I H ( z ) ⋅ 2 H z − 2 H − ( q z ) − 2 H = ( 1 − q − 2 H ) I H ( z ) K H ( z ) ≤ M ( 1 − q − 2 H ) z − 1/3 < M . (39)
The statement (iv) follows immediately by combining the previous results. In view of (39) we have
2 H I H ( z ) K H ( z ) ∫ z ∞ K H 2 ( s ) s d s = lim q → ∞ 2 H I H ( z ) K H ( z ) ∫ z q z K H 2 ( s ) s d s ≤ M . 2H\frac{I_H(z)}{K_H(z)}\int_z^\infty\frac{K_H^2(s)}{s}\,ds=\lim_{q\to\infty}2H\frac{I_H(z)}{K_H(z)}\int_z^{qz}\frac{K_H^2(s)}{s}\,ds\le M. 2 H K H ( z ) I H ( z ) ∫ z ∞ s K H 2 ( s ) d s = q → ∞ lim 2 H K H ( z ) I H ( z ) ∫ z q z s K H 2 ( s ) d s ≤ M .
From this with I H ( s ) , K H ( s ) > 0 I_H(s), K_H(s)>0 I H ( s ) , K H ( s ) > 0 , ∀ s > 0 \forall s>0 ∀ s > 0 we derive the following
R H ( z ) = log ( z ) + 2 H ( ∫ z ∞ I H ( s ) K H ( s ) s d s − I H ( z ) K H ( z ) ∫ z ∞ K H 2 ( s ) s d s ) ≥ log ( z ) − 2 H I H ( z ) K H ( z ) ∫ z ∞ K H 2 ( s ) s d s ≥ log ( z ) − M , \begin{aligned}
R_H(z)&=\log(z)+2H\left(\int_z^\infty\frac{I_H(s)K_H(s)}{s}\,ds-\frac{I_H(z)}{K_H(z)}\int_z^\infty\frac{K_H^2(s)}{s}\,ds\right)\\
&\ge\log(z)-2H\frac{I_H(z)}{K_H(z)}\int_z^\infty\frac{K_H^2(s)}{s}\,ds\ge\log(z)-M,
\end{aligned} R H ( z ) = log ( z ) + 2 H ( ∫ z ∞ s I H ( s ) K H ( s ) d s − K H ( z ) I H ( z ) ∫ z ∞ s K H 2 ( s ) d s ) ≥ log ( z ) − 2 H K H ( z ) I H ( z ) ∫ z ∞ s K H 2 ( s ) d s ≥ log ( z ) − M ,
which proves the last statement (v). As I H ( z ) I_H(z) I H ( z ) and K H ( z ) K_H(z) K H ( z ) are positive functions, for z > 1 z>1 z > 1 we have
∣ R H ( z ) ∣ ≤ log ( z ) + lim q → ∞ 2 H ∫ z q z I H ( s ) K H ( s ) s d s + lim q → ∞ 2 H I H ( z ) K H ( z ) ∫ z q z K H 2 ( s ) s d s . |R_H(z)|\le\log(z)+\lim_{q\to\infty}2H\int_z^{qz}\frac{I_H(s)K_H(s)}{s}\,ds+\lim_{q\to\infty}2H\frac{I_H(z)}{K_H(z)}\int_z^{qz}\frac{K_H^2(s)}{s}\,ds. ∣ R H ( z ) ∣ ≤ log ( z ) + q → ∞ lim 2 H ∫ z q z s I H ( s ) K H ( s ) d s + q → ∞ lim 2 H K H ( z ) I H ( z ) ∫ z q z s K H 2 ( s ) d s .
Through the previous evaluation of ∣ J 1 ∣ |\mathcal{J}_1| ∣ J 1 ∣ and ∣ J 3 ∣ |\mathcal{J}_3| ∣ J 3 ∣ we conclude that (vi) holds true. Therefore, the proof is complete. □
Lemma 4. (Birkhoff’s ergodic theorem) Let { X k } k = 0 ∞ \{X_k\}_{k=0}^{\infty} { X k } k = 0 ∞ be a stationary ergodic sequence and let g : R ⟶ R g:\mathbb{R}\longrightarrow\mathbb{R} g : R ⟶ R be a measurable function such that E ∣ g ( X 0 ) ∣ < ∞ \mathbb{E}|g(X_0)|<\infty E ∣ g ( X 0 ) ∣ < ∞ , then
1 N ∑ k = 0 N − 1 g ( X k ) → P - a . s . E g ( X 0 ) \frac{1}{N}\sum_{k=0}^{N-1}g(X_k)\xrightarrow{\mathrm{P\text{-}a.s.}} \mathbb{E}g(X_0) N 1 k = 0 ∑ N − 1 g ( X k ) P - a.s. E g ( X 0 ) as $N\to\infty.
Let us state the Gladyashev’s theorem which is essential in establishing the asymptotics given Section 5. Consider a process with stationary Gaussian increments with spectral decomposition of the form
Z ( t ) = Z ( 0 ) + ∫ − ∞ ∞ ( e i ω t − 1 ) d B ^ ( ω ) , (40) Z(t) = Z(0) + \int_{-\infty}^{\infty} (e^{i\omega t}-1)d\widehat{B}(\omega), \tag*{(40)} Z ( t ) = Z ( 0 ) + ∫ − ∞ ∞ ( e iω t − 1 ) d B ( ω ) , (40)
where B ^ ( ω ) \widehat{B}(\omega) B ( ω ) is a stochastic process with independent increments. If E [ ∣ d B ^ ( ω ) ∣ 2 ] = f Z ( ω ) d ω \mathbb{E}[|d\widehat{B}(\omega)|^2] = f_Z(\omega)d\omega E [ ∣ d B ( ω ) ∣ 2 ] = f Z ( ω ) d ω then f Z ( ω ) f_Z(\omega) f Z ( ω ) is said to be the spectral density of Z Z Z . Moreover, the following results are obtained (see [12 p. 57] )
Lemma 5. Let Z ( t ) Z(t) Z ( t ) be a real stochastic process with stationary Gaussian increments and mean 0, having the spectral density f Z ( ω ) f_Z(\omega) f Z ( ω ) such that
f Z ( ω ) ∣ ω ∣ − α = c + O ( ω − 2 ) , as ω → ∞ , (41) \frac{f_Z(\omega)}{|\omega|^{-\alpha}} = c + O(\omega^{-2}), \qquad\text{as } \omega\to\infty, \tag*{(41)} ∣ ω ∣ − α f Z ( ω ) = c + O ( ω − 2 ) , as ω → ∞ , (41) for some c > 0 c > 0 c > 0 and 1 < α < 3 1 < \alpha< 3 1 < α < 3 . Then we have almost surely
lim n → ∞ { 2 − 1 n log 2 ∑ k = 1 2 n [ Z ( k 2 n ) − Z ( k − 1 2 n ) ] 2 } = α (42) \lim_{n\to\infty}\left\{2-\frac{1}{n}\log_2\sum_{k=1}^{2^n}\left[Z\left(\frac{k}{2^n}\right)-Z\left(\frac{k-1}{2^n}\right)\right]^2\right\}=\alpha \tag*{(42)} n → ∞ lim { 2 − n 1 log 2 k = 1 ∑ 2 n [ Z ( 2 n k ) − Z ( 2 n k − 1 ) ] 2 } = α (42) and
lim n → ∞ { − cos ( α π 2 ) Γ ( α ) 2 n ( α − 2 ) − 1 ∑ k = 1 2 n [ Z ( k 2 n ) − Z ( k − 1 2 n ) ] 2 } = c (43) \lim_{n\to\infty}\left\{-\cos\left(\frac{\alpha\pi}{2}\right)\Gamma(\alpha)2^{n(\alpha-2)-1}\sum_{k=1}^{2^n}\left[Z\left(\frac{k}{2^n}\right)-Z\left(\frac{k-1}{2^n}\right)\right]^2\right\}=c \tag*{(43)} n → ∞ lim { − cos ( 2 α π ) Γ ( α ) 2 n ( α − 2 ) − 1 k = 1 ∑ 2 n [ Z ( 2 n k ) − Z ( 2 n k − 1 ) ] 2 } = c (43)
Appendix B.
Local injectivity of the mapping θ ↦ θ ϑ , 3 ( α ) \theta\mapsto\theta_{\vartheta,3}(\alpha) θ ↦ θ ϑ , 3 ( α )
Lemma 6. Let Θ ⊂ ( 0 , ∞ ) × ( 1 / 2 , ∞ ) × ( 0 , ∞ ) \Theta\subset(0,\infty)\times(1/2,\infty)\times(0,\infty) Θ ⊂ ( 0 , ∞ ) × ( 1/2 , ∞ ) × ( 0 , ∞ ) be the compact parameter space, then there is some α 0 > 0 \alpha_0 > 0 α 0 > 0 depending only on Θ \Theta Θ such that ∣ ∇ θ ϑ , 3 ( α ) ∣ < 0 |\nabla\theta_{\vartheta,3}(\alpha)| < 0 ∣∇ θ ϑ , 3 ( α ) ∣ < 0 , for all α ∈ ( α 0 , ∞ ) \alpha\in(\alpha_0,\infty) α ∈ ( α 0 , ∞ ) . In particular, the mapping θ ↦ θ ϑ , 3 ( α ) \theta\mapsto\theta_{\vartheta,3}(\alpha) θ ↦ θ ϑ , 3 ( α ) is locally injective.
Proof. Consider the notations (13), (18) and set Σ ~ 1 , 1 = ( y − 4 x + 3 u 0 ) \widetilde{\Sigma}_{1,1}=(y-4x+3u_0) Σ 1 , 1 = ( y − 4 x + 3 u 0 ) , Σ ~ 2 , 1 = ( z − 4 y + 7 x − 4 u 0 ) \widetilde{\Sigma}_{2,1}=(z-4y+7x-4u_0) Σ 2 , 1 = ( z − 4 y + 7 x − 4 u 0 ) and Σ ~ 3 , 1 = ( t − 4 z + 6 y − 4 x + u 0 ) \widetilde{\Sigma}_{3,1}=(t-4z+6y-4x+u_0) Σ 3 , 1 = ( t − 4 z + 6 y − 4 x + u 0 ) . Evaluating the determinant of ∇ θ ϑ , 3 ( α ) \nabla\theta_{\vartheta,3}(\alpha) ∇ θ ϑ , 3 ( α ) gives
∣ ∇ θ ϑ , 3 ( α ) ∣ = 2 k 3 σ 4 ∣ Σ ~ 1 , 1 ξ Σ ~ 1 , 1 + ( B y − 4 A x + 3 u ∗ ) ξ Σ ~ 1 , 1 − 2 α ( b y − 2 a x ) Σ ~ 2 , 1 ξ Σ ~ 2 , 1 + ( C z − 4 B y + 7 A x − 4 u ∗ ) ξ Σ ~ 2 , 1 − α ( 3 c z − 8 b y + 7 a x ) Σ ~ 3 , 1 ξ Σ ~ 3 , 1 + ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) ξ Σ ~ 3 , 1 − 4 α ( d t − 3 c z + 3 b y − a x ) ∣ = 2 k 3 σ 4 ∣ Σ ~ 1 , 1 ( B y − 4 A x + 3 u ∗ ) − 2 α ( b y − 2 a x ) Σ ~ 2 , 1 ( C z − 4 B y + 7 A x − 4 u ∗ ) − α ( 3 c z − 8 b y + 7 a x ) Σ ~ 3 , 1 ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) − 4 α ( d t − 3 c z + 3 b y − a x ) ∣ = − 2 α k 3 σ 4 ∣ Σ ~ 1 , 1 ( B y − 4 A x + 3 u ∗ ) 2 ( b y − 2 a x ) Σ ~ 2 , 1 ( C z − 4 B y + 7 A x − 4 u ∗ ) ( 3 c z − 8 b y + 7 a x ) Σ ~ 3 , 1 ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) 4 ( d t − 3 c z + 3 b y − a x ) ∣ = − 2 α κ 3 σ 4 ∣ Σ ~ 1 , 1 ( B − A ) y − 3 A ∗ ( 2 b − a ) y − 3 a u 0 Σ ~ 2 , 1 ( C − A ) z + 4 ( A − B ) y + 4 A ∗ ( 3 c − a ) z + 4 ( a − 2 b ) y + 4 a u 0 Σ ~ 3 , 1 ( D − A ) t + 4 ( A − C ) z + 6 ( B − A ) y − A ∗ ( 4 d − a ) t + 4 ( a − 3 c ) z + 6 ( 2 b − a ) y − a u 0 ∣ \begin{aligned}
\left|\nabla\theta_{\vartheta,3}(\alpha)\right|
=2k^3\sigma^4
\left|
\begin{array}{ccc}
\widetilde{\Sigma}_{1,1} & \xi\widetilde{\Sigma}_{1,1}+(By-4Ax+3u^*) & \xi\widetilde{\Sigma}_{1,1}-2\alpha(by-2ax)\\
\widetilde{\Sigma}_{2,1} & \xi\widetilde{\Sigma}_{2,1}+(Cz-4By+7Ax-4u^*) & \xi\widetilde{\Sigma}_{2,1}-\alpha(3cz-8by+7ax)\\
\widetilde{\Sigma}_{3,1} & \xi\widetilde{\Sigma}_{3,1}+(Dt-4Cz+6By-4Ax+u^*) & \xi\widetilde{\Sigma}_{3,1}-4\alpha(dt-3cz+3by-ax)
\end{array}
\right| \\
&=2k^3\sigma^4
\left|
\begin{array}{ccc}
\widetilde{\Sigma}_{1,1} & (By-4Ax+3u^*) & -2\alpha(by-2ax)\\
\widetilde{\Sigma}_{2,1} & (Cz-4By+7Ax-4u^*) & -\alpha(3cz-8by+7ax)\\
\widetilde{\Sigma}_{3,1} & (Dt-4Cz+6By-4Ax+u^*) & -4\alpha(dt-3cz+3by-ax)
\end{array}
\right| \\
&=-2\alpha k^3\sigma^4
\left|
\begin{array}{ccc}
\widetilde{\Sigma}_{1,1} & (By-4Ax+3u^*) & 2(by-2ax)\\
\widetilde{\Sigma}_{2,1} & (Cz-4By+7Ax-4u^*) & (3cz-8by+7ax)\\
\widetilde{\Sigma}_{3,1} & (Dt-4Cz+6By-4Ax+u^*) & 4(dt-3cz+3by-ax)
\end{array}
\right| \\
&=-2\alpha\kappa^{3}\sigma^{4}
\begin{vmatrix}
\widetilde{\Sigma}_{1,1} & (B-A)y-3A^{*} & (2b-a)y-3au_{0}\\
\widetilde{\Sigma}_{2,1} & (C-A)z+4(A-B)y+4A^{*} & (3c-a)z+4(a-2b)y+4au_{0}\\
\widetilde{\Sigma}_{3,1} & (D-A)t+4(A-C)z+6(B-A)y-A^{*} & (4d-a)t+4(a-3c)z+6(2b-a)y-au_{0}
\end{vmatrix}
\end{aligned} ∣ ∇ θ ϑ , 3 ( α ) ∣ = 2 k 3 σ 4 Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 ξ Σ 1 , 1 + ( B y − 4 A x + 3 u ∗ ) ξ Σ 2 , 1 + ( C z − 4 B y + 7 A x − 4 u ∗ ) ξ Σ 3 , 1 + ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) ξ Σ 1 , 1 − 2 α ( b y − 2 a x ) ξ Σ 2 , 1 − α ( 3 cz − 8 b y + 7 a x ) ξ Σ 3 , 1 − 4 α ( d t − 3 cz + 3 b y − a x ) = 2 k 3 σ 4 Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 ( B y − 4 A x + 3 u ∗ ) ( C z − 4 B y + 7 A x − 4 u ∗ ) ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) − 2 α ( b y − 2 a x ) − α ( 3 cz − 8 b y + 7 a x ) − 4 α ( d t − 3 cz + 3 b y − a x ) = − 2 α k 3 σ 4 Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 ( B y − 4 A x + 3 u ∗ ) ( C z − 4 B y + 7 A x − 4 u ∗ ) ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) 2 ( b y − 2 a x ) ( 3 cz − 8 b y + 7 a x ) 4 ( d t − 3 cz + 3 b y − a x ) = − 2 α κ 3 σ 4 Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 ( B − A ) y − 3 A ∗ ( C − A ) z + 4 ( A − B ) y + 4 A ∗ ( D − A ) t + 4 ( A − C ) z + 6 ( B − A ) y − A ∗ ( 2 b − a ) y − 3 a u 0 ( 3 c − a ) z + 4 ( a − 2 b ) y + 4 a u 0 ( 4 d − a ) t + 4 ( a − 3 c ) z + 6 ( 2 b − a ) y − a u 0
A ∗ = A u 0 − u ∗ = − 2 α κ 3 σ 4 ∣ ( y − 4 x ) + 3 u 0 ( B − A ) y − 3 A ∗ ( 2 b − a ) y − 3 a u 0 ( z − 9 x ) + 8 u 0 ( C − A ) z − 8 A ∗ ( 3 c − a ) z − 8 a u 0 Σ ~ 3 , 1 ( D − A ) t + 4 ( A − C ) z + 6 ( B − A ) y − A ∗ ( 4 d − a ) t + 4 ( a − 3 c ) z + 6 ( 2 b − a ) y − a u 0 ∣ (44) \begin{aligned}
A^{*}=Au_{0}-u^{*} \\
&=-2\alpha\kappa^{3}\sigma^{4}
\begin{vmatrix}
(y-4x)+3u_{0} & (B-A)y-3A^{*} & (2b-a)y-3au_{0}\\
(z-9x)+8u_{0} & (C-A)z-8A^{*} & (3c-a)z-8au_{0}\\
\widetilde{\Sigma}_{3,1} & (D-A)t+4(A-C)z+6(B-A)y-A^{*} & (4d-a)t+4(a-3c)z+6(2b-a)y-au_{0}
\end{vmatrix} \tag*{(44)}
\end{aligned} A ∗ = A u 0 − u ∗ = − 2 α κ 3 σ 4 ( y − 4 x ) + 3 u 0 ( z − 9 x ) + 8 u 0 Σ 3 , 1 ( B − A ) y − 3 A ∗ ( C − A ) z − 8 A ∗ ( D − A ) t + 4 ( A − C ) z + 6 ( B − A ) y − A ∗ ( 2 b − a ) y − 3 a u 0 ( 3 c − a ) z − 8 a u 0 ( 4 d − a ) t + 4 ( a − 3 c ) z + 6 ( 2 b − a ) y − a u 0 (44)
= − 2 α κ 3 σ 4 ∣ ( y − 4 x ) + 3 u 0 ( B − A ) y − 3 A ∗ ( 2 b − a ) y − 3 a u 0 ( z − 9 x ) + 8 u 0 ( C − A ) z − 8 A ∗ ( 3 c − a ) z − 8 a u 0 ( t − 16 x ) + 15 u 0 ( D − A ) t − 15 A ∗ ( 4 d − a ) t − 15 a u 0 ∣ = − 2 α κ 3 σ 4 { ∣ ( y − 4 x ) ( B − A ) y ( 2 b − a ) y ( z − 9 x ) ( C − A ) z ( 3 c − a ) z ( t − 16 x ) ( D − A ) t ( 4 d − a ) t ∣ + ∣ ( y − 4 x ) ( B − A ) y − 3 a u 0 ( z − 9 x ) ( C − A ) z − 8 a u 0 ( t − 16 x ) ( D − A ) t − 15 a u 0 ∣ (45) \begin{aligned}
=-2\alpha\kappa^{3}\sigma^{4}
\begin{vmatrix}
(y-4x)+3u_{0} & (B-A)y-3A^{*} & (2b-a)y-3au_{0}\\
(z-9x)+8u_{0} & (C-A)z-8A^{*} & (3c-a)z-8au_{0}\\
(t-16x)+15u_{0} & (D-A)t-15A^{*} & (4d-a)t-15au_{0}
\end{vmatrix} \\
&=-2\alpha\kappa^{3}\sigma^{4}
\left\{
\begin{vmatrix}
(y-4x) & (B-A)y & (2b-a)y\\
(z-9x) & (C-A)z & (3c-a)z\\
(t-16x) & (D-A)t & (4d-a)t
\end{vmatrix}
+
\begin{vmatrix}
(y-4x) & (B-A)y & -3au_{0}\\
(z-9x) & (C-A)z & -8au_{0}\\
(t-16x) & (D-A)t & -15au_{0}
\end{vmatrix}
\right. \tag*{(45)}
\end{aligned} = − 2 α κ 3 σ 4 ( y − 4 x ) + 3 u 0 ( z − 9 x ) + 8 u 0 ( t − 16 x ) + 15 u 0 ( B − A ) y − 3 A ∗ ( C − A ) z − 8 A ∗ ( D − A ) t − 15 A ∗ ( 2 b − a ) y − 3 a u 0 ( 3 c − a ) z − 8 a u 0 ( 4 d − a ) t − 15 a u 0 = − 2 α κ 3 σ 4 ⎩ ⎨ ⎧ ( y − 4 x ) ( z − 9 x ) ( t − 16 x ) ( B − A ) y ( C − A ) z ( D − A ) t ( 2 b − a ) y ( 3 c − a ) z ( 4 d − a ) t + ( y − 4 x ) ( z − 9 x ) ( t − 16 x ) ( B − A ) y ( C − A ) z ( D − A ) t − 3 a u 0 − 8 a u 0 − 15 a u 0 (45)
+ ∣ ( y − 4 x ) − 3 A ∗ ( 2 b − a ) y ( z − 9 x ) − 8 A ∗ ( 3 c − a ) z ( t − 16 x ) − 15 A ∗ ( 4 d − a ) t ∣ + ∣ 3 u 0 ( B − A ) y ( 2 b − a ) y 8 u 0 ( C − A ) z ( 3 c − a ) z 15 u 0 ( D − A ) t ( 4 d − a ) t ∣ } = − 2 α κ 3 σ 4 { ∣ ( y / x − 4 ) ( B − A ) ( 2 b − a ) y ( z / x − 9 ) ( C − A ) z / y ( 3 c − a ) z ( t / x − 16 ) ( D − A ) t / y ( 4 d − a ) t ∣ + x y ∣ ( y / x − 4 ) ( B − A ) − 3 a u 0 ( z / x − 9 ) ( C − A ) z / y − 8 a u 0 ( t / x − 16 ) ( D − A ) t / y − 15 a u 0 ∣ \begin{aligned}
\left.
+
\begin{vmatrix}
(y-4x) & -3A^{*} & (2b-a)y\\
(z-9x) & -8A^{*} & (3c-a)z\\
(t-16x) & -15A^{*} & (4d-a)t
\end{vmatrix}
+
\begin{vmatrix}
3u_{0} & (B-A)y & (2b-a)y\\
8u_{0} & (C-A)z & (3c-a)z\\
15u_{0} & (D-A)t & (4d-a)t
\end{vmatrix}
\right\} \\
&=-2\alpha\kappa^{3}\sigma^{4}
\left\{
\begin{vmatrix}
(y/x-4) & (B-A) & (2b-a)y\\
(z/x-9) & (C-A)z/y & (3c-a)z\\
(t/x-16) & (D-A)t/y & (4d-a)t
\end{vmatrix}
+xy
\begin{vmatrix}
(y/x-4) & (B-A) & -3au_{0}\\
(z/x-9) & (C-A)z/y & -8au_{0}\\
(t/x-16) & (D-A)t/y & -15au_{0}
\end{vmatrix}
\right.
\end{aligned} + ( y − 4 x ) ( z − 9 x ) ( t − 16 x ) − 3 A ∗ − 8 A ∗ − 15 A ∗ ( 2 b − a ) y ( 3 c − a ) z ( 4 d − a ) t + 3 u 0 8 u 0 15 u 0 ( B − A ) y ( C − A ) z ( D − A ) t ( 2 b − a ) y ( 3 c − a ) z ( 4 d − a ) t ⎭ ⎬ ⎫ = − 2 α κ 3 σ 4 ⎩ ⎨ ⎧ ( y / x − 4 ) ( z / x − 9 ) ( t / x − 16 ) ( B − A ) ( C − A ) z / y ( D − A ) t / y ( 2 b − a ) y ( 3 c − a ) z ( 4 d − a ) t + x y ( y / x − 4 ) ( z / x − 9 ) ( t / x − 16 ) ( B − A ) ( C − A ) z / y ( D − A ) t / y − 3 a u 0 − 8 a u 0 − 15 a u 0
+ A ∗ x y ∣ ( y / x − 4 ) − 3 ( 2 b − a ) ( z / x − 9 ) − 8 ( 3 c − a ) z / y ( t / x − 16 ) − 15 ( 4 d − a ) t / y ∣ + x y ∣ 3 u 0 ( B − A ) y / x ( 2 b − a ) 8 u 0 ( C − A ) z / x ( 3 c − a ) z / y 15 u 0 ( D − A ) t / x ( 4 d − a ) t / y ∣ } = − 2 α κ 3 σ 4 { o ( x y ) + O ( x y ) + 87 A ∗ x y ( 1 + o ( 1 ) ) + o ( x y ) } , as α → ∞ ∼ − 2 α κ 3 σ 4 A ∗ x y , as α → ∞ . \begin{aligned}
\left.
+A^{*}xy
\begin{vmatrix}
(y/x-4) & -3 & (2b-a)\\
(z/x-9) & -8 & (3c-a)z/y\\
(t/x-16) & -15 & (4d-a)t/y
\end{vmatrix}
+xy
\begin{vmatrix}
3u_{0} & (B-A)y/x & (2b-a)\\
8u_{0} & (C-A)z/x & (3c-a)z/y\\
15u_{0} & (D-A)t/x & (4d-a)t/y
\end{vmatrix}
\right\} \\
&=-2\alpha\kappa^{3}\sigma^{4}\{o(xy)+O(xy)+87A^{*}xy(1+o(1))+o(xy)\},\quad\text{as }\alpha\to\infty\\
&\sim-2\alpha\kappa^{3}\sigma^{4}A^{*}xy,\quad\text{as }\alpha\to\infty.
\end{aligned} + A ∗ x y ( y / x − 4 ) ( z / x − 9 ) ( t / x − 16 ) − 3 − 8 − 15 ( 2 b − a ) ( 3 c − a ) z / y ( 4 d − a ) t / y + x y 3 u 0 8 u 0 15 u 0 ( B − A ) y / x ( C − A ) z / x ( D − A ) t / x ( 2 b − a ) ( 3 c − a ) z / y ( 4 d − a ) t / y ⎭ ⎬ ⎫ = − 2 α κ 3 σ 4 { o ( x y ) + O ( x y ) + 87 A ∗ x y ( 1 + o ( 1 )) + o ( x y )} , as α → ∞ ∼ − 2 α κ 3 σ 4 A ∗ x y , as α → ∞.
We justify the above asymptotics by means of (33) , (34) and Lemma 3. For example y / x = g H ( 2 λ α ) / g H ( λ α ) = O ( e − λ α ) y/x=g_{H}(2\lambda\alpha)/g_{H}(\lambda\alpha)=O(e^{-\lambda\alpha}) y / x = g H ( 2 λ α ) / g H ( λ α ) = O ( e − λ α ) , a = r H ( λ α ) = 1 + o ( 1 ) a=r_{H}(\lambda\alpha)=1+o(1) a = r H ( λ α ) = 1 + o ( 1 ) , b = r H ( 2 λ α ) = 1 + o ( 1 ) b=r_{H}(2\lambda\alpha)=1+o(1) b = r H ( 2 λ α ) = 1 + o ( 1 ) and A ∗ = A u 0 − u ∗ = u 0 r H ( λ α ) − u ∗ ∼ u 0 log ( λ α ) A^{*}=Au_{0}-u^{*}=u_{0}r_{H}(\lambda\alpha)-u^{*}\sim u_{0}\log(\lambda\alpha) A ∗ = A u 0 − u ∗ = u 0 r H ( λ α ) − u ∗ ∼ u 0 log ( λ α ) , etc, as α → ∞ \alpha\to\infty α → ∞ . Although ∣ ∇ ϱ θ , 3 ( α ) ∣ |\nabla\varrho_{\theta,3}(\alpha)| ∣∇ ϱ θ , 3 ( α ) ∣ converges rapidly to zero, as α → ∞ \alpha\to\infty α → ∞ , we can find some α \alpha α large enough so that ∣ ∇ ϱ θ , 3 ( α ) ∣ < 0 |\nabla\varrho_{\theta,3}(\alpha)|<0 ∣∇ ϱ θ , 3 ( α ) ∣ < 0 . This proves the identifiability condition for α \alpha α large enough. For a practical purpose, we need to specify the range of α \alpha α , for which ∣ ∇ ϱ θ , 3 ( α ) ∣ < 0 |\nabla\varrho_{\theta,3}(\alpha)|<0 ∣∇ ϱ θ , 3 ( α ) ∣ < 0 .
Observe that each column in (45) can be expressed as ( F ( q , ξ ) ) q = 2 , 3 , 4 (F(q,\xi))_{q=2,3,4} ( F ( q , ξ ) ) q = 2 , 3 , 4 with ξ = λ α \xi=\lambda\alpha ξ = λ α . So, we simplify notations by denoting the determinant as det { ⋯ } \det\{\,\cdots\,\} det { ⋯ } where each expression inside it represents a column vector. The following additional notations are considered
g ( ξ ) = g H ( ξ ) , g ′ ( ξ ) = g H − 1 ( ξ ) , R ( ξ ) = R H ( ξ ) , g(\xi)=g_{H}(\xi),\quad g'(\xi)=g_{H-1}(\xi),\quad R(\xi)=R_{H}(\xi), g ( ξ ) = g H ( ξ ) , g ′ ( ξ ) = g H − 1 ( ξ ) , R ( ξ ) = R H ( ξ ) ,
φ q ( ξ ) = g ( q ξ ) g ( ξ ) , φ q ′ ( ξ ) = g ′ ( q ξ ) g ′ ( ξ ) \varphi_{q}(\xi)=\frac{g(q\xi)}{g(\xi)},\quad\varphi'_{q}(\xi)=\frac{g'(q\xi)}{g'(\xi)} φ q ( ξ ) = g ( ξ ) g ( q ξ ) , φ q ′ ( ξ ) = g ′ ( ξ ) g ′ ( q ξ )
x ′ = g ′ ( ξ ) , y ′ = g ′ ( 2 ξ ) , z ′ = g ′ ( 3 ξ ) , t ′ = g ′ ( 4 ξ ) . x'=g'(\xi),\quad y'=g'(2\xi),\quad z'=g'(3\xi),\quad t'=g'(4\xi). x ′ = g ′ ( ξ ) , y ′ = g ′ ( 2 ξ ) , z ′ = g ′ ( 3 ξ ) , t ′ = g ′ ( 4 ξ ) .
By definition (see Lemma 3) of r H ( ⋅ ) r_H(\cdot) r H ( ⋅ ) we have a = r H ( ξ ) = ξ ( x ′ / x ) a=r_H(\xi)=\xi(x'/x) a = r H ( ξ ) = ξ ( x ′ / x ) . Similarly, b = 2 ξ ( y ′ / y ) b=2\xi(y'/y) b = 2 ξ ( y ′ / y ) , c = 3 ξ ( z ′ / z ) c=3\xi(z'/z) c = 3 ξ ( z ′ / z ) and d = 4 ξ ( t ′ / t ) d=4\xi(t'/t) d = 4 ξ ( t ′ / t ) . Therefore, equation (45) becomes
∇ θ , 3 ( α ) = − 2 α κ 3 σ 4 ∣ ( y − 4 x ) + 3 u 0 ( B − A ) y − 3 A ∗ ( 4 ξ y ′ y − ξ x ′ x ) y − 3 u 0 ξ x ′ x ( z − 9 x ) + 8 u 0 ( C − A ) z − 8 A ∗ ( 9 ξ z ′ z − ξ x ′ x ) z − 8 u 0 ξ x ′ x ( t − 16 x ) + 15 u 0 ( D − A ) t − 15 A ∗ ( 16 ξ t ′ t − ξ x ′ x ) t − 15 u 0 ξ x ′ x ∣ = − 2 α κ 3 σ 4 ξ x ∣ ( y − 4 x ) + 3 u 0 ( B − A ) y − 3 A ∗ ( 4 x y ′ − x ′ y ) − 3 u 0 x ′ ( z − 9 x ) + 8 u 0 ( C − A ) z − 8 A ∗ ( 9 x z ′ − x ′ z ) − 8 u 0 x ′ ( t − 16 x ) + 15 u 0 ( D − A ) t − 15 A ∗ ( 16 x t ′ − x ′ t ) − 15 u 0 x ′ ∣ = − 2 α κ 3 σ 4 ξ x ∣ ( y − 4 x ) + 3 u 0 ( B − A ) y − 3 A ∗ 4 ( x y ′ − x x ′ ) ( z − 9 x ) + 8 u 0 ( C − A ) z − 8 A ∗ 9 ( x z ′ − x x ′ ) ( t − 16 x ) + 15 u 0 ( D − A ) t − 15 A ∗ 16 ( x t ′ − x x ′ ) ∣ = − 2 α κ 3 σ 4 ξ ∣ ( y − 4 x ) + 3 u 0 ( B − A ) y − 3 A ∗ 4 ( y ′ − x ′ ) ( z − 9 x ) + 8 u 0 ( C − A ) z − 8 A ∗ 9 ( z ′ − x ′ ) ( t − 16 x ) + 15 u 0 ( D − A ) t − 15 A ∗ 16 ( t ′ − x ′ ) ∣ = − 2 α κ 3 σ 4 ξ det { [ g ( q ξ ) − q 2 g ( ξ ) ] + ( q 2 − 1 ) u 0 ; [ R ( q ξ ) − R ( ξ ) ] g ( q ξ ) + ( 1 − q 2 ) A ∗ ; q 2 [ g ′ ( q ξ ) − g ′ ( ξ ) ] } = − 2 α κ 3 σ 4 ξ g ( 2 ξ ) g ′ ( ξ ) A ∗ det { C 1 + C 2 ; C 3 + C 4 ; C 5 } = − 2 α κ 3 σ 4 ξ g ( 2 ξ ) g ′ ( ξ ) A ∗ [ det { C 1 ; C 3 ; C 5 } + det { C 1 ; C 4 ; C 5 } + det { C 2 ; C 3 ; C 5 } ] (46) \begin{aligned}
\nabla_{\theta,3}(\alpha)&=-2\alpha\kappa^3\sigma^4
\begin{vmatrix}
(y-4x)+3u_0 & (B-A)y-3A^* & \left(4\xi\frac{y'}{y}-\xi\frac{x'}{x}\right)y-3u_0\xi\frac{x'}{x}\\
(z-9x)+8u_0 & (C-A)z-8A^* & \left(9\xi\frac{z'}{z}-\xi\frac{x'}{x}\right)z-8u_0\xi\frac{x'}{x}\\
(t-16x)+15u_0 & (D-A)t-15A^* & \left(16\xi\frac{t'}{t}-\xi\frac{x'}{x}\right)t-15u_0\xi\frac{x'}{x}
\end{vmatrix}\\
&=-2\alpha\kappa^3\sigma^4\frac{\xi}{x}
\begin{vmatrix}
(y-4x)+3u_0 & (B-A)y-3A^* & (4xy'-x'y)-3u_0x'\\
(z-9x)+8u_0 & (C-A)z-8A^* & (9xz'-x'z)-8u_0x'\\
(t-16x)+15u_0 & (D-A)t-15A^* & (16xt'-x't)-15u_0x'
\end{vmatrix}\\
&=-2\alpha\kappa^3\sigma^4\frac{\xi}{x}
\begin{vmatrix}
(y-4x)+3u_0 & (B-A)y-3A^* & 4(xy'-xx')\\
(z-9x)+8u_0 & (C-A)z-8A^* & 9(xz'-xx')\\
(t-16x)+15u_0 & (D-A)t-15A^* & 16(xt'-xx')
\end{vmatrix}\\
&=-2\alpha\kappa^3\sigma^4\xi
\begin{vmatrix}
(y-4x)+3u_0 & (B-A)y-3A^* & 4(y'-x')\\
(z-9x)+8u_0 & (C-A)z-8A^* & 9(z'-x')\\
(t-16x)+15u_0 & (D-A)t-15A^* & 16(t'-x')
\end{vmatrix}\\
&=-2\alpha\kappa^3\sigma^4\xi\det\left\{[g(q\xi)-q^2g(\xi)]+(q^2-1)u_0;\,[R(q\xi)-R(\xi)]g(q\xi)+(1-q^2)A^*;q^2[g'(q\xi)-g'(\xi)]\right\}\\
&=-2\alpha\kappa^3\sigma^4\xi g(2\xi)g'(\xi)A^*\det\{C_1+C_2;C_3+C_4;C_5\}\\
&=-2\alpha\kappa^3\sigma^4\xi g(2\xi)g'(\xi)A^*\left[\det\{C_1;C_3;C_5\}+\det\{C_1;C_4;C_5\}+\det\{C_2;C_3;C_5\}\right] \tag*{(46)}
\end{aligned} ∇ θ , 3 ( α ) = − 2 α κ 3 σ 4 ( y − 4 x ) + 3 u 0 ( z − 9 x ) + 8 u 0 ( t − 16 x ) + 15 u 0 ( B − A ) y − 3 A ∗ ( C − A ) z − 8 A ∗ ( D − A ) t − 15 A ∗ ( 4 ξ y y ′ − ξ x x ′ ) y − 3 u 0 ξ x x ′ ( 9 ξ z z ′ − ξ x x ′ ) z − 8 u 0 ξ x x ′ ( 16 ξ t t ′ − ξ x x ′ ) t − 15 u 0 ξ x x ′ = − 2 α κ 3 σ 4 x ξ ( y − 4 x ) + 3 u 0 ( z − 9 x ) + 8 u 0 ( t − 16 x ) + 15 u 0 ( B − A ) y − 3 A ∗ ( C − A ) z − 8 A ∗ ( D − A ) t − 15 A ∗ ( 4 x y ′ − x ′ y ) − 3 u 0 x ′ ( 9 x z ′ − x ′ z ) − 8 u 0 x ′ ( 16 x t ′ − x ′ t ) − 15 u 0 x ′ = − 2 α κ 3 σ 4 x ξ ( y − 4 x ) + 3 u 0 ( z − 9 x ) + 8 u 0 ( t − 16 x ) + 15 u 0 ( B − A ) y − 3 A ∗ ( C − A ) z − 8 A ∗ ( D − A ) t − 15 A ∗ 4 ( x y ′ − x x ′ ) 9 ( x z ′ − x x ′ ) 16 ( x t ′ − x x ′ ) = − 2 α κ 3 σ 4 ξ ( y − 4 x ) + 3 u 0 ( z − 9 x ) + 8 u 0 ( t − 16 x ) + 15 u 0 ( B − A ) y − 3 A ∗ ( C − A ) z − 8 A ∗ ( D − A ) t − 15 A ∗ 4 ( y ′ − x ′ ) 9 ( z ′ − x ′ ) 16 ( t ′ − x ′ ) = − 2 α κ 3 σ 4 ξ det { [ g ( q ξ ) − q 2 g ( ξ )] + ( q 2 − 1 ) u 0 ; [ R ( q ξ ) − R ( ξ )] g ( q ξ ) + ( 1 − q 2 ) A ∗ ; q 2 [ g ′ ( q ξ ) − g ′ ( ξ )] } = − 2 α κ 3 σ 4 ξ g ( 2 ξ ) g ′ ( ξ ) A ∗ det { C 1 + C 2 ; C 3 + C 4 ; C 5 } = − 2 α κ 3 σ 4 ξ g ( 2 ξ ) g ′ ( ξ ) A ∗ [ det { C 1 ; C 3 ; C 5 } + det { C 1 ; C 4 ; C 5 } + det { C 2 ; C 3 ; C 5 } ] (46)
where
A ∗ = u 0 R ( ξ ) − u ∗ , u 0 = 2 H − 1 Γ ( H ) , C 1 = g ( ξ ) [ φ q ( ξ ) − q 2 ] , C 2 = ( q 2 − 1 ) u 0 , (47) A^*=u_0R(\xi)-u^*,\quad u_0=2^{H-1}\Gamma(H),\quad C_1=g(\xi)[\varphi_q(\xi)-q^2],\quad C_2=(q^2-1)u_0, \tag*{(47)} A ∗ = u 0 R ( ξ ) − u ∗ , u 0 = 2 H − 1 Γ ( H ) , C 1 = g ( ξ ) [ φ q ( ξ ) − q 2 ] , C 2 = ( q 2 − 1 ) u 0 , (47)
C 3 = g ( q ξ ) A ∗ [ R ( q ξ ) − R ( ξ ) ] , C 4 = ( 1 − q 2 ) , C 5 = [ q 2 φ q ′ ( ξ ) − q 2 ] / g ( 2 ξ ) (48) C_3=\frac{g(q\xi)}{A^*}[R(q\xi)-R(\xi)],\quad C_4=(1-q^2),\quad C_5=[q^2\varphi'_q(\xi)-q^2]/g(2\xi) \tag*{(48)} C 3 = A ∗ g ( q ξ ) [ R ( q ξ ) − R ( ξ )] , C 4 = ( 1 − q 2 ) , C 5 = [ q 2 φ q ′ ( ξ ) − q 2 ] / g ( 2 ξ ) (48)
Assume that H ∈ K : = [ H , H ‾ ] H\in\mathcal{K}:=[H,\overline{H}] H ∈ K := [ H , H ] and λ ∈ [ λ ‾ , λ ‾ ] \lambda\in[\underline{\lambda},\overline{\lambda}] λ ∈ [ λ , λ ] . This hypothesis is technical and used to define the constants M ∗ , M j M^*,M_j M ∗ , M j , j = 0 , 1 , … , 5 j=0,1,\ldots,5 j = 0 , 1 , … , 5 by maximizing functions in H H H or λ \lambda λ . So, it can be dropped by considering the same functions in θ = ( σ 2 , H , λ ) \theta=(\sigma^2,H,\lambda) θ = ( σ 2 , H , λ ) and maximizing them over the parameter space Θ \Theta Θ . Let δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) to be specified later. We shall start the evaluation of the determinants appearing in (46).
i-Evaluating the determinant det { C 1 ; C 3 ; C 5 } \det\{C_1;C_3;C_5\} det { C 1 ; C 3 ; C 5 } . We apply (33) and (34) to get
{ u 0 e − q ξ < g ( q ξ ) ≤ u 0 ( 1 − δ ) − H e − q δ ξ , ( 1 − δ ) H e − ( q − δ ) ξ < φ q ( ξ ) < q H e − ( q − 1 ) ξ , 0 < φ q ′ ( ξ ) < q H − 1 e − ( q − 1 ) ξ (49) \begin{cases}
u_0e^{-q\xi}<g(q\xi)\leq u_0(1-\delta)^{-H}e^{-q\delta\xi},\\
(1-\delta)^He^{-(q-\delta)\xi}<\varphi_q(\xi)<q^He^{-(q-1)\xi},\\
0<\varphi'_q(\xi)<q^{H-1}e^{-(q-1)\xi} \tag*{(49)}
\end{cases} ⎩ ⎨ ⎧ u 0 e − q ξ < g ( q ξ ) ≤ u 0 ( 1 − δ ) − H e − q δ ξ , ( 1 − δ ) H e − ( q − δ ) ξ < φ q ( ξ ) < q H e − ( q − 1 ) ξ , 0 < φ q ′ ( ξ ) < q H − 1 e − ( q − 1 ) ξ (49)
for all ξ , H > 0 \xi,H>0 ξ , H > 0 and q = 2 , 3 , 4 q=2,3,4 q = 2 , 3 , 4 . In particular, for all ξ > 0 \xi>0 ξ > 0
1 / φ 2 ( ξ ) ≤ ( 1 − δ ) − H e ( 2 − δ ) ξ (50) 1/\varphi_2(\xi)\leq(1-\delta)^{-H}e^{(2-\delta)\xi} \tag*{(50)} 1/ φ 2 ( ξ ) ≤ ( 1 − δ ) − H e ( 2 − δ ) ξ (50)
and
{ g ( q ξ ) / g ( 2 ξ ) = 1 , for q = 2 g ( q ξ ) / g ( 2 ξ ) = φ q ( ξ ) / φ 2 ( ξ ) ≤ q H ( 1 − δ ) − H e − [ q − 3 + δ ] ξ , for q = 3 , 4. (51) \begin{cases}
g(q\xi)/g(2\xi)=1, & \text{for }q=2\\
g(q\xi)/g(2\xi)=\varphi_q(\xi)/\varphi_2(\xi)\leq q^H(1-\delta)^{-H}e^{-[q-3+\delta]\xi}, & \text{for }q=3,4. \tag*{(51)}
\end{cases} { g ( q ξ ) / g ( 2 ξ ) = 1 , g ( q ξ ) / g ( 2 ξ ) = φ q ( ξ ) / φ 2 ( ξ ) ≤ q H ( 1 − δ ) − H e − [ q − 3 + δ ] ξ , for q = 2 for q = 3 , 4. (51)
Keep in mind that the inequalities (49)–(51) will be systematically used. Straightforward computations with the use of Lemma 3 gives
∣ C 3 g ( 2 ξ ) ∣ = 1 ∣ A ∗ ∣ ∣ [ R ( q ξ ) − R ( ξ ) ] ∣ ⋅ g ( q ξ ) g ( 2 ξ ) ≤ 1 ∣ A ∗ ∣ [ log ( 4 ) + 3 M ( 2 H + 1 ) ] 4 H ( 1 − δ ) − H ≤ M 0 ∣ A ∗ ∣ , for all ξ > 1 , and q = 2 , 3 , 4. (52) \left|\frac{C_{3}}{g(2\xi)}\right|=\frac{1}{|A^{*}|}\left|[R(q\xi)-R(\xi)]\right|\cdot\frac{g(q\xi)}{g(2\xi)}
\leq\frac{1}{|A^{*}|}\left[\log(4)+3M(2H+1)\right]4^{H}(1-\delta)^{-H}
\leq\frac{M_{0}}{|A^{*}|},\quad\text{for all }\xi>1,\text{ and }q=2,3,4. \tag*{(52)} g ( 2 ξ ) C 3 = ∣ A ∗ ∣ 1 ∣ [ R ( q ξ ) − R ( ξ )] ∣ ⋅ g ( 2 ξ ) g ( q ξ ) ≤ ∣ A ∗ ∣ 1 [ log ( 4 ) + 3 M ( 2 H + 1 ) ] 4 H ( 1 − δ ) − H ≤ ∣ A ∗ ∣ M 0 , for all ξ > 1 , and q = 2 , 3 , 4. (52)
where M 0 = sup H ∈ K { 4 H ( 1 − δ ) − H [ log ( 4 ) + 3 M ( 2 H + 1 ) ] } M_{0}=\sup_{H\in\mathcal{K}}\left\{4^{H}(1-\delta)^{-H}[\log(4)+3M(2H+1)]\right\} M 0 = sup H ∈ K { 4 H ( 1 − δ ) − H [ log ( 4 ) + 3 M ( 2 H + 1 )] } and M M M being defined by (36).
But we know that
det { C 1 ; C 3 ; C 5 } = det { g ( ξ ) [ φ q ( ξ ) − q 2 ] ; C 3 ; [ q 2 φ q ′ ( ξ ) − q 2 ] / g ( 2 ξ ) } = q ( ξ ) [ det { φ q ( ξ ) ; C 3 g ( 2 ξ ) ; q 2 φ q ′ ( ξ ) } + det { φ q ( ξ ) ; C 3 g ( 2 ξ ) ; − q 2 } + det { − q 2 ; C 3 g ( 2 ξ ) ; q 2 φ q ′ ( ξ ) } ] . \begin{aligned}
\det\{C_{1};C_{3};C_{5}\}=\det\left\{g(\xi)[\varphi_{q}(\xi)-q^{2}];C_{3};[q^{2}\varphi_{q}'(\xi)-q^{2}]/g(2\xi)\right\} \\
&=q(\xi)\left[\det\left\{\varphi_{q}(\xi);\frac{C_{3}}{g(2\xi)};q^{2}\varphi_{q}'(\xi)\right\}+\det\left\{\varphi_{q}(\xi);\frac{C_{3}}{g(2\xi)};-q^{2}\right\}+\det\left\{-q^{2};\frac{C_{3}}{g(2\xi)};q^{2}\varphi_{q}'(\xi)\right\}\right].
\end{aligned} det { C 1 ; C 3 ; C 5 } = det { g ( ξ ) [ φ q ( ξ ) − q 2 ] ; C 3 ; [ q 2 φ q ′ ( ξ ) − q 2 ] / g ( 2 ξ ) } = q ( ξ ) [ det { φ q ( ξ ) ; g ( 2 ξ ) C 3 ; q 2 φ q ′ ( ξ ) } + det { φ q ( ξ ) ; g ( 2 ξ ) C 3 ; − q 2 } + det { − q 2 ; g ( 2 ξ ) C 3 ; q 2 φ q ′ ( ξ ) } ] .
Throughout the evaluation of this determinant at least one of the terms: φ q ( ξ ) \varphi_{q}(\xi) φ q ( ξ ) and φ q ′ ( ξ ) \varphi_{q}'(\xi) φ q ′ ( ξ ) arises with C 3 / g ( 2 ξ ) C_{3}/g(2\xi) C 3 / g ( 2 ξ ) . So, we use (49) and (52) to arrive at
∣ det { C 1 ; C 3 ; C 5 } ∣ ≤ 16 ⋅ 3 ! M 0 ⋅ g ( ξ ) ∣ A ∗ ∣ ( 4 2 H − 1 e − 2 ξ + 4 H e − ξ + 16 ⋅ 4 H − 1 e − ξ ) \left|\det\{C_{1};C_{3};C_{5}\}\right|\leq16\cdot3!M_{0}\cdot\frac{g(\xi)}{|A^{*}|}\left(4^{2H-1}e^{-2\xi}+4^{H}e^{-\xi}+16\cdot4^{H-1}e^{-\xi}\right) ∣ det { C 1 ; C 3 ; C 5 } ∣ ≤ 16 ⋅ 3 ! M 0 ⋅ ∣ A ∗ ∣ g ( ξ ) ( 4 2 H − 1 e − 2 ξ + 4 H e − ξ + 16 ⋅ 4 H − 1 e − ξ )
≤ 96 u 0 ( 1 − δ ) − H M 0 ⋅ e − ( 1 + δ ) ξ ∣ A ∗ ∣ ( 4 2 H − 1 e − ξ + 5 ⋅ 4 H ) ≤ M 1 ∣ A ∗ ∣ e − ( 1 + δ ) ξ . (53) \leq96u_{0}(1-\delta)^{-H}M_{0}\cdot\frac{e^{-(1+\delta)\xi}}{|A^{*}|}\left(4^{2H-1}e^{-\xi}+5\cdot4^{H}\right)
\leq\frac{M_{1}}{|A^{*}|}e^{-(1+\delta)\xi}. \tag*{(53)} ≤ 96 u 0 ( 1 − δ ) − H M 0 ⋅ ∣ A ∗ ∣ e − ( 1 + δ ) ξ ( 4 2 H − 1 e − ξ + 5 ⋅ 4 H ) ≤ ∣ A ∗ ∣ M 1 e − ( 1 + δ ) ξ . (53)
where M 1 : = 96 M 0 sup H ∈ K { u 0 ( 1 − δ ) − H ( 4 2 H − 1 + 5 ⋅ 4 H ) } M_{1}:=96M_{0}\sup_{H\in\mathcal{K}}\left\{u_{0}(1-\delta)^{-H}(4^{2H-1}+5\cdot4^{H})\right\} M 1 := 96 M 0 sup H ∈ K { u 0 ( 1 − δ ) − H ( 4 2 H − 1 + 5 ⋅ 4 H ) } .
ii-Evaluating the determinant det { C 2 ; C 3 ; C 5 } \det\{C_{2};C_{3};C_{5}\} det { C 2 ; C 3 ; C 5 } . In similar fashion we obtain
det { C 2 ; C 3 ; C 5 } = det { ( q 2 − 1 ) u 0 ; C 3 ; [ q 2 φ q ′ ( ξ ) − q 2 ] / g ( 2 ξ ) } = u 0 ( det { ( q 2 − 1 ) ; C 3 g ( 2 ξ ) ; q 2 φ q ′ ( ξ ) } + det { ( q 2 − 1 ) ; C 3 g ( 2 ξ ) ; − q 2 } ) . \begin{aligned}
\det\{C_{2};C_{3};C_{5}\}=\det\left\{(q^{2}-1)u_{0};C_{3};[q^{2}\varphi_{q}'(\xi)-q^{2}]/g(2\xi)\right\} \\
&=u_{0}\left(\det\left\{(q^{2}-1);\frac{C_{3}}{g(2\xi)};q^{2}\varphi_{q}'(\xi)\right\}+\det\left\{(q^{2}-1);\frac{C_{3}}{g(2\xi)};-q^{2}\right\}\right).
\end{aligned} det { C 2 ; C 3 ; C 5 } = det { ( q 2 − 1 ) u 0 ; C 3 ; [ q 2 φ q ′ ( ξ ) − q 2 ] / g ( 2 ξ ) } = u 0 ( det { ( q 2 − 1 ) ; g ( 2 ξ ) C 3 ; q 2 φ q ′ ( ξ ) } + det { ( q 2 − 1 ) ; g ( 2 ξ ) C 3 ; − q 2 } ) .
Thus,
∣ det { C 2 ; C 3 ; C 5 } ∣ ≤ 16 × 15 ⋅ 3 ! u 0 ⋅ M 0 ∣ A ∗ ∣ ( 4 H − 1 e − ξ + 1 ) ≤ M 2 u 0 ∣ A ∗ ∣ . (54) \left|\det\{C_{2};C_{3};C_{5}\}\right|\leq16\times15\cdot3!u_{0}\cdot\frac{M_{0}}{|A^{*}|}\left(4^{H-1}e^{-\xi}+1\right)
\leq M_{2}\frac{u_{0}}{|A^{*}|}. \tag*{(54)} ∣ det { C 2 ; C 3 ; C 5 } ∣ ≤ 16 × 15 ⋅ 3 ! u 0 ⋅ ∣ A ∗ ∣ M 0 ( 4 H − 1 e − ξ + 1 ) ≤ M 2 ∣ A ∗ ∣ u 0 . (54)
where M 2 : = 1440 M 0 sup H ∈ K { ( 4 H − 1 + 1 ) } M_{2}:=1440M_{0}\sup_{H\in\mathcal{K}}\{(4^{H-1}+1)\} M 2 := 1440 M 0 sup H ∈ K {( 4 H − 1 + 1 )} . Now, we recall that u ∗ = u 0 ( log ( 2 ) + ψ ( H ) ) u^{*}=u_{0}(\log(2)+\psi(H)) u ∗ = u 0 ( log ( 2 ) + ψ ( H )) , u 0 = 2 H − 1 Γ ( H ) u_{0}=2^{H-1}\Gamma(H) u 0 = 2 H − 1 Γ ( H ) and A = R ( ξ ) A=R(\xi) A = R ( ξ ) . From (v)-Lemma 3 we deduce that
A ∗ = u 0 R ( ξ ) − u ∗ ≥ u 0 log ( ξ ) − u 0 [ M + log ( 2 ) + ψ ( H ) ] , (see (36)) ≥ u 0 [ log ( ξ ) − M ∗ ] , for all ξ > e M ∗ , \begin{aligned}
A^{*}=u_{0}R(\xi)-u^{*}\geq u_{0}\log(\xi)-u_{0}[M+\log(2)+\psi(H)],\quad\text{(see (36))} \\
&\geq u_{0}[\log(\xi)-M^{*}],\quad\text{for all }\xi>e^{M^{*}},
\end{aligned} A ∗ = u 0 R ( ξ ) − u ∗ ≥ u 0 log ( ξ ) − u 0 [ M + log ( 2 ) + ψ ( H )] , (see (36)) ≥ u 0 [ log ( ξ ) − M ∗ ] , for all ξ > e M ∗ ,
with M ∗ = sup H ∈ K { M + log ( 2 ) + ψ ( H ) + 1 / ω 0 } M^{*}=\sup_{H\in\mathcal{K}}\{M+\log(2)+\psi(H)+1/\omega_{0}\} M ∗ = sup H ∈ K { M + log ( 2 ) + ψ ( H ) + 1/ ω 0 } . Consequently, from (54) we derive
∣ det { C 2 ; C 3 ; C 5 } ∣ ≤ M 2 log ( ξ ) − M ∗ , for all ξ > 1 ∨ e M ∗ . (55) \lvert\det\{\mathcal{C}_{2};\mathcal{C}_{3};\mathcal{C}_{5}\}\rvert\le\frac{M_{2}}{\log(\xi)-M^{*}},\qquad\text{for all }\xi>1\vee e^{M^{*}}. \tag*{(55)} ∣ det { C 2 ; C 3 ; C 5 }∣ ≤ log ( ξ ) − M ∗ M 2 , for all ξ > 1 ∨ e M ∗ . (55)
Note also that by definition of the constant M ∗ M^{*} M ∗ , we have
A ∗ ≥ 1 , for all ξ > 1 ∨ e M ∗ . (56) A^{*}\ge1,\qquad\text{for all }\xi>1\vee e^{M^{*}}. \tag*{(56)} A ∗ ≥ 1 , for all ξ > 1 ∨ e M ∗ . (56)
iii-Evaluating the determinant det { C 1 ; C 4 ; C 5 } \det\{\mathcal{C}_{1};\mathcal{C}_{4};\mathcal{C}_{5}\} det { C 1 ; C 4 ; C 5 } . Once again, we proceed in similar fashion and write
det { C 1 ; C 4 ; C 5 } = det { g ( ξ ) [ φ q ( ξ ) − q 2 ] ; ( 1 − q 2 ) ; [ q 2 φ q ′ ( ξ ) − q 2 ] / g ( 2 ξ ) } = det { [ φ q ( ξ ) − q 2 ] ; ( 1 − q 2 ) ; [ q 2 φ q ′ ( ξ ) − q 2 ] } ⋅ [ g ( ξ ) / g ( 2 ξ ) ] = [ det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ ) } + det { φ q ( ξ ) ; ( 1 − q 2 ) ; − q 2 } + det { − q 2 ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ ) } ] / φ 2 ( ξ ) = [ det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ ) } + det { φ q ( ξ ) ; 1 ; − q 2 } + det { − q 2 ; 1 ; q 2 φ q ′ ( ξ ) } ] / φ 2 ( ξ ) = [ det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ ) } − det { φ q ( ξ ) ; 1 ; q 2 } + det { q 2 φ q ′ ( ξ ) ; 1 ; q 2 } ] / φ 2 ( ξ ) = [ det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ ) } + det { Φ q ; 1 ; q 2 } ] / φ 2 ( ξ ) , \begin{aligned}
\det\{\mathcal{C}_{1};\mathcal{C}_{4};\mathcal{C}_{5}\}
&=\det\left\{g(\xi)[\varphi_{q}(\xi)-q^{2}];(1-q^{2});[q^{2}\varphi_{q}'(\xi)-q^{2}]/g(2\xi)\right\}\\
&=\det\left\{[\varphi_{q}(\xi)-q^{2}];(1-q^{2});[q^{2}\varphi_{q}'(\xi)-q^{2}]\right\}\cdot[g(\xi)/g(2\xi)]\\
&=\left[\det\{\varphi_{q}(\xi);(1-q^{2});q^{2}\varphi_{q}'(\xi)\}+\det\{\varphi_{q}(\xi);(1-q^{2});-q^{2}\}+\det\{-q^{2};(1-q^{2});q^{2}\varphi_{q}'(\xi)\}\right]/\varphi_{2}(\xi)\\
&=\left[\det\{\varphi_{q}(\xi);(1-q^{2});q^{2}\varphi_{q}'(\xi)\}+\det\{\varphi_{q}(\xi);1;-q^{2}\}+\det\{-q^{2};1;q^{2}\varphi_{q}'(\xi)\}\right]/\varphi_{2}(\xi)\\
&=\left[\det\{\varphi_{q}(\xi);(1-q^{2});q^{2}\varphi_{q}'(\xi)\}-\det\{\varphi_{q}(\xi);1;q^{2}\}+\det\{q^{2}\varphi_{q}'(\xi);1;q^{2}\}\right]/\varphi_{2}(\xi)\\
&=\left[\det\{\varphi_{q}(\xi);(1-q^{2});q^{2}\varphi_{q}'(\xi)\}+\det\{\Phi_{q};1;q^{2}\}\right]/\varphi_{2}(\xi),
\end{aligned} det { C 1 ; C 4 ; C 5 } = det { g ( ξ ) [ φ q ( ξ ) − q 2 ] ; ( 1 − q 2 ) ; [ q 2 φ q ′ ( ξ ) − q 2 ] / g ( 2 ξ ) } = det { [ φ q ( ξ ) − q 2 ] ; ( 1 − q 2 ) ; [ q 2 φ q ′ ( ξ ) − q 2 ] } ⋅ [ g ( ξ ) / g ( 2 ξ )] = [ det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ )} + det { φ q ( ξ ) ; ( 1 − q 2 ) ; − q 2 } + det { − q 2 ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ )} ] / φ 2 ( ξ ) = [ det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ )} + det { φ q ( ξ ) ; 1 ; − q 2 } + det { − q 2 ; 1 ; q 2 φ q ′ ( ξ )} ] / φ 2 ( ξ ) = [ det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ )} − det { φ q ( ξ ) ; 1 ; q 2 } + det { q 2 φ q ′ ( ξ ) ; 1 ; q 2 } ] / φ 2 ( ξ ) = [ det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ )} + det { Φ q ; 1 ; q 2 } ] / φ 2 ( ξ ) ,
where Φ q = q 2 φ q ′ ( ξ ) − φ q ( ξ ) \Phi_{q}=q^{2}\varphi_{q}'(\xi)-\varphi_{q}(\xi) Φ q = q 2 φ q ′ ( ξ ) − φ q ( ξ ) . The last determinant simplifies to
det { Φ q ; 1 ; q 2 } = ∣ Φ 2 1 4 Φ 3 1 9 Φ 4 1 16 ∣ = Φ 2 ∣ 1 9 1 16 ∣ − Φ 3 ∣ 1 4 1 16 ∣ + Φ 4 ∣ 1 4 1 9 ∣ = 7 Φ 2 − 12 Φ 3 + 5 Φ 4 . \begin{aligned}
\det\{\Phi_{q};1;q^{2}\}
&=\begin{vmatrix}\Phi_{2}&1&4\\\Phi_{3}&1&9\\\Phi_{4}&1&16\end{vmatrix}
=\Phi_{2}\begin{vmatrix}1&9\\1&16\end{vmatrix}-\Phi_{3}\begin{vmatrix}1&4\\1&16\end{vmatrix}+\Phi_{4}\begin{vmatrix}1&4\\1&9\end{vmatrix}\\
&=7\Phi_{2}-12\Phi_{3}+5\Phi_{4}.
\end{aligned} det { Φ q ; 1 ; q 2 } = Φ 2 Φ 3 Φ 4 1 1 1 4 9 16 = Φ 2 1 1 9 16 − Φ 3 1 1 4 16 + Φ 4 1 1 4 9 = 7 Φ 2 − 12 Φ 3 + 5 Φ 4 .
Thereby,
det { C 1 ; C 4 ; C 5 } = 7 Φ 2 / φ 2 ( ξ ) + det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ ) } / φ 2 ( ξ ) − ( 12 Φ 3 − 5 Φ 4 ) / φ 2 ( ξ ) = : 7 Φ 2 / φ 2 ( ξ ) + R 1 ′ − R 2 ′ . (57) \det\{\mathcal{C}_{1};\mathcal{C}_{4};\mathcal{C}_{5}\}=7\Phi_{2}/\varphi_{2}(\xi)+\det\{\varphi_{q}(\xi);(1-q^{2});q^{2}\varphi_{q}'(\xi)\}/\varphi_{2}(\xi)-(12\Phi_{3}-5\Phi_{4})/\varphi_{2}(\xi)
=:\,7\Phi_{2}/\varphi_{2}(\xi)+R_{1}'-R_{2}'. \tag*{(57)} det { C 1 ; C 4 ; C 5 } = 7 Φ 2 / φ 2 ( ξ ) + det { φ q ( ξ ) ; ( 1 − q 2 ) ; q 2 φ q ′ ( ξ )} / φ 2 ( ξ ) − ( 12 Φ 3 − 5 Φ 4 ) / φ 2 ( ξ ) =: 7 Φ 2 / φ 2 ( ξ ) + R 1 ′ − R 2 ′ . (57)
According to (49) and (50) one has
∣ R 1 ′ ∣ ≤ 16 ⋅ 15 ⋅ 3 ! ⋅ 4 2 H ‾ − 1 e − 2 ξ ⋅ ( 1 − δ ) − H ‾ e ( 2 − δ ) ξ = : M 3 e − δ ξ , (58) \lvert R_{1}'\rvert\le16\cdot15\cdot3!\cdot4^{2\overline{H}-1}e^{-2\xi}\cdot(1-\delta)^{-\overline{H}}e^{(2-\delta)\xi}=:M_{3}e^{-\delta\xi}, \tag*{(58)} ∣ R 1 ′ ∣ ≤ 16 ⋅ 15 ⋅ 3 ! ⋅ 4 2 H − 1 e − 2 ξ ⋅ ( 1 − δ ) − H e ( 2 − δ ) ξ =: M 3 e − δ ξ , (58)
∣ R 2 ′ ∣ ≤ [ 12 ∣ φ 3 ′ ( ξ ) − φ 3 ( ξ ) ∣ + 5 ∣ 16 φ 4 ′ ( ξ ) − φ 4 ( ξ ) ∣ ] / φ 2 ( ξ ) ≤ ( 108 ⋅ 3 H ‾ − 1 + 12 ⋅ 3 H ‾ + 80 ⋅ 4 H ‾ − 1 + 5 ⋅ 4 H ‾ ) e − 2 ξ ⋅ ( 1 − δ ) − H ‾ e ( 2 − δ ) ξ = : M 4 e − δ ξ , (59) \begin{aligned}
\lvert R_{2}'\rvert
&\le[12\lvert\varphi_{3}'(\xi)-\varphi_{3}(\xi)\rvert+5\lvert16\varphi_{4}'(\xi)-\varphi_{4}(\xi)\rvert]/\varphi_{2}(\xi)\\
&\le\left(108\cdot3^{\overline{H}-1}+12\cdot3^{\overline{H}}+80\cdot4^{\overline{H}-1}+5\cdot4^{\overline{H}}\right)e^{-2\xi}\cdot(1-\delta)^{-\overline{H}}e^{(2-\delta)\xi}=:M_{4}e^{-\delta\xi}, \tag*{(59)}
\end{aligned} ∣ R 2 ′ ∣ ≤ [ 12 ∣ φ 3 ′ ( ξ ) − φ 3 ( ξ )∣ + 5 ∣ 16 φ 4 ′ ( ξ ) − φ 4 ( ξ )∣] / φ 2 ( ξ ) ≤ ( 108 ⋅ 3 H − 1 + 12 ⋅ 3 H + 80 ⋅ 4 H − 1 + 5 ⋅ 4 H ) e − 2 ξ ⋅ ( 1 − δ ) − H e ( 2 − δ ) ξ =: M 4 e − δ ξ , (59)
while
7 Φ 2 / φ 2 ( ξ ) = 7 ( 4 φ 2 ′ ( ξ ) φ 2 ( ξ ) − 1 ) = 7 ( 4 g ′ ( 2 ξ ) g ′ ( ξ ) ⋅ g ( ξ ) g ( 2 ξ ) − 1 ) = 7 ( 4 ( 2 ξ ) H − 1 K H − 1 ( 2 ξ ) ξ H − 1 K H − 1 ( ξ ) ⋅ ξ H K H ( ξ ) ( 2 ξ ) H K H ( 2 ξ ) − 1 ) = 7 ( 2 K H − 1 ( 2 ξ ) K H ( 2 ξ ) ⋅ K H ( ξ ) K H − 1 ( ξ ) − 1 ) = 7 ( 2 r H ( 2 ξ ) r H ( ξ ) − 1 ) > 7. (60) \begin{aligned}
7\Phi_{2}/\varphi_{2}(\xi)=7\left(4\frac{\varphi_{2}'(\xi)}{\varphi_{2}(\xi)}-1\right) \\
&=7\left(4\frac{g'(2\xi)}{g'(\xi)}\cdot\frac{g(\xi)}{g(2\xi)}-1\right)
=7\left(4\frac{(2\xi)^{H-1}K_{H-1}(2\xi)}{\xi^{H-1}K_{H-1}(\xi)}\cdot\frac{\xi^{H}K_{H}(\xi)}{(2\xi)^{H}K_{H}(2\xi)}-1\right) \\
&=7\left(2\frac{K_{H-1}(2\xi)}{K_{H}(2\xi)}\cdot\frac{K_{H}(\xi)}{K_{H-1}(\xi)}-1\right) \\
&=7\left(2\frac{r^{H}(2\xi)}{r^{H}(\xi)}-1\right)>7. \tag*{(60)}
\end{aligned} 7 Φ 2 / φ 2 ( ξ ) = 7 ( 4 φ 2 ( ξ ) φ 2 ′ ( ξ ) − 1 ) = 7 ( 4 g ′ ( ξ ) g ′ ( 2 ξ ) ⋅ g ( 2 ξ ) g ( ξ ) − 1 ) = 7 ( 4 ξ H − 1 K H − 1 ( ξ ) ( 2 ξ ) H − 1 K H − 1 ( 2 ξ ) ⋅ ( 2 ξ ) H K H ( 2 ξ ) ξ H K H ( ξ ) − 1 ) = 7 ( 2 K H ( 2 ξ ) K H − 1 ( 2 ξ ) ⋅ K H − 1 ( ξ ) K H ( ξ ) − 1 ) = 7 ( 2 r H ( ξ ) r H ( 2 ξ ) − 1 ) > 7. (60)
where r H ( ξ ) = K H − 1 ( ξ ) / K H ( ξ ) r_{H}(\xi)=K_{H-1}(\xi)/K_{H}(\xi) r H ( ξ ) = K H − 1 ( ξ ) / K H ( ξ ) . In the last inequality above we used the fact that r H ( ⋅ ) r_{H}(\cdot) r H ( ⋅ ) is strictly increasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) for H > 1 / 2 H>1/2 H > 1/2 (see [10 Lemma 2.2] ).
Conclusions
Putting the previous results (53), (55), (56), (57), (58), (59) and (60) together with (46), we obtain
{ ∣ ∇ θ , 3 ( α ) ∣ = − 2 α κ 3 σ 4 ξ g ( 2 ξ ) g ′ ( ξ ) A ∗ [ det { C 1 ; C 3 ; C 5 } + det { C 2 ; C 3 ; C 5 } + det { C 1 ; C 4 ; C 5 } ] ∣ det { C 1 ; C 3 ; C 5 } ∣ ≤ M 1 A ∗ e − ( 1 + δ ) ξ and A ∗ ≥ 1 , ∣ det { C 2 ; C 3 ; C 5 } ∣ ≤ M 2 / ( log ( ξ ) − M ∗ ) , det { C 1 ; C 4 ; C 5 } = 7 Φ 2 / φ 2 ( ξ ) + R 1 ′ − R 2 ′ ∣ R 1 ′ ∣ ≤ M 3 e − δ ξ , ∣ R 2 ′ ∣ ≤ M 4 e − δ ξ and 7 Φ 2 / φ 2 ( ξ ) > 7 , \left\{
\begin{aligned}
|\nabla_{\theta,3}(\alpha)| &= -2\alpha\kappa^{3}\sigma^{4}\xi g(2\xi)g'(\xi)A^{*}\left[\det\{\mathcal{C}_{1};\mathcal{C}_{3};\mathcal{C}_{5}\}+\det\{\mathcal{C}_{2};\mathcal{C}_{3};\mathcal{C}_{5}\}+\det\{\mathcal{C}_{1};\mathcal{C}_{4};\mathcal{C}_{5}\}\right] \\
|\det\{\mathcal{C}_{1};\mathcal{C}_{3};\mathcal{C}_{5}\}| &\le\frac{M_{1}}{A^{*}}e^{-(1+\delta)\xi}\quad\text{and}\quad A^{*}\ge1, \\
|\det\{\mathcal{C}_{2};\mathcal{C}_{3};\mathcal{C}_{5}\}| &\le M_{2}/(\log(\xi)-M^{*}), \\
\det\{\mathcal{C}_{1};\mathcal{C}_{4};\mathcal{C}_{5}\} &= 7\Phi_{2}/\varphi_{2}(\xi)+\mathcal{R}'_{1}-\mathcal{R}'_{2} \\
|\mathcal{R}'_{1}| &\le M_{3}e^{-\delta\xi},\ |\mathcal{R}'_{2}|\le M_{4}e^{-\delta\xi}\quad\text{and}\quad7\Phi_{2}/\varphi_{2}(\xi)>7,
\end{aligned}
\right. ⎩ ⎨ ⎧ ∣ ∇ θ , 3 ( α ) ∣ ∣ det { C 1 ; C 3 ; C 5 } ∣ ∣ det { C 2 ; C 3 ; C 5 } ∣ det { C 1 ; C 4 ; C 5 } ∣ R 1 ′ ∣ = − 2 α κ 3 σ 4 ξ g ( 2 ξ ) g ′ ( ξ ) A ∗ [ det { C 1 ; C 3 ; C 5 } + det { C 2 ; C 3 ; C 5 } + det { C 1 ; C 4 ; C 5 } ] ≤ A ∗ M 1 e − ( 1 + δ ) ξ and A ∗ ≥ 1 , ≤ M 2 / ( log ( ξ ) − M ∗ ) , = 7 Φ 2 / φ 2 ( ξ ) + R 1 ′ − R 2 ′ ≤ M 3 e − δ ξ , ∣ R 2 ′ ∣ ≤ M 4 e − δ ξ and 7 Φ 2 / φ 2 ( ξ ) > 7 ,
for all ξ > 1 ∨ e M ∗ \xi>1\vee e^{M^{*}} ξ > 1 ∨ e M ∗ . To conclude we choose ( δ , ξ ) ∈ ( 0 , 1 ) × ( 1 ∨ e M ∗ , ∞ ) (\delta,\xi)\in(0,1)\times(1\vee e^{M^{*}},\infty) ( δ , ξ ) ∈ ( 0 , 1 ) × ( 1 ∨ e M ∗ , ∞ ) so that
M 1 e − ( 1 + δ ) ξ + [ M 2 / ( log ( ξ ) − M ∗ ) ] + ( M 3 + M 4 ) e − δ ξ ≤ 6. M_{1}e^{-(1+\delta)\xi}+[M_{2}/(\log(\xi)-M^{*})]+(M_{3}+M_{4})e^{-\delta\xi}\le6. M 1 e − ( 1 + δ ) ξ + [ M 2 / ( log ( ξ ) − M ∗ )] + ( M 3 + M 4 ) e − δ ξ ≤ 6.
It suffices to choose
δ ∈ ( 0 , 1 ) and ξ = λ α > 1 ∨ e M ∗ + M 2 / 3 ∨ 1 δ log ( ∑ j ≠ 2 M j 3 ) . (61) \delta\in(0,1)\quad\text{and}\quad\xi=\lambda\alpha>1\vee e^{M^{*}+M_{2/3}}\vee\frac{1}{\delta}\log\left(\frac{\sum_{j\ne2}M_{j}}{3}\right). \tag*{(61)} δ ∈ ( 0 , 1 ) and ξ = λ α > 1 ∨ e M ∗ + M 2/3 ∨ δ 1 log ( 3 ∑ j = 2 M j ) . (61)
Under this condition we obtain ∣ ∇ θ , 3 ( α ) ∣ < − 2 α κ 3 σ 4 ξ g ( 2 ξ ) g ′ ( ξ ) A ∗ < 0 |\nabla_{\theta,3}(\alpha)|<-2\alpha\kappa^{3}\sigma^{4}\xi g(2\xi)g'(\xi)A^{*}<0 ∣ ∇ θ , 3 ( α ) ∣ < − 2 α κ 3 σ 4 ξ g ( 2 ξ ) g ′ ( ξ ) A ∗ < 0 . An equivalent condition for α \alpha α can be obtained. In fact, we choose α > ξ 0 / λ = : α 0 \alpha>\xi_{0}/\lambda=:\alpha_{0} α > ξ 0 / λ =: α 0 . This proves the first part in Lemma 6, while the second one follows by local univalent theorem (e.g., [23 p. 2] ). Therefore, the proof is complete.
Evaluating the principal minors of the Jacobian matrix ∇ θ , 3 ( α ) \nabla_{\theta,3}(\alpha) ∇ θ , 3 ( α )
In what follows ⟨ K ⟩ Σ \langle K\rangle_{\Sigma} ⟨ K ⟩ Σ will denote the principal minor of ∇ θ , 3 ( α ) \nabla_{\theta,3}(\alpha) ∇ θ , 3 ( α ) obtained by omitting rows and columns of indices in K ⊆ { 1 , 2 , 3 } K\subseteq\{1,2,3\} K ⊆ { 1 , 2 , 3 } . So, the principal minors of order 1 × 1 1\times1 1 × 1 are ⟨ 2 , 3 ⟩ Σ = Σ 1 , 1 \langle2,3\rangle_{\Sigma}=\Sigma_{1,1} ⟨ 2 , 3 ⟩ Σ = Σ 1 , 1 , ⟨ 1 , 3 ⟩ Σ = Σ 2 , 2 \langle1,3\rangle_{\Sigma}=\Sigma_{2,2} ⟨ 1 , 3 ⟩ Σ = Σ 2 , 2 and ⟨ 1 , 2 ⟩ Σ = Σ 3 , 3 \langle1,2\rangle_{\Sigma}=\Sigma_{3,3} ⟨ 1 , 2 ⟩ Σ = Σ 3 , 3 , while the principal minors of order 2 × 2 2\times2 2 × 2 and order 3 × 3 3\times3 3 × 3 are
⟨ 3 ⟩ Σ = ∣ Σ 1 , 1 Σ 1 , 2 Σ 2 , 1 Σ 2 , 2 ∣ ; ⟨ 2 ⟩ Σ = ∣ Σ 1 , 1 Σ 1 , 3 Σ 3 , 1 Σ 3 , 3 ∣ ; ⟨ 1 ⟩ Σ = ∣ Σ 2 , 2 Σ 2 , 3 Σ 3 , 2 Σ 3 , 3 ∣ . \langle3\rangle_{\Sigma}=\begin{vmatrix}\Sigma_{1,1}&\Sigma_{1,2}\\\Sigma_{2,1}&\Sigma_{2,2}\end{vmatrix};\quad
\langle2\rangle_{\Sigma}=\begin{vmatrix}\Sigma_{1,1}&\Sigma_{1,3}\\\Sigma_{3,1}&\Sigma_{3,3}\end{vmatrix};\quad
\langle1\rangle_{\Sigma}=\begin{vmatrix}\Sigma_{2,2}&\Sigma_{2,3}\\\Sigma_{3,2}&\Sigma_{3,3}\end{vmatrix}. ⟨ 3 ⟩ Σ = Σ 1 , 1 Σ 2 , 1 Σ 1 , 2 Σ 2 , 2 ; ⟨ 2 ⟩ Σ = Σ 1 , 1 Σ 3 , 1 Σ 1 , 3 Σ 3 , 3 ; ⟨ 1 ⟩ Σ = Σ 2 , 2 Σ 3 , 2 Σ 2 , 3 Σ 3 , 3 .
and
⟨ ∅ ⟩ Σ = ∣ Σ 1 , 1 Σ 1 , 2 Σ 1 , 3 Σ 2 , 1 Σ 2 , 2 Σ 2 , 3 Σ 3 , 1 Σ 3 , 2 Σ 3 , 3 ∣ . \langle\varnothing\rangle_{\Sigma}=
\begin{vmatrix}
\Sigma_{1,1}&\Sigma_{1,2}&\Sigma_{1,3}\\
\Sigma_{2,1}&\Sigma_{2,2}&\Sigma_{2,3}\\
\Sigma_{3,1}&\Sigma_{3,2}&\Sigma_{3,3}
\end{vmatrix}. ⟨ ∅ ⟩ Σ = Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 Σ 1 , 2 Σ 2 , 2 Σ 3 , 2 Σ 1 , 3 Σ 2 , 3 Σ 3 , 3 .
where Σ j , k \Sigma_{j,k} Σ j , k are the entries of ∇ θ , 3 ( α ) \nabla_{\theta,3}(\alpha) ∇ θ , 3 ( α ) defined by equations (14)-(18).
Lemma 7. There exists α ∗ > 0 \alpha^{*} > 0 α ∗ > 0 depending only on the compact parameter space
Θ ⊂ ( 0 , ∞ ) × ( 1 / 2 , ∞ ) × ( 0 , ∞ ) \Theta\subset(0,\infty) \times(1/2,\infty) \times(0,\infty) Θ ⊂ ( 0 , ∞ ) × ( 1/2 , ∞ ) × ( 0 , ∞ ) such that the principal minors of the Jacobian matrix ∇ Θ ϱ θ , 3 ( α ) \nabla_{\Theta}\varrho_{\theta,3}(\alpha) ∇ Θ ϱ θ , 3 ( α ) satisfy
⟨ 2 ⟩ Σ , ⟨ 3 ⟩ Σ , ⟨ 2 , 3 ⟩ Σ > 0 and ⟨ 1 , 2 ⟩ Σ , ⟨ ∅ ⟩ Σ < 0 , (62) \langle2\rangle_{\Sigma},\ \langle3\rangle_{\Sigma},\ \langle2,3\rangle_{\Sigma} > 0\quad\text{and}\quad\langle1,2\rangle_{\Sigma},\ \langle\emptyset\rangle_{\Sigma} < 0, \tag*{(62)} ⟨ 2 ⟩ Σ , ⟨ 3 ⟩ Σ , ⟨ 2 , 3 ⟩ Σ > 0 and ⟨ 1 , 2 ⟩ Σ , ⟨ ∅ ⟩ Σ < 0 , (62) { ⟨ 1 ⟩ Σ , ⟨ 1 , 3 ⟩ Σ < 0 , if λ < e ( ψ ( H + 1 / 2 ) + ψ ( H ) ) / 2 , ⟨ 1 ⟩ Σ , ⟨ 1 , 3 ⟩ Σ > 0 , if λ > e ( ψ ( H + 1 / 2 ) + ψ ( H ) ) / 2 , (63) \begin{cases}
\langle1\rangle_{\Sigma},\ \langle1,3\rangle_{\Sigma} < 0, & \text{if }\lambda< e^{(\psi(H+1/2)+\psi(H))/2},\\
\langle1\rangle_{\Sigma},\ \langle1,3\rangle_{\Sigma} > 0, & \text{if }\lambda> e^{(\psi(H+1/2)+\psi(H))/2}, \tag*{(63)}
\end{cases} { ⟨ 1 ⟩ Σ , ⟨ 1 , 3 ⟩ Σ < 0 , ⟨ 1 ⟩ Σ , ⟨ 1 , 3 ⟩ Σ > 0 , if λ < e ( ψ ( H + 1/2 ) + ψ ( H )) /2 , if λ > e ( ψ ( H + 1/2 ) + ψ ( H )) /2 , (63) for all α ∈ ( α ∗ , ∞ ) \alpha\in(\alpha^{*},\infty) α ∈ ( α ∗ , ∞ ) .
Proof. First, observe that the unique principal minor of order 3 × 3 3 \times3 3 × 3 is the determinant of the Jacobian matrix ∇ Θ ϱ θ , 3 ( α ) \nabla_{\Theta}\varrho_{\theta,3}(\alpha) ∇ Θ ϱ θ , 3 ( α ) . It is already investigated and shown to be negative when α > α 0 \alpha> \alpha_{0} α > α 0 (see Lemma 6). Throughout this proof δ ∈ ( 2 / 3 , 1 ) \delta\in(2/3,1) δ ∈ ( 2/3 , 1 ) and the supremum with respect to H H H and/or λ \lambda λ should be understood as being taken over their own spaces (say H ∈ [ H ‾ , H ‾ ] H \in[\underline{H},\overline{H}] H ∈ [ H , H ] , λ ∈ [ λ ‾ , λ ‾ ] \lambda\in[\underline{\lambda},\overline{\lambda}] λ ∈ [ λ , λ ] and ( H , λ ) ∈ [ H ‾ , H ‾ ] × [ λ ‾ , λ ‾ ] (H,\lambda) \in[\underline{H},\overline{H}] \times[\underline{\lambda},\overline{\lambda}] ( H , λ ) ∈ [ H , H ] × [ λ , λ ] ).
i-Evaluating the principal minors of order 1 × 1 1 \times1 1 × 1 . First observe that ⟨ 2 , 3 ⟩ Σ = Σ 1 , 1 = 2 κ ( g H ( 2 ξ ) − 4 g H ( ξ ) + 3 g H ( 0 ) ) \langle2,3\rangle_{\Sigma} = \Sigma_{1,1} = 2\kappa(g_{H}(2\xi) - 4g_{H}(\xi) + 3g_{H}(0)) ⟨ 2 , 3 ⟩ Σ = Σ 1 , 1 = 2 κ ( g H ( 2 ξ ) − 4 g H ( ξ ) + 3 g H ( 0 )) , where κ > 0 \kappa> 0 κ > 0 is independent of ξ = λ α \xi= \lambda\alpha ξ = λ α (see notations (15) and (18)). Letting α \alpha α to vary in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and consider the function ξ ↦ ⟨ 2 , 3 ⟩ Σ = : f 0 ( ξ ) \xi\mapsto\langle2,3\rangle_{\Sigma} =: f_{0}(\xi) ξ ↦ ⟨ 2 , 3 ⟩ Σ =: f 0 ( ξ ) . This function is continuously differentiable with derivative
∂ ξ f 0 ( ξ ) = 2 κ ( − 2 ( 2 ξ ) H K H − 1 ( 2 ξ ) + 4 ξ H K H − 1 ( ξ ) ) = 8 κ ξ ( g H − 1 ( ξ ) − g H − 1 ( 2 ξ ) ) > 0 \partial_{\xi}f_{0}(\xi) = 2\kappa\left(-2(2\xi)^{H}K_{H-1}(2\xi) + 4\xi^{H}K_{H-1}(\xi)\right)
= 8\kappa\xi\left(g_{H-1}(\xi) - g_{H-1}(2\xi)\right) > 0 ∂ ξ f 0 ( ξ ) = 2 κ ( − 2 ( 2 ξ ) H K H − 1 ( 2 ξ ) + 4 ξ H K H − 1 ( ξ ) ) = 8 κ ξ ( g H − 1 ( ξ ) − g H − 1 ( 2 ξ ) ) > 0
and lim ξ ↓ 0 f 0 ( ξ ) = 0 \lim_{\xi\downarrow0} f_{0}(\xi) = 0 lim ξ ↓ 0 f 0 ( ξ ) = 0 . In the above inequality we used the fact that ξ ↦ g H − 1 ( ξ ) \xi\mapsto g_{H-1}(\xi) ξ ↦ g H − 1 ( ξ ) is strictly decreasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Thereby, f 0 f_{0} f 0 is strictly increasing and ⟨ 2 , 3 ⟩ Σ = f 0 ( ξ ) > 0 \langle2,3\rangle_{\Sigma} = f_{0}(\xi) > 0 ⟨ 2 , 3 ⟩ Σ = f 0 ( ξ ) > 0 , ∀ ξ > 0 \forall\xi> 0 ∀ ξ > 0 .
The principal minor ⟨ 1 , 3 ⟩ Σ \langle1,3\rangle_{\Sigma} ⟨ 1 , 3 ⟩ Σ can be rewritten as (see notations (16) and (18))
⟨ 1 , 3 ⟩ Σ = Σ 2 , 2 = κ σ 2 { − 4 ( u 0 ξ + u ∗ ) + ξ ( z − 4 y + 7 x ) + ( C z − 4 B y + 7 A x ) } = κ σ 2 u 0 { − 4 [ − log ( λ 2 ) + ψ ( H + 1 / 2 ) + ψ ( H ) ] + R 1 ( ξ ) u 0 } . \begin{aligned}
\langle1,3\rangle_{\Sigma} = \Sigma_{2,2} = \kappa\sigma^{2}\left\{-4(u_{0}\xi+ u^{*}) + \xi(z - 4y + 7x) + (Cz - 4By + 7Ax)\right\} \\
&= \kappa\sigma^{2}u_{0}\left\{-4\left[-\log(\lambda^{2}) + \psi(H + 1/2) + \psi(H)\right] + \frac{R_{1}(\xi)}{u_{0}}\right\}.
\end{aligned} ⟨ 1 , 3 ⟩ Σ = Σ 2 , 2 = κ σ 2 { − 4 ( u 0 ξ + u ∗ ) + ξ ( z − 4 y + 7 x ) + ( C z − 4 B y + 7 A x ) } = κ σ 2 u 0 { − 4 [ − log ( λ 2 ) + ψ ( H + 1/2 ) + ψ ( H ) ] + u 0 R 1 ( ξ ) } .
It is not hard to see that
∣ R 1 ( ξ ) u 0 ∣ ≤ 1 u 0 [ ∣ ξ ∣ ( z + 4 y + 7 x ) + ( C z + 4 B y + 7 A x ) ] ≤ x u 0 ( 12 ∣ ξ ∣ + C + 4 B + 7 A ) ≤ 12 g H ( ξ ) u 0 [ ∣ − log ( 2 λ 2 ) + ψ ( H + 1 / 2 ) ∣ + log ( 3 ξ ) + M ( 6 H + 1 ) ] \begin{aligned}
\left|\frac{R_{1}(\xi)}{u_{0}}\right| \le\frac{1}{u_{0}}\left[|\xi|(z + 4y + 7x) + (Cz + 4By + 7Ax)\right] \\
&\le\frac{x}{u_{0}}(12|\xi| + C + 4B + 7A) \\
&\le\frac{12g_{H}(\xi)}{u_{0}}\left[|-\log(2\lambda^{2}) + \psi(H + 1/2)| + \log(3\xi) + M(6H + 1)\right]
\end{aligned} u 0 R 1 ( ξ ) ≤ u 0 1 [ ∣ ξ ∣ ( z + 4 y + 7 x ) + ( C z + 4 B y + 7 A x ) ] ≤ u 0 x ( 12∣ ξ ∣ + C + 4 B + 7 A ) ≤ u 0 12 g H ( ξ ) [ ∣ − log ( 2 λ 2 ) + ψ ( H + 1/2 ) ∣ + log ( 3 ξ ) + M ( 6 H + 1 ) ]
≤ 12 ( 1 − δ ) − H e − δ ξ [ ∣ − log ( 2 λ 2 ) + ψ ( H + 1 / 2 ) ∣ + log ( 3 ξ ) + M ( 6 H + 1 ) ] ≤ e − δ ξ ( L 0 + L 1 log ( ξ ) ) , ξ = λ α , \begin{aligned}
\le12(1-\delta)^{-H}e^{-\delta\xi}\left[|-\log(2\lambda^{2}) + \psi(H + 1/2)| + \log(3\xi) + M(6H + 1)\right] \\
&\le e^{-\delta\xi}(L_{0} + L_{1}\log(\xi)),\qquad\xi= \lambda\alpha,
\end{aligned} ≤ 12 ( 1 − δ ) − H e − δ ξ [ ∣ − log ( 2 λ 2 ) + ψ ( H + 1/2 ) ∣ + log ( 3 ξ ) + M ( 6 H + 1 ) ] ≤ e − δ ξ ( L 0 + L 1 log ( ξ )) , ξ = λ α ,
for all ξ > e M \xi> e^{M} ξ > e M , where M M M is the numerical constant given in (37) and used in (v),(vi)-Lemma 3, L 1 = 12 ( 1 − δ ) − H L_1 = 12(1-\delta)^{-H} L 1 = 12 ( 1 − δ ) − H and
L 0 = sup H , λ { 12 ( 1 − δ ) − H [ ∣ − log ( 2 λ 2 ) + ψ ( H + 1 / 2 ) ∣ + log ( 3 ) + M ( 6 H + 1 ) ] } . L_0 = \sup_{H,\lambda} \left\{12(1-\delta)^{-H}\left[\left|-\log(2\lambda^2)+\psi(H+1/2)\right|+\log(3)+M(6H+1)\right]\right\}. L 0 = H , λ sup { 12 ( 1 − δ ) − H [ − log ( 2 λ 2 ) + ψ ( H + 1/2 ) + log ( 3 ) + M ( 6 H + 1 ) ] } .
In the first four inequalities we used the following facts: g H ( q ξ ) , R H ( q ξ ) > 0 g_H(q^\xi), R_H(q^\xi)>0 g H ( q ξ ) , R H ( q ξ ) > 0 , for all ξ > 1 ∨ e M \xi>1\vee e^M ξ > 1 ∨ e M and q = 1 , 2 , 3 q=1,2,3 q = 1 , 2 , 3 , the function ξ ↦ g H ( q ξ ) \xi\mapsto g_H(q^\xi) ξ ↦ g H ( q ξ ) is decreasing, (vi)-Lemma 3 and the first part in (49) , respectively. We can find some ξ 1 \xi_1 ξ 1 (depending only on the parameter space) so that
{ ⟨ 1 , 3 ⟩ Σ < 0 , if λ < e ( ψ ( H + 1 / 2 ) + ψ ( H ) ) / 2 , ⟨ 1 , 3 ⟩ Σ > 0 , if λ > e ( ψ ( H + 1 / 2 ) + ψ ( H ) ) / 2 , for all ξ > ξ 1 . \begin{cases}
\langle1,3\rangle_\Sigma< 0, & \text{if }\lambda<e^{(\psi(H+1/2)+\psi(H))/2},\\
\langle1,3\rangle_\Sigma> 0, & \text{if }\lambda>e^{(\psi(H+1/2)+\psi(H))/2},
\end{cases}
\qquad\text{for all }\xi>\xi_1. { ⟨ 1 , 3 ⟩ Σ < 0 , ⟨ 1 , 3 ⟩ Σ > 0 , if λ < e ( ψ ( H + 1/2 ) + ψ ( H )) /2 , if λ > e ( ψ ( H + 1/2 ) + ψ ( H )) /2 , for all ξ > ξ 1 .
Formally, ξ 1 \xi_1 ξ 1 can be defined as
ξ 1 : = inf H , λ { ξ > e M ∨ e 1 / δ : e − δ ξ ( L 0 + L 1 log ( ξ ) ) < 4 ∣ ψ ( H ∗ + 1 / 2 ) + ψ ( H ∗ ) − log ( λ ∗ 2 ) ∣ } , \xi_1:=\inf_{H,\lambda}\left\{\xi>e^M\vee e^{1/\delta}:e^{-\delta\xi}(L_0+L_1\log(\xi))<4\left|\psi(H_*+1/2)+\psi(H_*)-\log(\lambda_*^2)\right|\right\}, ξ 1 := H , λ inf { ξ > e M ∨ e 1/ δ : e − δ ξ ( L 0 + L 1 log ( ξ )) < 4 ψ ( H ∗ + 1/2 ) + ψ ( H ∗ ) − log ( λ ∗ 2 ) } ,
with ( H ∗ , λ ∗ ) : = arg min H , λ ∣ ψ ( H + 1 / 2 ) + ψ ( H ) − log ( λ 2 ) ∣ (H_*,\lambda_*):=\underset{H,\lambda}{\arg\min}\left|\psi(H+1/2)+\psi(H)-\log(\lambda^2)\right| ( H ∗ , λ ∗ ) := H , λ arg min ψ ( H + 1/2 ) + ψ ( H ) − log ( λ 2 ) .
Note that ξ ↦ e − δ ξ ( L 0 + L 1 log ( ξ ) ) \xi\mapsto e^{-\delta\xi}(L_0+L_1\log(\xi)) ξ ↦ e − δ ξ ( L 0 + L 1 log ( ξ )) is strictly decreasing on ( e 1 / δ , ∞ ) (e^{1/\delta},\infty) ( e 1/ δ , ∞ ) . The principal minor ⟨ 1 , 2 ⟩ Σ \langle1,2\rangle_\Sigma ⟨ 1 , 2 ⟩ Σ has the form (see notations (17) and (18) )
⟨ 1 , 2 ⟩ Σ = Σ 3 , 3 = κ σ 2 { ζ u 0 + ζ ( t − 4 z + 6 y − 4 x ) − 4 α ( d t − 4 c z + 6 b y − 4 a x ) } = κ σ 2 u 0 ( − 2 H λ + R 2 ( ξ ) u 0 ) . \begin{aligned}
\langle1,2\rangle_\Sigma=\Sigma_{3,3}=\kappa\sigma^2\{\zeta u_0+\zeta(t-4z+6y-4x)-4\alpha(dt-4cz+6by-4ax)\} \\
&=\kappa\sigma^2u_0\left(-\frac{2H}{\lambda}+\frac{R_2(\xi)}{u_0}\right).
\end{aligned} ⟨ 1 , 2 ⟩ Σ = Σ 3 , 3 = κ σ 2 { ζ u 0 + ζ ( t − 4 z + 6 y − 4 x ) − 4 α ( d t − 4 cz + 6 b y − 4 a x )} = κ σ 2 u 0 ( − λ 2 H + u 0 R 2 ( ξ ) ) .
Let H > 1 / 2 H>1/2 H > 1/2 . Since ξ ↦ r H ( ξ ) = K H − 1 ( ξ ) / K H ( ξ ) \xi\mapsto r_H(\xi)=K_{H-1}(\xi)/K_H(\xi) ξ ↦ r H ( ξ ) = K H − 1 ( ξ ) / K H ( ξ ) is continuous and increasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) (e.g., [10 p. 376] ) with
lim ξ → 0 r H ( ξ ) = 0 , lim ξ → ∞ r H ( ξ ) = 1 \lim_{\xi\to0}r_H(\xi)=0,\qquad\lim_{\xi\to\infty}r_H(\xi)=1 ξ → 0 lim r H ( ξ ) = 0 , ξ → ∞ lim r H ( ξ ) = 1
we get
0 < r ∗ : = inf H r H ( 1 ) < a < b < c < d < 1 , for all ξ > 1. (64) 0<r^*:=\inf_H r_H(1)<a<b<c<d<1,\qquad\text{for all }\xi>1. \tag*{(64)} 0 < r ∗ := H inf r H ( 1 ) < a < b < c < d < 1 , for all ξ > 1. (64)
Moreover,
∣ R 2 ( ξ ) u 0 ∣ ≤ 1 u 0 [ ∣ ζ ∣ ( t + 4 z + 6 y + 4 x ) + 4 λ ξ ( d t + 4 c z + 6 b y + 4 a x ) ] ≤ 30 x u 0 ( H λ + 2 ξ λ ) \begin{aligned}
\left|\frac{R_2(\xi)}{u_0}\right|
\leq\frac{1}{u_0}\left[|\zeta|(t+4z+6y+4x)+\frac{4}{\lambda}\xi(dt+4cz+6by+4ax)\right] \\
&\leq\frac{30x}{u_0}\left(\frac{H}{\lambda}+\frac{2\xi}{\lambda}\right)
\end{aligned} u 0 R 2 ( ξ ) ≤ u 0 1 [ ∣ ζ ∣ ( t + 4 z + 6 y + 4 x ) + λ 4 ξ ( d t + 4 cz + 6 b y + 4 a x ) ] ≤ u 0 30 x ( λ H + λ 2 ξ )
≤ 30 λ − 1 ( 1 − δ ) − H e − δ ξ ( H + 2 ξ ) ≤ e − δ ξ ( L 2 + L 3 ξ ) , \begin{aligned}
\leq30\lambda^{-1}(1-\delta)^{-H}e^{-\delta\xi}(H+2\xi) \\
&\leq e^{-\delta\xi}(L_2+L_3\xi),
\end{aligned} ≤ 30 λ − 1 ( 1 − δ ) − H e − δ ξ ( H + 2 ξ ) ≤ e − δ ξ ( L 2 + L 3 ξ ) ,
for all ξ > 1 \xi>1 ξ > 1 , where L 2 = sup H , λ { 30 H λ − 1 ( 1 − δ ) − H } L_2=\sup_{H,\lambda}\{30H\lambda^{-1}(1-\delta)^{-H}\} L 2 = sup H , λ { 30 H λ − 1 ( 1 − δ ) − H } and L 3 = sup H , λ { 60 λ − 1 ( 1 − δ ) − H } L_3=\sup_{H,\lambda}\{60\lambda^{-1}(1-\delta)^{-H}\} L 3 = sup H , λ { 60 λ − 1 ( 1 − δ ) − H } . Once again, we can find some ξ 2 \xi_2 ξ 2 (depending only on the parameter space) so that ⟨ 1 , 2 ⟩ Σ < 0 \langle1,2\rangle_\Sigma<0 ⟨ 1 , 2 ⟩ Σ < 0 , for all ξ > ξ 2 \xi>\xi_2 ξ > ξ 2 .
ii-Evaluating the principal minors of order 2 × 2 2 \times2 2 × 2 . In the sequel we let ξ > 1 ∨ e M ∗ \xi> 1 \vee e^{M^*} ξ > 1 ∨ e M ∗ so that A ∗ ≥ 1 A^* \ge1 A ∗ ≥ 1 (as shown in (56)). Keep in mind notations (15)-(18) and set Σ ~ 1 , 1 = ( y − 4 x + 3 u 0 ) \widetilde{\Sigma}_{1,1}=(y-4x+3u_0) Σ 1 , 1 = ( y − 4 x + 3 u 0 ) , Σ ~ 2 , 1 = ( z − 4 y + 7 x − 4 u 0 ) \widetilde{\Sigma}_{2,1}=(z-4y+7x-4u_0) Σ 2 , 1 = ( z − 4 y + 7 x − 4 u 0 ) , Σ ~ 3 , 1 = ( t − 4 z + 6 y − 4 x + u 0 ) \widetilde{\Sigma}_{3,1}=(t-4z+6y-4x+u_0) Σ 3 , 1 = ( t − 4 z + 6 y − 4 x + u 0 ) . Then, by straightforward computations we get
⟨ 3 ⟩ Σ = ∣ 2 κ Σ ~ 1 , 1 2 κ σ 2 [ ξ Σ ~ 1 , 1 + ( B y − 4 A x + 3 u ∗ ) ] κ Σ ~ 2 , 1 κ σ 2 [ ξ Σ ~ 2 , 1 + ( C z − 4 B y + 7 A x − 4 u ∗ ) ] ∣ = 2 κ 2 σ 2 ∣ Σ ~ 1 , 1 ξ Σ ~ 1 , 1 + ( B y − 4 A x + 3 u ∗ ) Σ ~ 2 , 1 ξ Σ ~ 2 , 1 + ( C z − 4 B y + 7 A x − 4 u ∗ ) ∣ = 2 κ 2 σ 2 ∣ Σ ~ 1 , 1 B y − 4 A x + 3 u ∗ Σ ~ 2 , 1 C z − 4 B y + 7 A x − 4 u ∗ ∣ = 2 κ 2 σ 2 ∣ y − 4 x + 3 u 0 B y − 4 A x + 3 u ∗ z − 4 y + 7 x − 4 u 0 C z − 4 B y + 7 A x − 4 u ∗ ∣ = 2 κ 2 σ 2 ∣ y − 4 x + 3 u 0 ( B − A ) y − 3 A ∗ z − 4 y + 7 x − 4 u 0 ( C − A ) z − 4 ( B − A ) y + 4 A ∗ ∣ , A ∗ = A ω 0 − u ∗ = 2 κ 2 σ 2 ∣ ( y − 4 x ) + 3 u 0 ( B − A ) y − 3 A ∗ ( z − 9 x ) + 8 u 0 ( C − A ) z − 8 A ∗ ∣ = 2 κ 2 σ 2 ⋅ { ∣ y − 4 x ( B − A ) y z − 9 x ( C − A ) z ∣ + ∣ y − 4 x − 3 A ∗ z − 9 x − 8 A ∗ ∣ + ∣ 3 u 0 ( B − A ) y 8 u 0 ( C − A ) z ∣ } = 2 κ 2 σ 2 A ∗ x ⋅ { ∣ ( y / x − 4 ) ( B − A ) y / A ∗ ( z / x − 9 ) ( C − A ) z / A ∗ ∣ + ∣ ( y / x − 4 ) − 3 ( z / x − 9 ) − 8 ∣ + ∣ 3 u 0 / A ∗ ( B − A ) y / x 8 u 0 / A ∗ ( C − A ) z / x ∣ } = : 2 κ 2 σ 2 A ∗ x ∑ j = 3 5 R j ( ξ ) . \begin{aligned}
\langle3\rangle_{\Sigma}
&=
\begin{vmatrix}
2\kappa\widetilde{\Sigma}_{1,1} & 2\kappa\sigma^2[\xi\widetilde{\Sigma}_{1,1}+(By-4Ax+3u^*)]\\
\kappa\widetilde{\Sigma}_{2,1} & \kappa\sigma^2[\xi\widetilde{\Sigma}_{2,1}+(Cz-4By+7Ax-4u^*)]
\end{vmatrix}\\
&=2\kappa^2\sigma^2
\begin{vmatrix}
\widetilde{\Sigma}_{1,1} & \xi\widetilde{\Sigma}_{1,1}+(By-4Ax+3u^*)\\
\widetilde{\Sigma}_{2,1} & \xi\widetilde{\Sigma}_{2,1}+(Cz-4By+7Ax-4u^*)
\end{vmatrix}\\
&=2\kappa^2\sigma^2
\begin{vmatrix}
\widetilde{\Sigma}_{1,1} & By-4Ax+3u^*\\
\widetilde{\Sigma}_{2,1} & Cz-4By+7Ax-4u^*
\end{vmatrix}\\
&=2\kappa^2\sigma^2
\begin{vmatrix}
y-4x+3u_0 & By-4Ax+3u^*\\
z-4y+7x-4u_0 & Cz-4By+7Ax-4u^*
\end{vmatrix}\\
&=2\kappa^2\sigma^2
\begin{vmatrix}
y-4x+3u_0 & (B-A)y-3A^*\\
z-4y+7x-4u_0 & (C-A)z-4(B-A)y+4A^*
\end{vmatrix},
\qquad A^*=A\omega_0-u^*\\
&=2\kappa^2\sigma^2
\begin{vmatrix}
(y-4x)+3u_0 & (B-A)y-3A^*\\
(z-9x)+8u_0 & (C-A)z-8A^*
\end{vmatrix}\\
&=2\kappa^2\sigma^2\cdot
\left\{
\begin{vmatrix}
y-4x & (B-A)y\\
z-9x & (C-A)z
\end{vmatrix}
+
\begin{vmatrix}
y-4x & -3A^*\\
z-9x & -8A^*
\end{vmatrix}
+
\begin{vmatrix}
3u_0 & (B-A)y\\
8u_0 & (C-A)z
\end{vmatrix}
\right\}\\
&=2\kappa^2\sigma^2 A^*x\cdot
\left\{
\begin{vmatrix}
(y/x-4) & (B-A)y/A^*\\
(z/x-9) & (C-A)z/A^*
\end{vmatrix}
+
\begin{vmatrix}
(y/x-4) & -3\\
(z/x-9) & -8
\end{vmatrix}
+
\begin{vmatrix}
3u_0/A^* & (B-A)y/x\\
8u_0/A^* & (C-A)z/x
\end{vmatrix}
\right\}\\
&=:2\kappa^2\sigma^2 A^*x\sum_{j=3}^{5}\mathcal{R}_j(\xi).
\end{aligned} ⟨ 3 ⟩ Σ = 2 κ Σ 1 , 1 κ Σ 2 , 1 2 κ σ 2 [ ξ Σ 1 , 1 + ( B y − 4 A x + 3 u ∗ )] κ σ 2 [ ξ Σ 2 , 1 + ( C z − 4 B y + 7 A x − 4 u ∗ )] = 2 κ 2 σ 2 Σ 1 , 1 Σ 2 , 1 ξ Σ 1 , 1 + ( B y − 4 A x + 3 u ∗ ) ξ Σ 2 , 1 + ( C z − 4 B y + 7 A x − 4 u ∗ ) = 2 κ 2 σ 2 Σ 1 , 1 Σ 2 , 1 B y − 4 A x + 3 u ∗ C z − 4 B y + 7 A x − 4 u ∗ = 2 κ 2 σ 2 y − 4 x + 3 u 0 z − 4 y + 7 x − 4 u 0 B y − 4 A x + 3 u ∗ C z − 4 B y + 7 A x − 4 u ∗ = 2 κ 2 σ 2 y − 4 x + 3 u 0 z − 4 y + 7 x − 4 u 0 ( B − A ) y − 3 A ∗ ( C − A ) z − 4 ( B − A ) y + 4 A ∗ , A ∗ = A ω 0 − u ∗ = 2 κ 2 σ 2 ( y − 4 x ) + 3 u 0 ( z − 9 x ) + 8 u 0 ( B − A ) y − 3 A ∗ ( C − A ) z − 8 A ∗ = 2 κ 2 σ 2 ⋅ { y − 4 x z − 9 x ( B − A ) y ( C − A ) z + y − 4 x z − 9 x − 3 A ∗ − 8 A ∗ + 3 u 0 8 u 0 ( B − A ) y ( C − A ) z } = 2 κ 2 σ 2 A ∗ x ⋅ { ( y / x − 4 ) ( z / x − 9 ) ( B − A ) y / A ∗ ( C − A ) z / A ∗ + ( y / x − 4 ) ( z / x − 9 ) − 3 − 8 + 3 u 0 / A ∗ 8 u 0 / A ∗ ( B − A ) y / x ( C − A ) z / x } =: 2 κ 2 σ 2 A ∗ x j = 3 ∑ 5 R j ( ξ ) .
We can evaluate the last three determinants R j ( ξ ) \mathcal{R}_j(\xi) R j ( ξ ) , in terms of g ( ξ ) : = g H ( ξ ) g(\xi):=gH(\xi) g ( ξ ) := g H ( ξ ) and φ q ( ξ ) = g ( q ξ ) / g ( ξ ) \varphi_q(\xi)=g(q\xi)/g(\xi) φ q ( ξ ) = g ( q ξ ) / g ( ξ ) , R ( ξ ) : = R H ( q ξ ) \mathcal{R}(\xi):=\mathcal{R}_H(q\xi) R ( ξ ) := R H ( q ξ ) , q = 1 , 2 , 3 q=1,2,3 q = 1 , 2 , 3 as
∣ R 3 ( ξ ) ∣ = ∣ z A ∗ x ( C − A ) ( y / x − 4 ) − y A ∗ x ( B − A ) ( z / x − 9 ) ∣ ≤ ∣ g ( 3 ξ ) A ∗ [ R ( 3 ξ ) − R ( ξ ) ] ∣ ( φ 2 ( ξ ) + 4 ) + ∣ g ( 2 ξ ) A ∗ [ R ( 2 ξ ) − R ( ξ ) ] ∣ ( φ 3 ( ξ ) + 9 ) ≤ M 0 A ∗ g ( 2 ξ ) ∣ φ 2 ( ξ ) + 4 + φ 3 ( ξ ) + 9 ∣ , (By (52)) ≤ M 0 A ∗ u 0 ( 1 − δ ) − H e − 2 δ ξ ( 13 + 2 H e − ξ + 3 H e − 2 ξ ) , (By (49)) ≤ M 0 e − 2 δ ξ sup H { u 0 ( 1 − δ ) − H ( 13 + 2 H + 3 H ) } , for all ξ > e M ∗ . \begin{aligned}
|\mathcal{R}_3(\xi)|
&=
\left|
\frac{z}{A^*x}(C-A)(y/x-4)-\frac{y}{A^*x}(B-A)(z/x-9)
\right|\\
&\le
\left|
\frac{g(3\xi)}{A^*}[R(3\xi)-R(\xi)]
\right|(\varphi_2(\xi)+4)
+
\left|
\frac{g(2\xi)}{A^*}[R(2\xi)-R(\xi)]
\right|(\varphi_3(\xi)+9)\\
&\le\frac{M_0}{A^*}g(2\xi)\left|\varphi_2(\xi)+4+\varphi_3(\xi)+9\right|,
\qquad\text{(By (52))}\\
&\le\frac{M_0}{A^*}u_0(1-\delta)^{-H}e^{-2\delta\xi}(13+2^He^{-\xi}+3^He^{-2\xi}),
\qquad\text{(By (49))}\\
&\le M_0e^{-2\delta\xi}\sup_H\left\{u_0(1-\delta)^{-H}(13+2^H+3^H)\right\},
\qquad\text{for all }\xi>e^{M^*}.
\end{aligned} ∣ R 3 ( ξ ) ∣ = A ∗ x z ( C − A ) ( y / x − 4 ) − A ∗ x y ( B − A ) ( z / x − 9 ) ≤ A ∗ g ( 3 ξ ) [ R ( 3 ξ ) − R ( ξ )] ( φ 2 ( ξ ) + 4 ) + A ∗ g ( 2 ξ ) [ R ( 2 ξ ) − R ( ξ )] ( φ 3 ( ξ ) + 9 ) ≤ A ∗ M 0 g ( 2 ξ ) ∣ φ 2 ( ξ ) + 4 + φ 3 ( ξ ) + 9 ∣ , (By (52)) ≤ A ∗ M 0 u 0 ( 1 − δ ) − H e − 2 δ ξ ( 13 + 2 H e − ξ + 3 H e − 2 ξ ) , (By (49)) ≤ M 0 e − 2 δ ξ H sup { u 0 ( 1 − δ ) − H ( 13 + 2 H + 3 H ) } , for all ξ > e M ∗ .
On the other hand, R 4 ( ξ ) = − 8 ( y / x − 4 ) + 3 ( z / x − 9 ) = 5 − 8 φ 2 ( ξ ) + 3 φ 3 ( ξ ) \mathcal{R}_4(\xi)=-8(y/x-4)+3(z/x-9)=5-8\varphi_2(\xi)+3\varphi_3(\xi) R 4 ( ξ ) = − 8 ( y / x − 4 ) + 3 ( z / x − 9 ) = 5 − 8 φ 2 ( ξ ) + 3 φ 3 ( ξ ) with
∣ − 8 φ 2 ( ξ ) + 3 φ 3 ( ξ ) ∣ ≤ 8 φ 2 ( ξ ) + 3 φ 3 ( ξ ) ≤ 2 H + 3 e − ξ + 3 H + 1 e − 2 ξ , (By the second part in (49)) ≤ ( 2 H + 3 + 3 H + 1 ) e − ξ , for all ξ > 1. \begin{aligned}
|-8\varphi_2(\xi)+3\varphi_3(\xi)|
&\le8\varphi_2(\xi)+3\varphi_3(\xi)\\
&\le2^{H+3}e^{-\xi}+3^{H+1}e^{-2\xi},
\qquad\text{(By the second part in (49))}\\
&\le(2^{H+3}+3^{H+1})e^{-\xi},
\qquad\text{for all }\xi>1.
\end{aligned} ∣ − 8 φ 2 ( ξ ) + 3 φ 3 ( ξ ) ∣ ≤ 8 φ 2 ( ξ ) + 3 φ 3 ( ξ ) ≤ 2 H + 3 e − ξ + 3 H + 1 e − 2 ξ , (By the second part in (49)) ≤ ( 2 H + 3 + 3 H + 1 ) e − ξ , for all ξ > 1.
and for ξ > 1 ∨ e M ∗ \xi> 1 \lor e^{M^*} ξ > 1 ∨ e M ∗ (see (56) ) we have
∣ R 5 ( ξ ) ∣ = u 0 A ∗ ∣ 3 ( C − A ) z x − 8 ( B − A ) y x ∣ ≤ u 0 [ 3 φ 3 ( ξ ) ∣ R ( 3 ξ ) − R ( ξ ) ∣ + 8 φ 2 ( ξ ) ∣ R ( 2 ξ ) − R ( ξ ) ∣ ] ≤ u 0 [ log ( 3 ) + 3 M ( 2 H + 1 ) ] ⋅ [ 8 φ 2 ( ξ ) + 3 φ 3 ( ξ ) ] , (By (iv)-Lemma 3) ≤ u 0 [ log ( 3 ) + 3 M ( 2 H + 1 ) ] ⋅ [ 2 H + 3 e − ξ + 3 H + 1 e − 2 ξ ] ≤ L 4 e − ξ , \begin{aligned}
|\mathcal{R}_5(\xi)| &= \frac{u_0}{A^*}\left|3(C-A)\frac{z}{x}-8(B-A)\frac{y}{x}\right| \\
&\le u_0[3\varphi_3(\xi)|R(3\xi)-R(\xi)|+8\varphi_2(\xi)|R(2\xi)-R(\xi)|] \\
&\le u_0[\log(3)+3\mathcal{M}(2H+1)]\cdot[8\varphi_2(\xi)+3\varphi_3(\xi)],\quad\text{(By (iv)-Lemma 3)} \\
&\le u_0[\log(3)+3\mathcal{M}(2H+1)]\cdot[2^{H+3}e^{-\xi}+3^{H+1}e^{-2\xi}] \\
&\le L_4e^{-\xi},
\end{aligned} ∣ R 5 ( ξ ) ∣ = A ∗ u 0 3 ( C − A ) x z − 8 ( B − A ) x y ≤ u 0 [ 3 φ 3 ( ξ ) ∣ R ( 3 ξ ) − R ( ξ ) ∣ + 8 φ 2 ( ξ ) ∣ R ( 2 ξ ) − R ( ξ ) ∣ ] ≤ u 0 [ log ( 3 ) + 3 M ( 2 H + 1 )] ⋅ [ 8 φ 2 ( ξ ) + 3 φ 3 ( ξ )] , (By (iv)-Lemma 3) ≤ u 0 [ log ( 3 ) + 3 M ( 2 H + 1 )] ⋅ [ 2 H + 3 e − ξ + 3 H + 1 e − 2 ξ ] ≤ L 4 e − ξ ,
where L 4 = sup H { u 0 [ log ( 3 ) + 3 M ( 2 H + 1 ) ] ⋅ [ 2 H + 3 + 3 H + 1 ] } L_4=\sup_H\{u_0[\log(3)+3\mathcal{M}(2H+1)]\cdot[2^{H+3}+3^{H+1}]\} L 4 = sup H { u 0 [ log ( 3 ) + 3 M ( 2 H + 1 )] ⋅ [ 2 H + 3 + 3 H + 1 ]} . Putting the previous estimates of R j ( ξ ) \mathcal{R}_j(\xi) R j ( ξ ) together, we can find some ξ 3 \xi_3 ξ 3 (depending only on the parameter space) so that ⟨ 3 ⟩ Σ > 0 \langle3\rangle\Sigma>0 ⟨ 3 ⟩ Σ > 0 , for all ξ > ξ 3 \xi>\xi_3 ξ > ξ 3 .
Similarly, we evaluate the principal minor ⟨ 2 ⟩ Σ \langle2\rangle_\Sigma ⟨ 2 ⟩ Σ as
⟨ 2 ⟩ Σ = ∣ 2 κ Σ ~ 1 , 1 2 κ σ 2 [ Σ ~ 1 , 1 − 2 α ( b y − 2 a x ) ] κ Σ ~ 3 , 1 κ σ 2 [ Σ ~ 3 , 1 − 4 α ( d t − 3 c z + 3 b y − a x ) ] ∣ = 2 κ 2 σ 2 ∣ Σ ~ 1 , 1 ξ Σ ~ 1 , 1 − 2 α ( b y − 2 a x ) Σ ~ 3 , 1 ξ Σ ~ 3 , 1 − 4 α ( d t − 3 c z + 3 b y − a x ) ∣ = 2 κ 2 σ 2 ∣ Σ ~ 1 , 1 − 2 α ( b y − 2 a x ) Σ ~ 3 , 1 − 4 α ( d t − 3 c z + 3 b y − a x ) ∣ = − 4 α κ 2 σ 2 ∣ y − 4 x + 3 u 0 b y − 2 a x t − 4 z + 6 y − 4 x + u 0 2 d t − 6 c z + 6 b y − 2 a x ∣ = − 4 α κ 2 σ 2 { ∣ y − 4 x b y t − 4 z + 6 y − 4 x 2 d t − 6 c z + 6 b y ∣ + ∣ y − 4 x − 2 a x t − 4 z + 6 y − 4 x − 2 a x ∣ + ∣ 3 u 0 b y u 0 2 d t − 6 c z + 6 b y ∣ + ∣ 3 u 0 − 2 a x u 0 − 2 a x ∣ } = − 4 α κ 2 σ 2 ⋅ ( a u 0 x ) { 1 a u 0 ∣ y / x − 4 b y / x t − 4 z + 6 y − 4 x 2 d t − 6 c z + 6 b y ∣ − 2 u 0 ∣ y − 4 x 1 t − 4 z + 6 y − 4 x 1 ∣ + 1 a ∣ 3 b y / x 1 ( 2 d t − 6 c z + 6 b y ) / x ∣ − 2 ∣ 3 1 1 1 ∣ } = − 4 a u 0 α κ 2 σ 2 x ( ∑ j = 6 8 R j ( ξ ) − 4 ) , ξ = λ α . \begin{aligned}
\langle2\rangle_\Sigma
&=\left|\begin{array}{cc}
2\kappa\widetilde{\Sigma}_{1,1} & 2\kappa\sigma^2[\widetilde{\Sigma}_{1,1}-2\alpha(by-2ax)]\\
\kappa\widetilde{\Sigma}_{3,1} & \kappa\sigma^2[\widetilde{\Sigma}_{3,1}-4\alpha(dt-3cz+3by-ax)]
\end{array}\right|\\
&=2\kappa^2\sigma^2\left|\begin{array}{cc}
\widetilde{\Sigma}_{1,1} & \xi\widetilde{\Sigma}_{1,1}-2\alpha(by-2ax)\\
\widetilde{\Sigma}_{3,1} & \xi\widetilde{\Sigma}_{3,1}-4\alpha(dt-3cz+3by-ax)
\end{array}\right|\\
&=2\kappa^2\sigma^2\left|\begin{array}{cc}
\widetilde{\Sigma}_{1,1} & -2\alpha(by-2ax)\\
\widetilde{\Sigma}_{3,1} & -4\alpha(dt-3cz+3by-ax)
\end{array}\right|\\
&=-4\alpha\kappa^2\sigma^2\left|\begin{array}{cc}
y-4x+3u_0 & by-2ax\\
t-4z+6y-4x+u_0 & 2dt-6cz+6by-2ax
\end{array}\right|\\
&=-4\alpha\kappa^2\sigma^2\left\{\left|\begin{array}{cc}
y-4x & by\\
t-4z+6y-4x & 2dt-6cz+6by
\end{array}\right|+\left|\begin{array}{cc}
y-4x & -2ax\\
t-4z+6y-4x & -2ax
\end{array}\right|\right.\\
&\qquad\left.+\left|\begin{array}{cc}
3u_0 & by\\
u_0 & 2dt-6cz+6by
\end{array}\right|+\left|\begin{array}{cc}
3u_0 & -2ax\\
u_0 & -2ax
\end{array}\right|\right\}\\
&=-4\alpha\kappa^2\sigma^2\cdot(au_0x)\left\{\frac{1}{au_0}\left|\begin{array}{cc}
y/x-4 & by/x\\
t-4z+6y-4x & 2dt-6cz+6by
\end{array}\right|-\frac{2}{u_0}\left|\begin{array}{cc}
y-4x & 1\\
t-4z+6y-4x & 1
\end{array}\right|\right.\\
&\qquad\left.+\frac{1}{a}\left|\begin{array}{cc}
3 & by/x\\
1 & (2dt-6cz+6by)/x
\end{array}\right|-2\left|\begin{array}{cc}
3 & 1\\
1 & 1
\end{array}\right|\right\}\\
&=-4au_0\alpha\kappa^2\sigma^2x\left(\sum_{j=6}^{8}\mathcal{R}_j(\xi)-4\right),\quad\xi=\lambda\alpha.
\end{aligned} ⟨ 2 ⟩ Σ = 2 κ Σ 1 , 1 κ Σ 3 , 1 2 κ σ 2 [ Σ 1 , 1 − 2 α ( b y − 2 a x )] κ σ 2 [ Σ 3 , 1 − 4 α ( d t − 3 cz + 3 b y − a x )] = 2 κ 2 σ 2 Σ 1 , 1 Σ 3 , 1 ξ Σ 1 , 1 − 2 α ( b y − 2 a x ) ξ Σ 3 , 1 − 4 α ( d t − 3 cz + 3 b y − a x ) = 2 κ 2 σ 2 Σ 1 , 1 Σ 3 , 1 − 2 α ( b y − 2 a x ) − 4 α ( d t − 3 cz + 3 b y − a x ) = − 4 α κ 2 σ 2 y − 4 x + 3 u 0 t − 4 z + 6 y − 4 x + u 0 b y − 2 a x 2 d t − 6 cz + 6 b y − 2 a x = − 4 α κ 2 σ 2 { y − 4 x t − 4 z + 6 y − 4 x b y 2 d t − 6 cz + 6 b y + y − 4 x t − 4 z + 6 y − 4 x − 2 a x − 2 a x + 3 u 0 u 0 b y 2 d t − 6 cz + 6 b y + 3 u 0 u 0 − 2 a x − 2 a x } = − 4 α κ 2 σ 2 ⋅ ( a u 0 x ) { a u 0 1 y / x − 4 t − 4 z + 6 y − 4 x b y / x 2 d t − 6 cz + 6 b y − u 0 2 y − 4 x t − 4 z + 6 y − 4 x 1 1 + a 1 3 1 b y / x ( 2 d t − 6 cz + 6 b y ) / x − 2 3 1 1 1 } = − 4 a u 0 α κ 2 σ 2 x ( j = 6 ∑ 8 R j ( ξ ) − 4 ) , ξ = λ α .
The terms R j ( ξ ) \mathcal{R}_j(\xi) R j ( ξ ) can be controlled to be small. In fact, by using (64) and the fact: t < z < y < x t<z<y<x t < z < y < x (inherited from the monotony of ξ ↦ g H ( ξ ) \xi\mapsto g_H(\xi) ξ ↦ g H ( ξ ) ) we obtain
∣ R 6 ( ξ ) ∣ = 1 r ∗ u 0 ∣ ( 2 d − b ) y t x + ( − 6 c + 4 b ) y z x − 8 d t + 24 c z − 20 b y ∣ ≤ 1 r ∗ u 0 ( y x ( 2 t + 6 z ) + 8 t + 24 z + 20 y ) ≤ 1 r ∗ u 0 ( 8 y + 8 t + 24 z + 20 y ) ≤ 60 y r ∗ u 0 ≤ 60 r ∗ u 0 g ( 2 ξ ) ≤ 60 r ∗ ( 1 − δ ) − H ‾ e − 2 δ ξ , (By the first part in (49)), \begin{aligned}
|\mathcal{R}_6(\xi)| &= \frac{1}{r^*u_0}\left|(2d-b)\frac{yt}{x}+(-6c+4b)\frac{yz}{x}-8dt+24cz-20by\right|\\
&\le\frac{1}{r^*u_0}\left(\frac{y}{x}(2t+6z)+8t+24z+20y\right)\\
&\le\frac{1}{r^*u_0}(8y+8t+24z+20y)\le\frac{60y}{r^*u_0}\\
&\le\frac{60}{r^*u_0}g(2\xi)\le\frac{60}{r^*}(1-\delta)^{-\overline{H}}e^{-2\delta\xi},\qquad\text{(By the first part in (49)),}
\end{aligned} ∣ R 6 ( ξ ) ∣ = r ∗ u 0 1 ( 2 d − b ) x y t + ( − 6 c + 4 b ) x y z − 8 d t + 24 cz − 20 b y ≤ r ∗ u 0 1 ( x y ( 2 t + 6 z ) + 8 t + 24 z + 20 y ) ≤ r ∗ u 0 1 ( 8 y + 8 t + 24 z + 20 y ) ≤ r ∗ u 0 60 y ≤ r ∗ u 0 60 g ( 2 ξ ) ≤ r ∗ 60 ( 1 − δ ) − H e − 2 δ ξ , (By the first part in (49)),
∣ R 7 ( ξ ) ∣ = 2 u 0 ∣ ( y − 4 x ) − ( t − 4 z + 6 y − 4 x ) ∣ = 2 u 0 ∣ − t + 4 z − 5 y ∣ ≤ 20 g ( 2 ξ ) u 0 ≤ 20 ( 1 − δ ) − H ‾ e − 2 δ ξ , (By the first part in (49)) \begin{aligned}
|\mathcal{R}_7(\xi)| &= \frac{2}{u_0}|(y-4x)-(t-4z+6y-4x)|=\frac{2}{u_0}|-t+4z-5y|\\
&\le\frac{20g(2\xi)}{u_0}\le20(1-\delta)^{-\overline{H}}e^{-2\delta\xi},\qquad\text{(By the first part in (49))}
\end{aligned} ∣ R 7 ( ξ ) ∣ = u 0 2 ∣ ( y − 4 x ) − ( t − 4 z + 6 y − 4 x ) ∣ = u 0 2 ∣ − t + 4 z − 5 y ∣ ≤ u 0 20 g ( 2 ξ ) ≤ 20 ( 1 − δ ) − H e − 2 δ ξ , (By the first part in (49))
and
∣ R 8 ( ξ ) ∣ = 1 a ∣ 6 d t x − 18 c z x + 17 b y x ∣ ≤ 41 y a x ≤ 41 r ∗ φ 2 ( ξ ) ≤ 41 r ∗ 2 H ‾ e − ξ , (By the second part in (49)) \begin{aligned}
|\mathcal{R}_8(\xi)| &= \frac{1}{a}\left|6\frac{dt}{x}-18\frac{cz}{x}+17\frac{by}{x}\right|\le\frac{41y}{ax}\\
&\le\frac{41}{r^*}\varphi_2(\xi)\le\frac{41}{r^*}2^{\overline{H}}e^{-\xi},\qquad\text{(By the second part in (49))}
\end{aligned} ∣ R 8 ( ξ ) ∣ = a 1 6 x d t − 18 x cz + 17 x b y ≤ a x 41 y ≤ r ∗ 41 φ 2 ( ξ ) ≤ r ∗ 41 2 H e − ξ , (By the second part in (49))
Combining the previous estimates of R j ( ξ ) \mathcal{R}_j(\xi) R j ( ξ ) we can find some ξ 4 > 0 \xi_4>0 ξ 4 > 0 (depending only on the parameter space), so that ⟨ 2 ⟩ Σ > 0 \langle2\rangle_\Sigma>0 ⟨ 2 ⟩ Σ > 0 , for all ξ > ξ 4 \xi>\xi_4 ξ > ξ 4 .
Now, let us examine the principal minor ⟨ 1 ⟩ Σ \langle1\rangle_\Sigma ⟨ 1 ⟩ Σ . Recall that Σ ~ 2 , 1 = ( z − 4 y + 7 x − 4 u 0 ) \widetilde{\Sigma}_{2,1}=(z-4y+7x-4u_0) Σ 2 , 1 = ( z − 4 y + 7 x − 4 u 0 ) , Σ ~ 3 , 1 = ( t − 4 z + 6 y − 4 x + u 0 ) \widetilde{\Sigma}_{3,1}=(t-4z+6y-4x+u_0) Σ 3 , 1 = ( t − 4 z + 6 y − 4 x + u 0 ) , ξ = − log ( 2 λ 2 ) + ψ ( H + 1 / 2 ) \xi=-\log(2\lambda^2)+\psi(H+1/2) ξ = − log ( 2 λ 2 ) + ψ ( H + 1/2 ) , ζ = − 2 H / λ \zeta=-2H/\lambda ζ = − 2 H / λ and write
ξ Σ ~ 2 , 1 + ( C z − 4 B y + 7 A x − 4 u ∗ ) = − 4 ( u 0 ξ + u ∗ ) + ξ ( z − 4 y + 7 x ) + ( C z − 4 B y + 7 A x ) = : f 1 ( ξ ) \xi\widetilde{\Sigma}_{2,1}+(Cz-4By+7Ax-4u^*)=-4(u_0\xi+u^*)+\xi(z-4y+7x)+(Cz-4By+7Ax)=:f_1(\xi) ξ Σ 2 , 1 + ( C z − 4 B y + 7 A x − 4 u ∗ ) = − 4 ( u 0 ξ + u ∗ ) + ξ ( z − 4 y + 7 x ) + ( C z − 4 B y + 7 A x ) =: f 1 ( ξ )
ζ Σ ~ 2 , 1 − α ( 3 c z − 8 b y + 7 a x ) = − 4 u 0 ζ − 7 a α x + ζ ( z − 4 y + 7 x ) − α ( 3 c z − 8 b y ) = : f 2 ( ξ ) \zeta\widetilde{\Sigma}_{2,1}-\alpha(3cz-8by+7ax)=-4u_0\zeta-7a\alpha x+\zeta(z-4y+7x)-\alpha(3cz-8by)=:f_2(\xi) ζ Σ 2 , 1 − α ( 3 cz − 8 b y + 7 a x ) = − 4 u 0 ζ − 7 a α x + ζ ( z − 4 y + 7 x ) − α ( 3 cz − 8 b y ) =: f 2 ( ξ )
ξ Σ ~ 3 , 1 + ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) = ( u 0 ξ + u ∗ ) + ξ ( t − 4 z + 6 y − 4 x ) + ( D t − 4 C z + 6 B y − 4 A x ) = : f 3 ( ξ ) \xi\widetilde{\Sigma}_{3,1}+(Dt-4Cz+6By-4Ax+u^*)=(u_0\xi+u^*)+\xi(t-4z+6y-4x)+(Dt-4Cz+6By-4Ax)=:f_3(\xi) ξ Σ 3 , 1 + ( D t − 4 C z + 6 B y − 4 A x + u ∗ ) = ( u 0 ξ + u ∗ ) + ξ ( t − 4 z + 6 y − 4 x ) + ( D t − 4 C z + 6 B y − 4 A x ) =: f 3 ( ξ )
ζ Σ ~ 3 , 1 − 4 α ( d t − 3 c z + 3 b y − a x ) = u 0 ζ + 4 a α x + ζ ( t − 4 z + 6 y − 4 x ) − 4 α ( d t − 3 c z + 3 b y ) = : f 4 ( ξ ) \zeta\widetilde{\Sigma}_{3,1}-4\alpha(dt-3cz+3by-ax)=u_0\zeta+4a\alpha x+\zeta(t-4z+6y-4x)-4\alpha(dt-3cz+3by)=:f_4(\xi) ζ Σ 3 , 1 − 4 α ( d t − 3 cz + 3 b y − a x ) = u 0 ζ + 4 a α x + ζ ( t − 4 z + 6 y − 4 x ) − 4 α ( d t − 3 cz + 3 b y ) =: f 4 ( ξ )
Then the principal minor ⟨ 1 ⟩ Σ \langle1\rangle_{\Sigma} ⟨ 1 ⟩ Σ simplifies to
⟨ 1 ⟩ Σ = κ 2 σ 4 ∣ − 4 ( u 0 ξ + u ∗ ) + f 1 ( ξ ) − 4 u 0 ζ − 7 a α x + f 2 ( ξ ) ( u 0 ξ + u ∗ ) + f 3 ( ξ ) u 0 ζ + 4 a α x + f 4 ( ξ ) ∣ = κ 2 σ 4 { ∣ − 4 ( u 0 ξ + u ∗ ) − 4 u 0 ζ − 7 a α x ( u 0 ξ + u ∗ ) u 0 ζ + 4 a α x ∣ + ∣ − 4 ( u 0 ξ + u ∗ ) f 2 ( ξ ) ( u 0 ξ + u ∗ ) f 4 ( ξ ) ∣ + ∣ f 1 ( ξ ) − 4 u 0 ζ − 7 a α x f 3 ( ξ ) u 0 ζ + 4 a α x ∣ + ∣ f 1 ( ξ ) f 2 ( ξ ) f 3 ( ξ ) f 4 ( ξ ) ∣ } = κ 2 σ 4 ⋅ ( a α x ) { ∣ − 4 ( u 0 ξ + u ∗ ) − 7 ( u 0 ξ + u ∗ ) 4 ∣ + ∣ − 4 ( u 0 ξ + u ∗ ) f 2 ( ξ ) / ( a α x ) ( u 0 ξ + u ∗ ) f 4 ( ξ ) / ( a α x ) ∣ + ∣ f 1 ( ξ ) / x − 4 u 0 ζ / ( a α ) − 7 x f 3 ( ξ ) / x u 0 ζ / ( a α ) + 4 x ∣ + ∣ f 1 ( ξ ) f 2 ( ξ ) / ( a α x ) f 3 ( ξ ) f 4 ( ξ ) / ( a α x ) ∣ } = a α κ 2 σ 4 x { − 9 ( u 0 ξ + u ∗ ) + ( u 0 ξ + u ∗ ) ∣ − 4 f 2 ( ξ ) / ( a α x ) 1 f 4 ( ξ ) / ( a α x ) ∣ + ∣ f 1 ( ξ ) / x − 4 u 0 ζ / ( a α ) − 7 x f 3 ( ξ ) / x u 0 ζ / ( a α ) + 4 x ∣ + ∣ f 1 ( ξ ) f 2 ( ξ ) / ( a α x ) f 3 ( ξ ) f 4 ( ξ ) / ( a α x ) ∣ } = : κ 2 σ 4 ( − 9 ( u 0 ξ + u ∗ ) + ∑ j = 9 11 R j ( ξ ) ) . \begin{aligned}
\langle1\rangle_{\Sigma}
&=\kappa^2\sigma^4\left|\begin{array}{cc}
-4(u_0\xi+u^*)+f_1(\xi) & -4u_0\zeta-7a\alpha x+f_2(\xi)\\
(u_0\xi+u^*)+f_3(\xi) & u_0\zeta+4a\alpha x+f_4(\xi)
\end{array}\right|\\
&=\kappa^2\sigma^4\left\{
\left|\begin{array}{cc}
-4(u_0\xi+u^*) & -4u_0\zeta-7a\alpha x\\
(u_0\xi+u^*) & u_0\zeta+4a\alpha x
\end{array}\right|
+\left|\begin{array}{cc}
-4(u_0\xi+u^*) & f_2(\xi)\\
(u_0\xi+u^*) & f_4(\xi)
\end{array}\right|\right.\\
&\qquad\left.
+\left|\begin{array}{cc}
f_1(\xi) & -4u_0\zeta-7a\alpha x\\
f_3(\xi) & u_0\zeta+4a\alpha x
\end{array}\right|
+\left|\begin{array}{cc}
f_1(\xi) & f_2(\xi)\\
f_3(\xi) & f_4(\xi)
\end{array}\right|\right\}\\
&=\kappa^2\sigma^4\cdot(a\alpha x)\left\{
\left|\begin{array}{cc}
-4(u_0\xi+u^*) & -7\\
(u_0\xi+u^*) & 4
\end{array}\right|
+\left|\begin{array}{cc}
-4(u_0\xi+u^*) & f_2(\xi)/(a\alpha x)\\
(u_0\xi+u^*) & f_4(\xi)/(a\alpha x)
\end{array}\right|\right.\\
&\qquad\left.
+\left|\begin{array}{cc}
f_1(\xi)/x & -4u_0\zeta/(a\alpha)-7x\\
f_3(\xi)/x & u_0\zeta/(a\alpha)+4x
\end{array}\right|
+\left|\begin{array}{cc}
f_1(\xi) & f_2(\xi)/(a\alpha x)\\
f_3(\xi) & f_4(\xi)/(a\alpha x)
\end{array}\right|\right\}\\
&=a\alpha\kappa^2\sigma^4x\left\{-9(u_0\xi+u^*)+(u_0\xi+u^*)\left|\begin{array}{cc}
-4 & f_2(\xi)/(a\alpha x)\\
1 & f_4(\xi)/(a\alpha x)
\end{array}\right|\right.\\
&\qquad\left.
+\left|\begin{array}{cc}
f_1(\xi)/x & -4u_0\zeta/(a\alpha)-7x\\
f_3(\xi)/x & u_0\zeta/(a\alpha)+4x
\end{array}\right|
+\left|\begin{array}{cc}
f_1(\xi) & f_2(\xi)/(a\alpha x)\\
f_3(\xi) & f_4(\xi)/(a\alpha x)
\end{array}\right|\right\}\\
&=:\kappa^2\sigma^4\left(-9(u_0\xi+u^*)+\sum_{j=9}^{11}\mathcal{R}_j(\xi)\right).
\end{aligned} ⟨ 1 ⟩ Σ = κ 2 σ 4 − 4 ( u 0 ξ + u ∗ ) + f 1 ( ξ ) ( u 0 ξ + u ∗ ) + f 3 ( ξ ) − 4 u 0 ζ − 7 a α x + f 2 ( ξ ) u 0 ζ + 4 a α x + f 4 ( ξ ) = κ 2 σ 4 { − 4 ( u 0 ξ + u ∗ ) ( u 0 ξ + u ∗ ) − 4 u 0 ζ − 7 a α x u 0 ζ + 4 a α x + − 4 ( u 0 ξ + u ∗ ) ( u 0 ξ + u ∗ ) f 2 ( ξ ) f 4 ( ξ ) + f 1 ( ξ ) f 3 ( ξ ) − 4 u 0 ζ − 7 a α x u 0 ζ + 4 a α x + f 1 ( ξ ) f 3 ( ξ ) f 2 ( ξ ) f 4 ( ξ ) } = κ 2 σ 4 ⋅ ( a α x ) { − 4 ( u 0 ξ + u ∗ ) ( u 0 ξ + u ∗ ) − 7 4 + − 4 ( u 0 ξ + u ∗ ) ( u 0 ξ + u ∗ ) f 2 ( ξ ) / ( a α x ) f 4 ( ξ ) / ( a α x ) + f 1 ( ξ ) / x f 3 ( ξ ) / x − 4 u 0 ζ / ( a α ) − 7 x u 0 ζ / ( a α ) + 4 x + f 1 ( ξ ) f 3 ( ξ ) f 2 ( ξ ) / ( a α x ) f 4 ( ξ ) / ( a α x ) } = a α κ 2 σ 4 x { − 9 ( u 0 ξ + u ∗ ) + ( u 0 ξ + u ∗ ) − 4 1 f 2 ( ξ ) / ( a α x ) f 4 ( ξ ) / ( a α x ) + f 1 ( ξ ) / x f 3 ( ξ ) / x − 4 u 0 ζ / ( a α ) − 7 x u 0 ζ / ( a α ) + 4 x + f 1 ( ξ ) f 3 ( ξ ) f 2 ( ξ ) / ( a α x ) f 4 ( ξ ) / ( a α x ) } =: κ 2 σ 4 ( − 9 ( u 0 ξ + u ∗ ) + j = 9 ∑ 11 R j ( ξ ) ) .
Simple computations yield
∣ R 9 ( ξ ) ∣ ≤ ∣ u 0 ξ + u ∗ ∣ ⋅ ∣ 4 f 4 ( ξ ) + f 2 ( ξ ) ∣ a α x ≤ ∣ u 0 ξ + u ∗ ∣ r ∗ ( ∣ ζ ( 4 t − 15 z + 20 y − 9 x ) ∣ α x + ∣ α ( 16 d t − 45 c z + 40 b y ) ∣ α x ) ≤ ∣ u 0 ξ + u ∗ ∣ r ∗ [ λ ∣ ζ ∣ ξ ( 4 t x + 15 z x + 20 y x + 9 ) + ( 16 d t x + 45 c z x + 40 b y x ) ] , ξ = λ α ≤ ∣ u 0 ξ + u ∗ ∣ r ∗ ( 48 λ ∣ ζ ∣ ξ + 101 y x ) = ∣ u 0 ξ + u ∗ ∣ r ∗ ( 48 λ ∣ ζ ∣ ξ + 101 φ 2 ( ξ ) ) ≤ ∣ u 0 ξ + u ∗ ∣ r ∗ ( 48 λ ∣ ζ ∣ ξ + 101 ⋅ 2 H e − ξ ) ≤ L 5 ξ , \begin{aligned}
|\mathcal{R}_9(\xi)|
&\le|u_0\xi+u^*|\cdot\frac{|4f_4(\xi)+f_2(\xi)|}{a\alpha x}\\
&\le\frac{|u_0\xi+u^*|}{r^*}\left(\frac{|\zeta(4t-15z+20y-9x)|}{\alpha x}+\frac{|\alpha(16dt-45cz+40by)|}{\alpha x}\right)\\
&\le\frac{|u_0\xi+u^*|}{r^*}\left[\frac{\lambda|\zeta|}{\xi}\left(\frac{4t}{x}+\frac{15z}{x}+\frac{20y}{x}+9\right)+\left(\frac{16dt}{x}+\frac{45cz}{x}+\frac{40by}{x}\right)\right],\qquad\xi=\lambda\alpha\\
&\le\frac{|u_0\xi+u^*|}{r^*}\left(48\frac{\lambda|\zeta|}{\xi}+101\frac{y}{x}\right)
=\frac{|u_0\xi+u^*|}{r^*}\left(48\frac{\lambda|\zeta|}{\xi}+101\varphi_2(\xi)\right)\\
&\le\frac{|u_0\xi+u^*|}{r^*}\left(48\frac{\lambda|\zeta|}{\xi}+101\cdot2^H e^{-\xi}\right)\le\frac{L_5}{\xi},
\end{aligned} ∣ R 9 ( ξ ) ∣ ≤ ∣ u 0 ξ + u ∗ ∣ ⋅ a α x ∣4 f 4 ( ξ ) + f 2 ( ξ ) ∣ ≤ r ∗ ∣ u 0 ξ + u ∗ ∣ ( α x ∣ ζ ( 4 t − 15 z + 20 y − 9 x ) ∣ + α x ∣ α ( 16 d t − 45 cz + 40 b y ) ∣ ) ≤ r ∗ ∣ u 0 ξ + u ∗ ∣ [ ξ λ ∣ ζ ∣ ( x 4 t + x 15 z + x 20 y + 9 ) + ( x 16 d t + x 45 cz + x 40 b y ) ] , ξ = λ α ≤ r ∗ ∣ u 0 ξ + u ∗ ∣ ( 48 ξ λ ∣ ζ ∣ + 101 x y ) = r ∗ ∣ u 0 ξ + u ∗ ∣ ( 48 ξ λ ∣ ζ ∣ + 101 φ 2 ( ξ ) ) ≤ r ∗ ∣ u 0 ξ + u ∗ ∣ ( 48 ξ λ ∣ ζ ∣ + 101 ⋅ 2 H e − ξ ) ≤ ξ L 5 ,
where L 5 = sup H , λ { ∣ u 0 ξ + u ∗ ∣ r ∗ ( 48 λ ∣ ζ ∣ + 101 ⋅ 2 H ) } L_5=\sup_{H,\lambda}\left\{\frac{|u_0\xi+u^*|}{r^*}(48\lambda|\zeta|+101\cdot2^H)\right\} L 5 = sup H , λ { r ∗ ∣ u 0 ξ + u ∗ ∣ ( 48 λ ∣ ζ ∣ + 101 ⋅ 2 H ) } . Also we have
∣ R 10 ( ξ ) ∣ = ∣ f 1 ( ξ ) x ( u 0 ζ a α + 4 x ) + f 3 ( ξ ) x ( 4 u 0 ζ a α + 7 x ) ∣ = u 0 λ ∣ ζ ∣ a ξ ( ∣ f 1 ( ξ ) ∣ x + 4 ∣ f 3 ( ξ ) ∣ x ) + 4 ∣ f 1 ( ξ ) ∣ + 7 ∣ f 3 ( ξ ) ∣ , ξ = λ α \begin{aligned}
|\mathcal{R}_{10}(\xi)|
&=\left|\frac{f_1(\xi)}{x}\left(\frac{u_0\zeta}{a\alpha}+4x\right)+\frac{f_3(\xi)}{x}\left(\frac{4u_0\zeta}{a\alpha}+7x\right)\right|\\
&=\frac{u_0\lambda|\zeta|}{a\xi}\left(\frac{|f_1(\xi)|}{x}+4\frac{|f_3(\xi)|}{x}\right)+4|f_1(\xi)|+7|f_3(\xi)|,\qquad\xi=\lambda\alpha
\end{aligned} ∣ R 10 ( ξ ) ∣ = x f 1 ( ξ ) ( a α u 0 ζ + 4 x ) + x f 3 ( ξ ) ( a α 4 u 0 ζ + 7 x ) = a ξ u 0 λ ∣ ζ ∣ ( x ∣ f 1 ( ξ ) ∣ + 4 x ∣ f 3 ( ξ ) ∣ ) + 4∣ f 1 ( ξ ) ∣ + 7∣ f 3 ( ξ ) ∣ , ξ = λ α
with
∣ f 1 ( ξ ) ∣ x ≤ ∣ ξ ∣ ( z x + 4 y x + 7 ) + ( ∣ C ∣ z x + 4 ∣ B ∣ y x + 7 ∣ A ∣ ) ≤ 12 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] , ∣ f 3 ( ξ ) ∣ x ≤ ∣ ξ ∣ ( t x + 4 z x + 6 y x + 4 ) + ( ∣ D ∣ t x + 4 ∣ C ∣ z x + 6 ∣ B ∣ y x + 4 ∣ A ∣ ) ≤ 15 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] , \begin{aligned}
\frac{|f_1(\xi)|}{x} &\le|\xi|\left(\frac{z}{x}+4\frac{y}{x}+7\right)+\left(\frac{|C|z}{x}+4\frac{|B|y}{x}+7|A|\right)\\
&\le12\left[|\xi|+\log(4\xi)+M(6H+1)\right],\\
\frac{|f_3(\xi)|}{x} &\le|\xi|\left(\frac{t}{x}+4\frac{z}{x}+6\frac{y}{x}+4\right)+\left(\frac{|D|t}{x}+4\frac{|C|z}{x}+6\frac{|B|y}{x}+4|A|\right)\\
&\le15\left[|\xi|+\log(4\xi)+M(6H+1)\right],
\end{aligned} x ∣ f 1 ( ξ ) ∣ x ∣ f 3 ( ξ ) ∣ ≤ ∣ ξ ∣ ( x z + 4 x y + 7 ) + ( x ∣ C ∣ z + 4 x ∣ B ∣ y + 7∣ A ∣ ) ≤ 12 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] , ≤ ∣ ξ ∣ ( x t + 4 x z + 6 x y + 4 ) + ( x ∣ D ∣ t + 4 x ∣ C ∣ z + 6 x ∣ B ∣ y + 4∣ A ∣ ) ≤ 15 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ,
∣ f 1 ( ξ ) ∣ = ∣ f 1 ( ξ ) ∣ x ⋅ x ≤ 12 u 0 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( 1 − δ ) − H e − δ ξ ∣ f 3 ( ξ ) ∣ ≤ 15 u 0 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( 1 − δ ) − H e − δ ξ \begin{aligned}
|f_1(\xi)|&=\frac{|f_1(\xi)|}{x}\cdot x\le12u_0\left[|\xi|+\log(4\xi)+M(6H+1)\right](1-\delta)^{-H}e^{-\delta\xi}\\
|f_3(\xi)|&\le15u_0\left[|\xi|+\log(4\xi)+M(6H+1)\right](1-\delta)^{-H}e^{-\delta\xi}
\end{aligned} ∣ f 1 ( ξ ) ∣ ∣ f 3 ( ξ ) ∣ = x ∣ f 1 ( ξ ) ∣ ⋅ x ≤ 12 u 0 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( 1 − δ ) − H e − δ ξ ≤ 15 u 0 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( 1 − δ ) − H e − δ ξ
It follows that
∣ R 10 ( ξ ) ∣ ≤ [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( 72 u 0 λ ∣ ξ ∣ r ∗ ξ + 153 u 0 ( 1 − δ ) − H e − δ ξ ) ≤ ξ − 1 / 3 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( 72 u 0 λ ∣ ξ ∣ r ∗ + 153 u 0 δ ( 1 − δ ) H ) ξ − 2 / 3 ≤ [ ∣ ξ ∣ + 3 4 3 + M ( 6 H + 1 ) ] ( 72 u 0 λ ∣ ξ ∣ r ∗ + 153 u 0 δ ( 1 − δ ) H ) ξ − 2 / 3 ≤ L 6 ξ − 2 / 3 , for all ξ > 1 , \begin{aligned}
|\mathcal{R}_{10}(\xi)|&\le\left[|\xi|+\log(4\xi)+M(6H+1)\right]\left(\frac{72u_0\lambda|\xi|}{r^*\xi}+153u_0(1-\delta)^{-H}e^{-\delta\xi}\right)\\
&\le\xi^{-1/3}\left[|\xi|+\log(4\xi)+M(6H+1)\right]\left(\frac{72u_0\lambda|\xi|}{r^*}+\frac{153u_0}{\delta(1-\delta)^H}\right)\xi^{-2/3}\\
&\le\left[|\xi|+3\sqrt[3]{4}+M(6H+1)\right]\left(\frac{72u_0\lambda|\xi|}{r^*}+\frac{153u_0}{\delta(1-\delta)^H}\right)\xi^{-2/3}\\
&\le L_6\xi^{-2/3},\qquad\text{for all }\xi>1,
\end{aligned} ∣ R 10 ( ξ ) ∣ ≤ [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( r ∗ ξ 72 u 0 λ ∣ ξ ∣ + 153 u 0 ( 1 − δ ) − H e − δ ξ ) ≤ ξ − 1/3 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( r ∗ 72 u 0 λ ∣ ξ ∣ + δ ( 1 − δ ) H 153 u 0 ) ξ − 2/3 ≤ [ ∣ ξ ∣ + 3 3 4 + M ( 6 H + 1 ) ] ( r ∗ 72 u 0 λ ∣ ξ ∣ + δ ( 1 − δ ) H 153 u 0 ) ξ − 2/3 ≤ L 6 ξ − 2/3 , for all ξ > 1 ,
where L 6 = sup H , λ { [ ∣ ξ ∣ + 3 4 3 + M ( 6 H + 1 ) ] ( 72 u 0 λ ∣ ξ ∣ r ∗ + 153 u 0 δ ( 1 − δ ) H ) } L_6=\sup_{H,\lambda}\left\{\left[|\xi|+3\sqrt[3]{4}+M(6H+1)\right]\left(\frac{72u_0\lambda|\xi|}{r^*}+\frac{153u_0}{\delta(1-\delta)^H}\right)\right\} L 6 = sup H , λ { [ ∣ ξ ∣ + 3 3 4 + M ( 6 H + 1 ) ] ( r ∗ 72 u 0 λ ∣ ξ ∣ + δ ( 1 − δ ) H 153 u 0 ) } . Finally, we evaluate the last determinant as
∣ R 11 ( ξ ) ∣ = ∣ f 1 ( ξ ) f 4 ( ξ ) α α x f 1 ( ξ ) f 2 ( ξ ) α α x ∣ ≤ ∣ f 1 ( ξ ) ∣ r ∗ [ ∣ ξ ∣ α ( t x + 4 z x + 6 y x + 4 ) + 4 ( d t x + 3 c z x + 3 b y x ) ] + ∣ f 3 ( ξ ) ∣ r ∗ [ ∣ ξ ∣ α ( z x + 4 y x + 7 ) + ( 3 c z x + 8 b y x ) ] ≤ ∣ f 1 ( ξ ) ∣ r ∗ ( 15 λ ∣ ξ ∣ ξ + 28 φ 2 ( ξ ) ) + ∣ f 3 ( ξ ) ∣ r ∗ ( 12 λ ∣ ξ ∣ ξ + 11 φ 2 ( ξ ) ) , φ 2 ( ξ ) = y / x ≤ ∣ f 1 ( ξ ) ∣ r ∗ ( 15 λ ∣ ξ ∣ ξ + 28 ⋅ 2 H e − ξ ) + ∣ f 3 ( ξ ) ∣ r ∗ ( 12 λ ∣ ξ ∣ ξ + 11 ⋅ 2 H e − ξ ) \begin{aligned}
|\mathcal{R}_{11}(\xi)|&=\left|\begin{array}{cc}
f_1(\xi)&\dfrac{f_4(\xi)}{\alpha\alpha x}\\
f_1(\xi)&\dfrac{f_2(\xi)}{\alpha\alpha x}
\end{array}\right|\\
&\le\frac{|f_1(\xi)|}{r^*}\left[\frac{|\xi|}{\alpha}\left(\frac{t}{x}+4\frac{z}{x}+6\frac{y}{x}+4\right)+4\left(\frac{dt}{x}+3\frac{cz}{x}+3\frac{by}{x}\right)\right]\\
&\quad+\frac{|f_3(\xi)|}{r^*}\left[\frac{|\xi|}{\alpha}\left(\frac{z}{x}+4\frac{y}{x}+7\right)+\left(\frac{3cz}{x}+8\frac{by}{x}\right)\right]\\
&\le\frac{|f_1(\xi)|}{r^*}\left(15\frac{\lambda|\xi|}{\xi}+28\varphi_2(\xi)\right)+\frac{|f_3(\xi)|}{r^*}\left(12\frac{\lambda|\xi|}{\xi}+11\varphi_2(\xi)\right),\qquad\varphi_2(\xi)=y/x\\
&\le\frac{|f_1(\xi)|}{r^*}\left(15\frac{\lambda|\xi|}{\xi}+28\cdot2^He^{-\xi}\right)+\frac{|f_3(\xi)|}{r^*}\left(12\frac{\lambda|\xi|}{\xi}+11\cdot2^He^{-\xi}\right)
\end{aligned} ∣ R 11 ( ξ ) ∣ = f 1 ( ξ ) f 1 ( ξ ) α α x f 4 ( ξ ) α α x f 2 ( ξ ) ≤ r ∗ ∣ f 1 ( ξ ) ∣ [ α ∣ ξ ∣ ( x t + 4 x z + 6 x y + 4 ) + 4 ( x d t + 3 x cz + 3 x b y ) ] + r ∗ ∣ f 3 ( ξ ) ∣ [ α ∣ ξ ∣ ( x z + 4 x y + 7 ) + ( x 3 cz + 8 x b y ) ] ≤ r ∗ ∣ f 1 ( ξ ) ∣ ( 15 ξ λ ∣ ξ ∣ + 28 φ 2 ( ξ ) ) + r ∗ ∣ f 3 ( ξ ) ∣ ( 12 ξ λ ∣ ξ ∣ + 11 φ 2 ( ξ ) ) , φ 2 ( ξ ) = y / x ≤ r ∗ ∣ f 1 ( ξ ) ∣ ( 15 ξ λ ∣ ξ ∣ + 28 ⋅ 2 H e − ξ ) + r ∗ ∣ f 3 ( ξ ) ∣ ( 12 ξ λ ∣ ξ ∣ + 11 ⋅ 2 H e − ξ )
By using the fact that log ( w ) ≤ 3 w 1 / 3 \log(w)\le3w^{1/3} log ( w ) ≤ 3 w 1/3 , for all w > 0 w>0 w > 0 we obtain
∣ f 1 ( ξ ) ∣ ∨ ∣ f 3 ( ξ ) ∣ ≤ 15 u 0 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( 1 − δ ) − H e − δ ξ ≤ 15 u 0 [ ∣ ξ ∣ + 3 4 3 + M ( 6 H + 1 ) ] ( 1 − δ ) − H ξ 1 / 3 e − δ ξ ≤ L 7 ξ 1 / 3 e − δ ξ , for all ξ > 1 , \begin{aligned}
|f_1(\xi)|\vee|f_3(\xi)|&\le15u_0\left[|\xi|+\log(4\xi)+M(6H+1)\right](1-\delta)^{-H}e^{-\delta\xi}\\
&\le15u_0\left[|\xi|+3\sqrt[3]{4}+M(6H+1)\right](1-\delta)^{-H}\xi^{1/3}e^{-\delta\xi}\\
&\le L_7\xi^{1/3}e^{-\delta\xi},\qquad\text{for all }\xi>1,
\end{aligned} ∣ f 1 ( ξ ) ∣ ∨ ∣ f 3 ( ξ ) ∣ ≤ 15 u 0 [ ∣ ξ ∣ + log ( 4 ξ ) + M ( 6 H + 1 ) ] ( 1 − δ ) − H e − δ ξ ≤ 15 u 0 [ ∣ ξ ∣ + 3 3 4 + M ( 6 H + 1 ) ] ( 1 − δ ) − H ξ 1/3 e − δ ξ ≤ L 7 ξ 1/3 e − δ ξ , for all ξ > 1 ,
where
L 7 : = sup H , λ { 15 u 0 [ ∣ ξ ∣ + 3 4 3 + M ( 6 H + 1 ) ] ( 1 − δ ) − H } . L_7:=\sup_{H,\lambda}\left\{15u_0\left[|\xi|+3\sqrt[3]{4}+M(6H+1)\right](1-\delta)^{-H}\right\}. L 7 := H , λ sup { 15 u 0 [ ∣ ξ ∣ + 3 3 4 + M ( 6 H + 1 ) ] ( 1 − δ ) − H } .
Hence,
∣ R 11 ( ξ ) ∣ ≤ L 7 r ∗ ( 27 λ ∣ ζ ∣ ξ + 39 ⋅ 2 H e − ξ ) ξ 1 / 3 e − δ ξ ≤ L 8 ξ − 2 / 3 e − δ ξ , |\mathcal{R}_{11}(\xi)| \le\frac{L_7}{r^*}\left(27\frac{\lambda|\zeta|}{\xi}+39\cdot2^H e^{-\xi}\right)\xi^{1/3}e^{-\delta\xi}
\le L_8\xi^{-2/3}e^{-\delta\xi}, ∣ R 11 ( ξ ) ∣ ≤ r ∗ L 7 ( 27 ξ λ ∣ ζ ∣ + 39 ⋅ 2 H e − ξ ) ξ 1/3 e − δ ξ ≤ L 8 ξ − 2/3 e − δ ξ ,
where
L 8 : = sup H , λ { L 7 r ∗ ( 27 λ ∣ ζ ∣ + 39 ⋅ 2 H ) } . L_8:=\sup_{H,\lambda}\left\{\frac{L_7}{r^*}\left(27\lambda|\zeta|+39\cdot2^H\right)\right\}. L 8 := H , λ sup { r ∗ L 7 ( 27 λ ∣ ζ ∣ + 39 ⋅ 2 H ) } .
According to the previous estimates of R j ( ξ ) \mathcal{R}_j(\xi) R j ( ξ ) , j = 9 , 10 , 11 j=9,10,11 j = 9 , 10 , 11 we can find some ξ 5 > 0 \xi_5>0 ξ 5 > 0 (depending only on the parameter space), so that for all ξ > ξ 5 \xi>\xi_5 ξ > ξ 5 we have
{ ⟨ 1 ⟩ Σ < 0 , if u 0 ξ + u ∗ > 0 , ⟨ 1 ⟩ Σ > 0 , if u 0 ξ + u ∗ < 0. \begin{cases}
\langle1\rangle_{\Sigma}<0, & \text{if }u_0\xi+u^*>0,\\
\langle1\rangle_{\Sigma}>0, & \text{if }u_0\xi+u^*<0.
\end{cases} { ⟨ 1 ⟩ Σ < 0 , ⟨ 1 ⟩ Σ > 0 , if u 0 ξ + u ∗ > 0 , if u 0 ξ + u ∗ < 0.
Or equivalently,
{ ⟨ 1 ⟩ Σ < 0 , if λ < e ( ψ ( H + 1 / 2 ) + ψ ( H ) ) / 2 , ⟨ 1 ⟩ Σ > 0 , if λ > e ( ψ ( H + 1 / 2 ) + ψ ( H ) ) / 2 . \begin{cases}
\langle1\rangle_{\Sigma}<0, & \text{if }\lambda<e^{(\psi(H+1/2)+\psi(H))/2},\\
\langle1\rangle_{\Sigma}>0, & \text{if }\lambda>e^{(\psi(H+1/2)+\psi(H))/2}.
\end{cases} { ⟨ 1 ⟩ Σ < 0 , ⟨ 1 ⟩ Σ > 0 , if λ < e ( ψ ( H + 1/2 ) + ψ ( H )) /2 , if λ > e ( ψ ( H + 1/2 ) + ψ ( H )) /2 .
iii-Conclusion. All the proper principal minors are evaluated according to the values of ξ = λ α \xi=\lambda\alpha ξ = λ α in the intervals ( ξ j , ∞ ) (\xi_j,\infty) ( ξ j , ∞ ) , $j=1,2,3,4,5. It is then clear that both (62) and (63) hold for all α ∈ ( α ∗ , ∞ ) \alpha\in(\alpha^*,\infty) α ∈ ( α ∗ , ∞ ) , where α ∗ : = α 0 ∨ sup 1 ≤ j ≤ 5 { ξ j / λ } \alpha^*:=\alpha_0\vee\sup_{1\le j\le5}\{\xi_j/\lambda\} α ∗ := α 0 ∨ sup 1 ≤ j ≤ 5 { ξ j / λ } and α 0 \alpha_0 α 0 is due to Lemma 6. The proof is then complete.
Injectivity of the mapping θ ↦ B ϱ θ ( α ) \theta\mapsto\mathbb{B}_{\varrho_\theta}(\alpha) θ ↦ B ϱ θ ( α )
Let the assumptions of Theorem 2 be fulfilled, and let θ , θ ′ ∈ Θ \theta,\theta'\in\Theta θ , θ ′ ∈ Θ such that B ϱ θ ( α ) = B ϱ θ ′ ( α ) \mathbb{B}_{\varrho_\theta}(\alpha)=\mathbb{B}_{\varrho_{\theta'}}(\alpha) B ϱ θ ( α ) = B ϱ θ ′ ( α ) , where ϱ θ ( α ) = ( γ θ ( 0 ) , γ θ ( α ) , γ θ ( 2 α ) , γ θ ( 3 α ) ) ′ \varrho_\theta(\alpha)=(\gamma_\theta(0),\gamma_\theta(\alpha),\gamma_\theta(2\alpha),\gamma_\theta(3\alpha))' ϱ θ ( α ) = ( γ θ ( 0 ) , γ θ ( α ) , γ θ ( 2 α ) , γ θ ( 3 α ) ) ′ . Set
γ θ ( j α ) − γ θ ′ ( j α ) = η j , for all j = 0 , 1 , 2 , 3. (65) \gamma_\theta(j\alpha)-\gamma_{\theta'}(j\alpha)=\eta_j,\qquad\text{for all }j=0,1,2,3. \tag*{(65)} γ θ ( j α ) − γ θ ′ ( j α ) = η j , for all j = 0 , 1 , 2 , 3. (65)
Suppose that η j \eta_j η j are not all equal, that is, there is some j ∗ ≠ 0 j^*\ne0 j ∗ = 0 such that η 0 ≠ η j ∗ \eta_0\ne\eta_{j^*} η 0 = η j ∗ . Without loss of generality, say η 0 ≠ η 1 \eta_0\ne\eta_1 η 0 = η 1 . The other cases (j ∗ = 2 j^*=2 j ∗ = 2 or 3 3 3 ) can be treated in a similar way to produce the same result. Let u u u and v v v be two reals such that
{ u ≠ 0 , v = − u η 0 / η 1 , if η 1 ≠ 0 , v ≠ 0 , u = 0 , if η 1 = 0. \begin{cases}
u\ne0,\ v=-u\eta_0/\eta_1, & \text{if }\eta_1\ne0,\\
v\ne0,\ u=0, & \text{if }\eta_1=0.
\end{cases} { u = 0 , v = − u η 0 / η 1 , v = 0 , u = 0 , if η 1 = 0 , if η 1 = 0.
It follows that u ( γ θ ( 0 ) − γ θ ′ ( 0 ) ) + v ( γ θ ( α ) − γ θ ′ ( α ) ) = u η 0 + v η 1 = 0 u(\gamma_\theta(0)-\gamma_{\theta'}(0))+v(\gamma_\theta(\alpha)-\gamma_{\theta'}(\alpha))=u\eta_0+v\eta_1=0 u ( γ θ ( 0 ) − γ θ ′ ( 0 )) + v ( γ θ ( α ) − γ θ ′ ( α )) = u η 0 + v η 1 = 0 , or equivalently,
u γ θ ( 0 ) + v γ θ ( α ) = u γ θ ′ ( 0 ) + v γ θ ′ ( α ) . u\gamma_\theta(0)+v\gamma_\theta(\alpha)=u\gamma_{\theta'}(0)+v\gamma_{\theta'}(\alpha). u γ θ ( 0 ) + v γ θ ( α ) = u γ θ ′ ( 0 ) + v γ θ ′ ( α ) .
Combining this with the equality B ϱ θ ( α ) = B ϱ θ ′ ( α ) \mathbb{B}_{\varrho_\theta}(\alpha)=\mathbb{B}_{\varrho_{\theta'}}(\alpha) B ϱ θ ( α ) = B ϱ θ ′ ( α ) yields
B ~ ϱ θ ( α ) = B ~ ϱ θ ′ ( α ) , B ~ : = [ B w ] , (66) \widetilde{\mathbb{B}}_{\varrho_\theta}(\alpha)=\widetilde{\mathbb{B}}_{\varrho_{\theta'}}(\alpha),\qquad\widetilde{\mathbb{B}}:=\begin{bmatrix}\mathbb{B}\\w\end{bmatrix}, \tag*{(66)} B ϱ θ ( α ) = B ϱ θ ′ ( α ) , B := [ B w ] , (66)
where w w w is the row vector defined as w = ( u , v , 0 , 0 ) w=(u,v,0,0) w = ( u , v , 0 , 0 ) . Clearly, the determinant of B ~ \widetilde{\mathbb{B}} B has the form ∣ B ~ ∣ = 32 ( u + v ) ≠ 0 |\widetilde{\mathbb{B}}|=32(u+v)\ne0 ∣ B ∣ = 32 ( u + v ) = 0 . Thus B ~ \widetilde{\mathbb{B}} B is invertible and (66) becomes ϱ θ ( α ) = ϱ θ ′ ( α ) \varrho_\theta(\alpha)=\varrho_{\theta'}(\alpha) ϱ θ ( α ) = ϱ θ ′ ( α ) . Finally, by injectivity of the mapping θ ↦ ϱ θ ( α ) \theta\mapsto\varrho_{\theta}(\alpha) θ ↦ ϱ θ ( α ) (inherited from the injectivity of θ ↦ ϱ θ , 3 ( α ) \theta\mapsto\varrho_{\theta,3}(\alpha) θ ↦ ϱ θ , 3 ( α ) ) we get θ = θ ′ \theta=\theta' θ = θ ′ . We conclude that θ ↦ B ϱ θ ( α ) \theta\mapsto\mathbb{B}\varrho_{\theta}(\alpha) θ ↦ B ϱ θ ( α ) is injective on the set of parameters whose covariance vectors cannot differ by a common constant, that is, η j \eta_j η j are not all equal. The case where η j \eta_j η j are all equal is not possible for large values of α \alpha α , unless all η j \eta_j η j are identically zero which in turn implies ϱ θ , 3 ( α ) = ϱ θ ′ , 3 ( α ) \varrho_{\theta,3}(\alpha)=\varrho_{\theta',3}(\alpha) ϱ θ , 3 ( α ) = ϱ θ ′ , 3 ( α ) . Hence, θ = θ ′ \theta=\theta' θ = θ ′ . This is due to the asymptotic behaviour of g H ( ⋅ ) g_H(\cdot) g H ( ⋅ ) at infinity. As previously done for ∇ ϱ θ , 3 ( α ) \nabla\varrho_{\theta,3}(\alpha) ∇ ϱ θ , 3 ( α ) , a detailed investigation of the principal minors of ∇ B ϱ θ ( α ) = B ∇ ϱ θ ( α ) \nabla\mathbb{B}\varrho_{\theta}(\alpha)=\mathbb{B}\nabla\varrho_{\theta}(\alpha) ∇ B ϱ θ ( α ) = B ∇ ϱ θ ( α ) is possible, and an appropriate use of the existing results on univalence [9 , 23 , 21 , 14 ] will lead to global injectivity on the parameter space Θ \Theta Θ . The next lemma is very crucial for the statement of Theorem 3, since it justifies the existence of the matrix C ( θ ) C(\theta) C ( θ ) .
Lemma 8. There is some α ~ > 0 \widetilde{\alpha}>0 α > 0 (depending only on the parameter space Θ \Theta Θ ), such that ∣ B ∇ ϱ θ ( α ) ∣ < 0 \lvert\mathbb{B}\nabla\varrho_{\theta}(\alpha)\rvert<0 ∣ B ∇ ϱ θ ( α )∣ < 0 , for all α > α ~ \alpha>\widetilde{\alpha} α > α . Moreover, the mapping θ ↦ B ϱ θ ( α ) \theta\mapsto\mathbb{B}\varrho_{\theta}(\alpha) θ ↦ B ϱ θ ( α ) is locally injective.
Proof. We are concerned with the Jacobian matrix of ϱ θ ( α ) \varrho_{\theta}(\alpha) ϱ θ ( α ) with respect to θ \theta θ given as
∇ ϱ θ ( α ) = ( Σ 1 , 1 Σ 1 , 2 Σ 1 , 3 Σ 2 , 1 Σ 2 , 2 Σ 2 , 3 Σ 3 , 1 Σ 3 , 2 Σ 3 , 3 Σ 4 , 1 Σ 4 , 2 Σ 4 , 3 ) \nabla\varrho_{\theta}(\alpha)=
\begin{pmatrix}
\Sigma_{1,1} & \Sigma_{1,2} & \Sigma_{1,3}\\
\Sigma_{2,1} & \Sigma_{2,2} & \Sigma_{2,3}\\
\Sigma_{3,1} & \Sigma_{3,2} & \Sigma_{3,3}\\
\Sigma_{4,1} & \Sigma_{4,2} & \Sigma_{4,3}
\end{pmatrix} ∇ ϱ θ ( α ) = Σ 1 , 1 Σ 2 , 1 Σ 3 , 1 Σ 4 , 1 Σ 1 , 2 Σ 2 , 2 Σ 3 , 2 Σ 4 , 2 Σ 1 , 3 Σ 2 , 3 Σ 3 , 3 Σ 4 , 3
where its components Σ i , j \Sigma_{i,j} Σ i , j are given in (15)-(17), (18) and
Σ 4 , 1 : = ∂ σ 2 γ ( 3 α ) = κ ( u − 4 t + 6 z − 4 y + x ) Σ 4 , 2 : = ∂ H γ ( 3 α ) = σ 2 κ [ ξ ( u − 4 t + 6 z − 4 y + x ) + ( F u − 4 D t + 6 C z − 4 B y + A x ) ] , Σ 4 , 3 : = ∂ λ γ ( 3 α ) = σ 2 κ [ ξ ( u − 4 t + 6 z − 4 y + x ) − α ( 5 f u − 16 d t + 18 c z − 8 b y + a x ) ] \begin{aligned}
\Sigma_{4,1}&:=\partial_{\sigma^2}\gamma(3\alpha)=\kappa\left(u-4t+6z-4y+x\right)\\
\Sigma_{4,2}&:=\partial_H\gamma(3\alpha)=\sigma^2\kappa\left[\xi\left(u-4t+6z-4y+x\right)+\left(Fu-4Dt+6Cz-4By+Ax\right)\right],\\
\Sigma_{4,3}&:=\partial_{\lambda}\gamma(3\alpha)=\sigma^2\kappa\left[\xi\left(u-4t+6z-4y+x\right)-\alpha\left(5fu-16dt+18cz-8by+ax\right)\right]
\end{aligned} Σ 4 , 1 Σ 4 , 2 Σ 4 , 3 := ∂ σ 2 γ ( 3 α ) = κ ( u − 4 t + 6 z − 4 y + x ) := ∂ H γ ( 3 α ) = σ 2 κ [ ξ ( u − 4 t + 6 z − 4 y + x ) + ( F u − 4 D t + 6 C z − 4 B y + A x ) ] , := ∂ λ γ ( 3 α ) = σ 2 κ [ ξ ( u − 4 t + 6 z − 4 y + x ) − α ( 5 f u − 16 d t + 18 cz − 8 b y + a x ) ]
with u = g H ( 5 λ α ) u=g_H(5\lambda\alpha) u = g H ( 5 λ α ) , F = R H ( 5 λ α ) F=R_H(5\lambda\alpha) F = R H ( 5 λ α ) and f = r H ( 5 λ α ) f=r_H(5\lambda\alpha) f = r H ( 5 λ α ) . Set P = B ∇ ϱ θ ( α ) \mathbf{P}=\mathbb{B}\nabla\varrho_{\theta}(\alpha) P = B ∇ ϱ θ ( α ) . It is 3 × 3 3\times3 3 × 3 -matrix with entries P i , j \mathbf{P}_{i,j} P i , j given as
P 1 , 1 = κ ( − 2 t + 8 z − 8 y − 8 x + 10 u 0 ) P 2 , 1 = κ ( 2 t − 16 z + 56 y − 112 x + 70 u 0 ) P 3 , 1 = κ ( 2 u − 12 t + 26 z − 16 y − 28 x + 28 u 0 ) P 1 , 2 = σ 2 ξ P 1 , 1 + 2 κ σ 2 ( − D t + 4 C z − 4 B y − 4 A x + 5 u ∗ ) P 2 , 2 = σ 2 ξ P 2 , 1 + 2 κ σ 2 ( D t − 8 C z + 28 B y − 56 A x + 35 u ∗ ) P 3 , 2 = σ 2 ξ P 3 , 1 + 2 κ σ 2 ( F u − 6 D t + 13 C z − 8 B y − 14 A x + 14 u ∗ ) P 1 , 3 = σ 2 ξ P 1 , 1 + 8 κ σ 2 α ( d t − 3 c z + 2 b y + a x ) P 2 , 3 = σ 2 ξ P 2 , 1 − 8 κ σ 2 α ( d t − 6 c z + 14 b y − 14 a x ) P 3 , 3 = σ 2 ξ P 3 , 1 − 2 κ σ 2 α ( 5 f u − 24 d t + 39 c z − 16 b y − 14 a x ) \begin{aligned}
\mathbf{P}_{1,1}&=\kappa(-2t+8z-8y-8x+10u_0)\\
\mathbf{P}_{2,1}&=\kappa(2t-16z+56y-112x+70u_0)\\
\mathbf{P}_{3,1}&=\kappa(2u-12t+26z-16y-28x+28u_0)\\
\mathbf{P}_{1,2}&=\sigma^2\xi\mathbf{P}_{1,1}+2\kappa\sigma^2(-Dt+4Cz-4By-4Ax+5u^*)\\
\mathbf{P}_{2,2}&=\sigma^2\xi\mathbf{P}_{2,1}+2\kappa\sigma^2(Dt-8Cz+28By-56Ax+35u^*)\\
\mathbf{P}_{3,2}&=\sigma^2\xi\mathbf{P}_{3,1}+2\kappa\sigma^2(Fu-6Dt+13Cz-8By-14Ax+14u^*)\\
\mathbf{P}_{1,3}&=\sigma^2\xi\mathbf{P}_{1,1}+8\kappa\sigma^2\alpha(dt-3cz+2by+ax)\\
\mathbf{P}_{2,3}&=\sigma^2\xi\mathbf{P}_{2,1}-8\kappa\sigma^2\alpha(dt-6cz+14by-14ax)\\
\mathbf{P}_{3,3}&=\sigma^2\xi\mathbf{P}_{3,1}-2\kappa\sigma^2\alpha(5fu-24dt+39cz-16by-14ax)
\end{aligned} P 1 , 1 P 2 , 1 P 3 , 1 P 1 , 2 P 2 , 2 P 3 , 2 P 1 , 3 P 2 , 3 P 3 , 3 = κ ( − 2 t + 8 z − 8 y − 8 x + 10 u 0 ) = κ ( 2 t − 16 z + 56 y − 112 x + 70 u 0 ) = κ ( 2 u − 12 t + 26 z − 16 y − 28 x + 28 u 0 ) = σ 2 ξ P 1 , 1 + 2 κ σ 2 ( − D t + 4 C z − 4 B y − 4 A x + 5 u ∗ ) = σ 2 ξ P 2 , 1 + 2 κ σ 2 ( D t − 8 C z + 28 B y − 56 A x + 35 u ∗ ) = σ 2 ξ P 3 , 1 + 2 κ σ 2 ( F u − 6 D t + 13 C z − 8 B y − 14 A x + 14 u ∗ ) = σ 2 ξ P 1 , 1 + 8 κ σ 2 α ( d t − 3 cz + 2 b y + a x ) = σ 2 ξ P 2 , 1 − 8 κ σ 2 α ( d t − 6 cz + 14 b y − 14 a x ) = σ 2 ξ P 3 , 1 − 2 κ σ 2 α ( 5 f u − 24 d t + 39 cz − 16 b y − 14 a x )
Taking into account the following asymptotics: g H ( ξ ) ∼ π 2 ξ H − 1 / 2 e − ξ g_H(\xi)\sim\sqrt{\frac{\pi}{2}}\xi^{H-1/2}e^{-\xi} g H ( ξ ) ∼ 2 π ξ H − 1/2 e − ξ , R H ( ξ ) ∼ log ( ξ ) R_H(\xi)\sim\log(\xi) R H ( ξ ) ∼ log ( ξ ) and r H ( ξ ) = 1 + o ( 1 ) r_H(\xi)=1+o(1) r H ( ξ ) = 1 + o ( 1 ) , as ξ → ∞ \xi\to\infty ξ → ∞ (see Lemma 3) we can write
P 1 , 1 = κ ( − 8 y − 8 x + 10 u 0 ) + o ( y ) P 2 , 1 = κ ( 56 y − 112 x + 70 u 0 ) + o ( y ) P 3 , 1 = κ ( − 16 y − 28 x + 28 u 0 ) + o ( y ) P 1 , 2 = 10 u 0 κ σ 2 ξ + 2 κ σ 2 ( − 4 A x + 5 u ∗ ) + o ( A x ) P 2 , 2 = 70 u 0 κ σ 2 ξ + 2 κ σ 2 ( − 56 A x + 35 u ∗ ) + o ( A x ) P 3 , 2 = 28 u 0 κ σ 2 ξ + 2 κ σ 2 ( − 14 A x + 14 u ∗ ) + o ( A x ) P 1 , 3 = κ σ 2 ξ ( − 8 y − 8 x + 10 u 0 ) + 8 κ σ 2 α ( 26 b y + a x ) + o ( y ) P 2 , 3 = κ σ 2 ξ ( 56 y − 112 x + 70 u 0 ) − 8 κ σ 2 α ( 14 b y − 14 a x ) + o ( y ) P 3 , 3 = κ σ 2 ξ ( − 16 y − 28 x + 28 u 0 ) − 2 κ σ 2 α ( − 16 b y − 14 a x ) + o ( y ) \begin{aligned}
\mathrm{P}_{1,1} &= \kappa(-8y-8x+10u_0)+o(y)\\
\mathrm{P}_{2,1} &= \kappa(56y-112x+70u_0)+o(y)\\
\mathrm{P}_{3,1} &= \kappa(-16y-28x+28u_0)+o(y)\\
\mathrm{P}_{1,2} &= 10u_0\kappa\sigma^2\xi+2\kappa\sigma^2(-4Ax+5u^*)+o(Ax)\\
\mathrm{P}_{2,2} &= 70u_0\kappa\sigma^2\xi+2\kappa\sigma^2(-56Ax+35u^*)+o(Ax)\\
\mathrm{P}_{3,2} &= 28u_0\kappa\sigma^2\xi+2\kappa\sigma^2(-14Ax+14u^*)+o(Ax)\\
\mathrm{P}_{1,3} &= \kappa\sigma^2\xi(-8y-8x+10u_0)+8\kappa\sigma^2\alpha(26by+ax)+o(y)\\
\mathrm{P}_{2,3} &= \kappa\sigma^2\xi(56y-112x+70u_0)-8\kappa\sigma^2\alpha(14by-14ax)+o(y)\\
\mathrm{P}_{3,3} &= \kappa\sigma^2\xi(-16y-28x+28u_0)-2\kappa\sigma^2\alpha(-16by-14ax)+o(y)
\end{aligned} P 1 , 1 P 2 , 1 P 3 , 1 P 1 , 2 P 2 , 2 P 3 , 2 P 1 , 3 P 2 , 3 P 3 , 3 = κ ( − 8 y − 8 x + 10 u 0 ) + o ( y ) = κ ( 56 y − 112 x + 70 u 0 ) + o ( y ) = κ ( − 16 y − 28 x + 28 u 0 ) + o ( y ) = 10 u 0 κ σ 2 ξ + 2 κ σ 2 ( − 4 A x + 5 u ∗ ) + o ( A x ) = 70 u 0 κ σ 2 ξ + 2 κ σ 2 ( − 56 A x + 35 u ∗ ) + o ( A x ) = 28 u 0 κ σ 2 ξ + 2 κ σ 2 ( − 14 A x + 14 u ∗ ) + o ( A x ) = κ σ 2 ξ ( − 8 y − 8 x + 10 u 0 ) + 8 κ σ 2 α ( 26 b y + a x ) + o ( y ) = κ σ 2 ξ ( 56 y − 112 x + 70 u 0 ) − 8 κ σ 2 α ( 14 b y − 14 a x ) + o ( y ) = κ σ 2 ξ ( − 16 y − 28 x + 28 u 0 ) − 2 κ σ 2 α ( − 16 b y − 14 a x ) + o ( y )
With these arguments we evaluate the determinant of P \mathbf{P} P and find that the first leading terms, such as, those of order O ( α l A k x j ) O(\alpha^l A^k x^j) O ( α l A k x j ) , k , l = 0 , 1 k,l=0,1 k , l = 0 , 1 , j = 0 , 1 , 2 , 3 j=0,1,2,3 j = 0 , 1 , 2 , 3 are vanishing. So, we arrive at
∣ P ∣ ∼ κ 3 σ 4 ∣ − 8 y − 8 x + 10 u 0 10 ( u 0 ξ + u ∗ ) − 8 A x ( 16 b α − 8 ξ ) y + ( 8 a α − 8 ξ ) x + 10 u 0 ξ 56 y − 112 x + 70 u 0 70 ( u 0 ξ + u ∗ ) − 112 A x ( − 112 b α + 56 ξ ) y + ( 112 a α − 112 ξ ) x + 70 u 0 ξ − 16 y − 28 x + 28 u 0 28 ( u 0 ξ + u ∗ ) − 28 A x ( 32 b α − 16 ξ ) y + ( 28 a α − 28 ξ ) x + 28 u 0 ξ ∣ ∼ − 5376 u 0 κ 3 σ 4 A α x y , ( A = R H ( λ α ) ∼ log ( λ α ) , x = g H ( λ α ) , y = g H ( 2 λ α ) ) , \begin{aligned}
\lvert\mathbf{P}\rvert\sim\kappa^3\sigma^4
\begin{vmatrix}
-8y-8x+10u_0 & 10(u_0\xi+u^*)-8Ax & (16b\alpha-8\xi)y+(8a\alpha-8\xi)x+10u_0\xi\\
56y-112x+70u_0 & 70(u_0\xi+u^*)-112Ax & (-112b\alpha+56\xi)y+(112a\alpha-112\xi)x+70u_0\xi\\
-16y-28x+28u_0 & 28(u_0\xi+u^*)-28Ax & (32b\alpha-16\xi)y+(28a\alpha-28\xi)x+28u_0\xi
\end{vmatrix} \\
&\sim-5376u_0\kappa^3\sigma^4 A\alpha xy,\qquad\left(A=R_H(\lambda\alpha)\sim\log(\lambda\alpha),\ x=g_H(\lambda\alpha),\ y=g_H(2\lambda\alpha)\right),
\end{aligned} ∣ P ∣ ∼ κ 3 σ 4 − 8 y − 8 x + 10 u 0 56 y − 112 x + 70 u 0 − 16 y − 28 x + 28 u 0 10 ( u 0 ξ + u ∗ ) − 8 A x 70 ( u 0 ξ + u ∗ ) − 112 A x 28 ( u 0 ξ + u ∗ ) − 28 A x ( 16 b α − 8 ξ ) y + ( 8 a α − 8 ξ ) x + 10 u 0 ξ ( − 112 b α + 56 ξ ) y + ( 112 a α − 112 ξ ) x + 70 u 0 ξ ( 32 b α − 16 ξ ) y + ( 28 a α − 28 ξ ) x + 28 u 0 ξ ∼ − 5376 u 0 κ 3 σ 4 A α x y , ( A = R H ( λ α ) ∼ log ( λ α ) , x = g H ( λ α ) , y = g H ( 2 λ α ) ) ,
as α → ∞ \alpha\to\infty α → ∞ . Note that u 0 , κ , σ 2 > 0 u_0,\kappa,\sigma^2>0 u 0 , κ , σ 2 > 0 . This means that we can find some α ~ > 0 \widetilde{\alpha}>0 α > 0 (depending only on the parameter space Θ \Theta Θ ), so that ∥ B ∇ θ ( α ) ∥ < 0 \lVert\mathbb{B}\nabla_{\theta}(\alpha)\rVert<0 ∥ B ∇ θ ( α )∥ < 0 , for all α > α ~ \alpha>\widetilde{\alpha} α > α (to specify α ~ \widetilde{\alpha} α , just apply the same procedures as in Subsection B.1). We conclude that B ∇ θ ( α ) \mathbb{B}\nabla_{\theta}(\alpha) B ∇ θ ( α ) is invertible, and by local univalent theorem the mapping θ ↦ B θ ( α ) \theta\mapsto\mathbb{B}_{\theta}(\alpha) θ ↦ B θ ( α ) is locally injective. Therefore, the proof is complete. □ \square □
Disclosure statement
Both authors have no conflicts of interest.
Acknowledgments
The authors would like to express their gratitude to the anonymous reviewers for their valuable comments and constructive suggestions, which have significantly improved the quality of the paper.